Multi-score Learning for Affect Recognition: The Case of Body Postures

An important challenge in building automatic affective state recognition systems is establishing the ground truth. When the ground-truth is not available, observers are often used to label training and testing sets. Unfortunately, inter-rater reliability between observers tends to vary from fair to moderate when dealing with naturalistic expressions. Nevertheless, the most common approach used is to label each expression with the most frequent label assigned by the observers to that expression. In this paper, we propose a general pattern recognition framework that takes into account the variability between observers for automatic affect recognition. This leads to what we term a multi-score learning problem in which a single expression is associated with multiple values representing the scores of each available emotion label. We also propose several performance measurements and pattern recognition methods for this framework, and report the experimental results obtained when testing and comparing these methods on two affective posture datasets.