Multi-View 3D Human Pose Estimation with Self-Supervised Learning

Modern 3D human pose estimation builds on a deep learning network, requiring expensive amounts of training data that contain pairs of 2D and 3D pose annotations. In this paper, we propose a self-supervised 3D human pose estimation without 3D annotations. Instead, we exploit multi-view images and camera parameters to make the network learn 3D human pose based on geometric consistency. The merit of the proposed method is validated via experiments.