TransCycle: A Data Augmentation Method for 3D Human Pose Estimation

Abstract

The existing 3D human pose estimation solutions rely heavily on the human pose data collected by motion capture devices to train models. However, motion capture devices are expensive and their use is often limited to indoor environments, making the final human pose data unsatisfactory. Therefore, we propose a single-cycle 3D human pose model training scheme called TransCycle. This scheme includes a human motion generation adversarial network called Trans-PoseGAN, which allows a 3D human pose estimation model and a 3D human pose sequence generator learn together. These two components form a complementary cycle training process, and the generated data does not rely on any other motion capture devices. In experiments, our proposed method achieved better results than the original 3D human pose estimation method on HumanPose3.6M and MPI-INF-3DHP.