face-parsing 0.2

This repo is used to generate semantic segmentation for face image

README_ZH.md | README_EN.md

项目描述

该仓库用于生成人脸图像的语义分割。

安装

下载 https://drive.google.com/open?id=154JgKpzCPW82qINcVieuPH3fZ2e0P812 的权重文件 79999_iter.pth

# pip install -e git+https://github.com/sword4869/face_parsing.git#egg=face_parsing
pip install face_parsing

输入和输出

├── pretrain
│   └── 79999_iter.pth      # ckpt
├── test_img                # 输入
│   ├── 00000.jpg
│   └── 116_ori.png
└── test_res                # 输出
    ├── chosen_merge_00000.png      # 指定部分融合mask
    ├── chosen_merge_116_ori.png
    ├── merge_00000.png             # 融合mask
    ├── merge_116_ori.png
    ├── weighted_00000.png          # 叠加原图
    ├── weighted_116_ori.png
    ├── parsing_00000.png           # 分类结果，每个像素的值是[0, 18]
    ├── parsing_116_ori.png
    └── masks                       # 各部分mask
        ├── 00000
        │   ├── 00_background.png
        │   ├── 01_skin.png
        │   ├── 02_l_brow.png
        │   ├── 03_r_brow.png
        └── 116_ori
            ├── 00_background.png
            ├── 01_skin.png
            ├── 02_l_brow.png
            ├── 03_r_brow.png

usage: face_parsing [-h] [--res_path RES_PATH] [--img_path IMG_PATH] [--ckpt CKPT] [--chosen_parts CHOSEN_PARTS [CHOSEN_PARTS ...]] [--reverse] [--color_style {face-parsing-style,CelebAMask-HQ-style}]

options:
  -h, --help            show this help message and exit
  --res_path RES_PATH   results path
  --img_path IMG_PATH   data path
  --ckpt CKPT           checkpoint path
  --chosen_parts CHOSEN_PARTS [CHOSEN_PARTS ...]
                        chosen parts
  --reverse             reverse the chosen parts
  --color_style {face-parsing-style,CelebAMask-HQ-style}
                        color style

# 在face_parsing下
$ face_parsing

# 在face_parsing路径外
$ face_parsing --ckpt ~/79999_iter.pth --res_path ~/test_res --img_path ~/test_img

Index	Name	face-parsing.PyTorch Style RGB	CelebAMask-HQ Style RGB
0	background	[255, 0, 0]	[0, 0, 0]
1	skin	[255, 85, 0]	[204, 0, 0]
2	l_brow	[255, 170, 0]	[0, 255, 255]
3	r_brow	[255, 0, 85]	[255, 204, 204]
4	l_eye	[255, 0, 170]	[51, 51, 255]
5	r_eye	[0, 255, 0]	[204, 0, 204]
6	eye_g	[85, 255, 0]	[204, 204, 0]
7	l_ear	[170, 255, 0]	[102, 51, 0]
8	r_ear	[0, 255, 85]	[255, 0, 0]
9	ear_r	[0, 255, 170]	[0, 204, 204]
10	nose	[0, 0, 255]	[76, 153, 0]
11	mouth	[85, 0, 255]	[102, 204, 0]
12	u_lip	[170, 0, 255]	[255, 255, 0]
13	l_lip	[0, 85, 255]	[0, 0, 153]
14	neck	[0, 170, 255]	[255, 153, 51]
15	neck_l	[255, 255, 0]	[0, 51, 0]
16	cloth	[255, 255, 85]	[0, 204, 0]
17	hair	[255, 255, 170]	[0, 0, 204]
18	hat	[255, 0, 255]	[255, 51, 153]

face-parsing的模型效果

参考资料

folk from https://github.com/zllrunning/face-parsing.PyTorch, https://github.com/dw-dengwei/face-seg

1. Code: BiSeNet
2. Paper: BiSeNetV1
3. Paper: BiSeNetV2