mtcnn-train-gluon

Implement MTCNN train pipeline with MXNet gluon

Introduction

This repo mainly implement MTCNN(Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks) with mxnet gluon framework. I train a faster network instead of original version(only replace pooling with conv + stride), just design your own network and you may get faster and more accurate face detection model.

How to run

python MTCNN.py

How to train

The WIDER-FACE dataset is needed, download and unzip to datset directory such as:

---dataset   
  ---WIDER_FACE
    ---wider_face_split    
    ---WIDER_train  
    ---WIDER_val   

First train PNet, then RNet and finally ONet. The dataset process part reference at repo. Steps list:
1. Generate PNet data

python generate_data\Generate_PNet_data.py

2. Train PNet

python train_pnet.py

3. Generate RNet data with trained PNet

python generate_data\Generate_RNet_data.py

4. Train RNet

python train_rnet.py

5. Generate ONet data with trained PNet and RNet

python generate_data\Generate_ONet_data.py

6. Train ONet

python train_onet.py

Demo

Reference

Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks
mtcnn-pytorch