Reproduce “Self-Ensembling Gaussian Splatting for Few-Shot Novel View Synthesis” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chen Zhao, Xuan Wang, Tong Zhang, Saqib Javed, Mathieu Salzmann
Self-Ensembling Gaussian Splatting (SE-GS) introduces a self-ensembling framework with an uncertainty-aware perturbation strategy to enhance 3D Gaussian Splatting for novel view synthesis under few-shot conditions. The method significantly improves rendering quality and generalization by mitigating overfitting, consistently outperforming existing state-of-the-art techniques across multiple datasets.
Key insights
- Uncertainty-aware perturbation effectively balances model diversity and supervision reliability, enabling robust self-ensembling by selectively perturbing Gaussians based on rendering uncertainty.
- Self-ensembling can be efficiently integrated into the 3DGS framework by dynamically perturbing a single base model, avoiding the high computational cost of training multiple independent models.
- The self-supervised regularization process eliminates the dependency on potentially noisy auxiliary data, contributing to the method's consistent performance and stability in few-shot scenarios.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.