Organizations: University College London, London, UK · Karlsruhe Institute of Technology, Karlsruhe, Germany · University of Alberta, Edmonton, Canada · Shenzhen University, Shenzhen, China · Hunan University, Changsha, China
Cross-view Video Geo-localization (CVG) aims to localize ground-view videos by retrieving their corresponding geo-tagged aerial images. However, CVG approaches rely on fixed-length inputs and post-hoc refinement, hindering online-oriented localization under partial or dynamic observations. In this work, we formulate Progressive Cross-view Video Geo-localization (PCVG) as a deployment-oriented extension and evaluation protocol of CVG, enabling localization under varying temporal budgets, prefix-based inference, random-start evaluation, and long-range localization with interruptions. To explore PCVG, we introduce X2Localizer, a cross-grained alignment framework that jointly supervises global prefix-to-aerial retrieval and token-aggregated frame--aerial-tile matching with a budget-dependent asymmetric objective. Furthermore, we introduce a Sliding-Window Re-Localization (SWRL) strategy that dynamically refreshes candidate regions for failure recovery and long-range deployment without full-sequence reprocessing. Extensive experiments show that X2Localizer preserves conventional full-video performance, with marginal gains of +0.1 Recall@1 and +0.3 Recall@10, while substantially improving early localization. In the challenging single-frame setting, X2Localizer improves coarse retrieval by +4.7 Recall@1 and +11.5 Recall@10 over the previous state-of-the-art method. With SWRL, our approach further enables robust progressive localization under random-start and long-distance scenarios, narrowing the gap between benchmark evaluation and real-world deployment.
Figures & tables
Figure 1: Compared to traditional cross-view video geo-localization (CVG), our proposed Progressive CVG (PCVG) enables precise localization (1) at arbitrary timestamps, (2) for varying video lengths, and (3) remains robust to interruptions or missing frames.
Figure 2: (a) GeoAdapter adapts frozen pretrained dual-tower encoders from image-tile matching to video-aerial matching. (b) For each temporal budget, our objective combines global video-to-aerial alignment, frame-tile alignment, and ranking distillation; shorter prefixes receive stronger supervision, while longer prefixes emphasize global alignment.
Figure 3: (a) Sliding-window Re-localization (SWRL) refreshes the candidate region set Acand every Δ frames to support subsequent incremental refinement. (b) It improves robustness in long-range localization and enables failure recovery.
Model
Backbone
#params
Latency
R@1
R@5
R@10
R@1%
Full 40s video-to-global aerial image τ=8
TimeSFormer [ 2 ]
ViT-B
243M
14.2
20.1
44.5
55.6
83.5
VideoSWIN [ 14 ]
Swin-B
175M
14.3
20.4
45.9
59.9
88.0
CVLNet [ 26 ]
VGG16
17M
23.9
0.4
1.3
2.7
15.4
GAMa [ 30 ]
Mixed
23M
32.7
12.2
-
35.3
49.3
DeiT [ 29 ]
DeiT-S
45M
2.9
20.5
49.9
63.6
83.2
Table 1: Coarse retrieval performance under varying prefix budgets τ . Matching is evaluated between V(τ) and Aglobal , reported in Recall@ k (%) and inference latency (ms/video). Recall@1% denotes top 31 of the full gallery. The best results are highlighted in bold .
Model
Backbone
#params
Latency
R@1
R@5
R@10
R@1%
Full 40s video-to-global aerial image τ=8
Shi et al Shi et al. (2020)
VGG16
18M
2.0
9.6
18.1
26.6
71.9
L2LTR Yang et al. (2021a)
ViT-B
196M
12.7
11.7
20.8
28.2
87.1
GAMa Vyas et al. (2022)
Mixed
49M
4.3
15.2
27.2
33.8
91.9
GAMa ⋆ Vyas et al. (2022)
Mixed
72M
41.3
18.3
27.6
32.7
-
GAReT Pillai et al. (2024)
DeiT-S
45M
3.6
46.8
71.8
81.1
91.4
Table 2: Frame-level localization performance with different prefix budgets. Retrieval is performed for all frames after coarse retrieval, reported in Recall@ k (%) and inference latency (ms/frame). Following GAReT Pillai et al. (2024) , Recall@1% at this stage uses the full tile gallery and equals candidate-gallery Recall@All. Bold denotes the best results under the same inference strategy, and results in red indicate performance improvements brought by SWRL.
Method
R@1/5/10/1%
Prefix Coarse Retrieval
GAReT
20.0/46.9/58.9/74.4
X 2 Localizer
24.5 / 55.0 / 66.9 / 82.3
Incremental Fine-Grained Retrieval
GAReT
35.6/51.4/56.2/61.2
X 2 Localizer
39.0 / 57.5 / 63.1 / 69.3
Table 3: Random-start recovery. The system is initialized from a random frame or clip, and we evaluate both coarse and subsequent incremental frame-level retrieval.
Figure 4: Long-range localization on the challenging subset. We compare our proposed method against GAReT, both with and without SWRL, which supports short warm-start and periodic re-localization via SWRL. Dashed lines indicate warm-up and backward refinement performance, whereas solid lines denote incremental frame-level refinement after activation.
Figure 5: Qualitative visualization of frame-level localization cues. We show representative key frames, the corresponding aerial reference image, and the heatmaps produced by GAReT and X 2 Localizer. X 2 Localizer yields sharper and more temporally consistent responses across key frames, suggesting that asymmetric cross-grained alignment helps aggregate reliable local evidence for progressive localization.
Variant
τ=1
τ=2
τ=4
τ=8
Avg.
w/o X obj.
16.9
25.9
41.4
50.2
33.6
Lalign=Lg(τ) only
20.3
28.6
43.2
50.4
35.6
Lalign=Lf(τ) only
21.3
28.5
41.6
50.7
35.5
λf=λg (symmetric)
21.3
28.4
42.3
47.2
34.8
w/o Distill Lself
20.9
28.7
42.4
49.8
35.5
w/o Distill Lteacher
21.6
28.6
42.5
49.3
35.5
Table 4: Ablation studies of X 2 Localizer. (a) Component analysis of the asymmetric cross-grained alignment objective under different temporal budgets in Recall@1. (b) Inference-time analysis of the SWRL strategy under the single-frame setting ( τ=1 ). Gate@ Kc denotes whether the ground-truth global aerial region is included in the top- Kc coarse candidates.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Architecture of GeoAdapter.
Prefix τ
γτ
λg(τ)
λf(τ)
1
0.05
1.0
2.0
2
0.10
1.0
1.0
4
0.25
1.0
0.5
8
2.00
1.0
0.0
Appendix
Table 5: Weights used for progressive asymmetric training.
Figure 7: Examples from the constructed long-range validation subset.
Figure 8: Long-range localization on the challenging subset. We compare X 2 Localizer and GAReT under different SWRL re-localization budgets (1 frame, approximately 5 seconds, and approximately 20 seconds), together with the no-SWRL baseline. Vertical drops correspond to candidate-gallery refresh events. Longer windows generally improve re-localization reliability, while X 2 Localizer consistently achieves stronger recovery under short observation budgets.
Group
Setting
τ=1
τ=2
τ=4
τ=8
Avg.
λf for τ={1,2,4}
(1,0.5,0.25)
21.3
28.2
42.6
49.5
35.4
(1,1,1)
21.0
28.8
42.7
50.3
35.7
(4,2,1)
21.4
29.3
41.9
49.3
35.5
{\color[rgb]{1,0,0}(2,1,0.5)}
21.6
29.1
42.5
50.3
35.9
τf for s⋆
0.005
21.4
29.2
42.5
50.2
35.8
{\color[rgb]{1,0,0}0.01}
21.6
29.1
42.5
50.3
35.9
Appendix
Table 6: Sensitivity to asymmetric training weights and temperature parameters. We report Recall@1 (%) for each temporal budget, and Avg. denotes the mean across budgets. Red indicates the default setting used by X 2 Localizer.
Group
Setting
τ=1
τ=2
τ=4
τ=8
Avg.
ηself for Lself
0.1
20.9
28.3
42.3
50.1
35.4
0.2
21.6
29.1
42.5
50.3
35.9
0.4
21.4
28.9
42.1
49.9
35.6
ηteacher for Lteacher
0.5
20.3
28.1
41.9
49.8
35.0
1.0
21.6
29.1
42.5
50.3
35.9
1.5
20.9
28.8
42.0
50.2
35.5
Appendix
Table 7: Sensitivity to ranking distillation hyperparameters. We vary the weights of prefix-to-full self-distillation and early-teacher distillation, as well as the ranking-distillation temperature. Results are Recall@1 (%) under coarse retrieval. Red indicates the default setting.
Method
Sim.
R@1
R@5
R@10
τ=8
X 2 Localizer
sglobal(τ)
49.4
82.7
90.2
smix(τ)
50.3
83.9
91.0
τ=4
X 2 Localizer
sglobal(τ)
41.1
76.6
85.3
smix(τ)
42.5
78.2
86.7
Appendix
Table 8: Effect of inference similarity for coarse retrieval. The mixed similarity smix(τ) combines global prefix-to-aerial similarity with token-aggregated fine-grained similarity, and consistently improves retrieval across temporal budgets.
Variant
Sim.
K
Gate@K
R@1
R@5
R@10
τ=4
w/o SWRL
smix(τ)
5
78.2
48.9
70.3
76.8
SWRL
smix(τ)
5
78.2
48.7
67.3
71.7
w/o SWRL
smix(τ)
10
86.7
46.3
70.3
78.8
SWRL
smix(τ)
10
86.7
47.4
70.0
77.3
τ=2
Appendix
Table 9: Effect of SWRL candidate refresh in the recovery setting. K denotes the number of coarse candidate places used for frame-level reranking, and Gate@K measures whether the correct place is included in the candidate set. Frame-level recall is computed with the 80-meter GPS threshold.
Figure 9: Additional qualitative heatmap comparisons between GAReT and X 2 Localizer. Each example includes eight sampled video keyframes, the reference high-resolution aerial image, frame-level heatmaps, and the mean aggregated heatmap. X 2 Localizer produces more coherent responses across keyframes, indicating improved temporal consistency of local evidence.
Figure 10: More qualitative examples of frame-level localization heatmaps. Compared with GAReT, X 2 Localizer tends to generate sharper and more spatially stable activations around road-like regions in the aerial image, supporting the effectiveness of token-aggregated frame–aerial-tile alignment.
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
Cross-view geo-localization (CVGL) estimates a camera's location by matching a street-view image to geo-referenced overhead imagery, enabling GPS-denied localization and navigation. Existing methods almost universally formulate CVGL as an image-retrieval problem in a contrastively trained embedding space. This ties performance to large batches and hard negative mining, and it ignores both the geometric structure of maps and the coverage mismatch between street-view and overhead imagery. In particular, salient landmarks visible from the street view can fall outside a fixed satellite crop, making retrieval targets ambiguous and limiting explicit spatial inference over the map. We propose Just Zoom In, an alternative formulation that performs CVGL via autoregressive zooming over a city-scale overhead map. Starting from a coarse satellite view, the model takes a short sequence of zoom-in decisions to select a terminal satellite cell at a target resolution, without contrastive losses or hard negative mining. We further introduce a realistic benchmark with crowd-sourced street views and high-resolution satellite imagery that reflects real capture conditions. On this benchmark, Just Zoom In achieves state-of-the-art performance, improving Recall@1 within 50 m by 5.5% and Recall@1 within 100 m by 9.6% over the strongest contrastive-retrieval baseline. These results demonstrate the effectiveness of sequential coarse-to-fine spatial reasoning for cross-view geo-localization.
Yunus Talha Erzurumlu, Jiyong Kwag, Alper Yilmaz
Photogrammetric Computer Vision Lab, The Ohio State University
Cross-View Geo-Localization (CVGL) in remote sensing aims to locate a drone-view query by matching it to geo-tagged satellite images. Although supervised methods have achieved strong results on close-set benchmarks, they often fail to generalize to unconstrained, real-world scenarios due to severe viewpoint differences and dataset bias. To overcome these limitations, we present VFM-Loc, a training-free CVGL framework that leverages the generalizable visual representations from vision foundational models (VFMs). VFM-Loc identifies and matches discriminative visual clues across different viewpoints through a progressive alignment strategy. First, we design a hierarchical clue extraction mechanism using Generalized Mean pooling and Scale-Weighted R-MAC to preserve distinctive visual clues across scales while maintaining hierarchical confidence. Second, we introduce a statistical manifold alignment pipeline based on domain-wise PCA and Orthogonal Procrustes analysis, linearly aligning heterogeneous feature distributions in a shared metric space. Experiments demonstrate that VFM-Loc exhibits high accuracy on standard benchmarks and surpasses supervised methods by over 20% in Recall@1 on the challenging LO-UCV dataset with large oblique angles. This work highlights that principled alignment of pre-trained features can effectively bridge the cross-view gap, establishing a robust and training-free paradigm for real-world CVGL. The relevant code is made available at: github.com/DingLei14/VFM-Loc.
Jun Lu, Zehao Sang, Haoqi Wei +5
Information Engineering University, Zhengzhou, China · Chinese Academy of Sciences, Beijing, China