Recent advancements in 3D scene reconstruction and novel view synthesis have been marked by two prominent techniques: Neural Radiance Fields (NeRF) and Gaussian Splatting. While NeRF employs an implicit representation using neural networks to model 3D scenes with high fidelity, Gaussian Splatting utilizes an explicit representation with 3D Gaussian primitives, offering real-time rendering capabilities. In either of these two methods, we have to make a tradeoff between rendering speed and surface mesh accuracy. This project explores the extension of TetSphere Splatting, a method similar to Gaussian splatting that preserves surface mesh quality, with volumetric rendering to enable rendering of 3D reconstructions with internal surfaces, a limitation of the current 3D Gaussian Splatting approach. The combination of TetSpheres with volumetric rendering can potentially give us high-quality reconstructions like NeRF while offering faster rendering times similar to Gaussian Splatting. A robust and fast 2D-3D reconstruction algorithm paves the way for generation of 3D models that can be used to train 3D generative models at scale.

Through this SuperUROP project, I hope to delve deeper into computer graphics while applying some of the machine learning and computer vision theory I’ve learned from my previous classes and UROP experiences. The combination of optimization theory and the use of lower-level graphics algorithms to achieve more performant solutions is truly exciting to me. It is an excellent application of many of the concepts of software performance I’m currently learning about to do very impactful research. Additionally, I want to explore the uncertainty that comes with doing novel research and learn valuable experimentation and research techniques from my supervisor, ultimately culminating in a useful research paper.