What energy functions can be minimized via graph cuts
Summary (3 min read)
1 Introduction and summary of results
- Many of the problems that arise in early vision can be naturally expressed in terms of energy minimization.
- Researchers typically use general purpose global optimization techniques such as simulated annealing [3, 11] , which is extremely slow in practice.
- The experimental results produced by these algorithms are also quite good; for example, two recent evaluations of stereo algorithms using real imagery with ground truth found that a graph cut method gave the best overall performance [23, 25] .
- Minimizing an energy function via graph cuts, however, remains a technically difficult problem.
- The authors results provide a significant generalization of the energy minimization methods used in [4-6, 8, 13, 17, 24] , and show how to minimize an interesting new class of energy functions.
1.1 Summary of our results
- The main result in this paper is a precise characterization of the functions in F 3 that can be minimized using graph cuts, together with a graph construction for minimizing such functions.
- Note that in this paper the authors only consider binary-valued variables.
- As an example, the authors will show in section 4.1 how to use their results to solve the pixel-labeling problem, even though the pixels have many possible labels.
- The authors also identify an interesting class of class of energy functions that have not yet been minimized using graph cuts.
- In the language of Markov Random Fields [11, 19] , these methods consider first-order MRF's.
1.2 Organization
- Section 5 contains their main theorems for other classes.
- Detailed proofs of their theorems, together with the graph constructions, are deferred to section 6.
2 Overview of graph cuts
- The minimum s-t-cut problem is to find a cut C with the smallest cost.
- Due to the theorem of Ford and Fulkerson [10] this is equivalent to computing the maximum flow from the source to sink.
- There are many algorithms which solve this problem in polynomial time with small constants [1, 12] .
3 Defining graph representability
- Each cut on G has some cost; therefore, G represents the energy function mapping from all cuts on G to the set of nonnegative real numbers.
- Thus a natural question to ask is what is the class of energy functions for which the authors can construct a graph that represents it.
- Above the authors used each node (except the source and the sink) for encoding one binary variable.
- The authors will summarize graph constructions that they allow in the following definition.
4.1 Example: pixel-labeling via expansion moves
- In this section the authors show how to apply this theorem to solve the pixel-labeling problem.
- The authors will show how their method can be used to derive the expansion move algorithm developed in [8] .
- Note that the key technical step in this algorithm can be naturally expressed as minimizing an energy function involving binary variables.
- In their paper, it is not clear whether this is an accidental property of the construction (i.e., they leave open the possibility that a more clever graph cut construction may overcome this restriction).
- Using their results, the authors can easily show this is not the case.
6.2 Proof of theorems 3 and 6: the constructive part
- In this section the authors will give the constructive part of the proof: given a regular energy function from class F 3 they will show how to construct a graph which represents it.
- First the authors will consider regular functions of two variables, then regular functions of three variables and finally regular functions of the form as in the theorem 6.
- This will also prove the constructive part of the theorem 3.
- Indeed, suppose a function is from the class F 2 and each term in the sum satisfies the condition given in the theorem 3 (i.e. regular).
- Then each term is graph-implementable (as the authors will show in this section) and, hence, the function is graph-implementable as well according to the lemma 10.
Functions of two variables Let E(x 1 , x 2 ) be a function of two variables represented by a table
- Now the authors can easily constuct a graph G which represents this function.
- Note that the authors did not introduce any additional nodes for representing binary interactions of binary variables.
- This is in contrast to the construction in [8] which added auxiliary nodes for representing energies that the authors just considered.
- The authors construction yields a smaller graph and, thus, the minimum cut can potentially be computed faster.
Functions of three variables Now let us consider a regular function E of three variables. Let us represent it as a table
- It's easy to check that these transformations preserve the functional π.
- The authors need to show that all terms here are graph-representable, then lemma 10 will imply that E is graph-representable as well.
- The first three terms are regular functions depending only on two variables and thus are graph-representable as was shown in the previous section.
- The graph G that represents this term can be constructed as follows.
Functions of many variables Finally let us consider a regular function E
- Each term in the sum need not necessarily be regular.
- This can be done using the following lemma and a trivial induction argument.
- Therefore the authors did not introduce any nonregular projections for these terms.
6.3 Proof of theorem 7
- Hence, all terms E i,j are regular, i.e. they satisfy the condition in the theorem 3.
- The following sequence of operations shows one possible way to push the maximum flow through this graph.
Did you find this useful? Give us your feedback
Citations
5,670 citations
4,463 citations
Cites background or methods from "What energy functions can be minimi..."
...The question of what energy functions can be minimized via graph cuts was addressed in [25]....
[...]
...After [15, 31, 19, 8, 25, 5] minimum cut/maximum flow algorithms on graphs emerged as an increasingly useful tool for exact or approximate energy minimization in low-level vision....
[...]
...However, the results in [25] apply only to energy functions of binary variables with double and triple cliques....
[...]
3,699 citations
3,233 citations
Cites methods from "What energy functions can be minimi..."
... CRFs on these images [17]. The MCMC procedure was run for 36 hours and only partially converged for the bottom image. We have also experimented with graph cut inference in the fully connected models [11], but it did not converge within 72 hours. In contrast, a single-threaded implementation of our algorithm produces a detailed pixel-level labeling in 0.2 seconds, as shown in Figure 1(e). A quantitati...
[...]
3,099 citations
References
18,761 citations
"What energy functions can be minimi..." refers background in this paper
...However, researchers typically have needed to rely on general purpose optimization techniques such as simulated annealing [3], [16], which requires exponential time in theory and is extremely slow in practice....
[...]
...Energy functions of the form (3) can be justified on Bayesian grounds using the wellknown Markov Random Fields (MRF) formulation [16], [31]....
[...]
8,496 citations
"What energy functions can be minimi..." refers background in this paper
...Energy functions of the form (3) can be justified on Bayesian grounds using the wellknown Markov Random Fields (MRF) formulation [16], [31]....
[...]
7,458 citations
7,413 citations
[...]
4,341 citations