A Comparison of the Instructor-Trainee Dance Dataset Using Cosine similarity, Euclidean distance, and Angular difference

Abstract

The COVID-19 outbreak has restricted most outdoor activities, leads to increasing interest in exercise at home with online trainers. One issue of online exercise technology is the safety since improper motion might result in injury. As a basis to prevent improper motion, methods for evaluating the motion similarity between an instructor and a trainee are essential. Cosine similarity, Angular difference, and Euclidean distance are three general ways for the motion evaluation. This study aimed to determine the most effective way for analyzing the similarity of human motion on the dataset of instructor-led dances. We first experimented with the data to find the appropriate cut-off value for classifying posture into two classes based on the similarity score. Confusion matrix, precision, recall, F1-score, accuracy of the results were then used to compare the efficiency. We discovered that Cosine similarity had the highest accuracy, 82.77 percent at cut-off 93.