Perceiving heavily occluded human poses by assigning unbiased score

The problem of human pose estimation has been largely solved by the prevailing Deep Convolutional Neural Networks (DCNNs). However, heavily occluded human poses still represent great challenges. In this paper, we propose a new scoring method to perceive the detection of heavily occluded poses unbiasedly. The typical way of assigning scores to detected poses is to use the mean confidence of each joint. This makes poses with occlusion suppressed durng evaluation since invisible joints may have a much lower confidence than visible joints. We address this by identifying the visibility of each joint, an occlusion aware network is designed to predict both heatmaps and visible values of joints simultaneously. Thus, the degree of occlusion of a pose can be grasped, and a much fairer score is able to be set. Furthermore, a KS-net is proposed to predict the KS (Keypoint Similarity) between each estimated joint location and its matched ground-truth. The predicted KS calibrates localization accuracy better than the maximum heat value in heatmap. Pose score is calculated using the predicted visibility and KS value of each joint. The efficacy of our method is demonstrated on the most widely used MS-COCO pose dataset. Extensive experiments show that using our scoring approach can significantly improve the average precision of heavily occluded poses for the provided detections. (C) 2020 Elsevier Inc. All rights reserved.