UGOD: Uncertainty-Guided Differentiable Opacity and Soft Dropout for Enhanced Sparse-View 3DGS

TL;DR

UGOD introduces an uncertainty-guided approach with differentiable opacity and soft dropout to improve sparse-view 3D Gaussian Splatting, leading to better rendering quality with fewer Gaussians.

Contribution

The paper proposes a novel uncertainty-guided weighting and regularization method for 3D Gaussian Splatting, enhancing sparse-view 3D synthesis performance.

Findings

Outperforms existing methods in sparse-view 3D reconstruction.

Achieves 3.27% PSNR improvement on MipNeRF 360 dataset.

Requires fewer Gaussians for high-quality rendering.

Abstract

3D Gaussian Splatting (3DGS) has become a competitive approach for novel view synthesis (NVS) due to its advanced rendering efficiency through 3D Gaussian projection and blending. However, Gaussians are treated equally weighted for rendering in most 3DGS methods, making them prone to overfitting, which is particularly the case in sparse-view scenarios. To address this, we investigate how adaptive weighting of Gaussians affects rendering quality, which is characterised by learned uncertainties proposed. This learned uncertainty serves two key purposes: first, it guides the differentiable update of Gaussian opacity while preserving the 3DGS pipeline integrity; second, the uncertainty undergoes soft differentiable dropout regularisation, which strategically transforms the original uncertainty into continuous drop probabilities that govern the final Gaussian projection and blending process for rendering. Extensive experimental results over widely adopted datasets demonstrate that our method outperforms rivals in sparse-view 3D synthesis, achieving higher quality reconstruction with fewer Gaussians in most datasets compared to existing sparse-view approaches, e.g., compared to DropGaussian, our method achieves 3.27\% PSNR improvements on the MipNeRF 360 dataset.