Research map / 3D vision / Gaussian splatting
Quantifying and Alleviating Co-Adaptation in Sparse-View 3D 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.
Defines a co-adaptation score (CA) for sparse-view 3DGS via pixel variance across dropout renderings, and proposes Gaussian dropout and multiplicative opacity noise as plug-and-play regularizers that reduce CA and improve novel-view quality on LLFF, DTU, and Blender.
Problem
Sparse-view 3D Gaussian Splatting produces appearance artifacts (color outliers) in novel views even though training views are fit well; the cause is not well explained by existing geometry-focused fixes.
Approach
Attributes sparse-view appearance artifacts to co-adaptation among Gaussians, quantifies it with a CA score (pixel variance over random Gaussian-subset renderings), and mitigates it with two simple plug-and-play regularizers: random Gaussian dropout and multiplicative opacity noise.
Key results
- Dropout improves Binocular3DGS on LLFF from 21.44 to 22.12 PSNR and reduces train CA from 0.001845 to 0.000875.
- Dropout improves DNGaussian on DTU from 18.91 to 19.86 PSNR with lower test CA (0.005744 to 0.001917).
- Opacity noise with sigma=0.8 gives the best Binocular3DGS LLFF PSNR (22.119) in the sigma ablation, and both strategies lower CA across all five methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add dropout regularization | 3DGS-based sparse-view methods · LLFF and DTU | PSNR: 3DGS LLFF 19.36 -> 20.20; Binocular3DGS LLFF 21.44 -> 22.12; Binocular3DGS DTU 20.71 -> 21.03 | Helped |
| add opacity noise | 3DGS-based sparse-view methods · LLFF and DTU | PSNR: 3DGS LLFF 19.36 -> 19.91; 3DGS DTU 17.30 -> 17.27 | Mixed |
| add dropout and opacity noise together (ablation) | Binocular3DGS · LLFF and DTU | PSNR: LLFF 21.44 -> 22.11; DTU 20.71 -> 21.05 | Mixed |
| vary dropout probability from 0.0 to 0.2 (ablation) | Binocular3DGS · LLFF | PSNR: 21.440 -> 22.123 | Helped |
| increase dropout probability above 0.2 (ablation) | Binocular3DGS · LLFF | PSNR: 22.123 -> 21.793 | Hurt |
| use inference-time random dropout (Strategy A) instead of baseline rendering (ablation) | Binocular3DGS trained with dropout p=0.2 · LLFF | PSNR: 21.440 -> 21.977 | Helped |
| average five stochastic dropout renderings (Strategy B) rather than Strategy A (ablation) | Binocular3DGS trained with dropout p=0.2 · LLFF | PSNR: 21.977 -> 22.124 | Helped |
| scale all Gaussian opacities by (1-p) for single-pass rendering (Strategy C) (ablation) | Binocular3DGS trained with dropout p=0.2 · LLFF | PSNR: 21.440 -> 22.123 | Helped |
| vary opacity noise scale from 0.0 to 0.8 (ablation) | Binocular3DGS · LLFF | PSNR: 21.440 -> 22.119 | Helped |
| use opacity noise scale below 0.8 rather than 0.8 (ablation) | Binocular3DGS · LLFF | PSNR: 22.119 at 0.8; 21.647 at 0.1 | Mixed |
| inject scale noise at scale 0.2 (ablation) | Binocular3DGS · LLFF | PSNR: 21.440 -> 21.760 | Helped |
| inject position noise at scale 0.1 (ablation) | Binocular3DGS · LLFF | PSNR: 21.440 -> 21.093 | Hurt |
Surprising findings
- Lower CA does not always yield better rendering: CA keeps dropping for higher dropout rates or opacity noise scales, but PSNR peaks at moderate values.
- Applying dropout and opacity noise together gives no further improvement over either alone.
Limitations
- Theoretical CA derivation relies on a first-order approximation of transmittance, neglecting higher-order terms.
- CA is not strictly monotonic with reconstruction quality, so it is an imperfect proxy beyond a threshold.
- Experiments are limited to a few sparse-view benchmarks with 3 or 8 training views.
Experimental setup
Datasets: LLFF DTU Blender
Benchmarks: LLFF DTU Blender
Metrics: PSNR SSIM LPIPS Co-Adaptation score (CA)
Baselines: 3DGS DNGaussian FSGS CoR-GS Binocular3DGS AbsGS
Builds on: 3D Gaussian Splatting Dropout DNGaussian Binocular3DGS FSGS
Closest papers by meaning
- DropGaussian: Structural Regularization for Sparse-view Gaussian Splatting (CVPR 2025 · similarity 0.92)
- Generative Sparse-View Gaussian Splatting (CVPR 2025 · similarity 0.92)
- 3DGS-Enhancer: Enhancing Unbounded 3D Gaussian Splatting with View-consistent 2D Diffusion Priors (NeurIPS 2024 · similarity 0.92)
- FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training (NeurIPS 2024 · similarity 0.92)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.92)
- Perceptual-GS: Scene-adaptive Perceptual Densification for Gaussian Splatting (ICML 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.