From retina to silicon
The core idea is one line long: the retina detects change in light, not absolute brightness. ON cells fire when light increases, OFF cells when it decreases, and contrast and edges are picked out before the signal ever reaches the brain.
How the retina works
Photoreceptors
Rods and cones absorb photons and turn them into a graded voltage signal.
Horizontal cells
Sharpen contrast through lateral feedback across neighbouring receptors.
Bipolar cells
Split the signal into ON and OFF pathways — brighter versus darker.
Amacrine cells
Add temporal processing, emphasising how fast the light is changing.
Ganglion cells
Convert everything into spikes sent down the optic nerve.
Adaptation
Sensitivity re-tunes over time, so the eye works in sun and in shadow.
Working of an event-based camera
- 01
Photoreceptor stage
A photodiode converts light into current with a logarithmic response, which is where the huge dynamic range comes from.
- 02
Change stage
A circuit continuously compares the current signal with the last value that pixel recorded.
- 03
Comparator stage
If the change crosses a contrast threshold, the pixel fires an ON or OFF event — nothing else on the chip is disturbed.
- 04
Output
Each event is (x, y, timestamp, polarity): asynchronous, sparse, and timestamped to the microsecond. There are no frames.
The numbers
| Metric | Event camera | Conventional camera |
|---|---|---|
| Dynamic range | 120–143 dB | ~60–70 dB |
| Temporal resolution | Microsecond-level | Frame-limited (30–60 fps) |
| Power consumption | 10–100 mW | 1–3.5 W |
| Data bandwidth | ~0 on a static scene | ~27 MB/s (1280×720 @ 30 fps) |
A static scene generates essentially zero output: bandwidth scales with actual motion, unlike a frame camera that streams constantly regardless of what is in front of it. For scale, one human eye has about 100 million photoreceptors feeding roughly 1 million ganglion cells — the same compression trick, in wetware.
Pros and cons
Pros
- • Ultra-low latency — microsecond response.
- • Extremely high dynamic range, up to 120–143 dB.
- • Low power and bandwidth — only changed pixels emit data.
- • No motion blur, thanks to asynchronous per-pixel operation.
- • Ideal for high-speed tracking: drones, particles, robotics.
Cons
- • No absolute intensity or colour in pure DVS mode.
- • Adding intensity readout doubles pixel size and raises latency (ATIS).
- • Data is hard to compress — it does not fit standard video codecs.
- • Immature software ecosystem; legacy computer vision does not transfer.
- • Expensive, niche, and noisy in low-contrast or flickering scenes.
Research timeline
- 1991–92
Silicon retina
Mahowald & Mead (Caltech) build the first analog silicon retina.
- 2008
DVS128
Lichtsteiner, Posch & Delbruck publish the first 128×128 asynchronous DVS — 120 dB, 15 µs latency.
- 2011
ATIS
Posch, Matolin & Wohlgenannt add time-based intensity readout, 143 dB dynamic range.
- 2014
DAVIS
Brandli et al. combine DVS events with global-shutter frames — 130 dB, 3 µs latency.
- 2017
640×480 DVS
Son et al. (Samsung/ISSCC) reach a 9 µm pixel and 300 MEPS readout.
- 2019
Field survey
Gallego et al. publish the field-defining “Event-Based Vision” survey.
- 2020
HD stacked sensor
Finateu et al. (Prophesee/Sony) ship a 1280×720 back-illuminated stacked sensor.
- 2025
State-of-the-art review
Cimarelli et al. review hardware, algorithms and applications across the field.
Every section here is drawn from the project presentation.
View the slides