Research map / 3D vision / Depth and stereo
Weakly-supervised Camera Localization by Ground-to-satellite Image Registration
Yujiao Shi, HONGDONG LI, Akhil Perincherry, Ankit Vora
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace pose regressor with spatial correlation for translation estimation (ablation) | weakly supervised G2S localization · KITTI Test-2 lateral d=1 | % within 1m: 5.29 -> 45.11 (lateral longitudinal row Grd Reg. 5.04 -> 45.11) | Improved |
| add confidence-guided similarity matching (ablation) | weakly supervised G2S localization · KITTI Test-2 lateral d=1 | % within 1m: 45.11 -> 62.73 | Improved |
| add L3 loss when moderately accurate labels are available (lambda=1) (ablation) | weakly supervised G2S localization · KITTI Test-1 lateral d=1 | % within 1m: 59.58 -> 66.07 | Improved |
| use ground-plane homography instead of geometry-guided cross-view transformer (ablation) | overhead-view projection · KITTI Test-2 lateral d=1 (lambda=0) | % within 1m: 52.69 -> 62.73 | Improved |
| share feature extractor between satellite and pinhole ground images (ablation) | weakly supervised G2S localization · KITTI Test-2 lateral d=1 (lambda=0) | % within 1m: 54.49 -> 62.73 | Improved |
| use separate feature extractors for rotation and translation stages (ablation) | weakly supervised G2S localization · KITTI Test-2 lateral d=1 (lambda=0) | % within 1m: 32.18 -> 62.73 | Improved |
| stage-wise instead of end-to-end training (ablation) | weakly supervised G2S localization · KITTI Test-2 lateral d=1 | % within 1m: 32.53 -> 62.73 | Improved |
| project features instead of images to overhead view (ablation) | weakly supervised G2S localization · KITTI Test-2 lateral d=1 (lambda=0) | % within 1m: 48.24 -> 62.73 | Improved |
| regress angle directly instead of rotation matrix or sin/cos (ablation) | rotation estimator · KITTI Test-2 azimuth 1 degree | % within 1 degree: 5.07 -> 99.99 (sin/cos 86.52) | Improved |
| fine-tune with small fraction of accurately labeled data (semi-supervised) (ablation) | weakly supervised G2S localization · KITTI lateral d=1 | % within 1m: 62.73 -> 39.49 at 0.2 data; 62.73 -> 83.86 at full | Mixed |
| weak supervision vs full supervision in cross-area generalization | weakly supervised G2S localization · KITTI Test-2 lateral d=1 | % within 1m: 57.72 -> 64.74 (vs Shi et al.) | Improved |
| learn a confidence map for satellite images (ablation) | satellite branch · KITTI | localization accuracy | Hurt |
Closest papers by meaning
- Accurate 3-DoF Camera Geo-Localization via Ground-to-Satellite Image Matching (TPAMI 2022 · similarity 0.88)
- Boosting 3-DoF Ground-to-Satellite Camera Localization Accuracy via Geometry-Guided Cross-View Transformer (ICCV 2023 · similarity 0.88)
- BevSplat: Resolving Height Ambiguity via Feature-Based Gaussian Primitives for Weakly-Supervised Cross-View Localization (NeurIPS 2025 · similarity 0.87)
- Lending Orientation to Neural Networks for Cross-View Geo-Localization (CVPR 2019 · similarity 0.87)
- Unleashing Unlabeled Data: A Paradigm for Cross-View Geo-Localization (CVPR 2024 · similarity 0.85)
- From Coarse to Fine: A Matching and Alignment Framework for Unsupervised Cross-View Geo-Localization (AAAI 2025 · similarity 0.84)
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.
