Research map / 3D vision / Gaussian splatting
From Sparse to Dense: Camera Relocalization with Scene-Specific Detector from Feature Gaussian Splatting
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 STDLoc, a full camera relocalization pipeline that renders Feature Gaussian feature fields, samples matching-oriented scene landmarks, detects them with a self-supervised scene-specific CNN detector, and refines poses via sparse-to-dense feature matching.
Problem
Camera relocalization without pose priors in indoor and outdoor scenes, where feature matching struggles in weak texture and SCR methods struggle at large outdoor scale, and radiance-field methods lack full pipelines or rely on photometric refinement.
Approach
Use Feature Gaussian (3DGS with distilled local feature field) as scene representation; select evenly distributed, multi-view-consistent landmarks via a matching score; detect query keypoints with a scene-specific CNN detector for sparse PnP initialization; then refine by coarse-to-fine dense feature matching against rendered feature maps instead of photometric alignment.
Key results
- On 7-Scenes, STDLoc reports average median error 0.76cm/0.24° and recall 99.1% at 5cm5° and 90.9% at 2cm2°.
- On Cambridge Landmarks, STDLoc achieves average median error 10.1cm/0.14°, outperforming all compared methods on average.
- Sparse stage alone reaches 99.6% recall at 5m10° on Cambridge Landmarks versus 48.2% for image retrieval.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use STDLoc instead of DSAC* | 7-Scenes localization · 7-Scenes | median translation error (cm): 1.07 -> 0.76 | Helped |
| Use STDLoc instead of ACE | 7-Scenes localization · 7-Scenes | median rotation error (°): 0.34 -> 0.24 | Helped |
| Use STDLoc instead of HLoc (SP+SG) | Cambridge Landmarks localization · Cambridge Landmarks | median translation error (cm): 11.0 -> 10.1 | Helped |
| Use STDLoc instead of HLoc (SP+SG) | Cambridge Landmarks localization · Cambridge Landmarks | median rotation error (°): 0.21 -> 0.14 | Helped |
| Add matching-oriented sampling to FPS (ablation) | FPS landmark sampling with SuperPoint · Cambridge Landmarks | recall at 5m 10°: 57.9% -> 90.5% | Helped |
| Add matching-oriented sampling to random sampling (ablation) | Random landmark sampling with SuperPoint · Cambridge Landmarks | recall at 5m 10°: 95.0% -> 98.2% | Helped |
| Replace SuperPoint with the scene-specific detector (ablation) | Random landmark sampling · Cambridge Landmarks | recall at 5m 10°: 95.0% -> 99.0% | Helped |
| Use matching-oriented sampling and the scene-specific detector (ablation) | STDLoc sparse stage · Cambridge Landmarks | recall at 5m 10°: 95.0% -> 99.6% | Helped |
| Replace image retrieval with the sparse localization stage (ablation) | STDLoc pipeline · Cambridge Landmarks | average median translation error (cm): 586 -> 13.8 | Helped |
| Add dense feature-map refinement to the sparse stage (ablation) | STDLoc sparse localization · Cambridge Landmarks | average median translation error (cm): 13.8 -> 10.1 | Helped |
| Add dense RGB-rendering refinement to the sparse stage (ablation) | STDLoc sparse localization · Cambridge Landmarks | average median translation error (cm): 13.8 -> 15.3 | Hurt |
| Replace 3DGS with 2DGS (ablation) | STDLoc with SuperPoint · Cambridge Landmarks | average median translation error (cm): 10.1 -> 10.4 | Hurt |
Surprising findings
- Rendering RGB images for dense matching performs worse (15.3cm) than directly rendering the feature map (10.1cm), attributed to low-quality RGB renders under illumination changes.
Limitations
- Feature Gaussian training limits scalability to very large scenes.
- Floating artifacts in 3DGS can degrade localization.
Experimental setup
Datasets: 7-Scenes Cambridge Landmarks
Benchmarks: 7-Scenes Cambridge Landmarks
Metrics: median translation error (cm) median rotation error (deg) recall at 5cm 5deg recall at 2cm 2deg recall at 5m 10deg
Baselines: HLoc (SP+SG) AS (SIFT) DSAC* ACE GLACE NeuMap CROSSFIRE PNeRFLoc NeRFMatch
Builds on: 3D Gaussian Splatting Feature 3DGS LoFTR HLoc SuperPoint
Closest papers by meaning
- Hierarchical Visual Relocalization with Nearest View Synthesis from Feature Gaussian Splatting (CVPR 2026 · similarity 0.91)
- Gaussian Splatting Feature Fields for (Privacy-Preserving) Visual Localization (CVPR 2025 · similarity 0.85)
- Reloc3r: Large-Scale Training of Relative Camera Pose Regression for Generalizable, Fast, and Accurate Visual Localization (CVPR 2025 · similarity 0.85)
- LightLoc: Learning Outdoor LiDAR Localization at Light Speed (CVPR 2025 · similarity 0.84)
- GS-SLAM: Dense Visual SLAM with 3D Gaussian Splatting (CVPR 2024 · similarity 0.84)
- GS-CPR: Efficient Camera Pose Refinement via 3D Gaussian Splatting (ICLR 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.