Joint relation based human pose estimation

With the increasing application of computer vision technology in real life, human pose estimation task becomes more and more important. However, inferencing accurate coordinates of limb joints or invisible joints is still difficult for even state-of-the-art approaches. The positions of limb joints are diversified, and a percentage of the joints are occluded. In this paper, we aim to solve such problem by proposing joint relation based human pose estimation framework. Joint relation is the spatial relation between selected neighbor joints which can imitate human body structure and localize a complex joint with the help of its neighbor joint. We evaluate the joint relation module on challenging dataset and demonstrate its effectiveness by accuracy and visualization results. The proposed joint relation based human pose estimation method achieves state-of-the-art performance.