IXMAS Actions – New Views and Occlusions

The dataset on this page contains new recordings of the 11 actions in the IXMAS dataset [1]. The dataset was recorded with different actors, cameras, and viewpoints, and 2/3 of the actions were recorded with objects in the scene partially occluding the actors. Overall it contains 1148 sequences.


 

We provide full sequences, and data in the format that we used for our experiments in [2]: one avi per action showing a static ROIs of size (64×48) centered around the person. In addition we provide a short annotation file for every avi in Matlab format.

For convenience we also provide extracted ROIs from the original IXMAS data in the same format, as well as the artificially occluded sequences from the IXMAS data used for training. For the complete, original IXMAS dataset please refer to the 4drepository webpage.

You are free to use the new sequences for research purposes. If you use them to publish results, please cite reference [2].

Download

Data as used in paper

new EPFL IXMAS Recordings ROIs(191 MB)
original IXMAS ROIs (251 MB)
occluded IXMAS ROIs (208 MB)

Other

For questions, please contact weinland { at } gmail.com

References

Weinland, Daniel and Ronfard, Remi and Boyer, Edmond Free Viewpoint Action Recognition using Motion History Volumes. Computer Vision and Image Understanding (CVIU), Volume 104, Number 2-3, 2006

Making Action Recognition Robust to Occlusions and Viewpoint Changes

D. Weinland; M. Özuysal; P. Fua 

2010. European Conference on Computer Vision, Heraklion, Greece, September, 2010. p. 635–648. DOI : 10.1007/978-3-642-15558-1_46.