UniABG: Unified Adversarial View Bridging and Graph Correspondence for Unsupervised Cross-View Geo-Localization

2025-11-15

AI Summary

Cross-view geolocalization (CVGL) faces two major challenges in the unsupervised setting: significant inter-domain distribution shifts and severe noise in pseudo-labels. To address these, we propose a two-stage unsupervised framework. In Stage I, we design a view-aware adversarial bridging module that learns generative invariant features to mitigate cross-view distribution shift. In Stage II, we construct a heterogeneous dual-view structural graph and introduce graph filtering and correspondence calibration mechanisms to denoise pseudo-labels and enhance cross-view matching. Our method requires no paired annotations. On University-1652 and SUES-200, it achieves absolute mAP improvements of 10.63% and 16.73%, respectively, for satellite-to-UAV retrieval—surpassing existing supervised methods. These results demonstrate its robustness and state-of-the-art performance in feature alignment and cross-view correspondence learning.

Technology Category

• Computer Vision: Image and Video Retrieval
• Machine Learning: Multi-instance/Multi-view Learning
• Intelligent Robots: Localization, Mapping, and Navigation

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Cross-view geo-localization (CVGL) matches query images ($ extit{e.g.}$, drone) to geographically corresponding opposite-view imagery ($ extit{e.g.}$, satellite). While supervised methods achieve strong performance, their reliance on extensive pairwise annotations limits scalability. Unsupervised alternatives avoid annotation costs but suffer from noisy pseudo-labels due to intrinsic cross-view domain gaps. To address these limitations, we propose $ extit{UniABG}$, a novel dual-stage unsupervised cross-view geo-localization framework integrating adversarial view bridging with graph-based correspondence calibration. Our approach first employs View-Aware Adversarial Bridging (VAAB) to model view-invariant features and enhance pseudo-label robustness. Subsequently, Heterogeneous Graph Filtering Calibration (HGFC) refines cross-view associations by constructing dual inter-view structure graphs, achieving reliable view correspondence. Extensive experiments demonstrate state-of-the-art unsupervised performance, showing that UniABG improves Satellite $ ightarrow$ Drone AP by +10.63% on University-1652 and +16.73% on SUES-200, even surpassing supervised baselines. The source code is available at https://github.com/chenqi142/UniABG

Problem

• Unsupervised cross-view geo-localization without pairwise annotations
• Reducing noisy pseudo-labels from cross-view domain gaps
• Improving drone-satellite image matching through adversarial bridging

Innovation

• Unified adversarial view bridging for cross-view feature modeling
• Graph-based correspondence calibration refining cross-view associations
• Dual-stage unsupervised framework enhancing pseudo-label robustness