Single-View Clothed Human Reconstruction With Multi-View Consistency Representation.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Machine Learning

Background:

• Single-view clothed human reconstruction faces challenges with depth ambiguity and missing invisible information.
• Existing PIFu-like frameworks rely on pixel-aligned features, leading to representation inaccuracies.

Purpose of the Study:

• To address depth ambiguity and the lack of invisible information in single-view clothed human reconstruction.
• To enhance the accuracy and detail of 3D human models generated from single images.

Main Methods:

• Proposed a depth-guided pixel-aligned feature to filter features based on depth discrepancies.
• Introduced multi-view consistency prior using contrastive learning and masked autoencoder for SMPL vertex features.
• Constrained representation learning to focus on visible parts and infer invisible information.

Main Results:

• The proposed method significantly improves feature representation for surface details.
• Achieved more reasonable and accurate representation learning in single-view reconstruction.
• Demonstrated superior performance compared to previous implicit representation-based methods on public and commercial datasets.

Conclusions:

• The novel approach effectively tackles key limitations in single-view clothed human reconstruction.
• The method provides a more robust and accurate way to generate detailed 3D human models.
• This work advances the state-of-the-art in implicit 3D human shape and appearance reconstruction.