3D Clothing Wrinkle Prediction Based on GCN and Human Keypoint Constraints

This paper proposes a three-dimensional clothing wrinkle prediction framework that integrates dynamic human keypoint constraints with adaptive graph convolution to achieve high-fidelity end-to-end wrinkle generation. Considering that accurate three-dimensional geometric representation is essential for digital modeling, intelligent visualization, and geometry-aware simulation in advanced engineering applications, a pose-aware constraint field is introduced to explicitly guide deformation, while an adaptive graph convolution module dynamically adjusts adjacency weights according to local curvature and keypoint proximity and employs a gated attention mechanism for feature fusion. This design enables effective modeling of the coupling relationship between human posture and garment deformation while preserving geometric consistency and fine wrinkle details. Experimental evaluations on the CAPE dataset and a self-constructed MPG dataset demonstrate that the proposed method achieves a vertex error of 3.82 mm, a Chamfer distance of 4.04 mm, a penetration rate of 0.71%, and a 15.8% improvement in local curvature standard deviation compared with ClothSim. The framework significantly enhances geometric fidelity, reduces non-physical artifacts, and improves wrinkle detail representation, providing an effective graph-based modeling strategy for highprecision three-dimensional digital reconstruction and geometry-aware computational applications relevant to intelligent electromagnetic and virtual simulation environments.