Conference Agenda

We present CARS, a multiview stereo pipeline developed by CNES. This pipeline will be integrated into the CO3D mission processing chain, a mission whose goal is to generate a 3D model of the Earth in less than four years. Because this is an operational mission involving massive production, particular attention has been paid to ensuring that the software is robust, efficient and includes a set of advanced automatic processing features. The paper will provide a comprehensive overview of all the features developed since its creation to achieve this goal.

We present SatGeo-NeRF, a geometrically regularized NeRF for satellite imagery that mitigates overfitting-induced geometric artifacts observed in current state-of-the-art models using three model-agnostic regularizers. Gravity-Aligned Planarity Regularization aligns depth-inferred, approximated surface normals with the gravity axis to promote local planarity, coupling adjacent rays via a corresponding surface approximation to facilitate cross-ray gradient flow. Granularity Regularization enforces a coarse-to-fine geometry-learning scheme, and Depth-Supervised Regularization stabilizes early training for improved geometric accuracy. On the DFC2019 satellite reconstruction benchmark, SatGeo-NeRF improves the Mean Altitude Error by 13.9% and 11.7% relative to state-of-the-art baselines such as EO-NeRF and EO-GS.

High dynamic range (HDR) variations in satellite optical imagery arise from extreme differences in surface reflectance and illumination conditions. Conventional satellite NeRF frameworks are typically trained on tone-mapped or radiometrically enhanced images, where nonlinear preprocessing alters the physical relationship between measured pixel values and true scene radiance. This leads to biased photometric optimization and loss of geometric fidelity, especially under strong illumination contrasts. To address these limitations, we propose an HDR-consistent learning framework that integrates RawNeRF-style radiance supervision with shadow regularization. The method trains directly on raw satellite imagery using a logarithmic, tone mapping–aware loss that preserves linear radiance and stabilizes optimization under high dynamic range conditions. In parallel, a soft shadow regularization constrains network-predicted shadows using geometric cues derived from solar ray casting, promoting physically consistent irradiance decomposition. Experiments on four AOIs from the DFC2019 dataset demonstrate that HDR-aware radiance learning significantly improves DSM accuracy by maintaining linear radiometric consistency. The proposed shadow regularization also improves geometric consistency in structure-dominated urban scenes, although its effect is limited in vegetation-dominant areas where shadow cues are less informative. Although performance gains are smaller in vegetation-dominant areas, the results confirm that combining HDR radiance learning with geometric shadow regularization yields more radiometrically consistent and geometrically accurate 3D reconstruction from satellite imagery.

Recently, 3D Gaussian Splatting has been introduced as a compelling alternative to NeRF for Earth observation, offering competitive reconstruction quality with significantly reduced training times.

In this work, we extend the EOGS framework to propose \namemodel, a novel method tailored for satellite imagery that directly operates on raw high-resolution panchromatic data %and multispectral data

without requiring external preprocessing.

Furthermore, we embed bundle adjustment directly within the training process with optical flow techniques, avoiding reliance on external optimization tools while improving camera pose estimation.

We also introduce several improvements to the original implementation, including early stopping and TSDF post-processing, all contributing to sharper reconstructions and better geometric accuracy.

Experiments on the IARPA 2016 and DFC2019 datasets demonstrate that EOGS++ achieves state-of-the-art performance in terms of reconstruction quality and efficiency, outperforming the original EOGS method and other NeRF-based methods while maintaining the computational advantages of Gaussian Splatting. Our model demonstrates an improvement from 1.33 to 1.19 mean MAE errors on buildings compared to the original EOGS models.

The emergence of satellite constellations enables near-synchronous multi-view optical imaging, offering new opportunities for large-scale 3D city modeling. Yet a practically promising configuration, in which a primary near-nadir view is complemented by multiple oblique side-looking viewpoints, remains under-examined. This study develops a controlled semi-simulation framework to analyze how multi-view imaging geometry affects the recoverability of urban 3D structures. Under idealized conditions with imaging perturbations removed, e.g., radiometric, illumination, and sensor model errors, the experiments focus on three practical factors: the number of side-looking views, view obliqueness, and the constellation’s azimuthal orientation relative to the scene. With parameter sweep analysis, it reveals an asymmetric U-shaped trend between reconstruction performance and both the view count and the obliqueness: moderate angular diversity markedly strengthens urban scene recoverability. In contrast, large obliqueness reduces inter-view overlap and destabilizes matching, while excessive redundancy introduces consistency issues that ultimately degrade reconstruction performance. Furthermore, the results shows that geometric accuracy, completeness, and texture appearance each peak at different parameter combinations, revealing intrinsic trade-offs in multi-view urban reconstruction, as different evaluation criteria favor distinct optimal configurations. The study provides practical guidance for the geometric design and mission planning of multi-satellite constellations aimed at improving satellite-based 3D modeling in urban areas.

This work presents an illumination-prior-based reconstruction model for high-resolution DEM generation from single-view lunar imagery, developed for the extreme illumination conditions and rugged terrain of the lunar south pole. The model integrates an initial DEM prior with multi-scale monocular image features and incorporates illumination priors derived from solar geometry to enhance stability in shadowed, low-texture, and terrain-transition regions. Through cross-modal feature fusion, it effectively aligns geometric structure with shading and photometric cues, enabling accurate recovery of fine-scale topography even when visual information is severely degraded. Experimental evaluations across multiple south-polar regions show that the proposed reconstruction model outperforms existing deep learning approaches and the classical Shape-from-Shading method in elevation, slope, and aspect accuracy, with independent validation using LOLA laser altimetry points confirming its improved geometric reliability. Visual comparisons demonstrate clear advantages in reconstructing crater rims, steep slopes, and permanently shadowed areas where conventional methods often fail or produce blurred terrain structures. The model also maintains robust performance under varying solar azimuths, highlighting the effectiveness of incorporating illumination priors to improve generalization in challenging environments. Overall, the proposed reconstruction model provides a reliable and effective solution for detailed lunar terrain recovery from monocular images and offers valuable support for scientific investigation, resource assessment, landing-site evaluation, and mission planning in the lunar south polar region.

To address the critical demand for high-precision mapping of the Lunar South Pole (LSP)—a region pivotal for deep space resource utilization yet plagued by extreme illumination variations, extensive permanent shadow regions (PSRs), and weak texture—this study proposes a control network construction method for multi-temporal Lunar Reconnaissance Orbiter (LRO) Narrow Angle Camera (NAC) images, anchored in lunar impact crater matching. Leveraging the morphological stability and spatial consistency of impact craters, we first created a dedicated dataset: 94 multi-temporal LSP orthophotos (1 meter/pixel resolution) with manual annotations, allocating 70% for YOLOv8 model training and 30% for validation to ensure accurate crater detection (extracting center coordinates and semi-major/semi-minor axes). For virtual feature point matching, we integrated crater geometric attributes (coordinates, aspect ratio) and inter-crater topological relationships (distance, azimuth angle) to build local descriptors, enhanced by KD-tree indexing for efficient neighborhood queries, multi-attribute similarity measurement, and bidirectional voting to eliminate mismatches. For large craters, normalized cross-correlation (NCC) was used for secondary matching to refine accuracy. Post-matching, tie points were back-projected from orthophoto to original image space via ground coordinates. Experiments on 1,208 LRO NAC images showed the method outperforms SIFT and SuperPoint: it generated 938,029 tie points (even in dark shadows) with 2,347,629 measurements, and bundle adjustment achieved a sigma naught of 0.68. This work enables automatic high-quality control network construction, supporting reliable LSP topographic mapping for deep space exploration.