Abstract
dc:description.abstractOver the past year, 3D Gaussian Splatting has become a widespread research topic of interest. It has proven itself against other popular Novel View Synthesis algorithmslike Neural Radiance Fields (NeRF) by replacing the approximation function witha modified point cloud. This modified point cloud allows for the rendering of novelviews in real time. While this gives Gaussian Splatting an advantage over algorithmslike NeRF, it shares the same issue of requiring prior camera intrinsics such as a PointCloud and Camera Poses. Most Gaussian Splatting pipelines use Colmap Sparse, aStructure-from-Motion (SfM) technology that efficiently registers the necessary pointclouds and camera poses. However, these features require a surplus number of images that are often unavailable. In this thesis, we investigate and compare other types of SfM and Multi-View-Stereo (MVS) technologies that use both traditional and deep-learning-based algorithms. The objective is to see if these methods’ increased point cloud density results in better Gaussian Splatting results, provided a low number of images. This evaluation is performed with quantitative metrics (PSNR, SSIM, PointCloud Size, Runtime), and qualitative assessment (rendering Gaussian Splats andPoint Clouds).
Degree
thesis:*- Name thesis:degree_name
- MS in Computer Science
- Discipline thesis:degree_discipline
- Computer Science
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Levin, Daniel L
- Contributors dc:contributor
-
- Jonathan Ventura
- Computer Science
- College of Engineering
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://digitalcommons.calpoly.edu/theses/2842