Research map / 3D vision / Gaussian splatting
DropGaussian: Structural Regularization for Sparse-view Gaussian Splatting
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 DropGaussian, a prior-free sparse-view 3D Gaussian splatting regularizer that randomly drops Gaussians during training with a progressively increasing drop rate.
Problem
3D Gaussian splatting overfits to the few training views in sparse-view novel view synthesis, and existing fixes rely on strong priors (depth, 2D generative models, optical flow) that add cost or error.
Approach
Randomly drop Gaussians during training like dropout, compensating opacity by 1/(1-r), so remaining Gaussians (including distant ones) receive larger gradients; the drop rate grows linearly with iteration (r_t = gamma * t/T_total) since overfitting appears late in training.
Key results
- On LLFF 3-view, DropGaussian reaches 20.76 PSNR / 0.713 SSIM / 0.200 LPIPS, the best PSNR among compared methods (CoR-GS 20.45).
- On Mip-NeRF360 12-view, reaches 19.74 PSNR vs. 19.52 for CoR-GS; 24-view 24.13 PSNR.
- On Blender 8-view, reaches 25.42 PSNR, the highest, though SSIM (0.888) trails CoR-GS (0.896).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply DropGaussian to 3DGS | 3DGS · LLFF 3-view | PSNR: 19.22 -> 20.76 | Helped |
| Apply DropGaussian to 3DGS | 3DGS · Mip-NeRF360 12-view | PSNR: 18.52 -> 19.74 | Helped |
| Apply DropGaussian to 3DGS | 3DGS · Blender | PSNR: 21.56 -> 25.42 | Helped |
| Progressively increase the drop rate rather than fixing it at 0.2 (ablation) | DropGaussian · LLFF 3-view | PSNR: 20.16 -> 20.56 | Helped |
| Use a fixed drop rate of 0.1 instead of progressive adjustment (ablation) | DropGaussian · LLFF 3-view | PSNR: 20.56 -> 20.28 | Hurt |
| Use a fixed drop rate of 0.3 instead of progressive adjustment (ablation) | DropGaussian · LLFF 3-view | PSNR: 20.48 -> 20.15 | Hurt |
| Randomly drop Gaussians instead of selectively dropping by gradient (ablation) | DropGaussian · LLFF 3-view | PSNR: 19.62 -> 20.76 | Helped |
| Randomly drop Gaussians instead of selectively dropping by distance (ablation) | DropGaussian · LLFF 3-view | PSNR: 17.49 -> 20.76 | Helped |
| Use temporary Gaussian dropping instead of gradient-based L1 regularization (ablation) | DropGaussian · LLFF 3-view | PSNR: 19.97 -> 20.76 | Helped |
| Use temporary Gaussian dropping instead of distance-based L1 regularization (ablation) | DropGaussian · LLFF 3-view | PSNR: 19.73 -> 20.76 | Helped |
Surprising findings
- Randomly dropping Gaussians outperforms selective dropping by gradient magnitude or distance, and outperforms L1 regularization for pruning.
- A fixed high drop rate (0.2 to 0.3) degrades performance versus lower rates, while progressive scheduling improves results.
Limitations
- Sensitivity to the scaling factor gamma, which requires dataset-specific tuning.
- Not uniformly best on all metrics; Blender SSIM and LPIPS trail some baselines.
Experimental setup
Datasets: LLFF Mip-NeRF360 Blender Replica
Benchmarks: LLFF Mip-NeRF360 Blender Replica
Metrics: PSNR SSIM LPIPS
Baselines: 3DGS DNGaussian FSGS CoR-GS Mip-NeRF DietNeRF RegNeRF FreeNeRF SparseNeRF pixelSplat MVSplat FreeSplat
Builds on: 3D Gaussian splatting FSGS CoR-GS DNGaussian
Closest papers by meaning
- DropoutGS: Dropping Out Gaussians for Better Sparse-view Rendering (CVPR 2025 · similarity 0.95)
- Generative Sparse-View Gaussian Splatting (CVPR 2025 · similarity 0.93)
- FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training (NeurIPS 2024 · similarity 0.93)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.92)
- Quantifying and Alleviating Co-Adaptation in Sparse-View 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.92)
- S2Gaussian: Sparse-View Super-Resolution 3D Gaussian Splatting (CVPR 2025 · similarity 0.92)
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.