# A Selective Overview of Graph Cut Energy Minimization A Selective Overview of Graph Cut Energy Minimization Algorithms Ramin Zabih Computer Science Department Cornell University Joint work with Yuri Boykov, Vladimir Kolmogorov and Olga Veksler Outline 1. 2. 3. Philosophy and motivation Graph cut algorithms Using graph cuts for energy minimization in vision What energy functions can be minimized via graph cuts?

Problem reductions Suppose youre given a problem you dont know how to solve Turn it into one that you can solve If any instance of a hard problem can be turned into your problem, then your problem is at least as hard Pixel labeling problem Given S 1,, m 2

Find N S S Labeling f f1 , , f m 5 2 1 5 1 3 4 L l1 , ,lk

Assignment cost for giving a particular label to a particular node. Written as D. Separation cost for assigning a particular pair of 3 4 Such that the sum of the assignment costs and separation costs (the energy E) is small Solving pixel labeling problems We want to minimize the energy

E(f) arg min f D ( p, f p p assignment costs V ( f , fq ) p , qN

p separation costs Problem show up constantly in vision ) And in other fields as well Bayesian justification (MRFs) Pixel labeling for stereo Stereo

Labels are shifts (disparities, hence depths) At the right disparity, I1(p) I2(p+d) Assignment cost is D(p,d) = (I1(p) I2(p+d))2 Neighboring pixels should be at similar depths Except at the borders of objects! Classical solutions No good solutions until recently General purpose optimization methods

(correlation) (annealing) Fastcuts What do graph cuts provide? For less interesting V, polynomial algorithm for global minimum! For a particularly interesting V, approximation algorithm Proof of NP hardness

For many choices of V, algorithms that find a strong local minimum Very strong experimental results Part A: Graph Cuts Everything you never wanted to know about cuts and flows (but were afraid to ask) Maximum flow problem a flow F Max flow problem: source sink

S T A graph with two terminals Each edge is a pipe Find the largest flow F of water that can be sent from the source to the sink along the pipes Edge weights give the pipes capacity Minimum cut problem

a cut C Min cut problem: source sink S T A graph with two terminals Find the cheapest way to cut the edges so that the source is

completely separated from the sink Edge weights now represent cutting costs Max flow/Min cut theorem Max Flow = Min Cut: source S sink T

A graph with two terminals Maximum flow saturates the edges along the minimum cut. Ford and Fulkerson, 1962 Problem reduction! Ford and Fulkerson gave first polynomial time algorithm for globally optimal solution Fast algorithms for min cut Max flow problem can be solved fast

This is not at all obvious Many algorithms, well sketch one Variants of min cut are NP-hard Multiway cut problem More than 2 terminals Find lowest cost edges separating them all Augmenting Path

algorithms source S Find a path from S to T along nonsaturated edges Increase flow along this path until some edge saturates sink T A graph with two terminals

Augmenting Path algorithms source S Find a path from S to T along nonsaturated edges Increase flow along this path until some edge saturates Find next path

Increase flow sink T A graph with two terminals Augmenting Path algorithms source S Find a path from S to T along nonsaturated edges

Increase flow along this path until some edge saturates sink T A graph with two terminals Iterate until all paths from S to T have at least one saturated edge Implementation notes There are many fast flow algorithms

Augmenting paths depends on ordering Breadth first = Edmonds-Karp Vision problems have many short paths Subtleties needed due to directed edges [BK 04] gives an algorithm especially for vision problems Software is freely available Part B: Graph cuts in vision How you can turn vision

problems into min cut problems for fun and profit (or, not) A surprising result Minimizing E in vision is difficult With 2 labels, can find the global min! Huge search space Many local minima

[Greig, Porteus, Shehult, 1986] Problem reduction to min cut Construction Exactly 1 green link is cut, for every pixel 0 1 Cuts are labelings Two obvious

encodings If two adjacent pixels end up linked to different terminals, the black link between them must be cut Cost of cut is energy of labeling Smoothness term matters V determines what the solution prefers

Consider uniform D Comes from the MRFs prior Convex V over-penalizes discontinuities |d1 d2| Non-convex V is important T[d1 != d2], called the Potts model min(|d1 - d2|,K) Why is the problem hard? Minimizing the energy for the interesting choices of V is NP-hard

Problem reduction from multiway cut Result is somewhat recent [BV&Z 01] Requires exponential search Dimension = number of pixels Convex V is easy Graph cuts can rapidly compute the global minimum

Convex V, contiguous integer labels [Ishikawa 04] Another amazing result Not of practical interest (IMHO) You can really see the over-smoothing Local minima and moves A local minimum is all we can hope for Minimum relative to a set of moves

For the important class of V Better than any nearby solution We will compute a local minimum with respect to very powerful moves Green-yellow swap move Starting point Red expansion move Swap move algorithm 1. Start with an arbitrary labeling 2. Cycle through every label pair (,)

in some order 2.1 Find the lowest E labeling within a single ,-swap 2.2 Go there if this has lower E than the current labeling 3. If E did not decrease in the cycle, were done. Otherwise, go to step 2 Expansion move algorithm 1. Start with an arbitrary labeling 2. For each label in some order

2.1 Find the lowest E labeling within a single -expansion 2.2 Go there if this has lower E than the current labeling 3. If E did not decrease in the cycle, were done. Otherwise, go to step 2 Algorithm properties In a cycle the energy E doesnt increase

These are greedy methods Convergence guaranteed in O(n) cycles In practice, termination occurs in a few cycles When the algorithms converge, the resulting labeling is a local minimum Even when allowing an arbitrary swap move or expansion move Strong local minima A local minimum with respect to these moves is the best answer in a very large

neighborhood For example, there are O(k 2n) labelings within a single expansion move Starting at an arbitrary labeling, you can get to the global minimum in k expansion moves The expansion move algorithm yields a 2-approximation for Potts model V Binary sub-problem The input problem involves k labels, but the key sub-problem involves only 2

Minimize E over all O(2n) labelings within a single -expansion move from f Each pixel p either keeps its old label fp, or acquire the new label Classical problem reduction To min cut problem Part C: What energy functions can graph cuts minimize? Or, what else can we do with this?

Different D example Birchfield-Tomasi method Compute an intensity interval Use this as basis for D(p,d) Handles sampling error Different V example Stereo Image: White rectangle moves one pixel, background is stationary

Cuts are binary labelings 0 C D B vp A vq F 1 E G

vr S {0, v p } T {1, vq , vr } labeling {v p 0; vq 1; vr 1} cost A B D E Recent progress Over the last 4 years several such problem reductions have been done Graph constructions have been specialized to a

particular E, and quite complex You cant tell by looking at E whether such a construction is possible, let alone how to do it We now have a much more general result for energy functions of binary-valued variables Graph cuts invariably use binary variables Energy functions and graphs Consider E which assigns to any cut (binary labeling) the cost of that cut

Computing the min cut on G is a fast way to find the labeling that minimizes E We will say that G represents E Every weighted graph with two terminals represents some energy function What we really want is to go in the opposite direction! Basic questions For what energy functions E can we construct a graph G that represents E?

I.e., what energy functions can we efficiently minimize using graph cuts? How can we easily construct the graph G that represents E? I.e., given an arbitrary E that we know to be graph-representable, how do we find G? Question #1 What class of energy functions can we efficiently minimize using graph cuts? The classes F2 and F3 Consider functions of n binary variables Functions in F2 can be written as

i E ( v ) E i, j i i ( vi , v j ) i j Functions in F3 can be written as i i, j E (

v ) E i (vi , v j ) i i j i , j ,k E (vi , v j , vk ) i j k Regularity All functions E of 1 binary variable

are defined to be regular A function E of 2 binary variables is regular if E (0,0) E (1,1) E (1,0) E (0,1) A function E of more than 2 binary variables is regular if all its projections are regular Regularity theorem A graph-representable function of binary variables must be regular In fact, minimizing an arbitrary non-regular functions in F2 is NPhard

Reduction from independent set problem F3 Theorem Any regular function in F3 is graphrepresentable With the regularity theorem, this completely characterizes the energy functions E that can be efficiently minimized with graph cuts Assuming E has no terms with >3 variables Question #2 Given: an arbitrary graphrepresentable E

Question: How do we find Desired construction Input: an arbitrary regular EF3 Output: the graph that represents E Additivity theorem The sum of two graph-representable energy functions is itself graphrepresentable Assume that the two graphs are defined on the same sets of vertices, i.e. they

differ only in the edge weights We simply add the edge weights together If there is no edge, we can pretend there is one with weight 0 Regrouping theorem Any regular function in F3 can be re-written as the sum of terms, where each term is regular and involves 3 or fewer variables Combined with the additivity theorem, we need only build a graph for an arbitrary regular term involving 3 or fewer variables

Construction for F2 Consider an arbitrary regular E in F2 We only need to look at a single term, whose form is like D(vp) or V(vp,vq) Example: expansion moves for stereo We will show how the construction works for both types of terms Each term is known to be regular! Data terms

We need a graph construction that imposes one penalty if vp = 0, but another if vp = 1 The penalty can be positive or negative! v p 0 D( p, f p ) v p 1 D( p, ) E (0) E (1) Rewriting the penalty

v p 0 E ( 0) v p 1 E (1) E (1) E (1) E (0) E (1) 0

The penalty Z = E(0) E(1), imposed iff vp = 0, can be positive or negative But graph has non-negative edge weights! Case 1: Z > 0 v p 0 Z v p 1 0 0 Not cut iff vp = 0

(source) v2 vp Z 1 (sink) Case 2: Z < 0 v p 0 Z v p 1 0

Z Z 0 0 (source) -Z Z v2 vp

Not cut iff vp = 1 1 (sink) Smoothness terms vq 0 v p 0 v p 1 vq 1 V ( f p , f q ) V ( f p , ) V ( , f q ) V ( , ) A B C D

We will assume A>0, C>0, D-C>0 and construct the appropriate graph The other cases are very similar Rewriting the term A B A A 0 0 0 D C C D 0 0 C C 0 D C 0 B C A D

0 0 All the entries are non-negative by assumption except the last Which is non-negative by regularity! 0 Not cut iff vp = 0 vq 0 v p 0 v p 1 vq 1 A A 0 0

vq vp A 1 0 C vq 0 v p 0 v p 1 vq 1 0 0 C C vq

vp Not cut iff vp = 1 1 0 D-C vq 0 v p 0 v p 1 vq 1 0 D C 0 D C vq vp

Not cut iff vq = 1 1 vq 0 Non-negative by regularity vq 1 v p 0 0 B C A D v p 1 0 0 0

Not cut iff vp= 0 vp B+C-A-D vq Not cut iff vq = 1 1 Putting it all together 0 vq 0 v p 0 v p 1 vq 1 A B C D

C D-C B+C-A-D vp A 0, C 0 D C 0 A 1 vq One of the other cases 0 vq 0 v p 0

v p 1 vq 1 A B C D C B+C-A-D vp vq A 0, C 0 D C 0 A C-D 1

What energy functions? Partial characterization of the energy functions of binary variables that can be minimized with graph cuts General-purpose construction for an arbitrary regular function in F3 It is no longer necessary to explicitly build the graph (do the problem reduction) Instead, simply check the regularity condition and apply our construction Conclusions

Problem reductions are powerful Beware of general-purpose solutions! Special-purpose ones are fragile Graph cuts have both theoretical and practical interest They are now (relatively) easy to use