Abstract

We present Poly-GAN, a novel conditional GAN architecture that is motivated by different Image generation and manipulation applications like Fashion Synthesis, an application where garments are automatically placed on images of human models at an arbitrary pose, image inpainting, an application where we try to recover a damaged image using the edges or a rough sketch of the image. While different applications use different GAN setup for image generation, we propose only one architecture for multiple applications with little to no change in the pipeline. Poly-GAN allows conditioning on multiple inputs and is suitable for many different tasks. Our novel architecture enforces the conditions at all layers of the encoder and utilizes skip connections from the coarse layers of the encoder to the respective layers of the decoder. Coarse layers are easier to manipulate in shape change using condition, which results in higher level change in the result. Our system achieves state-of-the-art quantitative results on Fashion Synthesis based on the Structural Similarity Index metric and Inception Score metric using the DeepFashion dataset. For the image inpainting task we are achieving competitive results compared to current state of the art methods.

Library of Congress Subject Headings

High performance computing; Neural networks (Computer science); Machine learning; Computer architecture

Publication Date

12-2019

Document Type

Thesis

Student Type

Graduate

Degree Name

Computer Engineering (MS)

Department, Program, or Center

Computer Engineering (KGCOE)

Advisor

Andreas Savakis

Advisor/Committee Member

Raymond Ptucha

Advisor/Committee Member

Alexander Loui

Campus

RIT – Main Campus

Plan Codes

CMPE-MS

Download

Share

COinS