OperatorNet: Recovering 3D Shapes From Difference Operators

This is our implementation of OperatorNet, a network that reconstructs shapes from shape difference operators.

This code was written by Ruqi Huang and Marie-Julie Rakotosaona.

Prerequisites

CUDA and CuDNN (changing the code to run on CPU should require few changes)
Python 2.7
Tensorflow

Setup

Install required python packages, if they are not already installed:

pip install numpy
pip install plyfile

Clone this repository:

git clone https://github.com/mrakotosaon/operatornet.git
cd operatornet

Shape difference operators

To generate demo shape difference matrices: run demo_compute_shape_diff.m with Matlab.

Models

Download pretrained models:

cd models
python download_models.py

Data

A demo dataset can be found here: https://nuage.lix.polytechnique.fr/index.php/s/BqiX5rcWszkKT9N

It contains shape differences and labels as Matlab matrices. This dataset is a simplified version of the one used in the paper.

Download pretrained models:

cd Data
python download_demo_data.py

Training

To train OperatorNet with the default settings and demo data:

python train.py

Test

To test OperatorNet with the default settings and demo shapes:

python test.py

From generated demo shapes, run reconstruction, interpolation or analogy (see code).

Produced shapes are saved in a results directory. Please create this directory if it does not exist.

Citation

If you use our work, please cite our paper.

@article{huang2019operatornet,
  title={OperatorNet: Recovering 3D Shapes From Difference Operators},
  author={Huang, Ruqi and Rakotosaona, Marie-Julie and Achlioptas, Panos and Guibas, Leonidas and Ovsjanikov, Maks},
  journal={ICCV},
  year={2019}
}

Contact

If you have any problem about this implementation, please feel free to contact via:

rqhuang88 AT gmail DOT com or mrakotosaon AT gmail DOT com