Learning Manifold Patch-Based Representations of Man-Made Shapes
Dmitriy Smirnov, Mikhail Bessmeltsev, Justin Solomon
International Conference on Learning Representations (ICLR) 2021

To install the code, run:

conda create -n learningpatches python=3.6 -y
conda activate learningpatches
conda install pytorch=1.3.1 torchvision cudatoolkit=10.2 -c pytorch -y
pip install -r requirements.txt

Also, be sure to execute export PYTHONPATH=:$PYTHONPATH prior to running any of the scripts.

First, download the pretrained models for each shape category:

wget -O models.zip https://www.dropbox.com/s/ntt1ytpjwx2385i/learningpatches_models.zip?dl=0
unzip models.zip

Then, run the following to generate an OBJ file with the 3D model for a given input sketch PNG image:

python scripts/run.py demo/airplane.png airplanes out.obj --no-turbines

Make sure to specify airplanes, bathtubs, bottles, cars, guitars, guns, knives, or guns as the shape category. Optionally, for the airplanes category, the --no-turbines flag does not output the turbine patches in the 3D model.

The demo directory contains PNGs of some sample input sketches.

Note that the meshes output by the demo script may have non-manifold discontinuities between patches due to discretization artifacts. This can be avoided by choosing the number of subdivisions based on patch boundary arc lengths. The results shown in the paper are all computed in this way.

@inproceedings{smirnov2021patches,
  title={Learning Manifold Patch-Based Representations of Man-Made Shapes},
  author={Smirnov, Dmitriy and Bessmeltsev, Mikhail and Solomon, Justin},
  year={2021},
  booktitle={International Conference on Learning Representations (ICLR)}
}