⚽ Soccer Object Detection with RF-DETR
Professional-grade object detection for soccer videos using RF-DETR-Large model.
Model: julianzu9612/RFDETR-Soccernet
- Architecture: RF-DETR-Large (128M parameters)
- Performance: 85.7% mAP@50, 49.8% mAP
- Dataset: SoccerNet-Tracking 2023 (42,750 images)
- Classes: Ball, Player, Referee, Goalkeeper
Upload a soccer image to detect objects
0.1 1
Detection Results
Example Images (Upload your own!)
Upload a soccer video to track objects frame by frame
0.1 1
1 10
⚡ Performance Tips:
- CPU: 2-3 FPS (slow) - Use frame_skip=5 and limit frames
- GPU: 12-30 FPS (fast) - Can process full videos
- Quick test: Use 300 frames with frame_skip=5
Detection Results
About This Model
🎯 Detected Classes
| Class | Color | Precision | Description |
|---|---|---|---|
| 🔴 Ball | Red | 78.5% | Soccer ball detection |
| 🟢 Player | Green | 91.3% | Field players from both teams |
| 🟡 Referee | Yellow | 85.2% | Match officials |
| 🔵 Goalkeeper | Blue | 88.9% | Specialized goalkeeper detection |
📊 Model Performance
- mAP@50: 85.7%
- mAP: 49.8%
- mAP@75: 52.0%
- Parameters: 128M
- Training Time: ~14 hours on NVIDIA A100 40GB
🎓 Training Details
- Dataset: SoccerNet-Tracking 2023
- Images: 42,750 annotated images
- Source: Professional soccer broadcasts
- Input Resolution: 1280x1280 pixels
- Optimizer: AdamW (lr=1e-4)
💡 Best Practices
Confidence Threshold:
- Use 0.5 for general detection
- Use 0.7+ for high-precision applications
Video Quality:
- Works best on 720p+ broadcast footage
- Standard broadcast camera angles preferred
Frame Processing:
- frame_skip=1: Every frame (best accuracy, slow)
- frame_skip=5: Every 5th frame (good balance)
- frame_skip=10: Every 10th frame (fast, lower accuracy)
🚨 Limitations
- Optimized for professional broadcast footage
- May have reduced accuracy in poor lighting
- Small balls may be missed when heavily occluded
- Camera angle dependency
📚 Use Cases
- Sports Analytics: Player tracking, formation analysis
- Broadcast Enhancement: Automatic highlighting, statistics overlay
- Research: Tactical analysis, computer vision benchmarking
- Video Analytics: Automated video processing pipelines
🔗 Links
📄 Citation
@misc{rfdetr-soccernet-2025,
title={RF-DETR SoccerNet: High-Performance Soccer Object Detection},
author={Computer Vision Research Team},
year={2025},
publisher={Hugging Face},
url={https://huggingface.co/julianzu9612/rf-detr-soccernet}
}
License: Apache 2.0