Diploma Thesis — GANs for Pose & Style Selection in Fashion

By Athanassios Charisoudis· Supervisor: Prof. Pericles Mitkas, Dr. Antonios Chrysopoulos· Type: courseFeatured
GANsComputer VisionImage-to-ImageStyle Transfer
Diploma Thesis — GANs for Pose & Style Selection in Fashion

Highlights

  • Supervisors: Prof. Pericles Mitkas, Dr. Antonios Chrysopoulos
  • Conducted qualitative user study and ablations.

Abstract

Generative Modelling, a branch of Machine Learning that focuses on generating realistic looking samples, has traditionally constituted the upper bound of what Machine and Deep Learning models can achieve. This regime has completely changed the past years, especially after 2014, when I. Goodfellow presented his idea for a generative model comprising two competing neural networks: the Generative Adversarial Network of GAN for short. Subsequently, a plethora of models based on GAN have been proposed with impressive results, some of which, principally in the context of image generation, surprise even an experienced human vision system.

Concurrently, more and more research is devoted during the last decades around the development of techniques for demystifying the notion of fashion and fashion trends. Among its purposes, is creating artificial intelligence systems that provide help in the process of designing new garments as well as in the process of conducting better and more well targeted purchases. In an endeavour to apply modern machine learning techniques to automate generation and editing of fashion images, in this project we employ Generative Adversarial Networks. In particular, we design and utilize a multi-tool for automatic editing of fashion images, equipped with four (4) fundamental operations: pose change, cloth extraction, style matching and on-demand realistic fashion images generation.

In order to achieve our goals, we train four models based on the Generative Adversarial Network in fashion image (i.e. images of garments as well as human models advertising them) datasets, giving the corresponding outcomes at the end. It is our firm belief that further developments of such models will play a central role in fashion design and especially in clothes distribution through e-commerce systems in the near future, which has made us focus zealously on implementing an effective intelligent tool for fashion image editing in this work.

Keywords

Generative Modelling, Intelligent fashion systems, Generative Adversarial Networks, Noise-to-image generation, Image-to-image translation, DeepFashion, StyleGAN, CycleGAN, Computer Vision, Artificial Neural Networks

Report

PDF preview isn’t supported in this browser. Download the file to view it locally.