Eigenexpressions: Emotion Recognition Using Multiple Eigenspaces

This paper presents an appearance-based holistic method for expression recognition. A two stage supervised learning approach is used. At the first stage, training images are used to compute one subspace per expression. At the second stage, the same images are used to train a classifier. In this step, Euclidean distances from each image to each particular subspace are used as the input to the classifier. The resulting system significantly outperforms the baseline eigenfaces method on the Cohn-Kanade data set, with performance gains in the range 10%-20%.