Action Detection and Annotation in Videos

Abstract

Action detection has been a hot topic for last many years. Despite the recent advances in computer vision especially deep learning techniques, action detection in realistic videos is still a problem far from being solved. The lack of availability of spatio-temporally annotated datasets makes it di cult to train and test various deep learning based models. The manual methods for obtaining these bounding box annotations are time consuming, cumbersome and include human biases. Moreover, annotation needs to be done from scratch for any new dataset. The above mentioned problems create the need for an automatic annotation method. Most of the existing action localization techniques make use of the action proposals. The supervised action proposal method generally need action labels as well as annotated boxes, limiting the scalability of method. Whereas unsupervised methods usually generate large number of action proposals which contain many noisy, inconsistent, and unranked action proposals. In this thesis a new scheme for selection of fewer but properly ranked action proposals is proposed to solve the issue of time complexity with improved proposal ranking. These action proposals when fed into any system can improve the e ciency as well as accuracy of the existing techniques. The action proposals ranked are compared with existing work and results show the superiority of our scheme. Moreover, the proposed method is generic and independent of action proposal method.