HAA500 stands for Human-Centric Atomic Action Dataset, which is a human-centric atomic action dataset with selected videos. This dataset is used for action recognition of more than 591k annotated frames, involving 500 categories in total. Unlike existing atomic action datasets, HAA500 contains fine-grained atomic actions, where only consistent actions belong to the same label to minimize ambiguity in action classification. HAA500 is carefully planned to capture human actions with less spatiotemporal label noise, thereby greatly improving the training of deep neural networks.