Performance of computer vision algorithms is contingent upon the quality of the input. With uncorrupted data most of the algorithms perform well, however, once noise is present the results may become unreliable. The noise can appear as uncertainty about the input and/or incorrect assumptions about the underlying model structure. A robust computer vision algorithm should tolerate both types of noise. The methods borrowed from statistics or estimation theory were not designed for visual data where resistance to data uncertainty and tolerance of incorrect assumptions are of paramount importance. Good performance for noisy visual data can only be achieved through robust detection of the consensus among the outputs of a large number of identical non-robust processes, with a methodology called the Consensus Paradigm. A similarity exist between the computational steps of recently proposed 3D object recognition algorithms and the robust estimation methods. The research will exploit this similarity and develop object recognition algorithms with improved noise resistance. Results of the research will contribute to the development of techniques for autonomous decision making in unfamiliar visual environments.
