Rectification and pose recovery for spherical images

Abstract

The experience of navigation in a virtual environment using image-based rendering can be greatly improved by the use of spherical images or 360° panoramas. These images carry more information and can be represented in different convenient formats. This thesis addresses classical vision problems, in those of rectification and pose recovery, applied to cubic panoramas or cubes, one of the possible formats of spherical images. As an introduction to the concept of rectification, we explore a solution to the problem of rectifying triplets for limited field-of-view (FOV) cameras using the fundamental matrix and a homography composition. A solution to cubic panorama rectification is then given as an extension of the latter and based on the essential matrix. Finally, the study of cubic panoramas geometry is completed by a two-stage algorithm to solve the pose recovery problem for a given set of cubes based on error minimization procedures. Results are obtained for different sets of images and different configurations to corroborate the fore mentioned algorithms. In a more practical way, the navigation of a user in sphericalimage-based models of environments can be improved by the results of the presented study.