1Xiamen University2Tsinghua University3University of Macau
Fine-grained cross-view localization aims to precisely localize a ground camera against geo-referenced aerial imagery, but current methods typically assume a narrow geographic prior, substantially limiting their applicability when upstream localization remains coarse. To bridge this gap, we study this task over much larger aerial regions. However, this seemingly simple extension fundamentally changes the localization problem: the growing disparity in spatial extent between ground and aerial views makes reliable cross-view matching difficult to establish directly, while the increasing number of look-alike locations in larger regions introduces substantial visual ambiguity.
To address these challenges, we propose RingMatch, a search-and-verification framework for wide-area cross-view localization. RingMatch first leverages predictive uncertainty to uncover plausible locations and progressively zooms into their surroundings to obtain increasingly fine-grained evidence, yielding a compact set of promising candidates. We then introduce the depth ring, a novel geometric representation that reveals subtle variations in the spatial layout around a location and serves as a view-shared spatial signature for an LVLM verifier to resolve residual ambiguity.
To benchmark this setting, we construct VIGOR-Pro by composing VIGOR aerial tiles into contiguous regions with spatial extents of up to 400 m. Experiments show that RingMatch maintains strong localization performance as the aerial region expands, consistently outperforming representative end-to-end and retrieval-based methods.
Expanding the aerial region introduces two coupled challenges: sparse ground–aerial correspondences and visual ambiguity among look-alike locations.
A ground panorama observes only the camera’s local surroundings. As the aerial extent grows, reliable cross-view evidence occupies a smaller, unknown part of the region, making direct localization increasingly difficult.
Recurring scene layouts create multiple plausible locations. Resolving them requires comparing subtle spatial differences across the ground–aerial viewpoint gap.
Given a ground panorama and a geo-referenced aerial region containing the camera, RingMatch combines multi-hypothesis search with geometric verification. We consider the known-orientation setting: the panorama heading is known relative to the north-up aerial image.
Sample multiple coordinate predictions and cluster nearby predictions into hypotheses, preserving distinct plausible locations.
Recursively narrow the region around each retained hypothesis to recover finer cross-view evidence.
Compare the ground depth ring with predicted aerial rings. An LVLM verifier uses ring patterns and visual context to select the final candidate.
A spatial signature that both views can share.
Search produces a shortlist of plausible locations, but look-alike candidates can remain difficult to distinguish. A depth ring records the estimated distance from a location to the first scene surface along each horizontal direction. Its profile captures fine-grained spatial structure around that location.
Ground and aerial rings express this structure in a shared, camera-centered representation. In the known-orientation setting, they share an azimuth reference and can be directly aligned.
Estimate metric depth from the panorama, then pool depth values within a narrow band around the horizon for each azimuth.
Predict a ring from each candidate-centered aerial patch. Ground-derived rings from paired training data provide supervision, requiring no additional annotation.
The ground depth ring is compared with predicted aerial rings at two candidate locations. Consistent profiles support the matching location, while differences in the surrounding layout reveal a look-alike.
Localization errors on VIGOR-Pro under same-area and cross-area evaluation. All experiments use the known-orientation setting. Mean and median errors are measured in meters; lower is better.
| Approach | Method | Mean (m) ↓ | Median (m) ↓ |
|---|
At 400 m under cross-area evaluation, RingMatch achieves 19.81 m mean and 2.04 m median error, compared with 26.38 m and 4.78 m for Sample4Geo + FG².
The 100, 200 and 400 m labels denote nominal aerial-region side lengths.
Depth-ring evidence improves the candidate verification accuracy of GPT-5.6 Terra and Gemini-3.7 Flash for both aerial-region sizes, without task-specific fine-tuning. Evaluation uses VIGOR-Pro same-area verification samples.
| Model | Evidence | Fine-tuned | 200 m | 400 m |
|---|---|---|---|---|
| Random | Random choice | - | 50.00 | 50.00 |
| Ring-only | Ring correlation | - | 67.29 | 69.42 |
| GPT-5.6 Terra | Visual only | - | 56.93 | 59.47 |
| Visual + depth ring | 67.58+10.65 pp | 66.21+6.74 pp | ||
| Gemini-3.7 Flash | Visual only | - | 66.70 | 72.56 |
| Visual + depth ring | 69.92+3.22 pp | 75.29+2.73 pp | ||
| RingMatch | Visual + depth ring | Yes | 81.92 | 82.81 |
@misc{hu2026localrefinementwideareaprecise,
title={Beyond Local Refinement: Wide-Area Precise Cross-View Localization},
author={Qingguo Hu and Xinyu Ma and Yujie Lin and Yaoxiang Wang and Qi Shi and Zonghao Guo and Chi Chen and Jinsong Su},
year={2026},
eprint={2611.04759},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2611.04759}
}