Garment sewing pattern estimation from a single image

Abstract

In this paper, we consider reconstructing a natural 3D garment according to any pose from a single human image. Previous works had limitations in that 3D garment and body can only be generated in a state being attached to each other, so when changing the pose, the garment depend on motions of the body. Also, they generated garment that corresponded only to the input pose. To address these problems, we propose generating a garment sewing pattern from a single image. By predicting sewing patterns instead of 3D garments and simulating the sewing patterns, we can generate 3D garments that can change with any pose. There are two processes to predict a garment sewing pattern from a single image. First, we use T-pose prediction network to transform the avatar and garment of an input image into T-pose that best represents the shape of the garment. Next, we use sewing pattern parameters network to extract garment sewing pattern parameters from a T-pose garment image. In addition, we construct the dataset of garment images in various poses and their corresponding garment sewing patterns. we visualize sewing patterns predicted through our framework and also conduct an ablation study on the effect of components of our framework.