---
license: agpl-3.0
tags:
  - object-detection
  - yolo
  - ultralytics
  - football
---

# gst-python-ml football models

Weights for the football broadcast demo in [gst-python-ml](https://github.com/collabora/gst-python-ml). `demo/football/fetch_models.py` downloads them into `models/football/`.

| file | classes | input | role |
|---|---|---|---|
| `football.pt`, `football.onnx`, `football_fp16.onnx`, `football_int8.onnx` | ball, goalkeeper, player, referee | 640 | player detector |
| `ball11n.pt` | ball | 1280 | ball detector, run after the player detector |

## ball11n.pt

yolo11n fine-tuned from the ultralytics COCO weights for 40 epochs at imgsz 1280 on a one-class ball dataset of 15,357 training images. The player detector's own ball class is dropped in the demo in favour of this model.

| set | imgsz | P | R | mAP50 | mAP50-95 |
|---|---|---|---|---|---|
| SoccerNet v3 H250 test | 1280 | 0.774 | 0.708 | 0.701 | 0.327 |
| football-player-detection valid+test, ball only | 1280 | 0.921 | 0.850 | 0.888 | 0.407 |

Training data, all CC BY 4.0:

- SoccerNet v3 H250, Zenodo record [7808511](https://zenodo.org/records/7808511), ball labels only
- [martinjolif/football-ball-detection](https://huggingface.co/datasets/martinjolif/football-ball-detection)
- [martinjolif/football-player-detection](https://huggingface.co/datasets/martinjolif/football-player-detection), held out as the second evaluation set

## Licence

The weights derive from ultralytics YOLO and are released under AGPL-3.0 like the ultralytics package.