Detects the 4 corners of a table-tennis table (far-left, far-right, near-right, near-left) in broadcast
main-camera frames, plus a "table present" score and per-corner visibility. Used by the PingPongTracker
match-recording tool (recorder_qt) and the Tabletennis_detr project.
| File | Use |
|---|---|
TableDetector.mlpackage/ |
Core ML (macOS / iOS), fp16, input 512x288 RGB, outputs sigmoid heatmaps (4x72x128), presence, visibility |
table_detector_v4.pt |
PyTorch checkpoint ({"model": state_dict, ...}) for TableKeypointNet (ResNet18 + 3 deconv), other platforms |
Model: SimpleBaseline-style heatmap keypoint net, input 512x288 (1280x720 scaled by 0.4), heatmap stride 4.
v3 (2026-11): trained with amodal targets for occluded corners + synthetic occlusion; 385 labelled frames from
9 venues / broadcasts. v4 (2026-11): v3 fine-tuned with 125 side-camera / replay negatives (presence = 0), so non-main-camera shots
no longer get a table; 510 frames total. Validation: visible-corner median error 2.0 px, occluded-corner median
2.8 px (1280x720); side-camera negatives judged "table present" 2/32 (v3: 32/32).
Older versions are kept as git tags (v3).
Decoding, tracking and the full pipeline: see the TableDetection/ folder in the PingPongTracker repository.
table_detector_v4.pt 62 MBTableDetector.mlpackage/Data/com.apple.CoreML/weights/weight.bin 31 MB