PCA = Gabor for Expression Recognition

We show that Gabor lter representations of facial images give quantitatively indistinguishable results for classi cation of facial expressions as local PCA representations, in contrast to other recent work. We then show that a linear discriminant analysis performed on the Gabor lter representation automatically locates the important regions corresponding to the facial actions involved in portraying emotion. Finally, we introduce a cognitively plausible method of \peeking" at an (unlabeled) test set to improve generalization.