Judo-AI 3 Classes Models (Weights from "best.pt")

Check GitHub.

**Classes**

- `referee`: the match's referee
- `athlete_blue`: the athletes wearing blue judogis
- `athlete_white`: the athletes wearing white judogis
- `athlete_red`: the athletes wearing white judogis + red belt (for differentiation)

---

## **Results**

All metrics are on the held-out **test** split, produced by `src/evaluate_models.py` (`eval_exp1.csv`, `eval_exp2.csv`). Every run below is the 3-class scenario. Scenarios 1 (2-class) and 2 (4-class) were exploratory and were not carried through to test evaluation.

### Experiment 1 — 1,193-frame dataset, 200 epochs

| Architecture | Model  | mAP50 | mAP50-95 | Precision | Recall | F1    |
|--------------|--------|-------|----------|-----------|--------|-------|
| YOLOv11      | Nano   | 0.955 | 0.818    | 0.963     | 0.892  | 0.926 |
| YOLOv11      | Small  | 0.968 | 0.843    | 0.964     | 0.907  | 0.935 |
| YOLOv11      | Medium | 0.963 | 0.831    | 0.966     | 0.909  | 0.937 |
| YOLOv26      | Nano   | 0.950 | 0.823    | 0.953     | 0.905  | 0.928 |
| YOLOv26      | Small  | 0.963 | 0.842    | 0.953     | 0.915  | 0.933 |
| YOLOv26      | Medium | 0.958 | 0.832    | 0.967     | 0.894  | 0.929 |

### Experiment 2 — expanded 1,860-frame dataset, 250 epochs

| Architecture | Model  | mAP50 | mAP50-95 | Precision | Recall | F1    |
|--------------|--------|-------|----------|-----------|--------|-------|
| YOLOv11      | Nano   | 0.966 | 0.846    | 0.950     | 0.927  | 0.939 |
| YOLOv11      | Small  | 0.967 | 0.868    | 0.944     | 0.939  | 0.941 |
| YOLOv11      | Medium | 0.968 | 0.858    | 0.961     | 0.912  | 0.936 |
| YOLOv26      | Nano   | 0.962 | 0.851    | 0.937     | 0.932  | 0.934 |
| YOLOv26      | Small  | 0.959 | 0.828    | 0.944     | 0.915  | 0.929 |
| YOLOv26      | Medium | 0.963 | 0.875    | 0.955     | 0.930  | 0.943 |