Soccer Player Segmentation (YOLOv8n-seg)
Instance segmentation model for detecting and segmenting soccer players in match footage.
Perfomance
Model Details
- Model type: YOLOv8n-seg (instance segmentation)
- Base model: ultralytics/yolov8n-seg
- Developed by: lifatsastain
- License: Apache 2.0
- Training compute: Kaggle Tesla T4
Dataset
- Source: FIFA World Cup Qatar 2022 — USA vs Netherlands (Dec 3, 2022)
- Images: 150 frames extracted from match footage
- Classes: Ball, Player, Ref
- Annotations: Bounding boxes converted to segmentation masks using SAM (Segment Anything Model)
- Split: 103 train / 28 valid / 19 test
Training
- Epochs: 100
- Image size: 640
- Batch size: 16
- Optimizer: AdamW
- Experiment tracking: W&B
- The chart
Evaluation Results
| Metric | Score |
|---|---|
| Box mAP50 | 0.919 |
| Box mAP50-95 | 0.725 |
| Mask mAP50 | 0.917 |
| Mask mAP50-95 | 0.539 |
| Precision | 0.945 |
| Recall | 0.886 |
How to Use
from ultralytics import YOLO
model = YOLO("lifatsastain/soccer-player-segmentation")
results = model("your_image.jpg")
results[0].show()