Real-time dance evaluation by markerless human pose estimation

This paper presents a unified framework that evaluates dance performance by markerless estimation of human poses. Dance involves complicated poses such as full-body rotation and self-occlusion, so we first develop a human pose estimation method that is invariant to these factors. The method uses ridge data and data pruning. Then we propose a metric to quantify the similarity (i.e., timing and accuracy) between two dance sequences. To validate the proposed dance evaluation method, we conducted several experiments to evaluate pose estimation and dance performance on the benchmark dataset EVAL, SMMC-10 and a large K-Pop dance database, respectively. The proposed methods achieved pose estimation accuracy of 0.9358 mAP, average pose error of 3.88 cm, and 98% concordance with experts’ evaluation of dance performance.