Robust Matching of Aerial Images with Low Overlap

This paper addresses the problem of aerial image matching. We analyze existing approaches to this problem and show that, though the modern algorithms cope with the task quite well, their results are deteriorated in case of low overlap and significant rotation angle between the images. A two-stage feature-based image matching scheme is presented. It is shown that preliminary stage of simplified (shift-rotation) model estimation is crucial in case of low overlap and influences significantly the whole matching scheme. We introduce a novel method of shift-rotation model estimation, based on the voting procedure in parameter space, which allows finding the correct model in case of extremely difficult input data (overlap area is less than 10%) without using any additional information. Finally, the experimental results of our method on both synthetic and real data are presented. We compare our results with one of the state of the art SAC-based model estimators and show that our algorithm outperforms existing methods in case of very small overlap.