From Text Vocabularies to Visual Vocabularies

What Basis?

Jean Martinet

LIFL/CNRS-UMR 8022, Universit´e Lille 1, Lille, France

Keywords:

Bag-of-Features, Quantization, Vocabulary, Zipf’s Law, Evaluation.

Abstract:

The popular ”bag-of-visual-words” approach for representing and searching visual documents consists in de-

scribing images (or video keyframes) using a set of descriptors, that correspond toquantized low-level features.

Most of existing approaches for visual words are inspired from works in text indexing, based on the implicit

assumption that visual words can be handled the same way as text words. More speciﬁcally, these techniques

implicitly rely on the same postulate as in text information retrieval, stating that the words distribution for

a natural language globally follows Zipf’s law – that is to say, words from a natural language appear in a

corpus with a frequency inversely proportional to their rank. However, our study shows that the visual words

distribution depends on the choice of low-level features, and also especially on the choice of the clustering

method. We also show that when the visual words distribution is close to this of text words, the results of an

image retrieval system are increased. To the best of our knowledge, no prior study has yet been carried out

to compare the distributions of text words and visual words, with the objective of establishing the theoretical

foundations of visual vocabularies.

1 INTRODUCTION

In image retrieval as in information retrieval in gen-

eral, the quality of document representation is ob-

viously central for the system efﬁciency. Popular

bag-of-visual-words approaches consist in represent-

ing images with sets of descriptors, that correspond to

quantized low-level features. The features are locally

extracted from image regions, yielding a large num-

ber of feature vectors – which are the equivalent to

text words.

A quantiﬁcation step enables to reduce this num-

ber by gathering vectors in clusters into the feature

space, and associating to each vector a representative

in a discrete set – equivalent to index terms. Images

can therefore be represented as sets, histograms, or

vectors of such terms, which play the role of an in-

dexing vocabulary.

This representation is inspired from the bag-of-

words approach in text indexing and information re-

trieval, where text documents are represented as sets

of terms taken from a vocabulary, that is built using

the corpus.

Since almost a decade, most content-based im-

age retrieval systems are based on this now popular

representation. A number of classical text indexing

techniques such as term selection or term weighting

are directly applied to visual vocabularies, so is the

matching method with set intersection or vector com-

parison.

However, visual and textual vocabularies are gen-

erated with processes of intrinsically different na-

tures, and to our knowledge, no study has yet dealt

with validating the direct application of text tech-

niques to images. The objective of this work is

to compare text and visual vocabularies, in order to

study and clarify, according to some criteria, the ap-

plicability of text techniques to visual vocabularies.

The remainder of this paper is organized as fol-

lows: Section 2 presents generalities about text in-

formation retrieval, and implicit hypotheses on which

traditional techniques in this domain are founded.

Section 3 presents the visual bag-of-words paradigm,

based on analogies between text and image. This sec-

tion also gives details about widely used low-level

features and clustering algorithms that we select in

our work. Section 4 presents the study of visual vo-

cabularies: it describes the vocabularies with the dif-

ferent generation steps, the experimental conditions,

and the results. We give a conclusion to this work and

mention about future research in Section 5.

668

Martinet J..

From Text Vocabularies to Visual Vocabularies - What Basis?.

DOI: 10.5220/0004749606680675

In Proceedings of the 9th International Conference on Computer Vision Theory and Applications (VISAPP-2014), pages 668-675

ISBN: 978-989-758-004-8

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

2 TEXT INFORMATION

RETRIEVAL

Text Information Retrieval (IR) approaches are devel-

oped since about 40 years, and they are based on key-

words found in the text collection(Baeza-Yates and

Ribeiro-Neto, 1999).

The advantage of these approaches is namely due

to their being efﬁcient and fast, as it can be noticed

with Web search engines that are able to retrieve very

quickly documents among hundreds of millions (Brin

and Page, 1998).

2.1 Classical Processing in Text IR

Among existing IR models, the Vector Space model

(Salton, 1971) gives good results from atomic pieces

of knowledge such as keywords widely used for text

indexing. A document content is expressed as a group

of keywords considered representative of its content,

the keywords are taken from an indexing vocabulary.

The indexing vocabulary contains all terms used for

documents description; it is built from words found

in the whole corpus, after possible operations and pro-

cessing described below.

2.1.1 Stop-word Filtering and Anti-dictionaries

Only words conveying semantics are usually kept in

the indexing vocabulary: the other so-called stop-

words (such as ”of”, ”the”, ”for” in english), stored

in an anti-dictionary, are not part of the indexing vo-

cabulary. In the same way, rare words are generally

removed from the indexing vocabulary.

2.1.2 Word Stemming

Besides, it is frequent to stem words, that is to say to

detect word spelling/lexical variations, such as plural

or verb conjugation, and to consider only a common

root, or semantical unit (Salton and McGill, 1983)

as an indexing term. For instance, the words ”writ-

ers”, ”writing”, and ”written” have the common root

”writ”, and only this root would be kept in the vocab-

ulary.

2.1.3 Term Weighting

Document descriptors can all be considered having

equal importance, i.e. being all equally representative

of the document semantics: solely the occurrence (or

absence) of a term in a document matters. However,

the descriptors do generally not have the same impor-

tance in documents. Hence, descriptors are weighted

in an index according to how well they describe the

document, in order to express the relative importances

and to reﬁne the indexing.

The popular TF × IDF weighting scheme (Salton

and Buckley, 1988) and its probabilistic variation

BM25 (Jones et al., 2000) consist in giving more im-

portance to terms that are frequent in given documents

(TF) and also not frequent in the whole corpus (IDF).

2.2 Zipf’s Law and Luhn’s Model

A central aspect of text IR approaches is that they

have as a foundation important characteristics of the

words distribution in a natural language. Zipf’s law

links word frequencies in a language to their ranks,

when ordered in decreasing frequency order (Zipf,

1932). This law stipulates that that words occurrences

follow this distribution model:

(1)

where P

is the occurrence probability of the word at

rank n, and a is a value close to 1. For instance, for the

english language, the term ”the” would be the most

frequent (it generally represents about 7% of the to-

tal word count) in a large text collection. The second

most frequent would be ”of”, with 3,5% of the total,

etc. As an illustration, Figure 1 shows the word dis-

tribution from Wikipedia articles in November 2006

in a log-log (bi-logarithmic) plot, which reveals the

logarithmic distribution.

Figure 1: Word distribution from Wikipedia article in

November 2006 (Source : Wikipdia).

This logarithmic distribution, interpreted through

information theory (Shannon, 1948), has namely

made it possible to establish the theoretical founda-

tions of word ﬁltering and weighting schemes. In par-

ticular, the selection techniques for signiﬁcant terms

are mainly grounded on hypotheses from Luhn’s

model (Luhn, 1958), and this model originates from

Zipf’s law.

FromTextVocabulariestoVisualVocabularies-WhatBasis?

669

This model indicates a relation between the rank

of a word and its resolving power, or discrimina-

tive power, that is to say its capacity of identifying

relevant documents (notion of recall) combined with

its capacity of distinguishing non relevant documents

(notion of precision); this relation is illustrated Figure

2. Hence, less discriminative words are those at low

Figure 2: Zipf’s law and Luhn’s model.

ranks – very frequent, and those at high ranks – very

rare. More discriminative words are those located in-

between, therefore these terms should be selected for

the indexing vocabulary.

3 BAGS OF VISUAL WORDS

In a trend similar to text documents, approaches of

bags of visual words (Sivic and Zisserman, 2003; Ju-

rie and Triggs, 2005; Grauman and Darrell, 2005;

Lazebnik et al., 2006) represent images with sets of

descriptors, corresponding to quantized low-level fea-

tures. Such approaches are founded on the obvious

analogy between text methods and visual methods.

3.1 Low-level Descriptors

Among low-level descriptors, SIFT (Scale-Invariant

Feature Transform) (Lowe, 2004) and SURF

(Speeded-Up Robust Features) (Bay et al., 2008) are

widely used; they represent an image with a set of

scale invariant local features. Both include an interest

point detector, and a descriptor for such points.

3.1.1 SIFT

SIFT detector (Lowe, 2004) uses a Difference of

Gaussians (DoG) ﬁlter to approximate the Laplacian

of Gaussians (LoG), to accelerate processing. This

interest point detector is scale invariant. The descrip-

tor calculates a local histogram of oriented gradient

in the neighborhood of the interest point, and stores

the bins in a 128-dimension vector (8 orientation bins

for each 4 × 4 position bins). A speciﬁcity of SIFT

is that it generates a large number of features, spread-

ing over the image with different scales and locations.

An image with resolution 500 × 500 pixels typically

generates about 2000 stables features.

3.1.2 SURF

The more recent detector-descriptor SURF (Bay et al.,

2008) is based on a Hessian matrix for detecting

interest points, that is estimated from integral im-

ages. The descriptor calculates the distribution of

Haar wavelet responses in the neighborhood of inter-

est points, stored in a 64-dimension vector.

3.2 Vector Quantization

The generation of a visual vocabulary requires to clus-

ter all features extracted from the image collection.

This step of vector quantization consists in represent-

ing each feature by the centroid of the cluster it be-

longs.

3.2.1 Vocabulary Size

In previous works, there has been many different at-

tempts with several vocabulary sizes. For example,

(Sivic and Zisserman, 2003) used a vocabulary size

of range 6000-10000. Later, other researchers used

a wide span of vocabulary sizes, e.g. (Csurka et al.,

2004) who used the range 500-2500, (Nowak et al.,

2006) who tried 300, 1000, and 4000, (Lazebniket al.,

2006) used 200-400, Zhang et al. (Zhang et al., 2007)

used 1000, and Li et al. (Li et al., 2008) in 2008 used

a vocabulary size less than 1500. Ballan et al. (Bal-

lan et al., 2012) have addressed the problem of vo-

cabulary generation in the context of video (i.e. with

space-time dimensions) by using radius-based clus-

tering with soft assignment. They have further per-

formed vocabulary compression in order to avoid a

too large size. There is no consensus which size or

range of vocabulary size performs best. In our study,

we have set the vocabulary size to be 10000.

3.2.2 Clustering with KMeans

KMeans is certainly the most widely used cluster-

ing algorithm. From a set of k centroids (generally

initialized randomly), this iterative partitioning algo-

rithm assigns to each point the closest centroid, and

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

670

re-estimates the center of each group in order to de-

termine new centroids. The algorithm stops when a

convergence criterion is satisﬁed, and the ﬁnal parti-

tioning determines the mapping between points and

centroids.

3.2.3 Clustering with SOM

Much less known are the Self-Organizing Maps

(SOM) (Kohonen, 2001; Xu et al., 2007), a kind

of artiﬁcial neural network, that are used for non-

supervised classiﬁcation. From data in a large dimen-

sion real space, SOM generates a space mapping by

assigning to each point a representative after a learn-

ing step. This partitioning of data enables to gather

points into clusters.

We chose to compare KMeans and SOM because

of their intrinsic difference in both processing the data

and behavior according to the type and size of data

(Abbas, 2008).

3.3 Text Techniques Applied to Images

Once the visual vocabulary is created, images can be

represented with sets, histograms, or vectors of vi-

sual terms. A great amount of work in image rep-

resentation and indexing are inspired from text tech-

niques, such as term ﬁltering (Sivic and Zisserman,

2003; Tirilly et al., 2008), where visual terms with

low interest are removed from the indexing vocabu-

lary. Besides, visual terms can be weighted to repre-

sent their importance. For instance, the system SIM-

PLIcity (Wang et al., 2001) implements a weighting

scheme named RF × IPF (Region Frequency et In-

verse Picture Frequency), directly inspired from the

widely known TF × IDF for text. Weights are given

to image regions according to their frequency in the

image and in the collection. Works in (Sivic and Zis-

serman, 2003; Tirilly et al., 2008; Philbin et al., 2008)

also adapt this weighting scheme to images.

If we go further into the text-image analogy, sev-

eral works (Bosch et al., ; Tirilly et al., 2008; Wu

et al., 2009) have adapted language models to images.

Some works even implement query expansion tech-

niques (Philbin et al., 2008) (originally designed for

text), by automatically adding selected visual terms to

the original query in order to get better search results.

4 VISUAL VOCABULARY STUDY

This section describes the experiments made in order

to compare the considered visual vocabularies.

4.1 Terminology Distinction

We start by introducing the following terminology

distinction. Following the text-image analogy, we call

visual words the image regions or local areas originat-

ing from segmentation, tessellation, or regions in the

neighborhood of interest points.

Therefore, visual words correspond to text words,

that is to say: occurrences of terms, with lexical vari-

ations. Visual terms correspond to text terms, that is

to say: elements of the indexing vocabulary.

As a consequence, a visual word appearing in an

image is an occurrence of a visual term, and the visual

variations of regions assigned to a given cluster can be

seen as the visual counterpart of lexical and syntactic

variations for text.

4.2 Description of Considered

Vocabularies

The two above-described features (SIFT and SURF),

combined with the two clustering algorithms

(KMeans and SOM) enable to generate four visual

vocabularies: KMSIFT, KMSURF, SOMSIFT,

SOMSURF. We aim at comparing these vocabularies

according to their term distributions, and according

to the results obtained in an image search task.

4.3 Image Collections

We selected Caltech-101 and Pascal image collections

for this study. Sample images from these collections

are given Figure 3. Caltech-101 database

contains

101 image categories, with about 50 images per cat-

egory (varying from 40 to 800). We have randomly

selected 10 categories in our experiments, resulting in

1345 images.

Figure 3: Image samples from Caltech-101 (left) and Pascal

(right).

Caltech-101, see URL : http://www.vision.caltech.edu/

Image Datasets/Caltech101/.

FromTextVocabulariestoVisualVocabularies-WhatBasis?

671

Figure 4: Vocabularies distributions for Caltech-101.

Pascal database

contains 20 image categories,

and we selected 1000 images uniformlysampled from

all categories in our experiments.

The average processing time for visual words

extraction is 3h13min with a 2.93GHz Intel Xeon

processor-based machine, with 4GB of RAM. The re-

quired time is slightly longer for SIFT than for SUFT.

We note that SIFT generates about twice as much vi-

sual words than SURF for both databases. This is

because the Difference of Gaussians detector extracts

much more interest points than the Hessian matrix de-

tector.

Table 1 shows the total number of visual words

generated for each database and each descriptor,

along with the average number of words per image.

Table 1: Overview of the generated visual words.

Database Caltech-101 Pascal

(# images) (1345) (1000)

Feature SIFT SURF SIFT SURF

# words in total 463K 204K 297K 174K

# words/image 344 151 297 174

4.4 Visual Words Distributions

Visual vocabularies have been generated for each

database, using alternatively KMeans and SOM for

clustering. As mentioned in Section 3.2, the obtained

vocabularies have a size of 10K terms.

The generation time required to generate the vo-

cabularies are drastically different: 4.5 days for

KMeans, and 2.5h for SOM with the above-described

machine, using a Matlab environment. Note that us-

Visual Object Classes Challenge 2009, see URL :

http://pascallin.ecs.soton.ac.uk/challenges/VOC/voc2009/.

ing compiled C/C + + implementations is likely to

speed up the processing time for both algorithms.

Figures 4 and 5 show the distributions for all vo-

cabularies for both databases, along with the words

distribution from Wikipedia, as shown Figure 1. It

can be easily noticed that for both databases, KM-

SIFT and KMSURF vocabularies have rather ﬂat dis-

tributions, compared with SOMSIFT and SOMSURF.

This indicates that KMeans generates vocabularies

that are more uniformly distributed than SOM. In

other words, it means that the cluster sizes are more

similar to one another with KMeans than with SOM.

However, the distributions of SOM vocabularies

are much similar to the english words distribution

(Wikipedia articles in 2006). This tendency is the

same for both databases.

We now focus on the slope of the vocabularies dis-

tributions in the log-log plot. For this purpose, we

have estimated the slope of the trend line calculated

using linear regression, with the least squares crite-

rion.

The slopes are given Table 2. The results conﬁrm

that SOM-generated vocabularies are more similar

to those of natural language than KMeans-generated

ones. The results also indicate that the choice of

detector-descriptor does not seem to greatly impact

the vocabularies distribution.

Table 2: Estimated slopes (in degrees) for visual distribu-

tions (slope for Wikipedia: -24.2

◦

Database Caltech-101 Pascal

Descriptor SIFT SURF SIFT SURF

KMeans -7.1

◦

-6.3

◦

-8.2

◦

-9.8

◦

SOM -19.0

◦

-21.5

◦

-18.7

◦

-20.1

◦

In order to explain this, it is useful to take a look at

the different clustering processes: KMeans randomly

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

672

Figure 5: Vocabularies distributions for Pascal.

initializes cluster centroids, and a convergencestate is

searched by progressively shifting centroids. A situ-

ation in which two centroids would be both initially

placed inside a given group of points close to one an-

other (thus a good candidate for a single cluster) is

likely to happen.

This way, instead of producing a single cluster re-

ﬂecting the reality of the group of points, KMeans

would artiﬁcially produce two ill-formed clusters, as

illustrated Figure 6. Because the cluster generation

process is fundamentally different with SOM, this al-

gorithm does not encounter this problem.

Figure 6: Illustration of clusterings: correct (left) and not

correct (right) for KMeans, after different initializations.

In the remainder of this section, we focus on

results obtained with the vocabularies in an image

search task.

4.5 Image Search Task

The second step in our work consists in evaluating the

precision that one can reach with each vocabulary in

an image search task. For the purpose of this eval-

uation, we have implemented an image retrieval sys-

tem based on the IR vector space model, integrating

ﬁltering and TF × IDF weighting scheme for visual

terms. For both databases, each image is used in turn

as a query, and the ground truth is made of all images

in the same category as the query. Recall-precision

curves are given Figures 7 and 8, for Caltech-101 and

Pascal, respectively.

We notice from the curves that the search preci-

sion seems to be higher for Caltech-101 than for Pas-

cal. This can be explained by the nature of images

in the database: while Caltech-101 contains focused

images generally showing a single well identiﬁed ob-

ject with a simple background, Pascal contains less

focused images often containing several objects, with

more complex backgrounds.

In Figure 7, we can see noticeable differences in

the precision values: SOMSURF yields a better pre-

cision than other vocabularies. Besides, KMSURF

yields lower results than others. This point conﬁrms

the low inﬂuence of the detector-descriptor. The curve

for SOMSIFT is slightly above KMSIFT, but the dif-

ference is too small to be signiﬁcant.

In Figure 8, the results are similar for all vocabu-

laries; the differences in the curves are much smaller

than those shown Figure 7. Even though the tendency

of better results with SOMSURF and lower results

with KMSIFT can be noticed, the difference is not

signiﬁcant and too small to draw conclusions. How-

ever, we note that SIFT is less sensitive than SURF to

the choice of clustering method.

As a general comment, even though KMeans is

more popular than SOM, the results of our image re-

trieval system with SOM-generated vocabularies are

better than KMeans-generated ones. According to our

study, a theoretical justiﬁcation could be that SOM

generates vocabularies whose distributions are more

similar to those of natural language words. By that

very fact, the implicit hypotheses upon which text

techniques are established (Zipf’s law, Luhn’s model)

are satisﬁed in this case for images, which would a

theoretical validation explaining the better results.

FromTextVocabulariestoVisualVocabularies-WhatBasis?

673

Figure 7: Recall-precision curves (Caltech-101).

Figure 8: Recall-precision curves (Pascal).

5 CONCLUSIONS

Most of text IR techniques are implicitly founded on

the postulate that natural language words globally fol-

low Zipf’s law, and on Luhn’s model. Applying such

techniques to visual vocabularies built from images

requires prior veriﬁcations of these postulates valid-

ity.

In this paper, we have presented a study about

visual vocabularies compared to text vocabularies,

in order to clarify conditions for applying text tech-

niques to images. Our study showed that visual words

distributions highly depend the clustering method –

the inﬂuence of the choice for low-level features is

limited. Indeed, the choice for the visual terms gen-

eration method is important, and determines the vo-

cabulary distribution. We also show that when the

visual words distribution is close to text words dis-

tributions, the results of an image retrieval system are

increased. Therefore, SOM clustering method, that

generates distributions similar to natural languages,

yields better search precision than the yet popular

KMeans. To our knowledge, no prior study has yet

been carried out aiming at validating the application

of text techniques for images.

Our future works will focus on generalizing this

comparative approach to other popular low-level de-

scriptors in image representation. We also wish to get

insights from the consequences of our results when

varyingthe size of the vocabularies. Besides, we work

toward making a more formal comparison of distribu-

tions with dedicated statistical tools. This work has

been done in a global initiative of applying text tech-

niques to images.

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

674

REFERENCES

Abbas, O. A. (2008). Comparisons between data clustering

algorithms. International Arab Journal of Information

Technology (IAJIT), 5(3):320.

Baeza-Yates, R. A. and Ribeiro-Neto, B. A. (1999). Modern

Information Retrieval. Addison-Wesley.

Ballan, L., Bertini, M., Del-Bimbo, A., Seidenari, L.,

and Serra, G. (2012). Effective codebooks for hu-

man action representation and classiﬁcation in uncon-

strained videos. IEEE Transactions on Multimedia,

14(4):1234–1245.

Bay, Ess, A., Tuytelaars, T., and Gool, L. V. (2008). Surf:

Speeded up robust features. Computer Vision and Im-

age Understanding (CVIU), 110(3).

Bosch, A., Zisserman, A., and Muoz, X. Scene classiﬁca-

tion via plsa. In ECCV’06, pages 517–530.

Brin, S. and Page, L. (1998). The anatomy of a large-scale

hypertextual Web search engine. In In journal of Com-

puter Networks and ISDN Systems (30), pages 107–

117, Brisbane.

Csurka, G., Dance, C. R., Fan, L., Willamowski, J., and

Bray, C. (2004). Visual categorization with bags of

keypoints. In In Workshop on Statistical Learning in

Computer Vision, ECCV, pages 1–22.

Grauman, K. and Darrell, T. (2005). The pyramid match

kernel: Discriminative classiﬁcation with sets of im-

age features. In ICCV, pages 1458–1465.

Jones, K. S., Walker, S., and Robertson, S. E. (2000). A

probabilistic model of information retrieval: devel-

opment and comparative experiments. Inf. Process.

Manage., 36:779–808.

Jurie, F. and Triggs, B. (2005). Creating efﬁcient codebooks

for visual recognition. In ICCV, pages 604–610.

Kohonen, T. (2001). Self-Organizing Maps. Springer Series

in Information Sciences, vol 30.

Lazebnik, S., Schmid, C., and Ponce, J. (2006). Beyond

bags of features: Spatial pyramid matching for recog-

nizing natural scene categories. In CVPR (2), pages

2169–2178. IEEE Computer Society.

Li, T., Mei, T., and Kweon, I. S. (2008). Learning optimal

compact codebook for efﬁcient object categorization.

In Proc. of WACV, pages 1–6.

Lowe, D. G. (2004). Distinctive image features from scale-

invariant keypoints. International Journal of Com-

puter Vision, 60(2):91–110.

Luhn, H. P. (1958). The automatic creation of literature

abstracts. IBM Journal of Research Development,

2(2):159165.

Nowak, E., Jurie, F., and Triggs, B. (2006). Sampling strate-

gies for bag-of-features image classiﬁcation. In In

Proc. ECCV, pages 490–503. Springer.

Philbin, J., Chum, O., Isard, M., Sivic, J., and Zisserman,

A. (2008). Lost in quantization: Improving partic-

ular object retrieval in large scale image databases.

In Computer Vision and Pattern Recognition, 2008.

CVPR 2008. IEEE Conference on, pages 1 –8.

Salton, G. (1971). The SMART Retrieval System. Prentice

Hall.

Salton, G. and Buckley, C. (1988). Term-weighting ap-

proaches in automatic text retrieval. In Information

Processing and Management, pages 513–523.

Salton, G. and McGill, M. (1983). Introduction to Modern

Information Retrieval. McGraw-Hill.

Shannon, C. E. (1948). A mathematical theory of commu-

nication. Bell System Technical Journal, 27:379–423,

623–656.

Sivic, J. and Zisserman, A. (2003). Video google: A text

retrieval approach to object matching in videos. In

ICCV, pages 1470–1477.

Tirilly, P., Claveau, V., and Gros, P. (2008). Language mod-

eling for bag-of-visual words image categorization. In

CIVR’08, pages 249–258.

Wang, J., J.L., and Wiederhold, G. (2001). SIMPLIcity:

Semantics-sensitive Integrated Matching for picture

LIbraries. IEEE Transactions on Pattern Analysis and

Machine Intelligence, 23(9):947–963.

Wu, L., Hu, Y., Li, M., Yu, N., and Hua, X.-S. (2009).

Scale-invariant visual language modeling for object

categorization. Multimedia, IEEE Transactions on,

11(2):286 –294.

Xu, X., Zeng, W., and Zhao, Z. (2007). A structural adapt-

ing self-organizing maps neural network. In ISNN (2),

pages 913–920.

Zhang, J., Lazebnik, S., and Schmid, C. (2007). Local fea-

tures and kernels for classiﬁcation of texture and ob-

ject categories: a comprehensive study. International

Journal of Computer Vision, 73:2007.

Zipf, G. K. (1932). Selective Studies and the Principle of

Relative Frequency in Language. Addison- Wesley,

Cambridge, MA, USA.

FromTextVocabulariestoVisualVocabularies-WhatBasis?

675