Dataset and method for evaluating optical-to-sar image registration algorithms based on keypoints

Image registration is the alignment (i.e. finding a general coordinate system) of two or more images of the same scene. Complex case of this task is the multimodal image registration, for example, optical-to-SAR. The need for such registration appeared in image fusion and object detection. The research on optical-to-SAR image registration shows that there are no publicly available datasets with the enough size for debugging and testing algorithms. In many works, testing is carried out on a several pairs of images that are not always publicly available. Moreover, different works use different datasets. In this paper we present a dataset of 100 aligned optical-SAR pairs of images. Additionally, we consider methods for evaluating keypoints repeatability, the accuracy of their matching, and a method for calculating the accuracy of image registration for optical-SAR data. Using these methods, we compare the results of these methods for the classical SIFT, YAPE and Harris detectors and SIFT, ORB and SURF descriptors on the presented dataset.