Projective invariants and the correspondence problem

The correspondence problem has been difficult for the computer vision community to automate robustly. Typically, correspondence can only be obtained if the epipolar geometry is known (stereo camera pair), or the images are very similar (a video sequence for example). This paper proposes an new algorithm for solving the correspondence problem of line matching for uncalibrated cameras and large motion between images. The algorithm uses the cross-ratio, an invariant from projective geometry, and the minimum line border intensity difference for line matching. The algorithm was tested on real and synthetic data. Line matches with very low error were achieved for camera relations of up to 45 degrees about an object with partial image overlap.