Image registration and a metric to assess the transformation function

Image registration has been a broadly applied topic across the photogrammetric/remote sensing and computer vision communities. It is a foundational step for many applications such geopositioning, data fusion, change detection, conflation, and object recognition and extraction. The efficacy of many automated geospatial processes can be limited or nullified by an inadequate registration process. The task of automated image registration presents two main challenges: 1) establishing image-to-image correspondence through feature matching, and 2) determining an appropriate transformation model for a given registration scenario. When imaging 3D environments, a goal of the transformation function is to accurately relate the 2D pixel spaces of candidate images with potential geometric distortions and surface discontinuities projected from a 3D object space. When sensor model metadata and 3D surface information is available (e.g. a digital surface model), a 3D-to-2D photogrammetric transformation will generally provide the most reliable registration solution. Moreover, photogrammetric solutions propagate error to provide a statistically rigorous estimation of registration accuracy. On the other hand, direct 2D-to-2D transformations such as affine, homographic, and polynomials are often used when sensor metadata and/or object space information is limited or unavailable. Owing to their convenience of use and implementation, direct 2D-to-2D registration methods abound in commercial software application. However, such registration solutions are generally more suspect in terms of accuracy and uncertainty estimation. Nonetheless, they do have practical utility, provided appropriate care is exercised in their application. The goal of this paper is to quantitatively demonstrate different scenarios and solutions that users should consider when applying 3D-to-2D photogrammetric versus direct 2D-to-2D image registration methods.