Daily Program

Driven by the increasing demand for high dynamic range (HDR) content, current display terminals urgently need real-time, lightweight, and effective inverse tone mapping (ITM) models. However, existing models can hardly be deployed on display terminals owing to high computational complexity and large footprints. In this paper, we propose a lightweight LUT-based model based on the $IC_TC_P$ color space to address ITM tasks for display terminals. Leveraging the perceptual uniformity and decorrelation of luma and chroma in the $IC_TC_P$ space, two independent, uniformly sampled LUTs are developed. The two LUTs are a 1D Luma LUT and a 2D Chroma LUT, which adapt to individual luma and chroma components and enable optimally suited mapping based on input frames. This model greatly reduces the number of LUT entries compared to general-purpose 3D LUTs. Our model exhibits superior efficacy, lightweight design, and real-time performance through extensive experiments, satisfying the demands of display terminal.

Skeleton-based human action recognition (HAR) has received significant attention in the industry and the research community. In particular, recent HAR methods have leveraged neural networks to capture spatial configurations and temporal dynamics encoded in skeleton data. Despite significant progress, the scarcity of real-world datasets poses a challenge in training HAR models. To address this challenge, we propose a new method of generating synthetic data for HAR, drawing inspiration from recent developments in text-conditioned generative models. Our framework, called text-to-motion for skeleton-based action recognition (T2M4AR), integrates text-to-motion generation into the training loop of the HAR model so that HAR performance can be improved without modifying network architectures or inference time. We verify the effectiveness of T2M4AR by applying it to three state-of-the-art HAR models. Experimental results demonstrate that T2M4AR can increase the performance of HAR for all baseline HAR models on five benchmark datasets.

Early-exit neural networks reduce inference cost by enabling confident predictions at intermediate layers. However, joint training often leads to gradient interference, with deeper classifiers dominating optimization. We propose Confidence-Gated Training (CGT), a paradigm that conditionally propagates gradients from deeper exits only when preceding exits fail. This encourages shallow classifiers to act as primary decision points while reserving deeper layers for harder inputs. By aligning training with the inference-time policy, CGT mitigates overthinking, improves early-exit accuracy, and preserves efficiency. Experiments on the Indian Pines, Fashion-MNIST and CIFAR-100 benchmarks show that CGT lowers average inference cost while improving overall accuracy, offering a practical solution for deploying deep models in resource-constrained environments.

In this study, we investigate a computationally efficient framework for real-time action recognition in Virtual Reality (VR) using a single-neuron-per-class paradigm. Our methodology introduces Image-Encoded Kinematics (IEK), a technique where complex 3D sports gestures are projected into compact 2D spatio-temporal representations. These features serve as inputs for a comparative evaluation between a bio-inspired Resonate-and-Fire (RAF) neural model and a standard Perceptron (Per) baseline. We employ a parallel One-vs-All architecture to distinguish specific sports patterns based on frequency-selective neural activation. The proposed shallow learning framework ensures high interpretability and sub-millisecond latency, making it suitable for resource-constrained hardware such as the Meta Quest 3. Experimental results demonstrate that while all models achieve near-perfect accuracy (above 99.7%) on clean data, the bio-inspired RAF model significantly outperforms the Per in noisy environments, leveraging its intrinsic frequency tuning to filter out sensor jitter. This work highlights the potential of combining topological data encoding with neuromorphic computing to create robust, low-latency immersive experiences. To facilitate reproducibility and further research, the dataset will be made publicly available upon acceptance.

Training Deep Neural Networks for tracking individual cells in biomedical videos requires a large amount of annotated data. The annotation of videos for cell tracking is very time consuming and often requires domain expertise; this explains the limited availability of public annotated data to address important medical problems like tissue repair or cancer treatment. Generating synthetic videos along with their Ground Truth annotations is a promising solution that relies, as a foundational first step, on the synthesis of single cell annotations (or phantoms). Phantoms need to be time consistent, as they have to replicate biological processes that are specific to the cell types. In this work, we propose a novel framework for generating videos of cell phantoms in the Elliptical Fourier Descriptors (EFDs) domain, a compact and geometrically interpretable representation for 2D closed contours. We represent the cell phantom evolution as a multivariate time series of EFD coefficients, introducing a strong prior for cell morphology and enabling the efficient generation of sequences that evolve coherently in time. Our experimental validation proves that modelling the temporal evolution in EFD space enables the generation of biologically plausible phantom videos. Our method can be used in generative pipelines for synthesizing annotated data for cell tracking, thus strongly mitigating the annotation effort for creating new datasets. Our code is available for download here: https://github.com/FrancescoBenedetto99/efd-cell-video-gen.

Falls represent a critical public health challenge, and accurate detection of the impact moment when an individual hits the ground is crucial for timely intervention. Existing skeleton-based methods rely on graph neural networks modeling only pairwise joint connections, failing to capture multi-joint coordination characteristic of fall impacts, while transformer-based temporal models suffer from quadratic complexity limiting real-time deployment. We propose FLASH, a novel framework integrating single-matrix hypergraph representations with Mamba's selective state-space models through adaptive feedback mechanisms for efficient impact detection. Our approach constructs biomechanically-grounded hyperedges to model functional joint coordination while leveraging Mamba's linear-time complexity to capture temporal dynamics. Experiments on UP-Fall and UMAFall datasets demonstrate that FLASH achieves state-of-the-art accuracy with real-time inference capability and strong zero-shot cross-dataset generalization, while significantly reducing computational cost compared to dual-representation and transformer-based methods. The model provides interpretable feedback through learned attention patterns aligned with biomechanical principles. Code is available at https://github.com/Tresor-Koffi/FLASH-Impact-Fall-Detection.

This paper explores the efficiency–accuracy trade-off in high-resolution document layout detection, introducing the Selective State-space Document Detector (SSDoc-Det), a compact detection framework designed for low-latency, resource-efficient deployment. Our method incorporates selective two-dimensional state-space modeling into both the feature extraction backbone and a hybrid encoder, enabling long-range dependency modeling while avoiding the quadratic overhead of self-attention. To facilitate effective multi-scale feature integration, a bottleneck-style convolutional alignment module allows the decoding stage to be streamlined with fewer layers. For bounding-box localization, SSDoc-Det replaces conventional IoU variants with MPDIoU and retains distribution-aware refinement and self-distillation mechanisms to enhance regression stability and precision. Extensive experiments on standard document layout benchmarks demonstrate that the proposed framework achieves strong detection accuracy while substantially reducing computational complexity and inference latency. These results show that combining selective state-space modeling with lightweight feature fusion and refined localization yields an effective solution for accurate, efficient document layout detection.

While high-capacity AI models have advanced state-of-the-art performance, their practical deployment is often hindered by high inference costs, environmental impact, and a "one-size-fits-all" approach that ignores varying sample complexity. In clinical settings for instance, the waste of computational resources on routine cases is a significant barrier to sustainable AI. In this paper, we introduce the Conformal Adaptive Decision System (CADS), a sequential multi-model algorithm designed to optimize resource allocation by efficiently sampling models based on the estimated data complexity. CADS leverages conformal prediction to quantify image uncertainty at runtime. CADS provides a mathematically grounded framework for balancing the cost-accuracy dilemma that dynamically routes samples through a model cascade, ranging from lightweight "Scout" models to high-capacity "Oracle" architectures. Validated on two datasets, CADS demonstrated superior efficiency and accuracy at a computational cost that can be up to 12 times lower than heavy-model inference. By accurately routing samples based on real-time complexity, CADS ensures high diagnostic reliability while drastically reducing the economic and environmental footprint of AI.

Facial Emotion Recognition (FER) is an important task that has significant implications across various fields such as biometrics, health, and human-computer interaction. Current Vision Transformer-based approaches display quadratic complexity $\mathcal{O}(N^2)$, with N being the input sequence length, making them cumbersome to deploy at the edge. In this paper, we hypothesize that the FER task does not necessarily require all facial information to correctly interpret emotional states, as specific regions such as the eyes, the mouth, and parts of the cheeks carry discriminative information that can be sufficient to recognize emotions. Based on this, we propose Sparse Attention to Emotion (SAE), a model that discards image tokens that have no added value to the emotional context, while preserving good accuracy and achieving a significant gain in computational cost. Surprisingly, even after suppressing 90\% of the image tokens, our model achieves competitive accuracy to state of the art methods at much lower cost, providing a lightweight Facial Emotion Recognition approach. Experimental results demonstrate that SAE achieves new state of the art results on the RAF-DB dataset while reducing the computational complexity by up to 90\%.

Deep neural networks have achieved remarkable success across a wide range of learning tasks, but their performance often comes at the cost of large training datasets and high computational overhead. To mitigate this, coreset generation techniques aim to construct compact yet representative subsets of data that enable efficient model training without compromising performance. However, many existing coreset methods rely on empirically tuned hyperparameters, which introduce additional computational complexity. In this work, we propose a novel coreset generation framework grounded in compressed sensing principles. Our approach identifies non-redundant and informative samples by selecting incoherent latent representations of data points, thereby preserving class diversity while significantly reducing dataset size. Extensive experiments on chest X-ray datasets demonstrate that the proposed method retains less than one-third of the original training data and achieves up to a fourfold reduction in per-epoch training time, while maintaining performance comparable to full-dataset training. Moreover, despite using fewer samples, our method achieves competitive results relative to existing coreset generation strategies.

Statistical dependence measures like mutual information is ideal for analyzing autoencoders, but it can be ill-posed for deterministic, static, noise-free networks. We adopt the variational (Gaussian) formulation that makes dependence among inputs, latents, and reconstructions measurable, and we propose a stable neural dependence estimator based on an orthonormal density-ratio decomposition. Unlike MINE, our method avoids input concatenation and product-of-marginals re-pairing, reducing computational cost and improving stability. We introduce an efficient NMF-like scalar objective and demonstrate empirically that assuming Gaussian noise to form an auxiliary variable enables meaningful dependence measurements and supports quantitative feature analysis.

Text image inpainting aims to restore clear and complete text images from degraded or incomplete inputs, with important applications in document digitization and preservation. Despite recent progress, existing methods still struggle with complex text structures and unique character styles. To address these challenges, we propose a novel framework that leverages a selective state space model to effectively capture textual continuity and structural features. Our approach integrates three key components: character style extraction for language-specific glyphs, StyleGAN-based structure prior generation, and prior-guided restoration. Extensive experiments demonstrate that our method significantly outperforms existing techniques in both quantitative metrics and visual quality. Code will be made publicly available upon publication.

Maritime unmanned aerial vehicle (UAV) object detection is critical for search-and-rescue and emergency monitoring, where fast and reliable discovery of open-water targets from high-altitude viewpoints directly affects the response time and mission outcomes. This work proposes Height-Aware Feature-Scale Adaptive Real-Time Detection Transformer (RT-DETR) to improve detection accuracy under altitude-dependent scale variations. The proposed Height-Aware Scale Selector (HASS) predicts height-conditioned weights to adaptively emphasize the primary feature scale for single-branch intra-scale interaction (between S4/S5) while retaining high-resolution cues for cross-scale fusion. To further reduce computation, we adopt MobileNetV4 as backbone and replace the multi-head attention in attention-based intra-scale feature interaction (AIFI) with an Efficient Multi-Scale Attention (EMA) module for lower-cost intra-scale feature enhancement. Experiments on the SeaDronesSeeV2 dataset show that the proposed method achieves 84.2 mAP50 and 52.3 mAP50:95, improving RT-DETR-R18 by +2.0 and +2.5, respectively, while reducing model complexity from 19.8M to 11.2M parameters and from 57.0G to 29.7G FLOPs.

Viewpoint change invariance and action temporal consistency are critical aspects for the effective deployment of human action detection of untrimmed videos. Existing appearance-based video detection methods often struggle with limited viewpoint diversity during training, while motion-based detection approaches frequently fail to model fine-grained temporal relationships across consecutive motion windows. This paper introduces a novel two-stage action detection approach designed to improve both view-invariance and global temporal coherence properties. In the first stage, we extract motion features from augmented virtual viewpoints, solely used at training. Then, the second stage introduces a new view-invariant, multi-scale temporal encoder based on selective state-space sequence modelling to aggregate information across viewpoints and time scales. Experiments on PKU-MMD and BABEL benchmarks demonstrate that this approach significantly outperforms state-of-the-art methods in all considered splits. Code and trained models are available at: https://icb-vision-ai.github.io/HydraView-TAD

We present a geometry-aware method for heterogeneous single-particle cryogenic electron microscopy (cryo-EM) re- construction that predicts atomic backbone conformations. To incorporate protein-structure priors, we represent the back- bone as a graph and use a graph neural network (GNN) autodecoder that maps per-image latent variables to 3D displacements of a template conformation. The objective combines a data-discrepancy term based on a differentiable cryo-EM forward model with geometric regularization, and it supports unknown orientations via ellipsoidal support lift- ing (ESL) pose estimation. On synthetic datasets derived from molecular dynamics trajectories, the proposed GNN achieves higher accuracy compared to a multilayer percep- tron (MLP) of comparable size, highlighting the benefits of a geometry-informed inductive bias.

Although deep multi-view subspace clustering (DMSC) methods have shown strong performance, in large-scale scenarios they are still limited by the overhead of constructing and processing multiple large self-representation matrices and by the high computational complexity of spectral clustering. Existing solutions mostly rely on specific model architectures or graph construction strategies, and lack a unified extension capability for existing DMSC methods. To address this issue, this paper proposes a scalable deep multi-view subspace clustering framework. The framework is generally applicable and can be plugged into existing DMSC methods, extending them to large datasets while retaining the original advantages as much as possible. Specifically, the framework learns low-rank factor matrices to compactly represent self-representation matrices, so that multi-view information fusion is performed in factor space. Furthermore, based on the structural properties of low-rank factor matrices, we reformulate the spectral clustering procedure by converting original large-scale matrix operations into eigendecomposition and matrix multiplication on small matrices, thereby significantly reducing storage and computational overhead. We validate the proposed framework on multiple datasets of different scales.

Skeleton-based action recognition has achieved remarkable progress through deep learning, yet its success remains tethered to the availability of massive, manually labeled datasets—an acquisition process that is notoriously labor-intensive. While Active Learning (AL) has emerged as a promising strategy to mitigate this dependency, existing state-of-the-art approaches typically rely on selection heuristics that lack a rigorous theoretical foundation. In this paper, we propose a novel, label-efficient framework that shifts the paradigm from heuristic sampling to a principled probabilistic display model. Our approach adaptively designs unlabeled exemplars by optimizing a global objective that balances data representativity, diversity, and uncertainty. To ensure these exemplars remain faithful to the underlying data distribution, this design is conducted within a latent space governed by Invertible Graph Convolutional Networks (GCNs). A primary contribution of this work is the introduction of a Composite Stability constraint; by regularizing the network’s global Jacobian, we preserve high expressive capacity while maintaining end-to-end stability. This allows the GCN to model complex, non-linear manifolds while ensuring the bi-directional stability required for in-distribution exemplar design. Extensive experiments on two benchmark skeleton recognition tasks showcase the effectiveness of our proposed method against the related work.

Facial Expression Recognition (FER) is crucial for healthcare emotional monitoring, but most fail on non-frontal views common in clinical settings. This study proposes a two-stage transfer learning framework to improve over the baseline with significant gains in fear recognition.

Drowsiness is a major cause of vehicular and industrial accidents, motivating the need for real-time monitoring systems that are accurate, non-intrusive, and deployable on low power platforms. While physiological signal–based approaches (e.g., EEG/EOG) provide high reliability, their dependence on sensor attachment makes them intrusive for everyday use. Camera-based solutions are more practical but face two key challenges. First, state-of-the-art deep learning–based face recognition and drowsiness detection models require substantial computational and memory resources, making them unsuitable for embedded deployment. Popular architectures such as DeepFace, Keras-Face, and VGG-Face are extremely accurate but Cloud-scale. Existing approaches for camera based drowsiness detection often rely on heavy convolutional or attention-based backbones, resulting in large parameter footprints and high operational complexity. Deploying these models on embedded systems typically leads to significant accuracy degradation. Second, traditional handcrafted pipelines, although lightweight, depend heavily on face detection and landmark extraction, making them fragile under illumination changes, occlusions, and multi-view variations. Our experiments show that even extended handcrafted feature pipelines struggle to exceed 75–82% accuracy under realistic conditions. To address these limitations, we propose a compact end-to-end CNN architecture enhanced with Convolutional Block Attention Module (CBAM). The proposed model achieves 99.05% accuracy on YAWDDand94.65%onNTHU-DDD,outperforming transfer-learning backbones such as MobileNetV2, ResNet-50, and EfficientNet-B0. Furthermore, component wise performance analysis confirms the effectiveness of CBAM-based attention refinement, while computational analysis shows that the model achieves superior predictive performance with only 0.539M parameters, enabling an effective balance between accuracy, compactness, and execution efficiency for resource-constrained drowsiness monitoring systems.

Unsupervised fine-grained image hashing aims to learn compact binary codes that preserve subtle visual differences among highly similar instances without manual annotations. However, most existing methods neglect collision resistance, leading to identical hash codes for slightly semantically different samples. In this paper, we propose Collision-Resistant Single-Pass Self-Supervised Semantic Hashing (CS3H), a collision-resistant framework that directly optimizes Hamming-space similarity via a single-pass normalized Hamming distance loss to produce well-separated binary representations. We further introduce a collision-sensitive attention module to emphasize rare and discriminative local patterns, reducing hash collisions and improving fine-grained discrimination. Experiments on multiple benchmarks show that CS3H consistently outperforms state-of-the-art methods in retrieval accuracy while achieving superior collision resistance with minimal computational overhead.

Learning with noisy labels is a fundamental challenge in visual recognition. Existing approaches often rely on heuristic statistics or computationally intensive multi-model training, lacking a principled foundation for distinguishing clean from noisy samples. We propose a lightweight, post-hoc detection framework guided by KL divergence and statistical features derived from loss dynamics. By modeling the temporal evolution of per-sample losses and predictive distributions, we extract compact descriptors capturing stability and uncertainty in training trajectories. Unlike methods requiring auxiliary networks or iterative refinement, our approach uses a single training run with no additional supervision. Extensive experiments on synthetic and real-world benchmarks demonstrate that our method consistently outperforms prior post-hoc detectors while offering improved interpretability through its clear statistical grounding.

Video anomaly detection (VAD) is the automated identification of abnormal events or patterns within a video, and is highly beneficial in settings such as security, healthcare, and content moderation. Despite recent advances, VAD continues to perform poorly on category-specific detection, easily confusing semantically similar classes, such as “stealing” and “robbery”. To remedy, we suggest integrating neuro-symbolic (NeSy) AI. Its core concept lies in combining current deep learning systems with symbolic reasoning systems to enforce human interpretable reasoning. We explore NeSy AI and its benefits over current DL models within the context of VAD. Particularly, we show that by incorporating human-readable symbolic attributes to a neural system, rather than solely relying on uninterpretable statistical patterns, allows for better discrimination of semantically similar classes. Experiments on UCF-Crime and XD-Violence demonstrate the potential of NeSy architectures to improve performance and explainability beyond the state-of-the-art.

Referring expression counting (REC) is an intention-driven task that requires context-aware visual reasoning. While recent vision–language models incorporate language for visual understanding, most existing REC methods rely on rulebased reinforcement learning with rewards focused primarily on final accuracy, overlooking the quality of intermediate reasoning. We propose REC-RL, a reinforcement learning framework that introduces a think–range–answer paradigm to explicitly optimize the visual reasoning process. RECRL employs Group Relative Policy Optimization and two lightweight rewards: an accuracy reward that combines range-based interval supervision with Gaussian-based precision guidance, and a format reward that enforces structured outputs. By modeling intermediate focus prediction as internal decision-making, REC-RL avoids additional annotations and better aligns with human perception. Extensive experiments demonstrate consistent improvements over strong baselines and robust generalization across benchmarks.

Open-Vocabulary Object Detection (OVOD) extends closed-set object detection to novel categories by leveraging vision-language alignment, enabling models to detect objects specified by natural language. While recent OVOD models achieved strong accuracy, their encoder and heavy fusion architectures impose limitations on real-time usability. Through this work, we present FM-OVD, a lightweight and efficient OVOD model that employs feature-wise modulation to condition multi-stage visual features on textual embeddings. Unlike previous models that rely on some form of embedding similarity, FM-OVD actively reshapes visual features with semantic guidance, reducing textual prompt sensitivity and improving generalization to novel categories. We propose the Vision-Language Feature-wise Linear Modulation (VL-FM) module, which adaptively scales and shifts fused multi-stage visual features according to natural language inputs. Extensive experiments on benchmark datasets demonstrate that FM-OVD achieves a superior balance between speed and accuracy, consistently outperforming existing real-time OVOD.

Hierarchical classification (HC) under real-world degradations (noise, occlusion, blur, low resolution) often hurts fine-grained accuracy and amplifies error propagation. We propose UGHM-Net, an HC framework that integrates uncertainty-quantification (UQ)-driven dynamic confidence adjustment, hierarchical backtracking, and a joint confidence--multi-level loss. UGHM-Net leverages a Vision Transformer (ViT) for global feature extraction and a Semantic-Guided Cross-Attention (SGCA) module for visual--semantic fusion between visual tokens and static label embeddings. Guided by UQ, the model adaptively selects reliable hierarchical paths and backtracks to higher levels when fine-level confidence is low, curbing over-classification and reducing cascading errors. The joint objective balances coarse- and fine-grained supervision. Experiments on degraded remote sensing and natural image datasets show that UGHM-Net consistently outperforms prior methods in $ISDL$, hierarchical precision ($P_H$), hierarchical recall ($R_H$), and level-wise accuracy, providing a robust and efficient HC solution via UQ-driven path control and visual--semantic fusion.

Despite their widespread adoption in image classification, saliency map-based post-hoc explanation methods lack rigorous quantitative evaluation of their ability to capture contrastive, class-dependent decision structures. Existing evaluation techniques rarely assess whether explanations meaningfully differentiate between opposing class predictions, particularly under adversarial perturbations, where perceptually similar inputs yield substantial prediction shifts. Framing image classification as a deductively closed process, we argue that explanations must justify the predicted class while meaningfully excluding alternatives. To address this gap, we introduce a novel metric for quantifying the contrastive power of saliency map-based explanations -- a rank-based Wasserstein distance defined over saliency maps of original inputs and their adversarial counterparts. Our analysis reveals that popular saliency methods yield invariant explanations across divergent class predictions, indicating limited sensitivity to class-specific decision boundaries. This limitation is effectively captured by the proposed metric, positioning contrastive evaluation as an important criterion and valuable extension to existing explanation sanity checks.

Recent advances in learned image compression (LIC) have enabled practical deployments, spurring active research into image compression for machines and progressive coding schemes. However, their integration remains under-explored: prior works on progressive machine codec predominantly target sample-level difficulty adaptation (i.e., easy-to-hard), without considering semantic-level scalability. In this work, we introduce a semantic hierarchy-aware progressive codec that enables semantic scalability (i.e., coarse-to-fine) from a single bitstream. We first systematically categorize ImageNet-1K classes into CLIP embedding-based semantic hierarchies. Based on a channel-wise autoregressive framework, we decompose latent representations into hierarchically ordered channel blocks, each explicitly optimized for its corresponding semantic hierarchy level. Extensive experiments demonstrate that our approach substantially improves coarse-level recognition at low bitrates while maintaining fine-grained accuracy at higher bitrates. By reframing progressive transmission through the lens of semantic scalability, our work provides an efficient and interpretable solution for task-adaptive image coding, outperforming existing progressive codecs under hierarchical evaluation.

Car-following modeling is fundamental to microscopic traffic simulation. Traditional theory-driven approaches offer interpretability but struggle to capture diverse driving behaviors. Data-driven approaches demonstrate accuracy but lack physical interpretability and may violate safety constraints. We propose SAPINN, a dual-branch architecture featuring LSTM encoder-decoder with self-attention for temporal modeling, adaptive IDM parameter learning for physics grounding, and scene-aware fusion for dynamic branch balancing. The scene-aware fusion mechanism computes danger and congestion scores from current traffic states to adaptively weight the contributions of data-driven and physics-informed branches. Experiments on multiple benchmarks demonstrate SAPINN achieves RMSE reductions of up to 31.6% compared to the methods while maintaining physical safety guarantees, showing superior performance in car-following prediction tasks.

Automated analysis of building facades from street-level imagery has great potential for urban analytics, energy assessment, and emergency planning. However, it requires reasoning over spatially arranged elements rather than solely isolated detections. In this work, we model each facade as a graph over window/door detections with a vertical prior on edges. Additionally, we introduce GATA2Floor, a multi-head Graph Attention v2 (GATv2) based model that predicts the global floor count of a building and, via learnable cross-attention queries, softly assigns elements to latent floor slots, yielding interpretable outputs and robustness to irregular designs. To mitigate the lack of labeled datasets, we demonstrate that the proposed graph-based reasoning can be applied without annotations by leveraging a lightweight label-free proposal mechanism based on self-supervised features and vision–language scoring. Our approach demonstrates the value of graph-attention-based relational reasoning for facade understanding.

Recent advances in image classification have been driven by architectural innovations and large-scale pretraining. However, once a model is trained, improving recognition performance without additional learning or architectural modification remains challenging. Although test-time augmentation (TTA) is a common strategy for boosting accuracy, it often incurs substantial evaluation-time overhead. This paper introduces Contribution-Aware Spatial Recalibration, a training-free post-hoc approach that enhances image classification by exploiting model-internal signals inherent in pretrained networks. The proposed method derives a spatial contribution map from internal representations, such as class-token attention in Vision Transformers or activation magnitudes in convolutional networks, to estimate the relative importance of local regions. Based on this representation, a spatially focused view is generated and re-evaluated by the same model, and the resulting prediction is combined with the original global prediction via a confidence-aware fusion mechanism. Extensive experiments on ImageNet-1K and multiple fine-grained recognition benchmarks demonstrate consistent accuracy improvements over standard single-pass evaluation using identical trained models. In particular, larger gains are observed on fine-grained datasets, where localized visual cues are critical for correct recognition. Owing to its architecture-agnostic design and training-free nature, Contribution-Aware Spatial Recalibration offers a practical and computationally efficient alternative or complement to conventional TTA-based evaluation.

This paper presents Attentive Graph-based Feature Learning from Image Markers (A-GFLIM), a flyweight graph convolutional feature extraction strategy that operates without backpropagation while preserving user control. The method maintains kernel–class correspondence and introduces a new aggregation function by discarding neighbor ordering without information loss, alongside a class-specific self-attention mechanism that dynamically weights node channels based on channel-wise importance and spatial consistency. Integrated with an SVM classifier, A-GFLIM is evaluated on imbalanced biomedical image datasets, demonstrating effective feature extraction with one training image per class. Although it may not surpass traditional deep architectures under identical conditions, A-GFLIM offers an interpretable and computationally efficient alternative for scenarios with limited resources.

Video Object Segmentation (VOS) has recently benefited from foundation models such as SAM2. However, these models still struggle in scenarios where objects undergo drastic physical or topological transformations. In this work, we show that the causes of this performance decrease are largely due to the progressive degradation of internal representations caused by erroneous mask propagation over time. Motivated by this insight, we propose SAM2TC, a lightweight extension of SAM2 that introduces a Transformation Compensation (TC) module operating directly on the model’s internal representations. The TC is trained using privileged information derived from an oracle configuration, learning to correct degraded features and bring them closer to transformation-consistent representations. Experiments on popular benchmarks demonstrate that SAM2TC significantly outperforms SAM2 and recent state-of-the-art methods, achieving improved segmentation accuracy to extreme object deformations and state changes.

Recent continual learning benchmarks for vision-language models involve sequentially learning over different datasets containing identical label names, semantically related classes, visually similar concepts, or multiple objects in the same image. These are evaluated on both previously seen and unseen datasets in a zero-shot manner. Currently used standard metrics like top-$1$ accuracy penalize models for semantically meaningful predictions from different datasets, even if the underlying concepts are correctly retained or transferred. In this work, we propose \textit{SRank}, a semantic-aware evaluation metric which constructs a semantic ranking of class names centered on the ground-truth class and compares it against the model’s top-$k$ predictions, assigning a graded credit based on semantic distances. SRank extends beyond binarized accuracy and provides a fine-grained measure of prediction quality. It is model-agnostic, and reduces to standard accuracy measure as a special case. Evaluation across multiple continual learning methods under recent dataset-level continual learning protocols demonstrates that SRank can represent failure cases and transfer behaviors that are not captured by baseline metrics, showing a higher correlation with human judgment.

Backdoor attacks, in which a model’s behavior is altered by embedding a trigger into the input to cause targeted misclassifications, pose an increasing security threat to deep neural networks. This is particularly concerning given the widespread reliance on open-source pre-trained models to overcome data and computational constraints. We propose a novel, attack-agnostic, and computationally efficient method for detecting poisoned models. Our approach uses activation optimisation to generate per-class optimised logit outputs that serve as model signatures. Instead of training a meta-classifier on the signatures, our method leverages statistical features for anomaly detection. We demonstrate that our method is robust to various trigger types, attack strategies, including BadNet, Sinusoidal, and WaNet, as well as model architectures such as VGG and ResNet, without requiring retraining. We validate our approach on three popular image-classification Trojan model benchmarks: the ULP CIFAR-10 and Tiny-ImageNet, and the ImageNet Vision Transformer-based TAT dataset.

We propose an imaging method for translucent objects using a high-temporal-resolution event camera and a line-scan projector for fast and efficient internal measurement. In translucent objects, the spatial spread of the Line Spread Function (LSF) varies depending on the internal structure and is observed as the temporal width of generated events at each pixel during line scanning. In the proposed method, we visualize internal patterns using this temporal width. Experimental results demonstrate that internal patterns inside translucent objects, which are difficult to observe using conventional cameras, can be visualized with only a single line scan.

Creating annotations for new fine-grained classes is expensive, particularly when pretrained models lack knowledge of the target category. We propose a human-in-the-loop pipeline for large-scale data harvesting from unlabeled crowdsourced data. Candidate objects are first extracted using segmentation or detection models, then filtered with CLIP \cite{clip} embedding similarity retrieval. A lightweight classifier trained on manually verified positives and negatives further refines the retrieval results, after which the collected samples are used to fine-tune the final production model. The retrieval, verification, and retraining stages are executed iteratively until convergence. The proposed approach significantly reduces manual annotation effort while enabling efficient extension of fine-grained recognition systems with new classes.

Modeling hand-object interaction requires not only accurate hand-object pose estimation but also explicit reasoning about contact. Most existing contact formulations treat contact as isolated points and ignore the patch-based nature of grasping. To address this limitation, we introduce an Elastomer Simulator for physically plausible hand-object interaction. The simulator models contact as deformable patches and defines a differentiable energy that balances attractive contact and repulsive interpenetration. Building on this formulation, we develop a Physically-grounded Elastomer Network that predicts the simulator-specific quantities and enables end-to-end pose optimization. Experiments on the HO3D and DexYCB datasets show improvements in both pose accuracy and physics-based quality over prior methods.

Mamba-based models have emerged as a promising alternative for salient object detection (SOD), offering significant advantages in modeling long sequences. However, existing models often fail to explore contextual information and the depth of the entire architecture. This paper introduces U$^2$Mamba, a powerful and innovative U-structured network for salient object detection. We propose multiscale Mamba U-blocks (MMUBs) that enhance the model depth to improve local feature extraction capabilities. Our newly developed nested U-structure, incorporating MMUBs, enables the network to integrate various receptive fields from shallow and deep layers, thereby collecting richer contextual information and longer-range data without being constrained by resolution. Instead of using the traditional deep supervision scheme and top-level supervised training, we propose a hierarchical training supervision method where the loss is computed at each level during the training process. Extensive experiments demonstrate that U$^2$Mamba achieves highly competitive performance against state-of-the-art methods. The source code is available at \url{https://github.com/JL021/U2Mamba}.

The democratization of deepfakes necessitates detectors that remain effective against unseen attacks and robust to media degradation. While many detectors rely on low-level noise artifacts, these "hidden signals" are often fragile. This work investigates whether Facial Action Units (FAUs) constitute discriminative features for deepfake detection. The rationale is twofold: generative models often struggle to consistently reproduce complex physiological muscle dynamics, and unlike pixel-level noise, FAUs are high-level semantic features potentially more resistant to compression or resizing. To determine whether FAUs are truly distinguishable features, we conduct an analysis across different learning configurations, comparing One-Class (training on real data only) and Binary classifiers using SVMs, DeepSVDD and Normalizing Flows models, across both identity-specific and identity-agnostic configurations. We evaluate their utility in discerning real videos from deepfakes across those multiple scenarios using the FakeAVCeleb dataset, aiming to answer whether FAUs offer a generalized and robust physiological signature for detection.

Reliable out-of-distribution detection is pivotal for deploying deep learning in open-world scenarios. However, existing methods often rely on a single feature modality, either semantic direction (confidence) or activation intensity (norm), limiting their robustness against diverse shifts. To bridge this gap, we propose a unified framework that harmonizes representation quality with decision certainty. We introduce the angular-magnitude likelihood, which jointly evaluates semantic alignment and signal typicality. To ensure the reliability of the magnitude component, we integrate convolutional block attention modules into discriminative layers. This approach purifies feature norms without diluting low-level texture information. Experiments demonstrate that our framework achieves state-of-the-art performance.

The exponential growth of multimedia content highlights the importance of specific methods to automatically extract the most representative key images, called key-frame - which we will refer to hereafter as tag-images - from videos. Unlike video surveillance, where the objects searched for are clearly defined, cinema offers a wide variety of visual and narrative styles, making the selection of a tag-image per shot particularly complex. In this paper, we propose a new method based on object detection, specially adapted to this context. The proposed method introduces an innovative strategy based on statistical analysis. This analysis relies on histograms of occurrences of detected object classes, enabling a more semantically consistent selection. This method is enriched with spatial, temporal, and sharpness weightings to optimize the relevance of a selected tag-image per shot. A qualitative evaluation conducted with film analysis experts shows that our method outperforms state-of-the-art approaches in terms of direct preference and robustness across various types of shots.

While one-stage detectors like YOLOv8 offer fast training speed, they often under-perform on detecting small objects as a trade-off. This becomes even more critical when detecting tiny objects in aerial imagery due to low-resolution targets and cluttered backgrounds. To address this, we introduce four enhancement strategies—input image resolution adjustment, data augmentation, attention mechanisms, and an alternative gating function for attention modules—that can be easily implemented on YOLOv8. We demonstrate that image size enlargement and the proper use of augmentation can lead to enhancement. Additionally, we designed a Mixture of Orthogonal Neural-modules Network (MoonNet) pipeline which consists of multiple attention-module-augmented CNNs. Two well-known attention modules, Squeeze-and-Excitation (SE) Block and Convolutional Block Attention Module (CBAM), were integrated into the backbone of YOLOv8 to form the MoonNet design, and the MoonNet backbone obtained improved detection accuracy compared to the original YOLOv8 backbone and single-type attention-module-augmented backbones. MoonNet further proved its adaptability and potential by achieving state-of-the-art performance on a tiny-object benchmark when integrated with the YOLC model. Our code is available at: https://github.com/Kihyun11/MoonNet

Reliable object detection is critical for automated driv- ing, yet even state-of-the-art detectors inevitably make errors. Introspection methods that predict when a detector is likely to fail enable safer deployment by triggering fallback mech- anisms or alerting human operators. Existing approaches rely solely on hand-crafted specific layer features, discarding potentially valuable information from earlier layers that cap- ture different levels of visual abstraction. We propose Layer Feature Attention (LFA), an introspection method that learns to aggregate features from multiple backbone layers through an attention mechanism. Our key insight is that detection errors manifest differently across feature hierarchies—low- level layers capture fine-grained details useful for detecting small or occluded objects, while high-level layers encode semantic information for scene understanding. LFA learns layer importance weights end-to-end, enabling both better error prediction and interpretable analysis of which feature levels are most indicative of detector failures. Experiments on KITTI and BDD100K demonstrate that LFA achieves state-of-the-art introspection performance, outperforming single-layer baselines.

Automated animal re-identification (Re-ID) is vital for longitudinal ecological monitoring. However, current benchmarks largely overlook two critical generalization challenges: performance degradation across unseen ecological contexts and drastic optical shifts. To bridge this gap, we introduce MYST (MalaYsian Sea Turtle), a novel benchmark designed to evaluate model robustness to the ecological gap (captivity-to-wild) and the medium gap (aquatic-to-terrestrial). We further propose a Chronological Open-Set (COS) Protocol that utilizes the DIR@FPIR metric to rigorously couple identification accuracy with impostor rejection. Benchmarking state-of-the-art models reveals critical generalization failure modes. While the MegaDescriptor excels on its source domain ($96\%$ Rank-1), it fails to generalize across the ecological gap, underperforming hand-crafted SIFT descriptor ($70\%$ vs. $76\%$), despite being trained on the same species in rehabilitation settings. More critically, under the medium shift of the nesting environment, MegaDescriptor collapses to $22\%$ DIR@1\%FPIR, indicating a high risk of identity merging. In contrast, the learned local descriptor ALIKED demonstrates superior robustness, achieving not only high $97\%$ Rank-1 accuracy but also robust open-set reliability ($91.7\%$ DIR@1\%FPIR) by exploiting the pose stability of nesting data. MYST provides a standardized testbed to diagnose these failures, pushing the evaluation of animal Re-ID systems towards the ecological and optical robustness required for real-world deployment.

Early-exiting neural networks enable adaptive inference by allowing inputs to exit at intermediate classifiers, reducing computation for easy samples while maintaining high accuracy. In practice, exits can be trained sequentially by incrementally adding them to a shared backbone; however, this sequential training can cause newly introduced exits to interfere with previously learned ones, degrading the performance of earlier classifiers. We address this problem by retaining the knowledge embedded in existing exits while allowing new ones to specialize. We propose two alternative approaches that operate at different levels of the model. The first constrains learning by protecting parameters that are important for previously trained exits, while the second preserves the output distributions of earlier exits as the network adapts. These alternatives directly reflect the stability-plasticity trade-off studied in continual learning. Accordingly, we leverage Elastic Weight Consolidation to constrain critical weights and Learning without Forgetting to preserve output distributions. Experiments on standard benchmarks show that our approaches consistently improve early-exit performance, achieving higher accuracy over existing sequential training methods and significant performance speedups at low computational budgets.

Accurate detection of camouflaged crops, where objects share visual characteristics with their surrounding foliage, is critical for modern agricultural tasks. However, the high computational demands of state-of-the-art deep learning detectors often preclude their deployment on resource-constrained edge devices. This paper presents a systematic evaluation of three lightweight design strategies comprising backbone replacement, pruning, and knowledge distillation applied to SSD, YOLOv8l, and RT-DETR architectures. Using the TomatOD and CCrop datasets, we investigate the trade-offs between mean Average Precision and computational efficiency. Our findings reveal divergent architectural resilience where RT-DETR and SSD demonstrate improved precision under moderate global pruning, with RT-DETR achieving a +7.30 gain in mAP50 at 10% sparsity. Conversely, YOLOv8l exhibits superior stability during extreme architectural compression via backbone substitution but suffers sharp performance decline under pruning levels exceeding 5%. Furthermore, we show that knowledge distillation allows lightweight students to occasionally outperform their teachers in specific agricultural domains. These results provide guidance for optimizing object detectors for real-time agricultural applications in complex and resource-limited environments.

Deep learning has achieved remarkable success in deformable image registration, yet the visual information that drives deformation estimation remains poorly understood. Rather than pursuing incremental performance improvements, this work investigates the fundamental source of robustness in deep registration models. Using diverse, domain-agnostic synthetic datasets, we decouple deformation learning from application-specific appearance and show that domain-shift immunity is an inherent, largely architecture-agnostic property of deep deformable registration when trained with a robust pipeline. To identify the mechanism underlying this immunity, we compare models trained on raw image intensities with models operating exclusively on fixed, pre-defined local feature representations. The comparable performance of these models provides strong empirical evidence that deformation estimation is governed primarily by local structural features, rather than global, domain-specific appearance cues. These findings offer a principled explanation for the cross-domain generalizability of deep registration networks and point toward feature-centric designs for domain-independent registration.

Object detection in images still remains one of the central problems in image analysis. The You Only Look Once (YOLO) detectors have become a de facto standard for this task in a very wide range of applications of diverse image types and qualities. Extensive research is devoted to the introduction of attention models in these detectors for performance improving. We propose an alternative to attention models and introduce Explainable AI (XAI) methods, Explainability Based Learning (XBL), into the training and generalisation for object detection. Two configurations are studied; the best one follows a teacher-student scheme. Experiments show an improvement in performance of more than 0.1 points on the challenging Railway defect dataset and on the more diverse SALICON dataset, compared to the best-performing and most recent attention models for YOLO: CBAM, SE and GAM. The results thus indicate that XBL is a promising alternative to the attention modules in the task of object detection.

Recent video saliency prediction (VSP) methods typically focus on modeling visual saliency of low-resolution (LR) frames, overlooking the information loss caused by downsampling. To address this issue, we propose a model paradigm for high-resolution (HR) video saliency prediction in this work. Specifically, the input video is first processed into two parts: an LR frame sequence containing semantic information and an HR frame preserving fine-grained details. Based on the different emphasis of information in the two parts, specialized two-stream encoders then process each part according to its distinct characteristics. We further propose an enhancement module to strengthen critical detail information. This module exploits the semantic cues encoded in low-resolution (LR) features to enhance key detailed representations in high-resolution (HR) features. Finally, a decoder utilizes these enhanced features to predict the saliency map. Extensive experiments on two high-resolution video saliency datasets demonstrate that our method achieves superior performance compared to state-of-the-art approaches.

Magnetic Resonance Imaging (MRI) plays a crucial role in detecting and classifying anatomical structures. However, automated analysis of medical magnetic resonance imaging remains challenging due to the many anatomical structures (e.g., the ACL and meniscus in the knee) and surrounding soft tissues that have very similar signal intensities, making their boundaries difficult to classify clearly. In this paper, we present a lightweight, end-to-end framework that bridges 2D and 3D analysis. By using a multi-scale center cropping strategy to capture anatomical details at differing resolutions and employing task-aware memory with depth-aware attention, our method automatically classifies the most diagnostic angle for each injury while maintaining high computational efficiency for clinical deployment. In evaluations on the MRNet dataset for the knee, our framework demonstrates excellent performance in labeling volumetric sequences, while also providing a mechanism to highlight the relevant anatomical features that contribute to diagnosis, thereby bridging the gap between automated scores and clinical reasoning. By combining multi-scale inputs with precise, context-aware feature encoding, our framework provides a robust, lightweight, and interpretable solution for diagnosing complex knee injuries.

Social group activity recognition is a task that detects humans and identifies multiple human groups and their activities in a scene. Existing methods rely on video input for this task. This paper extends social group activity recognition from video to still images to enable broader application. Since still images lack motion information, detecting humans and their interactions is challenging. As a result, reliable recognition usually requires large annotated datasets, which are not easy to obtain. To address this issue, our key idea is to train a model simultaneously on tasks with rich still-image datasets that localize objects, such as human detection and object detection, to strengthen the model’s ability to find and localize objects, including humans. To this end, we introduce a multitask conditional token sequence generation model that, given a token specifying the target task, outputs the task result as a token sequence. To represent social group activity recognition results, we propose an activity label token that simultaneously represents the activity label and the group boundary. The proposed method allows the model to be trained on multiple object localization tasks and improves performance. Experiments on public benchmark datasets demonstrate the effectiveness of our approach.

Bone age estimation from hand and wrist X-rays is a basic task in pediatric radiology, allowing for the detection of growth abnormalities like precocious puberty, but is difficult due to the subjective nature of skeletal maturity. Traditional deep learning models mainly rely on deterministic regression, which can be overly confident and less interpretable by ignoring predictive uncertainty. To overcome these issues, we introduce an end-to-end uncertainty-aware model that uses a ConvNeXt-Tiny backbone to predict a full age probability distribution. The key contribution of this paper is the Self-Consistent Distribution KL (SCD-KL) loss function, a new objective that promotes consistency between predictions and soft labels, while imposing constraints to ensure that the distributions are unimodal and well-calibrated around the model’s internal prediction. Experiments conducted on the RSNA Pediatric Bone Age Challenge dataset show that our approach is comparable to existing state-of-the-art models, achieving a Mean Absolute Error (MAE) of 4.17 months and an R2 score of 0.98. Our method provides a more robust, interpretable, and useful tool for pediatric radiology analysis by modeling ambiguity and predicting a set of possible ages.

Existing evaluations of fMRI-based visual brain decoding (VBD) primarily rely on pixel-level or semantic similarity, offering limited insight into whether attentional structure, a core aspect of visual processing, is preserved. This limitation is especially relevant for real-world and clinical applications, where how visual attention is encoded and expressed matters beyond visual appearance. In this work, we treat saliency as a proxy for visual attention and conduct a structure-aware analysis of attentional fidelity across VBD pipelines. Saliency maps are represented as graphs that capture the relational organization of attention, enabling comparison between stimulus images, brain-derived fMRI–CLIP embeddings, and reconstructed outputs using established graph similarity measures. This formulation allows to explicitly disentanglerepresentation loss introduced during brain-to-embedding mapping from generation loss arising in image synthesis. A multi-metric, graph-based analysis across subjects and reconstruction pipelines reveals systematic differences in how attentional structure is preserved or degraded, providing diagnostic insights beyond conventional evaluations

Learned image codecs such as JPEG AI offer strong rate-distortion performance but introduce computationally heavy decoding pipelines, which can be unnecessary when the end goal is automated visual analysis rather than human viewing. This paper studies object detection directly from JPEG AI pre-reconstruction latent representations, bypassing image synthesis. We build on YOLOv12 and propose an adaptation strategy that translates codec latents into task-compatible features by combining a lightweight latent-to-feature bridge with feature-level knowledge distillation from the early stem of a pixel-domain teacher. Experiments on COCO demonstrate that JPEG AI latents retain sufficient semantic information to support meaningful object detection. Although latent-domain inference incurs an accuracy gap relative to pixel-domain detection, the proposed method substantially reduces computational cost by replacing the decoder with shallow processing, enabling a favorable accuracy-efficiency trade-off for compressed-domain analytics.

This paper introduces a retinal-inspired neuromorphic edge detector by combining adjustable receptive fields, ON/OFF bipolar pathways, and threshold modulation, to closely reflect how the retina balances fine-detail acuity and broad structural context. By using Gaussian centre–surround filters with different $\sigma$ configurations, the proposed detector replicates foveal and parafoveal functions, preserving fine details while also capturing broader structural information. The ON- and OFF-bipolar pathways enhance local luminance increments and decrements, while adaptive threshold modulation of bipolar and amacrine responses refines sensitivity and selectivity under varying visual conditions. The results show that the proposed detector achieved superior edge continuity, noise suppression, and dynamic range handling compared to classical methods. This approach, which produces edge maps that closely approximate biological vision, offers robust pre-processing for tasks such as object recognition, scene analysis, and spiking neural network training.

Skeleton action recognition involves recognizing human action from human skeletons. The use of graph convolutional networks (GCNs) has driven major advances in this recognition task. In real-world scenarios, the captured skeletons are not always perfect or complete because of occlusions of parts of the human body or poor communication quality, leading to missing parts in skeletons or videos with missing frames. In the presence of such non-idealities, existing GCN models perform poorly due to missing local context. To address this limitation, we propose Parts-Mamba, a hybrid GCN-Mamba model designed to enhance the ability to capture and maintain contextual information from distant joints. The proposed Parts-Mamba model effectively captures part-specific information through its parts-specific scanning feature and preserves non-neighboring joint context via a parts-body fusion module. Our proposed model is evaluated on the NTU RGB+D 60 and NTU RGB+D 120 datasets under different occlusion settings, achieving up to 12.9% improvement in accuracy.

Steganography conceals secret information within innocuous carriers while preserving visual fidelity and enabling reliable recovery. Recent unified networks operate normally under untriggered conditions but switch to hidden steganographic tasks when triggered. PUSNet follows this paradigm by performing image purification during normal operation and steganographic embedding when activated. However, it supports only a single user with one key pair, limiting its applicability in multi-user settings. We propose PUSNet-MK, a multi-key extension that enforces strict key isolation via a mismatched-key isolation loss, effectively preventing cross-key decoding when a wrong key is applied. This design preserves the intended steganographic behavior while addressing a critical security limitation of PUSNet. Extensive experiments demonstrate that PUSNet-MK produces high-quality stego images and accurate secret recovery, while preventing unintended information leakage. Code is available at https://gitlab.pairlab.ai.nycu.edu.tw/pairlab/ai.phd/pusnet-mk.

Existing methods for instance-level 6D pose estimation typically rely on neural networks that either directly regress the pose in SE(3) or estimate it indirectly via local feature matching. The former struggle with object symmetries, while the latter fail in the absence of distinctive local features. To overcome these limitations, we propose a novel formulation of 6D pose estimation as a conditional flow matching problem in R3. We introduce Flose, a generative method that infers object poses through a denoising process conditioned on both geometric and appearance-based semantic features. Unlike prior conditional flow matching approaches, which rely solely on geometric guidance, Flose leverages semantic features to mitigate ambiguities caused by object symmetries while incorporating RANSAC-based registration for robustness against outliers. We validate Flose on five datasets from the established BOP benchmark, outperforming prior methods with an average improvement of +4.5 in Average Recall.

Large multimodal models (LMMs) provide rich image--text latent representations, but representing abstract concepts such as interior style within human-interpretable semantic spaces remains challenging. We propose a framework that constructs a two-dimensional interior style semantic space by linearly aligning image and language latent representations of an LMM using style descriptions. The resulting space represents abstract styles, such as Cool--Warm and Light--Heavy, as continuous semantic axes. Experiments on expert-labeled interior images show that the proposed method achieves higher cluster separability and style classification accuracy than direct projection onto pretrained vision--language spaces. We further demonstrate that the constructed space enables quantitative analysis of stylistic differences between current and ideal interiors.

While text-to-image diffusion models have made remarkable strides, generating accurate scene text remains a challenge, often suffering from spelling errors and semantic inconsistencies. Prior approaches typically rely on fine-tuning diffusion architectures or manual post-processing tools, lacking an effective verification and self-correction mechanism. To address this, we propose Text-Pilot, a unified and training-free framework that employs a Multimodal Large Language Model (MLLM) as an autonomous agent to reason about text-prompt inconsistencies and coordinate multiple visual foundation models for correction. Through iterative perception-reasoning-action cycles, the system automatically detects, edits, erases, and regenerates textual regions to ensure consistency and legibility. Extensive experiments and rating-based evaluations incorporating several Vision-Language Models (VLMs) demonstrate that Text-Pilot can be seamlessly integrated with various text-to-image models, significantly enhancing text accuracy while preserving overall image quality.

Per-pixel visual correspondence underlies tasks such as stereo matching, optical flow, and video frame interpolation. While learnable cost volumes are highly effective in stereo, extending them to 2D search spaces is computationally prohibitive, leading most methods to rely on channel-collapsing all-pairs correlations and iterative refinement. We propose Generalized Learnable Cost Volume (GLCV), a unified formulation that enables learnable cost aggregation for 2D correspondence. GLCV decomposes the 2D search space into orthogonal horizontal and vertical 1D subspaces via attention, transforming an intractable 5D volume into two efficient 4D learnable cost volumes while preserving channel-wise geometric information. Experiments on optical flow and video frame interpolation demonstrate that GLCV achieves higher accuracy with substantially fewer optimization iterations. GLCV thus bridges stereo-style learnable cost volumes and general 2D correspondence, providing an efficient and general solution for dense visual matching.

Video editing poses a significant challenge. While a series of tuning-free methods circumvent the need for extensive data collection and model training, they often underutilize the rich information embedded within noisy latent, leading to unsatisfactory results. To address this, we propose a \textit{tuning-free, instruction-based} video editing framework. We approach video editing from the perspective of noisy latent: we design a Structural Noise Initialization Strategy (SNIS) to secure a superior editing starting point by assigning higher noise levels to edited regions (to facilitate content change) and lower noise levels to unedited regions (to maintain content consistency). We introduce a Noise Guidance Mechanism (NGM), which leverages the video prior in the generative model and effectively integrates rich information within the noisy latent to guide the denoising process, thereby preserving unedited content and overall visual coherence. Experiments show that our proposed method achieves better visual quality and state-of-the-art performance.

Music style transfer blends source structure with reference style to enable personalized creation. However, existing zero-shot methods often struggle to capture fine-grained audio nuances, relying on coarse text descriptions or requiring expensive task-specific training. We propose Stylus, a training-free framework that repurposes pretrained image diffusion models for music style transfer in the Mel-spectrogram domain. By treating audio as structured time-frequency images, Stylus manipulates self-attention by injecting style keys and values while preserving source structural queries. To ensure high fidelity, we introduce a phase-preserving reconstruction strategy to mitigate spectrogram inversion artifacts, alongside a classifier-free-guidance-inspired control for adjustable stylization. Extensive evaluations including 2,925 human ratings demonstrate that Stylus outperforms state-of-the-art baselines, achieving 34.1% higher content preservation and 25.7% better perceptual quality. Our work validates that generic image priors can be effectively leveraged for the training-free transformation of structured Mel-spectrograms. Code and materials are available at https://github.com/Sooyyoungg/Stylus.git

Recent advances in scene-based video generation have enabled systems to synthesize coherent visual narratives from structured prompts. However, a crucial dimension of storytelling—character-driven dialogue and speech—remains underexplored. In this paper, we present a modular pipeline that transforms action-level prompts into visually and auditorily grounded narrative dialogue, enriching visual storytelling with natural voice and character expression. Our method takes as input a pair of prompts per scene, where the first defines the setting and the second specifies a character’s behavior. While a story generation model such as Text2Story produces the corresponding visual scene, we focus on generating expressive, character-consistent utterances grounded in both the prompts and the scene image. A pretrained vision-language encoder extracts high-level semantic features from a representative frame, capturing salient visual context. These features are then integrated with structured prompts to guide a large language model in synthesizing natural dialogue. To ensure contextual and emotional consistency across scenes, we introduce a Recursive Narrative Bank—a speaker-aware, temporally structured memory that recursively accumulates each character’s dialogue history. Inspired by Script Theory in cognitive psychology, this design enables characters to speak in ways that reflect their evolving goals, social context, and narrative roles throughout the story. Finally, we render each utterance as expressive, character-conditioned speech, resulting in fully-voiced, multimodal video narratives. Our training-free framework generalizes across diverse story settings—from fantasy adventures to slice-of-life episodes—offering a scalable solution for coherent, character-grounded audiovisual storytelling.

Large Vision-Language Models (LVLMs) integrate visual understanding and language generation, and have achieved strong performance. However, they still suffer from hallucinations, where objects or attributes not present in the image are generated. Recently, Contrastive Decoding (CD)-based methods have been proposed. These methods mitigate hallucinations by comparing output distributions conditioned on the original input and a deliberately constructed contrastive input that induces hallucinated predictions at inference time. Existing approaches rely on noise injection, which is insufficient for effective hallucination induction, and use a fixed hyperparameter to control the strength of CD, limiting adaptability across inputs. To address this, we propose a new decoding strategy that constructs structure-aware contrastive inputs by masking foreground components and adaptively weights CD based on output distributional differences. Specifically, we identify and mask foreground using Singular Value Decomposition (SVD), and measure the semantic discrepancy between the two distributions using an optimal transport distance and then use it to set the adaptive CD weight. Experiments on multiple hallucination and general LVLM benchmarks demonstrate that our method outperforms existing approaches without additional training.

Large Vision-Language Models (LVLMs) often generate hallucinations, producing textual responses misaligned with visual input. We introduce CATEM, Contrastive decoding with ATtention Enhancement and Masking, a training-free approach to improve response reliability. Noting that attention to image regions, even with existing enhancements, still decays over decoding steps, we design an attention enhancement module that fuses current and historical attention to maintain consistent focus on critical image regions. Furthermore, an attention-masking strategy is employed to perturb attention to induce spurious predictions, which are then contrasted against the enhanced outputs to suppress hallucinations. These two modules jointly amplify visually supported tokens while suppressing hallucination-prone ones. Experiments and analyses demonstrate that CATEM substantially reduces hallucinations in image captioning and achieves competitive performance on object presence judgment.

Virtual try-on has become a key component of online shopping, allowing users to assess clothing fit before purchase. However, existing methods often struggle to preserve fine garment details and achieve accurate body alignment, leading to unnatural results. To address these limitations, we propose a novel self-supervised framework that integrates both image- and video-based priors. Specifically, we employ self-supervised image representations to capture fine-grained textures and combine them with diffusion models for high-fidelity synthesis. In addition, pretrained video-based self-supervised models provide motion priors to guide garment warping, ensuring realistic adaptation to human poses. With these strategies, our method delivers highly realistic virtual try-on results, and extensive experiments confirm substantial improvements in detail preservation and overall realism, establishing a new state-of-the-art for virtual try-on.

Learned image compression faces a fundamental dilemma between pixel fidelity and perceptual realism. Existing perception-distortion (P-D) controllable frameworks are often restricted by retraining costs and architecture dependencies. We propose ADAPT, a training-free framework that enables P-D control for existing image codecs. The method utilizes a per-image optimization strategy at the encoder to estimate parameters based on specific quality preference. At the decoder, a two-stage diffusion process refines compressed image: the first stage suppresses compression artifacts to establish structural accuracy, while the second stage restores textures through guided denoising. Evaluations demonstrate that ADAPT achieves a P-D balance comparable to specialized state-of-the-art models across diverse codecs with a side-information overhead of 0.001 bpp.

Multimodal Large Language Models (MLLMs) have achieved remarkable success in Visual Question Answering (VQA), yet their ``black-box'' nature hinders deployment in critical domains. While Grounded Visual Reasoning (GVR) approaches attempt to improve interpretability by explicitly couple textual rationales with visual grounding information, typically textual coordinates. This mechanism lacks a learnable semantic link to the visual features, often resulting in a semantic-spatial gap where the model hallucinates coordinates that do not correspond to image evidence. In this work, we introduce Composer, an MLLM that leverages a novel visual grounding mechanism based on learned proxy-tokens to promote faithful interpretability. These discrete symbolic pointers explicitly index the image latent space, allowing the model to manipulate visual regions as addressable, semantically manipulable sets. To rigorously validate our novel grounding mechanism, we constructed ComposerGCoT, a dataset synthesized to enable holistic assessment of reasoning consistency and grounding accuracy. Experiments demonstrate that Composer achieves performance parity with strong baseline on general VQA tasks while significantly outperforming its coordinate-based variant in grounding precision (+9.0 points). By demonstrating that discrete proxy-tokens capture spatial semantics more effectively than naive coordinates, we establish that building explicit, learnable links between modalities is a critical requirement for trustworthy and reliable MLLMs.

Flow Matching (FM) models have achieved remarkable success in generative tasks. However, a critical limitation persists: users often desire to modify specific image regions while retaining others. Current approaches force users to regenerate the entire image, discarding partially-acceptable results. To address this, we propose RPO (Regional Preference Optimization), a novel training-free framework for precision refinement. RPO introduces Regional Preference Loss, enabling direct optimization of the source noise vector based on localized user preferences. Users simply mark satisfactory regions to retain and unsatisfactory ones to regenerate. RPO intelligently optimizes the initial noise: it preserves visual features in preferred regions while synthesizing coherent replacements for undesirable areas. Extensive experiments across various models highlight RPO's advantages in generation quality and adaptability.

In the area of customization of text-to-image generation, previous methods rely on costly finetuning of the model parameters per subject or require heavy training on large datasets. Although training-free customization methods have emerged recently, image layout preservation has been an issue in the multi-subject setting. In this work, we present FreeInstance, a novel training-free method capable of multi-subject customization while preserving the original image layout. We introduce simple but effective techniques of rescaling attention score and applying layout instance mask. FreeInstance outperforms existing training-based and training-free methods on a range of metrics, highlighting its clear advantages. In addition, our method can also be extended for real image editing by customizing the image with the reference appearances without affecting the original image layout. Code will be available in https://github.com/fengming001ntu/FreeInstance.

We present SIGHTA-AI, a two-stage on-device architecture for visual assistance that reduces user-perceived latency for blind and low-vision (BLV) users. Stage 1 (FastVLM, 0.5B parameters) generates a scene description from a wearable camera using a fixed prompt; Stage 2 (Qwen3-0.6B) answers the user’s spoken question from this description. The key insight is speech-overlapped inference: because Stage 1’s prompt is fixed, it triggers the instant voice input is detected and runs in parallel with the user’s speech. For natural- length queries (≥2 s), Stage 1 completes before the user finishes speaking, reducing perceived post-speech latency from ∼2.6 s (monolithic) to ∼0.7 s—a 3.2× improvement. A default analysis mode with proactive caching eliminates perceived latency entirely. Both stages run on a commodity smartphone (iPhone 14 Pro) without cloud dependency.

Gaseous fluid flow reconstruction has significant practical value in monitoring and quantifying gaseous fluid leakage, emissions, and evaporation. Compared to traditional methods, video-based gaseous fluid flow reconstruction offers a promising approach for understanding fluid dynamics and demonstrating fluid flows. However, it faces several challenges, including the absence of persistent geometric structures and the emergence and disappearance of visual features over time. In this paper, we present a physics-guided gaseous fluid flow reconstruction method, PhysGasFluid, to enhance physical plausibility. Specifically, a Cubic Spline kernel, a Lagrangian baroclinic turbulence model, and a divergence-free loss are incorporated to address existing limitations. We evaluate our method on widely used benchmarks across three real-world datasets. Experimental results demonstrate that our method excels at reconstructing gaseous fluid flow, enabling physically plausible novel view synthesis and re-simulation.

SpheroSeg is the open-source, browser-based platform behind our accepted paper — turning manual wound-healing analysis into a fast, reproducible, installation-free workflow. In this live demo, attendees drive the full pipeline themselves. We will (1) upload brightfield microscopy images directly in the browser; (2) run prompt-free segmentation with the supervised model from the paper — a regularized U-Net with a hierarchical transformer encoder that reached 90.0% IoU under domain shift on the external Löwenstein dataset, outperforming heuristic and foundation-model baselines; (3) inspect and, where needed, correct predictions in the integrated human-in-the-loop annotation editor, closing the ground-truth loop live; and (4) export per-image quantitative read-outs (masks plus wound-area and closure metrics) as ready-to-analyse files. Visitors can bring or select their own images to test generalization on the spot, across cell lines. The demo shows how one command-free tool covers the entire annotate-to-quantify loop and how it generalizes beyond its training data. Code, weights, and the 446-image multi-cell-line dataset are released under open licences. Live at spherosegapp.utia.cas.cz. A ready-to-use demo account is preloaded with the walkthrough data — sign in with [email protected] (password icip2026) to open the "Wound Healing Demo (ICIP 2026)" project of 12 brightfield scratch-assay images, already segmented and ready to inspect, correct, and export.

Early and automatic wildfire detection is critical for minimizing environmental damage, infrastructure loss, and threats to human life. However, real-time wildfire detection and monitoring remain challenging, caused by atmospheric distortion, motion, and illumination variability. To address these challenges, we present GWFP (Global Wildfire Prevention Dataset), a large-scale, open-source, multimodal dataset designed to support robust, efficient, real-time deep learning-based Artificial Intelligence (AI) detection models. The GWFP dataset is compiled from publicly available sources, including the University of California San Diego (UCSD) HPWREN, the General Directorate of Forestry-Turkey camera networks, and drone-based recordings. The GWFP dataset includes seven classes: Flames, Smoke, Negative Samples, Waterdogs, Near-Infrared (NIR) Fire, NIR No Fire, and Ember. The inclusion of NIR imagery enables cross-spectral analysis and supports multimodal fusion under challenging visibility conditions. Firefighters often report excessive false alarms from video-based smoke detection systems triggered by clouds, fog, or changes in lighting, reducing trust in automated alerts and wasting critical response resources. This dataset and real-time wildfire detection demonstration aim to advance robust video-based smoke detection and efficient edge AI systems, designed with low-cost FPGA-based deployment in mind, to enable more reliable early wildfire detection and prevention.

Tamper with our image — we'll show you exactly where you touched it. This demo lets attendees interact live with Pixel Block Chain (PBC), a learning-free watermarking scheme that embeds a grid of independent hash-linked integrity chains into an image's least-significant bits. Visitors pick a photo (or bring their own), which we protect in under a second on an ordinary CPU laptop. They then edit it freely in a standard paint application: draw, erase, clone-stamp, or paste over any region. Our verifier immediately renders a tile-level verdict map — GREEN for intact regions, RED for every 128x128 tile they modified — demonstrating exact spatial tamper localization with zero false positives and no cascade to untouched regions, a property the accompanying paper proves formally. Two further interactions follow. First, a compliant-editor workflow appends entries to the in-image Edit Ledger, showing how legitimate edits remain distinguishable (YELLOW) from unexplained tampering (RED). Second, attendees crop the image arbitrarily; PBC-Forest, our scatter mode, recovers surviving genesis anchors from the fragment (41-45% after severe non-aligned crops, where regular grids recover nothing). The entire pipeline is open source (NumPy + Pillow, no GPU, no network) and runs live on the presenter's laptop, mirrored to the provided display. Code: github.com/flegare/pixel-block-chain.

This interactive demo presents PCAR, a profile-aware Planner–Critic Agentic Remediation system for trustworthy recipe generation from food images. Attendees can select or upload a food image and specify a user profile containing allergies, dietary requirements, and preparation-safety constraints, such as nut allergy, vegan or halal diets, pregnancy-safe preparation, or low-sodium needs. The Planner Agent first identifies likely ingredients and generates a personalized recipe draft. The Safety Critic Agent then independently audits the draft for allergen leakage, dietary conflicts, hidden derivative risks, and unsafe preparation steps. When a violation is detected, the interface visualizes the critic’s structured feedback and the Planner Agent revises the recipe in real time. The process continues until the recipe is approved or reaches a fail-safe termination. The demo allows visitors to compare one-pass generation with PCAR, inspect each remediation step, and test increasingly complex profiles. It demonstrates how agentic safety verification can convert an unsafe but plausible recipe into a practical, image-grounded, and profile-compatible alternative.

Multimodal Emotion Recognition (MER) integrates visual and audio cues to more accurately identify human emotions. Most existing work uses traditional unimodal backbones without systematically analyzing encoder impact. In this work, we evaluate the performance of unimodal encoders and an attention gate fusion-based multimodal framework. The best model (DenseNet + HuBERT) outperforms unimodal baselines and achieves an accuracy of 84.36%. Future work will explore transformer-based encoders and additional modalities to improve recognition robustness further.

Multi-modal object Re-Identification (ReID) seeks to retrieve objects by exploiting complementary cues from different modalities. However, most existing approaches emphasize feature fusion while indiscriminately aggregating patch tokens, which mixes identity-relevant regions with background noise. Moreover, they typically apply uniform fusion weights and thus cannot adapt to modality reliability that varies with imaging conditions. To address these issues, we propose CoRe-Net, a multi-modal feature learning framework that couples stable consensus-based token selection with reciprocal reliability-guided fusion. Specifically, we introduce a Cross-Modal Consensus Selection (CMCS) module that filters modality-specific noise by selecting patches that remain consistently discriminative across modalities via stability-aware scoring and cross-modal attention. We further propose a Reciprocal Reliability Modulation (RRM) module that adapts features to modality-specific distributions and performs reliability-aware fusion by jointly accounting for intrinsic evidence strength and inter-modal consistency. Together, CMCS and RRM form a coherent pipeline that suppresses modality-specific noise and enables reliability-aware aggregation. This design yields robust multi-modal representations in challenging scenarios. Extensive experiments on three multi-modal object ReID benchmarks validate the effectiveness of CoRe-Net.

Audio-visual quality assessment (AVQA) is essential for streaming, teleconferencing, and immersive media. In realistic streaming scenarios, distortions are often asymmetric, where one modality may be severely degraded while the other remains clean. Still, most contemporary AVQA metrics treat audio and video as equally reliable, causing confidence-unaware fusion to emphasize unreliable signals. This paper proposes MCM-AVQA, a multimodal confidence-aware AVQA framework that explicitly estimates modality-specific confidence and injects it into a dedicated Audio–Visual Mixer for cross-modal attention. The Audio–Visual Mixer utilizes frame-level, confidence-guided channel attention to gate fusion, modulating feature interaction between modalities so that high-confidence streams dominate while unreliable inputs are suppressed, preserving temporal degradation patterns. A multi-head visual confidence estimator turns frame-level artifact probabilities into temporally smoothed, clip-level visual confidence scores, while an audio confidence module derives confidence from speech-quality cues without requiring a clean reference. Experiments on multiple AVQA benchmarks show that MCM-AVQA, and specifically its confidence-guided Audio–Visual Mixer, improve correlation with human mean opinion scores and yield more stable behavior under real-world asymmetric audio-visual distortions.

In recent years, long-form video captioning has been an important task for indexing, accessibility, and downstream video analytics. However, it becomes challenging when videos are untrimmed and extend over minutes or hours. In these settings, visual evidence is often incomplete: objects are occluded, viewpoints shift, and crucial semantics may be implied rather than directly observable. While recent hierarchical captioning models improve temporal coverage, they still rely primarily on the available visual stream and can produce captions that omit key entities or relations. External knowledge retrieved from training data or knowledge bases can improve the visual signal, but existing integrations are often shallow and are either added late at the decoder input or applied without accounting for retrieval reliability making them sensitive to noisy evidence and limiting cross-modal interaction. To solve this issue, we propose a video-aware knowledge fusion approach that integrates retrieved knowledge at the feature level prior to generation. The method combines (i) a contextual gating mechanism that modulates knowledge strength using retrieval confidence signals together with pooled visual context, and (ii) a cross-modal fusion transformer that refines the joint sequence of video and gated-knowledge query tokens through self-attention before decoding. We evaluate on challenging datasets, such as YouCook2 and Ego4D‑HCap, and observe measurable improvements over strong baselines, indicating feature-level knowledge fusion can enhance semantic coverage in long-form caption generation.

Egocentric activity recognition can benefit from integrating multiple sensory cues, yet practical settings are challenged by modality gaps: real audio is not always available and can be corrupted by background noise, and real depth is rarely available at scale. To address these limitations, we augment video with synthetic audio and synthetic depth, and study their impact together with three native modalities: visual appearance, motion, and real audio. We then propose Appearance-Routed Fusion (ARF), a routing-based fusion module that performs sample-specific modality weighting. ARF uses appearance-conditioned cross-attention to capture interaction-centric cross-modal cues and predicts per-modality weights to suppress weak or noisy signals before forming the fused representation for classification. Experiments on Epic-Kitchens and Ego4D demonstrate that synthetic modalities help mitigate missing or noisy inputs and that ARF consistently improves performance over standard fusion baselines.

Object hallucination remains a critical challenge in multimodal large language models (MLLMs). While contrastive decoding methods effectively suppress hallucination at inference time, they require repeated multi-pass comparison and do not explicitly model hallucination-sensitive representations. We propose COHESION that reinterprets contrastive decoding as a representation-level hallucination probing mechanism. COHESION estimates hallucination-sensitive subspaces directly from contrastive visual views and integrates complementary signals from heterogeneous probes through consensus estimation. The resulting consensus subspace is applied through representation-level editing, enabling single-pass inference without requiring hallucination-specific datasets or LLM-generated hallucinated/faithful response pairs. Experiments on CHAIR, OPOPE, and MME show that COHESION improves hallucination mitigation across multiple MLLMs while preserving general multimodal ability and reducing inference overhead. Code: https://github.com/aiiu-lab/Cohesion

Cell segmentation in single-shot quantitative phase microscopy (ssQPM) faces challenges from traditional thresholding methods that are sensitive to noise and cell density,while deep learning approaches using simple channel concatenation fail to exploit the complementary nature of polarized intensity images and phase maps. We introduce DM-QPMNet, a dual-encoder network that treats these as distinct modalities with separate encoding streams. Our architecture fuses modality-specific features at intermediate depth via multi-head attention, enabling polarized edge and texture representations to selectively integrate complementary phase information. This content-aware fusion preserves training stability while adding principled multi-modal integration through dual-source skip connections and per-modality normalization at minimal overhead. Our approach demonstrates substantial improvements over monolithic concatenation and single-modality baselines, showing that modality-specific encoding with learnable fusion effectively exploits ssQPM’s simultaneous capture of complementary illumination and phase cues for robust cell segmentation.

Visible-Thermal (RGB-T) object tracking benefits from complementary properties of both modalities, yet remains challenging due to the scarcity of annotated multi-modal data and the domain gap between heterogeneous modalities. Existing RGB-T trackers predominantly focus on explicit inter-modal fusion through prompts or adapters, while largely neglecting intra-modal adaptation of pre-trained RGB backbone to the thermal domain. In this work, we propose Hybrid Adapter Tracker (HAT), an efficient RGB-T tracker that integrates both intra-modal adaptation and inter-modal fusion in a unified architecture. Specifically, we introduce an Intra-modal Low-Rank Adapter (IM-LoRA) to adapt the frozen backbone with minimal trainable parameters, enabling effective domain alignment within each modality. Furthermore, we design a Conditional Inter-Modal Adapter (CIMA) that performs effective cross-modal interaction via conditional feature modulation, thereby achieving dynamic multi-modal feature interaction. Extensive experiments on existing benchmarks demonstrate that HAT achieves superior tracking performance. Compared to state-of-the-art trackers, HAT achieves a performance gain of 1.4% and 0.5% in terms of success rate on LasHeR and RGBT234, respectively.

Vision-language models (VLMs) have achieved impressive performance; however, the large number of visual tokens incurs substantial computational overhead compared to text tokens. Existing attention-based visual token pruning methods reduce this cost by retaining only high-attention tokens. Although effective, these approaches typically aggregate attention across all heads, overlook spatial structure, and ignore information in discarded tokens, leading to unreliable token selection and potential loss of important semantic content. To address these issues, we propose HSSP, a visual token pruning framework that improves selection reliability through head-selective attention scoring, preserves spatial completeness via region-level constraints, and maintains semantic diversity by recycling pruned tokens using a maximum-minimum distance strategy. Experiments across multiple tasks demonstrate that our method effectively reduces computational overhead while preserving, or even improving, model performance compared to existing pruning methods.

Vision-Language Models (VLMs) achieve strong performance across a wide range of tasks but incur high computational costs. Existing VLM compression methods primarily target the pretraining stage, requiring large-scale datasets and often retaining information irrelevant to downstream tasks. We propose TaperVLMs, an efficient pruning framework that compresses pretrained VLMs directly for downstream tasks. Grounded in Information Bottleneck (IB) theory, TaperVLMs prunes components based on their relevance to the target task, while freezing pretrained weights and introducing trainable LoRA modules to ensure efficient adaptation. Leveraging IB and LoRA parameters, TaperVLMs further integrates an efficient knowledge distillation mechanism to retain essential pretraining information during pruning. Experiments on text-to-image and image-to-text retrieval demonstrate that TaperVLMs achieves substantial compression with minimal loss in accuracy, even in data-limited settings.

Multi-modal remote sensing segmentation benefits from complementary cues (e.g., RGB, SAR, DSM) but often degrades under cross-modal mis-registration, large intra-class variation, and rotation/scale changes across acquisitions. HPA-Seg improves robustness by coupling three lightweight components: Correlation-Constrained Cross-modal Gating (C3-Gate), a dependence-regularized sparse cross-attention that suppresses spurious fusion; Cross-granularity Prototype Memory with Distribution-Robust updates (CPM-DR), which maintains class/subclass/morphology prototypes via reliability-weighted exponential moving average updates; and an Equivariant Multi-Group Context module (EMGC) that enforces discrete rotation/scale consistency at mid/high layers. Experiments on ISPRS Vaihingen/Potsdam and SEN12MS show consistent mIoU gains over strong fusion baselines, stronger robustness under controlled geometric perturbations and synthetic mis-registration, and statistically significant improvements under paired t-tests across multiple seeds.

Vision Language models (VLMs) have demonstrated strong performance across a wide range of benchmarks, yet they often suffer from modality dominance, where predictions rely disproportionately on a single modality. Prior approaches primarily address this issue by steering model's attention allocation, implicitly assuming that all modalities provide sufficient information. However, attention only determines where the model focuses, and cannot enrich information that is missing or ambiguous. In the real world, input modalities often differ in information density and their signal-to-noise ratios. In such cases, simply adjusting model's attention does not resolve the underlying lack of information. In this paper, we propose \textsc{MoIR}: \textit{Multi-modal Information Router}, an information-level fusion method that explicitly reduces information disparity prior to fusion. \textsc{MoIR} identifies less informative tokens and routes complementary information from a stronger modality, constructing information-dense token representations before they are processed by a large language model. By modifying information availability, \textsc{MoIR} enables reliable shifts in modality dominance, even when one modality is degraded. We evaluate \textsc{MoIR} on three widely used multi-modal benchmarks across multiple model backbones. Experimental results show that \textsc{MoIR} consistently demonstrates more balanced modality contribution, and improves robustness and downstream performance, particularly even under modality degradation. These findings demonstrate that explicitly modifying cross-modal information is an effective and complementary strategy for mitigating modality dominance in multi-modal reasoning models.

Gloss-free Sign Language Translation (SLT) translates sign language videos into spoken-language sentences without gloss annotations, avoiding costly labeling but requiring fine-grained modeling of hands, body, and facial cues. Existing methods often use single-modality or weakly fused features, limiting performance. We propose ViPo-MLLM, a framework that integrates spatio-temporal RGB and human pose features. Dedicated encoders model intra-modal dynamics and cross-modal attention captures long-range dependencies. The fused representation is conditioned with a structured prompt and processed by an LLM trained with contrastive and language modeling objectives. The proposed model was evaluated on the PHOENIX14T and CSL-Daily datasets and achieved new state-of-the-art results on both datasets. Moreover, the ViPo-MLLM model attained competitive performance compared to gloss-based recognition approaches, confirming the effectiveness of the proposed pose cues and cross-modal attention mechanisms. Code is available at https://github.com/gufranSabri/ViPo-SLT.

Vision-language models (VLMs) encode images into a large number of visual tokens, leading to substantial computational overhead during inference. Training-free visual token pruning offers a simple and practical solution to improve efficiency, but designing effective pruning strategies under aggressive token reduction remains challenging. In particular, existing methods often struggle to simultaneously preserve informative visual content and control redundancy among retained tokens. In this work, we propose ID-Pruner (Importance–Diversity Pruner), a training-free visual token pruning framework that explicitly disentangles importance preservation from redundancy control in a coarse-to-fine manner. ID-Pruner first performs importance filtering using an image-intrinsic signal derived from masked autoencoding to remove globally predictable visual tokens. It then enforces diversity on the remaining candidates via K-Center selection in the multimodal embedding space used by the language model, ensuring sufficient coverage under strict token budgets. The proposed method introduces no additional learned components and remains fully training-free. Extensive experiments on multiple VLMs, including LLaVA-1.5-7B, LLaVA-NeXT-7B, Qwen-VL-7B, and InternVL-Chat-13B, demonstrate that ID-Pruner consistently outperforms existing training-free pruning methods across a wide range of benchmarks and pruning ratios, particularly under aggressive token reduction.

Accurately tracking difficult-to-heal wounds is vital for effective treatment, yet manual assessment remains slow and subjective. To bridge this gap, we propose a deep learning pipeline using Generative Adversarial Networks (GANs) designed to automate wound segmentation despite real-world challenges including data heterogeneity and class imbalance. Our approach was validated on the WoundsDB benchmark dataset, comparing standard RGB imaging against a multi-modal strategy that integrates color channels with thermography. The multi-modal GANs outperformed other deep learning architectures, achieving a mean sensitivity of 0.515 and specificity of 0.994. By leveraging non-invasive, multi-modal data, this method paves the way for objective, high-precision wound monitoring and personalized clinical care.

Ship hull corrosion inspection remains challenging because corrosion is often hidden beneath paint layers, limiting the effectiveness of visual inspection methods. This work presents a robotic-arm-based multimodal framework for corrosion segmentation using multispectral and thermal cameras. In the thermal modality, active thermography is performed using four halogen lamps to enhance the detection of subsurface defects. A dataset of 800 multimodal acquisitions from real ship hulls is collected, calibrated, aligned, and manually annotated into corrosion and non-corrosion regions. The proposed framework explores CNN-, transformer-, and foundation-model-based segmentation architectures combined with cross-attention fusion strategies to improve corrosion characterization under challenging industrial inspection conditions.

Low-light image enhancement is challenging due to complex degradations, including amplified noise, artifacts, and color distortion. While Retinex-based deep learning methods have achieved promising results, they primarily rely on single-modality RGB information. We propose M2Retinexformer (Multi-Modal Retinexformer), a novel framework that extends Retinexformer by incorporating depth cues, luminance priors, and semantic features within a progressive refinement pipeline. Depth provides geometric context that is invariant to lighting variations, while luminance and semantic features offer explicit guidance on brightness distribution and scene understanding. Modalities are extracted at multiple scales and fused through cross-attention, with adaptive gating dynamically balancing illumination-guided self-attention and cross-attention based on the reliability of auxiliary cues. Evaluations on the LOL, SID, SMID, and SDSD benchmarks demonstrate overall improvements over Retinexformer and recent state-of-the-art methods. Code and pretrained weights are available at https://github.com/YoussefAboelwafa/M2Retinexformer .

Poor lighting conditions adversely affect image quality, creating significant challenges for image editing and visualization. Many existing enhancement methods emphasize complex models while neglecting the intrinsic information contained within low-light images. In this work, we propose the Self-Information Mining (SIMI) network, an innovative unsupervised framework that decomposes low-light images into multiple components based on bit-plane decomposition. Our approach allows to mine intrinsic information without relying on external data. This not only accelerates model convergence but also improves performance and reduces computational overhead. Taking a step toward real-world applicability, we focus on a zero-reference setting within an unsupervised framework. Experiments conducted on standard benchmarks demonstrate that SIMI achieves state-of-the-art performance. The code will be publicly released.

Reconstructing relightable 3D scenes from in-the-wild photo collections is highly ill-posed due to uncontrolled illumination and inconsistent capture conditions. Existing inverse rendering-based frameworks capture complex illumination with environment maps, which remains fundamentally under-determined and unstable for in-the-wild setting. To this end, we introduce Multi-Light Relightable Gaussian Splatting (MR-GS) - a physically grounded relighting framework from in-the-wild images - that approximates illumination with multiple light sources and represents scene attributes with the simple yet effective Phong Reflectance model. Additional physics-aware regularizations are introduced to ensure correct and realistic material-lighting decomposition. Experiments on the NeRF-OSR benchmark demonstrate that MR-GS outperforms state-of-the-art methods in both quantitative and qualitative evaluations, while achieving a significantly higher rendering speed.

Color-polarization imaging using a color-polarization filter array (CPFA) sensor captures both texture (color intensity) and physical (polarization) information of the scene in a single shot, enabling various applications in computer vision. However, the raw mosaic output from a CPFA sensor often suffers from severe noise and resolution loss, especially under low-light conditions. Existing methods generally focus on either denoising or demosaicking tasks, failing to capture the coupling between them and neglecting shared low-level features. In this paper, we propose a color-polarization denoising and demosaicking network (CPDDNet), which is a joint framework that performs noise removal and CPFA interpolation using a feature fusion module that retains the features from the CPFA raw data at both the denoising and the demosaicking stages. Experimental results demonstrate that CPDDNet significantly enhances image quality and polarization parameter accuracy, outperforming existing approaches on a real dataset. The source code of CPDDNet will be publicly available upon the paper acceptance.

Low-light video enhancement (LLVE) is challenging due to noise, low contrast, and color degradation. While learning-based methods enable fast inference, they often fail under heavy real-world noise because they do not sufficiently exploit long-term temporal cues. We propose DWTA-Net, a novel deep-learning recurrent LLVE framework with a recurrent design. DWTA-Net adopts an integrated two-stage architecture: Stage I restores local structure and color via multi-frame alignment for temporally consistent Mamba-based enhancement, while Stage II performs recurrent refinement using a novel dynamic weight-based temporal aggregation guided by optical flow, functioning as a recurrent denoiser that adapts to motion. We further introduce a texture-adaptive loss that preserves fine details in textured regions while suppressing noise in homogeneous areas. Experiments on real-world low-light footage show that DWTA-Net achieves stronger noise suppression and fewer artifacts, delivering superior visual quality compared with state-of-the-art methods.

In practical applications, low-light images are often compressed for efficient storage and transmission. Most existing methods disregard compression artifacts removal or hardly establish a unified framework for joint task enhancement of low-light images with varying compression qualities. To address this problem, we propose an efficient hybrid priors-guided network (HPGN) that enhances compressed low-light images by integrating both compression and illumination priors. Our approach fully utilizes the JPEG quality factor (QF) and DCT quantization matrix (QM) to guide the design of efficient plug-and-play modules for joint tasks. Additionally, we employ a random QF generation strategy to guide model training, enabling a single model to enhance low-light images with different compression levels. Experimental results demonstrate the superiority of our proposed method.

Under low-light conditions, degradation and subsequent amplification are often applied uniformly across channels, causing noise to exhibit strong cross-channel correlation in feature space. As a result, correlation-driven attention may preserve such noise, leading to over-smoothing or noise amplification. To address this issue, we propose Differential-Regulated Attention (DRA), which suppresses cross-channel correlated noise by emphasizing differential responses and thus preserving discriminative structural information for robust enhancement. Moreover, we introduce a Feature Fusion Bridge Module (FFBM) to promote semantic consistency between the encoder and decoder for structure-preserving texture restoration. Furthermore, we employ an Asymmetric Depthwise Convolution (ADWC) module to enhance directional structural cues, providing reliable priors that support DRA in suppressing correlated noise. Extensive experiments on three public datasets demonstrate that DiRA-Net consistently outperforms existing approaches by effectively suppressing cross-channel correlated noise and preserving discriminative structures under severe low-light conditions.

Low-light image enhancement (LLIE) often struggles with noise and blurred textures due to the intrinsic coupling of color and intensity in standard color spaces. We propose TaiChiNet, a novel dual-branch CNN-Transformer architecture that effectively addresses this by leveraging color space factorization. TaiChiNet decouples image intensity and color into dual branches, enabling multiscale deep interactions through dense hybrid cross-fusion modules. The core TaiChi Block employs channel splitting and dense connections to progressively learn multiscale residual maps, stabilizing regional features while preserving crisp textures. By integrating cross-attention and convolution, the model robustly captures both long-range contextual dependencies and local biases. Extensive experiments demonstrate that TaiChiNet achieves state-of-the-art performance on paired benchmarks and superior visual quality on unpaired datasets with high parameter efficiency. Code is available at https://github.com/NoahHU901/TaiChiNet.

This paper presents a novel training-free low-light image enhancement framework that combines sparse representations with multi-scale fusion. Multiple exposure candidates are generated patch-wise using sparse coding under controlled sparsity constraints, followed by patch aggregation and denoising to mitigate noise amplification. The enhanced result is obtained via cumulative aggregation and a final multi-scale fusion stage designed to balance detail preservation and global illumination consistency. The proposed method is mathematically explainable, transparent and interpretable, avoiding the data and computational requirements of learning-based pipelines, which is beneficial for hardware-constrained and data-scarce applications. Evaluations on LOL-v1 and LOL-v2 (real and synthetic) show that the proposed approach achieves competitive or superior performance across standard full-reference metrics (e.g., PSNR/SSIM/MS-SSIM) and widely used no-reference perceptual metrics (e.g., BRISQUE/NIQE/PIQE).

Severe image degradation under low-light nighttime conditions constitutes a core bottleneck preventing all-day applications for UAV-based single object tracking. Existing image enhancement methods often struggle to distinguish between target and background regions, which can easily lead to amplified background noise or compromise target features. To overcome this limitation, we propose TAE, a target-aware low-light enhancement framework tailored for nighttime object tracking. Guided explicitly by weak supervisory signals from tracking bounding boxes, the framework performs region-aware enhancement to ensure operations focus on the target area. It further adopts an adaptive RGB multi-curve fusion mechanism to achieve refined modeling and adaptive adjustment across different regions. To facilitate research in this domain, we also contribute DarkSOT, a new benchmark for nighttime UAV tracking, comprising 268 sequences across 9 target categories. Experimental results on the DarkSOT and UAVDark135 demonstrate that TAE significantly improves tracking performance in low-light nighttime scenarios, exhibiting strong robustness and generalization. The DarkSOT dataset is available at https://github.com/Fu0511/DarkSOT-Dataset.

Low-light image enhancement (LLIE) is essential for improving visibility in challenging illumination conditions. Existing methods often apply uniform adjustments or rely solely on spatial or frequency cues, leading to exposure imbalance, detail loss, and limited generalization. To overcome these issues, we propose TriFIRNet, a tri-stage frequency domain interactive restoration network for adaptive LLIE. Specifically, the framework begins with Retinex-based pre-enhancement to stabilize illumination, followed by a Wavelet-based module to enhance brightness and refine local textures, and a Fourier-based module to enforce global consistency and suppress artifacts through spatial–frequency interaction. Additionally, a pixel-wise luminance-difference guided mechanism adaptively modulates enhancement strength across regions with varying illumination, effectively balancing brightness. Extensive experiments on multiple benchmarks demonstrate that TriFIRNet outperforms state-of-the-art LLIE methods in both quantitative metrics and visual quality.

This paper proposes a neural network for low light image enhancement (LLIE) based on retinex theory to make LLIE robust for various dynamic range scenes. The retinex theory is an image formulation model inspired by a human color perception hypothesis, where a low light image is decomposed into intrinsic color context (i.e., reflectance map) and scene-dependent illumination (i.e., illumination map). Due to non-uniqueness of its decomposition, existing retinex-based LLIE methods often fail to achieve stable decomposition, which lead to over-enhancement. Typically, they are sensitive to the dynamic ranges that vary in different lighting conditions. To tackle this issue, we propose WREN: An LLIE neural network with double U-Net-like structures. WREN consists of two U-Net-like sub-networks. The first network has one encoder and two decoders that decompose an input image into the reflectance and illumination maps. The second network with a customized Transformer block between an encoder and a decoder only enhances the illumination map obtained from the first network: This completely follows the assumption of the retinex theory. Finally, the enhanced illumination map is recombined with the reflectance map. The network is trained end-to-end with a scale-invariant loss function, which gives robustness against the illumination scaling. Numerical results show that our method achieves the state-of-the-art performance across multiple datasets. Our code is available online.

Low-light image enhancement aims to recover visual details from images captured under insufficient illumination. Existing methods always overlook global illumination correction and local detail preservation, and show limited performance when enhancing dark regions and suppressing highlight regions. To address these issues, we propose a novel low-light enhancement network, termed CETNet. The input image is first decomposed into I-channel and HV-channel to reduce color distortion and artifacts, followed by a multi-scale cross-update mechanism to maintain semantic consistency between luminance and structural information. Light Modulator is then introduced to adaptively correct global illumination, while Dual Contrast Enhancement Module separately enhances dark regions and suppresses highlight regions. Extensive experiments on the LOL, LoLI-Street, and RE2L datasets demonstrate that CETNet outperforms existing methods in PSNR and SSIM, and user study further confirms its superior visual quality.

Low-light image enhancement remains difficult under severe real-world degradations, where uniform treatment of color channels and simplistic fusion in existing methods cause color shifts and exposure artifacts. We propose a low-light image enhancement framework with a prior-guided diffusion model that separately models luminance and chrominance noises with structured contrast priors to preserve illumination with structural details. Multi-path image fusion aggregates complementary enhancements, and multi-feature contrastive learning aligns illumination, noise level, and content with normal-light references. Experiments on benchmark datasets show that our framework outperforms the state-of-the-art methods by a very large margin across metrics, reducing color distortion and preserving illumination and structural details.

Low-light image enhancement remains challenging due to noise amplification, color distortion, and structural detail loss, limiting its applications in surveillance, autonomous navigation, and medical imaging. We propose a novel depth-guided, flash-simulated framework that estimates scene depth to apply spatially-adaptive virtual flash illumination, combined with transformer-based multi-scale feature fusion. Despite recent advancements, most methods apply uniform processing without considering the variation of light attenuation and sensor noise with scene depth. Our pipeline includes: (1) a low-light depth estimation network supervised by pretrained MiDaS, (2) differentiable depth-based clustering via soft K-means, (3) adaptive flash simulation per depth cluster, (4) dedicated flash feature encoder, and (5) transformer fusion blocks within the main enhancement network. Extensive experiments on LOL demonstrate state-of-the-art PSNR/SSIM and competitive visual quality against recent methods.

Low-light image enhancement methods are often evaluated under benchmarks with limited control over illumination conditions. This makes it difficult to study how specific lighting factors affect restoration performance in a reproducible way. To address this limitation, we introduce CID-LIE (Controlled Illumination Dataset for Low-Light Image Enhancement), a dataset built around physically controlled lighting protocols. CID-LIE contains 1250 paired RGB images from 125 indoor scenes acquired under systematically varied illumination settings. The dataset supports controlled variation of lighting dimensions, including lamp type, illumination intensity, and number of active light sources. Each low-light image is paired with a reference captured under standardized full illumination. Unlike datasets based on synthetic degradation, CID-LIE generates low-light conditions physically during acquisition using a professional lighting system. Benchmarking results show that controlled illumination reveals distinct restoration behaviors and robustness limitations that conventional datasets often fail to expose. Overall, CID-LIE provides a reproducible and physically grounded benchmark for illumination-aware low-light image enhancement.

The restoration and enhancement of images and video is a challenging inverse problem as insufficient illumination and predominant noise are deeply coupled. Treating them sequentially often leads to noise amplification. Unfolding networks, such as URetinex-Net++, jointly optimize these tasks with an iterative deep learning model. However, they struggle to disentangle noise from structural details, propagating some residual artefacts across the unfolding stages. Self-attention mechanisms capture long-range dependencies but bring a prohibitive quadratic computational cost. In this context the challenge lies in creating an unfolding framework performing frequency-aware noise separation in linear time.

Nuclear fuel images acquired during post-irradiation examinations in hot-cell environments with 'end-of-life' cameras are strongly degraded by noise, low contrast, illumination heterogeneity, and acquisition artifacts. In this work, we propose a novel joint restoration framework based on a dual-branch SWT-WKAN Unet architecture dedicated to nuclear fuel imaging. The proposed method introduces two main contributions.  First, we extend the WKAN Unet architecture by integrating a shared encoder stem initialized with Stationary Wavelet Transform (SWT) filters, enabling better preservation of structural and  high-frequency details during feature extraction. The architecture is then divided into two complementary branches dedicated to image enhancement and segmentation tasks. Second, we propose an original training strategy that addresses the absence of paired ground-truth data in post-irradiation imaging. Instead of relying on manually annotated references, the enhancement branch is supervised using 'beginning-of-life' images restored by a non-AI classical pipeline combining percentile normalization, CLAHE, black-hat filtering, Telea inpainting, and unsharp masking. Experiments conducted on a dataset of 64 high-dynamic-range nuclear fuel images show that the proposed method reduces estimated noise while improving visibility and preserving relevant structures. Preliminary results demonstrate the complementarity between classical restoration and deep-learning-based enhancement approaches.

Multi-Object Tracking (MOT) in dynamic environments relies on robust temporal reasoning to maintain consistent object identities over time. Transformer-based end-to-end MOT models achieve strong performance by explicitly modeling temporal dependencies, yet training them requires extensive bounding-box and identity annotations. Given the high labeling cost and strong redundancy in videos, Active Learning (AL) is an effective approach to improve annotation efficiency. However, existing AL methods for MOT primarily operate at the frame level, which is structurally misaligned with modern end-to-end trackers whose inference and training rely on multi-frame clips. To bridge this gap, we formulate clip-level active learning and propose Clip-level Uncertainty and Temporal-aware Active Learning (CUTAL). In contrast to frame-based approaches, CUTAL scores each clip using uncertainty metrics derived from multi-frame predictions to capture inter-frame correspondence ambiguities, while enforcing temporal diversity to select an informative and non-redundant subset. Experiments show that CUTAL achieves stronger overall performance than baselines at the same label budgets across MeMOTR and SambaMOTR. Notably, CUTAL achieves performance comparable to full supervision for MeMOTR on both datasets using only 50% of the labeled training data.

Flood forecasting from satellite imagery is often challenged by severe class imbalance and label noise. The Global Flood Forecasting (GFF) benchmark provides a standardized framework for end-to-end inundation prediction, but its data characteristics raise questions about the choice of training objectives. In this work, we provide a systematic and reproducible empirical evaluation of commonly used loss functions on the GFF benchmark within a fixed forecasting pipeline. Experimental results show that hybrid region-based losses reduce false negatives on the full test set, yielding modest but stable F1-score improvements under noisy and highly imbalanced conditions. On a cleaner and hand-annotated subset, flood prediction performance differences across loss functions diminish and no universal ordering emerges. Overall, this study clarifies how loss selection shapes the precision–recall tradeoff on GFF and provides practical guidance for selecting imbalance-aware training objectives.

Diffusion-based video editing has made significant progress; however achieving precise and temporally consistent object-level control, especially in multi-object scenarios, remains challenging due to attention leakage, identity drift, and unstable temporal dynamics. In this work, we propose IDAG-Edit, a unified training-free framework for fine-grained multi-object video editing with strong temporal consistency. The framework adopts Layout-guided Attention Modulation to facilitate coherent multi-object editing, while Instance-level Masks are introduced to preserve individual object identity and enforce localized attention within each object region, thereby enabling fine-grained, object-level editing. Extensive qualitative and quantitative evaluations demonstrate that our method improves temporal stability and multi-object controllability over state-of-the-art video editing approaches.

Recent advances in generative AI have democratized video creation at scale. AI-generated videos, including partially manipulated clips across visual and audio channels, pose escalating risks of semantic distortion and misuse, which motivates the need for reliable detection tools. Most existing AI-generated video detectors remain limited by single- or partial-modality of data modeling and the lack of fine-grained temporal forgery localization. To address these challenges, our primary novelty introduces a core architecture that jointly integrates an LMM semantic branch with a spatio-temporal (ST) visual branch and a multi-scale partial-spoof (PS) audio branch. This multi-modal approach enables simultaneous detection and fine-grained temporal localization of partially manipulated AI-generated video forgeries. Extensive experiments show that this approach outperforms existing state-of-the-art methods.

Text-Based Person Search aims to retrieve target person images from large galleries using natural language descriptions, requiring robust cross-modal alignment and fine-grained discrimination. Existing methods often optimize global and local objectives independently, lacking explicit mechanisms to coordinate identity-level semantics with fine-grained reasoning. We propose Global--Local Representation Alignment (GLoRA), a unified framework that integrates global semantic aggregation with difficulty-aware local refinement. At the global level, GLoRA aggregates multiple textual descriptions to construct a stable identity-centric semantic space. At the local level, a coarse--fine slot-based alignment mechanism anchors fine-grained representations to global semantics, improving robustness against semantic drift. Moreover, a difficulty-aware control signal derived from global matching ambiguity adaptively regulates the contribution of local alignment, emphasizing fine-grained refinement for ambiguous samples. Extensive experiments on CUHK-PEDES, ICFG-PEDES, and RSTPReid demonstrate that GLoRA achieves competitive performance across multiple retrieval metrics.

Recently, text-to-video generation has developed rapidly due to the crucial role of video diffusion models (VDMs). Although the customization of text-to-video generation models has gained growing attention, customizing motion dynamics still poses a significant challenge. The existing methods are either limited by their inability to decouple appearance information or are prone to be affected by background variations. To address these issues, we propose a novel one-shot tuning framework, enabling it to generate videos that maintain motion consistency with a single reference video across different contexts. Specifically, our method focuses on the temporal layers in VDM and introduces a Spatial-Attention-Guided motion learning objective (SAG), which first leverages spatial attention maps extracted from the spatial transformer blocks in VDM to construct Motion-Mask that highlights foreground dynamics. Secondly, the Motion-Mask is employed to constrain the learning of motion patterns, driving the model to focus on semantically meaningful dynamics while suppressing background variations. Extensive experiments on diverse datasets demonstrate that our approach outperforms multiple state-of-the-art methods in both quantitative and qualitative evaluations.

Video anomaly detection (VAD) in nighttime traffic scenarios remains challenging due to low illumination, overexposure, motion blur, and reflection noise. To address these issues, we propose a Hybrid CNN–Swin Temporal Attention Network (HSTA) for unsupervised nighttime video anomaly detection based on future-frame prediction. HSTA integrates convolutional and transformer-based branches to jointly capture local texture details and global contextual dependencies. A Random Temporal Shift (RTS) module enhances temporal robustness, while a Temporal-Aware Channel Attention Block (TACAB) emphasizes motion-sensitive channels for refined spatial-temporal modeling. Experimental results on real nighttime traffic data demonstrate the effectiveness of HSTA, achieving AUC scores of 76.4\%, which is 1.5\% higher than the state-of-the-art method. Furthermore, additional evaluations on three daytime benchmarks indicate that HSTA exhibits promising generalization and robustness across different lighting conditions.

Multi-Object Tracking and Segmentation is a crucial vision task that combines detection, segmentation, and tracking for a deeper understanding of dynamic scenes. With the rapid growth of foundation models, Segment Anything Model 2 (SAM2) has been proposed to provide a zero-shot image segmentation. In this paper, we present MASAM, a memory-augmented framework that extends SAM2 to multi-object tracking and segmentation without the need for fine-tuning. We propose Motion-guided Object Memory (MOM) module that leverages motion information for robust object tracking. With the proposed object-level design and feature management strategy, our method possesses high identity consistency and supports an eidetic-memory-like ability. To avoid propagation of false detections, we introduce a short-term memory mechanism to suppress unreliable detections by using a finite-state transition system. Our experimental results demonstrate that MASAM achieves state-of-the-art performance on KITTI MOTS benchmark and the lowest ID switches on both MOTSChallenge and KITTI MOTS datasets. In addition, our method achieves SOTA MOT performance on DanceTrack benchmark.

Adapting RGB-pretrained object detection models to the infrared (IR) domain remains challenging due to the inherent modality gap and the limited availability of paired RGB--IR data. While existing approaches aim to mitigate this gap, they often suffer from poor inter-class separability and increased false positives, primarily due to the absence of explicit semantic guidance. To overcome these limitations, we propose a CLIP-Driven Semantic Alignment (CDSA) framework. Our approach leverages the text encoder of CLIP as a set of semantic anchors, enabling RoI-level alignment of IR features with an RGB-optimized semantic space without requiring paired RGB source images. In addition, we introduce a False Positive Filtering (FPF) that suppresses spatially misaligned detections during training, preventing unreliable regions from corrupting the semantic alignment process. Extensive experiments on the FLIR and DroneVehicle datasets demonstrate the effectiveness and superiority of our proposed method.

Synthesizing images of objects not seen during training remains a major challenge in computer vision. In this paper,we introduce a method for synthesizing images of objects not seen during training from text prompts (zero-shot text-to-image synthesis), utilizing various attributes of objects as auxiliary information. This is motivated by the fact that a superior correlation between the text input and object attributes may facilitate superior image synthesis. We synthesize high fidelity images using a latent diffusion model, requiring only thousands of training samples compared to the billions of samples required by existing large-scale models. To that end, we use salient visual features of images extracted by a vector-quantized variational autoencoder as input to the diffusion model for training. To ensure the quality of the synthetic images, we introduce a novel classifier that produces attributes from images. An attribute-based self-consistency loss for the classifier helps in improving the synthetic images.Quantitative and qualitative evaluations demonstrate the effectiveness of the proposed method. Our results also highlight the complementary nature of textual and attribute-based representations for an efficient zero-shot image synthesis. This opens a new avenue for generative models, especially in problems with limited visual examples.

Multi-Object Tracking (MOT) in complex dynamic scenes remains challenging due to frequent occlusions, non-linear motion, and appearance drift, which make motion and appearance cues highly non-stationary in reliability. Most existing trackers rely on fixed heuristics or static cue fusion, and standard temporal modeling often over-smooths short-term discriminative dynamics, leading to identity switches under sudden maneuvers. To address these issues, we propose CG-Track, a robust MOT framework that adaptively regulates information flow along both temporal and modality dimensions. Specifically, a Temporal Feature Enhancement Network (TFEN) captures multi-scale temporal context while selectively emphasizing salient short-term fragments to suppress noisy history. In addition, a Cue-Gated Attention Fusion (CGAF) module dynamically re-weights appearance and motion features based on their instantaneous reliability under local spatial context, mitigating failures caused by occlusion or aggressive motion. Extensive experiments on DanceTrack and SportsMOT demonstrate consistent improvements on association-oriented metrics---achieving 67.0 and 81.2 Identification F1 scores (IDF1), respectively---as well as enhanced robustness in crowded, fast-changing scenarios.

Robust semantic segmentation of road scenes under adverse illumination, lighting, and shadow conditions remain a core challenge for autonomous driving applications. RGB-Thermal fusion is a standard approach, yet existing methods apply static fusion strategies uniformly across all conditions, allowing modality-specific noise to propagate throughout the network. Hence, we propose CLARITY that dynamically adapts its fusion strategy to the detected scene condition. Guided by vision-language model (VLM) priors, the network learns to modulate each modality's contribution based on the illumination state while leveraging object embeddings for segmentation, rather than applying a fixed fusion policy. We further introduce two mechanisms - one which preserves valid dark-object semantics that prior noise-suppression methods incorrectly discard, and a hierarchical decoder that enforces structural consistency across scales to sharpen boundaries on thin objects. Experiments on the MFNet dataset demonstrate that CLARITY establishes a new state-of-the-art (SOTA), achieving 62.3% mIoU and 77.5% mAcc.

Cloth-changing person re-identification (CC-ReID) remains a significant challenge as traditional models often overfit to transient clothing features, weakening their capacity to identify individuals across varying clothes. To address this, we propose a framework that disentangles identity from clothing attributes within the embedding space of a fixed CLIP encoder. We leverage an Invertible Interpretation Network to break down CLIP embeddings into distinct components of identity, upper wear, lower wear, and residual factors. By utilizing training pairs that share specific factors and imposing a Gaussian prior on the latent space, the model learns to extract clothing-invariant identity features without retraining the CLIP backbone. Additionally, we also present an adaptive prompt learning module that matches text-based queries with disentangled factor representations, making text-guided retrieval more reliable. Tests on standard CC-ReID benchmarks of PRCC and LTCC show that our method performs effectively and remains efficient. This work advances interpretable and cloth invariant person re-identification by bridging visual disentanglement with language guided retrieval.

Visible–infrared person re-identification aims to retrieve the same person across time, cameras, and modalities, yet it remains challenging due to the inherent heterogeneity between modalities. Existing methods often employ a shared backbone to extract modality-shared information; however, visible and infrared images differ significantly in color and appearance attributes, and the learned modality-shared features tend to be biased toward modality-specific cues. To address this issue, we propose MoE-driven Modality-invariant Feature Learning (M2FL), which strengthens the learning of modality-shared features via multi-modal feature disentanglement. Specifically, we propose MoE-driven Modality Disentanglement Learning (MMDL) framework that applys Frequency-Spatial Guided Experts (FSGE) to leverage both Mixture-of-Experts (MoE) mechanism and frequency-spatial domain cues for disentangling modality-shared and modality-specific features. In addition, we propose a Modality-Invariant Representation Learning (MIRL) module that eliminates modality-specific information from modality-shared features, thereby generating pure, highly discriminative modality-invariant representations. Extensive experiments on SYSU-MM01 and LLCM demonstrate that M2FL achieves competitive performance against state-of-the-art methods.

Shuttlecock tracking is challenging due to the object’s small size, extreme speed, and severe motion blur. While current models like TrackNetV4 achieve high precision, they rely on 2D convolutions to implicitly handle temporal relationships, failing to capture complex spatiotemporal dynamics. We propose TrackNetV5, which introduces multi-scale R(2 + 1)D blocks for explicit spatiotemporal feature extraction and a gated residual motion prompt layer for high-quality motion attention. Furthermore, a SpatialChannelGate module adaptively fuses motion priors with visual features across multiple decoder layers. Experimental results show that TrackNetV5 achieves a state-of-the-art F1-Score of 98.47. We also present TrackNetV5-Lite, which maintains a high F1-Score of 98.06 with only 2.58M parameters, demonstrating significant potential for real-time edge deployment.

No-reference video quality assessment (NR-VQA) is challenging due to large content variability and complex real-world distortions. Existing deep NR-VQA approaches mainly rely on feedforward fusion of multi-scale spatiotemporal features, which can struggle to disentangle distortion from content. We propose a hierarchical feedback NR-VQA framework built on a generic spatiotemporal encoder and a feature pyramid network (FPN)-style hierarchy. Multi-level features are extracted from video clips and combined to capture distortions at different spatial and temporal scales. Global high-level context is then fed back to refine lower-level representations via lightweight channel-wise gating, enhancing distortion sensitivity and reducing content bias. Clip-level predictions are aggregated to obtain video quality scores. Experiments on multiple representative UGC-VQA databases demonstrate competitive performance, and ablation studies validate the benefits of hierarchical feedback over feedforward multi-scale baselines.

Infrared-visible (IR–VIS) multispectral object detection is robust under adverse illumination, but deployment is often hindered by inter-modal misalignment from sensor discrepancies and calibration errors. Most existing approaches rely on explicit pre-alignment or fusion strategies sensitive to spatial misregistration. We reinterpret misalignment as an end-to-end representation learning problem rather than a preprocessing alignment task. We propose an end-to-end framework that implicitly learns cross-modal correspondence via deformable cross-modal sampling. Without enforcing explicit geometric alignment, the network adaptively samples semantically corresponding regions across misaligned modalities, allowing misalignmentinvariant detection representations to emerge from the detection objective. The learned sampling offsets serve as task-driven latent correspondence variables jointly optimized with detection. To validate this formulation, we perform targeted ablations against standard fusion baselines and analyze learned offsets under controlled translation misalignment. Experiments on FLIR and LLVIP show that our approach maintains performance from moderate to severe misalignment while preserving real-time speed. These results indicate that modeling inter-modal misalignment as an end-to-end correspondence learning problem provides a principled and practical representation framework for robust multispectral object detection in real-world settings.

Deepfake detection systems remain vulnerable to common post-processing operations such as blur, compression, and noise. We propose FreqStack, a robustness-oriented Mixture-of-Experts framework that adaptively combines complementary global Fourier and localized wavelet representations for robust zero-shot deepfake detection. The framework fuses a frequency-aware Fourier expert and a localized wavelet expert through a reliability-weighted gating mechanism that dynamically adjusts expert contributions according to corruption conditions. We evaluate FreqStack on ForenSynths and perturbed variants containing Gaussian blur, JPEG compression, and additive Gaussian noise without corruption-aware retraining or augmentation. Experimental results show consistent robustness improvements across all perturbation categories while preserving competitive clean performance. Under combined perturbations, FreqStack improves Average Precision from 68.9 to 79.6 compared to FreqNet, corresponding to a 15.5% relative improvement. These findings suggest that adaptive fusion of complementary spectral representations mitigates spectral cue collapse more effectively than single-domain frequency detectors.

Commercial camera systems capture raw sensor data through a Bayer pattern. In typical neural network-based camera image signal processing (ISP) pipelines, the captured raw image is either demosaiced into an RGB image or repacked into four half-resolution channels (e.g., RGGB) before feature extraction. However, such pre-processing weakens the spatial and statistical structure encoded in the Bayer pattern, highlighting the need for a dedicated pre-processing method in the raw domain. To address these limitations, we propose a simple yet effective convolution operation called Bayer convolution (BayerConv), which explicitly exploits the inherent structure of the color filter array (CFA). BayerConv can be seamlessly integrated into existing ISP networks in a plug‑and‑play manner, enhancing performance under comparable computational cost. Moreover, BayerConv is applicable to generic CFA patterns, offering a practical advantage for real‑world deployment.

Sliding-window image denoising in the discrete cosine transform (DCT) domain, combining hard-thresholding with empirical Wiener shrinkage, remains attractive due to its computational efficiency, analytical transparency, and strong baseline performance. This paper revisits the classical sliding DCT Wiener denoising framework, with particular emphasis on stabilized overlap-add reconstruction to improve robustness across overlapping blocks and diverse image content. Building on this analytical backbone, we introduce a lightweight hybrid refinement strategy in which a compact convolutional neural network augments the Wiener estimate. The network predicts a residual correction together with a spatially adaptive, noise-aware gating function that selectively controls the refinement strength. This design preserves the stability and interpretability of the Wiener estimator.  Experiments on standard benchmark datasets demonstrate consistent improvements in PSNR and SSIM over strong DCT-based baselines and competitive CNN denoisers, with gains that increase at higher noise levels.

Event data is becoming increasingly useful for multiple computer vision tasks. This paper proposes the first of its kind Compressed Domain Event Data Classifier (CD-EDCL), that operates directly on a compressed domain (latent) representation of event data. The proposed solution is composed by two main, novel components: a Partial Classifier, based on the well-known EST event classifier, and a Bridge model, which adapts the event compressed latent information, created by a state-of-the-art DL-based codec (DL-JEC), to the Partial Classifier. Moreover, latent filtering is proposed to eliminate latents which degrade classification performance. Experimental results show that the proposed CD-EDCL solution improves the Top-1 classification performance by up to 27 percentage points over relevant decompressed domain classification solutions at the lowest rates, while simultaneously reducing the model weights by up to 4.4%, when compared with the original EST classifier and 14.6% when the DL-JEC decoder model is also considered.

Mobile image sensors now exceed hundreds of megapixels and employ complex color filter array (CFA) patterns, making image signal processing (ISP) challenging under strict on-sensor constraints. While conventional filtering performs adequately across most of the image, visually complex regions such as text and fine textures require more expressive solutions, such as deep convolutional-neural-networks (CNNs), whose computational demands exceed budgets for tightly constrained hardware. We present a hybrid system that enables CNN-based ISP directly on-sensor by selectively applying a compact CFA-specific network where needed, as determined by a lightweight learned classifier. The classifier and CNN are trained jointly through a novel iterative supervision strategy, allowing specialization in challenging regions while keeping the model size within several thousand parameters. We evaluate the proposed framework on green-channel demosaicing for Tetra^2 CFA sensor data, using supervision from a high-capacity teacher network adapted to the target sensor domain. Training and evaluation use RAW images from a proprietary CMOS sensor. Experimental results show improvements in complex regions, including text, edge-intersections, and high-frequency patterns. Quantitative evaluation on real-sensor data demonstrates gains over a classical demosaicing pipeline, while comparisons on public datasets adapted to Tetra^2 sampling show competitive performance relative to state-of-the-art deep demosaicing methods. These results demonstrate that high-quality neural demosaicing can be achieved under stringent hardware constraints by combining classical interpolation with selectively applied learned reconstruction. The proposed hybrid strategy provides a practical path toward deploying neural ISP components on-sensor and can be extended to additional ISP tasks such as denoising, HDR reconstruction, and bad-pixel correction.

Hyperspectral imaging is increasingly important in remote sensing, but transmitting massive data from satellites under strict power constraints is challenging. While compressive sensing with metalenses enables compact acquisition, efficient transmission requires compatibility with standard low-power codecs like JPEG. However, standard quantization tables are not necessarily optimal for compressive measurements. We propose an end-to-end framework that integrates the compressive acquisition model with a differentiable JPEG-based codec and jointly trains quantization parameters and the reconstruction network to minimize distortion. Experiments on benchmark datasets show consistent gains over a conventional decoupled baseline, yielding an average BD-rate reduction of 39.5\% in the wide-quality regime. Notably, the proposed end-to-end learning discovers non-trivial quantization tables that strategically prioritize specific Bayer channels to serve as reconstruction anchors---a non-intuitive strategy that is difficult to design manually. This approach effectively leverages the spatial structure of metalens measurements, offering a practical path for efficient hyperspectral transmission.

This paper presents a high-level synthesis (HLS) implementation of integer motion estimation (IME) on a field-programmable gate array (FPGA). The proposed IME architecture implements a hardware-oriented Multi-Resolution Motion Estimation Algorithm (MMEA), which uses a hierarchical search strategy on down-sampled frames to gradually refine motion vectors (MVs) from coarse to fine resolutions across multiple levels. The proposed strategy makes it possible to extend search range of IME to ±592 pixels and thereby capturing fast motion dynamics not typically reached by fixed-window hardware approaches. Our IME core supports all square motion partition modes ranging from 8×8 to 64×64, as well as symmetric and asymmetric motion partitioning (SMP, AMP) modes. When integrated into the open-source Kvazaar HEVC encoder, the design achieves coding efficiency gains of -1.0% against fast motion estimation (ME) algorithms on software, and a throughput rate of 4K@30fps.

Super-pixel segmentation is a fundamental preprocessing step in many computer vision and image analysis tasks, where both object boundary adherence and computational efficiency are critical. Simple Non-Iterative Clustering (SNIC) is an effective super-pixel generation method that produces compact, high-quality, boundary-preserving super-pixels through a work queue driven region growing process. However, the inherently sequential nature of its work queue prevents efficient data-parallel implementations on GPUs. We present Relaxed Simple Non-Iterative Clustering (ReSNIC) algorithm, a data-parallel version of SNIC that relaxes the strict sequential work queue dependency of the original algorithm, while preserving its segmentation behavior, and an efficient CUDA implementation of ReSNIC. Our experimental evaluations demonstrate that the proposed CUDA implementation achieves up to 10 times speedup compared to reference CPU implementations of SNIC on 1920 by 1080 sized images. Importantly, on standard image segmentation datasets the accelerated method maintains high boundary fidelity, matching the original SNIC super-pixel contours with F1-scores of up to 0.94.

Panoptic segmentation is a key enabler for robotic perception, as it unifies semantic understanding with object-level reasoning. However, modern query-based models are often computationally and memory intensive, which can make deployment difficult on power-constrained robotic platforms. We propose LiPS (Lightweight Panoptic Segmentation), an efficient query-based panoptic segmentation framework that preserves masked transformer decoding while reducing the computational cost of upstream processing. LiPS combines selective feature routing, spatial compression before multi-scale fusion, and a shallow reduced-width pixel decoder. Experiments on ADE20K and Cityscapes show that LiPS preserves most panoptic and semantic performance while reducing computation by up to 6.8x and improving embedded throughput by up to 4.5x. Additional 15 W measurements on Jetson AGX Orin confirm lower latency and GPU memory usage, making LiPS suitable for robotic perception pipelines where throughput and memory footprint are prioritized over fine instance boundaries.

The recent increase in mobile and immersive devices has created a growing demand for faster transmission of complex point cloud sequences. MPEG AI-PCC has adopted an Inter prediction with motion estimation and compensation performed in latent domain. Due to its simplified multiscale coding design, the Inter prediction is applied independently at each octree level using an identical process. Since the latent motion information was extracted via convolution-based network independently per octree level, the motion range to be covered is limited by the kernel sizes. Hence, it is impossible to fully cover large motion range especially for the last few octree levels. To address the limitation, we propose propagating the latent motion from the parent octree level to the current octree level. The proposed hierarchical motion propagation mechanism can better cover large motions and enable more accurate motion representation. The proposed "Hierarchical Motion for Dynamic Point Cloud Compression (HM-DPCC)" produces -8.7% BD-Rate savings on point-to-point geometric error compared to MPEG AI-PCC V3.

Raw image compression, which encodes Bayer raw data into a bitstream for sRGB reconstruction, is essential for addressing bandwidth bottlenecks in modern photography. While recent learning-based end-to-end methods have achieved promising results, existing pipelines often suffer from insufficiently constrained latent representations and fail to fully exploit intrinsic raw-domain characteristics, leading to suboptimal compression efficiency. To address these challenges, we propose a novel framework that integrates an Image Signal Processor (ISP)-prior preprocessing module and compact raw side information. Specifically, we introduce a learnable ISP module to form an intermediate representation better aligned with the sRGB domain, mitigating the raw–sRGB gap and facilitating subsequent compression. Furthermore, we propose a raw-modulated Entropy Model (REM) that conditions entropy modeling on the compact raw side information, enabling more accurate probability estimation. Extensive experiments demonstrate that our method achieves superior rate–distortion performance over compared methods.

Existing computer vision processing pipeline suffers from two efficiency problems, including image signal processor (ISP) that first converts Bayer pixel data to RGB on a pixel by pixel basis, and computation redundancy within video data, leading to high power consumption and latency. In this paper, we propose a novel framework that eliminates the ISP and leverages motion estimation to accelerate video vision tasks directly in the Bayer domain. We introduce Motion Estimation-based Video Convolution (MEVC), which integrates sliding-window motion estimation into each convolutional layer, enabling prediction and residual-based refinement that reduces redundant computations across frames. This design bridges the structural gap between block-based motion estimation and spatial convolution, enabling accurate, low-cost processing. Our end-to-end pipeline supports raw Bayer input and achieves over 70\% reduction in FLOPs with minimal accuracy degradation across video semantic segmentation and depth estimation benchmarks.

Although many model-driven methods can perform fairly accurate blur kernel estimation, they are usually computationally intensive and time-consuming. This paper proposes a new approach by embedding a Local Boosting process in the latent image estimation module of blur kernel estimation to accelerate the convergence speed of latent image estimation. In addition, we develop an efficient method to estimate the blur kernel by using a smaller latent image. Experimental results show that using Local Boosting and subimage can significantly improve the computational efficiency.

Video frame interpolation under motion blur is particularly challenging when the camera exposure duration is unknown. In this setting, accurate motion estimation becomes difficult, and the inherent imbalance between the exposure and readout phases further complicates the reconstruction of intermediate frames. To address these challenges, we propose an Exposure-aware Video Deblurring and Interpolation (EVDI) approach, which performs joint video deblurring and interpolation under unknown exposure. EVDI leverages a pre-trained video deblurring network as an exposure-aware feature encoder to capture both inter- and intra-frame motion while providing high-frequency details. These representations are then organized into two phase-specific feature anchors corresponding to the exposure and readout phases, and a multi-flow guided feature enhancement mechanism is further introduced to mitigate the phase imbalance. Based on the anchors, sharp and temporally consistent intermediate frames are reconstructed. Extensive experiments demonstrate that the proposed method achieves superior performance in both deblurring and interpolation quality.

In this work, we propose PatchCompressor , a resource-efficient video streaming framework, which uses lightweight edge-side processing to optimize bandwidth consumption for real-time video analytics. PatchCompressor is motivated by the observation that inference-critical information is spatially sparse and localized, while most of the frame consists of non-critical regions. Based on this, PatchCompressor employs a lightweight, class-agnostic region selection mechanism at the edge to identify patches of interest. The region-selection module guides the video encoder to encode the patches of interest at a higher quality, while aggressively compressing the remaining patches. We implement PatchCompressor as a complete end-to-end system on Jetson Orin Nano edge device using the Kvazaar HEVC encoder, and evaluate it on two real-world traffic surveillance datasets for two video analytics tasks. Experimental results show that PatchCompressor reduces bandwidth consumption by up to ~38% as compared to uniform encoding, with negligible loss in detection performance, and achieves higher end-to-end throughput compared to prior region-based video streaming pipelines.

Compressed-domain motion vector analysis is well established for rigid or near-rigid motion, such as vehicles and walking pedestrians, but unvalidated for the non-rigid, articulated motion of animals against the wind-driven backgrounds typical of camera trap video. We classify wildlife presence directly from H.264 motion vectors, with no frame decoding and no pixel-domain detector in the inference chain. Twenty-four per-frame motion statistics, organized in four families and processed in 30-frame windows by an LSTM (210k parameters), reach E0.5 = 0.77 on the camera trap benchmark of Riechmann et al., exceeding their best MV+YOLOv4 hybrid pipeline (E0.5 = 0.73) using motion vectors alone.

To reduce the computational complexity of H.266/VVC, we propose an adaptive bypass method for affine motion estimation using a Gradient Boosting Decision Tree. By applying a separate classifier for each block size, the method achieved a 3.23% speedup while limiting the decrease in coding efficiency to 0.13%.

This demonstration complements the accepted paper (SAR-NEUS: Neural surface inverse rendering for 3D reconstruction and novel view synthesis from SAR imagery) by providing an interactive visualization of the complete SAR-NeuS pipeline and presenting extended experimental results that could not be included in the conference paper due to space limitations. Visitors will see the different stages of the reconstruction process, from multi-view SAR observations to the learned neural surface representation, reconstructed 3D geometry, and synthesized SAR images from previously unseen viewpoints. The demonstration also provides an intuitive overview of the proposed inverse rendering formulation, illustrating the SAR-specific differentiable rendering process that links the learned 3D representation to the observed SAR images. The demo includes a collection of realistic electromagnetic simulations covering multiple object categories as well as larger and more complex scenes than those presented in the paper. During the demo, the author will interactively inspect reconstructed meshes, compare rendered SAR views against the corresponding ground-truth images, and analyze the effect of different acquisition conditions, including varying view density and scene complexity

This demo showcases the pixel-to-geocoordinate mapping pipeline presented in the paper, as deployed within SkyScoutAi, a UAV-based wildfire risk intelligence platform. Attendees will observe how pixel-level vegetation detections from UAV oblique and nadir imagery are automatically geolocated using camera geometry, drone metadata, and a trained ML offset correction model, enabling precise GPS-coordinate assignment to field detections. The live demonstration walks through the full SkyScoutAi workflow: raw UAV imagery input in both nadir and oblique views, the angular projection and elevation correction pipeline, and the trained offset model achieving sub-11m accuracy (CEP50-validated). Geolocated detections feed directly into SkyScoutAi's multi-factor wildfire risk model, producing parcel-level fuel maps and hazard assessments. Attendees can explore the live platform at dashboard.skyscoutai.com, seeing how pixel-level computer vision outputs translate to actionable geospatial risk intelligence. The demo bridges the paper's core research contribution to a deployed real-world system, illustrating the practical value of accurate UAV geolocation in wildfire prevention and land management.

This demo is associated with our accepted regular paper focused on text image inpainting for cultural heritage document restoration. We build a novel framework leveraging selective state space models to capture text continuity and structural characteristics. Our pipeline integrates character style extraction, StyleGAN-based structural prior generation and prior-guided image restoration. In the live demonstration, we visually compare our method against state-of-the-art algorithms on synthetic degraded text images containing ink erosion, stains, blurring and structural corruption. Visual reconstruction results and quantitative metrics will be displayed to verify the effectiveness of our approach. The source code will be publicly released after the conference. This demonstration illustrates the core contributions of our paper and enables attendees to intuitively understand the workflow of damaged document restoration.

Bi-temporal remote sensing change detection is useful for urban monitoring, environmental surface-change analysis, and downstream hydrological decision support. Existing supervised change detectors often perform well in-domain, but can confuse real structural change with registration error, illumination variation, seasonal appearance, or background clutter under cross-dataset transfer. We present \textbf{RVLCD}, a lightweight reliability-aware vision--language framework for bi-temporal remote sensing change detection. RVLCD keeps the CLIP backbone frozen, learns lightweight text and image adapters, encodes the two temporal images with a shared-weight Siamese visual branch, and fuses absolute temporal context with a \emph{signed} residual feature. A reliability head estimates whether the temporal discrepancy is trustworthy and controls the residual contribution. Unlike absolute differencing, the signed residual preserves the direction of the $T_2-T_1$ change in the CLIP embedding space, which is important for matching changed and unchanged text prototypes. Experiments using LEVIR-CD as the source domain and WHU-CD, DSIFN-CD, and S2Looking as target domains show strong zero-shot transfer on WHU-CD and reveal clear limitations under larger domain shifts. Ablations further show that a short text-adapter schedule and constrained signed-residual fusion are more effective for localization than black-box MLP or gated fusion variants.

Electron microscopy (EM) is a critical tool for visualizing biological structures at nanometer resolution, but acquiring high-quality volumetric datasets remains challenging due to frequent image corruption caused by mechanical vibration and sample instability. In this paper, we propose a structure-aware blind inpainting framework to restore severely damaged regions in EM images; here, \textit{blind} indicates that neither the position nor the size of the missing areas is known a prior. We first propose a degradation simulation pipeline that mimics realistic image corruption by disentangling signal and damage patterns from real EM dataset. Based on this synthetic training data, we develop a hybrid restoration network consisting of two stages: an adaptive convolution-based interpolation network that estimates a coarse structural prior from adjacent slices, followed by a structure-aware refinement network that fills in fine-grained image details. The network restores missing regions while maintaining structural continuity and visual coherence with the surrounding context. Experiments conducted on the newly collected \textit{EM-Inpaint} dataset demonstrate that our method outperforms existing inpainting baselines in both quantitative metrics and visual fidelity, particularly in regions with large and irregular damage. This work enables more complete and accurate reconstruction of biological specimens from imperfect EM data. Code is available at https://github.com/jtshou/BEAM.

We present GRACE (Geometry-level Reasoning for 3D HumAn-scene Contact Estimation), a novel approach for estimating dense contact on unstructured 3D human point clouds. Unlike prior works that depend on fixed-topology meshes (e.g., SMPL), GRACE learns to fuse 2D interaction semantics with 3D geometric structures via a point-cloud encoder and hierarchical feature fusion. This topology-agnostic design enables the model to generalize across diverse human geometries. Experiments show that GRACE achieves stateof-the-art performance on standard benchmarks and demonstrates superior robustness on unstructured point clouds compared to existing template-based methods.

Gaze estimation aims to determine human gaze direction from camera-acquired facial and ocular features for driver monitoring in intelligent cockpits. However, in wide-angle imaging scenarios within compact cabins, existing methods rely on conventional 2D representations; nevertheless, as gaze estimation is inherently a 3D task, the lack of 3D spatial priors inevitably results in suboptimal performance. To address this, we propose a novel geometry-aware framework, Tri-plane Representation Gaze Estimation (TriGaze), which utilizes camera-parameter-guided 3D representations for robust gaze estimation. Specifically, a Tri-plane Representation Mechanism (TRM) is introduced to compress 3D spatial information into orthogonal feature planes. To effectively process the projected 3D information, a Camera-Parameter-Guided Feature Modulation (CFM) module is devised to dynamically calibrate feature distortion according to specific imaging geometries, thereby enabling precise gaze estimation. Comparative experiments on the IVGaze dataset demonstrate the superiority of the proposed method.

Granulometric analysis of wood strands ensures structural integrity and production efficiency of Oriented Strand Board (OSB). Current vision-based studies have demonstrated the potential of automated image analysis for particle size estimation, but their progress is significantly hindered by the absence of public, high-quality, and domain-specific datasets. This paper introduces Granulo-10k, the first curated, open dataset of multiple-view wood-strand imagery designed specifically for research on OSB strand segmentation and multiple-view granulometry, covering height, width, and thickness. The dataset includes about 10,000 high-resolution images of 200 wood strands captured under controlled acquisition conditions, along with granulometric ground truth. We also provide 3D point clouds to capture three-dimensional information. We describe the acquisition protocol and annotation methodology used to ensure representativeness across real production variability. Baseline evaluations using modern deep neural networks and foundation models demonstrate the dataset's utility for benchmarking while revealing open research challenges such as precise thickness estimation.

Reconstructing high-fidelity 3D furniture models from a single image is crucial for applications in e-commerce, augmented reality, and virtual interior design. However, existing feed-forward 3D generation methods are often limited by structural distortions and surface detail loss due to the lack of explicit geometric constraints, making it challenging to capture both rigid structures and natural folds in soft furniture. To address this pain point, we propose a geometry-aware latent diffusion framework for single-image 3D furniture reconstruction. Our approach consists of two key modules: Mesh-Regularized Latent Training (MRLT) and Normal-Guided Feature Feedback (NGFF). MRLT imposes mesh-level geometric constraints on predicted latent representations during training to preserve valid geometric structures in the latent space. NGFF iteratively refines surface details during inference by fusing normal features derived from multi-lighting renders. Experimental results on the 3D-FUTURE furniture subset demonstrate that our approach consistently outperforms state-of-the-art feed-forward methods, achieving superior performance in Chamfer Distance, F-score, and Normal Consistency.

We introduce VolHuMe, a dataset of high-quality 4D human scans captured with a state-of-the-art volumetric studio using 64 RGB and 32 depth cameras. VolHuMe contains individual captures of 104 subjects and provides extensive ground truth, including SMPL-X, high-resolution meshes, multi-view RGB/depth images, rigged meshes, point clouds, garment segmentation, and detailed hand and facial geometry. Unlike prior datasets that primarily rely on full-body imagery, VolHuMe uses a close-range, high-resolution capture setup that preserves fine-grained body-part details, improving geometric fidelity and texture resolution. We benchmark VolHuMe on state-of-the-art methods across 3D and 4D human reconstruction tasks, showcasing the dataset’s quality and exposing the limitations of current evaluation testbeds.

Most algorithms for hyperspectral image unmixing produce point estimates of fractional abundances of the materials to be separated. However, in the absence of reliable ground truth, the ability to perform abundance uncertainty quantification (UQ) should be an important feature of algorithms, e.g. to evaluate how hard the unmixing problem is and how much the results should be trusted. The usual modeling assumptions in Bayesian unmixing rely heavily on the Euclidean geometry of the simplex and typically disregard spatial information. In addition, to our knowledge, abundance UQ is close to nonexistent in the literature. In this paper, we propose to leverage Aitchison geometry used in compositional data analysis to provide practitioners with alternative tools for modeling prior abundance distributions. In particular, we show how to design simplex-valued Gaussian Process priors using this geometry. Then we link Aitchison geometry to constrained optimization and sampling algorithms, and propose UQ diagnostics that comply with the constraints on abundance vectors. We illustrate these concepts on real and simulated data.

Reinforcement Learning (RL) offers a promising pathway for aligning Multimodal Large Language Models (MLLMs) with the strict precision requirements of chest X-ray (CXR) analysis. However, even with supervised initialization, the complexity of CXR interpretation creates a sparse reward manifold, often leading to exploration stagnation and training instability. To address these challenges, we propose ChestR1, a sample-efficient RL framework grounded in Group Relative Policy Optimization (GRPO). We introduce a novel Ground-Truth Augmented sampling strategy to guide exploration, coupled with an Entropy-Aware Policy Modulation mechanism to prevent distribution collapse. Furthermore, a Dynamic Failure-Aware SFT mechanism tightly couples supervision with RL, specifically injecting guidance for hard instances where exploration fails. Validated on our curated multi-task dataset, ChestR1 achieves competitive performance with lower training consumption and a smaller model size compared to state-of-the-art baselines.

In response to the ever-growing data storage demands, research has been increasingly focusing on more sustainable alternatives to traditional storage methods (e.g. HDD). One promising approach is storing cold data, i.e. rarely accessed data, on synthetic DNA, which offers exceptional density, low energy consumption, and long-term durability. Within this context, the Joint Photographic Experts Group (JPEG) developed a project, JPEG DNA, to create a standard for efficient coding of images on synthetic DNA. The JPEG DNA Verification Model (VM) is composed of a binary encoder followed by a DNA quaternary encoder. This modular design enables a stand-alone binary coding stage prior to quaternary encoding into a DNA-compatible format. This paper compares the performance of conventional transform-based methods and state-of-the-art learning-based approaches within the processing pipeline of the JPEG DNA VM. The evaluation uses well-known objective image quality metrics and includes a study of the codecs' complexity.

Single-image camera calibration and lens distortion estimation are fundamental to geometry-aware computer vision. Accurate calibration is essential in real-world applications, where small errors can lead to significant geometric inconsistencies in downstream tasks. However, progress in learning-based methods remains limited by the lack of datasets providing dense, image-level ground truth for both intrinsics and distortion. Existing benchmarks typically expose only partial calibration metadata and rarely include explicit Brown-Conrady distortion coefficients with controlled variation. To the best of our knowledge, no existing dataset provides large-scale, image-level ground truth for full camera intrinsics and Brown-Conrady distortion suitable for supervised learning. We introduce Camera Parameter Search (CPS), a large-scale synthetic dataset generated with AILiveSim that provides, for every image, ground-truth horizontal field-of-view, principal point, and full Brown-Conrady radial and tangential distortion parameters (shared within each configuration folder and thus valid for every image instance). CPS is designed as a large-scale dataset for training and benchmarking: it includes controlled parameter coverage, geometric validity constraints, and standard evaluation tasks for single-image calibration. The current release contains approximately 1.5 million images across two resolutions and 794 camera configurations, covering horizontal fields of view from 30° to 150° with multiple distortion realizations. CPS focuses on urban driving scenes with strong geometric structure, providing informative cues for camera calibration. We describe the data generation pipeline, constrained parameter sampling, annotation schema, and dataset organization, and we report experimental results demonstrating that CPS enables systematic and reproducible evaluation of calibration models.

Copy Detection Patterns (CDPs) are structures printed on physical objects to enable cost-effective authentication. Verification is achieved by comparing a captured image with the digital template from which the CDP was printed. In practice, printer stochasticity and camera distortions hinder this comparison, limiting robustness against counterfeiting. Prior work addressed camera effects by synthesising reference images in the verification camera domain, but it ignored printing variability. We introduce an enrolment-based cross-camera dual-synthetic referencing framework. Each printed CDP is first captured by a controlled enrolment camera, and a deep-learning-based translator jointly exploits the digital template and the enrolled capture to generate a high-quality reference for the verification image. We provide an information-theoretic justification showing that the dual reference is more informative than template-based references. Experiments on heterogeneous mobile cameras demonstrate improved authentication performance, robustness to machine-learning-based copy attacks, and reliable verification from small CDP regions and on low-end devices.

Recent advances in image manipulation have enabled highly photorealistic content generation, but also lowered the barrier to arbitrary editing, raising concerns about multimedia authenticity and security. Existing Image Manipulation Detection and Localization (IMDL) methods mainly target splicing or copy-move forgeries, while benchmarks for inpainting-based manipulations remain limited. To bridge this gap, we present COCO-Inpaint, a comprehensive benchmark specifically designed for inpainting detection and localization, with three key contributions: 1) High-quality inpainting samples generated by six state-of-the-art inpainting models, 2) Diverse generation scenarios enabled by four mask generation strategies with optional text guidance, and 3) Large-scale coverage of 238,302 inpainted images with rich semantic diversity. Our benchmark is constructed to highlight intrinsic inconsistencies between inpainted and authentic regions, rather than superficial semantic artifacts such as object shapes. We further establish a rigorous evaluation protocol with three standard metrics to benchmark existing IMDL methods and reveal current trends and challenges.

Single-Image Defocus Deblurring (SIDD) aims to recover an all-in-focus image from a single defocused observation, but rigorous and reproducible evaluation remains challenging due to the scarcity of realistic, high-resolution datasets with well-aligned defocused/sharp pairs and standardized protocols. We build on RealDefocus, a benchmark derived from the real-world RealBokeh dataset originally proposed for Bokeh Rendering. RealDefocus provides paired defocused inputs and sharp ground truth images, predefined training/validation/test splits, and a unified evaluation framework for comparing image restoration and neural rendering approaches. We further outline a benchmarking protocol with cross-dataset validation to assess reconstruction quality and generalization.

Existing individual re-identification methods often struggle with the deformable nature of animal fur or skin patterns, which undergo local geometric distortions due to non-rigid body deformations. We propose a surface-consistent match verification framework that combines evidence from two complementary image representations: original segmented images and unwrapped pelage textures mapped to a canonical UV space. To obtain the unwrapped representation, we estimate surface normals from monocular images and compute a geometry-aware UV mapping that preserves isometric relationships between the 3D body surface and the 2D texture space, requiring no ground truth UV annotations. Feature matches extracted independently from both representations are linked through bidirectional geometric verification, producing compound similarity scores that are robust to pose variation and surface deformation. We evaluate our approach on four challenging species with distinctive yet highly deformable fur and skin patterns.

Saliency maps are one of the most widely used post-hoc approaches for interpreting the behavior of Deep Learning models. Yet, assessing their fidelity is difficult in the absence of ground-truth explanations. To address this, numerous fidelity metrics have been introduced. Previous studies have shown that fidelity metrics can behave inconsistently under different perturbations, and a recent work has attempted to estimate this effect. However, the underlying reasons behind these observations have not been systematically explained. In this work, we formalize the assumptions implicit in perturbation-based fidelity metrics and verify them analytically. We introduce two conformity measures, DROP and PSim, that directly test these assumptions without requiring saliency maps or supervised models. Our results show that these assumptions often break down, explaining the observed inconsistencies. We therefore recommend such conformity analysis as a precondition for fidelity evaluation and carefully selecting perturbation design in Explainable Artificial Intelligence (XAI).

As freeform displays—such as curved, foldable, and stretchable—become increasingly prevalent in consumer electronics, ensuring high perceptual image quality under real viewing conditions remains a critical challenge. Geometric distortions induced by 3D surface curvature are often imperceptible in flat-panel evaluations but become visually noticeable when users view the display from off-axis positions. To address this, we propose a systematic framework that bridges physical display geometry and human visual perception through a vantage-aware distortion modeling pipeline, integrating surface reconstruction, viewing-angle-dependent mapping, distortion factor extraction, subjective validation, and final quantification into a unified perception-correlated index. This work lays the foundation for objective, user-centric quality evaluation in next-generation display design.

We present an in-progress audio-visual quality database for benchmarking objective and subjective quality assessment under modern streaming conditions. The database will contain about 30 source audio -visual contents with resolutions from 720p to 4K, covering diverse scenes such as campus, news, music, animation, sports, and everyday activities, with paired audio tracks representative of typical online media. The sources are encoded using conventional video codecs like HEVC, AV1, VVC, and neural video codecs (SSF), combined with AAC and neural audio codecs (SNAC) at multiple bitrates, under controlled asymmetric audio and video distortion profiles across five quality bands. Distortions include spatial blur and desaturation, temporal frame dropping or freezing, and audio degradations such as additive noise, echo, and other effects. Perceptual ground truth will be obtained from a single stimulus ACR HR experiment, yielding mean opinion scores for about 720 distorted sequences to support the development of no reference audio visual quality models.

Accurate cavitation volume estimation in tunnel experiments requires precise stereo camera calibration. However, extreme operating conditions—such as high-speed flow, elevated pressure, and vibration—cause continuous camera displacement. To address this, we proposed the propeller's CAD model-based camera calibration. Although the propeller's low-texture surface, rotational symmetry, and partial occlusion cause recent deep learning methods to fail, our method overcomes these challenges with a geometry-based approach. Given a manual initialization, our pipeline iteratively refines the camera pose using 2D ICP followed by a PnP solver. The experimental results demonstrate that our method achieved more robust alignments.

Federated video action recognition enables collaborative model training without sharing raw video data, yet remains vulnerable to two key challenges: model exposure and communication overhead. Gradients exchanged between clients and the server can leak private motion patterns, while full-model synchronization of high-dimensional video networks causes significant bandwidth and communication costs. To address these issues, we propose Federated Differential Privacy with Selective Tuning and Efficient Communication for Action Recognition, namely FedDP-STECAR. Our FedDP-STECAR framework selectively fine-tunes and perturbs only a small subset of task-relevant layers under Differential Privacy (DP), reducing the surface of information leakage while preserving temporal coherence in video features. By transmitting only the tuned layers during aggregation, communication traffic is reduced by over 99% compared to full-model updates. Experiments on the UCF-101 dataset using the MViT-B-16x4 transformer show that FedDP-STECAR achieves up to 70.2% higher accuracy under strict privacy (ε=0.65) in centralized settings and 48% faster training with 73.1% accuracy in federated setups, enabling scalable and privacy-preserving video action recognition. Code available at https://github.com/izakariyya/mvit-federated-videodp

Spiking Neural Networks incorporating Spatio-Temporal Attention have emerged as an effective paradigm for neuromorphic vision, leveraging spatial and temporal dependencies to achieve strong performance and energy efficiency. However, such models often exhibit limited robustness, as subtle input perturbations can induce unstable spike dynamics and amplify output errors over time. This paper presents a theoretical analysis of robustness in STA-equipped SNNs, characterizing how perturbations propagate through attention--spike interactions. The analysis reveals that output deviations arise from two primary factors: spatial sensitivity of attention modulation and temporal drift in spike responses, which together can lead to quadratic error accumulation across time steps. Guided by these insights, two regularization techniques are introduced to improve robustness during training: Spatial Smoothness Regularization (SSR), which constrains attention sensitivity to spatial perturbations, and Temporal Consistency Regularization (TCR), which suppresses noise-induced temporal fluctuations in attention dynamics. Extensive experiments on CIFAR-10, CIFAR-100, and ImageNet demonstrate that the proposed training strategy improves average corrupted-data accuracy by over 1\% across all three benchmarks and by 4.90\% under CIFAR-100 shift noise, while preserving the intrinsic efficiency of spiking computation with no additional inference cost.

Federated learning (FL) is a machine learning method that enables multiple local clients to collaboratively train a global model. This is especially useful in the case of medical image segmentation, where data sharing can lead to privacy concerns. With all its advantages, FL is also subject to a class of attacks, called Byzantine attacks, wherein a subset of clients act as adversaries. These clients can seriously hamper global model performance by sending malicious updates to the global server. Our model, Pruning-based Entropy Aggregation (PEA), employs a detection-based algorithm to identify malicious clients, thereby providing extreme robustness. It is inspired by the fact that benign clients make predictions with high confidence. Hence, the algorithm utilises a Gaussian latent variable to generate a proxy dataset, which is used to calculate the entropy signatures of all client predictions and to identify and prune out the adversarial clients.

Reconstructing thin 3D structures is challenging due to their sparsity, scale variation, and complex geometry. Such structures arise in a wide range of domains, including medical imaging of vascular systems and industrial pipe systems. While recent neural methods perform well on dense surfaces, they often fail to recover fine thin geometries. We propose a reconstruction approach based on local depth projections, which provide an efficient and informative 2D representation of thin structures. Specifically, we traverse the 3D model with a sliding box to generate local orthographic depth projections, which are processed by a neural network to reconstruct missing thin structures in 2D. The local reconstructions are subsequently fused back into the 3D model to produce a coherent and detailed shape. Experiments on pulmonary artery reconstruction from CT volumes and industrial pipeline recovery from synthetic and real scans demonstrate improved preservation of fine structural details over existing methods.

Automatic image cropping is a pivotal technique for high quality image editing. However, existing methods often overlook latent compositional consistency of crops, which serve as an essential clue for cropping optimization. To address the challenge of modeling this implicit clue, this paper proposes a two-stage cropping model guided by the compositional consistency. In the first stage, a baseline network generates multiple candidate crops. Then, a compositional consistency judgment module (CCJM) is developed in the second stage to distinguish whether each candidate crop can maintain the compositional consistency compared with the original image or not. Consequently, our model can adaptively select different feature fusion strategies to refine the candidate crop. Moreover, we also design a boundary constraint loss to better regularize the coordinate deviation between the refined and initial crops. Extensive experiments demonstrate that our method effectively leverages compositional consistency to achieve better performances across benchmark datasets.

Accurate image segmentation remains challenging, particularly in generating sharp, confident boundaries. While modern architectures have advanced the field, many of them still rely on standard loss functions like Cross-Entropy and Dice, which often neglect the discriminative structure of learned features, leading to inaccurate boundaries. This work introduces Deep Discriminant Analysis (DDA), a differentiable, architecture-agnostic loss function that embeds classical discriminant principles for network training. DDA explicitly maximizes between-class variance while minimizing within-class one, promoting compact and separable feature distributions without increasing inference cost. Evaluations on the DIS5K benchmark demonstrate that DDA consistently improves segmentation accuracy, boundary sharpness, and model confidence across various architectures. Our results show that integrating discriminant analysis offers a simple, effective path for building more robust segmentation models.

State-of-the-art learned reconstruction methods often rely on black-box modules that, despite their strong performance, raise questions about their interpretability and robustness. Here, we build on a recently proposed image reconstruction method, which is based on embedding data-driven information into a model-based convolutional dictionary regularization via neural network-inferred spatially adaptive sparsity level maps. By means of improved network design and dedicated training strategies, we extend the method to achieve filter-permutation invariance as well as the possibility to change the convolutional dictionary at inference time. We apply our method to low-field MRI and compare it to several other recent deep learning-based methods, also on in vivo data, where the benefit of using a different dictionary is demonstrated. We further assess the method’s robustness when tested on in- and out-of-distribution data. When tested on the latter, the proposed method suffers less from the data distribution shift compared to the other learned methods, which we attribute to its reduced reliance on training data due to its underlying model-based reconstruction component.

Continual learning (CL) aims to learn tasks sequentially, reflecting real-world applications such as intelligent surveillance systems. Such practical situations are often resource-constrained, making sparse continual learning (sparse CL) particularly attractive as it ensures computational and memory efficiency by updating only a small subset of parameters without model expansion. However, existing sparse CL methods still struggle with the core challenge of CL: the stability--plasticity dilemma---retaining prior knowledge (stability) while acquiring new knowledge (plasticity). This paper proposes Adaptive Continual Sparse Learning (ACSL), a novel sparse CL method that adaptively controls both how many and which parameters are updated for each task. ACSL estimates the task familiarity of the model and dynamically allocates the number of parameters to update, thereby tailoring the stability--plasticity balance to each task. Furthermore, ACSL periodically reselects parameters to update in a stochastic manner, controlling plasticity at the parameter level. Experiments on multiple image classification benchmarks show that ACSL achieves higher accuracy than existing sparse CL methods across most datasets.

While Contrastive Language–Image Pre-training (CLIP) demonstrates strong zero-shot image classification, it remains limited in open-vocabulary semantic segmentation (OVSS) due to inaccurate spatial localization and ambiguous object boundaries. To overcome this, we propose SGCLIP, a training-free framework for OVSS. Specifically, we introduce Multi-scale Adaptive Feature Fusion and Entropy-Adaptive Top-K sparsification to suppress redundant attention and distill discriminative semantic cues, alongside a Multi-View Visual Consistency strategy to enhance the consistency of prediction responses across views. Furthermore, a Semantic-Geometric Uncertainty Fusion module is utilized, where zero-shot guidance from the Segment Anything Model (SAM) is exploited as a geometric prior for the refinement of coarse segmentation mask boundaries. Experimental results demonstrate that SGCLIP outperforms existing training-free and even training-based methods on multiple benchmark datasets.

Facial Expression Recognition (FER) adds a layer of emotional intelligence to technology, enabling digital systems to respond to user emotional states. While deep learning has advanced FER, practical application is constrained by the long-tail distribution of real-world data, where underrepresented minority classes (e.g., Fear, Disgust) are frequently misclassified due to feature ambiguity. To address this limitation, we present EMVMamba, a novel hybrid architecture designed to resolve the part-whole ambiguity in rare expressions. Specifically, EMVMamba framework employs a dual-stream paradigm: a Local Hierarchical Branch (CNN) to capture multi-level compositional structure and textural details, and a Global Topological Branch (VMamba) to model long-range semantic dependencies. We also introduce an Adaptive Fusion Gate that leverages global context to bridge these distinct feature domains. Combining this architectural advantage with a class-weighted loss, EMVMamba achieves a competitive 80.61% accuracy on RAF-DB. Notably, its performance surpasses DDAMFN and marginally exceeds the heavy Poster++ model with 40% fewer FLOPs (5.06G vs 8.4G), offering a state-of-the-art trade-off between computational efficiency and rare-class recognition performance. Additional experiments on the FER-2013 dataset further demonstrate its robust generalization capabilities.

Federated Learning (FL) methods typically require Deep Neural Network (DNN) weight transfer from FL clients to an FL aggregator (master) for centralized DNN aggregation. However, this may not be possible under strict privacy or network constraints. In this paper, we present Proto-SVDD, a fully decentralized federated learning framework for enabling collaborative DNN model training, without neural parameter sharing. Instead, Proto-SVDD employs a lightweight class-wise prototype learning mechanism based on Support Vector Data Description (SVDD), which is trained by each FL client DNN using their own local, private data. Collaboration between FL clients involves exchanging and aggregating only their SVDD class prototypes in a fully decentralized topology. Neither DNN weights nor training data are exchanged between FL nodes whatsoever. Proto-SVDD is evaluated under a six-client cross-domain protocol that simulates separate data sources across locations and disaster-related distribution shifts. Compared with FedAvg, which transmits the full YOLOv6-Small model in each communication round, Proto-SVDD exchanges only one compact class prototype per object class. In the two-class setting used in this work, this corresponds to about 0.016 MB per client per round, compared with about 68 MB for full YOLOv6-Small weight transmission. Thus, Proto-SVDD greatly reduces communication cost while retaining competitive object detection performance against prototype-based FL baselines.

Efficient rate-distortion optimization (RDO) is a key challenge in immersive applications involving 3D scenes with heterogeneous assets, where uniform bitrate allocation fails to account for the viewpoint-dependent perceptual contribution of each asset to the rendered view. This paper proposes a low-complexity, viewpoint-aware coding framework that uses a Lagrangian RDO formulation based on asset importance, to guide bitrate allocation under dynamic viewing of such scenes. A simulation framework is also developed to assemble multi-asset 3D scenes and render 2D projections, to demonstrate the effectiveness of the proposed approach. Experiments with standard MPEG point cloud and dynamic mesh datasets, encoded using state-of-the-art MPEG codecs, show that the proposed framework enables the usage of different importance measures, yielding markedly different bitrate allocation behaviors, demonstrating the flexibility of the proposed approach. Results further show that rate-distortion performance is strongly dependent on scene layout and camera motion, further highlighting the need for viewpoint-aware RDO in immersive systems.

Federated Learning (FL) has emerged as the standard paradigm for training computer vision models on distributed networks without compromising user privacy. One of the key challenges in federated learning is the non-stationarity of data, particularly distributed concept drift, where disjoint client groups hold contradictory label assignments for identical visual features over time. Another challenge is data heterogeneity, where different clients have different data distributions. Despite previous efforts, creating a lightweight framework capable of disentangling simultaneous data heterogeneity and distributed concept drift with negligible computational overhead remains a significant challenge. In this work, we propose a robust, adaptive clustered federated learning framework that efficiently disentangles client populations with conflicting semantic concepts. Our method utilizes a recursive bi-partitioning algorithm to efficiently identify latent cluster structures by analyzing the geometric congruence of local weight updates. We obtained encouraging results through experiments on different datasets, demonstrating the efficacy and robustness of our proposed framework.

Federated learning (FL) enables privacy-preserving medical image segmentation by allowing collaborative training across multiple institutions. However, most existing FL-based medical segmentation methods rely on federated averaging (FedAvg). In the presence of non-IID data distributions, FedAvg may degrade performance due to inconsistent updates from different clients. To address this limitation, we propose gradient-aware federated segmentation (GAFSeg), an FL framework for medical image segmentation. At the server, we introduce a gradient awareness factor (GAF) that adaptively weights client updates based on gradient alignment with the server's aggregated gradient. We also propose a client training technique, termed ProxEMA, that integrates exponential moving-average-based weight updates with a proximal regularization term to suppress model drift across communication rounds. Extensive experiments on multiple publicly available data sources show that the proposed method outperforms the state-of-the-art approaches. Code is available at https://github.com/Sayantatechgit/GAFSEG

Split Federated Learning (SplitFed) combines federated and split learning to preserve privacy while reducing client-side computation. However, in medical image segmentation, heterogeneous label quality across clients can significantly degrade performance. We propose SplitFed-CL, a co-learning framework where a global teacher guides local students to detect and refine unreliable annotations. Reliable labels supervise training directly, while unreliable labels are corrected via weighted student--teacher refinement. SplitFed-CL further incorporates consistency regularization for robustness to input perturbations and a trainable weighting module to balance loss terms adaptively. We also introduce a novel difficulty-guided strategy to simulate human-like boundary-centric annotation errors, where the degree of perturbation is governed by shape complexity and the associated annotation difficulty. Experiments on two multiclass segmentation datasets with controlled synthetic noise, together with a binary segmentation dataset containing real-world annotation errors, demonstrate that SplitFed-CL consistently outperforms seven state-of-the-art baselines, yielding improved segmentation quality and robustness.

Photometric inconsistency across indoor views often causes color drift and unstable geometry in multi-view 3D reconstruction. We propose LUCID-GS, a photometrically robust pipeline that jointly calibrates appearance and optimizes 3D Gaussian primitives. LUCID-GS learns per-image exposure gains and 3*3 color correction matrices, trained end-to-end with reprojection supervision and global histogram alignment to harmonize illumination and color across views. To strengthen geometric consistency, we integrate a multi-view depth module that combines cost-volume aggregation with cross-view attention, providing reliable depth constraints for accurate Gaussian initialization and refinement. Experiments on a challenging indoor multi-view dataset and additional scenes demonstrate improved appearance fidelity and structural reconstruction quality while preserving real-time rendering efficiency and strong generalization.

3D Gaussian Splatting (3DGS) has garnered considerable attention for photorealistic novel view synthesis, and is being introduced to 3D viewing applications. However, intentional non-photorealistic visualizations, such as rendering hierarchical structures semi-transparently or overlaying semantic information are also important for helping users understand scenes. This paper proposes a method that incorporates artificial semi-transparent effect into 3DGS scenes, together with an efficient implementation that preserves real-time rendering performance. Our approach adopts a point cloud-based semi-transparent representation and performs rendering that integrates point clouds with Gaussians. The visualization effects and rendering performance of the proposed method are evaluated by rendering multiple 3D objects semi-transparently within a 3DGS scene.

Dynamic Gaussian Splatting (often known as 4DGS) is emerging as a promising technique for representing real-world dynamic scenes. However, 4DGS representations are susceptible to photometric variations caused by exposure and white balance shifts inherent to physical cameras. These cross-view inconsistencies degrade rendering quality and destabilize training because the resulting 4DGS representation risks overfitting to view- and content-dependent photometric artifacts. To address this challenge in the offline setting, we introduce IC-4DGS, a plug-and-play framework with view-dependent illumination-compensated networks (IC-Nets) that disentangle photometric artifacts from the offline 4DGS representation, enabling stable optimization and compact, high-fidelity 4DGS models. Extensive experiments show that IC-4DGS is effective and broadly applicable across diverse offline 4DGS representations, consistently improving rendering quality without altering core optimization procedures.

3D Gaussian Splatting (3D-GS) enables high-quality novel view synthesis with real-time rendering, but typically requires millions of Gaussian primitives, resulting in substantial memory and storage footprints. Most existing compression methods rely on heuristic pruning or generic quantization, which are ineffective at preventing redundancy accumulation due to unstable early-stage optimization. In this work, we propose a compact training framework for 3D-GS that enables reliable compression during optimization. Our method employs progressive pyramidal training, allowing Gaussian primitives to stabilize under coarser scene representations before incorporating fine-scale appearance details. This formulation enables reliable estimation of Fisher Information Matrix-based sensitivity scores during training, which are subsequently used to prune redundant primitives. To further reduce storage, we introduce a correlation-aware quantization strategy. Experiments on standard novel view synthesis benchmarks demonstrate that our method achieves an average reduction of 95% in storage footprint while preserving high-quality rendering.

Recent advances in 3D Gaussian Splatting (3DGS) have demonstrated impressive performance in 3D reconstruction. However, artifacts such as spurious geometry and floaters remain annoying in practical 3DGS. While recent approaches increasingly leverage diffusion priors to improve visual fidelity, they often suffer from geometric hallucination due to their exclusive reliance on 2D image supervision. In this paper, we propose \textbf{DepthFix3D}, a geometry-aware diffusion framework designed for high-quality artifact removal in 3DGS renderings. The core of our approach introduces depth 3constraints in diffusion-based artifact removal, which consist of multi-source geometric cues extracted from both 3DGS and point cloud. The depth images from 3DGS and point cloud exhibit distinct artifacts-related patterns, enabling better identification of rendering artifacts via the comparison of their discrepancies. Moreover, we find that introducing point cloud representations provides complementary scene information, enabling more 3D consistent and higher-quality content generation. Multi-source geometric cues are conditioned into the diffusion process via a ControlNet, and a progressive view synthesis mechanism is further utilized to enhance 3DGS quality. Experiments on diverse datasets demonstrate that DepthFix3D significantly outperforms existing baselines in both perceptual quality (1.87 dB in PSNR) and geometric consistency.

Instruction-guided editing with 3D Gaussian Splatting (3DGS) enables easy asset modification, but introduces security risks. While 2D protection methods are well-established, they fail in 3D due to view-inconsistent perturbations that converge to an averaged representation during 3D reconstruction. Moreover, directly optimizing pre-trained 3DGS models lacks sufficient capacity to embed effective perturbations, resulting in weak protection. To address this, we propose \textbf{SplatShield}, a 3DGS protection framework that embeds view-consistent adversarial perturbations into 3D Gaussians through (i) contribution-aware initialization that strategically expands model capacity via Gaussian pruning and splitting to accommodate the embedding of subtle perturbations, and (ii) iterative adversarial optimization that injects perturbations while preserving multi-view consistency. Experiments demonstrate that SplatShield provides robust protection against unauthorized editing across both training and novel views while maintaining high visual quality.

Neural radiance field (NeRF) and 3D Gaussian splatting (3DGS) are two mainstream approaches for novel view synthesis. They often show complementary performance, i.e., 3DGS demonstrating faster rendering speed and NeRF demonstrating higher rendering quality. Motivated by this, we propose leveraging NeRF-rendered images for 3DGS. Specifically, we target street scenes and utilize a pre-trained street-specific NeRF method to produce training images for a target 3DGS method. In our 3DGS training, NeRF-rendered images are used to remove transient objects in street-level input views and to generate bird's-eye views as additional views, inheriting the higher-quality rendering of NeRF into 3DGS. We further incorporate a diffusion-based image enhancement to improve the image quality of the additional views. Experimental results on one synthetic and two real datasets demonstrate that our proposed method improves street-scene rendering while preserving the speed of 3DGS and the quality of NeRF

3D Gaussian Splatting (3DGS) has emerged as a highly efficient and high-quality approach to 3D scene reconstruction and rendering. However, it still relies on per-scene optimization, which typically takes tens of minutes. Generalizable 3DGS is proposed as a solution, while existing generalizable 3DGS methods still exhibit limited performance on unseen scenes. To address this challenge, we propose a generalizable 3DGS method guided by a Vision Foundation Model, named VFG-GS. VFG-GS consists of two core components: Dual Spade Modulation (DSM) and Cross View Feature Aggregation (CVFA). DSM adopts a symmetric dual-branch spatially-adaptive normalization (SPADE) architecture to modulate two types of features from the encoder and vision foundation model, thereby generating generalizable features with multi-view consistency. CVFA aggregates modulated features, capturing both view-specific details and global scene context. Experimental results on four datasets demonstrate that VFG-GS achieves competitive performance compared to existing methods. Moreover, after a brief per-scene optimization about one minute, VFG-GS outperforms the state-of-the-art methods in terms of PSNR, SSIM and LPIPS on the Real Forward-Facing and NeRF Synthetic datasets.

Recently, the field of novel view synthesis has witnessed remarkable advancement, where 3D Gaussian Splatting has gained considerable attention and spawned numerous applications due to its efficiency and performance. However, its effectiveness deteriorates significantly under sparse-view conditions. While existing approaches incorporate monocular depth maps as additional geometric information, they remain limited by inherent scale ambiguity. In this paper, we propose leveraging monocular normal maps, which provide rich geometric cues without scale ambiguity, to improve the reconstruction quality under sparse-view conditions. Two reconstruction-based losses based on normal maps are introduced to guide the orientations and undulation of 3D Gaussian primitives. We also incorporate a dense feature correspondence extractor to improves the density and coverage of the initial point cloud. Experiments on representative datasets demonstrate the advantages of the proposed framework compared to related methods.

3D Gaussian Splatting (3DGS) has emerged as an advanced technique for real-time novel view synthesis by representing scene geometry and appearance using differentiable Gaussian primitives. However, efficiently computing precise Gaussian–tile intersections remains a critical task in the rasterization pipeline. To this end, we propose QuadBox, a method that leverages four axis-aligned bounding boxes to tightly encapsulate projected Gaussians in a discrete manner. First, we derive a geometry-aware stretching factor that enables the construction of a tile-aligned QuadBox, which covers the elliptical projection and largely excludes irrelevant tiles. Second, we introduce QPass, a single-pass tile traversal algorithm that exhaustively exploits the discrete nature of QuadBox, ensuring that the tile intersection check is performed with simple interval tests. Experiments on public datasets show that our method accelerates the rendering speed of 3DGS by 1.85$\times$. Code is available at \href{https://github.com/Powertony102/QuadBox}{https://github.com/Powertony102/QuadBox}.

We present Cloth-HUGS, a Gaussian Splatting based neural rendering framework for photorealistic clothed human reconstruction that explicitly disentangles body and clothing. Unlike prior methods that absorb clothing into a single body representation and struggle with loose garments and complex deformations, Cloth-HUGS represents the performer using separate Gaussian layers for body and cloth within a shared canonical space. The canonical volume jointly encodes body, cloth, and scene primitives and is deformed through SMPL-driven articulation with learned linear blend skinning weights. To improve cloth realism, we initialize cloth Gaussians from mesh topology and apply physics-inspired constraints, including simulation-consistency, ARAP regularization, and mask supervision. We further introduce a depth-aware multi-pass rendering strategy for robust body-cloth-scene compositing, enabling real-time rendering at over 60 FPS. Experiments on multiple benchmarks show that Cloth-HUGS improves perceptual quality and geometric fidelity over state-of-the-art baselines, reducing LPIPS by up to 28% while producing temporally coherent cloth dynamics.

3D Gaussian Splatting (3DGS) is a major reference for learning 3D models of real scenes, enabling real-time novel view synthesis and fast model training. As part of the Elliptical Weighted Average (EWA) volume splatting methods, 3DGS assumes that the support areas of the Gaussians do not overlap, which – among other hypothesis – leads to a simple rendering by alpha blending. While this algorithm enables training a 3D model, theoretical analysis shows that alpha blending generates rendering misalignments for Gaussians belonging to the same surface, which may impair learning and eventually view synthesis. In this paper, we analyze the difference between alpha blending of Gaussian kernels and an accumulation by weighted average of their contributions, and we propose a simple algorithm for accumulating overlapping Gaussians while keeping alpha blending between surfaces. Experimental setup shows benefits of the proposed approach in terms of rendering quality

Although 3D Gaussian Splatting (3D-GS) achieves efficient rendering for novel view synthesis, extending it to dynamic scenes still results in substantial memory overhead from replicating Gaussians across frames. To address this challenge, we propose 4D Neural Voxel Splatting (4D-NVS), which combines voxel-based representations with neural Gaussian splatting for efficient dynamic scene modeling. Instead of generating separate Gaussian sets per timestamp, our method employs a compact set of neural voxels with learned deformation fields to model temporal dynamics. The design greatly reduces memory consumption and accelerates training while preserving high image quality. We further introduce a novel view refinement stage that selectively improves challenging viewpoints through targeted optimization, maintaining global efficiency while enhancing rendering quality for difficult viewing angles. Experiments demonstrate that our method outperforms state-of-the-art approaches with significant memory reduction and faster training, enabling real-time rendering with superior visual fidelity.

3D Gaussian Splatting (3DGS) achieves high-quality rendering but suffers from large memory footprints. Existing compression methods reduce storage through pruning, vector quantization, or redundancy-aware designs, but they often require scene-specific optimization, resulting in high computational cost, limited generalization, and poor real-time applicability. Recently, feed-forward 3DGS has emerged, generating Gaussian parameters in a single forward pass without scene-specific tuning, enabling fast and consistent inference. However, how to enable 3DGS compression in feed-forward 3DGS scenes in a general and training-free manner remains to be explored. To address this, we propose a feature-sorting-based compression method specifically optimized for feed-forward 3DGS. By exploiting the internal channel–view correlation of transformer-generated high-dimensional features, we develop a feature reordering strategy that yields a more temporally coherent sequence for compression. The reordered features are packaged into a feature sequence and compressed by an existing traditional video codec without any additional training. Our method significantly improves compression efficiency while preserving rendering quality and generalizing across diverse scenes.

While novel view synthesis methods like Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) have shown impressive performance on common scenes, they often struggle with capturing view-dependent specular effects, particularly near-field reflections. Existing methods utilize environment maps, which only represent lights from infinity, or environment Gaussians, which rely solely on indirect supervision from reflections, are insufficient for accurately modeling such effects. Here, we propose a framework based on 2D Gaussian Splatting (2DGS) that leverages directly observable environment information in input images. Our approach also integrates a hybrid reflection appearance representation and view-dependent specular modeling. Experiments on public benchmarks and a newly curated near-field reflection dataset demonstrate that our method achieves superior rendering quality and geometric accuracy.

3D Gaussian Splatting (3DGS) recently emerged as an effective representation of radiance fields. These scene representations define 3D models of a scene based on a set of multiple views, resulting in large amounts of data, therefore, their practical application can benefit from efficient coding and representation. Performance evaluation of different coding solutions requires new quality evaluation protocols, based on subjective and objective models. While prior work on subjective quality assessment of Gaussian splatting primarily focused on evaluating rendering quality, the perceptual impact of compression-induced distortions remains largely unexplored. This paper presents a subjective and objective quality assessment of state-of-the-art 3DGS codecs, explicitly targeting coding artifacts rather than rendering effects. Six different scenes depicting diverse scenarios were rendered using Scaffold-GS as a common anchor. Two Scaffold-GS-based codecs, namely HAC++ and ContextGS, were evaluated at four distinct quality levels. A subjective quality assessment study was conducted to assess the perceptual degradation of compression. Additionally, the performance of a set of objective image quality metrics was analyzed.

3D Gaussian Splatting (3DGS) is an efficient method for novel-view synthesis, but its performance depends on how camera relationships are modeled in complex scenes. Graph GS introduces a camera graph construction method, but it is primarily designed for linear camera trajectories and is less effective for 360° scenes. We propose a novel robust 360° camera graph construction method using Omnidirectional Neighbor Selection and Dynamic Quadrant Filtering. The camera graph is used to guide the 3DGS optimization process. Experiments on the Mip-NeRF 360 dataset show improved perceptual quality and rendering stability while preserving real-time performance.

Arbitrary-scale image super-resolution (SR) often suffers from geometric distortions under rotation, limiting its reliability in tasks that demand high structural fidelity. We propose an unsupervised rotation-equivariant arbitrary-scale SR framework that, for the first time, leverages 2D Gaussian splatting as the implicit neural representation. By structuring features along a discrete rotation group and predicting Gaussian parameters in a group-aligned manner, our pipeline guarantees provable end-to-end rotation equivariance without any rotation labels or data augmentation. The Gaussian representation naturally models oriented textures and delivers continuous, artifact-free reconstruction at any scale via local ensemble decoding. This unsupervised geometric prior consistently improves structural consistency, especially on rotated inputs. Experiments show clear PSNR gains over LIIF, LTE, OPE, and vanilla Gaussian, with our Gaussian-Equivariant achieving the best results and superior generalization. Visual results confirm practical advantages in preserving fine details across various imaging scenarios.

Image splicing is a challenging form of digital manipulation in which content from multiple sources is combined into a single image, often leaving subtle forensic traces. Prior work has increasingly highlighted that splicing localization differs from standard semantic segmentation in its objectives and the nature of the cues it relies on. Building upon these insights, this paper presents a lightweight hierarchical transformer-based framework for pixel-level image splicing localization, in which both the architecture and training strategy are deliberately aligned with the forensic characteristics of the task. The proposed approach emphasizes resolution preservation, multi-scale supervision, and boundary sensitivity to better capture subtle manipulation traces while limiting dominance of semantic image content. Unlike heavier multi-stream solutions, the method adopts a compact single-stream design focused on forensic signals. The approach is broadly evaluated using quantitative and qualitative analysis with a focus on splicing forgeries, demonstrating stable localization behavior across diverse datasets.

Multilingual Text-Based Person Search (TBPS) remains challenging in low-resource settings due to weak cross-modal alignment. While TBPS-mSigLIP improves robustness with sigmoid-based contrastive learning, it lacks effective discrimination for hard negatives. We propose mSigLIP-CLoRA, a parameter-efficient framework that enhances TBPS-mSigLIP by integrating Cross-modal Circle Loss for fine-grained discrimination and LoRA for stable, efficient training. Experiments show that our method achieves state-of-the-art performance in data-scarce scenarios, reaching 52.28% Rank@1 on 3000VnPersonSearch and significantly outperforming baselines with limited data, while remaining competitive on large-scale benchmarks with far fewer trainable parameters. The source code is available at our code repository.

The development of video game streaming has grown rapidly, with major platforms such as YouTube and Twitch using different codecs. To support quality assessment models that work consistently across any codec, it is necessary to have access to large, diverse subjective gaming quality datasets. Currently, there are only a few available, each having limitations. To address this gap, we present the largest gaming video quality dataset to date, incorporating both user-generated content (UGC) and professional-generated content (PGC) with extensive visual diversity. Our dataset covers the most widely used codecs—H.264, H.265, and AV1—and consists of 4,048 video samples, each annotated by an average of 37 mean opinion score (MOS) ratings. In addition to overall quality scores, we collect coarse-grained quality attributes, enabling a better understanding of perceptual factors. We study the performance of leading video quality assessment methods on this dataset, including a vision language model that outperforms all the benchmarks. To the best of our knowledge, this is the first dataset that comprehensively addresses gaming video quality assessment across multiple codecs and content types with quality attributes. Our dataset is publicly available at \url{https://rajeshsureddi.github.io/GameScope/}.

Driver fatigue remains a leading cause of road accidents, responsible for 24% of crashes. While yawning serves as an early behavioral indicator of fatigue, existing approaches face significant challenges due to the presence of systematic noise in video-annotated datasets arising from coarse temporal annotations. Training robust machine learning (ML) models requires rich supervisory labels that help learn salient features from the training data. Moreover, efficient on-device training and inference of models on edge devices is crucial in driver fatigue detection tasks to enable accurate real-time decisions on vehicles without reliance on cloud infrastructure. To address this issue, we develop a semi-automated labeling pipeline with human-in-the-loop verification to annotate YawDD videos to YawDD+ frame-level annotations, enabling more accurate model training on edge platforms such as NVIDIA Jetson NANO. Training the established MNasNet classifier and YOLOv11 detector architectures on YawDD+improves frame accuracy by up to 6% and mAP by 5% over video-level supervision, achieving 99.34% classification accuracy and 95.69% detection mAP on Jetson NANO and AGX. Moreover, MNasNet completed the epoch time in just 8.69 min/epoch while delivering up to 115 frames-per-second (FPS) inference time on AGX, confirming that enhanced data quality alone supports on-device driver fatigue monitoring systems without server-side computation. The YawDD+dataset and trained models are available online.

Video quality assessment (VQA) is a challenging research topic with broad applications. Traditional hand-crafted and discriminative learning-based VQA models mainly focus on pixel-level distortions and lack contextual understanding, while recent multimodal large language models (MLLMs) struggle with sensitivity to small distortions or handle quality scoring and description as separate tasks. To address these shortcomings, we introduce CP-LLM: a Context- and Pixel-aware Large Language Model. CP-LLM is a novel multimodal LLM architecture featuring dual vision encoders designed to independently analyze perceptual quality at both high-level (video context) and low-level (pixel distortion) granularity, along with a language decoder that subsequently reasons about the interplay between these aspects. This design enables CP-LLM to simultaneously produce robust quality scores and interpretable quality descriptions, with enhanced sensitivity to pixel distortions (e.g., compression artifacts). Experiment results demonstrate that CP-LLM achieves state-of-the-art cross-dataset performance on VQA benchmarks and superior robustness to pixel distortions.

This paper introduces a two-fold contribution built around the DARRL dataset, a multimodal collection of human demonstrations in an industrial manipulation setting designed to support action recognition and robot learning. In the first part, we present the structure and motivation of the dataset, highlighting the diversity of tasks performed, experimental design, and the extent of manual annotations. DARRL comprises RGB-D video recordings of human operators performing constrained and unconstrained manipulation scenarios, with detailed object and action annotations and auxiliary eye-tracking data collected from external observers. In the second part, we propose a structured representation of selected annotated sequences using frame-wise spatial scene graphs connected with temporal edges. Each graph encodes object-centric information: class labels, bounding boxes, segmentation masks, and predefined role categories, alongside a set of spatial and temporal relationships generated by algorithms that take node properties and centroid coordinates as input. Spatial edges are bidirectional and capture static scene layout, while temporal edges are directed and preserve object continuity across time. The resulting scene graphs are interpretable, reproducible, and tailored to the needs of robotics applications. They provide a foundation for symbolic reasoning and serve as direct input to graph-based learning models for downstream tasks such as activity recognition, affordance inference, and action segmentation.

Remote sensing semantic segmentation typically relies on pixel-level annotations, which are costly and semantically ambiguous under complex textures and unclear boundaries. Although weakly supervised methods reduce annotation cost, they remain dependent on coarse image-level supervision and error-prone pseudo masks. We propose a region-level text supervision paradigm that relocates semantic supervision from pixel space to region-level language descriptions, enabling fully mask-free semantic segmentation. Based on this paradigm, PaTSeg learns class-aware region representations via patch–text alignment and performs dense prediction through language-guided evidence projection. Experiments show over 30% mIoU improvement over image-level supervision and performance approaching pixel-level training without pixel-level annotations.

Automatic depression detection is important for mental health assessment. Recent multimodal approaches exploit cues from visual, audio and text, but lack explicit injection of clinically meaningful depression context. To address this limitation, we propose a context-aware depression detection method. A depressive personal feature (PF), aligned with nine symptoms from PHQ-9 questionnaire, is fused into a multimodal Transformer. We use GPT to provide soft PF labels, and fine-tune a LLaMA-based predictor with injected domain knowledge to obtain refined PF vectors, which explicitly encode symptom-level depressive context. To enable interpretable symptom-level contributions, dimension-wise gated fusion is employed to inject depressive PF into multimodal representation. Evaluated on the CMDC dataset, the proposed method outperforms baselines with 98.66% F1-score, with loss of interest, appetite change, sleep problem, low energy and suicidal ideation as the most influential PF dimensions, aligning with clinical findings. The results indicate improved performance with context-aware depressive PF injection, while offering better interpretability through symptom-level contributions. Our code is available at https://github.com/huiyu-yang/multimodal-depression-phq9-feature.

The pervasive growth of User-Generated Content (UGC), particularly Short-Form Video (SFV), has made robust Video Quality Assessment (VQA) an important research topic with high practical impact. Existing No-Reference based metrics are typically designed to measure absolute video quality and consequently become unreliable when tasked with distinguishing fine-grained quality differences, such as those introduced by compression, among variants of a single source video. Furthermore, traditional reference-based VQA approaches struggle with UGC because the source videos are often not of pristine quality. To drive forward research on fine-grained UGC quality assessment, we introduce the YouTube SFV Compression Dataset, a large-scale UGC/SFV quality dataset specifically annotated for the fine-grained quality variances resulting from video compression. Leveraging this novel dataset, we establish comprehensive dimensions for evaluating VQA metrics and propose Universal Video Quality (UVQ) model training framework for no-reference based metrics to efficiently measure both absolute perceptual quality and fine-grained quality distinctions. The dataset is available at media.withyoutube.com/ugc-fq.

Car sticker detection is a challenging small-object detection problem due to the tiny size of stickers, cluttered backgrounds, and limited annotated data. These challenges are further amplified in few-shot settings where only a small number of labeled instances are available for training. In this work, we propose Windshield-to-Sticker Detection (WiSDet), a region-guided detection pipeline that first localizes the windshield and then performs sticker detection within the cropped region to reduce background interference and improve effective sticker resolution. WiSDet is integrated with few-shot transfer learning using both FsDet and YOLOv8, and evaluated under 31-, 10-, 5-, and 2-shot settings. Experimental results on the proposed FewCarStickerPH dataset demonstrate consistent improvements across all models and K-shot splits, with gains of up to +21.58% AP for FsDet and +16.80% AP for YOLOv8n. Performance improvements are observed across different vehicle viewpoints and sticker placements, highlighting the robustness of the windshield-guided approach. To support this task, we also introduce FewCarStickerPH, a real-world crowdsourced dataset capturing diverse windshield viewpoints and conditions. The results show that incorporating structural priors through region-guided detection provides a simple yet effective strategy for improving few-shot tiny object detection.

Accurate scene text segmentation remains challenging because it typically requires dense pixel-level annotations that are costly and labor-intensive to obtain. In contrast, bounding boxes provide coarse supervision that is far cheaper to collect. We ask: can coarse box labels effectively boost fine-grained text segmentation? We present MDSeg, a multi-task learning framework that trains with multi-granularity supervision—binary masks and bounding boxes. The core is a detection-guided feature fusion module that distills localized text-region cues from the detection branch to assist the segmentation decoder. This design keeps detection and segmentation decoupled yet complementary: detection localizes and shapes region features, while segmentation refines boundaries. To further facilitate training, we curate a large-scale synthetic scene-text detection corpus called MTSD, covering diverse scenarios and text distributions, supplying abundant box-level supervision that indirectly benefits segmentation. Extensive experiments on challenging benchmarks show that MDSeg achieves state-of-the-art performance and consistently surpasses strong baselines under both full- and weak-supervision settings. The code and dataset are publicly available at: https://github.com/aTeppei/MDSeg.

No-reference video quality assessment (NR-VQA) for gaming videos is challenging due to limited human-rated datasets and unique content characteristics including fast motion, stylized graphics, and compression artifacts. We present MTL-VQA, a multi-task learning framework that uses full-reference (FR) quality metrics as supervisory signals to learn perceptually meaningful features without human labels during pretraining. By jointly optimizing multiple complementary proxy FR objectives with adaptive task weighting, our approach learns shared representations that transfer effectively to downstream NR-VQA. Experiments on gaming video datasets show that MTL-VQA achieves competitive performance against state-of-the-art methods in both mean opinion score-supervised and label-efficient or self-supervised settings.

Object detection and segmentation models are widely deployed in real-world applications, yet current benchmarks are almost exclusively limited to 8-bit Low Dynamic Range (LDR) imagery. These datasets fail to represent the challenging illumination conditions often encountered in practice, where LDR models struggle to capture both dark and bright regions simultaneously. To address this gap, we introduce HDRDCL (High Dynamic Range Detection in Challenging Lighting), a novel dataset of native HDR images annotated for both detection and segmentation. HDRDCL is composed of two subsets: HDRDCL-det, containing bounding box annotations, and HDRDCL-seg, which provides the first real-world HDR dataset with instance segmentation masks for diverse objects across indoor and outdoor scenes. Images were curated through an automated filtering pipeline that leverages multi-exposure tone mapping and object detection to select cases where HDR content provides tangible benefits. We benchmark six tone–mapping operators across multiple detection and segmentation models. By releasing HDRDCL, we aim to enable systematic evaluation of HDR-aware perception methods and to promote the development of robust computer vision systems capable of operating under extreme illumination conditions. Dataset information can be found at the following link: https://github.com/mlab-fau/HDRDCL

Assistive navigation systems for visually impaired people rely on accurate visual scene understanding. Developing these systems requires vision datasets that are representative of diverse real-world environments and outdoor pedestrian scenes. In many countries, pedestrian spaces sometimes feature uneven sidewalks, shared paths, and cluttered pedestrian zones that are under-represented in existing benchmarks. To address this gap, we introduce IndoNav, a segmentation benchmark dataset of pedestrian-centric scene understanding in Indonesian urban environments. IndoNav comprises 1,000 high-resolution RGB images, paired with corresponding depth data, and pixel-level annotations for 16 semantic classes. The dataset is designed to capture realistic, semantically complex pedestrian scenes that are critical for safety-aware navigation. Using IndoNav, we evaluate representative state-of-the-art semantic segmentation models and provide baseline results for pedestrian-centric scene understanding.

Annotating large-scale visual datasets for electrical grid inspection is labor-intensive, particularly if exhaustive labeling of all assets is required. However, for defect detection tasks, labeling errors in non-defective assets are more tolerable than errors on defective instances. We explore this error-tolerance asymmetry to propose a semi-supervised annotation pipeline that efficiently constructs high-quality training datasets from partial supervision. Given limitations on the available human experts and consequently on annotation resources, we propose labeling defects across thousands of images, while remaining assets can be labeled in only a small subset, used to train an asset detector that further scales asset-level annotations. Human and model-generated labels are then merged, using a supervision-aware strategy that prioritizes defect annotations. To address the challenges of long-distance image acquisition while avoiding downsampling high-resolution inspection imagery, supervision-aware image cropping is applied to focus training on regions of interest. Finally, coarse bounding boxes are refined into accurate polygonal instances using visual prompting with large vision models, while preserving supervised class identities. Ablation studies with state-of-the-art models demonstrate the effectiveness of each stage of the proposed pipeline.

This paper proposes a hybrid pipeline that combines multitask learning (MTL) with geometric analysis for the automatic quantification of capillary loops, arterial limbs, and venous limb diameters in nailfold capillaries. MTL U-Net simultaneously estimated vessel segmentation masks and keypoints for loop diameter calculation, whereas arterial and venous limb diameters were calculated through a geometric analysis of the segmentation results. The proposed method was statistically validated for reliability and agreement with expert measurements.

Accurate segmentation of 3D neurovascular images is crucial for vascular disease diagnosis, neurointervention planning, and navigation in conditions such as Cerebral Arteriovenous Malformation (cAVM). However, automatic segmentation algorithms often contain inaccuracies such as false positives or leakages at vessel crossings, where two physically separate vessels are incorrectly fused into a single connected component. In this work, we present a fully automatic framework to detect and correct false connections while keeping the correct vascular topology. Skeletonization of a binary vessel segmentation mask is used to create a graph representation, where all degree-four junctions are considered as potential false merging points. Each candidate resolves by finding the physically smoothest pairing of four incoming branches via A* (shortest-path graph search) path planning in a Voronoi-seeded local region of interest. A multi-gate classifier discards non-target anatomical patterns before any false connection removal is attempted. For the remaining candidates, an iterative min-cut procedure is applied to separate the incorrectly fused vessels. The pipeline resolved 10 (out of 19) targeted corrections (52% improvement).

Regular inspection of electrical equipment in hazardous (EX) environments is essential to prevent incidents such as fires and unplanned shutdowns. However, traditional inspections rely on manual visual checks, scaffolding, and high‑risk work at height, making the process costly and time‑consuming. We present an automated EX‑inspection pipeline combining robotics, machine‑vision segmentation, multimodal AI, and 3D spatial analysis to enable reliable remote inspection. Images from drones and 360° cameras are processed using deep‑learning models to detect key electrical components (e.g., lights, junction boxes, motors, cables). Segmentation outputs are converted into bounding boxes for seamless integration with asset management systems. Starting from a semantic segmentation baseline (~69% IoU, ~77% pixel accuracy), we transition to instance segmentation and hard‑negative mining, which significantly improve performance on challenging classes. Pixel accuracy increases to ~88% and IoU to ~77%, with notable reductions in false positives for small and ambiguous objects. A vision‑language model (GLM‑4) is applied to cropped detections to assess equipment condition and filter non‑damaged cases, reducing manual review effort. Integration with asset management systems supports tracking, SAP linkage, and historical condition monitoring. To handle complex scenes, we combine panoramic image flattening, point‑cloud alignment, and 3D localisation. 2D detections are projected into 3D space, clustered, and merged across multiple views, reducing inspection effort. Overall, the proposed pipeline reduces inspection time, HSSE exposure, and operational cost while improving consistency and scalability. Future work includes improving visual accuracy with multi‑view data, deeper integration with wIMS/Floc systems and EX registers, and enhanced 3D asset modelling for automated condition monitoring.

This demonstration presents an open-source end-to-end volumetric video streaming system based on standardized Video-based Point Cloud Compression (V-PCC) and secure Real-time Transport Protocol (RTP). Visitors can enter the volumetric video capture system built around a six-camera RGB-D rig and view their reconstructed point cloud live on a separate receiver computer after real-time point cloud generation, compression, network transmission, decoding, and rendering. This experience allows them to directly assess the end-to-end latency and subjective visual quality of the streaming pipeline. The system streams voxelized point clouds containing up to 280,000 points per frame at 30 fps. Its low-latency configuration achieves photon-to-photon latencies of 391–525 ms at transmission bitrates of 3.7–6.6 Mbps, while a low-bitrate configuration reduces the bitrate to approximately 2 Mbps. To the best of our knowledge, this is the first open-source real-time end-to-end volumetric streaming framework integrating V-PCC compression with RTP transport. Beyond showcasing the practical use of standardized V3C technologies, the demonstration provides attendees with a hands-on view of a complete volumetric communication pipeline and introduces a reproducible platform for further research on capture, coding, transport, and rendering.

This demo shows real-time playback of dynamic 3D Gaussian splat (4DGS) volumetric video on consumer hardware. A video-based implicit 3DGS compression pipeline factors Gaussian attributes into compact 2D feature planes coded with a standard video codec, while small decoder MLPs are coded with the Neural Network Codec. Applied to dynamic sequences (Bartender, Breakfast, Cinema) with per-frame, all-intra encoding, the resulting bitstreams are concatenated into looping videos. The full decoder runs live from bitstream to rendered view on a MacBook Air and a Samsung Galaxy S24+. Decoding and rendering happen entirely on the hardware video decoder and GPU, with no CPU–GPU transfers, sustaining real-time frame rates.

We present an interactive web-based visualization tool for cross-modal querying of multimodal Printed Circuit Board (PCB) data. The tool integrates four sensor modalities: RGB, VNIR hyperspectral (400–1000 nm, 224 bands), NIR/SWIR hyperspectral (900–1700 nm, 224 bands), and 3D depthmap, all captured from our custom industrial linear acquisition system. The core of the tool is built on UFM (Unified Flow Matching), a dense optical flow-based image registration framework that aligns all modalities to the RGB coordinate space. This allows users to simply click on any pixel in the RGB image and instantly retrieve the full reflectance spectrum from both the VNIR and NIR/SWIR hyperspectral cubes, as well as the corresponding metric depth value at that location. Users can also select individual modalities independently and query spectral or depth values from any point within that modality's own coordinate space. The tool is fully web-based. It is designed to accelerate diagnostic workflows in PCB analysis by removing the need to manually navigate high-dimensional hyperspectral cubes or separate depth maps. This demo supports our broader goal within the EU Horizon RETURN project: enabling fast, accurate material characterization of electronic waste components to support automated sorting and recycling of PCBs.

This demonstration presents the virtual reality environment developed for passive stress detection research, associated with our workshop paper "Temporal Pose Dynamics and Headset Telemetry for Passive Stress Detection in VR Environments." The environment is a room-scale interactive investigation scenario built in Unity, in which participants must locate specific items under time constraints. It operates in two distinct modes: a calm mode featuring a green timer and encouraging vocal feedback, and a stressful mode introducing time pressure through a red pulsing timer, sensory overload via intense music, and cognitive load through discouraging feedback. Conference attendees are invited to wear the VR headset and experience both modes firsthand. While doing so, they will generate the same two data streams used in our research: headset and controller telemetry, captured natively by the device, and body pose dynamics, captured via an external smartphone camera. This live setup illustrates the core thesis of our work — that the instrumentation already embedded in VR environments can serve as a passive, low-cost foundation for stress monitoring, without any contact-based physiological sensors. The demonstration offers a tangible and reproducible example of ecologically valid stress elicitation in VR, directly relevant to researchers working on human motion analysis, affective computing, and behavioral signal processing.

This Show & Tell demonstration presents a real-time RGB-D image-processing and AR system for non-contact livestock weight estimation. The system integrates RGB-D sensing, AI-based animal detection, 3D geometric analysis, and AR visualization to estimate the body weight of pigs and cattle without physical contact. Visitors can experience the complete processing pipeline using AR smart glasses equipped with an RGB-D camera. During the live demonstration, the system captures synchronized color and depth images, detects the animal using AI, reconstructs its 3D body shape, extracts geometric features, and estimates body weight in real time. The estimated weight and measurement results are immediately displayed within the user’s field of view, providing an intuitive understanding of the workflow. The framework has been extended from pigs to cattle by adapting the AI models and geometric feature-extraction methods. For cattle, the system analyzes standardized 3D point clouds to extract body height, body length, back profile, cross-sectional dimensions, and girth-related features for weight estimation and growth monitoring. Unlike conventional methods that require animal restraint or floor scales, the proposed system enables rapid, non-contact measurement while reducing animal stress and labor. The demonstration shows how image processing, 3D vision, and AI can be integrated into a practical smart livestock-monitoring system for real farm environments.

White blood cell classification across heterogeneous image datasets remains challenging due to variations in staining protocols, scan- ning devices, resolution, and class distribution. In this work, we propose a two-stage framework for robust cross-dataset classifica- tion. YOLOv11 is first applied for white blood cell localization and extraction, followed by a Swin Transformer for fine-grained classi- fication. Experimental results on multi-source datasets demonstrate that the proposed method outperforms ResNet50, VGG16-BN, and EfficientNet-B0. The best 9-class no-crop setting achieves 0.9709 macro-F1 on mixed-source testing, while maintaining 0.8188 macro- F1 on an unseen external test set.

Accurate recognition of endoscopic nasopharyngeal landmarks is vital for computer-aided diagnosis and surgical navigation. To address challenges posed by morphological similarity and intraoperative interference (e.g., reflections, bubbles), we propose GC-ResNet, an enhanced residual framework. GC-ResNet integrates an SE (Squeeze-and-Excitation) attention mechanism for adaptive channel calibration and incorporates a Global Convolutional bottleneck module to capture long-range spatial topologies. Experiments on clinical datasets demonstrate that GC-ResNet achieves an accuracy of 95.83%, significantly outperforming state-of-the-art models like Swin Transformer . Notably, it improves the F1-Score for highly confusable categories to 0.9581. While maintaining a high overall accuracy comparable to the baseline model, GC-ResNet demonstrates significantly enhanced robustness against multi-level interference. Grad-CAM visualizations and stress tests further confirm its superior anatomical localization precision under complex intraoperative conditions (p

As an extended paradigm of incremental learning, Incremental Implicitly-Refined Classification (IIRC) learns a model sequence from streaming data with hierarchical labels, leading to the confusion of hierarchical relation and mixup knowledge of superclasses and subclasses. We propose the Class Knowledge Capacity constrained optimal Transport (CKCT), by constructing dynamic hierarchical relation and utilizing it as the transport cost to guide reciprocal knowledge transport between old and new classes. During the reciprocal knowledge transport, we utilize a class knowledge capacity scheme to the source and target distributions to enforce class-relevant knowledge transfer. CKCT creates a feedback loop by refining the knowledge transfer and the hierarchical relation interactively. Experiments demonstrate the superior performance of CKCT over state-of-the-art methods. Code is available at https://anonymous.4open.science/r/CKCT-CAA4

Screen-shooting has become a major threat to image watermarking in the era of mobile devices. Most existing approaches focus on increasingly sophisticated distortion simulation during training, yet still adopt short-range network backbones. Such local-scale feature fusion is insufficient to model cross-region dependencies under nonlinear screen-shooting degradations, leading to unreliable decoding. In this paper, we propose WNet, an end-to-end watermarking framework based on a cascaded dual-UNet architecture. WNet establishes long-range dependencies to effectively capture global distortion patterns and enhance resilience against complex real-world degradations. To mitigate the attenuation of fine details introduced by long-range modeling, we design a Laplacian-pyramid-based high-frequency enhancement channel attention module to explicitly strengthen high-frequency structures. In addition, a texture edge guided loss is constructed to align watermark information with invisible texture-rich regions. Experimental results demonstrate that the proposed scheme significantly outperforms existing state-of-the-art methods. On average, our method achieves 1.28\% higher bit accuracy and a 0.52 dB gain in terms of PSNR compared with DoBMark.

Driven by the growing global burden of preventable vision impairment, MPS-RetNet is a semi-supervised framework for retinal disease classification that addresses annotation scarcity, image quality variability, and limited model transparency. It integrates a ConvNeXt backbone with multi-scale spatial pooling, spatial attention, prototype-guided representation learning, and a quality-aware supervision branch. Using consistency regularization with an exponential moving average teacher, it achieves 90.10% accuracy (Cohen’s κ = 87.04) on 4217 fundus images across four diagnostic categories (cataract, diabetic retinopathy, glaucoma, normal), with a mean cross-validation accuracy of 91.84%. External validation on 400 Messidor-2 images yields 95% accuracy. The quality estimation branch shows a correlation of 0.71 with BRISQUE metrics, confirming effective perceptual quality modeling. Clinical interpretability is validated by ophthalmologists, who rated saliency maps as anatomically relevant (mean relevance score: 4.36 ± 0.54). MPS-RetNet thus provides an annotation-efficient, quality-robust solution for scalable retinal screening.  Code: https://github.com/Maisamilens/MPSRET-Net-for-retina-Classification/tree/main

The coupled degradation of low illumination and motion blur significantly reduces image visibility and destroys structural details. Existing methods lack explicit modeling of the imaging mechanism, making it difficult to effectively characterize the impact of illumination changes on structure restoration. To address this, this paper proposes an illumination-aware frequency enhancement (IAFE) method. Starting from the physical imaging mechanism, under physical consistency constraints, it explicitly decomposes the degraded image into an illumination field and a structure-aware reflectivity component, thereby decoupling illumination attenuation and motion blur at the representation level. Furthermore, IAFE performs frequency-aware enhancement on high-frequency structures related to motion blur in the reflectivity component and introduces structured supervision for key edges in low-illumination regions to stabilize the structure restoration process and improve overall restoration quality. The proposed method achieves state-of-the-art performance on the LOL-Blur dataset and demonstrates good robustness and generalization ability on the LOL-V1, LOL-V2, and RealBlur datasets.

Retinal image-based eye tracking is widely used in ophthalmic imaging and vision science, and is a promising path to deliver higher gaze accuracy than the pupil- and cornea-based approaches commonly used in modern AR/VR devices. Nevertheless, existing retinal tracking algorithms still primarily rely on classical template-matching registration, which can be insufficiently robust to retinal feature variability and real-world imaging conditions. In this work, we propose a novel weakly-supervised, learning-based framework for robust retinal eye tracking. Initial studies demonstrate high accuracy, achieving the 95th-percentile gaze error

Diffusion models have become powerful priors for real-world image super-resolution. Most existing diffusion-based SR methods adopt Transformer architectures with fixed-resolution inputs, which are poorly suited to the arbitrary and often large image sizes required in practical super-resolution. As a result, these methods exhibit poor computational scalability and incur prohibitive memory and runtime costs at high resolutions. To mitigate this issue, existing approaches typically rely on patch-based processing, which is inflexible, inefficient, and unable to fully model interactions across patches. In this work, we show that this limitation can be naturally addressed by pure convolutional networks. We propose A$^2$SR, a fully convolutional diffusion framework that combines MeanFlow-based modeling with a pure ConvNet backbone. A$^2$SR supports arbitrary input resolutions and flexible sampling steps within a single model, enabling both one-step and few-step inference. Experiments on real-world benchmarks show that A$^2$SR achieves comparable or better perceptual quality than Transformer-based diffusion methods, while substantially improving computational efficiency.

We propose Temporally and Spatially Ordered Gaussians (TSOG), a format for efficient representation of 4D Gaussian Splatting (4DGS) content. TSOG extends the Spatially Ordered Gaussians (SOG) framework to the temporal domain by introducing a timeline attribute and temporal parameterization of geometry and appearance attributes. Similar to SOG, TSOG is a lossy format that assigns each Gaussian a unique index and encodes attribute values as index-aligned image data. TSOG is model-agnostic, extensible, and compatible with both discrete and continuous 4DGS representations. Evaluation using a PLYs sequence and FreeTimeGS as baselines, serving as simplistic and state-of-the-art 4DGS representations respectively, shows file size reductions exceeding 90%, with PSNR differences ranging between -0.42 and +0.85 dB. These results demonstrate substantial file size savings with minimal quality degradation, enabling efficient representation, storage, and delivery of dynamic scenes for next-generation 4D content. Related material can be found at https://github.com/xiaomi-research/tsog/.

Diabetic retinopathy (DR) grading systems suffer performance degradation across clinical imaging devices due to population domain shift. We present HyperNEST-TTA, a pipeline comprising encoders and combining hyperbolic geometry with nested multi-timescale learning for test-time adaptation in DR diagnosis. Operating on the Poincaré ball manifold, we employ three nested modules: a frozen module preserving universal disease hierarchy, a softly-adapted module capturing domain-specific patterns, and a fully-adapted module handling local variations. Evaluated on six cross-dataset grading tasks and one segmentation task, HyperNEST-TTA obtained 66.6% accuracy and 66.1% quadratic weighted kappa for grading, improving over Euclidean baselines by +4.3%, while reaching 0.452 average Dice coefficient for multi-lesion segmentation.

Reliable fundus image quality is critical for accurate retinal diagnosis and automated analysis. However, in real-world datasets, images labeled as high quality often contain residual artifacts, and truly paired low and high-quality samples are rarely available. These limitations challenge conventional restoration methods that assume clean ground truth references and supervised training conditions. In this work, we propose an Operational CycleGAN (OpCycle-GAN) framework tailored for fundus image restoration under unpaired and imperfect data settings. The model integrates Self-Organized Operational Neural Network (Self-ONN) layers into the CycleGAN architecture to enhance nonlinear representation capacity while maintaining a compact design. Additionally, a user-guided data refinement strategy is introduced to construct more reliable clean and corrupted domains, strengthening domain separation and stabilizing training. Extensive experiments on the DeepDRiD dataset demonstrate that the proposed method surpasses previous state-of-the-art restoration approach. The restored images not only exhibit improved illumination consistency and structural clarity, but also significantly enhance downstream diabetic retinopathy classification performance. Notably, in many cases, the generated outputs are even superior to the labeled high-quality reference images, exposing the limitations of existing ground truth annotations. These results establish the proposed framework as an effective and practical solution for high-fidelity fundus image restoration in real-world clinical settings.

Video periocular recognition is the task of recognizing an individual's identity based on the region around an individual's eyes. The periocular area is one of the most discriminative regions of the human face, making it suitable for recognition tasks. Its use as a biometric modality has emerged as an alternative, especially in surveillance scenarios where conventional biometric traits such as face or iris recognition become unfeasible due to unconstrained acquisition conditions. This paper proposes an attention-aware approach for video-based periocular recognition in surveillance environments. The framework consists of two main modules: feature embedding and aggregation. The feature embedding module is a deep convolutional neural network that maps periocular data to feature vectors. The aggregation module is an encoder-only transformer that adaptively learns to aggregate frame-level features into a single video representation and a feature vector for the still reference image. Experiments on the publicly available COX Face dataset show the robustness of the proposed method, consistently outperforming naive aggregation schemes. In the best scenario, the approach achieves 99.8% of TPR@1e-1 and 96.6% of Rank-5.

3D Gaussian Splatting (3D-GS) enables real-time 3D scene reconstruction but lacks robust segmentation for editing tasks such as object removal, extraction, and recoloring. Existing approaches that lift 2D segmentations to the 3D domain scene suffer from view inconsistencies and coarse masks. In this paper, we propose a novel framework that leverages the Segment Anything Model – High Quality (SAM-HQ) to generate accurate 2D masks, addressing the limitations of the standard SAM in boundary fidelity and fine-structure preservation. To achieve robust 3D segmentation of any target object in a given scene, we introduce a prior-guided label reassignment method that assigns labels to 3D Gaussians by enforcing multiview consistency with learned priors. Our approach achieves state-of-the-art segmentation accuracy and enables interactive, real-time object editing while maintaining high visual fidelity. Qualitative results demonstrate superior boundary preservation and practical utility in Virtual Reality (VR) and robotics, advancing 3D scene editing.

The exponential growth of visual data across digital systems offers challenging problems for efficient data processing, for scalable data storage and for semantic understanding in big image and video data. Image captioning, positioned at the intersection of computer vision and natural language processing, offers a route to convert large-scale image repositories into meaningful textual understandings. Nonetheless, existing vision language models frequently face scalability issues and struggle with fine-grained object recognition and relational reasoning in diverse domains. To overcome these limitations, we develop an attention-driven image captioning framework which is uniquely designed for large scale data environments. The proposed framework integrates scene attention, object attention, and graph interaction attention modules built on BLIP embeddings and a selectively fine-tuned GPT-2 language model, enabling efficient contextual learning and robust cross-domain generalization. We experimentally evaluate the proposed framework on MSCOCO and NoCaps benchmarks demonstrate a relative improvement of 10.8\% in CIDEr and a gain of 3.6\% in SPICE over ClipCap, with a relative increase of 39.7\% in CIDEr for out-of-domain generalization. Overall, the proposed framework shows a scalable solution for extracting structured, semantically rich information from large-scale multimodal datasets, advancing the integration of big data analytics and vision language modeling.

Probe-based confocal laser endomicroscopy (pCLE) enables in-vivo assessment of microscopic disease activity in ulcerative colitis (UC) by visualizing intestinal barrier impairment. Although deep learning offers potential for pCLE-based UC diagnosis, most existing methods operate at the frame level, ignoring the temporal structure inherent to video sequences. To address this limitation, we propose Multi-scale Temporal Semantic Prototypes for multiple instance learning (MTSP-MIL), a novel time- and semantically-aware weakly-supervised model for pCLE video classification. Specifically, we integrate multi-scale temporal aggregation, learnable visual-semantic prototypes, and a transformer encoder with attention pooling to yield a robust video representation capturing local, global, and semantic dynamics. MTSP-MIL operates on frame-level features extracted by a foundation model. Empirical results validate the superior encoding capabilities of DINOv2 for pCLE and demonstrate that MTSP-MIL outperforms state-of-the-art MIL methods. Qualitatively, the learned prototypes effectively decouple anatomical semantics from noise, validating model interpretability.

No-Reference Image Quality Assessment (NR-IQA) remains challenging because pristine references are typically unavailable in real-world scenarios. Inspired by the internal generative mechanism of the Human Visual System (HVS), we propose a pseudo Full-Reference Image Quality Assessment (FR-IQA) approach that replaces the missing ground truth with a restoration-generated proxy. Our framework employs a distortion-specific SwinIR backbone to estimate a proxy reference image, enabling the direct use of established FR metrics in a no-reference setting. This modular, plug-and-play design decouples restoration from quality scoring, allowing classical or learned FR-IQA metrics to be integrated without retraining the overall pipeline. We evaluate the approach on CSIQ, TID2013, and KADID-10k under Additive Gaussian Noise (AGN) and JPEG compression. Results show that FR metrics used with our proxy references consistently outperform strong NR-IQA baselines, with notably improved ranking consistency. Overall, the analysis confirms that even imperfect restoration proxies provide a sufficient structural anchor to stabilize quality prediction, bridging the gap between blind and full-reference assessment.

Unsupervised 3D LiDAR semantic segmentation enables scalable driving-scene understanding from abundant unlabeled point clouds. However, most existing methods process each scan independently and ignore temporal continuity. To address this issue, we propose TAUSS (Temporally Aligned Unsupervised Semantic Segmentation), which introduces temporal consistency by matching superpoints as stable regions across consecutive scans. TAUSS estimates these superpoint correspondences using geometric statistics and motion cues, then aligns semantic primitive assignment distributions of matched superpoints to stabilize pseudo-labels under viewpoint changes and partial observations. On SemanticKITTI and nuScenes LiDARSeg, TAUSS achieves state-of-the-art performance among unsupervised methods without distillation from pretrained foundation models, improving the scan-wise base framework by +3.3 and +4.0 mIoU points on the respective validation splits.

Contrastive Language-Image Pre-training (CLIP) has demonstrated impressive capabilities in zero-shot transfer but often struggles with dense prediction tasks due to low spatial resolution and the loss of structural information. To address these limitations, we propose MARS-CLIP (Multi-resolution and Attention Refined Segmentation for CLIP), a novel framework for zero-shot semantic segmentation. Our approach introduces two key strategies: (i) a multi-resolution feature extraction module that fuses local fine-grained features with global context to overcome input resolution constraints, and (ii) an attention refinement mechanism that injects spatial and color biases from intermediate layers into the final self-attention block to accurately restore object boundaries. A set of experiments on six public datasets demonstrates that MARS-CLIP significantly outperforms state-of-the-art methods.

Existing deep image compression methods typically employ a single encoding path to directly compress image and indiscriminately extract visual features without using frequent domain information. To this end, we propose a structure-texture guided image compression method based on cross-scale interaction and frequency enhancement. First, we introduce a multi-scale mean filtering to decompose image as structure and texture components. Second, we propose a triple-branch encoding architecture in the encoder network that explicitly leverages the decomposed structure and texture components as well as original image as the inputs for complementary multi-stage fusion. Third, we propose a dual-domain octave residual transform within encoder and decoder networks, in which group differentiation enhancement unit is designed for efficient feature interaction. Experimental results on benchmark datasets demonstrate that the proposed method achieves state-of-the-art performance, significantly improving objective reconstruction quality for image compression.

Feedforward 3D Gaussian Splatting enables fast and high-quality scene reconstruction, but its dense Gaussian representations incur substantial storage and transmission overhead, limiting deployment in bandwidth-constrained and mobile scenarios. Existing compression approaches, such as FCGS, rely on training-aware neural codecs and introduce non-negligible decoding complexity that is often impractical for edge devices. We present Stream-FCGS, a lightweight compression and streaming framework for feedforward 4D Gaussian Splatting. Stream-FCGS reduces redundancy via progressive pruning, then compresses Gaussian geometry and attributes with a disentangled, codec-friendly design. We further build a web-based 6DoF renderer to support interactive playback on mobile devices. Extensive experiments show that Stream-FCGS improves rate–distortion performance over state-of-the-art methods while preserving high-fidelity rendering quality. It also supports real-time 6DoF interaction on end devices, indicating strong potential for immersive 3D streaming in bandwidth-constrained mobile and XR settings.

Learned image compression has seen significant progress in recent years with the development of end-to-end learned models achieving higher compression rates than even the best conventional methods. Recently, Mixture of Experts (MoE) approaches have seen promising results in NLP and computer vision tasks. In this paper, we introduce the MoE approach to learned image compression. We propose a MoE-based Entropy model (MoEE) for learned image compression, allowing the model to selectively activate only the subset of parameters required for the current image. Our model achieves a BD-Rate improvement over VVC of -16.85% on the Kodak images.

The demand for high-quality talking head video compression has grown with the rise of remote conferencing and digital education. Generative Face Video Compression (GFVC) has recently been explored to enable fast encoding and high compression efficiency by using compact facial motion features. However, existing GFVC approaches fail to handle dynamic facial changes and rely on manually tuned rate control, limiting their generalizability in real-world scenarios. To address these problems, we propose a novel facial feature–guided adaptation for GFVC. Our method introduces Facial Feature–aware Keyframe Selection, which adaptively selects keyframes robust to facial expression changes. We also perform optimal bit allocation to reconstruct high-quality videos within the available bitrate.  Experimental results demonstrate that our method achieves average bitrate savings of 17.82% over the latest GFVC approach. Moreover, our rate control mechanism achieves a 94.9% reduction in encoding time compared with the conventional video codec VVC.

Autoencoder-based image codecs achieve state-of-the-art compression performance but often incur high computational complexity, particularly at decoding time. This work introduces a low-complexity learned image compression framework based on spatial competition between multiple specialized neural codecs. For each image region, the encoder selects the codec that best matches the local content according to a rate-distortion cost. A mode map is transmitted as side information to indicate the per-region codec selection. At decoding time, this mode map-based selection guides reconstruction while preserving the complexity of a single codec. This design enables per-image adaptation with low decoding complexity and fast encoding. On the CLIC 2020 dataset, our method achieves up to −14.5% rate reduction compared to a single codec and reaches HEVC-level performance with a decoding complexity of 1433 MACs per pixel.

The Video-based Dynamic Mesh Coding (V-DMC) standard, developed by the Moving Picture Experts Group (MPEG), represents a state-of-the-art approach for compressing dynamic meshes. A challenge within the V-DMC framework is the inherent trade-off between reconstruction quality and bitrate: while increasing the number of subdivisions improves the quality of the reconstructed mesh, it also leads to substantially higher bitrate requirements. This constraint hinders further quality enhancement through finer subdivision levels. To overcome this limitation, this paper proposes a lifting-based geometry optimization method that enables the reconstruction of high-quality meshes using higher subdivision levels while only requiring the encoding of data from a lower subdivision level. Specifically, an enhanced lifting scheme is introduced that utilizes information from a higher subdivision level to refine the mesh representation at a lower level. Experimental results demonstrate that the proposed method significantly improves the quality of the reconstructed mesh without substantially increasing the bitrate. The efficacy and efficiency of the algorithm have led to its adoption and integration into the MPEG V-DMC reference software.

Rate control (RC) is critical for achieving accurate bitrate in video compression and transcoding. However, existing RC methods for learned video compression (LVC) often rely on history-driven parameter updates, which can lead to large bitrate errors under abrupt content variations. Moreover, quality fluctuation at both frame and block levels is rarely addressed, despite its importance to the visual experience. To tackle these issues, a rate-quality (R-Q) prediction-based RC scheme for learned video transcoding is presented. Specifically, a Coding Priors-Guided R-Q Prediction Network (CP-RQNet) is designed to predict frame- and block-level R-Q relationships by leveraging coding priors (e.g., residuals and motion vectors) parsed from the input bitstream. Built on these predictions, a Quality Smoothing RC (QSRC) algorithm is proposed to allocate bits and refine parameters. Distinctively, this algorithm incorporates a pre-traverse pass for global quality smoothing across the entire GOP, and extends quality consistency to the block level. Furthermore, an error-adaptive weighted combination mechanism is utilized to dynamically balance historical information and current predictions, thereby enhancing the accuracy of RC parameters. Using H.264/AVC transcoding as a representative case, experiments on DCVC-DC and DCVC-HEM across multiple datasets demonstrate that the proposed scheme achieves consistently lower bitrate errors, improved rate-distortion performance, and reduced quality fluctuations compared with existing methods.

Generative real-time interaction systems, such as digital avatars and AI-mediated video calls, produce media content online under strict latency and throughput constraints. Unlike conventional real-time communication systems, frame generation is computationally intensive and capacity-bounded, while audio features become available earlier and provide limited lookahead for short-term control under network variability. We propose HGS, a Head-of-Queue (HoQ) slack-aware generation scheduling framework that incorporates content generation into the control loop to adapt to network variability. Trace-driven experiments on a prototype generative real-time interaction system show that, compared with the best fixed-rate baseline selected by P.1203 O46, HGS reduces danger rate and deadline miss rate by up to 69.4 pp and 11.4 pp, improves effective frame rate by up to 25.4%, improves P.1203 O46 by up to 6.5%, and remains robust across broader evaluation settings.

Rate control in mesh coding is usually achieved by varying the quantization bit depths of geometric positions and other attributes. In addition, mesh simplification and remeshing techniques help approximate meshes with fewer vertices and faces, thereby reducing bitrates. Although these methods offer some rate-distortion trade-off, a formal rate-distortion optimization (RDO) framework is absent in mesh codecs. This work introduces an RDO framework for mesh geometry coding. Inspired by RDO in video codecs, it provides a way to select a prediction mode that minimizes the RD cost among candidate modes. The proposed method is implemented in the AOMedia polygonal mesh codec (PMC) encoder with an average BD-Rate gain of more than -3\%. It can also be seamlessly integrated into any mesh encoder that uses entropy coding to encode vertex positions.

A parametric video coder-decoder (codec) is one where the performance is represented by a curve either in the plane or in higher-dimensional space. We propose a process for comparing parametric video compression systems through a rate-distortion-complexity (RDC) analysis. We discuss the popular 2D Bjontegaard delta metric and other correspondence-based curve-comparison methods, and their extensions to allow comparisons in RDC space. We argue that all these metrics in 3D may be ineffective and we propose an absolute cost for each RDC curve. Such a metric involves the computation of Lagrangian costs of RDC points in the form D+λ R + γC, and some integration. We also argue that any application may be associated to a (λ,γ) pair. Several state-of-the-art neural video codecs were compared, using the proposed metrics for the entire (λ,γ) plane for each codec through all the available dataset. We could, then, establish the best codecs for each application, i.e. for each (λ,γ) pair. Tests with 19 codecs (17 neural and 2 conventional video codecs) revealed that only 6 codecs would be of interest, each best suited for a range of applications.

Compression of 3D Gaussian Splatting (3DGS) is essential for efficient storage and delivery due to its large data size. Furthermore, scalability is a powerful feature for 3DGS compression in real-world applications with heterogeneous devices. Since more than 80\% of the raw 3DGS data size is occupied by Spherical Harmonic (SH) coefficients, this paper, for the first time, proposes an effective degree-wise scalability of SH coefficients based on the G-PCC v1 Amendment, an MPEG standard for 3DGS compression. The proposed approach includes a Base Layer (BL) for low-degree SH coefficients and several Enhancement Layers (ELs) for higher-degree SH coefficients. A key challenge is that directly truncating higher SH degrees leads to significant degradation in rendering quality. To address this, we design a Degree-Wise Compression-Aware Progressive (DWCAP) fine-tuning method for SH coefficients, which is applied layer by layer at the encoder and operates as a codec-agnostic pre-processing step}. The proposed DWCAP fine-tuning has three main features: (1) rendered images are used as supervision in a compression stage that is independent of the original 3DGS generation process; (2) a rate–distortion loss function is employed, jointly considering distortion and bitrate; and (3) the fine-tuned lower-degree coefficients are frozen when fine-tuning higher-degree SH coefficients, which ensures the feasibility of SH scalability. Evaluations on the MPEG test dataset demonstrate that the proposed approach achieves not only SH scalability but also significant coding gains compared to the single-layer G-PCC v1 Amendment.

Gaming and game streaming videos typically contain a dynamically changing background overlaid with relatively stable text while imposing stringent requirements on compression efficiency and low latency. Due to the distinct motion characteristics of text and background layers within the same coding block, conventional block-based coding schemes often fail to model motion accurately, leading to suboptimal coding performance. To address this issue, this paper proposes a text-aware layered inter prediction algorithm within the H.266/VVC framework. A binary text mask is first generated through text pixel detection, based on which independent motion estimation (ME) and motion compensation (MC) are performed for the text and background layers. The resulting prediction blocks are then fused under mask guidance. Integrated into H.266/VVC test model version 23.0 and evaluated under the Low Delay configuration, the proposed method achieves average bitrate savings of 2.32%, 3.57%, and 3.36% for the Y, U, and V components, respectively.

Scalable compression is essential for bandwidth-adaptive transmission, yet most learned codecs are optimized for a fixed rate-distortion point, making rate adaptation costly due to re-encoding or maintaining multiple bitstreams. In this work, we propose TAFA-GSGC, a scalable learned point cloud geometry codec that enables multi-quality decoding from a single bitstream and a single trained model. TAFA-GSGC combines layered residual refinement with channel-group entropy coding, and introduces a Target-Aligned Feature Aggregation module to reduce cross-layer redundancy in enhancement residuals. Our framework supports up to 9 decodable quality levels with monotonic quality improvement as more sub-bitstreams are received, while maintaining strong compression efficiency. Compared with the PCGCv2 baseline, TAFA-GSGC demonstrates improved RD performance, achieving average BD-rate reductions of 4.99% and 5.92% in terms of D1-PSNR and D2-PSNR, respectively.

Gaussian splatting is a fast-growing technology allowing for real-time and photorealistic rendering of captured scenes or objects with six-degrees-of-freedom (6DoF) interactivity. Recent advances show that mobile devices can not only render Gaussian splats in real-time but also capture and learn them through on-device processing or cloud services. This paves the way towards distribution and sharing via social networks of user-generated volumetric content with photorealistic quality. Key requirements include interoperability and high-fidelity, efficient compression to reduce bandwidth and storage demands for Gaussian splats. This paper presents the ongoing standardization efforts at MPEG for coding Gaussian splats using V3C/V-PCC and provides current compression performance results.

Non-reference metrics (NRMs) can assess the visual quality of images and videos without a reference, making them well-suited for the evaluation of user-generated content. Nonetheless, rate-distortion optimization (RDO) in video coding is still mainly driven by full-reference metrics, such as the sum of squared errors, which treat the input as an ideal target. A way to incorporate NRMs into RDO is through linearization (LNRM), where the gradient of the NRM with respect to the input guides bit allocation. While this strategy improves the quality predicted by some metrics, we show that it can yield limited gains or degradations when evaluated with other NRMs. We argue that NRMs are highly non-linear predictors with locally unstable gradients that can compromise the quality of the linearization; furthermore, optimizing a single metric may exploit model-specific biases that do not generalize across quality estimators. Motivated by this observation, we extend the LNRM framework to optimize ensembles of NRMs and, to further improve robustness, we introduce a smoothing-based formulation that stabilizes NRM gradients prior to linearization. Our framework is well-suited to hybrid codecs, and we advocate for its use with overfitted codecs, where it avoids iterative evaluations and backpropagation of neural network–based NRMs, reducing encoder complexity relative to direct NRM optimization. We validate the proposed approach on AVC and Cool-chic, using the YouTube UGC dataset. Experiments demonstrate consistent bitrate savings across multiple NRMs with no decoder complexity overhead and, for Cool-chic, a substantial reduction in encoding runtime compared to direct NRM optimization.

The Local Warped Motion Compensation (LWMC) tool in the SVT-AV1 encoder provides significant compression gains for sequences with coherent motion but introduces substantial computational overhead. The default LWMC selection is conservative, often underutilizing its potential to limit complexity. In this work, we propose a content-adaptive LWMC configuration strategy that balances compression efficiency and encoding cost. A Random Forest classifier trained on motion-related features categorizes each video sequence into one of four motion regimes (static, global, local, or composite dynamic) and drives a decision algorithm that selects the most appropriate LWMC configuration (OFF, DEFAULT, or ON) according to a quality-complexity tradeoff metric defined as the ratio of BD-rate gain to additional encoding cycles. Evaluated on a set of representative video sequences, the proposed New Mode achieves up to 81% of the VMAF BDRate gain of the full LWMC (ON mode), while consuming only 24–36% of its computational overhead. Compared to the Default mode, it delivers up to 2,6× more BD-Rate gain at a marginal additional cycle cost. These results demonstrate that contentadaptive tool  selection is an effective strategy for efficient video encoding in SVT-AV1.

Film grain preservation in cinematic video is increasingly handled through analysis–remove–encode–synthesize pipelines, where grain is removed prior to compression and restored at the decoder from signaled metadata. The technical bottleneck is not synthesis alone, but the accurate estimation of film grain parameters from real grainy content in a form that is standard-compliant, low complexity, and robust enough for practical deployment. This work presents a conventional film grain analysis (FGA) method that estimates standards-compatible film grain parameters directly from grainy video and outputs metadata in the Film Grain Characteristics SEI (FGC-SEI) format used by VVC frequency-filtering synthesis. The work emphasizes parameter estimation rather than grain generation alone, and targets faithful reconstruction of the original grain appearance together with substantial bitrate reduction.

This paper addresses the problem of selecting between LCEVC and its base codec by predicting the intersection point of their rate-distortion (RD) curves. Conventional approaches require multiple encoding trials to identify this switching point, resulting in high computational cost. We propose a content-adaptive method to predict the intersection point prior to encoding using video features. Experimental results show a strong negative correlation (r =−0.93) between Spatial Perceptual Information (SI) and the PSNR at the intersection point. The proposed approach provides a practical basis for efficient codec selection in adaptive video encoding systems.

We propose a scalable face avatar communication system combining MPEG Video-based Dynamic Mesh Coding (V-DMC) with the Avatar Representation Format (ARF). The system operates in two modes: a low-bitrate baseline streaming blendshape activation weights and a high-quality enhancement mode transmitting V-DMC inter-coded mesh sequences that capture fine details beyond the blendshape basis. V-DMC compression of ARF avatar assets provides significant size reduction at near-lossless quality. Built on web technologies, the system enables adaptive real-time face avatar communication with automatic mode switching based on available bandwidth. The high-quality streaming mode has been implemented; the full integrated dual-mode architecture is under development.

Missing MRI modalities are common in clinical imaging and significantly reduce performance in multi-modal segmentation. In this work, we propose GeoScaffold, a feature-space restoration framework that manages arbitrary missing-modality patterns without pixel-level synthesis. Our approach features three main components: (i) coordinate-conditioned geometric scaffolding that injects Fourier-encoded spatial priors at the bottleneck to stabilize topology when intensity cues are weak; (ii) hyperspherical directional distillation that aligns student-teacher features through scale-invariant cosine similarity, reducing sensitivity to magnitude attenuation; and (iii) a hybrid blindfold curriculum that balances full-modality accuracy with missing-modality robustness. On BraTS 2021, GeoScaffold achieves a mean Dice of 87.00% with complete inputs and 77.01% averaged over 14 missing-modality scenarios, surpassing standard CNN baselines by over 4 points. Code is available at https://github.com/David-SUSTech/GeoScaffold.git.

Magnetic Resonance Imaging translation is crucial for multi-modal data acquisition, yet existing methods are often limited in preserving fine anatomical details. To address this, a novel framework that integrates explicit texture and structure priors is proposed. First, an initial translation is generated by a regression model. Subsequently, a dynamically adaptive probability map is introduced to model complex textural and structural regions. This map, along with the initial translation, is utilized as a condition for a diffusion-based generative model to refine anatomical details. Finally, the outputs of both models are fused under the guidance of the probability map to produce the high-quality translated image. The superiority of the proposed method in preserving image fidelity and enhancing visual quality is demonstrated through extensive experiments. The code will be made publicly available.

Accurate parcellation of brain regions from diffusion MRI (dMRI) is fundamentally challenged by low tissue contrast and the scarcity of subject-level annotations. We propose RefParcel, a novel deep learning framework that leverages a single standard atlas to achieve precise, individualized fine-scale parcellation. By integrating nonlinearly registered spatial priors with multi-channel diffusion features into a U-Mamba network enhanced with Spatial-Adaptive Channel Attention, RefParcel rectifies registration-induced deformations. This learned capability is transferred to target atlases lacking individual gold standards, enabling accurate zero-label cross-atlas parcellation. RefParcel achieves a Dice score of 0.798 on the source atlas (outperforming SOTA by 7.98%, p

Advanced diffusion MRI-based microstructure imaging holds clinical promise but involves lengthy multi-shell acquisition protocols. Deep learning predictions overlook high-frequency details and perceptual quality, leading to a loss of detail in brain-structure reconstructions. This paper proposes a novel spatial-frequency information diffusion model that efficiently synthesizes multiple high-fidelity metrics from conventional single-shell dMRI. Our model is based on Denoising Diffusion Probabilistic Models and introduces two key innovations: 1) a streamlined wavelet sampler module that uses multi-scale spatial convolution and 3D discrete wavelet transform in parallel to capture macroscopic anatomical structure and critical high-frequency texture details simultaneously. 2) a structure-aware loss that integrates normalized MSE and HaarPSI loss to synergistically optimize quantitative accuracy and structural fidelity. Results show that SF-Diff achieves state-of-the-art performance across multiple metrics, demonstrating its effectiveness and superiority in the dMRI metric synthesis. Code is available at https://github.com/csx1009/SF-Diff.

Predicting fluid intelligence from brain connectivity is crucial for understanding the neural basis of human cognition and for the assessment of cognitive health in clinical and population studies. Recent multimodal learning approaches that integrate functional connectivity (FC) and structural connectivity (SC) have demonstrated strong predictive performance. However, the SC is often unavailable in clinical practice. Existing hallucination-based methods transfer multimodal knowledge to a unimodal network via distillation, but they suffer from two limitations: 1) Feature-based distillation treats all feature channels equally, causing the student to be distracted by task-irrelevant features thereby failing to capture important discriminative patterns. 2) Logit-based distillation often introduces supervision conflicts between the ground truth (GT) and the teacher’s predictions, confusing the student model's optimization process. In this work, we propose a Multimodal Brain Hallucination Network (MBHNet) to enable robust fluid intelligence prediction when SC is unavailable. Specifically, we introduce the Discrimination-Aware Module (DAM) and the Conflict-Free Module (CFM) to address the aforementioned limitations. First, the DAM employs a Top-K strategy with rank-adaptive scaling to identify and enhance dominant channels, strictly guiding the student to focus on discriminative patterns. Second, the CFM forwards student features through the teacher’s regression head to align with the teacher’s predictions, resolving the supervision conflict found in logit-based distillation. Extensive experiments on the UK Biobank dataset demonstrate that MBHNet outperforms state-of-the-art methods under missing SC, providing a robust solution for fluid intelligence analysis.

We introduce DINO-Detector, a novel framework for one-shot 3D craniofacial landmark detection. The proposed approach leverages the semantic representations of a pretrained vision foundation model to generate high-quality pseudo-labels. Specifically, DINO features are used to establish dense pixel-level correspondences through a dual-branch feature extractor that captures both global contextual information and fine-grained volumetric structure. A learnable volumetric registration model then efficiently produces pixel-level correspondences and pseudo-labels for 3D landmarks. To further improve robustness, we exploit the noise-tolerant learning behavior of neural networks and propose an landmark detector with adaptive spatial correlation regularization, enabling accurate and efficient 3D landmark inference. Experimental results on clinical craniofacial volumetric imaging datasets demonstrate consistent performance improvements in 3D craniofacial landmark detection.

Low-field (LF) brain Magnetic Resonance Imaging (MRI) suffers from degraded image quality that limits its clinical usability. Existing LF to high-field (HF) enhancement methods are mainly optimized for uniform reconstruction fidelity, which often oversmooths subtle and localized abnormalities. We propose a report-aware latent diffusion framework for LF brain MRI enhancement that incorporates high-level clinical semantics into the enhancement process. Diagnostic reports are used during training to learn a shared vision-language embedding space via CLIP-style alignment. The learned semantic representations are injected into a latent diffusion model through cross-attention to modulate the denoising trajectory in a context-dependent manner. This enables selective enhancement of diagnostically relevant structures while preserving global anatomical consistency. The proposed method operates solely on LF MRI at inference time and requires no reports or annotations. Experiments demonstrate improved perceptual quality and structural preservation over image-only diffusion baselines.

Cardiovascular pulsations drive both cerebral blood flow (CBF) and glymphatic cerebrospinal fluid (CSF) convection. While respiratory pulsations drive venous outflow from the brain and modulate cardiovascular pulsations according to an updated Monro-Kellie doctrine. Cardiovascular propagation driving CBF and glymphatic CSF are altered in Alzheimer's disease (AD), but it is unknown if respiratory pulsations influence pathological cardiovascular pulse propagation. Here we acquired ultrafast magnetic resonance encephalography (MREG) data from 31 AD and 27 control subjects and conducted a three-dimensional multi-resolution optical flow analysis to quantify cardiovascular pulse propagation speed (CPPS) at different respiratory phases and stages. Mean CPPS values during inspiration and expiration did not differ substantially between control and AD groups, but CPPS variation was higher in AD patients due to greater values around both inspiratory and expiratory peaks and troughs. No differences were detected during the middle stages of these respiratory phases. Variance changes did not significantly co-localize with gray matter atrophy but rather centered around periventricular structures of the glymphatic system. Therefore, respiration contributes to increased variance of cardiovascular brain impulse propagation in AD, indicating an interaction of cardiorespiratory brain pulsations in driving (peri)vascular CSF convection.

Autism Spectrum Disorder (ASD) exhibits substantial heterogeneity, with communication impairments strongly influencing functional outcomes. Conventional structural MRI (sMRI) studies typically rely on conventional ROI-level mean features that obscure localized morphological variations linked to behavioral severity. Utilizing the ABIDE-II dataset with FreeSurfer preprocessing, this study evaluates increasing cortical granularity—from ROI means and percentile-based descriptors to spatially-aware geodesic partitions—while concurrently comparing cortical morphology against subcortical volumetric features. We develop a complete pipeline including image preprocessing, feature extraction and selection, and multiple machine learning architectures. The highest performance is achieved using the ridge classifier. Spatially-aware cortical partitioning yields peak results with F1-scores (and AUCs) of 78.7% (95.8%) for mild, 72.3% (90.3%) for moderate, and 62.5% (81.3%) for severe categories, significantly outperforming conventional ROI-level mean representations. In contrast, subcortical features show limited predictive value (~61.5% AUC) and fail to improve combined models, indicating that the cortical surface represents the more informative anatomical substrate for communication-related severity in ASD and that topological subdivision of regions is critical to capturing the localized structural markers necessary for accurate severity stratification.

Multi-modal Magnetic Resonance Imaging (MRI), which is pivotal for comprehensive brain analysis, frequently suffers from missing modalities in clinical settings. Current synthesis methods, which often produce anatomical hallucinations, fail to preserve 3D structural consistency. To address these issues, we propose Swin-Control-LDM, which is a novel dual-stream latent diffusion framework. We introduce a Global-Local Spatial Adaptive Normalization (GL-SPADE) module, which explicitly disentangles anatomical structure from modality style in the latent space. Furthermore, we construct a 3D Swin-Control backbone, which integrates Swin Transformers for long-range dependency modeling. This backbone leverages a ControlNet with zero-convolution injection, which imposes strict structural constraints. Extensive experiments, which utilize the BraTS 2021, BraTS 2023, and IXI datasets, demonstrate that our method achieves state-of-the-art performance. Our approach generates high-fidelity missing modalities, which maintain anatomically consistent structures.

Accurate brain image segmentation is essential for computer-aided diagnosis but remains challenging due to low contrast, noise, and poorly defined boundaries. Existing convolutional and Transformer-based methods leverage multi-scale representations and attention mechanisms, but often suffer from limited long-range spatial reasoning, imprecise boundary localization, and reduced robustness across imaging conditions. To address these limitations, we propose HTGBNet, a hybrid Transformer–Graph framework for boundary-aware medical image segmentation. HTGBNet employs a Pyramid Vision Transformer (PVT) encoder to capture hierarchical multi-scale features, followed by a graph convolution module to model long-range spatial dependencies. In addition, a Boundary Feature Enhancement (BFE) module explicitly strengthens edge representations, while a multi-scale decoder with deep supervision improves optimization stability. Experiments on public benchmarks demonstrate that HTGBNet consistently outperforms state-of-the-art methods, achieving higher Dice scores and lower HD95 values, indicating improved boundary delineation and segmentation accuracy.

Abstract—Magnetic Resonance Imaging (MRI) plays a crucial role in brain disease diagnosis by providing high-resolution anatomical information. However, MRI scans are frequently affected by acquisition noise, low contrast, and reconstruction artifacts, which degrade visual quality and reduce diagnostic reliability. Conventional enhancement techniques based on spa- tial filtering and histogram equalization often fail to preserve fine anatomical structures while suppressing noise. To address these limitations, this paper presents HMDER-AttnNet, a hybrid attention-based deep learning framework for noise-aware brain MRI image enhancement and restoration. The proposed frame- work integrates classical preprocessing strategies, realistic MRI noise modeling, and an attention-augmented U-Net architecture incorporating Convolutional Block Attention Modules (CBAM) to improve feature discrimination and structural preservation. Ex- tensive experiments were conducted on three public benchmark datasets, namely BraTS, IXI, and BrainWeb. Quantitative results demonstrate that HMDER-AttnNet achieves a PSNR of 30.2 dB and SSIM of 0.892 on the BraTS dataset, 31.6 dB PSNR and 0.907 SSIM on the IXI dataset, and 33.1 dB PSNR with 0.924 SSIM on the BrainWeb dataset, outperforming conventional filtering techniques and state-of-the-art deep learning baselines. The experimental findings validate the effectiveness, robustness, and generalization capability of the proposed framework, highlighting its suitability as a reliable preprocessing solution for clinical brain MRI analysis and downstream diagnostic applications.

Segmenting brain tumors in magnetic resonance images is key in clinical care. However, if manual, it is time-consuming and prone to human bias. We tackle this issue, and propose a zero-click segmentation pipeline utilizing foundation models. It employs a 3D U-Net to generate initial segmentations, serving as automatic prompts for the Segment Anything Model 2, which independently refines tumor region boundaries for each modality. It treats axial slice stacks as video sequences. Modality-specific refined masks are subsequently integrated into a 3D segmentation through a feature aggregation network that resolves boundary conflicts and fuses the derived information. Quantitative evaluation over the benchmark demonstrates that 3D U-Net achieves a mean validation Dice of 0.625, over three tumor sub-regions, which increases to 0.713 using our method (p

Detection of alveolar fenestration and dehiscence with cone beam computed tomography is still very difficult. This is mainly because of very small sizes of the defects that are even less than one millimeter and there is also a variability of the interpretation of the results by different observers (κ = 0.42:0.68). We introduce a method that uses a texture-aware vision transformer which maximizes the mutual information between multi-scale local binary pattern features and the pathology labels via dual-radius fusion (D1=1, D2=2), thus producing an information-theoretically optimal three-channel representation. Evaluated on 775 mandibular canine cross-sections, our EVA-02-based framework achieves 98.54% accuracy (CI=[97.66%, 99.12%]) and 97.79% F1-score; surpassing ResNetV2-BiT by 0.79% while reducing variance by 33%. The model demonstrates clinical non-inferiority to 3 board-certified periodontists (p

Resting-state fMRI (rs-fMRI) provides rich temporal dynamics of brain activity. Component analysis methods, such as principal component analysis (PCA) and independent component analysis (ICA), are widely used to extract functional patterns, but they yield largely static representations and do not explicitly model the temporal evolution of the fMRI signal. Here, we investigate Dynamic Mode Decomposition (DMD) as a dynamic extension of conventional component analysis for characterizing Parkinson’s disease (PD)-related brain dynamics. DMD decomposes rs-fMRI time series into coherent spatial modes coupled with explicit temporal dynamics, enabling systematic interrogation of disease-sensitive spatio-temporal patterns. Experiments demonstrate that DMD-derived dynamic modes provide stronger discriminative power for PD detection than PCA/ICA-based representations, supporting the value of modeling temporal evolution in addition to spatial structure. Overall, DMD offers a promising tool for spatio-temporal analysis of brain functional dynamics.

While generative models enable encoding of complex neuroimaging data for feature generation and reconstruction, developing optimal architectural frameworks with appropriate encoding and latent space processes is crucial for studying structural and functional properties of the brain. We design a multimodal generative framework for structural and functional magnetic resonance imaging (MRI) features through systematic evaluation of encoding strategies, latent multimodal fusion, and generative model selection. Using structural gray matter volume (GMV) and static functional network connectivity (sFNC) features from a large neuroimaging dataset, we analyze generative frameworks involving variational autoencoders (VAEs), transformers, generative adversarial networks (GANs), and diffusion models. Architectures that employ modality-aware graph encoding of functional connectivity into a lower-dimensional latent space outperform vectorized encoders or direct data space approaches. The proposed multimodal graph VAE (gMMVAE) surpasses alternative generative variants across multiple metrics for generation fidelity, reconstruction quality, efficiency, and latent space discriminability, highlighting its potential for robust multimodal neuroimaging analysis.

We present computational single-shot quantitative phase and polarimetric imaging using only a polarization camera in a standard DIC microscope. Our new QPI algorithm enables artifact-free, single-shot phase maps, while an additional Stokes formalism recovers pixel-resolved retardance and molecular orientation. The method enables sustained, high-speed observation of dynamic cellular processes at quantitative contrast, without phototoxicity or temporal limitations.

Arterial spin labeling (ASL) magnetic resonance imaging is a noninvasive renal perfusion imaging technique that does not require exogenous tracers. However, due to the inherently weak perfusion signal, the original images generally suffer from extremely low signal-to-noise ratio (SNR) and spatial resolution, making it very difficult to extract the boundary of the Renal cortex. This study proposed a collaborative segmentation scheme. First, anisotropic diffusion filtering (ADF) is applied to suppress the background noise of ASL images, and then a two-stage cascaded nnUNet framework is used for segmentation. The research results show that this method can effectively improve segmentation accuracy.

Positronium lifetime tomography (PLT) estimates spatially varying ortho-positronium (o-Ps) lifetime distributions from time-of-flight (TOF) PET measurements. In this work, we develop a unified variance analysis framework for PLT using histogram-based and moment-based lifetime estimators. The framework is based on the ex-Gaussian mixture timing model commonly used in positronium lifetime estimation. We compare continuous-time Fisher-information bounds, histogram likelihood estimators, and generalized method-of-moments (GMM) estimators using both power moments and exponentially weighted Laguerre moments. The analysis shows that non-uniform histogram binning improves variance efficiency compared with uniform bins at fixed bin count, while weighted Laguerre moments achieve lower finite-order variance than raw power moments. Both histogram and moment estimators converge toward the continuous-time Fisher limit as the number of bins or moments increases. These results provide theoretical guidance for designing efficient and low-noise PLT reconstruction algorithms.

Automated knee cartilage segmentation in MRI is clinically important for osteoarthritis assessment but remains challenging because cartilage is thin, curved, and anatomically variable. We propose a hybrid framework combining deformable multi-atlas segmentation (MAS) and deep learning (DL) with Joint Label Fusion (JLF). Three registration-based MAS strategies, symmetric normalization (SyN), free-form deformation (FFD), and symmetric diffeomorphic demons (SDD), are integrated with multi-plananr (MP) Unet and 3D Unet predictions. Experiments on 10 subjects using 5-fold cross-validation demonstrate that the proposed fusion framework achieves the highest mean Dice Similarity Coefficient (DSC) of 0.860, outperforming all individual methods. The results demonstrate the complementary strengths of atlas-based and deep learning based segmentation approaches for automated cartilage analysis in MRI.

Cross-component prediction (CCP) is a promising coding tool that provides bitrate reductions by exploiting inter-channel correlations between luma (Y) and chroma (UV) components for next-generation video coding (NGVC). In fact, CCP is utilized in the exploration test model for NGVC, namely Enhanced Compression Model (ECM), developed by the Joint Video Experts Team (JVET). However, existing CCP methods assume fixed chroma sample positions, and this assumption does not hold for diverse YUV 4:2:0 video sequences in which the chroma components are subsampled with a different phase relative to luma. We propose an adaptive CCP framework that estimates suitable chroma sample positions from the first frame of each sequences and signals them using existing indices to avoid additional signaling overhead. We introduce the proposed method into seven CCP variants in ECM by modifying downsampling filters, coordinates, and gradients. Experiments on ECM-18 show significant bitrate reductions with negligible complexity impact for diverse test sequences of JVET's common test conditions and call for evidence, especially for gaming content with -2.84\% (U) and -2.09\% (V) reductions. The proposed method has been adopted in ECM-19 due to its favorable coding gains.

Most learned image compression systems rely on a single latent representation combined with a hyperprior, which limits their ability to efficiently capture image structure across spatial scales. In this work, we propose a hierarchical latent representation to improve the efficiency of the entropy model.Using multiple latents at different scales with their own entropy models, we aim to better capture the spatial structure of the latent representation. Our experiments show that this approach achieves a -17.9% reduction in BD-rate over VVC on Kodak, demonstrating the effectiveness of multi-scale latent representations. Furthermore, the approach is orthogonal to other advances in learned image compression, making it a versatile addition to existing methods.

In this paper, we propose a novel Low Operation Point (LOP) in-loop filter for VVC that leverages learning rate scheduling and single instruction multiple data (SIMD) acceleration. In the luma branch, the proposed LOP in-loop filter replaces the network structure of the LOP6 baseline with five TwinBlock modules, while in the chroma branch, it employs five Backbone Block (BBB) modules. Compared to the LOP6 baseline, the proposed network leverages $1\times5$ and $5\times1$ depth-wise convolutions to expand the receptive field of the BBB and limits the usage of BBB(1) to once, thereby effectively reducing the network complexity. Furthermore, the chroma branch adopts a variable channel configuration, which enriches input information while further optimizing model complexity. During the training phase, a novel learning rate scheduling strategy is utilized to enhance the learning efficiency and the performance of the proposed LOP in-loop filter. To further minimize model inference latency, multiple acceleration functions are integrated into the Small Ad-hoc Deep-learning Library (SADL). Experimental results demonstrate that under All Intra (AI) and Random Access (RA) configurations, the proposed LOP in-loop filter achieves BD-rate savings of \{-0.24\% (Y), -0.03\% (U), -0.10\% (V)\} and \{-0.35\% (Y), -0.57\% (U), -0.37\% (V)\}, respectively, compared to the original LOP6 network. Moreover, the decoding time is reduced by approximately 10\% (AI) and 14\% (RA), with computational complexity decreasing from 16.60 kMAC/pixel to 16.56 kMAC/pixel.

Dynamic point cloud attribute compression presents challenges due to the absence of one-to-one geometric correspondences across frames and the underutilization of geometry-attribute correlations. Registration methods such as recoloring and Iterative Closest Point (ICP), commonly used in prior work, introduce significant computational overhead. To address these issues, we propose a novel low-delay compression framework that introduces geometry-guided temporal prediction and bypasses explicit motion estimation/registration. Specifically, we establish soft inter-frame correspondences at the bottleneck scale using a Geometric Similarity Module (GSM) and Cross-Coordinate Attention (CCA), utilizing geometric cues to derive attention weights for efficient inter prediction. In addition, we adopt a lightweight Residual Channel Attention Block (RCAB) and a Residual Transform Module (RTM) to enhance feature representation and entropy coding efficiency. Experiments on public datasets demonstrate that our method significantly outperforms MPEG's AI-PCC reference software, TMAP, in terms of latency. In the Low-Delay configuration, average GPU decoding time is reduced by approximately 83.94% compared to TMAP while maintaining competitive Rate-Distortion performance.

Subblock Transform (SBT) is an effective tool in Versatile Video Coding (VVC) for improving the coding efficiency of inter-predicted residuals by enabling localized transform operations. However, its coding gains depend on explicit signaling of a finer-grained transform unit partitioning in terms of the partition direction and position, which introduces non-negligible overhead for each coding unit. To address this issue, we propose an implicit SBT (I-SBT) scheme that derives the partition direction and position without incurring any additional signaling overhead. Specifically, the proposed scheme exploits structural information inherently available in the prediction block by extracting the gradient energy features to characterize spatial activity. Based on these features, the direction and position of SBT are implicitly decided by detecting the regions with higher gradient energy. To maintain the standardized decoding parsing rule, a parsing-friendly two-stage decoder-side SBT derivation scheme is adopted, allowing implicit SBT decisions to be reproduced in a manner compatible with the VVC parsing process. The proposed method is implemented in VTM-23.0. Experimental results show that the proposed I-SBT can achieve, on average, 0.23% and 0.10% BD-rate reduction under LDB and RA configurations, respectively, with a negligible increase in time complexity.

The introduction of the Reference Picture Resampling (RPR) tool in Versatile Video Coding (VVC) has established a framework for resampling-based coding, significantly enhancing coding efficiency for high-resolution videos. Most existing studies focus on enhancing decoder side upsampling algorithms, resulting in non-standard-compliant decoder modifications. In this paper, an encoder side downsampling solution is proposed to both improve coding efficiency and preserve full decoder side compatibility. This integrated solution is made up of a content adaptive scale factor predictor and a learnable reconstruction quality oriented downsampling module. The content adaptive scale factor predictor first determines an appropriate resampling scale for each GOP based on both the quantization parameter (QP) and seven well designed image features. To further improve resampling quality under the selected scale factor, a multi-scale downsampling neural network module is employed. It generates downsampled pictures that are specifically optimized to the standard DCT-based upsampling filters, which are used in the decoder side of VTM, for enhanced reconstruction quality. Compared to VTM-23.13, experimental results demonstrate that the proposed method achieves average BD-rate savings of {-7.08%, 1.17%, -3.67%} and {-7.93%, -4.88%, -1.45%} for {Y, U, V} at five QP values of 32, 37, 42, 47, 52 under All Intra and Random Access configurations, respectively.

Versatile Video Coding (VVC), developed by the Joint Video Experts Team (JVET) and finalised in 2020, is the latest international video coding standard. Following its release, Neural Network (NN) based solutions have been actively explored to advance hybrid video coding. The Neural Network based in-Loop Filter (NNLF), a key development, performs image enhancement during encoding/decoding and operates in parallel to the De-Blocking Filter (DBF) to reduce artefacts from quantisation, partitioning, and other coding processes. In contrast to conventional signalling from encoder to decoder, this work proposes deriving the NNLF activation at the decoder from available coding information. Implemented in NN-based Video Coding (NNVC) software 15.0, the method achieves coding gains of 0.24% Y, 1.36% U and 1.23% V with 11% decoder runtime reduction in Low-Delay B (LDB). In Random Access (RA), gains are 0.02% Y, 0.12% U and 0.07% V with 5% decoder runtime reduction.

Hybrid video codecs, which combine conventional handcrafted components with learned modules, have recently emerged as a promising direction for future video coding standards. While deep non-linear transforms offer significant potential for exploiting complex statistical dependencies beyond the reach of traditional linear orthogonal transforms, they introduce significant challenges regarding Rate-Distortion (RD) optimization. Unlike orthogonal transforms, non-linear architectures do not preserve the l2 norm, preventing RD optimization in the transform domain. Previous attempts to integrate learned transforms into hybrid frameworks have generally relied on distortion-only criteria or bitrate proxies, failing to address the RD trade-off explicitly. In this paper, we bridge this gap by proposing a non-linear transform coding approach optimized under an explicit RD criterion. Building on principles from end-to-end learned image compression, we introduce a method to jointly optimize the transform with an external entropy model. Specifically, we design neural network architectures adapted to the specific residual characteristics (small blocks, noise-like statistics) which differs from usual end-to-end image compression. Experimental simulations performed on luma residuals extracted from the Versatile Video Coding standard demonstrate the effectiveness of the proposed framework.

Neural reference frame generation (NRFG) plays a critical role in neural network-based video coding (NNVC) by enhancing the Versatile Video Coding (VVC) inter prediction. Most existing studies primarily pursue coding gains through network architecture, while the impact of training data distribution on the performance and generalization of NRFG models remains unexplored. In this paper, we VVCadopt a unified reference frame synthesis (URFS) network and systematically investigate the effect of motion distribution in the training data on NRFG. Specifically, the URFS network is trained on the BVI-DVC/AOM and Vimeo-90K triplet datasets, respectively, each with distinct motion characteristics. Then, the two models are evaluated under the same common test conditions (CTC) to analyze their impact on VVC inter prediction enhancement. Experimental results show that the URFS network trained on the Vimeo-90K triplet dataset performs significantly better than the one trained on the BVI-DVC/AOM dataset. Although the same network architecture and training strategy are used, the trained models exhibit completely different performance depending on the training dataset. These results highlight the critical role of training data distribution for NRFG and provide practical guidelines for dataset selection when training NNVC tools.

Most neural video codecs rely on temporal conditioning, which makes them susceptible to error propagation over long sequences. While Transformer-based architectures like the Video Compression Transformer (VCT) offer a drift-free alternative, they suffer from high computational complexity and inferior Rate-Distortion (RD) performance. The recent Sliding Window Attention (SWA) addresses these shortcomings by reducing complexity and enhancing RD performance, yet it restricts decoding to a strictly sequential raster-scan order, creating a critical bottleneck in decoding latency. To resolve this, we propose Parallel Sliding Window Attention in Neural Video Coding (P-SWA), utilizing diagonal wavefronts to enable parallel decoding. By embedding a hyperprior and introducing an accumulator to fuse side information and local spatial context, our method increases decoding speed by 36 % over the parallel VCT. Simultaneously, it achieves Bjøntegaard Delta-rate savings of up to 10.0 % for I-frames and 7.1 % for P-frames over the SWA baseline.

The Low Operation Point (LOP) in-loop filter for VVC primarily focuses on enhancing the network backbone, while paying relatively little attention to enhancing the network head. In this paper, we propose a perceptually optimized LOP in-loop filter based on multi-scale head alignment. The core innovation lies in the multi-scale head alignment for multiple inputs that has parallel convolution layers at the network head. Specifically, we expand spatial context by combining spatial 3$\times$3 convolutions with complementary 1$\times$1 convolutions at the network head. The multi-scale head alignment significantly widens the receptive field and enriches feature diversity in the shallow layers without parameter increase. Simultaneously, we reduce the input channels of the first convolutional layer from 34 to 30 by refining the auxiliary input features, thereby reducing computational complexity while keeping channel dimensions as multiples of 16 for the hardware-friendly operation. Experimental results demonstrate that the proposed method achieves significant average Bj{\o}ntegaard Delta rate (BD-rate) gains of \{-0.72\% (Y), +0.22\% (U), -0.35\% (V)\} and \{-0.41\% (Y), +0.29\% (U), -0.03\% (V)\} under the All-Intra (AI) and Random Access (RA) configurations, respectively, while keeping computational complexity of 16.602 kMAC/pixel.

In this paper, a novel diamond shape filter design is proposed in low complexity neural network-based in-loop filtering in video coding. To enable the training of diamond shape filter in modern deep learning libraries which only support rectangle convolutional kernels, a new backbone block is proposed. It starts with a PReLu and a 1x1 convolution, and the output is then input to 1x5, 5x1 and 3x3 depth-wise separable convolution paths independently with the same number of channels. The outputs of the three paths are added together by element-wise addition and input to the last 1x1 convolution. At inference, the three 1x5, 5x1 and 3x3 parallel kernels could be merged into one diamond shape filter to reduce complexity. The model is implemented in JVET NNVC LOP6 by replacing the original backbone blocks in luma path. Its complexity is 16.59 kMac/Pixel after merging, which is slightly lower than 16.60 kMac/Pixel of LOP6. Experimental results show its BD-Rate for {Y, U, V, YUV} is {-0.30%, 0.45%, 0.43%, -0.26%} under RA and {-0.13%, 0.65%, 0.59%, -0.06%} under AI compared to NNVC LOP6. The BD-Rate compared to VTM anchor is {-6.77%, -13.28%, -11.68%, -6.95%} under RA and {-5.51%, -11.36%, -11.48%, -5.31%} under AI for {Y, U, V, YUV}. This shows superior efficiency over NNVC LOP6, achieving state-of-the-art coding performance at LOP category. The proposed training and merging of the new kernels can also be extended as a general method to enable the usage of various diamond shape filters in deep neural networks.

This paper introduces an optimized method for encoding 3D Gaussian Splat rotations within the MPEG V-PCC standard. Addressing the inefficiency of the current 4-component quaternion representation, which requires two video streams to be properly packed in the encoded bitstream, this approach reduces the rotation attribute to three components, allowing the data to fit within a single video sequence. The proposed technique employs a 24-rotation canonicalization strategy that exploits splats symmetry to maximize the scalar component, strictly bounding the remaining normalized values to a narrow range. Experimental results demonstrate superior coding efficiency, achieving an average bitrate reduction of approximately 7% compared to the anchor.

Full-precision floating-point neural image and video codecs pose significant challenges in power consumption, storage requirements, and cross-platform interoperability, particularly when deployed on resource-constrained devices. To address these issues, network quantization techniques have been extensively studied for neural image codecs. However, the quantization of neural video codecs remains largely unexplored. Unlike quantizing neural image codecs, quantizing neural video codecs requires significantly more effort. Many coding components operate on temporally correlated data and often rely on features propagated from previous frames, introducing additional sensitivity to both cross-platform round-off errors and network quantization. The major findings emphasize key trade-offs in video coding, especially the high quantization sensitivity of inter-frame decoding and the effectiveness of mixed-precision in balancing performance and complexity. This work presents the first systematic and algorithmic study of quantization effects across multiple neural video coding frameworks and temporal buffering strategies. This work offers actionable insights into the future development of neural video codecs.

LiDAR-acquired point clouds exhibit large variations in point density, which strongly affect geometry compression efficiency. Geometry-based codecs such as MPEG Geometry Point Cloud Coding (G-PCC) in angular mode are robust for sparse regions, while learning-based approaches such as the JPEG Pleno Verification Model (JP-PCC) are more effective for dense point distributions. This paper proposes a Density-adaptive LiDAR point cloud compression method that partitions each point cloud into dense and sparse subsets using a density threshold computed from the global density distribution. The dense subset is encoded with JP-PCC, while the sparse subset is encoded with MPEG G-PCC in angular mode. Experiments on the Ford, KITTI, and PandaSet datasets, restricted to bitrates below 4 bits per point, show consistent BD-Rate reductions relative to MPEG G-PCC. For PSNR Point-to-Point (D1), average bitrate savings reach up to 48% on KITTI and 26% on PandaSet, with gains observed across different density ranges. Similar trends are obtained for PSNR Point-to-Plane (D2). The proposed method significantly reduces runtime compared to JP-PCC while maintaining complexity close to that of MPEG G-PCC. These results demonstrate that density-adaptive coding provides an effective and computationally efficient strategy for LiDAR geometry compression.

The Lenslet Video Coding (LVC) project, launched by the Moving Picture Experts Group (MPEG) in June 2025, aims to develop codec-agnostic tools for efficiently compressing lenslet videos captured by plenoptic cameras or synthesized from multiview data. This paper introduces the architecture and core techniques of the first LVC Test Model (LVTM), and further explores optimization to enhance the performance: (1) the edge pixel prediction vectors are optimized via sequential search to refine reference selection and improve reconstruction quality; (2) an enhanced edge pixel fitting method using a bias-augmented linear model is introduced, achieving more accurate luminance recovery with minimal bitrate cost; and (3) an adaptive chroma upsampling scheme is adopted to further enhance color fidelity. Experimental results demonstrate that the proposed optimizations achieve a 54.76% bitrate reduction for lenslet quality while incurring only minimal multi-view quality degradation after rendering and negligible decoding time overhead. These techniques have been adopted by LVC in MPEG 153rd meeting and have been implemented to LVTM as the latest version LVTM 0.5.

Versatile Video Coding (VVC) introduces several new features; one is rectangular partitioning, which can divide an image into both rectangular and square blocks. We implemented inverse 2D DCT hardware for a VVC decoder, that includes a block transposition mechanism to equalize the computational complexities of the two 1D transform units. RTL simulations were performed, reducing the total number of cycles up to 9.82%.

Predictive intra coding is a key method for reducing bit-rate in video coding. While conventional methods seem to approach saturation in terms of bit-rate savings, learned prediction models have demonstrated promising gains. However, their high computational complexity makes hardware implementation at the decoder side extremely challenging. In this work, we propose a novel mixture-of-experts-based architecture that shifts most of the computational burden to the encoder, thereby keeping the decoder lightweight and hardware-friendly. We present preliminary experimental results as a proof of concept and demonstrate the potential of the proposed design for efficient learned intra-prediction.

This demo shows how spatially adaptive sparsity level maps for convolutional synthesis regularization can be learned using our open-source reconstruction package MRpro. In the presented method, sparsity level maps are estimated from a high-pass filtered input image and subsequently used to regularize an MR image reconstruction problem, which is solved using an accelerated proximal gradient descent scheme. These steps constitute an unrolled neural network that can be trained end-to-end. MRpro is a highly modular package that allows for easy prototyping of state-of-the-art image reconstruction methods and also application to real-world data. It consists of a large collection of MR-specific operators (e.g., Cartesian and non-Cartesian Fourier sampling operators, sensitivity operators, etc.) as well as widely used operators for model-based image reconstruction (finite differences, wavelets, convolutional dictionaries, etc). Further, some well-known algorithms (e.g., FISTA or PDHG) in their most general form are available. Additionally, it provides dedicated data containers to store raw scanner measurements, images, and quantitative parameters, together with their associated metadata. Further functionalities highly facilitate data loading and handling and allow researchers to quickly and easily work both with simulated and, more importantly, real raw MR measurement data. It is based on PyTorch and hence natively supports GPU acceleration for fast data processing.

The demo will provide an interactive GUI for strand measurement using the trained CNN-OSB regression model. The user selects a strand image from disk and specifies the acquisition view: frontal or sideways. The application then preprocesses the image with the same pipeline used during training, including resizing, tensor conversion and normalization, before passing it to the trained model launched from cnn_osb. The model, based on a CNN encoder and regression decoder, predicts the strand’s physical dimensions. For a frontal image, the GUI reports the visible measures such as strand width and height/length. For a sideways image, it estimates the thickness-related dimension, using the appropriate trained output depending on the selected view. After inference, the GUI displays the original strand image with the predicted measurements overlaid directly on it. Dimension annotations are drawn as visual guides, for example arrows or calibrated line segments, together with numerical values in physical units. This allows the user to immediately see both the image and the model interpretation. The demo therefore turns the trained CNN-OSB pipeline into a practical inspection tool: select image, choose view, run prediction, and visualize the estimated width, height and/or thickness on the strand.

We present an interactive demo of VolHuMe, a high-resolution large-scale dataset of volumetric human meshes designed to advance research on realistic digital humans. The demo offers attendees a hands-on exploration of the dataset, showcasing its rich ground-truth annotations, including multi-view RGB-D data, dense point clouds, high-resolution textured meshes, SMPL-X fittings, garment segmentation, and rigged animation-ready characters. By allowing users to inspect these modalities interactively, the demo highlights the level of geometric and appearance detail captured by VolHuMe, with particular emphasis on fine-grained body regions such as hands, face, clothing, and close-range surface details. Beyond dataset visualization, the demo lets participants directly experience the potential of VolHuMe for digital human animation. Attendees will be able to animate, in real time, a character from the dataset and observe how high-fidelity volumetric scans can be transformed into interactive avatars. This live interaction illustrates practical challenges in avatar animation, including rigging, motion transfer, temporal consistency, and the preservation of realistic human appearance. By combining dataset exploration with real-time character animation, the demo provides an accessible and engaging entry point to VolHuMe. It shows how comprehensive 4D ground truth can support future research in human reconstruction, neural rendering and digital human animation.

This interactive demo presents a cloud-gaming no-reference video quality assessment system based on MTL-VQA. Users can select a gaming clip and an encoding bitrate from 0.25, 0.5, 1, 2, or 5 Mbps, then inspect the reconstructed video together with its predicted perceptual quality. The system extracts multi-task learned video features and produces a MOS-calibrated quality score on a 0–100 scale, accompanied by an intuitive quality category: Poor, Fair, Good, or Excellent. Beyond a single clip-level score, the interface visualizes frame-level quality variation over time, enabling users to identify transient degradation caused by compression, motion, or scene complexity. A rate–quality curve further shows how predicted quality changes across bitrates for the same content. Optional VMAF and SSIM overlays allow visitors to compare the trend of the learned no-reference model with conventional full-reference metrics. The demo illustrates how perceptual quality telemetry can support bitrate selection, quality monitoring, and model analysis in practical cloud-gaming workflows, without requiring access to the original reference video during inference.

Monocular 3D Human Pose Estimation (3DHPE) remains challenging due to inherent depth ambiguities and occlusions. Existing video-based approaches often incur high computational costs by alternating spatiotemporal processing, while current diffusion-based methods typically rely on generic denoisers with naive input concatenation. To address these limitations, we propose a specialized diffusion framework incorporating Pose Patchification (PoPatch) and Adaptive Pose Modulation (AdaPoseMod). PoPatch extracts spatiotemporal features simultaneously, reducing computational complexity; our method reduces Multiply-Accumulate Operations (MACs) by over 200 times compared to state-of-the-art diffusion baselines. AdaPoseMod facilitates effective interaction between 2D observations and contaminated 3D poses through a dedicated modulation mechanism. Our approach achieves state-of-the-art performance on Human3.6M, MPI-INF-3DHP, and HumanEva datasets. Extensive ablation studies further validate the efficacy of our design choices in balancing efficiency and robustness.

Estimating the precise timing of batting impact is crucial for understanding the rapid sensorimotor control. However, this task is challenging for RGB cameras due to insufficient temporal resolution and motion blur. Similarly, Inertial Measurement Units (IMUs) are impractical for actual matches due to sensor intrusiveness and their limited temporal precision. To overcome these limitations, we propose a novel framework leveraging event-based cameras, which offer microsecond resolution and high dynamic range, to estimate impact timing based on the weighted centroid distance between the detected ball and bat. To address the domain gap between event frames and RGB images that degrades segmentation accuracy, we generate high-density event frames. We then introduce a mask refinement network that leverages these frames and bidirectional mask information, optimized using a novel loss function. Experiments on real-world datasets demonstrate that our method achieves superior accuracy under challenging conditions, including low-light environments and severe occlusions, outperforming baselines by reducing the Mean Absolute Error by approximately 63%.

Monocular egocentric human pose estimation is essential for ubiquitous activity monitoring. However, understanding the user's absolute location within the environment remains a challenge. Existing methods primarily focus on relative motion from an initial position, and tend not to account for the wearer's absolute location within an environment. Furthermore, inherent scale ambiguity in monocular vision leads to severe translational drift, limiting long-term tracking without specialized multi-sensor hardware. To address this, we propose Map-Mono-Ego, a novel framework achieving globally consistent human pose estimation solely from a monocular camera by leveraging a pre-scanned 3D point cloud. We also introduce AIST-Living dataset, a new dataset pairing egocentric video with ground-truth motion in a scanned environment. Experiments demonstrate that our approach significantly outperforms the state-of-the-art baseline, proving its utility for practical monitoring tasks without specialized hardware.

Due to the inherent scale ambiguity of monocular images, existing Vision-Language Models (VLMs) often struggle with quantitative spatial reasoning, particularly in tasks that require metric-scale depth understanding. While data-driven approaches could, in principle, enhance the ranging capability of VLMs, acquiring dense and high-quality ground-truth depth annotations with real-world physical units remains prohibitively expensive. More fundamentally, metric depth inference is challenged by the representation gap between continuous geometric quantities and the discrete token-based embedding space of language models. To address these issues, we propose MetricDepth-VLM, a geometry–semantic alignment framework that enables metric-consistent spatial reasoning in VLMs. Our approach calibrates relative monocular depth predictions into absolute metric depth representations and discretizes them into patch-wise numerical labels with explicit physical meaning, referred to as metric-aware bin tokens . These geometry tokens are aligned with the language model embedding space, allowing the model to reason about scene semantics while preserving precise metric-scale spatial information. Experimental results on MSMU-Bench and Q-Spatial++ demonstrate that the proposed approach significantly improves VLM performance in metric distance estimation and quantitative spatial reasoning tasks, enabling more accurate geometric assessment and more reliable semantic understanding.

Monocular 3D face reconstruction typically faces a trade-off: 3DMM-based regression ensures global stability but often misses high-frequency geometric details, whereas generative approaches excel in fidelity but suffer from geometric misalignment. To bridge this gap, we propose FD-DIFF, a novel framework integrating frequency decoupling with dual-stream collaborative diffusion. By explicitly disentangling facial features via haar wavelet transforms, FD-DIFF leverages low-frequency signals for robust coarse regression while utilizing high-frequency components to condition fine-grained synthesis. Our core innovation is a dual-stream diffusion mechanism: a spectral geometry stream synthesizes high-fidelity vertex displacements in the efficient Laplacian spectral domain, while a parallel landmark residual stream predicts 2D offsets to rectify projection errors under challenging conditions. Crucially, we enforce a projection-consistency constraint that synergistically couples 3D spectral deformations with 2D landmark observations. Extensive experiments demonstrate that FD-DIFF sets a new state-of-the-art in both joint 3D face reconstruction and dense alignment.

Reliable rebar spacing inspection is essential for construction quality assurance but remains challenging under varying viewpoints and complex geometric conditions. Existing approaches typically rely on a fixed measurement strategy, which becomes unreliable when sensing conditions vary across scenarios. This paper proposes a Mixture of Measurement Strategies (MoMS) framework that treats rebar spacing estimation as a reliability-aware selection problem. Given a single monocular image captured by a mobile device, the framework reconstructs rebar topology and estimates metric scale using lightweight geometric cues, while monocular depth estimation is exploited only when reliable. A Reliability-Aware Measurement Strategy (RAMS) adaptively selects between planar and depth-guided formulations. Experiments across diverse viewpoints and construction environments demonstrate robust performance, and system optimization enables efficient deployment in cloud-based mobile applications.

Egocentric human mesh recovery (HMR) from monocular head-mounted cameras is increasingly important for AR/VR applications, but remains challenging due to the lack of reliable ground-truth (GT) annotations based on parametric human body models such as SMPL and SMPL-X for real egocentric images. Existing egocentric HMR methods typically rely on pseudo-GT and focus on body pose estimation, which limits their ability to recover fine-grained whole-body details such as hands and face. We study egocentric whole-body human mesh recovery and propose a prior-guided learning framework that reconstructs whole-body meshes from a single egocentric image. We construct more accurate optimization-based pseudo-GT aligned with 3D joint supervision, and leverage multiple priors by adapting an exocentric HMR foundation model together with a diffusion-based pose prior. A deterministic undistortion module is further adopted to handle fisheye distortions in egocentric images. Experiments across multiple egocentric benchmarks demonstrate improved whole-body reconstruction compared to state-of-the-art methods, and show that our optimization-based pseudo-GT is substantially more accurate than existing regression-based pseudo-GT. To facilitate reproducibility, the code and dataset annotations are publicly available at \url{https://github.com/naso06/EgoSMPLX}.

Accurate multi-resolution surface defect detection is vital for industrial quality control, yet detecting multi-scale defects—from microscopic pinholes to large dents—under edge-computing constraints remains challenging. This paper proposes YOLOv8-HRP2, a framework optimized for high-resolution industrial inspection at $1280 \times 1280$. To preserve fine-grained spatial cues for tiny defects, we introduce a dedicated High-Resolution P2 (HRP2) layer. To mitigate sample scarcity and class imbalance, a two-stage curriculum training strategy with targeted Copy-Paste augmentation is implemented. Furthermore, a Test-Time Augmentation (TTA) engine is integrated to bolster inference robustness against complex environmental noise. The model establishes a superior trade-off between accuracy and efficiency, achieving a mAP@50 of 0.606 and 92.22 FPS with 25.21 M parameters. The proposed framework establishes a robust balance between high-resolution feature representation and computational efficiency, ensuring reliable performance for autonomous monitoring on resource-constrained industrial devices.

Traditional monocular Visual-Inertial Odometry (VIO) systems struggle in low-texture environments where sparse visual features are insufficient for accurate pose estimation. To address this, dense Monocular Depth Estimation (MDE) has been widely explored as a complementary information source. While recent Vision Transformer (ViT) based complex foundational models offer dense, geometrically consistent depth, their computational demands typically preclude them from real-time edge deployment. Our work bridges this gap by integrating learned depth priors directly into the VINS-Mono optimization backend. We propose a novel framework that enforces affine-invariant depth consistency and pairwise ordinal constraints, explicitly filtering unstable artifacts via variance-based gating. This approach strictly adheres to the computational limits of edge devices while robustly recovering metric scale. Extensive experiments on the TartanGround and M3ED datasets demonstrate that our method prevents divergence in challenging scenarios and delivers significant accuracy gains, reducing Absolute Trajectory Error (ATE) by up to 26.8%. Code is available at https://github.com/aalniak/vinsfusion-clean.

The Vision Transformer (ViT) achieves remarkable accuracy across visual tasks but remains computationally expensive for edge deployment. This paper presents MicroViTv2, a lightweight Vision Transformer optimized for real-device efficiency. Built upon the original MicroViT, the proposed model is designed based on reparameterized design, specifically Reparameterized Patch Embedding (RepEmbed) and Reparameterized Depth-Wise convolution mixer (RepDW) for faster inference, and introduces the Single Depth-Wise Transposed Attention (SDTA) to capture long-range dependencies with minimal redundancy. Despite slightly higher FLOPs, MicroViTv2 improves accuracy up to 0.5% compared to its predecessor and surpassing MobileViTv2, EdgeNeXt, and EfficientViT while maintaining fast inference and high energy efficiency on Jetson AGX Orin. Experiments on ImageNet-1K and COCO demonstrate that hardware-aware design and structural re-parameterization are key to achieving high accuracy and low energy consumption, validating the need to evaluate efficiency beyond FLOPs. Code is available at https://github.com/novendrastywn/MicroViT.

3D Gaussian Splatting has revolutionized the field of radiance field reconstruction with its real-time rendering capabilities. As a pivotal advancement, Mip-Splatting demonstrates superior potential in suppressing aliasing artifacts by introducing a multi-scale sampling mechanism. However, these methods predominantly rely on photometric supervision from RGB signals, leading to inherent ill-posedness in weakly textured regions. This limitation hinders the recovery of accurate geometric surfaces and results in noticeable floaters within the scene. To address this challenge, we propose a novel framework that integrates anti-aliased rendering with monocular geometric priors. Specifically, we adopt Mip-Splatting as the backbone and leverage high-quality monocular depth features extracted by Depth Anything V2 as geometric guidance. To mitigate the inherent scale-and-shift ambiguity of monocular depth, we innovatively introduce a Pearson Correlation Loss within the disparity space. This loss function is designed to maximize the linear correlation between the predicted depth and the prior depth distribution, rather than enforcing a rigid alignment of absolute values. Extensive evaluations on the Mip-NeRF 360 and Tanks \& Temples datasets demonstrate that our method not only maintains state-of-the-art rendering quality but also significantly suppresses background noise and recovers smooth, consistent geometric structures.

The Vision Transformer (ViT) architecture, leveraging self-attention, has achieved remarkable success across a wide range of tasks. Despite its strengths, ViT's optimization remains confined to modeling relationships within individual images, limiting its ability to capture the global geometric relationships between data points. To address this limitation, a novel framework is proposed to integrate ViT with proximal tools, enabling a unified geometric optimization approach to enhance feature representation and classification performance. This proceeds with interpreting individual attention heads in a ViT as tangent spaces to the data manifold, each providing a local representation, and collectively as a tangent bundle reflecting the totality of exploitable information in the data. Proximal iterations are then introduced to define sections within the tangent bundle and project data from tangent spaces onto the base space, achieving global feature alignment and optimization. Experimental results confirm that the proposed method outperforms conventional ViT in terms of classification accuracy and data distribution.

General Visual Inspection (GVI) of aircraft surfaces is critical for aviation safety, yet automating defect detection remains challenging due to the geometric diversity of surface anomalies. We present a systematic study revealing that CNN-based detectors excel at geometrically regular defects but consistently struggle with scratches-elongated, irregular anomalies that violate CNN locality assumptions. Our experiments demonstrate that CNN models plateau at 82-83% mAP on scratch detection, regardless of architecture or augmentation strategy, whereas the transformer-based DETA Swin achieves 87.4% mAP via global attention mechanisms, albeit with 130x slower inference. To address this accuracy-efficiency trade-off, we implement an Uncertainty-Guided Hybrid Architecture that routes inputs through CNN detection heads for regular defects while selectively activating transformer attention only for uncertain regions likely to contain high-aspect-ratio anomalies. Evaluated end-to-end on a live testbed, this adaptive approach achieves 91.3% overall mAP across all defect classes while reducing average inference time by 74% compared to transformer-only processing, enabling practical deployment for commercial aircraft maintenance.

This paper presents modifications to block importance mapping (BIM) for alignment with group-of-picture-based reference picture resampling (GOP-based RPR). The paper also introduces continuous BIM and includes some modifications to the motion compensated temporal filter (MCTF). All three encoder-only methods are part of the VVC reference software. The results indicate Bjøntegaard delta rate (BD-rate) savings of 1.4% on average for the standard dynamic range (SDR) and the high dynamic range (HDR) content that was used in the call for evidence (CfE) evaluation in Joint Video Experts Group (JVET), in the case when both the test and anchor have both BIM and GOP-based RPR enabled. The corresponding impact of those improved methods when the anchor does not use them is 8.0%. Furthermore, for content used in the common test conditions (CTC) in JVET, the effect of only the changes to BIM and MCTF gives a BD-rate saving of 2.5% for the random access (RA) configuration and 2.6% for the low-delay (LD) configuration compared to an anchor where BIM is turned off. The paper also presents an alternative configuration of BIM that enables the control of rate distortion optimization without adapting the quantization parameter (QP) locally. This, together with the MCTF changes against an anchor where BIM is turned off, gives a BD-rate saving of 2.2% and 2.1% respectively for CTC. Finally, the impact of only the changes to MCTF is a BD-rate saving of 0.5% for the RA configuration and a loss of 0.1% for the LD configuration.

Self-supervised monocular depth estimation (MDE) enables depth prediction from monocular images without requiring ground-truth supervision, making it attractive for large-scale and real-world applications. Despite steady improvements in accuracy, most existing methods remain difficult to interpret, as depth is inferred from RGB representations that obscure the impact of individual perceptual image components. This lack of transparency limits systematic analysis of failure cases and reduces confidence in safety-critical settings. This paper presents a self-supervised framework for perceptually interpretable monocular depth estimation (PIMDE), designed to associate depth predictions with distinct perceptual components of the input image. Rather than operating directly on RGB inputs, the proposed method decomposes each image into a set of perceptual feature maps (PFMs), each encoding a specific visual cue. Distinct depth estimation branches process these PFMs independently to produce depth estimates (PIDEs), which are subsequently combined through an explicit fusion strategy. This formulation allows us to examine directly the contribution of each perceptual cue to the final depth prediction. Experiments conducted on the KITTI benchmark dataset demonstrate that PIMDE achieves performance comparable to established self-supervised MDE methods while providing additional insight into how different perceptual cues influence depth estimation. These results indicate that perceptual decomposition can support interpretability without sacrificing depth estimation accuracy.

Understanding on-site conditions in complicated industrial and construction environments is essential for ensuring safety and operational efficiency. This requires multiple site visits, resulting in increased labor and cost. Indoor-street views that can be constructed from simple monocular 360-degree cameras are often used to enable remote inspection. However, the large number of images makes inspection and surveying difficult, even remotely. In this paper, we present a 3D object mapping system using a monocular 360-degree camera that can produce an intelligent "labelled street-view". The proposed approach integrates object detection and 360-degree monocular visual SLAM to reconstruct a semantically labelled 3D map of the environment. We achieve this by associating multi-view 2D bounding boxes to the 3D locations of each object in the reconstructed spatial map. The integration of object detections from multiple views allows for instances detected in only a few images to appear in the final map. This enables intuitive remote understanding of complex scenes where the surveyor can simply click on the desired object to "move" to that location in the street view. Experimental results using real-world 360-degree video data demonstrate the feasibility of our approach and highlight the potential for future extensions toward searchable street-view-style interfaces and natural-language-based inspection systems.

This paper presents LINDE (Lightweight INter-pixel DEnoiser), a compact neural network for non-blind denoising of remote sensing images corrupted by Additive White Gaussian Noise. LINDE incorporates non-local inter-pixel interactions into an efficient architecture that employs approximately 33 times fewer parameters than the state-of-the-art MambaIR. Despite its lightweight design, it outperformed other baselines, while requiring only 0.0072 seconds per image. LINDE also delivers competitive performance across multiple benchmark datasets, demonstrating strong robustness to severe noise, including meteorological radar images.

Compared to supervised cross-modal hashing (CMH), unsupervised CMH reduces the reliance on manual labeling by learning binary codes from unlabeled image-text pairs. However, existing unsupervised CMH methods often rely on large-scale image-text pairs, which are costly to collect. To address this limitation, we propose Global-Neighborhood Alignment Hashing (GNAH), a novel approach that preserves the semantic structure of vision–language foundation models within a compact binary Hamming space using only a limited number of image–text pairs. Specifically, GNAH captures global structural information from the continuous latent space and transfers it into the binary Hamming space through a Prototype-Anchored Global Alignment module. In addition, GNAH extends conventional pairwise contrastive learning by modeling stochastic neighborhood relationships via a Contrastive Stochastic Neighborhood Alignment module, thereby alleviating overfitting to sparse pairwise correlations. Extensive experiments demonstrate that GNAH consistently outperforms existing unsupervised cross-modal retrieval methods under data-constrained settings, offering a practical solution for real-world CMH applications.

While traditional visual compression techniques remain widely used, learned methods based on deep neural networks have recently surpassed conventional codecs in compression efficiency. However, these approaches often suffer from high computational cost, long inference times, and the need for separate models at different bitrates. In this paper, we make two main contributions. First, we propose an enhanced State-Space Model–Based Image Compression architecture with a Channel-Wise autoregressive entropy model, SSMIC_CW, achieving improved rate–distortion performance. Second, we introduce USSMIC_CW, a universal framework for continuous variable-rate compression using a single trained model, where the target bitrate or distortion is controlled by simple input scaling at inference. To support precise rate distortion targeting, we further develop a lightweight method to model and predict the required scaling factor, using an empirically derived exponential relationship refined via a fast iterative fitting procedure.

In much of the existing Dense Pose Estimation (DPE) literature, models are either trained from scratch or initialized with conventional pretrained backbones, while the potential of transfer learning from recent vision foundation models remains largely underexplored. In this work, we investigate transfer-learning-friendly architectures for DPE. To this end, we introduce SimpleDPE, an encoder-decoder architecture that employs Vision Transformers as the backbone encoder. We show that initializing SimpleDPE with pretrained weights from DPE-relevant tasks consistently outperforms existing baselines while maintaining comparable computational efficiency. Furthermore, we propose SimpleDPE-MS, an extension of SimpleDPE that incorporates multi-scale feature representations to enhance performance. Extensive experiments on two chart-based DPE benchmarks demonstrate that our models achieve strong performance-efficiency trade-offs, with several variants reaching state-of-the-art performance

State Space Models (SSMs) have emerged as powerful architectures in computer vision, yet computational efficiency remains crucial for scalable deployment. While token merging stands as an effective approach, applying it to SSMs requires careful consideration of their recurrent state dynamics. To address this, we propose MaMe, a token-merging strategy tailored for SSM-based vision models. Our approach identifies merge candidates based on state contribution and content similarity and arranges merged tokens to preserve sequential continuity. Extensive experiments demonstrate that MaMe achieves superior efficiency-performance trade-offs for both fine-tuned and off-the-shelf models.

The growing integration of Virtual and Augmented Reality (VR/AR) in various fields makes fast 3D scene reconstruction increasingly important. Although generalizable 3D synthesis methods offer promising performance in generating 3D scenes from multiple viewpoints, they remain limited by slow adaptation to new scene categories, which can take minutes. In this paper, we introduce PT-GS, a novel approach based on prompt tuning to enable rapid domain transfer in generalizable 3D synthesis models. By tuning only 3.18\% extra parameters and keeping the core network frozen, we achieve rapid domain transfer with significantly reduced training time. To further boost efficiency, we incorporate two prompt-side mechanisms: (i) Token Ranking, which selectively retains most relevant tokens to suppress redundant pre-trained information, and (ii) Runtime Token Filtering, which prunes low-contribution tokens on the fly. Experiments show that our method achieves up to 20.3x speedup in cross-domain adaptation and reaches over 29 dB PSNR within 10 seconds. Meanwhile, it reduces memory usage by up to 2.1x and adaptation energy by up to 7.0x compared with full fine-tuning, paving the way toward eventual real-time AR/VR applications.

Transformers have become the dominant backbone for video representation learning, yet their quadratic attention complexity remains a major bottleneck, especially for long-form videos where redundancy exists both spatially and temporally. Existing approaches mitigate this by pruning or merging tokens heuristically, often without a principled measure of representational fidelity. In this work, we revisit video token merging from a distributional perspective. We formulate token reduction as a coreset selection problem grounded in Optimal Transport (OT) theory. Specifically, we propose a novel OT-based token merging method, Coreview, that leverages the Fused Gromov–Wasserstein distance to jointly capture structural and semantic relationships between tokens. By selecting "core-tokensets" that best preserve the global distributional geometry of the input, our approach enables efficient yet faithful compression of spatiotemporal token sets. Extensive experiments on LVU, COIN and Breakfast datasets show that Coreview achieves 2.4% higher accuracy compared to state-of-the-art baselines. Code: https://github.com/SusimRoy/Coreview

Data scaling laws demonstrate the diminishing returns of indiscriminate dataset expansion, necessitating data governance to refine the pre-training corpora. Existing methods termed as DataSieves, primarily focus on sample-level coarse-grained pruning using heuristic ideas. However, they overlook the fundamental semantic asymmetry between modalities. By treating image-text pairs as indivisible units, they fail to address visual redundancy and textual semantic fragility, precluding the optimization of intra-sample compositions. In this work, we propose DataJuicer, shifting data governance from sample-level ``sieving'' to fine-grained token-level ``juicing''. DataJuicer leverages a dual-branch intra-sample architecture: the vision branch prunes redundant patches to distill salient visual semantics, while the text branch leverages visual cues as the semantic anchors to rectify and denoise captions. Such a coordinated process maximizes information density and significantly enhances image-text alignment. Extensive experiments show that DataJuicer consistently outperforms state-of-the-art DataSieves across diverse downstream tasks, enabling efficient vision-language pre-training for alignment.

Embedded devices typically lack the resources of GPU-equipped machines, and existing inference methods suffer from either high computational overhead (patch-based) or accuracy loss (approximation-based). We propose GaLe, a memory-efficient technique that enables deploying pretrained networks on constrained devices without retraining. GaLe partitions feature maps into two components: a local exact (LE) representation preserving fine details, and a global approximate (GA) representation retaining long-range dependencies. Unlike standard tiling, GaLe supports global operations and attention mechanisms found in hybrid CNN-transformer models. Validated on ImageNet, our method matches exact inference performance while achieving up to 65% speedup and 90% RAM reduction on a \texttt{Cortex-M33} compared to patch-based inference. We further demonstrate GaLe's versatility across classification and detection tasks, highlighting its potential as a foundation for designing resource-efficient architectures.

State Space Models (SSMs) have recently gained traction in remote sensing change detection (CD) for their favorable scaling properties. In this paper, we explore the potential of modern convolutional and attention-based architectures as a competitive alternative. We propose NeXt2Former-CD, an end-to-end framework that integrates a Siamese ConvNeXt encoder initialized with DINOv3 weights, a deformable attention-based temporal fusion module, and a Mask2Former decoder. This design is intended to better tolerate residual co-registration noise and small object-level spatial shifts, as well as semantic ambiguity in bi-temporal imagery. Experiments on LEVIR-CD, WHU-CD, and CDD datasets show that our method achieves the best results among the evaluated methods, improving over recent Mamba-based baselines in both F1 score and IoU. Furthermore, despite a larger parameter count, our model maintains inference latency comparable to SSM-based approaches, suggesting it is practical for high-resolution change detection tasks.

Automatic segmentation of multiple abdominal organs is crucial for precisely locating lesion areas. However, significant variations in structural morphology and axial dimensions among different organs pose substantial challenges for accurate segmentation. To tackle this issue, we introduce the Multi-axial Guidance Network. Specifically, the Multi-axial Channel Enhancement module is proposed to capture fine-grained features across different axes and adaptively adjust the combined results. To further enhance multi-scale information fusion, we have designed the Axial Guidance Multi-scale Fusion module for deep cross-axial interactions. These modules facilitate the entire process of fine-grained feature enhancement and fusion, with axial anisotropy playing a dominant role. Notably, our method achieves state-of-the-art performance on the mainstream abdominal multi-organ segmentation dataset.

Task-Free Online Continual Learning (TFOCL) requires models to adapt to streaming data without explicit task boundaries while mitigating catastrophic forgetting. Existing methods expand parameter-efficient modules to learn new concepts, but this often leads to uncontrolled module growth. Although weighted averaging across modules during inference can mitigate this issue, it tends to blur module specialization and compromise the integrity of learned knowledge. To address these issues, we introduce Flexible Expansion and Merging (FlexEM). During training, FlexEM employs an expand-and-merge strategy that adds new modules in response to distribution shifts and performs representation-aware merging to consolidate modules with similar feature representations, controlling model capacity while preserving learned knowledge. During inference, FlexEM adopts a dual-stage module retrieval strategy: it first selects candidate modules via feature affinity matching to the current data distribution, then verifies the selected module using an entropy-based criterion to prevent module mismatch. Experimental results demonstrate that FlexEM significantly improves adaptability in TFOCL while maintaining the stability of previously learned knowledge.

In this paper, we introduce GeoTopo, a novel method for 3D mesh generation that explicitly encourages geometric smoothness and improves topological consistency during the autoregressive generation process. By integrating a lightweight constraint module that refines each predicted vertex, GeoTopo produces meshes that are visually compelling and structurally more reliable for downstream use like gaming and VR. Our experiments provide quantitative evidence that GeoTopo significantly reduces topological errors, such as non-manifold edges and holes, while producing substantially smoother surfaces compared to state-of-the-art transformer-based methods. Unlike approaches that require computationally expensive post-processing, our method integrates these priors directly into the generation loop with minimal overhead, preserving an end-to-end workflow. Future work could explore optimizing the constraint module's efficiency to further reduce the small inference-time overhead, paving the way for real-time applications.

Surgical phase recognition plays a crucial role in advancing computer-assisted surgical systems by enabling the understanding of sequential procedural stages. However, existing methods often suffer from redundant spatiotemporal feature extraction and high computational costs throughout the entire surgical procedure, particularly when processing long surgical videos. In this paper, we introduce SurgMLPs, a novel MLP-like architecture that leverages Contrastive Language-Image Pre-training (CLIP) for surgical phase recognition. Our approach combines a frozen CLIP model for robust feature extraction with a two-stage, purely MLP-based network. The first stage, MLP-Mixer, captures spatiotemporal embeddings and identifies surgical phase segments. The second stage, MLP-Classifier, aggregates and classifies these segments for accurate phase prediction. This lightweight design, with an extremely small number of parameters, significantly reduces redundancy and improves computational efficiency. Extensive experiments on two benchmark datasets demonstrate the superiority of SurgMLPs over state-of-the-art methods, offering an effective and efficient solution for surgical phase recognition and positioning it as a promising tool for real-time surgical assistance.

To address the issues of insufficient global-local feature coordination, inefficient binocular fusion, and high computational complexity in existing gaze estimation models under unconstrained scenarios, this paper proposes an efficient gaze estimation model named EF-ViMGaze, based on the Vision Mamba (ViM) architecture. The model takes facial images and binocular local images as input, capturing both global facial context and fine-grained local eye features via a dual-branch feature extraction module. An EyeAttnPool attention pooling module is designed to adaptively fuse binocular features and focus on gaze-relevant key information. The ViM encoder is improved by introducing an FFN branch to enhance non-linear feature representation, while bidirectional state space modeling enables low-complexity capture of long-range dependencies. Finally, a concise gaze estimation head outputs a normalized 3D gaze direction vector. Experiments on the MPIIFaceGaze, EyeDiap, and Gaze360 datasets demonstrate that EF-ViMGaze achieves average angular errors of 3.74°, 4.79°, and 8.59°, respectively, reaching state-of-the-art performance on EyeDiap and Gaze360 while maintaining comparable accuracy on MPIIFaceGaze. Ablation studies verify the effectiveness of the FFN branch and the EyeAttnPool module, providing an efficient and accurate solution for end-side gaze estimation applications.

Ensuring the safety of surgical instruments requires reliable detection of visual defects. However, manual inspection is prone to error, and existing automated defect detection methods, typically trained on natural/industrial images, fail to transfer effectively to the surgical domain. We demonstrate that simply applying or fine-tuning these approaches leads to issues: false positive detections arising from textured backgrounds, poor sensitivity to small, subtle defects, and inadequate capture of instrument-specific features due to domain shift. To address these challenges, we propose a versatile method that adapts unsupervised defect detection methods specifically for surgical instruments. By integrating background masking, a patch-based analysis strategy, and efficient domain adaptation, our method overcomes these limitations, enabling the reliable detection of fine-grained defects in surgical instrument imagery.

Recently, Mamba-based architectures have emerged as a promising alternative to Transformers due to their linear complexity. Moreover, hybrid models like MambaVision have successfully integrated the local-feature extraction strengths of Transformers with the long-range modeling of State Space Models (SSMs), achieving state-of-the-art performance in vision tasks. However, their structural heterogeneity characterized by diverse non-uniform activations and fused convolutional weights presents significant challenges for post-training quantization. In this paper, we propose HyQuant, a systematic framework tailored for hybrid vision models. HyQuant employs an automatic framework to adaptively allocate quantization strategies, coupled with clustering-based weight quantization to handle fusion-induced distortions and a reparameterization technique to eliminate SSM inference bottlenecks. Our results show that HyQuant restores MambaVision-T’s 8-bit accuracy to 81.89% (only 0.35% drop) and recovers 6-bit performance from a collapsing 1% to 80.00% on ImageNet-1k. We further validate its generalizability on the VSSD architecture, demonstrating robust effectiveness across diverse hybrid models.

Over-segmentation is a widely used operation in computer vision pipelines, often employed as an intermediate spatial representation for efficient image perception, including superpixel-based tokenisation and interpretable region-based analysis. As these systems are increasingly deployed in real-time safety-critical applications such as medical imaging, industrial inspection and security, algorithms that are boundary-adherent, deterministic, and hierarchically-consistent become increasingly important. Existing methods typically trade off these properties, making it difficult to satisfy all four simultaneously. We propose Adaptive Zone Merging (AZM), a deterministic graph-based bottom-up over-segmentation algorithm designed to combine these properties while remaining amenable to parallel execution through local merge proposals. AZM exposes a single colour-distance threshold that induces a nested hierarchy obtained by monotonic region merging. This enables efficient multi-scale analysis with minimal re-computation overhead, providing stable region primitives for downstream learning and symbolic aggregation. We evaluate AZM on the BSDS500 dataset using boundary-focused metrics, demonstrating improved boundary F1-score compared to widely used baselines. Source code is available at: https://github.com/PetarKotsev/azm-segmentation.

Hybrid opto-electronic neural networks combine optical front-ends with electronic back-ends to perform vision tasks, but joint end-to-end (E2E) optimization of optical and electronic components is computationally expensive due to large parameter spaces and repeated optical convolutions. We propose Direct Kernel Optimization (DKO), a two-stage training framework that first trains a conventional electronic CNN and then synthesizes optical kernels to replicate the first-layer convolutional filters, reducing optimization dimensionality and avoiding hefty simulated optical convolutions during optimization. We evaluate DKO in simulation on a monocular depth estimation model and show that it achieves twice the accuracy of E2E training under equal computational budgets while reducing training time. Given the substantial computational challenges of optimizing hybrid opto-electronic systems, our results position DKO as a scalable optimization approach to train and realize these systems.

We propose EasyControlEdge, adapting an image-generation foundation model to edge detection. In real-world edge detection (e.g., floor-plan walls, satellite roads/buildings, and medical organ boundaries), crispness and data efficiency are crucial, yet producing crisp raw edge maps with limited training samples remains challenging. Although image-generation foundation models perform well on downstream tasks, their pretrained priors for data-efficient transfer and iterative refinement for high-frequency detail preservation remain underexploited for edge detection. To enable crisp and data-efficient edge detection using these capabilities, we introduce an edge-specialized adaptation of image-generation foundation models. To better specialize the foundation model for edge detection, we incorporate an edge-oriented objective with an efficient pixel-space loss. At inference, we introduce guidance based on unconditional dynamics, enabling a single model to control edge density through a guidance scale. Experiments on BSDS500, NYUDv2, BIPED, and CubiCasa show competitive or superior performance against state-of-the-art methods, particularly under no-post-processing crispness evaluation and with limited training data.

Sparse multi-view stereo (MVS) often suffers from severe artifacts due to the limited number of viewpoints. In con trast, multi-view photometric stereo (MVPS) can recover detailed surface geometry under sparse views by exploiting illumination variations, but its reliance on multi-illumination inputs limits practical use. In this paper, we bridge these two paradigms by converting an MVS input into an MVPS compatible setting via virtual illumination, enabling photo metric stereo to enhance MVS performance in sparse capture scenarios. We propose a two-stage framework in which an inverse neural rendering module first estimates surface bidi rectional reflectance distribution function parameters from sparse-view images. Based on these estimates, a multi illumination rendering module synthesizes images under varying lighting conditions, which are processed by a pre trained photometric stereo network to infer detail-rich normal maps. These normals are then used to jointly refine sur face material properties and fine-scale geometry, yielding high-fidelity reconstructions. Experiments on synthetic and real-world datasets demonstrate that the proposed method outperforms state-of-the-art approaches in reconstruction ac curacy and detail preservation.

Accurate cattle live weight estimation is vital for livestock management, welfare, and productivity. Traditional methods, such as manual weighing using a walk-over weighing system or proximate measurements using body condition scoring, involve manual handling of stock and can impact productivity from both a stock and economic perspective. To address these issues, this study investigated a cost-effective, non-contact method for live weight calculation in cattle using 3D reconstruction. The proposed pipeline utilized multi-view RGB images with SAM 3D-based agreement-guided fusion, followed by ensemble regression. Our approach generates a single 3D point cloud per animal and compares classical ensemble models with deep learning models under low-data conditions. Results show that within this dataset, SAM 3D with multi-view agreement fusion outperforms other 3D generation methods, while classical ensemble models provide the most consistent performance under the studied low-data setting (R$^2$ = 0.69 $\pm$ 0.10, MAPE = 2.22 $\pm$ 0.56 \%), indicating strong potential for practical on-farm use. These findings demonstrate that in this low-data setting, improving reconstruction quality appears more critical than increasing model complexity for downstream weight estimation.

Though there exists a reasonable forward model for blur based on optical physics, recovering depth from a collection of defocused images remains a computationally challenging optimization problem. In this paper, we show that with contemporary optimization methods and reasonable computing resources, a global optimization approach to depth from defocus is feasible. Our approach rests on alternating minimization. When holding the depth map fixed, the forward model is linear with respect to the all-in-focus image. When holding the all-in-focus image fixed, the depth at each pixel can be computed independently, enabling embarrassingly parallel computation. We show that alternating between convex optimization and parallel grid search can effectively solve the depth-from-defocus problem at higher resolutions than current deep learning methods. We demonstrate our approach on benchmark datasets with synthetic and real defocus blur and show promising results compared to prior approaches. Our code is available at github.com/hollyjackson/dfd.

Structure from Motion (SfM) is essential for multi-view 3D reconstruction, however, its accuracy heavily relies on the accuracy of image matching. While the recent correspondence matching method, MASt3R, enables robust matching even under challenging conditions, it tends to generate incorrect correspondences for non-overlapping image pairs. Consequently, existing SfM methods using MASt3R, such as MASt3R-SfM, suffer from significant degradation in pose estimation accuracy as they incorporate these unreliable matches directly into optimization. To address this issue, we propose G-MASt3R-SfM, a novel SfM pipeline that enhances robustness through two key modules. First, the Graph-based View Pruning (GVP) module constructs a scene graph from matching confidence and geometrically prunes outlier views. Second, the Multi-Stage Optimization (MSO) module progressively refines camera parameters by expanding the optimization scope from local consistency to the global consistency. Experiments on the ETH3D dataset demonstrate that our method achieves state-of-the-art accuracy in both camera pose estimation and 3D reconstruction, effectively suppressing noise caused by outliers.

In oral healthcare and digital dentistry, structured-light-based3D reconstruction is widely used for tooth scanning; however, the lack of distinctive textures and the high similarity among tooth morphologies make point-cloud registration particularly challenging, often leading to tracking loss. We present a robust and efficient global relocalization framework designed for tracking-loss recovery. It first enhances surface texture with adaptive histogram equalization so that ORB can reliably extract rich 2D keypoints; these keypoints are then lifted to produce a sparse, high-quality set of 3D candidate points. Building on this sampling strategy, we compute FPFH descriptors guided by ORB salient features and replace costly k-d tree queries with a grid-based indexing scheme and depth-consistency checks. This reduces neighborhood queries for M ORB-lifted 3D keypoints from O (M log N) to O(M), where M ≪ N. Finally, we employ TurboReg, a linear-complexity graph-based solver,to obtain deterministic, millisecond-level backend pose estimates. Experiments on dental datasets simulating severe tracking loss demonstrate an over 11 × acceleration of the full pipeline while preserving high local accuracy, with a 0.020 mm inlier RMSE and up to83.8% inlier ratio under large viewpoint variations.

Optical-SAR image registration remains a critical and challenging problem in remote sensing due to the substantial differences in content representation between the two modalities. Although recent advances in multimodal image registration have shown promising results, existing methods often struggle to generalize to optical-SAR scenarios, where meaningful structures in optical images may appear as noise in SAR imagery. To address this challenge, we propose a novel cross-modality and semantic-guided keypoint matching framework for optical-SAR image registration. Our approach first performs supervised cross-modality keypoint detection, followed by semantic-guided keypoint description to generate modality-invariant descriptors tailored for optical and SAR data. High-level semantic features are extracted at each keypoint location using Janus-Pro, a pretrained vision foundation model, and integrated with local features through a graph neural network to learn context-aware correspondences across modalities. Extensive experiments on benchmark optical-SAR datasets demonstrate that the proposed method consistently outperforms existing approaches on both easy and challenging registration tasks.

Visual localization seeks to recover the camera pose of a query image with respect to a 3D map, enabling applications such as indoor navigation and AR without GPS. In this work, we propose FMI2P-Loc, a two-stage framework that narrows the 2D–3D gap with foundation models. A 3D reconstruction foundation model is used to enable 3D-to-3D retrieval, avoiding cross-modality errors in the coarse stage. A depth estimation foundation model is used to build the depth map, and finishes pose regression while explicit correspondence building is not required. On the indoor benchmark 2D-3D-S, FMI2P-Loc achieves state-of-the-art localization accuracy, outperforms not only traditional image-based and image-to–RGB-D methods, but also recent image-to–point cloud approaches. Our method greatly improves accuracy at the strictest threshold over the latest image-to–point cloud method in the most challenging situation by 3.8% (85.24% --> 89.05%), evidencing that FMI2P-Loc has better stability and cross-area generalization.

Salmon re-identification in commercial net-pens is challenging due to large populations, which impose strict accuracy requirements and make large-scale labeled data acquisition infeasible. Trajectory IDs can be used as proxy labels, but this introduces trajectory-ID bias. To address these challenges, we propose a patch-based re-identification framework that fuses patch-level predictions into a salmon identity decision. A key component is the prediction of the salmon's lateral line, enabling extraction of texture-anchored patches and patch slices. To enable realistic evaluation, we introduce an experimental setup using multiple cameras placed 6 m apart, allowing the same fish to be recorded in different trajectories. This enables the construction of a cross-camera test set through manual match confirmation. Our ensemble approach outperforms the full-image baseline in same-trajectory validation (0.932 to 0.965 mAP) and cross-camera testing (0.609 to 0.860 mAP). The substantial improvements in the cross-camera setting demonstrate improved generalizability and robustness. Code and data: https://github.com/espenbh/salmon-reid-patch-ensemble.

The Video-based Point Cloud Compression (V-PCC) standard achieves state-of-the-art performance for volumetric video compression by projecting 3D point clouds into 2D videos. Current V-PCC implementations focus almost exclusively on geometry and texture, even though the V-PCC specification allows the signaling of additional attributes, such as point cloud normals that can enhance rendering quality. This paper studies the impact of encoding point cloud normals as an additional attribute in a practical V-PCC encoder. Point cloud normals are integrated for two purposes: 1) replacing encoder-side normal estimation and 2) enabling standard-compliant coding of normals within the bitstream. Our experiments show that replacing encoder-side normal estimation with point cloud normals reduces the bitstream up to 32% while encoding up to 1.40× faster, and even improves visual quality. When point cloud normals are also transmitted in the bitstream, perceived visual quality under lit rendering after decoding greatly improves compared with re-estimated normals. Despite the additional process, the encoder runtime is still sped up by 1.12×–1.18× and the impact on bitrate can limited while preserving quality.

Vision foundation models are widely used as frozen backbones across many downstream tasks, making them a single point of failure under adversarial attack. We study multi-level Floyd–Steinberg error-diffusion dithering as a lightweight, model-agnostic input transformation that disrupts adversarial perturbations while preserving semantic content. Unlike prior work, which was limited to binary dithering, grayscale CIFAR-10, and a single small model trained from scratch, we evaluate across six tasks (classification, segmentation, depth estimation, retrieval, captioning, visual question answering), two model families (DINOv2, PaliGemma), and three attacks of increasing strength (PGD, MI-FGSM, SIA), as well as an adaptive attacker using a straight-through estimator. Our results show that Floyd–Steinberg dithering at intermediate quantization levels, especially when combined with post-processing blur, exceeds or matches all tested baselines, including diffusion-based denoising, with substantially less degradation on clean inputs.

We present a non-learning stereo framework for disparity estimation from severely noisy images. Using the Field of Junctions (FoJ), it retains coarse visual features stable under severe noise for cost volume construction while discarding fine textures inseparable from photon noise. The resulting structural information guides boundary-aware Semi-Global Matching (SGM) that dynamically adapts smoothness penalties to preserve true disparity discontinuities. The output is a sparse disparity map more accurate than those of recent stereo algorithms over unmasked pixels on widely-used benchmark datasets

Point clouds are a fundamental representation for 3D perception in robotics, autonomous systems, and augmented reality; however, despite progress in single-label classification, multi-label object recognition in point clouds—where multiple object categories inherently co-exist—remains underexplored. This paper proposes Multi-ContexPoint-GCN, a context-aware framework that formulates point cloud classification as a multi-label problem by employing Binary Cross-Entropy loss with sigmoid activation and integrating a Graph Convolutional Network (GCN) branch — chosen for its ability to exploit sparse, structured label co-occurrence through localized message passing over a data-driven label graph — to model inter-label dependencies arising from the physical layout constraints of indoor 3D scenes. Extensive experiments on the S3DIS and ScanNet datasets demonstrate the effectiveness and generalizability of the proposed approach across different backbone architectures. On S3DIS, the GCNbranch improves mean Average Precision (mAP) by 3.18% with PointNet and by 1.39% with PointNet++, while on ScanNet, gains of 2.05% and 1.33% mAP are achieved using PointNet and PointNet++ backbones, respectively. These results confirm that explicitly modeling label correlations consistently enhances multi-label point cloud recognition, establishing a robust paradigm for context-aware 3D scene understanding.

Precise cross-domain matching between Unmanned Aerial Vehicle (UAV), imagery and satellite maps is essential for navigation in GPS-denied environments but remains challenging due to drastic variations in scale and rotation. This paper proposes a geometry-aware matching framework that decouples the process into specialized extraction, description, and navigation modules. By incorporating a local patching strategy and the Grid-based Random Sample Consensu (GRANSAC) algorithm, the framework effectively mitigates geometric biases and spatial feature clustering. Experimental results on the MDAS dataset demonstrate that our method significantly outperforms SIFT and SuperPoint in both accuracy and robustness, providing reliable localization precision for autonomous UAV flight.

In this paper, we present a network structure for classifying metadata of cuneiform tablets. The problem is of practical importance, as the size of the existing corpus far exceeds the number of experts available to analyze it. But the task is made difficult by the combination of limited annotated datasets and the high‑resolution point‑cloud representation of each tablet. To address this, we develop a convolution‑inspired architecture that gradually down-scales the point cloud while integrating local neighbor information. The final down-scaled point cloud is processed by computing neighbors in feature space to include global information. Our method is compared with state-of-the-art methods and obtains the best performance

This paper proposes an omnidirectional perception and mapless multi-goal navigation framework for complex indoor environments. The system combines noise-filtered 2D LiDAR with front and rear stereo cameras to improve low-height obstacle perception in the omnidirectional directions. Different to conventional robot navigation system, which uses a pre-build global map, the proposed method repeatedly builds a local map from real-time stereo camera and LiDAR scans, identifies the location, and performs A*-based path planning without any global map. Obstacle maps from both stereo cameras and LiDAR are fused together to generate the global map. Experiments in building hallways and lobbies demonstrate successful omnidirectional navigation to multiple predefined goals.

Neural Radiance Fields (NeRF) have revolutionized photorealistic 3D synthesis but are often limited by prohibitive storage requirements. We propose HyBiC, a novel compression framework that integrates a Hybrid Binary-Real representation with advanced contextual entropy modeling. While building on the efficiency of binary grids, our method overcomes the inherent precision limits of full binarization by introducing a real-valued 3D coarse grid. This hybrid hierarchy preserves critical geometric structures and enhances scene expressiveness without the storage overhead of traditional volumetric grids. To maximize coding efficiency, we introduce a grouped channel-wise autoregressive context model that explicitly targets both spatial and inter-channel redundancies. Experimental results demonstrate that HyBiC achieves a new state-of-the-art (SOTA) in NeRF compression, offering a balance between fidelity and storage efficiency.

To address the inherent challenges of complex pathological feature extraction in cardiac MRI, we propose SAMba-UNet, a pioneering dual-encoder framework that synergistically integrates SAM2, Mamba, and UNet. To bridge the natural-to-medical domain shift, we introduce a Dynamic Feature Fusion Refiner that incorporates multi-scale calibration, which may benefit the representation of subtle lesions. We leave rigorous validation of this effect to future work. Furthermore, we design a Heterogeneous Omni-Attention Convergence Module (HOACM) to effectively fuse SAM2's local spatial semantics with Mamba's long-range dependencies, ensuring a holistic understanding of global context and local details. Benchmarked on the ACDC dataset, SAMba-UNet achieves state-of-the-art performance with a Dice of 0.9103 and an HD95 of 1.0859 mm, significantly improving boundary localization for challenging structures such as the right ventricle. These high-fidelity results provide a robust perception engine for medical robotic systems, facilitating a seamless clinical loop from preoperative planning to intraoperative navigation. Our code is available at https://github.com/CodeHarcourt/SAMba-UNet.

Manual quantification of regulatory T-cells (FoxP3+) in immunohistochemistry images is time-consuming and subject to inter-observer variability. This paper presents a fully automated pipeline for cell segmentation and classification in histopathological images, without any dataset specific fine-tuning. The method combines PathoSAM, a foundation-model-based segmentation approach specialized for histopathology, with a semi-supervised hierarchical classification strategy based on agglomerative clustering and label propagation. All nuclei are segmented and classified as FoxP3-positive or negative using a semi-supervised hierarchical scheme, which supports optional expert refinement. \rev{Results demonstrate the feasibility of a zero-shot, low-annotation pipeline for cell quantification, achieving an F1-score of 0.88 for positive cells on LyNSeC and 0.78 on the FoxP3+ oral mucosa dataset, without dataset-specific fine-tuning.

Polyp segmentation in colonoscopy images is challenging due to low contrast, ambiguous boundaries, and large appearance variations. We propose a decoupled coarse-to-fine segmentation framework that separates semantic localization from boundary refinement to improve stability and correction flexibility. The first stage produces coarse predictions enhanced by a spatially-adaptive high-pass fusion (SA-HPF) module for multi-scale context modeling. A detached pixel-wise gating map then guides the second stage to focus on candidate regions while preventing gradient coupling. SA-HPF integrates fixed multi-scale features with boundary-aware high-frequency information through spatially adaptive fusion, enhancing detail preservation without additional tuning. A lightweight refinement network and learnable logit-level fusion further improve robustness. Experiments on CVC-ClinicDB, CVC-ColonDB, and ETIS-LaribPolypDB demonstrate state-of-the-art performance under challenging imaging conditions. The source code are available at https://github.com/Ger0224/DUNet-SAHPF.

Chromosome instance segmentation is a crucial technique in automated karyotype analysis. Existing methods still exhibit limited performance when dealing with clustered, occluded, and overlapping chromosomes. The primary challenges arise from boundary ambiguity caused by chromosome touch and overlap, as well as the high visual similarity among different instances. To tackle these challenges, we proposes ChromoSIS, a chromosome instance segmentation network based on shape constraints and spatially aware feature enhancement. To alleviate boundary ambiguity, we design a shape-constrained instance representation that explicitly models pixel level boundary discrepancies and enforces continuous geometric constraints. To address feature confusion induced by high inter-instance similarity, we introduce a spatially aware query contrastive learning, which enlarges the feature disparity between adjacent instances, thereby enhancing the discriminative capability for densely distributed instances. Extensive experiments on two representative datasets demonstrate that ChromoSIS consistently outperforms state-of-the-art (SOTA) approaches, highlighting its effectiveness and robustness in chromosome instance segmentation.

The Segment Anything Model (SAM) achieves strong performance on natural images, yet its accuracy degrades on medical image segmentation due to severe domain shifts. SAM-Med2D alleviates this issue through full-model fine-tuning on diverse medical datasets, but it struggles with small anatomical structures (e.g., lymph nodes or tumors) and precise boundary delineation, both of which are essential for clinical reliability. To address these limitations, we propose SAM-Med2D-GeoCrop, a small-target-oriented enhancement of SAM-Med2D. Our method introduces a Region-of-Interest (ROI) cropping mechanism that adaptively increases the effective resolution around small or fine-grained structures, enabling more discriminative local feature learning. In addition, we design a geometry-aware prompting strategy that provides stable geometric guidance to improve localization robustness and boundary accuracy. To further enhance adaptability with minimal overhead, we incorporate Low-Rank Adaptation (LoRA)-based parameter-efficient fine-tuning, allowing the model to better specialize in small-structure segmentation while preserving computational efficiency. Extensive experiments across eight benchmarks show that SAM-Med2D-GeoCrop consistently outperforms SAM-Med2D, achieving up to +28.7\% Dice on ToothF. (52.03 → 80.69), +15.5\% on BraTS25 (68.71 → 84.21), and +6.6\% on LNQ (83.63 → 90.21), demonstrating the strong effectiveness of SAM-Med2D-GeoCrop for small and fine-grained medical structures.

Neuron segmentation can effectively enhance its automatic reconstruction and aid in understanding brain functions. However, the existing neuron segmentation methods fail to incorporate prior knowledge about neuronal fine structures and imaging background, hindering their performance. In this paper, we propose TENS-LLM, a text-guided neuron segmentation method using large language models (LLMs), which employs two LLMs to respectively generate texts describing the dataset-level prior knowledge for neurons and extract the text features. These text features are then fused with the neuronal image features extracted from the encoder of segmentation model, guiding the decoder to focus on the linear fiber regions and improve the neuron segmentation. We also design a topology structure loss to train TENS-LLM. In practical, TENS-LLM extracts the text features in advance, thus incurring virtually no additional computational cost for the segmentation. The experiments demonstrate that TENS-LLM not only effectively enhances existing neuron segmentation models but also improves the downstream neuron reconstruction.

Segment Anything Model 2 (SAM2) demonstrated impressive zero-shot capabilities on natural images but faces challenges in biomedical segmentation due to significant domain shifts and prompt dependency. To address these limitations, we propose a prompt-free, parameter-efficient fine-tuning framework designed for multi-class segmentation on variable-sized inputs. We introduce a convolutional Positional Encoding Generator to adapt effectively to arbitrary aspect ratios and present a dual-adapter strategy: High-Performance Adapter utilizing deformable convolutions for precise boundary modeling and Lightweight Adapter employing structural re-parameterization to minimize inference latency. Experiments on ISBI 2012, Kvasir-SEG, Synapse, and ACDC datasets demonstrate that our approach significantly outperforms strong adaptation baselines. Specifically, our method improved segmentation accuracy by up to 19.66% over the vanilla SAM2, while reducing computational costs by approximately 87% compared to heavyweight medical SAM adaptations, establishing a superior trade-off between accuracy and efficiency.

Is Segment Anything Model 3 (SAM3) capable in segmenting Any Pathology Images? Digital pathology segmentation spans tissue-level and nuclei-level scales, where traditional methods often suffer from high annotation costs and poor generalization. SAM3 introduces Promptable Concept Segmentation, offering a potential automated interface via text prompts. With this work, we propose a systematic evaluation protocol to explore the capability space of SAM3 in a structured manner. Specifically, we evaluate SAM3 under different supervision settings including zero-shot, few-shot, and supervised with varying prompting strategies. Our extensive evaluation on pathological datasets including NuInsSeg, PanNuke and GlaS, reveals that: 1) text-only prompts poorly activate nuclear concepts; 2) performance is highly sensitive to visual prompt types and budgets; 3) few-shot learning offers gains, but SAM3 lacks robustness against visual prompt noise; and 4) a significant gap persists between prompt-based usage and task-trained adapter-based reference. Our study delineates SAM3’s boundaries in pathology image segmentation and provides practical guidance on the necessity of pathology domain adaptation.

Deep segmentation models trained on a labeled source domain often underperform when deployed on an unlabeled target domain due to domain shift, a challenge particularly pronounced in medical imaging. Unsupervised Domain Adaptation (UDA) methods tackle this issue via reducing cross-domain discrepancy and transferring semantic knowledge from source to target domain. However, they do not consider the structure of the representation space crucial for segmentation performance. In order to address this, we propose Equiangular Prototype Alignment (EPA), a feature-level alignment strategy based on a fixed set of equiangular and maximally separated class prototypes structured as a simplex Equiangular Tight Frame (ETF). Unlike existing prototype-based UDA methods that rely on dynamically estimated prototypes, EPA avoids prototype drift by enforcing structured class-level alignment to ETF prototypes in the representation space. Our proposed framework unifies EPA strategy with self training and Fourier-based data-level domain augmentation for improved performance on medical image segmentation. On the MM-WHS dataset, EPA outperforms state-of-the-art UDA methods in MR->CT segmentation by +1.1 mean Dice and 0.5 mean ASD (including a +7.5 Dice gain on the challenging MYO class). Similarly, CT->MR validation improves upon prior baselines by +1.3 mean Dice and 1.6 mean ASD, with a +2.0 MYO Dice gain.

Semantic segmentation of coronary arteries in invasive angiography is critical for clinical diagnosis but challenging, due to complex vascular branching and subtle distal structures. Current topology-aware methods often use post-processing, which decouples topological constraints from end-to-end segmentation learning and can cause fragmented segments or missing small branches. To address these limitations, we propose TCSeg, a topology-consistent framework that embeds anatomical prior knowledge directly into the training objective. Central to TCSeg is a novel graph prior loss, which enforces topological consistency by penalizing deviations from an anatomical graph on a fixed geometric manifold via soft-geodesic propagation. Integrating geometric constraints into optimization enables TCSeg to learn global vascular connectivity end-to-end without extra inference overhead. Experiments show TCSeg consistently improves segmentation performance across multiple CNN architectures and enhances distal branch recovery. This work presents a robust strategy for incorporating geometric priors into deep segmentation models, with potential to improve the reliability of automated clinical analysis for coronary arteries.

Automated polyp segmentation in colonoscopy images plays an essential role in the early detection and prevention of colorectal cancer, one of the leading causes of cancer-related deaths worldwide. Recent deep learning approaches have achieved remarkable segmentation accuracy, but require significant computational resources and a large number of parameters. To support real-time clinical decision making while lowering the computational burden, we thus propose Anisotropic Deformable Network (ADNet), a lightweight polyp segmentation framework incorporating three key principles in specifically designed modules: (1) deformable strip convolutions, (2) differential feature extraction and (3) cross-scale kernel attention mechanism. The first two components are integrated into each encoder layer, while the last is introduced into decoder. These modules support effective feature extraction and multi-scale interaction while maintaining low computational complexity. Extensive experiments demonstrate the competitive segmentation performance and inference time of ADNet compared to state of-the-art approaches, while reducing memory consumption, making it suitable for efficient real-time clinical deployment.

Semi-supervised segmentation is especially fragile in regions where the image signal is structurally ambiguous, such as fuzzy boundaries, small lesions, and intensity overlap between foreground and background. Most uncertainty-aware consistency methods often rely on stochastic sampling, multiple networks, or hard masking of unreliable predictions, which suppresses informative ambiguous regions. We propose LUNA, a method that learns from ambiguity by combining uncertainty-weighted consistency and structural-ambiguity contrastive objectives. LUNA introduces a deterministic, single-pass uncertainty weighting scheme that integrates complementary epistemic and aleatoric cues to modulate voxel-wise consistency, allowing ambiguous regions to contribute to learning instead of being discarded. Additionally, a structural-ambiguity contrastive objective emphasizes boundary-adjacent and gradient-rich regions through targeted hard negative mining, improving feature discrimination in confusing areas. Experiments on challenging medical image segmentation benchmarks (ISLES-2022 and ACDC) demonstrate that LUNA consistently improves boundary delineation and segmentation accuracy over strong semi-supervised baselines under low-label regimes.

Voxel-level annotation for volumetric medical imaging is expensive and difficult to scale, which makes training highcapacity 3-D segmentation models challenging in practice. Transfer learning (TL) from large public datasets is a common remedy, but it can underperform when the source domain differs from the target anatomy and acquisition characteristics, as is often the case for pulmonary nodules. In this work, we propose a masked autoencoder (MAE) pretraining-based approach to break the data efficiency wall of domain difference and present a focused empirical study of domain-specific self-supervised learning (SSL) for 3-D lung nodule segmentation. We evaluate two experimental settings: first, Masked Autoencoder (MAE) pretraining versus random initialization across representative baselines; second, MAE versus Decathlon TL for UNETR++ while testing whether MAE-based pretraining also benefits a CNN baseline (V-Net). MAE pretraining on target-domain CT volumes achieves a Dice Similarity Coefficient (DSC) of 0.307, outperforming random initialization (0.136) and Decathlon weights (0.257). In addition, MAE improves the stability of V-Net in a ”low-data” regime (i.e., with “insufficiently labeled” data), increasing DSC from 0.010 to 0.071. Overall, these results suggest that MAE-based pretraining can provide a practical and robust initialization strategy for volumetric segmentation when labeled data are limited.

Accurate segmentation of infected regions in chest X-rays is critical for diagnosis and treatment planning. Recently, multimodal language-guided segmentation methods have demonstrated superior performance by incorporating clinical text reports. However, these methods require reports at inference time, limiting their application. We propose HyperDistill, a knowledge distillation framework enabling text-free inference while preserving multimodal performance. Our Visual-to-Semantic Projector (VSP) generates pseudo-text embeddings from visual features, replacing the 110M-parameter BERT encoder with 3M parameters. The student learns from a hypergraph-based teacher through three distillation objectives: semantic embedding alignment, hypergraph attention distillation, and output-level soft supervision. Trained on QaTa-COV19 dataset, HyperDistill achieves 88.1% Dice and 97.2% accuracy without text input. With only 29.9M parameters, a 76% reduction from the teacher, it outperforms SGSeg (87.4% Dice, 186.6M parameters) while approaching the multimodal upper bound (90.7% Dice). Ablation studies confirm that hypergraph attention distillation provides the largest contribution.

Trustworthy computer-aided diagnosis requires not only accurate predictions but also reliable confidence estimates. Standard cross-entropy training often yields overconfident models, obscuring prediction errors. We propose an uncertainty estimation framework built on DualU-Net, a multitask architecture for cell instance segmentation, that unifies train-time calibration with single-pass uncertainty estimation. First, we replace the cross-entropy objective with the Brier-Score Cross-Entropy with Gradient Rescaling (BSCE--GRA) loss, improving probabilistic calibration without post-hoc adjustment. Second, we introduce an inference-time uncertainty formulation that fuses complementary cues from both decoders: Shannon entropy from the segmentation head and three Gaussian-based measures from the centroid head capturing mass consistency, spatial alignment, and peak confidence. A threshold--power fusion rule activates centroid-based uncertainty only when it is informative, preventing over-amplification of unreliable signals. Experiments on two histopathology datasets with diverse stains and tissue types show that the proposed approach preserves baseline segmentation, detection, and classification performance while substantially improving calibration and error--uncertainty alignment. The resulting single-model uncertainty achieves error-separation comparable to ten-member deep ensembles and Monte Carlo Dropout at a fraction of the computational cost, and qualitative maps highlight over-segmentation, missed detections, and ambiguous instances. Code and pretrained models will be made publicly available.

Accurate medical image segmentation remains challenging, particularly for clinical applications operating under constrained computational budgets. This paper introduces UNetv2-Lite, a lightweight residual attention segmentation framework inspired by UNetv2 and designed for real-time deployment. The proposed model replaces the transformer-based encoder of UNetv2 with a Res2Net backbone to capture fine-grained multi-scale representations at low computational cost. A Shared Interleaved Attention Module (SIAM) jointly models channel-wise and spatial importance with negligible overhead, while a shared Semantic--Detail Infusion (SDI) block aligns semantic and spatial details across scales and reduces parameter redundancy. UNetv2-Lite achieves an 89.2\% reduction in parameters, requires only 2.06 GFLOPs, and operates at approximately 129 FPS on a single GPU. Experiments on five public polyp segmentation benchmarks such as CVC-300, CVC-ClinicDB, Kvasir-SEG, CVC-ColonDB, and ETIS-LaribPolypDB demonstrate competitive and consistent performance, achieving Dice scores of 0.907, 0.944, 0.928, 0.808, and 0.800, respectively. These results highlight a strong efficiency--accuracy trade-off and support the generalization capability of the proposed architecture across multiple benchmark settings.

Foundation segmentation models such as the Segment Anything Model (SAM) exhibit strong generalization capabilities but often suffer performance degradation under domain shifts in medical imaging. Existing source-free domain adaptation methods primarily use foundation models to assist specialist models, while overlooking their own adaptation and sensitivity to noisy pseudo labels. We propose a source-free domain adaptation framework that explicitly adapts a foundation model through uncertainty-aware supervision and cross-model learning. A source-trained teacher generates filtered pseudo labels and refined prompts to guide adaptation on unlabeled target data. The adapted foundation model then supervises a student model under strong data augmentations, while an exponential moving average update mitigates confirmation bias. Experiments on multi-center benchmarks demonstrate state-of-the-art performance under challenging domain shifts without requiring source data or target annotations. Our study highlights the effectiveness of actively and jointly adapting foundation models in source-free medical imaging scenarios.

Skin cancer is among the most common cancers worldwide, and early detection of suspicious lesions is essential for improving clinical outcomes. Dermoscopic imaging provides a non-invasive tool for lesion assessment; however, accurate segmentation remains difficult due to low contrast, uneven illumination, hair artifacts, image noise, irregular borders, and variations in lesion color, texture, shape, and size. Although deep learning has advanced lesion segmentation, many state-of-the-art models rely on complex encoder, decoder,  architectures with large parameter counts, limiting their suitability for real-time clinical use and constrained deployment. This study presents a lightweight wavelet-assisted deep learning framework for robust and scalable dermoscopic lesion segmentation. The proposed method integrates the discrete wavelet transform (DWT), DWT-based multiscale preprocessing, with a compact deep neural network (DNN) to enhance lesion-specific structures while reducing computational complexity. During preprocessing, dermoscopic images are decomposed into multiresolution frequency components, suppressing high-frequency noise and artifacts while preserving low-frequency lesion information. The reconstructed wavelet-enhanced images provide cleaner and more discriminative inputs for segmentation. A compact DNN learns lesion-relevant features without relying on highly parameterized architectures, while a refinement stage improves boundary continuity, removes residual artifacts, and produces accurate binary lesion masks. The framework was evaluated on ISIC 2016, ISIC 2017, and ISIC 2018 datasets using accuracy, Dice coefficient, Jaccard index, sensitivity, and specificity. Results achieved accuracies of 96.5%, 96.0%, and 97.8%, respectively. Comparisons with U-Net, UNet++, FAT-Net, and Swin-Unet confirm a strong balance between segmentation accuracy, robustness, and computational efficiency, supporting potential use in teledermatology, mobile health, and real-time clinical decision-support systems.

Medical image segmentation is essential for computer-assisted diagnosis and clinical decision support but the limited interpretability of modern deep learning and vision-language models remains a major challenge for real-world clinical deployment. Although recent vision-language segmentation frameworks achieve strong performance, they often operate as black-box systems, reducing transparency and clinician trust. Existing explainability approaches, particularly gradient-based visualization methods, frequently suffer from computational instability, sensitivity to noise, and limited capability in modeling complex multi-modal interactions between image features and semantic prompts. To address these limitations, we propose a gradient-free explainability framework for medical image segmentation that generates anatomically aligned saliency maps by computing similarity relationships between image embeddings and reference feature representations across multiple embedding layers. The proposed framework reveals hierarchical reasoning behavior through a three-level interpretability pipeline consisting of embedding-level feature analysis, organ prototype similarity modeling, and prompt-spatial embedding integration. Unlike conventional saliency techniques that rely on unstable gradient propagation, our approach directly exploits embedding-space relationships to produce robust and consistent explanations of model behavior. Experiments on the CHAOS and FLARE22 datasets covering 13 anatomical organs demonstrate progressive reasoning patterns across network layers, where early layers capture broad anatomical attention, intermediate layers focus on organ-specific regions, and deeper layers produce precise boundary-aware localization. The resulting saliency maps provide clinically interpretable visualizations that enable verification of segmentation decisions against medical expertise, supporting the development of trustworthy and clinically deployable vision-language systems for healthcare applications.

Tooth crack detection using deep learning has recently attracted significant attention; however, existing approaches typically require highly detailed pixel-level annotations that are labor-intensive and time-consuming to generate. In clinical practice, coarse localization of crack regions is often sufficient to assist dentists in diagnosis, reducing the need for ultra-precise segmentation. To address this limitation, we propose a tooth crack segmentation framework trained using rapidly generated brushing-based annotations instead of precise pixel-wise labels. A dataset of 582 dental images, including 395 crack images and 187 normal images, was constructed and divided into training, validation, and testing subsets. The proposed model is based on U-Net++ augmented with Attention Gates (AG) for suppressing false positives and Atrous Spatial Pyramid Pooling (ASPP) for capturing crack continuity and directional information. A composite loss function combining Asymmetric Focal Loss, Dice Loss, Connectivity Penalty Loss, and Soft clDice Loss was employed to preserve thin crack topology while improving segmentation accuracy. Experimental results demonstrated that the proposed model achieved superior overall performance, obtaining an F1-score of 69.49, Crack Detection Rate of 97.5%, Normal Detection Rate of 100%, and clDice score of 76.89. These results indicate that clinically meaningful tooth crack detection can be achieved using rapidly generated coarse annotations, substantially reducing annotation burden while maintaining robust segmentation performance.

SBP-Net reconstructs thin 3D structures - pulmonary arteries, industrial pipes - that defeat standard neural methods, using local sliding-box depth projections completed in 2D and fused back into 3D. Our demo makes this pipeline interactive and live. Visitors orbit a real 3D volume on the provided display and click any point on the structure. Around that click, the system crops a local cube, projects it to six orthographic depth views, runs the actual reconstruction network, and patches the newly recovered voxels into the scene — completing missing thin branches in real time, live on the laptop. Nothing is pre-rendered; every click runs real inference, instantly. A before/after panel reveals what the network sees: the six 2D depth projections given to the model versus the completed output it produces. Visitors can also trigger full-volume reconstruction and watch per-region progress stream across the entire structure, then toggle between point-cloud and solid voxel views. The demo runs entirely on a single laptop driving the provided screen over HDMI - no special hardware - and lets attendees directly experience how projection-based completion recovers fine geometry that surface- and volume-based methods miss.

This demonstration presents MPS-RetNet, a semi-supervised deep learning framework for retinal disease classification designed to improve diagnostic performance under limited annotations while providing clinically meaningful explanations. Participants will interact with a live application that classifies color fundus images into cataract, diabetic retinopathy, glaucoma, and normal categories. The demo illustrates the complete inference pipeline, including image preprocessing, multi-scale feature extraction using a ConvNeXt backbone, prototype-guided representation learning, spatial attention, and quality-aware supervision. Users can upload retinal fundus images and observe real-time predictions together with confidence scores, estimated image quality, and Guided Grad-CAM visualizations highlighting disease-relevant anatomical regions. The system also demonstrates how prototype representations improve class separability and interpretability, allowing attendees to explore feature similarities between different retinal diseases. The demonstration highlights the advantages of semi-supervised learning by explaining how reliable pseudo-labels and quality-aware weighting improve robustness when labeled data are limited. Attendees can compare predictions on images with varying quality and observe how the quality estimation branch influences model behavior. The demo showcases a practical, explainable, and annotation-efficient AI system suitable for scalable retinal screening.

We propose a scalable face avatar communication system combining MPEG Video-based Dynamic Mesh Coding (V-DMC) with the Avatar Representation Format (ARF). The system operates in two modes: a low-bitrate baseline streaming blendshape activation weights and a high-quality enhancement mode transmitting V-DMC inter-coded mesh sequences that capture fine details beyond the blendshape basis. V-DMC compression of ARF avatar assets provides significant size reduction at near-lossless quality. Built on web technologies, the system enables adaptive real-time face avatar communication with automatic mode switching based on available bandwidth. The high-quality streaming mode has been implemented; the full integrated dual-mode architecture is under development. At the Show & Tell session, visitors experience this live in the mobile browsers of ordinary smartphones: a phone's front camera captures a face and transmits a high-quality moving 3D avatar within a fraction of a second, preserving expressions, head motion, and fine details such as smiles and blinks. On-screen indicators reveal live bandwidth, delay, and quality. Running in real time with no special plugins, accelerators, or apps, the demo shows that interoperable, standards-based avatar communication is already practical on today's hardware.

Rolling Shutter (RS) CMOS sensors are cost-effective but introduce severe geometric distortions and motion blur under high-speed motion, limiting their use in critical vision applications. Event Cameras provide crucial motion cues to address these artifacts due to their high temporal resolution. However, existing unified methods using Implicit Neural Representation (INR) often yield suboptimal results, suffering from residual blur and detail loss caused by information loss during implicit compression. We propose a novel framework that explicitly utilizes time-conditional optical flow as the central kinematic constraint to jointly perform RS correction, deblurring, and frame interpolation. Specifically, our model extracts optical flow based on query time and introduces a Flow-Constrained INR to enforce geometric consistency explicitly. By making optical flow the primary constraint, our design ensures the INR learns a physically consistent motion trajectory. Experiments demonstrate that our method achieves state-of-the-art performance, obtaining superior PSNR and SSIM scores across all temporal upsampling factors.

Multi-view Clustering (MvC) aims to leverage complementary information from diverse views to improve clustering performance. Contrastive learning has recently become a powerful approach for MvC, enforcing cross-view consistency while capturing heterogeneous information. However, existing contrastive methods often treat all samples equally when repelling negative pairs or attracting positive pairs, overlooking variations in inter- and intra-cluster similarity and sample confidence. This can lead to suboptimal clustering, particularly for ambiguous samples near cluster boundaries. To address these issues, we propose Adaptive Universum Representation Alignment (AURA), a framework that adaptively creates synthetic negative samples based on cluster similarity to an anchor and modulates attraction and repulsion according to sample-level confidence. AURA also incorporates a warm-up stage followed by hub-penalization to prevent the model from collapsing multiple samples toward a global hub. Experiments on three widely used multi-view datasets demonstrate that AURA improves clustering accuracy and robustness compared to eight competitive MvC baseline methods.

In multi-view 3D human pose estimation, the temporal resolution of reconstructed motion is fundamentally limited by the frame rate of optical cameras. Capturing fast human motion therefore requires high-speed imaging hardware, which dramatically increases system cost and complexity. Although motion capture (MoCap) systems can operate at much higher frame rates, existing datasets and learning pipelines enforce frame-rate alignment between image inputs and pose annotations, effectively discarding high-frequency motion information. To overcome this limitation, we introduce a new problem setting: multi-view frame-rate enhanced 3D human pose estimation, which aims to reconstruct high-frequency pose sequences from standard-frame-rate multi-view images. To enable this task, we build a million-scale dataset where MoCap annotations are recorded at 360Hz while optical cameras operate at conventional frame rates under soft synchronization. We further propose HyperDiffPose, a diffusion-based architecture that performs end-to-end dense pose reconstruction in the temporal domain. Experimental results show that our approach consistently outperforms interpolation-based pipelines and achieves accurate recovery of high-speed motion.

Mesh reconstruction from multi-view images is a fundamental problem in computer vision, but its performance degrades under sparse-view conditions, especially in unseen regions lacking ground-truth observations. Recent diffusion models can synthesize novel views from limited inputs, but their outputs often contain visual artifacts and lack 3D consistency, making them unreliable for mesh optimization. In this paper, we propose a framework that leverages diffusion models to enhance sparse-view mesh reconstruction with improved rendering quality and geometric fidelity. To address the uncertainty of diffusion outputs, we introduce a Consensus Diffusion Module that filters low-confidence generations via interquartile range (IQR) analysis and fuses consistent regions through variance-aware blending, providing stable, agreement-aligned pseudo-supervision. These fused images, combined with sparse ground-truth observations, jointly supervise a NeRF-based reconstruction model, enhancing consistency in both geometry and appearance. Extensive experiments show that our method significantly improves geometric fidelity and rendering quality under sparse-view conditions.

Generating coherent long-form video sequences from discrete text prompts remains challenging due to difficulties in maintaining temporal coherence, semantic consistency, and scene–action continuity across segments. We propose a novel storytelling framework that integrates scene and action prompts through dynamics-inspired prompt mixing. Our approach combines three key components: (i) a bidirectional time-weighted latent blending strategy that enforces temporal consistency between consecutive video segments, (ii) a dynamics-informed prompt weighting (DIPW) mechanism that adaptively balances scene and action prompts at each diffusion timestep based on CLIP-based alignment, narrative progression, and temporal smoothness, and (iii) a semantic action representation that encodes high-level action semantics to modulate transitions according to action similarity. Latent-space blending preserves spatial coherence within scenes, while time-weighted blending introduces bidirectional temporal constraints to prevent abrupt transitions. Together, these components enable fluid and coherent video narratives that faithfully reflect both scene context and action dynamics. Extensive experiments demonstrate that our method significantly outperforms baselines, producing temporally consistent and visually compelling long-form videos without any additional training, thereby bridging the gap between short clips and extended text-driven video storytelling.

Full-frame video stabilization aims to convert shaky videos into smooth sequences while preserving the entire field of view. However, this task faces significant challenges due to information loss and complex warping artifacts near frame boundaries caused by large camera motions. Conventional discriminative approaches generally lack the generative capability to hallucinate missing content, often resulting in sheared or distorted boundaries. To address this limitation, we propose \textbf{StabDiff}, a framework that leverages the strong generative priors of diffusion models for high-quality boundary reconstruction in full-frame stabilization. Rather than applying standard diffusion directly, our approach incorporates two motion-aware components to ensure temporal consistency: a Temporal Motion Guidance module that estimates smooth trajectories for latent-level alignment, and a Generative Boundary Inpainter that synthesizes spatially coherent content for missing areas. Extensive experiments demonstrate that StabDiff effectively mitigates boundary artifacts and outperforms existing baselines, producing visually plausible and stable video sequences. Code is available at: \url{https://github.com/HuyueZhu/StabDiff}.

Hierarchical 3D Gaussian Splatting methods like ScaffoldGS employ uniform voxelization to reduce storage costs.However, this ignores spatial complexity variations, causing detail loss in complex regions and computational waste in simple ones. We propose CA³ (complexity-aware adaptive anchor allocation) to address this limitation. Unlike prior work, CA³ enables spatially-adaptive anchor distribution through multi-dimensional complexity assessment.Our method computes three complementary metrics: geometric curvature, texture variation, and multi-view observation frequency. Each metric captures orthogonal scene properties. An information-theoretic stratification strategy leverages quantile-based thresholds to partition points into complexity levels, ensuring balanced resource distribution.Level-specific voxel sizes then adaptively allocate dense an chors for complex regions and sparse anchors for simple ones. A priority-based merging mechanism preserves high complexity anchors while removing boundary redundancy.Extensive experiments demonstrate that CA³ consistently outperforms existing methods, achieving superior rendering quality with reduced computational overhead.

Event camera data carries rich spatio-temporal cues, yet it presents Spiking Neural Networks with representational bottlenecks caused by spatio-temporal coupling. Recent SpikeFormer, which introduces Transformer into SNNs, has achieved low-power advantages through event-driven sparse computation, but it either focuses solely on the spatial dimension or only on the temporal dimension, making it difficult to fully exploit the joint spatio-temporal potential. To address this, this paper proposes a general ”decoupling-first-thenfusion” paradigm: two completely independent spike-driven pathways are constructed in parallel at the input end, which model long-range dynamic dependencies along the temporal dimension and capture global static structures along the spatial dimension respectively. Subsequently, the features from the two pathways are fused, and cross-dimensional cooperative attention is applied to achieve spatio-temporal complementary enhancement. Experiments on HMDB51-DVS, UCF101-DVS, DailyDvs200, and HARDVS human action recognition datasets demonstrate that this method achieves stable performance improvements compared to single-path SpikeFormer, validating the effectiveness of explicit spatiotemporal separation and late fusion for spike Transformers.

Reconstructing complete 3D geometry from monocular clinical endoscopic videos is challenging due to weak texture, repetitive tissue patterns, and severe illumination artifacts. Although emerging dense matching methods exhibit improved resilience to textureless regions, they often produce abundant spurious correspondences across non-overlapping views, which corrupts the correspondence graph and causes structure-from-motion (SfM) pipelines to fail. In this work, we propose a framework for the construction of dense correspondence graph that leverages explicit temporal locality, parallax-driven geometric constraints, and loop-closure revisiting to enable reliable SfM for monocular endoscopic videos. Instead of exhaustively connecting all frame pairs, the approach effectively suppresses invalid inter-frame correspondences while preserving essential long-range geometric relations critical for stable reconstruction. Combined with illumination-aware masking and SfM initialization adapted to endoscopy, the proposed framework achieves substantial improvements in registration robustness and reconstruction completeness for both phantom and real clinical datasets.

Event-camera information is carried by the joint spatio-temporal structure of asynchronous events rather than intensity frames, yet the underlying event distribution is rarely modeled explicitly. We propose a distributional framework in which events are represented as binary variables and modeled by an exponential family with explicit interaction terms. An unsupervised hierarchical dictionary learning procedure driven by information-geometric criteria controls combinatorial growth and yields a sparse, interpretable set of higher-order spatio-temporal interactions, each with a quantified contribution to the distribution. Experiments on a synthetic translating-bar dataset show that the learned dictionary naturally separates into spatial, temporal, and spatio-temporal atoms, with higher-order interactions capturing coherent motion trajectories. Experiments on real event-camera data confirm the emergence of similarly structured interactions, demonstrating an interpretable, distribution-level model without task-specific supervision or motion priors.

Training deep learning models to generate dynamic 3D point cloud sequences requires supervision that provides both per-frame geometric fidelity and temporal coherence. Loss functions such as the Adaptive Probabilistic Matching Loss (APML) are effective for reconstructing static scenes, but they operate frame-by-frame and do not prevent temporal artifacts like jitter and structural inconsistency in dynamic sequences. This work introduces t-APML, a loss function that promotes spatio-temporal consistency by augmenting APML with a temporally-aware guidance term. The method uses an extra ground truth (ExGT) frame from a temporal neighbor as an anchor to encourage smooth transitions. An adaptive gating mechanism dynamically modulates the influence of this temporal anchor by measuring the true motion between ground truth frames, enforcing smoothness during slow movements while permitting rapid changes. To evaluate temporal performance, we propose a set of Optimal Transport-based metrics that quantify the alignment of predicted motion with ground truth dynamics. Experiments on the MM-Fi dataset, generating human point clouds from WiFi-CSI signals, show that t-APML improves the results on temporal metrics, while also improving the spatial accuracy per-frame compared to the original APML and other state-of-the-art losses.

Spatiotemporal fusion (ST fusion) is a promising approach to address a trade-off between spatial and temporal resolutions in satellite images. However, most existing methods assume cloud-free observations, even though clouds frequently contaminate satellite images in real-world scenarios. In this paper, we propose a cloud-robust ST fusion method that directly operates on cloud-contaminated observations. We introduce observation models for cloud-contaminated images, and design regularizations and constraints that separately characterize the cloud component and the underlying high spatial and high temporal resolution images. Based on these components, we formulate ST fusion with simultaneous cloud removal as a constrained convex optimization problem and develop an efficient algorithm based on a preconditioned primal–dual splitting method. Our method does not require precomputed cloud masks as the algorithm generates the cloud masks automatically during the iterations. Experiments on simulated data demonstrate that our algorithm provides more cloud-robust ST fusion than two-stage pipelines that sequentially apply cloud removal and ST fusion.

3D Gaussian Splatting (3DGS) has emerged as a prominent framework for real-time, photorealistic scene reconstruction, offering significant speed-ups over Neural Radiance Fields (NeRF). However, the fidelity of 3DGS representations remains heavily dependent on the quality of the initial point cloud. While standard Structure-from-Motion (SfM) pipelines using COLMAP provide adequate initialisation, they often suffer from high computational costs and sparsity in textureless regions, which degrades subsequent reconstruction accuracy and convergence speed. In this work, we introduce an AV1-based feature detection and matching pipeline that significantly reduces SfM processing overhead. By leveraging motion vectors inherent to the AV1 video codec, we bypass computationally expensive exhaustive matching while maintaining geometric robustness. Our pipeline produces substantially denser point clouds, with up to eight times as many points as classical SfM. We demonstrate that this enhanced initialisation directly improves 3DGS performance, yielding an 9-point increase in VMAF and a 63% average reduction in training time required to reach baseline quality.

Remote photoplethysmography (rPPG) allows for the contactless estimation of physiological signals from facial videos by analyzing subtle skin color changes. However, rPPG signals are extremely susceptible to illumination changes, motion, shadows, and specular reflections, resulting in low-quality signals in unconstrained environments. To overcome these issues, we present a Reliability-Aware Weighted Multi-Scale Spatio-Temporal (WMST) map that models pixel reliability through the suppression of environmental noises. These noises are modeled using different weighting strategies to focus on more physiologically valid areas. Leveraging the WMST map, we develop an SSL contrastive learning approach based on Swin-Unet, where positive pairs are generated from conventional rPPG signals and temporally expanded WMST maps. Moreover, we introduce a new High-High-High (HHH) wavelet map as a negative example that maintains motion and structural details while filtering out physiological information. Here, our aim is to estimate heart rate (HR), and the experiments on public rPPG benchmarks show that our approach enhances motion and illumination robustness with lower HR estimation error and higher Pearson correlation than existing Self-Supervised Learning (SSL) based rPPG methods.

Automated quantification of dental plaque removal is an important tool for evaluating toothbrush cleaning efficacy in simulated in-vitro trials. However, precise assessment faces two persistent challenges in video-based analysis: topological ambiguity caused by inter-dental adhesion and temporal instability due to rapid cleaning dynamics. Existing methods often rely on heuristic post-processing masks or implicit feature learning, failing to enforce anatomical constraints during plaque residue detection. In this paper, we propose the Topology-Prompted Spatio-Temporal TransUNet (TP-ST-TransUNet), a framework that reformulates anatomical priors as learnable geometric prompts. Specifically, inter-dental boundary heatmaps are transformed into high-dimensional geometric embeddings and injected through a Topology-Cross-Attention (TCA) module. This mechanism enables visual features to actively query geometric structure, adaptively suppressing artifacts in gap regions while preserving residual stained plaque on valid tooth surfaces. A ConvLSTM bottleneck is further integrated to model the temporal evolution of the cleaning process. Experiments on the dynamic Dental Plaque Video dataset demonstrate that TP-ST-TransUNet achieves superior segmentation accuracy and temporal consistency over recent baselines, providing a reliable automated metric for toothbrush quality control.

Quanta image sensors provide single-photon sensitivity by outputting high-frame-rate binary measurements. A common reconstruction strategy is to aggregate binary observations over time and/or space to form lower-dimensional summary representations, thereby improving robustness. However, such early aggregation can discard spatio-temporal information contained in the raw measurements. In this paper, we propose an objective-driven reconstruction framework that avoids early aggregation by modeling the entire binary frame sequence as a third-order binary tensor. The proposed method estimates a multi-bit intensity sequence via probabilistic binary tensor decomposition, imposing a low-dimensional spatio-temporal structure directly on the raw bit-plane tensor. Experimental results demonstrate the potential utility of the proposed tensor-based approach in photon-limited imaging.

This paper proposes a Crop-Location-Aware loss gating method between random crops in self-supervised contrastive learning. The method gates the contrastive loss between two crops based on their spatial distance in the original image, so that spatially distant crops are not forced to be attracted in the feature space. Evaluated on three histopathology classification benchmarks, the proposed method outperforms the baseline, demonstrating the effectiveness of gating the contrastive loss based on the spatial distance between crops.

Human-machine interaction (HMI) requires low-latency interfaces and reliable hand tracking. Current models like the Segment Anything Model (SAM) provide robust segmentation but have too much computational complexity for real-time use. On the other hand, frameworks like MediaPipe track landmarks but struggle to distinguish between very close postures (e.g., pinch vs. near-pinch), creating unnecessary computational overhead when only a few key points are needed for specific tracking tasks. To overcome these limitations, we propose a lightweight, real-time multi-task network dedicated to hand image analysis. Using a pre-trained ResNet-50 encoder and a fine-tuned U-Net decoder, the architecture splits into three decoding branches. These branches jointly predict a hand segmentation map, a hand 'cursor' coordinate (the hand centroid), and a posture classification (pinch, almost-pinch, open, or closed). The model was trained on 7,000 RGB images using ground truth masks genrated with SAM2 and centroids calculated with three MediaPipe landmarks. We optimized the network with a joint loss function, weighting the regression MSE loss to prevent the segmentation gradient from dominating. Evaluation was made using Binary IoU for the segmentation and MAE for the centroid, and shows that this approach suits targeted HMI applications. Preliminary results are as follows: the model achieves a Binary Intersection of Union of 97.85 % for segmentation and a landmark Mean Absolute Error of 3.45 10-2.

Since event cameras offer high temporal resolution, high dynamic range, and robustness against motion blur, they have been widely used in many vision tasks, especially in stereo depth estimation. In this paper, we propose an event-image deep stereo framework using multi-scale cross-modal attention, named MS-CMAStereo. To learn the complementary features from both modalities, we design a cross-modal Attention architecture that exchanges global complementary features between two modalities. Then, we perform multi-scale feature fusion to learn the interaction from the exchanged features across multiple scales. Experiments on both real-world and synthetic datasets demonstrate that our network effectively captures multi-scale and cross-modal interactions, producing both global and local fused features, and achieves outstanding performance in terms of quantitative metrics and visual quality.

The task of Long-tailed Class Incremental Learning (LT-CIL) addresses the sequential learning of new classes from datasets with imbalanced class distributions. This scenario intensifies the fundamental problem of catastrophic forgetting, inherent to continual learning, with the dual challenges of under-learning minority classes and overfitting majority classes. To tackle these combined issues, this paper proposes two main techniques. First, we introduce gradient consistency regularization, which leverages the moving average of gradients to suppress abrupt fluctuations and stabilize the training process, thereby preventing early-task bias that typically degrades minority class recognition. Second, we dynamically adjust the weight of the distillation loss based on training progression and the degree of class imbalance measured by normalized entropy across all observed tasks. This adaptive weighting establishes an optimal balance between retaining old knowledge and acquiring new information, particularly preventing majority class dominance in the distillation signal. Experiments on the CIFAR-100-LT, ImageNetSubset-LT, and Food101-LT benchmarks show that our method achieves consistent accuracy improvements of up to 5.0\%. Furthermore, we demonstrate dramatic gains in the challenging 'In-ordered' setting, where tasks progress from majority to minority classes, highlighting our method's robustness in mitigating forgetting under unfavorable learning dynamics. This enhanced performance is achieved without a significant increase in computational overhead, demonstrating the practicality of our framework. Our method is compatible with existing distillation-based approaches, providing a lightweight and generalizable enhancement for LT-CIL. Our method shows consistent improvements under both random initialization and partially pre-trained scenarios.

Current knowledge distillation (KD) methods for semantic segmentation focus on distilling the teacher's knowledge via logit and feature-based techniques. Recent work explored the improvement of knowledge distillation methods by incorporating the uncertainty of the teacher in dense prediction tasks, primarily in object detection. Yet, its application in knowledge distillation for semantic segmentation has received limited attention. Moreover, utilizing the uncertainty on the student side remains largely underexplored. We posit that student-side uncertainty can serve as a valuable signal for guiding the distillation process in semantic segmentation. To this end, we propose Focus on the Fog (FOTF), which leverages student-side uncertainty estimated via Monte Carlo dropout to dynamically modulate the distillation loss. This amplifies the signal in "foggy" low-certainty regions, providing targeted guidance for semantic segmentation. Experiments on Cityscapes, CamVid, and Pascal VOC show that FOTF substantially strengthens simple KD and can complement stronger distillation objectives in several settings. On Cityscapes, Vanilla KD + FOTF reaches 77.37 mIoU, corresponding to a +3.16 improvement over the student trained from scratch without distillation and outperforming Af-DCD in this setting. Our implementation is available at: https://github.com/medem23/FOTF.

Vision Foundation Models (VFMs) show superior performance in visual tasks. To learn their superior representation capabilities in lightweight CNNs, knowledge distillation is a promising approach. However, architectural gaps cause attention mismatch. Class Activation Map(CAM) serves as an architecture-agnostic semantic bridge due to their fixed category dimension. We conducted an in-depth investigation into the generation mechanism of CAM, revealing that both feature generation and CAM alignment exert influence on heterogeneous distillation. To address this, we propose \textbf{D}ual \textbf{A}ttention \textbf{T}ransfer (DAT). DAT consists of two synergistic components: 1) Proxy Cross-Attention Transfer (PCAT) maps student features to a teacher-aligned proxy space to generate refined, teacher-informed representations;2) Salient Class-wise Attention Transfer (SCAT) builds on these high-quality features, using teacher-derived saliency weights to guide CAM alignment toward critical categories. Extensive experiments show that DAT consistently improves CNN student performance, achieving state-of-the-art results in VFM-to-CNN transfer.

While existing large scale deep learning models exhibit excellent semantic representation capabilities in semantic communication scenarios, the transmitter is constrained by limited computing resources and strict power consumption. To this end, this paper proposes a robust knowledge distillation powered lightweight semantic communication method for remote sensing image. The proposed method consists of three core components: a knowledge distillation module, a lightweight semantic encoder and an SNR adaptive channel encoder-decoder module. Specifically, the knowledge distillation module enables effective knowledge transfer from a large scale model to a compact lightweight model, which serves as the lightweight semantic encoder.The lightweight semantic encoder is optimized to extract discriminative semantic features and accurately recover semantic categories.The SNR adaptive channel encoder-decoder is designed to enhance system robustness against diverse channel impairments and noise interference.Extensive experiments conducted on benchmark NWPU-RESISC45 dataset validates that our method achieve higher transmission accuracy and robustness than state-of-the art approaches under limited computing resources.

Zero-shot cross-modal retrieval represents an paradigm in cross-modal information retrieval, aiming to enable models to effectively retrieve across different modalities on previously unseen instances during training. A key challenge lies in learning the generalization capability without explicit knowledge of the connection between seen and unseen semantics, e.g., attribute labels. To this end, we propose a novel framework that inputs category information into LLMs to obtain class-related adjectives and nouns as attribute, correspondingly introducing attribute-specific prompts into existing vision language models to capture more fine-grained and transferable knowledge. Furthermore, a textual constraint term is incorporated to strengthen semantic alignment. Extensive experiments on three multimodal datasets demonstrate that our approach delivers both efficient and robust performance for zero-shot cross-modal retrieval tasks.

In this paper, we propose a learning-based method to improve the robustness of 3D pose lifting models against noise in 2D pose detections. Our approach leverages knowledge distillation with privileged information, where a student model is trained to replicate the internal representations of a teacher model that has access to noiseless 2D poses during training. Specifically, temporal latent features extracted from the teacher model are transferred to the student model, enabling it to learn noise-corrected representations from noisy inputs. This method addresses two major types of noise in 2D pose detection: jitter noise and structured mislocalization (SM noise). While jitter noise consists of small, frame-wise fluctuations, SM noise involves systematic errors at specific incorrect locations. Unlike conventional filtering-based approach, i.e., PoseFormerV2, our method can distinguish and correct both types of noise. We validate the effectiveness of our method on recent state-of-the-art architectures, including PoseFormerV1, KTPFormer, and D3DP. Experimental results show that our method improves performance on sequences affected by SM noise by up to 7.7 millimeters in Mean Per Joint Position Error (MPJPE).

Live streaming presents diverse video qualities and complex degradations, making image restoration and quality enhancement essential yet challenging since the algorithm must run in real-time.  While existing high-performance image restoration methods achieve good adaptability, they often rely on heavy computational costs. Knowledge distillation can transfer capabilities to lightweight networks. However, Standard Knowledge Distillation (SKD) often weakens the model's ability to adapt to the variance of the input content, especially in terms of image quality.  To address this issue, we present an effective real-time adaptive image restoration method for live streaming, featuring a novel Decoupled Knowledge Distillation (DKD) strategy. Unlike existing distillation-based methods, the DKD strategy decouples the knowledge distillation process into two stages, targeting lower-quality and higher-quality input separately, enabling the lightweight model to learn adaptive capabilities effectively. Furthermore, we propose a subject-oriented training strategy tailored for improving the model's performance on faces and text, both of which are of particular interest in live streaming applications. Based on these strategies, we develop a compact network named DKDSoftNet. Extensive experiments show that DKDSoftNet performs favorably against state-of-the-art methods on real-world live video benchmark datasets, while achieving a $389\times$ speed-up over Real-ESRGAN.

Image denoising methods often suppress semantic features, leading to degraded performance in semantic segmentation. Existing segmentation-aware denoising approaches typically rely on joint training of denoising and segmentation models, which often compromises visual fidelity to prioritize segmentation. To address this limitation, we propose LR-RAM, a backbone-agnostic test-time adaptation framework specifically designed to recover segmentation-relevant semantics by adapting the restored image. Our approach combines (1) logit refinement (LR), which corrects uncertain predictions using a lightweight residual module, and (2) region-constrained activation maximization (RAM), which modifies the input pixels to be more optimal using activation maximization. Extensive experiments demonstrate that LR-RAM consistently improves semantic segmentation accuracy across diverse backbones, while preserving restoration quality.

Existing knowledge transfer methods often overlook the representativeness of sampled training data and the negative impact of missing classes within mini-batches, both of which can hinder the optimization process. In this paper, we introduce a prototype-based knowledge transfer framework that leverages representations of informative samples, referred to as prototypes, during the transfer process, and aligns the distribution of data in the feature space rather than individual representations. Prototypes are carefully selected to encapsulate the defining characteristics of their respective classes, ensuring that the student model captures essential structural properties of the teacher's knowledge. This approach enables a more balanced and informative transfer across mini-batches, improving generalization while maintaining computational efficiency. Experimental evaluation across multiple datasets and architectures demonstrates that the proposed method consistently improves over baseline knowledge transfer approaches, achieving superior performance for lightweight student models.

Wildfire damage assessment plays a critical role in natural disaster response, where the ability to retrieve semantically similar regions is essential for situational awareness and planning. While real-time retrieval systems rely on compact and efficient Deep Neural Network (DNN) models, existing approaches often overlook the importance of learning discriminative, retrieval-oriented feature embeddings that capture subtle visual differences across affected areas. In this work, we enhance deep embedding learning for image retrieval through a novel triplet-based correlation-aware knowledge distillation (CAKD) framework that captures semantic structure and visual correlation patterns. Our strategy optimizes the Student DNN feature space by distilling structural dependencies through triplets while aligning spatial feature maps between teacher and student, enabling retrieval-effective embeddings. Extensive experiments on the BLAZE+ dataset demonstrate that CAKD achieves a 3.6\% mAP improvement over standard distillation methods, effectively transferring structural dependencies required for fine-grained disaster assessment.

Epistemic uncertainty estimation is essential for identifying regions where deep learning system outputs may be unreliable. However, existing approaches require computationally expensive ensemble methods or multiple stochastic forward passes, limiting their scalability to dense prediction tasks like segmentation. We propose Resolution-Aggregated Decoder Mutual Information (RADMI), a single-pass method that estimates prediction uncertainty by measuring mutual information (MI) between consecutive decoder layers in segmentation networks. We observe that elevated inter-layer MI correlates with prediction uncertainty, as the network must integrate conflicting contextual information at ambiguous regions such as class boundaries. Evaluating on a seismic facies segmentation benchmark, RADMI achieves the highest correlation with deep ensemble uncertainty among all single-pass methods, outperforming the next-best baselines by 5.5% in Pearson and 10.7% in Spearman correlation coefficients. Compared to baselines that either lack spatial precision or demand significant computational overhead, RADMI yields sharp, boundary-localized uncertainty maps without architectural modifications. Our results suggest that linear aggregation of normalized information flow provides a principled and efficient proxy for prediction uncertainty in encoder-decoder architectures.

Distilling knowledge from large Vision-Language Models (VLMs) into lightweight networks is crucial yet challenging in Fine-Grained Visual Classification (FGVC), due to the reliance on fixed prompts and global alignment. To address this, we propose PAND (Prompt-Aware Neighborhood Distillation), a two-stage framework that decouples semantic calibration from structural transfer. First, we incorporate Prompt-Aware Semantic Calibration to generate adaptive semantic anchors. Second, we introduce a neighborhood-aware structural distillation strategy to constrain the student’s local sample-pair logit relation structure. PAND consistently improves over representative KD and VLM-based distillation baselines on four FGVC benchmarks. Notably, our ResNet-18 student achieves 76.09% accuracy on CUB-200, surpassing the strong baseline VL2Lite by 3.4%. Code is available at https://github.com/LLLVTA/PAND.

Semantic segmentation is a key component of autonomous driving perception, yet state-of-the-art models remain difficult to deploy on vehicles due to strict latency and compute constraints. Knowledge distillation can transfer segmentation capability from a high-capacity teacher to a lightweight student, but traditional logits-based KD implicitly assumes that all teacher predictions are equally reliable. This assumption is violated in dense prediction, where uncertainty varies spatially across the scene, and uniformly distilling such pixels can propagate noisy supervision into the student. To remedy this limitation, we propose an uncertainty-aware KD objective that reweights the per-pixel KL divergence using reliability scores derived from the teacher's predictive entropy. As a result, the student more strongly imitates the teacher in low-uncertainty regions while reducing the influence of high-uncertainty pixels, without introducing auxiliary networks or architectural modifications. On Cityscapes, our method improves a DeepLabV3-ResNet18 student baseline by +1.87 mIoU and outperforms several state-of-the-art approaches by clear margins. The proposed framework is lightweight, stable, and plug-and-play, making it well suited for resource-constrained autonomous driving systems.

Adversarial training (AT) is an effective defense against adversarial examples, but often suffers from robust overfitting, where robustness degrades in the later stages of training, particularly after learning rate decay. In this work, we show that relying exclusively on ground-truth (GT) labels in AT leads to divergent prediction behaviors for clean inputs and adversarial examples, contributing to this degradation. Motivated by this observation, we propose Predictive Label Consistency Adversarial Training (PLCAT), which leverages predicted labels of adversarial examples to complement GT supervision during training. Experiments across three datasets, two architectures, and six AT methods demonstrate that PLCAT consistently mitigates robust overfitting, achieving an average improvement of 12\% in robustness and a 4\% reduction in robustness degradation. Moreover, PLCAT alleviates the trade-off between clean accuracy and adversarial robustness.

Large deep learning models have many applications, but most cannot be deployed on resource-constrained edge devices. A solution is knowledge distillation (KD), which compresses a large teacher model into a smaller student; however, it remains unclear whether students inherit teachers’ explanatory behavior, which is important for regulatory compliance. We introduce KD-Ex, a benchmark designed to evaluate explainability alignment between teacher and student models in knowledge distillation. We also propose KL-XAI, a metric that quantifies divergence between teacher and student Grad-CAM attributions. We evaluate six distillation methods on Flowers-102 and ImageNet-1K. Results show that accuracy alignment does not imply explainability alignment. On Flowers-102, Class Attention Transfer (CAT-KD) achieves the highest accuracy (+2.45%) but weakest explainability transfer, whereas on ImageNet-1K it achieves both. We posit this difference relates to dataset scale or training regime. Existing metrics fail to reliably capture explanation alignment compared to KL-XAI, highlighting the importance of evaluating explainability alongside accuracy for efficient and trustworthy deployment.

Few-shot class-incremental learning (FSCIL) is a paradigm where a model, initially trained on a dataset of base classes, must adapt to an expanding problem space by recognizing novel classes with limited data. We focus on the challenging FSCIL setup where a model receives only a single sample (1-shot) for each novel class and no further training or model alterations are allowed after the base training phase. This makes generalization to novel classes particularly difficult. We propose a novel approach predicated on the hypothesis that base and novel class embeddings have structural similarity. We map the original embedding space into a residual space by subtracting the class prototype (i.e., the average class embedding) of input samples. Then, we leverage generative modeling via a diffusion model to learn the multi-modal distribution of residuals over the base classes, and we use this as a valuable structural prior to improve recognition of novel classes. Our approach, Gen1S, consistently improves novel class recognition over the state of the art across multiple benchmarks and backbone architectures.

Inter-observer variability in crowdsourced and multi-expert annotationintroduces latent label noise into training data, hindering reliablelearning of AI models across many domains.We propose NL-EKD, which incorporates Negative Learning (NL) intoan Ensemble Knowledge Distillation (EKD) framework.The proposed method leverages classes selected by no annotatoras complementary labels, reducing reliance on positive-classinformation prone to inter-annotator disagreement.We validate NL-EKD on two benchmarks: a controlled simulation on theCRC-100K colorectal cancer dataset with symmetric label noise,and a real crowdsourced setting using the CrowdGleason prostate cancerdataset annotated by seven pathologists.NL-EKD consistently outperforms all baselines across both settings.

Degree of sea ice deformation (DoD), in addition to sea ice concentration (SIC) and sea ice thickness (SIT),  is an important sea ice parameter for multiple purposes. Baltic Sea with seasonal sea ice is one of the most heavily trafficked sea ice areas during winters and timely and accurate sea ice information is necessary for safe winter navigation. In this study DoD over the Baltic Sea is estimated from dual-polarized synthetic aperture radar (SAR) data using a U-net neural network model with a self-attention block in the bottleneck. The training data are derived from the daily digitized Baltic Sea ice charts by refining the ice charts based on SAR data. Two year data are used for training  and the estimation accuracy is evaluated for an independent test data set of one additional year. The estimation results are promising and the automated DoD estimation.will later be integrated in the operational Copernicus Marine Service. The major advantages of the automated method compared to the digitized manually drawn ice chart DoD are better timeliness and higher level of detail.

Joint Embedding Predictive Architectures (JEPAs) learn visual representations by predicting latent target embeddings from contextual views, without reconstruction or contrastive negatives. While effective, existing JEPA predictors are shallow feed-forward networks applied once per training step, offering limited control over multi-step latent dynamics and stability. We revisit this design choice and introduce Learnable Iterated Function Systems (LIFS), a structured predictor composed of a contractive mixture of affine maps applied recursively in latent space. LIFS integrates seamlessly into JEPA without modifying the encoder or objective, and introduces explicit geometric inductive bias through spectral control, soft mixture routing, and multi-step refinement. We provide sufficient conditions for contraction under input-dependent mixing and empirically verify these conditions during training. Experiments on CIFAR-10/100 and ImageNet-1K with ResNet-18 and ViT-S/B/16 encoders show that LIFS improves training stability, accelerates predictive alignment, and yields consistent linear-probe gains at matched capacity. Analyses of routing entropy, operator specialization, and FLOPs demonstrate that these improvements stem from predictor geometry rather than increased depth or parameters. Our results highlight structured, contractive predictors as a promising and underexplored design axis for latent-space self-supervised learning.

Remote sensing object detection (RSOD) faces challenges such as difficulty in image acquisition, high labeling costs, and long-tailed distribution, resulting in poor recognition performance for tail categories. Existing oversampling methods fail to provide novel learning samples, while loss optimization strategies often introduce training instability and overfitting issues. To create valuable and realistic objects, we propose a novel Contextual Copy-Paste sample augmentation method for Multi-Class Remote Sensing (CCP-MCRS). This approach synthesizes few-shot and hard samples while preserving the contextual coherence between objects and their backgrounds. First, to retain spatial contextual information of objects, we introduce density clustering method to generate replication region. Furthermore, we design a semantic-context-guided region matching mechanism based on a texture-semantic feature extractor. Extensive experiments are conducted from multiple perspectives on three representative datasets including DOTA-1.0, NWPU VHR-10, and PASCAL VOC. The plug-and-play CCP-MCRS method achieves a 1.12% mAP improvement over the baseline on DOTA-1.0. Quantitative and visualization results demonstrate the effectiveness of our approach in mitigating class imbalance.

Kolmogorov-Arnold Networks (KAN) can achieve superior accuracy over Multi-Layer Perceptrons (MLP) in vision tasks. However, their extensive hyperparameter space (grid size, spline order, base activation, etc.) makes manual architecture selection difficult. We present a Neural Architecture Search (NAS) benchmark dataset of over 150,000 candidate Conv-KAN architectures, combining convolutional feature extractors with terminal KAN layers for image classification on five benchmark vision datasets. Unlike existing NAS benchmarks that sample discrete architectures, our dataset exhaustively covers all valid configurations through two principled constraints–spatial validity and non-decreasing kernel. The five benchmark datasets are MNIST, CIFAR- 10, CIFAR-100, STL-10, and SVHN; all architectures are trained from scratch under a fixed, dataset-specific protocol. We further fine-tune a pretrained autoregressive transformer (Mistral-7B-Instruct) with low-rank adapters that invert the typical NAS structure: given hardware constraints, it directly generates best-candidate architectures, which on average achieve 0.5529% better accuracy with fewer floating-point operations than architectures selected using existing NAS strategies.

With the rapid development of Generative AI making image forgery increasingly realistic, Image Forgery Detection and Localization (IFDL) has become critical for precisely detecting and localizing forged regions. However, current IFDL methods struggle to achieve satisfactory localization precision and robustness, primarily due to their reliance on either physical traces or semantic context alone. Physical-based methods can be confused by complex textures, while semantic-based methods may overlook subtle manipulation traces. To address this, we propose PSCA-Net, a unified framework comprising a Texture Stream and a Semantic Stream that integrates physical traces and semantic contexts to capture complementary cues. Crucially, we introduce a Semantic-Texture Consistency (STC) module that explicitly calculates cross-modal conflicts, utilizing these discrepancies to verify and refine feature representations for precise localization. Extensive experiments on multiple benchmarks demonstrate that PSCA-Net not only achieves SOTA performance in both detection and localization tasks, but also exhibits superior robustness compared to existing methods.

Accurate and automatic pancreas segmentation from computed tomography (CT) images is crucial for the subsequent diagnosis and qualitative treatment of pancreatic diseases. However, accurate pancreas segmentation is a challenging task due to the low-intensity contrast between the pancreas and surrounding organs in CT images. To address this issue, we propose a Prior-Guided Mamba Network (PGM-Net) that leverages dynamic prior information to guide pancreas segmentation in low-intensity contrast. The PADC module exploits pancreas-specific semantic priors and dynamically generates semantically aware convolution weights based on this information. Through prior-guided adaptive convolution, this method effectively activates potential pancreas regions and mitigates ambiguity caused by weak intensity contrast. We propose bidirectional Mamba to capture long-range dependencies efficiently and preserve more spatial relationship information. The comprehensive experiments on the MSD dataset and the NIH dataset have verified the superior performance of the proposed method and its effectiveness in improving accuracy in challenging segmentation scenarios.

Existing model-agnostic training set search techniques primarily rely on deep feature similarity between training samples and target inputs. While effective for high-level vision tasks, the applicability to low-level vision tasks remains largely unverified. In this paper, we propose a target-aware training set search method, HDR-TS, specifically tailored for HDR video reconstruction. The search process follows a coarse-to-fine paradigm. First, a shallow feature-based clustering and matching module is designed to generate a coarse searched set. Target scene inputs are leveraged to match training data clusters in a shallow feature space, enabling the capture of low-level attributes like luminance and texture. Samples are also selectively retained within each cluster based on calculated matching scores. Second, a self-searching strategy incorporating motion-exposure cues is employed to identify samples with maximal diversity to satisfy a predefined size budget. Experiments demonstrate that HDR-TS achieves superior PU21-PSNR and comparable HDR-VDP and HDR-VQM scores. These results indicate that our method effectively identifies representative samples that maintain model performance even with a significantly reduced training set, offering a practical path toward data-efficient training for the low-level image processing community.

Reconstructing ocean surface dynamics from sparse satellite observations remains a central challenge in space oceanography. While traditional variational and optimal interpolation methods provide a principled framework, they are computationally demanding and generally lack guarantees of optimality. Recent end-to-end deep learning approaches address these limitations by directly learning neural mappings from gappy observations to gap-free states. In this work, we propose a unified class of end-to-end mapping schemes that bridges variational data assimilation and state-of-the-art U-Net–based unrolled architectures. Our approach generalizes existing schemes, including score-based, and energy-based models. We report an application to real ocean remote sensing data for the space-time interpolation of the global satellite-derived sea level anomaly. These experiments support the state-of-the-art performance of the proposed physics-informed schemes.

Deep learning models, despite their impressive achievements, suffer from high computational cost and memory requirements, limiting their deployment in resource-constrained environments. Sparse neural networks substantially alleviate these constraints by reducing parameter count and computation. However, existing sparse training methods often struggle to maintain stable optimization and generalization at high sparsity levels. To address this challenge, we propose Zero-Order Sharpness-Aware Minimization (ZO-SAM), a single-pass sharpness-aware optimization framework tailored for sparse training. Unlike standard Sharpness-Aware Minimization (SAM), which requires two backward passes per iteration, ZO-SAM applies lightweight zeroth-order gradient estimation only in the perturbation step, while preserving exact first-order gradients for parameter updates. This design eliminates SAM’s extra backward pass and substantially reduces training overhead without introducing additional heuristics or gradient reuse mechanisms. Extensive experiments on convolutional and transformer-based architectures across CIFAR-10/100 and ImageNet-1K demonstrate that ZO-SAM consistently improves test accuracy when integrated into a wide range of static and dynamic sparse training methods, while recovering a large fraction of the training throughput lost by SAM-based optimizers. As a result, ZO-SAM provides a favorable accuracy--efficiency trade-off, making sharpness-aware optimization practical for sparse training in resource-constrained settings.

Structured illumination (SI) enhances image resolution and contrast by projecting patterned light onto a sample. In two-phase optical sectioning SI (OS-SI), reduced acquisition time introduces reconstruction-based residual artifacts that conventional denoisers struggle to suppress. Deep learning offers a data-driven alternative; however, supervised training is limited by the lack of clean, optically sectioned ground-truth data. We investigate encoder-decoder networks for artifact reduction in two-phase OS-SI, using synthetic training pairs formed by applying real artifact fields to synthetic images. An asymmetrical denoising autoencoder (DAE) and a U-Net are trained on synthetic data, then evaluated on real OS-SI images. Both networks reduce artifacts, exhibiting complementary strengths across two different reconstruction approaches. These results demonstrate that training with synthetic data enables supervised denoising of OS-SI images and highlight the potential of encoder-decoder networks to streamline reconstruction workflows.

Semantic segmentation of high-resolution remote sensing images remains highly challenging due to the confficting demands for global context, precise boundaries, and computational  efffciency. We propose a learnable constrained convolution (LCC) operator that injects multiple constraints, including spectral, spatial, and geometric prior information, into the convolution kernel. The relative importance of these constraints is learned end-to-end. LCC achieves better segmentation of ffne structures without signiffcantly increasing the number of parameters or ffoating-point operations. To further enhance decoding performance, we propose an LCMK module that combines the long-range linear complexity modeling of the Mamba state-space model with the strong nonlinear expressive power of the Kolmogorov-Arnold network. The resulting network model is trained using a twostage training strategy—”constraint warm-up” and ”global fine-tuning”—effectively mitigates overfftting. Extensive experiments on the LoveDA, Vaihingen, and UAVid datasets demonstrate that our model outperforms state-of-the-art models  with comparable parameter budgets.

Hierarchical vision transformers employ shallow feedforward networks (FFNs) with uniform pointwise nonlinearities, limiting channel-wise nonlinear diversity and structured feature interactions. This constraint is particularly restrictive in low-data and fine-grained medical imaging settings. We propose DeKAHT, which replaces the standard Swin FFN with a Grouped Nonlinear KAN (GN-KAN) module. GN-KAN introduces structured nonlinear decomposition via channel grouping, depthwise spatial coupling, and gated re-composition, while preserving Swin’s hierarchical attention design. We investigate two variants: (i) GN-KANGELU, which retains GELU with architectural modification, and (ii) GN-KAN-KANAct, which replaces GELU with a learnable spline-based channel-wise activation inspired by Kolmogorov–Arnold networks. Across Tiny, Small, and Base scales, DeKAHT consistently outperforms Swin. On ImageNet-10, it achieves up to 15% absolute Top-1 accuracy gain, while maintaining comparable computational cost on ImageNet-100. On medical imaging benchmarks, including 4-class and 44-class brain tumor MRI datasets, DeKAHT achieves near-saturated performance, demonstrating strong robustness in low-data regimes. These results suggest that structured nonlinear token mixing, rather than increased model capacity, is key to improving hierarchical transformer performance.

Snapshot Compressed Imaging (SCI) offers high-speed, low-bandwidth, and energy-efficient image acquisition, but remains challenged by low-light and low signal-to-noise ratio (SNR) conditions. Moreover, practical hardware constraints in high-resolution sensors limit the use of large frame-sized masks, necessitating smaller, low-bandwidth designs. In this work, we present an SCI-based computer vision framework using pseudo-random binary masks of only 8$\times$8 in size, aligned with existing CMOS implementations. At its core is CompDAE, a Compressive Denoising Autoencoder with a spatiotemporal Transformer encoder and lightweight task-specific decoders, designed to perform downstream tasks, such as edge detection and depth estimation, directly from noisy compressive raw pixel measurements without image reconstruction. Key innovations include a Noise Level Mapping (NLM) module for unified noise-adaptive training across varying illumination conditions, and hardware-constrained tiled binary masks compatible with existing CMOS designs. A shared encoder paired with lightweight task-specific decoders enables a unified multi-task platform. Extensive experiments across multiple datasets demonstrate that CompDAE achieves strong performance with significantly lower complexity, particularly under ultra-low-light conditions where traditional CMOS and SCI pipelines struggle.

Semantic segmentation is a fundamental task in image processing, with applications, for example, in medical imaging, remote sensing, video analysis, and intelligent transportation systems. The effectiveness of current supervised deep learning approaches is often hindered when the training set is sparse or noisy. This paper proposes KANU-Net-CRF, a novel framework that integrates convolutional Kolmogorov–Arnold networks (CKANs) with conditional random fields (CRFs) for robust structured prediction. CKANs provide a mathematically rigorous and interpretable alternative to convolutional layers, while CRFs enforce spatial coherence and mitigate label imperfections. In particular, the proposed framework is analytically proved to be equivalent, under suitable assumptions, to a CRF model learned nonparametrically through a neural training. We experimentally evaluate KANU-Net-CRF on the ISPRS Vaihingen aerial imagery dataset, comparing it against U-Net, SegNet, UNetFormer, and KANU-Net. The results highlight the potential of integrating interpretable deep architectures based on CKANs with structured probabilistic models, offering a principled and effective solution for semantic segmentation, particularly under the realistic case of limited or imperfect training set.

Vector field reconstruction from sparse sensor data is a fundamental challenge in domains like fluid dynamics and meteorology, where full sensor coverage is often infeasible. Inspired by the emerging generative pretraining paradigm that predicts continuous patch embeddings instead of discrete visual tokens, we propose REPA (Random-order Embedding Predictive Autoregression) for sparse vector field reconstruction. REPA replaces the fixed raster-scan prediction order used in prior formulations with a randomized prediction order. This design helps address the pretraining—finetuning mismatch that arises in reconstruction tasks where predictions benefit from bidirectional context. To evaluate REPA on the vector field reconstruction task, we represent vector fields as multi-channel images and cast the task as a vision problem. In our primary evaluations, REPA consistently improves over the reconstruction baselines.

Class imbalance in LiDAR point clouds poses challenges for semantic segmentation in autonomous navigation and urban mapping. While 2D vision has numerous mitigation techniques, their effectiveness in 3D remains unclear. We benchmark six reweighting schemes and five imbalance-aware losses across three datasets (DALES, S3DIS, STPLS3D) using two architectures (KPConv, RandLA-Net). Inverse-frequency weighting degrades performance by up to 12\% compared to uniform weighting, with catastrophic failures in minority classes. Uniform weighting performs within 2\% of complex losses for structured sampling (KPConv) but benefits less for random sampling (RandLA-Net, up to 4.6\% gap). Loss landscape analysis reveals a complex interplay: for structured sampling, imbalance ratio determines landscape geometry on real LiDAR data but decouples from it on synthetic data; for random sampling, landscapes show high sensitivity to dataset geometry regardless of imbalance ratio. For the two evaluated point-based architectures, these results suggest that the interaction between sampling strategy (structured vs.\ random), imbalance severity, and data acquisition characteristics shapes which mitigation approaches are effective.

Super-resolution and semantic segmentation are known to benefit one another, especially in the Earth observation context. However, learning both tasks in a joint model often requires both task annotations, which is impractical and expensive. In this paper, we study the multi-task partially supervised learning paradigm for both tasks, where each example is assumed to have only a single-task annotation. To that end, we examine two multi-task architectural variations, the sequential and shared variants, then propose a hybrid variant and a re-projection loss to benefit from the shared representation and enforce image quality of super-resolution when training with semantic segmentation. Experiments show favorable results compared to the SOTAs sequential variant.  Source codes will be published at https://github.com/lhoangan/munera.

The rapid advancement of image manipulation techniques has significantly increased the prevalence of highly realistic fake imagery, posing critical challenges for applications relying on the integrity of remote sensing data. This study investigates the detection of manipulated Sentinel-2 RGB imagery from a representation-driven perspective. Challenging the assump- tion that all forgeries share common patterns, we demonstrate that GAN-generated images and copy-move forgery (CMF) exhibit distinct characteristics requiring fundamentally dif- ferent feature representations. Through systematic analysis using frequency-domain (FFT), wavelet-based, and subspace learning methods (Geo-DefakeHop), we show that individual model performance is strictly dependent on the alignment between representation and manipulation characteristics. Ex- perimental results reveal that specialized models, includ- ing representation-specific ResNet architectures and Geo- DefakeHop variants, exhibit limited generalization due to representation mismatches. To address this, a representation- aware hierarchical ensemble framework is proposed. The system integrates these specialized models through a condi- tional decision mechanism that adaptively selects representa- tions based on input characteristics. The proposed ensemble achieves a detection accuracy of 0.715, significantly outper- forming all individual specialized configurations, including the best standalone Geo-DefakeHop (0.645) and various ResNet-based models using FFT and wavelet decomposi- tions (0.513–0.535). By explicitly accounting for representa- tion compatibility, the framework leverages complementary strengths without requiring prior knowledge of the forgery type. Overall, this research establishes that fake image de- tection is inherently representation-dependent and provides a principled, adaptive framework for robust and generalized remote sensing image forensics.

Earth Observation missions generate satellite imagery at scales that outpace conventional pixel-domain processing pipelines; enforcing full reconstruction before inference adds unnecessary overhead, despite image data being routinely compressed for transmission. JPEG AI produces latent representations that encode features that are applicable to vision tasks. While latent domain inference has been demonstrated for face detection, semantic segmentation of remote sensing (RS) imagery, a more demanding task, remains unexplored. We show that a network operating directly on JPEG AI latents can outperform a baseline UNet while utilising fewer parameters.

Accurate estrogen receptor (ER) status determination is essential for breast cancer therapy, yet present multimodal computational pathology approaches only perform cross-modal fusion at the slide level, thereby failing to utilize semantic-visual alignment in a fine-grained manner. We introduce a cross-modal enhancement approach that allows patches to be the unit of interaction and enables individual tissue regions to query pathology reports actively and integrate relevant clinical semantics before diagnostic aggregation. Traditionally, image and text modalities are handled separately in our method, however, it introduces an innovative enhancement feature where each histopathology patch focuses on the report embedding and obtains a contextualized semantic vector via residual integration. The resultant patch representations are semantically enriched and hence carry both morphological features and clinical concepts. Subsequently, a hierarchical dual-stage attention model is employed to distinguish between alignment-based enhancement (α-attention) and diagnostic importance weighting (β-attention), thereby delivering multi-level interpretability. An adaptive gating network dynamically balances the enhanced visual features with global textual representations on a per-sample basis. Evaluated on 58 breast tissue cases using 5- fold stratified cross-validation repeated across 10 independent runs, the method achieves 93.0% ± 4.0% balanced accuracy with an AUC of 0.959 ± 0.024, outperforming the best single-modality baseline (SVM WSI-only: 87.0% ± 2.5%) and the best concatenation fusion approach (SVM Concat: 89.2% ± 0.7%) (both p

Foundation models pretrained on diverse histopathology data are increasingly used for clinical decision support from whole-slide images, yet their fairness across demographic attributes remains insufficiently characterized. Performance disparities across patient subgroups may compromise reliability and generalizability in clinical deployment. In this work, we study fairness in pathology foundation models from an evaluation and experimental design perspective. We introduce a patient-level, demographic-aware framework that enforces balanced test sets and systematically controls training demographic composition under fixed sample budgets. Using four cancer classification tasks, five demographic attributes, and three foundation models, we evaluate fairness with equalized odds and quantify demographic attribute leakage from learned representations. We show that strong predictive performance can coexist with substantial demographic disparities, that training on a single demographic group often improves in-group performance, and that age consistently emerges as the most sensitive attribute. We further demonstrate that classifier design and operating-point selection strongly influence observed fairness gaps, and that simple validation-based threshold optimization can effectively mitigate disparities.

Medical image classification systems often face the dual challenge of data scarcity for training and the need for interpretability for inference. In the fields of cancer and oncology for the early diagnosis of lesions, these problems are even more recurrent. In this paper, we propose OCC-MTL, a knowledge-guided multi-task learning framework designed to enhance classification performance by leveraging domain-specific semantic priors. Unlike standard black-box models, our architecture jointly optimizes lesion classification and visual concept identification. Crucially, we introduce a novel clinical knowledge-guided loss that acts as a semantic regularizer, penalizing predictions that violate expert-defined logic constraints. We evaluate this approach on both a proprietary histopathologically-validated dataset of oral cancer lesions and public benchmarks (SkinCon). Results demonstrate that injecting structured domain knowledge into the loss landscape significantly improves generalization and interpretability compared to standard ViT and concept bottleneck baselines.

Pathological diagnosis is central to cancer diagnosis, but pathologist shortages increase workload, motivating artificial intelligence (AI)-assisted support. Although training models require high-quality labels, expert annotations can be inconsistent. Such inconsistency arises from inter-observer variability, where experts disagree, and intra-observer variability, where an expert's decisions change across repeated reads. We propose the Bias-Aware Adaptive Model Integration Network (BA-AMIN) to enable stable learning under this variability. To mitigate intra-observer noise, BA-AMIN trains one teacher per pathologist. To mitigate inter-observer disagreement, it adds a teacher trained on labels estimated by a probabilistic aggregation method that jointly infers latent labels and annotator-specific error tendencies. Because reliability varies by image, BA-AMIN uses image-wise credibility weights to integrate teachers and distills into a single student for inference. On a 5-class lung adenocarcinoma dataset annotated by five pathologists, BA-AMIN improves Macro-F1 by 0.04 and increases invasive mucinous adenocarcinoma class F1 from 0.71 to 0.84.

Medical anomaly detection is severely constrained by the scarcity of pathological samples. To alleviate this limitation, we propose a text-conditioned and controllable framework for high-fidelity medical anomaly synthesis. The framework adopts a two-stage strategy: first, normal–anomalous image pairs are constructed using Stable Diffusion Inpainting; second, a diffusion-based image editing model is trained to transform normal images into semantically consistent anomalous counterparts guided by textual prompts. The model is trained on six large-scale datasets and evaluated on BrainMRI, CheXpert, and BUSI benchmarks. Experimental results show that the generated anomalies substantially improve downstream anomaly detection performance, outperforming existing synthesis methods. Ablation studies further demonstrate the critical role of text conditioning and diffusion-based editing in preserving image quality and diagnostic relevance. Overall, this work offers an effective and controllable data augmentation solution for medical imaging under data-scarce conditions.

Preserving subtle spectral variations and temporal consistency is critical for reliable endoscopic video denoising. However, most existing methods process RGB channels independently and emphasize spatial noise suppression, often causing color distortion and temporal incoherence. In this paper, we propose STQFormer, a spatio-temporal denoising framework that models color channels as a unified vector entity and captures long-range temporal dependencies via quaternion-based representations. STQFormer integrates Spatio-Temporal Quaternion Transformer blocks with a frame-wise Swin encoder and is trained in an unsupervised manner without paired clean–noisy data. Experimental results on endoscopic video datasets demonstrate that STQFormer achieves state-of-the-art denoising performance while preserving diagnostic color fidelity and temporal coherence.

Evaluation of medical image enhancement remains challenging, as conventional full-reference metrics often correlate poorly with perceptual quality and clinical usability. We propose CRISP, a clinically relevant and interpretable scoring protocol that uses vision-language models (VLMs) to approximate radiologist-style assessment in a scalable and reproducible manner. CRISP formulates evaluation as an input-conditioned triplet comparison, jointly considering the degraded input, enhanced output, and reference image. In addition to an overall acceptability score, CRISP provides interpretable attribute-level feedback reflecting clinically relevant cues. Reliability is further improved through evaluator qualification tests that filter inconsistent VLM evaluators. Experiments on low-field to high-field brain MRI enhancement demonstrate that CRISP provides interpretable VLM-based judgments that show consistent agreement trends with radiologist preferences, with the leading VLM evaluator achieving up to 77% Top-1 agreement with expert judgments, while remaining modality-agnostic.

Artificial intelligence in healthcare requires models that are accurate and interpretable. We advance mechanistic interpretability in medical vision by applying Medical Sparse Autoencoders (MedSAEs) to the latent space of MedCLIP, a vision-language model trained on chest radiographs and reports. To quantify interpretability, we propose an evaluation framework that combines correlation metrics, entropy analyses, and automated neuron naming via the MedGemma foundation model. Experiments on the CheXpert dataset show that MedSAE neurons achieve higher monosemanticity and interpretability than raw MedCLIP features. Our findings bridge high-performing medical AI and transparency, offering a scalable step toward clinically reliable representations. The source code supporting the findings of this study is available at https://github.com/EIDOSLAB/MedSAE.

Non-muscle-invasive bladder cancer (NMIBC) carries a high postoperative recurrence risk, yet existing clinical risk stratification tools demonstrate limited discriminatory performance. In this study, we propose CRI-Net (Clinical-Radiologic Integrative Network), a multimodal deep learning framework that combines deep embeddings extracted from T2-weighted (T2W) MRI with structured clinical variables for recurrence prediction. The radiologic path processes the ResNet-18 embeddings using a multilayer perceptron, and the clinical path uses univariate feature selection and logistic regression. These two branches are merged by a novel Confidence-Guided Fusion mechanism which provides adaptive weights to modality-specific predictions depending on the certainty of the radiologic output, defaulting to clinical evidence in the context of imaging uncertainty. Evaluated on a cohort of 63 NMIBC patients using stratified 5-fold cross-validation, CRI-Net achieved an AUROC of 0.756 and an F1-score of 0.722, outperforming the single-modality paths and the different fusion strategies. These findings underscore the effectiveness of confidence-guided multimodal fusion as a reliable and novel approach for recurrence risk stratification in small sample clinical settings.

Multimodal medical image fusion aims to integrate complementary information from different imaging modalities. This paper proposes a Spatial-Frequency Cooperative Fusion Network, termed SFCFNet, for more effective multimodal medical image fusion. SFCFNet follows an encoder-decoder structure, with two image encoders and one decoder. It employs a Spatial-Frequency Cooperative Fusion (SFCF) block to connect the two image encoders within each layer. Specifically, the SFCF block comprises a frequency-interaction module and a spatial-compensation module. The frequency interaction module employs a discrete Fourier transform to extract phase and amplitude components from both modalities, and then aggregates them separately across modalities using modality-guided gating units to capture global semantics. To extract the local details, the spatial compensation module formulates a refinement mechanism in the spatial domain that adaptively compensates for texture discrepancies between different modalities. Extensive experiments show the effectiveness and superiority of the proposed network over state-of-the-art methods.

Laryngoscopic reports are critical clinical documents in otolaryngology, providing essential records of laryngeal examinations to support diagnosis and treatment planning. However, their manual generation—requiring clinicians to review entire laryngoscopy videos and compose free-text summaries—is labor-intensive and leads to inconsistent documentation in content and style. To overcome these limitations, we present LARES, an end-to-end automated system for generating structured laryngoscopic reports directly from video inputs. LARES first identifies 16 standardized anatomical regions by leveraging temporal stability analysis and deep visual feature extraction, ensuring region-aware representation learning. These visual cues are then integrated into a multimodal large vision-language model to produce draft reports that are structured and clinically meaningful, while remaining editable by physicians. Evaluated on a real-world laryngoscopy dataset from the West China Tianfu Hospital, LARES demonstrates strong agreement with expert annotations in region localization and report generation, while significantly reducing average documentation time compared to manual reporting workflows.

Automatic segmentation and classification of cells in histology images are essential for medical diagnosis and brain cytoarchitectural analysis, yet the availability of large-scale annotated datasets for training deep learning models remains limited. We introduce HistoSmith, a conditional latent diffusion model for generating paired image–label samples to augment histology datasets for cell segmentation and classification. In contrast to prior approaches that decouple label and image generation, our model jointly synthesizes both, enabling controllable data generation through conditioning on user-defined parameters such as cell types. Trained on the H&E CoNIC and Nissl-stained CytoDArk0 datasets, HistoSmith produces realistic and diverse labeled samples that improve segmentation and classification performance while mitigating data scarcity.

Perineural invasion (PNI) is an important pathological indicator associated with poor prognosis in several cancers, yet its detection in histopathologic images remains challenging due to sparse annotations, inter-observer variability, extreme class imbalance, and ambiguous tissue interactions. Existing approaches typically rely on hard morphological dilation of junction annotations or multi-stage pipelines that infer invasion from tumor–nerve proximity, both of which are sensitive to annotation noise and incomplete supervision. In this work, we propose an uncertainty-aware, interaction-driven framework for PNI junction detection that directly models junction likelihood while respecting the limitations of available annotations. Sparse PNI polylines are converted into soft, distance-based labels that encode spatial uncertainty, enabling probabilistic supervision instead of brittle hard boundaries. To further distinguish true invasion from mere tissue adjacency, we introduce an Interaction Fusion Module that conditions predictions on explicit nerve–tumor relational cues. In addition, counterfactual negative augmentation exposes the model to challenging non-invasion scenarios during training. Extensive experiments using a structure-aware, distance-based junction evaluation protocol show that the proposed method consistently outperforms strong single-model and ensemble baselines by 8–16% under a unified setting. Qualitative results further show improved localization, reduced false positives, and smoother junction representations. These findings highlight the importance of uncertainty modeling and interaction-aware learning for reliable PNI detection under realistic annotation constraints.

Federated learning (FL) holds great potential for medical applications. However, statistical heterogeneity across healthcare institutions poses a major challenge for FL, as the global model struggles both to generalize across unseen patient populations and to adapt to the unique data distributions of individual hospitals. This heterogeneity also exacerbates forgetting at both the global and local level, resulting in previous learned patient patterns to be misclassified after model updates. While prior work has largely treated generalization and personalization as separate challenges, we show that a better balance between the two can be achieved through selective alignment with the global model and a modified aggregation scheme, which together mitigate the effects of statistical heterogeneity. Specifically, we introduce FedKPer, which introduces knowledge personalization into the training stage of each local device. Afterwards, generalization is considered via the global model aggregation process, where local updates that are reliable and label-diverse are emphasized. We evaluate the performance of FedKPer, devising additional metrics that relate to common consequences of forgetting. Overall, we demonstrate FedKPer improves the generalization-personalization trade-off without sacrificing retention.

Accurate subtyping of renal cell carcinoma (RCC) remains a critical clinical challenge. Significant morphological overlap between histological variants leads to diagnostic uncertainty and suboptimal treatment decisions. This paper proposes a novel computer-aided diagnosis system that combines rotation-invariant texture analysis with hybrid deep learning for robust RCC classification. The framework introduces an optimized Local Binary Pattern (LBP) algorithm that captures rotation-invariant texture heterogeneity across circular neighborhoods at multiple scales. The algorithm generates discriminative feature maps that encode fine-to-coarse morphological patterns characteristic of RCC subtypes. These texture biomarkers are fused with RGB histopathology images to create enriched multichannel inputs that preserve both chromatic and spatial information. Classification relies on a cascaded MobileNetV3–Transformer architecture. The CNN backbone hierarchically extracts local tissue patterns. The Transformer encoder then captures long-range spatial dependencies and global architectural context through self-attention mechanisms. The evaluation uses 298,071 patches from 383 whole slide images that span six diagnostically challenging renal tumor subtypes. The system achieves a weighted F1-score of 91.20% at patch level and 94.90% at WSI level, with improvements of 3.75– 6.30% over state-of-the-art models. External validation confirms robust generalization (TCGA-RCC: 91.83%; DHMC: 90.66%). These results highlight the potential of the proposed approach as a reliable diagnostic aid for precise RCC classification.

Multimodal medical image fusion aims to combine complementary information from computed tomography (CT), which highlights dense anatomical structures, and magnetic resonance imaging (MRI), which provides superior soft-tissue contrast, to improve clinical interpretation. Despite its potential, effective CT–MRI fusion remains challenging due to the lack of ground-truth fused images and the difficulty of accurately modeling and integrating modality-specific complementary features. In this paper, we propose ESCAN, a novel fusion framework that employs dual modality-specific encoders and a shared decoder built on Enhanced Self-Attention (ESA) blocks to efficiently model both global and local contextual dependencies via channel-wise attention. To further exploit cross-modal complementarity, a Multi-level Adaptive Complementary Feature Fusion (MACF) Module is introduced, which adaptively integrates MRI and CT features across multiple spatial scales. The proposed framework is trained end-to-end using a composite unsupervised loss that preserves intensity consistency, structural similarity, and edge information without requiring ground-truth fused images. Experiments on the SynthRAD 2023 dataset, along with cross-dataset evaluations on the Harvard CT–MRI dataset, demonstrate that ESCAN achieves superior quantitative performance, visual quality, and computational efficiency, while exhibiting strong generalization to unseen data.

Accurate annotation of medical images by domain experts is essential for training reliable deep learning (DL) models, yet labeling errors are inevitable in clinical practice. This study investigates the effect of label noise on transfer learning-based DL classification models applied to two dental radiographic tasks. Label noise was simulated by inverting training labels at rates of 0%, 10%, 20%, 30%, 40%, and 50%. Results showed that classification performance, measured by AUC, degraded markedly at label noise rates of 40% or higher. These preliminary findings suggest that transfer learning-based DL classification models may tolerate moderate label noise, while performance degraded substantially at label noise rates of 40% or higher. This may offer initial guidance for clinical dataset annotation.

We evaluate whether a binary mammography classifier trained on public full-field digital mammography (FFDM) datasets can be redeployed on a private computed radiography (CR) cohort without target-domain weight updates or malignant calibration examples. An EfficientNet-B4 pretrained on VinDr-Mammo and fine-tuned on CDD-CESM low-energy images was evaluated on 468 accessions and 1,936 held-out images using deterministic seven-variant test-time augmentation, post-hoc temperature scaling, and a laterality-aware per-side mean voting rule. At a fixed threshold of 0.5, the pipeline achieved an exam-level AUC of 0.761, balanced accuracy of 0.711, sensitivity of 0.683, and specificity of 0.740. Exam-level multi-view aggregation contributed most of this gain, demonstrating feasible cross-domain deployment from public FFDM to private CR mammography.

We propose a multi-backbone ensemble pipeline for peritoneal metastasis classification from 64×64 endoscopic patches acquired during staging laparoscopy. Four backbone architectures (ConvNeXt-Tiny, CAFormer-S18, Swin-T, DaViT-S) are pretrained via SimCLR on ∼34,000 unlabeled crops from spatial neighborhoods of annotated lesion bounding boxes, then fine-tuned on 365 biopsy-confirmed lesions using 5-fold cross-validation. Ensemble predictions aggregate per-lesion probabilities across models with complementary inductive biases, exploiting architectural diversity spanning convolutional and attention-based representations. The four-model ensemble achieves a mean AUC of 0.813 ± 0.007 across five random seeds, outperforming the prior baseline (AUC 0.785) by +0.028. Cross-seed aggregation of all 20 checkpoints further improves AUC to 0.820.

Training vision–language models (VLMs) for whole-slide image (WSI) analysis is constrained by scarce paired slide report data, especially for tasks such as prostate cancer grading. We ask whether training on translated copies of clinical captions acts as an implicit regularizer improving stability and generalization of WSI–VLMs in the low-data regime using a SlideChat-style pipeline. On DiagSet, bilingual English–Polish training cuts checkpoint volatility 2–3× and improves Polish open-ended METEOR by a relative +35% over monolingual training, while same-language paraphrase controls do not

How can three ordinary photographs become an estimate of a cow’s live weight? Agreement Ranch turns our ICIP 2026 paper into a nine-station interactive journey requiring no 3D vision background. Visitors start beside left, right and top cameras, then watch the animal segmentation from a busy farm scene. At reconstruction, intermediate states recorded from agreement-driven MVSAM3D reveal a process normally hidden inside the model: a rough 3D body emerges point by point, stays visible, and is refined into a colored animal beside it. Visitors replay both stages, orbit the cattle, and inspect how shape and color develop instead of seeing only the final output. At fusion, our main idea becomes visual. Each camera sees different body parts and may produce missing or uncertain geometry. We give more influence to 3D evidence supported by several views and less to conflicting evidence. Visitors watch agreement change during reconstruction and compare single-view, simple averaging, entropy-based and agreement-driven results side by side. The final stations place intuitive body measurements, length, width, height, volume, and shape on the reconstructed animal, then show how they become a live-weight estimate. Visitors can revisit any station, rotate and zoom models, replay reconstruction, and connect better 3D evidence to the paper’s best result (R² = 0.69 ± 0.10; MAPE = 2.22 ± 0.56%).

This demo presents DFSI, a LiDAR distance-field safety plug-in for diffusion-based visual navigation. From an attendee's perspective, the demo addresses a practical deployment question: how can a camera-based diffusion navigator remain safe when obstacle distances are unreliable under scene changes or perception drift? DFSI adds LiDAR-derived distance-to-obstacle safeguarding at execution time, without modifying or retraining the pre-trained backbone policy. Attendees will observe a mobile robot equipped with an RGB camera and a 2D LiDAR navigating indoor and outdoor environments with newly introduced and dynamic obstacles. The demo will show how the Reliability-aware Distance-Field Mapper (R-DFM) constructs a robot-centric Euclidean Signed Distance Field through TSDF-to-ESDF conversion and maintains reliability via the Adaptive Refresh-Management Finite-State Machine (ARM-FSM). It will also illustrate how the Dual-Stage Safety Gate (DSG) uses ESDF-derived clearance to rerank diffusion-sampled trajectories and supervise execution through execute, slow-down, and recovery modes. Through side-by-side comparisons with a vision-only diffusion navigator and a distance-guided baseline, attendees will see how DFSI changes robot behavior near walls, newly placed obstacles, and dynamic objects. The demo highlights reduced collisions and safety interventions, increased minimum observed obstacle clearance, and preserved navigation efficiency with modest latency.

This demo presents Adaptive Zone Merging, a graph-based algorithm for hierarchical over-segmentation. Images will be captured live at the stand, and attendees will be able to interact with the algorithm by adjusting the scale of the resulting superpixels in real time. Starting from an initial over-segmentation, the method adaptively merges zones based on graph structure to produce superpixels that better preserve object boundaries and reduce redundancy at each scale. The demonstration illustrates how this hierarchical approach yields compact, perceptually meaningful visual primitives suitable for downstream vision tasks.

This demonstration showcases a GPU-based simulation framework for time-of-flight (ToF) imaging, presented alongside the corresponding research paper. The simulation generates realistic ToF measurements from virtual scenes, enabling direct comparisons between simulated and real sensor data. During the demo, attendees will be shown a series of representative scenes together with their simulated ToF outputs, illustrating the simulation's ability to reproduce characteristic imaging effects observed in real cameras. Since the simulation is designed for high-throughput offline data generation rather than interactive rendering, the demonstration focuses on the resulting data and visual comparisons. In addition, a physical ToF camera and a small real-world scene will be displayed alongside their virtual counterparts. The simulated scene has been modeled to match the physical setup, allowing attendees to compare captured camera measurements with the corresponding simulated results. This side-by-side presentation highlights the realism of the simulation and demonstrates its potential for algorithm development, validation, and synthetic dataset generation for ToF imaging applications.

Surface defect detection models trained on a source domain often experience substantial performance degradation when applied to new industrial environments due to domain shift. Source-free domain adaptation (SFDA) and test-time adaptation (TTA) aim to alleviate this issue without accessing source data, yet their unsupervised nature makes them vulnerable to unstable optimization, particularly when target domains contain unseen defect categories or shifted label distributions. This paper investigates a few-shot source-free domain adaptation setting for surface defect detection, in which a small number of labeled target samples per class are available while the source data remain inaccessible. We analyze the challenges of adapting high-capacity inspection models under extreme label scarcity and present a simple yet effective Few-shot SFDA framework. The framework integrates stochastic weight reversion to regularize adaptation toward the source-pretrained solution and partial parameter updating to control model plasticity. Experimental results demonstrate consistent improvements over existing source-free and test-time adaptation baselines across different few-shot settings. The study highlights the importance of parameter selection and regularization in achieving stable and reliable adaptation for industrial surface defect detection.

We propose Person–Object Relationship Consistency Learning (PORCL), a simple yet effective framework for enhancing generalization in zero-shot spatio-temporal action detection. Existing methods rely heavily on observed person–object interactions, but often struggle to generalize to new contexts where unseen or additional objects appear. PORCL addresses this challenge with two components: an object-aware input augmentation, which generates augmented inputs by inserting irrelevant objects from other videos, and a relational consistency loss that enforces prediction agreement between original and augmented inputs. This design suppresses irrelevant objects and highlights meaningful person–object relationships, thereby improving robustness in novel contexts. Experiments on JHMDB and UCF101-24 show that PORCL consistently improves baseline models in zero-shot settings, and further boosts performance in few-shot scenarios.

Reliable image recognition systems should both classify known categories and detect novel classes in open-set settings, especially under zero-shot constraints where no training examples are available. We propose a zero-shot OOD detection method that enriches each known label with a semantic hierarchy of fine-grained subcategories. Hierarchies are generated via structured LLM prompts and filtered with a lexical ontology for domain alignment, then integrated into CLIP to exploit coarse-to-fine semantic consistency. This training-free design improves the ability to reject inputs that do not match any known class. Experiments on standard OOD benchmarks show competitive performance and provide a more structured, interpretable prediction space.

Zero-shot anomaly detection has achieved remarkable success in the 2D domain but remains under-explored for 3D point clouds. Most existing 3D approaches rely on projecting data into 2D images, inadvertently discarding intrinsic geometric manifold information required to identify structural defects. To bridge this gap, we propose ULIPAD, a novel framework extending zero-shot anomaly detection capabilities to the 3D domain by leveraging the pre-trained ULIP model to align 3D geometry directly with natural language. Our method adopts a dual-branch design that jointly leverages 3D point cloud features and multi-view image information for effective anomaly localization and classification. To enhance the semantic alignment between textual descriptions and visual features, we introduce a text optimization module that employs similarity-based outlier filtering to refine textual prompts. Extensive experiments on MVTec3D-AD and Eyecandies datasets demonstrate that ULIPAD achieves an i-AUROC of 74.5% and 69.3% respectively, outperforming state-of-the-art 2D zero-shot methods by significant margins.

Industrial defect inspection is challenging due to limited expert-labeled data and significant variability in defect appearance across product categories. In this study, we propose a simple multimodal approach for few-shot defect classification that leverages pretrained vision and language models without extensive full model fine-tuning. Visual features are extracted using DINOv2, adapted via LoRA, then grouped with slot attention into semantically meaningful region representations. Textual descriptions are encoded with CLIP using learnable prompts, enabling class-specific alignment between visual slots and defect-relevant semantic concepts. All pretrained backbones are frozen, and only lightweight, parameter-efficient modules are trained, ensuring high data efficiency. Experiments on industrial inspection benchmarks demonstrate improved multiclass defect classification performance, particularly in few-shot settings with limited data.

Training-free anomaly detection eliminates the need for category-specific training data, dedicated training infrastructure, and per-category model maintenance, making it attractive for industrial inspection with rapidly changing product lines.Among training-free methods, MuSc achieves competitive performance through transductive mutual scoring on unlabeled test sets.However, MuSc fuses multi-layer anomaly maps with fixed uniform weights, discarding layer-wise reliability information that could improve localization.Different transformer layers exhibit varying sensitivity to fine-grained defects versus background clutter, yet this heterogeneity remains unexploited.This paper proposes Adaptive Layer Fusion (ALF), a sharpness-aware module that estimates per-image, per-layer reliability weights directly from the anomaly response distribution.Layers producing concentrated, peaky responses receive higher weights, while those with diffuse activations are down-weighted.ALF requires no learnable parameters and introduces negligible computational overhead.We integrate ALF into MuSc and evaluate on MVTec AD, VisA, and BTAD using CLIP and DINOv3 backbones.Under the token-aligned protocol, ALF improves mean AUPRO from 93.80\% to 93.92\% on MVTec AD, from 92.70\% to 92.92\% on VisA, and from 83.41\% to 84.01\% on BTAD with CLIP; with DINOv3, improvements reach +0.17, +0.69, and +0.70 points respectively.As detection performance on MVTec AD approaches saturation, even modest AUPRO gains correspond to improved localization quality.The largest per-category gain is +4.71 AUPRO on VisA (DINOv3).The code will be available at https://github.com/Deyu008/ALF.

Student-teacher distillation frameworks have emerged as the dominant paradigm in unsupervised anomaly detection. However, existing methods struggle to achieve fine-grained segmentation, particularly in preserving the structural integrity of subtle defects and localizing multiple instances, due to the low-order manifold alignment and inherent denoising-preservation conflict. To address this, we propose DASR-Net. First, we adopt a distribution-aware consistency paradigm to enforce strict high-order manifold alignment by matching statistical distributions between student and teacher features, establishing a clean feature baseline. Then, a selective feature reconstruction module is proposed to decouple anomaly filtering from detail preservation via a complementary semantic compression and texture reflow mechanism. Finally, a residual dense scale aggregation head is introduced to eliminate blind spots and ensure topological continuity. Experiments on MVTec-AD demonstrate that our method achieves 98.1% pixel-level AUC and 77.6% instance-level average precision, validating its superiority in precision industrial inspection.

Zero-shot anomaly detection is critical for scenarios where target-domain data is unavailable. Although existing CLIP-based methods have demonstrated promise, they often neglect the inherent limitations of the visual encoder in modeling local anomalies. Primarily, CLIP's pre-training prioritizes aligning the global class token with textual descriptions, leaving local patch features without fine-grained supervision. Moreover, vision transformers suffer from insufficient local inductive bias for high-frequency defects and rely heavily on implicit self-attention mechanisms, which could further dilute discriminative anomaly signals. To address these issues, we propose CoBi-CLIP, a framework synergizing convolutional inductive biases with explicit global-local interactions. Specifically, we incorporate lightweight convolutional adapters to inject local spatial awareness into the visual encoder, coupled with a bidirectional alignment mechanism that fosters a dynamic interaction between global semantics and local details. Complemented by parameter-efficient bias tuning that harmonizes textual embeddings with visual updates, CoBi-CLIP establishes precise alignment between holistic semantics and fine-grained structural anomalies. Extensive experiments across 11 industrial and medical datasets demonstrate that CoBi-CLIP achieves superior performance compared to existing methods.

Industrial defect detection is a crucial task in manufacturing, associated with many critical applications. Recent approaches adopt the Zero-Shot Anomaly Detection (ZSAD) paradigm to address the task, utilizing an auxiliary training dataset to overcome the scarcity of labeled defective data in real-world applications, with the vast majority of them employing CLIP-based models. In this paper, we propose a novel method for improving the performance of a downstream CLIP-based model for ZSAD, leveraging LMMs as inverse reasoners. Particularly, we first prompt an LMM specialized for anomaly detection, in order to extract counterfactual defect explanations for the normal images of the auxiliary dataset. Subsequently, we introduce this knowledge to the CLIP-based model for ZSAD, in order to improve its defect discrimination ability. To do so, we extract the text embeddings from the LMM-generated counterfactual explanations using the text encoder of CLIP, and introduce an auxiliary objective that minimizes the cosine similarity between the CLIP's image embeddings of normal images and their corresponding text embeddings that convey defect semantics. We apply the proposed method on a state-of-the-art model for ZSAD, considerably improving its performance. The proposed method can be extended to generic detection and classification tasks for improving the baseline performance of CLIP-based models.

The detection of sensational content in media items can be a critical filtering mechanism for identifying check-worthy content and flagging potential disinformation, since such content triggers physiological arousal that often bypasses critical evaluation and accelerates viral sharing. In this paper we introduce the task of sensational image detection, which aims to determine whether an image contains shocking, provocative, or emotionally charged features to grab attention and trigger strong emotional responses. To support research on this task, we create a new benchmark dataset (called Sens-VisualNews) that contains 9,576 images from news items, annotated based on the (in-)existence of various sensational concepts and events in their visual content. Finally, using Sens-VisualNews, we study the prompt sensitivity, performance and robustness of a wide range of open SotA Multimodal LLMs, across both zero-shot and fine-tuned settings.

Test-time adaptation (TTA) improves resilience against distribution shifts by updating models online using unlabeled test data, but fails in practical scenarios when test streams comprise a mix of in-distribution (ID) and out-of-distribution (OOD) samples. Existing noisy approaches rely on sample filtering or auxiliary detectors, which implies that reliability signals stay steady throughout adaptation. This assumption, as we show, fails when the model is updated online. The reliability estimate drifts, causing acceptance criteria to loosen and progressively admit more OOD samples, which amplifies failure. We present ARTTA (Anchored Reliability Test-Time Adaptation), a lightweight noisy technique for CLIP that clearly separates reliability estimation and model adaptation. By limiting changes to LayerNorm parameters and keeping a frozen anchor copy for margin-based confidence calculation, our method minimizes adaptation-induced drift in the reliability signal. Experiments on mixed ID-OOD streams show enhanced ID accuracy with significant OOD robustness, even without learned detectors or pseudo-labels. Our results establish frozen reliability with lightweight adaptation as a simple and stable baseline for noisy test-time adaptation in foundation-model pipelines.

A fundamental challenge in deploying vision models is domain shift, which arises when training and test data follow different distributions, leading to degraded performance. This challenge is amplified when the same semantic concept appears under distinct visual forms, such as photographs and sketches, where visual similarity is weak despite semantic correspondence. Existing unsupervised domain-adaptation methods aim to align distributions across domains but often ignore semantic relationships among samples of the same class. To address this issue, this paper introduces SPACE, a method that exploits the semantic structure of CLIP's vision-language space for domain adaptation. The key idea is to use text descriptions as semantic anchors by applying Singular Value Decomposition to CLIP embeddings of class descriptions, yielding an orthogonal basis that captures semantic relationships among categories. Visual features from both domains are projected into this semantic subspace, aligning images based on meaning rather than appearance.

Large Vision-language models (VLMs) achieve strong zero-shot generalization but struggle to adapt to few-shot training and cross-domain scenarios due to limited labeled data and distribution shifts. Earlier adaptation strategies either rely on gradient-based methods, which are computationally expensive and prone to overfitting, or adopt training-free approaches, such as ProKeR, which ignore correlations between output classes. We propose Bilateral Regularized Kernel Adaptation (BiReg), a closed-form and training-free framework that constructs few-shot adaptation as a multi-output kernel regression problem. BiReg uses bilateral regularization to constrain both the input feature space and the semantic output embedding space, enabling the model to capture inter-class relationships while remaining anchored to the zero-shot CLIP baseline. Experiments across ten diverse benchmarks demonstrate that BiReg achieves superior accuracy, robustness, and scalability compared to prior methods.

Surface anomaly detection of nuclear fuel rod claddings is critical to nuclear safety. In practice, illumination, noise, and reflection cause severe domain shift. With scarce source-domain samples, static-prompt methods cannot adapt to such domain shift, leading to semantic-visual misalignment and degraded detection performance. We propose a Residual-guided Dynamic Prompts framework (ReDyPrompt) for robust anomaly detection under few-shot domain shift. First, we introduce an Prompt Separability Constraint (PSC) that enforces a minimum margin between normal and anomaly semantic, which improves the stability of the decision boundary. Second, we propose Residual-Guided Dynamic Prompts (RGDP), which estimate domain shift using the residual between a test sample and its support prototype, and retrieve domain-relevant descriptors to calibrate the prompts without target-domain annotations or retraining. Extensive experiments on a fuel-rod dataset under four corruptions settings demonstrate strong robustness. ReDyPrompt achieves an image-level AUROC of 99.28%, significantly outperforming static-prompt baselines. Code is available at https://github.com/xwshishuaige/ReDyPrompt.

As a challenging task, the Few Shot Anomaly Detection (FSAD) stems from the real-world industrial scenario where only a few anomaly-free samples are available for training. In this work, we propose to perform the FSAD in an online-adaptive style. In specific, a feature-matching based AD model, pre-trained on the few-shot training set, is continuously adapted on each test image. To avoid over-fitting problem which is crucial to an online adapted model, the online training data is extracted by using a novel and robust sampling method. In this way, our algorithm can effectively absorb the information of the test samples, which is usually ignored in existing SOTA methods, at the cost of slightly lower inference speed. The proposed method is termed Robustly Adaptive Feature maTching (RAFT) and the extensive experiments in this work verifies the superiority of our algorithm.

Zero-shot anomaly detection aims to identify defects in unseen categories without target-specific training. Existing methods usually apply the same feature transformation to all samples, treating normal and anomalous data uniformly despite their fundamentally asymmetric distributions, compact normals versus diverse anomalies. We instead exploit this natural asymmetry by proposing AVA-DINO, an anomaly-aware vision-language adaptation framework with dual specialized branches for normal and anomalous patterns that adapt frozen DINOv3 visual features. During training on auxiliary data, the two branches are learned jointly with a text-guided routing mechanism and explicit routing regularization that encourages branch specialization. At test time, only the input image and fixed, predefined language descriptions are used to dynamically combine the two branches, enabling an asymmetric activation. This design prevents degenerate uniform routing and allows context-specific feature transformations. Experiments across nine industrial and medical benchmarks demonstrate state-of-the-art performance, achieving 93.5\% image-AUROC on MVTec-AD and strong cross-domain generalization to medical imaging without domain-specific fine-tuning. https://github.com/aqeeelmirza/AVA-DINO

Vedic Sanskrit manuscripts on ochre parchment pose a dual challenge for Document Image Analysis: standard binarisation conflates chromatic staining with ink, and no annotated OCR corpus exists for this tradition. We present MorphBG, a CPU-only colour aware pipeline using per-channel morphological background estimation and Weber-contrast normalisation, achieving DRD 2.38 and F- measure 0.981. We release VedOCR: 57 expert-annotated pages with 681 line-level transcriptions and 41,286 character instances across 87 grapheme classes, including archaic conjuncts and Vedic tone marks absent from standard training sets. Zero-shot engines yield 71–78% CER; fine-tuning trocr-base handwritten on VedOCR reduces CER to 48.89% (−26.85 pp), establishing the first deep-learning OCR baseline for this script.

Whole-image CLIP retrieval returns entire frames rather than the queried object, providing no spatial localization. Region-aware encoders address this but retrain the model, confounding indexing granularity with representation learning. We isolate the indexing variable by keeping PE-CLIP frozen and comparing a SAM2-segment FAISS index (55.2 M vectors) against a whole-image index from the same 1.74 M Open Images V7 frames, evaluating with strict bounding-box IoU over N=1160 queries. At K=10, segment indexing delivers oracle localization precision of 0.534 ± 0.029 with +12% relative hit-rate gain at τ=0.7 and median zero-shot query latency of 34 ms on CPU. PE-CLIP text-to-image is the strongest single localizer (HR@10(0.5)=0.551 ± 0.029, +27% over OpenCLIP text, disjoint 95% Wilson CIs).

Accurate detection of bridge surface defects is essential for structural health monitoring and condition assessment. However, existing deep learning-based methods remain limited in complex environments, especially for irregular cracks, small defects, and precise localization. To this end, this paper proposes DSHF, a bridge defect detection framework based on Dynamic Snake Convolution and hierarchical feature fusion. DSHF employs a Dynamic Snake Convolution-enhanced feature extraction module to capture slender and curvilinear defect structures, and designs a three-branch hierarchical fusion strategy to integrate multi-scale semantic and spatial features for improved small-defect detection. Moreover, MPDIoU loss is introduced to enhance bounding box regression by jointly considering overlap, center distance, and scale variation. Experiments on public datasets show that DSHF improves mAP by 3.4 percentage points, with notable gains in precision and recall, demonstrating its effectiveness and robustness.

Night Photography Rendering (NPR) poses a significant challenge due to the extreme contrast between dark and illuminated areas in scenes, stemming from concurrent capture of severely dark regions alongside intense point light sources. Existing methods, which are mainly tailored for fidelity metrics, reveal considerable perceptual gaps and often detract from visual quality. We introduce pHVI-ISPNet, a novel end-to-end RAW-to-RGB framework built on the robust HVI color space. Our network integrates four distinct key refinements: RAW-domain feature processing and Wavelet-based feature propagation to mitigate high-frequency detail loss; sample-based dynamic loss coefficients to ensure stable learning across varying exposure levels; and Feature Distribution Matching (FDM) Loss to maintain rigorous color constancy. Evaluations on the dataset introduced in the NTIRE 2025 challenge on NPR confirm our approach achieves competitive fidelity while establishing new state-of-the-art results in both CIE2000 color difference and LPIPS. This validates our perceptually-driven design for high-quality nighttime imaging.

Color transfer adjusts the colors of a source image to match the color characteristics of a reference image, often for artistic or visual enhancement purposes. While most existing methods operate purely on color statistics, the influence of intrinsic scene cues on perceptual quality remains underexplored. This paper investigates the impact of incorporating semantics, illumination, and geometry into color transfer algorithms. To enable this analysis, we introduce a dataset of photorealistic indoor scenes rendered with controlled variations in viewpoint, object arrangement, illumination, and color distribution. Each scene includes intrinsic scene representations capturing illumination, semantic, and geometric properties. A modular framework is proposed to integrate this information into existing methods. Subjective user studies show that intrinsic scene cues significantly influence perceived color transfer quality.

Deep learning-based Low Dynamic Range (LDR) to High Dynamic Range (HDR) conversion demands significant computational resources, limiting its use on personal devices. Although a lightweight A2C-based reinforcement learning framework was introduced, it was built upon a standard Markov Decision Process formulation and thus included a critic network. This is fundamentally misaligned with the 'Single-Step' characteristics of the HDR conversion task, introducing unnecessary complexity and potential instability. In this paper, we overcome this limitation by reframing the problem as a contextual bandit. We propose a novel Critic-Free method that replaces the unstable learned critic with an unbiased Empirical Mean Baseline (EMB) derived from multiple trials (-trials) on the same state. This approach results in theoretically more grounded and significantly more stable optimization process. Experimental results demonstrate that the proposed method achieves superior training stability and faster convergence compared to the A2C baseline, while preserving its lightweight design. Moreover, although it requires orders of magnitude fewer computational resources than state-of-the-art deep learning methods, it still attains competitive perceptual image quality and demonstrates promising generalizability across the evaluated configurations.

Full-color e-paper displays provide a paper-like viewing experience with advantages such as low power consumption and reduced eye strain. However, their limited contrast ratio and narrow color gamut compared to conventional LCDs often lead to muted colors and reduced visual quality. In this paper, we investigate image enhancement methodologies for improving color reproduction on limited-gamut e-paper displays. We focus on learnable global curve-based color transformations and evaluate the effectiveness of various curve formulations across multiple color spaces to identify the most suitable enhancement strategy. Furthermore, we compare a single global curve trained on a large-scale image dataset with individually optimized per-image curves. Experimental results show that a single global curve achieves performance comparable to per-image curves, indicating that the enhancement task can be effectively addressed using a simple global model. Additionally, the proposed approach significantly outperforms existing e-paper enhancement algorithms, yielding statistically significant improvements in visual quality.

Computational Color constancy (CC) aims to recover the true colors of a scene despite changes in illumination. While recent cross-camera deep learning methods have made impressive progress, their training relies on relatively small RAW datasets. In this work, we introduce a zero-shot color constancy framework grounded in intrinsic image decomposition (IID). Our method estimates the albedo and infers the illumination map from a pre-trained IID model, followed by chromaticity clustering. This design allows us to avoid any fine-tuning on CC datasets. Experiments on the Gehler-Shi and NUS-8 datasets show that our approach achieves performance comparable to cross-dataset SOTA performance, while maintaining robustness to camera and dataset variations. These results highlight the potential of transferring IID-learned illumination priors to achieve truly sensor-agnostic color constancy.

With the continuous development of artificial intelligence, deep learning techniques have shown strong performance in many multimedia forensic tasks, especially in image steganalysis. Nevertheless, most current deep learning-based steganalysis approaches are still formulated as binary classification problems, which aim to determine only whether an image contains hidden information. Although such methods achieve strong detection capability, they cannot estimate how much information is embedded in a stego image. To address this limitation, this paper introduces a quantitative deep learning-based steganalysis framework for embedding-rate estimation, where the network directly predicts the embedding rate of an input image. In the proposed framework, embedding rates are used as soft-label supervision signals. The framework is optimized using a Soft-label Cross Entropy (SCE) loss together with two independent L2 loss terms for cover and stego samples, respectively. Experimental results on the BOSSBase and BOWS datasets demonstrate that the proposed framework can effectively estimate embedding rates and achieve reliable quantitative steganalysis performance.

The paper deals with the recognition of color images that were degraded by blur and linear channel mixing with an unknown matrix. We propose invariants to this compound degradation in the Fourier domain by means of projection operators. We show an equivalent definition of the invariants in the image domain by means of moment expansion, which is more stable and faster when evaluated numerically. The experiment demonstrates a good performance of the invariants and their advantages over a recently proposed method of a similar nature. Possible modifications of the proposed invariants are discussed in the conclusion.

Camouflage pattern generation is a critical task designed to enhance concealment across diverse environments. Traditional approaches primarily depend on manual design or basic image processing algorithms, lacking adaptability in complex environments. Although recent studies have employed neural networks for environmental feature extraction in camouflage pattern generation, they remain inadequate not only in addressing foreground objects interference but also in generating high-quality camouflage patterns. To overcome these limitations, we propose a novel two-stage generator that systematically (1) utilizes adversarial autoencoders with partial convolution networks to extract environment-adaptive texture maps, then (2) processes these texture maps through a style transfer network to synthesize high-quality camouflage patterns. Experimental results demonstrate the outperformance of our generator in effectively generating high-quality environment-adaptive camouflage patterns.

Autonomous driving semantic segmentation has mostly been based on RGB imagery, which can be easily confused by similar materials that have a similar appearance even in adverse lighting conditions. Hyperspectral imaging (HSI) provides enhanced spectral information that can complement RGB data. However, fusing these heterogeneous modalities remains challenging due to differences in dimensionality and feature representation. In this paper, we apply the SIGMA multi-modal fusion framework to the problem of RGB-HSI semantic segmentation by adding a simple adaptation module that encodes both modalities into a common embedding space. The adapter allows stable learning of the shared backbone and retains modality-specific properties. We test the suggested methodology on several publicly available hyperspectral driving datasets and present similar gains in mean Intersection-over-Union and pixel accuracy compared to the unimodal and naive fusion baselines. The significance of the design of the adapters is confirmed by the extensive ablation research that proves that hyperspectral information is effective in material-conscious perception in autonomous driving.

Confocal microscopy is used to capture high-quality, high-resolution 3D images of biological structures. Generally, markers with distinct fluorescence emission spectra are used to label different subcellular entities, limiting the number of structures that can be imaged simultaneously. In this study, we propose a deep learning approach to separate the signals of different structures from a single 3D fluorescence channel, based on a modified version of the well-known U-Net model. The proposed architecture features a regression dual decoder and a custom penalty term in the loss function to explicitly discourage the spatial overlap between the structures in the output images. We demonstrate how this can be used to separate membranes and nuclei intensity signals in mono-channel 3D images of living plant roots. Our method reduces acquisition time and phototoxicity while freeing up channels to image additional subcellular entities.

Image vignetting, characterized by spatially varying brightness attenuation toward image boundaries, is a common degradation caused by optical limitations and camera design. This phenomenon not only reduces visual quality but also negatively impacts downstream vision tasks that rely on photometric consistency. While recent learning-based methods have improved devignetting performance, many existing approaches operate directly in the RGB color space and implicitly model illumination correction, which often entangles luminance and chrominance information and leads to color distortion, noise amplification, or over-smoothing under varying vignetting strengths. In this work, we propose a dual-branch UNet framework that explicitly separates luminance and chrominance processing in the YCbCr color space. The proposed architecture employs a Cross-Attention Enhancement (CAE) module to enable effective interaction between the luminance and chrominance branches at multiple scales, facilitating illumination correction while preserving color fidelity. In addition, a lightweight gated convolutional block is introduced for efficient feature refinement without relying on increased network capacity. Extensive experiments on the VigSet benchmark demonstrate that the proposed method consistently outperforms existing state-of-the-art approaches. Our dual-branch framework achieves significant absolute gains over the strongest baseline, improving PSNR by up to $+0.91$ dB and SSIM by up to $+0.15$ across multiple resolutions. Code is available at \Url{https://github.com/MariamHossam00/IDD-UNet}.

Recovering the past from present observations is an intriguing challenge with potential applications in forensics and scene analysis. Thermal imaging, operating in the infrared range, provides access to otherwise invisible information. Since humans are typically warmer ($\approx$ 37 °C / 98.6 °F) than their surroundings, interactions such as sitting, touching, or leaning leave residual heat traces. These fading imprints serve as passive temporal codes, allowing for the inference of recent events that exceed the capabilities of RGB cameras. We propose a time-reversed reconstruction framework that uses paired RGB and thermal images to recover scene states from a few seconds earlier. The proposed approach couples Visual–Language Models (VLMs) with a constrained diffusion process, where one VLM generates scene descriptions and another guides image reconstruction, ensuring semantic and structural consistency. The method is evaluated on a purpose-built dataset consisting of controlled human–environment interactions. Experimental results demonstrate the feasibility of reconstructing plausible RGB scene states up to 120 seconds prior to observation, establishing a first step toward practical time-reversed imaging based on thermal traces.

Sensitive images in medical, military, and IoT applications are often transmitted by resource-limited edge devices over latency- and bandwidth-constrained networks. Classical secret image sharing (SIS) is typically all-or-nothing, while progressive secret image sharing (PSIS) aims to provide usable intermediate outputs as shares accumulate. We present a secure, scalable, and lossless PSIS design whose progressiveness is \emph{certified}: after the threshold, the decoder incrementally reveals a growing set of \emph{decidable} bits that are guaranteed correct, and achieves exact reconstruction when all $n$ shares are collected. The scheme is built on a near-balanced constant-composition bit-sharing primitive (CCBS) together with threshold-protected PRG masking, so both sharing and reconstruction rely only on bitwise XOR and simple counting, with $\mathcal O(n) complexity. Experiments demonstrate fast lossless recovery for RGB images (0.05\,s per image at $n=100$) and throughput up to 53,MP/s.

Deep learning–based compressive sensing (CS) methods typically learn sampling operators using convolutional or block-wise fully connected layers, which limit receptive fields and scale poorly for high-dimensional data. We propose MTS-CSNet, a CS framework based on Multiscale Tensor Summation (MTS) factorization, a structured operator for efficient multidimensional signal processing. MTS performs mode-wise linear transformations with multiscale summation, enabling large receptive fields and effective modeling of cross-dimensional correlations. In MTS-CSNet, MTS is first used as a learnable CS operator that performs linear dimensionality reduction in tensor space, with its adjoint defining the initial back-projection, and is then applied in the reconstruction stage to directly refine this estimate. This results in a simple feed-forward architecture without iterative or proximal optimization, while remaining parameter and computation efficient. Experiments on standard CS benchmarks show that MTS-CSNet achieves state-of-the-art reconstruction performance on RGB images, with notable PSNR gains and faster inference, even compared to recent diffusion-based CS methods, while using a significantly more compact feed-forward architecture.

Blind reconstruction of low-dose computed tomography scans has become an important but challenging medical image analysis method due to spatially correlated noise and scanner-specific artifacts. They often lead to structural distortion during denoising. Existing self-supervised approaches achieve strong performance under relatively simple or well-characterized noise conditions but frequently struggle to maintain structural stability when confronted with unknown noise characteristics. In this work, we propose a blind reconstruction framework that integrates learned non-local filtering with attention-driven latent similarity modeling under structural consistency objectives. The method exploits spatial redundancy to adapt to structured noise without requiring paired supervision with a decoupled learning strategy that separates structural representation learning from noise adaptation. Extensive experiments on multiple CT datasets demonstrate consistent noise suppression while preserving structural integrity, yielding approximately 3-5 dB improvements in PSNR over representative self-supervised methods alongside favorable SSIM performance

Mahjong tile recognition is challenging when tiles differ only in character color. This paper proposes a four-channel RGB-a* input extension for YOLO11m by appending the a* channel from the L*a*b* color space to conventional RGB input. The additional chromaticity information improves discrimination between red and black tile characters. Experiments demonstrate a 13.3% improvement in mAP50 over the RGB baseline with negligible computational overhead.

Metalenses are increasingly used in compact optical devices, such as AR/VR and quantum computing systems. This work introduces a complete design simulation pipeline for a metasurface system, from initially setting target phase profiles to the subsequent stage of choosing meta-atom material parameters and shape optimization, and finally characterizing optical performance. We demonstrate the functionality of our developed simulation platform on an RGB metalens system design.

Near-infrared (NIR) imaging provides robust structural and texture information under challenging environmental conditions such as low illumination, haze, and reduced visibility. However, NIR images are typically single-channel and lack natural color information, limiting their interpretability for human observers and computer vision systems. In this study, we propose a multi-stage deep learning framework for NIR-to-RGB image reconstruction that aims to preserve structural details while generating visually realistic RGB images. The proposed architecture consists of a Content Network for luminance extraction, a Color Generator for chrominance prediction in the YCbCr color space, a Guided YCbCr Fusion module, and a multi-scale PatchGAN discriminator for adversarial learning. Experimental results demonstrate that the proposed framework successfully reconstructs structurally consistent RGB images while improving perceptual image quality. Quantitative evaluations achieved a PSNR value of 16.199 dB and an SSIM score of 0.6386 on paired RGB-NIR datasets. The proposed approach has potential applications in autonomous driving, nighttime surveillance, remote sensing, and intelligent vision systems.

Diffusion models have recently achieved remarkable success in image restoration, particularly in image super-resolution. However, blind image denoising remains much less explored, despite being a fundamental problem in low-level vision. In this work, we propose an adaptive noise and timestep diffusion model for blind image denoising, named ANTD, which augments a pre-trained diffusion backbone with an auxiliary adaptive noise and timestep branch, referred to as ANT. Specifically, given a noisy input image, the ANT branch jointly predicts a noise compensation map and an appropriate diffusion timestep. The diffusion model then performs one-step denoising using the compensated noisy image and the predicted timestep. By training the framework end-to-end with a fixed diffusion backbone, ANTD inherits the rich generative prior of the pre-trained diffusion model and adapts it to blind image denoising. Extensive experiments on synthetic benchmarks demonstrate that ANTD achieves higher PSNR/SSIM and substantially improves perceptual quality, while also generalizing well to real-world image denoising.

Video colorization aims to generate semantically appropriate and temporally consistent colors for grayscale videos. Existing diffusion-based approaches often face challenges in high training resources or poor cross-frame consistency due to their reliance on stochastic denoising processes. To address these limitations, this paper proposes an optical flow-guided diffusion framework that explicitly utilizes inter-frame motion to enhance temporal coherence. We explicitly inject optical flow into the network during training and introduce an optical flow-guided cross-frame attention mechanism, which leverages precise optical flow correspondences to modulate attention between frames during sampling. This ensures stable and artifact-free color propagation under complex motion patterns. Moreover, our framework incorporates reference images as additional guidance to improve controllability and color fidelity. To reduce dependence on large-scale real-world video data, we design a hybrid data synthesis pipeline that first pre-trains the model using synthesized optical flow-image pairs and then fine-tunes it on a limited set of high-quality videos generated by modern video synthesis models. Experiments demonstrate that our method achieves superior performance in terms of training efficiency, temporal consistency, color realism, and user controllability compared to existing state-of-the-art approaches.

While conventional adversarial examples (AEs) rely on adding imperceptible perturbations to existing images, Natural Adversarial Examples (NAEs) leverage generative models to synthesize adversarial images that appear more realistic and transfer more effectively to black-box models. In practice, however, existing NAE generation methods suffer from severe instability: generated samples range from high-fidelity adversarial images to semantically corrupted outputs that diverge from their ground-truth labels, inflating reported attack success rates. We show that this instability can be effectively addressed through controlled perturbation injection during the generative process. We propose \textbf{Stable-NAE}, a lightweight framework that employs an adaptive control strategy and momentum-based adversarial guidance to stabilize NAE generation. When evaluated across multiple benchmarks including VQA and CLIP scores, and attack success rate, Stable-NAE surpasses existing methods in visual semantic fidelity while achieving over 99\% white-box attack success rate, all while requiring 30\% less time than the fastest prior NAE generation method.

Real-world image super-resolution is a challenging image generation task which aims to achieve high fidelity and perceptual quality. Diffusion-based approach is currently one of the most promising techniques, yet the computational complexity remains high due to the fundamental large U-Net backbones, which hinders practical deployment. Recent research on neural network compression showed that uniform channel pruning is an effective method to control a trade-off between efficiency and image quality. However, existing works did not systematically evaluate the significance of each layer. Moreover, the resolution-dependent redundancy inherent in the U-Net architecture is often being neglected. This paper presents the first study to systematically quantify resolution-stage pruning sensitivity of a Stable-Diffusion-based SR U-Net under real-world image SR. We develop the proposed method based on a recent state-of-the-art, AdcSR. We systematically evaluate the sensitivity of each resolution stage of the denoising U-Net and develop a stage-wise pruning strategy that decouples channel reduction ratios while preserving encoder–decoder symmetry. Our experiments conducted using publicly available datasets, DRealSR, RealSR, and DIV2K_Val suggest that the deep bottleneck stage contains significant redundancy and can be pruned aggressively. High-resolution stages are more critical for maintaining perceptual details. Our proposed method achieves a reduction of the number of parameters by 34% and a reduction of computational time of 10.3% in student U-Net latency evaluated on a Tesla V100, while maintaining comparable visual quality. Codes and models are available at: https://github.com/khangnguyen-sci/non-uniform-pruning

This paper presents VLM-Dreamer, a novel three-stage pipeline for generating an enriched and consistent 360-degree 3D scene from a single image. VLM-Dreamer addresses the limitations of existing approaches by integrating a Vision- Language Model (VLM) to overcome the constraints of fixed text prompts. By leveraging the VLM’s reasoning capa- bilities, we can imagine plausible surrounding objects that naturally co-occur with the input, allowing us to generate semantically rich and spatially consistent scene descriptions. Specifically, VLM-Dreamer first synthesizes a consistent 360-degree scene image. It then uses an inpainting diffusion model, guided by the VLM-generated descriptions, to pro- duce a coarse 3D Gaussian Splatting representation. At the final stage, it introduces bi-directional inpainting to refine the scene, enhancing cross-view alignment and semantic coher- ence. Through extensive experiments, we demonstrate that VLM-Dreamer achieves superior performance across multi- ple metrics, effectively balancing creative imagination with spatial fidelity to produce diverse and high-quality 3D scenes.  The source code is available at https://github.com/Lalalalex/VLMDreamer.git.

The perception-distortion trade-off poses a fundamental challenge in single-image super-resolution (SR). Although diffusion-based SR methods excel at generating perceptually realistic images, achieving high fidelity remains a key limitation. Recent advances in diffusion-based SR have shown promise in improving fidelity, but these methods often compromise perceptual quality due to their high reliance on a high-fidelity image. To address this, we introduce UGDiff, a novel diffusion guidance paradigm designed to further improve the perception-distortion balance. In particular, we first estimate the reconstruction uncertainty of the latent features corresponding to a high-fidelity image. This uncertainty is then used to guide the diffusion process to selectively restore high-frequency details in high-uncertainty regions, while preserving fidelity elsewhere. Furthermore, our guidance method adaptively identifies the high-uncertainty regions by considering not only the estimated uncertainty but also the posterior variance of the diffusion sampler at each timestep. This relaxes the reliance on the high-fidelity image in the later stages of sampling, thereby achieving a better perception-distortion balance. Extensive experimental results demonstrate that our method performs favorably against state-of-the-art diffusion-based SR methods.

Deep neural networks (DNNs) have become the dominant paradigm for synthetic aperture radar (SAR) target recognition, but their vulnerability to adversarial examples poses significant security risks. However, most existing adversarial attack methods generate pixel-level perturbations under $L_p$ constraints, causing the generated adversarial examples to violate physical consistency and become easily detectable by distribution-based defenses. To this end, a physics-constrained diffusion attack framework against SAR target recognition is proposed, which initially integrates adversarial perturbation generation into the reverse denoising process of the diffusion model. Specifically, an attack-guided reverse denoising mechanism is designed to progressively drive samples away from the correct decision boundary. Meanwhile, physical constraints are imposed to preserve both global intensity distributions and local scattering center structures. By jointly optimizing the diffusion prior, attack guidance, and physical constraints, the proposed method achieves both effective misclassification and high physical realism. Experimental results demonstrate that the generated adversarial SAR samples can mislead recognition networks while maintaining statistical and scattering consistency.

abstract: Diffusion models have achieved remarkable progress in text-to-image generation, but suffer from memorization, where training data is reproduced either entirely (i.e., global memorization) or partially (i.e., local memorization). In this paper, we highlight the different influence of prompts on these two types of memorization and argue that this needs to be considered in mitigating them. Motivated by this, we propose the first framework to classify local versus global memorization. Then, we show that simple parameter adjustments in mitigation based on the classification results improve mitigation performance. This work provides a practical step toward effective mitigation of memorization in diffusion models without requiring access to training data.

With the significant advances in AI-generated contents, watermarking has aroused great attention to address the copyright and security concerns. Recently, in-generation watermarking via diffusion models has demonstrated a great success as it enables an efficient pipeline of watermark embedding and image generation. However, they often struggle to balance the competing goals of a watermarking system, i.e., robustness and imperceptibility, since the watermark has to survive the diffusion or its reverse denoising process. In addition to this, the embedding strength in current schemes is universally set for different carrier images. To address these challenges, we propose a two-stage training framework for ingeneration watermarking via rectified flow diffusion models. First, the watermark encoder and decoder are jointly trained. Then the parameters of the watermark encoder is frozen while the watermark decoder and latent adapter are fine-tuned to adjust the fusion of latent code and feature representation of the watermark. Experimental results demonstrate that our method is able to adaptively balance imperceptibility and robustness with the aid of the adapter. Compared to the state-of-the-art methods, the proposed method enables greater capacity while maintaining superior fidelity of the watermarked images. Besides, it also exhibits stable robustness against various types of attacks.

In the last decade, neural network (NN) models have been successfully utilized in Artificial Intelligence (AI) systems for various applications. One weakness of these models is that they tend to memorize the training data, leading to the potential privacy concerns. Various techniques address these concerns by intentionally perturbing the data and system characteristics. However, they can reduce the utility, which makes the data harder to use for machine learning (ML) tasks. In this paper, we introduce a framework that estimates the optimal privacy from data to improve the utility for NN models in ML tasks. All models employed in our framework are differentiable, allowing us to optimize the privacy-utility trade-off in an end-to-end model training scheme. The experimental results demonstrate that our framework significantly improves the utility for image classification tasks while ensuring similar or stronger privacy protection against data reconstruction and data leakage attacks compared to the state-of-the-art baseline.

We present Zero‑Shot Memorability Control for Text‑to‑Image Diffusion (ZSMC), a plug‑and‑play framework that endows existing latent diffusion models with the ability to generate intrinsically more memorable images without any task‑specific fine‑tuning. Central to ZSMC is a Memorable Mapping Network (MMN) that transforms the [EOS] token in CLIP text embeddings into a memorability‑aware token embedding. At inference time, this single token injection conditions downstream diffusion models on abstract cues—such as object size, shape, and color—while preserving the semantic and spatial fidelity specified by the prompt. Extensive experiments on the MemCat and LaMem benchmarks show that ZSMC consistently attains state‑of‑the‑art memorability scores (IMS‑V/R) across Stable Diffusion, ControlNet, and Instance Diffusion backbones, without compromising diversity or CLIP alignment. A 32‑participant user study further confirms that ZSMC images are recalled around 20 \% more frequently than those from competitive baselines. Because ZSMC operates entirely at the token level, it can be dropped into any off‑the‑shelf text‑to‑image generator in a true zero‑shot fashion, making memorability control both practical and widely accessible.

The rapid rise of AI-generated images (AIGIs) from text-to-image (T2I) models calls for effective quality assessment. Traditional IQA methods lack semantic alignment with text prompts, while existing AIGIQA approaches, especially CLIP-based ones, overemphasize semantics and ignore perceptual quality, relying on fixed prompts and static scoring. To bridge these gaps, we propose a dual-attribute prompt learning framework with progressive tuning based on CLIP that jointly models semantic consistency and perceptual quality. Our method employs dual-attribute learnable prompts and a dynamic patch-level regression head, trained through a three-stage strategy for improved stability and generalization. Experiments on multiple AIGIQA datasets show that our approach outperforms or matches state-of-the-art methods. The code of the proposed method is available at: https://github.com/Handvery/PPTCLIP.

Recently advanced image generation foundation models, such as Stable Diffusion, sometimes have an overcapacity and are unsuitable for user-specific tasks. Hence, there is still a demand for training one’s own image generation model from scratch that only generates images of designated concepts, with only a limited number of training samples. Nevertheless, no existing work has proposed a few-shot learning method for Latent Diffusion Models (LDMs) working with around a hundred training images. In this paper, we propose a few-shot learning method for unconditional LDMs based on a domain adaptation technique. The proposed method leverages not only a few target domain images but also lots of source domain images. Based on the assumption that the source and target domains can share a domain-independent latent space defined by a single VAE, we first train such a VAE. Then, we train a U-Net in the shared latent space, and finally fine-tune only the VAE decoder with the target domain images using an adversarial training strategy. Our experimental results demonstrated that the proposed method successfully trains an LDM with around 100 training images.

Positron Emission Tomography (PET) imaging entails an inherent trade-off between image quality and radiation exposure. While reducing tracer dosage mitigates patient health risks, it inevitably degrades image resolution. To address this ill-posed inverse problem, we propose a novel coarse-to-fine cascaded framework leveraging a medically pre-trained PMC-CLIP image encoder. Instead of relying on complex iterative sampling, our method effectively leverages medical visual priors from PMC-CLIP to guide a deterministic CNN decoder. This design not only surpasses state-of-the-art (SOTA) methods in reconstruction fidelity but also significantly reduces inference latency and eliminates the stochastic instability associated with diffusion models. Extensive experiments validate the superior performance of the proposed method in recovering high-quality PET images. Code is available at https://github.com/nahida444/CLIP-PET}.

Neural watermarking has been on the rise as a simple tool for marking generated multimedia content in a robust way. This simplicity, however, comes at a cost. While it is easy to enforce robustness in a training loss through various data augmentations, it is currently unknown how to include the security aspect of watermarking. Consequently, these black-box schemes are easily breakable by targeted attacks. In this work, we show how to remove a watermark from a recently proposed WAM image watermarking model. Consistent with Kerckhoffs's principle, we show that the absence of a secret key, combined with the requirement for robustness, creates an exploitable weakness. This manifests as near-periodic patterns in one principal color component of the image. We show that these patterns can be accurately estimated even from a single image, and erased by manipulating a single component of the image in its Fourier representation. Furthermore, we demonstrate that the same watermark can be manually injected, yielding noticeably higher image fidelity than when using WAM for watermarking. The code used in this work will be made available upon acceptance of the paper.

This paper investigates the detectability of image-in-image steganography schemes. In this paradigm, the payload is usually an image of the same size as the cover image, leading to very high embedding rates. We first show that the embedding yields a mixing process that is easily identifiable by independent component analysis. We then propose a simple, interpretable steganalysis method based on the first four moments of the independent components estimated from the wavelet decomposition of the images, which are used to distinguish between the distributions of cover and stego components. Experimental results demonstrate the efficiency of the proposed method, with eight-dimensional input vectors attaining up to $83\%$ accuracy. This vulnerability analysis is supported by two other facts: the use of keyless extraction networks and the high detectability w.r.t. classical steganalysis methods, such as the SRM combined with support vector machines, which attains over $99\%$ accuracy. Code is available at https://github.com/malletan/Detecting-image-in-image-steganography.

This paper presents a watermarking assessment in the context of the emerging JPEG Trust Part 3: Media Asset Watermarking (ISO/IEC 21617-3). The specification complements the JPEG Trust family of standards by enabling a robust content–metadata binding through watermarking, while remaining algorithm-agnostic via explicit signalling of watermark presence, identification of the embedding/extraction technique, and references to external repositories that store Trust Records. It further describes a recommended payload structure and a benchmarking methodology covering imperceptibility and robustness, including modern attacks such as advanced compression (e.g., JPEG XL, JPEG AI) and AI-driven object manipulation. Four state-of-the-art methods from industry and academia (InvisMark, Watermark Anything, TrustMark, and VINE-R) are then evaluated on COCO dataset, highlighting strengths on classical perturbations and persistent vulnerabilities to emerging AI-based attacks, underscoring the need for standard and effective evaluation.

Image super-resolution (SR) demands both local detail restoration and long-range contextual modeling. While window-based Transformers such as SwinIR improve efficiency, their limited receptive fields restrict global dependency capture. We propose Window-based Linear Attention (WLA), which applies linear attention within local windows for stable and efficient spatial modeling. Building on WLA, we introduce Efficient Multi-Region Attention (EMRA), which reuses window-level KV buffers to perform unified attention over local, stripe, and global regions with learnable fusion. This design avoids multi-branch overhead while preserving linear complexity. Extensive experiments on standard benchmarks demonstrate that WLA+EMRA achieves strong SR performance with substantially reduced computational cost.

Rain streaks inherently exhibit spatial heterogeneity. Although recent hybrid CNN–Transformer single-image deraining(SID) methods benefit from complementary local and global modeling to recover clear images from the degraded ones. However, these methods consistently suffer from module redundancy and suboptimal feature fusion, due to the co-designed CNN–Transformer architecture and the resulting misalignment in learned representations. In this paper, we propose a Multi-Scale Large-Kernel Attention (MLA) module and a Ghost Kernel Gate (GKG) to jointly model long-range dependencies and local details within a unified convolutional space. Specifically, MLA captures rich multi-scale contextual features, while GKG enables local features selection. The unified-space multi-scale feature extraction and fusion strategy effectively handles spatially-varying rain distribution and reduces model parameters. Extensive experiments demonstrate that our method achieves state-of-the-art performance, yielding consistent improvements in both quantitative metrics and visual quality.

Single image dehazing remains a challenging problem due to spatially varying atmospheric scattering and the degradation of structural details. While recent deep learning methods have achieved notable progress, most operate purely in the spatial domain and often struggle to preserve global geometric consistency under real-world haze. In this paper, we propose a frequency-aware dual-branch framework for image dehazing that explicitly leverages spectral representations. Our key contribution is the Frequency-Adaptive Depth-Haze Consensus (FADHC) block, which operates in the Frequency domain to synchronize phase information between depth and dehazing branches, enforcing geometry-consistent structure recovery while adaptively restoring high-frequency components suppressed by haze. In addition, a CLIP-based semantic encoder is incorporated to provide robust semantic priors under dense and non-homogeneous haze. Experiments demonstrate that the proposed method achieves strong performance on both synthetic and real-world benchmarks, particularly on challenging datasets such as O-Haze and Dense-Haze. Project Code is available on https://github.com/AhmedSakr54/FADHC

In underwater imaging, light scattering and absorption easily degrade image quality, resulting in color distortion, low contrast, and blurred details. Most existing underwater image enhancement (UIE) methods focus on mining spatial domain features, but either ignore the frequency domain entirely or overlook the distinct roles of different frequency bands. To exploit the strengths of both domains simultaneously and activate the functions of all frequency bands, we propose a multi-scale cross-domain fusion network (MSCDF) based on a U-net-style encoder-decoder architecture. The proposed MSCDF comprises three scale levels, and implements bidirectional spatial-frequency interaction via the elaborately designed spatial-frequency fusion block (SFF) at each level. Notably, a multi-band decomposition unit (MBD) is embedded into SFF block prior to spatial-frequency interaction to help capture both global illumination patterns and local structural details. Extensive experiments on both synthetic and real-world underwater datasets demonstrate that MSCDF outperforms existing state-of-the-art methods.

To address the limitations of task decoupling and insufficient utilization of multi-scale features in infrared and visible light image fusion, this paper proposes a task-driven frequency-aware fusion network. Within this framework, we construct a frequency-aware feature extraction module that accurately captures the thermal target radiation features of infrared images and the high-frequency texture details of visible light images by performing multi-band decomposition and information integration on the source images. We introduce a dilated spatial convolution pooling structure, utilizing parallel multi-scale receptive fields to capture extensive contextual spatial dependencies, thereby enhancing the robustness of feature representation. Furthermore, we design a joint module driven by the detection task, using semantic gradient feedback generated by downstream detection branches to guide the fusion network, ensuring that the generated fused image maintains excellent visual quality while significantly improving semantic awareness. Qualitative and quantitative analyses on various datasets demonstrate that our proposed method outperforms state-of-the-art methods.

Electrical substation infrastructure plays a paramount role in the power grid, is particularly susceptible to thermally-induced faults, making infrared thermography an ideal method for monitoring its operational conditions. Deploying lightweight, automated perception models on explosion-proof smartphones can mitigate the risks for workers operating in close proximity to apparatus. However, existing infrared detection models often lack robustness in dynamic and complex real-world scenarios, and three-phase electrical equipment is prone to occlusion. To address this, we have designed a denoising and perception enhancement module to counteract infrared image degradation, along with an occlusion-aware detection head. For the enhancement stage, we designed a Parallel mixed frequency-space Domain Gating (PDG) module, inspired by vertebrate visual perception. It utilizes interacting dual-domain information from horizontal and vertical orientations to achieve infrared image denoising and simulates the optic chiasm to enhance the model's perceptual and inference capabilities. Furthermore, we adopted a parallel structure to design the Occlusion-aware Head (OHead) , which can reason from local information of the occluded object in conjunction with global information. We conducted comprehensive experiments on a real-world substation and standard infrared datasets. The experimental results fully demonstrate the effectiveness of our method. The source code is available at https://github.com/4everWZ/PO-YOLO.

Vision Transformers (ViTs) performance degrades on imbalanced, low-resolution data, where the loss of fine detail and minority-class distributions limits their effectiveness. We propose Hierarchical Patch and Encoding Representation Vision Transformer (HiPerViT), a novel hybrid CNN-Transformer architecture designed to overcome these challenges. HiPerViT combines the strengths of CNNs for hierarchical local feature extraction and Transformers for global context modelling, enhanced by a multi-scale patch-embedding and fusion strategy that preserves fine-grained lesion details while maintaining efficiency. We introduce SkewMod, an imbalance-aware modulation module that adaptively amplifies minority-class features during token interactions. To validate HiPerViT, we conduct experiments on five skin lesion datasets with varying resolutions and class imbalance ratios, as high as 1:10 for melanoma versus non-melanoma. HiPerViT outperforms existing ViT-based and CNN-Transformer hybrid models, achieving the highest F1 and AUC scores on minority classes. It opens new avenues for enhancing feature extraction and imbalance-aware modulation in hybrid CNN-Transformer architectures, enabling more robust performance on imbalanced and low-resolution datasets.

Underwater structured-light systems project multiple laser lines to compensate for texture scarcity and illumination degradation, where reliable 3D recovery relies on continuous laser line detection. However, treating this task as conventional pixel-wise segmentation with CNNs can lead to fractures and fragmentation under scattering noise, reflections, and local occlusions, undermining downstream triangulation and data association. We propose SSM-UNet, a structure-aware laser line detector that couples local convolutional enhancement with global state-space modeling. A Mamba-based bridge captures long-range dependencies, while a cross-scale guided SSM module leverages shallow structural guidance to complete weak or missing line segments. To match the constrained X-shaped topology of cross-line lasers, we further introduce DBSS, which serializes 2D features along the two diagonals and performs bidirectional scans to mitigate causal bias, followed by routed fusion for two orthogonal laser branches. Experiments on real underwater datasets show improved line integrity and connectivity over representative CNN-based, SSM-based, and thin-structure baselines, leading to more stable and accurate structured-light reconstruction.

LiDAR super-resolution enables high-quality 3D perception from cost-effective, low-resolution sensors. While transformer-based approaches like TULIP show promise, they remain limited to spatial-domain processing with restricted receptive fields. We introduce FLASH, a novel framework featuring two key innovations: (i) Frequency-Aware Window Attention combining local spatial attention with global frequency-domain analysis via Fast Fourier Transform, and (ii) Adaptive Multi-Scale Fusion replacing conventional skip connections with learned position-specific feature aggregation enhanced by Convolutional Block Attention Module. Experiments on KITTI demonstrate state-of-the-art performance across all metrics, outperforming even TULIP with Monte Carlo Dropout while maintaining single-pass efficiency at 66 FPS. Our dual-domain approach effectively handles uncertainty through architectural design, making it practical for real-time autonomous systems.

Underwater images suffer from severe wavelength-dependent light absorption and scattering, and turbidity due to suspended particles, degrading visual quality for applications in autonomous underwater vehicles (AUVs), marine biology, archaeology, and offshore infrastructure inspection. Classical IFM inadequately capture nonlinear underwater light behavior, while purely data-driven methods lack physical interpretability. This paper proposes a three-stage network named ADR, that extends the underwater image formation model with additional terms to perform underwater dehazing, followed by Retinex-based enhancement and attention-enabled U-Net++ refinement. Experiments on UIEB and UFO-120 benchmark datasets demonstrate competitive performance with state-of-the-art methods. The code should be available upon publication at https://github.com/SahanaRay/ADR.

Single-image dehazing is commonly modeled by the atmospheric scattering model, where accurate recovery depends critically on the transmission map and atmospheric light. However, estimating these terms from a single hazy image is ill-posed and becomes unstable under dense haze, often causing color shift, over-enhancement, or residual veil. We present a physics-guided dehazing framework that explicitly separates physical parameter estimation from image refinement. A lightweight TauNet predicts a non-negative haze-thickness map, which is mapped to transmission via a learnable exponential attenuation, while a compact atmospheric-light estimator predicts global atmospheric light conditioned on the hazy image and the estimated thickness. These estimates yield a closed-form physics reconstruction that serves as a reliable baseline. To suppress inversion artifacts and recover fine details, we employ a compact refiner that predicts a residual correction and blends it with the physics reconstruction through a learnable gate, preventing over-correction in low-transmission regions. With only 2.43M parameters, the proposed method achieves the best PSNR among compared methods on SateHaze1k across thin, moderate, and thick haze, maintains competitive SSIM, and shows consistent gains on RSID and NID.

Reconstructing complex scenes in participating media, such as foggy or underwater environments, remains challenging due to severe light scattering and absorption. Recent hybrid representations that combine explicit surface modeling and implicit volumetric rendering have shown promising results, yet they still suffer from two critical issues: multi-view inconsistency caused by view-dependent scattering, and frequency-domain representation conflicts arising from the uncontrolled coupling of high- and low-frequency signals. In this paper, we address these problems from a signal-processing perspective and propose a unified framework for multi-view consistency modeling and frequency-aware representation decoupling in scattering scenes. First, we introduce a multi-view scattering consistency model that jointly enforces color and geometric coherence across viewpoints. Specifically, we extend haze-lines from single-image enhancement to a multi-view color subspace alignment mechanism, and incorporate a Microflake-based directional scattering prior to constrain view-invariant volumetric scattering behavior. We further introduce a frequency-aware representation decoupling strategy that explicitly separates low-frequency volumetric scattering from high-frequency surface geometry, mitigating frequency aliasing during joint optimization.Extensive experiments on synthetic and real-world datasets show consistent improvements in reconstruction quality and stability, achieving up to 1.5 dB PSNR, higher SSIM, lower LPIPS, and reduced cross-view color and depth variance.

With the rapid development of aerospace technology, the number of space debris in Earth's orbit has been continuously increasing, posing a severe threat to the safe operation of on-orbit spacecraft. Consequently, Space Situational Awareness (SSA) has become increasingly crucial. Optical imaging systems are widely used in space target detection due to multiple advantages. However, constrained by limited spatial resolution and extremely long observation distances, space debris appear as "point-like" targets highly similar to stars. Moreover, the extremely low luminous flux of these targets makes them easily submerged in noise under low signal-to-noise ratio (SNR) conditions, rendering real-time and robust extraction a bottleneck in the SSA field. To address these challenges, this paper proposes a temporal consensus feature fusion network, termed temporal consensus UNet (TC-UNet), for extracting faint small space targets. Specifically, TC-UNet adopts UNet as the backbone model architecture and embeds the core Temporal Consensus Feature Fusion (TCFF) module, which can effectively aggregate and enhance the temporal features of faint targets while suppressing background noise interference.The performance of the proposed algorithm is evaluated using simulated and real-shot star maps. The results demonstrate that when detecting with consecutive 4-frame images, the limit detection magnitude of the algorithm can reach 10.00 magnitudes (Mv), representing an improvement of 0.60 Mv compared with the optimal result of single-frame detection.

Deep Unfolding Networks (DUNs) have emerged as an advanced framework for image Compressive Sensing (CS) reconstruction tasks. However, most existing methods perform reconstruction predominantly in the spatial domain, which limits their ability to exploit frequency-domain structure-detail priors. To address this limitation, this paper proposes a Bidirectional Wavelet Cross-Attention Unfolding Network (BWCA-Net). The core of BWCA-Net is a Discrete Wavelet Transform (DWT)-based bidirectional cross-attention prior. Specifically, DWT decouples features into low-frequency components, which carry coarse structures, and high-frequency components, which contain edges and textures. BWCA then performs a two-stage cross-frequency interaction: low-frequency features first guide high-frequency restoration, and the refined high-frequency features subsequently feed back detail cues to enhance the low-frequency representation. The experimental results demonstrate that BWCA-Net achieves competitive reconstruction accuracy and visual quality, reaching 30.61 dB in PSNR and 0.9052 in SSIM on the Set11 dataset at CS ratio 0.1, 0.46 dB above the second-best method.

This paper proposes Laplacian Pyramid Gated Fusion Network (LP-GFNet), a model that decomposes features into low- and high-frequency components and integrates high- frequency components using a channel-wise High-to-Low gate. On CIFAR-100, LP-GFNet achieves 71.35% accuracy with 16% fewer parameters compared with DAWN.

Industrial X-ray CT images suffer from mixed Poisson-Gaussian noise from photon starvation and electronic readout, which adversely affects the X-ray visual inspection process. We propose a physics-aware pipeline with a Realistic Noise Simulation (RNS) module that explicitly models the photon-counting and sensor readout statistics to synthesize high-fidelity training pairs. A DnCNN baseline validated on synthetic and real industrial PCB CT data achieves 28.3 dB PSNR, providing a foundation for future Transformer-based X-ray image restoration.

This work extends context-based multi-sensor fusion (contextual fusion) for 3D object detection in adverse operating conditions. Contextual fusion combines LiDAR point cloud and RGB image information while considering domain knowledge about the sensors behaving differently across lighting and weather variations. However, contextual fusion is limited to channel-level gating and cannot consider spatially local variations that are of particular relevance in night and rainy conditions. Therefore, we propose a fine-grained entropy-enhanced contextual fusion network that can adapt to spatially local variations. Currently, entropy-based contextual fusion can provide a relative increase of almost 10% of the average precision (AP) for hard rainy conditions when compared to non-contextual fusion.

We present a commercial-ready AR head-mounted device (HMD) built on PetaRay's patented LiFiD™ light field technology, which reconstructs physical light rays to generate the visual experience of arbitrary focal depth. By matching the focal depth of digital and physical elements, LiFiD™ resolves the vergence-accommodation conflict (VAC), reducing eye strain and enhancing visual comfort during extended use. In the demo, attendees wear the headset and scan a printed QR code to activate an interactive scene. Users can experience accurate spatial alignment, continuous focus, and realistic depth perception through the HMD. Beyond comfortable everyday AR, the same LiFiD™ core also points toward medical and surgical applications, including 3D stomach reconstruction, physician gaze-pattern visualization, and AR-guided minimally invasive intervention. This demo highlights a practical path toward clinically valuable surgical guidance and training tools.

We have developed a fully functional client–server prototype that implements the proposed framework in an interactive and accessible form. The system is designed to support the practical analysis of manipulated Earth Observation imagery and to illustrate the behavior of representation-dependent forensic models in a real-world scenario. The demo provides a web-based interface through which users can upload Sentinel-2 RGB images. The backend processes the input using our ensemble of specialized detectors, including frequency-domain (FFT)-based models, wavelet-based representations, and subspace learning approaches (Geo-DefakeHop variants), combined through a representation-aware hierarchical decision mechanism. For each input image, the system returns: * An estimated probability of authenticity (real vs. manipulated), * and a localization heatmap highlighting regions potentially affected by manipulation. The framework is implemented as a client–server architecture (available at: https://github.com/hilalyldz/Fake_Remote_Sensing_Image_Detection) and is designed to be lightweight, reproducible, and easily extensible to other remote sensing datasets and manipulation types.

We demonstrate a compact reflective lensless microscopy setup for through-silicon imaging. Illuminated by a laser source, it retrieves amplitude and phase information using a single-shot iterative phase retrieval algorithm without lenses or moving components. This approach is relevant to semiconductor manufacturing, advanced packaging, and silicon photonics, where through-silicon inspection is required for alignment and quality control. Experimental validation was performed using resolution targets, photonic integrated circuits, and stacked samples. The reconstructed phase images demonstrated substantially higher contrast and improved visibility compared with amplitude images, enabling observation of waveguide features, metallization patterns, and alignment marks through silicon substrates. Modulation transfer function analysis confirmed superior contrast preservation of the phase channel compared with conventional near-infrared through-silicon microscopy. The recovered complex optical field further enabled numerical propagation and computational refocusing. Experiments with stacked chips demonstrated depth-selective imaging and quantitative axial discrimination of structures located at different planes within a multilayer assembly. The results indicate that reflective lensless through-silicon imaging provides a compact platform for semiconductor metrology, combining phase-enhanced inspection and computational depth analysis within a single hardware architecture.

This Show & Tell demo presents a practical framework for combining datasets acquired sequentially with two independent optical microscopy systems: a confocal microscope and a Fourier light field microscope (FLFM). The demo focuses on visualizing the complete registration workflow and illustrating the challenges associated with aligning datasets that differ in field of view, resolution, and imaging geometry. Attendees will explore representative confocal and FLFM data, including high-resolution confocal volumes, reconstructed FLFM volumes, and intermediate projection images used during registration. The demonstration showcases the coarse-to-fine alignment strategy, where a larger confocal field of view serves as the spatial reference and the corresponding FLFM region is automatically localized through correlation-based template matching. Visualizations shown during the demo will highlight key processing stages, including FLFM volume reconstruction, obtaining projection images, correlation map computation, region extraction from the confocal volume, and final alignment results. Examples of successful registrations and challenging cases will be presented to illustrate the robustness and limitations of the approach. The demo provides insight into practical considerations for multimodal microscopy workflows and demonstrates how accurate correspondence can be established between independently acquired confocal and light field data without requiring hardware-level co-registration.

Ptychography is a computational imaging technique that reconstructs the amplitude and phase information of a specimen from its diffraction patterns. Conventional ptychography typically records a series of diffraction patterns from overlapping sample regions and relies on the Fourier transform to model the propagation from the object plane to the detector plane. However, this approach struggles with ill-posedness under low-overlap sampling in the far-field regime. Moreover, in near-field regimes, it faces inherent limitations in numerical aperture (NA) and propagation distance. To address these challenges, we propose a novel imaging modality, Fractional Fourier Near-field Ptychography (FrF-NFP), designed to relax the constraints on both high overlap ratio and low NA. Specifically, we formulate a ptychographic forward model based on the fractional Fourier transform and propose the Fractional Fourier Ptychographic Iterative Engine (FrPIE) for robust image reconstruction. Comparative simulations demonstrate that the proposed method achieves superior reconstruction quality compared to the conventional ptychography under low-overlap sampling conditions. Furthermore, it extends beyond the capabilities of the angular spectrum method by enabling a larger effective NA and a substantially broader range of valid propagation distances.

Gravimetry inversion aims to recover 3D subsurface density-contrast models from surface gravity anomalies, which are fundamental to resource exploration. These measurements are typically sparse and irregularly distributed relative to the 3D model's dimensionality. Classical inversion approaches stabilize the solution through regularization but depend strongly on the assumed prior and careful hyperparameter tuning and tend to degrade under extreme sparsity; moreover, many alternative pipelines rely on conventional interpolation schemes, such as kriging or RBF-based methods, or directly assume samples on a grid can represent the field without explicitly enforcing physical validity. When a pre-interpolation step is performed, the resulting field is rarely constrained to be consistent with the gravity forward operator, which can introduce artifacts in unobserved regions that propagate to the subsequent inversion. To address these limitations, this work introduces a physics-guided closed-loop that (i) replaces the conventional two-stage pipeline with a \emph{joint interpolation-inversion} optimization, where a continuous gravity field and the 3D density model are co-estimated under the differentiable forward physics and (ii) introduces \emph{auxiliary coordinates} to enforce physics consistency beyond the unobserved samples, effectively constraining the reconstructed field to remain within the range of physically plausible forward responses and mitigating non-physical solutions in extremely sparse regimes. The proposed method reduces reconstruction error by 38\%–67\% compared to a neural baseline and by about 95\% compared to a classical inversion approach, yielding more coherent reconstructions with fewer artifacts under undersampling.

Remote photoplethysmography (rPPG) enables non-contact measurement of cardiac pulse signals by analyzing subtle color changes in facial videos. Nevertheless, extracting rPPG signals remains challenging because of their extremely weak signal strength and susceptibility to illumination noise. In this paper, we propose an rPPG signal extraction method that exploits the quasi-periodic characteristics of rPPG signals. Our approach models quasi-periodicity of the rPPG signal, which arises from the stable cardiac cycle, as a block-sparse structure in the time-frequency domain. To incorporate a block-sparse model and enable adaptive signal separation under illumination fluctuations, we construct a time-varying signal separation framework. Experiments using a public dataset demonstrate the effectiveness of our method.