MultiGO++: Monocular 3D Clothed Human Reconstruction via Geometry-Texture Collaboration.

Area of Science:

• Computer Vision
• Computer Graphics
• 3D Reconstruction

Background:

• Monocular 3D clothed human reconstruction aims to create realistic 3D avatars from single images.
• Current methods often rely on multi-view supervision and struggle with data scarcity, inaccurate priors, and biased supervision, leading to suboptimal results.

Purpose of the Study:

• To propose a novel framework, MultiGO++, for effective geometry-texture collaboration in monocular 3D clothed human reconstruction.
• To address limitations in texture quality, geometric accuracy, and systematic supervision.

Main Methods:

• Developed a multi-source texture synthesis strategy creating over 15,000 3D scans for improved texture estimation.
• Introduced a region-aware shape extraction module and a Fourier geometry encoder to bridge the modality gap for geometry learning.
• Utilized a dual reconstruction U-Net for collaborative feature refinement to generate high-fidelity textured 3D meshes.

Main Results:

• The MultiGO++ framework demonstrates superior performance in generating high-fidelity textured 3D human meshes.
• Experiments on benchmarks and in-the-wild cases show significant improvements over state-of-the-art methods.
• The method effectively addresses challenges in texture quality and geometric accuracy.

Conclusions:

• MultiGO++ offers a robust solution for monocular 3D clothed human reconstruction by enabling systematic geometry-texture collaboration.
• The proposed techniques significantly enhance the realism and accuracy of generated 3D avatars.
• This work advances the field by overcoming key limitations in existing reconstruction approaches.