Geometry image super-resolution with AnisoCBConvNet architecture for efficient cloth modeling.

Area of Science:

• Computer Graphics
• Artificial Intelligence
• Computational Physics

Background:

• Cloth simulation is crucial for realistic computer graphics (CG) and visual effects (VFX).
• Existing methods for high-resolution cloth simulation can be computationally expensive and prone to oscillations or noise.
• Generating high-quality meshes from low-resolution data presents a significant challenge.

Purpose of the Study:

• To propose a novel convolutional neural network, Anisotropic Constrained-Boundary Convolutional Neural Network (AnisoCBConvNet), for stable super-resolution of cloth meshes.
• To improve the quality and detail of cloth meshes generated from low-resolution simulation data.
• To alleviate surface wrinkling and noise issues common in existing geometry image-based approaches.

Main Methods:

• Developed AnisoCBConvNet, applying super-resolution to low-resolution cloth meshes.
• Trained the network using pairs of low-resolution (LR) and high-resolution (HR) cloth simulation data, converted to 2D geometry images.
• Employed anisotropic weight control near boundaries to prevent oscillations and surface wrinkling.

Main Results:

• AnisoCBConvNet successfully converts LR geometry images to HR geometry images, producing detailed cloth meshes.
• The method effectively alleviates surface wrinkling and noise, outperforming existing geometry image approaches.
• Stable and efficient performance was confirmed across various cloth simulation scenarios.

Conclusions:

• AnisoCBConvNet provides a stable and efficient solution for generating high-quality cloth meshes via super-resolution.
• The proposed method enhances the realism of CG VFX in games and movies by improving cloth simulation fidelity.
• This approach offers a significant advancement in creating detailed and artifact-free cloth dynamics.