Research map / Vision-language and multimodal / Human and embodied modeling
GeoRanker: Distance-Aware Ranking for Worldwide Image Geolocalization
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 GeoRanker, an LVLM-based distance-aware candidate re-ranker trained with a multi-order (first- and second-order) partial Plackett-Luce distance loss, plus the GeoRanking dataset, achieving SOTA worldwide geolocalization on IM2GPS3K and YFCC4K.
Problem
Worldwide image geolocalization in the two-stage retrieve-then-select pipeline suffers because candidate selection relies on independent cosine similarity and point-wise supervision, ignoring spatial relations among candidates.
Approach
A Qwen2-VL-based ranker jointly encodes the query image with each candidate (and negative candidates) in one prompt and outputs a scalar proximity score, trained with a loss that ranks both absolute distances (first-order) and pairwise distance gaps (second-order) via partial Plackett-Luce.
Key results
- GeoRanker reaches 18.79% street-level (1km) accuracy on IM2GPS3K vs 16.65% for G3, a 12.9% relative gain.
- On YFCC4K, GeoRanker reaches 32.94% at 1km vs 23.99% for G3, a 37.3% relative gain.
- Removing the second-order distance loss lowers IM2GPS3K 1km accuracy from 18.79% to 18.48% and continent-level from 89.29% to 88.28%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add second-order distance loss (ablation) | GeoRanker · IM2GPS3K, 1km | accuracy: 18.48 -> 18.79 | Helped |
| add negative candidates (ablation) | GeoRanker · IM2GPS3K, 1km | accuracy: 17.35 -> 18.79 | Helped |
| include candidate textual descriptions (ablation) | GeoRanker · IM2GPS3K, 1km | accuracy: 18.02 -> 18.79 | Helped |
| include candidate image data (ablation) | GeoRanker · IM2GPS3K, 1km | accuracy: 15.58 -> 18.79 | Helped |
| include generated candidates at inference (ablation) | GeoRanker · IM2GPS3K, 1km | accuracy: 18.21 -> 18.79 | Helped |
| add second-order distance loss (ablation) | GeoRanker · YFCC4K, 1km | accuracy: 31.97 -> 32.94 | Helped |
| add negative candidates (ablation) | GeoRanker · YFCC4K, 1km | accuracy: 31.57 -> 32.94 | Helped |
| include candidate textual descriptions (ablation) | GeoRanker · YFCC4K, 1km | accuracy: 31.70 -> 32.94 | Helped |
| include candidate image data (ablation) | GeoRanker · YFCC4K, 1km | accuracy: 15.81 -> 32.94 | Helped |
| include generated candidates at inference (ablation) | GeoRanker · YFCC4K, 1km | accuracy: 32.60 -> 32.94 | Helped |
Surprising findings
- Removing the image modality of candidates causes the largest drop, e.g. YFCC4K 1km falls from 32.94% to 15.81%.
- Increasing the first-order top-K (K^(1)) consistently degrades performance, attributed to train-test mismatch.
Limitations
- Adds a ranking stage that increases inference computation compared to pure embedding retrieval.
- Performance on coarse-grained ranking is less improved than fine-grained; hyperparameter sensitivity to lambda and K^(1).
Experimental setup
Datasets: MP16-Pro IM2GPS3K YFCC4K GeoRanking
Benchmarks: IM2GPS3K YFCC4K
Metrics: accuracy at 1km accuracy at 25km accuracy at 200km accuracy at 750km accuracy at 2500km Recall@K NDCG@K
Baselines: [L]kNN PlaNet CPlaNet ISNs Translocator GeoDecoder GeoCLIP Img2Loc PIGEON G3
Builds on: G3 Img2Loc GeoCLIP Plackett-Luce loss LoRA
Closest papers by meaning
- GeoRouter: Dynamic Paradigm Routing for Worldwide Image Geolocalization (NeurIPS 2026 · similarity 0.92)
- PIGEON: Predicting Image Geolocations (CVPR 2024 · similarity 0.90)
- G3: An Effective and Adaptive Framework for Worldwide Geolocalization Using Large Multi-Modality Models (NeurIPS 2024 · similarity 0.90)
- Around the World in 80 Timesteps: A Generative Approach to Global Visual Geolocation (CVPR 2025 · similarity 0.89)
- Scaling Image Geo-Localization to Continent Level (NeurIPS 2025 · similarity 0.88)
- Vision-Language Reasoning for Geolocalization: A Reinforcement Learning Approach (AAAI 2026 · similarity 0.88)
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.