Dance Self-learning Application and Its Dance Pose Evaluations

In this study, we propose a variety of dance pose evaluation methods required for a smartphone-based application that helps dance self-learning. Previously, various dance-pose-evaluation methods and dance-learning applications have been proposed for the same purpose, but some problems were identified in the actual application process, such as the absence of joints due to occlusion, the overfitting of the torso, and the evaluation of rotational poses. To mitigate these problems, this study proposes three methods to a) solve the view occlusion problem that precludes evaluation by creating a state variable referred to as visibility, b) solve the torso overfitting problem by creating a keypoint referred to as the core, and c) estimate the rotation of the body based on a vector obtained between the core and torso joints. Furthermore, we conducted experiments using a dance self-learning application running on a real smartphone and various real-world datasets. Based on the experimental results, we demonstrate that the newly proposed method can provide effective feedback to students who try to learn to dance alone, by performing robust dance pose evaluations in different scenarios.