pygco

Python wrappers for GCO alpha-expansion and alpha-beta-swaps. These wrappers provide a high level interface for graph cut inference for multi-label problems.

See this blog for examples and comments: peekaboo-vision.blogspot.com

Installation

With Pip

Run pip install git+git://github.com/JianchaoTan/gco_python

For Linux

Download and install Cython (use your package manager).
run make
Run example.py for a simple example.

Troubleshooting

There have been some problems compiling gco (not my wrappers) using gcc4.7. Please install gcc-4.6 and adjust the call in Makefile accordingly.

Usage

GCO implements alpha expansion and alpha beta swaps using graphcuts. These can be used to efficiently find low energy configurations of certain energy functions. Note that from a probabilistic viewpoint, GCO works in log-space.

Note that all input arrays are assumed to be in int32. This means that float potentials must be rounded!

These algorithms can only deal with certain energies. Unfortunately I have not figured out yet how to convert C++ errors to Python. If an unknown error is raised, it probably means that you used an invalid energy function. Look at the gco README for details.

This package gives a high level interface to gco, providing the following functions:

cut_simple: Graph cut on a 2D grid using a global label affinity-matrix.

cut_VH: NOT DONE YET Graph cut on a 2D grid using a global label affinity-matrix and edge-weights. V contains the weights for vertical edges, H for horizontal ones.

cut_from_graph: Graph cut on an arbitrary graph with global label affinity-matrix.

cut_from_graph_weighted: NOT DONE YET Graph cut on an arbitrary graph with global label affinity-matrix and edgeweights.