A Dual Ascent Framework for Lagrangean Decomposition of Combinatorial Problems

12/16/2016 ∙ by Paul Swoboda, et al. ∙ 0

We propose a general dual ascent framework for Lagrangean decomposition of combinatorial problems. Although methods of this type have shown their efficiency for a number of problems, so far there was no general algorithm applicable to multiple problem types. In his work, we propose such a general algorithm. It depends on several parameters, which can be used to optimize its performance in each particular setting. We demonstrate efficacy of our method on graph matching and multicut problems, where it outperforms state-of-the-art solvers including those based on subgradient optimization and off-the-shelf linear programming solvers.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 1

page 2

page 3

page 4

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

Computer vision and machine learning give rise to a number of powerful computational models. It is typical that inference in these models reduces to non-trivial combinatorial optimization problems. For some of the models, such as conditional random fields (CRF), powerful specialized solvers like [46, 47, 11, 50] were developed. In general, however, one has to resort to off-the-shelf integer linear program (ILP) solvers like CPLEX [2] or Gurobi [35]. Although these solvers have made a tremendous progress in the past decade, the size of the problems they can tackle still remains a limiting factor for many potential applications, as the running time scales super-linearly in the problem size. The goal of this work is to partially fill this gap between practical requirements and existing computational methods.

It is an old observation that many important optimization ILPs can be efficiently decomposed into easily solvable combinatorial sub-problems [31]. The convex relaxation, which consists of these sub-problems coupled by linear constraints is known as Lagrangean or dual decomposition [30, 48]. Although this technique can be efficiently used in various scenarios to find approximate solutions of combinatorial problems, it has a major drawback: In the most general setting only slow (sub)gradient-based techniques [49, 55, 48, 40, 59] can be used for optimization of the corresponding convex relaxation.

In the area of conditional random fields, however, it is well-known [39] that message passing or dual (block-coordinate) ascent algorithms (like e.g. TRW-S [46]) significantly outperform (sub)gradient-based methods. Similar observations were made much earlier in [57] for a constrained shortest path problem.

Although dual ascent algorithms were proposed for a number of combinatorial problems (see the related work overview below), there is no general framework, which would (i) give a generalized view on the properties of such algorithms and more importantly (ii) provide tools to easily construct such algorithms for new problems. Our work provides such a framework.

Related Work

Dual ascent algorithms optimize a dual problem and guarantee monotonous improvement (non-deterioration) of the dual objective. The most famous examples in computer vision are block-coordinate ascent (known also as message passing) algorithms like TRW-S [46] or MPLP [27] for maximum a posteriori inference in conditional random fields [39].

To the best of our knowledge the first dual ascent algorithm addressing integer linear programs belongs to Bilde and Krarup [10] (the corresponding technical report in Danish appeared ). In that work an uncapacitated facility location problem was addressed. A similar problem (simple plant location) was addressed with an algorithm of the same class in [29]. In Fisher and Hochbaum [21] constructed a dual ascent-based algorithm for a problem of database location in computer networks, which was used to optimize the topology of Arpanet [1], predecessor of Internet. The generalized linear assignment problem was addressed by the same type of algorithms in [22]. The Authors considered a Lagrangean decomposition of this problem into multiple knapsack problems, which were solved in each iteration of the method. An improved version of this algorithm was proposed in [33]. Efficient dual ascent based solvers were also proposed for the min-cost flow in [24], for the set covering and the set partitioning problems in [23] and the resource-constrained minimum weighted arborescence problem in [34]. The work [32] describes basic principles for constructing dual ascent algorithms. Although the authors provide several examples, they do not go beyond that and stick to the claim that these methods are structure dependent and problem specific.

The work [17] suggests to use the max-product belief propagation [71] to decomposable optimization problems. However, their algorithm is neither monotone nor even convergent in general.

In computer vision, dual block coordinate ascent algorithms for Lagrangean decomposition of combinatorial problems were proposed for multiple targets tracking [7], graph matching (quadratic assignment) problem [76] and inference in conditional random fields [46, 47, 27, 72, 73, 60, 36, 54, 70]. From the latter, the TRW-S algorithm [46] is among the most efficient ones for pairwise conditional random fields according to [39]. The SRMP algorithm [47] generalizes TRW-S to conditional random fields of arbitrary order. In a certain sense, our framework can be seen as a generalization of SRMP to a broad class of combinatorial problems.

Contribution.

We propose a new dual ascent based computational framework for combinatorial optimization. To this end we:
(i) Define the class of problems, called integer-relaxed pairwise-separable linear programs (IRPS-LP), our framework can be used for. Our definition captures Lagrangean decompositions of many known discrete optimization problems (Section 2).
(ii) Give a general monotonically convergent message-passing algorithm for solving IRPS-LP, which in particular subsumes several known solvers for conditional random fields (Section 4).
(iii) Give a characterization of the fixed points of our algorithm, which subsumes such well-known fixed point characterizations as weak tree agreement [46] and arc-consistency [72] (Section 5).

We demonstrate efficiency of our method by outperforming state-of-the-art solvers for two famous special cases of IRPS-LP, which are widely used in computer vision: the multicut and the graph matching problems. (Section 6).

A C++-framework containing the above mentioned solvers and the datasets used in experiments are available under http://github.com/pawelswoboda/LP_MP.

We give all proofs in the supplementary material.

Notation.

Undirected graphs will be denoted by , where is a finite node set and is the edge set. The set of neighboring nodes of w.r.t. graph is denoted by . The convex hull of a set is denoted by . Disjoint union is denoted by .

2 Integer-Relaxed Pairwise-Separable Linear Programs (IRPS-LP)

Combinatorial problems having an objective to minimize some cost over a set

of binary vectors often have a decomposable representation as

for being sets of binary vectors, typically corresponding to subsets of the coordinates of . This decomposed problem is equivalent to the original one under a set of linear constraints , which guarantee the mutual consistency of the considered components. Replacing by its convex hull and therefore switching to real-valued vectors from binary ones one obtains a convex relaxation111More precisely, this is a linear programming relaxation, since a convex hull of a finite set can be represented in terms of linear inequalities of the problem, which reads:

(1)
(2)

Here are called factors of the decomposition and are called coupling constraints. The undirected graph is called factor graph. We will use variable names whenever we want to emphasize and whenever , .

Definition 1 (Irps-Lp).

Assume that for each edge the matrices of the coupling constraints are such that and for some , analogously for . The problem is called an Integer-Relaxed Pairwise-Separable Linear Program, abbreviated by IRPS-LP.

In the following, we give several examples of IRPS-LP. To distinguish between notation for the factor graph of IRPS-LP, where we stick to bold letters (such as , , ) we will use the straight font (such as , , ) for the graphs occurring in the examples.

Example 1 (MAP-inference for CRF).

A conditional random field is given by a graph , a discrete label space , unary and pairwise costs for , . We also denote . The associated maximum a posteriori (MAP)-inference problem reads

(3)

where and denote the components corresponding to node and edge respectively. The well-known local polytope relaxation [72] can be seen as an IRPS-LP by setting , that is associating to each node and each edge a factor, and introducing two coupling constraints for each edge of the graphical model, i.e. . For the sake of notation we will assume that each label is associated a unit vector with dimensionality equal to the total number of labels and on the -th position. Therefore, the notation makes sense as a convex hull of all such vectors. After denoting an -dimensional simplex as the resulting relaxation reads

(4)

in the overcomplete representation [69] and is defined as

(5)

Here and denote those coordinates of vectors and , which correspond to the label and the pair of labels respectively.

Example 2 (Graph Matching).

The graph matching problem, also known as quadratic assignment [12] or feature matching, can be seen as a MAP-inference problem for CRFs (as in Example 1) equipped with additional constraints: The label set of belongs to a universe , i.e. and each label can be assigned at most once. The overall problem reads

(6)

Graph matching is a key step in many computer vision applications, among them tracking and image registration, whose aim is to find a one-to-one correspondence between image points. For this reason, a large number of solvers have been proposed in the computer vision community [17, 74, 76, 68, 51, 67, 61, 28, 77, 37, 52, 15]. Among them two recent methods [68, 76] based on Lagrangean decomposition show superior performance and provide lower bounds for their solutions. The decomposition we describe below, however, differs from those proposed in [68, 76].

Our IRPS-LP representation for graph matching consists of two blocks: (i) the CRF itself (which further decomposes into node- and edge-subproblems with variables and (ii) additional label-factors keeping track of nodes assigned the label . We introduce these label-factors for each label . The set of possible configurations of this factor consists of those nodes which can be assigned the label and an additional dummy node . The dummy node denotes non-assignment of the label and is necessary, as not every label needs to be taken. As in Example 1, we associate a unit binary vector with each element of the set , and denotes the convex hull of such vectors. The set of factors becomes , with the set of the factor-graph edges. The resulting IRPS-LP formulation reads

(7)

Here we introduced (i) auxiliary variables for all variables and (ii) auxiliary node costs , which may take other values in course of optimization. Factors associated with the vectors and correspond to the nodes and edges of the graph (node- and edge-factors), as in Example 1 and are coupled in the same way. Additionally, factors associated with the vectors  ensure that the label can be taken at most once. These label-factors are coupled with node-factors (last line in (7)).

Example 3 (Multicut).

The multicut problem (also known as correlation clustering) for an undirected weighted graph is to find a partition , , of the graph vertexes, such that the total cost of edges connecting different components is minimized. The number of components is not fixed but is determined by the algorithm. See Fig. 1 for an illustration. Although the problem has numerous applications in computer vision [3, 4, 5, 75] and beyond [6, 58, 13, 14]

, there is no scalable solver, which could provide optimality bounds. Existing methods are either efficient primal heuristics 

[64, 56, 26, 18, 19, 8, 9] or combinatorial branch-and-bound/branch-and-cut/column generation algorithms, based on off-the-shelf LP solvers [41, 42, 45, 75]. Move-making algorithms do not provide lower bounds, hence, one cannot judge their solution quality or employ them in branch-and-bound procedures. Off-the-shelf LP solvers on the other hand scale super-linearly, limiting their application in large-scale problems.

Instead of directly optimizing over partitions (which has many symmetries making optimization difficult in a linear programming setting), we follow [16] and formulate the problem in the edge domain. Let , denote the cost of graph edges and let be the set of all cycles of the graph . Each edge that belongs to different components is called a cut edge. The multicut problem reads

(8)

Here signifies a cut edge and the inequalities force each cycle to have none or at least two cut edges. The formulation (8) has exponentially many constraints. However, it is well-known that it is sufficient to consider only chordless cycles [16] in place of the set in (8). Moreover, the graph can be triangulated by adding additional edges with zero weights and therefore the set of chordless cycles reduces to edge triples. Such triangulation is refered to as chordal completion in the literature [25]. The number of triples is cubic, which is still too large for practical efficiency and therefore violated constraints are typically added to the problem iteratively in a cutting plane manner [41, 42]. To simplify the description, we will ignore this fact below and consider all these cycles at once. Assuming a triangulated graph and redefining as the set of all chordless cycles (triples) we consider the following IRPS-LP relaxation of the multicut problem 222One can show that this relaxation coincides with the standard LP relaxation for the multicut problem [16]:

(9)
(10)

For the sake of notation we shortened a feasible set definition to . Here is the relaxed (potentially non-integer) variable corresponding to . Variable is a copy of , which corresponds to the cycle . Therefore, each gets as many copies , as many chordless cycles contain the edge . For each cycle the set of binary vectors satisfying the cycle inequality is considered. For a cycle with edges this set can be written explicitly as in (10). Along with copies of , we copy the corresponding cost and create auxiliary costs for each cycle containing the edge . During optimization, the cost will be redistributed between itself and its copies , . The factors of the IRPS-LP are associated with each edge (variable ) and each chordless cycle (variable ). Coupling constraints connect edge-factors with those cycle-factors, which contain the corresponding edge (see the last constraint in (10)). An in-depth discussion of message passing for the multicut problem with tighter relaxations can be found in [66].

Figure 1: Illustration of Example 3. A multicut of a graph induced by three connected components (green). Red dotted edges indicate cut edges .

3 Dual Problem and Admissible Messages

Since our technique can be seen as a dual ascent, we will not optimize the primal problem (1) directly, but instead maximize its dual lower bound.

Dual IPS-LP

The Lagrangean dual to (1) w.r.t. the coupling constraints reads

(11)

Here for for some . The function is called lower bound and is concave in . The modified primal costs are called reparametrizations of the potentials . We have duplicated the dual variables by introducing to symmetrize notation. In practice, only one copy is stored and the other is computed on the fly. Note that in this doubled notation the reparametrized node and edge potentials of the CRF from Example 1 read

It is well-known for CRFs that cost of feasible solutions are invariant under reparametrization. We generalize this to the IRPS-LP-case.

Proposition 1.

, whenever obey the coupling constraints.

Admissible Messages

While Proposition 1 guarantees that the primal problem is invariant under reparametrizations, the dual lower bound is not. Our goal is to find such that is maximal. By linear programming duality, will then be equal to the optimal value of the primal (1).

First we will consider an elementary step of our future algorithm and show that it is non-decreasing in the dual objective. This property will ensure the monotonicity of the whole algorithm. Let  be any reparametrization of the problem and be the corresponding dual value. Let us consider changing the reparametrization of a factor by a vector with the only non-zero components  and  . This will change reparametrization of the coupled factors (such that ) due to . The lemma below states properties of which are sufficient to guarantee improvement of the corresponding dual value :

Lemma 1 (Monotonicity Condition).

Let be a pair of factors related by the coupling constraints and be a corresponding dual vector. Let and satisfy

(12)

Then implies .

Example 4.

Let us apply Lemma 1 to Example 1. Let correspond to , where is some node and is any of its incident edges. Then corresponds to a locally optimal label and . Therefore we may assign to any value from . This assures that (20) is fulfilled and remains a locally optimal label after reparametrization even if there are multiple optima in .

Lemma 1 can be straightforwardly generalized to the case, when more than two factors must be reparametrized simultaneously. In terms of Example 1 this may correspond to the situation when a graph node sends messages to several incident edges at once:

Definition 2.

Let be a factor and be a subset of its neighbors. Let , satisfies (20) for all and all other coordinates of are zero. If there exists such that , the dual vector is called admissible. The set of admissible vectors is denoted by .

Lemma 2.

Let then .

1 Input:Factor , neighboring factors , dual variables
(13)
(14)
Maximize admissible messages to :
(15)
Output: .
Procedure 1 Message-Passing Update Step.

Message-Passing Update Step

To maximize , we will iteratively visit all factors and adjust messages connected to it, monotonically increasing the lower bound (11). Such an elementary step is defined by Procedure 1.

Procedure 1 is defined up to the vector , which satisfies (14) (see Proc. 1). Usually, is a good choice. Although different may result in different efficiency of our framework, fulfillment of (14) is sufficient to prove its convergence properties.

The reparametrization adjustment problem (15) serves an intuitive goal to move as much slack as possible from the factor to its neighbors . For example, for the setting of Example 4 its solution reads . Depending on the selected it might correspond to maximization of the dual objective in the direction defined by admissible reparametrizations. Although maximization (15) is not necessary to prove convergence of our method (as we show below, only a feasible solution of (15) is required for the proof), (i) it leads to faster convergence; (ii) for the case of CRFs (as in Example 1) it makes our method equivalent to well established techniques like TRW-S [46] and SRMP [47], as shown in Section 4.1.

The following proposition states that the elementary update step defined by Procedure 1 can be performed efficiently. That is, the size of the reparametrization adjustment problem (15) grows linearly with the size of the factor and its attached messages:

Proposition 2.

Let with . Let the messages in problem (15) have size . Then (15) is a linear program with variables and constraints.

4 Message Passing Algorithm

Now we combine message passing updates into Algorithm 2. It visits every node of the factor graph and performs the following two operations: (i) Receive Messages, when messages are received from a subset of neighboring factors, and (ii) Send Messages, when messages to some neighboring factors are computed and reweighted by . Distribution of weights may influence the efficiency of Algorithm 2 just like it influences the efficiency of message passing for CRFs (see [47]). We provide typical settings in Section 4.1. Usually, factors are traversed in some given a-priori order alternately in forward and backward direction, as done in TRW-S [46] and SRMP [47]. We refer to [47] for a motivation for such a schedule of computations.

1 for  in some order do
2       Receive Messages:
3       Choose a subset of connected factors
4       for  do
5             Compute with Procedure 1.
6             Set  .
7            
8       end for
9       
10       Send Messages:
11       Choose partition .
12       for  do
13             Compute with Procedure 1.
14       end for
15      Choose weights such that .
16       for  do
17             Set .
18            
19       end for
20      
21 end for
Algorithm 2 One Iteration of Message-Passing

We will discuss parameters of Algorithm 2 (factor partitioning , weights ) right after the theorem stating monotonicity for any choice of parameters.

Theorem 1.

Algorithm 2 monotonically increases the dual lower bound (11).

4.1 Parameter Selection for Algorithm 2

There are the following free parameters in Algorithm 2: (i) The order of traversing factors of ; (ii) for each factor the neighboring factors from which to receive messages ; (iii) the partition of factors to send messages to and (iv) the associated weights for messages.

Although for any choice of these parameters Algorithm 2 monotonically increases the dual lower bound (as stated by Theorem 1), its efficiency may significantly depend on their values. Below, we will describe the parameters for Examples 1-3, which we found the most efficient empirically. Additionally, in the supplement we discuss parameters, which turn our algorithm into existing message passing solvers for CRFs (as in Example 1).

Sending a message by some factor automatically implies receiving this message by another, coupled factor. Therefore, usually there is no need to go over all factors in Algorithm 2. It is usually sufficient to guarantee that all coupling constraints are updated by Procedure 1. Formally, we can always exclude processing some factors by setting and , to the empty set. Instead, we will explicitly specify, which factors are processed in the loop of Algorithm 2 in the examples below.

Parameters for Example 1, MAP-inference in CRFs.

Pairwise CRFs have the specific feature that node factors are coupled with edge factors only. This implies that processing only node factors in Algorithm 2 is sufficient. Below, we describe parameters, which turn Algorithm 2 into SRMP [47] (which is up to details of implementation equivalent to TRW-S [46] for pairwise CRFs). Other settings, given in the supplement, may turn it to other popular message passing techniques like MPLP [27] or min-sum diffusion [62].

We order node factors and process them according to this ordering. The ordering naturally defines the sets of incoming and outgoing edges for each node . Here is incoming for if and outgoing if . Each node receives messages from all incoming edges, which is . The messages are send to all outgoing edges, each edge in the partition in line 2 of Algorithm 2 is represented by a separate set. That is, the partition reads . Weights are distributed uniformly and equal to , . After each outer iteration, when all nodes were processed, the ordering is reversed and the process repeats. We refer to [47] for substantiation of these parameters.

Parameters for Example 2, Graph Matching.

Additionally to the node and edge factors, the corresponding IRPS-LP has also label factors (7). To this end all node factors are ordered, as in Example 1. Each node factor receives messages from all incoming edge factors and label factors and sends them to all outgoing edges and label factors. The corresponding partition reads . The weights are distributed uniformly with . The label factors are processed after all node factors were visited. Each label factor receives messages from all connected node factors and send messages back as well: . We use the same single set for sending messages, i.e. . After each iteration we reverse the factor order.

Parameters for Example 3, Multicut.

Similarly to Example 1, it is sufficient to go only over all edge factors in the loop of Algorithm 2, since each coupling constraint contains exactly one cycle and one edge factor. Each edge factor receives messages from all coupled cycle factors and sends them to the same factors. As in Example 1, each cycle factor forms a trivial set in the partition in line 2 of Algorithm 2, the partition reads . Weights are distributed uniformly with . After each iteration the processing order of factors is reversed.

4.2 Obtaining Integer Solution

Eventually we want to obtain a primal solution of (1), not a reparametrization . We are not aware of any rounding technique which would work equally well for all possible instances of IRPS-LP problem. According to our experience, the most efficient rounding is problem specific. Below, we describe our choices for the Examples 13.

Rounding for Example 1

coincides with the one suggested in [46]: Assume we have already computed a primal integer solution for all and we want to compute . To this end, right before the message receiving step of Algorithm 2 for we assign

(16)

Rounding for Example 2

is the same except that we select the best label among those, which have not been assigned yet, to satisfy uniqueness constraints:

(17)

Rounding for Example 3.

We use the efficient Kernighan&Lin heuristic [43] as implemented in [44]. Costs for the rounding are the reparametrized edge potentials.

5 Fixed Points and Comparison to Subgradient Method

Algorithm 2 does not necessarily converge to the optimum of (1). Instead, it may get stuck in suboptimal points, similar to those correspoding to the ”weak tree agreement” [46] or ”arc consistency” [72] in CRFs from Example 1. Below we characterise these fixpoints precisely.

Definition 3 (Marginal Consistency).

Given a reparametrization , let for each factor a non-empty set , be given. Define . We call reparametrization marginally consistent for on if

(18)

If is marginally consistent for on all , we call marginally consistent for .

Note that marginal consistency is necessary, but not sufficient for optimality of the relaxation (1). This can be seen in the case of CRFs (Example 1), where it exactly corresponds to arc-consistency. The latter is only necessary, but not sufficient for optimality [72].

Theorem 2.

If is marginally consistent, the dual lower bound cannot be improved by Algorithm 2.

Comparison to Subgradient Method.

Decomposition IRPS-LP and more general ones can be solved via the subgradient method [48]. Similar to Algorithm 2, it operates on dual variables and manipulates them by visiting each factor sequentially. Contrary to Algorithm 2, subgradient algorithms converge to the optimum. Moreover, on a per-iterations basis, computing subgradients is cheaper than using Algorithm 2, as only (13) needs to be computed, while Algorithm 2 needs to solve (15) additionally. However, for MAP-inference, the study [39] has shown that subgradient-based algorithms converge much slower than message passing algorithms like TRWS [46]. In Section 6 we confirm this for the graph matching problem as well.

The reason for this large empirical difference is that one iteration of the subgradient algorithm only updates those coordinates of dual variables that are affected by the current minimal labeling (i.e. coordinates ), while in Algorithm 2 all coordinates of are taken into account. Also message passing implicitly chooses the stepsize so as to achieve monotonical convergence in Algorithm 1, while subgradient based algorithms must rely on some stepsize rule that may either make too large or too small changes to the dual variables .

6 Experimental Evaluation

Our experiments’ goal is to illustrate applicability of the proposed technique, they are not an exhaustive evaluation. The presented algorithms are only basic variants, which can be further improved and tuned to the considered problems. Both issues are addressed in the specialized studies [65, 66] which are appended. Still, we show that the presented basic variants are already able to surpass state-of-the-art specialized solvers on challenging datasets. All experiments were run on a computer with a 2.2 GHz i5-5200U CPU and 8 GB RAM.

Figure 2: Runtime plots comparing averaged values on car, motor and worms graph matching datasets. Both axes are logarithmic.
Figure 3: Runtime plots comparing averaged primal/dual values on the three knott-3d-{150|300|450} multicut datasets. Values are averaged over all instances in the dataset. The x-axis are logarithmic. Continuous lines are dual lower bounds while corresponding dashed lines show primal solutions obtained by rounding.

6.1 Graph Matching

Solvers.

We compare against two state-of-the-art algorithms: (i) the subgradient based dual decomposition solver [68] abbreviated by DD and (ii) the recent “hungarian belief propagation” message passing algorithm [76], abbreviated as HBP. While the authors of [76] have embedded their solver in a branch-and-bound routine to produce exact solutions, we have reimplemented their message passing component but did not use branch and bound to make the comparison fair. Both algorithms DD and HBP outperformed alternative solvers at the time of their publication, hence we have not tested against [17, 74, 51, 67, 61, 28, 77, 37, 52, 15]. We call our solver AMP.

Datasets.

We selected three challenging datasets. The first two are the standard benchmark datasets car and motor, both used in [53], containing 30 pairs of cars and 20 pairs of motorbikes with keypoints to be matched 1:1. The images are taken from the VOC PASCAL 2007 challenge [20]. Costs are computed from features as in [53]. Instances are densely connected graphs with nodes. The third one is the novel worms datasets [38], containing 30 problem instances coming from bioimaging. The problems are made of sparsely connected graphs with up to nodes and up to labels. To our knowledge, the worms dataset contains the largest graph matching instances ever considered in the literature. For runtime plots showing averaged logarithmic primal/dual gap over all instances of each dataset see Fig. 2.

Results.

Our solver AMP consistently outperforms HBP and DD w.r.t. primal/dual gap and anytime performance Most markedly on the largest worms dataset, the subgradient based algorithm DD struggles hard to decrease the primal/dual gap, while AMP gives reasonable results.

6.2 Multicuts

Solvers.

We compare against state-of-the-art multicut algorithms implemented in the OpenGM [39] library, namely (i) the branch-and-cut based solver MC-ILP [42] utilizing the ILP solver CPLEX [2], (ii) the heuristic primal “fusion move” algorithm CC-Fusion [8]

with random hierarchical clustering and random watershed proposal generator, denoted by the suffixes

-RHC and -RWS and (iii) the heuristic primal “Cut, Glue & Cut” solver CGC [9]. Those solvers were shown to outperform other multicut algorithms [8]. Algorithm MC-ILP provides both upper and lower bounds, while CC-Fusion and CGC are purely primal algorithms. We call our message passing solver with cycle constraints added in a cutting plane fashion MP-C.

Datasets.

A source of large scale problems comes from electron microscopy of brain tissue, for which we wish to obtain neuron segmentation. We have selected three datasets

knott-3d-{150|300|450} of increasingly large size [39], each consisting of 8 instances. Instances have , and nodes and , , and edges respectively.

Results.

For plots showing dual bounds and primal solution objectives over time see Figure 3. Our algorithm MP-C combines advantages of LP-based techniques awith those of primal heuristics: It delivers high dual lower bounds faster than MC-ILP. Its has fast primal convergence speed and delivers primal solutions comparable/superior to CGC’s and CC-Fusion’s.

7 Acknowledgments

The authors would like to thank Vladimir Kolmogorov for helpful discussions. This work is partially funded by the European Research Council under the European Unions Seventh Framework Programme (FP7/2007-2013)/ERC grant agreement no 616160.

References

  • [1] https://en.wikipedia.org/wiki/ARPANET.
  • [2] IBM ILOG CPLEX Optimizer. http://www-01.ibm.com/software/integration/optimization/cplex-optimizer/.
  • [3] A. Alush and J. Goldberger. Break and conquer: Efficient correlation clustering for image segmentation. In E. R. Hancock and M. Pelillo, editors, SIMBAD, volume 7953 of Lecture Notes in Computer Science, pages 134–147. Springer, 2013.
  • [4] B. Andres, J. H. Kappes, T. Beier, U. Köthe, and F. A. Hamprecht. Probabilistic image segmentation with closedness constraints. In D. N. Metaxas, L. Quan, A. Sanfeliu, and L. J. V. Gool, editors, ICCV, pages 2611–2618. IEEE Computer Society, 2011.
  • [5] B. Andres, T. Kröger, K. L. Briggman, W. Denk, N. Korogod, G. Knott, U. Köthe, and F. A. Hamprecht. Globally optimal closed-surface segmentation for connectomics. In A. W. Fitzgibbon, S. Lazebnik, P. Perona, Y. Sato, and C. Schmid, editors, ECCV (3), volume 7574 of Lecture Notes in Computer Science, pages 778–791. Springer, 2012.
  • [6] A. Arasu, C. Ré, and D. Suciu. Large-scale deduplication with constraints using dedupalog. In Y. E. Ioannidis, D. L. Lee, and R. T. Ng, editors, ICDE, pages 952–963. IEEE Computer Society, 2009.
  • [7] C. Arora and A. Globerson. Higher order matching for consistent multiple target tracking. In Proceedings of the IEEE International Conference on Computer Vision, pages 177–184, 2013.
  • [8] T. Beier, F. A. Hamprecht, and J. H. Kappes. Fusion moves for correlation clustering. In CVPR, pages 3507–3516. IEEE Computer Society, 2015.
  • [9] T. Beier, T. Kröger, J. H. Kappes, U. Köthe, and F. A. Hamprecht. Cut, glue & cut: A fast, approximate solver for multicut partitioning. In CVPR. Proceedings, 2014.
  • [10] O. Bilde and J. Krarup. Sharp lower bounds and efficient algorithms for the simple plant location problem. Annals of Discrete Mathematics, 1:79–97, 1977.
  • [11] Y. Boykov, O. Veksler, and R. Zabih. Fast approximate energy minimization via graph cuts. IEEE Transactions on pattern analysis and machine intelligence, 23(11):1222–1239, 2001.
  • [12] R. E. Burkard, E. Çela, P. M. Pardalos, and L. S. Pitsoulis. The Quadratic Assignment Problem, pages 1713–1809. Springer US, Boston, MA, 1999.
  • [13] Y. Chen, S. Sanghavi, and H. Xu. Clustering sparse graphs. In P. L. Bartlett, F. C. N. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger, editors, NIPS, pages 2213–2221, 2012.
  • [14] F. Chierichetti, N. Dalvi, and R. Kumar. Correlation clustering in mapreduce. In Proceedings of the 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’14, pages 641–650, New York, NY, USA, 2014. ACM.
  • [15] M. Cho, J. Lee, and K. M. Lee. Reweighted random walks for graph matching. In K. Daniilidis, P. Maragos, and N. Paragios, editors, ECCV (5), volume 6315 of Lecture Notes in Computer Science, pages 492–505. Springer, 2010.
  • [16] S. Chopra and M. R. Rao. The partition problem. Mathematical Programming, 59(1):87–115, 1993.
  • [17] J. Duchi, D. Tarlow, G. Elidan, and D. Koller. Using combinatorial optimization within max-product belief propagation. In NIPS, pages 369–376. MIT Press, 2006.
  • [18] M. Elsner and E. Charniak. You talking to me? a corpus and algorithm for conversation disentanglement. In K. McKeown, J. D. Moore, S. Teufel, J. Allan, and S. Furui, editors, ACL, pages 834–842. The Association for Computer Linguistics, 2008.
  • [19] M. Elsner and W. Schudy. Bounding and comparing methods for correlation clustering beyond ILP. In Proceedings of the Workshop on Integer Linear Programming for Natural Langauge Processing, ILP ’09, pages 19–27, Stroudsburg, PA, USA, 2009. Association for Computational Linguistics.
  • [20] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The Pascal visual object classes (VOC) challenge. International Journal of Computer Vision, 88(2):303–338, June 2010.
  • [21] M. L. Fisher and D. S. Hochbaum. Database location in computer networks. Journal of the ACM (JACM), 27(4):718–735, 1980.
  • [22] M. L. Fisher, R. Jaikumar, and L. N. Van Wassenhove. A multiplier adjustment method for the generalized assignment problem. Management Science, 32(9):1095–1103, 1986.
  • [23] M. L. Fisher and P. Kedia. Optimal solution of set covering/partitioning problems using dual heuristics. Management science, 36(6):674–688, 1990.
  • [24] D. Gamarnik, D. Shah, and Y. Wei. Belief propagation for min-cost network flow: Convergence & correctness. In SODA, pages 279–292. SIAM, 2010.
  • [25] M. R. Garey and D. S. Johnson. Computers and Intractability: A Guide to the Theory of NP-Completeness. W. H. Freeman & Co., New York, NY, USA, 1979.
  • [26] A. Gionis, H. Mannila, and P. Tsaparas. Clustering aggregation. ACM Trans. Knowl. Discov. Data, 1(1):4, 2007.
  • [27] A. Globerson and T. S. Jaakkola. Fixing max-product: Convergent message passing algorithms for MAP LP-relaxations. In NIPS, pages 553–560, 2007.
  • [28] S. Gold and A. Rangarajan. A graduated assignment algorithm for graph matching. IEEE Trans. Pattern Anal. Mach. Intell., 18(4):377–388, 1996.
  • [29] M. Guignard. A Lagrangean dual ascent algorithm for simple plant location problems. European Journal of Operational Research, 35(2):193–200, 1988.
  • [30] M. Guignard and S. Kim. Lagrangean decomposition: A model yielding stronger Lagrangean bounds. Mathematical programming, 39(2):215–228, 1987.
  • [31] M. Guignard and S. Kim. Lagrangean decomposition for integer programming: theory and applications. Revue française d’automatique, d’informatique et de recherche opérationnelle. Recherche opérationnelle, 21(4):307–323, 1987.
  • [32] M. Guignard and M. B. Rosenwein. An application-oriented guide for designing Lagrangean dual ascent algorithms. European Journal of Operational Research, 43(2):197–205, 1989.
  • [33] M. Guignard and M. B. Rosenwein. Technical note-an improved dual based algorithm for the generalized assignment problem. Operations Research, 37(4):658–663, 1989.
  • [34] M. Guignard and M. B. Rosenwein. An application of Lagrangean decomposition to the resource-constrained minimum weighted arborescence problem. Networks, 20(3):345–359, 1990.
  • [35] Gurobi Optimization, Inc., 2015. http://www.gurobi.com.
  • [36] J. Jancsary and G. Matz. Convergent decomposition solvers for tree-reweighted free energies. In AISTATS 2011, 2011.
  • [37] B. Jiang, J. Tang, C. Ding, and B. Luo. A local sparse model for matching problem. In

    Proceedings of the Twenty-Ninth AAAI Conference on Artificial Intelligence

    , AAAI’15, pages 3790–3796. AAAI Press, 2015.
  • [38] D. Kainmueller, F. Jug, C. Rother, and G. Myers. Active graph matching for automatic joint segmentation and annotation of C. elegans. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 81–88. Springer, 2014.
  • [39] J. H. Kappes, B. Andres, F. A. Hamprecht, C. Schnörr, S. Nowozin, D. Batra, S. Kim, B. X. Kausler, T. Kröger, J. Lellmann, N. Komodakis, B. Savchynskyy, and C. Rother. A comparative study of modern inference techniques for structured discrete energy minimization problems. International Journal of Computer Vision, 115(2):155–184, 2015.
  • [40] J. H. Kappes, B. Savchynskyy, and C. Schnörr. A bundle approach to efficient MAP-inference by Lagrangian relaxation. In

    Computer Vision and Pattern Recognition (CVPR), 2012 IEEE Conference on

    , pages 1688–1695. IEEE, 2012.
  • [41] J. H. Kappes, M. Speth, B. Andres, G. Reinelt, and C. Schnörr. Globally optimal image partitioning by multicuts. In EMMCVPR. Springer, Springer, 2011.
  • [42] J. H. Kappes, M. Speth, G. Reinelt, and C. Schnörr. Higher-order segmentation via multicuts. CoRR, abs/1305.6387, 2013.
  • [43] B. Kernighan and S. Lin. An efficient heuristic procedure for partitioning graphs. The Bell Systems Technical Journal, 49(2), 1970.
  • [44] M. Keuper, E. Levinkov, N. Bonneel, G. Lavoué, T. Brox, and B. Andres. Efficient decomposition of image and mesh graphs by lifted multicuts. In ICCV, 2015.
  • [45] S. Kim, S. Nowozin, P. Kohli, and C. D. Yoo. Higher-order correlation clustering for image segmentation. In J. Shawe-Taylor, R. S. Zemel, P. L. Bartlett, F. C. N. Pereira, and K. Q. Weinberger, editors, NIPS, pages 1530–1538, 2011.
  • [46] V. Kolmogorov. Convergent tree-reweighted message passing for energy minimization. IEEE Trans. Pattern Anal. Mach. Intell., 28(10):1568–1583, 2006.
  • [47] V. Kolmogorov. A new look at reweighted message passing. IEEE Trans. Pattern Anal. Mach. Intell., 37(5):919–930, 2015.
  • [48] N. Komodakis, N. Paragios, and G. Tziritas. MRF energy minimization and beyond via dual decomposition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(3):531–552, March 2011.
  • [49] C. Lemaréchal. Lagrangian decomposition and nonsmooth optimization: Bundle algorithm, prox iteration, augmented Lagrangian. Nonsmooth Optimization: Methods and Applications, pages 201–216, 1992.
  • [50] V. Lempitsky, C. Rother, S. Roth, and A. Blake. Fusion moves for Markov random field optimization. IEEE transactions on pattern analysis and machine intelligence, 32(8):1392–1405, 2010.
  • [51] M. Leordeanu and M. Hebert. A spectral technique for correspondence problems using pairwise constraints. In ICCV, pages 1482–1489. IEEE Computer Society, 2005.
  • [52] M. Leordeanu, M. Hebert, and R. Sukthankar. An integer projected fixed point method for graph matching and MAP inference. In Y. Bengio, D. Schuurmans, J. D. Lafferty, C. K. I. Williams, and A. Culotta, editors, NIPS, pages 1114–1122. Curran Associates, Inc., 2009.
  • [53] M. Leordeanu, R. Sukthankar, and M. Hebert. Unsupervised learning for graph matching. International Journal of Computer Vision, 96(1):28–45, 2012.
  • [54] T. Meltzer, A. Globerson, and Y. Weiss. Convergent message passing algorithms - a unifying view. In UAI, pages 393–401. AUAI Press, 2009.
  • [55] I. Necoara and J. A. Suykens. Application of a smoothing technique to decomposition in convex optimization. IEEE Transactions on Automatic Control, 53(11):2674–2679, 2008.
  • [56] V. Ng and C. Cardie. Improving machine learning approaches to coreference resolution. Proceedings of the 40th Annual Meeting on Association for Computational Linguistics - ACL ’02, (July):104, 2001.
  • [57] C. Ribeiro and M. Minoux. Solving hard constrained shortest path problems by Lagrangean relaxation and branch-and-bound algorithms. Methods of Operations Research, 53:303–316, 1986.
  • [58] E. Sadikov, J. Madhavan, L. Wang, and A. Halevy. Clustering query refinements by user intent. In World Wide Web Conference (WWW). ACM Press, April 2010.
  • [59] B. Savchynskyy, J. Kappes, S. Schmidt, and C. Schnörr. A study of Nesterov’s scheme for Lagrangian decomposition and MAP labeling. In Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on, pages 1817–1823. IEEE, 2011.
  • [60] B. Savchynskyy, S. Schmidt, J. H. Kappes, and C. Schnörr. Efficient MRF energy minimization via adaptive diminishing smoothing. In UAI, pages 746–755. AUAI Press, 2012.
  • [61] C. Schellewald and C. Schnörr. Probabilistic subgraph matching based on convex relaxation. In EMMCVPR, volume 3757 of Lecture Notes in Computer Science, pages 171–186. Springer, 2005.
  • [62] M. I. Schlesinger and K. V. Antoniuk. Diffusion algorithms and structural recognition optimization problems. Cybernetics and Systems Analysis, 47(2):175–192, 2011.
  • [63] D. Sontag and T. S. Jaakkola. Tree block coordinate descent for map in graphical models. In AISTATS, volume 5 of JMLR Proceedings, pages 544–551, 2009.
  • [64] W. M. Soon, H. T. Ng, and D. C. Y. Lim. A machine learning approach to coreference resolution of noun phrases. Computational Linguistics, 27(4):521–544, 2001.
  • [65] C. Swoboda, Paul Rother, A. A. Hassan, and Kainm\̇lx@bibnewblockStudy of Lagrangean decomposition and dual ascent solvers for graph matching.
  • [66] P. Swoboda and B. Andres. A message passing algorithm for the minimum cost multicut problem.
  • [67] P. H. S. Torr. Solving Markov random fields using semi definite programming. In In: AISTATS, 2003.
  • [68] L. Torresani, V. Kolmogorov, and C. Rother. A dual decomposition approach to feature correspondence. IEEE Trans. Pattern Anal. Mach. Intell., 35(2):259–271, 2013.
  • [69] M. J. Wainwright and M. I. Jordan. Graphical models, exponential families, and variational inference. Foundations and Trends in Machine Learning, 1(1-2):1–305, 2008.
  • [70] H. Wang and D. Koller. Subproblem-tree calibration: A unified approach to max-product message passing. In 30th International Conference on Machine Learning (ICML-13), pages 190–198, 2013.
  • [71] Y. Weiss and W. T. Freeman. On the optimality of solutions of the max-product belief-propagation algorithm in arbitrary graphs. IEEE Transactions on Information Theory, 47(2):736–744, 2001.
  • [72] T. Werner. A linear programming approach to max-sum problem: A review. IEEE Trans. Pattern Analysis and Machine Intelligence, 29(7):1165–1179, 2007.
  • [73] T. Werner. Revisiting the linear programming relaxation approach to Gibbs energy minimization and weighted constraint satisfaction. IEEE Trans. Pattern Anal. Mach. Intell., 32(8):1474–1488, 2010.
  • [74] J. Yarkony, C. C. Fowlkes, and A. T. Ihler. Covering trees and lower-bounds on quadratic assignment. In CVPR, pages 887–894. IEEE Computer Society, 2010.
  • [75] J. Yarkony, A. Ihler, and C. C. Fowlkes. Fast Planar Correlation Clustering for Image Segmentation, pages 568–581. Springer Berlin Heidelberg, Berlin, Heidelberg, 2012.
  • [76] Z. Zhang, Q. Shi, J. McAuley, W. Wei, Y. Zhang, and A. van den Hengel. Pairwise matching through max-weight bipartite belief propagation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016.
  • [77] F. Zhou and F. D. la Torre. Factorized graph matching. In CVPR, pages 127–134. IEEE Computer Society, 2012.

8 Supplementary Material

Proof of Proposition 1

Proposition.

, whenever obey the coupling constraints.

Proof.

where due to and . ∎

Proof of Proposition 2

Proposition.

Let with . Let the messages in problem (15) have size . Then (15) is a linear program with variables and constraints.

Proof.

From LP-duality we know that iff and . Hence, (15) can be rewritten as

(19)

is a linear expression and is constant during the computation, hence (19) is a LP. ∎

Proof of Lemma 1 and Lemma 2

Lemma.

Let be a pair of factors related by the coupling constraints and be a corresponding dual vector. Let and satisfy

(20)

Then implies .

Proof.

Let be a solution of (13) at which the dual lower bound (11) is attained before the update and be an integral solution at which the dual lower bound is attained after has been updated. Variable as chosen in (13) is optimal for and for by construction. We need to prove

(21)

We shuffle all terms with variables , to the right side and all other terms to the left side.