The last decades have shown an increasing interest in studying how to automatically capture the likeness or proximity among data objects due to its importance in machine learning and pattern recognition. Under this scope, two major approaches have been followed that use either feature-based or distance-based representations to perform learning and classification tasks. This paper presents the first results of a comparative experimental study between these two approaches for computing similarity scores using a classification-based method. In particular, we use the Support Vector Machine, as a flexible combiner both for a high dimensional feature space and for a family of distance measures, to finally learn similarity scores in a CBIR context. We analyze both the influence of the different input data formats and the training size on the performance of the classifier. Then, we found that a low dimensional multidistance-based representation can be convenient for small to medium-size training sets whereas it is detrimental as the training size grows.