Efficacy of Monocular Depth Estimation in Lunar Environments

Michael W. Lanighan, Kenji Brameld, Seth GeearXiv 2026 · arXiv

TL;DR

This paper evaluates whether monocular depth estimation (MDE) models—trained almost exclusively on terrestrial data—can accurately reconstruct the alien, unstructured landscape of the Moon. The researchers found that Vision Transformer (ViT)-based models generalize remarkably well to lunar environments, outperforming traditional stereogrammetry by successfully navigating harsh shadows and textureless regolith while potentially cutting downlink bandwidth requirements in half.

What Problem Does This Paper Address?

Remote robotic operations in space, such as those planned for NASA’s Artemis missions, face a dual challenge: high cognitive load for operators and severe bandwidth constraints. Currently, operators must mentally fuse multiple 2D camera feeds to understand a robot's 3D surroundings. While 3D reconstruction helps, traditional stereogrammetry requires two cameras (doubling bandwidth) and often fails in the lunar environment due to extreme lighting, deep shadows, and the "featureless" nature of lunar dust (regolith).

The authors investigate if MDE—which predicts 3D structure from a single 2D image—can provide a robust, low-bandwidth alternative for situational awareness in these "out-of-distribution" (OOD) environments.

Methods

The study benchmarks several state-of-the-art MDE approaches against a traditional baseline, the NASA Ames Stereo Pipeline (ASP). The MDE models were categorized into two groups:

  1. Single-View Reconstruction: Models like Marigold (diffusion-based), Depth Anything V2, and MoGe-2 (ViT-based).
  2. Multi-View Reconstruction: Models like VGGT, MapAnything, and OmniVGGT, which can fuse information from multiple frames.

To evaluate these models fairly, the authors used the Scale-Invariant Logarithmic (SILog) error. This metric is crucial because many MDE models predict "relative" depth (the shape of the scene) rather than "absolute" metric depth (exact meters). SILog measures how well the predicted structure aligns with reality, regardless of the global scale.

Datasets

The researchers employed two distinct datasets to bridge the gap between simulation and historical reality:

  • NASA Ames POLAR Stereo Dataset: A collection of images from simulated lunar terrains in a laboratory setting, complete with ground-truth LiDAR scans. This allowed for rigorous quantitative analysis across various lighting conditions and rock distributions.
  • Apollo 17 Surface Imagery: Historical photos from the final moon landing. Since no ground-truth depth exists for these, they were used for qualitative "analog-to-field" validation to see if the models could handle real lunar artifacts like lens flares and astronaut equipment.

Key Insights and Ideas

  • ViT Generalization: Models using Vision Transformer (ViT) backbones (like MoGe-2 and VGGT) showed incredible zero-shot generalization. Despite never seeing the Moon during training, their learned geometric priors were robust enough to interpret the lunar surface.
  • Diffusion Failure: Conversely, the diffusion-based model (Marigold) performed poorly. The authors suggest this is because diffusion models rely too heavily on "generative priors"—essentially, they try to make the Moon look like the terrestrial environments they were trained on, which leads to failure when the distribution shifts so radically.
  • Robustness to Artifacts: A standout finding was that MDE models handled lens flares and deep shadows much better than traditional stereogrammetry. While the ASP baseline often produced "holes" or extreme errors in shadowed regions, MDE models consistently estimated depth across the entire image.
Fig. 1: Input Apollo 17 image (left) and the resulting depth images and point clouds (middle and right columns) generated by the specified monocular depth estimation techniques. Note the consistent reconstruction of the barren, textureless lunar landscape, even in the presence of shadows.

Key Results

  • Quantitative Top Performer: MoGe-2 achieved the best performance with a mean SILog of 2.31, corresponding to a 2σ error factor of 1.047. This means 95% of its depth predictions were within ~5% of the ground truth scale-aligned depth.
  • Stereo Baseline Struggles: The traditional ASP baseline performed the worst, with a mean SILog of 109.99. Its reliance on matching identical pixels across two images made it highly susceptible to failure in textureless areas and regions with lens flares.
  • Qualitative Success: In the Apollo 17 "fender repair" scene, MoGe-2 successfully separated the complex geometry of the Lunar Roving Vehicle (LRV) from the background, whereas other models tended to "bleed" the foreground into the background.
Fig. 8: Reconstructions of the Apollo 17 fender repair incident. MoGe-2 correctly separates the rover from the background.

Limitations and Future Work

The primary limitation identified is metric scaling. While the models are excellent at capturing the shape of the scene, they struggle to determine the exact size (e.g., is that rock 1 meter away or 1.1 meters?). The authors suggest that future systems could "anchor" these relative reconstructions using known dimensions of rover components (like a robotic arm) visible in the frame.

Future work will also look into multi-view MDE to create internally consistent, fused maps of large lunar areas, providing a richer 3D context for remote operators than single-frame snapshots can offer.

Summary updated

Abstract

Recent advances in monocular depth estimation have demonstrated highly accurate 3D reconstructions within in-distribution terrestrial domains. This study investigates the efficacy of several state-of-the-art approaches to reconstruct out-of-distribution lunar environments using the NASA Ames Polar Optical Lunar Analog Reconstruction (POLAR) stereo dataset and Apollo 17 surface imagery. By leveraging these datasets, we establish a robust analog-to-field validation pipeline that evaluates quantitative performance against ground-truth physical analogs and qualitative performance against historical field data. Our results indicate that select state-of-the-art techniques can successfully generalize to produce high-fidelity relative 3D reconstructions of the unstructured lunar surface.
BibTeX
@misc{lanighan2026efficacy,
  author = {Michael W. Lanighan and Kenji Brameld and Seth Gee},
  title = {{Efficacy of Monocular Depth Estimation in Lunar Environments}},
  year = {2026},
  eprint = {2610.38080},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.38080},
}
Figure