Image Segmentation

Image Segmentation
Introduction. The goal of image segmentation is to cluster pixels into
salient image regions, i.e., regions corresponding to individual surfaces,
objects, or natural parts of objects.

A segmentation could be used for object recognition, occlusion boundary estimation within motion or stereo systems, image compression,
image editing, or image database look-up.
We consider bottom-up image segmentation. That is, we ignore (topdown) contributions from object recognition in the segmentation process.
For input we primarily consider image brightness here, although similar techniques can be used with colour, motion, and/or stereo disparity
information.
Reading on Segmentation: See Chapter 14 of the text.
2503: Segmentation

c A.D. Jepson and D.J. Fleet, 2007

Page: 1

Example Segmentations: Simple Scenes
Segmentations of simple gray-level images can provide useful information about the surfaces in the scene.
Original Image

Segmentation (by SMC)

Note, unlike edge images, these boundaries delimit disjoint image regions (i.e. they are closed).
2503: Segmentation

Page: 2

Siren Song of Segmentation
Why would a good segmentation be useful? Imagine...
Parent to baby: “Look, there is a baby horse with its mommy!”
Baby:
Reasoning
1.
2.
3.
4.
5.

Image

Follow pointing gesture.
Acquire image.
horse is an animal
animal ; quadruped
baby horse ; small horse

Visual Task: Seek correlates
of two similar quadrupeds in image,
one smaller than the other.

Bottom-Up Segmentation

Parse of Two Quadrupeds

Baby: “Gaaa.” (Translation: “Eureka, I can see!”)

2503: Segmentation

Page: 3

Key Questions
1. How well can we expect to segment images without recognizing
objects (i.e. bottom-up segmentation)?
2. What determines a segment? How can we pose the problem mathematically?
3. How do we solve the speciﬁed problem(s)?
4. How can we evaluate the results?

2503: Segmentation

Page: 4

Example Segmentations: Horses Image
Original Image

LV

SMC

H

ED

NC

Which is the best segmentation? Why?
2503: Segmentation

Page: 5

Example Segmentations: Tiger Image
Original Image

LV

SMC

H

ED

NC

Group these into K categories based on quality. (K = 2?)
2503: Segmentation

Page: 6

Observations on Example Segmentations
The previous segmentations were done by the local variation (LV) algorithm [7], spectral min-cut (SMC) [6], human (H) [11, 9], edgeaugmented mean-shift (ED) [4, 3], and normalized cut (NC) [13, 5].
• The quality of the segmentation depends on the image. Smoothly
shaded surfaces with clear gray-level steps between different surfaces are ideal for the above algorithms.
• Humans probably use object recognition in conjunction with segmentation, although the machine algorithms exhibited above do
not.
• For relatively simple images it is plausible that machine segmentations, such as those shown on p.2, are useful for several visual
tasks, including object recognition.
• For more complex images (pp. 5, 6), the machine segmentations
provide a less reliable indicator for surface boundaries, and their
utility for subsequent processing becomes questionable.
• While many segmentation algorithms work well with simple examples, they will all break down given examples with enough clutter and camouﬂage. The assessment of segmentation algorithms
therefore needs to be done on standardized datasets.
2503: Segmentation

Page: 7

Current Goals
• Provide a brief introduction to the current image segmentation literature, including:
– Feature space clustering approaches.
– Graph-based approaches.
• Discuss the inherent assumptions different approaches make about
what constitutes a good segment.
• Emphasize general mathematical tools that are promising.
• Discuss metrics for evaluating the results.

2503: Segmentation

Page: 8

Clustering in Feature Space
Given an image I(x), consider feature vectors F (x) of the form


x


 I(x)  .
F (x) = 

L(x)

Here L(x) is a vector of local image features, perhaps bandpass ﬁlter
responses. For colour images, F (x) would also include information
about the colour at pixel x.
In order to segment the image we might seek a clustering of the feature
vectors F (x) observed in that image. A compact region of the image
having a distinct gray-level or colour will correspond to a region in the
feature space with a relatively high density of sampled feature vectors.

2503: Segmentation

Page: 9

Mixture of Gaussians Model
A natural approach is then to model the observed feature vector distribution using a mixture of Gaussians (MoG) model M ,
K

p(F |M ) =

πk g(F | mk , Σk ).
k=1

Here πk ≥ 0 are the mixing coefﬁcients, with

K
k=1 πk

= 1, and mk ,

Σk are the means and covariances of the component Gaussians.
For a given K, the parameters {(πk , mk , Σk )}K of the MoG model
k=1
can be ﬁt to the data {F (x)}x∈X using maximum-likelihood (here X
denotes the set of all pixels).
Penalized likelihood (aka minimum description length (MDL)) can be
used to select the number of components, K.

2503: Segmentation

Page: 10

Maximum Ownership Labelling
The segment label c(x) = k for a pixel x is the k which maximizes the
ownership of F (x) in the MoG model M . That is,
c(x) = arg max

πk g(F (x) | mk , Σk )

k

p(F (x) | M )

.

Here K = 3 (above right). The maxownership image was post-processed using connected components and small regions were discarded (gray). The average colour of the remaining large components is shown (right). The width of
the segment boundaries is due to the use
of a spatial texture feature.

From Blobworld [2].

Variations: The MoG model can be replaced by K-means (see text), or
restricted to use low-dimensional parameterizations for Σk (eg. block
diagonal).
2503: Segmentation

Page: 11

Assumptions Come Home to Roost
The quality of the resulting segmentation depends on the degree to
which the given image matches the (implicit) assumptions we began
with, namely:
1. Different segments form compact, well-separated clusters in F .
2. Gaussian components in M correspond to salient regions.

From Blobworld [2].

Nevertheless, this feature space clustering can be useful for extracting rough summaries of image content suitable for querying image
databases [2].
2503: Segmentation

Page: 12

Mean-Shift Segmentation
The mean-shift segmentation algorithm [4] also considers the probability density of feature vectors F (x) obtained from a given image.
However, a non-parametric model of the density is used instead of
an MoG. In particular, a kernel-density estimate is used, with
p K (F ) ≡

1
|X|

K(F − F (x)), with F ∈ RD ,
x∈X

where X is the set of all pixels in the image, |X| is the number of
pixels, and K(e) is a kernel.
Common choices for K(e) have the form
K(e) = k(e T Σ−1 e),

(1)

where k(s) is a concave decreasing function of the squared deviation
s ≡ e T Σ−1e ≥ 0. For example,
k(s) = ce−s/2 ,
k(s) = c 1 − s

for a Gaussian kernel,
+,

(2)

for an Epanechnikov kernel.

(3)

Here c = c(Σ) is a normalizing constant which ensures K(e) integrates
to one, and z

+

denotes positive rectiﬁcation, i.e. z

+

≡ max(z, 0).

We show an example next.
2503: Segmentation

Page: 13

Example Feature Density

From Comaniciu and Meer [4].

The kernel density estimate using the Epanechnikov kernel (c) on the
2d feature points in (a). The covariance parameter Σ of the kernel K(e)
determines the smoothness of the density estimate pK (F ). The tradeoff is between sampling artifacts (kernel too narrow) versus loss of resolution in pK (F ) (kernel too broad).
2503: Segmentation

Page: 14

Mean-Shift Iterations
We will use the modes (i.e. peaks) of pK (F ) to be segmentation labels,
replacing the use of the component labels in the previous MoG model.
That is, we wish to locally solve
F∗ = arg max pK (F ).
F

This is similar to robust M-estimation, although here we are maximizing the objective function pK (F ), not minimizing it. A similar derivation to the one for M-estimation shows F∗ must satisfy
F∗ =

w(F (x) − F∗)F (x) /
x∈X

w(F (x) − F∗)
x∈X

where w(e) = −k (e T Σ−1e) and k (s) =

dk
ds (s).

In words, F∗ must be

the weighted mean of F (x) using the weights w(F (x) − F∗) centered
on F∗.
The analogue of the iterative reweighting idea used in M-estimation is
to solve for F∗ here by iterating the mean-shift equation
Fj+1 =

x∈X

w(F (x) − Fj )F (x)

x∈X

w(F (x) − Fj )

.

(4)

Note Fj+1 is just the weighted mean of the feature points F (x), with
the weights w(F (x) − Fj ) centered on the previous guess Fj .
2503: Segmentation

Page: 15

Watersheds of Mean-Shift
The label for an arbitrary pixel x0 denotes the mode that the mean shift
iterations (4) converge to, when started at the feature F0 = F (x0). That
is, the segments produced by mean-shift are defined to be the domains
of convergence (aka watersheds) of the mean-shift iterations.
In the figure on p.14 the trajectories of mean-shift are shown in (c). The
labelling resulting from the watersheds is shown by the colours in (b).
Properties:
1. Convergence: The mean-shift iterations converge to a stationary
point of pK (F ) (see [4]).
2. Anti-edge Detection: The mean-shift iterations are repelled from
local maxima of the norm of the gradient (wrt x) of F T (x)Σ−1F (x).
This occurs, for example, at strong edges in the image I(x).
3. Fragmentation of Constant Gradient Regions: The density pK (F )
is constant (up to discretization artifacts) in regions where the gradient of F T (x)Σ−1F (x) is constant. For example, when

I(x) is

constant, every point of pK (F ) is stationary and the mean-shift iterations stall (see figure p.14, part c). Postprocessing is required to
keep only salient local maxima (see [4]).
2503: Segmentation

Page: 16

Example Mean-Shift Segmentations
Segmentations from the basic mean-shift algorithm:

The scale of the mean-shift kernel (controlled by Σ) roughly controls
the size and shape of the extracted regions. There is a trade-off between
maintaining the salient boundaries but suffering over-segmentation, versus missing some of the important boundaries and under-segmenting
the image. The segmentations above illustrate a typical compromise.
An enhanced system (EDISON [3]) combines the mean-shift algorithm
with image edge information. An edge-saliency measure is used to
modify the weight function used in the mean-shift equation (4). This
eases the above trade-off, allowing weak boundaries to be kept in the
segmentation without incurring as much over-segmentation. Image
segmentation results using the EDISON system are shown on pp. 56 (labelled ED). The use of salient-edge information signiﬁcantly improves the results.
2503: Segmentation

Page: 17

Similarity Graph Based Methods
Graph-based methods provide an alternative to feature space clustering.
A weighted undirected graph G = (V, E) is formed, with the set of
vertices V corresponding to the pixels x in the image. Edges E in
the graph are taken between any two pixels xi and xj within a small
distance of each other.
The edge weight w(xi, xj ) ≥ 0 reflects the dissimilarity (alternatively,
the similarity) between the two image neighbourhoods centered on pixels xi and xj . A common form of the weight function is to use w(xi, xj ) =
1 − a(xi, xj ) where the affinity a(xi, xj ) is given by
1

a(xi, xj ) ≡ e− 2 (F (xi )−F (xj ))

T Σ−1 (F (x )−F (x ))
i
j

.

Here F (x) is a feature vector associated with pixel x, for example:
1. F (x) = I(x), so the affinity is determined only by the grey-level
difference between neighbouring pixels,
2. F (x) = I(x), the RGB values for a colour image, or some mapping
of the RGB values to a more uniform colour space (eg. L*u*v*).
3. F (x) includes texture primitives, such as local filter responses,
along with the brightness and/or colour at pixel x.
2503: Segmentation

Page: 18

Connected Components (Not Robust)
A simple approach is to delete all edges between dissimilar pixels (i.e.,
with weights w(xi , xj ) > τ ), and then seek connected components
(CCs) in the remaining graph.
Note that a single edge with w(xi, xj ) ≤ τ would be sufficient to cause
two desired regions to be merged. Therefore CCs are not robust to stray
links (aka “leaks”) between regions. The consequence is that there is
often no suitable value of τ which gives a useful segmentation.
Kruskal’s Algorithm. In passing, it is useful to point out that an efficient way to do CC clustering, with a variable τ , is to first build a
minimal spanning tree (MST) of the graph. Kruskal’s algorithm can be
used, which is a greedy approach guaranteed to give an optimal MST.
Beginning with the completely disconnected graph, edges are added
one at a time in increasing order of their weights, so long as adding an
edge does not introduce cycles in the current sub-graph.
The CCs of the decimated graph (with edges having w(xi, xj ) > τ
removed) are then efficiently computed by deleting these same edges
from the MST. The trees in the resulting forest provide the desired CCs.
A modified version of Kruskal’s algorithm is considered next.
2503: Segmentation

Page: 19

Local Variation Method
Felzenszwalb and Huttenlocher [7] introduce a simple but effective
modiﬁcation of Kruskal’s algorithm. As in Kruskal’s algorithm, it begins with the completely disconnected graph, edges are added one at a
time in increasing order of their weights, maintaining a forest of MSTs
for the current components.
During processing, each MST Ci is associated with a threshold
T (Ci) = w(Ci) + k/|Ci|

(5)

where w(Ci) is the maximum weight in the spanning tree Ci (i.e. the
local variation of Ci). Also k > 0 is a constant, and |Ci| is the number
of pixels in Ci.
Suppose the edge (xk , xl ) is to be processed next, and its two endpoints
are in two separate MSTs Ci and Cj . Then these MSTs are merged by
adding the edge (xk , xl ) only if
w(xk , xl ) ≤ min(T (Ci), T (Cj )).

(6)

Note that, as the size of Ci increases, (5) and (6) dictate an increasingly
tight upper bound T (Ci) (compared to the largest weight w(Ci) in Ci)
for the acceptable afﬁnity of an edge merging Ci with another region.

2503: Segmentation

Page: 20

Examples of Local Variation Segmentation
Sorting the edges according to weight causes the algorithm to grow
relatively homogeneous regions first.
The parameter k in (5) roughly controls the size of the regions in the
resulting segmentation. Larger k provides a looser constraint (6), and
allows more merging.
k = 50

k = 150

k = 250

The merging is sensitive to the local variation within the regions being
merged. Due to the increasingly tight bound (5), a large homogeneous
region Ci is only merged using edges with weights at most fractionally
larger than w(Ci), the largest affinity in the MST Ci. However, this
bound is much looser for small regions Ci, encouraging their growth.
The approach has a tendency to produce narrow regions along ‘true’
segment boundaries (see examples above).
The approach is very efficient computationally, requiring O(e log(e))
operations where e is the number of edges.
2503: Segmentation

Page: 21

Source-Sink Minimum Cut
An alternative graph-based approach makes use of efficient solutions
of the max-flow/min-cut problem between source and sink nodes in
directed graphs.
S-T Min Cut

From Boykov and Kolmogorov [1].

S-T Min-Cut Problem. An S-T graph is a weighted directed graph
with two identified nodes, the source s and the sink t. We seek a minimum cut separating s and t. That is, we seek a partioning of the graph
into two sets of nodes F and G, with G = V − F , s ∈ F , and t ∈ G,
such that the linkage
L(F, G) =

a(xi, xj ).

(7)

xi ∈F,xj ∈G

is minimized.
2503: Segmentation

Page: 22

Source-Sink Minimum Cut (Cont.)
Efficient algorithms have recently been developed that solve the S-T
min-cut problem (see [1]).
The S-T min-cut problem is computationally much simpler than the
more general graph partitioning problem, which is to find a (nonempty) partition F and G = V − F which minimizes L(F, G) (i.e.,
without any further constraints, such as s ∈ F and t ∈ G.)
To take advantage of an efficient solution to the S-T min-cut problem,
we need to generate an S-T graph. Given two disjoint sets of pixels
S and T , we form a weighted directed graph as follows. For each
edge (xi, xj ) in the previous undirected graph, the two directed edges
xi, xj and the reverse edge xj , xi are included. Both of these edges
are weighted by the affinity a(xi, xj ). In addition, two additional nodes
s and t are created, namely the source and sink nodes, respectively. Finally, infinitely weighted directed links s, xi and xj , t are included
for each xi ∈ S and xj ∈ T .
The resulting S-T min-cut then provides a globally minimum cost cut
between the sets of pixels S and T .

2503: Segmentation

Page: 23

Seed Regions for S-T Min Cut
The sets S and T (connected to the source and sink, respectively)
should satisfy:
1. Each S and T generated must be sufficiently large (otherwise the
minimum cut is either S and V − S, or T and V − T ),
2. Each S and T should be contained within different ‘true’ segments
(due to the infinite weights, neither S or T will be partitioned),
3. Enough pairs S and T should be generated to identify most of the
salient segments in the image.
One suitable generation process is discussed in Estrada et al [6]. It
is based on spectral properties of a matrix representing the affinities.
Sample results are given in segmentations labelled SMC on pp.2, 5,
and 6 above.
The process is much more computationally intensive than the previous
ones. Several hundred min-cut problems are typically solved for different S, T , and these alone require several minutes on relatively small
images (eg. 40K pixels).
The intriguing property of this approach is that the S-T min-cut algorithm computes globally optimal cuts (subject to the proposals for S
and T ).
2503: Segmentation

Page: 24

Normalized Cut
Finally we outline the normalized cut approach of Shi and Malik [13].
Here we seek a partition F and G = V − F of the afﬁnity weighted,
undirected graph (without source and sink nodes). In order to avoid
partitions where one of F or G is a tiny region, Shi and Malik propose
the normalized cut criterion, namely that F and G should minimize
N (F, G) ≡ L(F, G)

1
1
+
,
L(F, V ) L(G, V )

(8)

where L is the linkage deﬁned in (7).
Unfortunately, the resulting graph partitioning problem,
F = arg min N (F, V − F ),
F ⊂V

(9)

is computationally intractable [13]. Therefore we must seek algorithms
which provide approximate solutions of (9).

Note any segmentation technique can be used for generating proposals for suitable
regions F , for which N (F, V − F ) could be evaluated. Indeed, the SMC approach
above can be viewed as using S and T to provide lower bounds on the terms L(F, V )
and L(G, V ) (namely L(S, V ) and L(T, V ), respectively), and then using the S-T
min cut to globally minimize L(F, G) subject to S ⊂ F and T ⊂ G.

2503: Segmentation

Page: 25

Discrete Rayleigh Quotient
Shi and Malik [13] prove that (9) is equivalent to the discrete optimization problem
y T (D − A)y
subject to yi ∈ {1, −b} and d T y = 0.
arg min
y
y T Dy

(10)

Here A is the N × N symmetric matrix of afﬁnities a(xi, xj ), which
is arranged (say) according to the raster ordering of the pixels xi , i =
1, . . . , N . Also, d = A1, where 1 is the N -vector of all ones, b > 0, and
D is the diagonal matrix with Di,i = di.
Given a solution y of (10), the corresponding solution F of (9) is then
obtained by setting F = {xi | yi > 0}. And, vice versa, given F we set
yi = 1 for each xi ∈ F , and set the other elements of y to −b, where
b > 0 is chosen such that d T y = 0.

2503: Segmentation

Page: 26

Spectral Approximation for Normalized Cut
Equation (10) is a discrete version of a standard eigenvector formulation, namely the Rayleigh quotient. This suggests using the tractable
approximation obtained by temporarily allowing y to be a real-valued
vector (instead of 2-valued). By setting y = D −1/2u we ﬁnd
u T (I − B)u
arg min
subject to d T /2u = 0.
Tu
u
u

(11)

with B = D −1/2AD −1/2, a symmetric matrix. This is a standard eigenvalue problem in linear algebra!
Equation (11) can be simpliﬁed further by noting that u = d 1/2 is an
eigenvector of B with eigenvalue 1. It therefore must be an eigenvector
of I − B with eigenvalue 0. Moreover, it can be shown that all the
eigenvalues of I − B are in the interval [0, 2]. Thus (11) is the standard
Rayleigh quotient form for the eigenvector u of I − B with the second
smallest eigenvalue.

2503: Segmentation

Page: 27

Spectral Approximation (Cont.)
In the original Ncut algorithm [13], an approximation to a discrete solution of (10) is then obtained by thresholding D 1/2u at each of a set
of values. For each threshold τ , the Ncut objective function (8) is evaluated, and the best value of τ is selected. This produces two regions
F and V − F . Regions are then recursively partitioned using the same
approach, until a user-speciﬁed number of segments is obtained. See
pp.5-6 for examples.

The step of thresholding the second largest eigenvector to provide a
partitioning proposal is a key limitation of the approach. In practice,
the approximation only appears to be consistently reliable when there
is exactly one obvious way to partition the data. More recently, Yu and
Shi [14] attempt to alleviate this problem by extracting K segments
from the subspace spanned by the K eigenvectors of I − B having the
smallest eigenvalues.
For further information, see the reading list in the CVPR 2004, graphbased segmentation tutorial [12].

2503: Segmentation

Page: 28

Natural Image Boundaries
As we have seen, segmentation involves finding salient regions and
their boundaries.
A boundary in an image is a contour that represents the change from
one object or surface to another. This is distinct from image edges,
which mark rapid changes in image brightness (say), but may or may
not correspond to salient boundaries.
The previous segmentation techniques could be (and some have been)
usefully coupled with a bottom-up boundary detector. For example:
1. The EDISON mean-shift segmentation algorithm [3] illustrated one
example of this in reweighting the mean-shift iterations based on
salient edge information.
2. The affinities used in the Ncut algorithm [13] use intervening edge
information to reduce the affinities between pairs of pixels [8].
The development of local, bottom-up, boundary detectors is an important problem, complimentary to the segmentation approaches discussed
here. See Martin et al [10] for recent work.

2503: Segmentation

Page: 29

Berkeley Segmentation Database
The Berkeley Segmentation Dataset [9, 11] provides image segmentations done by humans. As stated on the dataset’s webpage:
The goal of this work is to provide an empirical and scientiﬁc
basis for research on image segmentation and boundary detection.
The public portion of this dataset consists of the segmentations of 300
images by roughly 5 humans each, done separately for greylevel and
colour versions of the images. Three examples from one image are
shown below:

Note these segmentations appear to be consistent, except different subjects have decided to resolve particular regions into more or less detail.
This variability should be taken into account in a quantitative comparison of two segmentations.
2503: Segmentation

Page: 30

Benchmarking Segmentation
Segmentation algorithms have been benchmarked on synthetic fractal images. The precision-recall curves for the detection of segment
boundary points were computed (Estrada and Jepson, submitted).

Here the rows correspond to the test image, the true segmentation, and
the results from SE-MinCut, NCut, LocalVariation, and MeanShift.
2503: Segmentation

Page: 31

Precision-Recall on Fractal Images
Tuning Curves for all Algorithms
1
0.9
0.8
0.7

Recall

0.6
0.5
0.4
0.3
0.2
0.1
0
0

SE Min−Cut
Mean Shift
Local Variation
Normalized Cuts
0.2

0.4

0.6

0.8

1

Precision

2503: Segmentation

Page: 32

Precision-Recall on Berkeley Dataset
Tuning Curves for all Algorithms
1
0.9
0.8
0.7

Recall

0.6
0.5
0.4
0.3
0.2
0.1
0
0

SE Min−Cut
Mean Shift
Local Variation
Normalized Cuts
Human
Human Median
0.2

0.4

0.6

0.8

1

Precision

2503: Segmentation

Page: 33

References
[1] Y. Boykov and V. Kolmogorov. An experimental comparison of min-cut/max-ﬂow algorithms for
energy minimization in vision. submitted IEEE Trans. Pattern Anal. and Machine Intell., 2004.
[2] C. Carson, S. Belongie, H. Greenspan, and J. Malik. Blobworld: Image segmentation using
expectation-maximization and its application to image querying. IEEE Trans. Pattern Anal. and
Machine Intell., 24(8):1026–1038, 2002.
[3] C. M. Christoudias, B. Georgescu, and P. Meer. Synergism in low level vision. In 16th International Conference on Pattern Recognition., Quebec City, Canada, volume IV, pages 150–155,
2002.
[4] D. Comaniciu and P. Meer. Mean shift: A robust approach toward feature space analysis. IEEE
Trans. Pattern Anal. and Machine Intell., 24:603–619, 2002.
[5] T. Cour, S. Yu, and J. Shi. Normalized cuts matlab code. Computer and Information Science,
Penn State University. Code available at http://www.cis.upenn.edu/˜jshi/software/.
[6] F.J. Estrada, A.D. Jepson, and C. Chennubhotla. Spectral embedding and min-cut for image
segmentation. In British Machine Vision Conference, 2004.
[7] P.F. Felzenszwalb and D.P. Huttenlocher. Efﬁcient graph-based image segmentation. Int. J. of
Comp. Vis., 59(2):167–181, 2004.
[8] J. Malik, S. Belongie, T. Leung, and J. Shi. Contour and texture analysis for image segmentation.
Int. J. of Computer Vision, 43(1):7–27, 2001.
[9] D. Martin and C. Fowlkes.
The Berkeley Segmentation Dataset and Benchmark.
http://www.cs.berkeley.edu/projects/vision/grouping/segbench/.
[10] D. Martin, C. Fowlkes, and J. Malik. Learning to detect natural image boundaries using local
brightness, color, and texture cues. IEEE Trans. Pattern Anal. and Machine Intell., 26(5):530–
549, 2004.
[11] D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images
and its application to evaluating segmentation algorithms and measuring ecological statistics. In
Proc. 8th Int’l Conf. Computer Vision, volume 2, pages 416–423, July 2001.
[12] J. Shi, C. Fowlkes, D. Martin, and E. Sharon. Graph based image segmentation tutorial. CVPR
2004. http://www.cis.upenn.edu/˜jshi/GraphTutorial/.
[13] J. Shi and J. Malik. Normalized cuts and image segmentation. IEEE Trans. Pattern Anal. and
Machine Intell., 22(8):888–905, 2000.
[14] S. Yu and J. Shi. Multiclass spectral clustering. In Proc. Int’l Conf. Computer Vision, 2003.

2503: Segmentation

Notes: 34

Image Segmentation

More Related Content

What's hot (20)

Viewers also liked (7)

Similar to Image Segmentation (20)

More from Syed Muhammad Hammad (15)

Recently uploaded (20)

Image Segmentation