Detecting conversational gaze aversion using unsupervised learning

The aversion of gaze during dyadic conversations is a social signal that contains information relevant to the detection of interest, turn-taking cues, and conversational engagement. The understanding and modeling of such behavior has implications for the design of embodied conversational agents, as well as computational approaches to conversational analysis. Recent approaches to extracting gaze directions from monocular camera footage have achieved accurate results. We investigate ways of processing the extracted gaze signals from videos to perform gaze aversion detection. We present novel approaches that are based on unsupervised classification using spectral clustering as well as optimization methods. Three approaches that vary in their input parameters and their complexity are proposed and evaluated.