This is the authors' abstract. We don't add key points for this paper.
As high-resolution satellite remote sensing images become essential tools for understanding geospatial information, the large-scale 3D reconstruction of Earth's surface using these images has emerged as a significant research area in computer vision, photogrammetry, and remote sensing. However, satellite-based 3D reconstruction is highly sensitive to image changes arising from the multitemporal acquisition of images. These changes are primarily caused by varying shadows, reflections, and transient objects (e.g., vegetation), which complicate accurate modeling. Neural radiance fields (NeRFs), utilizing differentiable rendering to learn implicit scene representations, offer a novel approach to generating 4D products from multiview images without requiring additional data, gaining considerable attention in 3D scene reconstruction and rendering. Building on this, we propose a method for generating digital elevation models (DEMs) from satellite images, leveraging NeRF to create 3D scenes from a set of images captured at different times while addressing the challenges posed by lighting variations and transient objects. Our experiments demonstrate that our approach can generate high-quality DEMs and corresponding mesh models, outperforming both traditional and recent methods in qualitative and quantitative evaluations.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Wang et al. (2024) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: