Industry Expert Talks

The following industry expert talks will be offered at ICIP 2026:

• Triplane-Video based Gaussian Splatting Compression and Applications, Guan-Ming Su (Dolby)
• Foundation Models for Population Brain Health: From Trustworthy Multimodal AI to Digital Biomarkers, Ngai-Man (Man) Cheung (Singapore University of Technology and Design)
• World’s first filter-free color splitting image sensor: from concept to silicon to first results, Gerd Van den Branden (Eyeo)
• From Semiconductor Manufacturing to Vision-Language AI: Evidence-Grounded Industrial Intelligence for Trustworthy Anomaly Reasoning at Scale, Leo Lin (Nanya TC)
• Brain-computer interfaces: are we there yet?, Ivan Tashev (Microsoft)
• Enabling Next-Generation Image Sensors Through Back-End Integration of IGZO Electronics and Metasurfaces, Stefano Guerrieri (imec)
• Automotive Imaging: The Need for Image Quality KPIs in Development and Sensor Degradation Detection, Sven Fleck (SmartSurv)
• Vision-Language Models for Manufacturing Inspection: Bridging the Synthetic-to-Real Gap Through Domain Knowledge Injection, Aditi Roy (Siemens)

Time: Monday, 14 September 2026, 12:30 PM – 1:00 PM

Abstract: Triplane-based implicit representations offer a highly promising structured framework for modeling static and dynamic 3D scenes. To meet the rigorous bandwidth and latency demands of immersive real-time streaming with practical deployment considerations, pioneering methods leverage established video coding standards like HEVC and VVC to compress these high-dimensional implicit representations. This talk focuses on state-of-the-art technologies engineered for compressing dynamic 3D scenes using Triplane-Video 3D Gaussian Splatting (TVGS), while exploring their extended applications across various domains.

The presentation begins with the introduction of 3D Gaussian Splatting foundations, covering primitive attributes, scene construction, parallel rendering, and end-to-end optimization pipelines. We then examine how dynamic scenes are efficiently represented using three interconnected components: coarse geometry anchor points, triplane feature vectors, and MLP decoders. We also detail the bit-plane splitting and packing strategies used to map these heterogeneous components into standard 2D video frames, fully exploiting hardware-accelerated video compression infrastructures. To ensure robust spatial-temporal stability, satisfactory visual fidelity, and optimal rate-distortion performance, several critical design paradigms will be presented, including channel-to-spatial mapping, anchor tracking, coordinate remapping, and dynamic range reshaping.

Furthermore, we demonstrate diverse downstream applications built on top of the TVGS framework. Specifically, we discuss converting conventional physically-based rendering (PBR) 3D mesh assets into fully relightable TVGS representations to achieve better visual quality, faster rendering, and higher data compactness. We also address a novel framework leveraging TVGS to represent multiple articulated 3D objects simultaneously and show the high effectiveness on rate-distortion-joint accuracy metrics. The talk concludes with a comprehensive overview of the latest international standardization efforts and critical future research trajectories. Ultimately, this presentation aims to provide ICIP 2026 attendees with a deep understanding of TVGS technology, motivating them to identify new research topics and inspiring innovative solutions in this rapidly evolving domain.

Bio: Guan-Ming Su received the Ph.D. degree from the University of Maryland, College Park. He is currently the Director of Research with Dolby Laboratories, Sunnyvale, CA, USA. He is the inventor of more than 240 U.S./international patents and pending applications. He is one of the recipients of the 2020 (72nd) Technology and Engineering Emmy Award and the 2021 (73rd) Engineering Emmy Award Philo T. Farnsworth Award for the contribution to high dynamic range (HDR) and wide color gamut (WCG) video as the Dolby Vision format. He received the 2025 University of Maryland ECE Distinguished Alumni Award and the 2025 APSIPA Industrial Distinguished Leader Award. His co-authored paper won the Best Industry Paper Award at IEEE ICIP 2025. He served in multiple IEEE international conferences, such as the TPC Co-Chair of ICME 2021, the Industry Innovation Forum Chair of ICIP 2023 and 2025, and the General Co-Chair of MIPR 2024 and 2025. He served as a VP for industrial relations and development in APSIPA from 2018 to 2019. He has been serving as the Vice Chair for Conference in the IEEE Technical Committee on Multimedia Computing (TCMC) since 2021. He served as an Associate Editor for APSIPA Transactions on Signal and Information Processing, IEEE Multimedia Magazine, and now IEEE Transactions on Circuits and Systems for Video Technology.

Time: Tuesday, 15 September 2026, 8:30 AM – 9:00 AM

Abstract: Foundation models are reshaping artificial intelligence and are beginning to transform healthcare and brain health research. Unlike traditional AI systems that primarily focus on individual modalities or narrowly defined tasks, next-generation foundation models for brain health aim to integrate heterogeneous and multimodal data sources to enable population-scale assessment and personalized monitoring of cognitive health.

This talk provides an overview of emerging trends in foundation models for brain health and discusses several key challenges, including learning from heterogeneous and largely unlabeled medical data, ensuring trustworthy multimodal reasoning, and identifying scalable digital signals for the early detection of cognitive decline. Illustrative examples from recent work on neuroimaging foundation models, medical multimodal large language models, and digital phenotyping will be presented to highlight both the opportunities and limitations of current approaches. The talk concludes with a perspective on how trustworthy multimodal foundation models may enable continuous, preventive, and personalized brain healthcare at population scale.

Bio: Ngai-Man (Man) Cheung is Associate Professor and Associate Pillar Head (Education) at the Singapore University of Technology and Design (SUTD), Singapore. He was a core team member of Singapore’s National Research Foundation (NRF) Foundational Research Capabilities programme for Artificial Intelligence and previously served as an AI Advisor to the Smart Nation and Digital Government Office (SNDGO), Singapore. His research focuses on trustworthy and efficient AI systems for perception and real-world deployment, with recent work spanning multimodal foundation models, medical AI, and digital health.

Prof. Cheung has published more than 200 papers in leading international venues and holds 14 U.S. patents, several of which have been licensed and translated into real world applications. His work has received international recognition, including a CVPR Best Paper Finalist Award and selection as a Finalist at the World Artificial Intelligence Conference (WAIC), Shanghai, China, for AI-driven healthcare innovation. He currently serves as Senior Area Editor of IEEE Transactions on Image Processing and IEEE Signal Processing Letters, and is the General Chair of IEEE VCIP 2026.

Time: Tuesday, 15 September 2026, 9:00 AM – 9:30 AM

Abstract: Color imaging has relied on the same fundamental principle for more than half a century. The Bayer color filter array, used in virtually all CMOS image sensors today, applies a mosaic of red, green, and blue filters across the pixel array. Each filter transmits only a portion of the light spectrum while absorbing the rest, meaning that up to two thirds of incoming photons never reach the photodetector. This inherent loss sets a fundamental limit on sensor sensitivity and signal-to-noise ratio, impacting image quality across applications, from smartphones to industrial vision systems. eyeo introduces a fundamentally different approach. Instead of filtering light, its sensors use NCOS® (Nanophotonic Color Splitting), a waveguide-based architecture that separates light by wavelength and directs it to the appropriate pixels. This eliminates the need for color filters and microlenses, ensuring that no photons are discarded and preserving full spectral information.

This presentation introduces the first filter-free color splitting image sensor and shares initial results from a fabricated prototype. Beyond the sensor itself, it focuses on the image processing challenges and opportunities enabled by this new architecture. Color splitting produces a different kind of raw data than a traditional Bayer sensor, with adapted color conversion, demosaicing, and ISP processing developed alongside the sensor. The results highlight a clear path toward significantly improved image quality through enhanced light sensitivity, with strong potential for applications in smartphones, XR devices, and machine vision.

Bio: Gerd Van den Branden is the CPO and co-founder of eyeo, a Dutch startup developing image sensors powered by advanced nanophotonic color-splitting technology. The company aims to redefine image sensing by overcoming the limitations of traditionalcolor filters and achieving higher image quality, resolution, and low-light performance.

Time: Tuesday, 15 September 2026, 9:30 AM – 10:00 AM

Abstract: Industrial artificial intelligence is undergoing a fundamental transition from traditional vision-based inspection systems toward multimodal reasoning frameworks that integrate vision, language, and structured decision-making. While modern deep learning systems achieve strong performance in anomaly detection and segmentation, they often fail to provide consistent, explainable, and spatially grounded reasoning required for deployment in real manufacturing environments.

In this talk, I present a unified perspective bridging over 20 years of semiconductor manufacturing and automation experience with recent advances in vision-language models (VLMs) for industrial anomaly reasoning. The focus is not only on improving detection accuracy, but on building trustworthy industrial AI systems that align perception, reasoning, and decision consistency.

We introduce the concept of evidence-grounded industrial intelligence, where AI systems are required to:

1. Localize visual evidence of defects,
2. Generate structured diagnostic reasoning aligned with visual cues,
3. Ensure consistency between answers, spatial localization, and textual explanations.

A representative framework, EGVLR (Evidence-Grounded Vision–Language Reinforcement), is presented as a case study. EGVLR introduces a unified Evidence-Driven Diagnostic Protocol (EDDP) that enforces structured output alignment across reasoning traces, bounding boxes, and final decisions. The system further incorporates staged optimization including visual pre-alignment, instruction tuning, and geometry- and semantics-aware preference optimization to reduce hallucinated reasoning and spatial inconsistency in multimodal industrial tasks. Experimental studies on multimodal industrial anomaly benchmarks demonstrate that enforcing structured consistency significantly improves localization reliability, reduces
false-positive spatial grounding on normal samples, and enhances reasoning stability compared to standard vision-language models. Beyond methodology, this talk emphasizes real-world deployment considerations in semiconductor manufacturing environments, including AOI (Automatic Optical Inspection) systems, AMHS logistics automation, and large-scale fab-level AI integration. We highlight practical constraints such as latency, robustness, and interpretability that are often overlooked in academic benchmarks.

The goal of this talk is to provide a practical roadmap for researchers and industry practitioners to design deployable, interpretable, and reliable multimodal AI systems for industrial inspection and beyond.

Bio: Leo Lin is a Ph.D. candidate in Artificial Intelligence at National Tsing Hua University, Taiwan, and a senior program manager at Nanya Technology Corporation, with over 20 years of experience in semiconductor manufacturing systems, including MES, AMHS, and AOI deployment. His current research focuses on vision-language models for industrial anomaly reasoning, multimodal learning, and structured reasoning in industrial AI systems. He has published multiple works in CVPR/ECCV/ICIP workshops on anomaly detection, representation learning, and multimodal reasoning systems. He specializes in bridging academic AI research with real-world semiconductor manufacturing deployment, with direct experience in large-scale production systems and AI-driven automation.

Time: Tuesday, 15 September 2026, 10:00 AM – 10:30 AM

Abstract: A brain-computer interface (BCI) is a direct connection between the human brain and a computer. It records central nervous system activity and interprets it to assess brain states or decode commands. This talk will explore methods for detecting and interpreting brain activity, highlight notable research achievements, and review the emerging BCI market, including devices and algorithms developed at Microsoft Research. The topic connects to image processing in several ways: brain signals are often represented as spatial grids that can be analyzed with image-processing techniques; neuroscience offers insights into human vision that can inspire new computational methods; human and computer vision can complement each other in symbiotic systems; and computer vision and neuroscience are likely to converge in future high-definition sensory prosthetics for augmenting human vision.

Bio: Dr. Ivan Tashev is a Partner Software Architect at Microsoft Research (MSR) in Redmond, WA, USA, where he leads the Audio and Acoustics Group and coordinates MSR’s Brain-Computer Interfaces project. His research interests include multichannel signal processing with machine learning and AI-based approaches. Dr. Tashev is an Affiliate Professor in the Department of Electrical and Computer Engineering at the University of Washington in Seattle and an Honorary Professor at the Technical University of Sofia, Bulgaria. His technologies have been incorporated into many Microsoft products, and he served as the audio architect for Kinect and HoloLens. He is an IEEE Fellow and a member of AES and ASA. More information is available on his web page https://www.microsoft.com/en-us/research/people/ivantash/.

Time: Wednesday, 16 September 2026, 8:30 AM – 9:00 AM

Abstract: The continuous scaling of pixel size is increasingly constrained by the minimum area needed for pixel circuitry, while emerging applications demand simultaneously higher resolution, wider dynamic range, lower noise, and enhanced optical functionality. To address these challenges, imec is exploring a new image sensor architecture that combines back-end-of-line (BEOL) integrated Indium-Gallium-Zinc-Oxide (IGZO) thin-film transistors with advanced meta-optical components. BEOL-compatible IGZO devices can be deposited at temperatures below 400°C, enabling monolithic integration above CMOS circuitry without disrupting conventional image sensor manufacturing flows. This approach allows pixel transistors to be relocated from the silicon substrate into the metallization stack, further reducing area competition within the pixel and creating a pathway toward unprecedent pixel pitches while reducing the complexity and cost associated with 3D stacked wafers. Complementing these advances in electronics, flat optics offer a compelling pathway toward highly miniaturized optical systems with enhanced functionality. Nanostructured metasurfaces operating in the visible spectrum, combined with refractive optics, can significantly reduce the complexity of conventional multi-element refractive optical stacks while integrating additional optical functions, potentially at wafer level. This approach enables thinner camera modules, improved spectral and polarization selectivity, and novel computational imaging capabilities. At imec, we believe that the combination of BEOL-integrated IGZO electronics and wafer-level meta-optics establishes a compelling platform for next-generation visible, SWIR, multispectral, and polarization-sensitive image sensors. Together, these innovations enable higher performance, greater functional integration, and a new level of miniaturization to support future portable, wearable, robotic, and other consumer imaging applications where device size and sensing performance are critical differentiators.

Bio: Stefano Guerrieri received his M.Sc. degree in Solid State Physics from the University of L’Aquila, Italy, in 1999, with a thesis focused on DRAM technology development at Micron Technology Italy. He started his career at Applied Materials as a System Expert before joining STMicroelectronics’ Central R&D organization in Agrate Brianza, Italy, where he contributed to the development of advanced NAND technology nodes. In 2008 he joined Micron Technology Italy as a Principal Engineer, leading the development of innovative image sensor technologies, including Backside Illumination (BSI) and light-pipe architectures. In 2014, Stefano joined imec, where he contributed in advancing next-generation pixel technologies and imaging solutions, working on both silicon-based and emerging material platforms. In September 2019, Stefano joined ams AG as Engineering Fellow, focusing on innovation in sensing and image sensor technologies. In July 2025, he returned to imec and joined the Technology Portfolio Office, where he is responsible for shaping the strategy of technology developments of advanced technologies for consumer and industrial applications. Throughout his career, Stefano has authored or co-authored numerous patents and scientific publications.

Time: Wednesday, 16 September 2026, 12:00 PM – 12:30 PM

Abstract: Cameras have become safety-critical sensors in automated driving, ADAS, and in-cabin monitoring. Yet the industry still often treats them as static components: qualified at start of production, integrated into perception stacks, and then trusted throughout the vehicle lifetime. This talk will raise awareness of what can go wrong when automotive imaging performance is not rigorously specified, reviewed, or monitored.

The presentation will raise awareness of how image quality and camera fitness relate to concrete safety scenarios in exterior perception and in-cabin sensing. In automated driving and ADAS, degraded or miscalibrated cameras can contribute to missed pedestrians, delayed AEB reactions, phantom braking, incorrect time-to-collision estimates, and poor object localization.

A key challenge is the need for robust, application-relevant KPIs to measure camera image quality and link to perception performance. The talk will discuss the role of IEEE P2020 in defining automotive image-quality KPIs. Once in the field, sensor system degradation detection is still unaddressed and is another motivator for KPI driven design.

Bio: Dr. Sven Fleck is an imaging scientist and consultant working through SmartSurv on automotive camera systems, image quality, and perception-related sensor performance. He is co-founder of obsurver and co-founder and Vice Chair of the IEEE P2020 Working Group on automotive image-quality KPIs. He earned his M.S. and Ph.D. in Computer Science from the University of Tübingen. He has worked with automotive OEMs and suppliers on camera system development and imaging performance.

Time: Wednesday, 16 September 2026, 12:30 PM – 1:00 PM

Abstract: Deploying visual defect detection in manufacturing confronts three challenges familiar across image processing domains from medical imaging to semiconductor fabrication: extreme scarcity of real defective training data, the need to generalize across product variants without retraining, and regulatory demands for explainable decisions. This talk presents our industry experience addressing these challenges through a two-stage cascade framework combining supervised anomaly detection trained on synthetic-only data with vision-language model (VLM) reasoning, and the practical lessons learnt from deployment in production.

Stage 1 of our framework employs CLIP ViT-L-14 with Defect-Aware Prompt Optimization, trained entirely on synthetic data, to generate defect probabilities and class activation heatmaps under 200ms. A confidence gate routes ambiguous cases to Stage 2, where an 8B-parameter VLM reasons over a visual triangle: the test image, the Stage 1 heatmap, and a optional-good reference image.

The key enabler for real-world generalization is domain knowledge injection at inference time. Manufacturing engineers author a structured defect taxonomy which is semantic description for each defect type with visual identification criteria (e.g., “the rectangular hole behind the hexbolt is partially visible when the washer is missing”). Embedded in the VLM prompt, this taxonomy grounds reasoning in domain-specific cues and constrains outputs to valid categories, requiring zero model retraining for new product lines. The VLM’s pre-trained visual representations simultaneously bridge the synthetic to-real domain gap: when Stage 1 predictions degrade on real production images, the VLM compensates by reasoning over actual pixel content guided by the injected defect taxonomy. This enables deployment with zero real defective training samples across three product families spanning 3–8 defect classes.

We discuss deployment tradeoffs from production experience, including confidence threshold tuning, the cost-accuracy balance of the two-stage gate, and why natural-language reasoning traces proved essential for operator trust in regulated environments.

Bio: Dr. Aditi Roy is a Senior Key Expert in the Smart Machine Vision Group at Siemens Technology, Princeton, NJ, with 25+ peer-reviewed publications (1300+ citations, h-index 16) and multiple patents. She has served as PI/Co-PI on $3M+ in federally funded research (NSF, DARPA, DOE, MxD, ARM Institute), including the MxD Smart-DARPQuA program which deployed a real-time welding inspection system achieving 99.15% measurement accuracy in production. She received the Women in IT “Rising Star” award in 2020.