we are
Metric
Hive
A production-ready tracking system built for the toughest conditions in computer vision: high-speed play, heavy motion blur, and aggressive camera pans.
Detection & Recognition
Using a fine-tuned RF-DETR (State-of-the-Art object detection), the system reliably detects players, referees, and jersey number regions even during fast transitions. We implemented a hybrid ORB/ECC alignment pipeline to stabilize the feed, ensuring the tracker focuses on player movement rather than camera shake.
For identity verification, we deployed a custom PARSeq model trained specifically on ice hockey fonts. Combined with confidence-weighted voting across frames, the system achieves >99% recognition accuracy, identifying players even when individual frames are blurred or unreadable.
Sports Analytics, Computer Vision
30th September 2025
Stability in Chaos
To handle the extreme occlusion of hockey (players piling up), we moved beyond simple bounding boxes. The system uses SAM (Segment Anything Model) for initial masks and CUTIE for temporal propagation. This allows us to track player silhouettes, preventing identity switches during high-speed crossovers.
This is a production-ready pipeline, not just a demo. It solves the specific "Re-Identification" challenges of sports broadcasting, providing stable IDs for automated match analysis, fan engagement overlays, and performance insights without manual correction.
(Project)
