UCF-101 is a dataset with 101 action categories collected from YouTube, which is an extension of the UCF-50 dataset with 50 action categories.

The UCF-10 dataset has 101 action categories and a total of 13,320 videos. It provides diversity in action and has changes in camera motion, object appearance, object scale, posture, etc. It is also the most challenging dataset to date. This dataset aims to accelerate behavior recognition by learning and exploring new display actions.

The 101 action categories in this dataset are divided into 25 groups, each containing 4-7 action videos. The videos in the same group will have similar scenes. The action categories are mainly divided into the following five types:

• Human-object interaction
• Human body movements
• Interaction between people
• Playing an instrument
• sports