Indoor Reconstruction and Visualization Based on 3D Gaussian Model

Abstract

Computer vision and graphics 3D reconstruction has a wide usage. The existing methods to have a few problems handling complex geometrics, changes in the illumination and space insecurity. In order to address these issues, the present paper presents a new algorithm of 3D reconstruction by relying on the model of Gaussian distribution and features multi-scale extraction. Spatial uncertainty and geometric variation of point clouds is represented in the form of 3D Gaussian and the multi-scale division of the feature extraction assists in enhancing the ability to grasp the fine-grained problems and the broad structures. It is also demonstrated that adaptive multi-scale loss is introduced to present powerful supervision during dynamic illumination and obstruction. As demonstrated in the experimental results of DTU and LLFF datasets, the proposed method has superior reconstruction accuracy, completeness, and robustness compared to the existing methods. This work has contributed in the following way: (1) a system of reconstruction and visualization of the indoor environment based on 3D-Gaussian models, (2) a 3-scale feature extraction design to increase the capability of resisting the complex geometry and illumination and (3) the adaptive loss design system, which creates a balance between reconstruction quality and efficiency.