CGS: Continual Gaussian Splatting for Evolving 3D Scene Reconstruction
Loading...
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
The Eurographics Association
Abstract
3D Gaussian Splatting (3DGS) has gained significant attention for its fast optimization and high-quality rendering capabilities. However, in the context of continual scene reconstruction, optimizing newly observed regions often leads to degradation in previously reconstructed areas due to changes in camera viewpoints. To address this issue, we propose Continual Gaussian Splatting (CGS)-an efficient incremental reconstruction method that updates dynamic scenes using only a limited amount of new data while minimizing computational overhead. CGS is composed of three core components. First, we introduce a similarity-based registration algorithm that leverages the strong semantic understanding and translation invariance of pretrained Transformers to identify and align similar regions between new and existing scenes. These regions are then modeled as Gaussian Mixture Models (GMMs) to handle sparsity and outliers in point clouds, ensuring geometric consistency across scenes. Second, we propose Continual Gaussian Optimization (CGO), an importance-aware optimization strategy. By computing the Fisher Information Matrix, we evaluate the significance of each Gaussian point in the old scene and automatically restrict updates to those deemed critical, allowing only non-sensitive components to be adjusted. This ensures the preservation of the original scene while efficiently integrating new content. Finally, to address remaining issues such as geometric inconsistencies, blurring, and ghosting artifacts during optimization, we introduce a series of geometric regularization techniques. These terms guide the optimization toward geometrically coherent 3D structures, ultimately enhancing rendering quality. Extensive experiments demonstrate that CGS effectively mitigates forgetting and significantly improves overall reconstruction fidelity.
Description
CCS Concepts: Computing methodologies → Reconstruction
@inproceedings{10.2312:pg.20251284,
booktitle = {Pacific Graphics Conference Papers, Posters, and Demos},
editor = {Christie, Marc and Han, Ping-Hsuan and Lin, Shih-Syun and Pietroni, Nico and Schneider, Teseo and Tsai, Hsin-Ruey and Wang, Yu-Shuen and Zhang, Eugene},
title = {{CGS: Continual Gaussian Splatting for Evolving 3D Scene Reconstruction}},
author = {Yang, Shuojin and Chen, Haoxiang and Mu, Taijiang},
year = {2025},
publisher = {The Eurographics Association},
ISBN = {978-3-03868-295-0},
DOI = {10.2312/pg.20251284}
}