Sat3DGen Generates Street-Level 3D Scenes from Single Satellite Image
Researchers have introduced Sat3DGen, a novel method for generating comprehensive street-level 3D scenes from a single satellite image. The approach addresses the limitations of existing techniques, which either produce high geometric fidelity but limited semantic diversity (geometry-colorization models) or yield rich content but coarse and unstable geometry (proxy-based models). Sat3DGen employs a geometry-first methodology that integrates novel geometric constraints with a perspective-view training strategy to counter the extreme viewpoint gap and sparse supervision inherent in satellite-to-street data. The method enhances feed-forward image-to-3D frameworks by jointly learning geometry and texture while explicitly targeting geometric errors. This work is published on arXiv under identifier 2605.14984.
Key facts
- Sat3DGen generates street-level 3D scenes from a single satellite image.
- Current methods show a trade-off between geometric fidelity and semantic diversity.
- Geometry-colorization models achieve high geometric fidelity but lack semantic diversity.
- Proxy-based models produce rich content but coarse and unstable geometry.
- Sat3DGen uses a geometry-first methodology with novel geometric constraints.
- A perspective-view training strategy is employed to counter geometric errors.
- The method addresses extreme viewpoint gap and sparse supervision.
- The paper is available on arXiv with identifier 2605.14984.
Entities
Institutions
- arXiv