Skip to main navigation menu Skip to main content Skip to site footer

Spatial Refinement for 3D Human Mesh Recovery in Ice Hockey Broadcast Videos

Abstract

Incorporating 3D Human Mesh Recovery (HMR) is crucial for understanding player actions and game scenarios. However, challenges remain, including vertical drift along the Z-axis and the lack of a mechanism to align frame-based X--Y positions with a top-down rink template. In this work, we address these issues by introducing spatial refinement strategies within the HMR pipeline. First, we reduce vertical drift by estimating a temporally consistent ground plane and correcting each reconstructed mesh's Z-axis trajectory. Second, we link 3D predictions to tactical analysis by projecting players onto a standardized top-down rink template via homography estimation. Integrated into the GVHMR framework with a hockey-specific tracking model, our refined pipeline yields more stable and physically plausible meshes and enables accurate reconstruction of player trajectories. Experiments on NHL and AHL broadcast videos demonstrate substantial improvements in temporal consistency and spatial alignment, facilitating reliable downstream analysis of player behavior and game strategy.
PDF