Exploring the Potential of Unsupervised Image Synthesis for SAR-Optical Image Matching

We consider SAR-optical image matching problems, where correspondences are acquired from a pair of SAR and optical images. Recent methods for such problem typically simplify the to SAR-SAR or optical-optical matchings using supervised-image-synthesis methods. However, training needs plenty aligned pairs while gathering sufficient amounts multi-modal is challenging in remote sensing. In this work, we investigate applicability unsupervised-image-synthesis that unaligned images could be used. To end, apply feature loss well known method, i.e., CycleGAN, enforce consistency. Moreover, develop shared-matching-strategy improve results further. Qualitative comparisons against StarGAN, DualGAN demonstrate superiority our approach. Quantitative show that, compared with DualGAN, method obtains at least 2.6 times more qualified matchings.