HACS, the full name of which is Human Action Clips and Segments, is a video dataset for human action recognition.

The dataset contains 200 action categories. ActivityNet-v1.3 The classification of the dataset is the same. The videos come from YouTube, with a size of 504K. Each video is less than 4 minutes long, with an average length of 2.6 minutes. The author uses a method based on uniform randomness and consistency/inconsistency of image classifiers to sample 1.5M video clips with a length of 2 seconds. The 0.6M-sized clips are labeled as positive samples, and the 0.9M-sized clips are labeled as negative samples. The dataset includes a training set of 1.4M, a validation set of 50K, and a test set of 50K, which are sampled from 492K, 6K, and 6K videos respectively.