RegGR1945 & CycleGAN Training Patches: Datasets for Cross-Temporal Image Registration and Appearance Translation between Historical Aerial Photographs and Satellite Imagery

This deposit contains three datasets accompanying the paper "A Window to the Past: Connecting Historical Aerial Photographs to Satellite Images with the use of Image Registration and Enhancement." Together they support research on cross-temporal image registration, historical image enhancement, and unpaired image-to-image translation between historical aerial photographs and modern satellite imagery. The deposit consists of three zip files: RegGR1945_2015 is a benchmark dataset for cross-temporal image registration between historical aerial photographs (1945) and satellite imagery (2015) of Greek areas. It contains 50 image pairs with ground-truth registration coordinates and visualization overlays. Transformations between image pairs are translational only. The recommended evaluation threshold is τ = 0.93. Historical aerial photographs and satellite images were obtained from the Hellenic Cadastre Geoportal (https://www.ktimatologio.gr). RegGR1945_2025 is a benchmark dataset for cross-temporal image registration between historical aerial photographs (1945) and satellite imagery (2025) of Greek areas. It contains 30 image pairs with ground-truth registration coordinates. Due to licensing considerations regarding the modern imagery, this dataset is distributed without the 2025 satellite images themselves, for which a shapefile and txt coordinates of the bounding boxes of the images are provided. This dataset presents a higher difficulty level than RegGR1945_2015, as image pairs involve combined transformations of shift, anisotropic scale, and rotation. The recommended evaluation threshold is τ = 0.90. Historical aerial photographs were obtained from the Hellenic Cadastre Geoportal (https://www.ktimatologio.gr). The modern satellite imagery is not redistributed with this dataset; users must retrieve it themselves from a source of their choice and comply with the terms of that source. The two registration benchmark datasets together form a two-level benchmark allowing direct comparison of method performance across difficulty levels. Ground-truth coordinates are provided as 8 floating-point values representing the 4 corner coordinates of the aerial photograph's bounding box within the satellite image coordinate system, listed clockwise from top-left. All images in both datasets were used exclusively for registration evaluation. CycleGAN_training_patches contains the unpaired image patches used for training the CycleGAN appearance translation model described in the paper. It includes 1,000 grayscale patches of 256×256 pixels from historical aerial photographs of 1945 (trainA) and 1,000 color patches of 256×256 pixels from satellite images of 2015 (trainB). The patches in trainA and trainB were extracted exclusively from aerial photographs and satellite images that are not included in either registration benchmark dataset, ensuring fully independent evaluation with no data leakage between training and test data. Each zip file contains a detailed README.txt file with full documentation of the folder structure, file formats, ground truth format, evaluation metrics, data sources, and intended use. The 1945 aerial photographs and 2015 satellite images are redistributed under the Hellenic Cadastre terms of use (https://www.ktimatologio.gr/oroi-xrisis). Required attribution: "Data source: Hellenic Cadastre" (in Greek: "Πηγή δεδομένων: Ν.Π.Δ.Δ. ΕΛΛΗΝΙΚΟ ΚΤΗΜΑΤΟΛΟΓΙΟ"). License (for the derived works — registration ground truth, training patches, splits, and documentation): Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). If you use this dataset, please cite: Christopoulou, A., Gatos, B. & Kakogeorgiou, I. (2026). A Window to the Past: Connecting Historical Aerial Photographs to Satellite Images with the use of Image Registration and Enhancement. In International Conference on Pattern Recognition. Cham: Springer Nature Switzerland.