We focus on the practical challenge of segmenting new query fundus images that may be dissimilar to existing training images. It is addressed in this paper by a supervised learning pipeline. The core lies in the construction of a synthetic fundus image dataset by our proposed R-sGAN simulation method. The images are realistic-looking in terms of the query images, while maintaining the annotated vessel structures of the existing set. This helps to bridge the mismatch between testing and training images, thus enables the direct utilization of existing supervised fundus segmentation techniques on the query images, even without the use of corresponding manual annotations. Extensive experiments are carried out on distinct fundus image datasets, %including in particular those poorer quality images acquired by low-cost mobile fundus device.where experimental results demonstrate the competitiveness of the proposed approach in dealing with a diverse range of mismatch settings.