GROUPING OBJECTS BASED ON THEIR APPEARANCE

International Journal of Artificial Intelligence & Applications (IJAIA), Vol. 4, No. 4, July 2013
DOI : 10.5121/ijaia.2013.4408 77
GROUPING OBJECTS BASED ON THEIR APPEARANCE
Altamirano Robles Luis Carlos1
, Zacarías Flores Fernando2
, De León Gómez
Oscar Antonio3
and Sánchez López Abraham4
1,2,3,4
ComputerScience, AutonomousUniversity of Puebla (BUAP), Puebla, México
fzflores@yahoo.com.mx, altamirano@cs.buap.mx, oleongomez@gmail.com
asanchez@cs.buap.mx
ABSTRACT
The use of clustering algorithms for partition to establish a hierarchical structure in a library of object
models based on appearance is deployed. The main contribution corresponds to a novel and intuitive
algorithm for clustering of models based on their appearance, closer to “human behavior”. This divides
the complete set into subclasses. Immediately, divides each of these in a number of predefined groups to
complete the levels of hierarchy that the user wants. Whose main purpose is to obtain a competitive
classification compared to what a human would perform.
KEYWORDS
Vision, Appearance, Models,Clustering and Hierarchy
1. INTRODUCTION
At the beginning of object recognition through digital images, was used the well-known
geometric paradigm. This was mainly to the wide availability of algorithms for detecting edges
and corners. Despite the good results obtained, is shown that using only the way to recognize
objects has its limitations. For instance, if we have two equal cubes in shape and size but one red
and the other blue, all systems based on geometric characteristicswill say that it is the same object
because both objects have the same shape. This happens because there is a substantial information
loss,inherent to the object in question.
Other side, appearance-based approach has been proposed as an alternative to geometrical
approaches. However, this approach also has advantages (increases recognition rates, principally)
and disadvantages (depends on illumination conditions, pose and other parameters, y of course,
requires a lot of images in order to build models). But, since appearance models offer a promises
alternative to traditional models, they are been studying widely[10, 13, 14].
Because it is not easy to establish a hierarchy among models learnt by the machine, most authors
prefer to works in different directions, such that: improve recognition rate, improve lost of
information, compression methods (PCA, for example); one exception are works made by Nelson
and others [11, 12]. However, they use just contour of objects for establish any kind of hierarchy
no clear and depends of primitives previously selected.
In this paper, we present a novel and intuitive approach to establish a hierarchy which does not
depend of any primitive or uses just contour of objects. Instead, we use all visual information
about object appearance: of course contour, but also, color, texture, pose, etc. Our approach is
intuitive because is closer to “human behavior”: our proposal uses a not supervised approach in

78
order to build several different clusters and each one has an explanation in terms of human
behavior. They establish a hierarchy beginning with objects and finalizing with 1, 2, or the
number of trees that we give a priori, like we want to divide a group between men and women:
we establish the only parameter for our algorithm equal to 2 and we can obtain this grouping with
2 classes, but which classes? Answer: several different 2 classes, one pair for each possible
duality: men/women, black/white, short/tall, etc., like human does.
So, we developeda more natural object classifier that considers the following aspects:
• Appropriate ways to differentiate between objects modeled by sets of digital images
(measures of similarity/dissimilarity).
• Clustering algorithms able to put objects models in class,
• Classification based not only on points, as the traditional algorithms.
Different experiments (shown below) shows that our approach exhibits results close to “human
results”, in the sense above explained.
2. PRELIMINARIES
The following subsections present basic concepts about processing objects.
2.1. Models based on appearance
The object appearance is given by: the combination of shape, its reflectance properties, and its
position in front the camera in scene and the illumination present in this [6]. Should be noted that
the first two are inherent to the object, however, in the third, the object’s position in front of the
sensor and illumination may vary from scene to scene[6].Likewise, is clear that this definition
implies that the appearance of an object at a particular time can be derived from a digital image in
which that element is captured in a specific scene. Similarly, then may be used a set of images to
generate a model of the object under different circumstances. In figure 1, you can see a graphical
representation of an object's appearance[6].
Figure 1. An object and its model
Usually, appearance-based approach works as follows: are taken a number of images of the object
you want to build the model,an image for each of the possible ways in which the object is
intended that may appear in an imageand combined with the various lighting conditions that
impinge upon it.Next, these images are stored in a format chosen (gif, tiff, bmp, etc.),for later use;
these images themselves can form the model of the object or, they can extract some features that
you think describe the object accurately, thus, these constitute the model. Finally, the recognition
process itself,is to compare the model with a new image that is likely to contain the object (of
course, this latter process requires some extra processing on the new image, such as object
segmentation of interest in the image, interpolating between images, etc.).
2.2. Creating models using Hotelling transform
Due to the need for a large number of images to a single object model and in practice it is
common to have large quantities of different object, we need to find a compressed representation

of all the images. Normally, It is possible to use a common technique in pattern recognition
named “Hotelling transform” or commonly known as principal component analysis [
Figure
The general procedure is as follows: [
1. Each image is stacked by columns
2. Each image is normalized by dividing by their respective norm
࢏࢐
′
ൌ
ܑ
⋮ܑ
3. We obtain the average vector of all images
࢏࢓ ൌ
4. Each image is subtracted from the average
5. Matrix P is obtained, where each column is one of the vectors previously calculated
6. We obtain the covariance matrix implied.
*T
Q P P=
7. Are obtained the eigenvalues
covariance matrix,
i iλ λ=
1/2
i i ie Peλ
−
=
Next, we obtain the set of eigenvectors or eigenimages, these vectors and eigenvalues
dimensional eigenspace. The dimension of the eigenspace corresponding to the number of
principal components you want to use. 20 are normally sufficient as maximum. We consider that
this should include all the images of all objects. Immediately, we solv
equations
eA P=
We obtain the matrix A containing necessary coefficients to reconstruct any image of any
pattern.This compression process shows insignificant losses of information for problem that
concerns us. In figure 3 the result of the manifolds extraction is shown for two objects and, in
figure 2 some of the eigenimages can be observed obtained by means of this method.
2.3. Similarity measures
Typically when you have a set of models and a test object b
some overlap in these, global data can differ greatly the strategy is to find the common view and
measure the similarity of those parts of the data.
as follows:
1 2 2( , ) ( , )d M M min d C M=
Normally, It is possible to use a common technique in pattern recognition
commonly known as principal component analysis [
e1. Eigen-image and a reconstructed image
The general procedure is as follows: [6], [4]:
Each image is stacked by columns
Each image is normalized by dividing by their respective norm
ܑ‫ܒ‬
ܑ‫ ܒ‬⋮
the average vector of all images
∑ ܑ‫ܒ‬
′‫ܚ‬
‫ܒ‬ୀ૚
‫ܚ‬
Each image is subtracted from the average
Matrix P is obtained, where each column is one of the vectors previously calculated
We obtain the covariance matrix implied.
obtained the eigenvalues and eigenvectors of which are obtained corresponding to the
Next, we obtain the set of eigenvectors or eigenimages, these vectors and eigenvalues
this should include all the images of all objects. Immediately, we solve the following system of
In figure 3 the result of the manifolds extraction is shown for two objects and, in
Typically when you have a set of models and a test object but with different views that contain
measure the similarity of those parts of the data. Then, we define the distance between manifolds
1 2 2( , ) ( , )id M M min d C M (1)
79
Normally, It is possible to use a common technique in pattern recognition
commonly known as principal component analysis [6].
Matrix P is obtained, where each column is one of the vectors previously calculated
and eigenvectors of which are obtained corresponding to the
Next, we obtain the set of eigenvectors or eigenimages, these vectors and eigenvalues form a low-
e the following system of
In figure 3 the result of the manifolds extraction is shown for two objects and, in
ut with different views that contain
we define the distance between manifolds

where 1i r= … , is the number of patches that make up the first manifold
between manifolds is calculated as the best patch similarity adjusted
For this calculation it is necessary to
2.3.1. Measure distance between subspaces
The measure of distance between subspaces is denoted by
unified definition. In this paper we used the concept of
[1].With these concepts are defined the following measures distance. Main idea is that the
nonlinear manifolds are composed of linear subspaces and the distance of these components
defines the similarity between two ma
M is a manifold consisting of linear subspaces.
Once partitioned manifolds, it is only necessary to calculate the distance among subspaces to
proceed. A way to make it is to calculate the main angles that refle
subspaces without taking the data into account
Other methods take into account the similarity through the averages of each local pattern in a
similarity measure among exemplary
measures is better for the problem than we want to solve
The distance from a subespace to a manifold is denoted as
following way
( , ) ( , )d S M mind S C
This means to find the minimum distance between a subspace and the lineal components of
manifold.
2.3.3. Distance from a manifold to a manifold
The distance of a manifold to a manifold is denoted as
way:
( , ) ( , )
i jC M C M
d M M mind C M
min min d C C
∈ ∈
=
Clearly, the similarity between M
they adapt. In [Wang2008] this distance is used in recognition of faces through sets of images. In
this paper we will use it to get grouping of models of objects based apparently.
is the number of patches that make up the first manifold. Thus, the distance
as the best patch similarity adjusted.
For this calculation it is necessary to obtain other measures presented below.
The measure of distance between subspaces is denoted by 1 2( , )d S S and currently there is not a
unified definition. In this paper we used the concept of principal angles for good performance
].With these concepts are defined the following measures distance. Main idea is that the
defines the similarity between two manifolds. Formally denoted as: 1 2, , ,M C C C= …
is a manifold consisting of linear subspaces.
is to calculate the main angles that reflect the variations among the
ng the data into account [1].
Figure 3. Examples calculated
similarity measure among exemplary. It is easy to understand that to use the coalition of both
measures is better for the problem than we want to solve.
The distance from a subespace to a manifold is denoted as ( , )d S M and is defined in the
( , ) ( , )
i
i
C M
d S M mind S C
∈
=
2.3.3. Distance from a manifold to a manifold
The distance of a manifold to a manifold is denoted as 1 2( , )d M M and is defined in the following
1
1 2
1 2 2( , ) ( , )
( , )
i
i j
i
C M
i j
C M C M
d M M mind C M
min min d C C
∈
′∈ ∈
=
′
1M and 2M is calculated as similarity of local models that better
this paper we will use it to get grouping of models of objects based apparently.
80
Thus, the distance
and currently there is not a
principal angles for good performance
].With these concepts are defined the following measures distance. Main idea is that the
, , , mM C C C= … where
ct the variations among the
It is easy to understand that to use the coalition of both
and is defined in the
and is defined in the following
rity of local models that better

81
2.4. Distance among manifolds
As it was already defined previously, Matrix A contains manifolds of the objects in the library.
These define a trajectory parameterized in k
R where k is the number of selected main
components. Main problem of grouping models of objects based on apparently is that it should be
treated with this compact representation. It is essential to find a measure or set of measures of
similarity are necessary to make a classification of objects similar to what a human would. In the
literature it is common to find similarity measures between objects represented as a vector
(column orrow) withd characteristics, stored in database of N objects (dataset) [2]. The strength
of the relationship between two points of a dataset is indicated by a similarity coefficient, the
larger this coefficient, said objects are related more.
Next, we define the distance between manifolds as a measure of similarity between models based
on appearance.
2.4.1. Maximal Lineal Patch
After obtaining the Mainfolds, as mentioned in previous section, is necessary that each manifold
is divided into a series of subspaces called maximal lineal patches.The concept of maximal lineal
patche (MLP) is defined as a set of points defining a maximal linear subspace, where the linear
perturbation of these points is given naturally as the deflection between the Euclidean distance
and geodesic distance.
Figure 4. Examples of Manifolds obtained from COIL-20
The main idea is to split the data set into several MLPs such that sets of images containing very similar to
each other. These patches are calculated from a random point of dataset until the condition of linearity is
lost. This procedure guarantees the local linearity of models and adaptively controls the number of models
calculated. Formally, [7] be 1,..., NX x x= a data set, in our case are points that make up manifolds, where
D
ix R∈ and Nis the number of points in X. we divide X into a disjoint collection of MLPs iC such that:
1
( ) ( ) ( )
1 2
1
,
( , , 1,2, , )
, , , ( )
N
i
i
i j
m
i i i
i Ni i
i
X C
C C i j i j m
C x x x N N n
=
=
=
=
∪ ∅ ≠ = …
= … =∑
U
where m is the number of patches and iN is the number of points in the patch.

82
Figure 5. Partitioning of a model based on appearance
For partitioning is necessary to calculate the Euclidean distance matrix ED and the matrix of geodesic
distances GD . Calculation of the MLPs is used as an intermediate step to find the distance between two
manifolds. This distance has been used in face recognition through image sets [7]. As mentioned above,
the idea is to split the set of images in several patches which are embedded in a linear subspace. The
following shows the algorithm developed.
Initially
1i = , iC = ∅, TX = ∅ , TX = ∅ , RX X= .
1. While TX ≠ ∅
2. Randomly select a point in RX as seed ( )
1
i
x ,
2.1. ( ) ( )
1 1{ }, { }i i
i R RC x X X x= = −
2.2. For i
m ix C∀ ∈
2.2.1. Identify each of k nearest neighbors cx as candidate
2.2.2. If c Rx X∈ and ( ) ( )
( ( , ) / ( , ))i i
G c n E c nD x x D x x θ〈
2.2.3. { }i i cC C x= U
2.2.4. { }R R cX X x= −
EndIf
EndFor 3
1
, ; 1,
i
T j R T i
j
X C X X X i i C
=
= = − ← + =∅U
EndWhile
After calculating a linear patch, its average is calculated. This is taken as a representative of set
and summarizes information of selected patch that is formed by numerically similar images. The
importance of these representatives is reflected in the reduction of the number of images required
to represent the object appearance and is taken into account for the calculation of the similarity of
a complete model with other.
2.4.2. Measure distance between subspaces by variations
For two subspaces iC and jC we define “a distance measure by variations” as their average
canonical correlations as follows:
1
( , ) /
r
V i j k
k
d C C r σ
=
= ∑

83
Where 1 rσ σ… are the canonical correlations and are obtained by the method described in [1]
and r is the minimum sizes of linear subspaces.
2.4.3. Measure distance between copies
To complement the above measure is defined as measured between copies of the duplicate
correlation ie and je :
ࢊࡱ(ࢋ࢏, ࢋ࢐) ൌ
⋮ ࢋ࢏ ⋮ . ⋮ ࢋ࢐ ⋮
ࢋ࢏
ࢀ
ࢋ࢐
2.4.4. Similarity between models based on appearance
Finally, the distance between subspaces is obtained as the weighted average of the two previous
measures as follows:
( , ) (1 ) ( , ) ( , )i j E i j V i jd C C d C C d C Cα α= − +
When comparing two sets of images the first shows how similar the appearance of objects, while
the other is so close that common variation trends of the two sets. Thus, we can easily define the
distance between manifolds as in equation (1).
2.5. Clustering algorithms
The K-means algorithm is one of the most used due to its simplicity and speed. The core idea is to
group the data repeatedly around the centroids calculated from an initial allocation of the data in a
fixed number of k groups. The class membership changes according to the evaluation of an
function error until this error does not change significantly or data belonging either. The error
function is defined as follows [5]:
Let D be a data set with n instances, and be 1 2, , , kC C C… k cluster disjoint from D. The error
function is calculated as:
1
( , ( ))
i
k
i
i x C
E d x Cµ
= ∈
= ∑∑
where ( )iCµ is the centroid of cluster iC and (.,.)d is the Euclidean distance typically.
The Keywords section begins with the word, “Keywords” in 13 pt. Times New Roman, bold
italics, “Small Caps” font with a 6pt. spacing following. There may be up to five keywords (or
short phrases) separated by commas and six spaces, in 10 pt. Times New Roman italics. An 18
pt. line spacing follows.
2.5.1. Spectral clustering
Graph theory can be used to partition data. In these techniques we use a measure of similarity
between the data to form a weighted adjacency matrix. Intuitively, we want find a partition of the
graph where the edges of the nodes that are in different groups have low weight and have high
weights when they are in the same group [3], [8].
Let ( , )G V E= be anundirected graph with a set of vertices 1, , nV v v= … . We assume that each
edge has an associated positive weight. The heavy adjacency matrix , , 1, ,ijW w i j n= = … , where
if 0ijw = , vertices i and j are not connected. Since G is not addressed must be ij jiw w= .
Furthermore, degree of vertex is defined as follow:

84
1
n
i ij
j
d w
=
= ∑
The diagonal matrix having the degrees of the vertices is called degree matrix and is denoted as
D . Likewise define heavy adjacency matrix and degree matrix. Next, we define the un-
normalized Laplacian matrix of a graph as:
L D W= −
This matrix has several interesting properties and. is positive semi-definite and the number of
eigenvectors of this matrix is directly related to the number of connected components of the graph
[3]. Besides, this method is closely related to the method kmeans kernelizado and cuts of graphs
normalized [9].
2.5.2. Normalized Laplacian Matrices
In the literature there are two standard Laplacian matrices. Each one is closely related to each
other and is defined as:
1/2 1/2
1
sym
rw
L D LD
L D L−
=
=
Matrix symL is denoted so because it is symmetrical and matrix rwL is intimately related to
random walks.
As in the case of the non-normalized matrix, the multiplicity of the first Laplacian eigenvalue of
the matrix is related to the number of connected components of the graph of heavy adjacencies
[8].In this study we used the non-standard matrix and matrix symL for the calculation of the
groupings.
3. Our proposal
The intention of using different clustering methods for partitioning is to achieve an initial division
of the set of models and then apply the same algorithm again, dividing the classes obtained in the
previous step in another hierarchical level. Besides using the similarity measure manifolds also be
transformed so that they can be manipulated as a single object and not as a set of points. This was
achieved by calculating the centroid of each manifold and calculating the standard deviation.This
defines a sphere that contains almost all points of the manifold. This defines then a sphere that
contains almost all points of the manifold.
Immediately show the method generates a hierarchy of a set of object models.
The proposed method can be summarized as follows:
1. It creates an initial partition of the original set.
2. Each class obtained is divided k into groups, the parameter k is user defined for each
level of the hierarchy
3. Results are stored
4. Repeat from step 2 to complete the required levels
Finally, we obtained a tree structure that shows how have been divided every class obtained,in
each of the iterations calculated. The rates of each object displayed in the tree nodes.
4. TEST AND RESULTS
Tests were conducted on the classic-looking set of models supplied by Columbia University,
COIL-20. The results obtained were compared with those validated by a user visually, this in

85
order to check that the classification obtained is similar or not. The results obtained are presented
in a tree structure where nodes tags are interpreted according to the following scheme:
Figure 6. scheme of object index
The system was executed several times for each of the algorithms and with different numbers of
levels and Partitions by level, found empirically that in the particular case of the library used,
three levels and no more than three partitions per group are sufficient to achieve good rankings.
In figure 7, you can see one of the groups calculated by the algorithm. Note that the classification
of objects 4.8 and 16 places in the same class and this will not change using other methods
proposed, change only the objects that are grouped into higher levels of the hierarchy This is
because the calculation of the distance between manifolds gives greater weight to objects
appearance so those in which the dominant gray tone is very similar are grouped in the same class
as evidenced in the case of the wood pieces. This measure can also reflect some special
characteristics as how objects in the case of figure 7 the majority shares (objects are higher than
wide, almost the same gray and have almost the same kind of symmetry) which may be an
indication that the results can be improved somewhat by mixing these results to those obtained
using only borders.
Make a change of models representation calculating the centroids allows us to use unmodified
existing algorithms but at the same time preserves all the innate drawbacks of these same. To
avoid the particular algorithms derived using k-means, we will use the spectral clustering on the
centroids.
4.1. Spectral clustering by manifolds centroids
Below are shown some of the results obtained in testing algorithm. For this particular test we used
three levels and a vector (3,3,2)v = .
Figure 7. One group obtained with the proposal.

86
5. CONCLUSIONS AND FUTURE WORK
Although algorithm is the type of divide and conquer, its complexity is polynomial since the
algorithms used (kmeans, spectral) assure us that the groups formed are very similar to each other
and very different from the other groups thereby avoiding combinatorial check what is the best
partition. However, when an object has been misclassified, is impossible to assign another class
and in the best case can be segregated from the group to the next level calculated. On the other
hand,using a measure of similarity between objects, results were very similar to those obtained by
other methods since the model library allows. In addition, it was found that the system is able to
match the results obtained by the user and also is able to detect similarities not evident at a
glance.In figure 8 we show some results obtained with our proposal.
The results have demonstrated the usefulness of similarity measures alternative to classic and the
data processing as a means to obtain better results. However, much work remains to be done in
this field in which this research aims to pioneer.Some things to incorporate into future research
are:
Figure 8. An unexpected group?.
• Use models based on appearance only for objects borders. With the aim of using as a
complementary method.
• Select different algorithms to apply to each level. The results may be improved if
different algorithms are used at each level of the hierarchy.
• Testing with other models libraries and compare results. It is essential to experiment
with different libraries to delimit the scope of the proposal developed in this work and and
modify it to make it applicable to more general problems.
• Automatically analyze the results. To quantitatively compare the classifications obtained
by the system with those made by a human.
• Improving the representation of hierarchical structure.
• Explore the possibility of calculating the tree dendrogram.
• Investigate possible applications of this proposal.
All these points can be taken as a basis for further work in order to find new applications, detect
problems andand generate other proposals that serve to extend knowledge currently hasin the
areas of computer vision and clustering algorithms.
ACKNOWLEDGEMENTS
The authors would like to thank to Autonomous University of Puebla for their financial
support.Óscar De León Gómez would like to thank to CONACyT by the scholarship No. 368458.
REFERENCES
[1] Å.BjörckandG.H.Golub.Numerical methods forcomputing angles between linear subspaces.
Mathematics of Computation, 27:579594, 1973.

87
[2] Clustering. IEEE Press, 2009.
[3] Ulrike Luxburg. A tutorial on spectral clustering. Statistics and Computing, 17(4):395–416,
December 2007.
[4] Taylor C.J. Cootes,T.F. Statistical model of appearance for computer vision. Technical report,
Imaging Science Biomedical Engineering, University of Manchester,March, 2004.
[5] Jianhong Wu Guojun Gan, Chaoqun Ma. Data Clustering: Theory, Algorithms, and applications.
ASA-SIAM Series on Statistics and Applied Probability, SIAM, 2007.
[6] Nene,Nayar, Murase. Parametric appearance representation. Nayar, S. K., Poggio, T., (Eds.) Early
Visual Learning ,OxfordUniversity Press. NewYork, 1:131–160, 1996.
[7] Xilin Chen WenGao Ruiping Wang, Shiguang Shan. Manifold-manifold distance with application to
face recognition based on image set. IEEE,2008.
[8] Francis R. Bach y Michael I. Jordan.Learning spectral clustering.Technical Report UCB/CSD-03-
1249, EECS Department, Universityof California, Berkeley, Jun 2003.
[9] Inderjit S. Dhillon. Kernel k-means, spectral clustering and normalized cuts. In pp. 551–556. ACM
Press, 2004.
[10] Mundy, J., Liu, A., Pillow, N., et. al., “An experimental comparison ofappearance and geometric
model based recognition”, Object Representationin Computer Vision II, in Proceedings of ECCV96
International Workshop,Cambridge U.K., pp. 247-269. April, 1996.
[11] Andrea Selinger and Randal C. Nelson, ``A Perceptual Grouping Hierarchy for Appearance-Based 3D
Object Recognition'', Computer Vision and Image Understanding, vol. 76, no. 1, October 1999, pp.83-
92.
[12] Randal C. Nelson and Andrea Selinger ``A Cubist Approach to Object Recognition'', International
Conference on Computer Vision (ICCV98), Bombay, India, January 1998, 614-621.
[13] Zaman Khan and Adnan Ibraheem. Hand Gesture recognition: A literature review. International
journal of artificial Intelligence & Applications, Vol 3, No. 4, July 2012.
[14] Bimi Jain, M.K. Gupta and JyotiBharti.Efficient iris recognition algorithm using method of
moment.International journal of artificial Intelligence & Applications, Vol 3, No. 5, September 2012.
Authors Biography
Dr. Fernando Zacarias is full time professor in the Computer Science Department at
Autonomous University of Puebla. He has directed and participated in research projects
and development in this area from 1995, with results that they have been reported in more
than 50 publications of international level. And one of the most important is “Answer set
programming and applications”. Conacyt project, Reference number: 37837-A. He has
been a member of several international committees such as: The IEEE Latin America Transaction,
IAENG – International journal of Computer Science, International conference on Advances in Mobile
Computing and Multimedia, etc.
Dr. Luis Carlos Altamirano received his Ph.D. degree at National Polytechnic Institute
(México) in 2002. His interest areas include: Computer Vision, Image Processing,
Remote Sensing and Artificial Intelligence. He has participated in several different
projects including: Oil projects and Computer Vision Projects. Currently, he is full
time professor in the Computer Science Department at Autonomous University of
Puebla and head of Postgraduate Department at the same Institution.
Dr. Abraham Sánchez López is full time professor in the Computer Science Department at Autonomous
University of Puebla. He was member of SNI (National System of Researchers) from 2003 until 2012
(Level 1). His main areas include: Artificial Intelligence, Robot planning and simulation. He received his
Ph. D. degree in 2003.
M.C. Óscar Antonio De León Gomez received his Master´s degree at Computer Science Department
atAutonomous University of Puebla (2012). He has been developing several different projects on
Computer Vision at INAOE (National Institute of Astrophysics, Optical and Electronics), where currently
he is working

GROUPING OBJECTS BASED ON THEIR APPEARANCE

More Related Content

What's hot (18)

Similar to GROUPING OBJECTS BASED ON THEIR APPEARANCE (20)

Recently uploaded (20)

GROUPING OBJECTS BASED ON THEIR APPEARANCE