3D Dynamic wrinkling and deformation inference model of knitted garments integrating Graph Neural networks and node topological constraints

To address the obstacles in inferring the dynamics of knitted garments in 3D virtual try-on, namely the high computational latency caused by traditional physical iteration and the mesh clipping problem caused by the defects of existing graph networks due to unconstrained topological regression and soft physical penalties, this paper proposes a Topologically Constrained Spatio-Temporal Graph Convolutional Network (TC-STGCN). Methodologically, the model first encodes the spatio-temporal tensor of the human skeleton into global action features. Secondly, the core construction is a spatio-temporal graph convolution guided by local structure tensors, aggregating graph features along the anisotropic force directions of the fabric to infer coordinate offsets. Finally, a Bounding Volume Hierarchy (BVH) and a normal repulsion gradient field are introduced at the prediction end as a post-inference physical verification module that applies hard truncation penalties to nodes that exceed physical boundaries. Compared to unregistered, traditional ICP, and leading-edge baselines, TC-STGCN reduces vertex RMSE from 3.12 ± 0.54 mm to 1.68 ± 0.25 mm and MAE to 1.32 ± 0.18 mm (R2>0.95, p< 0.01) in the extreme large-scale jumping action test on the custom Knit-Subset-HD dataset, and keeps the penetration error rate below 1% within a high-frame time window with a 2.0 mm tolerance. Spectral analysis shows that this physical constraint mechanism alleviates high-frequency energy dissipation and achieves physical boundary mapping of micro-wrinkles while maintaining computational efficiency.