基于mxnet的ssh人脸检测算法（改进版）

2019.02.27: 初始版本，优化误检

1.介绍

1. Reproduce SSH (Single Stage Headless Face Detector) with MXNet.

2. Original Caffe code: [https://github.com/deepinsight/insightface/tree/master/SSH]

3. Evaluation on WIDER FACE（原版的结果，改进后的没做测试，应该会更好）:

Impelmentation	Easy-Set	Medium-Set	Hard-Set
Original Caffe SSH	0.93123	0.92106	0.84582
Our SSH Model	0.93489	0.92281	0.84525

4. Evaluation on fddb = 98.7%

5. 优化原始版本边缘人脸漏检

2.安装

环境

ubuntu16.04 cuda cudnn mxnet以及python的依赖项等

配置

1. Type make to build necessary cxx libs（需要更改python版本时需在Makefile修改对应py的版本）

2. Download MXNet VGG16 pretrained model from here and put it under model directory.

3. 编译，在rcnn/config.py里修改参数配置

config.BBOX_MASK_THRESH = 20 #add mask with in train for little size faces
# config.COLOR_JITTERING = 0.125
config.COLOR_JITTERING = 0 # add augmentation for bright and so on

config.TEST.SCORE_THRESH = 0.5

# scale changed as smallhard face
config.TEST.SCALES = [50, 500, 1000]
config.TEST.PYRAMID_SCALES = [0.75, 1.5]

default.base_lr = 0.004
default.e2e_epoch = 40

3.Training

python train_ssh.py

算法对代码中blur>1, invalid>0, occlusion>1的人脸都加上mask，这样会减少误捡，但是同样造成漏检

人脸识别数据集组成

.
├── WIDER_train
|   └── images
│       ├── .... 
│       ├── .... 
└── ...