This is work in progress of a rewrite of my uni project I did in 2010 for image processing class to not use (ancient!) OpenCV 1.x and use Eigen. It probably is not optimal and might be buggy.
The grabcut algorithm works by first split foreground/background areas into GMM color models (I use orchard-bauman clustering method). Next the maxflow/mincut algorithm is used to refine edges. Lastly, the GMM is then refined by using new segmentation mask and maxflow/mincut is applied again - we continue that for n steps or until convergence.
- Input image
- User rectangle selection
- Running grabcut (1 iteration)
At the moment all batteries are included for library/tests but it requires OpenCV for user-facing app.
cmake -B build -DBUILD_TESTS=ON -DCMAKE_BUILD_TYPE=Release -DBUILD_GRABCUT_APPS=OFF
cmake --build build -j
for building library and tests, and if you have installed OpenCV findable by CMake:
cmake -B build -DCMAKE_BUILD_TYPE=Release -DBUILD_GRABCUT_APPS=ON
cmake --build build -j
which will build app which runs OpenCV and this implementation for comparision (this is slower and results may vary)
build/apps/run_grabcut <img.jpg|png>
A few examples to compare with OpenCV (5 iterations for each implementation):
Flower:
People:
