Lunar-G2R dataset

Lunar-G2R Training  Dataset: DEM-Images for computer vision applications

 Paper for explanation on the construction and the usage of this dataset: https://arxiv.org/abs/2601.10449

The Lunar-G2R training dataset is designed to support learning-based approaches for lunar surface understanding, geometry–appearance modeling, and reflectance estimation. It provides paired samples associating local lunar terrain geometry with real optical observations, together with detailed acquisition metadata.

An example of usage of this dataset in présented on our paper for Lunar SVBRDF estimation

If you find our work useful, please cite:

@misc{grethen2026lunarg2rgeometrytoreflectancelearninghighfidelity,
      title={Lunar-G2R: Geometry-to-Reflectance Learning for High-Fidelity Lunar BRDF Estimation}, 
      author={Clementine Grethen and Nicolas Menga and Roland Brochard and Geraldine Morin and Simone Gasparini and Jeremy Lebreton and Manuel Sanchez Gestido},
      year={2026},
      eprint={2601.10449},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2601.10449}, 
}

Code https://clementinegrethen.github.io/publications/Lunar-G2R

1) Dataset Construction

Each sample consists of a Digital Elevation Model (DEM) crop and a corresponding orthorectified lunar image:

• DEM patches are extracted from a large-scale DEM covering the Tycho crater region.

  • Patch size: 128 × 128 pixels

  • Ground Sampling Distance (GSD): 5 m/px

  • Physical coverage: approximately 0.4 km² per patch

• For each DEM crop, a real lunar image acquired by the Lunar Reconnaissance Orbiter (LRO) is selected.

  • Source images come from the LRO Narrow Angle Camera (NAC)

  • Native resolutions range from 0.5 m/px to 2 m/px

Each selected LRO image is:

1. Orthorectified onto the local DEM,

2. Cropped to exactly match the DEM spatial extent,

3. Stored as a ground-truth appearance image aligned with the terrain geometry.

This process yields precisely aligned geometry–appearance pairs, suitable for supervised learning.

2) Dataset Size and Splits

In total, the dataset contains 83,614 DEM–image pairs, each of size 128 × 128 pixels, distributed as follows:

• Training set: 66,662 samples

• Validation set: 8,615 samples

• Test set: 8,337 samples

To prevent spatial leakage, dataset splits are performed geographically after pair generation:

• The global DEM is divided into geographic tiles

• All samples whose centers fall within the same tile are assigned to the same split

• This ensures strict spatial separation between training, validation, and test data

The dataset therefore contains over 80,000 samples, with a robust and physically meaningful split strategy.

Three text files are provided:

• train.csv

• val.csv

• test.csv

Each file lists the samples belonging to the corresponding split and explicitly documents the geographic partitioning strategy used.

3) Data Organization

The dataset is organized into the following main components:

• dem.zip

  • Contains all DEM patches (128 × 128 px, 5 m/px)

  • Each DEM corresponds to a unique terrain location

• render.zip

  • Contains the orthorectified and cropped LRO images

  • Each image is aligned pixel-wise with its associated DEM patch

• metadata/

  • Contains a JSON file per DEM–image pair

  • Stores full geometric, photometric, and acquisition metadata

Metadata Description

For each sample, the associated metadata provides detailed information on:

Image and Projection Parameters

• Source LRO image file (e.g. NAC image identifier)

• Field of view (fov)

• Image resolution and original image size

• Projection model (equirectangular)

• Projection parameters (latitude/longitude bounds, scale, translation)

Camera Geometry

• Camera position (cam_pos)

• Camera orientation as a quaternion (cam_att)

• Image footprint on the lunar surface (top-left, top-right, bottom-left, bottom-right)

• Sub-spacecraft latitude and longitude

Illumination Geometry

• Sun position in 3D space (sun_pos)

• Incidence angle

• Emission angle

• Phase angle

• Solar azimuth and north azimuth

• Sub-solar latitude and longitude

LRO / PDS Metadata

• LRO image ID and name

• Mission phase

• Orbit number and orbit node

• Acquisition start and stop time

• NAC image resolution

• Image dimensions

• Slew angle and flight direction

An excerpt of a typical metadata JSON includes fields such as:

• texture_file

• cam_pos, cam_att

• sun_pos

• incidence_angle, emission_angle, phase_angle

• center_lat, center_lon

• Full PDS metadata from the original LRO product

This rich metadata enables precise photometric modeling, BRDF learning, geometry-aware vision tasks, and physically grounded rendering or simulation pipelines.

Intended Use

The Lunar-G2R training dataset is suitable for:

• Geometry-to-appearance learning

• Lunar BRDF estimation

• Shape-from-shading and reflectance modeling

• Vision-based navigation (VBN)

• DEM-conditioned image synthesis

• Cross-modal learning between terrain geometry and optical imagery

Additional implementation details and preprocessing steps are described in Section 1 of the supplementary material of the associated paper!

Contact: clementine.grethen@irit.fr