Information-Guided 3D Gaussian Splatting

Abstract

This thesis presents an intrinsically informed segmentation method using clustering for 3D Gaussian Splatting (3DGS) scenes to reconstruct solely the object of interest (OOI). We developed a pipeline that leverages the underlying distribution and density of Gaussians initialized from Structure from Motion (SfM) to segment the OOI. While the results do not show comprehensive segmentation of the OOI, the results are promising with room for improvement in future work. The pipeline is evaluated on a subset of the MiP-NeRF 360 dataset, for which the ground truth segmentation masks have been manually created. This work contributes to creating 3D Gaussian Splats of solely an object at the center of the scene and is the first to the authors known method that allows 3DGS reconstruction on a specific object without foundational models. Furure research directions include incorporating the clustering pipeline into the training loop of 3DGS to enable more detailed segmentation of the OOI.