Spectrum-Enhanced Graph Attention Network for Garment Mesh Deformation.

Area of Science:

• Computer Graphics
• Computational Physics
• Machine Learning

Background:

• Existing mesh deformation methods require separate training for each garment/body type.
• Current approaches often fail to generate detailed folds and realistic dynamics.
• Physics-based simulations offer high quality but lack efficiency.

Purpose of the Study:

• To develop a unified model for efficient and high-quality mesh-based deformation simulation.
• To overcome limitations of existing methods in generalization and dynamic realism.
• To bridge theoretical analysis of neural networks and garment deformations.

Main Methods:

• Developed a spectrum-enhanced deformation network focusing on learning spectral information for intricate deformations.
• Introduced target-aware temporal skinning weights building on standard blend skinning.
• Weights dynamically adjust mesh vertex influence based on garment, body shape, and motion state.

Main Results:

• Achieved physics-based simulation quality with superior efficiency in a unified model.
• Demonstrated effective learning of spectral information within a specific frequency band for complex deformations.
• Validated generalization across diverse garments, bodies, and motions through ablation studies.

Conclusions:

• The proposed method significantly outperforms state-of-the-art techniques in generalization, deformation quality, and performance.
• The spectrum-enhanced network and target-aware temporal skinning weights are key innovations.
• This approach offers a more versatile and efficient solution for realistic mesh deformation simulation.