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

We propose an anisotropic constrained-boundary convolutional neural networks (hereafter, AnisoCBConvNet) that can stably express high-quality meshes without oscillation by applying super-resolution operations to low-resolution cloth meshes. As a training set for the neural network, we use a pair between simulation data of low resolution (LR) cloth and data obtained by applying the same simulation to high resolution (HR) cloth with increased quad mesh resolution of LR cloth. The actual data used for training are 2D geometry images converted from 3D meshes. The proposed AnisoCBConvNet is used to train an image synthesizer that converts LR geometry images to HR geometry images. In particular, by controlling the weights anisotropically near the boundary, the problem of surface wrinkling caused by oscillation is alleviated. When the HR geometry image obtained through AnisoCBConvNet is converted back to the HR cloth mesh, details including wrinkles are expressed better than the input cloth mesh. In addition, our results improved the noise problem in the existing geometry image approach. We tested AnisoCBConvNet-based super-resolution in various simulation scenarios, and confirmed stable and efficient performance in most of the results. By using our method, it will be possible to effectively produce CG VFX created using high-quality cloth simulation in games and movies.