A SURVEY ON GENERATIVE NEURAL RADIANCE FIELDS

PhD Qualifying Examination


Title: "A SURVEY ON GENERATIVE NEURAL RADIANCE FIELDS"

by

Mr. Hao OUYANG


Abstract:

Generating photorealistic 3D-consistent images has been a challenge in computer 
vision community. Although fast progress has been made on 3D-aware image 
synthesis using generative models and neural rendering, the inconsistent 
rendering and the nonexpressive 3D representations results in low image 
quality. A new 3D representation, neural radiance field which represents a 3D 
scene using a multi-layer perceptron with volume rendering, has greatly 
improved the quality of the generation. The key idea is to learn to generate a 
neural radiance field from limited multi-view supervision or only single-view 
2D images. In this survey, we present a comprehensive review of generative 3D 
models using neural radiance fields, which has achieved stunning visual 
quality. We first introduce the prior works of 3D generation without radiance 
fields and then provide a summary of state-of-the-art. We then categorize the 
methods into fully 3d-consistent and hybrid types. We will also describe the 
possible applications such as 3D Gan inversion and editing based on the 
generative radiance fields. Finally, we conclude the survey with the discussion 
of general unsolved problem and future research directions.


Date:  			Monday, 4 July 2022

Time:                  	10:00am - 12:00noon

Zoom Meeting:		https://hkust.zoom.us/j/2693115905

Committee Members:	Dr. Qifeng Chen (Supervisor)
 			Prof. Chiew-Lan Tai (Chairperson)
 			Dr. Xiaojuan Ma
 			Dr. Dan Xu


**** ALL are Welcome ****