An implementation of soccer match game state reconstruction, inspired by the SoccerNet GSR Challenge. Here is a video introducing this implementation of the project: Video Introduction.
This project uses a multimodal architecture to track key objects (players, referees, ball, pitch keypoints) and perform various annotations and visualizations. Functionality comprises of:
- object detection: YOLOv5-detect
- tracking: roboflow supervision
- team assignment: k-means clustering
- jersey number detection: MMOCR
- pitch keypoint detection: YOLOv5-pose; TVCalib
- pitch localization: homography transformations
TODO
- object detection:
- pitch localization: position smoothing
- visualizations: player heatmaps, speed estimators

