Consistent Multi-view Reconstruction from Epipolar Geometries with Outliers

This paper presents a method for automatic 3D-reconstruction from a set of images. The main contribution of this paper is a robust integration of the partial correspondences from image pairs provided by an existing correspondence estimator into a reconstruction consistent with all images of the scene. Projective shape and motion is estimated by factorization of a matrix containing the images of all scene points. Outlier detection is based on ransac paradigm and the trifocal tensors. Compared to previous methods, this method can handle perspective views, occlusions, and outliers in image correspondences jointly. The main novelty of this paper is the fusion of a correspondence estimator [9] for wide base-line stereo and the method for outlier detection [7]. It appears that the method is able to detect outliers that cannot be detected using the epipolar geometry alone and therefore it is suitable for integrating with wide base-line stereo from image pairs. The new method is demonstrated by experiments with laboratory and outdoor image sets and some results on metric reconstruction are shown.