Hockey Rink Keypoint Detection

🔗 This model is trained on the HockeyRink dataset.

📊 Access the dataset used for training here: https://huggingface.co/datasets/SimulaMet-HOST/HockeyRink
🚀 Try the model in action with our interactive Hugging Face Space: https://huggingface.co/spaces/SimulaMet-HOST/HockeyRink

This repository contains a YOLOv8-based model for detecting and mapping keypoints on ice hockey rinks. The model is trained on the HockeyRink dataset, which comprises precise annotations of hockey rink landmarks.

Features

Accurate detection of 56 keypoint landmarks on hockey rinks
Real-time keypoint visualization with confidence scores
Support for various camera angles and lighting conditions
Handles player occlusions and dynamic game situations
Trained on diverse SHL (Swedish Hockey League) game footage

Model Details

Architecture: YOLOv8-Large pose estimation
Input: RGB images (any resolution)
Output: 56 keypoint coordinates with confidence scores
Average Performance:
- [email protected]: 97.48%
- [email protected]:0.95: 76.45%
- Precision: 96.21%
- Recall: 96.24%

Applications

Camera calibration and homography estimation
2D/3D scene mapping
Player tracking and analysis
Broadcast overlay generation
Game analytics and statistics
AR/VR applications

Model Performance

Performance tested across different hardware setups
13.64 FPS on Tesla T4 GPU
6.4 FPS on M3 MacBook Pro
Handles varying lighting conditions and occlusions

📩 For any questions regarding this project, or to discuss potential collaboration and joint research opportunities, please contact:

Mehdi Houshmand: [email protected]
Cise Midoglu: [email protected]
Pål Halvorsen: [email protected]