Cloth and Skin Deformation with a Triangle Mesh Based Convolutional Neural Network

How to read this

Category

Method: a neural network architecture for mesh-defined data

Contributions

• Introduces a triangle-mesh-based CNN for problems where input and/or output live on a manifold triangle mesh, with or without boundary
• Applies it to cloth upsampling, restoring detail to PCA-compressed cloth, regressing clothing deformation from poses, and hand skin deformation from joint angles
• Trains on XPBD and offline FEM simulation data and reports large inference speedups over a GPU simulator (only for camera-visible meshes)

Context

Sits in the learning-based deformation lineage, related to Santesteban et al.'s 'Learning-Based Animation of Clothing for Virtual Try-On', generalizing such regressors via a mesh-native convolution operator.Builds on: Learning-Based Animation of Clothing for Virtual Try-On

Correctness

Assumes the mesh connectivity is fixed/manifold and that simulation-generated training data is representative; demonstrated across several deformation tasks with reported speedups, but accuracy is bounded by the training simulator and the visible-mesh-only inference is an approximation.

Clarity

Moderately technical; a first pass conveys the applications and speed claims, a second pass is needed for the mesh-convolution operator definition.

How to read it

Focus on how convolution is defined on the triangle mesh and which tasks it handles; second pass for the operator and data-generation pipeline if you plan to reuse the architecture.