Research map / Vision-language and multimodal / Human and embodied modeling
Semantic Ambiguity Modeling and Propagation for Fine-Grained Visual 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 SAMP, which models semantic ambiguity of semi-positive aerial references as 2D Gaussian offset uncertainty and propagates it into the retrieval similarity metric to balance joint retrieval and offset regression for fine-grained cross-view geo-localization.
Problem
Joint retrieval-and-calibration for fine-grained visual cross-view geo-localization suffers from semantic ambiguity in semi-positive reference images (low overlap, dynamic foregrounds, viewpoint changes, perceptual aliasing), causing negative transfer and dominating gradients between tasks.
Approach
Learn a per-sample 2D Gaussian covariance over the query offset via Cholesky decomposition and use it both to down-weight the regression loss under high uncertainty and to define an uncertainty-aware similarity metric that attenuates the ratio loss gradient for ambiguous semi-positive samples; uncertainty also selects the best top-K candidate at inference.
Key results
- On VIGOR same-area, retrieval top-1 reaches 80.61% and hit rate 93.89% versus Sample4Geo's 77.86% and 89.82%, with practical localization mean error 4.91m vs CCVPE's 6.12m.
- On CVACT practical setting, top-1 recall 91.02% and hit rate 94.78%, outperforming Sample4Geo (87.96% top-1, 91.63% hit rate).
- Ablation on VIGOR same-area: uncertainty-aware ratio loss raises hit rate from 82.8-level baseline to 91.5% (9.7 points gain); top-K selection by lowest uncertainty raises top-1 to 80.6%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the full ambiguity-aware joint method rather than Sample4Geo | VIGOR same-area retrieval · VIGOR same-area | Hit Rate: 89.82 -> 93.89 | Helped |
| Use the full ambiguity-aware joint method rather than Sample4Geo | VIGOR cross-area retrieval · VIGOR cross-area | Hit Rate: 69.87 -> 72.91 | Helped |
| Use the full ambiguity-aware joint method rather than Sample4Geo | CVACT-validation practical retrieval · CVACT-validation, practical | Top-1 recall: 87.96 -> 91.02 | Helped |
| Use the full ambiguity-aware joint method rather than CCVPE | VIGOR same-area practical localization · VIGOR same-area, practical | Mean localization error: 6.12 -> 4.91 | Helped |
| Use the full ambiguity-aware joint method rather than CCVPE | VIGOR cross-area practical localization · VIGOR cross-area, practical | Mean localization error: 7.29 -> 6.28 | Helped |
| Add ambiguity-aware regression loss L′reg instead of Lreg (ablation) | VIGOR same-area joint model · VIGOR same-area | Hit Rate: 67.2 -> 78.4 | Helped |
| Add ambiguity-aware regression loss L′reg instead of Lreg (ablation) | VIGOR same-area joint model · VIGOR same-area | Mean-Error: 10.08 -> 7.83 | Helped |
| Add uncertainty-aware similarity metric Mu in the ratio loss (ablation) | VIGOR same-area joint model · VIGOR same-area | Hit Rate: 81.8 -> 91.5 | Helped |
| Select the top-K candidate with lowest predicted uncertainty (ablation) | VIGOR same-area inference · VIGOR same-area | Top-1 recall: 80.1 -> 80.6 | Helped |
Surprising findings
- Modeling ambiguity in the regression loss alone improves retrieval hit rate by 11.2 points, showing regression-retrieval interference can be mitigated by uncertainty.
Limitations
- Uncertainty-aware metric hyperparameters (alpha, beta, gamma) must be set manually.
- Evaluated only on VIGOR and CVACT; practical setting for fine-grained baselines is approximated using Sample4Geo retrieved candidates.
Experimental setup
Datasets: VIGOR CVACT
Benchmarks: VIGOR CVACT
Metrics: top-1 recall top-5 recall top-10 recall hit rate mean localization error (m) median localization error (m)
Baselines: TransGeo FRGeo Sample4Geo MCC SliceMatch CCVPE VIGOR Shi L2LTR GeoDTR UCVGL Ours
Builds on: VIGOR SAFA Sample4Geo CCVPE
Closest papers by meaning
- From Coarse to Fine: A Matching and Alignment Framework for Unsupervised Cross-View Geo-Localization (AAAI 2025 · similarity 0.90)
- Aligning Geometric Spatial Layout in Cross-View Geo-Localization via Feature Recombination (AAAI 2024 · similarity 0.88)
- On the Estimation of Image-matching Uncertainty in Visual Place Recognition (CVPR 2024 · similarity 0.87)
- RELOCATE: A Simple Training-Free Baseline for Visual Query Localization Using Region-Based Representations (CVPR 2025 · similarity 0.86)
- Unleashing Unlabeled Data: A Paradigm for Cross-View Geo-Localization (CVPR 2024 · similarity 0.86)
- Weatherproofing Retrieval for Localization with Generative AI and Geometric Consistency (ICLR 2024 · 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.