From Horizontal to Rotated: Cross-View Object Geo-Localization with Orientation Awareness

Chenlin Fu; Ao Gong; Yingying Zhu

cs.CVMar 16, 2026

From Horizontal to Rotated: Cross-View Object Geo-Localization with Orientation Awareness

Authors: Chenlin Fu, Ao Gong, Yingying Zhu

Organizations: College of Computer Science and Software Engineering, Shenzhen University, China

Abstract

Cross-View object geo-localization (CVOGL) aims to precisely determine the geographic coordinates of a query object from a ground or drone perspective by referencing a satellite map. Segmentation-based approaches offer high precision but require prohibitively expensive pixel-level annotations, whereas more economical detection-based methods suffer from lower accuracy. This performance disparity in detection is primarily caused by two factors: the poor geometric fit of Horizontal Bounding Boxes (HBoxes) for oriented objects and the degradation in precision due to feature map scaling. Motivated by these, we propose leveraging Rotated Bounding Boxes (RBoxes) as a natural extension of the detection-based paradigm. RBoxes provide a much tighter geometric fit to oriented objects. Building on this, we introduce OSGeo, a novel geo-localization framework, meticulously designed with a multi-scale perception module and an orientation-sensitive head to accurately regress RBoxes. To support this scheme, we also construct and release CVOGL-R, the first dataset with precise RBox annotations for CVOGL. Extensive experiments demonstrate that our OSGeo achieves state-of-the-art performance, consistently matching or even surpassing the accuracy of leading segmentation-based methods but with an annotation cost that is over an order of magnitude lower.

Figures & tables

Figure 1: Illustration of the CVOGL task and a comparison of solution paradigms. (a) The Cross-View Object Geo-Localization (CVOGL) Task: A query object, indicated by a click point (red dot) in a drone or ground-view image, is to be precisely localized on a georeferenced satellite image. (b) Comparison of Localization Granularity: We compare four paradigms. Retrieval-based: Matches to a coarse grid cell. HBox-based (Detection): Localizes with a loose, axis-aligned Horizontal Bounding Box (HBox). RBox-based (Ours): Localizes with a tight, oriented Rotated Bounding Box (RBox). Segmentation-based: Localizes with a pixel-perfect (but costly) mask.

Figure 2: Illustration of the necessity of RBoxes. The star denotes the object center. The red rectangle represents the predicted box, while the green rectangle denotes the ground truth box.

Figure 3: Overall architecture of our proposed OSGeo. The architecture consists of a Pre-processing for input preparation and click point modeling, an Encoder for feature extraction, an MCP for multi-scale feature fusion, and a Geo-Localization Decode Head for prediction.

Figure 4: Visualization of data annotation for cross-view object geo-localization. Click points are indicated by red dots in the query images. Green boxes indicate different annotation formats.

	Method	Drone $\rightarrow$ Satellite				Ground $\rightarrow$ Satellite
		Validation		Test		Validation		Test
		Acc@50	Acc@25	Acc@50	Acc@25	Acc@50	Acc@25	Acc@50	Acc@25
Ret.	L2LTR [ 27 ]	2.00%	24.32%	1.32%	21.84%	0.51%	4.99%	0.33%	4.76%
	TransGeo [ 37 ]	2.01%	21.48%	1.45%	20.12%	0.70%	5.03%	0.40%	6.53%
	SAFA [ 20 ]	2.39%	22.40%	1.51%	21.47%	0.71%	4.89%	0.41%	6.51%
	Sample4Geo [ 2 ]	6.07%	32.72%	5.34%	30.32%	0.76%	5.93%	0.72%	6.75%

Table 1: Performance comparison on the proposed CVOGL-R dataset. OSGeo-R is against retrieval (Ret.), detection (Det.), and segmentation (Seg.) based approaches. OST denotes TROGeo [ 35 ] leveraging segmentation tasks. The

*

indicates baseline models adapted for RBox prediction.

\spadesuit

denotes our model equipped with a more powerful backbone. The best and second-best results are highlighted.

	Method	Drone $\rightarrow$ Satellite				Ground $\rightarrow$ Satellite
		Validation		Test		Validation		Test
		Acc@50	Acc@25	Acc@50	Acc@25	Acc@50	Acc@25	Acc@50	Acc@25
Ret.	L2LTR [ 27 ]	5.96%	38.68%	6.27%	38.95%	1.84%	12.24%	2.16%	10.69%
	TransGeo [ 37 ]	5.42%	34.78%	6.37%	35.05%	3.25%	21.67%	2.88%	21.17%
	SAFA [ 20 ]	6.39%	36.19%	6.58%	37.41%	3.25%	20.59%	3.08%	22.20%
	Sample4Geo [ 2 ]	16.25%	52.87%	18.40%	52.83%	4.33%	24.27%	5.24%	25.80%

Table 2: Performance comparison on the standard CVOGL dataset, which utilizes HBox annotations.

Method	Drone $\rightarrow$ Satellite								Ground $\rightarrow$ Satellite
	Validation				Test				Validation				Test
	mIoU $\uparrow$	mDice $\uparrow$	AAE $\downarrow$	ME $\downarrow$	mIoU $\uparrow$	mDice $\uparrow$	AAE $\downarrow$	ME $\downarrow$	mIoU $\uparrow$	mDice $\uparrow$	AAE $\downarrow$	ME $\downarrow$	mIoU $\uparrow$	mDice $\uparrow$	AAE $\downarrow$	ME $\downarrow$
	(%)	(%)	( $m^{2}$ )	( $m$ )	(%)	(%)	( $m^{2}$ )	( $m$ )	(%)	(%)	( $m^{2}$ )	( $m$ )	(%)	(%)	( $m^{2}$ )	( $m$ )
Sample4Geo [ 2 ]	16.83	25.41	3782.95	62.96	16.62	25.29	3910.25	57.49	5.30	8.56	3782.69	127.92	5.92	9.56	3910.39	128.81
Sample4Geo + SPS [ 2 ]	19.21	25.37	2270.92	62.51	18.67	24.83	2178.15	57.47	5.03	6.94	2403.32	128.58	6.26	8.57	2347.28	130.22
DetGeo [ 22 ]	30.81	39.40	2885.87	80.69	30.50	39.12	2517.70	85.35	25.77	32.65	2510.07	100.89	25.40	32.27	3005.94	103.71

Table 3: Comparison with previous works on the CVOGL-Seg dataset. “+ SPS” means that SAM Prompt Stage (SPS) from TROGeo [ 35 ] is added to obtain the segmentation mask using the box output as a prompt. The best and second-best results are highlighted.

Figure 5: Visual comparison with state-of-the-art methods.

MCP	$\mathcal{L}_{o}$	$\mathcal{L}_{IoU}$	Drone $\to$ Satellite			Ground $\to$ Satellite
MCP	$\mathcal{L}_{o}$	$\mathcal{L}_{IoU}$	Acc@75	Acc@50	Acc@25	Acc@75	Acc@50	Acc@25
$\times$	$\times$	$\times$	10.38%	50.26%	58.99%	8.94%	41.83%	47.17%
$\times$	✓	$\times$	31.86%	67.42%	74.00%	22.51%	44.09%	48.82%
$\times$	$\times$	✓	29.91%	65.16%	72.87%	19.22%	43.17%	47.79%
✓	$\times$	✓	36.07%	69.06%	75.64%	28.67%	47.56%	50.27%
✓	✓	$\times$	42.24%	73.48%	78.42%	29.39%	50.36%	52.42%

Table 4: Ablation study on different components of OSGeo.

\mathcal{L}_{o}

and

\mathcal{L}_{IoU}

represent the losses in Geo-Localization Decode Head.

Method	Shared.	Param.(M) $\downarrow$	Drone $\to$ Satellite			Ground $\to$ Satellite
Method	Shared.	Param.(M) $\downarrow$	Acc@75	Acc@50	Acc@25	Acc@75	Acc@50	Acc@25
DetGeo ∗ [ 22 ]	$\times$	70.38	10.38%	50.26%	58.99%	8.94%	41.83%	47.17%
VAGeo ∗ [ 13 ]	$\times$	70.41	9.66%	55.50%	64.23%	9.66%	43.27%	48.41%
OSGeo-R (ours)	$\times$	56.74	38.13%	67.83%	71.84%	26.98%	49.19%	51.68%
OSGeo-R (ours)	✓	30.63	42.24%	73.48%	78.42%	29.39%	50.36%	52.42%

Table 5: The effect of shared backbone and neck. “Shared.” indicates shared parameters, “Param.” indicates the parameter count.

Backbone	Drone $\to$ Satellite			Ground $\to$ Satellite
Backbone	Acc@75	Acc@50	Acc@25	Acc@75	Acc@50	Acc@25
ResNet-18	33.71%	65.57%	70.61%	24.56%	48.20%	52.42%
ResNet-50	42.24%	73.48%	78.42%	29.39%	50.36%	52.42%
ResNet-101	47.28%	74.31%	78.62%	32.27%	51.39%	53.34%

Table 6: Ablation study of different backbones for OSGeo-R.

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 94 papers

Beyond 2D Matching: A Unified Single-Stage Framework for Geometry-Aware Cross-View Object Geo-Localization
Jun 29, 2026Liyao Wang, Ruipu Wu, Haojun Xu +3Cross-View Geo-Localization Camera Pose Estimation
Beihang University, Beijing, China · Meituan, Beijing, China
MOGeo: Beyond One-to-One Cross-View Object Geo-localization
Mar 14, 2026Bo Lv, Qingwang Zhang, Le Wu +2Cross-View Geo-Localization
College of Computer Science and Software Engineering, Shenzhen University
Multi-View Mixture-of-Experts with Vision-Language Reranking for Cross-View Object Geo-Localization
Sep 16, 2026Xuyu Fan, Qi Ming, Zhu Han +6Multi-View Contrastive Learning Cross-View Geo-Localization
College of Computer Science, Beijing University of Technology · Zhengzhou University · Zhejiang University +3

Jun 29, 2026J/K move · Enter open · S save

Beyond 2D Matching: A Unified Single-Stage Framework for Geometry-Aware Cross-View Object Geo-Localization

Liyao Wang, Ruipu Wu, Haojun Xu, Lei Shi+2

Beihang University, Beijing, China · Meituan, Beijing, China

Cross-View Geo-Localization Camera Pose Estimation

Open paper AskPDF arXiv

Cross-view object geo-localization (CVOGL) aims to locate a target object from a query view (e.g., ground or drone) within a geo-tagged reference image (e.g., satellite). Existing approaches heavily rely on 2D appearance matching and are constrained by limited datasets lacking geometric metadata, diverse prompts, and standard field-of-view imagery. To address these intertwined challenges, we first introduce \dataset, a large-scale, high-fidelity building dataset comprising over 220,000 ground-satellite and drone-satellite pairs. It provides multi-modal prompts (points, boxes, masks) and camera poses to enable flexible target referring and explicit spatial modeling. Furthermore, we propose a novel single-stage Geometry-Aware Geo-localization framework (GAGeo), built upon the permutation-equivariant 3D foundation model

π^3

. By seamlessly integrating visual features, referring prompts, and learnable task tokens, our model adapts the inherited 3D prior to jointly predict bounding boxes, segmentation masks, and camera poses in a single forward pass. Additionally, we introduce a contrastive loss that utilizes the satellite view as a universal anchor, implicitly aligning ground and drone representations to enable zero-shot ground-to-drone localization without requiring triplet training data. Extensive experiments demonstrate that our approach significantly outperforms state-of-the-art methods, exhibiting exceptional generalization ability in unseen scenes and novel cross-view setups.

Abstract

Cross-View Geo-Localization

Explore similar work

Beyond 2D Matching: A Unified Single-Stage Framework for Geometry-Aware Cross-View Object Geo-Localization

MOGeo: Beyond One-to-One Cross-View Object Geo-localization

Multi-View Mixture-of-Experts with Vision-Language Reranking for Cross-View Object Geo-Localization