Deep Sketch-Based Modeling: Tips and Tricks

Contents

• Introduction
• Requirements
• Download Dataset
• Results

Introduction

This repository contains the Pytorch implementation of Deep Sketch-Based Modeling: Tips and Tricks, including binary mask prediction and 3D shape reconstruction.

You can find detailed usage instructions for training and evaluation below.

If you use our code or dataset, please cite our work:

@inproceedings{deepsketch2020,
    title = {Deep Sketch-Based Modeling: Tips and Tricks },
    author = {Yue, Zhong and Yulia, Gryaditskaya and Honggang, Zhang and Yi-Zhe, Song},
    booktitle = {Proceedings of International Conference on 3D Vision (3DV)},
    year = {2020}
}

Requirements

First you have to make sure that you have all dependencies in place. The simplest way to do so, is to use anaconda. sss Please refer the README file in each sub-task for detailed instruction.

Download Dataset

We use two datasets in this paper: the ProSketch dataset and a dataset of synthetic sketches.

ProSketch is a dataset of human sketches, and is a part of this publication. The synthetic data can be generated for other shapes and categories as described below.

Most of our experiments are conducted on the modelsfrom a chair category of the ShapeNetCore dataset2, complemented by two additional categories: planes and lamps. We selected these categories guided by the next principles: 1) Easy to sketch. 2) Generality. 3) View differentiability. 4) Shape genius higher than 1。 5) Large inter-category variance. We generate three categories with distinctive styles, whichwe refer to as naive, stylized and style-unified. Please refer paper for further details.

Generate your own dataset

To generated your own dataset, simply run the code below to generate the synthetic dataset.

Generate your own dataset of synthetic sketches.

python dataset/run.py

Note: you need to change the *.csv file according to your own dataset.

Then, to stylised the genertaed dataset, run the code from SynDraw

python dataset/svg_tools_svg_disturber.py -a -c -n 1.3 -r 2.5 -sl 0.9 -su 1.1 -t 2 -min 1 -max 2 -os 1 -pen 2.5 -penv 1.5 -bg -u

Results

We identify key differences between sketch and image inputs, driving out important insights and proposing the respective solutions, we show an improved performance of deep image modeling.