Model card¶
Models in this repository¶
| Model | Source | Licence | Used for |
|---|---|---|---|
| YOLOv8 n/s/m, YOLO11 n/s/m | Ultralytics, COCO-pretrained | AGPL-3.0 | person detection (class 0) |
| OSNet x0.25 / x0.5 (MSMT17) | Zhou et al. / torchreid authors, via Hugging Face kaiyangzhou/osnet |
MIT | appearance embeddings |
| ResNet18 (ImageNet) | torchvision | BSD-3-Clause | baseline generic embeddings |
web/public/models/yolo11n-*.onnx |
exports of YOLO11n | AGPL-3.0 (derivative) | in-browser detection |
No model is trained or fine-tuned in this repository. Checkpoint hashes are pinned in emot/reid/embedders.py.
Intended use¶
Counting and flow analysis of people in fixed or hand-held video; research on tracker design; edge-deployment studies.
Out-of-scope use¶
Identifying individuals; cross-camera tracking; surveillance of persons without a lawful basis; any decision about a person based on tracker output alone. See Privacy & ethics.
Performance¶
On MOT17 train (7 sequences), HOTA 0.35-0.40 averaged over detectors for the trackers here (Results); detection AP50 0.54-0.64 at 640 px depending on the backbone. Performance degrades with small/distant people, heavy occlusion, night scenes and low resolution; raising input size helps more than raising model size.
Known limitations¶
COCO-pretrained detectors, no fine-tuning; person class only; accuracy numbers are for this protocol and not leaderboard-comparable; fairness across demographics, clothing or lighting was not evaluated.