Research map / 3D vision / Gaussian splatting
DNGaussian: Optimizing Sparse-View 3D Gaussian Radiance Fields with Global-Local Depth Normalization
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 DNGaussian, a depth-regularized 3D Gaussian Splatting framework for real-time few-shot novel view synthesis, using hard/soft depth regularization on Gaussian centers/opacity and global-local depth normalization.
Problem
Few-shot novel view synthesis from sparse input views with 3D Gaussian radiance fields suffers from geometry degradation, while existing NeRF-based sparse-view methods have high training cost and slow rendering.
Approach
Freeze Gaussian scale and rotation and regularize only centers and opacities with monocular depth (hard depth renders nearest Gaussians with large fixed opacity; soft depth freezes centers), plus a Global-Local Depth Normalization that normalizes depth patches locally and globally to emphasize small local depth changes.
Key results
- On LLFF 3-view, DNGaussian reaches PSNR 19.12, LPIPS 0.294, SSIM 0.591, comparable to FreeNeRF (19.63) and SparseNeRF (19.86) while training in 3.5 min on 2GB GPU memory.
- Rendering at 300 FPS, over 3000x faster than FreeNeRF/SparseNeRF, and about 25x reduction in training time under limited resources.
- On Blender 8-view, DNGaussian achieves PSNR 24.305, SSIM 0.886, LPIPS 0.088, the best among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DNGaussian depth regularization to 3DGS | 3D Gaussian Splatting · LLFF, 3 views | LPIPS: 0.405 -> 0.294 | Helped |
| use DNGaussian instead of 3DGS | 3D Gaussian Splatting · DTU, 3 views | SSIM: 0.585 -> 0.790 | Helped |
| use DNGaussian instead of 3DGS | 3D Gaussian Splatting · Blender, 8 views | PSNR: 22.226 -> 24.305 | Helped |
| add shape freezing during depth regularization (ablation) | DNGaussian · LLFF, 3 views | PSNR: 17.96 -> 19.12 | Helped |
| add center freezing in soft depth calculation (ablation) | DNGaussian · LLFF, 3 views | PSNR: 18.87 -> 19.12 | Helped |
| add Hard Depth Regularization (ablation) | DNGaussian · LLFF, 3 views | PSNR: 17.90 -> 18.31 | Helped |
| add Soft Depth Regularization alongside Hard Depth Regularization (ablation) | DNGaussian · LLFF, 3 views | PSNR: 17.90 -> 18.31 | Helped |
| add Global Depth Normalization (ablation) | DNGaussian depth regularization · LLFF, 3 views | PSNR: 18.31 -> 18.68 | Helped |
| add Local Depth Normalization (ablation) | DNGaussian depth regularization · LLFF, 3 views | LPIPS: 0.324 -> 0.294 | Helped |
Surprising findings
- Raw 3DGS with the same hyperparameters performs poorly on sparse views, and fixed-global-scale depth losses are shown to overlook small local depth errors.
- On DTU, DNGaussian has lower PSNR than the best NeRF baselines despite better LPIPS and SSIM, attributed to scale variance and background noise.
Limitations
- PSNR lags some NeRF baselines on LLFF and DTU, partly due to background color in invisible regions for Gaussian methods.
- Relies on monocular depth estimates (DPT) which can be coarse or inaccurate.
Experimental setup
Datasets: LLFF DTU NeRF Blender Synthetic
Benchmarks: LLFF DTU NeRF Blender Synthetic
Metrics: PSNR SSIM LPIPS AVGE FPS training time GPU memory
Baselines: SRF PixelNeRF MVSNeRF Mip-NeRF DietNeRF RegNeRF FreeNeRF SparseNeRF 3DGS
Builds on: 3D Gaussian Splatting SparseNeRF FreeNeRF DPT
Closest papers by meaning
- Compact 3D Gaussian Representation for Radiance Field (CVPR 2024 · similarity 0.90)
- 3D Convex Splatting: Radiance Field Rendering with 3D Smooth Convexes (CVPR 2025 · similarity 0.89)
- Structure Consistent Gaussian Splatting with Matching Prior for Few-shot Novel View Synthesis (NeurIPS 2024 · similarity 0.89)
- Gaussian Herding across Pens: An Optimal Transport Perspective on Global Gaussian Reduction for 3DGS (NeurIPS 2025 · similarity 0.89)
- MaskGaussian: Adaptive 3D Gaussian Representation from Probabilistic Masks (CVPR 2025 · similarity 0.89)
- FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training (NeurIPS 2024 · similarity 0.89)
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.