Research map / Vision-language and multimodal / Multimodal LLMs
GeoReasoner: Geo-localization with Reasoning in Street Views using a Large Vision-Language Model
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 GeoReasoner, a Qwen-VL-based LVLM fine-tuned in two LoRA stages (reasoning from geo-game clues, then location on locatability-filtered street views) for street-view geo-localization with reasoning.
Problem
Street-view geo-localization is typically a black-box classification or retrieval task, and existing street-view data contains many low-quality images without visual clues and lacks reasoning annotations.
Approach
Introduce a "locatability" metric (segmentation-mask area weighted by CLIP/Sentence-BERT similarity to geo-clue text) to filter street views, and fine-tune an LVLM in two stages: reasoning tuning on geo-game text-image clue pairs, then location tuning on the filtered GSV images, with stacked LoRA adapters.
Key results
- GeoReasoner reaches country-level F1 0.9033 and city-level F1 0.8585, versus 0.7225 and 0.5270 for Qwen-VL-7B, per the paper (about 25% and 38% gains).
- GeoReasoner (F1 0.8585 city-level) slightly exceeds StreetCLIP (0.8543) while being trained on 70K images versus StreetCLIP's 1.1M.
- On Im2GPS3k with filtered Flickr fine-tuning, GeoReasoner with reasoning reaches 0.83 at country level (750km), matching GeoCLIP's 0.83 under the same setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase the share of high-locatability GSV images from 0% to 100% | GeoReasoner location component · GSV test set, country-level | accuracy: 0.63 to 0.72 | Helped |
| Increase the share of high-locatability GSV images from 0% to 100% | GeoReasoner location component · GSV test set, city-level | accuracy: 0.47 to 0.51 | Helped |
| Add location tuning (ablation) | GeoReasoner · Country-level geo-localization | F1: 0.8215 -> 0.9033 | Helped |
| Add location tuning (ablation) | GeoReasoner · City-level geo-localization | F1: 0.5813 -> 0.8585 | Helped |
| Add reasoning tuning (ablation) | GeoReasoner · Country-level geo-localization | F1: 0.8766 -> 0.9033 | Helped |
| Add reasoning tuning (ablation) | GeoReasoner · City-level geo-localization | F1: 0.8255 -> 0.8585 | Helped |
| Train on filtered, highly locatable Flickr images | GeoReasoner · Im2GPS, Street 1km | accuracy: 0.10 -> 0.13 | Helped |
| Train on filtered, highly locatable Flickr images | GeoReasoner · Im2GPS, City 25km | accuracy: 0.41 -> 0.44 | Helped |
| Train on filtered, highly locatable Flickr images | GeoReasoner · Im2GPS, Country 750km | accuracy: 0.82 -> 0.86 | Helped |
| Train on filtered, highly locatable Flickr images | GeoReasoner · Im2GPS3k, Street 1km | accuracy: 0.09 -> 0.10 | Helped |
| Train on filtered, highly locatable Flickr images | GeoReasoner · Im2GPS3k, City 25km | accuracy: 0.35 -> 0.38 | Helped |
| Train on filtered, highly locatable Flickr images | GeoReasoner · Im2GPS3k, Country 750km | accuracy: 0.74 -> 0.83 | Helped |
Surprising findings
- GPT-4V frequently refuses to answer geo-localization queries, yielding a country-level recall of only 0.34.
- The model maps Eiffel Tower replicas in New York and Hangzhou to Paris, relying on architectural style.
Limitations
- Reasoning data is small (about 3K pairs) and only provides country-level reasoning.
- Misled by architectural replicas (e.g., Eiffel Tower copies) and reliance on learned architectural cues.
- Comparisons with LVLMs use recall-conditioned accuracy and the sample size of the test set is small (1K).
Experimental setup
Datasets: Google Street View images (130K collected, 70K filtered) GeoGuessr and Tuxun textual clues (3K) Flickr images for Im2GPS and Im2GPS3k fine-tuning
Benchmarks: Im2GPS Im2GPS3k GSV test set (1K sampled)
Metrics: precision recall F1 accuracy at street/city/country thresholds
Baselines: StreetCLIP LLaVA Qwen-VL GPT-4V ViT PlaNet CPlaNet ISNs Translocator GeoDecoder GeoCLIP
Builds on: Qwen-VL LoRA StreetCLIP MaskFormer Sentence-BERT
Closest papers by meaning
- Recognition through Reasoning: Reinforcing Image Geo-localization with Large Vision-Language Models (NeurIPS 2025 · similarity 0.93)
- GRE Suite: Geo-localization Inference via Fine-Tuned Vision-Language Models and Enhanced Reasoning Chains (NeurIPS 2025 · similarity 0.92)
- Vision-Language Reasoning for Geolocalization: A Reinforcement Learning Approach (AAAI 2026 · similarity 0.91)
- ERGeoBench: A Comprehensive Benchmark for Embodied Reasoning and Geo-localization in Multimodal Large Language Models (ICML 2026 · similarity 0.89)
- GTR-Bench: Evaluating Geo-Temporal Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.89)
- TimeSpot: Benchmarking Geo-Temporal Understanding in Vision–Language Models in Real-World Settings (ICML 2026 · similarity 0.89)
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.