On the Utility of Pose Estimation Models for Golf Swing Understanding

Human pose estimation has shown increasing potential in sports analytics, particularly for evaluating and improving athletic motion. However, existing models are typically optimized for general-purpose or stationary scenarios and struggle when applied to high-speed, occlusion-prone sports such as golf. This study investigates the performance of two state-of-the-art pose estimation models—YOLO Pose and MediaPipe Pose—in analyzing golf swings from video data. A custom dataset was developed, consisting of golf swing recordings across diverse players, backgrounds, and lighting conditions. Each video was segmented into frames, and a subset was manually annotated to create ground-truth keypoints for evaluation. Model performance was assessed using the Object Keypoint Similarity (OKS) metric. Results show that while MediaPipe Pose achieved higher average accuracy (mean OKS = 0.636) compared to YOLO Pose (mean OKS = 0.604), YOLO demonstrated more consistent predictions with lower variance. Qualitative analysis further revealed that MediaPipe better handles partial occlusions but is more sensitive to environmental factors. These findings highlight trade-offs between model precision, consistency, and robustness in dynamic sports contexts, suggesting the need for domain-specific adaptations to improve accuracy in golf swing analysis. Such insights further underscore how pose-based motion understanding can serve as a foundation for developing intelligent feedback systems, bridging the gap between traditional coaching and automated performance analytics.