Human Identification versus Expression Classification via Bagging on Facial Asymmetry

Abstract

We demonstrate a dual usage of quantified facial asymmetry for (1) human identification under expression variations and (2) expression classification across different human subjects. Our experiments show the effectiveness of using statistical bagging and feature subspace selection BEFORE applying classiffers such as Linear Discriminant Analysis. This preprocessing allows the same type but different dimensions of image features to be discriminative for two seemingly conflicting classiffcation goals. Statistically significant improvements are found when facial asymmetry features are combined into classical classifiers.