The PKU-MMD dataset is a skeleton-based action recognition dataset. The dataset consists of 1,076 uncut video sequences of 66 participants shot from 3 camera positions. There are 51 action categories annotated in the dataset, producing nearly 20,000 action instances and 5.4 million frames. Similar to the NTU RGB+D dataset, this dataset also has two recommended evaluation schemes, namely cross-subject and cross-view.