Reproduce “Gaussian Splatting for Efficient Satellite Image Photogrammetry” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Luca Savant Aira, Gabriele Facciolo, Thibaud Ehret
EOGS (Earth-observation Gaussian Splatting) adapts 3D Gaussian Splatting for satellite image photogrammetry, generating 3D models and Digital Surface Models (DSMs) from multi-date satellite imagery. The framework achieves a 300x speedup in training time, completing scene reconstruction in 3 minutes compared to 15 hours for state-of-the-art NeRF-based methods like EO-NeRF, while maintaining competitive elevation accuracy (1.46m MAE).
Key insights
- 3D Gaussian Splatting can be effectively adapted to satellite photogrammetry, delivering a significant speedup (300x faster training) compared to NeRF-based methods while maintaining high accuracy.
- Explicitly modeling satellite-specific factors like solar illumination via shadow mapping and introducing robust regularization terms (sparsity, view consistency, opaqueness) are crucial for accurate and physically consistent 3D surface reconstruction from sparse satellite views.
- The combination of a simplified affine camera model and tailored regularizations successfully bridges the gap between the efficiency of 3DGS and the demands of Earth observation data, making rapid 3D mapping feasible.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.