Research map / 3D vision / Depth and stereo
AutoCompass: Accurate Visual Localization on Public Maps by Learning from Weak Labels
Javier Tirado-Garín, Alan Savio Paul, Shuai Chen, Axel Barroso-Laguna, Tommaso Cavallari, Daniyar Turmukhambetov, Victor Adrian Prisacariu, Eric Brachmann
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 |
|---|---|---|---|
| train with raw GPS and chunk marginalization instead of strong geo-referenced 3-DoF labels | OrienterNet (ResNet-101) · KITTI (no orientation prior) | lateral recall@1m: 37.7 -> 43.1 | Improved |
| add relative-pose supervision instead of raw GPS only | AutoCompass (ResNet-101) · KITTI (no orientation prior) | lateral recall@1m: 43.1 -> 56.6 | Improved |
| swap ResNet-101 for DINOv2-B backbone | AutoCompass w/ rel. poses · KITTI (no orientation prior) | lateral recall@1m: 56.6 -> 70.8 | Improved |
| add relative-pose supervision instead of raw GPS only | AutoCompass (ResNet-101) · LaMAria | XY recall@5m: 26.3 -> 33.0 | Improved |
| add relative-pose supervision instead of raw GPS only | AutoCompass (ResNet-101) · Oxford Day-and-Night | XY recall@5m: 52.2 -> 56.4 | Improved |
| train with raw GPS instead of strong supervision on ODN | OrienterNet (ResNet-101) · Oxford Day-and-Night | XY recall@1m: 9.8 -> 10.4 | No clear effect |
| add GPS neighborhood marginalization (r=5 m vs r=0) (ablation) | AutoCompass GPS · Average over benchmarks | XY recall@5m: 37.1 -> 44.4 | Improved |
| increase chunk radius from 20 m to 30 m (ablation) | AutoCompass GPS · Average over benchmarks | XY recall@5m: 45.0 -> 37.6 | Hurt |
| vary norm binning size from 1 to 10 m (ablation) | AutoCompass dTheta+norm · Average over benchmarks | XY recall@5m: 46.5 -> 44.9 | No clear effect |
| remove GPS chunk loss from combined supervision (ablation) | AutoCompass · Average over benchmarks | XY recall@5m: 48.1 -> 50.8 | Improved |
| fuse single-view estimates over under 10 frames | AutoCompass · LaMAria (sequential) | XY recall@5m: 33.0 -> 92.6 | Improved |
| add relative rotation loss to chunk GPS (ablation) | AutoCompass (ResNet-101) · LaMAria | orientation recall@5deg: 28.7 -> 31.9 | Improved |
Closest papers by meaning
- Geometry-Aware Learning of Maps for Camera Localization (CVPR 2018 · similarity 0.86)
- VectorReLoc: Reliable Vectorized SD Map Visual Re-localization with Contrastive Feature Alignment (ECCV 2026 · similarity 0.86)
- Weakly-supervised Camera Localization by Ground-to-satellite Image Registration (ECCV 2024 · similarity 0.85)
- SNAP: Self-Supervised Neural Maps for Visual Positioning and Semantic Understanding (NeurIPS 2023 · similarity 0.85)
- GeoDistill: Geometry-Guided Self-Distillation for Weakly Supervised Cross-View Localization (ICCV 2025 · similarity 0.85)
- Change-Robust Online Topological Memory for Long-Term Relocalization and Semantic Navigation (NeurIPS 2026 · 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.
