Visually Similar K-poselets Based Human Pose Recognition

In the paper, we propose the visually similar k-poselets to recognize human poses (e.g., stoop, squat) in still images. Compared with the original k-poselets that are collected according to similar keypoints configurations, we further introduce appearance similarity constraints to generate visually similar k-poselets. The number of selected visually similar k-poselets for each pose category is iteratively decreased based on discriminative criterion. The pose dictionary, constructed with learned visually similar and discriminative k-poselets of different poses, is applied in pose recognition. The experimental results on our released human pose database verify the effectiveness of the proposed visually similar k-poselets based pose recognition method.