Training an Xception network for facial landmark detection using Microsoft Synthesis data. Main contribution is the discovery of some nuanced parameters and relations during training which affect the final accuracy.

This project makes use of the following repositories:

https://github.com/braindotai/Facial-Landmarks-Detection-Pytorch

https://microsoft.github.io/FaceSynthetics/

Python3

opencv-python
dlib
numpy
matplotlib
imageio
scikit-image
torchsummary
tqdm
torch
imutils

Two directories should be set up:

/Landmark Dataset <-- custom name
    /ibug_300W_large_face_landmark_dataset <-- ibug 300W dataset
    /dataset_5000 <-- MS synthesis dataset
        00000.png
        00000_ldmks.txt
        00001.png
        00001_ldmks.txt
    labels_ibug_300W_train.xml
    labels_ms_synthesis_mix.xml
    labels_ms_synthesis_landmark.xml
    labels_ms_synthesis_dlib.xml

The 300W dataset can be downloaded directly from here, or take a look at here.

The MS synthesis dataset should be direclty downloaded from here. Due to limit on training system, we choose 5000 image sets randomly. You can choose any subset of your choice.

The XML files contains file names for each dataset and the corresponding landmark labels and face boundary box (bbox). The 300W xml is directly taken from the provided dataset; the synthesis xmls are generated by script /helper/build_synthesis_xml.py. The _mix, _landmark, _dlib represents the bbox generation method used. dlib means bbox are created using the dlib frontal face detector, which has only about 60% of face detection rate on the synthesis dataset. landmark means that we use provided landmark labels to infer face boundaries. We find the min and max values for both axes, and add a little offset (15 pixels). The _mix label means the bbox is created with dlib package whenever possible, then with landmarks if dlib fails.

For training, just use the /src directory and the train.py script.

Run the command train.py --folder $train_subdir$ to start training. Make sure:

/src
    train.py
    /train_set1
        train.json

Look at the exmaple_train_folder for reference.

The training will run automatically and add 3 files under the train_subdir: model.epoch, model.pt, and datalist.pickle. These are model checkpoint files and used for evaluation and resumption. Read the train.py comments for more information.

To resume training, run: train.py --folder $train_subdir$ --resume.

Testing directory is the same as training direcoty. You should use Testing.ipynb notebook file.

{
    "data_dir": "/Users/dongxuhuang/Downloads/Landmark",
    "bbox_option": "dlib",  # dlib, landmark, mix
    "size": "2000",
    "ft_ratio": "0.7",
    "train_test_split": "1.0",
    "image_dim": "128",
    "brightness": "0.24",
    "saturation": "0.3",
    "contrast": "0.15",
    "hue": "0.14",
    "angle": "14",
    "face_offset": "32",
    "crop_offset": "16",
    "epochs": "100"
}

The training configration file train.json in each training set subdir is used to initialize the training. The data_dir is the absolute path to the data directory you defined. The bbox_option tells the training script which bbox method to use (which xml file to pull, see previous section for explanation). The size tells how many images should be contained the training dataset. Please note only a limited number of synthesis images can be processed by dlib for bbox; if the requested size exceed this limit, mix option will be used. The ft_ratio is the ratio of synthesis-to-real faces in training data. Use 1.0 if you want to train with fully synthetic data. Otherwise the script will randomly pick some real-faces from the 300W dataset. epochs defines number of epochs used for training.

All other parameters are preprocessing parameters defined in model. Please read the model to understand.