AI - face

Information

1. MTCNN Face detection
2. insightface Face recognition (mainly based on MXNet)
3. convert mxnet to caffe

insightface

Environmental installation

1. pip3：sudo apt install python3-pip
2. numpy：pip3 install --user numpy scipy matplotlib
3. pip install mxnet-cu101
   [1] cuda The version is 10.1
4. pip install opencv-python
5. insightface download

Training

Before processing

1. Detect alignment based on detection network , Such as MTCNN
2. Remove wearing sunglasses based on attribute network 、 Mask
3. The deletion angle is greater than 60 Degree
4. Delete id Number of files less than 5 Of

pack

1. mxnet Scripts in the framework im2rec.py
2. Generate a list of files in the folder (name.lst)
   (1) python im2rec.py --list --recursive name data_dir
3. Generate binaries (name.rec, name.idx)
   (1) python im2rec.py name.lst data_dir

Training

1. Modify configuration information config.py

dataset.emore.dataset_path = './alg_test'  #  Default name  train.rec, train.idx
dataset.emore.num_classes = 1193
dataset.emore.image_shape = (128, 128, 3)
 
default.end_epoch = 200
default.per_batch_size = 40
default.ckpt = 3
default.lr_steps = '65682,112680,150240,187870'
default.models_root = './models'

2. Training orders
   CUDA_VISIBLE_DEVICES=‘0’ python -u train.py --network r100 --loss arcface --dataset emore