Precomputed features were mapped into a 95 dimensional space using Fisher linear discriminant analysis. The mean and standard deviation of each class in the LDA space was estimated from the training data set. Classification of the input vector was done by mapping the input to the LDA space and calculating the distance from the input vector to the means of all 3 classes. The input vector was then classified based on which distance/(standard deviation) ratio was smallest (i.e.: the input was assigned to the class whose mean divided by std was closest to the input).