Face Parsing via SegNeXt

Introducion

本项目在 CelebAMask-HQ 数据集上训练了用于人脸解析（人脸分割）的SegNeXt模型。在指标表现上远远优于之前常用的 BiSeNetv2 模型。
依据下面的步骤可以方便的使用该模型，获得良好的人脸解析结果。
The project provides SegNeXt for face parsing, trained on the CelebAMask-HQ dataset.
The results are as follows, which are much better than previously widely-used BiSeNetv2.

IoU of 19 classes

Model	skin	nose	eye_g	l_eye	r_eye	l_brow	r_brow	l_ear	r_ear	mouth	u_lip	l_lip	hair	hat	ear_r	neck_l	neck	cloth	background
SegNeXt	93.69	89.29	87.12	82.59	82.55	76.81	76.75	80.86	79.30	87.74	82.41	84.83	91.98	81.80	57.74	22.07	84.88	80.70	93.87
BiSeNetv2	92.79	88.40	83.51	34.72	33.13	35.91	25.45	43.11	4.26	83.26	78.30	82.06	90.58	74.23	46.40	0	82.04	71.81	92.15

mIou and mAcc of all classes

Model	mIou	mAcc
SegNeXt	79.83	86.63
BiSeNetv2	60.12	69.94

Examples of result images with labels

Contributors

• Beyondzjl
• fei-aiart

Prerequisites

• Linux
• Python >=3.6
• Anaconda or miniconda

Quick start

Preparation

• Clone this repo

  git clone https://github.com/Beyondzjl/segmentation-CelebAMask-HQ-SegNeXt.git
  cd segmentation-CelebAMask-HQ-SegNeXt
  

• Download dataset CelebAMask-HQ from Google Drive.
  I have divided the original dataset into following structure.

  CelebAMask-HQ
  |
  |-train
  |      |-images
  |      |-labels
  |-test
  |      |-images
  |      |-labels
  |-val
  |      |-images
  |      |-labels
  

• Prepare OpenMMLab dependences

  pip install -U openmim
  mim install mmcv-full==1.6.0
  pip install timm
  

• Prepare project dependences

  pip install -r requirements.txt
  

Test

• Get train-weight from Google Drive.

• Run
  

  export PYTHONPATH="${PYTHONPATH}:/xxx/mmsegmentation-master" ##指示所有文件的搜索范围
  python tools/test.py ${配置文件} ${检查点文件} [--out ${结果文件}] [--eval ${评估指标}]
  

  For example:
  

  python /xxx/segmentation-CelebAMask-HQ-SegNeXt/tools/test.py /xxx/segmentation-CelebAMask-HQ-SegNeXt/mysegconfig/segnext_CelebAMask_test.py /xxx/segmentation-CelebAMask-HQ-SegNeXt/iter_160000.pth --eval mIoU
   ## give you the evalution results
  python /xxx/segmentation-CelebAMask-HQ-SegNeXt/tools/test.py /xxx/segmentation-CelebAMask-HQ-SegNeXt/mysegconfig/segnext_CelebAMask_test.py /xxx/segmentation-CelebAMask-HQ-SegNeXt/iter_160000.pth --show-dir <results_path> --gpu-id x
   ## save results to the path（覆盖原图的结果）
  python /xxx/segmentation-CelebAMask-HQ-SegNeXt/tools/test.py /xxx/segmentation-CelebAMask-HQ-SegNeXt/mysegconfig/segnext_CelebAMask_test.py /xxx/segmentation-CelebAMask-HQ-SegNeXt/iter_160000.pth  --show-dir <results_path> --opacity 1
   ## 得到单独的分割结果（没有原图）
  

• Tips
  If you want to use your own dataset, you need to write new config giving the proper form and path of your dataset. You can get the example of config from mysegconfig.

Train

• Get pretrain model from Google Drive.
• Run with one GPU
  

  python tools/train.py ${CONFIG_FILE} [optional parameters]
  

Acknowledgments

The project is based on OpenMMLab. Thanks for the excellent work of OpenMMLab, CelebAMask-HQ and SegNeXt.