Research map / Label-efficient and robust learning / Recognition and long-tail
From Coarse to Fine: A Matching and Alignment Framework for Unsupervised Cross-View Geo-Localization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes an unsupervised cross-view geo-localization framework combining DBSCAN-based cluster contrastive learning, local-to-global cross-view cluster matching, and a non-learning color-transfer intermediate domain, evaluated on University-1652 and SUES-200.
Problem
Cross-view geo-localization (drone-satellite) without location labels or matched image pairs, where direct instance-level cross-view matching suffers from large view discrepancies and noisy pseudo-labels.
Approach
Coarse-to-fine alignment: cluster-based contrastive learning assigns pseudo-labels within each view; a local-to-global cluster matching strategy (instances voting for centroids) links clusters across views; and a non-learning Reinhard-style color transfer builds intermediate-domain states to which both views are pulled simultaneously.
Key results
- On University-1652, drone-to-satellite R@1 of 85.95 and AP of 90.33 vs 70.29/74.93 for the prior unsupervised method (Li et al. 2024).
- On University-1652 satellite-to-drone, R@1 94.01 and AP 82.66, outperforming the prior unsupervised method (79.03/61.03).
- On SUES-200, the unsupervised method reaches drone-to-satellite R@1 of 76.90 to 95.10 across altitudes, comparable to supervised methods though below MCCG in several settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add cross-view location alignment to the baseline (ablation) | Cluster contrastive learning baseline · University-1652 drone to satellite | R@1: 59.10 -> 84.55 | Helped |
| add cross-view location alignment to the baseline (ablation) | Cluster contrastive learning baseline · University-1652 drone to satellite | AP: 68.63 -> 89.33 | Helped |
| add cross-view location alignment to the baseline (ablation) | Cluster contrastive learning baseline · University-1652 satellite to drone | R@1: 72.61 -> 93.01 | Helped |
| add cross-view location alignment to the baseline (ablation) | Cluster contrastive learning baseline · University-1652 satellite to drone | AP: 49.60 -> 80.04 | Helped |
| add intermediate state transition to the baseline (ablation) | Cluster contrastive learning baseline · University-1652 drone to satellite | R@1: 59.10 -> 80.15 | Helped |
| add intermediate state transition to the baseline (ablation) | Cluster contrastive learning baseline · University-1652 drone to satellite | AP: 68.63 -> 86.10 | Helped |
| add intermediate state transition to the baseline (ablation) | Cluster contrastive learning baseline · University-1652 satellite to drone | R@1: 72.61 -> 91.44 | Helped |
| add intermediate state transition to the baseline (ablation) | Cluster contrastive learning baseline · University-1652 satellite to drone | AP: 49.60 -> 76.62 | Helped |
| add intermediate state transition to cross-view location alignment (ablation) | Baseline + cross-view location alignment · University-1652 drone to satellite | R@1: 84.55 -> 85.95 | Helped |
| add intermediate state transition to cross-view location alignment (ablation) | Baseline + cross-view location alignment · University-1652 drone to satellite | AP: 89.33 -> 90.33 | Helped |
| add intermediate state transition to cross-view location alignment (ablation) | Baseline + cross-view location alignment · University-1652 satellite to drone | R@1: 93.01 -> 94.01 | Helped |
| use local-to-global cluster matching instead of instance-only matching (ablation) | Cross-view location alignment · University-1652 drone to satellite | R@1: 79.35 -> 84.55 | Helped |
Surprising findings
- The unsupervised method matches or exceeds several supervised methods on University-1652 (e.g., beats LPN, PCL, FSRA on drone-to-satellite R@1).
Limitations
- Performance remains below the best supervised method (MCCG) on several SUES-200 and University-1652 metrics, attributed to clustering errors and data discrepancy.
- Unsupervised method depends on DBSCAN clustering quality, which can introduce noisy pseudo-labels.
- Compared against only one prior unsupervised method on University-1652 because other unsupervised code was unavailable.
Experimental setup
Datasets: University-1652 SUES-200
Benchmarks: University-1652 SUES-200
Metrics: Recall@1 Recall@5 Recall@10 AP
Baselines: Li et al. 2024 (unsupervised EM) Zheng et al. 2020 LCM LPN PCL F3-net FSRA MCCG ViT (SUES-200)
Builds on: ClusterNCE AGW Reinhard color transfer MoCo DBSCAN
Closest papers by meaning
- Cross-View Geo-Localization via Learning Disentangled Geometric Layout Correspondence (AAAI 2023 · similarity 0.92)
- Aligning Geometric Spatial Layout in Cross-View Geo-Localization via Feature Recombination (AAAI 2024 · similarity 0.91)
- Unleashing Unlabeled Data: A Paradigm for Cross-View Geo-Localization (CVPR 2024 · similarity 0.91)
- Semantic Ambiguity Modeling and Propagation for Fine-Grained Visual Cross View Geo-Localization (AAAI 2025 · similarity 0.90)
- FG^2: Fine-Grained Cross-View Localization by Fine-Grained Feature Matching (CVPR 2025 · similarity 0.89)
- View From Above: Orthogonal-View aware Cross-view Localization (CVPR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.