Jointly Learning Multi-view Features for Human Action Recognition

Classical methods of multi-view human action recognition focus on constructing similar feature between each view so that videos under different view can be classified identically. However, the features from a certain action under different views are variable especially when the view changes sharply. Using the similarity from different views solely can't get a desirable result. In this paper we proposed a joint learining model for jointly learning features from different views. Our model explores the shared information of an action under different views and the specific information of an action under a certain view. The combination of the shared and specific information shows a distinguishable feature. The features from different views show are reconstructed by a linear projection matrix so that they can show a same structure. To obtain a optimal solution under a certain convergence, the model consists of a three-step iterative optimization process. The effectiveness of our method has been verified on WVU dataset.