Research map / Label-efficient and robust learning / Recognition and long-tail
Scaling Image Geo-Localization to Continent Level
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 a hybrid geo-localization method that fuses learned ground-view cell prototypes (from a classification-style proxy task) with aerial tile embeddings, enabling fine-grained (~200 m) retrieval over continent-sized regions.
Problem
Fine-grained (meter-to-100m) image geo-localization at continent scale, where global classification is coarse (10+ km), ground-to-ground retrieval has prohibitive database sizes, and cross-view retrieval suffers domain gap and has been studied only at small scales.
Approach
Learn l2-normalized ground-view cell prototypes via a contrastive/classification proxy over S2 cells, then combine them additively (with a calibration factor) with aerial tile embeddings into a single cell-code database, so retrieval is fine-grained yet database size scales with cells rather than images.
Key results
- Hybrid model (DINOv3-L) reaches 68.7% top-1 recall within 200m on EuropeWest, versus 57.5% for the ViT-B hybrid and 32.6% for the Fervers et al.-style baseline.
- On BEDENL, the full hybrid model (iBOT-B16 at 224px) achieves 60.3% top-1 recall at 200m, improving over SALAD-Aerial cross-view retrieval (39.2%) and classification baselines (CosFace 7.4%).
- Ours with DINOv3-L and 448px inputs reaches 76.1% top-1 recall at 200m on BEDENL, and prototypes-only trained on BEDENL get 47.6% on Portugal vs 27.3% for official SALAD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine cell prototypes with aerial embeddings | Ours (full) · BEDENL | Recall@1@200m: 56.3 -> 64.0 | Helped |
| add ground-prototype supervision (ablation) | Hybrid model · BEDENL | Recall@1@200m: 47.2 -> 60.3 | Helped |
| add ground-aerial supervision (ablation) | Hybrid model · BEDENL | Recall@1@200m: 47.2 -> 58.3 | Helped |
| add aerial-prototype supervision (ablation) | Hybrid model · BEDENL | Recall@1@200m: 58.3 -> 60.3 | Helped |
| use cell level L=16 rather than L=15 (ablation) | Hybrid model · BEDENL | Recall@1@200m: 60.3 -> 63.3 | Helped |
| use coarser cell level L=12 (ablation) | Hybrid model · BEDENL | Recall@1@200m: 60.3 -> 37.8 | Hurt |
| use aerial interpolation instead of nearest-neighbor sampling (ablation) | Hybrid model with ground nearest-neighbor sampling · BEDENL | Recall@1@200m: 57.8 -> 58.5 | Helped |
| use ground interpolation instead of nearest-neighbor sampling (ablation) | Hybrid model with aerial interpolation · BEDENL | Recall@1@200m: 58.5 -> 57.7 | Hurt |
| select all cells covered by the ground camera frustum (ablation) | Hybrid model with aerial interpolation · BEDENL | Recall@1@200m: 57.7 -> 60.3 | Helped |
| increase input image size to 448 px (ablation) | DINOv3-L16 model · BEDENL | Recall@1@200m: 70.2 -> 76.1 | Helped |
| use a larger DINOv3-L backbone (ablation) | Ours (full) · EuropeWest | Recall@1@200m: 57.5 -> 68.7 | Helped |
| sample training images per cell proportional to sequence counts (ablation) | Ours (full), trained on BEDENL+ · Trekker | Recall@1@200m: 18.7 -> 30.3 | Helped |
Surprising findings
- Selecting all overlapping cells for query frustum works best, while bilinear interpolation of query prototypes hurts recall.
- Multi-similarity loss outperforms cross-entropy losses such as InfoNCE and DCL for this setting.
Limitations
- Training and evaluation rely on proprietary Google StreetView data, which limits reproducibility.
- Prototype-based retrieval degrades in rural areas with sparse ground data and on pedestrian viewpoints (Trekker), requiring retraining for new areas.
- Model weights are withheld for misuse reasons, and training requires 128 TPUs.
Experimental setup
Datasets: Google StreetView (BEDENL, EuropeWest) aerial and satellite overhead tiles UK+IE Trekker GoogleUrban
Benchmarks: BEDENL EuropeWest UK+IE Trekker GoogleUrban Portugal cross-area VPR
Metrics: Recall@K@200m Recall@K@1km Recall@K@100m
Baselines: SALAD-Aerial Fervers et al. CosFace loss classification Hierarchical loss (OSV-5M) Haversine loss (PIGEON) SALAD ground retrieval Gemini 2.5 Pro
Builds on: SALAD Fervers et al. cross-view localization multi-similarity loss iBOT PIGEON
Closest papers by meaning
- GeoRanker: Distance-Aware Ranking for Worldwide Image Geolocalization (NeurIPS 2025 · similarity 0.88)
- PIGEON: Predicting Image Geolocations (CVPR 2024 · similarity 0.87)
- RANGE: Retrieval Augmented Neural Fields for Multi-Resolution Geo-Embeddings (CVPR 2025 · similarity 0.87)
- G3: An Effective and Adaptive Framework for Worldwide Geolocalization Using Large Multi-Modality Models (NeurIPS 2024 · similarity 0.87)
- OpenStreetView-5M: The Many Roads to Global Visual Geolocation (CVPR 2024 · similarity 0.86)
- GeoViS: Geospatially Rewarded Visual Search for Remote Sensing Visual Grounding (CVPR 2026 · similarity 0.86)
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.