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

Area of Science:

• Computer Vision
• 3D Reconstruction
• Machine Learning

Background:

• Single-image 3D clothed human reconstruction is difficult, especially with complex poses and loose garments.
• Current methods lack sufficient 3D geometry priors and effective prior exploitation.

Purpose of the Study:

• To propose GeoCoHuman, a novel method for high-quality 3D clothed human reconstruction from single images.
• To leverage complementary geometry priors for enhanced reconstruction accuracy and detail.

Main Methods:

• Utilizes an oriented point cloud as a clothes geometry prior for unconstrained topology.
• Employs the parametric SMPL mesh as a body geometry prior for shape regularization.
• Introduces a Parametric Model-guided Diffusion Model (PMDM) for depth map inference.
• Develops Proximity Feature Aggregation (PFA) for robust feature extraction.
• Implements a Collaboration Perception Implicit Function (CPIF) for fine-grained reconstruction.

Main Results:

• GeoCoHuman achieves superior performance in 3D clothed human reconstruction compared to state-of-the-art methods.
• The method demonstrates significant improvements on benchmark datasets and in-the-wild images.
• Reconstructions exhibit enhanced detail and accuracy, particularly for challenging cases.

Conclusions:

• GeoCoHuman effectively integrates multiple geometry priors for superior 3D human reconstruction.
• The proposed approach advances the state-of-the-art in single-image clothed human modeling.
• This work offers a robust solution for complex human shape and clothing reconstruction.