Reconstruction of structured scenes from two uncalibrated images

This paper investigates a practical heuristics method for reconstruction of structured scenes from two uncalibrated images. The method is based on an initial estimation of principal homographies from 2D initial point matches that may contain some outliers, and the homographies are refined recursively by incorporating the supporting matches of both points and lines on principal space surfaces. Then epipolar geometry is recovered directly from the refined homographies and cameras are calibrated from three orthogonal vanishing points and the recovered infinite homography. There are several points of novelty. First, a simple homography-guided method for matching line segments between two views is proposed. Second, under the assumption of zero-skew, the cameras are auto-calibrated with all the four intrinsic parameters varying between the two views. The advantages of the method is that it can build more realistic models with minimal human interactions, and it also allows us to reconstruct more visible surfaces on the detected planes than traditional methods which can only reconstruct the overlapping parts, since the homography provides a one-to-one mapping of points and lines between different views. Extensive experiments with real images illustrate the validity and advantages of the proposed method.