Validation of Effective Active Learning Method in Semantic Learning of Context-dependent Motions

Abstract

Since the intelligent systems require a huge dataset of motion and label to recognize the meaning (label) of the body motion, we consider active learning in which the systems ask the label to users. We aim to realize an effective learning and question management method by considering the context in motion performance. In this paper, we use VR avatars that perform motions in different contexts, and define the context by tools and places used in the motion performance. Active learning was performed by combining each method concerning three points of context selection method, selection of Open/Close question, and label estimation method. We showed that the combination of margin sampling as context selection, naive Bayes as label estimation method, and performing open question at the beginning of the question and close question at latter term, is most efficient.