Object Recognition and Pose Estimation in Eigenspace Using aRBF Network

This correspondence addresses the problem of learning object models for recognition and pose estimation. The problem is stated in terms of an input-output mapping. The input space, Rn, is a speci c eigenspace representation recently introduced by Murase and Nayar [7]. The output space, Rm, in our application consisted of three spaces with the range of [0; 1]. A function whose graph is a manifold Rn+m, that corresponds to this input-output mapping, is approximated. A regularization network, speci cally a RBF network, is used to approximate the function. The ability of the network to perform the mappings demonstrates that this is a promising technique for object recognition and pose estimation. We conclude with a discussion of some the issues that need to be resolved to make RBF networks a more attractive solution to this particular problem as well as input-output mappings in general.