In learning detectors, there are two primary competing paradigms: generative and discriminative. Generative methods learn an underlying representation of image patches that can best generate the object appearance in all its manifestations. Discriminative methods, on the other hand, learn a representation that best discriminates instances of the object from other objects that the model will likely encounter when deployed. Generative methods have the advantage that the resulting model encodes properties specific to the object, allowing novel instances of the object to be visually inspected. A popular approach that falls within the paradigm of generative methods is the famous Eigenfaces method. Discriminative methods have the advantage that the full capacity of the model is geared directlytowards the problem at hand; discriminating instances of the object from all others.
Perhaps the most well-known of all discriminative methods is the support vector machine. Although both paradigms can work well in many situations, we will see that when modeling facial features as an image patch, the discriminative paradigm is far superior.