There are many available datasets that we can use in the training process, as follows:

• UCF101 (http://crcv.ucf.edu/data/UCF101.php) is an action recognition dataset of realistic action videos with 101 action categories. There are 13,320 videos in total for the 101 action categories, which makes this dataset a great choice for many research papers.
• ActivityNet (http://activity-net.org/) is a large-scale dataset for human activity understanding. There are 200 categories with over 648 hours of video. Each category has about 100 videos.
• Sports-1M (http://cs.stanford.edu/people/karpathy/deepvideo/) is another large-scale dataset for sports recognition. There are 1,133,158 videos in total, annotated with 487 sports labels.

In this chapter, we will use UCF101 to perform the training process. We also recommend that you try to apply the techniques discussed in this chapter to a large-scale dataset to take full advantage of multiple-GPU training.