FaceForensics is a video dataset consisting of more than 500,000 frames containing faces from 1004 videos that can be used to study image or video forgeries. To create these videos we use an automatated version of the state of the art Face2Face approach. All videos are downloaded from Youtube and are cut down to short continuous clips that contain mostly frontal faces. In particular, we offer two versions of our dataset: source-to-target, where we reenact over 1000 videos with new facial expressions extracted from other videos, which e.g. can be used to train a classifier to detect fake images or videos, and self-reenactment, where we use Face2Face to reenact the facial expressions of videos with their own facial expressions as input to get pairs of videos, which e.g. can be used to train supervised generative refinement models. Consult our paper for more information or watch our youtube video.
If you would like to download the FaceForensics data, please fill out an agreement to the FaceForensics Terms of Use and send it to us at faceforensics@googlegroups.com.
If you have not received a response within a week, it is likely that your email is bouncing - please check this before sending repeat requests.
Once you have our script, you can choose to download H.264 lossless compressed and raw videos, the original videos as well as the cropped self-reenactment images that were used for our refinement task. Note that the dataset size is ~130gb for lossless compressed and ~3,5tb for raw videos. For more information about the download script usage, use the '-h' option.
Video origin
All videos were downloaded from youtube. We used the youtube8m dataset to filter
for videos with tags "face", "newscaster" and "newsprogram". In addition we also
used videos that we found on youtube with these tags. Afterwards, we used the
Viola-Jones face detector to gather sequences of minimum length 300 that contain a single
face. These sequences were then checked manually for occlusions.
Folder structure
We provide the split we used for the final numbers in our paper which divides the 1004 videos in
704 videos for train, 150 videos for validation and 150 videos for test, where we made sure that
all videos come from different original video sources.
FaceForensic dataset
|-- test
|-- altered
Contains videos that were altered by Face2Face
|-- mask
Contains the masks video that were produced by the Face2Face algorithm
|-- original
Contains the original videos
|-- train
-- same as test --
|-- val
-- same as test -- Video format
All videos have a constant frame rate of 30 fps and are either written raw with the 'DIB ' codec
or have been lossless compressed with H.264. All videos are at least 480p, at most 1080 and at least
300 frames long.
File format template
All video files are formatted as follows
<target youtube id>_<number>_<source youtube id>_<number>.<file extension>
where a youtube id is always 11 characters long.
Stats
For stats about the dataset please consult files in the stats folder.
Video origins
You can download the youtube id's together with additional information (like fps)
as well as the coordinates of our extracted faces under this link.
If you use the FaceForensics data or code please cite:
@article{roessler2018faceforensics,
author = {Andreas R\"ossler and Davide Cozzolino and Luisa Verdoliva and Christian Riess and Justus Thies and Matthias Nie{\ss}ner},
title = {FaceForensics: A Large-scale Video Dataset for Forgery Detection in Human Faces},
journal={arXiv},
year={2018}
}
If you have any questions, please contact us at faceforensics@googlegroups.com
The data is released under the FaceForensics Terms of Use, and the code is released under the MIT license.
Copyright (c) 2018