Methods

We utilize semantic and geometric priors to train a 3D Gaussian Splatting (3DGS) scene, enabling the learning of 3D-consistent embeddings. These embeddings help in effectively grouping Gaussians that belong to common planes.

Figure 5: By applying the Push-Pull loss function, we effectively bring the embeddings of Gaussians belonging to the same plane closer together, while pushing apart the embeddings of Gaussians from different planes

Figure 6: Using RANSAC, we fit planes to the grouped Gaussians and subsequently sample points from these planes, which helps in the reinitialization of the 3DGS process.