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

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

1)Introduction

MoonAnything, a unified benchmark for lunar surface perception that addresses the lack of comprehensive datasets
combining geometric and photometric supervision. By unifying and extending our prior works (Lunar-G2R, StereoLunar) MoonAnything provides stereo imagery of unprecedented dataset size with dense depth maps across two lunar regions (LunarGeo) and spatially-varying BRDF parameters with multi-illumination renderings (LunarPhoto). Together, these sub-datasets offer the first benchmark enabling research on 3D reconstruction, reflectance estimation, and illumination-robust perception within a consistent lunar context.

2) LunarGeo: 

LunarGeo provides stereo image pairs with dense depth supervision for 3D reconstruction research and training. It is generated
using real Lunar data that covers both the South Pole and Tycho crater regions, using the common rendering pipeline with regionspecific camera, illumination, and reflectance configurations. For the South Pole, we render using the Hapke BRDF with constant albedo, providing consistent photometric behavior across the region. For Tycho crater, we employ both the base Hapke model and the SVBRDF by using LunarG2R framework

It covers a broad range of lunar terrains and illumination conditions, offering diverse altitudes, viewpoints, and camera trajectories around the lunar South Pole, & Tycho crater. The dataset contains stereo image pairs, each with dense ground-truth supervision, including pixel-level depth maps and accurate camera poses. These pairs were generated using a combination of:

• High-resolution Digital Elevation Models (DEMs),

• A Bidirectional Reflectance Distribution Function (BRDF) for surface reflectance,

• Realistic solar illumination configurations,

• A parametric camera model simulating various orbital and descent trajectories.

Key Features:

• Physical Realism: The dataset employs the Hapke BRDF model to simulate lunar surface reflectance, accounting for unique phenomena like the opposition effect and anisotropic scattering. The illumination setup includes side, overhead, and backlighting to generate varying shadow patterns and photometric changes, mimicking real-world conditions.

• Geometric Supervision: Every stereo pair comes with dense ground-truth depth maps and precise camera extrinsic parameters, providing fully supervised data for 3D reconstruction tasks.

• Diverse Trajectories: The dataset includes three types of camera motion:

  • Nadir (vertical descent) with minimal parallax.

  • Oblique (tilted camera views) with varying altitudes and baselines.

  • Dynamic motion featuring additional camera variations, such as altitude and viewpoint changes, to simulate real lunar descent sequences.

• High Coverage: Stereo pairs are distributed across 10 altitude bands,

Geometric Perception Tasks: The stereo pairs with dense depth supervision support stereo matching, multi-view 3D reconstruction, and camera pose estimation. The diverse viewing configurations (nadir, oblique, dynamic) and altitude ranges make LunarGeo suitable for evaluating algorithms under realistic descent conditions. These capabilities directly support mission-critical applications, such as hazard detection and avoidance, where accurate depth estimation enables the identification of rocks, craters, and unsafe slopes, and terrain-relative navigatio

More details on the readme_geo.txt.

3) LunarPhoto

LunarPhoto provides paired geometry and reflectance data for appearance modeling research. It extends the original LunarG2R dataset , which provided geometry and reflectance pairs but only single-illumination observations per sample. To the best of our knowledge, this is the first dataset providing spatially-varying BRDF parameters with diverse multi-illumination supervision. The sub-dataset covers the Tycho crater region, combining real LRO observations with physically-grounded multi-lighting augmentation

Data Description

Each sample consists of:

• A DEM patch

  • Size: 128 × 128 pixels

  • Ground Sampling Distance: 5 m/px

  • Physical coverage: ~0.4 km²

  • Extracted from a large-scale DEM covering the Tycho crater region

• A corresponding real lunar image

  • Acquired by the LRO Narrow Angle Camera (NAC)

  • Native resolution: 0.5–2 m/px

  • Orthorectified onto the local DEM

  • Cropped to exactly match the DEM spatial extent

  • Stored as a ground-truth appearance image, precisely aligned with the terrain geometry

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

Additional Modalities and Renderings

For each DEM–LRO image pair, we extract a BRDF map using the method described in
https://arxiv.org/html/2601.10449v1, enabling physically based multi-lighting rendering.

We provide also :

• SVBRDF map

• Normal map

• Depth map

• Line-of-sight (LOS) map

• Multi-lighting rendered images

  • 9 solar illumination conditions sampled using SPICE

  • Rendered with both Hapke BRDF and SVBRDF

  • Total: 18 rendered images per sample

• Photometric Perception Tasks. The geometry vs. appearance

Supported Tasks

LunarPhoto supports a wide range of photometric perception and geometry–appearance learning tasks, including:

• Monocular depth estimation

• Crater detection

• Semantic segmentation

• Normal estimation via photometric stereo

• Illumination-robust learning through principled data augmentation

• Reflectance and BRDF estimation, enabled by pixel-wise correspondence between real LRO observations and local DEM geometry