Research map / 3D vision / 3D generation
Aligning Geometric Spatial Layout in Cross-View Geo-Localization via Feature Recombination
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 FRGeo, a cross-view geo-localization network with a parameter-free Feature Recombination Module that aligns ground and aerial features by region, plus a weighted (B+1)-tuple loss, achieving SOTA on CVUSA, CVACT and VIGOR.
Problem
Cross-view geo-localization must match ground and aerial images despite large viewpoint and appearance differences; existing methods either ignore geometric spatial layout correspondence or rely on costly polar transforms or strict center-alignment assumptions.
Approach
Divide ground and aerial feature maps into four directional regions (N/E/S/W aligned FoV) and recombine spatially average-pooled region features to align layout at the region level directly on feature maps, paired with a weighted (B+1)-tuple loss that contrasts an anchor against all in-batch negatives.
Key results
- FRGeo reaches 97.06% R@1 on CVUSA, versus 94.08% for TransGeo and 93.76% for GeoDTR.
- On VIGOR cross-area, FRGeo obtains 37.54% R@1 (+18.55% over the prior best), and 71.26% same-area R@1.
- Ablation on CVUSA: Baseline 94.10% R@1 rises to 96.70% with FRM alone and 97.06% with FRM+WBL; FRGeo reaches 10-epoch performance that other methods need about 100 epochs to match.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add FRM (ablation) | ConvNeXt-T baseline · CVUSA | R@1: 94.10% -> 96.70% | Helped |
| add WBL (ablation) | ConvNeXt-T baseline · CVUSA | R@1: 94.10% -> 95.27% | Helped |
| add FRM and WBL (ablation) | ConvNeXt-T baseline · CVUSA | R@1: 94.10% -> 97.06% | Helped |
| add FRM (ablation) | ConvNeXt-T baseline · CVACT val | R@1: 84.77% -> 88.79% | Helped |
| add WBL (ablation) | ConvNeXt-T baseline · CVACT val | R@1: 84.77% -> 87.42% | Helped |
| add FRM and WBL (ablation) | ConvNeXt-T baseline · CVACT val | R@1: 84.77% -> 90.35% | Helped |
| add FRM (ablation) | ConvNeXt-T baseline · VIGOR Same-area | R@1: 56.25% -> 67.38% | Helped |
| add WBL (ablation) | ConvNeXt-T baseline · VIGOR Same-area | R@1: 56.25% -> 66.58% | Helped |
| add FRM and WBL (ablation) | ConvNeXt-T baseline · CVACT test | R@1: 59.19% -> 72.15% | Helped |
Surprising findings
- FRGeo outperforms polar-transform-based methods without using polar transform, despite being only a region-pooling recombination of features.
Limitations
- Region division is fixed to four quadrants based on a FoV assumption, which may not generalize to other view settings.
- Only evaluated on three cross-view datasets with ConvNeXt-T and a TransGeo variant as backbones.
Experimental setup
Datasets: CVUSA CVACT VIGOR
Benchmarks: CVUSA CVACT val CVACT test VIGOR
Metrics: R@1 R@5 R@10 R@1% hit rate
Baselines: SAFA DSM CDE L2LTR TransGeo SEH GeoDTR Siamese-VGG SAFA+Mining VIGOR
Builds on: weighted soft margin ranking loss TransGeo ConvNeXt Multi-class N-pair loss
Closest papers by meaning
- Cross-View Geo-Localization via Learning Disentangled Geometric Layout Correspondence (AAAI 2023 · similarity 0.93)
- From Coarse to Fine: A Matching and Alignment Framework for Unsupervised Cross-View Geo-Localization (AAAI 2025 · similarity 0.91)
- Semantic Ambiguity Modeling and Propagation for Fine-Grained Visual Cross View Geo-Localization (AAAI 2025 · similarity 0.88)
- Unleashing Unlabeled Data: A Paradigm for Cross-View Geo-Localization (CVPR 2024 · similarity 0.88)
- CVGL: Causal Learning and Geometric Topology (NeurIPS 2025 · similarity 0.88)
- Enhancing Visual Localization with Cross-Domain Image Generation (ICML 2025 · similarity 0.87)
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.