Reproduce “Real Eyes Realize Faster: Gaze Stability and Pupil Novelty for Efficient Egocentric Learning” 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
Ajan Subramanian, Sumukh Bettadapura, Rohan Sathish
A two-stage frame curator uses gaze stability to filter low-quality egocentric video frames, then pupil variation to prioritize potentially informative moments. On activity recognition, it reached macro F1 0.228 with 10% of training frames, slightly exceeding the all-frame score of about 0.224; gaze-only selection was more effective for scene recognition, showing that useful curation signals depend on the downstream task.
Key insights
- Gaze stability and pupil variation serve different purposes: gaze helps identify reliable frames, while pupil changes can prioritize moments associated with visual or behavioral transitions.
- Applying the signals sequentially worked better for activity recognition than simply combining them into one score; equal-weight fusion was close to random selection.
- The best selection signal varies by task and budget: pupil ranking helped most for activity recognition at tight budgets, whereas stable gaze was more useful for scene recognition.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.