MTCNN C++ Implementation

This is a C++ project to implement MTCNN, a perfect face detect algorithm, on different DL frameworks.
The most popular frameworks: caffe/mxnet/tensorflow, are all suppported now.

Build

• Bulid caffe, mxnet or tensorflow first   Please edit makefile.mk (set xxx_ON flags to enable corresponding dp framework) to select one or more to be supported

  • Build CaffeOnACL, refer to CaffeOnACL Release notes

  • Build MxnetOnACL, refer to MxnetOnACL release notes

  • Build tensorflow, to generate libtensorflow.so, please use:

    bazel build --config=opt //tensorflow/tools/lib_package:libtensorflow

    the tarball, bazel-bin/tensorflow/tools/lib_package/libtensorflow.tar.gz, includes the libtensorflow.so and c header files

• Install opencv package

• Edit Makefile to set CAFFE_ROOT, MXNET_ROOT or TENSORFLOW_ROOT to the right path in your machine. For example : CAFFE_ROOT=~/oaid/caffeOnACL.

• make -j4

Run

If the basic work is ready (build caffe/Mxnet/Tensorflow sucessfully) followed by above steps. You can run the test now.

1. Test on single picture:

./test -f photo_fname [ -t DL_type] [-s] 
  -f photo_fname  picture to be  detected
  -t DL_type      DL frame: "caffe" , "mxnet"(default) or "tensorflow"
  -s              Save face chop into jpg files

The new picture, which boxed face and 5 landmark points will be created and saved as "new.jpg"

2. Test on camera (DL Framework is caffe)

./camera -t caffe

Credit

MTCNN algorithm

https://github.com/kpzhang93/MTCNN_face_detection_alignment

MTCNN C++ on Caffe

https://github.com/wowo200/MTCNN

MTCNN python on Mxnet

https://github.com/pangyupo/mxnet_mtcnn_face_detection

MTCNN python on Tensorflow

FaceNet uses MTCNN to align face

https://github.com/davidsandberg/facenet

From this directory:

facenet/src/align