CloTH-VTON: Clothing Three-dimensional reconstruction for Hybrid image-based Virtual Try-ON

Matiur Rahman Minar, Heejune Ahn; Proceedings of the Asian Conference on Computer Vision (ACCV), 2020

Virtual clothing try-on, transferring a clothing image onto a target person image, is drawing industrial and research attention. Both 2D image-based and 3D model-based methods proposed recently have their benefits and limitations. Whereas 3D model-based methods provide realistic deformations of the clothing, it needs a difficult 3D model construction process and cannot handle the non-clothing areas well. Image-based deep neural network methods are good at generating disclosed human parts, retaining the unchanged area, and blending image parts, but cannot handle large deformation of clothing. In this paper, we propose CloTH-VTON that utilizes the high-quality image synthesis of 2D image-based methods and the 3D model-based deformation to the target human pose. For this 2D and 3D combination, we propose a novel 3D cloth reconstruction method from a single 2D cloth image, leveraging a 3D human body model, and transfer to the shape and pose of the target person. Our cloth reconstruction method can be easily applied to diverse cloth categories. Our method produces final try-on output with naturally deformed clothing and preserving details in high resolution.