Research map / 3D vision / Depth and stereo
View From Above: Orthogonal-View aware Cross-view 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 View From Above (VFA), a pixel-wise fine-grained cross-view localization method that adds top-to-ground attention aggregation, a cycle domain adaptation loss, and an equidistant re-projection loss on top of PureACL.
Problem
Fine-grained cross-view vehicle localization from ground images to satellite maps is hurt by the ground-satellite appearance gap: homography-based methods ignore off-ground features and occlusions and suffer from domain shifts and orientation bias toward nearby keypoints.
Approach
Aggregate features of elevated pixels down onto the on-ground pixel beneath them via column-wise attention (T2GA), enforce view-invariant features with a bidirectional cycle domain adaptation loss (CycDA), and reweight the re-projection error by keypoint distance (ERP) so orientation error does not depend on keypoint distance.
Key results
- Under (±45°,±20m) noise on KITTI-CVL cross split, mean lateral error 0.21m and mean longitudinal error 0.12m vs PureACL 3.89m/4.24m (approx.), mean orientation 2.88° vs 10.96° for PureACL.
- Ablation on KITTI cross: adding T2GA, CycDA and ERP progressively lowers mean orientation error from 10.96° (baseline) to 2.88° (full model).
- Continual pose estimation completes 100% of routes on KITTI-CVL and Ford-CVL, whereas PureACL single-camera covers 5.01% on F-1C and BoostAcc under 5% on KITTI-CVL cross.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add T2GA to PureACL (ablation) | PureACL · KITTI-CVL K-Cross (±45°, ±20m) | mean lateral error: 4.24 -> 1.29 | Helped |
| add T2GA to PureACL (ablation) | PureACL · KITTI-CVL K-Cross (±45°, ±20m) | mean longitudinal error: 3.89 -> 1.27 | Helped |
| add CycDA to T2GA (ablation) | T2GA variant · KITTI-CVL K-Cross (±45°, ±20m) | mean longitudinal error: 1.27 -> 0.12 | Helped |
| add ERP to T2GA + CycDA (ablation) | T2GA + CycDA variant · KITTI-CVL K-Cross (±45°, ±20m) | mean orientation error: 3.38 -> 2.88 | Helped |
| add T2GA, CycDA, and ERP | PureACL · KITTI-CVL K-Cross (±45°, ±20m) | mean lateral error: 4.24 -> 0.17 | Helped |
| use the full method instead of PureACL | KITTI-CVL K-Cross · Accumulated pose estimation | running percentage: 10.31 -> 100.00 | Helped |
Surprising findings
- Pixel-wise baseline PureACL shows large gap between median and mean errors under high initial noise, indicating convergence failures that the proposed modules reduce.
Limitations
- Relies on the assumption that ground and satellite views are orthogonal, which may fail on uphill/downhill or turns (partially tested with affine warping).
- Still depends on loop closure for long routes; future SLAM integration planned to remove this dependency.
Experimental setup
Datasets: KITTI-CVL Ford Multi-AV Seasonal (FMAVS) Ford-CVL Google Maps satellite images
Benchmarks: KITTI-CVL Ford-CVL
Metrics: mean lateral error mean longitudinal error mean orientation error recall at 0.25m/0.5m/1m orientation recall at 1/2/4 degrees running percentage in accumulated pose estimation
Baselines: PureACL HighlyAcc SliceMatch BoostAcc CCVPE HC-Net DSM SIBCL CVGL
Builds on: PureACL HighlyAcc U-Net
Closest papers by meaning
- Where Am I Looking At? Joint Location and Orientation Estimation by Cross-View Matching (CVPR 2020 · similarity 0.91)
- FG^2: Fine-Grained Cross-View Localization by Fine-Grained Feature Matching (CVPR 2025 · similarity 0.88)
- From Coarse to Fine: A Matching and Alignment Framework for Unsupervised Cross-View Geo-Localization (AAAI 2025 · similarity 0.88)
- BevSplat: Resolving Height Ambiguity via Feature-Based Gaussian Primitives for Weakly-Supervised Cross-View Localization (NeurIPS 2025 · 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 0.87)
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.