AutoFACS is a tool to automatically annotate video data using the Facial Action Coding System.
The scripts currently in this repo are a proof of concept that closely follows this paper.
The general process is to take a frame, predict facial landmarks, normalize those landmarks using procrustes analysis, and then feed the distance between landmarks into a Support Vector Machine to detect the presence or absence of an Action Unit.
Several improvements are currently underway including:
- An improved facial landmark detector that can predict more points with higher accuracy
- Training the Action Unit classifiers on a much larger dataset that was assembled from several different datasets
- Adding a GUI for ease of use
The result of these improvements will be a fully featured GUI based tool for psychology researchers with an improved model for Action Unit prediction.