Graph Cut with CUDA

This is my project for my CUDA courses.

It's purpose was to implement the max-flow/min-cut algorithm known as Push-Relabel in order to segment images.

It's implemented in CXX17 and CUDA.

The program take two images in order to work:

• The main image to segment
• The seed image with background (red) and foreground (blue) labels

Installation

Clone the repository

git clone git@github.com:AlexandreGIRARD/graph_cuts.git
cd graph_cuts
mkdir build
cd build
cmake ..

Usage

The images should be at .jpg format.

./graph_cut [-cg] <img> <seeds_img>

-c for CPU -g for GPU

Benchmark

The benchmarcking is made with Google Benchmark

./bench

Testsuite

The testsuite compare the results obtained with the the ground truth given by my teacher. It used the mean dice score over the 15 given images in segmentation_dataset.

Actually the mean dice score over the 15 images is 0.91 using the GPU implementation

make test