Worldwide image geolocalization aims to predict precise GPS coordinates for images captured anywhere on Earth, which is challenging due to the large visual and geographic diversity. Recent methods mainly follow two paradigms: retrieval-based approaches that match queries against a reference database, and generation-based approaches that directly predict coordinates using Large Vision-Language Models (LVLMs). However, we observe distinct error profiles between them: retrieval excels at fine-grained instance matching, while generation offers robust semantic reasoning. This complementary heterogeneity suggests that no single paradigm is universally superior. To harness this potential, we propose GeoRouter, a dynamic routing framework that adaptively assigns each query to the optimal paradigm. GeoRouter leverages an LVLM backbone to analyze visual content and provide routing decisions. To optimize GeoRouter, we introduce a distance-aware preference objective that converts the distance gap between paradigms into a continuous supervision signal, explicitly reflecting relative performance differences. Furthermore, we construct GeoRouting, the first large-scale dataset tailored for training routing policies with independent paradigm predictions. Extensive experiments on IM2GPS3k and YFCC4k demonstrate that GeoRouter significantly outperforms state-of-the-art baselines.
Figures & tables
Figure 1: Paradigm Complementarity in Worldwide Image Geolocalization.
Figure 2: Overview of the dynamic routing framework – GeoRouter.
Methods
IM2GPS3K
YFCC4K
Street 1km
City 25km
Region 200km
Country 750km
Continent 2500km
Street 1km
City 25km
Region 200km
Country 750km
Continent 2500km
[L]kNN , sigma=4 [ 43 ]
ICCV’17
7.2
19.4
26.9
38.9
55.9
2.3
5.7
11.0
23.5
42.0
PlaNet [ 48 ]
ECCV’16
8.5
24.8
34.3
48.4
64.6
5.6
14.3
22.2
36.4
55.8
CPlaNet [ 49 ]
ECCV’18
10.2
26.5
34.6
48.6
64.6
7.9
14.8
21.9
36.4
55.5
ISNs [ 50 ]
ECCV’18
10.5
28.0
36.6
49.7
66.0
6.5
16.2
23.8
37.4
55.0
Translocator [ 51 ]
ECCV’22
11.8
31.1
46.7
58.9
80.1
8.4
18.6
27.0
41.1
60.4
Table 1: Main results on the IM2GPS3K and YFCC4K benchmarks. Higher values indicate better performance. The best and second-best results are highlighted in bold and underlined , respectively. Δ denotes the relative improvement of GeoRouter over the best baseline.
Methods
Street 1km
City 25km
Region 200km
Country 750km
Continent 2500km
Average
w/o DisPO
20.85
49.38
64.26
79.05
89.76
60.66
w/o C
20.69
49.18
64.60
79.75
90.09
60.86
w/o C & gret
19.22
46.11
61.76
77.11
89.29
58.69
w/o ggen
20.22
48.31
64.06
78.75
89.92
60.25
w/o Context
19.29
45.85
61.46
77.41
89.16
58.63
Ours
20.82
50.48
65.73
80.35
90.66
61.60
Table 2: Ablation study on IM2GPS3K.
Figure 3: Effect of the hyperparameter α on geolocalization accuracy. The subfigures present the results evaluated on the IM2GPS3K dataset across distance thresholds ranging from 1km to 2500km.
Dataset
Method
Geolocalization Accuracy (%)
Routing Accuracy (%)
1km
25km
200km
750km
2500km
Average
1km
25km
200km
750km
2500km
Average
IM2GPS3K
Pure Retrieval
18.42
43.14
59.29
75.31
88.36
56.90
57.28
48.44
48.48
47.57
48.37
50.03
Pure Generation
17.55
47.91
63.30
78.85
88.59
59.24
42.72
51.56
51.52
52.43
51.63
49.97
GeoRouter
20.82
50.48
65.73
80.35
90.66
61.61
65.52
63.36
62.09
61.61
61.30
62.78
Oracle
24.29
54.55
71.37
85.09
93.89
65.84
100.00
100.00
100.00
100.00
100.00
100.00
YFCC4K
Pure Retrieval
32.47
42.70
53.13
69.95
82.67
56.18
85.23
73.34
67.51
65.47
64.44
71.20
Table 3: Comparison of geolocalization and routing accuracy across different distance thresholds. The best performance among the realistic (non-Oracle) methods is highlighted in bold , while the Oracle results are presented in italics as the theoretical upper bound.
Figure 7
Figure 6: Case studies illustrating GeoRouter’s routing decisions across two paradigms.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Representative examples of GeoRouting dataset entries. Each row shows a query image with its ground-truth GPS coordinates, predictions from the retrieval-based and generation-based paradigms with their respective distance errors, retrieved candidate images with their GPS coordinates, and the binary routing label (1 = generation-based paradigm is more accurate; 0 = retrieval-based paradigm is more accurate).
Parameter
Value
Batch Size per Device
3
Batch Size
24
Dataset Size
100K
DeepSpeed Stage
2
GPU
NVIDIA H800 × 8
Training Time
50 min / epoch
Appendix
Table 4: Detailed experimental configurations.
Figure 8: Qualitative examples of GeoRouter predictions across different geolocalization accuracy levels
Methods
IM2GPS3K
YFCC4K
Street 1km
City 25km
Region 200km
Country 750km
Continent 2500km
Average
Street 1km
City 25km
Region 200km
Country 750km
Continent 2500km
Average
w/o DisPO
20.85
49.38
64.26
79.05
89.76
60.66
33.05
45.75
57.08
71.65
82.80
58.06
w/o C
20.69
49.18
64.60
79.75
90.09
60.86
33.58
45.55
56.42
71.94
83.53
58.20
w/o C & gret
19.22
46.11
61.76
77.11
89.29
58.69
29.83
41.34
52.05
68.67
81.50
54.67
w/o ggen
20.22
48.31
64.06
78.75
89.92
60.25
33.20
45.08
55.62
71.25
82.94
57.61
w/o Context
19.29
45.85
61.46
77.41
89.16
58.63
30.14
41.60
52.98
69.91
82.34
55.39
Appendix
Table 5: Ablation study on IM2GPS3K and YFCC4K.
Figure 9: Effect of the hyperparameter α on geolocalization accuracy across all five distance thresholds on both IM2GPS3K (blue) and YFCC4K (yellow). Each column corresponds to a distance threshold from Street level (1km) to Continent level (2500km).
Base Generative Model
Method
Distance Thresholds
Δ Mean
1km
25km
200km
750km
2500km
Mean
Gemini 2.5 Flash
Base Model
17.55
47.91
63.30
78.85
88.59
59.24
+4.00%
+ GeoRouter
20.82
50.48
65.73
80.35
90.66
61.61
Gemini 2.5 Flash (Thinking)
Base Model
18.59
48.88
62.96
78.61
88.99
59.61
+3.37%
+ GeoRouter
21.15
50.98
64.83
80.35
90.79
61.62
Gemini 2.0 Flash
Base Model
15.42
45.38
61.23
78.14
87.92
57.62
+4.84%
Appendix
Table 6: Transferability of the GeoRouter across different generative LVLMs on the IM2GPS3K dataset. All metrics are reported as geolocalization accuracy (%). The Δ Mean column indicates the performance gain achieved by integrating our routing mechanism.
Backbone Model
IM2GPS3K
YFCC4K
1km
25km
200km
750km
2500km
Mean
1km
25km
200km
750km
2500km
Mean
Qwen2-VL-2B
20.35
49.02
64.36
79.35
90.06
60.63
33.49
45.50
56.97
71.38
83.00
58.07
Qwen2-VL-7B
20.82
50.48
65.73
80.35
90.66
61.61
32.98
46.01
57.52
72.02
83.02
58.31
Qwen3-VL-2B
20.95
49.68
64.80
79.91
90.72
61.21
33.55
45.90
57.30
71.89
83.40
58.41
Qwen3-VL-4B
21.02
49.98
65.37
79.98
90.72
61.41
33.82
46.45
57.76
72.53
83.44
58.80
Qwen3-VL-8B
20.99
50.18
65.33
80.28
90.66
61.49
33.71
46.45
57.80
73.06
84.37
59.08
Appendix
Table 7: Performance scaling across different sizes of the routing backbone model. The evaluation is conducted using Qwen2-VL and Qwen3-VL families on the IM2GPS3K and YFCC4K datasets. All results are presented as geolocalization accuracy (%).
Figure 10: Detailed data efficiency analysis across five geographic levels on the IM2GPS3K (top row) and YFCC4K (bottom row) datasets. The solid blue line represents the accuracy of GeoRouter trained on varying proportions of the dataset, while the dashed yellow line indicates the performance of the state-of-the-art baseline GeoRanker.
GPU
Batch Size
IM2GPS3K
YFCC4K
GPU Memory Consumption
Inference Speed
GPU Memory Consumption
Inference Speed
NVIDIA H800 × 1
16
Avg 65 GiB
0.1054s per figure
Avg 65 GiB
0.0551s per figure
Appendix
Table 8: Inference overhead of GeoRouter on the IM2GPS3K and YFCC4K datasets.
Large language model routing selects models, tools, experts, or inference strategies to balance answer quality, cost, latency, and deployment constraints.
Department of Data Science, City University of Hong Kong · Department of Computer Sciences, University of Wisconsin-Madison · Department of Geography, University of Wisconsin-Madison