Research map / 3D vision / Gaussian splatting
VCR-GauS: View Consistent Depth-Normal Regularizer for Gaussian Surface 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 VCR-GauS, a 3D Gaussian Splatting surface reconstruction method that uses a depth-normal (D-Normal) regularizer coupling normals to Gaussian positions, a view-consistency confidence weight on monocular normal priors, and a new densify-and-split strategy.
Problem
Extracting high-quality surfaces from 3D Gaussian Splatting is hard for large scenes; normal supervision from rendered Gaussian normals mainly updates rotations, and monocular normal priors are inconsistent across views.
Approach
Derive a depth from the ray-plane intersection of flattened Gaussians so that normals (via back-projected depth finite differences) supervise all geometric parameters, weight the loss with a confidence term from the cosine agreement between rendered and predicted normals, and densify/split large Gaussians to reduce depth error.
Key results
- On Tanks and Temples, mean F1 of 0.40 versus 0.30 for 2DGS (0.36 in the ablation's w/o-confidence variant) and 0.09 for 3DGS, with ~53 min optimization and 145 FPS.
- On Replica, F1 of 78.17 versus 64.36 for 2DGS and 50.79 for 3DGS, close to MonoSDF's 81.64.
- On Mip-NeRF 360 it reaches 128 FPS with PSNR/SSIM/LPIPS comparable to other Gaussian methods (indoor PSNR 30.53 vs 30.41 for 3DGS).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add D-Normal regularization while retaining rendered-normal supervision (ablation) | VCR-GauS · Tanks and Temples | F-score: 0.30 -> 0.40 | Helped |
| add confidence weighting to the D-Normal regularizer (ablation) | VCR-GauS · Tanks and Temples | F-score: 0.36 -> 0.40 | Helped |
| use intersection depth calculation (ablation) | VCR-GauS · Tanks and Temples | F-score: 0.35 -> 0.40 | Helped |
| add the proposed densification and splitting strategy (ablation) | VCR-GauS · Tanks and Temples | F-score: 0.33 -> 0.40 | Helped |
| add semantic surface trimming (ablation) | VCR-GauS · Tanks and Temples | F-score: 0.38 -> 0.40 | Helped |
| add scale regularization (ablation) | VCR-GauS · Tanks and Temples | F-score: 0.36 -> 0.40 | Helped |
| use scale regularization instead of setting scale s3 to zero (ablation) | VCR-GauS · Tanks and Temples | F-score: 0.37 -> 0.40 | Helped |
Surprising findings
- Setting the minimum scale to zero (2DGS-style) gives 0.37 F1 versus 0.40 for scale regularization, and removing the D-Normal regularizer drops F1 from 0.40 to 0.30.
Limitations
- Fails when almost all predicted normals across views are wrong.
- Cannot reconstruct beyond the observed scene.
- Cannot capture semi-transparent surfaces such as windows.
Experimental setup
Datasets: Tanks and Temples Replica Mip-NeRF 360 DTU COLMAP sparse point clouds
Benchmarks: Tanks and Temples Replica Mip-NeRF 360 DTU
Metrics: F1-score Chamfer distance PSNR SSIM LPIPS FPS optimization time
Baselines: NeuS MonoSDF Geo-NeuS SuGaR 3DGS 2DGS Neuralangelo DN-Splatter Mip-NeRF360 BakedSDF
Builds on: 3D Gaussian Splatting NeuSG Mip-Splatting DSINE GeoWizard
Closest papers by meaning
- Normal-GS: 3D Gaussian Splatting with Normal-Involved Rendering (NeurIPS 2024 · similarity 0.91)
- VA-GS: Enhancing the Geometric Representation of Gaussian Splatting via View Alignment (NeurIPS 2025 · similarity 0.90)
- Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization (CVPR 2025 · similarity 0.89)
- Effective Rank Analysis and Regularization for Enhanced 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.89)
- DNGaussian: Optimizing Sparse-View 3D Gaussian Radiance Fields with Global-Local Depth Normalization (CVPR 2024 · similarity 0.88)
- GaussianPainter: Painting Point Cloud into 3D Gaussians with Normal Guidance (AAAI 2025 · similarity 0.88)
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.