Putting poses on manifold for action recognition

In action recognition, bag of words based approaches have been shown to be successful, for which the quality of codebook is critical. This paper proposes a novel approach to select key poses for the codebook, which models the descriptor space utilizing manifold learning to recover the geometric structure of the descriptors on a lower dimensional manifold space. A PageRank based centrality measure is developed to select key poses on the manifold. In each step, a key pose is selected and the remaining model is modified to maximize the discriminative power of selected codebook. In classification, the ambiguity of each action couple is evaluated through cross validation. An additional subdivision will be executed for ambiguous pairs. Experiments on ut-tower dataset showed that our method is able to obtain better performance than the state-of-the-art methods.