Transformation of Self Affine Features due to Windowing Effects

Descriptor-based image registration consists of many processing stages. Errors introduced by the keyframe localization stage (whose magnitude correlates with the disparity between the images) are usually neglected, or corrected by computationally intensive optical flow methods. Results presented by this paper target a faster solution to this problem. It is based on the investigation, how the Self Affine Feature Transform (SAFT descriptor) changes due to affine transformations of the window of interest. The analytic nature of SAFT makes it attractive for this purpose. For example, it is possible to apply coordinate transformation after descriptor extraction. Window and coordinate transformation rules of a descriptor need to be combined to approximate the descriptor of transformed image details. Using Gaussian windows helps to achieve our goal, due to the special behavior of these windows respect to slight affine transformations. This ensures that the transformation dependence can be efficiently approximated linearly which enables to approximate the affine transformation parameters between two image details by solving a least-squares (LS) problem. The paper shows results obtained by the introduced algorithm.