Facial action unit detection using variable decision thresholds

Detection of facial action units (AUs) is an important research field for recognizing emotional states in facial expressions. Here, we propose a novel, yet effective method, that utilizes variable decision thresholds at the prediction stage of a binary learning method for AU detection. The method performs a thresholding technique to find optimum values for each AU and make use of these thresholds as the decision threshold of the support vector machine (SVM) algorithm. Our experiments on Extended Cohn-Kanade (CK+) dataset show significant improvements on most of the AUs with an average F1 score of 6.383% compared with the baseline method.