Abstract
The ability to accurately judge another persons’ higher-level attributes with a short duration of observations is a unique perceptual mechanism for humans, termed as the thin-slice theory of judgment in psychology. In this work, we propose a computational framework based on behavior-emotion mutual information and behavior density to extract the “thin-sliced local emotion-rich and non-outlier behavior segments” within each session to be used as the data to train the global affect recognizers. We achieve the global emotion recognition accuracy of 0.722, 0.834, and 0.822 for activation, dominance, and valence respectively, which improves 0.338, 0.159 and 0.251 absolute over using behavior data of the entire session. The significant improvement in the global emotion recognition rate reinforces the thin-slice nature of human emotion perception. Furthermore, our detailed analyses for the “thin-sliced segments” indicate that the major changes within the selected thin slice data are along the following aspects:(1)time distribution and (2)behavior distribution. It effectively reduces the complexity of the training data and thus enhance the prediction accuracy; furthermore, our human perceptual experiment demonstrates that the framework indeed does make an impact and change for affect perception. By properly extracting the thin-slice segments, we obtain not only improved global emotion recognition rates but also bring additional insights into this emotion thin-slice perception mechanism.