/line-drawing-vectorization-polyvector-flow

Primary LanguageC++

Keypoint-Driven Line Drawing Vectorization via PolyVector Flow

This repository contains:

  • C++ code with vectorization algorithm
  • Matlab code for debugging polyvector flow
  • Python code for keypoint inference

Contact me if you need code for speed benchmarks. This release was tuned for better vectorization accuracy.

Citation
@article{Puhachov2021KeypointPolyvector,
    author = {Ivan Puhachov and William Neveu and Edward Chien and Mikhail Bessmeltsev},
    title = {Keypoint-Driven Line Drawing Vectorization via PolyVector Flow},
    journal = {ACM Transactions on Graphics (Proceedings of SIGGRAPH Asia)},
    volume = {40}, number = {6}, year = {2021}, month = dec,
    doi = {10.1145/3478513.3480529}
}

Installation

Cmake build

mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
make

Using Docker (alternative)

Requires Docker and NVIDIA Container Toolkit (for GPU acceleration).

Retrieve a Gurobi WLS license. This can be done by running:

grbgetkey

This will produce a gurobi.lic file. Place this into the repository directory.

To build the Docker image using docker compose:

docker compose build

Alternatively, build it using docker build:

docker build -t line-drawing-vectorization-polyvector-flow-gpu .

Then, run a Docker container based on the built image using docker compose:

docker compose run gpu inputs/example.png

Alternatively, run it using docker run:

docker run --rm -v ./gurobi.lic:/opt/gurobi/gurobi.lic:ro -v ./inputs:/workspace/line-drawing-vectorization-polyvector-flow/inputs --gpus all line-drawing-vectorization-polyvector-flow-gpu inputs/example.png

Usage

Running build/vectorize will start vectorization on default input parameters: inputs/example.png and inputs/example.pts

It will output a debug window:

To suppress debug window, provide output destination as third parameter:

build/vectorize image.png image.pts image.svg

This bash script runs both keypoint prediciton and vectorization give a single image as input:

bash run_vectorization.sh inputs/example.png

Keypoints prediction

Uses PyTorch (tested on 1.10) and CUDA. Can work with CPU, but expect longer inference time.

To predict keypoints run:

 python prediction/usemodel.py --model prediction/best_model_checkpoint.pth --input image.png --output image.pts

Illustrations Copyrights

  • 'Dracolion', 'Mouse', 'Muten', 'Sheriff' are from [Noris et al. 2013]. Please ask the authors for the .pngs.

  • 'Banana-Tree', 'Elephant', 'Hippo', 'Kitten', 'Puppy', 'Leaf', 'Dog06', 'Dog14': (c) Ivan Huska. https://www.easy-drawings-and-sketches.com/

Build Instructions

Gurobi

Tested with Gurobi 9.1.1, 9.0.3, 8.1.1. Install Gurobi and activate license, then update GUROBI_HOME in CMakeLists.txt:

set(GUROBI_HOME "/opt/gurobi911/linux64/")

Alternativelly, you can set your own path in cmake/FindGUROBI.cmake

Qt

Instructions: ubuntu

You may want to:

sudo apt-get install libz-dev
sudo apt-get install libbz2-dev
sudo apt install libatlas-base-dev

Boost

https://www.boost.org/

sudo apt-get install libboost-all-dev

OpenMP

sudo apt-get install libomp-dev

Paal

http://paal.mimuw.edu.pl/

Then move paal/include/paal/ folder here to paal/

Gurobi

We use environment variable, don't forget to setup GUROBI_HOME and

Linux: you may need to recompile (see https://support.gurobi.com/hc/en-us/articles/360039093112-How-do-I-resolve-undefined-reference-errors-while-linking-Gurobi-in-C-)

cd /opt/gurobi903/linux64/src/build
make
cp libgurobi_c++.a ../../lib/

OpenCV

This one can be finicky. Here's what worked for me (William, Ubuntu 18.04).

Install the OpenCV library this way. Make sure to do sudo make install at the very end.

Make sure that you have a FindOpenCV.cmake file in polyvector_flow/cpp/cmake.

Make sure that the CMakelists.txt file in polyvector_flow/cpp specifies the path to the OpenCV static libraries (the .a ones). Example : find_package(OpenCV REQUIRED PATHS "/usr/lib/x86_64-linux-gnu")

Building will take at least 30 minutes

Eigen

sudo apt install libeigen3-dev

Miscellaneous

-- Qt found
-- Found OpenMP_C: -fopenmp (found version "4.5") 
-- Found OpenMP_CXX: -fopenmp (found version "4.5") 
-- Found OpenMP: TRUE (found version "4.5")  
-- Found OpenCV: /usr/local (found version "4.5.0") 
-- Found Boost: /usr/lib/x86_64-linux-gnu/cmake/Boost-1.71.0/BoostConfig.cmake (found suitable version "1.71.0", minimum required is "1.48")  
-- Boost_INCLUDE_DIRS = /usr/include
-- GUROBI_HOME = /opt/gurobi911/linux64/
-- Found GUROBI: /opt/gurobi911/linux64/lib/libgurobi_c++.a  
-- Gurobi found
-- Paal = /home/ivan/projects/polyvector_cpp/paal
-- Configuring done