A Bayesian framework for panoramic imaging of complex scenes

Abstract

This thesis presents a Bayesian framework for generating a panoramic image of a scene from a set of images, where there is only a small amount of overlap between adjacent images. Dense correspondence is computed using loopy belief propagation on a pair-wise Markov random field, and used to resample and blend the input images to remove artifacts in overlapping regions and seams along the overlap boundaries. Bayesian approaches have been used extensively in vision and imaging, and involve computing an observational likelihood from the input images and imposing a priori constraints. Photoconsistency or matching cost computed from the images is used as the likelihood in this thesis. The primary contribution of this thesis is the use of and efficient belief propagation algorithm to yield the piecewise smooth resampling of the input images with the highest probability of not producing artifacts or seams.