Abstract
To model the dynamics of social interaction, it is necessary both to detect specific Action Units (AUs) and variation in their intensity and coordination over time. An automated method that performs well when detecting occurrence may or may not perform well for intensity measurements. We compared two dimensionality reduction approaches - Principal Components Analysis with Large Margin Nearest Neighbor (PCA+LMNN) and Laplacian Eigenmap - and two classifiers, SVM and K-Nearest Neighbor. Twelve infants were video-recorded during face-to-face interactions with their mothers. AUs related to positive and negative affect were manually coded from the video by certified FACS coders. Facial features were tracked using Active Appearance Models (AAM) and registered to a canonical view before extracting Histogram of Oriented Gradients (HOG) features. All possible combinations of dimensionality reduction approaches and classifiers were tested using a leave-onesubject-out cross-validation. For detecting consistency (i.e. reliability as measured by ICC), PCA+LMNN and SVM classifiers gave best results.