|Description (include details on usage, files and paper references)||This is a custom generated dataset designed for the task of action co-segmentation in pairs of action sequences.
The dataset contains 101 pairs of action sequences that were generated based on single action clips of the original Berkeley-MHAD dataset (http://tele-immersion.citris-uc.org/berkeley_mhad/).
Two types of data are provided based on motion capture data and video data noted as the MHAD101-s and MHAD101-v variants.
We provide 3D skeletal features for the MHAD101-s based on 3D joint angles extracted from the original motion capture data of MHAD action clips.
MHAD101-v contains video-based extracted features based on optical flow and Dense Trajectories.
In 50 of the paired sequences, each sequence consists of 3 concatenated action clips and the paired sequences have exactly 1 in common. In 17 pairs,
each sequence consists of 3-7 actions and the two sequences have 2 in common. In 17 pairs, each sequence consists of 3-7 actions and the paired sequences have 3 actions in common. Finally, in 17 pairs, each sequence consists of 4-7
actions and paired sequences have 4 in common. It is also guaranteed that (a) a sequence contains actions of the same subject (b) to promote style and duration variability, for every pair, the two sequences involve different subjects and (c) the placement of the common actions in the sequences is random.
More details are provided in the provided link of the dataset
as well as in the publication:
 K. Papoutsakis, C. Panagiotakis and A.A. Argyros, "Temporal Action Co-Segmentation in 3D Motion Capture Data and Videos", In IEEE Computer Vision and Pattern Recognition (CVPR 2017),