OSDataset2.0: Optical-SAR Image Matching Dataset and Evaluation Benchmark

OSDataset2.0, a large-scale benchmark dataset designed for optical-SAR image matching tasks. The dataset consists of two components: a local training dataset and a full-scene test set. The local training dataset provides 6,343 registered image patch pairs of 512×512 pixels (with a spatial resolution of 1 meter), covering nearly 20 countries worldwide. The dataset was generated through multi-method matching combined with manual verification. The full-scene test set includes a pair of full-scene optical and SAR images in GeoTIFF format, with a spatial resolution of 0.43 meters and data sizes of 35,783x35,783 and 35,507x27,298 pixels. The ground truth data was collected based on the consistency principle of imaging mechanisms for streetlight pole structures in both optical and SAR payloads, resulting in 72 high-precision ground truth corresponding point pairs. We also provide accompanying general evaluation code, supporting quantitative analysis of registration accuracy for any matching algorithm.