eMotion-GAN

[Colab Demo 😀]

This repository contains the source code for our paper:

eMotion-GAN: A Motion-based GAN for Photorealistic and Facial Expression Preserving Frontal View Synthesis

Check out our Supplementary Video for more details & results.

Animation:

Updates

• ...

Installation

Create and activate conda environment:

conda create -n eMotionGAN python=3.10
conda activate eMotionGAN

Install all dependencies:

pip install -r requirements.txt

Install Jupyter Lab to visualize demo:

conda install -c conda-forge jupyterlab

Download the pre-trained weights file from Google Drive. Move it to the ./weights folder.

Demos

Run our interactive visualization demo using Colab notebook (no GPU needed).

Frontal View Synthesis (FVS) & Facial Expression Preserving

Cross-subject Facial Motion Transfer

Training

By default, the training datasets are structured as follows:

├── datasets

    ├── CK+
        ├── sub_id
            ├── seq_id
              ├── img_001.png
              ├── img_002.png
              ...
              ├── img_010.png
              
    ├── ADFES
        ├── sub_id
            ├── seq_id
              ├── img_001.png
              ├── img_002.png
              ...
              ├── img_010.png
              
    ...

Motion Calculation

python generate_data_files.py --dataset_name SNAP
                              --data_dir_F ./datasets/SNAPcam1/ 
                              --data_dir_NF ./datasets/SNAPcam2 
                              --save_dir_F ./datasets/optical_flows/ 
                              --save_dir_NF ./datasets/optical_flows/ 
                              --emotions_file_path ./datasets/emotions/SNAP_emotions.csv 
                              --flow_algo Farneback

Training

Download DMUE and its pre-trained model.

python train.py --data_path ./datasets/data_file.txt 
                --proto_path ./datasets/protocols/SNAP_proto.csv 
                --fold 1

Evaluation

Frontal View Synthesis

python FVS_demo.py --img1 images/images_1.png 
                   --img2 images/images2.png
                   --model_path weights/eMotionGAN_model.pth

Motion Transfer

python motion_transfer_demo.py --img1 images/images_1.png 
                               --img2 images/images2.png 
                               --neutrl_img images/neutral_img.png
                               --model_path weights/eMotionGAN_model.pth

Citation

If you find this repo useful, please consider citing our paper

@article{ikne2024emotion,
  title={eMotion-GAN: A Motion-based GAN for Photorealistic and Facial Expression Preserving Frontal View Synthesis},
  author={Ikne, Omar and Allaert, Benjamin and Bilasco, Ioan Marius and Wannous, Hazem},
  journal={arXiv preprint arXiv:2404.09940},
  year={2024}
}