UW-HSPPE: Leveraging Human Skeleton Priors for Robust Underwater Pose Estimation

Human pose estimation has achieved remarkable progress in recent years, enabling various real-world applications. However, extending these achievements to underwater scenarios remains challenging yet crucial for applications such as aquatic sports training and underwater rescue. The unique characteristics of underwater environments—particularly semi-transparent occlusions caused by water splashes and dynamic surface in-terference—severely impair the performance of traditional pose estimation methods. This paper proposes UW-HSPPE (Underwater Human Skeleton Prior Pose Estimation), a lightweight generative framework that enhances underwater pose estimation by incorporating human structural priors. Our method takes both keypoint positions and their confidence scores from conventional pose detectors as input, and leverages a Conditional Variational Autoencoder (CVAE) architecture to refine these predictions. The core of our approach is the Certainty-aware Keypoint Prior Refinement (CKPR) module, which encodes human structural constraints in the latent space and employs an innovative nonlinear reparameterization technique to adaptively adjust predictions based on their uncertainty levels. A certainty-aware adaptive fusion mechanism then optimally combines the original predictions with the structure-aware refinements, ensuring both local detection accuracy and global pose plausibility. To address the scarcity of underwater pose data, we introduce UW-Human, a comprehensive dataset encompassing diverse underwater scenarios and human poses. Extensive experiments demonstrate that our approach significantly improves the robustness of state-of-the-art pose estimators while maintaining real-time performance. And as a plug-and-play module, UW-HSPPE shows consistent improvements across different evaluation metrics on both general test cases and underwater-specific scenarios.