Learning a similarity metric discriminatively for pose exemplar based action recognition

Exemplar-based action recognition has the advantages of being compact and time-invariant. But how to select suitable exemplars and measure the pose similarities between frames and exemplars are no easy tasks. In this paper, we propose an approach to efficiently select pose exemplars and learn a pose similarity metric between frames and pose exemplars. First, a subset of training frames is mapped into pose space, where clustering is performed to select pose exemplars. Second, a pose similarity metric between frames and pose exemplars is learned based on exemplar classifiers. Finally, both training and testing videos are embedded into a space defined by similarities to pose exemplars, where action classifiers are trained to recognize actions from videos. To test our method, we have used a publicly available dataset which demonstrates that , using very simple features and fewer exemplars, our method can achieve the same or better recognition rate as the state-of-the-art methods.