Comparing OpenFace to Manual Annotations of Communicative Facial Signals
This thesis assesses the differences between the output of OpenFace and manual annotations of communicative and holistic facial signals. OpenFace is a software program that detects facial signals in videos of human faces as Action Units. These unit of facial movement are not always of interest for research. Human coders might only want to annotate communicative and holistic facial signals, instead of all visible signals. Video annotation is a time-consuming process to do manually, so automation is desired. This thesis explains how the output of OpenFace and annotations of communicative signals differ on conceptual level, goal, and features. These differences should be considered when using OpenFace for annotation of communicative and holistic facial signals. An attempt is made to transform the output of OpenFace into annotations of frowns, blinks, smiles, and gaze aversion by manually finding thresholds and constraints. A minimal agreement is reached between the transformed output and the manual annotations. The conclusion is that OpenFace can be used to automate the annotation of communicative facial signals, but only with the help of machine learning. Unbiased data is required for training, together with objective definitions of communicative facial signals.
Faculteit der Sociale Wetenschappen