Reproduce “Retina : Low-Power Eye Tracking with Event Camera and Spiking Hardware” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Pietro Bonazzi, Sizhen Bian, Giovanni Lippolis, Yawei Li, Sadique Sheik, Michele Magno
Retina uses real event-camera recordings and a compact spiking network to predict pupil boxes, combining spike activity over time with a temporal filter. On Ini-30, it reports centroid error of 3.24 (± 0.79) pixels, below 3ET’s 4.48 (± 1.94). On synthetic LPW, it is worse than 3ET; Speck tests demonstrate on-chip execution, not continuous closed-loop tracking.
Key insights
- On Ini-30, Retina achieves 3.24 ± 0.79 pixels of pupil-centroid error, compared with 4.48 ± 1.94 pixels for 3ET; on the synthetic LPW event dataset, Retina performs worse than 3ET.
- Event-count-based windows often improve localization over fixed-time windows of similar average duration, but the gain comes with increased hardware cost in the reported comparison.
- Temporal integration and bounding-box prediction are central to performance: removing the temporal filter raises error to 24.46 pixels, while direct coordinate prediction yields 5.89 pixels versus 3.24 for boxes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.