Which Strategy to Combine Face Identiﬁcation Tools

with Clothing Similarity: Contesting or Reinforcing?

Sa¨ıd Kharbouche and Michel Plu

Orange Labs 2, Av Pierre Marzin, 22307 Lannion, France

Abstract. This paper describes a novel and efﬁcient approach that integrates

clothing similarity into face identiﬁcation process in personal photos. The in-

formation extracted from people’s clothes would be helpful if they are dissimilar,

however, this information could make errors and noise if we have some people

with similar clothes. To resolve this problem, we propose here a better method-

ology that exploits clothing similarity. The main idea is summarized as follows:

if a person is well identiﬁed in a detected face, instead to reinforce this person in

every face (in other photo) with similar clothes, we contest her/him in every face

with dissimilar clothes. The weight and the inﬂuence of the information extracted

from a face in a photo to another face depend on the spatiotemporal distance be-

tween photos, the similarity degree between the clothes and the incertitude level

about their real identities. We utilize belief functions theory in order to man-

age efﬁciently the imprecision and the uncertainty. Besides, the results obtained

showed off the interest of our approach.

1 Introduction & Previous Work

Every user of digital photo technologies regularly found himself with large collection

of images to annotate for browsing and later photo retrieval. So, lot of systems have

been conceived with the aim of helping users to index and annotate their personal pho-

tos. Some systems are web based, as they permit the users to share online the photos

with their communities. Also, other systems are accessible by mobile devices (GSM,

PDA,...) which allow users a realtime photo sharing [6,9,13]. In order to retrieve these

shared photos, there are many ways to organize them but the most current ways are:

– Indexing by dates: time-taken;

– Indexing by location: place-taken;

– Indexing by people’s identities: photographed people.

The date and location are encoded in photo as a timestamp of creation date and GPS

coordinates, so the indexing by date or location can be an automatic and quit reliable

process. Also in the case of cameraphone, the radio cell id can be used as local iden-

tiﬁcation [6]. But the indexing by people’s identities is very difﬁcult to make it as an

automatic process. In personal photos the faces have various sizes, positions and image

qualities. To resolve people’s identities in photo, some systems use contextual infor-

mation (e.g. date, local, events, comments analysis, social data, co-presence detection

Kharbouche S. and Plu M. (2008).

Which Strategy to Combine Face Identiﬁcation Tools with Clothing Similarity: Contesting or Reinforcing?.

In Metadata Mining for Image Understanding, pages 78-89

DOI: 10.5220/0002338500780089

 SciTePress

using bluetooth...) either lonely or combining with face recognition outputs [6,12].

Clothes appearance has been investigated in many work. The authors in [1] describe

a technique for building a composite photo that represents an event.Their approach clas-

siﬁes the photos in four classes using K-NN algorithm and clothing similarity (color and

texture). The clothes are region of upper body that can be localized by a face detection.

The representative photo of a folder includes four sub photos of faces (cluster centers)

and a main photo that contains the maximum number of faces. In [18] a technique

of semi-automated face annotation was detailed. The proposed technique combines the

both recognition results (face recognition, and clothing similarity), and deals with miss-

ing features of face or clothes. The latter is named contextual features and they are ex-

tracted from the extended face region. The authors used a fusion technique based on

conditional probability. In [3], a system of face semi-automated annotation combines

the face similarity, clothing similarity and uniqueness constraint (no person appears in

the same photo twice) by a loopy belief propagation (LBP). The clothing similarity is

used for the photos of the same event (interval of 4 or 6 hours) and clothing features

(color and texture) are extracted from the upper body region. The position of this re-

gion is relative to the detected frontal face. In this last work, in order to avoid the errors

engendered by clothing similarity, the latter is less weighted compared to face recog-

nition. The paper [16] describes a personal photos clustering approach based on face

recognition and clothing similarity. The clothes pieces are localized by face detector,

and the vector feature is based on descriptors of colors and texture. In addition, it pro-

poses a method for skin detection that exploits face color, and the detected skin will not

be used for further computing. The scores of face recognition and clothing similarity

are combined linearly, so the result score is integrated in logistic function (likelihood

measuring) in order to compute the similarity between faces. The uniqueness constraint

is integrated in the clustering algorithm (K-means) as a hard constraint.

The aim of our system is to suggest the metadata list (people’s identities) of each

detected face in the image. Given two faces which are detected in two different photos

with a small spatiotemporal gap, then their clothes features are extracted. In previous

work [18, 3,16], the use of clothing similarity exploits essentially this idea: If clothing

similarity is high, the identities of faces are probably the same. So, for a face to iden-

tify, every person who has already been detected with similar clothes in other photo of

the same event (same date and place) will be reinforced, thus pushed forward in the

candidates list. This method is called reinforcing strategy. Another strategy (our con-

tribution in this paper) is called contesting strategy that exploits the following idea: if

clothing dissimilarity is high, the identities of faces are probably different. So, for a face

to identify, every person who has already been detected with dissimilar clothes in other

photo will be contested, thus pushed back into candidates list. The contesting strategy is

more logic than reinforcing strategy because, photographed people would be with sim-

ilar color and texture of their upper bodies (similar clothes and skin). Then, we can’t

identify with certainty a face only via its clothing similarity. However, the use of cloth-

ing similarity could provide us with certainty the non-matching between a detected face

and a given person identity that we know his/her clothes. In [16], the clothing similarity

becomes misleading in the case of people with similar clothes. Also, for this reason, the

7979

clothing similarity is less weighted than face recognition in [3]. The reinforcing strategy

results important errors in the case of people with similar clothes.

In our previous work [11] , we proposed a method based on belief functions for

combining face detection, face recognition and gender recognition, and we introduced

a logic-based information named Inter-Zones (or uniqueness constraint). The later con-

sists to contest every person who has already recognized in another face of the same

photo. The thesis [10] details more our system and it explains how we have combined

the information extracted from audio-textual comments, spatiotemporal data and the

information extracted from image analysis tools. Our personal-photos indexing system

was integrated in a web-based system ”Someone” [2] of multimedia documents index-

ing and sharing, and it includes two major information sources:

IA: Image Analysis it includes three tools: Face Detection −F D−, Face Recognition

−F R− and Gender recognition −GR−;

CA: Comments Analysis it consists of recognizing the people who were quoted in

textual and vocal comments.

However, in our precedent work, we identify people’s identities in photo without taking

into account the information that could be extracted from clothes.

In this paper we won’t use ”comment analysis” CA, but we will focus on ”images

analysis” IA and a tool of Clothing Similarity CS. So, we integrate the information is-

sued from others photos of the same collection for face identiﬁcation process. Thereby,

each face in a given photo will take part in the decision for resolving the identity of

all faces of other photos. It depends on spatiotemporal distance between photos, image

analysis (IA) results and the clothing similarity. Belief functions theory has been cho-

sen in our fusion process in order to manage imprecision and uncertainty. Many works

attempt to compare the formal framework of this theory to several other methods of

information fusion, like the probabilistic methods, the possibility theory and to linear

combination. For the multimodality fusion, the belief functions are introduced in many

systems, and it was compared with success to many others techniques of fusion like

possibility and linear combination [4].

This paper is organized into the following sections. Second section gives a sum-

mary of belief function theory, particularly the rules that will be called in our proposed

approach. Third section describes our fusion process, followed by a section that gives

our evaluation and preliminary results. Lastly, we draw our conclusion.

2 Belief Functions Theory

Here, we recall basics of this theory, especially the tools that we will use in our fusion

process. Dempster’s work on lower and upper limits of probability distributions [7]

allows Shafer to build the fundamental of belief functions theory [14].

Let X be an unknown form and, Ψ a non empty ﬁnite set of q assumptions (about

the reality of X), also called the frame of discernment:

Ψ = {h

, h

, . . . , h

}. (1)

8080

Let S be an information source (or piece of evidence) which gives data about the mem-

bership between X and the elements of 2

. A basic probability assignment or mass

function m

[X] maps 2

to the interval [0, 1] and satisﬁes this condition:

A∈2

[X](A) = 1. (2)

[X](A) represents the degree of belief attributed exactly to the subset A of Ψ , this

can’t be divided between the elements of A. The belief function bel

[X] maps 2

[0,1], it can be derived from the mass function using this formula:

bel

[X](A) =

B/B⊆A

[X](B) ∀ A ∈ 2

. (3)

This measure may be interpreted as the minimum degree of belief attributed to A.

2.1 Data Combining Rules

Given two informationsources S

and S

(or two pieces of evidence). To identify X, we

derived two mass functions m

[X] and m

[X] from S

and S

respectively. These

functionscan be incorporatedin one mass function by the conjunctivecombination rule,

noted by

∩

. It is deﬁned for all A ∈ 2

as follows:

[X](A) = (m

[X]

∩

[X])(A); (4)

B∩C=A

[X](B).m

[X](C). (5)

Hence, the degree of conﬂict between S

and S

is the mass attributed to ∅. In order to

work in a closed world so that Ψ must be exhaustive (include all possible assumptions)

and the mass attributed to ∅ must be null, we use Dempster’s rule of combination noted

by ⊕, also called orthogonal rule. It is deﬁned for all A ∈ 2

− ∅ as follows:

[X](A) = (m

[X] ⊕ m

[X])(A); (6)

[X](A)

1 − m

[X](∅)

; (7)

[X](∅) = 0. (8)

Dempster’s rule is possible only if S

and S

are not in total conﬂict, in other words if

[X](∅) < 1. The function m

[X](Ψ) = 1 is a neutral element of ”⊕”:

[X] ⊕ m

[X] = m

[X] ∀S. (9)

And the function m

[X](h

) = 1 is an absorber element of ”⊕”:

[X] ⊕ m

[X] = m

[X] ∀S, with m

[X]

∩

[X](∅) < 1. (10)

8181

2.2 Decision Making

There are several kinds of decision rules employed in the evidence theory. One of those

methods was proposed by Ph. Smets [15]. It transforms the evidence model to a proba-

bility model by a function noted BetP

[X]

(.). It is called pignistic probability func-

tion, which is formalized for all h

as follows:

Be tP

[X]

) =

1 − m

[X](∅)

A∈2

∈A

[X](A)

|A|

. (11)

|A| represents the cardinality of A.

2.3 Processing Time

Evidently, working with 2

|Ψ |

= 2

sets of Ψ expands the fusion processing time. In

order to resolve this problem we have developed an algorithm that deals only with non-

null values of mass function (focal elements). This algorithm will be explained in a

future paper.

3 Fusion Process

This section explains how we integrate the clothing similarity in the process of people’s

identities identiﬁcation in personal-photos. We will try to combine the result of cloth-

ing similarity (CS) with one of our current image analysis IA using two strategies:

contesting and reinforcing. To start with, we will extract two mass functions from the

information sources IA and CS. Then we combine them into a unique mass function

for metadata (people’s identities) relevant measuring.

3.1 Notations

We introduce now some basic annotation that we will use in the rest of this paper.

Given a user U and its address book B, the later represent a set of N known people to

the user U . Each person is represented in this set by: its names (last name, ﬁrst name

and nicknames), its face images, its gender and its id.

B = {P

, P

, ..., P

...., P

}. (12)

Then, to identify any face we use these exhaustive and exclusive discernment frame:

Ω = {P

, P

, ..., P

...., P

, UM, UF, noF ace}. (13)

UM and U F denote unknownmale and unknown female respectively. This frame is ex-

haustive and exclusive because the reality of each detected face is an unique assumption

of Ω. I symbolizes a photos collection:

I = {I

, I

, ..., I

}. (14)

8282

Each photo I

is accompanied with two metadata, photo-taken time t

and photo-taken

place coordinates (GPS coordinates). F

denotes the set, if the faces that have been

detected by F D (face detector) in photo I

= {F

s,1

, F

s,2

, ..., F

s,f

, ..., F

s,z(s)

}. (15)

z(s) is the number of faces that have been detected in photo I

. So, we determine the

clothes detached to these faces:

= {R

s,1

, R

s,2

, ..., R

s,f

, ..., R

s,z(s)

}. (16)

s,f

denotes a part of the clothes attached to the face F

s,f

(a part of upper body region).

3.2 Images Analysis: IA

We seek to identify a face F

s,f

. So, for this face we have four information sources:

face detection (F D) , face recognizer (F R), gender recognizer (GR) and uniqueness

constraint (IZ). All these information are aggregated in IA as follows:

Ω

s,f

] = m

Ω

F D

s,f

] ⊕ m

Ω

F R

s,f

] ⊕ m

Ω

s,f

] ⊕ m

Ω

s,f

]. (17)

For more detail we refer the reader to [11].

3.3 Clothing Similarity: CS

In photos collection I, there are r detected faces, thus r associated clothing parts (with

r = (z(1) + ... + z(S))). Each face and its clothing in a photo different to I

will

be considered as an information source for identifying the face Z

s,f

, consequently we

obtain r − z(s) information sources.

Contesting Strategy. Always, we try to identify the face F

s,f

. Let F

′

be another

face (with s 6= s

′

and f

′

∈ {1, ..., z(s

′

)}) and R

′

be its related clothes, the masses

function issued from R

′

about the metadata P

can be as follows:

Ω

′

s,f

](P

) = α

(s,f),(s

′

)

.bel

Ω

′

](P

); (18)

Ω

′

s,f

](Ω) = 1 −

Ω

′

s,f

](P

). (19)

With bel

Ω

′

](P

, ) be the credibility measure about the assumption that the person

is the real identity of the face F

′

. It is derived from the mass function 17 using the

equation 3. P

represents the union of all elements of Ω except P

. This mass function

associates belief degree to two sets; the incertitude Ω and to opposite of a person P

. It is

called contesting strategy because it could reduce the pertinence degree of the metadata

and can’t increase it. The coefﬁcient α

(s,f),(s

′

)

is computed as follows:

(s,f),(s,f

′

)

= Φ

(d(I

, I

′

)).Φ

(t(I

, I

′

)).(1 − Φ

(h(R

s,f

, R

′

))). (20)

With:

8383

– d(I

, I

′

) be the geographical distance between I

and I

′

– t(I

, I

′

) be the time between time-taken of I

and time-taken of I

′

– h(R

s,f

, R

′

) be the similarity degree between R

s,f

and R

′

– Φ

be an increasing function that maps d(I

, I

′

) to interval [0, 1]. The more the

distance is small, the more its value tends towards 1.

– Φ

be an increasing function that maps t (I

, I

′

) to interval [0, 1]. The more the

temporal gape is small, the more its value tends towards 1.

– Φ

be an increasing function that maps h(R

s,f

, R

′

) to interval [0, 1]. The more

similarity measure is high, the more its value tends towards 1.

The functions Φ

, can be a logistic regression functions as follows:

(dist(I

, I

′

)) =

1 + exp(−λ

.d(I

, I

′

) + δ

)

. (21)

The parameters λ

and δ

permit us to determine a fuzzy threshold between the two

opposite situations, near and far geographically. Besides, it is easy to optimize analyt-

ically these parameters. We can say the same thing for Φ

and Φ

. Theses parameters

can be updated and learned by a relevance feedback. The mass function

Ω

′

s,f

]

veriﬁes these axioms:

– The more t(I

, I

′

) or d(I

, I

′

) or h(R

s,f

, R

′

) is high or bel

Ω

′

](P

) is

low, the weak contestation. In other words, the belief degree allocated to incertitude

(Ω) become high, as a result the function

Ω

′

s,f

] become more neutral (no

information).

– The more t(I

, I

′

) and d(I

, I

′

) are low h(R

s,f

, R

′

) and bel

Ω

′

](P

) is

high, the high contestation.

The masse function 18 is issued from R

′

but it is speciﬁc to the metadata P

. So the

absolute mass function issued from R

′

could be computed as follows:

Ω

′

s,f

] =

i=1

Ω

′

s,f

]. (22)

Finally, the mass function issued from all clothes (to identify F

s,f

) is the fusion of all

mass function issued from other faces in other photos:

Ω

s,f

] =

′

=1,...N ; f

′

=1,...z(s

′

))/(s

′

6=s)

Ω

′

s,f

]. (23)

Reinforcing Strategy. The belief function theory allows us without difﬁculty to ex-

change the contesting strategy to reinforcing strategy. The later is used in several works

[3,16, 18] but with other fusion techniques like linear combining and conditional prob-

ability. If we want working with reinforcing strategy, the equations 18 and 19 become:

Ω

′

s,f

](P

) = α

′

(s,f),(s

′

)

.bel

Ω

′

](P

); (24)

Ω

′

s,f

](Ω) = 1 −

Ω

′

s,f

](P

). (25)

8484

This mass function associates belief degrees to two sets: the incertitude Ω and a person

}. It is called reinforcing strategy because it could increase the pertinence degree of

the metadata P

and it can’t never reduce it. The coefﬁcient α

′

(s,f),(s

′

)

is computed as

follows:

′

(s,f),(s

′

)

= Φ

(dist(I

, I

′

)).Φ

(t(I

, I

′

)).Φ

(h(R

s,f

′

)). (26)

In the reinforcing strategy the mass function

Ω

′

s,f

] veriﬁes these axioms:

– The more t(I

, I

′

) or d(I

, I

′

) or h(R

s,f

, R

′

) is high or be l

Ω

′

](P

) is

low, the weak reinforcing (neutral mass function).

– The more t(I

, I

′

) and d(I

, I

′

) are low and h(R

s,f

, R

′

) and bel

Ω

′

](P

)

are high, the high reinforcing.

Then, we compute the unique mass function issued from clothes recognition CS as

follows (as equations 22 and 23):

Ω

s,f

] =

′

=1,...N ; f

′

=1,...z(s

′

))/(s

′

6=s)

(

i=1

Ω

′

s,f

]). (27)

3.4 Suggested Metadata List

Finally, For a face F

s,f

, we combine the mass functions issued from IA (equation

17) and from CS (equation 23 if we use contesting strategy, or equation 27 if we use

reinforcing strategy):

Ω

IACS

s,f

] = m

Ω

s,f

] ⊕ m

Ω

s,f

]. (28)

Then, we compute the probability pignoistic for each element of Ω using the equation

11. The suggested metadata list will be ranked according to these probability values. In

other words, the metadata P

is high ranked than P

′

(with i

′

∈ {1, ...N }) if:

Be tP m

Ω

IACS

s,f

]

) ≥ BetP m

Ω

IACS

s,f

]

′

). (29)

This list will be suggested to user for the face F

s,f

annotation.

4 Experimentation

In this section we present the technique that we have used for clothes detections and

segmentation, followed by experiment and evaluation results.

4.1 Clothes Detection and Segmentation

This paper focuses on a new method for combining face recognition and clothing sim-

ilarity. But, we can integrate any available clothing detection and segmentation in our

fusion process. We present a simple technique in order to extract the feature vectors of

8585

clothes and computing the similarity between them. If we want to photograph a person,

the face is the most interesting and the upper body appears often. That is because we

have chosen to utilize the upper body. We use a tool to detect only the frontal faces [17],

hence we localize easily the upper body region if a face is detected. For this goal we use

a simple technique that is shown in ﬁgure 1, it is inspirited from [3] and [16]. The face

detector provides us the coordinates ((x

, y

), (x

, y

)) of two extreme points of face

bounding box. Thus, the coordinates of two extreme points of bounding box of upper

body region are computed as follows:

= max{0, x

− ρ

.(x

− x

)}; (30)

= min{H, y

+ ρ

.(y

− y

)}; (31)

= min{W, x

+ ρ

.(x

− x

)}; (32)

= min{H, y

+ ρ

.(y

− y

)}. (33)

W and H denote the width and the height of the photo. The positive values ρ

, ρ

and

allow us to determinate the position and dimensions of upper body region according

to the ones of the detected face (ρ

< ρ

Fig.1. A detected face and its rectangular part of upper body (a clothes part).

In our experiment, the feature vector of R

s,f

is based on color descriptors. The color

are quantiﬁed into HSV color space (hue, saturation, value) in 168 bins in such a way

that 18 bins for hue, 3 bins four saturation, 3 bins for value and 4 bins for gray level

[8]. The percentages of these quantiﬁed colors represent the elements of clothes feature

vector. To measure the similarity between R

s,f

and R

′

, we employed this function

[5]:

h(R

s,f

, R

′

) =

168

c=1

min(p

s,f

(c), p

′

(c))

min(

168

c=1

s,f

(c),

168

c=1

′

(c))

. (34)

With p

s,f

′

(c)) represents the percentage of color c in R

s,f

(resp. R

′

This similarity function values belong to the interval [0, 1]. The big value of this func-

tion is, the high similarity between clothes.

8686

4.2 Results

The objective of this evaluation is to validate the choice between reinforcing and con-

testing strategies. We used an address book that contains 4 people. 90 photos have been

collected, each photo showing only one known person. 12 images have been used as

model for face recognition tool (F R) such that 3 images for each person. In fact we

will produce the metadata lists for 78 = 90 − 12 faces (all have been detected and,

there is no false positive or false negative of face detection tool F D). All images are

taken in the same day and in the same building (The photographed persons were in dis-

similar clothes). The clothes in the photos model of F R are not taken into consideration.

We produced the metadata for the 78 faces. From this initial collection, we extracted

another one, where we change the colors of clothes (the extracted rectangular region

of upper body). We did that in order to get people with similar clothes (two person

with a red clothes and two others with green clothes); in fact, we have used two narrow

Gaussian distributions for clothes painting, the center of the ﬁrst was ﬁxed on the red

color and the center of the second was ﬁxed on the green color. Now, we have two col-

lections, the ﬁrst one with similar clothing and the other being dissimilar. We use each

collection separately in order to evaluate these strategies: ”without clothing similarity

(images analysis IA only)”, ”using clothing similarity with reinforcing strategy”, and

”using clothing recognition with contesting strategy”. All photo of our corpus belong

to the same event (at the same day and in the same place). So, we believe that for these

two collections, the approaches detailed in [18,3,16] will give the approximate results

to our reinforcing strategy, because they are based only on reinforcing strategy and we

have only one event.

The ROC curves are computed according to metadata scores (pignistic probability).

The ﬁgure 2 shows the ROC related to the ﬁrst collection (people in dissimilar clothes).

We observe that the use of clothing similarity improves the accuracy of face identiﬁca-

tion in personal photos and there isn’t a big difference between the results of the two

strategies reinforcing and contesting. However, the later is better in the beginning (in-

terval [0, 0.05] of false positive rate). The ﬁgure 3 shows the ROC curves related to

the second collection (people in similar clothes). We can say that the use of reinforc-

ing strategy is negatively inﬂuenced and reduce signiﬁcantly the system accuracy. But

contesting strategy resists more to the noise engendered by similar clothes, further it

becomes better from the point ”false positive rate 0.22”. After all, contesting strategy

showed the best results and it is better adapted to photo albums, specially, when we

have people with similar appearance.

5 Conclusions

In this paper, we have proposed an original approach that uses an external information

source for face identiﬁcation process. On one hand, we attempted to take advantage

from the similarity between clothes and the spatiotemporal distances between photos.

On the other hand, we tried to use this external information source correctly in order to

avoid maximum generation and accumulation of errors. We have seen that, the problem

of persons with similar clothes can be reduced without either making clothing similarity

unavailable (hard constraint) or less weighted. In addition, the spatiotemporal data was

8787

Fig.2. ROC curves according to thresholds on metadata score, and using the photo collection that

contains people in dissimilar clothes.

Fig.3. ROC curves according to thresholds on metadata score, and using the photo collection that

contains people in similar clothes.

exploited with a ﬂexible manner and without quantiﬁcation (no events detection). Note

that belief function theory allowed us to deal with our problem efﬁciently thanks to its

capacity of uncertainty modeling and representation. The experimentations showed off

clearly the beneﬁt of our approach, the results can be re-ameliorate with the use of an

advanced technique of clothes segmentation. In future, we will use this approach for

video annotation and present our relevance feedback for parameters updating.

References

1. M. Abdel-Mottaleb and L. Chen. Content-based photo album management using faces’

arrangement. In IEEE International Conference on Multimedia and Expo, ICME04, pages

2071–2074, Taipei, Taiwan, June 2004. IEEE.

2. L. Agosto, M. Plu, P. Bellec, and L. Vignollet. Someone : A cooperative system for personal-

ized information exchange. In International Conference of Enterprise Information Systems,

pages 71–78, Angers, France, 2003.

8888

3. D. Anguelov, K. Lee, S. B. Gokturk, and B. Sumengen. Contextual identity recognition

in personal photo albums. In IEEE Computer Society Conference on Computer Vision and

Pattern Recognation, pages 01–07, June 2007.

4. Y. A. Aslandogan and C. Yu. Multiple evidence combination in image retrieval: Diogenes

searches for people on the web. In Proceedings of the 23rd Annual International ACM/SIGIR

Conference on Research and Development in Information Retrieval, pages 88–95, Athens,

Greece, June 2000.

5. A.W.M.Smeulders, M.Worring, S. Santini, A.Gupta, and R. Jain. Content-based image re-

trieval at the end of the early years. IEEE Transactions on Pattern Analysis and Machine

Intelligence, 22(12):1349–1380, December 2000.

6. M. Davis, M. Smith, F. Stentiford, A. Bambidele, J. Canny, N. Good, S. King, and R. Janaki-

raman. Using context and similarity for face and location identiﬁcation. In Proceedings

of the IS& T/SPIE 18th Annual Symposium on Electronic Imaging Science and Technology

Internet Imaging VII, page cdrom, San Jose, California, 2006. IS& T/SPIE Press.

7. A. Dempster. Upper and lower probabilities induced by a multivalued mapping. Annals of

Mathematical Statistics, AMS 38:325–339, 1967.

8. Y. Gong, C. Chuan, and G. Xiaoyi. Image indexing and retrieval using color histograms.

Multimedia Tools and Applications, 2(2):133–156, March 1996.

9. R. Gossweiler and J. Tyler. PLOG: Easily create digital picture stories through cell phone

cameras. In First International Workshop on Ubiquitous Computing (IWUC 2004), pages

94–103, Porto, Portugal, April 2004. INSTICC Press.

10. S. Kharbouche. Fonctions de croyance et indexation multimodale, application `a

l’identiﬁcation de personnes dans des albums. Master’s thesis, University of Rouen, Rouen,

Normandy, France, December 2006.

11. S. Kharbouche and M.Plu. Combination of face detection, face recognition and gender recog-

nition using belief function using evidence theory. In IEEE International Conference on

Information Reuse and Integration, pages 526–531, Las Vegas, USA, August 2007.

12. M. Naaman, R. Yeh, H. Garcia-Molina, and A. Paepcke. Leveraging context to resolve

identity in photo albums. In Proceedings of the Fifth ACM/IEEE-CS Joint Conference on

Digital Libraries (JCDL 2005), pages 178–187, Denver, Colorado, USA, June 2005.

13. A. Pigeau and M. Gelgon. Incremental statistical geo-temporal structuring of a personal

camera phone image collection. In Proceedings of the 17th International Conference on

Pattern Recognition, ICPR’2004, volume 3, pages 878–881, 2004.

14. G. Shafer. A mathematical theory of evidence. Princeton University Press, 1976.

15. P. Smets. Belief induced by the partial knowledge of the probabilities. In D. Heckerman and

Al., editors, Uncertainty in Artiﬁcial Intelligence, UAI’94, pages 523–530, San Mateo, 1994.

Morgan Kaufmann.

16. Y. Song and T. Leung. Context-aided human recognition - clustering. In 9th European Con-

ference on Computer Vision ECCV 2006, pages 382–395, Graz, Austria, May 2006. Springer,

LNCS.

17. M. Visani, C. Garcia, and C. Laurent. Comparing robustness of two-dimensional PCA and

eigenfaces for face recognition. In Proceedins of the International Conference on Image

Analysis and Recognition, volume 3212/2004, pages 717–724, 2004.

18. L. Zhang, L.Chen, M. Li, and H. Zhang. Automated annotation of human faces in family

albums. In MULTIMEDIA ’03: Proceedings of the eleventh ACM international conference

on Multimedia, pages 355–358, New York, NY, USA, 2003. ACM Press.

8989