Sports analytics increasingly relies on machine learning models to understand and predict player behavior, strategy, and outcomes. In team sports like rugby, pose estimation models can extract player keypoints (e.g., body joint locations) from match footage or sensor data, enabling automated performance analysis.
However, despite recent progress in pose-based classification tasks (e.g., identifying tackles, passes, or formations), the models’ decision-making processes often remain opaque. Understanding which body keypoints or movements are most salient for a given prediction could offer coaches, analysts, and sport scientists valuable insights — and increase trust in automated analytics systems.
This project aims to apply similar techniques to sports data. Specifically, it will use feature attribution methods to evaluate the relative importance of keypoints in classifying rugby-specific actions.