LayerNet: High-Resolution Semantic 3D Reconstruction of Clothed People.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Existing 3D reconstruction methods often require garment-specific models.
• Representing diverse clothing topologies and body shapes in 3D remains challenging.

Purpose of the Study:

• Introduce SMPLicit, a generative model for unified body, pose, and clothing geometry representation.
• Develop LayerNet, a deep network for simultaneous 3D reconstruction of body and clothes from single images.
• Enable garment-agnostic 3D reconstruction and facilitate 3D virtual try-on.

Main Methods:

• SMPLicit unifies representation of various garment topologies, sizes, and tightness.
• LayerNet employs a coarse-to-fine strategy, predicting smooth cloth geometry then refining details with an image-guided network.
• The approach recovers high-frequency details like wrinkles and fits challenging body poses.

Main Results:

• LayerNet achieves competitive accuracy against state-of-the-art garment-agnostic methods in controlled settings.
• The method surpasses existing approaches in challenging poses and uncontrolled environments.
• Demonstrates high-fidelity 3D reconstruction of clothed humans.

Conclusions:

• SMPLicit and LayerNet offer a unified and robust solution for 3D clothed human reconstruction.
• The semantically rich 3D outputs are suitable for direct 3D virtual try-on.
• Advances garment-agnostic 3D reconstruction for diverse scenarios.