Cross-view Action Recognition via Dual-Codebook and Hierarchical Transfer Framework

In this paper, we focus on the challenging cross-view action recognition problem. The key to this problem is to find the correspondence between source and target views, which is realized in two stages in this paper. Firstly, we construct a Dual-Codebook for the two views, which is composed of two codebooks corresponding to source and target views, respectively. Each codeword in one codebook has a corresponding codeword in the other codebook, which is different from traditional methods that implement independent codebooks in the two views. We propose an effective co-clustering algorithm based on semi-nonnegative matrix factorization to derive the Dual-Codebook. With the Dual-Codebook, an action can be represented based on Bag-of-Dual-Codes (BoDC) no matter it is in the source view or in the target view. Therefore, the Dual-Codebook establishes a sort of codebook-to-codebook correspondence, which is the foundation for the second stage. In the second stage, we observe that, although the appearance of action samples will change significantly with viewpoints, the temporal relationship between atom actions within an action should be stable across views. Therefore, we further propose a hierarchical transfer framework to obtain the feature-to-feature correspondence at atom-level between source and target views. The framework is based on a temporal structure that can effectively capture the temporal relationship between atom actions within an action. It performs transfer at atom levels of multiple timescales, while most existing methods only perform video-level transfer. We carry out a series of experiments on the IXMAS dataset. The results demonstrate that our method obtained superior performance compared to state-of-the-art approaches.