IMPSAC: Synthesis of importance sampling and random sample consensus

Abstract

This paper proposes a new method for effecting feature correspondence between images. The method operates from coarse to fine and is superior to previous methods in that it can solve the wide baseline stereo problem, even when the image has been deformed or rotated. At the coarsest level a RANSAC-style estimator is used to estimate the two viewimage constraint R which is then used to guide matching. The two viewrelation is an augmented fundamental matrix, being a fundamental matrix plus a homography consistent with that fundamental matrix. This is akin to the plane plus parallax representation with the homography being used to help guide matching and to mitigate the effects of image deformation. In order to propagate the information from coarse to fine images, the distribution of the parameters Θ of R is encoded using a set of particles and an importance sampling function. It is not known in general how to choose the importance sampling function, but a new method “IMPSAC” is presented that automatically generates such a function. It is shown that the method is superior to previous single resolution RANSAC-style feature matchers.