Classifying Actions and Measuring Action Similarity by Modeling the Mutual Context of Objects and Human Poses

In this paper, we consider two action recognition problems in still images. One is the conventional action classification task where we assign a class label to each action image; the other is a new problem where we measure the similarity between action images. We achieve the goals by using a mutual context model to jointly model the objects and human poses in images of human actions. Experimental results show that our method not only improves action classification accuracy, but also learns a similarity measure that is largely consistent with human perception.