We present in this paper a new multi-class Bayes classifier that permits using separate feature vectors, chosen specifically for each class. This technique extends previous work on feature Class Specific Classifier to kernel methods, using a new class of Gibbs probability distributions with nonlinear kernel mapping as energy function. The resulting method, that we call Kernel Class Specific Classifier, permits using a different kernel and a different feature set for each class. Moreover, the proper kernel for each class can be learned by the training data with a leave-one-out technique. This removes the ambiguity regarding the proper choice of the feature vectors for a given class. Experiments on appearance-based object recognition show the power of the proposed approach.