Get Better 1 Pixel PCK: Ladder Scales Correspondence Flow Networks for Remote Sensing Image Matching in Higher Resolution

Weitao Chen, Zhibin Wang, Hao Li; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2021, pp. 742-751

Recently, remote sensing image matching by deep learning reaches competitive performance evaluated by Probability of Correct Keypoints(PCK). Percentage of image size is often used as the threshold of PCK. Even though it can achieve a good 1% PCK in high resolution by regression of transformer parameters,the value will be reduced by using the absolute 1 pixel as threshold in the higher resolution. Inspired by the flow-based methods used in natural image matching tasks, we convert the transformer to correspondence flow and propose ladder scales correspondence flow networks(LSCFN) to get better 1 pixel PCK in higher resolution.Input images are resized to multi scales and then sent to network backbone to generate multi feature pyramids. These pyramids are linked and effectively pull up the highest resolution of original backbone just like a ladder when the global correlation scale is fixed.LSCFN regress correspondence flow in ladder scales by a dense cascade way.We build LSCFN-b and LSCFN-s based on the degree of semantic change between compared images. One with only global correlation is used for the big change, another with global and local correlation is used for the opposite one.The proposed LSCFN achieve state-of-the-art performance evaluated by 1% of image size PCK and absolute 1 pixel PCK on google earth dataset.