Lunar X-view dataset

The "Lunar X-view" dataset is the official dataset of ESA's Kelvins Lunar X-view challenge, created in a collaborative effort by researchers of Institute of Science Tokyo, University of Adelaide and Politecnico di Milano. It consists of computer generated aerial maps of two regions of interest, accompanied by 360° panoramic groundview images, depicting a lunar-like surface. Each region's map and panoramas were rendered from the same underlying terrain model using the Planet and Asteroid Natural Scene Generation Utility (PANGU) developed by the University of Dundee's Space Technology Centre. All 10 808 panoramas of the train folder of this data come with a precise position label, while the 10 000 panoramas of the test folder are provided without any label. The panoramas viewpoint is always 2.5m above ground with zero pitch and roll, while the aerial maps provide a 5m/pixel ground resolution. The goal of the related competition is to find cross-view correspondences allowing to pinpoint all the viewpoint location of the panoramas for the unlabeled region of interest (test). Competition rules, submission instructions and evaluation details are available on the Kelvins competition platform. An earlier and unpublished prototype of this dataset has been documented by the following publication: H.M. Nguyen, M. Märtens, TJ. Chin., “Learning Cross-view Correspondences for Geo-localization on Planetary Surfaces,” Proceedings of the 3rd Conference on AI in and for Space (SPAICE 2026), European Space Research and Technology Centre (ESTEC), Noordwijk, Netherlands, 2026, pp. 91–95. Intended Purpose The primary purpose of this dataset is to provide the foundation for the Lunar X-view challenge. Consequently, not all details and labels can be publicly disclosed. The dataset could additionally facilitate various scientific studies related to computer vision, rover navigation, domain gaps and more. Details about dataset generation The geometrical model of this data is based on high-resolution LROC/NAC digital elevation models (DEMs) of the Moon. The geographical location of the DEMs remains undisclosed and several undisclosed perturbations have been applied to reduce the possibility of identifying the source terrain. To provide more realistic horizon lines, the NAC DEMs were embedded into lower resolution LROC/WAC DEMs that extend beyond the labeled regions, mitigating border effects. Aerial map data has been generated with an orthographic camera. PANGU was used to generate small-scale features such as rocks and synthetic craters not present in the DEM to enrich representation. The surface has been textured using a texture file based on the Yutu Moon Rover, displaying original lunar regolith. Surface reflectance was simulated using a weighted combination of Hapke (0.9) and Lambertian (0.1) reflectance models (BRDF), including opposition-effect behaviour. The illumination of the panoramas varies based on the local time. Solar azimuth and elevation were calculated using SPICE kernels for an assumed near-polar location. These illumination settings do not disclose the geographic origin of the underlying terrain. Both aerial maps have a fixed sun azimuth of 90° and sun elevation of 20°. The inconsistent lighting between aerial maps and panoramas is intended to stimulate the development of robust cross-view localization approaches. Each panorama was generated by stitching 16 separate images into a 360° representation by cylindrical projection. The central image column of each panorama corresponds to the map north. To simulate sensor noise, all images have been degraded using a mix of photon shot noise, readout noise, and dark current noise. Known limitations In order to ensure a fair competition, several details such as the underlying DEMs and labels are withheld The dataset contains synthetic imagery from two terrain regions, with labels provided for one region. Performance on this dataset alone does not establish generalization to other lunar terrains or real rover imagery. The rendering of the sun, planets and stars in the sky is kept simplistic. No lens flares or other optical disturbances have been inserted. Very far features of the horizon might not be captured due to the limited size of the LROC/WAC DEM. The texture for the moon regolith has been used repeatedly, leading to an overly monotonous look. Repeating patterns can be observed due to repetition of the texture file. The aerial maps include noticeable seamlines where the LROC/NAC and LROC/WAC DEM are stitched together. Due to the rendering with different resolution layers, certain regions in the aerial map may appear blurry or with more small-scale craters than others. Seamlines between such resolution layers might also be visible in some panoramas. Minor distortions (such as repeated pixels) due to the stitching of the panoramas with cylindrical projection could not be avoided. Isolated groups of a few small white pixels may be found in panoramas as a result of rendering issues. Ethical Use and Non-Endorsement This dataset has been developed and released for the advancement of scientific research, education, and peaceful technological development. The authors strongly oppose the use of this dataset, directly or indirectly, for military purposes, including weapons development, targeting systems, military surveillance, and other applications intended to facilitate armed conflict. We explicitly request that researchers, institutions, and commercial entities refrain from using this dataset for such purposes.