New Method for Evaluation of Video Segmentation Quality
Mahmud Abdulla Mohammad, Ioannis Kaloskampis and Yulia Hicks
School of Engineering, Cardiff University, Queen’s Buildings, The Parade, Cardiff, U.K.
Keywords:
Image/Video Segmentation, Object Segmentation, Image Processing, Edge Detection, Statistical Measures.
Abstract:
Segmentation is an important stage in image/video analysis and understanding. There are many different
approaches and algorithms for image/video segmentation, hence their evaluation is also important in order
to assess the quality of segmentation results. Nonetheless, so far there was little research aimed specifically
at evaluation of video segmentation quality. In this article, we propose the criteria of good quality of video
segmentation suitable for assessment of video segmentations by including a requirement for temporal region
consistency. We also propose a new method for evaluation of video segmentation quality on the basis of the
proposed criteria. The new method can be used both for supervised and unsupervised evaluation. We designed
a test video set specifically for evaluation of our method and evaluated the proposed method using both this
set and segmentations of real life videos. We compared our method against a state of the art supervised
evaluation method. The comparison showed that our method is better at evaluation of perceptual qualities of
video segmentations as well as at highlighting certain defects of video segmentations.
1 INTRODUCTION
Image segmentation means subdividing images into
meaningful segments or regions (Dey et al., 2010;
Morris et al., 1986). Image sequence (video) segmen-
tation can be viewed as an expansion of single image
segmentation (Charron and Hicks, 2010; Greenspan
et al., 2004). The additional challenge in video seg-
mentation is to cater for segmentation consistency
throughout the video (Kaloskampis and Hicks, 2014).
There are many different approaches and algorithms
for image/video segmentation hence it is important to
be able to evaluate the quality of the segmentations
produced by different methods. To date, the evalu-
ation methods are divided into three classes (Zhang
et al., 2008; Correia and Pereira, 2003).
Subjective Evaluation is the process in which hu-
man observers quantify the quality of segmentation
results on the basis of visual description. This is a
complicated and time consuming process, and the re-
sult varies from one observer to another.
Supervised Evaluation is the process in which
a segmented image/frame is compared against a
manually-segmented (ground truth) reference im-
age/frame. Producing the ground truth images is also
a time consuming process, with a degree of disagree-
ment between different people.
Unsupervised Evaluation, which is also known
as stand-alone evaluation or empirical goodness eval-
uation. It works automatically without any extra re-
quirements such as the ground truth images. The
methods in this evaluation class use only low-level
features and do not incorporate semantic information.
Most of evaluation methods are subjective or re-
lated to specific applications. The majority of pro-
posed objective evaluation methods is in the area of
supervised evaluation, with the area of unsupervised
evaluation receiving the least attention (Zhang et al.,
2008). Evaluation is usually based on several crite-
ria, with each criterion considering the quality of the
segmentation from a different aspect. A number of
researchers considered which aspects of the segmen-
tation quality should be evaluated. In the remainder of
this section, we will review the existing criteria along
with metrics used to evaluate them.
Levine and Nazif (Levine and Nazif, 1985) sug-
gested that to design a measure for evaluating the
quality of image segmentation, it is necessary to con-
sider the following: (i) uniformity within regions, (ii)
contrast across regions, (iii) provision for lines and
texture. Haralick and Shapiro (Haralick and Shapiro,
1985) proposed four criteria to evaluate image seg-
mentation: (i) the regions must be uniform and ho-
mogeneous, (ii) adjacent regions should have signifi-
cant differences with respect to the characteristic on
523
Mohammad M., Kaloskampis I. and Hicks Y..
New Method for Evaluation of Video Segmentation Quality.
DOI: 10.5220/0005306205230530
In Proceedings of the 10th International Conference on Computer Vision Theory and Applications (VISAPP-2015), pages 523-530
ISBN: 978-989-758-089-5
Copyright
c
2015 SCITEPRESS (Science and Technology Publications, Lda.)
which they are uniform, (iii) region interiors should
be simple and without holes, (iv) boundaries should
be smooth and accurate. Most of the previous evalu-
ation methods and metrics incorporate the above cri-
teria, either directly or indirectly (Levine and Nazif,
1985; Liu and Yang, 1994; Borsotti et al., 1998; Chen
and Wang, 2004; Zhang et al., 2004; Chabrier et al.,
2006).
Consequently (Zhang et al., 2008) classified the
previous work according to the criteria proposed in
(Haralick and Shapiro, 1985). The classification also
covers unsupervised metrics proposed for evaluation
of image and video segmentation. They concluded
that these criteria had become the de f acto standard
for unsupervised evaluation of image segmentation.
They suggested that the first two criteria were more
characteristic rather than semantic and hence incor-
porated first and second criteria in their work. Zhang
et al. (Zhang et al., 2008) conducted a comparative
evaluation of different approaches after which they
concluded that previous unsupervised approaches for
evaluation of image segmentation methods are insuf-
ficient for comparison of segmentation produced by
different algorithms. The criteria discussed above
have been applied for evaluation of the quality of im-
age segmentation. The previous unsupervised meth-
ods for evaluation of video segmentation methods
(Correia and Pereira, 2003; Erdem et al., 2004) are
limited and not designed for general purpose applica-
tions, with the former method manually labeling data,
and the latter being designed for evaluating video ob-
ject segmentation and tracking algorithms. Likewise,
the metric proposed in (Gelasca and Ebrahimi, 2006)
is based on spatial and temporal accuracy and de-
signed for evaluating video object segmentation.
In addition to the above, there are several popu-
lar supervised evaluation methods based on the im-
age/frame boundaries as opposed to regions. The
boundary precision-recall is used in (Martin et al.,
2001) as a supervised metric for evaluation of image
segmentation. Galasso et al. (Galasso et al., 2013) in-
troduced the volume precision-recall metric for evalu-
ation of video segmentation quality, Xu et al. (Xu and
Corso, 2012) proposed 3D volumetric quality metrics
to evaluate super-voxel methods, which they based on
the boundaries, without taking into account the region
uniformity and consistency.
To summarise the current state of art in the area
of evaluation of the video segmentation quality, the
following is noted: (i) there is no established criteria
for evaluation of overall video segmentation as op-
posed to image segmentation or video object segmen-
tation, (ii) there is a limited number of unsupervised
evaluation methods of video segmentation and they
are not designed for overall video segmentation, (iii)
supervised evaluation methods of video segmentation
consider the boundaries of the segmentations without
taking into account region interiors.
In this work, we propose new criteria of good
quality of video segmentation to include temporal re-
gion consistency. On the basis of the new criteria,
we propose an online method for evaluation of video
segmentation quality, which takes into account the
characteristics of both boundaries and regions. On-
line evaluation can be used to control the parame-
ters of online video segmentation in real-time appli-
cations (Zhang et al., 2008). Proposed method can
be used both for supervised and unsupervised eval-
uation. We design a test video set specifically for
evaluation of the quality of video segmentation and
evaluate the proposed method using both this set and
segmentations of real life videos. We compare our
method against a state of the art supervised evaluation
method both in supervised and unsupervised modes.
The remainder of the paper is organized as follows.
In Section 2, we discuss proposed criteria and met-
rics. In Section 3, we give a detailed overview of the
proposed method. Section 4 provides evaluation and
results. Finally, conclusion is discussed in Section 5.
2 PROPOSED CRITERIA AND
METRICS
An initial step for any evaluation process is deter-
mining the criteria of evaluation. As discussed in
the previous section, a number of researchers pro-
posed several criteria for evaluation of image segmen-
tation. However, there are some differences between
image segmentation and video segmentation, and thus
it is necessary to propose new criteria by consider-
ing additional characteristic relating to good quality
of video segmentation.
For evaluating video segmentation quality, the sta-
bility of the boundaries and consistent region iden-
tity between consequent frames should be evalu-
ated (Grundmann et al., 2010). Given this, and in the
light of the criteria proposed by Haralick and Shapiro,
we propose the following set of criteria:
1. The regions must be uniform, homogeneous, sim-
ple and without holes.
2. Adjacent regions should have significant differ-
ences with respect to the characteristic on which they
are uniform.
3. Corresponding regions between consequent frames
should be consistent.
4. Boundaries of segmented frame should be smooth,
stable and accurate when compared with the bound-
VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications
524
aries of original frame.
This work is based on low level image features
in line with previous unsupervised methods (Zhang
et al., 2008). For this reason, we will not use the
second criterion when evaluating the quality of video
segmentation, as it is difficult to find meaningful ad-
jacent segments without semantic information. In the
next section, we will consider the metrics for mea-
suring the quality of the video segmentation accord-
ing to the remaining three criteria: those for measur-
ing intra-region uniformity and homogeneity (crite-
rion 1), those for measuring region consistency be-
tween consequent frames (criterion 3), and those for
measuring boundary accuracy (criterion 4).
2.1 Intra-region Uniformity and
Homogeneity
The uniformity of regions can be divided into two cat-
egories, color uniformity and texture uniformity. The
former means that the pixel colors of a region should
have similar values, the latter means that each re-
gion should have consistent texture. In (Zhang et al.,
2008) the intra region uniformity metrics are classi-
fied into four classes: based on color error, squared
color error, texture and entropy. We select two sim-
ple and easy to understand metrics, the first is F
RC
(Rosenberger and Chehdi, 2000), which is based on
squared color error and measures the intra region
color disparity, and the second is texture variance
Tex var (Correia and Pereira, 2003), which measures
the texture uniformity and is based on the variance of
Y, U and V layers.
2.2 Temporal Region Consistency
There is a number of methods which can be used to
evaluate consistency between two regions. In addi-
tion, a number of metrics have been designed specif-
ically to measure the similarity between two ground
truth images, such as Variation of Information (VI)
(Meil
˘
a, 2003; Unnikrishnan et al., 2005), Global Con-
sistency Error GCE (Martin et al., 2001) and proba-
bilistic rand index PRI (Unnikrishnan et al., 2005).
GCE and VI have been designed to compare two seg-
mentations, while PRI has been designed to com-
pare more than two segmentations. VI is an infor-
mation based metric, which considers mutual infor-
mation between two segmentations, whilst GCE is
a region based metric, and designed to quantify the
consistency between image segmentations of differ-
ent granularities (Unnikrishnan et al., 2007). In video
segmentation corresponding regions from consecu-
tive frames should have consistent color and granular-
ity. Here, we propose to evaluate consistency between
two consecutive frames by combining GCE with pos-
itive correlation. Although GCE has been used for
image segmentation evaluation, no one has combined
it with positive correlation. Our approach to com-
bine GCE with positive correlation has the advantage
of taking into account both the consistency of region
granularity and the color consistency of regions be-
tween consecutive frames.
2.3 Boundary Stability and Accuracy
To measure the accuracy between the boundaries
of segmented and original frames, we will use F-
measure metric, (Martin et al., 2001) which is the
most popular boundary based metric for evaluation
of image segmentation (Galasso et al., 2013). We
develop a method to detect boundaries of original
frames in case of unsupervised evaluation, and use
ground truth boundaries in case of supervised evalua-
tion. In the next Section we will explain the proposed
method.
3 PROPOSED METHOD
In this work, we propose a method on the basis of the
new criteria proposed in Section 2. Our method can
be used both for supervised and unsupervised evalua-
tion. The former evaluation uses the boundaries of the
ground truth, the latter uses the boundaries found in
the original frame, for which we use a combination of
low-pass filtering to remove noise and multiscale edge
detection. Our method uses the found boundaries at
two different stages. First, our method uses them as
a map controlling the regions of the segmented frame
for measuring intra region uniformity as outlined in
Section 3.1. Second, it uses the detected boundaries
to evaluate the accuracy of the boundaries of the seg-
mented frames.
3.1 Intra-region Uniformity
Selecting semantic regions composing an image
needs either human help or ground truth template,
which are not available for unsupervised segmenta-
tion. We overcome this issue by detecting and using
the boundaries of the original video frames. Thus
our process of evaluation of intra-region uniformity
consists of the following three steps.
1. Detecting Boundaries. We use our own method
relying on a combination of low-pass filtering to
remove noise and multiscale edge detection.
NewMethodforEvaluationofVideoSegmentationQuality
525
2. Selecting Regions f rom Segmented Frame. The
detected boundaries produced from previous step
are used to select the regions of segmented frame.
We will use quad-tree image decomposition to
separate the segmented frame into a number of
rectangular areas not containing any boundaries
from the original video frame. The uniformity of
each of the rectangular areas is then evaluated in
the next step.
3. Evaluating Intra region Uni f ormity. The se-
lected regions produced from the previous step
are evaluated using two metrics selected in Sec-
tion 2.1, namely F
RC
and Tex var.
We will explain the metrics F
RC
and Tex var individ-
ually. Let N be the total number of regions of seg-
mented image I, with the height I
x
and width I
y
, j is
the index of regions j (1, 2, 3,..., N), R
j
represent set
of pixels in the region j where R
j
N
j=1
(R
j
)
, S
j
denotes the area of the region j, C
x
(P) is the color in-
tensity value for pixel P (x red, green, or blue com-
ponent), and the area of full image is S
I
= I
x
× I
y
.
The mean value of component x in region j can be
defined as the following:
b
C
x
(R
j
) =
1
S
I
PR
j
C
x
(P) (1)
F
RC
is based on the squared color error and measures
the intra region color disparity, squared color error
can be defined as the following:
e
2
x
(R
j
) =
PR
j
(C
x
(P)
b
C
x
(R
j
))
2
(2)
F
RC
metric can be explained as the following:
D(I) =
1
N
N
j=1
S
j
S
I
× e
2
x
(R
j
) (3)
where D(I) is the F
RC
color disparity of image/frame
I, and e
2
x
(R
j
) is squared color error of region R
j
.
Tex var (Correia and Pereira, 2003) is defined as the
following:
Tex var(I) =
1
N
N
j=1
1
5
3 × σ
2
y(R
j
)
+ σ
2
u(R
j
)
+ σ
2
v(R
j
)
(4)
where Tex var(R
j
) is the texture variance of the
region(R
j
), σ
y
, σ
u
and σ
v
are the variance of the Y ,
U and V components in R
j
region, respectively. Both
D(I) and Tex var(I) metrics are normalized to intra
region uniformity I
U
and texture uniformity T
U
re-
spectively, by the following function:
η =
1
1 +
ν
0.5
128
0.5
!
× 2 (5)
Figure 1: Result of boundary detection for frame number
one, from left to right: Soccer sequence and Ice sequence.
where η and ν represent normalized value (between 0
and 1) and initial value (between 0 and 128
2
) of the
metrics respectively.
A real scene can consist of both color and texture re-
gions, and it is difficult to decide which region cat-
egory is predominant in the scene. For this reason,
both color uniformity and texture uniformity to mea-
sure the region uniformity are used, and we take their
average to take them both into account.
3.2 Region Consistency
The content of consecutive frames in video sequences
is usually not exactly the same, but still there is con-
sistency and similarity between them, which is de-
pendent on the complexity of the sequence. As we
discussed previously, according to the listed criteria
in Section 2, identical regions between consequent
frames should be consistent both in terms of color and
granularity. In this work, we want to ensure the con-
sistency according to both metrics. To obtain this, we
use the minimum value between global consistency
index GCI used to evaluate granularity consistency
and positive correlation to evaluate color consistency.
GCI can be explained as follows.
Let \ denote set difference, and
|
x
|
the cardinality
of set x. Let S
1
and S
2
be two segmentations. For a
given pixel p
i
, consider the segments that contain p
i
in
S
1
and S
2
. We denote these sets of pixels by R(S
1
, p
i
)
and R(S
2
, p
i
) respectively, the local refinement error
is defined as:
E(S
1
, S
2
, p
i
) =
|
R(S
1
, p
i
) \ R(S
2
, p
i
)
|
|
R(S
1
, p
i
)
|
(6)
VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications
526
Figure 2: Sample of the synthetic video, first row is the ground truth and the second is segmented, both sequences are in the
same order.
(a) (b)
Figure 3: Result of: (a) F and F
0
; (b) P and P
0
.
GCE(S
1
, S
2
) =
1
n
min
i
E(S
1
, S
2
, p
i
),
i
E(S
2
, S
1
, p
i
)
!
(7)
GCE(S
1
, S
2
) is the global consistency error between
frame S
1
and S
2
, and n is the number of pixels.
GCI(S
1
, S
2
) = 1 (GCE(S
1
, S
2
)) (8)
In our case, we applied the GCI to each frame,
which means that S
1
and S
2
represent two consecutive
frames.
The positive correlation between consecutive
frames can be calculated as the following:
Corr(S
1
, S
2
) =
r(S
1
, S
2
) r(S
1
, S
2
) 0
0 r(S
1
, S
2
) < 0
(9)
where r(S
1
, S
2
) is the Pearson’s correlation between
frame S
1
and S
2
, and can be defined as the following:
r(S
1
, S
2
) =
n
(S
1
S
2
) (
S
1
)(
S
2
)
q
(n
S
2
1
(
S
1
)
2
)(n
S
2
2
(
S
2
)
2
)
(10)
3.3 Boundary Assessment
F-measure (Martin et al., 2001) is the most popular
metric in this area, as we discussed in Section 2, we
will use F to refer to F-measure.
F =
2 P R
P + R
(11)
where P is the precision of the boundaries, and R is
the recall of the boundaries.
3.4 Combining Metrics
The selected metrics explained in the previous sec-
tions are combined as a version of F formula. F is
the harmonic mean of precision and recall, with pre-
cision penalising oversegmentation and recall penal-
ising undersegmentation, both of which are important
for evaluating the quality of video segmentation. In
this work we will update precision P to P
0
to include
the metrics evaluating region uniformity and consis-
tency, which also play important role in penalising
over- and under-segmentation.
F
0
=
2 P
0
R
P
0
+ R
(12)
P
0
=
P + α
2
(13)
where P
0
is the updated precision, and it is the average
between precision P and α. We define α as:
α =
2 U C
U +C
(14)
where α is the harmonic mean between intra-region
uniformity U and consistency C. We will define them
as the following:
NewMethodforEvaluationofVideoSegmentationQuality
527
Figure 4: Segmentation evaluation result using F and F
0
for the supervised and unsupervised cases, for six video sequences.
Each video has 7 degrees of segmentation, ground truth, 3 degrees of under segmentation and 3 degrees of oversegmentation.
U =
I
U
+ T
U
2
(15)
C = min(GCI,Corr) (16)
where I
U
is the minimum value of normalized intra re-
gion uniformity among R, G, and B layers, obtained
using Eqs. 3 and 5, T
U
is the normalized texture uni-
formity obtained using Eqs.4 and 5, GCI is the mini-
mum value of global consistency index among R, G,
and B layers (Eq. 8), and Corr is the minimum value
of positive correlation among R, G, and B (Eq. 9).
4 EVALUATION AND RESULTS
4.1 Synthetic Data
We created two synthetic videos of length 100 frames
each. The first video depicts for differently colored
circles moving from different corners towards each
other, meeting in the middle and then moving to the
opposite corners (Figure 2). The second video repre-
sents different defects in the segmentation of the first
video, such as over- and undersegmentation, unde-
tected objects, inconsistent object identity (swapping
of identity between objects), etc. We also insert “cor-
rectly segmented” frames between the “defective seg-
mentations” to represent inconsistent temporal seg-
mentations. The dataset is available by contacting the
authors.
4.2 Real Video Data
The real video dataset is from (Chen and Corso,
2010) and is a subset of the Xiph.org videos. The
selected dataset used in this work can be divided into
three groups: ground truth, oversegmented and under-
segmented. We selected six different videos labelled
with a 24-class semantic pixel labeling as a ground
truth (Chen and Corso, 2010). For each video, from
ground truth frames, we created three degrees of un-
dersegmentation. Also for each video, we created
three degrees of oversegmentation using the hierar-
chical graph-based method (Grundmann et al., 2010).
The length of the videos varied from 69 to 86 frames.
4.3 Results on Synthetic Video
This example explains the ability of F and F
0
to eval-
uate different types of segmentation defects. We ap-
plied F and F
0
to the synthetic video data set described
in the previous section. The results of F are accurate
in most of the cases, but it is not as strict as the F
0
in
VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications
528
Figure 5: Visual comparison of the six degrees of segmentation quality and ground truth.
penalising inconsistent object identity and underseg-
mention. Figure 3 shows the differences between F
and F
0
, and P and P
0
. The frames between number 53
to 57 are undersegmented and frames between 94 and
97 contain inconsistent object identity.
4.4 Results on Real Video
Alongside the synthetic video we evaluated our
method on six real videos, as outlined in Section
4.2. For each real video from the dataset we created
Figure 6: Result of real video evaluation using F and F
0
met-
rics, for the supervised and unsupervised case. We took the
average score of all videos for each segmentation quality.
seven segmentations of different quality, containing
three degrees of oversegmentation, ground truth, and
three degrees of undersegmentation (Figure 5). Fig-
ures 4 and 6 provide the comparative information on
F and F
0
over these segmentations. Whilst F and F
0
are approximately the same for the undersegmented
and ground truth segmentations, their behavior for the
oversegmented areas is different. F
0
is more consis-
tent with the perceptual quality of the segmentations
Over 3, Over 2 and Over 1 than F, where we can ob-
serve significant difference in the quality of Over 2
and Over 1 (Figure 4 and 5).
5 CONCLUSION
The main contributions of this work are the proposal
of the new criteria of good video segmentation qual-
ity, proposal of a new evaluation method based on this
criteria, which can be used both for supervised and
unsupervised evaluation, and creation of a synthetic
video test set specifically for the purpose of evalua-
tion of the performance of the proposed method. The
results showed that our method can evaluate video se-
quences quality better than F, on different types of
NewMethodforEvaluationofVideoSegmentationQuality
529
content. This is due to our method taking into account
region uniformity and consistency between consecu-
tive frames, which we included in the new criteria.
REFERENCES
Borsotti, M., Campadelli, P., and Schettini, R. (1998).
Quantitative evaluation of color image segmentation
results. Pattern Recognition Letters, 19(8):741 – 747.
Chabrier, S., Emile, B., Rosenberger, C., and Laurent, H.
(2006). Unsupervised performance evaluation of im-
age segmentation. EURASIP Journal on Applied Sig-
nal Processing, 2006:217–217.
Charron, C. and Hicks, Y. (2010). An evolving mog for
online image sequence segmentation. In ICIP, pages
2189–2192.
Chen, A. Y. C. and Corso, J. J. (2010). Propagating multi-
class pixel labels throughout video frames. In Proc. of
Western New York Image Processing Workshop.
Chen, H.-C. and Wang, S.-J. (2004). The use of visible
color difference in the quantitative evaluation of color
image segmentation. Proc. IEEE Int. Conf. on Acous-
tics, Speech, and Signal Processing (ICASSP), 3:iii–
593–6 vol.3.
Correia, P. and Pereira, F. (2003). Objective evaluation of
video segmentation quality. IEEE Trans. on Image
Processing, 12(2):186–200.
Dey, V., Zhang, Y., and Zhong, M. (2010). A review on im-
age segmentation techniques with remote sensing per-
spective. In Proc. of the Int. Society for Photogram-
metry and Remote Sensing Symposium (ISPRS), vol-
ume 38, pages 5–7.
Erdem, C¸ . E., Sankur, B., and Tekalp, A. M. (2004).
Performance measures for video object segmentation
and tracking. IEEE Trans. on Image Processing,
13(7):937–951.
Galasso, F., Nagaraja, N. S., C
´
ardenas, T. J., Brox, T.,
and Schiele, B. (2013). A unified video segmenta-
tion benchmark: Annotation, metrics and analysis. In
IEEE Int. Conf. on Computer Vision (ICCV).
Gelasca, E. D. and Ebrahimi, T. (2006). On evaluating met-
rics for video segmentation algorithms. In Second Int.
Workshop on Video Processing and Quality Metrics
for Consumer Electronics (VPQM ).
Greenspan, H., Dvir, G., and Rubner, Y. (2004). Context-
dependent segmentation and matching in image
databases. Computer Vision and Image Understand-
ing, 93(1):86–109.
Grundmann, M., Kwatra, V., Han, M., and Essa, I. (2010).
Efficient hierarchical graph-based video segmenta-
tion. In IEEE Conf. on Computer Vision and Pattern
Recognition (CVPR), pages 2141–2148.
Haralick, R. M. and Shapiro, L. G. (1985). Image segmen-
tation techniques. Computer vision, graphics, and im-
age processing, 29(1):100–132.
Kaloskampis, I. and Hicks, Y. (2014). Estimating adap-
tive coefficients of evolving gmms for online video
segmentation. In 6th Int. Symp. on Communications,
Control and Signal Processing (ISCCSP), pages 513–
516.
Levine, M. D. and Nazif, A. M. (1985). Dynamic measure-
ment of computer generated image segmentations.
IEEE Trans. on Pattern Analysis and Machine Intel-
ligence, PAMI-7(2):155–164.
Liu, J. and Yang, Y.-H. (1994). Multiresolution color image
segmentation. IEEE Trans. on Pattern Analysis and
Machine Intelligence, 16(7):689–700.
Martin, D., Fowlkes, C., Tal, D., and Malik, J. (2001). A
database of human segmented natural images and its
application to evaluating segmentation algorithms and
measuring ecological statistics. In Proc. Eighth IEEE
Int. Conf. on Computer Vision (ICCV), volume 2,
pages 416–423.
Meil
˘
a, M. (2003). Comparing clusterings by the variation of
information. In Learning theory and kernel machines,
volume 2777, pages 173–187.
Morris, O., Lee, M. d. J., and Constantinides, A. (1986).
Graph theory for image analysis: an approach based
on the shortest spanning tree. Communications, Radar
and Signal Processing, IEE Proc. F, 133(2):146–152.
Rosenberger, C. and Chehdi, K. (2000). Genetic fusion:
application to multi-components image segmentation.
In Proc. IEEE Int. Conf. on Acoustics, Speech, and
Signal Processing (ICASSP), volume 6, pages 2223–
2226.
Unnikrishnan, R., Pantofaru, C., and Hebert, M. (2005). A
measure for objective evaluation of image segmenta-
tion algorithms. In IEEE Computer Society Conf. on
Computer Vision and Pattern Recognition-Workshops
(CVPR) Workshops, pages 34–34.
Unnikrishnan, R., Pantofaru, C., and Hebert, M. (2007). To-
ward objective evaluation of image segmentation al-
gorithms. IEEE Trans. on Pattern Analysis and Ma-
chine Intelligence, 29(6):929–944.
Xu, C. and Corso, J. J. (2012). Evaluation of super-voxel
methods for early video processing. In IEEE Conf.
on Computer Vision and Pattern Recognition (CVPR),
pages 1202–1209.
Zhang, H., Fritts, J. E., and Goldman, S. A. (2004). An
entropy-based objective evaluation method for image
segmentation. In Electronic Imaging, pages 38–49.
Zhang, H., Fritts, J. E., and Goldman, S. A. (2008). Image
segmentation evaluation: A survey of unsupervised
methods. Computer Vision and Image Understanding,
110(2):260–280.
VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications
530