Research map / 3D vision / Gaussian splatting
PlanarGS: High-Fidelity Indoor 3D Gaussian Splatting Guided by Vision-Language Planar Priors
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 PlanarGS, a 3DGS framework that uses language-prompted planar priors from GroundedSAM (refined by DUSt3R-based cross-view fusion and geometric inspection) plus depth/normal priors and co-planarity losses for indoor surface reconstruction.
Problem
3D Gaussian Splatting trained with photometric loss yields ambiguous geometry and inaccurate surfaces in indoor scenes dominated by large, low-texture planar regions.
Approach
Use a pretrained vision-language segmentation model prompted with indoor class words, refined by multi-view DUSt3R depth fusion and normal/plane-distance geometric inspection, to obtain planar priors that drive a co-planarity constraint, Gaussian flattening, and depth/normal priors in 3DGS optimization.
Key results
- On MuSHRoom, PlanarGS reaches CD 4.49 cm and F1 77.14 versus 61.86 F1 for the best prior baseline DN-Splatter.
- On ScanNet++, PlanarGS achieves CD 3.66 cm and F1 82.78 compared with 4.16 cm and 75.86 for DN-Splatter.
- On Replica, PlanarGS achieves CD 4.13 cm and F1 81.90 versus 5.60 cm and 68.12 for DN-Splatter.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PlanarGS instead of DN-Splatter | Indoor surface reconstruction · MuSHRoom | CD: 5.77 -> 4.49 | Helped |
| Use PlanarGS instead of DN-Splatter | Indoor surface reconstruction · ScanNet++ | CD: 4.16 -> 3.66 | Helped |
| Use PlanarGS instead of DN-Splatter | Indoor surface reconstruction · Replica | CD: 5.60 -> 4.13 | Helped |
| Use LP3 planar priors instead of ZeroPlane (ablation) | PlanarGS · MuSHRoom coffee room | CD: 5.25 -> 2.98 | Helped |
| Use LP3 planar priors instead of GroundedSAM without LP3 (ablation) | PlanarGS · MuSHRoom coffee room | CD: 3.43 -> 2.98 | Helped |
| Add the co-planarity constraint (ablation) | PlanarGS · MuSHRoom coffee room | CD: 3.18 -> 2.98 | Helped |
| Add the geometric prior constraint (ablation) | PlanarGS · MuSHRoom coffee room | CD: 4.82 -> 2.98 | Helped |
| Add both co-planarity and geometric prior constraints (ablation) | PlanarGS · MuSHRoom coffee room | CD: 6.59 -> 2.98 | Helped |
| Remove the depth-normal consistency constraint (ablation) | PlanarGS · MuSHRoom coffee room | CD: 2.98 -> 2.87 | Helped |
| Replace DUSt3R with VGGT (ablation) | PlanarGS · Replica | CD: 4.13 -> 5.14 | Hurt |
| Replace GroundedSAM with YoloWorld+SAM (ablation) | PlanarGS · Replica | CD: 4.13 -> 4.24 | Hurt |
| Add “blackboard, sofa” prompts (ablation) | PlanarGS · Replica | CD: 4.13 -> 4.19 | Hurt |
Surprising findings
- Using a single-image-based zero-shot plane detector (ZeroPlane) as planar prior yields worse results than GroundedSAM without the LP3 pipeline in some settings, showing that prior quality dominates.
Limitations
- Planar priors only help detected planar regions; no benefit for curved walls or natural outdoor scenes.
- Depends on quality of pretrained foundation models (GroundedSAM, DUSt3R) and prompt choice.
Experimental setup
Datasets: Replica ScanNet++ MuSHRoom
Benchmarks: Replica ScanNet++ MuSHRoom
Metrics: Acc Comp CD F1 NC PSNR SSIM LPIPS
Baselines: 3DGS 2DGS GOF PGSR QGS DN-Splatter ZeroPlane FDS DUSt3R
Builds on: 3D Gaussian Splatting PGSR DN-Splatter GroundedSAM DUSt3R
Closest papers by meaning
- IndoorGS: Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction (CVPR 2025 · similarity 0.93)
- SceneSplat++: A Large Dataset and Comprehensive Benchmark for Language Gaussian Splatting (NeurIPS 2025 · similarity 0.92)
- AtlasGS: Atlanta-world Guided Surface Reconstruction with Implicit Structured Gaussians (NeurIPS 2025 · similarity 0.91)
- PlanarSplatting: Accurate Planar Surface Reconstruction in 3 Minutes (CVPR 2025 · similarity 0.90)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.90)
- HybridGS: Decoupling Transients and Statics with 2D and 3D Gaussian Splatting (CVPR 2025 · similarity 0.90)
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.