Last week, I decided to try out the IBFS maximum flow algorithm for vision applications for the first time. The algorithm is due to A.V. Goldberg, S.Hed, H. Kaplan, R.E. Tarjan, and R.F. Werneck, who have released an implementation which uses the same interface as the commonly used algorithm by Boykov and Kolmogorov.
My test was very simple, I generated a 128×128 8-connected grid graph, with source and sink arcs computed from the ‘cameraman’ image. I then let the weights of the arcs between the nodes vary several orders of magnitude. The result is quite telling:
The BK algorithm takes over 100 seconds to compute the maximum flow when the regularization is high. By contrast, the IBFS algorithm always uses less than 1 second. Here is the logarithmic plot: