MoonAnything: Unified Lunar Vision Benchmark

• MoonAnything is a unified lunar vision benchmark built on real lunar topography and physically-based rendering, offering dense geometric and photometric supervision.
• It comprises two sub-datasets, LunarGeo and LunarPhoto, designed to evaluate stereo 3D reconstruction, pose estimation, and BRDF estimation under varied lunar illumination.
• The benchmark targets critical lunar challenges such as low-texture terrains, extreme lighting contrasts, and the failure of terrestrial algorithms, supporting safe lunar missions.

Searching arXiv for the specified paper to ground the article in the primary source. MoonAnything is a unified lunar vision benchmark built on real lunar topography with physically-based rendering, providing comprehensive geometric and photometric supervision under diverse illumination with large scale (Grethen et al., 1 Apr 2026). It comprises two complementary sub-datasets, LunarGeo and LunarPhoto, and offers over 130K samples with dense geometric supervision, photorealistic imagery, multi-illumination renderings, camera calibration, and spatially-varying BRDF supervision. The benchmark is designed for modern lunar exploration missions in which terrain-relative navigation (TRN), hazard detection and avoidance (HDA), and 3D surface reconstruction are critical, and it is explicitly motivated by the Moon’s airless environment, pitch-black shadows, extreme contrast, and largely textureless regolith, conditions under which terrestrial models fail.

1. Benchmark motivation and problem setting

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.

MoonAnything is presented as the first large-scale, unified lunar vision benchmark to provide both dense geometric supervision, namely stereo imagery with ground-truth depth, and physically-based photometric supervision, namely multi-illumination renderings and spatially-varying BRDFs, under realistic lunar lighting. The benchmark closes a gap left by two common dataset regimes: real orbital imagery without ground truth, and small laboratory mockups under single lighting. Synthetic renderings are also described as often ignoring spatial BRDF variation and true solar geometry.

The technical relevance of this design follows directly from the stated lunar failure modes. Feature detectors collapse on smooth plains, stereo matching “flat-lines,” and reflectance variations such as opposition surge and anisotropic scattering go unmodeled. This suggests that MoonAnything is intended not only as a domain-specific dataset for lunar descent and mapping, but also as a stress test for algorithms operating in low-textured, high-contrast imaging regimes.

2. Dataset composition and source data

MoonAnything comprises two complementary sub-datasets, LunarGeo and LunarPhoto, built on high-resolution real lunar digital elevation models (DEMs) and rendered using the SurRender physically-based ray-tracer (Grethen et al., 1 Apr 2026). In both, terrain comes from NASA’s LOLA altimeter and SELENE TC (South Pole at 5 m/px) and Airbus Pixel Factory LRO stereo (Tycho crater at 1 m/px).

The reflectance model is Hapke’s analytical BRDF, with constant parameters over the South Pole and learned spatially-varying BRDF (SVBRDF) over Tycho, thus capturing opposition surge, backscattering, and local albedo variation. Solar positions are sampled via JPL’s SPICE toolkit to span real lunar days and multi-angle illumination.

The benchmark is therefore organized around a clear division of supervision. LunarGeo targets stereo and multi-view geometry under realistic descent conditions. LunarPhoto targets photometric modeling, reflectance estimation, and illumination-aware learning. Taken together, these two components combine real lunar topography, physically-based rendering, and supervision modalities that are usually separated across different datasets.

3. LunarGeo

LunarGeo is the geometry benchmark. Its stated goal is to train and evaluate stereo/multi-view 3D reconstruction and pose estimation under realistic descent. The imagery consists of 512 × 512 px stereo pairs with Gaussian PSF blur and field of view values of 45° for the South Pole or 30° for Tycho (Grethen et al., 1 Apr 2026).

The benchmark defines three trajectory families:

• Nadir: cameras looking straight down; baselines 2–10 % of altitude.
• Oblique: tilt 20°–35°, simulating lateral scans.
• Dynamic: roll ±10°, altitude ±15 %, baselines up to 22 % of altitude.

Altitude bands are region-specific. For the South Pole they are 3.5–30.5 km, corresponding to ground sampling distance (GSD) 5.7–49.3 m/px. For Tycho they are 3.5–9.5 km, corresponding to GSD 3.7–9.3 m/px.

Ground truth is provided per pair. The benchmark includes intrinsics

$$K = \begin{bmatrix} f & 0 & c_x\ 0 & f & c_y\ 0 & 0 & 1 \end{bmatrix},$$

the baseline vector $b$, PSF parameters, 4×4 Moon-fixed camera poses, dense depth maps (EXR) computed by ray-tracing the DEM along each camera ray, stereo geometry in the form of inter-camera rotation $R$ and translation $t$, and metadata comprising latitude/longitude, altitude, GSD, and sun azimuth/elevation for three lighting configurations per sample.

Scale and split design are explicit: approximately 38 000 pairs for the South Pole and 20 000 pairs for Tycho, with an 80/10/10 train/val/test partition. In practical terms, LunarGeo is structured to support both region-specific evaluation and cross-region generalization, since the benchmark later reports seen South Pole results and unseen Tycho results.

4. LunarPhoto

LunarPhoto is the photometric benchmark. Its stated goal is to learn and evaluate spatially-varying BRDF estimation, photometric stereo, appearance-robust perception, and illumination-aware augmentation (Grethen et al., 1 Apr 2026). The unit sample is a DEM crop of 128 × 128 px at 5 m/px (0.4 km²) in Tycho, aligned with orthorectified LRO NAC images (0.5–2 m/px).

For each patch, the benchmark includes one true ortho-image with camera pose, sun direction, and footprint. DEM normalization is defined as subtracting patch mean elevation $\mu$ and dividing by global DEM standard deviation $\sigma$, so

$z' = (z - \mu)/\sigma.$

The learned SVBRDF is represented as per-pixel Hapke parameters $\{w, b, c, h, B0,\ldots\}$ estimated via the Lunar-G2R pipeline and stored as 128 × 128 × $N$ maps. Multi-illumination renderings are generated from nine sun positions sampled uniformly through local lunar daytime, and reflectance images are rendered under both constant Hapke and learned SVBRDF.

Ground truth per sample includes dem.tif for normalized elevation, depth.tif for per-pixel ray depth, normal.tif for surface normals, real_image.tif, brdf_map.tif, 9× multi-light TIFFs per reflectance model, and metadata.json containing camera, sun, selenographic coordinates, and timestamps. The scale and split are 84 000 samples partitioned into 67 000 train, 8 500 validation, and 8 500 test, with total rendered images of approximately 750 000 per BRDF model.

LunarPhoto is therefore not limited to image synthesis. It couples real data alignment, normalized geometry, explicit BRDF parameter maps, and multi-light renderings in a format suited to intrinsic decomposition, reflectance estimation, and illumination-conditioned training.

5. Physically-based rendering model

MoonAnything’s rendering core uses SurRender coupled with the Hapke BRDF. For each surface element, outgoing radiance $L$ at view direction $b$0 and illumination direction $b$1 is defined as

$b$2

where the Hapke reflectance function $b$3 is

$b$4

with $b$5, $b$6, $b$7 the phase angle, $b$8 the single-scattering albedo, $b$9 the opposition effect, $R$0 the particle phase function, $R$1 the multiple scattering function, and $R$2 macroscopic roughness correction (Grethen et al., 1 Apr 2026).

In LunarPhoto, constant parameters such as $R$3, $R$4, $R$5, and $R$6 are replaced with learned spatial maps, enabling SVBRDF. Realistic shadows and penumbrae arise naturally from high-resolution DEM ray-tracing.

This rendering formulation is central to the benchmark’s scope. It binds geometric truth to a lunar-specific photometric model rather than a generic Lambertian approximation. A plausible implication is that algorithms evaluated on MoonAnything are exposed to both geometric ambiguity and reflectance-driven appearance variability, including opposition surge, backscattering, and local albedo variation.

6. Experimental protocols, metrics, and baseline behavior

MoonAnything demonstrates benchmark value via stereo 3D reconstruction with two state-of-the-art networks, MASt3R and VGGT. Both are pretrained on terrestrial datasets and then fine-tuned on the South Pole split of LunarGeo (Grethen et al., 1 Apr 2026).

Evaluation is performed with three geometric metrics. Geometry accuracy (ACC) is the mean distance from reconstructed points to ground truth DEM, with lower being better. Completeness (Compl) is the mean distance from ground truth points to nearest reconstructed point. Chamfer distance is defined as ACC+Compl. Metrics are reported separately for nadir, oblique, and dynamic trajectories, and on seen (South Pole) versus unseen (Tycho) regions to assess cross-region generalization.

The reported baseline behavior is explicit. Pretrained terrestrial models produce flat or highly erroneous reconstructions, for example ACC ≈236 m under nadir in South Pole. Fine-tuning on just 30 000 LunarGeo pairs drops ACC to ≈43 m in nadir and yields consistent gains across all trajectories. On the unseen Tycho split, MASt3R FT surpasses VGGT FT under oblique/dynamic (ACC ≈72 m vs 76 m) and dynamic (34 m vs 47 m), which the source describes as demonstrating transferable geometry priors.

The benchmark also identifies key failure modes. Extreme shadowing under low-Sun angles leads to large dark regions, stereo matching holes, and incomplete reconstructions. Texture-less plains introduce minimal photometric variation and confound feature matching. These observations are consistent with the original problem statement: the most difficult lunar scenes are precisely those in which conventional terrestrial priors are least reliable.

7. Applications, retargeting, and tooling

Beyond stereo reconstruction, MoonAnything supports terrain-relative navigation, hazard detection and avoidance, photometric stereo and intrinsic decomposition, and relighting and augmentation (Grethen et al., 1 Apr 2026). The benchmark specifically describes the following uses:

• Terrain-relative navigation: train networks to regress lander pose from descent imagery.
• Hazard detection and avoidance: use depth+reflectance to segment unsafe slopes or rocks.
• Photometric stereo & intrinsic decomposition: recover normals and albedo from multi-lighting lunar patches.
• Relighting & augmentation: synthetically generate low/high Sun-angle views for robust vision training.

Because its core relies on airless-body physics, specifically Hapke BRDF, real-geometry DEMs, and solar ephemerides, MoonAnything can be retargeted to Mercury, Phobos, Deimos, or asteroids by replacing the DEM and re-estimating BRDF parameters. This suggests that the benchmark is positioned as a general methodology for airless celestial bodies rather than only a fixed lunar dataset.

All data, metadata, rendering scripts, and trained SVBRDF estimators are open-source at the project repository. The distribution includes generation commands for LunarGeo and LunarPhoto, supports custom DEMs and BRDF models through editable configuration files in configs/, and provides data loaders and evaluation scripts in examples/, including PyTorch dataloaders for both sub-datasets and built-in metric computation for ACC, completeness, Chamfer, normal error, and albedo L1. In that form, MoonAnything combines benchmark definition, reproducible rendering, and extension infrastructure within a single research artifact.