Integrating local Action Elements using Implicit Shape Model for Action Matching

In this paper we propose an approach for analyzing human action in video, and demonstrate its application to several related tasks: action retrieval, action classification and action localization. In our work, actions are rep-resented as a set of local space time features, or Action Elements, then an Implicit Shape Model of the action is built to integrate spatial and temporal correlations of these local Action Elements. In particular, we propose two different approaches to extracting Action Elements: a discriminative and a generative approach, respectively. Action Elements detected from either of the approaches are then used to build an Implicit Shape Model for current action of interest. We apply our action matching algorithm to the application of action recognition and carry out thorough experiments on the two well-known action datasets, KTH and Weizmann. Results obtained from those experiments demonstrate the highly competitive performance of our action matching algorithm compared to state-of-the-arts, and also give evaluative insights on the two proposed feature extraction approaches