We use the dense landmarks extracted by Face++ Dense Facial Landmarks to align images and train our FlowNet. You can process your images to get the landmarks by:
# Change the image path in get_landmarks.py and add the face++ keys
# You may also need to change the API url in face_plus_plus.py to yours, or you will get an Error 401.
python get_landmarks.py
To speed up the processing, we only extract the landmarks under 07 illumination condition from each pose. The naming format is xxx_xx_xx_xxx.json (no illumination).
We use face-parsing.PyTorch to extract the facial masks. The mask is composed of the segmentation of face, ear, hair, and neck. To speed up the processing, we only extract the masks under 07 illumination condition from each pose. The naming format is xxx_xx_xx_xxx_07.png.
The landmarks and masks we used to process multipie dataset can download from GoogleDrive or BaiduNetDisk (l98p).
Removing the comments and change the path in process.py, then run
python process.py
The final dataset folder is structured as follows:
dataset
├── multipie
│ ├── train
│ │ ├── images
│ │ ├── masks
│ │ └── landmarks.npy
│ └── test
│ ├── images
│ ├── gallery_list.npy (optional)
│ └── visual_list.npy (optional)
└── lfw
├── images
└── pairs.txt