An Embeddable Implicit IUVD Representation for Part-Based 3D Human Surface Reconstruction.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Machine Learning

Background:

• Reconstructing 3D human surfaces from single images requires integrating pose, shape, and clothing details.
• Current methods combine parametric body models (e.g., SMPL) with neural implicit functions, but face computational redundancy and loss of body shape prior.

Purpose of the Study:

• To develop a novel representation for efficient and accurate 3D human surface reconstruction.
• To address computational inefficiencies and preserve body shape priors in existing methods.

Main Methods:

• Introduced the IUVD-Feedback representation, comprising an IUVD occupancy function and a feedback query algorithm.
• Replaced signed distance calculations with linear transformations in the IUVD space, utilizing SMPL UV maps.
• Implemented a feedback mechanism to reduce redundant query points and enhance feature extraction.

Main Results:

• Achieved a three-fold acceleration in the query-and-infer process compared to existing methods.
• Improved the robustness of 3D human surface reconstruction results.
• Demonstrated preservation of the underlying parametric body shape prior.

Conclusions:

• The IUVD-Feedback representation offers a more efficient and accurate approach to 3D human surface reconstruction.
• This method can be seamlessly integrated into existing implicit human reconstruction pipelines.
• The representation shows potential for generative applications due to its inherent semantic information.