two stages strategies\ stage one: each gt matched anchor number:7\ stage one: each gt matched anchor number:2\ stage one: each gt matched anchor number:1\ stage one: each gt matched anchor number:1\ stage one: each gt matched anchor number:1\ stage one: each gt matched anchor number:3\ stage one: each gt matched anchor number:1\ stage one: each gt matched anchor number:9\ stage one: each gt matched anchor number:7\ stage one: each gt matched anchor number:9\ stage one: each gt matched anchor number:14\ the ground truth number:11\ the averge anchors matched number:5\ deal with tiny and outer face\ stage two: each gt matched anchor number:7\ stage two: each gt matched anchor number:5\ stage two: each gt matched anchor number:5\ stage two: each gt matched anchor number:5\ stage two: each gt matched anchor number:5\ stage two: each gt matched anchor number:5\ stage two: each gt matched anchor number:5\ stage two: each gt matched anchor number:9\ stage two: each gt matched anchor number:7\ stage two: each gt matched anchor number:9\ stage two: each gt matched anchor number:14

S3FD: Single Shot Scale-invariant Face Detector

Getting started

• You will need python modules: cv2, matplotlib and numpy. If you use mxnet-python api, you probably have already got them. You can install them via pip or package managers, such as apt-get:

  sudo apt-get install python-opencv python-matplotlib python-numpy

  Note The scale compensation anchor matching strategy is written into multibox_target.cu

• Copy multibox_target_operator/multibox_target.cc multibox_target.cu to mxnet/src/operator/contrib to cover original multibox_target.cc multibox_target.cu

• Build MXNet: Follow the official instructions

Train the model

This example only covers training on Wider Face dataset. Other datasets should

• Download the converted pretrained vgg16_reduced model here, unzip .param and .json files into model/ directory by default.
• Download the Wider Face dataset, skip this step if you already have one.
• Extra data in data/widerface/WIDER_train WIDER_val wider_face_split
• Convert voc format: cd data/widerface python widerface_voc.py
• Convert .rec data python tools/prepare_dataset.py --dataset widerface --set train --target ./data/train.lst python tools/prepare_dataset.py --dataset widerface --set val --target ./data/val.lst --shuffle False

• Start training:

  
  python train.py

NOTE!!!!!!!!

By default,this example use data_shape=608, if you have enough GPU memory, you should set data_shape=640.

I only have one GTX1080, so I don't have enough time to train.I use 0.001 learning rate for 7 epochs, the mAP achieved 64% in all validation set.You can try it at will.