Kinetics Human Behavior Dataset

Kinetics is a large-scale, high-quality human action-focused YouTube video link dataset, containing 650,000 video clips. Depending on the version of the dataset, it covers 400/600/700 human behavior categories.

These actions are mainly people-oriented and cover a wide range of content, including interactions between people and objects, such as playing musical instruments, and interactions between people, such as shaking hands. Each video is a manually annotated action classification and lasts about 10 seconds.

This dataset includes the following subsets:

• Kinetics 700-2020
• Countix
• Kinetics 700
• Kinetics 600
• Kinetics 400