Research map / Point clouds / Point cloud classification
FG^2: Fine-Grained Cross-View Localization by Fine-Grained Feature Matching
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 FG^2, a fine-grained cross-view localization method that lifts ground features to 3D, learns height-wise feature selection to pool them into a BEV point plane, and estimates a 3-DoF pose by Procrustes alignment over sparse matches, trained only with camera pose supervision.
Problem
Estimating the 3-DoF pose of a ground-level image within an aerial image by fine-grained local feature matching, without ground-truth cross-view local correspondences.
Approach
Learn a feature selection along height when pooling 3D ground points to a BEV point plane (instead of sum/max), so each BEV point traces back to ground image features; match sparse BEV descriptors between ground and aerial point planes and solve pose via differentiable Kabsch/Procrustes, supervised only by camera pose.
Key results
- On VIGOR cross-area known orientation, mean localization error 2.41 m vs 3.35 m for HC-Net (28% reduction).
- On VIGOR same-area known orientation, mean localization error 1.95 m vs 2.65 m for HC-Net (26% reduction).
- On KITTI cross-area, mean localization error 7.45 m, lower than HC-Net's 8.47 m, with median 4.03 m vs 4.57 m.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FG² instead of HC-Net | VIGOR same-area, known orientation localization · VIGOR same-area test | Mean localization error (m): 2.65 -> 1.95 | Helped |
| Use FG² instead of HC-Net | VIGOR cross-area, known orientation localization · VIGOR cross-area test | Mean localization error (m): 3.35 -> 2.41 | Helped |
| Use two-step inference instead of single-step FG² | FG² on VIGOR same-area, unknown orientation · VIGOR same-area test | Mean localization error (m): 8.95 -> 3.78 | Helped |
| Use two-step inference instead of single-step FG² | FG² on VIGOR cross-area, unknown orientation · VIGOR cross-area test | Mean localization error (m): 10.02 -> 5.95 | Helped |
| Use FG² instead of the strongest listed baseline | KITTI same-area localization · KITTI same-area test | Mean localization error (m): 0.80 -> 0.75 | Helped |
| Use FG² instead of DenseFlow | KITTI cross-area localization · KITTI cross-area test | Mean localization error (m): 7.97 -> 7.45 | Helped |
| Increase the point grid from N=21 × 21 to N=41 × 41 (ablation) | FG² · VIGOR same-area validation, known orientation | Mean localization error (m): 2.50 -> 2.17 | Helped |
| Increase the point grid from N=41 × 41 to N=51 × 51 (ablation) | FG² · VIGOR same-area validation, known orientation | Mean localization error (m): 2.17 -> 2.19 | Hurt |
| Add matching loss LM (β=1 rather than β=0) (ablation) | FG² · VIGOR same-area validation, known orientation | Mean localization error (m): 2.63 -> 2.17 | Helped |
| Increase matching-loss weight from β=1 to β=10 (ablation) | FG² · VIGOR same-area validation, known orientation | Mean localization error (m): 2.17 -> 2.54 | Hurt |
| Sample a subset of correspondences instead of using all correspondences (ablation) | FG² Procrustes pose estimation · VIGOR same-area validation, known orientation | Mean localization error (m): 2.55 -> 2.17 | Helped |
| Use learned feature selection along height instead of summing features (ablation) | FG² BEV pooling · VIGOR same-area validation, known orientation | Mean localization error (m): 2.34 -> 2.17 | Helped |
Surprising findings
- Under unknown orientation the method is weaker than CCVPE and DenseFlow (same-area mean 8.95 m vs 3.74 m for CCVPE), and its orientation errors on KITTI are worse than baselines.
Limitations
- Performance degrades substantially with unknown orientation, requiring a two-step inference to approach state of the art.
- Orientation prediction on KITTI is worse than baselines because the method cannot exploit the road-direction bias.
- Dense matching from all correspondences hurts accuracy, so sparse sampling is required, adding design complexity.
Experimental setup
Datasets: VIGOR KITTI
Benchmarks: VIGOR KITTI
Metrics: mean localization error median localization error mean orientation error median orientation error recall at 1m and 5m (lateral/longitudinal)
Baselines: SliceMatch CCVPE GGCVT DenseFlow HC-Net
Builds on: DinoV2 BEVFormer LoFTR Deformable attention VCE loss (Cross-view localization via virtual correspondence)
Closest papers by meaning
- From Coarse to Fine: A Matching and Alignment Framework for Unsupervised Cross-View Geo-Localization (AAAI 2025 · similarity 0.89)
- BevSplat: Resolving Height Ambiguity via Feature-Based Gaussian Primitives for Weakly-Supervised Cross-View Localization (NeurIPS 2025 · similarity 0.88)
- View From Above: Orthogonal-View aware Cross-view Localization (CVPR 2024 · similarity 0.88)
- Seeing Across Skies and Streets: Feedforward 3D Reconstruction from Satellite, Drone, and Ground Images (NeurIPS 2026 · similarity 0.87)
- Aligning Geometric Spatial Layout in Cross-View Geo-Localization via Feature Recombination (AAAI 2024 · similarity 0.86)
- 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.