Leveraging Planar Prior Knowledge for Regularization in 3D Gaussian Splatting Optimization

• Jisoo Han,
• Jongwon Jeong,
• I-gil Kim,

Vol. 19, No. 7, July 31, 2025
10.3837/tiis.2025.07.010, Download Paper (Free):

• 3D gaussian splatting
• AR/VR
• computer vision
• Plane Detection
• Ransac
• regularization
• Scene Reconstruction

Abstract

This study introduces a novel approach to improving 3D Gaussian splatting (3DGS) optimization by integrating prior knowledge of planar structures commonly found in indoor environments. Traditional 3DGS methods often struggle to reconstruct complex indoor scenes accurately due to the lack of structural guidance. To overcome this limitation, we propose a regularization framework that incorporates predefined plane equations to constrain the positions of Gaussian means during optimization. The proposed method comprises three key strategies. First, a distance loss ensures that the means of initial 3D Gaussians identified as belonging to a plane remain close to that plane throughout the optimization process. Second, the plane equations are dynamically updated at specific optimization steps based on the means of the Gaussians associated with each plane, allowing the model to adapt effectively to the data. Third, a dot loss enforces consistency in the angles between the normals of different planes, preventing significant deviations from their initial orientations. We validate our approach through extensive experiments on synthetic and real-world datasets. Results demonstrate that the proposed regularization strategies significantly enhance the reconstruction quality of indoor scenes by leveraging planar prior knowledge. The method improves the structural integrity of reconstructed models and remains effective even when plane information is partially available or estimated through methods such as RANSAC. This work highlights the value of incorporating geometric constraints into optimization algorithms and establishes a foundation for future research in structured 3D reconstruction.

Statistics

Cite this article