Pose Knowledge Transfer for multi-person pose estimation

Multi-person pose estimation is a fundamental yet challenging research topic for many computer vision applications. In this paper, to relieve the problem of variable pose structure and occlusion in complex scenes, we propose a novel Pose Knowledge Transfer approach for multi-person pose estimation, which doesn’t take into account deeper and wider network structure design. This approach contains a Keypoint Region Erasing (KRE) scheme and a Bi-directional Pose Knowledge Transfer (BPKT) model learning strategy for this task. The KRE encourages human pose estimator to explicitly focus learning on keypoints connectivity to robustly localize the occluded keypoint via its adjacent visible body patches. Specifically, without additional model parameters involvement, the BPKT effectively transfers the local connectivity knowledge and the global body configuration knowledge between two network with the same structure to encounter variable pose structure. Extensive experiments demonstrate that the BPKT and the KRE significantly improve the performance of a range of state-of-the-art human pose estimation models, consistently validating the effectiveness and generalization property of our model-agnostic approach on the MPII human pose dataset and the COCO keypoint benchmark.