MP-Loc: a multi-step progressive localization framework for UAV-satellite imagery

UAV-satellite visual localization is challenging under cross-scale and cross-view conditions because differences in scale, resolution, viewing geometry, and acquisition time weaken feature consistency. Existing retrieval, object-level, and pixel-level matching methods often struggle to maintain both accuracy and robustness under large viewpoint and scale changes. To address these challenges, we propose MP-Loc, a multi-step progressive framework that uses progressive refinement to rectify the satellite image, regenerate correspondences, and update the geometric transformation for more accurate and robust localization. The main contributions of this work are summarized as follows: (1) A Multi-step Progressive Localization framework (MP-Loc) is proposed as a robust solution for UAV absolute visual localization. (2) We construct MP-Dataset, comprising 249 annotated UAV-satellite localization instances from five geographical regions with diverse scale and viewpoint variations. (3) We propose the Normalized Precise Localization Score (NPLS), which jointly measures localization coverage and error among successful localizations. On MP-Dataset, MP-Loc achieves an LSR of 84.10%, a CLA@2 of 59.41%, and an MAE of 1.93 m among successful localizations, increasing NPLS from 45.98 to 58.38 over the MINIMA–RoMa baseline. On the public MTA-Dataset, progressive refinement improves NPLS at all six tested cycle-consistency thresholds. The dataset and code will be made publicly available.