Wide-baseline 3d Reconstruction from Digital Stills

The last decade saw tremendous progress in the area of structure-and-motion. Techniques have been developed to compute the camera trajectory and reconstruct scenes in 3D based on nothing but a video or a set of closely spaced images. No extra information or calibration is needed to perform the reconstruction. This paper presents extensions on the traditional shape-from-video pipeline in order to deal with wide-baseline conditions, i.e. views that are much farther apart. The algorithms for both the sparse and dense correspondence search need to be upgraded for the system to deal with these wide-baseline views. These new techniques are discussed in this paper and results are shown.