3D Gaussian Splatting for Isolated Objects

Abstract

This thesis outlines a new scene object isolation method that is capable of isolating centre scene objects from a set of images that captures the scene. Doing this it shows some interesting quirks in the SMF data. The thesis experiments with the different configurations for the object isolation method, and presents these results. Beside this the thesis proposes some changes and extensions to Gaussian splatting to enable it to optimize the isolated objects, and re-implements depthregularized Gaussian splatting. The modifications are explored by tweaking their hyper parameters, and in doing so showing their influence on the accuracy of the final trained object. The thesis is closed by speculating in possible improvements that could still be made to the object isolation method, and quickly proposes a method that could be used to allow for multiple objects, from the same scene, to be trained at the same time and be combined into one scene.