Breaking01 / AI
E2VID yields 40 person detections in reconstruction test
Researchers reported a detection gap between reconstructed images and direct representations even in a retrospective comparison that capped background false positives.

Researchers reported 40 person detections with E2VID in an experiment using decision thresholds set on development data.
According to the paper's abstract, event cameras record changes in brightness asynchronously. The researchers compared ways of turning these measurements into inputs that a visual model can interpret.
The abstract describes an evaluation of 192 segments selected from DSEC recordings. The researchers reported that E2VID detected people in 40 of 52 segments containing people. It also returned positive signals in 12 of 140 background segments without people.
The researchers reported 21 person detections with FireNet+ at a fixed decision threshold. They observed no false positives in the background segments at that threshold.
The abstract also presents a retrospective comparison with the proportion of background false positives capped at 10%. Under that condition, the researchers reported 40 detections for E2VID and at most 19 for any direct representation without image reconstruction.
The paper's abstract does not disclose specific figures for the cost of continuous recording.