The inverse compositional algorithm for parametric registration

Abstract

We present an implementation of the inverse compositional algorithm for parametric motion estimation. It computes a global motion between two images using a non-linear least square technique. Our implementation allows computing several types of planar transformations, such as translations, similarities, affinities or homographies. The algorithm is iterative so it typically yields solutions with high accuracy. The use of robust error functions, different from the L2 norm, improves the stability of the method under the presence of noise and occlusions, and allows it to detect the predominant motion, even if there are several types of displacements. The method works with multi-channel images and makes use of a coarse–to–fine strategy for dealing with large displacements.