acceptodds
Withdrawn from ICLR 2027

Towards Diverse Global Image Geolocalization: Baselines, Dataset, and Benchmark

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We introduce Diverse Global Image Geolocalization, an extended task that generalizes image geolocalization beyond the street-view domain to ten distinct domains. Inputs mainly originate from continuous domains, such as street-view, remote sensing, terrain and plant, and maps, or from discrete domains, such as landmark, road-network, indoor, shapes, space, and UAV imagery. We provide a complete suite for this task, comprising GeoVerse100K, a training dataset, a ten-domain evaluation benchmark, and a baseline. GeoVerse100K is constructed with a geographic-attribute sampling method to ensure balanced geographic coverage and a reward-based sampling method to better support RL-based training. We further divide the benchmark into 37 subcategories and conduct a comprehensive evaluation of both general and specialized models. To establish a strong baseline, we propose two training strategies tailored to the multi-domain setting: domain-balanced sampling and a difficulty-guided curriculum. We then finetune Qwen3.5-9B using GRPO with a distance-aware reward. Experiments show that current specialized models suffer from severe generalization limitations, whereas our baseline trained on GeoVerse100K achieves balanced performance across all ten domains. The code and dataset will be made publicly available.

Primary area:
datasets and benchmarks
Keywords:
Global Image Geolocalization, Multi-Domain Datasets and Benchmarks, Vision-Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_VejvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Diverse Global Image Geolocalization: predicting coordinates for images from ten domains, some continuous, the others discrete. The contributions are the dataset with weights provided by external information or other VLMs, and a baseline.

Critical strengths

Extending geolocation beyond street-view to remote sensing, maps, astronaut photos, UAV and indoor images is interesting. The 37-subcategory and per-continent breakdowns allow fine-grained analysis.

The dataset has also been carefully constructed with geographic rebalancing and shared anchor locations across the continuous domains.

Critical weaknesses

Proprietary VLMs (GPT-5, Gemini 2.5/3, Claude) are missing. They are the strongest general-purpose geolocators and essential for a benchmark paper. Specialized SOTA geolocators are missing: PIGEON/PIGEOTTO, OSV-5M baselines, GeoRanker, GeoAgent, GRE Suite, Around the World, LocDiff and GeoReasoner, several of which the paper cites.

GeoVerse-Val comes from the same sources and pipeline as the training set, so GeoVerse is tested in-distribution. The baseline is not evaluated on any standard external benchmark: IM2GPS3k, YFCC4k, OSV-5M. So this work can not conclude that GeoVerse training leads to good geographical knowledge or just learning a distribution. Also, comparing to state of the art geolocation methods that do not require a VLM, especially flow-matching ones, might be interesting to evaluate the gap between specialized methods and VLMs.

LLMs may be fond of decimals, but reporting GeoScore to two decimal places is meaningless here

Questions for authors

How does GeoVerse perform on IM2GPS3k, YFCC4k and OSV-5M test, compared with methods from "Around the World in 80 Timesteps: A Generative Approach to Global Visual Geolocation"?

How do frontier proprietary VLMs score on GeoVerse-Val?

What are the licenses of each source, and what exactly will be released?

Reviewer_RtMhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents GeoVerse100K and GeoVerse-Val, a dataset and benchmark for global image geolocalization across ten visual domains. It also develops a training baseline using geographic and reward-aware data selection, followed by GRPO with domain interleaving and difficulty scheduling. Alongside the benchmark comparisons, the paper studies how training on one domain affects performance on others.

Critical strengths
  1. The main value of this work is the breadth of the benchmark. Bringing together street-view, remote sensing, map, shape, indoor, and other images opens up questions that are difficult to study with a street-view dataset alone. A carefully validated resource covering these domains would be useful to the community.
  2. The paper includes more analysis than a standard benchmark comparison. The transfer matrix in Figure. 7 is particularly helpful for understanding where domain-specific training helps and where it introduces interference. The additional InternVL experiment in the appendix also provides some evidence that the training approach is useful beyond the primary backbone.
Critical weaknesses
  1. I am not convinced that the same point-distance evaluation has a consistent meaning across all ten domains. Section.5 converts predicted place names into coordinates using OpenCage. For an administrative silhouette, however, the answer is naturally a geographic entity. For a large map or remote sensing image, it may be an extended region. A correct place name could receive a poor score if its geocoded point is far from the annotated image center. Similarly, being within 750 km does not necessarily mean that the predicted country is correct. Since the benchmark is a central contribution, I would like clearer definitions of the targets in each domain and a check of whether entity-based or region-aware evaluation changes the conclusions.
  2. The geographic train–test separation needs to be more explicit. The continuous domains use shared geographic anchors and multiple spatial scales, so independently sampling evaluation images does not necessarily rule out nearby locations or overlapping image footprints across splits and domains. The UAV cluster exclusion is a useful precaution, but I could not find an equally clear account for the remaining domains. This does not establish that leakage occurred but leaves an important uncertainty. Also because the final model is trained on all ten evaluated domains and tested mainly on GeoVerse-Val, the results provide stronger evidence for adaptation to this benchmark than for the broader generalization claims.
Questions for authors
  1. How are ground-truth locations defined for silhouettes and images covering large regions? Have you checked the error introduced by geocoding a correct place name?
  2. Are training and test data separated jointly across domains by location or geographic entity? Please clarify how shared anchors, zoom levels, overlapping footprints, and repeated landmarks are handled.
  3. Do you have results on an external benchmark, a held-out geographic region, or a domain excluded from training? Any of these would help assess the scope of the generalization claims.
Reviewer_dLETOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper extends global image geolocalization to ten visual domains (four continuous: street-view, remote sensing, terrain-and-plant, map; six discrete: landmark, road-network, indoor, shape, space, UAV) and provides a suite for the task: (1) GeoVerse100K, a ~100K-image dataset curated from public sources via geographic attribute sampling (country quotas by area/population/road length, log-population grid sampling) and reward-aware sampling, which scores each image by the standard deviation and maximum of distance-aware rewards over 3 models × 8 rollouts to favor samples that yield informative GRPO signals; (2) GeoVerse-Val, a 9,000-image benchmark with 37 subcategories and continent labels, evaluated by geocoded predictions, threshold accuracy, and a GeoScore metric; (3) GeoVerse, a baseline trained from Qwen3.5-9B with GRPO, domain-ratio interleaving, and a difficulty-guided curriculum. Experiments compare four general VLMs and four specialized models, report a 10×10 single-domain transfer matrix, and ablate the sampling and training strategies.

Critical strengths
  1. The multi-domain, geographically rebalanced benchmark fills a genuine gap, and the 10×10 cross-domain transfer matrix (Fig. 7) is an informative empirical contribution: it documents both positive transfer and widespread interference from single-domain training, which no prior geolocalization benchmark could show.

  2. Reward-aware sampling is a well-motivated data-curation idea for verifiable-reward RL, targeting the known failure mode of near-zero within-group reward variance in GRPO, and is supported by reward-distribution statistics (Fig. 5).

Critical weaknesses
  1. The central empirical claims rest on a circular comparison with an outdated and incomplete baseline set. GeoVerse is trained and evaluated in distribution on the authors’ own pipeline, while the specialized baselines are evaluated out of distribution, so “specialized models suffer severe generalization limitations, whereas our baseline achieves balanced performance” is largely tautological. The specialized baselines moreover stop at 2025 (GeoClip, G3, GLOBE, Geo-R1, all ≤10B open models): the 2026 reasoning- and tool-based methods the paper itself cites (GeoSeek, Geo-R, agentic pipelines) and frontier proprietary VLMs, typically the strongest geolocalizers, are entirely absent. For a benchmark paper, this evaluation gap undermines both headline conclusions. Train/eval decontamination is also documented for only some domains; nothing is stated to prevent near-duplicates from GLDv2, Hotels-50K, or OSM-derived renderings crossing the split, making the in-domain numbers hard to interpret.

  2. The ablations avoid the questions that matter most. All ablation tables report only 3-4 hand-picked domains at a single threshold, never the weighted all-domain metric used for the headline claims, and the domain subsets differ across Tabs. 1-3. There is no GRPO-vs-SFT comparison on the same data, although the dataset’s stated purpose is suitability for RL; if SFT benefits equally, the reward-aware design loses its justification. Eq. 4 is never decomposed (σ-only vs r_max-only; α, β fixed at 1.0), the reference models driving the sampling are never identified or varied, raising a circularity concern if the base model is among them, difficulty scheduling is tested only as on/off despite Tab. 1 showing it hurts some domains, and no data-scale ablation shows curated 100K beats a larger uncurated sample. The core methodological contributions are therefore asserted more than validated.

Questions for authors
  1. Which three generation models are used in reward-aware sampling? Is Qwen3.5-9B among them, and how do Fig. 5 and Tab. 2 change with a different reference set?
  2. Is near-duplicate removal between training data and GeoVerse-Val enforced for all ten domains? Please give the per-domain protocol, especially for GLDv2, Hotels-50K, and OSM-derived renderings.
  3. Can you evaluate at least one 2026 reasoning/tool-based method and one frontier proprietary VLM, or justify their exclusion? How should readers separate the baselines’ “generalization failure” from ordinary distribution shift, given only GeoVerse is in distribution?
  4. Please report the Tab. 1 ablation on all ten domains with the weighted GeoScore, and add an SFT baseline on the same data. Does reward-aware sampling help under SFT, or only under GRPO?
  5. What fraction of textual predictions fail or misresolve under OpenCage geocoding, how are failures scored, and does this disadvantage models that output place names rather than coordinates?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.