3DFaceReconstruction

The implementation of the 3D face reconstruction through Resnet and BFM model, for the course project.

Preparing environment

use Singularity on linux(with GPU) /scratch/work/public/singularity/cuda10.2-cudnn8-devel-ubuntu18.04.sif Need python=3.7.0, pytorch=1.1.0, torchvision=0.3.0, cudatoolkit=10.0 In singularity container, you may need

Singularity>git clone https://github.com/timesler/facenet-pytorch.git facenet_pytorch

Singularity>pip install h5py

Preparing Dataset

Download and unzip the CACD original face from

Original face images (detected and croped by openCV face detector) can be downloaded here (3.5G)

Modify the corrsponding path of the dataset input image_list = glob.glob("../carc/CACD2000/*.jpg") and dataset out put save_path=./data/CACD2000_{}.hdf5in the Dataprepocessing.py. Modify the SHORT_LEN to get a smaller subdataset.

Run python Dataprepocessing.py

Training the model

Check the path in dataloader=("./data/CACD2000_train.hdf5")and you will get the trained model in ./model_trained/, with the visualized output images in ./result, and the loss curve in train.png.

Default NUM_EPOCH=25, BATCH_SIZE=8

Run python trainnet.py

Testing the model and get outcome

Download our trained model form GoogleDrive , put it in the MODEL_LOAD_PATH ="./modelload/demo.pth"

Check the path in dataloader=("./data/CACD2000_test.hdf5")and you need to put the model into MODEL_LOAD_PATH ="./modelload/demo.pth", and you will get all of the visualized output images in ./test

Run python testnet.py

Sample Output

There are some output of our training and testing in the output folder, including the trainging curve, some loss in the training epoches, the contrast of ori/reconstructed images of the testing, the reconstructed mesh.