Silicon Retina: History, Live Demo, and Whiteboard Pixel Design
Rockwood Memorial Lecture 2023: Tobi Delbruck, Institute of Neuroinformatics, UZH-ETH Zürich
Event Camera Silicon Retina; History, Live Demo, and Whiteboard Circuit Design
Rockwood Memorial Lecture 2023 (11/20/23)
https://inc.ucsd.edu/events/rockwood/
Hosted by: Terry Sejnowski, Ph.D. and Gert Cauwenberghs, Ph.D.
Organized by: Institute for Neural Computation, https://inc.ucsd.edu
Abstract: Event cameras electronically model spike-based sparse output from biological eyes to reduce latency, increase dynamic range, and sparsify activity in comparison to conventional imagers. Driven by the need for more efficient battery powered, always-on machine vision in future wearables, event cameras have emerged as a next step in the continued evolution of electronic vision. This lecture will have 3 parts: 1. A brief history of silicon retina development starting from Fukushima’s Neocognitron and Mahowald and Mead’s earliest spatial retinas; 2: A live demo of a contemporary frame-event DAVIS camera that includes an inertial measurement unit (IMU) vestibular system, 3: (targeted for neuromorphic analog circuit design students in the BENG 216 class), a whiteboard discussion about event camera pixel design at the transistor level, highlighting design aspects of event camera pixels which endow them with fast response even under low lighting, precise threshold matching even under large transistor mismatch, and temperature-independent event threshold.
Is here scientific grade event cameras ?
ReplyDeleteYes, at around 45:30 he talks about SciDVS.
ReplyDeletethank you
DeleteWhat are principal applications for which event sensors are must to have please? thanks
ReplyDeleteYou might want to watch the lecture to answer that!
DeleteI mean the real commercial applications ...
ReplyDeleteI think Car industry main driver. May be military little bit. Someone who have very high dynamic range scenes, need low power consumption, automatic movement detection without complex programming
DeleteThe ability to stream visual information at milli-second increments and at very low power opens many potential applications in which Frame based approaches are limited. Blurr correction is an imediate application that could be applied on Mobile cameras, high speed imaging in automation is another. In the future, event cameras may be the most awaited breaktrough for AR/VR to interface correctly with human eyes and also the outside world. The same could be said in the context of Automotive and Autonomous Driving.
DeleteWonderful to see the link to the Toby Delbruck lecture here. I would additionally suggest "The Silicon Eye" by George F. Gilder - well written book that allows great insights into the neuromorphic imagers development and people connected to it. It has always been marvelous, how early on the key components have been invented and how long it takes mainstream products to develop from those inventions.
ReplyDeleteMaybe the MP won't come until spectacles finally come to MP.... we'll see. But obviously visual prosthetics for the blind that have to have low latency, HDR and low power consumption along with a sparse output that can drive inference hardware that can exploit this sparsity and a form of data that is closer to what the brain uses already.
ReplyDeleteBut main Problems eyes implants not sensors. Rather cost and connection to neurones.
DeleteI want Eric to comment on what I said about his definitions of the "perfect image sensor" around slide 4... what do you say Eric?
ReplyDeletehi Tobi, I think the main thing I would like to say is that I think it is wonderful that after all these years working on this technology, it is going niche-mainstream. Congratulations! Regarding off-hand silly comments on a "perfect image sensor," we can add zero power and zero latency to the list. I appreciate just being mentioned in your talk! Take care Tobi. Looking forward to our next meeting. We missed you in Scotland.
Delete50：00，cancel the offset of comparator with preamp. But why not to use the Auto zero method?
ReplyDeletethis should be for reducing kick-off
Delete