Human action recognition with salient trajectories and multiple kernel learning
In this paper, we propose Motion Dense Sampling (MDS) for action recognition, which detects very informative interest points from video frames. MDS has three advantages compared to other existing methods. The first advantage is that MDS detects only interest points which belong to action regions of all regions of a video frame. The second one is that it can detect the constant number of points even when the size of action region in an image drastically changes. The Third one is that MDS enables to describe scale invariant features by computing sampling scale for each frame based on the size of action regions. Thus, our method detects much more informative interest points from videos unlike other methods. We also propose Category Clustering and Component Clustering, which generate the very effective codebook for action recognition. Experimental results show a significant improvement over existing methods on YouTube dataset. Our method achieves 87.5 % accuracy for video classification by using only one descriptor.