Note: This repo was forked from the original on bitbucket: https://bitbucket.org/ericyang3721/headpose_forensic.git

HeadPose Forensics

This repository is the implementation of the work used in our ICASSP paper Exposing Deep Fakes Using Inconsistent Head Poses

Environment

• Ubuntu 16.04
• tqdm 4.28.1
• numpy 1.15.4
• dlib 19.16.0
• opencv-python 3.4.3.18

Test

python run_test.py --input_dir=debug_data --classifier_path=path/to/trained/model --save_file=path/to/output/results

This will examine all images and videos in the folder of 'debug_data', print results in terminal, and the probability of being fake images/video could be saved in --save_file in the project root folder.

The result include a optout attribute, which indicate whether a image is not used for classification when it is true.

Train

There are 3 steps to train the classifier:

step 1: extract landmarks of real and fake data

python train_step1_landmarks.py --real_video_dir=dir/to/real/videos --fake_video_dir=dir/to/fake/videos --output_landmark_path=path/to/save/landmarks

step 2: extract head poses

python train_step2_headposes.py --landmark_info_path=path/to/landmarks/in/step1 --headpose_save_path=path/to/save/headpose/data

step 3: train svm model

python train_step3_training.py --headpose_path=path/to/headposes/in/step2 --model_save_path=path/to/save/trained/model

Citation

Please cite our paper in your publications if it helps your research.

@inproceedings{yang2019exposing,
  title={Exposing Deep Fakes Using Inconsistent Head Poses},
  author={Yang, Xin and Li, Yuezun and Lyu, Siwei},
  booktitle={IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  year={2019}
}

Notice

This repository is NOT for commecial use. It is provided "as it is" and we are not responsible for any subsequence of using this code.