Research map / 3D vision / Gaussian splatting
IndoorGS: Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction
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 IndoorGS, a 3DGS-based indoor reconstruction method that initializes and guides Gaussian densification with 3D line cues, filtered SfM points, and plane-like cues, plus tangent-space densification for line Gaussians.
Problem
3D Gaussian Splatting performs poorly on indoor scenes because textureless regions yield unreliable SfM seeds and weak geometric constraints, leading to local optima, holes, and floaters in geometry and rendering.
Approach
Extracts three geometric cue types (3D line segments matched from RGB and normal maps via DeepLSD/GlueStick, SOR-filtered SfM points, and plane-like points back-projected from SAM-segmented weakly textured regions with Metric3D depth), uses them for initialization, and applies geometric-cue-guided ADC where line Gaussians get tangent-space densification and plane-like Gaussians are not split or cl
Key results
- On mesh reconstruction, IndoorGS averages Chamfer-L1 0.038, F1 0.801, NC 0.909 across Replica and ScanNet++, versus 0.064 / 0.722 / 0.837 for PGSR.
- Average PSNR across Replica, ScanNet++ and Deep Blending is 33.88 dB, 0.35 dB above the second-best method (RaDeGS at 33.53).
- Ablation on ScanNet++ shows line-cue-guided ADC has the largest impact among the ADC strategies, and normal regularization gives the largest gain among losses.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add line, plane-like, and filtered SfM cues (ablation) | IndoorGS · Scannet++ | PSNR: 30.308 -> 30.453 | Helped |
| add line and plane-like cues (ablation) | IndoorGS with filtered SfM cues · Scannet++ | PSNR: 30.513 -> 30.453 | No clear effect |
| add plane-like cues (ablation) | IndoorGS with line and SfM cues · Scannet++ | PSNR: 30.583 -> 30.453 | No clear effect |
| add customized adaptive density control for all cue types (ablation) | IndoorGS · Scannet++ | PSNR: 30.499 -> 30.453 | No clear effect |
| add line-cue-guided adaptive density control (ablation) | IndoorGS · Scannet++ | PSNR: 30.595 -> 30.453 | No clear effect |
| add SfM-cue-guided adaptive density control (ablation) | IndoorGS · Scannet++ | PSNR: 30.573 -> 30.453 | No clear effect |
| add plane-cue-guided adaptive density control (ablation) | IndoorGS · Scannet++ | PSNR: 30.390 -> 30.453 | Helped |
| add monocular depth regularization loss (ablation) | IndoorGS · Scannet++ | PSNR: 30.553 -> 30.453 | No clear effect |
| add monocular normal regularization loss (ablation) | IndoorGS · Scannet++ | PSNR: 30.510 -> 30.453 | No clear effect |
| add normal consistency loss (ablation) | IndoorGS · Scannet++ | PSNR: 30.505 -> 30.453 | No clear effect |
Surprising findings
- Test views sampled sequentially from the original datasets are close to training views, so high PSNR can coexist with inaccurate geometry, undermining the test set's ability to assess true capability.
Limitations
- Test split is sequential and close to training views, limiting evaluation of generalization to diverse viewpoints.
- Relies on several external pretrained models (SAM, Metric3D, DeepLSD, GlueStick) whose errors propagate into cues.
- Plane-like cues depend on thresholds and monocular depth alignment, and plane-like Gaussians are not split or cloned.
Experimental setup
Datasets: Replica ScanNet++ Deep Blending
Benchmarks: Replica ScanNet++ Deep Blending
Metrics: PSNR SSIM LPIPS Chamfer-L1 F-score Normal Consistency
Baselines: 3DGS GOF RaDeGS PGSR
Builds on: 3D Gaussian Splatting RaDeGS LIMAP GeoGS PGSR
Closest papers by meaning
- PlanarGS: High-Fidelity Indoor 3D Gaussian Splatting Guided by Vision-Language Planar Priors (NeurIPS 2025 · similarity 0.93)
- AtlasGS: Atlanta-world Guided Surface Reconstruction with Implicit Structured Gaussians (NeurIPS 2025 · similarity 0.91)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.91)
- SfM-Free 3D Gaussian Splatting via Hierarchical Training (CVPR 2025 · similarity 0.90)
- HybridGS: Decoupling Transients and Statics with 2D and 3D Gaussian Splatting (CVPR 2025 · similarity 0.90)
- GS-2DGS: Geometrically Supervised 2DGS for Reflective Object Reconstruction (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.