Cross-Dataset Learning And Person-Specific Normalisation For Automatic Action Unit Detection

Automatic detection of Facial Action Units (AUs) is crucial for facial analysis systems. Due to the large individual differences, performance of AU classifiers depends largely on training data and the ability to estimate facial expressions of a neutral face. In this paper, we present a real-time Facial Action Unit intensity estimation and occurrence detection system based on appearance (Histograms of Oriented Gradients) and geometry features (shape parameters and landmark locations). Our experiments show the benefits of using additional labelled data from different datasets, which demonstrates the generalisability of our approach. This holds both when training for a specific dataset or when a generic model is needed. We also demonstrate the benefits of using a simple and efficient median based feature normalisation technique that accounts for person-specific neutral expressions. Finally, we show that our results outperform the FERA 2015 baselines in all three challenge tasks - AU occurrence detection, fully automatic AU intensity and pre-segmented AU intensity estimation.