Implicit 3D human body reconstruction method from geometric diffusion of dense surface correspondences

To address the limitations of insufficient topological constraints and feature redundancy in existing high-resolution implicit geometric representation methods for 3D human body reconstruction, we propose an implicit reconstruction framework based on geometric diffusion of dense surface correspondences. The framework consists of three-stage. First, PoseCorr-Net nonlinearly maps clothed human scan point cloud into the canonical space, mitigating pose distortion and self-occlusion. Second, a bidirectional diffusion mechanism is introduced to enhance the hybrid point cloud representation, generating a balanced point set that preserves high-frequency surface details, maintains continuous internal structures, and reduces redundant sampling. In BodyOcc-Net, the part-aware feature aggregation module is embedded to infer occupancy field, enhancing the geometric representation of the human body. Finally, a SMPL-D-guided differentiable inverse mapping ensures topology-consistent reconstruction by transforming body point cloud from canonical space back to their the original pose space. Extensive quantitative and qualitative evaluations demonstrate that the proposed method achieves high-fidelity 3D human body reconstruction with strong computational efficiency.