The ActivityNet dataset contains 849 hours of YouTube videos of 200 different types of activities. In terms of activity categories and number of videos, ActivityNet is the largest temporal activity detection benchmark to date.

ActivityNet V1.3 contains a total of 19,994 untrimmed videos, which are divided into three subsets: training, validation, and testing, in a ratio of 2:1:1. Each activity category includes an average of 137 untrimmed videos, with an average of 1.41 activities per video, and these activities have temporal boundary annotations. The real annotations of the test videos are not public.