Model selection when sensor errors are unknown

All the criteria discussed in the last section assume that the sensor error model is known a priori. In computer vision problems, however, error distributions are often unknown and difficult to model accurately, making it crucial to develop model selection criteria that depend on only weak assumptions about error distributions. Two approaches have been used for model selection when sensor errors are unavailable or unreliable. The first approach uses a runs test to reject a lower order model for a higher order model [4,9]. However, this test is only valid in 2d. In the 3d case, a heuristics due to Besl [1] is generally used. The second approach is based on the bootstrap technique. The bootstrap is a method for estimating an unknown distribution from available data. The information theoretic model selection criteria can be adapted to use bootstrap estimates of distribution dependent terms. Such criteria have been formulated using Bayesian and MDL model selection criteria, for data contaminated with small scale noise, and data with both noise and outliers [6,5].