A curated list of papers, datasets, and benchmarks for worldwide image geolocalization: predicting the geographic location of an image taken anywhere on Earth, expressed as GPS coordinates, a geographic cell, or an administrative region. The task is open-world by nature: no reference collection provides complete imagery coverage of the planet, so a model must generalize to locations it has never observed.
This repository accompanies our survey "Worldwide Image Geolocalization in the Foundation Model Era: A Survey" and follows its taxonomy.
- Taxonomy
- Classical Methods
- Foundation-Model Methods
- Datasets
- Benchmarks
- Performance Summary
- Related Surveys
- Contributing
We organize existing studies into two broad groups. Classical methods mainly differ in how they represent and predict geographic locations: classification-based methods divide the Earth into discrete geographic classes, retrieval-based methods infer locations by searching over geo-tagged references or location representations, and probabilistic methods estimate a distribution over possible locations. Foundation-model methods are organized by how foundation models are used in the geolocalization process: retrieval-augmented generation (RAG), reasoning, agentic, and hybrid designs.
flowchart LR
W[Worldwide Image Geolocalization] --> C[Classical Methods]
W --> F[Foundation-Model Methods]
C --> C1[Classification]
C --> C2[Retrieval]
C --> C3[Probabilistic]
C1 --> C11[Geo-cell classification]
C1 --> C12[Zero-shot classification]
C11 --> C111[Partitioning: Flat / Hierarchical]
C11 --> C112[Auxiliary signals]
C11 --> C113[Prediction reliability]
C2 --> C21[Image-gallery]
C2 --> C22[Location-gallery]
F --> F1[RAG]
F --> F2[Reasoning]
F --> F3[Agentic]
F --> F4[Hybrid]
F2 --> F21[Reasoning SFT]
F2 --> F22[Reinforcement learning]
F2 --> F23[Training-free structured reasoning]
| Method | Paper | Venue | Year |
|---|---|---|---|
| MvMF | Exploiting the earth’s spherical geometry to geolocate images | ECML-PKDD | 2019 |
| SwC | Leveraging efficientnet and contrastive learning for accurate global-scale location estimation | ICMR | 2021 |
| PLONK | Around the world in 80 timesteps: A generative approach to global visual geolocation | CVPR | 2025 |
| LocDiff | LocDiff: Identifying Locations on Earth by Diffusing in the Hilbert Space | NeurIPS | 2026 |
| Method | Paper | Venue | Year |
|---|---|---|---|
| Thinking-with-Map | Thinking with Map: Reinforced Parallel Map-Augmented Agent for Geolocalization | arXiv | 2026 |
| SpotAgent | SpotAgent: Grounding visual geo-localization in large vision-language models through agentic reasoning | arXiv | 2026 |
| LocationAgent | LocationAgent: A Hierarchical Agent for Image Geolocation via Decoupling Strategy and Evidence from Parametric Knowledge | arXiv | 2026 |
| GeoVista | GeoVista: Web-Augmented Agentic Visual Reasoning for Geolocalization | arXiv | 2025 |
| smileGeo | Swarm intelligence in geo-localization: A multi-agent large vision-language model collaborative framework | KDD | 2025 |
| Method | Paper | Venue | Year |
|---|---|---|---|
| NAVIG | NAVIG: Natural language-guided analysis with vision language models for image geo-localization | arXiv | 2025 |
| GeoMR | GeoMR: Integrating Image Geographic Features and Human Reasoning Knowledge for Image Geolocalization | Knowledge-Based Systems | 2026 |
| GeoRanker | GeoRanker: Distance-aware ranking for worldwide image geolocalization | NeurIPS | 2026 |
| GeoRouter | GeoRouter: Dynamic Paradigm Routing for Worldwide Image Geolocalization | arXiv | 2026 |
| VLM-VPR | VLM-Guided Visual Place Recognition for Planet-Scale Geo-Localization | arXiv | 2025 |
Datasets commonly used for training and evaluation in worldwide image geolocalization, ordered by the year they were introduced. Usage indicates whether a dataset is used for training, testing, or both.
| Dataset | Year | Images | Source | Annotation | Usage | Availability |
|---|---|---|---|---|---|---|
| IM2GPS | 2008 | 237 | Flickr | GPS | Test | Link |
| MP-16 | 2016 | 4.72M | Flickr (YFCC100M) | GPS | Train | Link |
| IM2GPS3k | 2017 | 3K | Flickr | GPS | Test | Metadata, Image |
| YFCC4k | 2017 | 4K | Flickr (YFCC100M) | GPS | Test | Metadata, Image |
| YFCC26k | 2018 | 25,600 | Flickr (YFCC100M) | GPS | Test | Link |
| GWS15k | 2023 | 14,955 | Google Street View | GPS | Test | Not released |
| OSV-5M | 2024 | 5.1M | Mapillary street view | GPS + admin | Both | Link |
| MP16-Pro | 2024 | 4.72M | Flickr (MP-16) + text | GPS + text | Train | Link |
| GeoBench | 2025 | 1,142 | Photos, panoramas, satellite | GPS + admin | Test | Link |
| MP-Caption | 2026 | 4.12M | Flickr (MP16-Pro) | GPS + text | Train | On request |
Benchmarks used to evaluate worldwide image geolocalization methods, ordered by the year they were introduced. Evaluation Focus is summarized with keywords.
| Benchmark | Year | Source | Evaluation Focus |
|---|---|---|---|
| Deep VG Benchmark | 2022 | Existing datasets | Accuracy, Efficiency |
| CityBench | 2024 | Simulator | Urban world model |
| LLMGeo | 2024 | Street View | Accuracy |
| Charting New Territories | 2024 | Web images, Maps | Geospatial capabilities |
| WikiTiLo | 2024 | Wikimedia | Reasoning |
| Search-based Geo Study | 2024 | Web images | Human search behavior |
| Geo-Risk Assessment | 2025 | Existing datasets | Accuracy, Privacy |
| FairLocator | 2025 | Street View | Bias, Privacy |
| GSV Global Benchmark | 2025 | Street View | Accuracy, Privacy, Tool use |
| IMAGEO-Bench | 2025 | Street View, Web images | Accuracy, Reasoning, Bias |
| EarthWhere | 2025 | Street View, Web images | Accuracy, Reasoning, Tool use |
| KoreaGEO Bench | 2025 | Street View | Bias, Privacy |
| GeoChain | 2025 | Street View | Reasoning |
| MLLM Geo Evaluation | 2025 | Street View | Accuracy, Privacy |
| GeoBrowse | 2026 | Web videos, Wikipedia | Reasoning, Tool use |
| GeoArena | 2026 | Web images | Reasoning, Human preference |
| GeoRC | 2026 | GeoGuessr | Reasoning |
| ERGeoBench | 2026 | Street View | Accuracy, Reasoning, Embodied |
The tables below collect the threshold accuracy (%) reported in existing papers on four representative benchmarks. For each method, we report the paradigm assigned under our taxonomy, the publication year, and accuracy at five standard distance thresholds (1 km, 25 km, 200 km, 750 km, and 2500 km). Bold marks the best result in each column, italic marks the second best, and "-" indicates that the value is not reported. Because these numbers are copied from the papers rather than reproduced under one controlled protocol, differences in training data, backbone models, and evaluation setups also contribute to the gaps between methods.
Note that GeoSearch uses OSV-5M as its retrieval source, and the results of PlaNet are reproduced values reported in the CPlaNet paper.
| Method | Paradigm | Year | @1km | @25km | @200km | @750km | @2500km |
|---|---|---|---|---|---|---|---|
| PlaNet | Classification | 2016 | 8.5 | 24.8 | 34.3 | 48.4 | 64.6 |
| CPlaNet | Classification | 2018 | 10.2 | 26.5 | 34.6 | 48.6 | 64.6 |
| ISNs | Classification | 2018 | 10.5 | 28.0 | 36.6 | 49.7 | 66.0 |
| SwC | Probabilistic | 2021 | 15.0 | 30.0 | 38.0 | 52.3 | 67.6 |
| StreetCLIP | Classification | 2023 | - | 22.4 | 37.4 | 61.3 | 80.4 |
| GeoDecoder | Classification | 2023 | 12.8 | 33.5 | 45.9 | 61.0 | 76.1 |
| TransLocator | Classification | 2022 | 11.8 | 31.1 | 46.7 | 58.9 | 80.1 |
| GeoLocSFT | Reasoning | 2025 | 8.8 | 32.7 | 47.2 | 65.8 | 82.3 |
| GaGA | Reasoning | 2024 | 11.7 | 33.0 | 48.0 | 67.1 | 82.1 |
| NAVIG | Hybrid | 2025 | 5.5 | 28.9 | 49.1 | 68.3 | 84.0 |
| Concept-Aware | Retrieval | 2025 | 13.2 | 34.0 | 49.8 | 68.2 | 83.5 |
| GeoCLIP | Retrieval | 2023 | 14.1 | 34.5 | 50.7 | 69.7 | 83.8 |
| GRE Suite | Reasoning | 2025 | 11.3 | 35.3 | 51.7 | 69.3 | 85.7 |
| LocDiff | Probabilistic | 2025 | 10.9 | 34.0 | 53.3 | 72.5 | 85.2 |
| GeoBayes | Reasoning | 2026 | 6.3 | 34.7 | 53.6 | 73.7 | 85.9 |
| PIGEOTTO | Classification | 2024 | 11.3 | 36.7 | 53.8 | 72.4 | 85.3 |
| GeoToken (MLLM-free) | RAG | 2025 | 16.8 | 39.6 | 53.8 | 70.8 | 85.0 |
| GLOBE | Reasoning | 2025 | 9.8 | 40.2 | 56.2 | 71.5 | 82.4 |
| G3 | RAG | 2024 | 16.7 | 40.9 | 55.6 | 71.2 | 84.7 |
| GAEA | Reasoning | 2025 | - | 36.9 | 56.0 | 73.2 | 86.7 |
| DualGeo | RAG | 2026 | 17.3 | 41.5 | 55.8 | 71.7 | 85.1 |
| LocDiff-H | Probabilistic | 2025 | 15.3 | 36.5 | 56.4 | 75.2 | 87.4 |
| TransGeoCLIP | RAG | 2026 | 17.7 | 42.2 | 56.8 | 71.7 | 86.1 |
| SpotAgent | Agentic | 2026 | 14.1 | 40.4 | 57.8 | 73.4 | 85.8 |
| GeoMR | Hybrid | 2026 | 17.0 | 41.7 | 57.5 | 73.7 | 85.7 |
| Img2Loc | RAG | 2024 | 17.1 | 45.1 | 57.9 | 72.9 | 84.7 |
| HierLoc | Retrieval | 2026 | 11.3 | 43.8 | 58.4 | 74.1 | 85.1 |
| Geo-R | Reasoning | 2026 | 18.1 | 41.5 | 58.3 | 75.3 | 86.4 |
| GeoSURGE | Retrieval | 2026 | 17.2 | 42.5 | 58.1 | 74.6 | 87.6 |
| VLM-VPR | Hybrid | 2025 | 18.9 | 46.1 | 59.8 | 73.7 | 86.0 |
| GeoAgent | Reasoning | 2026 | - | 40.8 | 58.6 | 76.2 | 89.9 |
| GeoToken (MLLM-assisted) | RAG | 2025 | 19.0 | 46.0 | 60.1 | 76.6 | 88.8 |
| GeoRanker | Hybrid | 2025 | 18.8 | 45.1 | 61.5 | 76.3 | 89.3 |
| Geo-ADAPT | Reasoning | 2026 | 17.9 | 45.3 | 62.6 | 77.9 | 89.5 |
| Pinpoint | Retrieval | 2026 | 20.5 | 47.4 | 63.5 | 79.0 | 90.2 |
| GeoRouter | Hybrid | 2026 | 20.8 | 50.5 | 65.7 | 80.4 | 90.7 |
| GeoSearch+G3 | RAG | 2026 | 23.5 | 54.9 | 66.9 | 79.6 | 89.4 |
| GeoSearch+GeoRanker | RAG | 2026 | 23.6 | 55.1 | 67.1 | 79.8 | 89.6 |
Note that GeoSearch uses OSV-5M as its retrieval source, and that the results of PlaNet, ISNs, and GeoCLIP are reproduced values reported in the CPlaNet, TransLocator, and G3 papers, respectively.
| Method | Paradigm | Year | @1km | @25km | @200km | @750km | @2500km |
|---|---|---|---|---|---|---|---|
| CPlaNet | Classification | 2018 | 7.9 | 14.8 | 21.9 | 36.4 | 55.5 |
| PlaNet | Classification | 2016 | 5.6 | 14.3 | 22.2 | 36.4 | 55.8 |
| SwC | Probabilistic | 2021 | 7.9 | 14.3 | 21.9 | 37.4 | 56.5 |
| ISNs | Classification | 2018 | 6.5 | 16.2 | 23.8 | 37.4 | 55.0 |
| TransLocator | Classification | 2022 | 8.4 | 18.6 | 27.0 | 41.1 | 60.4 |
| GeoDecoder | Classification | 2023 | 10.3 | 24.4 | 33.9 | 50.0 | 68.7 |
| GeoBayes | Reasoning | 2026 | 4.9 | 16.1 | 30.9 | 55.8 | 75.4 |
| GeoCLIP | Retrieval | 2023 | 9.6 | 19.3 | 32.6 | 55.0 | 74.7 |
| SpotAgent | Agentic | 2026 | 7.3 | 21.5 | 36.2 | 55.0 | 70.8 |
| PLONK | Probabilistic | 2024 | - | 23.7 | 36.4 | 54.5 | 73.6 |
| Geo-R | Reasoning | 2026 | 10.5 | 22.7 | 40.0 | 60.8 | 75.8 |
| PIGEOTTO | Classification | 2024 | 10.4 | 23.7 | 40.6 | 62.2 | 77.7 |
| Img2Loc | RAG | 2024 | 14.1 | 29.6 | 41.4 | 59.3 | 76.9 |
| HierLoc | Retrieval | 2026 | 8.4 | 30.2 | 43.3 | 61.7 | 75.8 |
| DualGeo | RAG | 2026 | 27.5 | 36.5 | 45.0 | 61.6 | 75.9 |
| GeoToken (MLLM-free) | RAG | 2025 | 24.3 | 35.3 | 46.6 | 64.2 | 78.6 |
| G3 | RAG | 2024 | 24.0 | 35.9 | 47.0 | 64.3 | 78.2 |
| GeoSearch+G3 | RAG | 2026 | 17.5 | 35.2 | 48.2 | 63.5 | 79.9 |
| GeoSearch+GeoRanker | RAG | 2026 | 17.5 | 35.2 | 48.2 | 63.4 | 79.3 |
| GeoSURGE | Retrieval | 2026 | 19.9 | 33.6 | 48.7 | 67.4 | 82.0 |
| GeoMR | Hybrid | 2026 | 26.3 | 37.6 | 49.3 | 66.3 | 81.4 |
| GeoToken (MLLM-assisted) | RAG | 2025 | 25.4 | 38.5 | 51.4 | 68.0 | 81.0 |
| TransGeoCLIP | RAG | 2026 | 31.2 | 41.1 | 51.8 | 67.8 | 81.0 |
| Geo-ADAPT | Reasoning | 2026 | 32.5 | 39.1 | 55.4 | 70.8 | 84.5 |
| GeoRanker | Hybrid | 2025 | 32.9 | 43.5 | 54.3 | 69.8 | 82.5 |
| Pinpoint | Retrieval | 2026 | 33.0 | 44.4 | 57.5 | 71.8 | 84.5 |
| GeoRouter | Hybrid | 2026 | 33.0 | 46.0 | 57.5 | 72.0 | 83.0 |
Note that the results of GaGA are reported on a version of GWS15k rebuilt by its authors from the OSV-5M test set, which is not identical to the original GWS15k.
| Method | Paradigm | Year | @1km | @25km | @200km | @750km | @2500km |
|---|---|---|---|---|---|---|---|
| GeoDecoder | Classification | 2023 | 0.7 | 1.5 | 8.7 | 26.9 | 50.5 |
| GAEA | Reasoning | 2025 | - | 3.7 | 16.7 | 43.3 | 73.5 |
| GeoCLIP | Retrieval | 2023 | 0.6 | 3.1 | 16.9 | 45.7 | 74.1 |
| GRE Suite | Reasoning | 2025 | 0.9 | 4.1 | 18.9 | 54.8 | 78.3 |
| GeoSURGE | Retrieval | 2026 | 1.0 | 4.6 | 21.9 | 54.7 | 80.8 |
| GaGA | Reasoning | 2024 | 0.1 | 8.5 | 33.9 | 60.6 | 82.2 |
| PIGEOTTO | Classification | 2024 | 0.7 | 9.2 | 31.2 | 65.7 | 85.1 |
| VLM-VPR | Hybrid | 2025 | 0.5 | 9.6 | 32.3 | 65.2 | 87.2 |
| LocDiff-H | Probabilistic | 2025 | 0.9 | 7.4 | 33.5 | 66.2 | 85.0 |
| LocDiff | Probabilistic | 2025 | 2.1 | 12.4 | 33.7 | 67.0 | 85.0 |
| Method | Paradigm | Year | @1km | @25km | @200km | @750km | @2500km |
|---|---|---|---|---|---|---|---|
| GeoLocSFT | Reasoning | 2025 | 2.4 | 3.5 | 15.0 | 47.2 | 69.6 |
| Chen2025-SemiVar | Retrieval | 2025 | - | 21.5 | 52.1 | 72.1 | - |
| GaGA | Reasoning | 2024 | 0.1 | 8.0 | 40.1 | 68.0 | 85.4 |
| LocDiff | Probabilistic | 2025 | - | 11.0 | 46.3 | 77.0 | 88.2 |
| Pinpoint | Retrieval | 2026 | - | 35.6 | 67.5 | 83.7 | 93.2 |
- Visual place recognition: A survey (IEEE T-RO 2015) - Visual place recognition
- Visual place recognition: A tutorial (IEEE RA-M 2023) - Visual place recognition
- Visual place recognition for aerial imagery: A survey (Robotics and Autonomous Systems 2025) - Aerial visual place recognition
- Cross-view geo-localization: a survey (IEEE Access 2024) - Cross-view geo-localization
- Image and object geo-localization (IJCV 2024) - Image and object geo-localization
- Geospatial Representation Learning: A Survey from Deep Learning to The LLM Era (arXiv 2025) - Geospatial representation learning
- Towards vision-language geo-foundation model: A survey (arXiv 2024) - Vision-language geo-foundation models
Contributions are welcome. To add a paper, please open a pull request that places the paper in the category of our taxonomy that matches its main mechanism, using the same table format (method name, linked title, venue, year). If you believe a paper is categorized incorrectly or a reported number needs correction, please open an issue.