RGB to 3D garment reconstruction using UV map representations

Abstract

Predicting the geometry of a 3D object from just a single image or viewpoint is an intrinsic human feature extremely challenging for machines. For years, in an attempt to solve this problem, different computer vision approaches and techniques have been investigated. One of the domains in which there has been more research has been the 3D reconstruction and modelling of human bodies. However, the greatest advances in this field have concentrated on the recovery of unclothed human bodies, ignoring garments. Garments are highly detailed, dynamic objects made up of particles that interact with each other and with other objects, making the task of reconstruction even more difficult. Therefore, having a lightweight 3D representation capable of modelling fine details is of great importance. This thesis presents a deep learning framework based on Generative Adversarial Networks (GANs) to reconstruct 3D garment models from a single RGB image. It has the peculiarity of using UV maps to represent 3D data, a lightweight representation capable of dealing with high-resolution details and wrinkles. With this model and kind of 3D representation, we achieve state-of-the-art results on CLOTH3D dataset, generating good quality and realistic reconstructions regardless of the garment topology, human pose, occlusions and lightning, and thus demonstrating the suitability of UV maps for 3D domains and tasks.