rudyryk / deepparametricshapes

Code for Deep Parametric Shape Predictions using Distance Fields, in CVPR 2020.

DeepParametricShapes | Webpage | Paper | Video

Deep Parametric Shape Predictions using Distance Fields
Dmitriy Smirnov, Matthew Fisher, Vladimir G. Kim, Richard Zhang, Justin Solomon
Conference on Computer Vision and Pattern Recognition (CVPR) 2020

Set-up

To install the code, run:

sudo apt install libcairo2-dev pkg-config python3-dev
conda create -n dps python=3.6 -y
conda activate dps
conda install pytorch 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.

2D: Font Vectorization

Demo

First, download a pretrained font vectorization model:

mkdir -p models/dps_2d
wget -O models/dps_2d/ckpt.pth https://www.dropbox.com/s/46tp19h6npqhuuh/dps_2d.pth\?dl\=0

Then, run the following to generate a PDF file with the vectorization for a given input glyph PNG image:

python scripts/run_2d.py demo/P1.png P out.pdf

Make sure to specify the letter of the input glyph (in this case P). The demo directory contains PNGs of the GAN-generated glyphs used for Figure 13 of the paper.

Training

To prepare the training dataset, first download and extract the font TTF files:

wget -O fonts.tar.gz https://www.dropbox.com/s/7oreepk5gm0efj4/fonts.tar.gz?dl=0
tar -xvf fonts.tar.gz

Then, process the TTFs to generate the input images and target distance fields:

python scripts/generate_fonts.py

To train a model from scratch, run:

python scripts/train_2d.py --output models/model_name --data data/fonts

3D: Shape Abstraction

Pretrained Models

Run the following to download pretrained models for airplane and chair shape abstraction:

mkdir -p models/airplanes
wget -O models/airplanes/ckpt.pth https://www.dropbox.com/s/k91u1zjdywgqdga/dps_3d_airplanes.pth?dl=0
mkdir -p models/chairs
wget -O models/chairs/ckpt.pth https://www.dropbox.com/s/i4l0rrx6rbit0fa/dps_3d_chairs.pth?dl=0

BibTeX

@inproceedings{smirnov2020dps,
  title={Deep Parametric Shape Predictions using Distance Fields},
  author={Smirnov, Dmitriy and Fisher, Matthew and Kim, Vladimir G. and Zhang, Richard and Solomon, Justin},
  year={2020},
  booktitle={Conference on Computer Vision and Pattern Recognition (CVPR)}
}