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

3D Gaussian Splatting performs 3D reconstruction by densifying sparse point clouds into Gaussian ellipsoids of the order of 100,000, and the reconstruction results show excellent visual effects. However, the point cloud data derived from 3D Gaussian Splatting is not fully utilized in the reconstruction process. To this end, this paper proposes to optimize the point cloud data derived from 3D Gaussian Splatting to improve the rendering quality of 3D Gaussian Splatting. First, the sparse point cloud is input into 3D Gaussian Splatting for reconstruction. During the reconstruction process, the color of the center point of the ellipsoid is extracted from the spherical harmonic characteristic coefficients of the Gaussian ellipsoid, and then the color and Gaussian ellipsoid center coordinate attributes are stored as new point cloud data. The radius filtering method is used to denoise the point cloud. Finally, the denoised point cloud is re-input into 3D Gaussian Splatting for reconstruction to obtain the optimized rendering results. The experiments are conducted on three public datasets. The quantitative comparison results show that the proposed method is improved compared with the original 3D Gaussian Splatting, and the qualitative comparison results show that the proposed method optimizes the blurred area rendered by 3D Gaussian Splatting. The method proposed in this paper optimizes 3D Gaussian Splatting-derived point cloud data for 3D reconstruction, obtains higher quality rendering results, and can obtain reconstruction effects that are better than the original 3D Gaussian Splatting on public datasets.