ARC-Loc: Leveraging Azimuthal Ray Convergence as a Geometric Cue for Direct Cross-View Localization

Hyeongsik Kim; Mincheol Kim; Heejoon Moon; Je Hyeong Hong

Source-linked AI summary

ARC-Loc: Leveraging Azimuthal Ray Convergence as a Geometric Cue for Direct Cross-View Localization

Hyeongsik Kim, Mincheol Kim, Heejoon Moon, Je Hyeong Hong

arXiv:2609.04965v1Sep 4, 2026cs.CV

TL;DR

Cross-view localization must estimate ground-camera position despite severe viewpoint differences and the limitations of BEV, 3D lifting, and external depth models. ARC-Loc uses direct 2D correspondences converted into azimuthal rays, with a minimal ARC solver and ARC loss. On VIGOR and KITTI, it achieves competitive accuracy while reducing reliance on costly 3D processing.

Problem
Cross-view localization must bridge severe ground-to-aerial viewpoint differences, while existing BEV and depth-based methods face geometric distortions, computational costs, or sensitivity to noisy depth predictions.
Method
ARC-Loc converts direct ground-to-satellite correspondences into azimuthal ray-to-point constraints, estimates position with a closed-form two-point ARC solver, and trains matching with ARC loss.
Results
ARC-Loc achieves competitive localization accuracy on VIGOR and KITTI while reducing inference latency and memory consumption compared with frameworks using 3D lifting or BEV projections.
Takeaways & Limitations
Operating directly in the 2D image domain provides an interpretable and practical alternative to BEV transformations and external depth foundation models.
Takeaways & Limitations
Position accuracy degrades with larger heading uncertainty, and restricted fields of view can cause insufficient ray-direction diversity and geometrically degenerate intersections.

Abstract

from arXiv · show

Cross-view localization (CVL) estimates the pose of a ground image by matching it to a geo-referenced satellite image. To bridge the extreme viewpoint gap, mainstream pipelines rely on Bird's-Eye-View (BEV) transformations or 2D-to-3D lifting. However, deriving 3D structures from a single ground image is fundamentally ill-posed, causing these methods to endure geometric distortions and computational costs during 3D lifting or BEV projection. Furthermore, relying on external depth foundation models to resolve this introduces latency and remains susceptible to noisy predictions. In this work, we present a different approach inspired by a human navigation technique called resection, that can perform direct ground to satellite image matching and localization without relying on external depth foundation models. The key insights of our method are that (i) ground keypoints can be translated into azimuthal rays on the satellite map, and (ii) these rays ideally converge at the user location. Exploiting this geometric constraint through direct line-to-point correspondences, we introduce a minimal Azimuthal Ray Convergence (ARC) solver to identify the intersection, alongside an ARC loss to optimize the matching network. By eliminating dependencies on computationally heavy BEV transformations and external depth foundation models, our approach achieves faster, memory-efficient inference, while its explicit feature matching ensures straightforward compatibility with existing frameworks. Experiments on VIGOR and KITTI demonstrate that ARC-Loc maintains competitive localization accuracy compared to recent approaches, highlighting its practicality.

1 Introduction

Cross-view localization must bridge severe ground-to-aerial viewpoint differences while avoiding the distortions, computational cost, and depth dependence of existing approaches. ARC-Loc reframes localization as azimuthal-ray convergence from direct 2D correspondences, enabling efficient, interpretable position estimation.

Motivation: Cross-view localization remains difficult because ground and satellite images have drastically different viewpoints, while precise global positioning is challenging in GNSS-denied environments.The task is motivated by autonomous driving and urban settings affected by GNSS denial and multipath interference.
Limitations of prior approaches: BEV and 3D-lifting methods can introduce geometric distortions, computational cost, and sensitivity to external depth inaccuracies.Direct matching avoids BEV warping but prior methods still depend on external depth priors.
ARC-Loc: ARC-Loc uses known or estimated heading to map direct ground-to-satellite correspondences into azimuthal rays whose convergence identifies the camera position.The method is inspired by resection, where bearings to landmarks define rays on a 2D map.
ARC-Loc: The ARC solver estimates camera position from a minimal set of two correspondences, while the ARC loss trains feature matching through ray convergence without dense correspondence labels or depth priors.The solver is differentiable and supports robust estimation with RANSAC.
Results: Experiments on VIGOR and KITTI show competitive localization accuracy while eliminating BEV transformations and external depth foundation models.The reported benefits include faster inference and practical compatibility with existing frameworks.

2 Related work

Prior cross-view localization methods use global descriptors, BEV representations, or direct local matching to bridge ground-to-aerial viewpoint differences. These strategies trade off geometric precision, distortion, computational cost, and dependence on depth estimation.

Global descriptor-based approaches: Global descriptor methods compare compact representations but sacrifice fine-grained geometric details needed for precise localization.Later variants add grid-structured or cyclic matching mechanisms to improve spatial precision.
BEV-based approaches: BEV-based approaches transform perspective ground images into an orthographic representation, but inverse perspective mapping relies on a flat-world assumption that distorts or loses vertical structures.Buildings and trees can be important landmarks in urban environments.
Direct local feature matching-based approach: ARC-Loc instead establishes direct ground-aerial correspondences and estimates position from azimuthal-ray intersections without BEV transformation or external depth models.Its localization stage uses the intersection of rays derived from matched features.
Direct local feature matching-based approach: Loc2 performs direct local feature matching and uses RANSAC-compatible explicit correspondences, but lifts matched ground keypoints into 3D with an external depth foundation model.This reduces BEV-warping burden but introduces computational overhead and sensitivity to noisy depth predictions.

3 Proposed method

ARC-Loc directly matches ground and satellite features, converts matched ground points into azimuthal rays, and estimates camera position by weighted ray convergence. Its differentiable ARC solver supports robust inference and end-to-end training through geometric losses.

Framework overview: The framework assumes an approximately known ground heading and primarily estimates the camera’s precise 2D position.Heading can come from consumer-grade IMUs or orientation prediction networks.
Feature extraction: Dense ground and satellite features are processed by separate descriptor and keypoint heads to produce matching representations and confidence maps.The heads use convolutional layers and self-attention, outputting descriptor maps and keypoint confidence maps for both views.
Direct feature matching: Direct matching combines descriptor similarity with keypoint confidence, then selects the top-N ground–satellite correspondences with weights w_i.A dual-softmax matcher with a learnable dustbin handles unmatchable features before correspondence selection.
Azimuthal ray generation: Each matched satellite point defines an azimuthal ray whose direction comes from the corresponding ground keypoint and known orientation, making the camera position the rays’ ideal intersection.The ray is represented as a line through satellite keypoint x_i with direction r(ϕ_i).
ARC solver: The ARC solver estimates position by minimizing weighted squared perpendicular distances from the candidate point to all rays and solving the resulting weighted least-squares problem in closed form.The closed-form solution can use two minimal correspondences, while RANSAC samples hypotheses and selects a robust initial position before refining with inliers.
Loss function: The ARC loss combines direct position supervision with ray-distance supervision to train matching through geometrically consistent predictions.Its two terms are weighted by a hyperparameter α and penalize both camera-position error and ray inconsistency.

4 Experiments

Experiments on VIGOR and KITTI evaluate ARC-Loc across orientation conditions, localization accuracy, efficiency, and ablations. ARC-Loc delivers competitive accuracy while avoiding BEV and depth-based overhead, with robustness under moderate orientation noise.

Datasets and metrics: VIGOR evaluation covers known and noisy orientations across Same-Area and Cross-Area settings, using position and orientation errors alongside efficiency measures.The protocol also includes KITTI evaluations with recall thresholds and decomposed position errors.
VIGOR results: ARC-Loc reduces mean localization error versus Loc2 by 17.6% on VIGOR Same-Area and 9.3% on Cross-Area without external depth modules.It achieves the lowest position error among Non-BEV methods and the second-best overall mean error on Same-Area, behind FG2.
Orientation robustness: Mean position error remains within 2.65 m for orientation noise up to ±10°, then increases at larger noise bounds such as ±45°.Under practical ±10° noisy orientation, ARC-Loc has higher mean position error than FG2.
KITTI results: On KITTI, ARC-Loc is competitive with leading BEV methods on Same-Area and outperforms existing Non-BEV baselines, but has higher Cross-Area position error than Loc2.The authors attribute the Cross-Area difficulty to extracting precise azimuthal cues under restricted field of view and unseen layouts.
Efficiency: Direct native-domain matching gives ARC-Loc lower matching latency and a compact memory footprint than depth-dependent methods while maintaining competitive localization accuracy.The efficiency benefit comes from removing BEV construction and depth-based 3D lifting overhead.
Ablations: The default α = 0.5 and N = 512 configuration improves VIGOR mean error from 2.76 m with Lposition alone to 2.52 m using Lray−dist.N = 256 gives 4.57 m, whereas N = 1024 gives 2.67 m, indicating a trade-off between match diversity and outlier noise.

5 Conclusion

ARC-Loc reformulates cross-view localization as a 2D line-to-point problem, avoiding BEV transformations and external depth models while maintaining competitive accuracy. Its 2D design reduces inference latency and memory consumption, but performance remains sensitive to heading uncertainty and ray diversity.

Method: ARC-Loc estimates camera location directly in the native 2D image domain using azimuthal ray convergence from ground-to-aerial correspondences.The method introduces a closed-form 2-point ARC solver and enforces convergence through ARC Loss without dense correspondence labels or depth priors.
Method: The framework bypasses conventional Bird’s-Eye-View transformations and external depth foundation models.This removes dependencies on BEV projection and depth-based 3D lifting.
Results: ARC-Loc achieves competitive localization accuracy on VIGOR and KITTI compared with state-of-the-art methods.The evaluation covers both datasets named in the conclusion.
Results: Operating entirely in 2D reduces inference latency and memory consumption relative to frameworks using 3D lifting or BEV projections.The authors identify this as practical for resource-constrained autonomous platforms.
Limitations: Position accuracy degrades with larger heading uncertainty, while restricted fields of view can cause geometrically degenerate ray intersections.The framework is reported as robust up to approximately ±10° orientation noise, but narrow keypoint distributions weaken intersection constraints.