Noise-Resilient Training Method For Face Landmark Generation From Speech

The code for the following paper.

You can find the project page here.

Installation

Install the required Python packages:

pip install -r requirements.txt

It also depends on the following packages:

ffmpeg --- 3.4.1

The code has been tested on Ubuntu 16.04 and OS X Sierra and High Sierra.

Note that the new versions of the PyTorch will yield different results due to numerical differences.

Code Example

The generation code has the following arguments:

-i --- Input folder containing speech files
- See this link for supported audio formats.
-m --- Input pre-trained talking face landmarks model
-o --- Output path
-s --save_prediction save the predicted landmarks and speech array in the folder specified by the -o option and disable generation of animation
-l --load_prediction load predictions from the folder specified by the -i option and generate a painted face animation in the folder specified by the -o option. This option expects the input folder to contain pairs of files with the same name but different extensions - .wav and .npy

You can run the following code to test the system:

python generate.py -i ../speech_samples/ -m ../pre_trained/1D_CNN.pt -o ../results/1D_CNN/

python generate.py -i ../speech_samples/ -m ../pre_trained/1D_CNN_NR.pt -o ../results/1D_CNN_NR/

python generate.py -i ../speech_samples/ -m ../pre_trained/1D_CNN_TC.pt -o ../results/1D_CNN_TC/ --temporal_condition

Save the landmarks predicted and speech vector using the ID_CNN model from audio in ../speech_samples/ to an NPY format file in replic/pred_out/

python generate.py -i ../speech_samples/ -m ../pre_trained/1D_CNN.pt -o ../replic/pred_out/ -s

Load landmarks from external files in replic/samples/identity_removed/ and generate animation in replic/anim_out/

python generate.py -i ../replic/samples/identity_removed/ -m ../pre_trained/1D_CNN.pt -o ../replic/anim_out/ -l

Face landmarks Extraction from Videos

Please see the following links for extracting face landmarks:

3D face landmarks

DLIB

For face normalization, please refer to this repo.

Training

The training code has the following arguments:

-i --- Input hdf5 file containing training data
-n --- Input hdf5 file containing noise files (optional)
--temporal_condition --- boolen value to enable autoregression model (optional)
-o --- Output folder path to save the model

Usage:

Base model training:

python train.py -i path-to-hdf5-train-file/ -o output-folder-to-save-model-file

Base model noise_resilient training:

python train.py -i path-to-hdf5-train-file/ -n path-to-hdf5-noise-file/ -o output-folder-to-save-model-file

Autoregressive model training:

python train.py -i path-to-hdf5-train-file/ --temporal_condition -o output-folder-to-save-model-file

Autoregressive noise-resilient model training:

python train.py -i path-to-hdf5-train-file/ -n path-to-hdf5-noise-file/ --temporal_condition -o output-folder-to-save-model-file

Citation

@ARTICLE{seeskimeznr3dtface,
         author={S. E. {Eskimez} and R. K. {Maddox} and C. {Xu} and Z. {Duan}},
         journal={IEEE/ACM Transactions on Audio, Speech, and Language Processing},
         title={Noise-Resilient Training Method for Face Landmark Generation From Speech},
         year={2020},
         volume={28},
         pages={27-38},
         doi={10.1109/TASLP.2019.2947741},
         ISSN={2329-9304}
}

eeskimez/noise_resilient_3dtface