Class Specific Object Recognition using Kernel Gibbs Distributions

Feature selection is crucial for effective object recognition. The subject has been vastly investigated in the literature, with approaches spanning from heuristic choices to statistical methods, to integration of multiple cues. For all these techniques the final result is a common feature representation for all the considered object classes. In this paper we take a completely different approach, using class specific features. Our method consists of a probabilistic classifier that allows us to use separate feature vectors, selected specifically for each class. We obtain this result by extending previous work on Class Specific Classifiers and Kernel Gibbs distributions. The resulting method, that we call Kernel-Class Specific Classifier, allows us to use a different kernel for each object class by learning it. We present experiments of increasing level of difficulty, showing the power of our approach.