Skip to Main Content
Equipping a machine with emotional sensitivity is considered a major step toward more human-like man-machine interaction. Because emotions are subjective, designing experiments to acquire ground truth data for training and testing emotion-recognition components is challenging. This is all the more true for pervasive computing environments, where users are exposed to a more diverse of stimuli than under laboratory conditions. Here, the author identifies problems of current evaluation practices, explains why they aren't appropriate to assess the performance of emotion-recognition components under real-life conditions, and presents some first ideas for designing experiments that reflect the situation outside the lab.