Skip to content

DanielXu123/FaceAttribute-FAN

 
 

Repository files navigation

FaceAttribute-FAN

This respository includes a Caffe implementation of FAN that achieves state-of-the-art performance on CelebA and LFWA face attribute benchmark.

Citation

If this code is helpful for your research, please cite the following paper:

@article{he2018harnessing,
  title={Harnessing Synthesized Abstraction Images to Improve Facial Attribute Recognition.},
  author={Keke He, Yanwei Fu, Wuhao Zhang, Chengjie Wang, Yu-Gang Jiang, Feiyue Huang, Xiangyang Xue},
  journal={Proceedings of the Twenty-Seventh International Joint Conference on
           Artificial Intelligence, {IJCAI-18}},
  pages={733-740}
  year={2018}
}

Introduction

Our method leverages facial parts locations for better attribute prediction. A facial abstraction is generated by a Generative Adversarial Network (GAN). Then a dual-path facial attribute network is built which accepts inputs from original images and absraction images.

Prerequisites

  • Caffe
  • Linux
  • NVIDIA GPU + CUDA CuDNN

Getting Started

Setup

Clone the github repository:

git  clone https://github.com/TencentYoutuResearch/FaceAttribute-FAN
cd FaceAttribute-FAN

Model

Please download the trained models from Google drive or Baidu drive, and put it into outputs folder.

Demo

To test the dual-path model,

sh demo_dual_path.sh

If you want to test your own image without synthesized abstraction image, you can

sh demo_single_path.sh

The name of 40 attributes can be found at Appendix

Dataset

  1. To train and evaluation the model on the CelebA benchmark, please download the CelebA dataset from CelebA.

  2. Please download pretrained model and synthesized abstraction images of CelebA dataset from Google drive or Baidu drive.

  3. Put the original CelebA and synthesized abstraction image under the data/CelebA folder, for example:

    ├── CelebA
    │   ├── img_align_celeba
    │   ├── img_celeba_pix2pix
    
    

Evaluation

cd evaluation
sh test_dual_path_celeba.sh

Training

cd models/dual_path_parse_resnet
sh train.sh

Appendix

40 binary attributes in CelebA dataset. Output 0: without this attribute, 1: with this attribute.

Id Name Name in Chinese
0 5_o_Clock_Shadow 短胡子
1 Arched_Eyebrows 弯眉毛
2 Attractive 有吸引力
3 Bags_Under_Eyes 眼袋
4 Bald 秃顶
5 Bangs 刘海
6 Big_Lips 厚嘴唇
7 Big_Nose 大鼻子
8 Black_Hair 黑色头发
9 Blond_Hair 金色头发
10 Blurry 模糊
11 Brown_Hair 棕色头发
12 Bushy_Eyebrows 浓眉毛
13 Chubby 胖的
14 Double_Chin 双下巴
15 Eyeglasses 眼镜
16 Goatee 山羊胡子
17 Gray_Hair 灰白头发
18 Heavy_Makeup 浓妆
19 High_Cheekbones 高颧骨
20 Male 男性
21 Mouth_Slightly_Open 嘴巴微张
22 Mustache 胡子,髭
23 Narrow_Eyes 小眼睛
24 No_Beard 没有胡子
25 Oval_Face 鸭蛋脸
26 Pale_Skin 皮肤苍白
27 Pointy_Nose 尖鼻子
28 Receding_Hairline 发际线后移
29 Rosy_Cheeks 红润双颊
30 Sideburns 连鬓胡子
31 Smiling 微笑
32 Straight_Hair 直发
33 Wavy_Hair 卷发
34 Wearing_Earrings 戴耳环
35 Wearing_Hat 戴帽子
36 Wearing_Lipstick 涂唇膏
37 Wearing_Necklace 戴项链
38 Wearing_Necktie 戴领带
39 Young  年轻

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 87.2%
  • Shell 12.8%