Rank3DGAN: Semantic mesh generation using relative attributes

This repository provides Tensorflow implementations for Rank3DGAN paper.

This code is heavily based on Multi-chart Generative Surface Modeling implementation multichart3dgans.

Models

shape_version: the project code for watertight meshes (human and bird shapes) that requires triplet of landmarks face_version: the project code for face meshes that requires quadriplet of landmarks

Datasets

• HumanShape
• Basel Face Model
• 3D CUB

For all the details of the data prepration, please check the paper Appendix.

For the data pre/post-processing, please check matlab folder.

Prerequisites

• Python 3.5
• TensorFlow 1.6
• bunch and tqdm packages

Dataset creation

Check matlab/createDataset.m for the details. To transform the charts to tfrecords format use:

• For the case of single meshes dataset:

python3 convert_to_tfrecords.py --database_signature=<database_signature>

• For the case of pairwise comparison meshes dataset:

python3 convert_to_tfrecords.py --database_signature=<database_signature> -p

Rank3DGAN training

python3 gan_main.py -c=configs/<my_config>.json

Rank3DGAN testing

python3 evaluate_main.py -c=configs/config.json

After getting the generated charts. Use matlab/inspectGeneratedData.m to obtain the resulting meshes.

Mesh generation