A Comparative Study on Geometric and Holistic Representations for Facial Expression Recognition

Encoding facial expression data for classification purpose plays an important role in classifier’s performance. Holistic and analytic approaches are two common representations for encoding facial expression data. Although previous research shows that a combination of these types yields the highest performance, in some applications we are constrained to choose only one of them. In this article we will explore the classification cost of Holistic and Analytic representations, instantiated by PCA and geometric features respectively. Classification cost will be measured using a Fisher-like criterion. Since our results indicate that geometric features are more promising, we will also examine which components of geometric features are more informative for facial expression classification.