Estimating garment texture from a single image

Abstract

As interest in AR/VR increases with the advent of telepresence and metaverse, 3D human restoration research for the development of 3D avatars is actively underway. In addition, 3D restoration research can be applied to the fashion e-commerce market, and in this regard, there is a 3D VT(=virtual try-on) solution. The multi-view capture/scan method used for 3D VT could not be extended to the real industry because it took time. However, with the advent of deep learning, it was possible to shorten the time as garments in 2D images could be quickly reconstructed to 3D. Research in this field has been steadily conducted, but the results of existing studies have limitations that they are unnatural because they depend heavily on input images. And the problem can be divided into ’unnatural wrinkles and shapes’ and ’blurry or distorted texture’. We aim to develop an improved 3D garment digitization solution by overcoming the limitations of existing 3D garment restoration techniques. First of all, problems with unnatural wrinkles or shapes can be solved by creating natural 3D garments in any pose using ’sewing patterns’, which is the development of actual garments. In the case of a problem in which a blur or distorted texture is generated, it can be improved by predicting the original patterns before being distorted and then filling it in the sewing pattern. In addition to this, we developed an automatic texture garment data generator, which we used to build our own dataset required for training and experiment.