3D Colorization

This repository contains implementations of several algorithms for 3d volume colorization. Given a binary voxel, the model outputs the same shape with colorizations. To run the code, you should have the tensorflow installed in your python.

Dataset

Our model takes the 3d voxel representations as input. We used Shapenet Core dataset for training. Before training, the conversion from obj format to voxel representations is required and implenmented under the 3d-colorization/objLoader directory. The dataset can't be provided directly due to license limitations. Please contact the Shapenet team for dataset downloading.

Algorithms for 3d volume colorization

cwgan

Conditional Wasserstein GAN. To run:

cd src64/cwgan
python train.py

cwgan with surface kernel

Still in process. To run:

cd src64/cwgan-conv2_5d
python train.py

vae

Variational Autoencoder. To run:

cd src-vae/vae
python train.py

began

Boundary Equibilibrium GAN. To run:

cd src64/began
python train.py

hourglass

Stacked Hourglass Networks. Only two stacks due to the meomory limitation issue. To run:

cd src-vae
python train.py