Human Motion Generation Based on GAN Toward Unsupervised 3D Human Pose Estimation

In this paper, we propose a method for generating joint angle sequences toward unsupervised 3D human pose estimation. Many researchers have proposed human pose estimation methods. So far, however, most methods have problems that require a large amount of images with supervised pose datasets to learn pose estimation models. Building such datasets is a time-consuming task. Thus, we aim to propose a method that can estimate 3D human poses without requiring training data with known poses. Toward this goal, we propose a GAN-based method for human motion generation and an optimization-based human pose estimation method. The proposed method consists of a generator that generates human pose sequence, a renderer that renders human images by changing 3D meshes based on the pose sequences generated, and a discriminator that discriminates between generated images and training data. Through an experiment based on simulated walking images, we confirmed that the proposed method can estimate the poses of body parts that are not occluded.