INRIA Xmas Motion Acquisition Sequences (IXMAS). Multiview dataset for view-invariant human action recognition.

13 daily-live motions performed each 3 times by 11 actors. The actors choose freely position and orientation.

Downloadable files: original views of 5 cameras (23fps) in png-format (390x291), silhouettes in bpm-format (390x291), reconstructed volumes in matlab format (64x64x64), camera calibration data, framewise ground truth labeling: 0 - nothing, 1 - check watch, 2 - cross arms, 3 - scratch head, 4 - sit down, 5 - get up, 6 - turn around, 7 - walk, 8 - wave, 9 - punch, 10 - kick, 11 - point, 12 - pick up, 13 - throw (over head), 14 - throw (from bottom up).

For more details on the dataset and for recognition results in 3D please see the paper:
[1] Daniel Weinland, Remi Ronfard, Edmond Boyer, Free Viewpoint Action Recognition using Motion History Volumes, Computer Vision and Image Understanding (2006). pdf, experimental matlab code

Results for single view recognition are in:
[2] Daniel Weinland, Edmond Boyer, Remi Ronfard, Action Recognition from Arbitrary Views using 3D Exemplars, International Conference on Computer Vision (2007). pdf

Note: In both papers we were testing on the actions {1 2 3 4 5 6 7 8 9 10 12} and on all actors except "Pao" and "Srikumar". In [2] we were not testing on the actions that are irregularly performed with the opposite body part.

For questions and remarks directly related to the IXMAS sequences please contact

Update: added missing frame in sequence nicolas3_png.tar.gz (9.10.2007); corrected small fault in ciara1 and clare1 ground truth (22.1.2008)

Check and download what you wish

Mouse Over pictures to display check boxes

groupe fond


groupe fond


click above to download a wget (?) script.

Auto-Extract all archives at the end of the wget process?


or click here to download files manually.


loading gif sequence illustration