Files in this item



application/pdfQIN-THESIS-2019.pdf (2MB)Restricted Access
(no description provided)PDF


Title:Analysis of loss functions in XVAE-GAN: A novel two-view image generation network
Author(s):Qin, Zhen
Advisor(s):Schwing, Alexander
Department / Program:Electrical & Computer Eng
Discipline:Electrical & Computer Engr
Degree Granting Institution:University of Illinois at Urbana-Champaign
Computer Vision
Generative Model
Representation Learning
Image-to-Image Translation
Deep Learning
Abstract:Generating novel views from 2D images has been a popular topic in computer vision given its usefulness in a wide range of applications. However, the task is challenging due to its ill-posed nature. Recent attempts have been extensively focusing on modeling the transition between different views as an image-to-image translation problem using deep neural networks inspired from the breakthroughs of novel generative models such as generative adversarial networks (GAN). However, very few if any existing works provide an insight on the relation between representations learned in the latent spaces from different views. To complement this missing aspect in the problem space, we introduce a novel two-stream network based on variational autoencoders and GANs (XVAE-GAN) as an attempt to disentangle common features shared between two views and private features distinctively belong to each view in a pairwise two-view image synthesis setting. This thesis presents a survey on existing works targeting on multi-view image synthesis problem, then introduces our newly proposed XVAE-GAN network in detail. The rest of the thesis is focused on ablation study investigating the impact of different loss functions on our proposed model by experimenting with three different applications: face image rotation, frontal/lateral chest X-ray image synthesis and ground/aerial street-view synthesis.
Issue Date:2019-12-09
Rights Information:Copyright 2019 Zhen Qin
Date Available in IDEALS:2020-03-02
Date Deposited:2019-12

This item appears in the following Collection(s)

Item Statistics