AI-Powered Rugby Element Detection

Project Overview

A computer vision system that detects and classifies key rugby elements (players, referees, rucks, scrums, balls, line outs, and mauls) with high precision using YOLOv5 architecture. The system achieves 90%+ precision with real-time inference capabilities (8.8ms per image).

Technologies Used

• PyTorch
• YOLOv5
• Roboflow
• OpenCV
• Python 3.8+

Dataset

• Custom annotated dataset of 2,100+ rugby images
• Classes: ball, line out, maul, player, referee, ruck, scrum
• Precision bounding box annotations

Model Architecture

This project implements a fine-tuned YOLOv5 architecture with the following specifications:

• YOLOv5s as base model with transfer learning
• Input resolution: 640x640 pixels
• Batch size: 16
• Training for 100 epochs

Results & Performance

Overall Metrics

• Precision: 90%+ across all classes
• Inference Speed: 8.8ms per image
• mAP@0.5:0.95: Varied by class

Class-Specific Performance

• Strong performing classes:
  • Scrum (mAP: 94.0%, precision: 91.6%, recall: 89.1%)
  • Player (mAP: 90.2%, precision: 92.8%, recall: 82.1%)
  • Referee (mAP: 90.1%, precision: 95.3%, recall: 80.0%)
  • Ruck (mAP: 86.0%, precision: 81.7%, recall: 85.6%)
• Average performing classes:
  • Line out (mAP: 56.8%, precision: 55.3%, recall: 64.3%)
  • Ball (mAP: 46.1%, precision: 72.1%, recall: 25.4%)
• Weak performing classes:
  • Maul (mAP: 12.5%, precision: 12.4%, recall: 12.9%)

Implementation Workflow

1. Dataset Acquisition and Setup

• Acquired the dataset via Roboflow
• Set up YOLOv5 environment and properly configured data paths
• Created appropriate data.yaml configuration file

2. Model Training

• Implemented training pipeline with appropriate hyperparameters
• Trained using GPU acceleration
• Monitored training metrics and saved checkpoints

3. Model Evaluation

• Evaluated the model on a test dataset
• Generated confusion matrix and precision-recall curves
• Calculated class-specific performance metrics

4. Inference on New Images

• Created a testing pipeline for new rugby images
• Implemented visualization of detection results

Technical Challenges & Solutions

Challenge: Small Object Detection

Problem: The rugby ball was difficult to detect consistently due to its small size in many frames. Solution: Applied focused data augmentation techniques specifically for images containing balls, including random scaling and rotation.

Challenge: Class Imbalance

Problem: Player detections significantly outnumbered other classes like mauls and line outs. Solution: Implemented weighted loss function to give higher importance to underrepresented classes.

Challenge: Distinguishing Similar Classes

Problem: Mauls, rucks, and scrums can appear visually similar in certain camera angles. Solution: Enhanced the dataset with more diverse examples of each formation type and additional annotation quality checks.

Future Improvements

• Collect more training data for maul and ball classes
• Train with a larger model (YOLOv5m, YOLOv5l)
• Add more data augmentation to increase variety
• Fine-tune model parameters specifically for small object detection
• Implement tracking to maintain consistent detections across video frames

Installation & Usage

# Clone repository
git clone https://github.com/YOUR_USERNAME/rugby-element-detection.git
cd rugby-element-detection

# Set up environment
pip install -r requirements.txt

# Run inference on sample image
python detect.py --source sample.jpg --weights models/best.pt