Structure and motion from a sparse set of views

We address the problem of acquiring 3D information of an object from multiple images. While a long image sequence contains more clues about the motion of the object in the scene, it provides no more information about the object than a few images that show various aspect of the object. We propose an algorithm that uses nonlinear least squares fitting to compute structure and motion from a small number of images in which various aspect of an object is shown. The location of features that show up in different aspect of the object are computed with respected to a single reference frame. As with all other nonlinear problems, our algorithm requires initial guesses. While we adopted an analytical method in the initialization stage, experimental results on synthetic data and real images show that the quality of our solution does not degrade with the accuracy of the initial guesses