GeoCoHuman: Single-view 3D Clothed Human Reconstruction via Complementary Geometry Priors

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Single-image 3D clothed human reconstruction is complex, particularly for intricate poses and loose apparel.
• Current methods struggle with insufficient 3D geometry priors and limited exploitation capabilities, leading to suboptimal results.

Purpose of the Study:

• To introduce GeoCoHuman, a novel method for high-quality single-image 3D clothed human reconstruction.
• To leverage complementary geometry priors for improved accuracy and detail in human shape and clothing representation.

Main Methods:

• Utilized a Parametric Model-guided Diffusion Model (PMDM) to generate depth maps, projecting them into an oriented point cloud for clothing geometry.
• Employed Proximity Feature Aggregation (PFA) to extract aligned features from the point cloud and a parametric SMPL mesh for body shape regularization.
• Developed a Collaboration Perception Implicit Function (CPIF) with multiple prediction heads for enhanced feature utilization and fine-grained reconstruction.

Main Results:

• GeoCoHuman successfully captures unconstrained clothing topology and regularizes human shape effectively.
• The method demonstrates robust and consistent feature extraction, preserving local structures.
• Extensive experiments validated the significant superiority of GeoCoHuman over existing state-of-the-art methods on benchmark datasets and real-world images.

Conclusions:

• GeoCoHuman achieves state-of-the-art performance in single-image 3D clothed human reconstruction.
• The proposed approach effectively integrates multiple geometry priors for detailed and accurate human modeling.
• This work advances the field by enabling high-quality reconstruction from complex poses and loose clothing.