To faciliate the experiments, a module was created to handle the datasets. The module is located in the dataset folder contains a single code file (__init__.py) that expose the utilitary functions and folders to all available datasets, (e.g. lfw, vggface2).
Usage example:
from dataset import DATASET_LFW, DATASET_KIND_RAW, ls_imgs_paths
# List all images paths in the LFW dataset
imgs_paths = ls_imgs_paths(dataset=DATASET_LFW, kind=DATASET_KIND_RAW)When the module is imported by the first time, it will check for available datasets and generates an index file to improve operations in the future uses, if the dataset are changes or updated, the function get_dataset_index(recreate=True) shall be executed the regenerate the index file.
In order to prepare the dataset of images for the experiments, we need to align the faces in the imagens, crop to a squared size and ensure file type.
To accomplish this, the script run_align.py can be used. The script uses the dataset module to lookup the images and process them. The results are placed in the aligned_images folder inside of respective folder of the dataset following the same structure of the raw folder (person_name/imgs). The following image ilustrates the process result:
Under the hood, the script uses the DLIB face detector model to identify the faces landmarks and align the images. The aligning function was extracted from the FFHQ Dataset Pre-Processing and transformed into a utilitary function for the project.
Once all the images from dataset are well aligned and with the right sizes, we can advance to the face segmentation step, the face segmentation process will separate all the facial elements of the images (e.g. eyes, nose, mouth, etc.) into different images. The following image ilustrates the process result:
To execute this task, the script run_segmentation.py can be used. The script uses the dataset module to lookup all the aligned images into the dataset folders and process them. The results are placed in the segmented_images folder inside of respective folder of the dataset following the same structure of the raw folder (person_name/imgs), the segmentation maps are also saved in the segmentation_maps folder, under the dataset folder following the same structure.
To execute the face segmentation it is used a pre-trained model from the BiSeNet architecture.
Obs.: The BiSeNet model being used is implemented using PyTorch, so it is recommended to have a different environment to execute the segmentation process.
...
...
...