Abstract
The understanding of emotion perception is of fundamental importance for the advance of cognitive and vision science. Yet, research on the production and perception of facial expressions of emotion has focused on the representation and recognition of six basic emotions – happy, sad, angry, surprise, fear and disgust. Muscle groups (i.e., Action Units, AU) involved in each of these emotions have been identified. Also, two cognitive models have been proposed for the representation and recognition of these six emotions. The continuous model represents each facial expression as a feature vector in a common face space. The categorical model defines a classifier for each of the emotion labels. Unfortunately, the current definition of these models does not account for the representation and recognition of compound emotions, e.g., happily surprised, angrily surprised, fearfully surprised, or hatred (which is defined as feeling anger and disgust toward someone). We have collected a large dataset of 25 distinct facial expressions of emotion from a total of 100 individuals. We have identified the common AUs for each expression and defined a hybrid continuous-categorical model that explains how such a large number of expressions can be represented and recognized. In the proposed model we do not need to train 25 distinct classifiers as in the categorical view. Instead, we show that by linearly combining a small number of classifiers, the model can readily represent and recognize a very large number of emotion categories. This also resolves the problem of the continuous view, where happily surprised and angrily surprised would be represented as disjoint areas of a continuous face space even though they both express surprise.
Meeting abstract presented at VSS 2012