The screenshot below features Patrick Lichtsteiner and his work on mimicking retinal circuits in the design of the dynamic vision sensor (DVS), an event-based camera where the log difference of light intensity at time t and t-1 is emitted (the event), rather than a typical camera frame. This has extraordinary implications for visual processing, data transfer bandwidth and data storage.
An innovation I had missed: mixing in a DVS pixel with 15 other "normal" color pixels, in a 4x4 16-pixel macro-pixel configuration. And what for? To use the events from the fast, single DVS pixel to deblur and get sharper edges imaged by the other, slow 15 pixels. Can then use a 120 fps input plus the events to reconstruct the video at 5,000 fps (!).
Having determined that the DVS pixel noise is limited to 2x the shot noise, Tobi's group built a "Scientific DVS" targetting e.g., very fast imaging of neural activity with low noise. They've done it by tweaking the DVS pixel circuit and also binning 4 pixels together for spatial integration.
The result: 10x more sensitive.
Looking forward to seeing applications in neuronal activity imaging, which seems ideally suited for event-based imaging: large fields of view where largely nothing changes, with few, very sparse but fast changing pixels – where neurons are active.
Add comment