Real-time Rendering Optimization with Gaussian Splatting
Session Number
CMPS 32
Advisor(s)
Dr. Amit Trivedi, University of Illinois at Chicago
Discipline
Computer Science
Start Date
17-4-2025 10:45 AM
End Date
17-4-2025 11:00 AM
Abstract
Three-dimensional Gaussian Splatting (3D GS) is a state-of-the-art rendering method designed to optimize computational efficiency in three-dimensional scene representation by using learnable Gaussians. Its predecessor, Neral Radiance Fields (NeRF), offers high accuracy but is computationally intensive. 3D GS reduces the required computational power by simplifying NeRF’s fivedimensional coordinate system into Gaussians that represent color and density. This research investigates the 3D GS’s tolerance to optimization by implementing and testing Knowledge Distillation and a custom Loss Function to refine Gaussian weight distributions during rendering. We compare these optimization implementations against baseline tests to evaluate their effect on rendering accuracy and computational efficiency. Our findings contribute to the broader discussion on optimizing three-dimensional rendering techniques and future advancements in machine learning-driven visualization.
Real-time Rendering Optimization with Gaussian Splatting
Three-dimensional Gaussian Splatting (3D GS) is a state-of-the-art rendering method designed to optimize computational efficiency in three-dimensional scene representation by using learnable Gaussians. Its predecessor, Neral Radiance Fields (NeRF), offers high accuracy but is computationally intensive. 3D GS reduces the required computational power by simplifying NeRF’s fivedimensional coordinate system into Gaussians that represent color and density. This research investigates the 3D GS’s tolerance to optimization by implementing and testing Knowledge Distillation and a custom Loss Function to refine Gaussian weight distributions during rendering. We compare these optimization implementations against baseline tests to evaluate their effect on rendering accuracy and computational efficiency. Our findings contribute to the broader discussion on optimizing three-dimensional rendering techniques and future advancements in machine learning-driven visualization.