This repository contains the un-official reimplementation of several single shot faces detectors in TensorFlow.

• S³FD: Single Shot Scale-invariant Face Detector. Click here for official codes in Caffe (evaluation only).
• PyramidBox: A Context-assisted Single Shot Face Detector. Click here for official codes in PaddlePaddle.
• DAN(Dynamic Anchor routing Network): A novel framework try to conduct inplace cascaded detection for extreme tiny faces.

• Download WIDER Face Dataset and reorganize the directory as follows:

   WIDER Face Dataset/
      |->WIDER_train/
      |    |->images/
      |    |->...
      |->WIDER_val/
      |    |->images/
      |    |->...
      |->WIDER_test/
      |    |->images/
      |    |->...
      |->wider_face_split/
      |    |->wider_face_train.mat
      |    |->...
   '''
   WIDER_ROOT is your path of the WIDER Face Dataset.
  

• Run the following script to generate TFRecords.

   python dataset/convert_tfrecords.py --dataset_directory=WIDER_ROOT --output_directory=./dataset/tfrecords

• Download the pre-trained VGG-16 model (reduced-fc) from here and put them into one sub-directory named 'model' (we support SaverDef.V2 by default, the V1 version is also available for sake of compatibility).

• Compile custom operations, run following command in directory 'cpp/Deform' and 'cpp/ExtraLib':

   mkdir build
   cd build && cmake ..
   make

• Run the following script to start training:

   python train_dan.py (or python train_pb.py for PyramidBox, or python train_sfd.py for S³FD)

• Run the following script to get detection results in *.txt (modify the WIDER_ROOT in 'data_dir' firstly):

   python eval_dan.py (or python eval_pb.py for PyramidBox, or python eval_sfd.py for S³FD)

• Run the official MATLAB script for evaluation

All the codes was tested under TensorFlow 1.8, Python 3.5, Ubuntu 16.04 with CUDA 8.0. You need at least two decent GPUs for training by yourself.

The whole framework:

Low-level FPN:

Context module V2:

Context module V1:

Reproduced results are list as follows:

The performance of PyramidBox could be further improved by using transposed convolution in LFPN and tuning the weight of losses. While the training peroid is too long to perform more experiments.

Easy:

Medium:

Hard:

You can download the trained model of DAN from GoogleDrive.

Welcome to join in QQ Group(758790869) for more discussion

Apache License, Version 2.0