KDFReVS-3D

Repository for 《A Distillation Framework for Accurate and Robust Single View 3D Face Reconstruction with Video Supervision》

Dependencies

pytorch >= 1.0
tensorboardX
pyyaml
opencv
neural_renderer_pytorch

Note that neural_renderer_pytorch may fail if you install it on Win10, in this case, please refer to https://github.com/lstcutong/neural_renderer_pytorch-win10 for solution.

Datasets

300VW. Please download original video from 300VW.
YTF. Please download original video from youtube.
CelebA. Please download original images from CelebA.
synface. Please download original images from synface.
AFLW2000-3D. Please download original images from AFLW2000-3D.

Please remember to cite the corresponding papers if you use these datasets.

Data preprocessing

All preprocessing tools have been provided in https://github.com/lstcutong/video-face-3d, please refer to it for more details.

For training, after preprocessing 300VW and YTF datasets, put them in ./datasets/train, one video in one folder, we provide an example for your reference. You can add more videos in it if you like. Note that we need some single face images for evaluation during training, please put any images you like in ./datasets/val.

For testing, we have provided some images that used in paper in ./datasets/test, you can directly run our code to verify the performance that demonstrated in our paper. You can also put some other images and test your own.

Pretrained Models

Download pretrained models from models.zip (password: fvq1) and extract them into ./models

Training and Testing

for training

python train_test.py --config ./configs/train_video_gru.yml --gpu 0 --num_workers 4

for testing

python train_test.py --config ./configs/test_single_image.yml

the results are saved in ./results