The freenet from jclimma

freenet's Introduction

FReeNet — Official PyTorch Implementation

Official pytorch implementation of the paper "FReeNet: Multi-Identity Face Reenactment, CVPR'20".

Using the Code

Requirements

This code has been developed under Python3.7, PyTorch 1.5.1 and CUDA 10.1 on Ubuntu 16.04.

# Install python3 packages
pip3 install -r requirements.txt

Datasets in the paper

Download RaFD dataset to datasets/RaFD.
Use Face++ API to extract the landmark with 106 points for each face, and save corresponding information to landmakr.txt, e.g. datasets/RaFD/RaFD90/landmark.txt.

Preprocess RaFD dataset.

> Split RaFD dataset to different dirs based on angle, e.g. RaFD45/image, RaFD90/image, RaFD135/image.
python3 1-preprocess.py

Unified Landmark Converter

Train ULC model.

cd src/ULC
python3 train.py --data RaFD --name RaFD --save_every 10 --every 60 --epochs 200  # for RaFD dataset

Test ULC model.

cd src/ULC
python3 test.py --data RaFD --name RaFD --save_every 10 --every 60 --epochs 200 --resume  # for RaFD dataset

Geometry-aware Generator

Train GAG model.

cd src
python3 train.py --name RaFD-08-21 --gpu_ids 0 --model landmark_L64_Tri --netG resnet_9blocks_cat --dataset_mode RaFD90L64Tri --batch_size 3  # for RaFD dataset

Test GAG model.

cd src
python3 test_image.py --name RaFD-08-21 --gpu_ids 0 --model landmark_L64_Tri --netG resnet_9blocks_cat --dataset_mode RaFD90L64Tri  # for RaFD dataset

Citation

If our work is useful for your research, please consider citing:

@inproceedings{zhang2020freenet,
  title={FReeNet: Multi-Identity Face Reenactment},
  author={Zhang, Jiangning and Zeng, Xianfang and Wang, Mengmeng and Pan, Yusu and Liu, Liang and Liu, Yong and Ding, Yu and Fan, Changjie},
  booktitle={CVPR},
  pages={5326--5335},
  year={2020}
}