Automatic Bharatanatyam Dance Video Annotation Tool Using CNN

Dance video analysis and interpretation have been challenging tasks in computer vision due to the lack of availability of annotated data. We may get several videos available in the public domain but hardly annotated because it incurs much time, cost, and expert knowledge. This paper addresses a solution to this problem in the field of Bharatanatyam, an Indian Classical Dance (ICD) form. The video annotation itself has a broad area of coverage. It may include the annotation of objects, activities, hand gestures, facial expressions, the semantics of a video, and many more. However, the paper annotates the elementary postures of the Bharatanatyam dance videos. The proposed tool takes a video as an input and segments it into motion and stationary/non-motion frames. The non-motion frames are part of the elementary postures and our point of interest. After segmentation, the tool recognizes the postures and labels the postures’ duration of occurrence inform of frame number. The data set on which it is applied covers most basic dance variations in the Bharatanatyam, which was not addressed earlier on this large scale. The basic dance variations used to learn Bharatanatyam are called Adavus. The paper includes 13 Adavus and its 52 variations, which three dancers have performed. We use this as our data set. This annotation tool may significantly contribute to the state of the arts, saving the time and cost involved in manual annotation. The tool uses the deep learning technique for dance video annotation and gets an accuracy above 75%.