Research map / Label-efficient and robust learning / Recognition and long-tail
CVGL: Causal Learning and Geometric Topology
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 CLGT, a cross-view geo-localization framework combining frequency-domain causal intervention (CFE), BEV road-topology fusion (GT Fusion), and gated pooling (DA Pooling) on top of EP-BEV, reporting SOTA on CVUSA/CVACT and corrupted variants.
Problem
Cross-view geo-localization matches street-view images to aerial images under large viewpoint differences, confounding factors (lighting, background), and real-world corruptions.
Approach
Applies a DCT-domain causal intervention that perturbs low/high-frequency non-causal components with content-aware radii, fuses BEV road topology into street features via cross-attention with dual dynamic fusion, and aggregates features with a learned gate over max, average and GeM pooling.
Key results
- Reaches 98.85% R@1 on CVUSA and 73.22% R@1 on CVACT_test (gamma=0.5), versus EP-BEV's 97.41% and 71.41%.
- Improves R@1 on CVUSA-C-ALL by 6.62 points over EP-BEV (92.84 vs 86.22 with gamma=0.5).
- On CVUSA->CVACT_val transfer, R@1 reaches 64.82% with gamma=0.5 versus 59.32% for EP-BEV.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CLGT instead of EP-BEV | Cross-view retrieval · CVUSA | R@1: 97.41 -> 98.85 | Helped |
| Add CLGT instead of EP-BEV | Cross-view retrieval · CVACT_val | R@1: 90.61 -> 91.97 | Helped |
| Add CLGT instead of EP-BEV | Cross-view retrieval · CVACT_test | R@1: 1.81% points | Helped |
| Add CLGT instead of EP-BEV | Cross-view retrieval · CVUSA-C-ALL | R@1: 86.22 -> 92.64 | Helped |
| Add CLGT instead of EP-BEV | Cross-view retrieval · CVACT_val-C-ALL | R@1: 85.94 -> 88.68 | Helped |
| Add CLGT instead of EP-BEV | Cross-view retrieval · CVACT_test-C-ALL | R@1: 64.62 -> 69.06 | Helped |
| Remove GT Fusion (ablation) | CLGT · CVACT_val | R@1: 91.61 -> 91.24 | Hurt |
| Remove CFE (ablation) | CLGT · CVACT_val | R@1: 91.61 -> 91.01 | Hurt |
| Use causal learning only (ablation) | CLGT · CVACT_val-C-ALL | R@1: 85.94 -> 88.11 | Helped |
| Use causal learning only (ablation) | CLGT · CVACT_test-C-ALL | R@1: 64.62 -> 67.88 | Helped |
| Replace GeM pooling with DA Pooling (ablation) | CLGT · CVACT_val | R@1: +2.37% | Helped |
| Replace average pooling with DA Pooling (ablation) | CLGT · CVACT_val | R@1: 91.19 -> 91.61 | Helped |
Surprising findings
- Reported GFLOPs and inference time are nearly identical to EP-BEV and Sample4Geo, indicating marginal overhead.
Limitations
- BEV representations from geometric transformation contain considerable noise, limiting gains.
- The gamma value of 0.5 is reported as optimal but the default in the main setup is 0.1; collapse risk noted for large gamma.
Experimental setup
Datasets: CVUSA CVACT CVUSA-C-ALL CVACT_val-C-ALL CVACT_test-C-ALL
Benchmarks: CVUSA CVACT_val CVACT_test CVUSA-C-ALL CVACT_val-C-ALL CVACT_test-C-ALL CVUSA->CVACT cross-dataset
Metrics: Recall@1 Recall@5 Recall@10 Recall@1% GFLOPs inference time
Baselines: SAFA LPN DSM TransGeo GeoDTR Sample4Geo ConGeo EP-BEV L2LTR CVM-Net
Builds on: EP-BEV Sample4Geo InfoNCE GeM pooling Spatial Reduction Attention
Closest papers by meaning
- Geo2: Geometry-Guided Cross-view Geo-Localization and Image Synthesis (CVPR 2026 · similarity 0.90)
- UniGeoRS: A Unified Benchmark for Tri-view Geo-Localization (CVPR 2026 · similarity 0.88)
- Aligning Geometric Spatial Layout in Cross-View Geo-Localization via Feature Recombination (AAAI 2024 · similarity 0.88)
- Unleashing Unlabeled Data: A Paradigm for Cross-View Geo-Localization (CVPR 2024 · similarity 0.86)
- Vision-Language Reasoning for Geolocalization: A Reinforcement Learning Approach (AAAI 2026 · similarity 0.85)
- Semantic Ambiguity Modeling and Propagation for Fine-Grained Visual Cross View Geo-Localization (AAAI 2025 · similarity 0.85)
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.