Exemplar-based Likelihoods Using the PDF Projection Theorem

MSR-TR-2004-148 |

In computer vision it is common to define algorithms in terms of matching against exemplars. This paper describes a probabilistic framework for such algorithms. It allows you to assign a consistent likelihood for each exemplar, eliminating the normalization and bias problems which occur under the scheme of Toyama & Blake (2002).