MoonAnything: A Vision Benchmark with Large-Scale Lunar Supervised Data

Clémentine Grethen; Yuang Shi; Simone Gasparini; Géraldine Morin

cs.CVApr 1, 2026

MoonAnything: A Vision Benchmark with Large-Scale Lunar Supervised Data

Authors: Clémentine Grethen, Yuang Shi, Simone Gasparini, Géraldine Morin

Organizations: IRIT - Université de Toulouse, France · National University of Singapore, Singapore · IPAL, IRL2955, Singapore

Abstract

Accurate perception of lunar surfaces is critical for modern lunar exploration missions. However, developing robust learning-based perception systems is hindered by the lack of datasets that provide both geometric and photometric supervision. Existing lunar datasets typically lack either geometric ground truth, photometric realism, illumination diversity, or large-scale coverage. In this paper, we introduce MoonAnything, a unified benchmark built on real lunar topography with physically-based rendering, providing the first comprehensive geometric and photometric supervision under diverse illumination with large scale. The benchmark comprises two complementary sub-datasets : i) LunarGeo provides stereo images with corresponding dense depth maps and camera calibration enabling 3D reconstruction and pose estimation; ii) LunarPhoto provides photorealistic images using a spatially-varying BRDF model, along with multi-illumination renderings under real solar configurations, enabling reflectance estimation and illumination-robust perception. Together, these datasets offer over 130K samples with comprehensive supervision. Beyond lunar applications, MoonAnything offers a unique setting and challenging testbed for algorithms under low-textured, high-contrast conditions and applies to other airless celestial bodies and could generalize beyond. We establish baselines using state-of-the-art methods and release the complete dataset along with generation tools to support community extension: https://github.com/clementinegrethen/MoonAnything.

Figures & tables

Dataset	Type	Region	Stereo	Depth	Reflect.	Multi-Light	Phys. Render	Scale
LRO NAC ( Vondrak et al., 2010 )	Real/Orbital	Global	×	×	×	×	N/A	2M images
Chang’E ( Wang et al., 2024b )	Real/Surface	Global	×	×	×	×	N/A	7.5K images
DLR’s TRON ( Lebreton et al., 2024a )	Lab	Mockup	✓	×	×	×	×	7.2K images
POLAR ( Wong et al., 2017 )	Lab	Mockup	✓	✓	×	×	×	2.6K pairs
LuSNAR ( Liu et al., 2024 )	Synthetic	Simulated	✓	✓	×	×	×	13K images
StereoLunar ( Grethen et al., 2025 )	Synthetic	South Pole	✓	✓	×	×	✓	50K pairs

Table 1. Comparison of lunar perception datasets. MoonAnything is the first to provide comprehensive geometric and photometric supervision with diverse illumination.

Figure 1. Overview of the MoonAnything dataset generation pipeline.

South Pole
Alt. ( $\mathrm{km}$ )	3.5	6.2	9.5	12.8	16.1	19.4	22.7	26.0	29.2	30.5
GSD ( $\mathrm{m}\text{/}\mathrm{px}$ )	5.7	10.0	15.4	20.7	26.0	31.4	36.7	42.1	47.2	49.3
Tycho Crater
Alt. ( $\mathrm{km}$ )	3.5	4.1	4.7	5.3	5.9	6.5	7.1	7.7	8.3	8.9
GSD ( $\mathrm{m}\text{/}\mathrm{px}$ )	3.7	4.3	4.9	5.5	6.2	6.8	7.4	8.1	8.7	9.3

Table 2. Altitude versus effective GSD for LunarGeo .

Dataset	Method	Nadir			Oblique			Dynamic
Dataset	Method	ACC.(m) $\downarrow$	Compl.(m) $\downarrow$	Chamfer(m) $\downarrow$	ACC.(m) $\downarrow$	Compl.(m) $\downarrow$	Chamfer(m) $\downarrow$	ACC.(m) $\downarrow$	Compl.(m) $\downarrow$	Chamfer(m) $\downarrow$
Tycho (Unseen)	MASt3R	60	58	59	128	88	108	49	45	47
	VGGT	39	39	40	95	78	87	75	51	63
	VGGT FT	35	35	35	76	60	68	47	52	50
	MASt3R FT	37	37	37	72	57	64	34	32	33
S. Pole (Seen)	MASt3R	236	235	236	385	259	322	289	270	279

Table 3. Stereo 3D reconstruction results on LunarGeo . Best results are highlighted in green and second-best in orange .

Figure 2. Examples of LunarGeo illustrating the three trajectory types. The first row shows stereo pairs and their associated 3D scene over the lunar South Pole, rendered using the Hapke BRDF with constant parameters. The second row presents two stereo pairs from the Tycho crater: the first rendered using the classical Hapke BRDF, and the second using SVBRDF from ( Grethen et al., 2026 ) .

Region	Total	Train ( $80\text{\,}\mathrm{\%}$ )	Val ( $10\text{\,}\mathrm{\%}$ )	Test ( $10\text{\,}\mathrm{\%}$ )
South Pole	38K	30K	4K	4K
Tycho	20K	16K	2K	2K

Table 4. LunarGeo dataset statistics and train/val/test splits.

Item	Description
Total samples	84K samples
Split	67K train / 8.5K val / 8.5K test
Sun positions (multi-lighting)	9 per sample
Rendered images	750K per BRDF model
Geometry	dem.tif , depth.tif , normal.tif
Real data	real_image.tif (orthorectified LRO NAC image)

Table 5. LunarPhoto statistics and per-sample content.

Figure 4. Examples of challenging cases: (a) extreme shadowing observed in the multi-lighting data from LunarPhoto , (b) a LunarGeo pair from the Tycho under raking lighting, and (c) a flat stereo pair from the S. Pole captured by LunarGeo .

Explore similar work

SomBench: Benchmark Dataset for Advancing Machine Learning in Lunar Science
Sep 8, 2026Himanshu Patil, Gabby Nyirjesy, Rachel A. Slank +19Benchmark Design Scientific ML
The University of Alabama in Huntsville, Huntsville, AL, USA · IBM Research, Yorktown Heights, NY, USA · Science and Technology Institute, Universities Space Research Association (USRA), Huntsville, AL, USA +7
LLaVA-LE: Large Language-and-Vision Assistant for Lunar Exploration
Mar 25, 2026Gokce Inal, Pouyan Navard, Alper YilmazVision-Language Models Large Vision-Language Models
The Ohio State University Columbus, OH, USA
MoonGS: High-quality Representation of the Lunar Surface via Gaussian Splatting Using Robust Depth Features from Image Pairs
Oct 5, 2026Yun Jiang, Bo Zheng, Yingying Zhang +73D Gaussian Splatting 3D Reconstruction
Jilin University, Changchun, China · Shanghai Aerospace Control Technology Institute, Shanghai, China · Beijing Institute of Control Engineering, Beijing, China +2

Sep 8, 2026J/K move · Enter open · S save

SomBench: Benchmark Dataset for Advancing Machine Learning in Lunar Science

Himanshu Patil, Gabby Nyirjesy, Rachel A. Slank, Vishal Gaur+18

The University of Alabama in Huntsville, Huntsville, AL, USA · IBM Research, Yorktown Heights, NY, USA · Science and Technology Institute, Universities Space Research Association (USRA), Huntsville, AL, USA+7

Benchmark Design Scientific ML Remote Sensing

Open paper AskPDF arXiv

Lunar orbital missions, such as Lunar Reconnaissance Orbiter, Kaguya/SELENE, Gravity Recovery and Interior Laboratory, and Lunar Prospector, among others, provide rich multi-instrument observations, but their heterogeneity in sampling, projection, and conventions limits reproducible machine learning (ML). We introduce SomBench, a unified, spatially-aligned, ML-ready lunar dataset aggregating 30+ co-registered layers from ten instruments across four missions, spanning 1 meter to 20 kilometer/pixel and covering 82 degree latitude in 90 Lunar Transverse Mercator zones with two polar stereographic caps. An image-anchored tiling pipeline yields pretraining-ready multimodal tile views with leakage-safe splits, distributed as netCDF with Parquet catalogs. An application benchmark suite spans impact processes, volcanic history, and polar volatiles. Baseline experiments with ResNet-50 and SwinV2-B models confirm that each benchmark task is learnable from the released inputs, establishing reference points for future model development.

Abstract

Explore similar work

SomBench: Benchmark Dataset for Advancing Machine Learning in Lunar Science

LLaVA-LE: Large Language-and-Vision Assistant for Lunar Exploration

MoonGS: High-quality Representation of the Lunar Surface via Gaussian Splatting Using Robust Depth Features from Image Pairs