Improved Rank Pooling Strategy for Action Recognition

In the field of action recognition, it is crucial to capture the temporal evolution of video content compactly and effectively. One of the solutions is the rank pooling method that enables acquiring the evolution of video content. To further enhance temporal discrimination of the rank pooling, we proposed two improved rank pooling strategies, named the Minimum Volume Enclosing Ellipsoids (MVEE) and the Temporal Minimum Volume Enclosing Ellipsoids (TMVEE). The proposed methods are compatible with rank pooling and characterize the data distribution in other orthogonal directions to improve the temporal discrimination. We performed experiments on the ChaLearn gesture recognition and HMDB51 database, the results reveal that our proposed methods outperform other mainstreaming methos.