ActivityNet Challenge 2017 Summary

Most of traditional video action recognition methods are based on trimmed videos, which is only one action in one video. But most of videos in real world is untrimmed. In order to overcome the difficulty in some extent, we propose a method based on fusion of multiple features for untrimmed video classification task of ActivityNet challenge 2017. We use the CNN features, MBH features and stacked C3D features for classification. Then, we use one-vs-rest linear SVM to construct classifier respectively. Finally, we fuse the three results by voting to get the final