TurbidPose: Environment-Adaptive Underwater Pose Estimation Across Turbid and Diverse Conditions

Abstract

Underwater target pose estimation is vital for robot operations in aquatic environments. However, it remains severely challenged by water turbidity and environmental variability. In turbid water, suspended particles significantly degrade visual quality, while across regions, variations in turbidity, color, and illumination induce distribution shifts that hinder transferability. To address these issues, we propose TurbidPose, an environment-adaptive pose estimation pipeline that incorporates diffusion-based style injection. By injecting controllable target water environment style into synthetic training images, we generate environment-aware data that better reflects real deployment conditions. Furthermore, we introduce a training-time disturbance-resilient module based on a pre-trained variational autoencoder (VAE). Through feature alignment between the VAE and the pose estimator, the model is encouraged to capture environment-disturbance-resistant representations. The VAE is discarded during inference, introducing negligible extra cost. Comprehensive experiments on multiple underwater pose estimation benchmarks demonstrate that TurbidPose significantly outperforms existing methods in both accuracy and robustness under diverse and turbid conditions.