This repository contains a minimal-dependency facial alignment method. We re-implement a state-of-the-art landmark detection method (2D-FAN) using only OpenCV. Thus, we refer to the software in this repository as FAO (Facial Aligment with OpenCV).

The simplest way to install FAO is to use pip as follows. But, this will be slow (~7-8 FPS) because the default opencv-python package has no CUDA support, therefore FAO will run on CPU. (See more below for GPU processing)

git clone https://github.com/sariyanidi/face_alignment_opencv.git
cd face_alignment_opencv
pip install --upgrade pip
pip install -r requirements_cpu.txt 

If you want to use FAO with GPU support (~100 FPS), you need to manually compile OpenCV with CUDA and CUDNN support. In this case, you should make sure that the repository does not use the OpenCV installed through pip, but the one that you manually compiled. Once you manually install OpenCV, then the following lines will suffice to use FAO with cuda support.

git clone https://github.com/sariyanidi/face_alignment_opencv.git
cd face_alignment_opencv
pip install -r requirements_gpu.txt 

FAO comes with three scripts for testing. The first one is for single-image processing, which can be tested as

python3 process_image.py samples/corb.jpeg
python3 process_image.py samples/barca.jpg

The second script is for video processing, which can be tested as:

python3 process_video.py samples/la_boheme.mp4

The third script is for processing all images in a directory, which can be tested as:

python3 process_image_dir.py samples/testdir