azSegFace

A Semantic Segmentation Network trained to segment Human Faces in potrait photos.

Details

Framework: Dlib DNN

Network Type: FCN

Network was first trained on Pascal VOC then retrained on a custom dataset with masked faces in potrait images. The results are reasonable. The underlying network is an FCN with 200+ layers.

Usage

Build

The project uses CMake.

Build in Windows

Install Cmake, VS 2017 x64 with VC tools v140.
Also install CUDA and cuDNN from NVidia's website. This was tested on CUDA v9.0. (If CUDA install fails, disable VS integration ad install again)
In the repo directory create a new directory called build
Open a PS terminal, and type cmake-gui ../
In the cmake gui, select VS 2017 x64 as generator. Click configure.
Cmake will try to configure build environment. It will print a warning saying Dlib will not use cuda
After configure completes, edit DLIB_USE_CUDA and set it to ON.
Then edit CUDA_HOST_COMPILER and set it to C:/Program Files (x86)/Microsoft Visual Studio 14.0/VC/bin/cl.exe
Click configure again and this time it should complete without errors.
Click generate and cmake will generate project files.
Exit cmake-gui and in the PS terminal type cmake --build . --config Release
Build might take 15 to 20 minutes depending on your system configuration.

Build in Linux

Install cmake, makeutils etc.
Install Nvidia CUDA and cuDNN
In the repo directory create a new directory called build
Open a terminal, and type cmake-gui ../
In the cmake gui, select Unix Makefiles as generator. Click configure.
Cmake will try to configure build environment.
Click generate and cmake will generate project files.
Exit cmake-gui and in the terminal type cmake --build .
Build might take 15 to 20 minutes depending on your system configuration.

Inference

After build, you can run the application

>cd Release\
>.\dnn_semantic_face_ex.exe \path\to\images\directory

-- or in linux --
$./dnn_semantic_face_ex.exe path/to/images/directory

The application will scan and load the CNN from ..\..\nets\ directory.
It will show the output for each image in the directory. Hit enter for next image.

Training

Use the VOC annotation tool in voc_utl folder to annotate on your custom images.
The tool will organize the dataset as a directory

To train, use ./dnn_semantic_face_train_ex.exe path/to/dataset/dir
The app will first look for semantic_segmentation_voc2012net.dnn in the current directory and do transfer learning.
You may have to edit the src file with the correct path to the net or move the net to the same directory as the training executable.
After training completes, a new .dnn file will be saved that can be used with the inference executable.