GeoFlow: Real-Time Fine-Grained Cross-View Geolocalization via Iterative Flow Prediction
Authors: Ayesh Abu Lehyeh, Xiaohan Zhang, Ahmad Arrabi, Waqas Sultani, Chen Chen, Safwan Wshah
Organizations: Vermont Artificial Intelligence Lab, Department of Computer Science, University of Vermont · Intelligent Machines Lab, Information Technology University · Institute of Artificial Intelligence, University of Central Florida
Accurate and fast localization is vital for safe autonomous navigation in GPS-denied areas. Fine-Grained Cross-View Geolocalization (FG-CVG) aims to estimate the precise 2-Degree-of-Freedom (2-DoF) location of a ground image relative to a satellite image. However, current methods force a difficult trade-off, with high-accuracy models being slow for real-time use. In this paper, we introduce GeoFlow, a new approach that offers a lightweight and highly efficient framework that breaks this accuracy-speed trade-off. Our technique learns a direct probabilistic mapping, predicting the displacement (in distance and direction) required to correct any given location hypothesis. This is complemented by our novel inference algorithm, Iterative Refinement Sampling (IRS). Instead of trusting a single prediction, IRS refines a population of hypotheses, allowing them to iteratively 'flow' from random starting points to a robust, converged consensus. Even its iterative nature, this approach offers flexible inference-time scaling, allowing a direct trade-off between performance and computation without any re-training. Experiments on the KITTI and VIGOR datasets show that GeoFlow achieves state-of-the-art efficiency, running at real-time speeds of 29 FPS while maintaining competitive localization accuracy. This work opens a new path for the development of practical real-time geolocalization systems.
Figures & tables
Figure 1: Comparison of FG-CVG techniques. (a) Match-based methods search over discrete grids. (b) Regression-based methods directly or iteratively regress the pose. (c) Our method uses learned displacement predictions to iteratively guide multiple diverse hypotheses (in different colors) in continuous space towards the ground truth.
Figure 2: The top panel illustrates the architecture of our proposed GeoFlow. Ground ( Fg ) and satellite ( Fs ) features are extracted by ψg and ψs , then flattened, positionally encoded, and fused by a Cross-Attention block to produce a single visual representation, fvis . Separately, an initial location point q0=(x0,y0) is passed through a fully connected layer (FC). This point embedding is concatenated with fvis to form the joint representation s . Finally, two regression heads ( vϕ ) predict the parameters for the probabilistic displacement: distance (μr,σr) and direction (μθ,κ) . The bottom panel illustrates our Iterative Refinement Sampling (IRS) algorithm. (1) The process starts with N randomly sampled hypotheses (red dots), the green dot represents the ground truth location. (2-4) In each round, our model predicts the displacement (blue arrows) for all hypotheses, which are then updated and iteratively converge toward the target. (5) After R rounds, the final robust location (yellow dot) is computed as the mean of the entire converged population.
Efficiency
Same-area
Cross-area
FPS↑
↓ Loc. (m)
↑ Lateral (%)
↑ Long. (%)
↓ Loc. (m)
↑ Lateral (%)
↑ Long. (%)
Method
Mean
Median
R@1m
R@5m
R@1m
R@5m
Mean
Median
R@1m
R@5m
R@1m
R@5m
GGCVT [ 17 ]
4.17
–
–
76.44
98.89
23.54
62.18
–
–
57.72
91.16
14.15
45.00
CCVPE [ 26 ]
24.00
1.22
0.62
97.35
99.71
77.13
97.16
9.16
3.33
44.06
90.23
23.08
64.31
HC-Net [ 22 ]
25.00
0.80
0.50
99.01
99.73
92.20
99.25
8.47
4.57
75.00
97.76
58.93
76.46
DenseFlow [ 20 ]
7.30
1.48
0.47
95.47
99.79
87.89
94.78
7.97
3.52
54.19
91.74
23.10
61.75
Table 1: KITTI test results comparison. We report localization error (Mean/Median, m), lateral/longitudinal recall (R@1m/R@5m, %), and inference speed (FPS) for both Same-Area and Cross-Area splits. Best results are in bold .
Figure 4
Method
Area
↓ Mean (m)
↓ Med. (m)
↑ FPS
GGCVT [ 17 ]
Same
4.12
1.34
3.57
Cross
5.16
1.40
CCVPE [ 26 ]
Same
3.60
1.36
18.00
Cross
4.97
1.68
HC-Net [ 22 ]
Same
2.65
1.17
20.00
Cross
3.35
1.59
Table 2: Comparison on VIGOR dataset. Localization error (m) and inference speed (FPS) are reported. Best results are in bold .
Rounds ( R )
↓ Mean (m)
↓ Median (m)
↑ Speed (FPS)
1
10.69
9.95
32.55
3
8.47
5.88
31.23
5
8.42
5.60
29.49
10
8.41
5.59
26.23
Table 3: Impact of refinement rounds ( R ) on KITTI Cross-Area.
Seeds ( N )
↓ Mean (m)
↓ Median (m)
↑ Speed (FPS)
1
8.58
5.75
30.70
5
8.49
5.66
30.05
10
8.42
5.60
29.49
20
8.41
5.60
28.08
Table 4: Impact of initial seeds ( N ) on KITTI Cross-Area.
Method
↓ Params(M)
↓ Mem(MiB)
↑ FPS
↓ Mean(m)
CCVPE [ 26 ]
57.40
4730
24.00
9.16
HC-Net [ 22 ]
11.21
1900
25.00
8.47
Ours
7.38
686
29.49
8.42
Table 5: Comparison of computational efficiency.
IRS Config.
↓ Mean (m)
↓ Med (m)
Single-Pass ( N=1,R=1 )
12.47
11.79
Full IRS (Default)
8.42
5.60
Table 6: Efficacy of the proposed IRS algorithm.
Loc. (m)
Orientation ( γ ) Errors ( ∘ )
Inference
Dataset
Method
↓ Mean
↓ Median
↓ Mean
↓ Median
↑ R@1
↑ R@5
Speed (FPS)
Same Area
FG 2 [ 25 ]
0.75
0.52
1.28
0.74
61.17 %
95.65 %
4.20
GeoFlow (3-DoF)
1.03
0.69
2.51
1.93
27.30%
87.80%
29.49
Cross Area
FG 2 [ 25 ]
7.45
4.03
3.33
1.88
30.34 %
81.17 %
4.20
GeoFlow (3-DoF)
8.53
5.68
3.87
2.75
20.00%
72.30%
29.49
Table 7: 3-DoF Pose Estimation Comparison on KITTI. GeoFlow demonstrates competitive accuracy against state-of-the-art methods while maintaining a superior computational profile. Best results are in bold
Efficiency
Same Area (m)
Cross Area (m)
Backbone
↓ Params (M)
↓ Mem (MiB)
↑ Speed (FPS)
↓ Mean
↓ Median
↓ Mean
↓ Median
EffNet-B0 (Default)
7.38
686
29.49
0.98
0.68
8.42
5.60
EffNet-B5 (Large)
54.81
1169
15.70
0.86
0.60
7.93
5.56
Table 8: Ablation Study: Accuracy vs. Computational Overhead. We report Mean and Median localization error (m) for both Same and Cross-Area splits on KITTI.
Model
GFLOPs
VRAM
Inference Time
FPS
CCVPE
31.18
4730 MiB
41.7 ms
24.0
HC-Net
11.56
1900 MiB
40.0 ms
25.0
GeoFlow
7.65
686 MiB
26.0 ms
29.5
Gain (vs HC-Net)
-34%
-64%
-35%
+18%
Table 9: Unified efficiency comparison across representative FG-CVG methods.
Figure 5: Additional results for IRS refinement dynamics: Visualization of pose hypotheses distribution over refinement rounds R∈{1,2,3,4} , demonstrating convergence towards the ground truth and how hypothesis consensus reflects confidence.
Figure 6: Additional Qualitative localization examples by GeoFlow on the Same-Area VIGOR dataset. The figure visualizes the trajectories generated by the learned regression field by GeoFlow: trajectories show how five randomly sampled initial pose hypotheses (seeds) are transformed to their respective predicted locations (colored circular markers), relative to the ground truth (red X).
Figure 7: Additional Qualitative localization examples by GeoFlow on the Cross-Area VIGOR dataset. The figure visualizes the trajectories generated by the learned regression field by GeoFlow: trajectories show how five randomly sampled initial pose hypotheses (seeds) are transformed to their respective predicted locations (colored circular markers), relative to the ground truth (red X).
Figure 8: Failure Cases: Localization difficulties in densely vegetated forest environments. The top row shows the ground images, and the bottom row visualizes the refinement trajectories and final converged hypotheses on the corresponding satellite images.