Composable Activities Dataset Action Recognition Dataset

The Composable activities dataset consists of 693 videos containing 16 types of activities performed by 14 actors. Each activity consists of 3 to 11 atomic actions. The RGB-D data of each sequence is collected using a Microsoft Kinect sensor, and the positions of the relevant body joints are estimated.

The dataset provides activity annotations for each video and the movements of four human body parts (left/right arms and legs) for each frame in each video.