Torso orientation: A new clue for occlusion-aware human pose estimation

Self-occlusion is a challenging problem existing in human pose estimation. In this paper we exploit a new cue to solve this problem: the torso orientation. We describe a technique to automatically detect self-occlusion in training set without visibility label. Given this prior information, we are able to jointly learn an occlusion-aware model to capture the pattern of self-occluded body parts. We evaluate our model on two major datasets, which are both publicly available. The experiment result shows that our model is quite competitive in both of the datasets with the state-of-the-arts. By this way, we illustrate our model's robustness to the self-occlusion problem in human pose estimation.