Informal Seminar - Tobi Delbruck
Tobi Delbruck
Sensors Group at the Inst. of Neuroinformatics: www.ini.uzh.ch
UZH-ETH Zurich
The amazing evolution of electronic cameras is increasingly driven by the need for battery powered, always-on AI vision in future robots, spectacles and prosthetics. They must operate with uncontrolled high dynamic range lighting under severe power-latency tradeoff constraints. Neuromorphic "silicon retina" event cameras electronically model spike-based output from biological eyes to increase dynamic range by using pixel-level gain control and their quick, activity-driven sparse output enables AI vision systems with >10X improvement of power-latency tradeoffs compared with frame cameras, particularly when coupled to sparsity-aware neural accelerators. This talk covers the history of silicon retina development starting from Fukushima and Mahowald and Mead's earliest spatial retinas up to present-day industrial frame-event camera developments that have come full circle back to the basic message from biological evolution of dual-stream sustained-transient solutions. If time permits, this will be followed by a live demo of our contemporary frame-event CDAVIS camera that combines a vestibular sensor, a global-shutter image sensor, and a brightness-change event camera in one USB device.