Palette View Synthesis - Novel View Synthesis using Diffusion Probabilistic Modelling

Loading...
Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

Sähkötekniikan korkeakoulu | Master's thesis

Date

2023-12-11

Department

Major/Subject

Signal Processing and Data Science

Mcode

ELEC3049

Degree programme

CCIS - Master’s Programme in Computer, Communication and Information Sciences (TS2013)

Language

en

Pages

39+3

Series

Abstract

Novel view synthesis is a class of computer vision problems, in which one or multiple views of a scene or an object are provided. The goal is then to produce novel, previously unseen views of the given scene or object. Recently, the endeavors to solve such problems have gained significant traction in the generative deep learning domain. From Neural Radiance Field (NeRF) based approaches to encoder-decoder style architectures, various ways of performing novel view synthesis have been previously introduced. This work introduces Palette View Synthesis, an end-to-end diffusion probabilistic generative modelling approach for performing novel view synthesis which aims to resolve the drawbacks of previous approaches by extending the model's abilities to generalize across multiple classes, given only a single view and a target angle of the object as inputs, while simultaneously maintaining the quality of the generated samples. It shows that by employing a diffusion-based model, with a simple U-Net backbone that parameterizes the denoising function, and concatenation along the input channel dimension as a form of conditioning, it is possible to produce high quality, believable novel views while simultaneously generalizing across multiple different classes.

Description

Supervisor

Ilin, Alexander

Thesis advisor

Deny, Stéphane

Keywords

novel view synthesis, diffusion probabilistic modelling, generative modelling, deep learning, mental rotation

Other note

Citation