3DGS-to-PC: 3D Gaussian Splatting to Dense Point Clouds

Lewis A G Stuart, Andrew Morton, Ian Stavness, Michael P Pound; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 3789-3798

3D Gaussian Splatting (3DGS) excels at producing highly detailed 3D reconstructions, but these scenes often require specialised renderers for effective visualisation. In contrast, point clouds are a widely used 3D representation and are compatible with most popular 3D processing software packages. In this work, we introduce 3DGS-to-PC, a highly customisable framework for transforming 3DGS scenes into dense, accurate point clouds. We first render new Gaussian colours by calculating the visual contributions that each Gaussian made to each pixel, and then assigning these Gaussians with the colour of the pixel they contributed most to. We then remove Gaussians that contribute little to the scene, reducing noise. Finally, we generate new points by sampling from each Gaussian as a multivariate normal distribution, while removing erroneously generated points. The number of points per Gaussian is determined based on its relative scale and pixel contributions. The result is a dense point cloud that accurately represents each scene. At time of writing, there is no dedicated method for converting a 3DGS scene into a point cloud. We evaluate our method against 3DGS meshing techniques - from which points can be sampled - and show that our method is competitive with state-of-the-art approaches that require training of the scene, while performing more efficiently than standard meshing methods, such as Poisson Surface Reconstruction. 3DGS-to-PC is efficient, typically taking under a minute to process complex scenes, thus providing an effective tool for converting 3DGS data into robust point clouds. Our codebase can be accessed via https://github.com/Lewis-Stuart-11/3DGS-to-PC