Research map / 3D vision / Depth and stereo
R-SCoRe: Revisiting Scene Coordinate Regression for Robust Large-Scale Visual 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.
Revisits scene coordinate regression (SCR) for large-scale visual localization, adding covisibility-graph-based global encodings and augmentation, a depth-adjusted reprojection loss, and a refinement module, achieving strong night-time accuracy with small maps.
Problem
Scene coordinate regression offers small map sizes for visual localization but is less robust than feature matching under complex illumination changes and image-level ambiguities in large-scale scenes.
Approach
Learn image-level global encodings aligned with the covisibility graph via Node2Vec, augment training by swapping global encodings with covisible neighbors, use multiple retrieved global hypotheses at test time with RANSAC-based selection, and apply a depth-adjusted reprojection loss to counter bias against near points.
Key results
- On Aachen Day-Night nighttime queries, R-SCoRe reaches 64.3% at (0.25m, 2°) with a 47MB map, about 10x more accurate than prior SCR methods without depth supervision (ACE, GLACE).
- On Hyundai Department Store B1 test, R-SCoRe (DeDoDe) reaches 60.1% at (0.1m, 1°) versus 2.7% for ACE (x50) and 0.9% for GLACE.
- Ablation on Dept. 1F val: multiple hypotheses, covisibility augmentation, and covisibility encoding raise (0.1m, 1°) accuracy from 34.1% (R2Former with Gaussian noise) to 72.3%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace Gaussian noise with covisibility-based data augmentation (ablation) | R-SCoRe global encoding · Hyundai Department Store Dept. 1F Val | accuracy (0.1m, 1°): 42.1 -> 62.0 | Helped |
| learn global encoding directly from the covisibility graph (ablation) | R-SCoRe · Hyundai Department Store Dept. 1F Val | accuracy (0.1m, 1°): 62.0 -> 72.3 | Helped |
| add multiple global hypotheses at test time (ablation) | R-SCoRe with R2Former global encoding and Gaussian noise · Hyundai Department Store Dept. 1F Val | accuracy (0.1m, 1°): 34.1 -> 42.1 | Helped |
| use Dedode instead of LoFTR local encoding (ablation) | R-SCoRe · Hyundai Department Store Dept. B1 Val | accuracy (0.1m, 1°): 29.4 -> 57.7 | Helped |
| use off-the-shelf local feature extractors instead of the ACE backbone (ablation) | R-SCoRe · Hyundai Department Store Dept. 1F Val | accuracy (0.1m, 1°): 68.7 -> 70.6 | Helped |
| build the covisibility graph using feature matching instead of pose-based frustum overlap (ablation) | R-SCoRe · Aachen Day | accuracy (0.25m, 2°): 74.8 -> 75.4 | No clear effect |
| use depth-adjusted reprojection error supervision instead of original reprojection error supervision (ablation) | R-SCoRe · Hyundai Department Store Dept. 1F Val | accuracy (0.1m, 1°): 62.3 -> 70.6 | Helped |
| use depth-adjusted reprojection error supervision instead of original reprojection error supervision (ablation) | R-SCoRe · Hyundai Department Store Dept. 4F Val | accuracy (0.1m, 1°): 59.1 -> 63.9 | Helped |
| use ground-truth depth supervision instead of depth-adjusted reprojection error supervision (ablation) | R-SCoRe · Hyundai Department Store Dept. 1F Val | accuracy (0.1m, 1°): 70.6 -> 76.8 | Helped |
| use ground-truth depth supervision instead of depth-adjusted reprojection error supervision (ablation) | R-SCoRe · Hyundai Department Store Dept. 4F Val | accuracy (0.1m, 1°): 63.9 -> 68.5 | Helped |
Surprising findings
- Off-the-shelf feature extractors (DeDoDe, LoFTR) outperform the originally used ACE-trained backbone, contrary to prior investigations.
- Using a more accurate covisibility graph from feature matching yields no significant improvement over the simple frustum-overlap graph.
Limitations
- Still trails feature matching methods at the strictest error thresholds and highest accuracy.
- Out-of-distribution generalization remains challenging.
- Optional depth supervision gives gains but the main method avoids it; depth-free performance is somewhat lower.
Experimental setup
Datasets: Aachen Day-Night Hyundai Department Store
Benchmarks: Aachen Day-Night Hyundai Department Store
Metrics: accuracy at (0.25m, 2°), (0.5m, 5°), (5m, 10°) thresholds accuracy at (0.1m, 1°), (0.25m, 2°), (1m, 5°) thresholds map size (MB)
Baselines: HLoc+SPSG AS Cascaded QP+R.Sift Squeezer PixLoc Neumap HSCNet HSCNet++ ESAC ACE GLACE
Builds on: GLACE ACE DSAC* Node2Vec NetVLAD
Closest papers by meaning
- GLACE: Global Local Accelerated Coordinate Encoding (CVPR 2024 · similarity 0.92)
- MutualVPR: A Mutual Learning Framework for Resolving Supervision Inconsistencies via Adaptive Clustering (NeurIPS 2025 · similarity 0.87)
- Enhancing Visual Localization with Cross-Domain Image Generation (ICML 2025 · similarity 0.86)
- Focus on Local: Finding Reliable Discriminative Regions for Visual Place Recognition (AAAI 2025 · similarity 0.86)
- SuperVLAD: Compact and Robust Image Descriptors for Visual Place Recognition (NeurIPS 2024 · similarity 0.86)
- LiSA: LiDAR Localization with Semantic Awareness (CVPR 2024 · similarity 0.86)
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.