Abstract:
Neural Radiance Fields (NeRF) have revolutionized 3D reconstruction, offering photorealistic rendering from multi-view images. Since their introduction in 2020, NeRFs have evolved rapidly, particularly in remote sensing, where they outperform traditional photogrammetry by learning differentiable functions that map 3D coordinates to density and color. This flexibility makes NeRFs ideal for handling complex phenomena such as shadows, specularity, and seasonal variations in VHRS satellite imagery.
However, scaling NeRFs to large scenes remains a challenge due to memory constraints. In this talk, I introduce Snake-NeRF, a framework designed to overcome these limitations. By dividing scenes into non-overlapping 3D tiles and using a novel \(2\times 2\) tile progression strategy, Snake-NeRF enables efficient training on a single GPU without compromising quality. This approach ensures linear time complexity and eliminates the need to load all data simultaneously, making large-scale 3D reconstruction feasible.
We will also trace the evolution of NeRF applications in remote sensing, from early adaptations like Shadow-NeRF to recent breakthroughs. Finally, I will explore the future of 3D reconstruction in Earth observation, highlighting the potential of multi-modality to enhance generative power, computational efficiency, and reconstruction accuracy.
Suggested readings:
C. Billouard, D. Derksen, E. Sarrazin and B. Vallet, "SAT-NGP : Unleashing Neural Graphics Primitives for Fast Relightable Transient-Free 3D Reconstruction From Satellite Imagery," IGARSS 2024 - 2024 IEEE International Geoscience and Remote Sensing Symposium, Athens, Greece, 2024, pp. 8749-8753, doi: 10.1109/IGARSS53475.2024.10641775.
C. Billouard, D. Derksen, A. Constantin and B. Vallet, "Tile and Slide : A New Framework for Scaling NeRF from Local to Global 3D Earth Observation," ICCV 2025 - International Conference on Computer Vision, Workshop 3D-VAST: From street to space: 3D Vision AcrosS alTitudes, Honolulu, Hawaï, 2025.
Bio:
Camille earned his master’s degree in Artificial Intelligence from the Paul Sabatier Federal University of Toulouse in 2023. He is currently in the final year of his Ph.D. at the French Space Agency (CNES) in Toulouse, in collaboration with the LaSTIG lab at the French Mapping Institute (IGN). His research focuses on 3D computer vision applied to satellite imagery of earth observation, with a particular focus on neural rendering and the scaling of these models.
