In behavior recognition, every frame is extremely important. To fully understand the human behavior in the video, it is necessary to annotate each frame according to the action and annotate densely in a video. THUMOS datasetBorn on the basis of expansion.

This dataset is an action annotation collection, including 38,690 annotations of 65 action categories. THUMOS'14 Action Detection DatasetFor the 400 videos in , there are 1.5 labels per frame on average and 10.5 action classes per video.