Point2Vertex: Human Mesh Reconstruction With Hybrid Position Encoding and Dual Refinement From Point Clouds

Abstract

Human mesh reconstruction from point clouds has gained increasing attention in recent years. However, existing reconstruction pipelines predominantly focus on regressing human model parameters, which results in an indirect mapping between input observations and output meshes. In this letter, we introduce a point-to-vertex reconstruction method, which bypasses parameter spaces to reconstruct complete SMPL topology from partial point observations directly. Specifically, our framework employs a hierarchical approach to extract geometric information from the partial point cloud. The feature mapping process is realized via a transformer-based architecture, which integrates parametric human model priors and employs hybrid position encoding. To further improve accuracy, the dual refinement module progressively refines the reconstruction process through sequential optimization in the coordinate and vertex-level feature spaces. Experiments on SURREAL datasets demonstrate that our framework achieves state-of-the-art performance, surpassing previous methods by 18.4% in MPVE (from 19.6 to 16.0 mm) and 10.2% in MPJPE (from 17.7 to 15.9 mm).