fashion_attgan's Introduction

Fashion-AttGAN: Attribute-Aware Fashion Editing with Multi-Objective GAN

This repo provides the dataset used in our paper "Fashion-AttGAN: Attribute-Aware Fashion Editing with Multi-Objective GAN" (https://arxiv.org/abs/1904.07460).

Dataset

The dataset contains 14,221 images and corresponding predicted attribute values. Currently we publish two attribute categories among others: clothing colors and sleeve lengths (data/data.json). 22 attribute values in total are selected in this published dataset. Therefore each image is associated with 1 color value and 1 sleeve-length value (these two values can be empty since there are other values of these two categories that are not covered in this dataset).

The image data is adapted from the VITON dataset, and can be downloaded here

Result

The results of our proposed Fashion-AttGAN and the original AttGAN are as follows:

Figure 1: Clothing Attribute-Editing Results. Top: Attribute-Editing Results of AttGAN. Bottom: Attribute-Editing Results of Fashion-AttGAN. From left to right columns:(1) original image, (2) reconstructed image, (3-6)varied sleeve lengths, (7-24) varied colors.

Train

Prerequisites

Tensorflow

Data Preparation

Put all images under /data.

Training

bash train.sh

Cite

To use or cite this dataset, please use the following:

Qing Ping, Bing Wu, Wanying Ding, Jiangbo Yuan. 2019. Fashion-AttGAN: Attribute-Aware Fashion Editing with Multi-Objective GAN. FFSS-USAD Workshop at CVPR 2019.

Statistics

The statistics of different attribute values are summarized as follows:

Attribute Category	Attribute Values	Frequency
Colors	black	3581
	white	3264
	red	1870
	grey	1649
	navy-blue	1291
	blue	945
	pink	778
	green	696
	orange	252
	purple	217
	yellow	153
	dark-brown	85
	beige	54
	apricot	45
	camel-brown	41
	sky-blue	31
	medium-brown	26
	light-brown	15
Sleeve Length	long sleeve	4151
	no sleeve	2137
	cap sleeve	468
	below elbow/three-quarter/seven-eighths sleeve	152

Recommend Projects