Robust Vision

The computation of structure and motion from an image sequence is a fundamental problem in vision. A fully automatic approach requires not only an understanding of geometry, on which a wide range of work has been carried out in the past, but also the ability to deal with incorrect data (such as mismatched features) which will inevitably arise in a real system. It is often assumed that a standard least squares framework is sufficient to deal with outliers (data that does not agree with a postulated model). However, outliers can so distort a fitting process that the final result is arbitrary. We eschew the non-robust approach as unworkable except for carefully controlled scenes, and we present a system—consisting of feature detector, feature matcher, estimation of the Fundamental Matrix, and estimation of structure—that emphasizes robustness to outliers at each stage. The system is fully automatic, and results are shown for the computation of structurefrom-motion in a cluttered and unknown environment. Spetsakis and Aloimonos [9] divided research into the problem of structure from motion into three epochs. The first was spent finding out whether the problem presented had a solution—is it possible to make three dimensional inferences from multiple distinct digitized images of an object? Once it was ascertained that there was indeed a solution, the next period had researchers devising constructive proofs of uniqueness of the solution, involving the minimum number of points e.g. [5]. Unfortunately initial algorithms were highly sensitive to noise, leading to an erroneous belief that recovery of structure was essentially an ill-posed problem and that only 'qualitative' solutions were possible. A third period of research was then directed towards using redundant information in an optimal manner to minimize the effects of noise; methods have been proposed that use more correspondences [11] and more images [9] than are necessary. The goal of defeating noise inexorably leads to methodologies that combine many observations, and this is usually done in a least squares frame work. A major drawback, however, is that outliers, which are inevitably included in the initial fit, can so distort the fitting process that the result can be arbitrary. This is especially pertinent when the true data is degenerate but the solution appears non-degenerate due to a handful BMVC 1994 doi:10.5244/C.8.14