Research map / 3D vision / Gaussian splatting
GS-CPR: Efficient Camera Pose Refinement via 3D 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 GS-CPR, a one-shot test-time camera pose refinement method that renders RGB and depth from a 3DGS scene model, matches the query to the render with MASt3R, and refines the pose via PnP+RANSAC on 2D-3D correspondences.
Problem
Improving the accuracy of initial camera pose estimates from absolute pose regression (APR) and scene coordinate regression (SCR) methods for visual relocalization, while avoiding the slow iterative optimization and trained feature descriptors of prior NeRF-based refinement methods.
Approach
Render images and depth from a 3DGS scene at the initial pose, use the pretrained MASt3R matcher for direct RGB 2D-2D matching (no trained descriptors), lift matches to 3D via rendered depth, and solve a single PnP+RANSAC refinement; an exposure-adaptive affine color transform handles query-render appearance shifts.
Key results
- On 7Scenes, ACE + GS-CPR reaches 100% recall at 5cm/5deg and 0.8cm/0.25deg average median error, versus 97.1% and 1.1cm/0.34deg for ACE alone.
- On Cambridge Landmarks, DFNet + GS-CPR reduces average median error from 119cm/2.9deg to 26cm/0.51deg, outperforming NeFeS with 50 iterations (35cm/0.77deg).
- Runtime averages 0.19 s per query for ACE + GS-CPR on Cambridge versus 2.4 s for MCLoc and 10 s for DFNet + NeFeS50, on an RTX 4090.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| refine DFNet pose with GS-CPR | DFNet · 7Scenes | median translation/rotation error (cm/°): 6/1.93 -> 1.1/0.34 | Helped |
| refine Marepo pose with GS-CPR | Marepo · 7Scenes | median translation/rotation error (cm/°): 2.9/1.04 -> 0.9/0.28 | Helped |
| refine ACE pose with GS-CPR | ACE · 7Scenes | median translation/rotation error (cm/°): 1.1/0.34 -> 0.8/0.25 | Helped |
| refine DFNet pose with GS-CPR | DFNet · 7Scenes | accuracy at [5cm,5°]: 43.1 -> 94.2 | Helped |
| refine Marepo pose with GS-CPR | Marepo · 12Scenes | accuracy at [2cm,2°]: 50.4 -> 90.9 | Helped |
| refine ACE pose with GS-CPR | ACE · 12Scenes | median translation/rotation error (cm/°): 0.7/0.26 -> 0.5/0.21 | Helped |
| refine DFNet pose with GS-CPR | DFNet · Cambridge Landmarks | median translation/rotation error (cm/°): 119/2.9 -> 26/0.51 | Helped |
| refine ACE pose with GS-CPR-rel | ACE · Cambridge Landmarks | median translation/rotation error (cm/°): 21/0.47 -> 47/0.54 | Hurt |
| replace MASt3R matcher with LoFTR (ablation) | Marepo + GS-CPR · 7Scenes | median translation/rotation error (cm/°): 0.9/0.28 -> 1.5/0.40 | Hurt |
| replace MASt3R matcher with DUSt3R (ablation) | Marepo + GS-CPR · 7Scenes | median translation/rotation error (cm/°): 0.9/0.28 -> 2.1/0.7 | Hurt |
| replace MASt3R matcher with LoFTR (ablation) | ACE + GS-CPR · 7Scenes | median translation/rotation error (cm/°): 0.8/0.25 -> 1.0/0.31 | Hurt |
| replace MASt3R matcher with DUSt3R (ablation) | ACE + GS-CPR · 7Scenes | median translation/rotation error (cm/°): 1.1/0.34 -> 1.5/0.6 | Hurt |
Surprising findings
- The relative-pose variant GS-CPR-rel degrades ACE accuracy (79.9% vs 97.1% recall on 7Scenes), since MASt3R relative pose is less accurate when the initial pose is already close to GT.
Limitations
- Depends on accuracy of initial pose; fails when initial pose is too far from GT (rotation error above about 50 degrees or translation above about 8 m in perturbation tests).
- Reconstructing high-quality 3DGS for large scenes remains challenging; only small-to-medium benchmarks evaluated.
- Relative-pose variant struggles to refine accurate SCR poses.
Experimental setup
Datasets: 7Scenes 12Scenes Cambridge Landmarks
Benchmarks: 7Scenes 12Scenes Cambridge Landmarks
Metrics: median translation error (cm) median rotation error (degrees) recall at 5cm/5deg recall at 2cm/2deg runtime (s)
Baselines: NeFeS CrossFire HR-APR MCLoc NeRFMatch DFNet Marepo ACE
Builds on: Scaffold-GS MASt3R ACE NeFeS DFNet
Closest papers by meaning
- GS-SLAM: Dense Visual SLAM with 3D Gaussian Splatting (CVPR 2024 · similarity 0.88)
- KeyGS: A Keyframe-Centric Gaussian Splatting Method for Monocular Image Sequences (AAAI 2025 · similarity 0.88)
- Gaussian Splatting Feature Fields for (Privacy-Preserving) Visual Localization (CVPR 2025 · similarity 0.87)
- DG-SLAM: Robust Dynamic Gaussian Splatting SLAM with Hybrid Pose Optimization (NeurIPS 2024 · similarity 0.86)
- Gaussian Splatting SLAM (CVPR 2024 · similarity 0.86)
- PUP 3D-GS: Principled Uncertainty Pruning for 3D Gaussian Splatting (CVPR 2025 · 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.