Frustratingly Easy Personalization for Real-time Affect Interpretation of Facial Expression

摘要: In recent years, researchers have developed technology to analyze human facial expressions and other affective data at very high time resolution. This is enabling develop study interactive robots that are increasingly sensitive their interaction partners' states. However, typical planning models algorithms operate on timescales frequently orders of magnitude larger than the which real-time affect sensed. To bridge this gap between scales sensor collection modeling, must be aggregated interpreted over longer timescales. paper we clarify formalize computational task interpretation in context an educational game played by a robot, during expression sensed, interpreted, used predict partner's gameplay behavior. We compare different techniques for interpretation, generate sets labels modeling inference task, evaluate how generated each technique impact model training inference. show incorporating simple method personalization into process - dynamically calculating applying personalized threshold determining feature leads significant improvement quality inference, comparable performance gains from pre-processing steps such as smoothing via median filter. discuss implications these findings future development affect-aware propose guidelines use methods scenarios.