zqbai-jeremy / DFNRMVS

Source code for CVPR 2020 paper "Deep Facial Non-Rigid Multi-View Stereo"

Deep Facial Non-Rigid Multi-View Stereo

Source code for CVPR 2020 paper "Deep Facial Non-Rigid Multi-View Stereo" [paper] [supp] [video].

Installation

(1) Create an Anaconda environment with python 3.6.

conda create -n DFNRMVS python=3.6
source activate DFNRMVS

(2) Clone the repository and install dependencies.

git clone https://github.com/zqbai-jeremy/DFNRMVS.git
cd DFNRMVS
conda install --yes --file requirements_conda.txt
pip install -r requirements_pip.txt

(3) Setup 3DMM

• Clone this repository, which is forked and modified from YadiraF/face3d, to "<DFNRMVS directory>/external/".

mkdir external
cd external
git clone https://github.com/zqbai-jeremy/face3d.git
cd face3d

• Setup face3d as in YadiraF/face3d.

• Download "Exp_Pca.bin" from Guo et al. (in "CoarseData" link of their repository) and copy to "<DFNRMVS directory>/external/face3d/examples/Data/BFM/Out/".

• Download "std_exp.txt" from Deng et al. and copy to "<DFNRMVS directory>/external/face3d/examples/Data/BFM/Out/".

(4) Install face-alignment.

conda install -c 1adrianb face_alignment

(5) Download pre-trained model (2views_model.pth or 3views_finetune_model.pth; May be used for research purpose only) to "<DFNRMVS directory>/net_weights/". Need to create the folder.

Run Demo

• Modify directory paths in demo.py and run

cd <DFNRMVS_directory>
python demo.py

• All images in the input directory will be used for reconstruction. Per-view results will be saved to the output directory.

• Some examples are in "<DFNRMVS directory>/examples/". The corresponding outputs are in "<DFNRMVS directory>/out_dir/".

• The model usually gives good results for 2 views input with +-30 degree yaw angles.

Training

• Training requires 256x256 images with ground truth 3D scans. Loss functions and training parameters are provided in "<DFNRMVS directory>/train/losses.py"

• Need to setup torch-batch-svd to use the losses.

Citation

@inproceedings{bai2020deep,
  title={Deep Facial Non-Rigid Multi-View Stereo},
  author={Bai, Ziqian and Cui, Zhaopeng and Rahim, Jamal Ahmed and Liu, Xiaoming and Tan, Ping},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={5850--5860},
  year={2020}
}