Reconstruction of topologically consistent 3D fabric drape models

Abstract

In order to efficiently reconstruct 3D fabric drape models with consistent topology, a 3D fabric drape model dataset named Ddrape was created firstly.The dataset consists of topologically consistent 3D fabric drape models and their corresponding 2D binary images. Subsequently, a reconstruction model named Rec3FDNet, which incorporates a convolution module and a graph convolution model, was constructed and trained to reconstruct 3D fabric drape models with consistent topology based on a single binary image. During the training of Rec3FDNet, the loss function, the normalization of 3D fabric drape model, and the structure of the initial mesh of Rec3FDNet were explored. The reconstruction results were characterized using Hausdorff distance toolbox of Meshlab in Python. The results demonstrated that Rec3FDNet enables fast and accurate reconstruction of 3D fabric drape models with consistent topology. Under the optimal parameter combination, the average Hausdorff distance(error) of the reconstructed 3D fabric drape models is 1.752 mm. In addition to reconstructing results with a consistent topology, Rec3FDNet is also robust to random translation of the input image. This study provides a practical method for large-scale acquisition of 3D fabric drape models.