Computational Information Geometry Wonderland

Inference from data to deduce solution has been increasingly popular in computer vision. Specially with the success of AdaBoost. More and more algorithms are data-driven (thanks to manually prepared groundtruth databases). Today, I would like to mention the use of the optimal log-likelihood ratio test. The error rate is know to decrease exponentially according to Chernoff information. So suppose you have to label on-edge or out-edge pixels (say, for a road tracking system), you can learn the response distributions, and design a simple classification based on a calibrated test.

This was done in a CVPR'99 (later PAMI 2003) paper. The method is rather generic and allows one to measure the amount of information gained by each operator/filter.

Statistical edge detection: learning and evaluating edge cues
(PAMI 2003)
Frank.