Action Unit Detection

In this task the goal is to detect the presence or absence of individual FACS action units (AU).

AU Detection on the Bosphorus database

Benchmarking Protocol for AU Coding

There are three protocols to evaluate automatic AU detection and intensity estimation performances. High intensity AU detection (C, D and E level), low intensity AU detection (testing with B level) and AU intensity estimation (A to E levels).

For each protocol and for each AU, subject independent 10-fold cross validation is made. Folds are balanced according to the number of positive AU samples. Cross validation folds for 25 AUs over a dataset of 2902 faces are provided.

Performance evaluation: The detection performance of an AU is measured in terms of AuC: the area under the ROC curve (hit versus false alarm rate), calculated over the combination of 10 test folds for more accurate estimation with small number of positive samples. The intensity estimation is evaluated in terms of the correlation coefficient between the scores of the FACS expert and the score of the estimator.

Author Title Source

A. Savran, B. Sankur, M. T. Bilge

IEEE CVPR'10 Workshop on Human Communicative Behavior Analysis, San Francisco, California, USA, June 2010.

AU detection on the Extended Cohn-Kanade (CK+) database

The evaluation protocol for CK+ is described in Lucey et al.

Author Title Source

P. Lucey, J.F. Cohn, T. Kanade, J. Saragih, and Z. Ambadar

In Proceedings of the 3rd IEEE Workshop on CVPR for Human Communicative Behavior Analysis (CVPR4HB), CVPR 2010, 2010.

AU detection subtask in FERA2011

The benchmarking protocol can be found on the FERA2011 website.

Author Title Source

Michel F. Valstar, Bihan Jiang, Marc Méhu, Maja Pantic, and Klaus Scherer

In Proc. IEEE Int’l Conf. Automatic Face and Gesture Recognition, 2011.