Research map / 3D vision / Gaussian splatting
Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization
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 CarGS, a unified anchor-based 3D Gaussian Splatting model that jointly achieves high-quality rendering and surface reconstruction by learning per-primitive adaptive covariance contributions via a residual Lite-Geo MLP and geometry-guided densification.
Problem
Simultaneously achieving SOTA rendering and surface reconstruction with one 3DGS model is hard because geometry regularization conflicts with rendering contributions of Gaussian primitives, while dual-model approaches are costly in storage and compute.
Approach
Decouple geometry and rendering contributions by predicting a residual geometry covariance offset via a lightweight Geo-MLP initialized from the rendering covariance MLP (Lite-Geo), and densify Gaussians using both normal and SDF cues.
Key results
- On Tanks and Temples, CarGS reaches mean F1 0.50 (tied with PGSR and NeuralAngelo at 0.50) and mean PSNR 26.01, the best PSNR among compared methods.
- On Mip-NeRF360, CarGS achieves average PSNR 27.68 and SSIM 0.834, the best average among compared methods, with LPIPS 0.209 vs PGSR 0.193.
- On the Truck scene CarGS uses 0.16GB storage versus 0.42GB for PGSR and trains in 1.2h with 90 FPS, versus GSDF's 3h training and 1.7GB storage.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CarGS instead of PGSR | Tanks and Temples · Tanks and Temples (mean) | PSNR: 24.45 -> 26.01 | Helped |
| use CarGS instead of PGSR | Tanks and Temples · Tanks and Temples (mean) | F1: 0.50 -> 0.50 | No clear effect |
| use CarGS instead of Scaffold-GS | Mip-NeRF360 · Mip-NeRF360 average on all scenes | PSNR: 28.30 -> 27.68 | Hurt |
| use CarGS instead of Scaffold-GS | Mip-NeRF360 · Mip-NeRF360 average on all scenes | SSIM: 0.832 -> 0.834 | Helped |
| use CarGS instead of PGSR | Mip-NeRF360 · Mip-NeRF360 average on all scenes | LPIPS: 0.193 -> 0.209 | Hurt |
| add geometry loss to Scaffold-GS (ablation) | Scaffold-GS · Tanks and Temples Truck | F1 | Helped |
| add Geo-MLP (ablation) | Scaffold-GS + geometry loss · Tanks and Temples Truck | F1: 0.38 -> 0.44 | Helped |
| add Lite-Geo (ablation) | Geo-MLP · Tanks and Temples Truck | F1: 0.44 -> 0.62 | Helped |
| add geometry-guided densification (ablation) | Geo-MLP + Lite-Geo · Tanks and Temples Truck | F1: 0.62 -> 0.65 | Helped |
| add geometry-guided densification (ablation) | Geo-MLP + Lite-Geo · Tanks and Temples Truck | PSNR: 26.13 -> 26.41 | Helped |
Surprising findings
- Removing the opacity gradient from geometry supervision has minimal effect on reconstruction, while removing covariance gradients severely hurts geometry learning.
Limitations
- Geo-MLP alone overfits geometry loss without the residual Lite-Geo design
- Reconstruction F1 matches but does not exceed NeuralAngelo and PGSR on TnT mean
- Pipeline requires pre-training the rendering task for several epochs before adding Lite-Geo
Experimental setup
Datasets: Tanks and Temples Mip-NeRF360
Benchmarks: Tanks and Temples Mip-NeRF360
Metrics: PSNR SSIM LPIPS F1-score storage (MB) FPS training time
Baselines: NeuS Neuralangelo 2DGS GOF PGSR GSDF Scaffold-GS 3DGS RaDe-GS SuGaR
Builds on: Scaffold-GS PGSR 3D Gaussian Splatting GSDF
Closest papers by meaning
- CompactSplat: Spatially Adaptive Gaussian Distribution for Feedforward Scene Reconstruction (NeurIPS 2026 · similarity 0.92)
- Effective Rank Analysis and Regularization for Enhanced 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.91)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.91)
- Sparse2DGS: Geometry-Prioritized Gaussian Splatting for Surface Reconstruction from Sparse Views (CVPR 2025 · similarity 0.91)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.91)
- VA-GS: Enhancing the Geometric Representation of Gaussian Splatting via View Alignment (NeurIPS 2025 · similarity 0.91)
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.