I trained a dataset on our own match footage, labelled and corrected frame by frame, and it is now public:
YOLO26s, three classes: `person`, `ball`, `referee`. Trained on 402 hand-corrected 3840x2880 frames from a two-camera fisheye rig, cut into 804 tiles of 1920x1920, holding roughly 7900 person, 580 ball and 580 referee boxes.
Use it via the **AI model (ONNX)** field in the GUI, or `–model best.onnx`. Nothing to rename: the classes are already called `ball` and `person`.
| class | precision | recall | mAP50 | mAP50-95 |
|---------|-----------|--------|-------|----------|
| all | 0.883 | 0.817 | 0.865 | 0.648 |
| person | 0.956 | 0.929 | 0.965 | 0.730 |
| ball | 0.844 | 0.612 | 0.724 | 0.500 |
| referee | 0.849 | 0.910 | 0.905 | 0.713 |