Face_regognition_only_cpp

Inspired by https://github.com/vinayakkailas/Face_Recognition in python (for a working update, check my fork: https://github.com/timombre/Face_Recognition)

Face recognition using Tensorflow C++ API and OpenCV

Requirements :

1. Tensorflow built from sources :
Follow official guidelines
https://www.tensorflow.org/install/install_sources

Generate shared library: (bazel build //tensorflow:libtensorflow_cc.so)

2. OpenCV and contribs installed:

git clone https://github.com/opencv/opencv.git
cd opencv 
git checkout 3.3.1 
cd ..
git clone https://github.com/opencv/opencv_contrib.git
cd opencv_contrib
git checkout 3.3.1
cd ..
cd opencv
mkdir build
cd build
cmake -D CMAKE_BUILD_TYPE=RELEASE \
  -D CMAKE_INSTALL_PREFIX=/usr/local \
  -D INSTALL_C_EXAMPLES=ON \
  -D INSTALL_PYTHON_EXAMPLES=ON \
  -D WITH_TBB=ON \
  -D WITH_V4L=ON \
  -D WITH_QT=ON \
  -D WITH_OPENGL=ON \
  -D OPENCV_EXTRA_MODULES_PATH=../../opencv_contrib/modules \
  -D BUILD_JPEG:BOOL=ON -DWITH_JPEG:BOOL=ON\
  -D BUILD_EXAMPLES=ON ..
make
sudo make install
sudo sh -c 'echo "/usr/local/lib" >> /etc/ld.so.conf.d/opencv.conf'
sudo ldconfig

Works on Unix systems (Not tested on Mac though)

Using :

In Makefile, modify TF_BUILDDIR ?= with your Tensorflow directory.

make

Download Tensorflow weightfile (The accuracy is better with older 20170512-110547.pb than with 20180408-102900.pb), Landmarks and OpenCV Haarcascade:

wget https://gitlab.fit.cvut.cz/pitakma1/mvi-sp/raw/eb9c9db755077bd6fe0a61c1bbb1cced5f20d6d1/data/20170512-110547/20170512-110547.pb
wget https://raw.githubusercontent.com/opencv/opencv/master/data/haarcascades/haarcascade_frontalface_alt2.xml
wget https://raw.githubusercontent.com/kurnianggoro/GSOC2017/master/data/lbfmodel.yaml

Create a dataset of faces for each person :

./databasegenerator DATABASEFOLDER NAMEFORLABEL # you can tune the snapshot rate with the flag -fps "frequency", default is 1

Note : This is just for code testing, it is a very bad way to create a dataset as all pictures will have the same background and global features. If you already have independant pictures use them.

Arrange your dataset in below order :

root folder  
│
└───Person 1
│   │───IMG1
│   │───IMG2
│   │   ....
└───Person 2
|   │───IMG1
|   │───IMG2
|   |   ....

Create embeddings data base and run code :

./embed_database PATH_TO_YOUR_DATA_DIRECTORY 
# you can add the -gen_aligned_db flag to generate the intermediary aligned database
# you can also randomly (while conserving the global ratio within labels) split your dataset -splitdb "percentage" 
# and -filesname to generate the list of the files ending in either dataset

./face_recognition face_embeddings_database.txt
# you can add the -show_crop flag if you want to display the analyzed sub pictures
# you can also tune the minimum display distance -thresh "float" (default is set at 0.4) and relative probability threshold -rp "float" (default is 1.25)
# -mindist will give the original behaviour based on the minimum distance and not average and variance

# In both cases, you can use the -choose_pb -choose_lm -choose_casc followed by the appropriate files to use your own pb, landmark and Haarcascade files.

Stats :

With a split database, you can also somewhat evaluate your model

./stats face_embeddings_database.txt face_embeddings_database_testset.txt # -thresh "float"
./stats face_embeddings_database.txt -avg_db # will give you stats per label

WIP :

-Rework stats for Linear Algebra enabled

-Average the result on several consecutive frames