CVGL: Causal Learning and Geometric Topology

Songsong Ouyang; Yingying Zhu

cs.CVMar 13, 2026

CVGL: Causal Learning and Geometric Topology

Authors: Songsong Ouyang, Yingying Zhu

Organizations: College of Computer Science and Software Engineering Shenzhen University

Abstract

Cross-view geo-localization (CVGL) aims to estimate the geographic location of a street image by matching it with a corresponding aerial image. This is critical for autonomous navigation and mapping in complex real-world scenarios. However, the task remains challenging due to significant viewpoint differences and the influence of confounding factors. To tackle these issues, we propose the Causal Learning and Geometric Topology (CLGT) framework, which integrates two key components: a Causal Feature Extractor (CFE) that mitigates the influence of confounding factors by leveraging causal intervention to encourage the model to focus on stable, task-relevant semantics; and a Geometric Topology Fusion (GT Fusion) module that injects Bird's Eye View (BEV) road topology into street features to alleviate cross-view inconsistencies caused by extreme perspective changes. Additionally, we introduce a Data-Adaptive Pooling (DA Pooling) module to enhance the representation of semantically rich regions. Extensive experiments on CVUSA, CVACT, and their robustness-enhanced variants (CVUSA-C-ALL and CVACT-C-ALL) demonstrate that CLGT achieves state-of-the-art performance, particularly under challenging real-world corruptions. Our codes are available at https://github.com/oyss-szu/CLGT.

Figures & tables

Figure 1: Structural Causal Model (SCM) for cross-view geo-localization. Nodes represent variables and arrows denote dependencies.

Figure 2: Visualization of low, mid, and high frequency components of a street image. Low and high frequencies emphasize domain-specific information such as style, while the mid-frequency band retains domain-invariant cues such as structure and shape.

Figure 3: Overview of the proposed Causal Learning and Geometric Topology (CLGT) framework. The road topology information from BEV is fused via the GT Fusion module to obtain the fused features, which are then used for location matching with aerial image features. The causally enhanced street features from the CFE module provide causal supervision, and the DA Pooling module performs final feature extraction.

Figure 4: Illustration of the Causal Feature Extractor (CFE). The input image is transformed into the frequency domain via Discrete Cosine Transform (DCT). A Content-aware Mask (CaM) strategy dynamically separates mid-frequency causal components from low- and high-frequency non-causal components. A Gaussian function is applied to only to the non-causal components (e.g., lighting and brightness) to reduce their influence. Both causal and non-causal parts are then reconstructed via inverse (IDCT) to obtain the causally enhanced image.

Figure 5: Overview of the GT Fusion and DA Pooling modules. GT Fusion uses cross-attention to exchange semantic information between street and BEV features, then uses Dual Dynamic Fusion (DDF) to enhance fusion robustness. DA Pooling employs a gating mechanism to adaptively weight features, highlighting the most informative ones.

Model	CVACT_val-C-ALL			CVACT_test-C-ALL
Model	R@1	R@5	R@10	R@1	R@5	R@10
Ours	88.68	95.58	96.66	69.06	90.12	92.70
Only CL	88.11	94.91	96.04	67.88	89.31	91.85
Baseline	85.94	94.52	95.93	64.62	87.75	90.78

Table 1: Ablation study on causal learning: comparisons of performance on the CVACT_val-C-ALL and CVACT_test-C-ALL datasets.

Model	CVUSA				CVACT_val				CVACT_test
Model	R@1	R@5	R@10	R@1%	R@1	R@5	R@10	R@1%	R@1	R@5	R@10	R@1%
$\text{SAFA}^{{\dagger}}$ [ 14 ]	89.84	96.93	98.14	99.64	81.03	92.80	94.84	-	-	-	-	-
LPN [ 20 ]	85.79	95.38	96.98	99.41	79.99	90.63	92.56	-	-	-	-	-
$\text{LPN}^{{\dagger}}$ [ 20 ]	92.83	98.00	98.85	99.78	83.66	94.14	95.92	98.41	-	-	-
DSM [ 15 ]	91.96	97.50	98.54	99.67	82.49	92.44	93.99	97.32	-	-	-
TransGeo [ 36 ]	94.08	98.36	99.04	99.77	84.95	94.14	95.78	98.37	-	-	-

Table 2: Comparisons with state-of-the-art models on the CVUSA, CVACT_val and CVACT_test datasets. (†methods that use polar transformation.)

Model	CVUSA-C-ALL				CVACT_val-C-ALL				CVACT_test-C-ALL
Model	R@1	R@5	R@10	R@1%	R@1	R@5	R@10	R@1%	R@1	R@5	R@10	R@1%
CVM-Net [ 5 ]	6.09	16.05	23.14	52.51	-	-	-	-	-	-	-	-
OriCNN [ 9 ]	9.38	22.26	30.04	58.99	15.31	28.31	35.21	58.39	3.69	8.33	11.04	43.93
SAFA [ 14 ]	63.68	78.08	82.82	93.91	56.72	73.60	78.59	91.32	31.18	52.06	58.60	90.41
CVFT [ 16 ]	41.05	64.01	72.64	91.37	45.69	66.45	72.97	88.38	22.82	43.48	51.07	88.99
DSM [ 15 ]	75.27	86.26	89.42	95.07	70.04	82.81	85.86	93.51	47.13	68.41	73.52	93.18

Table 3: Comparisons with state-of-the-art models on the CVUSA-C-ALL, CVACT_val-C-ALL and CVACT_test-C-ALL datasets.

Table 9

Task	Method	R@1	R@5	R@10	R@1%
CVUSA $\rightarrow$ CVACT_val	L2LTR [ 25 ]	47.55	70.58	-	91.39
	$\text{L2LTR}^{{\dagger}}$ [ 25 ]	52.58	75.81	-	93.51
	GeoDTR [ 34 ]	47.79	70.52	-	92.20
	$\text{GeoDTR}^{{\dagger}}$ [ 34 ]	53.16	75.62	-	93.80
	Samp4G [ 1 ]	56.62	77.79	87.02	94.69
	EP-BEV [ 28 ]	59.32	80.79	86.02	94.69

Table 6: Results on different cross-view transfer tasks. Each task is evaluated with representative methods. (†Methods using polar transformation.)

Method	GFLOPs	Avg Inference Time (ms)	CVACT_val R@1
Sample4Geo	90.54	2367.55	90.81
EP-BEV	90.54	2396.53	90.61
Ours	90.56	2374.76	91.61

Table 7: Comparison of GFLOPs and average inference time per batch (batch size = 128) on CVACT_val. Avg inference time (ms) represents the mean time to process one batch.

Figure 6: Heatmap visualizations on CVUSA under clean and snow settings. Compared to the baseline, our method focuses more on task-relevant structural information. Under heavy snow conditions, our method remains highly robust, with attention regions largely unchanged, whereas the baseline’s attention is completely misaligned.

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

S $^3$ Geo: Structure-Semantic Synergistic Learning for Cross-View Geo-Localization
Oct 8, 2026Ziqian Mo, Hill Zhang, Haosheng Tan +2Cross-View Geo-Localization Geospatial Representation Learning
The Hong Kong University of Science and Technology (Guangzhou) Guangzhou, China · Claremont McKenna College Claremont, California, USA
MRGeo: Robust Cross-View Geo-Localization of Corrupted Images via Spatial and Channel Feature Enhancement
Mar 13, 2026Le Wu, Lv Bo, Songsong Ouyang +1Cross-View Geo-Localization Retrieval Robustness
College of Computer Science and Software Engineering, Shenzhen University, China
BGG: Bridging the Geometric Gap between Cross-View images by Vision Foundation Model Adaptation for Geo-Localization
May 11, 2026Wei Wang, Dou Quan, Ning Huyan +4Cross-View Geo-Localization Geospatial Representation Learning
Key Laboratory of Intelligent Perception and Image Understanding of Ministry of Education of China, Xidian University · Department of Automation, Tsinghua University, Beijing 100084, China · School of Telecommunications, Xidian University, Xi’an 710071, China

Oct 8, 2026J/K move · Enter open · S save

S $^3$ Geo: Structure-Semantic Synergistic Learning for Cross-View Geo-Localization

Ziqian Mo, Hill Zhang, Haosheng Tan, Ling Li+1

The Hong Kong University of Science and Technology (Guangzhou) Guangzhou, China · Claremont McKenna College Claremont, California, USA

Cross-View Geo-Localization Geospatial Representation Learning

Open paper AskPDF arXiv

Cross-view geo-localization (CVGL) aims to estimate geographic locations by matching images captured from different viewpoints, such as drone and satellite views. Existing methods mainly rely on visual representations, but often fail to jointly model fine-grained structural correspondences and semantic priors, making them prone to confusion between visually similar but semantically different regions, and thus limiting robustness under large viewpoint variations. To address these challenges, we propose \textbf{S

^3

Geo}, a structure-semantic synergistic learning framework for cross-view matching. Specifically, we first introduce a Decoupled Query Pooling (DQP) module to extract a compact set of region-aware features from dense tokens, enabling explicit modeling of local structural patterns. We then design a query-level contrastive learning scheme with an optimal transport (OT)-based formulation to establish soft correspondences under cross-view spatial misalignment. Furthermore, we incorporate a Semantic Knowledge Distillation (SKD) strategy from a frozen CLIP teacher to transfer semantic priors and relational structures, thereby improving discrimination on hard negatives. By operating synergistically, the semantic priors provide robust contextual filtering, which guides the structural module to establish precise spatial alignments. Experiments on the University-1652 and SUES-200 datasets demonstrate that \textbf{S

^3

Geo} consistently outperforms state-of-the-art approaches without increasing inference complexity, validating the effectiveness of jointly modeling structural and semantic information for CVGL.

Abstract

Cross-View Geo-Localization

Explore similar work

S3^33Geo: Structure-Semantic Synergistic Learning for Cross-View Geo-Localization

MRGeo: Robust Cross-View Geo-Localization of Corrupted Images via Spatial and Channel Feature Enhancement

BGG: Bridging the Geometric Gap between Cross-View images by Vision Foundation Model Adaptation for Geo-Localization

S $^3$ Geo: Structure-Semantic Synergistic Learning for Cross-View Geo-Localization