This page is for the paper titled above appeared in BigMM2019 (short).
Here is the example of our results.
Python 3.5+, Pytorch 1.1.0, CUDA10 Since this project heavily relies on maskrcnn-benchmark, building is a bit complicated process.
pip install -r requirements.txt
export INSTALL_DIR=$PWD
# install pycocotools
cd $INSTALL_DIR
git clone https://github.com/cocodataset/cocoapi.git
cd cocoapi/PythonAPI
python setup.py build_ext install
# Build mask-RCNN
cd $INSTALL_DIR
python setup.py build develop
For testing our model, please install gdown and then download pre-trained weights by
bash download_trained_models.sh
For testing our model, please do the following;
- Download datasets from vectornet and decompress it (e.g.,
data/DATASET/*
). - Do preprocessing (SVG to PNG format, getting list of image ids.)
python preprocess.py --root data
After the installation, please do $cd scripts
.
This script will test our model on some images from the test subset.
CUDA_VISIBLE_DEVICES=<gpu_id> python demo.py <config> <checkpoint>
- If you need the result in SVG format, please use
--mode svg
. - If you want to change the directory, please use
--output_dir <output_dir>
This script will compute evaluation metrics using our model.
CUDA_VISIBLE_DEVICES=<gpu_id> python eval.py <config> <checkpoint>
CUDA_VISIBLE_DEVICES=<gpu_id> python train.py <config>
If you find this code or dataset useful for your research, please cite our paper:
@inproceedings{inoue_2019_bigmm,
author = {Inoue, Naoto and Yamasaki, Toshihiko},
title = {Fast Instance Segmentation for Line Drawing Vectorization},
booktitle = {IEEE International Conference on Multimedia Big Data(BigMM)},
month = {September},
year = {2019}
}