Get the Whole Action Event by Action Stage Classification

Abstract

Spatiotemporal action localization in videos is a challenging problem which is also an essential and important part of video understanding. Impressive progress has been reported in recent literature for action localization in videos, however, current state-of-the-art approaches haven’t considered the scenario of broken actions, in which an action in an untrimmed video is not a continuous image series anymore because of occlusion, shot change, etc. So, one action is divided into two or more footages (sub-actions) and the existing methods localize each of them as an independent action. To overcome the limitation, we introduce two major developments. Firstly, we adopt a tube-based method to localize all sub-actions and discriminate them into three action stages with a CNN classifier: Start, Process and End. Secondly, we propose a scheme to link the sub-actions to a complete action. As a result, our system is not only capable of performing spatiotemporal action localization in an online-realtime style, but also can filter out irrelevant frames and integrate sub-actions into single tube that has better robustness than the existing method.