This dataset consists of depth imagery recordings of pommel horse routines. The pommel horse is a men's gymnastics apparatus where routines involve a number of moves but primarily consist of the gymnast performing spins as he holds himself above the horse on his hands.
The dataset totals 10115 frames collected with a Microsoft Kinect 2. These frames are divided into 39 sequences spread over 5 recording sessions. All frames are annotated with whether or not the gymnast is spinning - 6405 frames consist of spinning while 3710 frames do not.
The image below shows a typical frame from the dataset.


Matlab Code

A number of Matlab scripts are included to aid in the use of this dataset:
action_annotation.m - provides an array of 0s and 1s for each sequence to indicate which frames in that sequence contain a spinning gymnast.
play_seq.m - loads and displays each sequence (change path depending on location of data).

Download and Citation

The dataset can be downloaded in entirety or as smaller sections: entire dataset (1.2GB), section 1 (143 MB), section 2 (421 MB), section 3 (271 MB), section 4 (97 MB), section 5 (314 MB),
The dataset provided here is for non-commercial research/educational use only.

Please cite the following paper if you use this dataset - detailed information about our approaches are also available in the paper:
Brian Reily, Hao Zhang, and William Hoff, "Real-time gymnast detection and performance analysis with a portable 3D camera", Computer Vision and Image Understanding, 2016 (Under Review).

Questions and Suggestions

Please contact Brian Reily: breily -AT- mines -DOT- edu