PICA: Physics-Integrated Clothed Avatar.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Existing neural rendering methods often use a single model for clothed humans, leading to artifacts in complex garment dynamics.
• Data-driven approaches for garment animation struggle with out-of-distribution poses and lack physical accuracy.

Purpose of the Study:

• To introduce PICA, a novel representation for high-fidelity animatable clothed human avatars.
• To achieve physics-plausible dynamics for clothing, especially loose garments.
• To overcome limitations of single-model representations and purely data-driven dynamics.

Main Methods:

• Employing two separate 2D Gaussian Splatting (2DGS) models for body and clothing with distinct deformation characteristics.
• Integrating a graph neural network (GNN)-based module for physics-accurate clothing simulation.
• Developing carefully designed features to enhance representation and dynamics.

Main Results:

• Achieved high-fidelity rendering of clothed human bodies in complex poses.
• Demonstrated superior performance in representing garment dynamics compared to previous methods.
• Enabled physics-plausible animations for loose and sliding clothing.

Conclusions:

• PICA offers a significant advancement in creating realistic and dynamically accurate clothed human avatars.
• Separating body and clothing representations with integrated physics simulation is key to improved results.
• The method shows promise for applications requiring high-fidelity human avatar animation.