FaceForensics Data Processing

This repository provides an introduction of all data preprocessing and dataloaders for the FaceForensics image dataset that are sharable between multiple projects

Setup

• Poetry: https://github.com/sdispater/poetry (curl -sSL https://raw.githubusercontent.com/sdispater/poetry/master/get-poetry.py | python): poetry install
• pre-commit: https://pre-commit.com/ (pip install pre-commit): Use before committing

Usage

• extract_compressed_videos: extracts frames from videos

• extract_face_locations: uses dlib to extract face locations from videos for each frame. Saves it as .json to disk

• migrate_bounding_boxes_to_face_information: renames bounding_box folders to face_information_folders

• extract_faces_from_bounding_boxes: uses previously saved bounding boxes to extract face images from images

• extract_mask_bounding_boxes: extract bounding boxes from mask videos (used later for checking validity of bounding boxes in videos)

• extract_faces_tracked_from_bounding_boxes: same as other face extraction script but with tracking and can use mask information

• resample_videos: resamples videos to new frame rate

• create_file_list: creates a filelist for easier sharing and comparing of datasets

Download

Using the above commands the FF++ dataset was preprocessed and made available via this link: http://kaldir.vc.in.tum.de/ff_all_jsons.zip