Sparse Code Filtering For Action Pattern Mining

Action recognition has received increasing attention during the last decade. Various approaches have been proposed to encode the videos that contain actions, among which self-similarity matrices (SSMs) have shown very good performance by encoding the dynamics of the video. However, SSMs become sensitive when there is a very large view change. In this paper, we tackle the multi-view action recognition problem by proposing a sparse code filtering (SCF) framework which can mine the action patterns. First, a class-wise sparse coding method is proposed to make the sparse codes of the between-class data lie close by. Then we integrate the classifiers and the class-wise sparse coding process into a collaborative filtering (CF) framework to mine the discriminative sparse codes and classifiers jointly. The experimental results on several public multi-view action recognition datasets demonstrate that the presented SCF framework outperforms other state-of-the-art methods.