We ran an experiment recording face expressions from participants watching naturalistic video portraying a very broad range of arousing situations. We start from scratch building both the camera head-set system,and the data collection paradigm. All subsequent data preprocessing and step to clean the data were developed and validated in the lab.
Around 1100 videos were selected from the internet and adapted to the experimental video presentation. Every participant came to the lab from 10-12 sessions where 90 to 150 video were presented. Before starting with the experiment we projected an equispaced grid on the face of participants and applied 68 adhesive markers. The facial expressions were quantified as time series of physical displacement of dots (markers manually applied) in the recorded videos of participants’ faces. For each video, subjects had to signal every time they experienced an emotion, by pressing a key. If they signaled an “emotion transition”(i.e. a shift from baseline, or an evident change in the subjective emotional state) during the vision of the video, a second video reproduction would have started stopping automatically at the exact time they had previously pressed the button. subjects were then ask to produce a 1-to-5 free-choice emotional labels (free terms to describe the quality of the emotional state) and a rating of intensity on a 1-to-9 Likert scale, indicating how intense the emotion was perceived. We obtained a total of 32000 facial expression modelled as timeseries of displacement of a 68 vertex grid, projected on participants’ faces, moreover for each facial expression we have an affective label and an intensity rating associated. Automated machine learning routines were applied to discover how many cluster of face expression could be suffiecient to explain the variance across 32 (most used) affective labels.