# cosmo-face

**Repository Path**: syshensyshen/cosmo-face

## Basic Information

- **Project Name**: cosmo-face
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2020-07-23
- **Last Updated**: 2020-12-19

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# cosmo-face

## face detection and landmark detection

+ source data  
  https://pan.baidu.com/s/1jIp9t30oYivrAvrgUgIoLQ  
  提取码：ruck  
  annotations(face bounding boxes & five facial landmarks)
  https://pan.baidu.com/s/1Laby0EctfuJGgGMgRRgykA
  
+ pretrain model  
  https://pan.baidu.com/s/12h97Fy1RYuqMMIV-RpzdPg  
  提取码：fstq  
  
+ prepared data  
  + 具体的数据集 data/ 文件夹结构如下： 
    + widerface/  
      + train/  
        + images/  
          + 0-Parade/  
            + 0_Parade_marchingband_1_5.jpg  
        + label.txt  
      + val/  
        + images/  
          + 0-Parade/  
            + 0_Parade_marchingband_1_20.jpg  
        + wider_val.txt  
    + FDDB/  
      + images/  
        + 2002/
          + 07/ 
            + 19/
              + big/
                + img_18.jpg
      + img_list.txt
      
+ train data
  + centerface
    + 运行 train_centerface.py (backbone 支持 resnet18, resnet50, ghost, RFB)  
    
  + retinaface
    + 运行 train_retina.py (backbone 支持 mobilenet0.25, resnet50, ghost, RFB, slim)  
    
+ inference data  
  + centerface  
    + 运行 widerface_test_centerface.py (支持单张图片和文件夹)  
  + retinaface
    + 运行 widerface_test_retina.py (支持单张图片和文件夹)
    
+ calculate mAP
  + 计算某个模型的mAP步骤：  
    1. 运行 widerface_test_centerface.py  
       widerface_evaluate/文件夹 结构如下:  
       + ground_truth/
         + wider_easy_val.mat  
         + wider_face_val.mat  
         + wider_hard_val.mat  
         + wider_medium_val.mat  
       + widerface_txt/
         + 24--Soldier_Firing/  
           + 24_Soldier_Firing_Soldier_Firing_24_10.txt  
    2. 运行 ./widerface_evalueta/evaluation.py
         
      
## face recognition

+ source data   
  https://pan.baidu.com/s/1S6LJZGdqcZRle1vlcMzHOQ

+ prepared data   
  ` python prepare_data.py `  
  + 具体的数据集 data/ 文件夹结构如下：  
    + faces_emore/  
      + agedb_30/  
      + calfw/  
      + cfp_ff/  
      + cfp_fp/  
      + cplfw/  
      + imgs/  
      + lfw/  
      + vgg2_fp/  
      + agedb_30_list.npy  
      + agedb_30.bin  
      + calfw_list.npy  
      + calfw.bin  
      + cfp_ff_list.npy  
      + cfp_ff.bin  
      + cfp_fp_list.npy  
      + cfp_fp.bin  
      + cplfw_list.npy  
      + cplfw.bin  
      + lfw_list.npy  
      + lfw.bin  
      + property  
      + train.idx  
      + train.rec  
      + vgg2_fp_list.npy  
      + vgg2_fp.bin  

+ train data  
  
  ` python train.py -net ir_se -depth 50 `  
  + -net ir_se -depth 50  设置网络是ir_se50 (建议), 可调整;
  + -lr 0.1 设置初始学习率是0.1 (建议), 可调整;

+ inference data  
  1. 先将人脸数据放在./recognition/data/faces_database/data 文件夹下  
  2. 运行 ./alignment/face_align.py 产生人脸对齐后的数据  
  3. 如果有新的人脸数据, 需要先更新人脸库, 需设置args.update = True  
  4. 运行./recognition/face_verify.py