UAVPairs: A Dataset for Match Pair Retrieval of Large-scale UAV Images

The dataset contains 21,622 high-resolution UAV images collected from 30 representative remote sensing scenes, covering urban areas, rural farmland, river corridors, mountainous regions, building clusters, and mixed land-cover environments. All images are acquired from real UAV surveying projects with significant viewpoint and scale variations.

To ensure reliable geometric annotations, UAVPairs employs a Structure-from-Motion (SfM) reconstruction pipeline to generate sparse 3D models and image tracks. The geometric similarity of image pairs is defined by the number of shared reconstructed 3D points, providing accurate matchability labels for image retrieval and matching tasks. Compared with existing UAV retrieval datasets, UAVPairs offers high-resolution imagery, scene-level organization, scalable annotations, and direct applicability to image matching and 3D reconstruction workflows.