RF-DETR SoccerNet - Professional Soccer Object Detection

A state-of-the-art RF-DETR-Large model fine-tuned on the SoccerNet-Tracking dataset for detecting objects in soccer videos. This model achieves 85.7% mAP@50 and provides professional-grade analysis capabilities for soccer broadcasts.

🏆 Model Performance

🎯 Detected Classes

The model can detect 4 essential classes in soccer videos:

• ⚽ Ball - Soccer ball detection with high precision
• 🏃 Player - Field players from both teams
• 👨‍⚖️ Referee - Match officials
• 🥅 Goalkeeper - Specialized goalkeeper detection

🚀 Quick Start

Installation

pip install rfdetr pandas opencv-python pillow tqdm numpy torch torchvision

Basic Usage

from inference import RFDETRSoccerNet

# Initialize model (auto-detects CUDA/CPU)
model = RFDETRSoccerNet()

# Process video and get DataFrame
df = model.process_video('soccer_match.mp4', confidence_threshold=0.5)

# Display first 5 detections
print(df.head())

# Save results
model.save_results(df, 'match_analysis.csv')

Output DataFrame Format

The model returns a pandas DataFrame with comprehensive detection data:

📹 Video Processing Examples

Process Full Match

# Process entire match
df = model.process_video(
    'full_match.mp4',
    confidence_threshold=0.5,
    save_results=True
)

print(f"Processed {len(df):,} detections")
print(df['class_name'].value_counts())

Fast Processing (Every 5th Frame)

# Process every 5th frame for speed
df = model.process_video(
    'match.mp4',
    frame_skip=5,  # 5x faster processing
    confidence_threshold=0.6
)

Limited Frame Processing

# Process first 10 minutes only
df = model.process_video(
    'match.mp4',
    max_frames=18000,  # ~10 minutes at 30fps
    confidence_threshold=0.5
)

🖼️ Image Processing

# Process single image
df = model.process_image('soccer_frame.jpg', confidence_threshold=0.5)

# Display results
for _, detection in df.iterrows():
    print(f"{detection['class_name']}: {detection['confidence']:.2f}")

📊 Advanced Analysis

Ball Possession Analysis

# Analyze which players are near the ball
possession_df = model.analyze_ball_possession(
    df, 
    distance_threshold=100  # pixels
)

print(f"Found {len(possession_df)} possession events")

Filter and Analyze Results

# Get high-confidence ball detections
ball_df = df[(df['class_name'] == 'ball') & (df['confidence'] > 0.8)]

# Calculate average players per frame
avg_players = df[df['class_name'] == 'player'].groupby('frame').size().mean()

# Find frames with goalkeepers
goalkeeper_frames = df[df['class_name'] == 'goalkeeper']['frame'].unique()

# Analyze referee positioning
referee_df = df[df['class_name'] == 'referee']
referee_activity = referee_df.groupby('frame').size()

Export in Different Formats

# Save as CSV (recommended for analysis)
model.save_results(df, 'detections.csv', format='csv')

# Save as JSON (with metadata)
model.save_results(df, 'detections.json', format='json')

# Save as Parquet (for big data)
model.save_results(df, 'detections.parquet', format='parquet')

🎯 Use Cases

Sports Analytics

• Player Tracking: Monitor individual player movements
• Ball Possession: Calculate possession percentages
• Formation Analysis: Study team formations and positions
• Heat Maps: Generate player movement heat maps

Broadcast Enhancement

• Automatic Highlighting: Identify key moments
• Statistics Overlay: Real-time player/ball statistics
• Tactical Analysis: Formation and strategy analysis
• Performance Metrics: Player distance, speed analysis

Research Applications

• Tactical Research: Academic sports analysis
• Computer Vision: Object detection benchmarking
• Dataset Creation: Generate labeled training data
• Video Analytics: Automated video processing pipelines

📈 Performance Benchmarks

Processing Speed

• GPU (RTX 4070): ~12-15 FPS
• GPU (A100): ~25-30 FPS
• CPU: ~2-3 FPS

Memory Usage

• Model Size: 1.46 GB
• GPU Memory: ~4-6 GB
• RAM: ~2-4 GB

Accuracy by Class

🛠️ Advanced Configuration

Custom Confidence Thresholds

# Class-specific confidence tuning
df = model.process_video('match.mp4')

# Filter by class-specific confidence
high_conf_players = df[(df['class_name'] == 'player') & (df['confidence'] > 0.7)]
high_conf_ball = df[(df['class_name'] == 'ball') & (df['confidence'] > 0.5)]

Batch Processing

import os

# Process multiple videos
video_files = ['match1.mp4', 'match2.mp4', 'match3.mp4']

for video in video_files:
    print(f"Processing {video}...")
    df = model.process_video(video, save_results=True)
    print(f"Completed: {len(df)} detections")

📚 Integration Examples

With Pandas for Analysis

import pandas as pd
import matplotlib.pyplot as plt

# Process video
df = model.process_video('match.mp4')

# Create timeline analysis
timeline = df.groupby('timestamp')['class_name'].value_counts().unstack(fill_value=0)
timeline.plot(kind='line', figsize=(15, 8))
plt.title('Object Detection Timeline')
plt.show()

With OpenCV for Visualization

import cv2

# Load video and predictions
cap = cv2.VideoCapture('match.mp4')
df = model.process_video('match.mp4')

# Draw detections on video frames
for frame_num in range(100):  # First 100 frames
    ret, frame = cap.read()
    if not ret:
        break
    
    # Get detections for this frame
    frame_detections = df[df['frame'] == frame_num]
    
    # Draw bounding boxes
    for _, det in frame_detections.iterrows():
        x1, y1, x2, y2 = int(det['x1']), int(det['y1']), int(det['x2']), int(det['y2'])
        cv2.rectangle(frame, (x1, y1), (x2, y2), (0, 255, 0), 2)
        cv2.putText(frame, f"{det['class_name']}: {det['confidence']:.2f}", 
                   (x1, y1-10), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 2)
    
    cv2.imshow('Detections', frame)
    if cv2.waitKey(1) & 0xFF == ord('q'):
        break

cap.release()
cv2.destroyAllWindows()

🔧 Technical Details

Model Architecture

• Base: RF-DETR-Large (Real-time Detection Transformer)
• Backbone: DINOv2 with ResNet features
• Input Resolution: 1280x1280 pixels
• Output: 4 object classes with bounding boxes

Training Details

• Dataset: SoccerNet-Tracking 2023 (42,750 images)
• Hardware: NVIDIA A100 40GB
• Training Time: ~14 hours (4 epochs)
• Batch Size: 4
• Learning Rate: 1e-4
• Optimizer: AdamW

Data Preprocessing

• Augmentation: Random scaling, rotation, color jittering
• Normalization: ImageNet statistics
• Resolution: Multi-scale training (896-1280px)

🚨 Limitations and Recommendations

Known Limitations

• Optimized for broadcast footage: Best performance on professional soccer broadcasts
• Lighting sensitivity: May have reduced accuracy in poor lighting conditions
• Camera angle dependency: Trained primarily on standard broadcast angles
• Ball occlusion: Small ball may be missed when heavily occluded

Best Practices

• Confidence thresholds: Use 0.5 for general detection, 0.7+ for high precision
• Frame skipping: Use frame_skip=5 for fast processing without significant accuracy loss
• Resolution: Higher resolution videos (720p+) provide better results
• Preprocessing: Ensure good video quality and standard soccer broadcast setup

📄 Model Card

Model Details

• Developed by: Computer Vision Research Team
• Model type: Object Detection (RF-DETR)
• Language(s): N/A (Visual model)
• License: Apache 2.0
• Fine-tuned from: RF-DETR-Large (COCO pre-trained)

Intended Use

• Primary use: Soccer video analysis and sports analytics
• Primary users: Sports analysts, researchers, developers
• Out-of-scope: Non-soccer sports, amateur footage, real-time applications requiring <10ms latency

Training Data

• Dataset: SoccerNet-Tracking 2023
• Size: 42,750 annotated images
• Source: Professional soccer broadcasts
• Classes: 4 (ball, player, referee, goalkeeper)

Performance

• Test mAP@50: 85.7%
• Validation mAP: 49.8%
• Processing Speed: 12-30 FPS (GPU dependent)

Ethical Considerations

• Bias: Model trained on professional broadcasts may not generalize to amateur soccer
• Privacy: Ensure compliance with privacy laws when processing broadcast footage
• Fair use: Respect copyright and licensing of video content

📞 Support and Citation

Getting Help

• Issues: Report bugs and feature requests on GitHub
• Documentation: Comprehensive guides and examples included
• Community: Join our discussions for tips and best practices

Citation

If you use this model in your research, please cite:

@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/YOUR-USERNAME/rf-detr-soccernet}
}

Acknowledgments

• RF-DETR Architecture: Roboflow team for the excellent RF-DETR implementation
• SoccerNet Dataset: SoccerNet team for providing the comprehensive dataset
• Training Infrastructure: Google Colab Pro+ for A100 GPU access
• Community: Open source community for tools and feedback

---

🔄 Changelog

v1.0.0 (2025-07-29)

• ✅ Initial release with 85.7% mAP@50
• ✅ Complete DataFrame-based inference API
• ✅ Video and image processing capabilities
• ✅ Ball possession analysis tools
• ✅ Comprehensive documentation and examples
• ✅ Multi-format export (CSV, JSON, Parquet)

---

Ready to analyze soccer like never before? 🚀⚽

Get started with python example.py and explore the power of AI-driven sports analytics!