GarNet++: Improving Fast and Accurate Static 3D Cloth Draping by Curvature Loss.

Area of Science:

• Computer Graphics
• Virtual Reality
• Machine Learning

Background:

• 3D cloth draping on virtual humans is computationally intensive.
• Existing methods struggle with detail and collision detection.

Purpose of the Study:

• To develop an efficient deep learning model for static 3D cloth draping.
• To achieve visually plausible and detailed garment simulation.
• To reduce computational cost compared to physics-based methods.

Main Methods:

• A two-stream deep network extracting features from body and garment shapes.
• Physics-based simulation (PBS) inspired loss terms for plausibility and collision awareness.
• Novel curvature loss functions, including a detail-preserving loss using local covariance matrices and Rayleigh quotients.

Main Results:

• The model mimics PBS with two orders of magnitude less computation time.
• New curvature loss functions enhance garment detail and outperform mean curvature normal loss.
• Framework validated on diverse garment types, body shapes, and poses.

Conclusions:

• The proposed deep network offers a computationally efficient and effective solution for 3D cloth draping.
• The novel curvature loss significantly improves the detail and realism of draped garments.
• This approach surpasses existing data-driven methods in performance.