Research map / 3D vision / Gaussian splatting
DropoutGS: Dropping Out Gaussians for Better Sparse-view Rendering
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 DropoutGS, which adds random Gaussian dropout regularization (RDR) and an edge-guided splitting strategy (ESS) to 3DGS to reduce overfitting under sparse views.
Problem
Novel view synthesis from sparse input views with 3D Gaussian Splatting overfits, producing hollow artifacts and degraded generalization.
Approach
Treats sparse-view 3DGS degradation as overfitting from excess Gaussians; randomly drops Gaussians during training and supervises the sub-model with the full model's rendering (no ground-truth term), then uses edge-guided splitting to recover high-frequency detail. Plug-in with no extra depth model.
Key results
- On LLFF with 3 views, DropoutGS reaches 19.35 PSNR / 0.282 LPIPS versus DNGaussian's 19.12 / 0.294.
- On DTU with 3 views, DropoutGS reaches 20.22 PSNR and 0.150 LPIPS, against DNGaussian's 18.91 and 0.176.
- On Blender with 8 views, DropoutGS reaches 24.476 PSNR and 0.085 LPIPS, the best of the compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply DropoutGS instead of DNGaussian | DNGaussian · LLFF, 3 views | PSNR: 19.12 -> 19.35 | Helped |
| Apply DropoutGS instead of DNGaussian | DNGaussian · DTU, 3 views | PSNR: 18.91 -> 20.22 | Helped |
| Apply DropoutGS instead of DNGaussian | DNGaussian · Blender, 8 views | PSNR: 24.305 -> 24.476 | Helped |
| Add ESS (ablation) | DNGaussian baseline · LLFF, 3 views | PSNR: 18.76 -> 18.91 | Helped |
| Add RDR (ablation) | DNGaussian baseline · LLFF, 3 views | PSNR: 18.76 -> 19.26 | Helped |
| Set dropout rate to 0.2 (ablation) | DropoutGS · LLFF, 3 views | PSNR: 18.76 -> 19.20 | Helped |
| Set dropout rate to 0.4 (ablation) | DropoutGS · LLFF, 3 views | LPIPS: 0.300 -> 0.282 | Helped |
| Set dropout rate to 0.6 (ablation) | DropoutGS · LLFF, 3 views | PSNR: 19.35 -> 19.37 | Helped |
| Set dropout rate to 0.8 (ablation) | DropoutGS · LLFF, 3 views | LPIPS: 0.282 -> 0.302 | Hurt |
| Lower the edge threshold from 1 × 10−1 to 1 × 10−2 (ablation) | DropoutGS ESS · LLFF, 3 views | LPIPS: 0.316 -> 0.313 | Helped |
| Lower the edge threshold from 1 × 10−2 to 1 × 10−3 (ablation) | DropoutGS ESS · LLFF, 3 views | LPIPS: 0.313 -> 0.282 | Helped |
| Lower the edge threshold from 1 × 10−3 to 1 × 10−4 (ablation) | DropoutGS ESS · LLFF, 3 views | PSNR: 19.35 -> 19.17 | Hurt |
Surprising findings
- Using the full model's rendering rather than ground-truth colors as the dropout supervision target works better, and the dropout compensation strategy did not help.
- Models with fewer Gaussians overfit less under sparse views, and the optimal Gaussian count grows with the number of training views.
Limitations
- Gains shrink as the number of training views increases (e.g., 9 views on LLFF).
- Edge threshold is sensitive: overly low thresholds cause distortion and lower PSNR, and over-splitting can create artifacts.
- Pilot study on DTU and the 3-view setting are the main justification for the overfitting explanation.
Experimental setup
Datasets: LLFF DTU Blender
Benchmarks: LLFF DTU Blender
Metrics: PSNR SSIM LPIPS AVGE
Baselines: DNGaussian 3DGS FreeNeRF SparseNeRF RegNeRF DietNeRF Mip-NeRF MVSNeRF PixelNeRF SRF CoR-GS FSGS
Builds on: 3DGS DNGaussian dropout
Closest papers by meaning
- DropGaussian: Structural Regularization for Sparse-view Gaussian Splatting (CVPR 2025 · similarity 0.95)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.91)
- Generative Sparse-View Gaussian Splatting (CVPR 2025 · similarity 0.91)
- Quantifying and Alleviating Co-Adaptation in Sparse-View 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
- FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training (NeurIPS 2024 · similarity 0.91)
- LP-3DGS: Learning to Prune 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.90)
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.