Event-Based Vision

last updated 2026-05-04

Physics / mechanism

Event-based vision sensors (dynamic vision sensors, DVS) mimic retinal ganglion cells: each pixel fires asynchronously when local log-luminance change exceeds a threshold (~10–50 mV typically), rather than integrating full frames. Output is a sparse stream of (x, y, t, polarity) events at sub-microsecond temporal resolution. Latency is ~1 µs vs ~33 ms for 30 fps CMOS. Dynamic range reaches 120–140 dB (vs ~60 dB for conventional imagers). Leading silicon: Prophesee Metavision IMX636 (1280×720, ~4.86 M px), Samsung DVS, Sony IMX636-based derivatives, iniVation DAVIS346. Fabricated on standard CMOS nodes (180 nm–65 nm); pixel pitch typically 4.86–9 µm.

Competitive landscape

Event-based sensors compete directly with high-speed global-shutter CMOS (Sony Pregius, ON Semi Python) and neuromorphic frame-free alternatives. Adjacent approaches include SPAD arrays (single-photon, time-resolved but frame-based), LiDAR for depth tasks, and optical flow ASICs. Processing back-ends compete with conventional CNN inference pipelines vs spiking neural network (SNN) accelerators.

Companies using

Connected ideas

Sources

Frontier (open questions)

Frontier questions