ACCV workshop on sensors meet computer vision

Workshop page: https://sensorsmeetvision.github.io/smv2026/

Unconventional Image Sensors Meet Computer Vision

A workshop at ACCV 2026, the 18th Asian Conference on Computer Vision, held in conjunction with the International Workshop on Image Sensors and Imaging Systems (IWISS) 2026 — bringing the Asia-Pacific image-sensor community and the computer-vision community into the same room.

14 December 2026 from 9:00 – 17:15 in Room 1007 Grand Cube Osaka

Format Invited talks + posters 

Most computer-vision venues treat the image sensor as a fixed, conventional RGB camera. But a new generation of image sensors is rapidly reshaping what can be measured and, in turn, what vision algorithms can achieve.

This workshop focuses on the co-design of unconventional image sensors and the computer-vision and computational-imaging methods they enable. Its central premise is bidirectional: novel sensors give rise to novel CV/CI methods, and the demands of CV/CI in turn drive new sensor designs. Unusually for a computer-vision venue, sensor-hardware researchers are part of the organizing team, so the program reaches down to the device level rather than treating sensing as a fixed front end.

The workshop is the first of its kind at ACCV, and is held in conjunction with IWISS 2026, the biennial International Workshop on Image Sensors and Imaging Systems. Co-locating the two events connects an established sensor-side international community — much of the world's image-sensor industry sits in the Asia-Pacific region — with the computer-vision community.

09:00 – 09:10    Opening remarks
09:10 – 10:25    Session 1
09:10 – 09:35    Event-based Photometric ImagingBoxin Shi (Peking University, China)
09:35 – 10:00    A Fully Reconfigurable Hybrid SPAD Vision Sensor Featuring Time-Coded Dual ExposuresSeong-Jin Kim (Sogang University, Korea)
10:00 – 10:25    SPAD Vision Chips with Near-Sensor Spiking Neural Network ProcessingLiyuan Liu (Chinese Academy of Sciences, China)
10:25 – 10:50    Coffee break
10:50 – 12:05    Session 2
10:50 – 11:15    Design and Implementation of Task-Specific Lensless Cameras (tentative)Seung Ah Lee (Seoul National University, Korea)
11:15 – 11:40    3D Integrated Image Sensors and Flexible Image Sensors by Using SOI TechnologyMasahide Goto (Toyohashi University of Technology, Japan), Shigeyuki Imura and Hiroto Sato (NHK Science & Technology Research Laboratories, Japan)
11:40 – 12:05    Processing-in-Sensor: From Imaging to IntelligenceChih-Cheng Hsieh (National Tsing Hua University, Taiwan)
12:05 – 13:30    Lunch
13:30 – 14:20    Session 3
13:30 – 13:55    Thermal and Transient Imaging for Scene Understanding (tentative)Kenichiro Tanaka (Ritsumeikan University, Japan)
13:55 – 14:20    Securing Vision at the Sensor: PUFs and Adversarial Example DefenseShunsuke Okura (Ritsumeikan University, Japan)
14:20 – 15:50    Poster session (with coffee break)
15:50 – 17:05    Session 4
15:50 – 16:15    Computational Imaging with Single-Photon Cameras (tentative)Kaushik Mitra (Indian Institute of Technology Madras, India)
16:15 – 16:40    Title to be announcedYoungcheol Chae (Yonsei University, Korea)
16:40 – 17:05    A 1.494-µm-Pixel CMOS Image Sensor with Cluster-Level ADC and Motion Compensation for Global-Shutter-Equivalent Imaging in Mobile ApplicationsSanggwong Lee (Samsung Electronics, Korea)
17:05 – 17:15    Closing remarks 

Samsung to exit 8-inch foundry business

Link: https://www.thelec.net/news/articleView.html?idxno=15179

Samsung Electronics to Exit 8-Inch Foundry Business by 2030

Company notifies customers of shutdown plans, leaving fabless chipmakers scrambling for alternative production capacity

South Korean fabless semiconductor companies that rely on Samsung for chip manufacturing are now scrambling to secure alternative production capacity.

Industry participants are also calling for Samsung Electronics to delay the closure of its 8-inch production lines. They argue that government intervention may be necessary because individual fabless companies have little leverage to change Samsung’s decision to end production. 

Programmable readout pixel

In a paper titled "A Programmable Readout Pixel Image Sensor" Lazzaro et al. from Dartmouth write:

In this paper, a pixel architecture that can be re-programmed to favor low noise, high speed, high dynamic range, or other characteristics desired in scientific imaging is proposed. Termed a programmable readout (PRO) pixel, this novel design combines a pinned photodiode, a deep-buried-channel CCD router, and multiple in-pixel amplifiers of different characteristics. By clocking the CCD router with different sequences, photogenerated charge can be steered to an appropriate readout amplifier given the application. The CCD router can be further programmed to store charge from multiple frames. The concept of such a pixel and its operation are described in detail and TCAD simulations aid in this discussion. To demonstrate this concept of programmability, a test chip is made with a CCD router and two in-pixel amplifiers: a floating diffusion amplifier (FDA) for fast readout and high-illumination imaging and a floating gate amplifier (FGA) for a low-noise Skipper-in-CMOS readout for low-light imaging. The test chip contains a 36 × 94 array of 20 µm pixels and contains 9 different pixel variants for experimental analysis. At the end of the paper, preliminary images captured from the different amplifiers are presented which demonstrate the ability to move charge to the selected readout while showcasing the different output characteristics of the two readout amplifiers. Furthermore, the output image from the Skipper-in-CMOS is compared across different numbers of non-destructive samples to demonstrate its noise-reduction capability.

Open access link: https://www.mdpi.com/1424-8220/26/16/5245

Figure 1. (a) PRO pixel concept with an adjacent PPD. (b) PRO pixel concept with a centered PPD. Both designs allow for readout by four different amplifiers of different characteristics.

 

 Figure 2. Test chip 20 × 20 μm PRO pixel showing simplified plan-view layout.

 Figure 11. (a) Test image captured using a 100 ns exposure time and the floating diffusion amplifier. (b) Test image captured using a 500 µs exposure time and the floating diffusion amplifier. (c) Test image captured using a 100 ns exposure time and the floating gate amplifier. (d) Test image captured using a 500 µs exposure time and the floating gate amplifier. It is noted that the horizontal bands are due to different pixel designs on those rows. 

Eric Fossum IEEE Jun-ichi Nishizawa medal announcement video

The IEEE Jun-ichi Nishizawa Medal was presented to Eric R. Fossum at the 2026 IEEE Honors Ceremony for the invention, development, and commercialization of the CMOS image sensor. 

The ceremony was held on 24 April 2026 in New York City.

The IEEE Jun-ichi Nishizawa Medal is given for outstanding contributions to material and device science and technology, including practical application.

ST announces a 1.1MP automotive CIS

STMicroelectronics introduces automotive image sensor to help scale affordable in-cabin sensing  

Helps OEMs and Tier 1s bring driver and occupant monitoring to high-volume vehicle platforms with lower system cost 

Delivers 35% higher image sharpness and nearly 60% higher infrared sensitivity than the previous generation in a compact 1.1 MP automotive image sensor 

STMicroelectronics (NYSE: STM), a global semiconductor leader serving customers across the spectrum of electronics applications, introduces ST SafeSense VD56GA, a new 1.1 MP automotive image sensor for infrared in-cabin sensing that helps OEMs and Tier 1 suppliers scale driver and occupant monitoring across vehicle lines while balancing performance, integration, and cost. 

As driver and occupant monitoring moves into more vehicle platforms, automotive manufacturers and suppliers are looking for imaging solutions that support strong in-cabin performance without adding unnecessary system cost or design complexity. VD56GA meets that need with a compact sensor architecture, high infrared sensitivity, and embedded image-processing features that can help simplify camera design and integration. The sensor is designed to give camera developers the image quality needed for driver and occupant monitoring while preserving the size and cost advantages of a compact format. In many in-cabin use cases, this can help customers meet performance targets without moving to a larger, more expensive sensor. 

“In-cabin sensing is moving beyond premium platforms, and carmakers need cost-effective solutions they can deploy broadly,” said Alexandre Balmefrezol, Executive Vice President and General Manager of the Imaging Sub-Group at STMicroelectronics. “With VD56GA, we focused on the system-level tradeoffs that matter most: image quality in infrared conditions, compact integration, and a cost structure that supports high-volume programs. Equally important, the sensor benefits from ST’s Integrated Device Manufacturer model, with front-end manufacturing in Crolles, France, giving customers greater supply-chain resilience, quality control, and long-term support for automotive programs.” 

“As driver and occupant monitoring systems move into a broader range of vehicle platforms, OEMs are increasingly focused on solutions that can be deployed at scale, with volumes set to exceed 70 million units by 2031. Success in this market depends not only on imaging performance, but also on reducing system cost, simplifying integration, and enabling adoption across high-volume vehicle programs," (*) underlined Anas Chalak, Market & Technology Analyst, Imaging at Yole Group.  

Helping carmakers scale driver and occupant monitoring 

VD56GA is designed to help reduce the cost of the complete camera module. Its imaging performance enables the use of lower-cost optics, simplified infrared illumination, less demanding optical filtering, and no external image-processing. For OEMs and Tier 1s, that can translate into a more cost-effective path to deploying driver and occupant monitoring across a broader range of vehicle programs. 

The sensor leverages a new backside-illuminated pixel architecture, called ST DeepNIR, optimized for infrared imaging. With 35% higher modulation transfer function (MTF) sharpness and nearly 60% higher quantum efficiency than the previous generation, VD56GA is currently the only sensor in its class to achieve this level of combined optical resolution and infrared sensitivity, helping in challenging conditions like under display camera. 

By extracting more performance from a compact 1.1 MP architecture, VD56GA helps automotive designers achieve the image quality required for advanced sensing functions, while benefiting from the integration and cost advantages of a compact sensor architecture.  

Compact integration with embedded image processing 

Designed for modern automotive camera modules, VD56GA is housed in a compact 3.7 mm x 3.2 mm CSP package, making it suitable for space-constrained installations and next-generation in-cabin camera designs. 

The device also integrates a range of embedded image-processing functions, including mirror, crop, dark calibration, auto exposure, and piecewise-linear processing. These features remove the need for external image-processing resources, simplify camera design and speed development. 

Support for both RAW and YUV output formats give OEMs and Tier 1 suppliers flexibility across different electronic architectures and software environments, making it easier to adapt the device to a broad range of implementation strategies. 

Building on ST’s automotive imaging leadership 

VD56GA builds on ST’s established presence in automotive imaging and in-cabin sensing. Earlier this year, ST announced shipment of its 10 millionth automotive of SafeSense by ST image sensors, reflecting broad adoption across key automotive regions, including Europe, Asia, and the United States.  

Customers benefit from ST’s vertically integrated manufacturing model and secure supply chain, with front-end manufacturing of VD56GA takes place at ST’s facility in Crolles, France and packaging in Asia; reinforcing the company’s focus on industrial resilience, product quality, and supply continuity for automotive customers worldwide. 

Availability 

Samples are available now. Pricing information and sample requests are available from local ST sales offices. 

Learn more: https://www.st.com/en/imaging-and-photonics-solutions/vd56ga.html

Leopard-Lumotive collaboration

Link: https://www.prnewswire.com/news-releases/leopard-imaging-and-lumotive-introduce-sirius-max--a-long-range-software-defined-3d-perception-platform-for-robotics-and-ai-302889658.html 

Leopard Imaging® and Lumotive Introduce "Sirius Max" -- A Long-Range Software-Defined 3D Perception Platform for Robotics and AI

 FREMONT, Calif., Sept. 24, 2026  -- Leopard Imaging®, a global leader in AI perception and spatial intelligence solution design and manufacturing, together with Lumotive, a pioneer in programmable optical semiconductor technology, today announced Sirius Max (LI-AR0234-AF0130-TX10-GMSL2-140H), a next-generation solid-state 3D LiDAR designed for advanced AI perception, robotics, industrial automation, autonomous machines, and intelligent edge systems.

Powered by Lumotive's programmable beam-steering technology and Leopard Imaging's high-performance embedded AI Perception expertise, Sirius Max delivers a compact, low-power, and software-defined 3D sensing solution optimized for next-generation machine vision and AI perception applications.

Sirius Max is Leopard Imaging's next-generation long-range RGB + depth perception platform, combining high-performance imaging, intelligent iToF depth sensing and processing, embedded synchronization, and programmable beam-steered illumination into a fully integrated GMSL2 camera system capable of real-time RGB-depth fusion and long-range active depth perception.

Unlike conventional flood-illumination ToF systems, Sirius Max utilizes directional beam steering to dynamically control illumination and depth scanning. This architecture significantly improves optical efficiency, ambient light rejection, signal-to-noise ratio, and long-range sensing performance while maintaining a fully solid-state design with no moving mechanical components.

The platform enables software-defined scanning and programmable sensing regions, opening new possibilities for intelligent perception systems in humanoid robotics, autonomous navigation, automotive monitoring, industrial automation, and smart infrastructure.

"Sirius Max represents a major advancement in programmable optical perception systems," said Bill Pu, CEO of Leopard Imaging. "By combining Leopard Imaging's AI perception and spatial intelligence expertise with Lumotive's breakthrough beam-steering technology, Sirius Max delivers long-range, software-defined 3D perception in a compact embedded platform. We are enabling AI systems to perceive farther, react faster, and perform more intelligent tasks with a single sensor platform."

"Leopard Imaging has extensive experience delivering high-performance embedded camera systems for AI and edge computing applications," said Apurva Jain, Senior Vice President of Product & Marketing at Lumotive. "Together, Sirius Max combines Lumotive's programmable beam-steering technology with Leopard Imaging's high-performance imaging platform to extend iToF sensing performance beyond 20 meters in both indoor and outdoor environments. With 1MP depth resolution and software-defined scanning capabilities, customers can leverage a single sensor platform for applications such as SLAM, collision avoidance, object handling, and navigation. Sirius Max enables AI systems to detect objects farther away, operate at higher speeds, and perform more intelligent tasks while reducing overall system complexity."

Advanced AI Perception Architecture

Sirius Max integrates synchronized RGB imaging, intelligent iToF depth sensing and processing, embedded processing, and Lumotive's TX10 programmable beam-steering transmitter into a compact GMSL2 camera module platform optimized for real-time AI perception workloads.

The platform supports synchronized RGB video, depth data, and control transmission over a single coaxial cable, simplifying integration into advanced embedded computing and edge AI systems. Through programmable scanning and region-of-interest control, Sirius Max can consolidate multiple sensing functions into a single camera platform. Customers can leverage one sensor for applications such as SLAM, cliff detection, collision avoidance, visual inspection, object handling, and navigation, reducing system complexity while improving performance and deployment flexibility.

Core technologies include

-   High-performance RGB imaging with integrated ISP
-   Intelligent iToF depth sensing based on the onsemi AF0130 depth sensor
-   Lumotive's LCM programmable beam-steering module

Key Features of Sirius Max

-   Solid-state wide-field-of-view RGB-D perception architecture
-   High resolution extended-range depth perception in indoor and outdoor conditions
-   Software-defined sensing in regions of interest
-   Compact, low-power embedded camera module design
-   Enhanced dynamic range, reduced multipath, and reduced saturation noise in all conditions
-   Embedded processing and synchronization
-   Low-latency GMSL2 connectivity
-   Optimized for robotics, industrial automation, automotive, and edge AI applications

Cognex acquires RealSense

Link: https://investor.cognex.com/news/news-details/2026/Cognex-to-Acquire-RealSense-Expanding-Machine-Vision-Leadership-into-High-Growth-Robotic-Perception-Market/default.aspx

Cognex to Acquire RealSense, Expanding Machine Vision Leadership into High-Growth Robotic Perception Market

September 22, 2026

NATICK, Mass. and CUPERTINO, Calif., Sept. 22, 2026 /PRNewswire/ -- Cognex Corporation (NASDAQ: CGNX), the global technology leader in industrial machine vision, today announced that it has entered into a definitive agreement to acquire RealSense, the market leader in depth-sensing cameras and vision technology for robotic perception and Physical AI. The acquisition advances Cognex's strategy to diversify its growth engine by entering the robotic perception market, a high-growth adjacency that expands its served market opportunity.

"RealSense represents a compelling strategic expansion for Cognex into robotic perception, one of the most attractive adjacent markets in machine vision," said Matt Moschner, President and Chief Executive Officer of Cognex. "RealSense brings a leading 3D perception platform with proprietary technology, a strong developer community and a compelling value proposition for customers. Combined with Cognex's machine vision expertise and global reach, RealSense positions us to offer a full-stack visual intelligence platform, spanning industrial ID, 2D and 3D machine vision measurement to 3D depth perception and robotic navigation, built to serve the full spectrum of Physical AI."

"RealSense was built on a simple idea: robots and autonomous systems can't act intelligently in the physical world if they can't first perceive it accurately. We've called that mission the Visual Cortex of Physical AI. Joining Cognex puts that mission on an entirely new scale, giving us access to a global industrial customer base and go-to-market reach that would have taken us years to build on our own. This is a combination that makes both businesses stronger," said Nadav Orbach, Chief Executive Officer of RealSense.

Originally founded by Intel in 2014 and spun out in 2025, RealSense has established a leading position in 3D robotic perception, with applications across fixed arm perception-guided robotics, autonomous mobile robots, quadrupeds and humanoids. The acquisition expands Cognex's served market opportunity by adding exposure to a robotic perception market estimated at $600 million today and expected to grow more than 25% annually to approximately $1.6 billion by 2030. RealSense is expected to generate revenue of $80 million to $90 million in 2026, representing more than 50% growth over the prior year.

"This transaction is highly aligned with our disciplined M&A strategy," said Dennis Fehr, Chief Financial Officer of Cognex. "RealSense is expected to be strongly accretive to Cognex's growth profile. Over time, we expect the business to scale in line with Cognex's through-cycle financial framework, including attractive Adjusted EBITDA margins and strong free cash flow conversion."

Under the terms of the agreement, Cognex will acquire RealSense for approximately $500 million, financed entirely from existing cash and investments on its balance sheet. Cognex also expects to provide a three-year $56.5 million cash retention program at target for RealSense employees, subject to performance modifiers, and to grant restricted stock units valued at approximately $50 million under Cognex's existing 2023 Stock Option and Incentive Plan dependent on the grant-date share price. The transaction is expected to close in the fourth quarter of 2026, subject to customary closing conditions.

Prior to closing, RealSense will spin out its Facial Authentication product line into an independent company, including all necessary functions to continue its focused growth strategy in the biometrics market.

Dr. Robert Henderson elected to the Royal Academy of Engineering

Link: https://raeng.org.uk/about-us/fellowship/new-fellows-2026/professor-robert-henderson-freng-frse/

Professor Robert Henderson’s innovations in the design of single photon avalanche photodiode (SPAD) devices using a standard industrial CMOS (complementary metal-oxide semiconductor) fabrication process have had major impact for optical distance sensing and have been globally adopted in smartphones and other technologies. He designed the first large two-dimensional arrays of SPADs (single-photon avalanche diodes), together with associated photon-timing and signal-processing electronics that were fabricated and marketed by ST Microelectronics as a laser auto-focus assist component for mobile devices. Over two billion of these novel products have been shipped and are integrated into over 170 different smartphone models for range-finding, gesture recognition and biomedical applications.

DPReview article on LOFIC

https://www.dpreview.com/articles/cmos-inventor-eric-fossum-helps-us-explain-lofic/

CMOS inventor Eric Fossum helps us explain LOFIC

 We worked with Prof. Eric Fossum to help explain the latest step forward in sensor design

It’s increasingly likely that you’ll have heard the term LOFIC, and some of the excitement around the capability it brings. We spoke to Professor Eric Fossum, inventor of the modern CMOS image sensor, who turns out to have some involvement in the early development of the concept, to explain how it works and what it could mean for cameras in future.

LOFIC stands for Lateral OverFlow Integration Capacitor and is a different way of designing pixel circuitry that allows them to capture more light. Conceptually, it stems from a similar idea to the dual conversion gain sensors we’ve seen over the past 12 years, but the benefits it promises appear even greater.

Fossum, now a Professor at Dartmouth College, uses the metaphor that capturing light is like collecting rainwater in a series of containers or buckets.

We’re going to slightly complicate the metaphor to encompass the concept of “conversion gain,” which is the relationship between the number of photoelectrons coming from the light-sensitive photodiode in each pixel and the size of the electrical voltage signal that’s then read out (measured in Volts per electron). 

In CMOS sensors, photoelectrons captured in a small bucket (smaller capacitance) give a proportionally bigger signal when read out than the same number would, if captured in a larger bucket (larger capacitance).

So a small bucket is well suited to low light (fewer photoelectrons) but you need a much larger one for capturing a lot of light, so that you don’t overfill the bucket.

In a LOFIC sensor, instead of one, we have both a small and a large bucket. The charge from the photodiode starts to collect in the small bucket, but if it exceeds its capacity, it overflows into the second, larger bucket. (Fossum called this “cascaded integration” in an early patent)

When it comes to readout, we measure the small bucket first. Because the charge was captured in a small bucket (with high conversion gain) we get a large signal that overwhelms any other noise that might get added later in the process. This gives us nice clean shadows and low-light images.

In a LOFIC sensor, we can then connect the two buckets together, combining their capacity (so, lower conversion gain) and adding in any overflowed charge. We then read out this combined signal. If we didn’t capture very much light, we can just use the first readout result. But if the small bucket was full, we can then use the second, combined readout, instead, giving us much more capacity for highlights.

Unlike the dual conversion gain system, Fossum points to the key benefit: “you don’t have to choose in advance.” And, as he points out, there’s no reason this system has to be limited to two buckets.

In practice, this means sensors that can tolerate a lot more light. And its implementation in relatively small, high-resolution sensors, suggests that space constraints aren’t a problem, at least not once you can use stacked sensor fabrication.

In the Osmo 4P we’ve seen DJI claim 17 stops of DR, which is an extraordinary number for a Type 1 sensor. This puts its DR on a par with much larger sensors (though doesn’t mean its image quality will match them). In its implementation, you can see how the additional dynamic range can be used.

The 4P’s camera that uses LOFIC technology is the only one that needs to use the new D-Log2 profile: a profile designed to accommodate more dynamic range. Interestingly, the base ISO of D-Log2 (the setting at which it captures maximum DR) appears to be 1600, suggesting that it’s being exposed to capture two additional stops of highlights compared with the ISO 400 base sensitivity of D-Log profile used for the conventional sensors.

However, the D-Log2 mode also lets you expose this same sensor state all the way down to ISO 100: clipping those additional highlights but using more exposure to capture the remaining tones, and thus boosting tonal quality. 

This gives us an idea of what LOFIC could bring to dedicated cameras: a choice of greater highlight capture or higher quality captures with the same highlight cutoff point as existing sensors.

After nearly a decade of sensors getting faster but not really offering improved image quality, LOFIC technology might be the step forward we’ve been waiting for.

Interestingly, Fossum – who is responsible for inventing the active pixel CMOS image sensor that virtually all modern cameras are based on – invented an early version of this concept, but the successors to the company he’d worked for instead concentrated their efforts on the simpler dual conversion gain system. A team at Tohoku University in Sendai, Japan, independently developed LOFIC as we’re now seeing it, citing Fossum’s Micron patent in their LOFIC patent, but proposing a more modern implementation.
 

ISSCC 2027 Late Breaking Industry News

ISSCC 2027 is introducing a new Late Breaking News session to provide an opportunity for companies to showcase significant product announcements that missed the regular ISSCC submission deadline. This new session will feature four industry papers describing production-grade chips or IP blocks incorporated into products launching in 2027.

Please register your intention to submit by uploading a paper title and a short abstract by October 7, 2026 using the submission site below. The final manuscript is due by December 1, 2026, after which papers will undergo an expedited review process, with acceptance notifications issued on December 9, 2026. This track is limited to four accepted papers and is reserved exclusively for production-grade industry submissions covering products launching in 2027.

The submission site and key dates are summarized below.

Submission Site

https://submissions.mirasmart.com/ISSCC2027


Key Dates

    October 7, 2026 — Intent-to-submit deadline
    December 1, 2026 — Paper submission deadline (camera-ready format)
    December 9, 2026 — Industry committee review and selection; notifications sent for the four accepted papers
    December 14, 2026 — Final accepted papers approved by the ISSCC Editorial Team
    December 14, 2026 — Session summary due for the Advanced Program