3D Gaussian Splatting techniques have recently emerged as a promising approach to address the tasks of novel-view synthesis and six degree-of-freedom (6-DOF) tracking of dense scene elements in a dynamic scene, most notably the method presented by Luiten et al. Though this method produces state of the art results in both performance and quality for these tasks, it exhibits a significant limitation in that it cannot handle new objects entering into the scene due to the fact that all of the scene's gaussians are initialized from the first frame. This project aims to address this by introducing an additional step in the training pipeline to identify new objects via identifying relative dips in the reconstruction losses of segmentation masks, initialize new gaussians for them via monocular depth estimation maps, and optimize the scene reconstruction to effectively initialize objects into the scene.
tbd