Research map / 3D vision / 3D generation
Satellite to GroundScape - Large-scale Consistent Ground View Generation from Satellite Views
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.
Proposes Sat2GroundScape, a fixed-LDM framework with satellite-guided and satellite-temporal denoising to generate consistent multi-view ground images from satellite views, plus a 100k+ perspective satellite-ground dataset.
Problem
Generating photorealistic and mutually consistent ground-view images (and videos) from satellite imagery, where prior single-view methods lack consistency across neighboring ground views.
Approach
Keeps a pretrained Stable Diffusion v2-1 LDM fixed, adds a ControlNet-like satellite-layout encoder for per-view denoising, and a temporal-spatial extension conditioned on an initial ground view and ResNet-encoded camera motion from satellite renderings to enforce consistency across views.
Key results
- Achieves LPIPS 0.542, FID 159.636, DreamSIM 0.531, and FVD1 16.83 on Sat2GroundScape, the best on all perceptual and temporal metrics except PSNR (15.86, second to SceneScape's 16.32).
- Ablation shows removing satellite-temporal denoising (w/o temp) raises FVD1 from 16.83 to 19.21 and removing satellite-guided denoising (w/o sat) yields 22.19.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Sat2GroundScape instead of the strongest baseline, SceneScape | Cross-view ground generation · Sat2GroundScape test | PSNR: 16.32 -> 15.86 | Hurt |
| Use Sat2GroundScape instead of the strongest baseline, Sat2Ground | Cross-view ground generation · Sat2GroundScape test | SSIM: 0.216 -> 0.231 | Helped |
| Use Sat2GroundScape instead of the strongest baseline, GVG | Cross-view ground generation · Sat2GroundScape test | LPIPS: 0.581 -> 0.542 | Helped |
| Use Sat2GroundScape instead of the strongest baseline, GVG | Cross-view ground generation · Sat2GroundScape test | FID: 175.135 -> 159.636 | Helped |
| Use Sat2GroundScape instead of the strongest baseline, GVG | Cross-view ground generation · Sat2GroundScape test | DreamSIM: 0.600 -> 0.531 | Helped |
| Use Sat2GroundScape instead of the strongest baseline, GVG | Cross-view ground generation · Sat2GroundScape test | FVD1 ×100: 19.34 -> 16.83 | Helped |
| Use Sat2GroundScape instead of the strongest baseline, GVG | Cross-view ground generation · Sat2GroundScape test | FVD2 ×100: 19.39 -> 16.88 | Helped |
| Add satellite-guided denoising and satellite-temporal denoising (ablation) | Standard LDM ground-view generation · Sat2GroundScape test | SSIM: 0.106 -> 0.231 | Helped |
| Add satellite-guided denoising (ablation) | Ground-view generation with satellite-temporal denoising · Sat2GroundScape test | SSIM: 0.159 -> 0.231 | Helped |
| Add satellite-temporal denoising (ablation) | Ground-view generation with satellite-guided denoising · Sat2GroundScape test | FVD1 ×100: 19.21 -> 16.83 | Helped |
Surprising findings
- SceneScape achieves the highest PSNR (16.32) despite poor perceptual and temporal quality due to sequential artifact accumulation.
Limitations
- Low-resolution satellite data with spatial uncertainty limits geometric accuracy and detail.
- Sequential warping-based baselines accumulate artifacts; the proposed method's scalability to very long sequences is not evaluated.
- Evaluated on a single city dataset (Jacksonville) plus a qualitative-only generalization test on HoliCity.
Experimental setup
Datasets: Sat2GroundScape HoliCity
Benchmarks: Sat2GroundScape HoliCity
Metrics: PSNR SSIM LPIPS FID DreamSIM FVD
Baselines: Sat2Ground SceneScape GVG
Builds on: Stable Diffusion ControlNet GVG Stable Video Diffusion DDIM
Closest papers by meaning
- DiffusionSat: A Generative Foundation Model for Satellite Imagery (ICLR 2024 · similarity 0.86)
- Controllable Satellite-to-Street-View Synthesis with Precise Pose Alignment and Zero-Shot Environmental Control (ICLR 2025 · similarity 0.86)
- SGFormer: Satellite-Ground Fusion for 3D Semantic Scene Completion (CVPR 2025 · similarity 0.85)
- COSMIC: Compress Satellite Image Efficiently via Diffusion Compensation (NeurIPS 2024 · similarity 0.85)
- GeoViS: Geospatially Rewarded Visual Search for Remote Sensing Visual Grounding (CVPR 2026 · similarity 0.84)
- GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation (NeurIPS 2026 · similarity 0.84)
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.