Comparison of local descriptors for image registration of geometrically-complex 3D scenes

Image registration is an important process in a number of machine vision applications. It is often used as a pre-processing step to gain a better understanding of the images and many techniques have been proposed to better register a set of images without user intervention. The performance of these techniques are often scene-dependent and a technique designed for one application often performs less favourably under a different condition. In this paper, the performance of a number of local descriptor methods which have been proposed for image registration are studied, in particular, the main focus is on the techniques based on the SIFT descriptor for use with Maori artefacts. These techniques have been previously studied and shown to have good performance in the case of planar scenes or scenes which are far away from the camera, however little data exist for geometrically-complex 3D scenes. The experimental setup used in the work which allows for a fair and accurate comparison of the techniques under a number of different conditions is presented. The results from the work are presented and the reasons for the poor performance of the descriptors under the given scenes and conditions are discussed. Finally, based on the results obtained, the proposed approach for automatic registration of images of Maori artefacts are presented.