基于互信息的中层部件挖掘的行为分类技术

Mid-level parts have shown prominent performance for video based action classification. And some existing schemes mine the semantic parts by some predefined heuristic rules. However, these mined parts maybe intuitive but not discriminative. To solve this issue, this paper proposes a mutual information based selection strategy, which regards the parts as the visual words. Typically, the algorithm first densely samples some 3d cubes from training samples, which followed by conducting candidacate set. Then videos are presented into candidate feature space, where we conduct parts selection in the space by applying mutual information and the top ranked parts will be retrained. Finally, the selected parts are applied to extract mid-level features for the training and testing data, and a classifier is trained on the training data and test on the testing data. Experiments are conducted on two challenging dataset--KTH, and HMDB51 dataset. The results show that the proposed method achieve prominent performance.