torchmaxflow: Max-flow/Min-cut in PyTorch for 2D images and 3D volumes

Pytorch-based implementation of Max-flow/Min-cut based on the following paper:

• Boykov, Yuri, and Vladimir Kolmogorov. "An experimental comparison of min-cut/max-flow algorithms for energy minimization in vision." IEEE transactions on pattern analysis and machine intelligence 26.9 (2004): 1124-1137.

If you want same functionality in Numpy, then consider Numpy-based implementation

Current release of torchmaxflow does not provide GPU implementation of maxflow

Citation

If you use this code in your research, then please consider citing:

Asad, Muhammad, Lucas Fidon, and Tom Vercauteren. "ECONet: Efficient Convolutional Online Likelihood Network for Scribble-based Interactive Segmentation." Medical Imaging with Deep Learning (MIDL), 2022.

Installation instructions

pip install torchmaxflow

or

# Clone and install from github repo

$ git clone https://github.com/masadcv/torchmaxflow
$ cd torchmaxflow
$ pip install -r requirements.txt
$ python setup.py install

Example outputs

Maxflow2d

Interactive maxflow2d

Example usage

The following demonstrates a simple example showing torchmaxflow usage:

image = np.asarray(Image.open('data/image2d.png').convert('L'), np.float32)
image = torch.from_numpy(image).unsqueeze(0).unsqueeze(0)

prob = np.asarray(Image.open('data/image2d_prob.png'), np.float32)
prob = torch.from_numpy(prob).unsqueeze(0)

lamda = 20.0
sigma = 10.0

post_proc_label = torchmaxflow.maxflow(image, prob, lamda, sigma)

For more usage examples see:

2D and 3D maxflow and interactive maxflow examples: demo_maxflow.py

References

• OpenCV's Graphcut implementation
• SimpleCRF's maxflow implementation

This repository depends on the code for maxflow from latest version of OpenCV, which has been included.