Automated recognition of complex categorical emotions from facial expressions and head motions

Classifying complex categorical emotions has been a relatively unexplored area of affective computing. We present a classifier trained to recognize 18 complex emotion categories. A leave-one-out training approach was used on 181 acted videos from the EU-Emotion Stimulus Set. Performance scores for the 18-choice classification problem were AROC = 0.84, 2AFC = 0.84, F1 = 0.33, Accuracy = 0.47. On a simplified 6-choice classification problem, the classifier had an accuracy of 0.64 compared with the validated human accuracy of 0.74. The classifier has been integrated into an expression training interface which gives meaningful feedback to humans on their portrayal of complex emotions through face and head movements. This work has applications as an intervention for Autism Spectrum Conditions.