Generic problems of image understanding in the context of reconstructing the 3D visual environment with uncalibrated cameras are to be investigated using tools of modern statistics. These generic problems also raise challenging questions in statistics requiring analysis of real data of uncommon types not ordinarily found in the statistical literature. Introduction of robust estimators into computer vision led to a significant performance improvement. The currently popular estimators, however, are not optimal for image data since they tend not to be simultaneously efficient locally and with high breakdown points. Realistic models of noisy data should take into account the distributions arising from the image formation process, as well as the presence of errors in all the feature coordinates not only in the measurements. Redefining image understanding tasks as statistical problems, offers additional insight which can be converted into new algorithms with superior performance. The goal of this research is to exploit recent progress in statistics for the advance of computer vision, and to use the real problems of image understanding as a testing ground and catalyst for novel statistical methods.
