Abstract
Single-image 3D clothed human reconstruction is a challenging task, especially for complex poses and loose clothing. Existing methods suffer from suboptimal reconstruction due to insufficient 3D geometry prior and limited prior exploitation capability. In this paper, we propose GeoCoHuman, which employs two complementary geometry priors for high-quality human reconstruction. Its core is to leverage the constructed oriented point cloud as the clothes geometry prior to capture unconstrained topology, and the parametric SMPL mesh as the body geometry prior to regularize human shape. Specifically, we first propose a Parametric Model-guided Diffusion Model (PMDM) to infer front and back depth maps. We project the inferred depth maps into 3D space to construct the oriented point cloud, which presents dense 3D points with corresponding normals, covering the complete clothes surface. Then, we propose a Proximity Feature Aggregation (PFA) to extract geometry-aligned features from the constructed oriented point cloud and the parametric SMPL mesh. PFA aggregates multi-distance neighborhood information to preserve local structures, resulting in robust and consistent feature extraction. Finally, we propose a Collaboration Perception Implicit Function (CPIF), which applies multiple prediction heads to focus on learning different geometry feature domains and collaboratively enable the final prediction. CPIF can enhance the utilization of prior features for occupancy-related property prediction, achieving fine-grained human reconstruction. Extensive experiments on two benchmarks and in-the-wild images demonstrate the significant superiority of GeoCoHuman over state-of-the-art methods.