Mosaicing a Large Number of Widely Dispersed, Noisy, and Distorted Images: A Bayesian Approach

In this paper we extend existing mosaicing algorithms to deal with image sequences that are captured over a wide spatial area, exhibit large geometric and photometric distortions, and contain significant additive noise and other contaminations, such as light reflections. The paper focuses on three main contributions. (1) We extend the camera model used for mosaicing to deal, not only with geometric lens distortion, but with vignetting and permanent occluders as well. (2) We introduce a novel method for global image alignment based on a technique from the robotics literature, together with a novel optimization strategy, the folding algorithm, to guarantee global convergence. (3) We utilize techniques developed in the super-resolution literature for restoration of the final image mosaic from the contaminated input images. The estimation of camera parameters, the global alignment and the final mosaic are all derived within a unifying Bayesian framework, starting with the single objective of obtaining the maximum a posteriori estimate of the final mosaic, given the input images. Our approach is illustrated with results on a complex and challenging image sequence obtained from a state of the art robotics application.