Reproducing every claim¶
| Claim | How to regenerate / verify |
|---|---|
| Leaderboard, ablations, per-sequence tables | make bench then python scripts/make_results_docs.py |
| HOTA is correct | pytest tests/unit/test_hota.py; TrackEval cross-check procedure in Experiments, Phase 7 |
| OSNet is correct | strict state-dict load; cosine 1.0 vs an independent ONNX export; pytest tests/unit/test_reid.py |
| ONNX detector == PyTorch | pytest tests/integration (real YOLO11n, IoU > 0.97 on the Ultralytics bus image; > 0.999 on MOT17) |
| Browser tracker/analytics == Python | python scripts/make_ts_fixtures.py && (cd web && npm test) |
| Edge numbers | make edge, python scripts/profile_onnx_ops.py (idle machine; results are machine-specific) |
| Camera-motion labels | python scripts/build_sequence_table.py |
| Figures / GIFs / banner | make figures, scripts/make_gifs.py, scripts/make_brand_assets.py |
Everything under results/ is committed so that docs and the web app always match the data in the repository.