Three dimensional reconstruction from multiple images
Motilal Agrawal, Larry Steven Davis · 2002
The ability to automatically and accurately extract models of real life objects from images is an important problem in computer vision. Applications range from virtual reality, automated surgery, to the modelling of tissues and cells (through electron microscopic images). The problem is ill-posed if we are given only a single image of the scene to be reconstructed. In this thesis, we develop new algorithms for three dimensional reconstruction from two or more views. In order to build these models, the cameras have to be calibrated first. We describe a novel algorithm for camera calibration using spheres. We then describe a window-based discontinuity preserving binocular stereo algorithm. Next, we describe a feature-based stereo algorithm and a novel non-linear warping scheme which is used for reconstructing synaptic structures from transmission electron microscopy images. Our third algorithm uses three views in a right angled configuration. Finally, we present a more general probabilistic framework applicable for any number of arbitrary views under the assumption that the scene to be reconstructed is approximately Lambertian.