On-line learning of object representations

Radial Basis Function (RBF) networks have been proposed as suitable representations for 3-D objects, in particular, since they can learn view-based representations from a small set of training views. One of the basic questions that arises in the context of RBF networks concerns their complexity, i.e., the number of basis functions that are necessary for a reliable representation, which should balance the accuracy and the robustness. In this paper we propose a systematic approach for building object representations in terms of RBF networks. We studied and designed two procedures: the "off-line" procedure, where the network is constructed after having a complete set of training views of an object, and the "on-line" procedure, where the network is incrementally built as new views of an object arrive.