MendNet

Code for "MendNet: Restoration of Fractured Shapes Using Learned Occupancy Functions."
Published at SGP 2022.

@article {10.1111:cgf.14603,
  journal = {Computer Graphics Forum},
  title = {{MendNet: Restoration of Fractured Shapes Using Learned Occupancy Functions}},
  author = {Lamb, Nikolas and Banerjee, Sean and Banerjee, Natasha K.},
  year = {2022},
  publisher = {The Eurographics Association and John Wiley & Sons Ltd.},
  ISSN = {1467-8659},
  DOI = {10.1111/cgf.14603}
}

We are now accepting broken object donations via the form below!
Broken Object Donations

Installation

Code tested using Ubutnu 18.04 and python 3.8.0.

We recommend using virtualenv. The following snippet will create a new virtual environment, activate it, and install deps.

sudo apt-get install virtualenv && \
virtualenv -p python3.8 env && \
source setup.sh && \
pip install -r requirements.txt && \
./install.sh

Issues with compiling pyrender are typically solved by upgrading cython: pip install --upgrade cython.

If you want to run the fracturing and sampling code, you'll need to install pymesh dependencies:

./install_pymesh.sh

Quickstart Inference

If you just want to try out inference, run the following script with the example file. This will infer a restoration and create a gif.

cd mendnet
./scripts/infer_quick.sh experiments/mugs/specs.json ../example_files/fractured_mug.obj

Data Preparation

Downloading the data

To generate fractured shapes from scratch, download shapenet from https://shapenet.org/. Extract the data into a directory, e.g. "/path/to/above/ShapeNet".

Almost all scripts require that the following environment variable be set to the directory above ShapeNet.

export DATADIR="/path/to/above/ShapeNet"

TODO: Add fracturing procedure.

Training

Navigate into the mendnet directory.

cd mendnet

Each training experiment needs a corresponding directory with a "specs.json" file. You can find an example at mendnet/experiments/mugs.

To train our approach, run the training python script with the path to an experiment directory.

python python/train.py -e experiments/mugs

Inference

Navigate into the mendnet directory.

cd mendnet

Inference (and related operations) is done in four steps:

1. Infer latent codes.
2. Reconstruct meshes.
3. Generate renders.
4. Evaluate meshes.

To infer:

./scripts/infer.sh experiments/mugs

Data is saved in the experiment directory passed to the reconstruction script, under a Reconstructions subdirectory. For example, results generated for our approach will be stored in mendnet/experiments/mugs/Reconstructions/ours/. Meshes are stored in the Meshes subdirectory. A render of all the results is stored in the top-level reconstruction directory.