Volleyball AI

Offline / batch volleyball analysis: upload a court video, detect keypoints, calibrate the camera, track players (SAM 3.1) and ball (VballNet), detect actions, count score, and view a 3D court. Hard rule: no model inference on the laptop. Cloud Run Jobs orchestrate ffmpeg + stages; all neural nets run on Modal. Local v0 scaffolds Next.js web + worker polling with parallel court/player/ball stages after normalize.

Contributions

• Locked hybrid architecture: web control plane, Cloud Run job worker, Modal AI stages, optional Supabase later.
• Pipeline stages: ingest, normalize, detect_court (YOLOv11n-pose), track_players, track_ball, calibrate, actions, score, project_3d (+ optional splat scene).
• Docs for AI policy, jobs, architecture, data model, and decision log.

Outcomes

• Local v0 path: upload mp4 → Modal tracks → calibration → overlays + 3D positions.
• Quality-over-realtime batch design with cacheable stage artifacts.