Sarah Fan¹, Kevin Guo² and Yu Sun³, ¹USA, ²University of Southern California, USA, ³California State Polytechnic University, USA

Human Pose Estimation has proven versatility in improving real-world applications in healthcare, sports, etc. [1]. Proper stance, form and movement is instrumental to succeeding in these activities. This paper will explain the research process behind the deep learning mobile ballet app, LaBelle [2]. LaBelle takes in two short videos: one of a teacher, and one of a student. Utilizing MediaPipe Pose to identify, analyze, and store data about the poses and movements of both dancers, the app calculates the angles created between different joints and major body parts. The app’s AI Model uses a K-means clustering algorithm to create a group of clusters for both the student dataset and the teacher dataset [3]. Using the two sets of clusters, LaBelle identifies the key frames in the student-video and searches the teacher cluster set for a matching set of properties and frames. It evaluates the differences between the paired frames and produces a final score as well as feedback on the poses that need improving. We propose an unsupervised guided-learning approach with improved efficiency in video comparison, which is usually both time and resource consuming. This efficient model can be used not just in dance, but athletics and medicine (physical therapy like activities) as well, where stance, form, and movements are often hard to track with the naked eye.

Artificial Intelligence, Machine Learning, Deep Learning, K-means Clustering, Computer Vision, Ballet.