2D Face reconstruction using SFSNet using Tensorflow
This repository entirely reproduces the work done by Sengupta et. al. from University of Maryland and published in CVPR 2018. Please refer the Project Page and the Github repo for further details. The technical details can be found in the paper.
Features
- Face deconstruction - It provides the normal, shading, and albedo output for every input image.
- Face reconstruction - It reconstructs any given face using the normal, shading and albedo deconstruction results.
Installation and Usage
Requirements
All the requirements can be found in the environment.yml file above. To install all the required dependencies and setup the environment to use,
conda env create -f environment.yml
source activate 3dface
Data Loader
After the environment is setup, we require the datasets to be loaded, for the various training and testing steps. This data download and preparation can be done by,
python data_loader.py --skipnet_batch_size 10
Alternately, we can just run the setup.sh file to the same effect. Once the setup.sh file is run, the entire setup is complete.
Training
In order to train the network we can simply run the run.sh file. This considers a set of default training parameters. The values can be changed to suit requirements using the training help. Training help can be found by running,
python train.py --help
Testing
To be added(After Training Complete)
Uninstallation
In order to purge the entire conda env and restart with this repo from the beginning please run,
purge.sh
In order to just temporarily stop the environment and continue from the same point at a later stage we can instead use,
source deactivate 3dface