Research map / Label-efficient and robust learning / Semi-supervised learning
Unleashing Unlabeled Data: A Paradigm for 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 and semi-supervised framework for large-area cross-view geo-localization that trains without ground-satellite pairs, using projected fake satellite images for cold-start and adaptive mutual-matching with threshold filtering and curriculum learning for pseudo-label refinement.
Problem
Cross-view geo-localization normally needs precisely paired ground-satellite images and GPS labels, which are costly to collect; the paper asks how to train CVGL using unlabeled data when there is no cold-start signal.
Approach
A correspondence-free projection (spherical transform to BEV plus CycleGAN style transfer) turns ground panoramas into fake satellite images to provide initial positive pairs; an Adaptive-Mutual-Matching scheme (mutual nearest neighbours, similarity-margin threshold, and curriculum relaxing of the threshold) then refines pseudo-labels for semi-supervised training.
Key results
- Unsupervised model (Ours-base, 0% labels) reaches R@1 92.56 on CVUSA and 84.58 on CVACT Val, comparable to supervised methods.
- With 10% labels fine-tuning, Ours-base reaches R@1 94.88 on CVUSA and 87.89 on CVACT Val, close to Sample4Geo's 97.83 and 87.49 on CVACT Val.
- On CVACT, the semi-supervised model with 1% labels reaches R@1 68.29 on VIGOR? (table lists CVACT 1% R@1 68.29 and VIGOR 10% R@1 78.88).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add geometric projection (BEV) to the cold-start stage (ablation) | Unsupervised CVGL · CVACT | R@1: 0.0 -> 44.81 | Helped |
| Add fake-image imaging projection (ablation) | Unsupervised CVGL · CVACT | R@1: 22.78 -> 44.81 | Helped |
| Add intra-view contrastive learning (ablation) | Unsupervised CVGL · CVACT | R@1: 42.98 -> 44.81 | Helped |
| Add cross-view contrastive learning (ablation) | Unsupervised CVGL · CVACT | R@1: 0.011 -> 44.81 | Helped |
| Add mutual matching when assigning pseudo-labels (ablation) | Semi-supervised CVGL · CVUSA | R@1: 14.40 -> 87.90 | Helped |
| Add threshold filtering to pseudo-labels (ablation) | Semi-supervised CVGL · CVUSA | R@1: 36.52 -> 87.90 | Helped |
| Add curriculum learning with progressively harder pseudo-labels (ablation) | Semi-supervised CVGL · CVACT | R@1: 63.14 -> 82.96 | Helped |
| Add SAFA spatial attention (ablation) | Unsupervised CVGL · CVACT | R@1: 73.48 -> 82.96 | Helped |
| Use separate-weight rather than shared-weight encoders (ablation) | Unsupervised CVGL · CVACT | R@1: 81.06 -> 82.96 | Helped |
| Add label smoothing (ablation) | Unsupervised CVGL · CVACT | R@1: 71.88 -> 82.96 | Helped |
| Train with random-shift images (ablation) | Unsupervised CVGL · CVACT | R@1: 82.96 -> 54.49 | Hurt |
| Train with random-rotate images (ablation) | Unsupervised CVGL · CVACT | R@1: 82.96 -> 50.32 | Hurt |
Surprising findings
- Fixing the initial pseudo-labels instead of refining them causes a large performance drop, showing that progressively adding harder samples matters.
Limitations
- Fixed geometric projection cannot simulate varied real-world cross-view relationships.
- Homography-based projection is insufficient for complex city scenes such as VIGOR, so VIGOR is only evaluated in semi-supervised setting.
- Relies on localization and orientation alignment assumptions in datasets, though random shift/rotate tests show some robustness.
Experimental setup
Datasets: CVUSA CVACT VIGOR
Benchmarks: CVUSA CVACT Val CVACT Test VIGOR
Metrics: R@1 R@5 R@10 R@1%
Baselines: CVM-NET Liu SAFA DSM L2LTR GeoDTR TransGeo Sample4Geo
Builds on: Sample4Geo SAFA CycleGAN InfoNCE ConvNeXt
Closest papers by meaning
- From Coarse to Fine: A Matching and Alignment Framework for Unsupervised Cross-View Geo-Localization (AAAI 2025 · similarity 0.91)
- UniGeoRS: A Unified Benchmark for Tri-view Geo-Localization (CVPR 2026 · similarity 0.89)
- UniABG: Unified Adversarial View Bridging and Graph Correspondence for Unsupervised Cross-View Geo-Localization (AAAI 2026 · similarity 0.89)
- Aligning Geometric Spatial Layout in Cross-View Geo-Localization via Feature Recombination (AAAI 2024 · similarity 0.88)
- CVGL: Causal Learning and Geometric Topology (NeurIPS 2025 · similarity 0.86)
- Semantic Ambiguity Modeling and Propagation for Fine-Grained Visual Cross View Geo-Localization (AAAI 2025 · similarity 0.86)
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.