Research map / Image generation / Autoregressive and latent generation
PIGEON: Predicting Image Geolocations
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.
Introduces PIGEON and PIGEOTTO, CLIP-based planet-scale image geolocalization systems using semantic geocells, haversine-smoothed labels, multi-task contrastive pretraining, and cluster-based retrieval refinement.
Problem
Planet-scale image geolocalization that generalizes to unseen places and diverse images, rather than narrow landmark distributions.
Approach
Combines semantic geocells built from administrative boundaries and OPTICS clustering with Voronoi tessellation, a haversine-distance-smoothed classification loss sharing parameters across implicit geographic hierarchies, multi-task continued CLIP pretraining with synthetic geographic captions, and retrieval over location clusters to refine guesses across top-K geocells.
Key results
- PIGEON places 40.36% of guesses within 25 km on a 5,000-location GeoGuessr Street View holdout set, with median error 44.35 km.
- PIGEOTTO improves country-level accuracy on IM2GPS3k by 11.4 points and on GWS15k by 38.8 points over the prior SOTA.
- PIGEON ranks in the top 0.01% of GeoGuessr players in 458 blind matches and beats a professional player 6-0.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add freezing of the last CLIP layer after pretraining (ablation) | PIGEON · GeoGuessr holdout, 5,000 Street View locations | median error (km): 45.47 -> 44.35 | Helped |
| add hierarchical guess refinement (ablation) | PIGEON · GeoGuessr holdout, 5,000 Street View locations | median error (km): 50.01 -> 44.35 | Helped |
| add contrastive CLIP pretraining (ablation) | PIGEON · GeoGuessr holdout, 5,000 Street View locations | median error (km): 55.51 -> 44.35 | Helped |
| use semantic geocells instead of non-semantic geocells (ablation) | PIGEON · GeoGuessr holdout, 5,000 Street View locations | median error (km): 60.63 -> 44.35 | Helped |
| add multi-task prediction heads (ablation) | PIGEON · GeoGuessr holdout, 5,000 Street View locations | median error (km): 61.81 -> 44.35 | Helped |
| fine-tune the last CLIP layer (ablation) | PIGEON · GeoGuessr holdout, 5,000 Street View locations | median error (km): 60.81 -> 44.35 | Helped |
| use a four-image panorama instead of a single image (ablation) | PIGEON · GeoGuessr holdout, 5,000 Street View locations | median error (km): 131.1 -> 44.35 | Helped |
| add haversine smoothing (ablation) | PIGEON · GeoGuessr holdout, 5,000 Street View locations | median error (km): 148.0 -> 44.35 | Helped |
| train PIGEOTTO rather than GeoDecoder | PIGEOTTO · IM2GPS | country accuracy (% @ 750 km): +2.3 percentage points | Helped |
| train PIGEOTTO rather than GeoDecoder | PIGEOTTO · IM2GPS3k | country accuracy (% @ 750 km): +11.4 percentage points | Helped |
| train PIGEOTTO rather than GeoDecoder | PIGEOTTO · YFCC4k | country accuracy (% @ 750 km): +12.2 percentage points | Helped |
| train PIGEOTTO rather than GeoDecoder | PIGEOTTO · YFCC26k | country accuracy (% @ 750 km): +13.6 percentage points | Helped |
Surprising findings
- Four-image panorama input is the largest single contributor, more than halving median error (131.1 to 60.8 km in ablation).
- 92% of GWS15k locations are not within 100 m of any MediaEval 2016 training image, yet PIGEOTTO still generalizes to them.
Limitations
- Model weights withheld due to dual-use and misuse risks.
- Street-level predictions remain unreliable according to the authors.
- PIGEON requires four-image panoramas, which are often unavailable in general settings.
Experimental setup
Datasets: GeoGuessr Street View panoramas (collected) MediaEval 2016 (YFCC) Google Landmarks v2 GADM administrative boundaries IM2GPS IM2GPS3k YFCC4k YFCC26k GWS15k
Benchmarks: IM2GPS IM2GPS3k YFCC4k YFCC26k GWS15k
Metrics: median distance error (km) % at 1km/25km/200km/750km/2500km GeoGuessr score mean distance error
Baselines: PlaNet CPlaNet ISNs Translocator GeoDecoder
Builds on: CLIP PlaNet Translocator GeoDecoder Theiner et al. semantic geocells OPTICS
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 0.88)
- OpenStreetView-5M: The Many Roads to Global Visual Geolocation (CVPR 2024 · similarity 0.88)
- Vision-Language Reasoning for Geolocalization: A Reinforcement Learning Approach (AAAI 2026 · similarity 0.88)
- SatCLIP: Global, General-Purpose Location Embeddings with Satellite Imagery (AAAI 2025 · similarity 0.87)
- Scaling Image Geo-Localization to Continent Level (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.