An Introduction to Graph-Cut

By: Paul Scovanner

An Introduction to Graph-Cut
ƒ Overview
ƒ Min-cut
ƒ Normalized-Cut
ƒ Applications

1
An Introduction to Graph-Cut
ƒ Graph-cut is an algorithm that finds a globally optimal
segmentation solution.
ƒ Also know as Min-cut.
ƒ Equivalent to Max-flow. [1]

[1] Wu and Leahy: An Optimal Graph Theoretic Approach to Data Clustering:…

What is a “cut”?
A graph G = (V,E) can be partitioned into two disjoint
sets, A, B, A ∪ B = V , A ∩ B = 0
by simply removing edges connecting the two parts.

The degree of dissimilarity between these two pieces

can be computed as total weight of the edges that have
been removed. In graph theoretic language it is called
the cut:
cut ( A, B ) = ∑ w(u, v)
u∈ A,v∈B
[2].

[2] Shi and Malik: Normalized cuts and image segmentation.

2
Example cut

cut(A,B)=

∑(
G = (V,E)

)
A B

Finding the Minimum-cut

3
Finding the Minimum-cut

Finding the Minimum-cut

4
Finding the Minimum-cut

Finding the Minimum-cut

5
Finding the Minimum-cut

The problem with Min-cut

Min-cut 1

Min-cut 2

better cut

Need to account for cluster similarity

6
Normalized-cut
ƒ Instead use normalized cut (Ncut).

cut ( A, B ) cut ( A, B )
Ncut ( A, B ) = +
assoc ( A, V ) assoc ( B, V )

assoc( A, V ) = ∑ w(u, t )
u∈ A,t∈V

⎛ assoc ( A, A) assoc( B, B ) ⎞
Nassoc ( A, B ) = ⎜⎜ + ⎟⎟
⎝ assoc ( A, V ) assoc( B,V ) ⎠

Normalized-cut
cut ( A, B ) cut ( A, B)
Ncut ( A, B ) = +
assoc ( A,V ) assoc( B, V )

assoc ( A, V ) − assoc ( A, A) assoc ( B, V ) − assoc ( B, B )

= +
assoc( A, V ) assoc ( B, V )

⎛ assoc( A, A) assoc( B, B) ⎞
= 2 − ⎜⎜ + ⎟⎟
⎝ assoc( A, V ) assoc( B, V ) ⎠

= 2 − Nassoc( A, B)

7
 Pixel labeling problem

Given Find
Labeling f = (f1,…,fn)
2 5
2 5
1
1
3 4
3 4

Such that the sum of the

Assignment cost for giving a assignment costs and separation
particular label to a particular costs (the energy E) is small
node. Written as D.

Separation cost for assigning a

particular pair of labels to
neighboring nodes. Written as V.

Energy Minimization
ƒ Optimizing the labeling problem can be thought
of as minimizing some energy function.

measure of image discrepancy

measure of smoothness or
other visual constraints

8
The Labeling Problem

Common idea behind many Computer Vision problems

Assign labels to pixels based on noisy measurements (input images)
In the presence of uncertainties, find the best Labeling !

(Stereo, 3D Reconstruction, Segmentation, Image Restoration)

Choices of V
Robust Not robust
Potts Linear
model model
V (α , β ) V (α , β )

α − β α − β

Quadratic
Truncated
V (α , β ) model
linear model V (α , β )

α − β
α − β

9
What do graph cuts provide?

ƒ For less interesting V, polynomial

algorithm for global minimum!
ƒ For a particularly interestingV,
approximation algorithm
ƒ Proof of NP hardness
ƒ For many choices of V, algorithms that find
a “strong” local minimum
ƒ Very strong experimental results

Graph Cut based Segmentation

10
Graph Cut based Segmentation

User Guided Segmentation;

Specifies hard constraints.

hard
t
n-links a cut
constraint

hard
constraint
s

3D Applications

2D and Time

[3] Wang et al.: Interactive Video Cutout

11
 Some Results

Input Image Efros and Freeman '01 Graph-cut

(Image Quilting)

[3] Kwatra et al.: Graphcut Textures: Image and Video Synthesis Using Graph Cuts

Some Results

[4] Kwatra et al.: Graphcut Textures: Image and Video Synthesis Using Graph Cuts

12
Some Results

[5] Rother, Kolmogorov and Blake:

“GrabCut“ - Interactive Foreground Extraction using Iterated Graph Cuts

Some Results

13
Some Results

Some Results

14
zSo far weight measurement has been
distance, but could also use appearance,
texture, or other information to calculate
similarity measure.

15