A Two-Stage Learning Approach for Goalie, Net and Stick Pose Estimation in Ice Hockey

Abstract

Accurate pose estimation of ice hockey goaltenders presents a unique challenge due to the dynamic nature of the sport and the intricate interactions among the goalie, equipment, and net. This study introduces a comprehensive investigation into goalie pose estimation using both One-Stage and Two-Stage Learning GoalieNet architectures. The One-Stage Learning GoalieNet predicts all keypoints simultaneously, while the Two-Stage Learning GoalieNet employs a Keypoint Predictor Network (KPN) to predict 26 out of 29 keypoints and a Keyheatmap Fusion Network (KFN) to predict 3 stick-related keypoints. Evaluation on a NHL dataset underscores the effectiveness of both approaches in accurately predicting keypoints. Results on the test data reveal a median percentage of detected keypoints of 71% for the Two-Stage approach and 70% for the One-Stage approach, along with normalized localization errors on detected keypoints of 0.0187 for the Two-Stage and 0.0194 for the One-Stage approach. This work introduces the first-ever goalie pose estimation technique designed specifically for ice hockey, accompanied by a thorough analysis of the obtained results.