Abstract
Bayesian estimation has many applications in computer vision. A frequent objection to Bayesian estimation is that the probability density functions (pdfs) involved are usually not known exactly. In fact, exact knowledge of the pdfs is not important; it often suffices to know the pdfs approximately. Furthermore, it may even suffice if we have a family of pdfs one of which approximates the actual pdf, provided we specify a “second-stage” pdf on the family such that the approximation of the actual pdf has high probability.
