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

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

cs.CVSep 4, 2026

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

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

Organizations: Dept. of Artificial Intelligence Semiconductor Engineering, Hanyang University, Republic of Korea · Dept. of Artificial Intelligence, Hanyang University, Republic of Korea · Dept. of Electronic Engineering, Hanyang University, Republic of Korea

Abstract

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.

Figures & tables

Figure 1: (Left) Localization procedure of ARC-Loc , (Right) comparison of CVL approaches . Given the matched descriptors across the ground image and the aerial satellite image, each ground keypoint’s azimuth is mapped into a line constraint on the satellite map by drawing an azimuthal ray from the corresponding satellite keypoint along the ground-view azimuth. The user device must lie on this ray, and the location is obtained by the intersection point of these rays. Based on this mechanism, ARC-Loc does not require computationally heavy BEV transformations or external depth foundation models, achieving faster inference times and competitive localization accuracy.

Figure 2: Overview of ARC-Loc. The pipeline comprises two main stages: feature matching and localization. Initially, we extract dense features using a pretrained visual feature encoder and pass them through keypoint and descriptor heads. Then, by leveraging a matching matrix based on descriptor similarity and keypoint confidence, ARC-Loc establishes direct ground-aerial correspondences and transforms the ground matches into azimuthal rays. ARC solver then estimates the camera position by finding the intersection of these rays. During inference, as our solver requires a minimal set of just two correspondences, RANSAC can be equipped, enabling robust estimation.

Figure 3: ARC loss visualization. Our ARC loss aims to supervise the ARC-Loc framework in an end-to-end manner. (a) The ray-distance loss enforces the azimuthal lines to converge at the predicted location without requiring explicit matching labels, while (b) the position loss directly minimizes the discrepancy between the predicted and ground-truth camera positions. By backpropagating these geometric errors, the network implicitly learns feature matching by penalizing geometrically inconsistent outliers.

Type	Models	Same-Area		Cross-Area
		$\downarrow$ Localization (m)		$\downarrow$ Localization (m)
		Mean	Median	Mean	Median
BEV	GGCVT	4.12	1.34	5.16	1.40
	DenseFlow	3.03	0.97	5.01	2.42
	G2SWeakly	4.19	1.68	4.70	1.68
	HC-Net	2.65	1.17	3.35	1.59

Table 1: Localization results on the VIGOR dataset under the known orientation setting. The best and second-best results within each category (BEV and Non-BEV) are highlighted in bold and underlined , respectively.

Type	Models	Loc. (m) $\downarrow$		Lateral (%) $\uparrow$		Long. (%) $\uparrow$		Ori. ( ∘ ) $\downarrow$		Ori. $\uparrow$
Type	Models	Mean	Median	R@1m	R@5m	R@1m	R@5m	Mean	Median	R@ $1^{\circ}$	R@ $5^{\circ}$
Same-Area
BEV	GGCVT	-	-	76.44	98.89	23.54	62.18	-	-	99.10	100.0
	DenseFlow	1.48	0.47	95.47	99.79	87.89	94.78	0.49	0.30	89.40	99.31
	HC-Net	0.80	0.50	99.01	99.73	92.20	99.25	0.45	0.33	91.35	99.84
	FG 2	0.75	0.51	95.81	99.66	92.50	99.05	0.93	0.66	67.27	98.91

Table 2: Quantitative comparison on the KITTI dataset under orientation noise (

\pm 10^{\circ}

). The best and second-best results within each category (BEV and Non-BEV) are highlighted in bold and underlined , respectively. * denotes the use of the orientation estimator from G2SWeakly [ 18 ] .

Method	BEV	Depth Module	RANSAC	Matching $\downarrow$ Latency (ms)	Pose Est. $\downarrow$ Latency (ms)	Memory (MB) $\downarrow$
FG 2	✓	✗	✗	187.61	2.03	885
			Original	185.94	3630	910
			Parallel (ours)	187.34	39.67	1281
BevSplat	✓	✓	–	139.90 (end-to-end)		2577
Loc 2	✗	✓	✗	137.98	1.44	2131
			Original	136.13	3984	2168

Table 3: Inference latency and memory comparison with ViT-based models. Matching Latency and Pose Est. Latency denote the feature matching and solver/RANSAC stages, respectively. Original denotes each method’s official sequential RANSAC implementation, while Parallel (ours) denotes our GPU-parallel RANSAC implementation.

Figure 6: Visualization of feature matching and azimuthal rays. (a), (c) denote cross-view feature matching, and (b), (d) denote azimuthal rays for corresponding matches.

Table 9

Cross-View Geo-Localization

Cross-view geo-localization matches ground, drone, or aerial images to maps or satellite views to estimate where they were captured.

See all 95 papers

More Than Where You Are: Learning Semantics, Structure, and Geometry from Cross-View Localization
Jul 14, 2026Mao Chen, Xiangkai Zhang, Zhiyong Liu +2Multi-View Consistency Multi-View Geometry
State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, and also with the School of Artificial Intelligence, University of Chinese Academy of Sciences · Beijing Aerospace Control Center
Seeing Across Skies and Streets: Feedforward 3D Reconstruction from Satellite, Drone, and Ground Images
May 8, 2026Qiwei Wang, Zhongyao Tuo, Xianghui Ze +1Relative Pose Estimation Cross-View Geo-Localization
ShanghaiTech University Shanghai, China · Nanjing University of Science and Technology Nanjing, China
Just Zoom In: Cross-View Geo-Localization via Autoregressive Zooming
Mar 26, 2026Yunus Talha Erzurumlu, Jiyong Kwag, Alper YilmazCross-View Geo-Localization Map-Based Localization
Photogrammetric Computer Vision Lab, The Ohio State University

Jul 14, 2026J/K move · Enter open · S save

More Than Where You Are: Learning Semantics, Structure, and Geometry from Cross-View Localization

Mao Chen, Xiangkai Zhang, Zhiyong Liu, Chuankai Liu+1

State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, and also with the School of Artificial Intelligence, University of Chinese Academy of Sciences · Beijing Aerospace Control Center

Multi-View Consistency Multi-View Geometry Cross-View Geo-Localization Multi-View Learning

Open paper AskPDF arXiv

Consistent cross-view understanding under extreme viewpoint changes is essential for spatial intelligence, as it enables models to recognize the same scene across extreme viewpoint gaps. Cross-view localization naturally provides a promising pathway toward this ability, as it requires a model to align ground-view imagery with geo-referenced satellite-view imagery despite drastic appearance changes to estimate camera poses. Recent visual foundation models have made this long-standing localization problem increasingly feasible by providing rich 2D representations for cross-view matching. However, we argue that cross-view localization should not be viewed merely as 2D matching or pose estimation. In this work, we revisit cross-view localization as more than pose estimation and investigate how it can help the model develop consistent cross-view understanding under extreme viewpoint changes, including stable semantics, reliable structure, and transferable geometry. We identify three key limitations of existing methods that prevent them from achieving this. They usually lack explicit 3D grounding, rely on strict point-wise matching that can weaken semantic consistency, and learn from an absolute objective that provides limited guidance for geometric reasoning. To address these limitations, we propose CROSS, a unified cross-view localization framework built upon 3D-grounded alignment, structure-aware matching, and hypothesis ranking. This formulation makes structure learning an intrinsic requirement, encourages semantic representations to remain stable, and enables the model to acquire transferable geometry. Extensive experiments on the KITTI and VIGOR datasets show that CROSS achieves state-of-the-art performance in cross-view localization. More importantly, CROSS effectively learns stable semantics, reliable structure, and transferable geometry across extremely different viewpoints.

Abstract

Cross-View Geo-Localization

Explore similar work

More Than Where You Are: Learning Semantics, Structure, and Geometry from Cross-View Localization

Seeing Across Skies and Streets: Feedforward 3D Reconstruction from Satellite, Drone, and Ground Images

Just Zoom In: Cross-View Geo-Localization via Autoregressive Zooming