# Normalized volleyball digital twin

## Truth model

- **Observed:** source pixels and persistent scene landmarks.
- **Detected:** pose, YOLO sports-ball, motion, and optical-flow evidence.
- **Inferred:** interpolation, event-local identity, normalized 2.5D coordinates, contacts, and net crossings.
- **Unknown:** metric depth/height/speed, exact score, and point winner.

The camera is fixed beside an edge-on net. Because official court boundary lines are not visible, the tactical floor is a normalized observation space split by the observed net base—not a regulation-size metric court.

## H01 00:12:17–00:12:22

Player tracks: 4. Ball hypothesis: medium confidence 0.82, coverage 0.84, YOLO anchors 11. Contact candidates: 4. Net-crossing candidates: 0. Physics-fit segments: 5.

## H02 00:13:07–00:13:14

Player tracks: 4. Ball hypothesis: medium confidence 0.82, coverage 0.871, YOLO anchors 20. Contact candidates: 4. Net-crossing candidates: 2. Physics-fit segments: 5.

## H03 00:34:31–00:34:43

Player tracks: 4. Ball hypothesis: medium confidence 0.82, coverage 0.933, YOLO anchors 29. Contact candidates: 12. Net-crossing candidates: 2. Physics-fit segments: 13.

## H04 00:35:35–00:36:10

Player tracks: 10. Ball hypothesis: medium confidence 0.82, coverage 0.974, YOLO anchors 19. Contact candidates: 29. Net-crossing candidates: 2. Physics-fit segments: 29.

## H05 00:55:44–00:55:55

Player tracks: 4. Ball hypothesis: medium confidence 0.82, coverage 0.982, YOLO anchors 37. Contact candidates: 7. Net-crossing candidates: 4. Physics-fit segments: 9.

## H06 01:06:24–01:06:45

Player tracks: 5. Ball hypothesis: low confidence 0.396, coverage 0.962, YOLO anchors 0. Contact candidates: 7. Net-crossing candidates: 0. Physics-fit segments: 8.

## H07 01:12:38–01:12:45

Player tracks: 4. Ball hypothesis: medium confidence 0.82, coverage 0.943, YOLO anchors 16. Contact candidates: 3. Net-crossing candidates: 1. Physics-fit segments: 3.

## H08 01:12:57–01:13:02

Player tracks: 5. Ball hypothesis: medium confidence 0.82, coverage 0.84, YOLO anchors 9. Contact candidates: 6. Net-crossing candidates: 0. Physics-fit segments: 4.

## Public synchronized viewer

- Dashboard: https://volleyball-b66ab85af2dc1f63.z.variksoo.ee/web/index.html#digital-twin
- Eight event windows expose 1,030 synchronized 10 fps states, H.264 evidence overlays, original and motion video sources, top-down tracks, and a draggable normalized 3D hypothesis.
- Layer controls cover players, ball, trails, net, uncertainty, and low-confidence ball hypotheses. H06 has zero model ball anchors, is labeled low confidence, and is hidden by default unless explicitly enabled.
- Public handoff verification exercised H07 seek-to-frame 35 at 3.5 s, synchronized playback to frame 42, original-source switching, 3D orbit/zoom, H06 confidence gating, and all four evidence links. HTTP byte ranges return 206; non-allowlisted source paths return 404.
- Public and local SHA-256 hashes were identical for the generated HTML, JavaScript, and H07 event JSON during final verification.

## Limitations

- Single camera and absent official boundary lines prevent unique metric 3D reconstruction.
- Player identities are event-local track hypotheses and can switch under occlusion.
- Ball tracks fuse sparse model detections with motion candidates and are hypotheses, not ground truth.
- The 3D view uses normalized net-relative coordinates and a height proxy; metric speed/height are unknown.