DeSplat: Decomposed Gaussian Splatting for Distractor-Free Rendering

Yihao Wang,Marcus Klasson,Matias Turkulainen,Shuzhe Wang,Juho Kannala,Arno Solin

2024-11-29

Abstract:Gaussian splatting enables fast novel view synthesis in static 3D environments. However, reconstructing real-world environments remains challenging as distractors or occluders break the multi-view consistency assumption required for accurate 3D reconstruction. Most existing methods rely on external semantic information from pre-trained models, introducing additional computational overhead as pre-processing steps or during optimization. In this work, we propose a novel method, DeSplat, that directly separates distractors and static scene elements purely based on volume rendering of Gaussian primitives. We initialize Gaussians within each camera view for reconstructing the view-specific distractors to separately model the static 3D scene and distractors in the alpha compositing stages. DeSplat yields an explicit scene separation of static elements and distractors, achieving comparable results to prior distractor-free approaches without sacrificing rendering speed. We demonstrate DeSplat's effectiveness on three benchmark data sets for distractor-free novel view synthesis. See the project website at <a class="link-external link-https" href="https://aaltoml.github.io/desplat/" rel="external noopener nofollow">this https URL</a>.

Computer Vision and Pattern Recognition,Machine Learning