Kinematic Priors Benefit Skeleton-Based Action Recognition

Skeleton-based action recognition has long been a fundamental and compelling problem in computer vision. The key challenge lies in modeling human motion while faithfully respecting its physical laws. Existing methods often treat joints as free 3D points, allowing anatomically implausible movements to distort feature learning. Moreover, conventional time and frequency approaches struggle to capture cross-limb phase relations, which are essential for distinguishing actions with similar motion frequencies. To address these issues, we propose to explicitly encode anatomical constraints by modeling their skeletons with a DoF prior. Additionally, a phase-coupling prior is developed to capture cross-limb coordination through relative phase relations. Extensive experiments demonstrate that our approach achieves state-of-the-art performance on three benchmark datasets.