C++ implementation of Joint Cascade Face Detection and Alignment.
I recommend using Git to fetch the source code. If you are not familiar with Git, there is a tutorial you can follow.
$ git clone --recursive https://github.com/luoyetx/JDA.git
OR
$ git clone https://github.com/luoyetx/JDA.git
$ cd JDA
$ git submodule update --init
If you directly download the zip file, please remember to download luoyetx/liblinear and luoyetx/jsmnpp, then extract the source code to 3rdparty. liblinear is used for global regression training and jsmnpp is used for json config parsing.
We use CMake to build the project, I highly recommend you to use this build tool. We also need OpenCV. If you are on Windows, make sure you have set environment variable OpenCV_DIR to OpenCV's build directory like D:/3rdparty/opencv2.4.11/build. You may also need Visual Studio to compile the source code. If you are on Linux or Unix, install the development packages of OpenCV via your system's Package Manager like apt-get on Ubuntu or yum on CentOS. However, Compile the source code of OpenCV will be the best choice of all.
$ cd JDA
$ mkdir build && cd build
$ cmake ..
$ make
We use config.json for configuration. config.template.json is a template, please copy one and rename it to config.json. Attention, all relative path is start from build directory, and please use / instead of \\ even if you are on Windows platform.
You should prepare your own data. You need two kinds of data, face with landmarks and background images. You also need to create two text file face.txt and background.txt which can be changed in config.json. Every line of face.txt indicates a face image's path with its landmarks. The number of landmarks can be changed in config.json and the order of landmarks does not matter.
../data/face/00001.jpg x1 y1 x2 y2 ........
../data/face/00002.jpg x1 y1 x2 y2 ........
....
....
The face images should be resized to the pre-defined size and you should do any data augmentation by yourself, the code will exactly use the face images you provide. background.txt is much more simpler. Every line indicates where the background image in the file system.
../data/bg/000001.jpg
../data/bg/000002.jpg
../data/bg/000003.jpg
....
....
Background images should have no face and we will do data augmentation during the hard negative mining. Of course, you can use absolute path to indicate where is your face images and background images.
$ ./jda train
If you are using Visual Studio, make sure you know how to pass command line arguments to the program. All trained model file will be saved to model directory.
All model file is saved as a binary file. The model parameters have two data type, 4 byte int and 8 byte double, please pay attention to the endianness of you CPU.
|-- mask (int)
|-- meta
| |-- T (int)
| |-- K (int)
| |-- landmark_n (int)
| |-- tree_depth (int)
| |-- current_stage_idx (int) // training status
| |-- current_cart_idx (int)
|-- mean_shape (double, size = 2*landmark_n)
|-- stages
| |-- stage_1
| | |-- cart_1
| | |-- cart_2
| | |-- ...
| | |-- cart_K
| | |-- global regression weight
| |-- stage_2
| |-- ...
| |-- stage_T
|-- mask (int)
For more details of the model file layout, please refer to cascador.cpp and cart.cpp.
FDDB is widely used for face detection evaluation, download the data and extract to data directory.
|-- data
| |-- fddb
| |-- images
| | |-- 2002
| | |-- 2003
| |-- FDDB-folds
| | |-- FDDB-fold-01.txt
| | |-- FDDB-fold-01-ellipseList.txt
| | |-- ....
| |-- result
prepare your data and model file, you also need to set current_stage_idx and current_cart_idx correctly in config.json. All result text file used by npinto/fddb-evaluation is under result directory.
$ ./jda fddb
It's a shame that I still not get a fully trained model now. There is many detailed problems of JDA algorithm which need to be discussed. My code may have some hidden bugs. Welcome any bug report and any question or idea through the issues.
There is a QQ group 347185749. If you are a Tencent QQ user, welcome to join this group and we can discuss more there.
BSD 3-Clause
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent