3D human pose estimation based on monocular RGB images and domain adaptation

Abstract

Human pose estimation in images is an important and challenging problem in Computer Vision. Currently, methods that employ deep learning techniques excel in the task of 2D human pose estimation. 2D poses can be used in a diverse and broad set of applications, of great relevance to society. However, the use of 3D poses can bring even more accurate and robust results. The objective of this research project is to improve the approaches being employed to estimate 3D poses in monocular RGB images. To this end, we intend to use, in the training phase of the methods, datasets of properly labeled, synthetically generated 3D human poses, and to use domain adaptation techniques that allow transferring information from the domain of synthetic images to the domains of images obtained from real environments. To date, no other work has been found in the literature that has investigated the use of deep domain adaptation methods to assist in the estimation of 3D poses in monocular RGB images obtained from real-world scenarios, from 2D poses obtained from synthetic scenarios, using adversarial neural networks. We hope to develop such a method in this work. (AU)