Georgios Kopanas
Point-Based Representations for Novel View Synthesis
Research Scientist at Google
May 15, 2024 (Wed), 4:00 p.m. KST
Online (Zoom).
Guest Lecture at CS580: Computer Graphics
Minhyuk Sung, KAIST, Spring 2024
Abstract
Neural Radiance Fields(NeRFs) have revolutionized the field of Novel View Synthesis and 3D reconstruction from 2D images. The outstanding speed with which variants and improvements of the original NeRF paper are being published makes a concrete statement regarding its impact on the research community. The underlying representations that store the volumetric information of the scene have come a long way from the original paper. The community seems to converge on the consensus that a combination of a grid-based representation together with a shallow Neural Component provides a good trade-off between speed and quality. This consensus seems to be challenged from our recent work that focuses on point-based representations that encode the volumetric information in point samples. Most notably "3D Gaussian Splatting" shows that a point-based representations can surpass by a big margin all grid-based representations invented so far both in speed and quality, while using traditional graphics techniques and omitting all neural components. In this talk we are going to introduce point-based representations for Novel View Synthesis and how they addressed the different trade-offs between the representations of Radiance Fields.
Bio
Georgios Kopanas recently joined Google as a Research Scientist based in London; he also holds a PhD from Universite Cote d'Azur. He was first exposed to Machine Learning during an internship at SRI International where he worked in unsupervised object detection. Later he joined Inria as a research engineer contributing to multiple projects involving 3D Graphics. He then worked on drivers for mobile GPUs at ARM as a Software Engineer. In 2020, he returned to Inria as a PhD candidate under G.Drettakis, where he focused on point-based representations for Novel View Synthesis and he is one of the lead authors of the SIGGRAPH best paper award "3D Gaussian Splatting for Real-Time Radiance Field Rendering.”