Department of Information and Computing Sciences

Coert van Gemeren


A collection of Human interactions with accompanying skeleton metadata.

This data may only be used for research purposes. If you use this data in your work, please contact the authors.


There are 4 files that make up the ShakeFive2 dataset:

The 153 videos are encoded with ffmpeg x264 at a resolution of 1280x720.

The metadata is frame based xml data containing the skeleton joints of the actors involved in the scene. The data was collected using a Kinect2 sensor. A Skeleton rendering program (C/C++) is available with this dataset.

Related publications

Spatio-Temporal Detection of Fine-Grained Dyadic Human Interactions (pdf)
Coert van Gemeren, Ronald Poppe and Remco C. Veltkamp
7th International Workshop on Human Behavior Understanding 2016. pg.116--133.


Fist bump
Hand shake
High five
Pass object
Thumbs up

In total there are 8 interactions in the dataset:

For the older ShakeFive1 dataset, please look here.

Copyright (C) 2016 Coert van Gemeren