Research map / 3D vision / Gaussian splatting
Gaussian Splatting for Efficient Satellite Image Photogrammetry
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Adapts 3D Gaussian Splatting to multi-date satellite photogrammetry (EOGS), adding affine RPC approximation, shadow mapping, and sparsity, view-consistency and opacity regularizers to reach EO-NeRF-level DSM accuracy in minutes.
Problem
Digital surface modeling from multi-date, non-orthorectified satellite images is slow with NeRF-based Earth observation methods, which take many hours to train.
Approach
Adapt standard 3DGS to satellite imagery by approximating RPC camera models locally as affine cameras, rendering cast shadows via a shadow-mapping pipeline using sun cameras, and adding L1 opacity sparsity, local view-consistency and entropy-based opaqueness losses.
Key results
- EOGS reaches grand mean elevation MAE of 1.46 m on JAX/IARPA without mask in 3 minutes, versus 1.43 m-range EO-NeRF results requiring about 15 hours.
- With foliage mask, EOGS achieves mean MAE 1.37 m, equivalent to EO-NeRF's 1.38 m, in 3 minutes versus 15 hours.
- Shadow mapping is the most impactful component, improving MAE by about 3.16 m in the linear regression ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add shadow mapping (ablation) | EOGS · JAX and IARPA scenes | elevation MAE: 3.16 meters | Helped |
| add view-consistency regularization (ablation) | EOGS · JAX and IARPA scenes | elevation MAE: 0.20 meters | Helped |
| add opaqueness regularization (ablation) | EOGS · JAX and IARPA scenes | elevation MAE: 0.09 meters | Helped |
| add sparsity regularization (ablation) | EOGS · JAX and IARPA scenes | elevation MAE: 0.04 meters | Helped |
| add sparsity regularization (ablation) | EOGS · JAX and IARPA scenes | training time | Helped |
| use both altitude and color consistency regularization (ablation) | EOGS · JAX 260 | elevation MAE | Helped |
| use the full EOGS method instead of EO-NeRF | satellite photogrammetry datasets · JAX and IARPA, no mask | elevation MAE | Mixed |
| use EOGS instead of EO-NeRF | satellite photogrammetry datasets · JAX and IARPA, foliage mask | elevation MAE | No clear effect |
| use EOGS instead of EO-NeRF in regions with high image coverage | satellite photogrammetry datasets · regions visible in most images | elevation MAE | Helped |
| use EOGS instead of EO-NeRF in regions with low image coverage | satellite photogrammetry datasets · regions observed in only a few images | elevation MAE | Hurt |
Surprising findings
- The authors note EOGS struggles in regions observed by few images, even though it performs better in high-coverage regions.
Limitations
- Performs worse than EO-NeRF on the whole AOI without foliage mask (1.62 vs 1.51 m mean MAE).
- Struggles in regions visible in only a few images.
- Requires ground-truth semantic maps to match EO-NeRF in foliage-masked evaluation.
Experimental setup
Datasets: DFC2019 IARPA 2016 Multi-View Stereo 3D Mapping Challenge
Benchmarks: DFC2019 IARPA2016
Metrics: mean absolute error on elevation (m) training time (minutes)
Baselines: EO-NeRF SAT-NGP Sat-Mesh S2P
Builds on: 3D Gaussian Splatting EO-NeRF Sat-NeRF 3DGS-MCMC
Closest papers by meaning
- RPC-GS: Gaussian Splatting with native RPC Rendering for Satellite Imagery (NeurIPS 2026 · similarity 0.91)
- Superpoint Gaussian Splatting for Real-Time High-Fidelity Dynamic Scene Reconstruction (ICML 2024 · similarity 0.87)
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.87)
- GES : Generalized Exponential Splatting for Efficient Radiance Field Rendering (CVPR 2024 · similarity 0.86)
- DoF-Gaussian: Controllable Depth-of-Field for 3D Gaussian Splatting (CVPR 2025 · similarity 0.86)
- Fully Explicit Dynamic Gaussian Splatting (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.