Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Adapting Stereo Vision From Objects To 3D Lunar Surface Reconstruction with the StereoLunar Dataset

Clementine Grethen, Simone Gasparini, Geraldine Morin, Jeremy Lebreton, Lucas Marti, Manuel Sanchez-Gestido

TL;DR

The Moon presents a low-texture, high-contrast, and GNSS-denied setting that challenges conventional 3D reconstruction methods trained on Earth. The authors introduce LunarStereo, a physically grounded, large-scale stereo dataset rendered from high-resolution lunar DEMs with Hapke BRDFs and varied illumination, enabling dense depth maps and ground-truth camera poses for over 50k stereo pairs. They fine-tune the MASt3R model on this dataset, achieving substantial improvements in 3D geometry accuracy and pose estimation over zero-shot baselines, and demonstrate generalization to real descent imagery from Chang\'E-3. This work provides a path for robust cross-domain lunar perception, supporting safer descent, hazard avoidance, and autonomous navigation in challenging, texture-sparse environments.

Abstract

Accurate 3D reconstruction of lunar surfaces is essential for space exploration. However, existing stereo vision reconstruction methods struggle in this context due to the Moon's lack of texture, difficult lighting variations, and atypical orbital trajectories. State-of-the-art deep learning models, trained on human-scale datasets, have rarely been tested on planetary imagery and cannot be transferred directly to lunar conditions. To address this issue, we introduce LunarStereo, the first open dataset of photorealistic stereo image pairs of the Moon, simulated using ray tracing based on high-resolution topography and reflectance models. It covers diverse altitudes, lighting conditions, and viewing angles around the lunar South Pole, offering physically grounded supervision for 3D reconstruction tasks. Based on this dataset, we adapt the MASt3R model to the lunar domain through fine-tuning on LunarStereo. We validate our approach through extensive qualitative and quantitative experiments on both synthetic and real lunar data, evaluating 3D surface reconstruction and relative pose estimation. Extensive experiments on synthetic and real lunar data validate the approach, demonstrating significant improvements over zero-shot baselines and paving the way for robust cross-scale generalization in extraterrestrial environments.

Adapting Stereo Vision From Objects To 3D Lunar Surface Reconstruction with the StereoLunar Dataset

TL;DR

The Moon presents a low-texture, high-contrast, and GNSS-denied setting that challenges conventional 3D reconstruction methods trained on Earth. The authors introduce LunarStereo, a physically grounded, large-scale stereo dataset rendered from high-resolution lunar DEMs with Hapke BRDFs and varied illumination, enabling dense depth maps and ground-truth camera poses for over 50k stereo pairs. They fine-tune the MASt3R model on this dataset, achieving substantial improvements in 3D geometry accuracy and pose estimation over zero-shot baselines, and demonstrate generalization to real descent imagery from Chang\'E-3. This work provides a path for robust cross-domain lunar perception, supporting safer descent, hazard avoidance, and autonomous navigation in challenging, texture-sparse environments.

Abstract

Accurate 3D reconstruction of lunar surfaces is essential for space exploration. However, existing stereo vision reconstruction methods struggle in this context due to the Moon's lack of texture, difficult lighting variations, and atypical orbital trajectories. State-of-the-art deep learning models, trained on human-scale datasets, have rarely been tested on planetary imagery and cannot be transferred directly to lunar conditions. To address this issue, we introduce LunarStereo, the first open dataset of photorealistic stereo image pairs of the Moon, simulated using ray tracing based on high-resolution topography and reflectance models. It covers diverse altitudes, lighting conditions, and viewing angles around the lunar South Pole, offering physically grounded supervision for 3D reconstruction tasks. Based on this dataset, we adapt the MASt3R model to the lunar domain through fine-tuning on LunarStereo. We validate our approach through extensive qualitative and quantitative experiments on both synthetic and real lunar data, evaluating 3D surface reconstruction and relative pose estimation. Extensive experiments on synthetic and real lunar data validate the approach, demonstrating significant improvements over zero-shot baselines and paving the way for robust cross-scale generalization in extraterrestrial environments.
Paper Structure (10 sections, 1 equation, 4 figures, 4 tables)

This paper contains 10 sections, 1 equation, 4 figures, 4 tables.

Table of Contents

  1. Introduction & Background
  2. Related work
  3. Lunar images datasets
  4. 3D Reconstruction in Lunar Context
  5. Our Proposed Lunar Stereo Dataset
  6. Learning Moon 3D Reconstruction
  7. Experiments
  8. Pose estimation
  9. Qualitative study with real landing images
  10. Conclusion

Figures (4)

  • Figure 1: Top row: examples of lunar stereo image pairs of our dataset taken from three different trajectories, along with the corresponding ground-truth depth maps (viridis colormap) and 3D scene renderings. Middle row: for each view, it shows the ground truth for the depth map (viridis colormap) and the slope map (heat colormap), followed by the corresponding predictions of our method and the MASt3R baseline. Bottom row: the ground truth 3D scene (left) with the final 3D reconstructions for our method (center) and MASt3R (right), with View 1 in red and View 2 in green.
  • Figure 2: Overview of the dataset generation pipeline. The output is images captured from the three trajectory types, and offers ground truth pixel-level correspondences. For each pair, three variants with different illumination configurations have been generated.
  • Figure 3: Qualitative study using real descent imagery. Comparison between MASt3R (pretrained) and Ours on a Nadir baseline (left column) and an Oblique baseline (right column). Additional real-world examples are provided in the supplementary material.
  • Figure 4: 3D reconstruction comparison on lunar stereo pairs. Each row shows a stereo pair with both the reconstructed 3D point cloud and the corresponding central depth profile (ground truth in red, prediction in brown), for MASt3R (pretrained) and our fine-tuned model (Ours). Top: oblique trajectory with lateral motion. Bottom: nadir view with limited parallax—a challenging case for triangulation-based SfM methods. Ours yields denser and more accurate reconstructions, with better depth consistency, especially under oblique conditions. Additional results are provided in the supplementary material.