PyramidBox: A Context-assisted Single Shot Face Detector

Introduction

A MXNet implementation of PyramiBbox:A Context-assisted Single Shot Face Detector.
If you want to learn more details,please refer to the original paper.

@inproceedings{Tang2018PyramidBoxAC,
  title={PyramidBox: A Context-Assisted Single Shot Face Detector},
  author={Xu Tang and Daniel K. Du and Zeqiang He and Jingtuo Liu},
  booktitle={ECCV},
  year={2018}
}

I train PyramidBox with WIDER FACE dataset,results are as follows:

I think mainly reasons that this repo can not get the same precision as paper as follows:

• I use batch size 4 because of memory limitations,which is 16 in the paper
• some parameters are not metioned in the paper

Here are several examples of succesful detection outputs:

Details

I implement following structures metioned in the paper:

• Low-Level FPN
• max-in-out layer
• PyramidAnchors
• Context-sensitive Prediction Module
• Data-anchor sampling
• Learning rate warmup and cosine decay

  Dependencies

• Python 3.6.4
• MXNet 1.3.1
• gluon-cv 0.4.1

Implement Details

• Ubuntu 16.04 LTS
• CUDA 9.0
• cuDNN 7.0.5

Preparation

git clone git@github.com:JJXiangJiaoJun/gluon_PyramidBox.git
cd gluon_PyramidBox

Download and prepare data

1. download WIDER FACE dataset into widerface/downloads

   $$ tree widerface/downloads
   widerface/downloads
   ├── eval_tools.zip
   ├── Submission_example.zip
   ├── wider_face_split.zip
   ├── WIDER_test.zip
   ├── WIDER_train.zip
   └── WIDER_val.zip

2. Prepare data: unzip data, annotations and eval_tools to ./widerface

   python tool/prepare.py
   $$ tree widerface -L 1
   widerface
   ├── downloads
   ├── eval_tools
   ├── wider_face_split
   ├── WIDER_train
   └── WIDER_val

3. Prepare custom val dataset for quick validation (crop and resize to 640)

   $$ python tool/build_custom_val.py
   $$ tree widerface -L 1
   widerface
   ├── downloads
   ├── eval_tools
   ├── WIDER_custom
   ├── wider_face_split
   ├── WIDER_train
   └── WIDER_val

Train on WIDER FACE Datasets

train vgg16 based pyramidbox with 1 gpus.I only implement VGG16 as backbone currently:

python train_end2end.py --use-bn

or you can see more details:

python train_end2end.py --help

Evalution

eval your own model on WIDER FACE Datasets:

python eval.py --use-bn --model models/pyramidbox/pyramidbox_best.params

Reference

• PyramidBox with pytorch
• sfd.gluoncv