GeoAlign: Foundation Model-Driven Asymmetric Dual-Stream Manifold Alignment for Unsupervised Cross-View Geo-Localization
IEEE Transactions on Circuits and Systems for Video Technology · 2026-11 · 36 (11) · 17702-17717
DOI: 10.1109/tcsvt.2026.3716447 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Circuits and Systems for Video Technology, 2026-11. Verfasst von Juzheng Liu und 4 weiteren Autor:innen. Die Arbeit nennt 59 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Fundamental Research Funds for the Central Universities, Natural Science Basic Research Program of Shaanxi.
Abstract (Originalsprache)
Cross-view geo-localization (CVGL) is critical for autonomous drone navigation in GNSS-denied environments. Existing supervised frameworks rely on costly paired annotations, hindering open-world deployments. Unsupervised alternatives mitigate data dependency but struggle with severe non-linear geometric distortions, specifically perspective disparities and scale inconsistencies from drone altitude and attitude variations. Optimization instability can also arise from noisy pseudo-labels. Addressing these challenges, we present GeoAlign, a novel asymmetric dual-stream unsupervised framework. Observing that conventional symmetric architectures struggle to handle complex non-linear geometric distortions inherent in drone views, GeoAlign incorporates an Asymmetric Geometry-Rectified Adapter (AGRA). This module employs a Mixture-of-Experts exclusively on the drone branch, rectifying distorted manifolds towards the standard satellite feature space. To ensure robust convergence without ground-truth supervision, we introduce an Optimal Transport-driven Curriculum Alignment (OTCA) mechanism that uses the Sinkhorn algorithm to generate consistent soft pseudo-labels, which subsequently guide the optimization via bidirectional soft distillation and reliability-driven complex matching. Furthermore, the system integrates an Intrinsic Structure Mining Stream (ISMS) to mine compact intrinsic structures. By employing manifold-constrained whitening to enforce statistical independence, this stream eliminates channel redundancy and uncovers fine-grained discriminative cues. Extensive evaluations demonstrate that GeoAlign achieves state-of-the-art or highly competitive unsupervised performance. By freezing the foundation backbone and adapting cached descriptors, GeoAlign reduces repeated image-level training cost while remaining competitive with several supervised methods. This work validates the efficacy of integrating distinct geometric priors into vision foundation models for robust geospatial reasoning. Our code is available at https://github.com/JZLhope/GeoAlign
Förderung
- National Natural Science Foundation of China
- Fundamental Research Funds for the Central Universities
- Natural Science Basic Research Program of Shaanxi
- Ningbo Natural Science Foundation
- National Key Laboratory of Air-Based Information Perception and Fusion
- Ministry of Education Joint Foundation
- Shaanxi Province Science and Technology Plan Project
- Xi’an City Science and Technology Plan Project
Zitieren
APA
Liu, J., Qin, H., Zhang, X., Pang, Z., & Deng, C. (2026). GeoAlign: Foundation Model-Driven Asymmetric Dual-Stream Manifold Alignment for Unsupervised Cross-View Geo-Localization. IEEE Transactions on Circuits and Systems for Video Technology, 36(11), 17702-17717. https://doi.org/10.1109/tcsvt.2026.3716447
Vancouver
Liu J, Qin H, Zhang X, Pang Z, Deng C. GeoAlign: Foundation Model-Driven Asymmetric Dual-Stream Manifold Alignment for Unsupervised Cross-View Geo-Localization. IEEE Transactions on Circuits and Systems for Video Technology. 2026;36(11):17702-17717. doi:10.1109/tcsvt.2026.3716447
BibTeX
@article{Liu2026,
author = {Liu, Juzheng and Qin, Hanlin and Zhang, Xupei and Pang, Zibo and Deng, Chenguang},
title = {GeoAlign: Foundation Model-Driven Asymmetric Dual-Stream Manifold Alignment for Unsupervised Cross-View Geo-Localization},
journal = {IEEE Transactions on Circuits and Systems for Video Technology},
year = {2026},
volume = {36},
number = {11},
pages = {17702--17717},
doi = {10.1109/tcsvt.2026.3716447}
}
EUVIMED