Automatic annotation of dance videos based on foot postures

Abstract

Along with the advancements in the field of artificial intelligence, machine learning, and deep learning video annotation has become a more exciting and crucial problem, especially the video from medical, arts performance games and so on. In India, there is an increased demand for digitizing and archiving of arts and culture . This paper focuses on annotating dance videos based on foot postures (stanas) in an automatic manner. Using transfer-learning, the features from the images are extracted and a deep neural network is used for image classification. A trained model of Deep Stana Classifier is used for identifying stanas from the frames of a video. A complete annotation system comprises of a coupled architecture, one Deep Stana Classifier module, and an annotation module. The findings on the accuracy obtained for the Deep Stana Classifier produced positive results. In the second phase, the result of the annotations made on the video is kept as an index structure in JSON object format. The scope and opportunity of this work in the future is consistent and useful for annotating dance videos as well as videos from another domain.