The dataset is composed of 10 sessions. Each session provides RGB-D images and skeleton tracks of 2 long videos of different activities performed by two people. Each session is zipped in a separate file, which contains a folder that has skeleton tracks in a text format and RGB (24 bits) and depth images. Each row of the skeleton text file contains information about positions (6 DoF) of 25 joints.

If you use the dataset or our code for your research, please cite our RO-MAN 2017 paper that describes the data collection in detail.

@InProceedings{Coppola2017,
Title = {Automatic Detection of Human Interactions from RGB-D Data for Social Activity Classification},
Author = {C. Coppola and S. Cosar and D. Faria and N. Bellotto},
Booktitle = {IEEE Int. Symposium on Robot and Human Interactive Communication (RO-MAN)},
Year = {2017},
Pages = {871-876}
}