A Review of Scene Representations for Novel View Synthesis

PhD Qualifying Examination


Title: "A Review of Scene Representations for Novel View Synthesis"

by

Mr. Li MA


Abstract:

Novel view synthesis (NVS) is the task of interpolating or extrapolating 
views given images captured from one or multiple views in a scene. A 
common paradigm for NVS is to first construct an intermediate scene 
representation, and then render novel views with this scene parameters, 
which involves both computer vision (CV) technologies of image 
understanding and computer graphics (CG) that tries to render 
photorealistic images. One vital aspect that distinguishes most NVS 
algorithms and builds the bridge between CV and CG is how to parameterize 
the plenoptic function, i.e., the scene representation. In the past 
decade, explicit geometry and appearance modeling methods such as textured 
polygon meshes and point clouds have been well explored but they usually 
suffer from imperfect scene reconstruction due to complex geometry or 
view-dependent effect. With the emerging of neural network, there is a 
growing body of research into modeling the physical world incorporate with 
neural representations. Modeling appearance with neural textures 
compensates for broken reconstruction and can thus achieve better visual 
quality. We also review the literature on implicit scene representation. 
Particularly, the neural radiance field (NeRF), as a focused research 
area, is described in detail, along with an investigation of several 
follow-up works. To sum up, in this survey we systematically review 
existing works on novel view synthesis algorithms and particularly focus 
on the scene representations they use, so as to provide a guidance for 
designing new NVS algorithms or choosing the proper existing NVS 
algorithm.


Date:			Tuesday, 10 August 2021

Time:                  	2:00pm - 4:00pm

Zoom meeting:
https://hkust.zoom.us/j/95562010105?pwd=ejNJSXdhTVZ3QmlRM21zMElzRFphZz09

Committee Members:	Prof. Pedro Sander (Supervisor)
 			Prof. Chiew-Lan Tai (Chairperson)
 			Dr. Qifeng Chen
 			Dr. Xiaojuan Ma
 			Prof. Huamin Qu


**** ALL are Welcome ****