Research map / Vision-language and multimodal / Human and embodied modeling
G3: An Effective and Adaptive Framework for Worldwide Geolocalization Using Large Multi-Modality Models
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 G3, a RAG framework for worldwide image geolocalization that aligns image, GPS and text embeddings (Geo-alignment), ensembles multiple RAG prompts with varying reference counts (Geo-diversification), and selects the final coordinate via the learned image-GPS similarity (Geo-verification), plus the MP16-Pro dataset.
Problem
Worldwide image geolocalization at coordinate level struggles with subtle location-aware visual semantics and heterogeneous geographic distribution of image data, which makes fixed-reference RAG generation and visual-similarity retrieval unreliable.
Approach
Jointly align image embeddings with GPS (Mercator-projected, random Fourier features) and textual geocoded descriptions to get location-aware retrieval; generate candidates from RAG prompts with different numbers of references (including zero-shot) and pick the final coordinate by similarity to the query image in the learned multimodal space.
Key results
- G3 reaches 16.65% at 1km and 40.94% at 25km on IM2GPS3k, above the strongest baseline Img2Loc (15.34%, 39.83%).
- On YFCC4k, G3 reaches 23.99% at 1km versus 19.78% for Img2Loc (21.3% relative improvement).
- Replacing GPT-4V with LLaVA-Next-LLaMA3-8b lowers G3 to 14.31% at 1km on IM2GPS3k but it still outperforms Img2Loc with LLaVA (10.21%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use G3 instead of the strongest baseline | IM2GPS3K · IM2GPS3K | accuracy within 1km: 8.5% | Helped |
| Use G3 instead of the strongest baseline | IM2GPS3K · IM2GPS3K | accuracy within 25km: 2.8% | Helped |
| Use G3 instead of the strongest baseline | IM2GPS3K · IM2GPS3K | accuracy within 200km: 3.3% | Helped |
| Use G3 instead of the strongest baseline | YFCC4K · YFCC4K | accuracy within 1km: 21.3% | Helped |
| Use G3 instead of the strongest baseline | YFCC4K · YFCC4K | accuracy within 25km: 16.9% | Helped |
| Use G3 instead of the strongest baseline | YFCC4K · YFCC4K | accuracy within 200km: 13.5% | Helped |
| Use G3 instead of the strongest baseline | YFCC4K · YFCC4K | accuracy within 750km: 3.3% | Helped |
| Use G3 instead of the strongest baseline | YFCC4K · YFCC4K | accuracy within 2500km: 0.6% | Helped |
| Add Geo-alignment (ablation) | G3 · IM2GPS3K and YFCC4K | accuracy within 1km: 15.71 -> 16.65; 20.8 -> 23.99 | Helped |
| Add Geo-diversification (ablation) | G3 · IM2GPS3K and YFCC4K | accuracy within 1km: 16.35 -> 16.65; 20.28 -> 23.99 | Helped |
| Add Geo-verification (ablation) | G3 · IM2GPS3K and YFCC4K | accuracy within 1km: 14.98 -> 16.65; 19.03 -> 23.99 | Helped |
| Add GPS alignment to image-only representations (ablation) | Geo-alignment on IM2GPS3K · IM2GPS3K | accuracy within 1km: 15.71 -> 16.91 | Helped |
Surprising findings
- Zero-shot RAG prompts (0 reference coordinates) can outperform reference-based prompts for landmark images, since the LMM's world knowledge identifies landmarks directly.
- Adding neighborhood-level text descriptions (G3-N) reduces performance, attributed to the text encoder's weak modeling of fine-grained neighborhood names.
Experimental setup
Datasets: MP16-Pro IM2GPS3k YFCC4k
Benchmarks: IM2GPS3k YFCC4k
Metrics: accuracy within 1km 25km 200km 750km 2500km thresholds
Baselines: [L]kNN PlaNet CPlaNet ISNs Translocator GeoDecoder GeoCLIP Img2Loc PIGEON
Builds on: Img2Loc CLIP GeoCLIP retrieval-augmented generation
Closest papers by meaning
- GeoRanker: Distance-Aware Ranking for Worldwide Image Geolocalization (NeurIPS 2025 · similarity 0.90)
- GeoRouter: Dynamic Paradigm Routing for Worldwide Image Geolocalization (NeurIPS 2026 · similarity 0.89)
- Vision-Language Reasoning for Geolocalization: A Reinforcement Learning Approach (AAAI 2026 · similarity 0.88)
- GRE Suite: Geo-localization Inference via Fine-Tuned Vision-Language Models and Enhanced Reasoning Chains (NeurIPS 2025 · similarity 0.88)
- Around the World in 80 Timesteps: A Generative Approach to Global Visual Geolocation (CVPR 2025 · similarity 0.87)
- Recognition through Reasoning: Reinforcing Image Geo-localization with Large Vision-Language Models (NeurIPS 2025 · similarity 0.87)
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.