Point-Skeleton Attention for Efficient Dynamic Modeling of Animatable Humans with Loose Clothing

• 1、School of Artificial Intelligence, Beijing University of Posts and Telecommunications, Beijing 100876

Abstract：Addressing the issues of temporal instability and inconsistent deformation that often occur in loosely clothed areas of driveable digital humans in complex dynamic scenes, this paper studies a pose-driven human dynamic modeling method aimed at improving the appearance stability and reconstruction quality of digital humans under complex motions. To this end, a dynamic human modeling framework based on skeletal structure constraints is proposed. It structurally encodes human pose information and explicitly models the relationship between spatial points and the skeleton to enhance the characterization of pose-driven non-rigid deformations. Based on an efficient point-based representation, the method introduces a point-to-skeleton attention mechanism to achieve stable modeling of deformations in loosely clothed regions. Experimental results on the I3D-Human dataset show that the method achieves superior reconstruction performance compared to existing methods under both new viewpoints and new motion settings, particularly demonstrating better temporal consistency and structural stability in complex regions such as hems and cuffs. Further experimental analysis validates the effectiveness of the proposed modules in improving dynamic modeling quality. The results indicate that the method can effectively improve the stability of dynamic human reconstruction while maintaining computational efficiency, providing a feasible solution for practical applications of high-quality driveable digital humans.

Keywords： Computer graphics dynamic human modeling loose clothing modeling 3D Gaussian splatting