3D Human Pose Estimation by an Annealed Two-Stage Inference Method

This paper proposes a novel human motion capture method that locates human body joint position and reconstructs the human pose in 3D space from monocular images. We propose a two-stage framework including 2D and 3D probabilistic graphical models which can solve the occlusion problem for the estimation of human joint positions. The 2D and 3D models adopt directed acyclic structure to avoid error propagation of inference in the models. Both the 2D and 3D models utilize the Expectation Maximization algorithm to learn prior distributions of the models. An annealed Gibbs sampling method is proposed for the two-stage method to inference the maximum posteriori distributions of joint positions. The annealing process can efficiently explore the mode of distributions and find solutions in high-dimensional space. Experiments are conducted on the Human Eva dataset to show the effectiveness of the proposed method. The experimental data are image sequences of walking motion with a full 180° turn around a region, which causes occlusion of poses and loss of image observations. Experimental results show that the proposed two-stage approach can efficiently estimate more accurate human poses from monocular images.