1 of 19

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

Ayesh Abu Lehyeh, Xiaohan Zhang, Ahmad Arrabi, Waqas Sultani, Chen Chen, Safwan Wshah

CVPR 2026

2 of 19

Overview

Accurate and fast localization is vital for safe autonomous navigation in GPS-denied areas.

GeoFlow predicts dense cross‑view flow to align ground/drone view with its corresponding location on the map to give fine-grained location in real time.

Drone view

Satellite overhead map

3 of 19

The Core Problem (FG-CVG)

Query ground view

Satellite overhead map

4 of 19

Previous Works (Retrieval CVGL)

Early Cross-View Retrieval: Framed localization as an image retrieval problem like GeoDTR.

Xiaohan Zhang, Xingyu Li, Waqas Sultani, Chen Chen, and Safwan Wshah. Geodtr+: toward generic cross-view geolocalization via geometric disentanglement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

Mechanism: Learned global descriptors to find the best-matching satellite region from a database.

Limitation: Assumes the ground camera is exactly at the center of the satellite map, restricting precision to the coarse level of the entire satellite tile

5 of 19

Previous Works (Traditional Approaches)

Direct Regression (e.g., VIGOR [1]): Simple, but struggles to achieve high fine-grained accuracy.

Optimization-Based (e.g., Shi et al. [2]): Uses Levenberg-Marquardt (LM) optimization to iteratively refine poses via visual correspondence. Requires heavy geometric projections and relies on priors like camera intrinsics. Heavy and slow.

Matching based approaches (e.g., SliceMatch [3]): Discretizes the continuous pose search space into a finite grid of distinct patches, The model's fundamental precision is permanently capped by the physical size of the patches.

[1] Sijie Zhu, Taojiannan Yang, and Chen Chen. Vigor: Cross-view image geo-localization beyond one-to-one retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3640–3649, 701 2021.

[2] Yujiao Shi and Hongdong Li. Beyond cross-view image retrieval: Highly accurate vehicle localization using satellite image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17010– 627 17020, 2022.

[3] Ted Lentsch, Zimin Xia, Holger Caesar, and Julian FP Kooij. Slicematch: Geometry-guided aggregation for cross-view pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17225–17234, 2023.

6 of 19

Previous Works (Modern Approaches)

[1] Zimin Xia, Olaf Booij, and Julian F. P. Kooij. Convolutional cross-view pose estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(5):3813–3831, 2024.

[2] Xiaolong Wang, Runsen Xu, Zhuofan Cui, Zeyu Wan, and Yu Zhang. Fine-grained cross-view geo-localization using a correlation-aware homography estimator. Advances in Neural Information Processing Systems, 36:5301–5319, 2023.

[3] Zimin Xia and Alexandre Alahi. Fgˆ 2: Fine-grained cross-view localization by fine-grained feature matching. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 6362–6372, 2025.

I am considering those as the modern FG-CVG methods. They are recent in the literature, and they are leading in terms of localization accuracy and also they handle the orientation prediction. All of them can be considered as geometric based matching methods. First CCVPE is doing good job in terms of accuracy and handle orientation well, they rely on heavy matching decoders to handle this but still the model is heavy and lack the edge deployment characteristics especially in terms of memory consumption and GFLOPS.

Then, we have HC-Net which is also a geometrical based approach rely on homography estimations to align and match between the two views. It’s doing good especially in the cross area settings, it’s better than CCVPE in terms of speed but still not well suitable for edge deployment.

Then, we have the recent FG2 method which is relying on geomaterial dense feature matching between the two view (trying to build dense pixel correspondences between the two views), while it’s really SOTA in terms on mean accuracy but it’s still a very heavy model compared with HC-Net and CCVPE so it’s not good for efficient edge deployment.

7 of 19

Motivations

The problem: Current methods force a harsh trade-off between accuracy and efficiency. High-accuracy models are too slow for real-time use and lack real world deployment requirements.

Discretizes search space so high quantization error (lower accuracy)

Continuous but computationally heavy (often needing camera intrinsics to work as well).

GeoFlow: Direct probabilistic mapping via continuous flow.

8 of 19

The Edge Deployment Reality ("Why GeoFlow?)

Targeting edge devices (Ex., Sub-250g Micro-UAVs) that need real-time speed.

Strict hardware limits: 2GB to 4GB RAM.

Concurrent tasks (OS, other tasks) consume >2GB.

9 of 19

GeoFlow Architecture

10 of 19

Probabilistic Displacement Regression

11 of 19

Objective Functions

12 of 19

Iterative Refinement Sampling (IRS)

Inference time

We use the learned distance and direction to move from any point in the sat space to toward the ground truth (prediction) in the inference. Instead of trusting a single prediction during inference, IRS refines a population of hypotheses. They "flow" across the map from random starting points to a converged consensus.

Mimics human spatial convergence.

Initializes $N$ candidate poses.

Iteratively refines over $R$ rounds to find a robust consensus.

The IRS Algorithm:

Initialize a population of $N$ candidate poses $\mathcal{Q}_0$ uniformly across the map.

For $R$ rounds, feed every candidate back into the regression field $v^\phi$.

Update all candidates in parallel using the step refinement equation.

+1

Result: The population flows to a robust consensus. The final location is the mean of the converged set

13 of 19

Quantitative Results

Outperforms CCVPE in mean accuracy (8.42m vs 9.16m) on KITTI Cross-Area while achieving real time speed.

14 of 19

Inference-Time Scaling

Heavy lifting (visual context) is done once.

Dynamic trade-off between accuracy and speed.

15 of 19

State-of-the-Art Efficiency

VRAM: 686 MiB (7x reduction vs. CCVPE).

Speed: 26 ms (35% speedup in pure inference).

Compute: 7.65 GFLOPs (reduces thermal throttling and power consumption).

16 of 19

Ablation Studies

17 of 19

Qualitative Results

18 of 19

Qualitative Results (Confidence)

IRS actively sculpts predictive uncertainty

19 of 19

Conclusion & Broader Impact

Summary: GeoFlow successfully bridges the gap between SOTA benchmark accuracy and practical edge deployment constraints.

Innovation: The probabilistic displacement field combined with Iterative Refinement Sampling (IRS) actively models confidence and consensus.

Future Work: Extending the continuous flow principles to full 6-DoF estimation and downstream path-planning tasks for autonomous micro-robotics.