Salient Points

Frame Selection for Action Recognition on trimmed videos

Considers all frames at once instead of one frame at a time

If used as a pre-processing step to a complex model, we obtain state-of-the-art results on UCF101 and HMDB51

Outperforms other frame selection approaches on long untrimmed videos with respect to accuracy

Reduces cost of computation by 4-10 times with respect to full video, whilst outperforming full video in terms of accuracy