Compressed Video Action Recognition

Training robust deep video representations has proven to be much more challenging than learning deep image representations and consequently hampered tasks like video action recognition. This is in part due to the enormous size of raw video streams, the associated amount of computation required, and the high temporal redundancy. The ‘true’ and interesting signal is often drowned in too much irrelevant data. Motivated by the fact that the superfluous information can be reduced by up to two orders of magnitude with video compression techniques (like H.264, HEVC, etc.), in this work, we propose to train a deep network directly on the compressed video, devoid of redundancy, rather than the traditional highly redundant RGB stream. This representation has a higher information density and we found the training to be easier. In addition, the signals in a compressed video provide free, albeit noisy, motion information. We propose novel techniques to use them effectively. Our approach is about 4.6 times faster than a stateof-the-art 3D-CNN model, 2.7 times faster than a ResNet152, and very easy to implement. On the task of action recognition, our approach outperforms all the other methods on the UCF-101, HMDB-51, and Charades dataset.