Research on Performance Improvement of 3D Gaussian Splatting Model Driven by Optimizing Point Cloud Data

The use of scenes captured from multiple photos or videos for 3D reconstruction is a hot topic in the fields of image and vision. At present, 3D Gaussian Splatting(3D-GS) [1] represents scenes as explicit 3D Gaussian ellipsoids with specific properties for 3D reconstruction, and the reconstruction results demonstrate advanced visual quality and real-time rendering effects in the industry. 3D-GS uses the sparse point cloud generated by SfM as input, and through adaptive density control, densifies the sparse point cloud into a million level Gaussian ellipsoid to achieve scene reconstruction. However, during this process, the point cloud data derived from 3D-GS was not fully utilized. This article proposes an optimized point cloud data derived from 3D-GS to improve the rendering quality of 3D-GS. Firstly, the point cloud generated by SfM is input into 3D-GS for reconstruction. During the reconstruction process, the color of the ellipsoid center point is extracted from the spherical harmonic feature coefficients of the Gaussian ellipsoid. Then, the color attributes and the coordinates of the Gaussian ellipsoid center position are stored as new point cloud data. The radius filtering method is used to denoise the obtained point cloud data. Finally, the denoised point cloud set is re inputted into 3D-GS as new data for reconstruction, resulting in an optimized rendering effect. The experiment was compared with the original 3D-GS method on three publicly available datasets. Compared with the original 3D-GS method, SSIM increased by 0.013, PSNR increased by 0.379, and LPIPS decreased by 0.037; At the same time, qualitative comparisons were made in the rendered scenes, and the comparison results proved that the proposed method improved the rendering quality of 3D-GS. The method proposed in this article is to optimize the point cloud data derived from 3D-GS and re input it into 3D-GS for reconstruction, in order to obtain higher quality rendering results. Good results can be achieved on all three publicly available datasets. Optimize the point cloud data derived from 3D-GS to achieve iterative optimization of 3D-GS rendering quality.