GeoFlow: Real-Time Fine-Grained Cross-View Geolocalization via Iterative Flow Prediction

Ayesh Abu Lehyeh ¹

Xiaohan Zhang ¹

Ahmad Arrabi ¹

Waqas Sultani ²

Chen Chen ³

Safwan Wshah ¹

¹ Vermont Artificial Intelligence Lab, Department of Computer Science, University of Vermont

² Intelligent Machines Lab, Information Technology University

³ Institute of Artificial Intelligence, University of Central Florida

CVPR 2026

Paper Code arXiv Slides

Abstract

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.

Method

GeoFlow architecture and Iterative Refinement Sampling (IRS) illustration — The top panel illustrates the architecture of our proposed GeoFlow. Ground ( $\mathbf{F}_g$ ) and satellite ( $\mathbf{F}_s$ ) features are extracted by $\psi_g$ and $\psi_s$ , then flattened, positionally encoded, and fused by a Cross-Attention block to produce a single visual representation, $\mathbf{f}_{vis}$ . Separately, an initial location point $\mathbf{q}_0 = (x_0, y_0)$ is passed through a fully connected layer (FC). This point embedding is concatenated with $\mathbf{f}_{vis}$ to form the joint representation $\mathbf{s}$ . Finally, two regression heads ( $\mathbf{v}^\phi$ ) predict the parameters for the probabilistic displacement: distance $(\mu_r, \sigma_r)$ and direction $(\mu_\theta, \kappa)$ . 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.

Results

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.

Method	Efficiency	Same-area						Cross-area
	FPS $\uparrow$	Loc. (m) $\downarrow$		Lateral (%) $\uparrow$		Long. (%) $\uparrow$		Loc. (m) $\downarrow$		Lateral (%) $\uparrow$		Long. (%) $\uparrow$
	FPS $\uparrow$	Mean	Median	R@1m	R@5m	R@1m	R@5m	Mean	Median	R@1m	R@5m	R@1m	R@5m
GGCVT	4.17	–	–	76.44	98.89	23.54	62.18	–	–	57.72	91.16	14.15	45.00
CCVPE	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	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	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
FG $^2$	4.20	0.75	0.52	99.73	100.00	86.99	98.75	7.45	4.03	89.46	99.80	12.42	55.73
GeoFlow (Ours)	29.49	0.98	0.68	96.85	99.68	74.05	98.75	8.42	5.60	36.36	83.85	14.76	52.51

Ablations (KITTI Cross-Area)

Our inference is controlled by two inference-time knobs (no retraining required):

$N$ : number of initial location hypotheses (random seeds) sampled at the start.
$R$ : number of refinement rounds applied to all hypotheses.

Increasing $N$ improves robustness (more hypotheses), while increasing $R$ improves convergence (more refinement), at the cost of lower FPS.

Rounds ( $R$ )	Mean (m) $\downarrow$	Median (m) $\downarrow$	Speed (FPS) $\uparrow$
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

Impact of initial seeds ( $N$ )

Seeds ( $N$ )	Mean (m) $\downarrow$	Median (m) $\downarrow$	Speed (FPS) $\uparrow$
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

Efficiency

Unified efficiency comparison across representative FG-CVG methods. GeoFlow is designed for practical deployment on resource-constrained and edge devices, prioritizing low compute and memory footprint.

Model	GFLOPs $\downarrow$	VRAM $\downarrow$	Inference time $\downarrow$	FPS $\uparrow$
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%

Citation

@article{abulehyeh2026geoflow,
      title  = {{GeoFlow}: Real-Time Fine-Grained Cross-View Geolocalization via Iterative Flow Prediction},
      author = {Abu Lehyeh, Ayesh and Zhang, Xiaohan and Arrabi, Ahmad and Sultani, Waqas and Chen, Chen and Wshah, Safwan},
      journal={arXiv preprint arXiv:2603.21943},
      year={2026}
}