Research map / 3D vision / Depth and stereo
PlanarSplatting: Accurate Planar Surface Reconstruction in 3 Minutes
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.
Optimizes explicit 3D rectangular plane primitives by splatting them into depth and normal maps supervised by monocular foundation-model cues, reconstructing indoor scenes in about 3 minutes.
Problem
Multi-view indoor planar surface reconstruction has relied on 2D/3D plane detection, matching and tracking, and learned methods need plane annotations; existing pipelines are slow and coarse.
Approach
Directly optimize learnable 3D rectangular plane primitives (center, quaternion rotation, double-direction radii) with a rectangle-based plane splatting function that renders depth and normal maps, supervised by pretrained monocular depth/normal models (Metric3Dv2, Omnidata), with no plane annotations and no plane detection or matching; planes are merged after optimization.
Key results
- On ScanNetV2, Ours@Metric3Dv2 achieves Chamfer 4.23 and F-score 73.76, versus 5.30 Chamfer and 64.92 F-score for AirPlanes (which uses plane annotations).
- On ScanNet++, Ours@Metric3Dv2 reaches F-score 50.68 and Chamfer 9.01 versus 32.58 F-score and 13.75 Chamfer for AirPlanes.
- Initializing 3DGS from PlanarSplatting planes raises ScanNetV2 PSNR from 24.417 to 25.471 with total time 5.6 min versus 12.2 min for 3DGS alone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PlanarSplatting with Metric3Dv2 supervision instead of AirPlanes | Planar reconstruction · ScanNetV2 test | Chamfer: 5.30 -> 4.23 | Helped |
| Use PlanarSplatting with Metric3Dv2 supervision instead of AirPlanes | Planar reconstruction · ScanNetV2 test | F-score: 64.92 -> 73.76 | Helped |
| Use PlanarSplatting with Metric3Dv2 supervision instead of AirPlanes | Planar reconstruction · ScanNet++ | Chamfer: 13.75 -> 9.01 | Helped |
| Use PlanarSplatting with Metric3Dv2 supervision instead of AirPlanes | Planar reconstruction · ScanNet++ | F-score: 32.58 -> 50.68 | Helped |
| Initialize 3DGS from PlanarSplatting, fixing point positions and omitting densification | 3DGS · ScanNetV2 | PSNR: 24.417 -> 25.471 | Helped |
| Initialize 2DGS from PlanarSplatting, fixing point positions and omitting densification | 2DGS · ScanNetV2 | PSNR: 24.766 -> 25.380 | Helped |
| Initialize from a sphere instead of Metric3Dv2 (ablation) | PlanarSplatting · ScanNetV2 ablation scenes | Chamfer: 4.66 -> 8.39 | Hurt |
| Initialize from a sphere instead of Metric3Dv2 and optimize for 30K iterations (ablation) | PlanarSplatting · ScanNetV2 ablation scenes | Chamfer: 4.66 -> 4.58 | No clear effect |
| Add Double Radii instead of using single-direction radii (ablation) | PlanarSplatting · ScanNetV2 ablation scenes | Chamfer: 4.73 -> 4.66 | Helped |
| Add Plane Splitting (ablation) | PlanarSplatting · ScanNetV2 ablation scenes | Chamfer: 4.69 -> 4.66 | Helped |
Surprising findings
- Plane-based initialization with fixed point positions and no densification gives better novel view synthesis than SfM-initialized 3DGS/2DGS with densification, with fewer Gaussians (0.37M vs 1.27M).
Limitations
- Not suitable for complex or curved surfaces, since representation is limited to planar primitives.
Experimental setup
Datasets: ScanNetV2 ScanNet++
Benchmarks: ScanNetV2 ScanNet++
Metrics: Chamfer distance F-score VOI RI SC Planar Fidelity Planar Accuracy Planar Chamfer PSNR SSIM
Baselines: PlanarRecon AirPlanes 2DGS+RANSAC Metric3Dv2+RANSAC SR+RANSAC 3DGS 2DGS
Builds on: 3D Gaussian Splatting 2D Gaussian Splatting RANSAC
Closest papers by meaning
- PlanarGS: High-Fidelity Indoor 3D Gaussian Splatting Guided by Vision-Language Planar Priors (NeurIPS 2025 · similarity 0.90)
- PLANA3R: Zero-shot Metric Planar 3D Reconstruction via Feed-forward Planar Splatting (NeurIPS 2025 · similarity 0.90)
- AlphaTablets: A Generic Plane Representation for 3D Planar Reconstruction from Monocular Videos (NeurIPS 2024 · similarity 0.87)
- IndoorGS: Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction (CVPR 2025 · similarity 0.86)
- FatesGS: Fast and Accurate Sparse-View Surface Reconstruction Using Gaussian Splatting with Depth-Feature Consistency (AAAI 2025 · similarity 0.86)
- Behind the Veil: Enhanced Indoor 3D Scene Reconstruction with Occluded Surfaces Completion (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.