Wednesday, July 27, 2011

SeeBetter Supports Asynchronous Silicon Retina Development

SPIE: European SeeBetter program is aimed to design of an integrated vision sensor adding asynchronous retinal functionality to artificial vision sensors. Recent silicon retinas include event-based readouts to mimic the asynchronous spikes on the eye optic nerve and reduce the output data amount or increase the effective frame rate:

SeeBetter intends to build a hybrid vision sensor where pixel electronics die is bump-bonded to BSI photodiode array. This approach allows to implement a complex retina cell models while maintaining large fill factor of the photodiode array.

The article is written by the program participants:

Konstantin Nikolic
Imperial College
London, UK

David San Segundo Bello
Leuven, Belgium

Tobi Delbruck, Shih-Chii Liu
Institute for Neuroinformatics
Zurich, Switzerland

Botond Roska
Friedrich Miescher Institute
Basel, Switzerland


  1. what is the usefulness of such structure ?

  2. @ "what is the usefulness of such structure ?"

    The main use will probably be to get research funding for the investigators and their institutions.

    At the nuts and bolts level, it looks like there are two ideas. One is to mimic biological vision by taking a photocurrent, then calculating the logarithm of the photocurrent, then calculating the derivative of the logarithm, then amplifying the derivative, then providing binary flags whenever the amplified derivative deviates enough up or down from a starting value. The various processing stages are analogous to components of the human visual front-end, and the authors suggest that an innate superiority of biological vision over ordinary image sensor techniques will mean their approach will also be better. The other idea I see is to figure out how to put a photo-detecting chip of one's choice on an underlying processing chip of one's choice.

    To me, the second idea is more interesting than the first, because if you can stack enough per-pixel circuits under a pixel on one or more chips with acceptable connections you might be able to get rid of some key performance-limiting ADC bottlenecks. The first idea has an "It's biomimetic, so it's better!" flavor that seems like a lay sales pitch to me. If it is superior, show me the metrics.

  3. people has worked on this idea since 1970s. The first one is built by NHK guys with discret photodiodes and OP-Amps. No single industrial application has been found. Europ funds all these usefuless stuffs with a huge debt burden.


All comments are moderated to avoid spam.