A thin QPBO wrapper for Python

pip install thinqpbo==0.1.4


Thin wrapper for QPBO

Thin Python wrapper for a modified version of the quadratic pseudo-Boolean optimization (QPBO) algorithm by Vladimir Kolmogorov. The original source code by Vladimir Kolmogorov availbable at This wrapper uses a modified version with support for larger graphs and slightly lower memory usage. See submodule repository for more details.

QPBO vs. Maxflow

While the QPBO algorithm performs a s-t graph cut similar to Maxflow, it allows for non-submodular energy terms, which Maxflow doesn't. Amongst other things, this allows QPBO to solve optimization problems with exclusions terms, which can be very usefull. The graph constructed by the QPBO implementation is twice the size of a Maxflow graph for an equivalent problem. Thus, QPBO uses more memory and is slightly slower than Maxflow.


Install package using pip install thinqpbo or clone this repository (including submodule). Building the package requires Cython.

Graph types

Currently, there are three different types of graphs: QPBOInt, QPBOFloat and QPBODouble. The only difference is the underlying datatypes used for the edge capacities in the graph. For stability, it is recommended to use QPBOInt for integer capacities and QPBODouble for floating point capacities. However, in some cases, it maybe be favourable to use QPBOFloat to reduce memory consumption.

Advanced features (QPBO-P and QPBO-I)

The QPBO implementation has a few advanced extensions known as QPBO-P and QPBO-I. Currently, not all advanced functions have been wrapped. If you need to use features of the QPBO C++ library that are not wrapped by thinqpbo, please let me know by creating an issue on GitHub.

Tiny example

import thinqpbo as tq

# Create graph object.
graph = tq.QPBOInt()

# Number of nodes to add.
nodes_to_add = 2

# Add two nodes.
first_node_id = graph.add_node(nodes_to_add)

# Add edges.
graph.add_unary_term(0, 0, 5) # E1(0) = 5, s     --5->   n(0)
graph.add_unary_term(0, 1, 0) # E0(0) = 1, n(0)  --1->   t
graph.add_unary_term(1, 5, 0) # E0(1) = 5, n(1)  --5->   t
graph.add_pairwise_term(0, 1, 0, 7, 0, 4)   # E01(0,1) = 7, n(0)  --7->   n(1)
                                            # E11(0,1) = 4, Not possible with standard Maxflow

# Find maxflow/cut graph.
twice_energy = graph.compute_twice_energy()

for n in range(nodes_to_add):
    segment = graph.get_label(n)
    print('Node %d has label %d.' % (n, segment))
# Node 0 has label 0.
# Node 1 has label 0.
print('Twice energy/flow: %s' % twice_energy)
# Twice energy/flow: 12


As the QPBO implementation is distributed under the GPLv3 license, so is this package.