Registration, calibration and blending in creating high quality panoramas

This paper presents a system for creating a full 360-degree panorama from rectilinear images captured from a single nodal position. The solution to the problem is divided into three steps. The first step registers all overlapping images projectively. A combination of a gradient-based optimization method and a correlation-based linear search is found to be robust even in cases of drastic exposure differences and small amount of parallax. The second step takes the projective matrices and their associated hessian matrices us inputs, and calibrates the internal and external parameters of every images through a global optimization. The objective is to minimize the overall image discrepancies in all overlap regions while converting projective matrices into camera parameters such as focal length, aspect ratio, image center, 3D orientation, etc. The third step re-projects all images onto a panorama by a Laplacian-pyramid-based blending. The purpose of blending is to provide a smooth transition between images and eliminate smell residues of misalignments resulting from parallax or imperfect pairwise registrations. The blending masks are generated automatically through the grassfire transform. At the end, we briefly explain the necessary human interface for initialization, feedback and manual options.