Neural Contours: Learning to Draw Lines from 3D Shapes

This repository contains the PyTorch implementation for CVPR 2020 Paper "Neural Contours: Learning to Draw Lines from 3D Shapes" by Difan Liu, Mohamed Nabail, Aaron Hertzmann, Evangelos Kalogerakis.

[Arxiv]

Dependency

• The project is developed on Ubuntu 16.04 with cuda9.0 + cudnn7.0. The code has been tested with PyTorch 1.1.0 (GPU version) and Python 3.6.8.
• Python packages:
  • OpenCV (tested with 4.2.0)
  • PyYAML (tested with 5.3.1)
  • scikit-image (tested with 0.14.2)

Dataset and Weights

• Pre-trained model is available here, please put it in data/model_weights:

  cd data/model_weights
  unzip weights.zip
  

• download example testing data:

  cd data/example
  wget https://people.cs.umass.edu/~dliu/projects/NeuralContours/example.zip
  unzip example.zip
  

• training data is available here.

Differentiable Geometry Branch

• we use rtsc-1.6 to compute all the input geometric feature maps and lines. See here for details.
• run geometry branch without NRM (Neural Ranking Module), this script takes thresholds of geometric lines as input:

  python -m scripts.geometry_branch_demo -sc 10.0 -r 10.0 -v 10.0 -ar 0.1 -model_name bumps_a -save_name data/output/bumps_a.png

Testing with NRM and ITB (Image Translation Branch)

• Testing with NRM and ITB:

  python -m scripts.test -model_name bumps_a -save_name data/output/bumps_a_NCs.png

  Note that computation time depends on GPU performance, parameter setting and input 3D model. For reference, tested on GeForce GTX 1080 Ti, under default setting, Neural Contours of bumps_a takes about 12 minutes.

Cite:

@InProceedings{Liu_2020_CVPR,
author={Liu, Difan and Nabail, Mohamed and Hertzmann, Aaron and Kalogerakis, Evangelos},
title={Neural Contours: Learning to Draw Lines from 3D Shapes},
booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2020}
}

Contact

To ask questions, please email.