Spherical mosaics with quaternions and dense correlation

We describe an algorithm for generating spherical mosaics from a collection of images acquired from a common optical center. The algorithm takes as input an arbitrary number of partially overlapping images, an adjacency map relating the images, initial estimates of the rotations relating each image to a specified base image, and approximate internal calibration information for the camera. The algorithm's output is a rotation relating each image to the base image, and revised estimates of the camera's internal parameters. Our algorithm is novel in the following respects. First, it requires no user input. (Our image capture instrumentation provides both an adjacency map for the mosaic, and an initial rotation estimate for each image.) Second, it optimizes an objective function based on a global correlation of overlapping image regions. Third, our representation of rotations significantly increases the accuracy of the optimization. Finally, our representation and use of adjacency information guarantees globally consistent rotation estimates. The algorithm has proved effective on a collection of nearly four thousand images acquired from more than eighty distinct optical centers. The experimental results demonstrate that the described global optimization strategy is superior to non-global aggregation of pair-wise correlation terms, and that it successfully generates high-quality mosaics despite significant error in initial rotation estimates.