Geometric and photometric affine invariant image registration

Abstract

This thesis aims to present a solution to the correspondence problem for the registration of wide-baseline images taken from uncalibrated cameras. We propose an affine invariant descriptor that combines the geometry and photometry of the scene to find correspondences between both views. The geometric affine invariant component of the descriptor is based on the affine arc-length metric, whereas the photometry is analysed by invariant colour moments. A graph structure represents the spatial distribution of the primitive features; i.e. nodes correspond to detected high-curvature points, whereas arcs represent connectivities by extracted contours. After matching, we refine the search for correspondences by using a maximum likelihood robust algorithm. We have evaluated the system over synthetic and real data. The method is endemic to propagation of errors introduced by approximations in the system.