A Study on Different Architectures on a 3D Garment Reconstruction Network

Abstract

In this work, we perform a study of different network architectures for 3D garment reconstruction. In the study, five networks are compared both quantitatively and qualitatively. The dataset used to train each network is CLOTH3D, which consists of six different garment types: tshirt, top, jumpsuit, dress, skirt, and trousers. The baseline architecture is introduced in the CLOTH3D paper. Each of the other four networks are defined by introducing modifications to this baseline: 1) applying central difference convolution (CDC) to GNNs, 2) new pooling based on spectral clustering, 3) applying octave convolution to GNNs, and 4) combining the CDC and pooling networks (CDC-pool). We show that the CDC and pooling networks independently outperform the baseline, octave, and CDC-pool networks quantitatively. As well as, the pooling network is best suited for modelling the complex dynamics present in the dress and skirt garment types, while there is little difference between the networks for the top, tshirt, trousers, and jumpsuit garment types.