Mivia Action Dataset

The Mivia Action dataset is composed by 7 high-level actions performed by 14 subjects (7 males and 7 females). All the subjects performed 2 repetitions of each action and each action has a variable length, depending on the nature of the action itself.
The images have been acquired by using a Kinect sensor. In particular, the following images have been acquired for each sequence:

• Depth Image
• Depth Background

More information about the considered actions are shown in the table below.

A frame for the actions Drinking, Sitting and Brushing teeth are shown in the following:

If you use this dataset please cite:

• Recognition of human actions from RGB-D videos using a reject option.

  @inproceedings{Carletti13,
  author = {Carletti, V. and Foggia, P. and Percannella, G. and Saggese, A. and Vento, M.},
  title = {Recognition of human actions from RGB-D videos using a reject option},
  booktitle = {International Workshop on Social Behaviour Analysis (SBA-2013)},
  issue_date = {2013},
  }

• Recognizing Human Actions by a bag of visual words.

  @inproceedings{Foggia13,
  author = {Foggia, P. and Percannella, G. and Saggese, A. and Vento, M.},
  title = {Recognizing Human Actions by a bag of visual words},
  booktitle = {IEEE International Conference on Systems, Man and Cybernetics (SMC-2013)},
  issue_date = {2013},
  }

This dataset has been used by several methods. The results, obtained by a leave one out strategy, are reported below. If you use this dataset, please let us know! We will add your results in our web site.

In order to download the whole dataset click here.

If you have any problems, do not hesitate to contact us here.