Automatic Identification of Anthropological Face Landmarks for Emotion Detection

We propose a facial landmark system constructed on anthropological principles that is intended to substitute the (costly) use of Action Units (AU) in characterizing facial dynamics, for the purpose of automatic emotion detection. We demonstrate the suitability of the method by implementing a lightweight automatic Anthropological Face Landmarks (AFL) identification algorithm and compare the task of emotional valence classification using the proposed system to explicitly using AU labels. Despite the comparison being unfair to our method (AUs are labeled by experts, while the landmarks are detected automatically), we show that even in a Proof-of-Concept setting, our system follows closely the performance of detecting emotions with Action Units.