Computer vision for tracking social behavior in clownfish

Boston University

• Joined a computer vision project that tracks individual clownfish (Amphiprion percula) in underwater video to study social behavior and division of labor in size-ranked groups around host anemones
• Worked with a benchmark of 12,771 annotated frames and 37 fish trajectories, comparing SAM 3, YOLO + ByteTrack, GroundingDINO + ByteTrack and SU-T under occlusion and near-identical individuals, evaluated with HOTA, IDF1 and MOTA
• Contributing to Phase 2, which defines and evaluates behavior labels (23-behavior ethogram) with start and end times and actor/recipient roles