Attention-Guided Sequence Feature Fusion and Mixed-Space Metric Learning for Cross-View Image Sequence Geo-Localization
Authors: Nan Wu, Chunfang Yang, Jiangshan Li, Xiangyang Luo
Organizations: Information Engineering University, Zhengzhou, 450001, China, and the Henan Key Laboratory of Cyberspace Situation Awareness, Zhengzhou, 450001, China. · the College of Electronic Science and Technology, National University of Defense Technology, Changsha 410073, China.
Cross-view image sequence geo-localization aims to determine the geographic location of a sequence of ground images by matching it against geo-tagged satellite imagery. Although sequential observations provide complementary scene information, substantial overlap between adjacent images introduces feature redundancy, while conventional fusion strategies may weaken spatial relationships essential for cross-view matching. Therefore, we propose a cross-view image sequence geo-localization method via Attention-guided Sequence fusion and Metric Learning (ASML). Specifically, a Sequential Content Fusion (SCF) module models cross-image feature correlations to suppress redundant content and adaptively preserve geographically informative non-overlapping regions. Next, a Spatial Feature Aggregation (SFA) module is used to capture and aggregate the spatial and semantic information of local features, so as to maintain the spatial layout of the image sequence. Finally, a Geo-InfoNCE loss function is proposed, which comprehensively considers the feature distance and geographic distance for mixed-space metric learning, thus improving the ability of the model to distinguish hard negative samples. A new cross-view image sequence dataset CrossRoad-66K is established to provide more complex scenarios. Extensive experiments and visualizations on the SeqGeo and CrossRoad-66K datasets demonstrate that compared to state-of-the-art (SOTA) methods, the proposed method ASML achieves an improvement of 3.37% and 2.10% at R@1, respectively.
Figures & tables
Fig. 1: Quantitative and qualitative analysis of overlapping areas in cross-view image sequences. (a) The average ratio of overlapping areas among ground image sequences with different sequence lengths on the SeqGeo and CrossRoad-66K datasets. (b) Visualization of overlapping regions between adjacent ground images. This figure provides a dataset-level analysis of sequence redundancy rather than a performance comparison between different methods.
Fig. 2: A pair of samples in SeqGeo dataset. (a) shows the satellite image. (b)-(d) show the ground images captured along a straight road, whose locations are marked by green points in (a). (e)-(g) show the ground images captured around a corner, whose locations are marked by red points in (a).
Max Masks
Structural Similarity
KL Divergence
L2 Distance ↓
Cosine Similarity ↑
G2S( 10−5 ) ↓
S2G( 10−5 ) ↓
5
0.1033
0.0465
1.7548
1.7551
3
0.1019
0.0480
1.6820
1.6826
1
0.1021
0.0478
1.6935
1.6945
0
0.1024
0.0475
1.7083
1.7088
TABLE I: Comparative Analysis of Ground Image Sequence Feature with Satellite Image Feature Using Structural Similarity and KL Divergence. ↓ Indicates Lower Values with Greater Similarity and ↑ Indicates Higher Values with Greater Similarity.
Fig. 3: Overview architecture of ASML. Specifically, the Sequential Content Fusion Module (SCF) integrates the geographical semantics among the ground image sequence, and reduces the number of tokens in the subsequent processing. Spatial Feature Aggregation Module (SFA) mainly contains a Transformer Encoder, which is used to aggregate feature descriptors. The Geo-InfoNCE Loss is used to learn discriminative features in feature space and geographic space.
Fig. 4: The operational workflow of the Sequential Content Fusion Module (SCF). First, a similarity scores map Si is calculated between the adjacent local features. Next, an attention scores map Ai+1 is generated from similarity scores map, which is then translated to an attention mask Mi+1 . Finally, the attention mask is utilized to select the features in non-overlapping areas and eliminate redundant features.
Fig. 5: The main processing of the spatial feature aggregation module (SFA). (a) shows the workflow of SFA module. (b) shows the detailed structure of the Transformer Encoder in (a).
Fig. 6: Visualization of some hard negative samples in the SeqGeo dataset, which are easily confused. (a) shows one image within the ground image sequence. (b) shows the matched satellite image. (c) shows the unmatched satellite image. The left ground image and the right satellite image are a pair of difficult samples. Each image is labeled with the GPS coordinates of the capture location.
Fig. 7: The mixed distance metric optimization strategy for the samples within a batch. The samples matrix represents the relationship between the samples within a batch, where the elements on the diagonal are positive samples, while the others are negative samples. M refers to the feature distance matrix D minus the GPS distance matrix G .
Datasets
CVACT [ 42 ]
CVUSA [ 43 ]
VIGOR [ 44 ]
SeqGeo [ 20 ]
CrossRoad-66K
Satellite Images
44,416
44,416
90,618
38,863
21,320
Ground Images
44,416
44,416
105,214
118,549
66,384
Ground Images per Satellite Image
1
1
5
7
6
Seamless Sampling
×
×
✓
✓
✓
Sequential Ground Images
×
×
×
✓
✓
Full Panorama
✓
✓
✓
×
×
TABLE II: Comparison between the proposed CrossRoad-66K dataset and existing cross-view image datasets.
β
SeqGeo
CrossRoad-66K
R@1
R@5
R@1%
R@1
R@5
R@1%
8.3%
15.89%
34.61%
74.15%
4.57%
13.02%
38.30%
16.7%
21.53%
41.46%
79.04%
5.25%
15.75%
42.82%
25%
21.34%
41.69%
78.44%
4.99%
14.81%
41.76%
33.3%
19.97%
38.45%
76.37%
4.61%
14.24%
41.38%
Adap.
23.53%
44.30%
80.77%
6.20%
16.83%
43.94%
TABLE III: Hyperparameter tuning for the quantile threshold β .
Methods
SeqGeo
CrossRoad-66K
R@1
R@5
R@10
R@1%
R@1
R@5
R@10
R@1%
SeqGeo [ 20 ]
2.07%
8.12%
13.16%
40.10%
1.86%
7.69%
13.56%
35.65%
GeoDTR+ [ 45 ]
2.52%
7.29%
11.69%
36.85%
1.29%
3.89%
6.32%
17.93%
GeoDTR [ 46 ]
3.20%
9.03%
14.08%
39.98%
3.15%
9.14%
14.18%
27.33%
GAReT [ 19 ]
3.34%
11.19%
17.18%
44.39%
4.10%
14.24%
21.46%
40.48%
L2LTR [ 47 ]
3.78%
12.00%
18.21%
47.59%
2.53%
9.71%
15.74%
23.15%
TABLE IV: Comparison with existing methods. boldfaced texts denote the best results and underlined texts denote the suboptimal results.
Fig. 8: Visualization comparison of attention maps between the baseline method and ASML. (a) shows the satellite images, (b) presents the attention maps generated by the baseline method [ 33 ] , and (c) shows the attention maps produced by the proposed ASML. Compared with the baseline, ASML generates more concentrated responses on geographically discriminative regions while suppressing redundant regions, demonstrating its effectiveness in selecting informative features for cross-view matching.
Fig. 9: Qualitative comparison of ASML and other methods. (a) shows the first images of the query ground image sequences. (b) shows the ground-truth results. (c) shows the Top-3 satellite images returned from ASML. (d) shows the Top-3 satellite images returned from Sample4Geo. (e) shows the Top-3 satellite images returned from GAReT. The first and second rows show the geo-localization results of dense-building areas, the third and fourth rows show the results of sparse-building areas, and the last row shows the results of overpass-intersection areas. The satellite images marked with green boxes denote the ground-truth results, which provide a common reference for all compared methods. The images marked with red boxes represent incorrect retrieval results. The first and fourth rows are samples from the CrossRoad-66K dataset, while the rest are samples from the SeqGeo.
Ablations
SeqGeo
CrossRoad-66K
Fusion
Aggregation
Position Encoding
Loss Function
R@1
R@5
R@10
R@1%
R@1
R@5
R@10
R@1%
Concat
SFA
Compact
Geo-InfoNCE
3.95%
12.46%
19.29%
50.86%
2.24%
7.27%
11.94%
29.35%
Random Mask
SFA
Sparse
Geo-InfoNCE
20.61%
40.33%
49.21%
76.52%
5.01%
14.28%
21.12%
39.42%
SCF
GAP
Compact
Geo-InfoNCE
3.55%
12.18%
18.54%
50.08%
3.11%
10.90%
16.97%
38.89%
SCF
ViT-S
Compact
Geo-InfoNCE
12.24%
29.08%
38.79%
72.30%
3.18%
11.13%
16.81%
37.52%
SCF
SFA
Compact
Soft-Margin Triplet
13.58%
29.80%
38.17%
70.37%
1.69%
5.56%
9.79%
25.40%
TABLE V: Ablation study of each component on SeqGeo and CrossRoad-66K datasets.
Dataset
Selection Strategy
R@1
R@5
R@10
R@1%
SeqGeo
Random Mask
20.61%
40.33%
49.21%
76.52%
High-Similarity SCF
17.07%
35.85%
44.87%
72.77%
SCF
23.53%
44.30%
53.23%
80.77%
CrossRoad-66K
Random Mask
5.01%
14.28%
21.12%
39.42%
High-Similarity SCF
4.55%
14.99%
22.75%
41.41%
SCF
6.20%
16.83%
24.40%
43.94%
TABLE VI: Quantitative comparison of different feature-selection strategies on SeqGeo and CrossRoad-66K.
Fig. 10: Visualization of the attention masks on SeqGeo dataset. (a) shows the ground images. (b) shows the corresponding attention masks.
K
SeqGeo
CrossRoad-66K
R@1
R@5
R@1%
R@1
R@5
R@1%
2
12.45%
30.42%
69.45%
3.41%
10.73%
33.12%
3
15.86%
34.52%
74.07%
5.26%
14.61%
38.47%
4
19.11%
37.55%
76.29%
6.20%
16.83%
43.94%
5
23.53%
44.30%
80.77%
4.21%
12.48%
38.94%
6
20.49%
39.99%
77.62%
3.67%
11.23%
35.95%
TABLE VII: Ablation study on different numbers K of ground images in a sequence.
Methods
Architecture
Parameters
GFLOPs
R@1
L2LTR [ 47 ]
CNN+Transformer
195.9M
44.06
2.53%
TransGeo [ 33 ]
Transformer
44.9M
11.34
2.37%
Sample4Geo [ 37 ]
Transformer
87.6M
20.05
2.35%
SeqGeo [ 20 ]
CNN+Transformer
47.02M
10.30
1.86%
GeoDTR [ 46 ]
CNN+Transformer
48.5M
38.89
3.15%
FlexGeo [ 31 ]
Transformer
87.6M
140.14
3.82%
TABLE VIII: Comparison with existing methods on parameters, computational cost, and R@1 (on CrossRoad-66K dataset).
λg
SeqGeo
CVZY
R@1
R@5
R@1%
R@1
R@5
R@1%
0.25
18.11%
36.39%
73.85%
4.50%
13.61%
40.86
0.5
20.00%
39.60%
77.90%
5.47%
15.81%
44.65%
0.75
23.53%
44.30%
80.77%
6.20%
16.83%
43.94%
1
23.32%
44.04%
80.48%
6.11%
16.67%
43.92%
1.25
20.17%
39.47%
78.39%
5.75%
17.14%
44.27%
TABLE IX: Sensitivity analysis of varying scaling coefficient λg .
Fig. 11: Visualization of the last image in the ground image sequence and the satellite image. The red box in the left ground image represents irrelevant content between the ground image and the satellite image, while the points in the right satellite image represent the capture positions of each image in the ground image sequence.
Dataset
Scenario
Threshold τθ /Numbers
R@1
R@5
R@1%
SeqGeo
Corner
15∘ (264)
27.27%
46.21%
84.09%
20∘ (106)
17.93%
35.85%
84.91%
25∘ (64)
14.06%
31.25%
89.06%
30∘ (41)
17.07%
34.15%
90.24%
Straight
15∘ (7,508)
23.22%
43.33%
80.16%
20∘ (7,666)
23.42%
43.53%
80.22%
TABLE X: Comparison of corner-road and straight-road scenarios under varying trajectory-change thresholds τθ .
Fig. 12: Quantitative analysis of SCF mask sensitivity to controlled perturbations on the SeqGeo and CrossRoad-66K datasets. (a) Non-overlap Precision (NoP) measures the proportion of selected features located in geometrically non-overlapping regions. (b) Overlap False Selection Rate (O-FSR) measures the proportion of overlapping features incorrectly selected. (c) ROI Δ FSR measures the increase in false selection within the perturbed overlapping region relative to the clean input. (d) Aligned Mask IoU evaluates the spatial consistency between the clean and perturbed masks. Higher NoP and Mask IoU, and lower O-FSR and ROI Δ FSR indicate better selection performance and robustness.
Fig. 13: Visualization of module-specific failure cases. (a) shows the first images of the query ground image sequences. (b) shows the Top-3 results from ASML. (c) shows the Top-3 results from the version without SCF. (d) shows the ground-truth. (e) shows the Top-3 results from the version without SFA. (f) shows the Top-3 results from the version without Geo-InfoNCE. The green boxes indicate the ground-truth satellite images, while the red boxes denote incorrect retrieval results.
Cross-view geo-localization (CVGL) aims to estimate geographic locations by matching images captured from different viewpoints, such as drone and satellite views. Existing methods mainly rely on visual representations, but often fail to jointly model fine-grained structural correspondences and semantic priors, making them prone to confusion between visually similar but semantically different regions, and thus limiting robustness under large viewpoint variations. To address these challenges, we propose \textbf{S3Geo}, a structure-semantic synergistic learning framework for cross-view matching. Specifically, we first introduce a Decoupled Query Pooling (DQP) module to extract a compact set of region-aware features from dense tokens, enabling explicit modeling of local structural patterns. We then design a query-level contrastive learning scheme with an optimal transport (OT)-based formulation to establish soft correspondences under cross-view spatial misalignment. Furthermore, we incorporate a Semantic Knowledge Distillation (SKD) strategy from a frozen CLIP teacher to transfer semantic priors and relational structures, thereby improving discrimination on hard negatives. By operating synergistically, the semantic priors provide robust contextual filtering, which guides the structural module to establish precise spatial alignments. Experiments on the University-1652 and SUES-200 datasets demonstrate that \textbf{S3Geo} consistently outperforms state-of-the-art approaches without increasing inference complexity, validating the effectiveness of jointly modeling structural and semantic information for CVGL.
Ziqian Mo, Hill Zhang, Haosheng Tan +2
The Hong Kong University of Science and Technology (Guangzhou) Guangzhou, China · Claremont McKenna College Claremont, California, USA
Cross-view geo-localization is challenging due to drastic viewpoint changes and large appearance discrepancies between street-level and satellite imagery. Although existing methods often use geometric warping to expose co-visible cues, such transformations rely on restrictive spatial assumptions and inevitably introduce severe visual distortions under view-dependent visibility, yielding noisy supervision and fragile correspondences. To overcome this, we propose a novel joint-view consensus-guided learning framework that entirely bypasses explicit geometric warping. Instead of forcing rigid spatial alignment, we dynamically mine and adaptively strengthen a semantic consensus directly within the feature space. Specifically, an auxiliary joint-view pathway during training enables direct cross-view interaction, allowing each view to selectively aggregate corroborative evidence into a unified consensus representation. To resolve feature heterogeneity among the single- and joint-view streams, we introduce global pattern probes acting as a semantic dictionary to project divergent modalities into a strictly aligned metric space. Guided by a consensus-mediated contrastive objective, single-view embeddings are explicitly pulled toward the joint-view anchor during training, distilling this consensus-mining capability into the single-view encoders for robust retrieval at inference. Extensive experiments demonstrate that our method achieves state-of-the-art performance across four standard benchmarks, underscoring the importance of discovering cross-view semantic consensus for reliable geo-localization.
Zhuo Song, Lian Xu, Runqing Jiang +4
Sun Yat-sen University, Shenzhen, China · The University of Western Australia, Perth, Australia
Geometric differences between cross-view images, such as drone and satellite views, significantly increase the challenge of Cross-View Geo-Localization (CVGL), which aims to acquire the geolocation of images by image retrieval. To further enhance the CVGL performance, this paper proposes a parameter-efficient adaptation framework for bridging the geometric gap across images based on the vision foundation model (VFM) (e.g., DINOv3), termed BGG. BGG not only effectively leverages the general visual representations of VFM and captures the robust and consistent features from cross-view images, but also utilizes the generalization capabilities of the VFM, significantly improving the CVGL performance. It mainly contains a Multi-granularity Feature Enhancement Adapter (MFEA) and a Frequency-Aware Structural Aggregation (FASA) module. Specifically, MFEA enhances the scale adaptability and viewpoint robustness of features by multi-level dilated convolutions, effectively bridging the cross-view geometric gap with small training costs. Additionally, considering the [CLS] token lacks spatial details for precise image retrieval and localization, the FASA module modulates patch tokens in the frequency domain and performs adaptive aggregation for local structural feature enhancement. Finally, BGG fuses the enhanced local features with the [CLS] token for more accurate CVGL. Extensive experiments on University-1652 and SUES-200 datasets demonstrate that BGG has significant advantages over other methods and achieves state-of-the-art localization performance with low training costs.
Wei Wang, Dou Quan, Ning Huyan +4
Key Laboratory of Intelligent Perception and Image Understanding of Ministry of Education of China, Xidian University · Department of Automation, Tsinghua University, Beijing 100084, China · School of Telecommunications, Xidian University, Xi’an 710071, China