Align scales globally
Level-wise InfoNCE trains one dual encoder while keeping negatives within the same physical scale.
arXiv 2026 Cross-view geo-localization
GeoMoE
One Query, Many Scales: Sparse Mixture-of-Experts for Efficient Hierarchical Cross-View Geo-Localization
1Tsinghua Shenzhen International Graduate School, Tsinghua University
2School of Geospatial Engineering and Science, Sun Yat-sen University
62.39%
R@1 on VIGOR-M
95.78%
R@40m on JustZoomIn
5.27%
of exhaustive L3 matching cost
01 / Overview
Global embeddings.
Local search.
GeoMoE decouples multi-scale representation learning from hierarchical retrieval, so one query descriptor can travel from a city-scale gallery to the finest satellite tiles.
Cross-view geo-localization matches a ground query to geo-tagged satellite imagery. GeoMoE replaces fixed-resolution exhaustive search with a shared DINOv2 dual encoder trained across scales and a sparse Top-2 mixture of experts. At inference, every image is encoded once; a probabilistic K=4 beam follows parent-child links and scores only a compact candidate set. A calibrator trained on the official training split reranks final paths without test-time label adaptation.
Core idea
Learn descriptors that remain comparable everywhere, then use hierarchy only to decide where to spend the search budget.
02 / Geographic hierarchy
From city scale
to a single place.
VIGOR-M preserves each city as a continuous surface served by an aligned map-tile pyramid. GeoMoE follows this hierarchy while keeping one globally comparable descriptor at every scale.
Seattle, Washington
47.5922 N 122.3323 W
City-scale gallery
49,152 px source
03 / Method
One encoder across every scale.
A shared backbone learns a globally comparable space; sparse routing and coarse-to-fine search allocate capacity only where it matters.
Route content sparsely
The final Transformer FFN routes each token through the best two of five learned experts.
Search coarse to fine
A K=4 probabilistic beam expands parent-child paths before a train-only residual calibrator reranks the finalists.
04 / Results
Accurate without exhaustive search.
Released B11/E5 Top-2 checkpoint, fixed K=4 beam, and PRC fitted only on the official training split.
New benchmark
VIGOR-M
| R@1 | R@100m | R@200m | R@300m | Median |
|---|---|---|---|---|
| 62.39% | 77.82% | 84.30% | 86.86% | 52.33 m |
Large-area benchmark
JustZoomIn
| R@1 | R@40m | R@50m | R@100m | Median |
|---|---|---|---|---|
| 81.71% | 95.78% | 96.32% | 97.68% | 16.43 m |
05 / Dataset
VIGOR-M
A four-city benchmark built for single-resolution, cross-resolution, and hierarchical geo-localization.
VIGOR-M adds an explicit parent-child satellite hierarchy and held-out half-step galleries, making resolution transfer a measurable capability rather than an incidental side effect.
Explore the dataset4cities
75,684panoramas
17,472native tiles
3evaluation protocols
06 / Citation
Cite GeoMoE.
If GeoMoE or VIGOR-M supports your research, please cite the paper. Code is released under Apache-2.0; dataset licenses and third-party notices remain separate.
Recommended citation
One Query, Many Scales: Sparse Mixture-of-Experts for Efficient Hierarchical Cross-View Geo-Localization
@misc{fan2026queryscalessparsemixtureofexperts,
title = {One Query, Many Scales: Sparse Mixture-of-Experts for Efficient Hierarchical Cross-View Geo-Localization},
author = {Ruijie Fan and Junyan Ye and Qi Zhu and Weijia Li},
year = {2026},
eprint = {2608.01060},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2608.01060},
}