Panorama interpolation for image-based navigation

Abstract

This thesis presents methods for novel image synthesis from cubic panoramas taken with multi-sensor cameras. The pre-captured cubic panoramas are used to interpolate arbitrary views to allow a virtual walkthrough of the remote real environment. In our approach, the "transfer" and "triangulation" methods are adopted to analyse the geometry of cubic panoramas and recover an accurate essential matrix between cubes. To generate a novel view between two aligned cubes, a warping model is applied to warp cubes to approximate navigation. This technique, called cube warping, works by simplifying the model of pixel displacements between cubes. A new raytracing-like image-based interpolation method is also proposed for free-viewpoint cube synthesis. Instead of attempting to recover dense reconstruction precisely, our method tries to reconstruct colours with colour invariance constraints. Due to the fact that photo consistency has more to do with colour than shape, our algorithm can generate a complete novel scene view with maximized photo consistency.