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

Bo Lv; Qingwang Zhang; Le Wu; Yuanyuan Li; Yingying Zhu

cs.CVMar 14, 2026

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

Authors: Bo Lv, Qingwang Zhang, Le Wu, Yuanyuan Li, Yingying Zhu

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

Abstract

Cross-View Object Geo-Localization (CVOGL) aims to locate an object of interest in a query image within a corresponding satellite image. Existing methods typically assume that the query image contains only a single object, which does not align with the complex, multi-object geo-localization requirements in real-world applications, making them unsuitable for practical scenarios. To bridge the gap between the realistic setting and existing task, we propose a new task, called Cross-View Multi-Object Geo-Localization (CVMOGL). To advance the CVMOGL task, we first construct a benchmark, CMLocation, which includes two datasets: CMLocation-V1 and CMLocation-V2. Furthermore, we propose a novel cross-view multi-object geo-localization method, MOGeo, and benchmark it against existing state-of-the-art methods. Extensive experiments are conducted under various application scenarios to validate the effectiveness of our method. The results demonstrate that cross-view object geo-localization in the more realistic setting remains a challenging problem, encouraging further research in this area.

Figures & tables

Figure 1: Comparison of cross-view object geo-localization in single-object and multi-object scenarios. Click points represent query objects, while bounding boxes in geo-tagged satellite images indicate location information. Points and bounding boxes of the same color form an object pair, such as $p_{1}$ and $b_{1}$ , where $b_{1}$ ’s geographic location is considered the position of $p_{1}$ .

Figure 2: Attention map comparison: previous vs. ours. Previous smooth positional encodings lead to diffuse attention maps, whereas our proposed attention map successfully concentrates on the target location to provide highly discriminative features.

Data name	Split	Query	Reference	Instances
CMLocation-V1	total	12760	12760	31944
	training	8371	8371	20931
	validation	2093	2093	5209
	test	2296	2296	5804
CMLocation-V2	total	12760	12760	31944
	training	8371	8371	20931

Table 1: Size and partitioning information of the CMLocation-V1 and CMLocation-V2 datasets.

Figure 3: Examples from the CMLocation-V1 and CMLocation-V2 datasets. CMLocation-V1 is curated under strict alignment principles—specifically center alignment and northward orientation—which ensure a consistent spatial distribution across all instances. By contrast, CMLocation-V2 does not impose these constraints, resulting in more heterogeneous spatial configurations.

Figure 4: The size distributions of objects in the CMLocation-V1 and CMLocation-V2 datasets.

Figure 5: Overview of our proposed MOGeo model. Our method takes as input a query image that contains an arbitrary number of objects of interest (either ground-view images or drone-view images) and a reference image that contains the objects to be localized (satellite image). The input images are processed by our position encoding module, MOPE, to extract feature information for each query object. Then, query object features are fused with the reference image to obtain fused features $\mathbf{F^{\prime\prime}}_{r}$ , which are fed into the detection head to produce localization results. In the figure, query points and bounding boxes with the same color represent object pairs.

Figure 6: Cross-view multi-object feature fusion module. As shown in the figure, this module sequentially fuses query object features with the reference image to obtain an attention map for each object, and finally combines fused features with the attention map to further enhance localization in the reference image.

Method	Test				Validation
Method	[email protected]	[email protected]	[email protected]	[email protected]	[email protected]	[email protected]	[email protected]	[email protected]
CMLocation-V1
FRGeo [ 31 ]	8.06	1.05	2.74	0.17	7.20	0.92	2.53	0.05
GeoDTR+ [ 34 ]	7.44	0.45	3.83	0.17	8.62	0.77	4.30	0.53
Sample4Geo [ 4 ]	29.19	2.31	16.46	1.44	30.68	2.36	19.10	1.81
DetGeo [ 24 ]	56.03	55.00	39.68	38.32	56.29	55.33	39.80	38.93

Table 2: Performance ( accuracy %) comparison of different approaches on our dataset CMLocation.

Method	Test		Validation
Method	[email protected]	[email protected]	[email protected]	[email protected]
CVOGL-SVI
CVM-Net [ 7 ]	4.73	0.51	5.09	0.87
TransGeo [ 38 ]	21.17	2.88	21.67	3.25
FRGeo [ 31 ]	8.12	1.31	7.80	0.87
GeoDTR [ 35 ]	32.37	6.06	31.53	5.31

Table 3: Performance ( accuracy %) comparison of different approaches on CVOGL dataset.

CMLocation-V1
Method	Validation
	I ( $N\leq 3$ )		II ( $3<N\leq 6$ )		III ( $N>6$ )
	[email protected]	[email protected]	[email protected]	[email protected]	[email protected]	[email protected]
DetGeo	61.03	59.90	49.03	48.23	40.58	40.34
VAGeo	62.97	62.30	47.74	47.54	42.27	42.27
Ours	67.22	66.07	56.74	56.34	53.14	52.90

Table 4: Model performance ( accuracy %) under different numbers of query objects on the validation set.

Figure 7: Relationship between model parameters and inference time, measured on the validation set of the CMLocation-V1.

Method	Test		Validation
Method	[email protected]	[email protected]	[email protected]	[email protected]
	CVOGL-Drone
Ours	66.39	59.71	65.11	58.72
w/o $L_{s}$	65.98	59.30	61.21	54.82
w/o CVMF	65.78	59.92	62.62	58.40
w/o MOPE	54.88	50.05	54.60	48.21

Table 5: Ablation performance ( accuracy %) comparison on CVOGL-Drone and CMLocation-V2 datasets.

Figure 8: Detection result visualization. Green represents the GT boxes, while the red and blue bounding boxes correspond to the predictions of VAGeo and our method, MOGeo, respectively.

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

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
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
Towards Active Cross-View Object Geo-Localization
Sep 17, 2026Shunyu Yao, Xiaohan Zhang, Zhuoran Yang +5Cross-View Geo-Localization Next-Best-View Planning
College of Information Science and Electronic Engineering, Zhejiang University, Hangzhou 310027, China · College of Computer Science, Beijing University of Technology, Beijing 100124, China · State Key Laboratory of Intelligent Green Vehicle and Mobility, Tsinghua University, Beijing 100084, China +2

Sep 16, 2026J/K move · Enter open · S save

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

Xuyu Fan, Qi Ming, Zhu Han, Liuqian Wang+5

College of Computer Science, Beijing University of Technology · Zhengzhou University · Zhejiang University+3

Multi-View Contrastive Learning Cross-View Geo-Localization Multi-View Learning Mixture of Experts

Open paper AskPDF arXiv

Cross-view object geo-localization (CVOGL) locates a target in satellite imagery using drone or street-view queries. Existing methods train separate detectors for each viewpoint, leading to parameter redundancy and impeding cross-view knowledge sharing. Moreover, top-ranked satellite candidates are often visually similar, so visual appearance and categorical labels alone are insufficient to resolve such ambiguity. To address these, we propose MVLGeo, an efficient framework designed to unify multiple viewpoints and reduce model redundancy. First, we introduce environmental contextual text from the query view as cues to distinguish visually similar candidates via Vision-Language Reranking (VL-Rerank). Second, we design a multi-view Mixture-of-Experts architecture (MV-MoE) with a shared encoder and view-specific experts to reduce redundancy and promote knowledge sharing, while cross-view contrastive learning aligns their representations for consistency. Third, we introduce an adaptive elliptical prior (ESAM-Prior) as auxiliary positional encoding for anisotropic geometric perception. Extensive experiments on the CVOGL benchmarks confirm that MVLGeo, as a unified model for multiple query viewpoints, achieves state-of-the-art performance, demonstrating robustness to input degradation and generalization across viewpoints. Code and models will be available on GitHub to facilitate future work.

Abstract

Cross-View Geo-Localization

Explore similar work

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

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

Towards Active Cross-View Object Geo-Localization