[논문 리뷰] (MGS)$^2$-Net: Unifying Micro-Geometric Scale and Macro-Geometric Structure for Cross-View Geo-Localization

Minglei Li, Mengfan He외 4명|arXiv (Cornell University)|2026. 02. 11.

Robotics and Sensor-Based Localization인용 수 0

한 줄 요약

This paper presents (MGS)², a geometry-grounded cross-view geo-localization framework that combines Macro-Geometric Structure Filtering and Micro-Geometric Scale Adaptation with a Geometric-Appearance Contrastive Distillation loss to bridge oblique UAV views and satellite references, achieving state-of-the-art results on University-1652 and SUES-200.

ABSTRACT

Cross-view geo-localization (CVGL) is pivotal for GNSS-denied UAV navigation but remains brittle under the drastic geometric misalignment between oblique aerial views and orthographic satellite references. Existing methods predominantly operate within a 2D manifold, neglecting the underlying 3D geometry where view-dependent vertical facades (macro-structure) and scale variations (micro-scale) severely corrupt feature alignment. To bridge this gap, we propose (MGS)$^2$, a geometry-grounded framework. The core of our innovation is the Macro-Geometric Structure Filtering (MGSF) module. Unlike pixel-wise matching sensitive to noise, MGSF leverages dilated geometric gradients to physically filter out high-frequency facade artifacts while enhancing the view-invariant horizontal plane, directly addressing the domain shift. To guarantee robust input for this structural filtering, we explicitly incorporate a Micro-Geometric Scale Adaptation (MGSA) module. MGSA utilizes depth priors to dynamically rectify scale discrepancies via multi-branch feature fusion. Furthermore, a Geometric-Appearance Contrastive Distillation (GACD) loss is designed to strictly discriminate against oblique occlusions. Extensive experiments demonstrate that (MGS)$^2$ achieves state-of-the-art performance, recording a Recall@1 of 97.5\% on University-1652 and 97.02\% on SUES-200. Furthermore, the framework exhibits superior cross-dataset generalization against geometric ambiguity. The code is available at: \href{https://github.com/GabrielLi1473/MGS-Net}{https://github.com/GabrielLi1473/MGS-Net}.

연구 동기 및 목표

• oblique UAV 뷰와 정투영 위성 참조 간의 3D 기하학적 간격을 CVGL에서 해소한다.
• 매크로-기하 구조 정보와 미세 기하 정보를 명시적으로 필터링하고 정렬하여 교차 뷰 매칭을 개선한다.
• 깊이 가이드 다중 스케일 융합 및 기하 인식 손실을 통해 스케일 및 시점 변화를 완화한다.
• 데이터셋 간의 최첨단 성능 및 일반화 가능성을 입증한다.

제안 방법

• Macr o-Geometric Structure Filtering (MGSF) that computes dilated geometric gradients to suppress view-dependent vertical facade artifacts and emphasize horizontal planes.
• Micro-Geometric Scale Adaptation (MGSA) that uses depth priors to dynamically fuse multi-scale features and rectify scale discrepancies via depth-aware branches.
• Depth-Aware Scale Fusion (DASF) within MGSA that creates near/mid/far scale branches and learns pixel-wise scale weights from depth embeddings.
• Geometric-Appearance Contrastive Distillation (GACD) loss that enforces a margin between semantic activations of rooftops versus vertical facades for robust discrimination.
• Integration of MGSA, MGSF, and GACD into an end-to-end retrieval framework with a weighted triplet loss for cross-view localization.

실험 결과

연구 질문

• RQ1 How can explicit 3D macro-geometric information be leveraged to reduce cross-view misalignment between oblique UAV imagery and satellite orthophotos?
• RQ2 Can depth-guided micro-scale adaptation and geometry-aware filtering improve feature robustness to scale and viewpoint variations in CVGL?
• RQ3 Does a geometry-centric distillation objective better discriminate against oblique occlusions than traditional depth regression losses?
• RQ4 What is the generalization and robustness of the proposed approach across different datasets and altitude scenarios?

주요 결과

방법	Publication	Drone → Satellite R@1	Drone → Satellite AP	Satellite → Drone R@1	Satellite → Drone AP
(MGS)² (Ours)	—	97.50	97.97	98.57	97.27

• (MGS)² achieves state-of-the-art Recall@1 on University-1652 (97.50) and SUES-200 ( Drone→Satellite 97.02, Satellite→Drone 100.00).
• The Macro-Geometric Structure Filtering (MGSF) module effectively suppresses vertical facade artifacts and enhances view-invariant horizontal surfaces.
• MGSA with depth priors improves robustness to altitude-related scale changes, showing strong performance across four altitude levels (150–300m) on SUES-200.
• GACD loss enforces a margin that emphasizes geometric consistency (roof-like structures) over oblique occlusions, boosting ranking quality.
• Cross-dataset generalization experiments show substantial transferability, outperforming zero-shot baselines on DenseUAV by focusing on geometry rather than texture.