Weakly‐Supervised Dynamic Point Cloud Saliency Prediction Network

ABSTRACT

Point cloud saliency prediction, which models Human Visual System (HVS) attention, is essential for applications like Computer Animation and eXtended Reality (XR). However, most existing research focuses on static scenes, and extending these models to dynamic point clouds is hindered by motion‐induced attention shifts and a lack of ground‐truth dynamic datasets. To address this, we propose the Weakly‐Supervised Dynamic Point Cloud Saliency Prediction Network (WDPS), one of the first frameworks to utilize classification supervision for dynamic 3D saliency. WDPS integrates saliency‐oriented feature mining, multiscale semantic‐detail fusion, and a temporal smoothness constraint to generate stable dynamic saliency maps. Extensive experiments demonstrate that our method is preferred by observers in over 78% of comparisons against static baselines. Furthermore, WDPS achieves superior Velocity‐Weighted Saliency (VWS) scores on highly dynamic sequences and significantly enhances downstream tasks performance.