Probabilistic selection of frames for early action recognition in videos

Early action recognition seeks to recognize human actions in a video, while the video has been only partially observed. In this paper, we introduce an approach to this kind of recognition task. In some offline (non-early) recognition works, it has been proposed to sample frames of the video uniformly and use them in training of the model. However, there is no reason that uniform sampling should be optimal, so we propose a non-uniform sampling to make it more tailored to early recognition. The proposed method samples the frames in such a way that earlier frames are more likely to be chosen. These frames are then used in training a deep network architecture. We compare our sampling approach with a uniform sampling process, using HMDB51 dataset as a benchmark. We further compare our method with other state-of-the-art early recognition works. The experimental results suggest that our sampling process leads to better recognition accuracy than uniform sampling, at the early stages of the video, and that our proposed algorithm outperforms the state-of-the-art.