SODA-GS: Scene Object Decomposition and Adaptive Cleanup for 3D Gaussian Splatting
Year: 2026
Type: conference-paper
Abstract: 3D Gaussian Splatting (3DGS) has recently emerged as an expressive scene representation for rendering and physics-based simulation. Nevertheless, extracting clean object-level 3DGS representations from real-world scenes remains challenging. Recent approaches leverage large language models (LLMs) alongside open-vocabulary segmentation to address this problem; however, they frequently exhibit background leakage, instance confusion, and geometric outliers. These artifacts may be visually subtle but critically degrade downstream tasks such as physical simulation. We present a scene preprocessing pipeline for robust object-level decomposition and cleanup for 3D Gaussian Splatting. Starting from posed RGB images and a reconstructed 3DGS scene, our method combines vision-language models for semantic object discovery with open-vocabulary 3DGS segmentation. We introduce an adaptive threshold refinement strategy guided by visual feedback, followed by clustering-based instance separation and geometric cleanup to isolate coherent object instances. We further perform environment gap filling to produce geometrically consistent static scenes suitable for physics simulation after object removal. We evaluate our approach on both a synthetic outdoor dataset with ground-truth geometry and a real-world captured scene. Results show consistent improvements over a direct open-vocabulary segmentation baseline, particularly in tail-sensitive geometric error metrics, while qualitative results demonstrate the removal of structurally inconsistent Gaussians critical for simulation.
Source: Eurographics
Language: English
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7163065249