Maximum consensus with mixed integer linear programming

Abstract

Maximum consensus is fundamentally important in computer vision as a criterion for robust estimation, where the goal is to estimate the parameters of a model of best fit. It is computationally demanding to solve such problems exactly. Instead, conventional methods employ randomised sample-and-test techniques to approximate a solution, which fail to guarantee the optimality of the result. This thesis makes several contributions towards solving the maximum consensus problem exactly in the context of Mixed Integer Linear Programming. In particular, two preprocessing techniques aimed at improving the speed and efficiency of exact methods are proposed.