SimpleCRF

An open-source toolkit for conditional random field (CRF) and dense CRF


License
BSD-3-Clause
Install
pip install SimpleCRF==0.2.1.1

Documentation

SimpleCRF

Matlab and Python wrap of Conditional Random Field (CRF) and fully connected (dense) CRF for 2D and 3D image segmentation, according to the following papers:

[1] Yuri Boykov and Vladimir Kolmogorov, "An experimental comparison of min-cut/max-flow algorithms for energy minimization in vision", IEEE TPAMI, 2004.

[2] Philipp Krähenbühl and Vladlen Koltun, "Efficient inference in fully connected crfs with gaussian edge potentials", in NIPS, 2011.

[3] Kamnitsas et al in "Efficient multi-scale 3D CNN with fully connected CRF for accurate brain lesion segmentation", Medical Image Analysis, 2017.

maxflow densecrf1 densecrf2

Dependency

This repository depends on the following packages: Maxflow, DenceCRF and 3D Dense CRF

Installation

  1. Install by: pip install SimpleCRF

  2. Alternatively, you can compile the source files by the following two steps:

python setup.py build
python setup.py install

Examples

Some demos of using this package are:

  • examples/demo_maxflow.py: using maxflow for automatic and interactive segmentation of 2D and 3D images.

  • examples/demo_densecrf.py: using dense CRF for 2D gray scale and RGB image segmentation.

  • examples/demo_densecrf3d.py: using 3D dense CRF for 3D multi-modal image segmentation.

Modules

  1. maxflow has four functions as follows. Note that the current version only supports binary segmentation.
  • maxflow.maxflow2d() for 2D automatic segmentation.

  • maxflow.interactive_maxflow2d() for 2D interactive segmentation.

  • maxflow.maxflow3d() for 3D automatic segmentation.

  • maxflow.interactive_maxflow3d() for 3D interactive segmentation.

  1. denseCRF has the following function. It can deal with multi-class segmentation, and only supports RGB images.
  • denseCRF.densecrf() for 2D automatic segmentation.
  1. denseCRF3D has the following function. It can deal with multi-class segmentation. The input channel number can be 1-5.
  • denseCRF3D.densecrf3d() for 3D automatic segmentation.