Robust Recovery of the Epipolar Geometry for an Uncalibrated Stereo Rig

This paper addresses the problem of accurately and automatically recovering the epipolar geometry from an uncalibrated stereo rig and its application to the image matching problem. A robust correlation based approach that eliminates outliers is developed to produce a reliable set of corresponding high curvature points. These points are used to estimate the so-called Fundamental Matrix which is closely related to the epipolar geometry of the uncalibrated stereo rig. We show that an accurate determination of this matrix is a central problem. Using a linear criterion in the estimation of this matrix is shown to yield erroneous results. Different parameterization and non-linear criteria are then developed to take into account the specific constraints of the Fundamental Matrix providing more accurate results. Various experimental results on real images illustrates the approach.