Reconstructing Complex Shaped Clothing From a Single Image With Feature Stable Unsigned Distance Fields.

Area of Science:

• Computer Vision
• Computer Graphics
• 3D Reconstruction

Background:

• Traditional single-view clothing reconstruction methods use fixed templates, limiting shape diversity and contour complexity.
• Existing methods struggle with pose variations and occlusions, leading to simplified or incomplete reconstructions.

Purpose of the Study:

• To develop a novel method for reconstructing complex clothing shape contours and open clothing meshes from a single image.
• To improve pose robustness and generate diverse clothing shapes beyond the limitations of fixed templates.

Main Methods:

• Utilizing implicit unsigned distance fields based on clothing-oriented and pose-stable spatial features.
• Employing a type-generic clothing template derived from a mainstream generative model.
• Implementing a two-stage clothing mesh extraction method via point cloud representation for smooth and editable outputs.

Main Results:

• The proposed method successfully generates complex shape contours and open clothing meshes from single images.
• Achieved state-of-the-art performance in extensive experiments, demonstrating superior results.
• The approach provides spatially aligned clothing shape priors, enhancing pose robustness.

Conclusions:

• The novel method offers a simple, effective, and low-cost solution for reconstructing complex clothing shapes from single images.
• The technique overcomes limitations of fixed templates, enabling greater diversity and detail in reconstructed clothing.
• The developed dataset enhances supervision by increasing pose diversity and addressing occlusions.