Vocabulary length experiments for binary image classification using bov approach

Signal & Image Processing : An International Journal (SIPIJ) Vol.4, No.6, December 2013

VOCABULARY LENGTH EXPERIMENTS FOR BINARY
IMAGE CLASSIFICATION USING BOV APPROACH
S.P.Vimal1, Eshaan Puri2 and P.K.Thiruvikiraman3
1,2

Department of Computer Science and Information Systems
Birla Institute of Technology and Science, Pilani, Rajasthan, India
3
Department of Physics, Birla Institute of Technology and Science,
Hyderabad Campus, Andra Pradesh, India

ABSTRACT
Bag-of-Visual-words (BoV) approach to image classification is popular among computer vision scientists.
The visual words come from the visual vocabulary which is constructed using the key points extracted from
the image database. Unlike the natural language, the length of such vocabulary for image classification is
task dependent. The visual words capture the local invariant features of the image. The region of image
over which a visual word is constrained forms the spatial content for the visual word. Spatial pyramid
representation of images is an approach to handle spatial information. In this paper, we study the role of
vocabulary lengths for the levels of a simple two level spatial pyramid to perform binary classifications.
Two binary classification problems namely to detect the presence of persons and cars are studied. Relevant
images from PASCAL dataset are being used for the learning activities involved in this work

KEYWORDS
Image classification, Bag-of-Visual-words, Spatial Pyramid

1. INTRODUCTION
Image classification is a supervised learning task of classifying the image to belong to one of the
several known semantic categories. Object detection is also a supervised learning task of
detecting the presence of one or more objects in the given image and optionally to detect their
possible location and scale in the given image. Identifying the appropriate representation of
object(s) or the image(s), formulation of the learning problem and evaluation of the learned
hypothesis are key activities in any supervised learning problem. The approaches to image
classification/detection vary along these lines.
Bag-of-Visual-words (BoV) [1],[4] representation for images is a successful representation for
wide range of image analysis applications, which is inspired from the Bag-of-Words (BoW) [1],
[4] approach to text mining. The success of BoV approach largely lies in the appropriate
definition of visual vocabulary consisting of visual words extracted from the domain of image
classes / object classes being under consideration. The visual vocabulary is built from the key
point descriptors extracted from images by standard clustering techniques by explicitly choosing
the number of visual words required for the vocabulary. Once the vocabulary of visual words is
constructed, each image is then represented by a distribution of visual words present in the image.
The number of words in the visual vocabulary and the nature of description (color, intensity,
gradient information, texture etc.) available with each word in the vocabulary are task dependent
and they are generally chosen empirically. Preserving the spatial information, i.e. the spatial
region where each visual word in the visual vocabulary is typically found, is important for better
performance of the classification/ detection task. There are different approaches in the literature
addressing this aspect of representation. Once the representation of visual vocabularies and the
DOI : 10.5121/sipij.2013.4608

95


distribution of visual words in each image are arrived at, the learning and classification can
proceed as if it is for text documents.
In this paper, we study histogram tiling technique used for building the visual vocabulary for two
simple binary classification problems. The binary classification problems we considered are (i) to
detect if the image consists of person(s) (ii) to detect if the image contains a car. The visual
vocabulary for detecting the presence of a car(s) or person(s) in an image consists of visual words
from positive and negative examples of all car(s) or person(s) from VOC 2007. We construct
four types of histogram tiling inspired from [1] [3] [4]. We considered the maximum of two levels
of the spatial pyramid in which the second level represents the image tessellated into four quarters
and reported the relative importance of these levels. The relative importance of these
levels(scales) is studied with the help of the required vocabulary length for better classification.
The rest of the paper is organized as follows: Section-II briefly reviews the related works,
Section-III outlines the Bag-Of-Visual-words (BoV) Approach implemented in this paper,
Section-IV explains the histogram tiling and the variations considered in this paper, Section-V
explains the experiments conducted along with the explanation on the dataset, Section-VI
Discusses the results obtained and the brief discussions and the Section-VII concludes the paper.

2. RELATED WORKS
The analogy of text retrieval problem to the content based image retrieval was discussed in [5]
which motivated a lot of fruitful research in this direction. Sivic and Zisserman [6] extended the
Bag-of-Words (BoW) approach to text retrieval for the computer vision problem of matching
objects in videos. Their work provided systematic formulation of textual retrieval problem for
retrieving objects from videos. The visual words forming the visual vocabulary are constructed
using descriptors like Histogram of Oriented Gradients [7] , filter banks [8],local image patches
[9] and SIFT [10] features. Lowe’s SIFT [10] is among the most commonly used local
descriptor. The local features extracted are vector-quantized to form a visual vocabulary. The
BoV obtained by quantizing local features and representing image documents by the histograms
introduces sparcity and sacrifices the locality of local features [6].
The spatial information which is lost in the traditional BoV is vital for scaling the
classification/detection process for larger number of categories and Lazebnik et. al [1] proposed a
spatial pyramid based approach to address this. The spatial pyramid proposed by Lazebnik et. al.
constructs the pyramid by repeatedly subdividing the image to obtain the successive levels in the
pyramid instead of gaussian subsampling approach. The concatenated histogram at each level is
used for learning purpose. Also, Lazebnik et. al [1] proposed a matching scheme for such
concatenated histograms. Similar spatial pyramids have been used for the problem of indooroutdoor classification in [11] along with local features like color, textures and frequency
information. Bosch et. al [2] used a similar spatial pyramid scheme to represent the shape of
object categories by extending Histogram of Oriented Gradients [7] and reported considerable
improvements in performance when used along with appearance feature [12]. Histogram tiling
approach to address the loss of spatial information can be considered as a special case of spatial
pyramid approach [1] [2] in which we consider only scale by partitioning an image in a regular
fashion and the histograms for each partition are concatenated. Viitaniemi et al. [13]
experimented various partitioning techniques for spatial tiling along with the hard and soft tiling
approaches and hard and soft histograming approach. They confirm the usefulness of taking the
spatial distribution into account. The encoding of spatial information by means of SPM [1]
encodes each region on every scale of the pyramid from the vocabulary generated from the entire
image. The sky often falls in the top region of the image and similar spatial arrangements can be
seen from various semantic categories. Zhang et al. [3][4] proposed to encode the regions
individually within the spatial constraints. This requires learning the vocabularies for regions
96


independently, in an attempt to better capture the inherent discriminative features of the
individual regions concerned.
The role of various aspects of visual vocabulary building are studied in the literature [4][13][14].
In this paper, we study the relative importance of the vocabulary at level-0 to the vocabulary at
level-1 of the spatial pyramid representation for both SPM [1] and SPC [3] based BoV approach
for a binary classification problem. Our results reveal that the shorter vocabulary is needed for
level-0 relative to level-1 in the spatial pyramid. This implies the level-0 of the spatial pyramid
serves as the context for level-1.

3. BAG OF VISUAL WORDS APPROACH
There are many different approaches to the construction of visual vocabulary from the given
image database. We follow a simple approach to construct the visual vocabulary in which we
quantize the SIFT descriptors into visual words. We use Lowe’s SIFT descriptors [10] for this
purpose which are densely extracted from the image. Each SIFT descriptor is a 128 dimensional
vector. For PASCAL dataset, 1416 key points are typically extracted per image on an average
[14]. We use a random subset of key points extracted from all the images and then perform kmeans clustering for constructing varied lengths of vocabulary. The length of visual vocabulary is
the number of visual words present in the vocabulary. The length of visual vocabulary impacts
the learning performance. The vocabulary of length r leads to the r-dimensional histogram
representation for each image in the database representing the frequency of visual words in the
corresponding image.

4. SPATIAL INFORMATION WITH BOV APPROACH
Spatial extensions to the classic BoV approach yields better classification performance. There are
various approaches to incorporate spatial information in the feature vector for images [1] [2] [13]
[15] in the literature. Histogram tiling refers to forming the histogram representation for the
whole image from the histograms computed on parts of the image. Image is typically partitioned
in a specific spatial arrangement [13] and the histograms are computed for individual partitions.
Such histograms are concatenated to form the feature vector for the image. Histogram tiling can
also be obtained by applying various tiling masks on the image and combining the responses from
each such mask. The masks can be overlapping in which the histogram and the tiling process are
either hard or soft [13].
The histogram tiling approach employs various ways of partitioning the image and there is no
single partitioning method which applies to all the learning problems. For natural scenes, in
which the sky, water and the sand appear together, the partition 3x1 is more appropriate.
Partitioning the image in 2kx2k where k = 0, 1, 2 etc. is more common in the literature. To obtain
the best results in the classification, the histograms obtained in many resolutions are combined in
systematic way, i.e. combining the histograms obtained for k=0, 1, 2 etc. This approach to
incorporate the spatial information, results in the spatial pyramid [1] [2]. The vocabulary is
learned for the whole image and then the distribution of visual words for each partition is
computed systematically. Zhang et al. [3][4] computed vocabularies for individual partitions and
use them to find the distribution of visual words in the corresponding partitions. In this paper, we
consider two levels of resolutions namely level-0 and level-1 which is shown in the Figure. 1. The
level-0 considers the entire image and level-1 is a partitions level-0 into 2x2.

97


Figure. 1. Spatial pyramid with two levels. The image on the left is level-0 and the partitioned image on
the right is level-1.

We consider four cases for our comparison namely case-1, case-2, case-3 and case-4. The case-1
and case-3 considers coding only level-1 and these cases does not code level-0. For case-1, we
learned the visual vocabulary for the entire image and used this vocabulary to code the all four
partitions of level-1. This case reflects a simple histogram tiling. For case-3, we learned the
separate visual vocabularies from all four partitions of level-1 and used them to code the
respective partitions. Case-2 and case-4 involves both levels of the pyramid. The case-2 learns a
global vocabulary and we use it to code all the partitions of level-0 and level-1. Individual
vocabularies are learned for all partitions and the partitions are coded with the respective
vocabulary for case-4. The cases 1 and 2 are inspired from works on histogram tiling and SPM
[1][2] and the cases 3 and 4 are inspired from SPC [3] [4].

5. EXPERIMENTS AND RESULTS
We used the images from PASCAL-2007 [16] dataset for the classes, person and car. For training
purpose, we ensured the positive and negative are equal in numbers and we used the validation
images provided in the dataset for our testing purpose. The car class has 713 positive training
examples, and we picked 713 negative training examples which make 1426 images for training
the car class. The testing images for car class consisting of 1442 images out of which 721 images
are positive and 721 images are negative. Similarly, the person class has 4016 images for training
out of which 2008 are positive training examples and 2008 are negative training examples. The
testing images in person class are 4014 in which 2007 are positive testing images and 2007 are
negative testing images.
We experimented with the vocabulary length for all the four cases. We varied the vocabulary
length and computed the mean average precision for each case. For case-2 and case-4, we varied
the vocabulary length of level-1 for a range of vocabulary lengths of level-0. For case-2, fixing
the vocabulary length for level-0 and varying it for level-1 implies learning two vocabularies of
different lengths from the entire image and code two levels separately. For cases 1 and 3, the
length of the feature vector for each image is the length of the vocabulary multiplied by 4. For
cases 2 and 4, the length of the feature vector is the sum of vocabulary length for level-0 and the
length of feature vector for level-1. For cases 1 and 3, we varied the vocabulary lengths from 25
to 250, thereby varying the length of the feature vector from 100 to 1000. For cases 2 and 4, we
varied the vocabulary length for level-1 from 25 to 250 for the vocabulary length of level-0
varying from 50 to 250. For example, when the level-0 vocabulary length is 50 and level-1
vocabulary length is 25, we have the shortest feature vector of length 50 + (4 x 25) = 150 for
case-2 and case-4.
The classification is learned using SVM classifier with a linear kernel. Figure. 2 and Figure. 3
show the performance for detecting person and car respectively for various vocabulary lengths for
cases 1 and 3. For the problem of detecting persons, the performance of case-3 dominates the
case-1 for the entire range of vocabulary lengths. The mean average precision increases as the
vocabulary length increase from 25 for both case-1 and case-3 here. The performance gain is not
98


significant for case-1 beyond the vocabulary length 100. The performance for case-3 is
significantly higher for the vocabulary length of 150 than for case-1. This huge difference could
be due to the fact that case-3 captures the regional features better in the presence of a person in an
environment. The environment in a human being is present hardly contains any visual words
representing human face in the bottom partitions and similarly legs cannot be expected in the top
two partitions.

Figure. 2. The Mean Average Precision for various vocabulary lengths for detecting person for case-1 and
case-3

Figure. 3. The Mean Average Precision for various vocabulary lengths for detecting car case-1 and case-3

Figure. 3 indicates that the performance of case-1 and case-3 are almost similar. In contrast to
detecting persons, neither case-1 nor case-3 dominates in the range of vocabulary lengths tested
for detecting the presence of car. This implies that the vocabularies learned for the individual
regions of level-1 in case-3 were not discriminative as compared to a single vocabulary learned
from all the images of the training set in case-1. A relatively shorter vocabulary length of around
50 for case-2 has the highest performance on detecting the presence of a car. The performance on
case-3 is higher for vocabulary length of 75. In either case, we observed better performances in
relatively smaller vocabulary lengths. We also observed that the performance gets lower as we
increase the vocabulary lengths. We feel that this phenomenon is due to the fact that, the larger
vocabulary need not necessarily increase the detection performances for all the classes, especially
where the features are more uniform in nature. That is, the features in an environment in which a
car can be found is mostly constructed environment like road, near buildings.
99


Figure. 4. The Mean Average Precision for various vocabulary lengths of level-1 fixing the vocabulary
length of level-0 for detecting person for Case-2 (best viewed in color)

length of level-0 for detecting person for Case-4 (best viewed in color)

Figure. 4 shows the classification performances obtained for detecting the presence of persons in
an image with one vocabulary per level (i.e. case-2). Figure. 5 shows the classification
performances obtained for detecting the presence of persons in an image with one vocabulary per
region (i.e. case-4). In both cases, the classification performance peaks when the vocabulary
length of level-0 is 50 and the level-1 vocabulary length is 200. This implies that when the
vocabulary length of level-1 is sufficiently large, the vocabulary for level-0 acts as a context for
the words in level-1. Also for case-2, the classification performance for level-0 vocabulary length
of 200 (blue continuous line) , dominates others when the level-1 vocabulary lengths are 100 and
200.
The plots for case-4 clearly depict the relationship between the vocabulary lengths of level-0 and
1. Here, the curves for smaller vocabulary lengths of level-0 dominate the later region of the plot
and the longer vocabulary lengths of level 0 dominate the initial region. This clearly indicates
that the vocabulary for level-0 compensates for the loss of global context due to the locally
learned vocabularies for level-1. As we keep increasing the length of the vocabulary at level-1,
the required length of level-0 vocabulary comes down where the vocabularies are learned for
individual regions. However, the classification performance is better for relatively shorter length
of vocabulary at level-0. These observations on case-2 and case-4 are made for detecting the
presence of persons in a given image. We can also see that for the problem of detecting persons,
case-4 yields better performance comparing to case-2 with shorter overall feature vector length.
100


length of level-0 for detecting cars for Case-2 (best viewed in color)

length of level-0 for detecting cars for Case-4 (best viewed in color)

Figure. 6 and 7 are the classification performances obtained for detecting the presence of cars in
an image with one vocabulary for levels (case-2) and one vocabulary for regions (case-4)
respectively. In contrast to the person detection problem, learning vocabularies for levels
marginally performs better than learning vocabularies from regions for detecting the presence of
cars in the image. In case-4 for detecting cars, the peak classification performances are attained
for level-0 vocabulary length of 200 and level-1 vocabulary length of 150. Apart from this, for
case 2 for detecting cars, the classification performances are higher and better than case-4 when
the level-0 vocabulary length is 50 and 100 which are relatively shorter. We believe that the
benefit of learning more spatially discriminating features by the way of learning vocabularies for
regions separately is absent for detecting cars because the distribution of visual words in the
images containing cars is more uniform across the image.

6. CONCLUSIONS AND FUTURE WORK
We performed experiments to understand the role of vocabulary lengths of two levels and their
impact on the final classification performance. The experiments for both classification problems
suggest that an additional level to simple tiling improves the classification performances. When
the classification problem concerned requires learning spatially discriminating features, the
relatively shorter vocabulary length for level-0 is found to be more appropriate. For such cases,
we believe that the level-0 vocabulary provides the context in which the visual words of level-1
could be found. Also, the choice between learning vocabulary for levels and for regions is
category sensitive. We need to extend similar experiments over more classes and more levels in
101


the spatial pyramid to understand better the relationships of the vocabulary lengths of successive
levels to the level of information it provides to the classification. This relationship is category
sensitive as we uncovered with our experiments. Better understanding of this phenomenon should
help to optimize the lengths of feature vector and optimize the learning efforts.

REFERENCES
[1] S. Lazebnik, C. Schmid, and J. Ponce. Beyond bags of features: Spatial pyramid matching for
recognizing natural scene categories. In Proc. of IEEE CVPR, volume 2, pages 2169–2178, 2006.
[2] Bosch, Anna, Andrew Zisserman, and Xavier Munoz. "Representing shape with a spatial pyramid
kernel." Proceedings of the 6th ACM international conference on Image and video retrieval. ACM,
2007.
[3] Zhang, Chunjie, et al. "Image classification using spatial pyramid coding and visual word
reweighting." Computer Vision–ACCV 2010. Springer Berlin Heidelberg, 2011. 239-249.
[4] Yang, Jun, et al. "Evaluating bag-of-visual-words representations in scene classification." Proceedings
of the international workshop on Workshop on multimedia information retrieval. ACM, 2007.
[5] Squire, David McG, et al. "Content-based query of image databases: inspirations from text retrieval."
Pattern Recognition Letters 21.13 (2000): 1193-1198.
[6] Sivic, Josef, and Andrew Zisserman. "Video Google: A text retrieval approach to object matching in
videos." Computer Vision, 2003. Proceedings. Ninth IEEE International Conference on. IEEE, 2003.
[7] Dalal, Navneet, and Bill Triggs. "Histograms of oriented gradients for human detection." Computer
Vision and Pattern Recognition, 2005. CVPR 2005. IEEE Computer Society Conference on. Vol. 1.
IEEE, 2005.
[8] Fei-Fei, Li, and Pietro Perona. "A bayesian hierarchical model for learning natural scene categories."
Computer Vision and Pattern Recognition, 2005. CVPR 2005. IEEE Computer Society Conference on.
Vol. 2. IEEE, 2005.
[9] Varma, Manik, and Andrew Zisserman. "Texture classification: Are filter banks necessary?."
Computer vision and pattern recognition, 2003. Proceedings. 2003 IEEE computer society conference
on. Vol. 2. IEEE, 2003.
[10] Lowe, David G. "Object recognition from local scale-invariant features." Computer vision, 1999. The
proceedings of the seventh IEEE international conference on. Vol. 2. Ieee, 1999.
[11] Szummer, Martin, and Rosalind W. Picard. "Indoor-outdoor image classification." Content-Based
Access of Image and Video Database, 1998. Proceedings., 1998 IEEE International Workshop on.
IEEE, 1998.
[12] Bosch, Anna, Andrew Zisserman, and Xavier Munoz. "Scene classification via pLSA." Computer
Vision–ECCV 2006. Springer Berlin Heidelberg, 2006. 517-530.
[13] Viitaniemi, Ville, and Jorma Laaksonen. "Spatial extensions to bag of visual words." Proceedings of
the ACM International Conference on Image and Video Retrieval. ACM, 2009.
[14] Yang, Jun, et al. "Evaluating bag-of-visual-words representations in scene classification." Proceedings
of the international workshop on Workshop on multimedia information retrieval. ACM, 2007.
[15] Zhang, Chunjie, et al. "Image Classification Using Spatial Pyramid Robust Sparse Coding." Pattern
Recognition Letters (2013).
[16] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual
Object Classes Challenge 2007 (VOC2007) Results

102

Vocabulary length experiments for binary image classification using bov approach

More Related Content

What's hot (17)

Viewers also liked (19)

Similar to Vocabulary length experiments for binary image classification using bov approach (20)

Recently uploaded (20)

Vocabulary length experiments for binary image classification using bov approach