See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/312037041

A review and an approach for object detection in images

Article  in  International Journal of Computational Vision and Robotics · January 2017

DOI: 10.1504/IJCVR.2017.081234

CITATIONS READS

18 11,744

2 authors:

Kartik Umesh Sharma Nileshsingh V. Thakur

Prof Ram Meghe College of Engineering and Management Prof Ram Meghe College of Engineering and Management
9 PUBLICATIONS   26 CITATIONS    85 PUBLICATIONS   357 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Wireless Sensor Networks View project

All content following this page was uploaded by Kartik Umesh Sharma on 14 May 2018.

The user has requested enhancement of the downloaded file.

196 Int. J. Computational Vision and Robotics, Vol. 7, Nos. 1/2, 2017

A review and an approach for object detection

in images

Kartik Umesh Sharma* and

Nileshsingh V. Thakur
Department of PG Studies (Computer Science and Engineering),
Prof. Ram Meghe College of Engineering and Management,
New Express Highway Road,
Badnera-Amravati, Maharashtra, PIN-444701, India
Email: [email protected]
Email: [email protected]
*Corresponding author

Abstract: An object detection system finds objects of the real world present
either in a digital image or a video, where the object can belong to any class of
objects namely humans, cars, etc. In order to detect an object in an image or a
video the system needs to have a few components in order to complete the
task of detecting an object, they are a model database, a feature detector, a
hypothesiser and a hypothesiser verifier. This paper presents a review of
the various techniques that are used to detect an object, localise an object,
categorise an object, extract features, appearance information, and many more,
in images and videos. The comments are drawn based on the studied literature
and key issues are also identified relevant to the object detection. Information
about the source codes and online datasets is provided to facilitate the new
researcher in object detection area. An idea about the possible solution for the
multi class object detection is also presented. This paper is suitable for the
researchers who are the beginners in this domain.

Keywords: object detection; localisation; categorisation; object recognition.

Reference to this paper should be made as follows: Sharma, K.U. and

Thakur, N.V. (2017) ‘A review and an approach for object detection in
images’, Int. J. Computational Vision and Robotics, Vol. 7, Nos. 1/2,
pp.196–237.

Biographical notes: Kartik Umesh Sharma received his Bachelor of

Engineering degree in Information Technology from Sipna College of
Engineering and Technology, Amravati, India in 2012. He received his
Master of Engineering degree in Computer Science and Engineering from
Prof. Ram Meghe College of Engineering and Management, Badnera, India
under Sant Gadge Baba Amravati University, Amravati, India in 2014. His
research interest includes image processing, video processing, language
processing and algorithms. Currently, he is an Assistant Professor in
Department of Computer Science and Engineering at Prof. Ram Meghe College
of Engineering and Management, Badnera, India.

Copyright © 2017 Inderscience Enterprises Ltd.

 A review and an approach for object detection in images 197

Nileshsingh V. Thakur received his BE (CSE) and ME (CSE) degrees from

Government COE, Amravati and COE, Badnera under SGBAU in 1992 and
2005 respectively. He received his PhD in CSE from VNIT, Nagpur, India on
2010. His research interest includes image processing and advanced computing.
He is having over 23 years of teaching and research experience. Presently, he
is a Professor and Dean at PRMCEAM, Badnera, India. He is the author or
co-author of more than 60 scientific publications in international journals and
conferences. He is an editorial board member of eight international journals and
worked as reviewer for international journals and conferences.

1 Introduction

Object detection (OD) system finds objects in the real world by making use of the object
models which is known a priori. This task is comparatively difficult to perform for the
machines as compared to Humans who perform OD very effortlessly and instantaneously.
In this paper we will give a review of the various techniques and approaches that are used
to detect objects in images and videos.
Basically an OD system can be described easily by seeing Figure 1 which shows the
basic stages that are involved in the process of OD. The basic input to the OD system can
be an image or a scene in case of videos. The basic aim of this system is to detect objects
that are present in the image or scene or simply in other words the system needs to
categorise the various objects into respective object classes.

Figure 1 Basic OD model

The OD problem can be defined as a labelling problem based on models of known

objects. Given an image containing one or more objects of interest and a set of labels
corresponding to a set of models known to the system, the system is expected to assign
correct labels to regions in the image. The OD problem cannot be solved until the image
is segmented and without at least a partial detection, segmentation process cannot be
applied. The term detection has been used to refer to many different visual abilities
including identification, categorisation and discrimination.
198 K.U. Sharma and N.V. Thakur

Figure 2 Working of the OD system

2 General methodology

The OD system basically comprises of two main phases namely: the learning phase and
the testing phase which are shown in Figure 2 that shows the normal working of the OD
system. Learning phase is mainly meant for the classifier so that it recognises the objects
present in the image that is given as input to the system. Learning phase can be further
classified as learning through training and learning through validation. Learning through
training comprises mainly of the learning block where a proper learning scheme is
defined, it can be part-based or patch-based, etc. The object template block then makes
use of the learning’s that were done previously to represent the objects with various
representations like histogram representation, random forest representation, etc. Whereas
on the other hand, learning through validation block does not require any sort of training
as they are validated beforehand. Hence after preprocessing the image, directly template
matching is done which produces the features of an object in the image. The main
purpose of the testing phase is to decide whether an object is present in the image that is
given to the system as input and if yes then to which object class does it belongs to. Here
the image is searched for an object by various searching techniques like the sliding
window technique, and according to the output of the searching mechanism, a decision is
made on the object class.
 A review and an approach for object detection in images 199

3 Classification of OD mechanisms

This section classifies the various OD mechanisms based on search, feature

classification, template creation and based on matching. We have classified the OD types
as sliding window-based, contour-based, graph-based, fuzzy-based, context-based and
some other types. Here we will review the work carried out by various authors in the field
of OD.

3.1 Sliding window-based OD

Sliding window OD has received remarkable attention as it is considered as a
very basic method of detecting objects in an image or video. The sliding window
technique basically works by searching through the whole image or scene in order to find
the object that is of interest. This is the reason why it failed to meet the criteria of real
time applications due to higher execution times and inaccurate localisation. Localisation
accuracy is important especially while the OD process is to be followed by object
recognition.
Bergboer et al. (2007) have studied the various learning methods which are use for
realising context-based OD in paintings, namely the gradient method and the context
detection method. The gradient method is used to transform a spatial context into a
gradient towards an object, whereas the context detection method makes use of the
sliding window approach to search the image regions that are likely to contain the object
of interest. Basically the gradient method works totally on assumptions which
may lead to higher timing constraints when there is only single object that is to be
detected in an image. On the other hand, the context detection works based on sliding
window which again introduces the timing constraint as it searches each window for the
presence of an object. Clearly the issue of inaccurate localisation has been a concern
while using the sliding window-based OD technique; Segvic et al. (2011) have
explained how localisation accuracy could be achieved by removing the need
for spatial clustering of the nearby detection responses. This leads to three main goals
namely high recall, high precision and accurate localisation. Spatial clustering
could be used to suppress the number of false positives but at the price of localisation
uncertainty.
Sliding window technique initially fixes the size of the window in which it will be
searching for the object, but in order to increase the rate at which the detections must
happen, Comaschi et al. (2013) have proposed a sliding window approach that decides on
the step size of the window at run time, which helps to apply this technique of sliding
window to real time applications. They have also demonstrated that how this technique
improves the performance of Viola Jones OD, and also claimed to have achieved a
speedup of 2.03x in frames per second without compromising the accuracy factor. The
main issue being the space utilised. Divvala (2012) has studied the two factors that
influence the performance of sliding window technique for OD, namely context and
subcategories. The use of the first factor that is context shows how the performance of the
sliding window approach could be improved. As the siding window approach searches
the total image for the presence of an object in it, while the use of context can be made in
order to know whether a particular object is present in that region or not. The
subcategories factor is where the information within the sliding window is used to split
200 K.U. Sharma and N.V. Thakur

the training data into smaller groups which have reduced appearance diversity
which leads to simpler classification. Here he has discussed only about the two factors,
there could be many other factors too such as contour which could affect the
performance to a great deal. Subburaman et al. (2010) have presented a technique which
is used to reduce the number of miss detections while increasing the grid spacing while
the sliding window approach is used for OD. They have achieved it by using a patch to
predict the bounding box of an object within the search area, and in order to
improve the speed of estimating the bounding boxes the authors make use of the decision
tree with simple binary test at each node. Although they claim that their proposed system
works on a wide variety of images, still we feel that an occluded image could remain a
challenge.
Gualdi et al. (2011) have presented statistical-based search for sliding window
technique which makes use of Monte Carlo sampling for estimating the likelihood
density functions with Gaussian kernels, which is a multi stage strategy where the
proposal distribution is progressively refined by taking into account the feedback of the
classifier. Their proposed method exploits the presence of a basin of attraction around
true positives to drive an efficient exploration of the state space, using a multi-stage
sampling-based strategy. We feel that the concept of multi-stage particle windows can be
extended for the training phase as well. An object is salient if it differs from its
surroundings or if the object contains rare or outstanding details. Yanulevskaya et al.
(2013) have proposed an approach to detect salient object. Here proto-objects are
considered as the units of analysis, where a proto-object is a connected image region
which can be converted to an object or an object part. The process proceeds by
segmenting a complex image into proto-objects and then finding the saliency of each
proto-object, the most salient proto-object is considered as being the salient object.
Sudowe and Leibe (2011) have investigated how geometric constraints can be used for
efficient sliding window OD; here they derived a general algorithm for incorporating
ground plane constraints directly into the detector computation. They claimed that their
approach allows to effortlessly combining multiple different detectors which
automatically compute the region of interest (ROI). As we have seen earlier in Lampert
et al. (2008) that object localisation is an important issue that needs to be seen into, but
most OD systems rely on binary classification for locating the objects but not giving out
the information about the location of the object. Lampert et al. (2008) have tried to
achieve the same goal as that of the sliding window approach that is object localisation
and retrieving the localised object by proposing a simple yet powerful branch and bound
scheme. Table 1 gives an overview of the various approaches for detecting objects using
the sliding window approach. The parameters like, the dataset used, the concept that has
been made use of, the performance evaluation parameter, issues addressed, author’s
remarks and our findings are summarised in the Tables 1(a) and 1(b).
 A review and an approach for object detection in images 201

Table 1(a) Summary of sliding window-based OD work

Performance
Dataset Concept Issues Authors Our
Authors evaluation
used used addressed remark findings
parameter
Bergboer Self-created Gradient and Accuracy Time Fast object Accuracy is
et al. context and speed complexity detection is achieved at
methods achieved. the cost of
speed.
Segvic Self-created Binary Accuracy Localisation Avoids spatial Localisation
et al. classification clustering. uncertainty
exits.
Comaschi CMU/MIT Adaptive Speed Space High speed High space
et al. face dataset sliding utilisation when utilised.
window compared to
OpenCV.
Divvala PASCAL Context Accuracy Performance Subcategories Factors like
VOC modelling gains and used for contour,
dataset computational computational shape, etc.,
tractability tractability. not
considered.

Table 1(b) Summary of sliding window-based OD work

Performance
Dataset Concept Issues Authors Our
Authors evaluation
used used addressed remark findings
parameter
Subburaman CMU/MIT Use of Speed Reduction of This method Occluded
et al. face patch for miss works on a images
dataset prediction detections variety of could be an
images. issue.
Gualdi et al. INRIA Multi Detection Time This method MS-PW
dataset stage rate complexity increases the could be
particle localisation extended to
window accuracy training
also. phase.
Yanulevskaya Standard Salient Accuracy Saliency of Combination Highly time
et al. dataset object an object of two types consuming.
detection of saliency
improves
performance.
Sudowe et al. INRIA Geometric Speed Computation Their Still
dataset constraints of region of approach detection
interest allows rates are
combination slow as
of various whole
detectors. image is
searched.
Lampert et al. UIUC/ Sub- Robustness Object Global Different
PASCAL window and speed localisation optimality is other shapes
VOC 2006 search retained. like circles
datasets eclipses not
considered.
202 K.U. Sharma and N.V. Thakur

3.2 Contour-based OD
Stiene et al. (2006) have proposed an OD approach which is based on range images, as
range images are well suited to contour extraction. They have used a 3D laser scanner
and reliable contour extraction with floor interpretation for the process of OD. Although
this process has high performance while working on range images, converting the natural
images into range images is the overhead process that needs to be performed every time
an object is to be detected. Contour-based OD can be well formulated as a matching
problem between model contour parts and image edge fragments, and hence Yang et al.
(2012) have used this problem and have treated it as a problem of finding dominant sets
in weighted graphs, where the nodes of the graph are pairs composed of contour parts and
edge fragments and the weights between nodes are based on shape similarity. The main
advantage of this system is that it can detect multiple objects present in an image in one
pass. Still the question arises that can this system detect objects in an occluded image or
other types of images. Basically the objects in an image can be characterised on the basis
of their appearance and by the shape of their contours; Schlecht and Ommer (2005)
have investigated a local representation of contours for OD that complements
appearance-based information. They have combined contour and appearance information
into a general voting-based detection algorithm and have claimed that the combination
has significantly improved the performance compared to the other voting methods. The
combination of contour representation and appearance descriptor increases an additional
step in the OD process which increases the time to detect and localise an object. In order
to justify that the selected contour is exactly what the image was being searched for,
Zhu et al. (2008) have introduced a shape detection framework and have named it as
contour context selection. As the shape-based detection is invariant to changes of object
appearance, their approach makes use of salient contours as tokens for matching the
shape. They have termed the task of matching the contours as a set-to-set contour
matching problem and have claimed that their approach takes linear time to compare the
contours. Although they have claimed that their approach is able to detect objects in
cluttered images, they have not spoken about an occluded image which would give a hard
time to their approach to detect objects.
Arbeláez (2006) has presented an approach to boundary extraction which relies on the
problem formulation which occurs in the framework of hierarchical classification that
allows for region-based segmentation and edge detection as a single task. He has defined
generic ultra metric distances by integrating local contour cues along the region
boundaries and combining this information with region attributes. Although this approach
extracts boundaries of objects but still the problem of localisation remains an issue which
the authors have not addressed. Segmentation is the basic process that is involved while
detecting an object in an image; Amine and Farida (2012) have proposed an approach
which makes use of a deformable model ‘Snake’ which they have termed as an active
contour for segmenting the range images. The process is again restricted to range images;
the question still lies about the various types of images. Ferrari et al. (2008) have
presented a family of scale invariant local shape features formed by chains of k
connected, roughly straight contour segments (kAS), and their use for OD. Here they
have demonstrated kAS within a sliding window object detector, where windows are
subdivided into tiles, each described by a bag of kAS. The authors have limited the use of
kAS to a simple detection framework. Authors have not specified the range of k, whether
k could be varied greatly or not.
 Performance evaluation
Authors Dataset used Concept used Issues addressed Authors remark Our findings
parameter
Stiene et al. Various Insensitivity Speed (reduction in Feature extraction Performance Converting natural
datasets of range data number of false negatives) comparable to image to range data
used MPEG-7 standard. may consume time.
Yang et al. ETHZ shape Dominant set Scale and average Mapping for Multiple objects at Occluded or
dataset computation detection rates features extraction multiple scales can cluttered images
be detected in one may affect
pass. performance.
Schlecht and ETHZ shape Appearance Speed (false negatives) Characterisation of This approach can Combination of
Ommer dataset and shape of an image also find most contour
an object relevant contours representation and
and junctions in appearance
object hypothesis. descriptor increases
the time to detect.
Table 2(a) Summary of contour-based OD work

Zhu et al. ETHZ and Shape of an Precision Set to set matching Objects in cluttered Authors have not
INRIA object problem. images can be spoken about how
datasets detected easily. their approach
works on occluded
images.
Arbeláez BSDB Ultra metric Area and quadratic error Boundary Higher level Localisation
dataset distances extraction approaches can problem may exist.
benefit from this
A review and an approach for object detection in images

low-level
representation.
Amine and SASB Homogeneity Accuracy and speed Image Multi agent systems The process is
Farida dataset segmentation are used for restricted to range
additional iterations. images.
Ferrari et al. INRIA Grouping of Feature complexity Straight contour Contour segments Authors have not
dataset adjacent segments can be reused. spoken on the
contours range of k.
203
 204

Performance evaluation
Authors Dataset used Concept used Issues addressed Authors remark Our findings
parameter
Shotton et al. Several Contour Orientation specificity Multi scale Have build Number of
challenging features categorical class-specific fragment of
datasets used objects codebook of contours is not
recognition uncluttered specified.
contour
fragments.
Ravishankar et al. ETHZ dataset Deformation Detection rates (FPPI) Deformable Local scale Variation in
K.U. Sharma and N.V. Thakur

happens at object detection variations, objects not

high curvature bending addressed.
Table 2(b) Summary of contour-based OD work

points deformations
handled.
Schindler and ETHZ dataset Global shape Detection rates (FPPI) Shape similarity The method id Image
Suter invariant to deformation not
shape and considered.
rotation.
Lu et al. ETHZ and Particle filters Recall and precision Contour grouping Global shape Edge images have
PASCAL dataset explicitly not been
employed. addressed.
Maire et al. BSDS dataset Local and Recall and precision Localising Idealised line Contours’ passing
global cues junctions drawings are through an image
produced. is a query.
 A review and an approach for object detection in images 205

The system proposed by Shotton et al. (2008) not only recognises objects based on local
contour features but also is capable of localising the object in space and scale in the
image. Fragments of contours could be a good idea to guess the object but here lies a
question that how many fragments could be feasible.
Ravishankar et al. (2008) have proposed an efficient multi stage approach to object
recognition in real and cluttered images that is robust to scale and rotation. The system
makes use of the shape information to localise objects and detect their contours. They
claimed to extract object contours in the matching stage using dynamic programming and
also k-segment grouping together with centroid of the object can localise it. They have
not spoken about the variations in the objects or occlusion present in the images.
Schindler and Suter (2008) have presented a method for OD based on global shape.
Although this method applies well for natural images, its performance under occluded
images will be a certain issue that has to be considered and also if the image is deformed.
Lu et al. (2009) have defined the process of object detection and recognition as a contour
fragment grouping and labelling problem. The system works by first performing selection
of relevant contour fragments in edge images then grouping of the selected contour
fragments and finally their matching with the model contours; but the conclusion of all
these steps is reached by making use of particle filters (PF) with static observations, the
main advantage of using PF is the fact that global shape similarity can be explicitly
employed. The authors have only spoken about the edge images which can be the
limitation of this system. Maire et al. (2008) have presented a framework for contour and
junction detection. Basically the authors develop a contour detector using the
combination of local and global cues. Detecting junctions is a problem as the image
intensity surface is a bit confusing in the neighbourhood of a junction, and hence the right
approach to junction detection should take the advantage of the contours that are incident
on the junctions. Although the junctions are detected by making use of the contours
passing through them, but it is not clear that all the junctions in an image have contours
passing through it. Tables 2(a) and 2(b) summarise the approaches for OD based on
contours.

3.3 Graph-based OD
Model-based methods play a central role to solve different problems in computer vision.
A particular important class of such methods relies on graph models where an object is
decomposed into a number of parts, each one being represented by a graph vertex.
He et al. (2004) have presented a skeleton-based graph matching method for object
recognition and object localisation which makes use of the skeleton model and contour
segment model for this purpose. The use of these models helps reduce the matching space
comparatively. This method has worked with satisfaction in case of biomedical images
but still there remains a scope to implement this method on various other types of images
and see if we still get satisfactory results. Felzenszwalb and Huttenlocher (2004) have
addressed the problem of segmenting an image into regions; this is achieved by defining
a predicate in order to measure an evidence for a boundary between two regions by
making use of a graph-based representation of the image and by developing an efficient
segmentation algorithm based on the predicate defined earlier. However finding a
segmentation that is neither too coarse nor too fine is an NP-hard problem, hence there
remains a huge scope in redesigning this method of image segmentation and to get good
206 K.U. Sharma and N.V. Thakur

results. Dasigi and Jawahar (2008) have discussed a representation scheme for efficiently
modelling parts-based representation and matching them, as graphs can be used for
effective representation of images for detection and retrieval of objects, the problem of
finding a proper structure which can efficiently describe an image and can be matched in
low computational expense remains a problem. They in their discussion have compared
two graphical representations namely the nearest-neighbour graphs and the collocation
tress, for the goodness of fit and the computational expense involved in matching. A
graph model-based tracking algorithm which generate a model for a given frame termed
as reference frame was used to track a target object in the subsequent frames.
Paixão et al. (2008) have proposed a different method and have claimed to have
improved the recent algorithm in many ways, mainly instead of updating the model; each
analysed frame is back-mapped to the model space thus providing robustness to the
method as model parameters need not be modified each time. The variation in staining,
fixation and sectioning procedures during gland segmentation may lead to considerable
amount of artefacts and variances in tissue sections, which may result in variances in
gland appearance. Gunduz-Demir et al. (2010) have presented a new approach to gland
segmentation which decomposes the tissue image into a set of primitive objects and
segments glands making use of the organisational properties of these objects, which are
quantified with the definition of object-graphs.
Cyr and Kimia (2004) have presented a method to generate an aspect-graph
representation of complex shapes using the dissimilarity between neighbouring views to
generate aspects and to select prototypes for complex shape, and then the set of
prototypical views obtained from each 3D object is cast in hierarchy. In order to detect an
object, the unknown view is compared against the set of prototypes hierarchically. At the
coarse levels those prototypes whose distance to query is larger than the radius associated
with each prototype are pruned. Lin et al. (2009) have presented a model for representing
compositional object categories as an attribute grammar which are embedded in an
and-or graph for each compositional object category. The model combines the power of a
stochastic context free grammar (SCFG) to express the variability of part configurations,
and a Markov random field (MRF) to represent the pictorial spatial relationships between
these parts. They have also proposed a recursive inference algorithm which is used to
quickly constrain bottom-up detection while testing top-down constraints. Process is time
consuming. Yu et al. (2002) have proposed a mechanism based on spectral graph
partitioning that readily combine the process of segmentation and recognition into one.
Initially the part-based recognition system detects object patches. Patch grouping is used
to find set of patches that conform best to the object configuration. This process is
integrated with the pixel grouping based on low-level feature similarity, through
pixel-patch interactions and patch competition that is encoded as constraints in the
solution space. The globally optimal partition is obtained by solving a constrained eigen
value problem. Combining the two groupings may lead to extra overhead.
Hori et al. (2012) have made use of graph structural expression in order to develop a
method for generic object recognition by embedding a graph into the vector spaces. In
order to overcome the drawbacks of the previous methods where the location information
of the objects and the relationship between the key points is lost, they have proposed this
approach where the graph is constructed by connecting scale-invariant feature transform
(SIFT) key points with lines, as a result of this key points maintain their relationship and
then the structural representation with location information is achieved. They have not
 A review and an approach for object detection in images 207

specified on what kind of images this method is feasible. Till now we have seen how are
objects detected and localised, but Vajda et al. (2009) have studied the problem of object
duplicate detection and localisation. They have proposed a graph-based approach for 3D
object duplicate detection which represents the spatial information of the object in order
to avoid making an explicit 3D object model. Finite domain constraint satisfaction
problem (FDCSP) assumes that the matching problem between regions and labels is
bijective. In image interpretation the matching problem is often non-univocal. The non-
univocal matching between data and a conceptual graph was not possible until the
introduction of arc consistency with bi-level (FDCSPBC) constraint is introduced by
Deruyver et al. (2009). Lebrun et al. (2011) have addressed the problem of graph
matching during the process of OD and localisation using the kernel functions. Basically
they have shown in their paper that the similarity between two graphs can be efficiently
and effectively computed through a set of walks within the graphs. Firstly they propose
kernels on graphs and kernels on walks, and then they propose the solutions for exact and
appropriate computations of the kernels. Authors have not specified the range of the walk
in a particular graph.
Zhang and Chang (2005) have presented a report in which they present a model
named random attributed relational graph (RARG) using which the authors show how
part matching and model learning could be achieved by combining variation learning
methods with part-based representations. They have tried to solve the part matching
problem through the formulation of association graphs that characterises the
correspondences between parts in an image and nodes in the object model. For this
system to be accurate there is a need for labelling the regions which can be an overhead.
Triesch and Eckes (2005) had reviewed object representations based on deformable
feature graphs which describe particular views of an object as a spatial constellation of
image features; these representations are useful in situations of high clutter and partial
occlusions. These representations have a number of advantages namely: it allows
recognition without prior segmentation of the object of interest, robustness to small
variations in appearance tends to be very good if features are chosen properly and many
more. Partial occlusion is dealt with but what if the occlusion is full. Nam and Bao (2012)
have presented a method to distinguish the principal objects in an image by making use
of graph-based segmentation and normalised histograms (PODSH). Their approach
basically focuses on such objects where one might focus while taking images, this
approach basically supposes that the position of the main object is located at the centre of
an image and the main object holds a large area. Normalised histograms are used to gain
the edge and corner information. They claimed that their system has shortcoming
due to lack of information of edge and corner information. Liang et al. (2012)
have proposed an approach to detect salient objects by making use of adaptive
multi-scale colour image neighbourhood hyper-graph representation and spectral hyper
graph partitioning methods. Initially their approach extracts a polygonal potential
region-of-interest by analysing the edge distribution in an image, then the image is
represented by context sensitive hyper graph and then finally an incremental hyper graph
partitioning is used to generate the candidate regions for final salient OD. Natural images
are only considered, there remains the question whether this approach is feasible on other
types of images.
208 K.U. Sharma and N.V. Thakur

Siddiqi et al. (1999) have applied the theory for the generic representation of 2D
shapes, where structural descriptions are derived from shocks (singularities) of a curve
evolution process, to the problem of shape matching. The singularities are organised into
a directed, acyclic shock graph and the space of all such graphs is highly structured and
can be characterised by the rules of a shock graph grammar. They in their paper
introduced tree matching algorithm which finds the best set of corresponding nodes
between two shock trees in polynomial time. This system has less feasibility
when an image is occluded. Serratosa et al. (2003) have presented an article on model
function-described graph (FDG), which is a type of compact representation of a set of
attributed graphs (AGs) that borrow from random graphs the capability of probabilistic
modelling of structural and attribute information. They have defined FDGs, their
features and two distance measures between AGs (unclassified patterns) and FDGs
(models or classes). Two applications of FDGs are also presented: in the first, FDGs are
used for modelling and matching 3D-objects described by multiple views, whereas in the
second, they are used for representing and recognising human faces, by several views.
Shams et al. (2001) have developed an algorithm which is an extension to the
labelled graph matching (LGM) algorithm named LGM1 and have compared the
performance of their algorithm by using the state of the art statistical method which is
based on mutual information maximisation (MIM). The LGM1 algorithm replaced the
pixel values with a Gabor wavelet representation which made it perform superior to the
successful version of LGM. Tables 3(a) through 3(c) summarises the approaches for OD
based on graphs.

3.4 Fuzzy-based OD
Reyes and Dadios (2004) have developed a logit-logistic fuzzy colour constancy
(LLFCC) algorithm for dynamic colour object recognition. This approach focuses on
manipulating a colour locus which depicts the colours of an object. A set of adaptive
contrast manipulation operators is introduced and utilised in conjunction with a fuzzy
inference system and a new perspective in extracting colour descriptors of an object are
presented. Again the question here arises about what colour ranges can be detected
feasibly. Munoz-Salinas et al. (2004) make use of the information provided by the
camera of a robot in order to assign a belief degree on the existence of a door in it; this is
done by analysing the segments of the image. Several fuzzy concepts are defined to lead
the search process and find different cases in which doors can be seen. Features of the
segments like size, direction or the distance between them are measured and analysed
using fuzzy logic in order to establish a membership degree of the segments on the
defined fuzzy concepts. This work is purely restricted to indoor environments. Bernardin
et al. (2007) have presented an automatic system for the monitoring of the indoor
environments using the pan-tilt-zoomable cameras. The system makes use of Haar like
feature classifier and colour histogram filtering in order to achieve reliable initialisation
of person tracks. The system uses a combination of adaptive colour and KLT feature
trackers for face and upper body which allows for robust tracking and track recovery in
the presence of occlusion. What level of darkness can the system deal with remains a
question?
 Performance evaluation
Authors Dataset used Concept used Issues addressed Authors remark Our findings
parameter
He et al. MRI corpus Graph Error in matching Recovery of Works with satisfaction Variety of images
callosum image matching objects in case of biomedical not considered for
dataset images. testing this
approach.
Pedro et al. COIL dataset Pair wise Difference Constructing a Substantial This issue is an NP
region in intensity graph using improvement observed hard problem and
comparison predicates due to this method. hence an
improvement still
exists.
Dasigi et al. UKBench Scheme for Accuracy Image Computational time for Various other types
dataset modelling matching this process is less of images not
compared to others. considered.
Paixão et al. Sequences from Back mapping Time taken to Tracking an Model parameters are An occluded image
Table 3(a) Summary of graph-based OD work

CAVIAR project process a video object not to be modified each may cause a
time. concern to this
approach.
Demir et al. Colon biopsy Decomposition Accuracy (FPPI) Segmentation of More tolerance to Classification of
samples of images into colon glands artefacts and variances glands not
primitive in tissues. considered.
objects
Cyr et al. Self-constructed Similarity bed Sample rate Representation Correct object A 3D cluttered
A review and an approach for object detection in images

dataset aspect graph of complex prototypes are always image not

shapes picked. considered.
Lin et al. Caltech dataset Graph Accuracy of Compositional Computational Process may be too
grammar detection object efficiency is relatively time consuming.
representation slower than other
approaches.
Yu et al. Self-constructed Graph Speed Recognition and Eliminates local false Combining the two
dataset partitioning segmentation as positives. groupings may lead
a single process to extra overhead.
209
 210
Performance evaluation
Authors Dataset used Concept used Issues addressed Authors remark Our findings
parameter
Vajda et al. PASCAL Graph model Computational Object duplicate Robust when using The approach is
VOC dataset complexity detection only one or few restricted to still
images for training. images.

Deruyver et al. NMR images Image Time Detection of Inadequacy may Works on
interpretation unexpected prevail when assumptions which
objects unexpected objects may not be true
appear. always.

Lebrun et al. Multiple Inexact graph Similarity Object Due to weak learning Authors have not
datasets used matching retrieval there is no need for specified the range of
the user to indicate the walk in a graph.
the ROI.
K.U. Sharma and N.V. Thakur

Zhang et al. Multiple Random Speed Part-based OD Detection accuracy Need for labelling the
datasets used attributed achieved by the regions can be an
Table 3(b) Summary of graph-based OD work

relational single RARG model. overhead.

graph
Triesch Multiple Deformable Time Representation Efficient recognition Partial occlusion is
datasets used feature of objects in the presence of dealt with but what if
graphs clutter and the occlusion is full.
occlusions.
Nam et al. DOLSOFT Normalised Precision Distinguishing The normalised System may have
dataset histograms principle histogram is added to shortcomings due to
objects increase the effect of lack of information
the system. of edge and corner
information.
Liang et al. MSRA Context Precision and recall Image Precise object There remains the
dataset sensitive classification boundaries are question whether this
hyper graph obtained. approach is feasible
on other types of
images.
Siddiqi et al. Self-created Singularities Time Representation The grammar permits This system may
database of curves of 2D objects a reduction of a have less feasibility
shock graph to a when an image is
unique rooted shock occluded.
tree.
 Table 3(c)

Performance evaluation
Authors Dataset used Concept used Issues addressed Authors remark Our findings
parameter
Serratosa et al. COIL Function Accuracy Modelling of It is more effective Clustered
dataset described structural and attribute and robust to images can be
graphs information partition a single an issue to this
3D-object. approach.
Summary of graph-based OD work

Shams et al. Self-created Graph Accuracy and speed Pattern recognition LGM1 algorithm is Working on
database matching better than MIM. occluded images
may not be
feasible.
Hori et al. 10-class Graph Detection rate Expressing an image Structural Authors have
image expression as an appearance representation with not specified on
dataset frequency histogram location information what kind of
is achieved. images this
method is
A review and an approach for object detection in images

feasible.
211
212 K.U. Sharma and N.V. Thakur

Malaviya and Malaviya (1993) have discussed various methods for fuzzy logic-based
visual object recognition systems. Basically fuzzy logic facilitates the smooth translation
of image information into natural language which can be easily processed by fuzzy set
theory. They have also commented that in fuzzy type OD, there is a lack of semantic
knowledge which raises the question that “What should be actually obtained from a given
image?” Elbouz et al. (2011) have proposed and validated a surveillance video system
that detects various posture-based events. The system makes use of adapted Vander-Lugt
correlator (VLC) and joint-transfer correlator (JTC) techniques in order to make
decisions on the identity of a patient and his 3D positions. They also proposed a fuzzy
logic technique to get decisions on the objects behaviour and an adapted fuzzy logic
control algorithm in order to make a decision based on information given to the system.
Again the question arises to the situation where the level of brightness is less. Kaur and
Dhir (2013) have proposed the implementation of face detection methods by combining
skin detection methods with template matching, where skin detection is carried out using
the YCbCr method where as the template matching is performed in edge detected image.
This edge detection is performed using a fuzzy edge detection method which is used to
detect edges of an image without determining the threshold value. Feasibility of this
approach for various kinds of images remains a question.
Kim et al. (2009) have proposed an object recognition processor which lightens
the workload by estimating the global ROI. This estimation of ROI is performed by a
neuro-fuzzy controller and this controller also manages the processors overall pipeline
stages by using workload aware task scheduling. As pipelining is introduced here raises a
question of parallel pipelining. Lopes et al. (2013) have introduced an object tracking
approach which is based on fuzzy concepts. The tracking task is performed through the
fusion of these fuzzy models by means of an inference engine. Here the object properties
considered are very basic, the properties like shape and textures, etc., have not been
considered. Rajakumar et al. (2011) have proposed a fuzzy filtering technique for contour
detection; the fuzzy logic is basically applied to extract value for an image which is used
for edge detection. In their approach, the threshold parameter values are obtained from
the fuzzy histograms of an input image, and the fuzzy inference method selects the
complete information about the border of the object. Their proposed system works for
grey images, but the question whether this system is feasible under occlusion or cluttered
image remains a question. Maddalena and Petrosino (2010) have adopted existing
approach to background subtraction which is based on self-organisation through artificial
neural networks, they have proposed a spatial coherence variant to such approach in order
to enhance robustness against false detections and have formulated a fuzzy model just to
deal with decision problems. Various video sequences are used. Ma et al. (2012) have
combined a fuzzy support vector machine (FSVM) with template matching process to
improve the computational efficiency of the process of OD; and also have parallelised the
process of template matching on a multi core platform with OpenMP. The system works
by initially classifying the samples by template matching and then they are refined
by the FSVM classifier. Although the use of multi core platforms increases the
computational efficiency, but the use of it also increases the cost and space required.
Tables 4(a) and 4(b) summarises the approaches for OD based on fuzzy logic.
 Performance evaluation
Authors Dataset used Concept used Issues addressed Authors remark Our findings
parameter
Reyes and Self-created Fuzzy colour Hits and misses Multi-channel Proposed scheme The question of the
Dadios dataset constancy colour imaging tremendously cuts colour ranges that
processing time. could be detected
remains.
Oz-Salinas et al. Self-created Fuzzy logic Accuracy Navigation in Proved to Restricted to indoor
dataset indoor successfully detect environment.
environments doors under strong
perspective
deformations.
Bernardin et al. Self-created Boosted False positives per Monitoring of Quickly acquire and Level of darkness
dataset cascades image (FPPI) indoor track subjects. not addressed.
environments
Malaviya and Not Discussed Not applicable Fuzzy In fuzzy type object Hence what should
Malaviya applicable logic-based detection, there is a be actually obtained
Table 4(a) Summary of fuzzy-based OD work

visual object lack of semantic from a given image

recognition knowledge. remains a query.
Elbouz et al. Self-created Optical Speed Patient This method is The question arises to
dataset correlation monitoring applicable to a very the situation where
wide range of the level of
situations and is brightness is less
robust enough.
Kaur and Dhir GTVA Fuzzy edge Detection rate Detection in still The computation Feasibility of this
dataset detection coloured images time taken by this approach for various
A review and an approach for object detection in images

method is more than kinds of images

classical methods. remains a question.
Kim et al. Any standard Neuro Execution time Reducing the Workload aware Parallel pipelining
dataset fuzzy-based workload management is not considered.
pipelining performed to reduce
power consumption.
Lopes et al. PETS dataset Fusion of Accuracy Single and The proposed method Properties like shape
fuzzy models multiple tracking is robust. and textures, etc.,
have not been
considered.
213
 214

Performance evaluation
Authors Dataset used Concept used Issues addressed Authors remark Our findings
parameter
Rajakumar et al. Grey Fuzzy filtering PSNR value Shape detection Filter mask is used Cluttered and
K.U. Sharma and N.V. Thakur

images used for all kinds of occluded images not

images. considered.
Table 4(b) Summary of fuzzy-based OD work

Maddalena et al. Not Fuzzy spatial False detection Background Robust against false Cluttered image
applicable coherence foreground detections. scenes not
separation considered.
Ma et al. UIUC Multi-core Accuracy Template High computational Cost and space
dataset computation matching efficiency is increases.
achieved.
 A review and an approach for object detection in images 215

3.5 Context-based OD
Wolf and Bileschi (2006) have designed a detector for object context. By using
context, authors have demonstrated detection of locations that are likely to contain the
object of interest. Mainly they have shown that context may be determined from basic
visual features like colour and texture. Occlusion remains a point. Perko and Leonardis
(2010) have presented a framework for visual-context aware OD; authors have tried to
extract visual contextual information from images which can be used prior to the process
of OD. In addition, bottom-up saliency and object cooccurrences are used in order to
define auxiliary visual context. Finally all the individual contextual cues are integrated
with local appearance-based object detector by using a fully probabilistic framework.
This system is tested on still images, can it work on other types of images remains an
issue.
Kumar and Hebert (2005) had presented a two layer hierarchical formulation to
exploit different levels of contextual information in images for robust classification. He
has claimed that their approach has two main advantages namely: first, it encodes the
short-range interactions as well as the long-range interactions and second being that it can
be applied to different domains. Although this system has these advantages, the authors
have assumed that the images are labelled, which could not be a case always.
Peralta et al. (2012) have presented a method which learns adaptive conditional
relationships that depend on the type of scene being analysed. Basically they have
proposed a model-based on a conditional mixture of trees which is able to capture
contextual relationships among objects using global information about an image.
Relationships between objects in an image could be formed only when the image is clear
enough but what if the image is occluded. Object categorisation makes use of appearance
information and context information in order to improve the object recognition accuracy.
Galleguillos and Belongie (2010) have addressed the problem of incorporating different
types of contextual information for object categorisation and have also reviewed the
different ways of using contextual information for object categorisation. Contextual
information would be accurate, once the images are labelled which will not be the case
always hence efficiency of this approach could be an issue.
Bergboer et al. (2006) have presented a dual stage context-based (COBA) OD
process; in the first stage an object descriptor based on visual features id used to find the
object candidates present in the image while in the second stage the identified object
candidates are assigned a confidence value based on contextual information. If the image
is occluded or cluttered then in that case the descriptor may have an issue while finding
an object in an image. Torralba et al. (2003) have presented a context-based vision
system for place and object recognition, firstly the locations are identified in order to
categorise environments and then use this information in order to provide contextual
priors for object recognition. The main advantage of using this system is that it cuts down
the number of possible objects that are needed to be considered. But what if the image is
cluttered with various different classes of objects remains an issue. Kontschieder et al.
(2012) have presented context sensitive decision forests which is used to exploit
contextual information that would be helpful for solving the OD problem. Their system
has the ability to access the information about each of the samples in the training set
216 K.U. Sharma and N.V. Thakur

which is helpful to learn the contextual information throughout the growing process. In
addition to this, they have also introduced a novel split criterion which in combination
with a priority-based way of constructing the trees, allows more accurate regression mode
selection and hence improves the current context information. Problem may arise when
there are a number of object classes present in an image.
Torralba et al. (2004) have presented a method for both detecting and segmenting the
objects present in the image. Boosted random fields (BRFs) have been introduced in
order to exploit the contextual information, this BRF algorithm combines boosting and
conditional random field (CRF) which eases the task of training and inference. Basically
they have tried to show hoe contextual information could be helpful for OD. But the
question regarding the various kinds of images remains an issue. Torralba (2003) had
introduced a simple framework for modelling the relationship between context and object
properties which is based on the correlation between the statistics of low level features
across the entire image and the objects that it contains. But when the number of object
classes in an image increases then the working of this system may come under question.
Song et al. (2011) have proposed an iterative contextualisation scheme in order to
mutually boost the performance of both OD and classification tasks. For this scheme to
work efficiently the authors have initially proposed a contextualised support vector
machine (context-SVM) through which a context adaptive classifier is achieved; and then
this context-SVM was made use of to boost performance of OD and classification tasks.
The scheme boosts up the performance but its efficiency on various kinds of images still
remains a challenge.
Chen and Tian have presented a part-based hierarchical compositional model (HCM)
which makes use of context information from signage for door detection and
classification of the doors that are detected. The basic advantage of using their system is
that it can handle partially captured objects as well as large variations in object classes.
What if the image is totally occluded or cluttered? Wang et al. (2011) have presented a
method named feature context (FC) which is an extension to shape context (SC) method;
where FC is used to encode the spatial information of local image features. SC computes
histogram of the points that belong to the target shape but on the contrary FC can be
applied to the entire image. They have also introduced radial basis coding (RBC) in order
to encode the local image features. Applying this process to the entire image may
increase the computational cost. Parikh et al. (2008) have proposed a model for context
which includes relative location and scale information along with cooccurrence
information. Basically when given a segmentation of an image, this model assigns each
segment to an object category based on the appearance and contextual information of the
segment. Their method may not perform well in case of a cluttered image or an occluded
image.
Rutishauser et al. (2004) have tried to investigate to what extent pure bottom-up
approach can extract useful information about the location, size and shape of objects from
images and how this information is useful in learning objects in an unlabelled image. And
have commented that bottom-up attention is useful for variety of applications. He et al.
(2004) have proposed an approach which makes use of contextual features for labelling
images; these features are incorporated into a probabilistic framework which combines
the outputs of many components. Their model basically is a combination of individual
 A review and an approach for object detection in images 217

models where each model provides labelling information, a classifier that looks at local
image statistics, regional label features and global image features that look at the label
patterns at local and global levels. Using contextual features for labelling a cluttered
image or an occluded image could be a challenge. Fink and Perona have proposed a
method which basically generalises the efficient features suggested by Viola and Jones.
They have tried to justify the power of a single AdaBoost learner could be augmented
using mutual boosting. The concept of mutual boosting could cause computational
overhead.
Shi and Malik (2000) have proposed an approach to solve the grouping problem in
vision; this is achieved by extracting the global features of an image rather than focusing
on the local features of an image. They have treated the grouping problem as a graph
partitioning problem and based on it they have proposed the normalised cut criteria for
segmenting the graph. The basic purpose of this criterion is that it measures both
the total dissimilarity between the different groups as well as the total similarity within
the groups. When the global features are considered, there are objects which are not of
our interest that are needed to be processed hence increasing the computational cost.
Galleguillos et al. (2008) have developed an approach for object categorisation which
uses context cooccurrences, location and appearance of an object for that purpose, and
hence have named their approach as for cooccurrence, location and appearance (CoLA).
CoLA makes use of CRF to maximise object label agreement in the image according to
spatial and cooccurrence constraints. Working on different kinds of images could be an
issue.
Ramström and Christensen (2004) have proposed a basic model for context-based OD
where context was extracted as coherent regions with the help of a distributed
segmentation scheme; this helped finding conspicuous visual cues and salient regions. By
guiding the foveated part of an active vision system to salient regions and conspicuous
visual cues the complexity of the visual search was reduced. What if the background was
occluded? Liu et al. (2011) have addressed the problem when using photogrammetric
constraints in OD when camera poses are unknown, basically photogrammetric context
captures the relationship between object heights and camera viewpoint. They also
proposed a branch-bound and cut algorithm to solve the NP-hard problem in structured
predictions when cuts of latent variable are embedded into the branch and bound process.
What if the objects are cluttered and what if the image is occluded? Sun et al. (2011) have
argued that feature selection is an important problem in OD and have demonstrated that
genetic algorithms (GAs) provide a simple, general, and powerful framework for
selecting good subsets of features, which lead to improved detection rates. They have
considered PCA for feature extraction and support vector machines (SVMs) for
classi1cation. The goal is searching the PCA space using GAs to select a subset of
eigenvectors encoding important information about the target object of interest. Genetic
algorithms may increase the time to detect objects.
Sun et al. (2012) have presented a framework which jointly detects objects, estimates
the scene layout and segments the supporting surfaces holding these objects. The object
detector module is capable of adaptively changing its confidence in establishing whether
a certain ROI contains an object (or not) and is-based on iterative estimation procedure
by which the object detector becomes more and more accurate. The worst case running
218 K.U. Sharma and N.V. Thakur

time of this algorithm can be high which may increase the computational cost. Russell
et al. (2007) have build a system which recognises and localises various object categories
in complex images and have achieved it by matching the input image with the images
present in the large training set of labelled images. Since there are regularities in object
identities across similar images, the retrieved matches provide hypotheses for object
identities and location and therefore they use a probabilistic model to transfer the labels
from the retrieval set to the input image. Again the worst case to get all the objects
matched would have a high execution time.
Rabinovich et al. (2007) have proposed an approach which incorporates semantic
object context as a post-processing step into any object categorisation model. Authors
make use of CRF framework that maximises object label agreement according to
contextual relevance. Their approach compares two sources of context: one learned from
training data and another queried from Google Sets. Again the running time of this
process will be high as matching is done. Singhal et al. (2003) have presented an
approach to determine scene content, which is based on a set of individual material
detection algorithms, as well as probabilistic spatial context models. As the major
limitation to individual material detectors is that the number of misclassifications that
occur because of similarities in colour and texture of various materials in an image and to
reduce the number of misclassifications, authors have developed a context-aware material
detection system. The major challenge to this system will come when the image is
cluttered and consists of many materials at a time. Zheng et al. (2009) have proposed a
context modelling framework which works without the need of any prior scene
segmentation or context annotation. This is achieved by exploring a polar geometric
histogram descriptor for context representation. In order to quantify context, authors have
formulated a new context risk function and a maximum margin context (MMC) model to
solve the minimisation problem of the risk function. Working of this system may be
questioned when the image is cluttered or occluded.
Verbeek and Triggs (2007) have introduced a CRF-based scene labelling model
which incorporates local features and features aggregated over the whole image. They
also introduce a method for learning CRFs from datasets with many unlabelled nodes by
marginalising out the unknown labels so that the log-likelihood of the known ones can be
maximised by gradient ascent. Considering the features of the total image may cause
unnecessary waste of time in certain cases. Belongie et al. (2002) have introduced a shape
descriptor named SC which is used for correspondence recovery and shape-based OD.
The SC works by capturing the distribution over relative positions of other shape points
and finally summarises global shape in a local descriptor. Deforming objects could cause
some issues with this shape descriptor. Shotton et al. (2007) have proposed an approach
for learning discriminative model of object classes, incorporating texture, layout, and
context information efficiently. Basically the learned model is used for automatic visual
understanding and semantic segmentation of images. They also proposed a discriminative
model that exploits texture-layout filters, features based on textons, which jointly model
patterns of texture and their spatial layout. Unary classification and feature selection are
achieved using shared boosting to give an efficient classifier which can be applied to a
large number of classes. Accurate image segmentation is achieved by incorporating the
unary classifier in a CRF. Occlusion remains a problem.
 A review and an approach for object detection in images 219

Gepperth et al. (2012) have tried to explore the potential contribution of multimodal
context information to OD in an ‘intelligent car’. The used car platform incorporates
subsystems for the detection of objects from local visual patterns, as well as for the
estimation of global scene properties such as the shape of the road area or the 3D position
of the ground plane. In order to quantify the contribution of context information, they
have investigated whether it can be used to infer object identity with little or no reference
to local patterns of visual appearance. This system is restricted to car detection.
Challenges will be given when a cluttered image is given to this system. Oliva and
Torralba (2007) have reviewed the role of context in object recognition process. And
have presented a computational model of attention guidance that integrates context
information with image saliency to determine regions of interest. By comparing scan
patterns of different models to those of human observers, authors validate the proposition
that top-down information from visual context modulates the saliency of regions during
the task of OD. Saliency calculation for an occluded image can be a concern. Kruppa and
Schiele (2003) have explored the idea of using local context for face detection purpose;
making use of quantitative and qualitative analyses authors claimed that the detection of
the local context of faces in greyscale images is feasible, which is in contrast to the
traditional object-centred approach to face detection where the role of local context has
so far been neglected. This idea has been only performed on grey scale images; there can
be an issue while using various kinds of images.
Murphy et al. (2003) have presented a method to combine global and local image
features in order to solve the task of OD. Standard approaches to OD focus on local
patches of the image, and try to classify them as background or not. They proposed to use
the scene context as an extra source of global information, to help resolve local
ambiguities. Also a CRF is presented for jointly solving the tasks of OD and scene
classification. Large number of object classes may cause problems to this approach. Bar
(2004) had proposed a testable model for rapid use of contextual associations in
recognition in which an early projection of coarse information can activate expectations
about context and identity that, when combined, result in successful object recognition.
They also have highlighted some open questions such as how are context frames
represented in the cortex, and what triggers their activation? How is contextual
information translated into expectations and many more? Tables 5(a) through 5(f)
summarises the approaches for OD based on context.

3.6 Other Types of OD

Torrent et al. (2013) have proposed a framework to simultaneously perform OD and
segmentation on objects of different nature, which is based on a boosting procedure
which automatically decides – according to the object properties – whether it is better to
give more weight to the detection or segmentation process to improve both results. Their
approach allows information to be crossed from detection to segmentation and vice versa.
The timing of this task may increase if initially the object detected is not the one of
interest. GE et al. (2009) have studied the use of Asymmetric Adaptive Boosting
(AdaBoost) in the OD process. And have commented that Asym-Gentle AdaBoost
methods are more robust than Asym-Real AdaBoost and achieve better performance than
the previous symmetric and asymmetric AdaBoost algorithms on both face detection and
pedestrian detection.
 220

Performance
Authors Dataset used Concept used evaluation Issues addressed Authors remark Our findings
parameter
Wolf et al. StreetScenes Context Detection Context Determination of map of The case of occlusion
dataset feature rates detection object context in under is not addressed.
10 seconds.
K.U. Sharma and N.V. Thakur

Perko et al. Multiple Visual context Detection rate Visual context Integration is based on Variety of images has
demanding aware OD modelling. not been considered
datasets used for testing.
Table 5(a) Summary of context-based OD work

Kumar Multiple Hierarchical Accuracy Modelling The formulation is Authors have

datasets used field different types general enough to be assumed that the
formulation of contexts applied to different images are labelled,
domains. which could not be a
case always.
Galleguillos et al. Not Context and Not applicable Object Authors review different Use of appearance
applicable appearance categorisation ways of using contextual information which
information information. may improve the
result.
 Performance
Authors Dataset used Concept used evaluation Issues addressed Authors remark Our findings
parameter
Peralta et al. OUTDOOR Adaptive Accuracy Category level Improves object recognition Obtaining relationship
and SUN09 context (FPPI) recognition performance with respect to a between objects during
datasets used model single tree model. occlusion is not explored.

Bergboer Self-created COBA Number of Object validation COBA performs favourably when If the image is cluttered then
et al. dataset approach false positives compared with current methods. descriptor may have an issue
while finding an object.

Kontschieder TUD dataset Context Accuracy Classification Allows more accurate regression Problem may arise when
et al. sensitive and regression mode selection and hence there are a number of object
decision improves the current context classes present in an image.
forest information.
Table 5(b) Summary of context-based OD work

Torralba Self-created Visual Accuracy Place Global image representation that if the image is cluttered with
et al. dataset context recognition provides relevant information for various different classes of
and place recognition. objects remains an issue.
categorisation

Torralba Self-created Boosted Accuracy and Exploiting The BRF algorithm provides a The question regarding the
et al. dataset random fields speed image data and natural extension of the cascade of various kinds of images
contextual classifiers. remains an issue.
information
A review and an approach for object detection in images

Torralba Self-created Contextual Accuracy and Modelling a Object locations and scales can be Increase in no of classes
dataset priming speed relationship inferred from a simple holistic may be a concern.
between context representation of context.
and object
properties
Song et al. VOC 2010 Iterative Detection rates Boosting Context-SVM utilised to Efficiency on various kinds
dataset contextualisat object iteratively and mutually boost of images remains a
ion classification performance of object detection challenge.
and classification.
221
 222

Performance
Table 5(c)

Authors Dataset used Concept used evaluation Issues addressed Authors remark Our findings
parameter
Chen et al. Self-created Part-based Accuracy Detection and Incorporation of contextual What if the image is
dataset HCM classification of information brings significant totally occluded or
doors improvements. cluttered?
Wang et al. Caltech Context Accuracy Image Using sliding window strategy, Applying this process to
dataset feature classification FC can outperform more the entire image may
sophisticated object detectors. increase the
computational cost.
Parikh et al. MSRC and Appearance Accuracy Dense scene Difficult scenes may require Cluttered images may be
Corel and context labelling the inclusion of more objects. an issue.
datasets used
K.U. Sharma and N.V. Thakur

Rutishauser Self-created Bottom-up Speed Learning Other modes of operation, such Bottom-up approach may
et al. dataset attention multiple objects as learning multiple objects are not be feasible when an
Summary of context-based OD work

from unlabelled mentioned. image is cluttered.

images
He et al. Corel and Multi scale Accuracy and Image labelling The main reasons for our Using contextual features
Sowerby random fields speed model’s success are its direct for labelling a occluded
datasets used representation of large-scale image is not considered.
interactions.
Fink et al. CMU/MIT Mutual Accuracy Information Mutual Boosting could be The concept of mutual
datasets used boosting inference enhanced by unifying the boosting could cause
selection of weak-learners. computational overhead.
Shi et al. Self-created Graph Detection rate Grouping This approach segments static Non-ROI are processed
dataset partitioning problem images and is found that results thereby increasing cost.
are encouraging.
Galleguillos MSRC and Cooccurrence, Categorisation Object Spatial interactions among Working on different
et al. PASCAL location and accuracy categorisation different categories are rather kinds of images could be
datasets used. appearance sparse. an issue.
information
 Performance
Authors Dataset used Concept used evaluation Issues addressed Authors remark Our findings
parameter
Ramström Self-created Background Accuracy Staggered The presented methodology is What if the background was
et al. dataset context recognition evaluated in the context of a occluded?
table top scenario.
Liu et al. INRIA Photogrammetric False Object This model can get significantly What if the objects are
dataset context detections recognition better detection performance cluttered and what if the
than models. image is occluded?
Sun et al. Multiple Feature subset Rate of Object detection Experimental results illustrate Genetic algorithms may
datasets used selection detection significant performance increase the time to detect.
improvements
Sun et al. In-house Context Accuracy Detecting This approach is built upon an The worst case running time
dataset used feedback objects, iterative estimation procedure. of this algorithm can be high.
Table 5(d) Summary of context-based OD work

estimating scene
layout
Rabinovich PASCAL Semantic context False Object Authors have incorporated a The running time of this
et al. and MSRC detections categorisation parts-based generative model for process will be high as
datasets used categorisation. matching is done.
Singhal et Corel dataset Spatial context Accuracy Content This scheme addresses spatial The major challenge to this
al. model understanding constraints among multiple system will come when the
material types. image is cluttered.
A review and an approach for object detection in images

Zheng et al. PASCAL Contextual False positive Context Superior performance of the Working of this system may
VOC2005) information detections surrounding of MMC model through extensive be questioned when the
and i-LIDS an object evaluation. image is cluttered or
datasets used occluded.
Verbeek MSRC Conditional Accuracy Scene Partially labelled training images Considering the features of
et al. dataset random fields segmentation could be handled by maximising the total image may cause
image segmentations. unnecessary waste of time.
223
 224

Table 5(e)

Performance
Authors Dataset used Concept used evaluation Issues addressed Authors remark Our findings
parameter
Belongie MNIST Shape False Shape matching Shape context leads to a robust Deforming objects could cause
et al. dataset context positives and recognition score. some issues.
Shotton Corel and Texton Accuracy Image The proposed algorithm gives Occlusion remains a problem.
et al. Sowerby boosting understanding competitive and visually pleasing
datasets used. results.
Gepperth HRI road Context Accuracy OD in an Basis function representations Challenges will be given when a
K.U. Sharma and N.V. Thakur

et al. traffic information intelligent car allow the simplest learning cluttered image is given to this
dataset. methods to perform best. system.
Summary of context-based OD work

Oliva Not Context Not Object A natural way of representing the Reviewing the role of context in
et al. applicable applicable recognition context of an object is in terms of object recognition process is
its relationship to other objects. done.
Oliva Self-created Top-down Accuracy Determining Contextual information provides Saliency calculation for an
et al. dataset information (FPPI) region of interest a shortcut for efficient object occluded image can be a concern.
detection systems.
Kruppa CMU/MIT Local context Number of Locating faces Using local context yields correct Various kinds of images are not
et al. datasets used correct detections. considered.
detections
Murphy Self-created Combining Accuracy OD and scene Authors present a conditional Large number of object classes
et al. dataset local and and speed recognition random field for jointly solving may cause problems to this
global image the tasks of object detection. approach.
features
 Performance
Authors Dataset used Concept used evaluation Issues addressed Authors remark Our findings
parameter
Table 6(a) Summary of other OD work

Torrent LabelMe, Boosting training Detection rate Simultaneous The approach is valid for Timing of this task may
et al. TUD and segmentation performing increase if ROI is not
Weismann and detection semiautomatic object known a priori.
datasets used labelling.
Jun-Feng Not Asymm-etric Not applicable Application of Asymmetric extensions There may be a problem
et al. applicable AdaBoost AdaBoost to OD can be derived naturally. while dealing with
cluttered images.
Hsieh et al. Self-created Watershed-based Accuracy Detecting This approach is feasible Various types of images
dataset transformation objects with low and effective in detecting not considered.
contrast objects.
A review and an approach for object detection in images
225
 226

Performance
Authors Dataset used Concept used evaluation Issues addressed Authors remark Our findings
parameter
Nguyen et al. MIT and Shape-based Accuracy in Capturing shape This approach performed Occlusion can be an issue
INRIA pattern matching and appearance well when a cluttered for this approach.
datasets used descriptor background.
Mimaroglu et al. Multiple Clustering Execution Arbitrary shape ASOD detects the number This approach may not be
datasets used time OD of objects automatically feasible when image is
with respect to the rate. occluded.
Hussin et al. Not applicable Circular hough Not OD from complex This approach CHT may not exactly
transform applicable background automatically detects the detect the circular object.
Table 6(b) Summary of other OD work

(CHT) images desire object.

K.U. Sharma and N.V. Thakur

Pavani et al. MIT/CMU Haar features Accuracy and Fast OD Object detectors based on Creating rectangles can be
datasets used speed the proposed features are problematic.
more accurate and faster.
Laptev VOC2005) Boosting Detection Localisation Validation of the method Boosting histograms may
dataset histograms rate on recent benchmarks for not be feasible when the
object recognition shows image is occluded or
its superior performance. cluttered.
Bhanu et al. SAR dataset Genetic Accuracy Synthesising GP can synthesise It may be time consuming
programming composite effective composite as it finds the features of
operators operators. the total image rather than
ROI.
Malagon-Borja MIT Principal Accuracy Locating This approach can be Timing can be an issue
et al. pedestrian component (FPPI) pedestrians in a generalised to the because the classifier
dataset analysis still image detection of several examines each location in
different types of objects. the image.
Zhang et al. UIUC image Spatial Detection rates Detection of Feature selection methods There will certainly be an
dataset histogram and false distinguishable are efficient. issue if the image is
features detections parts of an object occluded.
 A review and an approach for object detection in images 227

Hsieh et al. (2006) have proposed an approach to the detection of small objects which
makes use of a watershed-based transformation. Their proposed detection system
basically includes two main modules, ROI locating and contour extraction. ROI is
generated by making use of an image differencing technique and the watershed-based
segmentation algorithm applied on the ROI to extract object contours. Performance of
this method would be questioned when applied on various other types of images. Nguyen
et al. (2013) have proposed an object descriptor which is able to capture the shape and
appearance information of an object in an image. Contours templates which represent
object shape are used to derive a set of key points at which the appearance feature named
non-redundant local binary pattern (NR-LBP), is computed. Finally an object descriptor
is formed that concatenates BR-LBP features and appearance of the object. When an
image is occluded then the working of this approach may have some issues. Mimaroglu
and Erdil (2011) have considered the problem of detecting arbitrary shape objects as a
clustering application by decomposing images into representative data points, and then
performing clustering on these points. Their method of ASOD is based on COMUSA
which is an algorithm for combining multiple clustering’s. Their approach may not work
where an image is occluded.
Hussin et al. (2012) have discussed about the various techniques on how to detect the
mango from a mango tree. The techniques are colour processing which is used as primary
filtering to eliminate the unrelated colour or object in the image. Besides that, shape
detection are been used where it will use the edge detection, circular Hough transform
(CHT). Pavani et al. (2010) have proposed a method for assigning optimal weights to the
rectangles of the Haar-like features so that the weak classifiers constructed based on them
give best possible classification performance. The optimal weights were computed in a
supervised manner using three different techniques namely: brute-force search, genetic
algorithms and Fisher’s linear discriminant analysis. Creating rectangles can be
problematic when an image is cluttered or highly occluded. Laptev (2009) presented a
method for OD that combines AdaBoost learning with local histogram features.
He had introduced a weak learner for multi valued histogram features and also analyse
various choices of image features. Histogram-based descriptors can be feasible only when
the image is natural and clear. It may not be feasible when the image is occluded or
cluttered.
Zhao et al. (2011) have proposed a background model named greyscale arranging
pairs (GAP) which is based on the statistical reach feature (SRF). This model makes use
of the multi point pairs that exhibit a stable statistical intensity relationship as a
background model. The intensity difference between pixels of the pair is much more
stable than the intensity of a single pixel, especially in varying environments. Occlusion
can be an issue for this model. Bhanu and Lin (2004) have made use of genetic
programming in order to synthesise composite operators and composite features to detect
potential objects in images. Genetic programming can synthesise effective composite
operators for OD by running on selected training regions of training images and the
synthesised composite operators can be applied to the whole training images and other
similar testing images. Malagon-Borja and Fuentes (2009) have presented an OD system
which works without assuming any prior knowledge about the image. Their system
works as follows: in the first stage a classifier examines each location in the image at
different scales. Then in a second stage the system tries to eliminate false detections
based on heuristics. The classifier is based on the idea that Principal Component Analysis
228 K.U. Sharma and N.V. Thakur

(PCA) can compress optimally only the kind of images that were used to compute the
principal components (PCs). Thus the classifier performs separately the PCA from the
positive examples and from the negative examples; when it needs to classify a new
pattern it projects it into both sets of PCs and compares the reconstructions, assigning the
example to the class with the smallest reconstruction error. Timing can be an issue
because the classifier examines each location in the image.
Zhang et al. (2006) have presented a spatial histogram feature-based OD approach
which automatically selects informative spatial histogram features and learns a
hierarchical classifier by combining cascade histogram matching and a SVM to detect
objects in images. There will certainly be an issue if the image is occluded. Ugolotti et al.
(2013) have presented a method for OD in images that is based on deformable
models and swarm intelligence algorithms. The task of OD is modelled as an
optimisation problem which is tackled using particle swarm optimisation (PSO) and
differential evolution (DE). Occlusion and cluttered images may cause an issue to this
approach.
Park (2001) had presented a criterion composed of the area variation rate and the
compactness of the segmented shape which is based on optimisation that is used to select
local optimum thresholds. This method shows to have the shape resolving property in the
subtraction image, so that overlapped objects may be resolved into bright and dark
evidences characterising each object. His approach may have an issue when an image is
cluttered and large number of objects overlaps each other. Nonato et al. (2008) have
developed a framework for triangle characterisation in 2D meshes which is applied to
the problem of OD and also helps in creating two dimensional models from images. The
main advantage of using this is that it removes the unimportant objects present in an
image. But if the image is cluttered, then in that case identifying the unimportant object
may be an issue with this system. Table 6(a) through 6(c) summarises the various other
types of approaches for OD.

4 Open and key issues in OD

Following are the issues in the field of OD:

• Is it necessary to scan the whole image in order to locate the object?,
i.e., speed up.
• How to combine the classifiers?, i.e., accuracy.
• Which are good sets of classifiers that are needed to be combined?, i.e., accuracy.
• When and how should the combined classifiers be trained?, i.e., accuracy.
• Should multi-class recognition be performed by detection or by classification?,
i.e., speed up and accuracy.
• How to evaluate the performance for an undefined class distribution?,
i.e., performance evaluation.
 A review and an approach for object detection in images 229

• How can different views of an object be identified as representing a single object?,

i.e., accuracy.
• How to handle the occluded objects from detection point of view?,
i.e., improvement and efficiency.

5 Online source codes and datasets

The details for the source codes and the standard datasets, generally, used in the process
of OD is listed in Tables 7 and 8 respectively. Tables 7 and 8 provide the detail of the
authors, the language in which the source code is written and the URL’s for those source
codes and datasets.

5.1 Online source codes

Table 7 Details of various source codes available

Code for Author Language URL

Colour detection Shemal Fernando C++ http://opencv-
and object srf.blogspot.in/2010/09/obj
tracking ect-detection-using-colour-
seperation.html
Face detection Alexander Mordvintsev Python http://opencv-python-
using HAAR and Abid K. tutroals.readthedocs.org/en/
cascades latest/py_tutorials/py_objde
tect/py_face_detection/py_f
ace_detection.html
Object detection OpenCV dev Team OpenCV http://docs.opencv.org/mod
ules/ocl/doc/object_detectio
n.html
Object detection Shamsheer Verma OpenCV, http://www.instructables.co
and tracking Visual Studio m/id/OBJECT-
C++ 2010 DETECTION-AND-
TRACKING-USING-
OPENCV-VISUAL-/
Training HAAR Thosten Ball OpenCV http://coding-
classifier robin.de/2013/07/22/train-
your-own-opencv-haar-
classifier.html
Real time object Chesnokov Yuriy C++ http://www.codeproject.co
tracker m/Articles/22243/Real-
Time-Object-Tracker-in-C
Part-based Daniel Rodríguez Molina C++ http://www.uco.es/~in1maji
object detection m/proyectos/libpabod/
Object detector Li Fei-Fei MATLAB http://people.csail.mit.edu/t
with boosting orralba/shortCourseRLOC/
boosting/boosting.html
230 K.U. Sharma and N.V. Thakur

5.2 Online datasets

Table 8 Details of various datasets available

Datasets URL
LabelMe toolbox http://people.csail.mit.edu/torralba/LabelMeToolbox/LabelMeToolb
ox.zip
CMU/MIT face dataset http://vasc.ri.cmu.edu/idb/images/face/frontal_images/images.tar
UIUC car dataset http://l2r.cs.uiuc.edu/~cogcomp/Data/Car/CarData.tar.gz
INRIA person dataset http://pascal.inrialpes.fr/data/human/INRIAPerson.tar
ETHZ dataset http://www.vision.ee.ethz.ch/datasets/downloads/ethz_shape_classes
_v12.tgz
PASCAL VOC 2010 http://pascallin.ecs.soton.ac.uk/challenges/VOC/voc2010/VOCtrainv
al_03-May-2010.tar
PASCAL VOC 2011 http://pascallin.ecs.soton.ac.uk/challenges/VOC/voc2011/VOCtrainv
al_25-May-2011.tar
PASCAL VOC 2012 http://pascallin.ecs.soton.ac.uk/challenges/VOC/voc2012/VOCtrainv
al_11-May-2012.tar
BSD dataset http://www.eecs.berkeley.edu/Research/Projects/CS/vision/bsds/BS
DS300-images.tgz

6 Possible solution

Although for human beings, the recognition of familiar objects of any type and in any
kind of environment may be a simple task, but the process of recognition is still a huge
difficulty for computers. Especially, the situations when there are changes in light or
there is some sort of movements in space, the images of same object look differently. On
the other hand, the number of instruments that are able to capture images from day to day
life has increased drastically. To make instruments capable to classify such huge amounts
of data, OD has become a real challenge for the researchers.
After studying the referred literature, it is identified that the accuracy and time are the
main key parameters to judge any approach of OD. A few of the research papers provide
the balancing solution between the accuracy and time. Possible approach to have the
good balance between the accuracy and time can be devised by using the concept of
Steiner tree. The task of detecting objects and classifying them into a particular class is
formulated as a Steiner tree problem. The Steiner tree problem basically deals with
finding the minimum path between the given set of vertices. As it is an optimisation
problem and NP-hard problem, the scope for research exists.
The basic sliding window approach for OD analyses a large number of image regions
(of the order of 50,000) regions for a 640 × 480 pixel image) to know which of the region
contain the object of interest. In many applications, there is a need for recognising
multiple object classes, and hence multiple binary classifier are required to run over each
region, for instance, if detected object to be classified into one of the ten object classes,
the sliding window approach may require to explore 500,000) regions to detect the
correct class. Hence there is a need to have the approach which can minimise the search
space (number of regions), i.e., the approach should explore only those regions where the
 A review and an approach for object detection in images 231

probability of occurrence of object of interest is more. Steiner tree-based approach can be

developed to address this issue. For this purpose, a Steiner tree-based classifier can be
used to classify the object(s) present in a particular image. Multi scale boosted detector
can be used for recognising the objects in the image.
Various parameters for a particular image can be identified and the range of values
for these parameters should be fixed. Identified parameters can be used to mark the
various levels of the Steiner tree. Then, the classifier has to be trained using the range of
values for particular class so that the classifier can classify the identified object into
particular class. During the evaluation phase the classifier uses the set of values that are
obtained for a particular image and matches those values with the range of values store in
the database for particular class. When the parameter value obtained for the given image
is the same as the fixed value for the parameter, then it follows the path through the node
of one level to other, but when the parameter value for the given image does not match
the fixed value of the parameter, then the intermediate Steiner node is to be created for
this value for the relevant level. For later case, the evaluation of the intermediate Steiner
node can be carried out using the Euclidian distance to check the relevance of this
intermediate node with the main nodes of that level. Finally, the Steiner tree can be
created based on the identified main nodes. Last node of the created Steiner tree is the
Steiner node for the class of the object. Execution flow diagram for discussed possible
solution is shown in Figure 3.

Figure 3 Flow diagram of the possible solution

Start

Yes If Training No

Enter No. of Enter No. of

Images for Images to Detect
Training Objects

Enter some Apply Single

Description Scale Boosted
Detector

Compute Features Detected Object

Compute Mean
Value of Compute Features
Features

Apply Single
Scale Boosted Classifier Runs
Detector

Class of the
Detected Object Object

Stop
232 K.U. Sharma and N.V. Thakur

7 Conclusions

This paper presents the review of the various methods for detecting objects in images as
well as in videos. The process of OD is classified into five major categories namely
Sliding window-based, contour-based, graph-based, fuzzy-based and context-based.
Apart from this, other approaches that are used for detecting objects like the shape-based
detection and Steiner tree-based are also summarised.
A review on the topic of OD has been carried out by Prasad (2012), Madaan and
Sharma (2012) and Karasulu (2010). Prasad (2012) has discussed the problem of OD in
real images and addressed the various aspects like the feature types, learning model,
object templates, matching schemes and boosting methods, where as Madaan and Sharma
(2012) have considered the same problem of OD in remote sensing images and explored
the concept of pre-segmentation for OD. Karasulu (2010) has reviewed and evaluated the
methods for moving OD in videos. Though, the different review papers are available for
OD, but, this paper is different from the existing papers. This paper provides the details
of the existing approaches based on the key concept which is used as the base for
development of the approach. Apart from this, the summary of the available source codes
and the datasets used for the evaluation of the OD approach is presented. This paper also
provides the idea to solve the multi class OD problem based on the Steiner tree. This
paper is useful for the study purpose as well as for the new researchers who want to
explore the OD research area.

References
Amine, K. and Farida, M.H. (2012) ‘An active contour for range image segmentation’, Signal
& Image Processing: An International Journal (SIPIJ), Vol. 3, No. 3, pp.17–29,
doi: 10.5121/sipij.2012.3302.
Arbeláez, P. (2006) ‘Boundary extraction in natural images using ultra metric contour maps’,
Proceedings of the Conference on Computer Vision and Pattern Recognition Workshop,
p.182.
Bar, M. (2004) ‘Visual objects in contexts’, Nature Reviews Neuroscience, Vol. 5, No. 8,
pp.617–629, doi: 10.1038/nrn1476.
Belongie, S., Malik, J. and Puzicha, J. (2002) ‘Shape context: a new descriptor for shape matching
and object recognition’, IEEE Transactions on Pattern Analysis and Machine Intelligence,
Vol. 24, No. 4, pp.509–522.
Bergboer, N., Postma, E. and Herik, J. (2007) ‘Accuracy versus speed in context based
object detection’, Pattern Recognition Letters Vol. 28, No. 6, pp.686–694,
doi: 10.1016/j.patrec.2006.08.004.
Bhanu, B. and Lin, Y. (2004) ‘Object detection in multi-modal images using genetic
programming’, Journal of Applied Soft Computing, Vol. 4, No. 2, pp.175–201.
Camp, K. and Stiefelhagen, R. (2007) ‘Automatic person detection and tracking using fuzzy
controlled active cameras’, IEEE Conference on Computer Vision and Pattern Recognition.
Chen, C. and Tian, Y. (2010) ‘Door detection via signage context-based hierarchical compositional
model’, Computer Vision and Pattern Recognition Workshop, pp.1–6.
Comaschi, F., Stuijk, S., Basten, T. and Corporaal, H. (2013) ‘RASW: a run-time adaptive sliding
window to improve Viola-Jones object detection’, Proceedings of Seventh International
Conference on Distributed Smart Cameras (ICDSC), 2013, pp.1–6, 29 October to
1 November, doi: 10.1109/ICDSC.2013.6778224.
 A review and an approach for object detection in images 233

Cyr, C.M. and Kimia, B.B. (2004) ‘A similarity based aspect graph approach to 3D object
recognition’, International Journal of Computer Vision, Vol. 57, No. 1, pp.5–22.
Dasigi, P. and Jawahar, C.V. (2008) ‘Efficient graph based image matching for recognition and
retrieval’, Proceedings of National Conference on Computer Vision, Pattern recognition.
Deruyver, A., Hodé, Y. and Brun, L. (2009) ‘Image interpretation with a conceptual graph-labeling
over segmented images and detection of unexpected objects’, Artificial Intelligence, Vol. 173,
No. 14, pp.1245–1265.
Divvala, S.K. (2012) Thesis on ‘Context and Subcategories for Sliding Window Object
Recognition, CMU-RI-TR, pp.12–17.
Elbouz, M., Ayman, A. and Brosseau, C. (2011) ‘Fuzzy logic and optical correlation-based face
recognition method for patient monitoring application in home video surveillance’, Optical
Engineering, Vol. 50, No. 6, pp.1–13.
Felzenszwalb, P.F and Huttenlocher, D.P (2004) ‘Efficient graph based image segmentation’,
International Journal of Computer Vision, Vol. 59, No. 2, pp.167–181.
Ferrari, V., Fevrier, L., Jurie, F. and Schmid, C. (2008) ‘Groups of adjacent contour segments for
object detection’, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 30
No. 1, pp.36–51.
Fink, M. and Perona, P. (2004) ‘Mutual boosting for contextual inference’, Proceedings on
Advances in Neural Information Processing Systems.
Galleguillos, C. and Belongie, S. (2010) ‘Context based object categorization: a critical survey’,
Computer Vision and Image Understanding, Vol. 114, No. 6, pp.712–722.
GE, J. and Luo, Y. (2009) ‘A comprehensive study for asymmetric AdaBoost and its application in
object detection’, Journal of Acta Automatica Sinica, Vol. 35, No. 11, pp.1403–1409.
Gepperth, A., Dittes, B. and Ortiz, M.G. (2012) ‘the contribution of context information:
a case study of object recognition in an intelligent car’, Journal on Neurocomputing, Vol. 94,
No. 1, pp.77–86, doi: 10.1016/j.neucom.2012.03.008.
Gualdi, G., Prati, A. and Cucchiara, R. (2011) ‘Multi-stage particle windows for fast and accurate
object detection’, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 34,
No. 8, pp.1589–1604.
Gunduz-Demir, C., Kandemir, M., Tosun, A.B. and Sokmensuer, C. (2010) ‘Automatic
segmentation of colon glands using object-graphs’, Medical Image Analysis, Vol. 14, No. 1,
pp.1–12.
He, L., Han, C.Y., Everding, B. and Wee, W.G. (2004) ‘Graph matching for object recognition and
recovery’, Pattern Recognition, Vol. 37 No. 7, pp.1557–1560.
He, X., Zemel, R.S. and Carreira-Perpinan, M.A. (2004) ‘Multiscale conditional random fields for
image labeling’, CVPR, Vol. 2, pp.695–702.
Hori, T., Takiguchi, T. and Ariki, Y. (2012) ‘Generic object recognition by graph structural
expression’, ICASSP, pp.1021–1024.
Hsieh, F., Han, C., Wu, N., Chuang, T.C. and Fan, K. (2006) ‘A novel approach to the detection of
small objects with low contrast’, Journal of Signal Processing, Vol. 86. No. 1, pp.71–83.
Hussin, R., Juhari, M.R., Kang, N.W., Ismail, R.C. and Kamarudin, A. (2012) ‘Digital image
processing techniques for object detection from complex background image’, in International
Symposium on Robotics and Intelligent Sensors, Vol. 41, pp.340–344.
Karasulu, B. (2010) ‘Review and evaluation of well-known methods for moving object detection
and tracking videos’, Journal of Aeronautics and Space Technologies, Vol. 4, No. 4,
pp.11–22.
Kaur, R. and Dhir, V. (2013) ‘Fuzzy logic based novel approach for face detection’, International
Journal of Latest Research in Science and Technology, Vol. 2, No. 1, pp.558–566.
Kim, J., Kim, M., Lee, S., Oh, J., Oh, S. and Yoo, H. (2009) ‘Real-time object recognition
with neuro-fuzzy controlled workload-aware task pipelining’, Micro, IEEE, Vol. 29, No. 6,
pp.28–43.
234 K.U. Sharma and N.V. Thakur

Kontschieder, P., Bul, S., Criminisi, A., Kohli, P., Pelillo, M. and Bischof, H. (2012)
‘Context-sensitive decision forests for object detection’, Proceedings on Advances in Neural
Information Processing Systems.
Kruppa, H. and Schiele, B. (2003) ‘Using local context to improve face detection’, Proceedings of
the BMVC, pp.3–12.
Kumar, S. and Hebert, M. (2005) ‘A hierarchical field framework for unified context-based
classification’, Tenth IEEE International Conference on Computer Vision, pp.1284–1291.
Lampert, C.H., Blaschko, M.B. and Hofmann, T. (2008) ‘Beyond sliding windows: object
localization by efficient sub-window search’, IEEE Conference on Computer Vision and
Pattern Recognition.
Laptev, I. (2009) ‘improving object detection with boosted histograms’, Journal of Vision
Computing, Vol. 27, No. 5, pp.535–544.
Lebrun, J., Gosselin, P. and Philipp-Foliguet, S. (2011) ‘Inexact graph matching based on kernels
for object retrieval in image databases’, Image and Vision Computing, Vol. 29, No. 11,
pp.716–729.
Liang, Z., Chi, Z., Fu, H. and Feng, D. (2012) ‘Salient object detection using context
sensitive hyper graph representation and partitioning’, Pattern Recognition, Vol. 45, No. 11,
pp.3886–3901.
Lin, L., Wu, T., Porway, J. and Xu, Z. (2009) ‘A stochastic graph grammar for compositional
object representation and recognition’, Pattern Recognition, Vol. 42, No. 7, pp.1297–1307.
Liu, Y., Wu, Y. and Yuan, Z. (2011) ‘Object detection using discriminative photogrammetric
context’, International Conference on Image Processing, pp.2405–2408.
Lopes, N.V., Couto, P., Jurio, A. and Melo-Pinto, P. (2013) ‘Hierarchical fuzzy logic based
approach for object tracking’, Knowledge-Based Systems, Vol. 54, No. 1, pp.255–268, doi:
10.1016/j.knosys.2013.09.014.
Lu, C., Latecki, L.J., Adluru, N., Yang, X. and Ling, H. (2009) ‘Shape guided contour grouping
with particle filters’, International Conference on Computer Vision.
Ma, Y., Wu, W. and He, Q. (2012) ‘Algorithm for object detection using multi-core parallel
computation’, International Conference on Medical Physics and Biomedical Engineering,
Vol. 33, pp.455–461.
Madaan, T. and Sharma, H. (2012) ‘Object detection in remote sensing images: a review’, IJSRP,
Vol. 2, No. 6, pp.1–3.
Maddalena, L. and Petrosino, A. (2010) ‘Fuzzy spatial coherence based approach to background
foreground separation for moving object detection’, Neural Comput & Applic, Vol. 19, No. 2,
pp.179–186.
Maire, M., Arbelaez, P., Fowlkes, C. and Malik, J. (2008) ‘Using contours to detect and localize
junctions in natural images’, IEEE Conference on Computer Vision and Pattern Recognition.
Malagon-Borja, L. and Fuentes, O. (2009) ‘Object detection using image reconstruction with PCA’,
Journal of Image and Vision Computing, Vol. 27, Nos. 1–2, pp.2–9.
Malaviya, A. and Malaviya, P. (1993) ‘Object recognition using fuzzy set theoretic techniques’,
SPIE Proceedings, Vol. 1962.
Mimaroglu, S. and Erdil, E. (2011) ‘ASOD: arbitrary shape object detection’, Journal of
Engineering Applications of Artificial Intelligence, Vol. 24, No. 7, pp.1295–1299.
Munoz-Salinas, R., Aguirre, E., Garcia-Silvente, M. and Gonzalez, A. (2004) ‘Door-detection
using computer vision and fuzzy logic’, Proceedings of the 6th WSEAS International
Conference on Mathematical Methods & Computational Techniques in Electrical
Engineering.
Murphy, K., Torralba, A. and Freeman, W.T. (2003) ‘Using the forest to see the trees: a graphical
model relating features, objects, and scenes’, Proceedings on Advances in Neural Information
Processing Systems.
 A review and an approach for object detection in images 235

Oliva, A., Torralba, A., Castelhano, M.S. and Henderson, J.M. (2003) ‘Top-down control of visual
attention in object detection’, IEEE Proceedings of the International Conference on Image
Processing, Vol. 1, pp.253–256.
Paixão, T.M., Graciano, A., Cesar Jr., R.M. and Hirata Jr., R. (2008) ‘A back-mapping approach
for graph based object tracking’, Brazilian Symposium on Computer Graphics and Image
Processing, pp.45–52, doi: 10.1109/SIBGRAPI.2008.32.
Parikh, D., Zitnick, L. and Chen, T. (2008) ‘From appearance to context-based recognition: dense
labeling in small images’, IEEE Conference on Computer Vision and Pattern Recognition.
Park, Y. (2001) ‘Shape-resolving local thresholding for object detection’, Pattern Recognition
Letters, Vol. 22, No. 8, pp.883–890.
Pavani, K., Gomez, D.D. and Frangi, A.F. (2010) ‘Haar-like features with optimally
weighted rectangles for rapid object detection’, Journal of Pattern Recognition, Vol. 43,
No. 1, pp.160–172.
Peralta, B., Espinace, P. and Soto, A. (2012) ‘Adaptive hierarchical contexts for object recognition
with conditional mixture of trees’, Proceedings British Machine Vision Conference,
pp.121.1–121.11.
Perko, R. and Leonardis, A. (2010) ‘A framework for visual-context-aware object detection in still
images’, Computer Vision and Image Understanding, Vol. 114, No. 6, pp.700–711.
Prasad, D. (2012) ‘Survey of the problem of object detection in real images’, IJIP, Vol. 6, No. 6,
pp.441–466.
Rabinovich, A., Vedaldi, A., Galleguillos, C., Wiewiora, E. and Belongie, S. (2007) ‘Objects in
context’, International Conference on Computer Vision.
Rajakumar, T.C., Perumal, S.A. and Krishnan, N. (2011) ‘A fuzzy filtering model for contour
detection’, ICTACT Journal on Soft Computing, Vol. 1, No. 4, pp.197–200.
Ramström, O. and Christensen, H.I. (2004) ‘Object detection using background context’,
International Conference on Pattern Recognition, pp.45–48.
Ravishankar, S., Jain, A. and Mittal, A. (2008) ‘Multi stage contour based detection of deformable
objects’, Computer Vision-European Conference on Computer Vision, pp.483–496.
Reyes, N.H. and Dadios, E.P. (2004) ‘Dynamic color object recognition using fuzzy logic’, Journal
of Advanced Computational Intelligence and Intelligent Informatics, Vol. 8, No. 1 pp.29–38.
Russakovsky, O. and Ng, A.Y. (2010) ‘A Steiner tree approach to efficient object detection’,
Proceeding of: the Twenty-Third IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), pp.1070–1077.
Russell, B.C., Torralba, A., Liu, C., Fergus, R. and Freeman, W.T. (2007) ‘Object recognition by
scene alignment’, Proceedings on Advances in Neural Information Processing Systems.
Rutishauser, U., Walther, D., Koch, C. and Perona, P. (2004) ‘Is bottom-up attention useful for
object recognition?’, Computer Vision and Pattern recognition, Vol. 2, No. 2, pp.37–44.
Schindler, K. and Suter, D. (2008) ‘Object detection by global contour shape’, Pattern Recognition,
Vol. 41, No. 12, pp.3736–3748.
Schlecht, J. and Ommer, B. (2005) ‘Contour based object detection’, International Conference on
Computer Vision, Vol. 1, pp.503–510.
Segvic, S., Kalafatic, Z. and Kovaˇcek, I. (2011) ‘Sliding window object detection without spatial
clustering of raw detection responses’, Proceedings of the Computer Vision Winter Workshop.
Serratosa, F., Alquezar, R. and Sanfeliu, A. (2003) ‘Function-described graphs for modeling
objects represented by sets of attributed graphs’, Pattern Recognition, Vol. 36, No. 3,
pp.781–798.
Shams, L.B., Brady, M.J. and Schaal, S. (2001) ‘Graph matching vs mutual information
maximization for object detection’, Neural Networks, Vol. 14, No. 3, pp.345–354.
Shi, J. and Malik, J. (2000) ‘Normalized cuts and image segmentation’, IEEE Transactions on
Pattern Analysis and Machine Intelligence, Vol. 22, No. 8, pp.888–905.
236 K.U. Sharma and N.V. Thakur

Shotton, J., Blake, A. and Cipolla, R. (2008) ‘Multi scale categorical object recognition using
contour fragments’, IEEE Transactions of Pattern Analysis and Machine Intelligence, Vol. 30
No. 7, pp.1270–1281.
Shotton, J., Winn, J., Rother, C. and Criminisi, A. (2007) ‘TextonBoost for image understanding:
multi-class object recognition and segmentation by jointly modeling texture, layout, and
context’, International Journal of Computer Vision, Vol. 81, No. 1, pp.2–23.
Siddiqi, K., Shokoufandeh, A., Dickinson, S. and Zucker, S. (1999) ‘Shock graphs and shape
matching’, International Journal of Computer Vision, Vol. 35, No. 1, pp.13–32.
Singhal, A., Luo, J. and Zhu, W. (2003) ‘Probabilistic spatial context models for scene content
understanding’, IEEE Conference on Computer Vision and Pattern Recognition, Vol. 1,
pp.235–241.
Song, Z., Chen, Q., Huang, Z., Hua, Y. and Yan, S. (2011) ‘Contextualizing object detection
and classification’, IEEE Conference on Computer Vision and Pattern Recognition,
pp.1585–1592.
Stiene, S., Lingemann, K., N¨uchter, A. and Hertzberg, J. (2006) ‘Contour based object detection in
range images’, 3D Data Processing, Visualization, and Transmission, Third International
Symposium, pp.168–175.
Subburaman, B., Venkatesh and Marcel, S. (2010) ‘Fast bounding box estimation based face
detection’, ECCV, Workshop on Face Detection.
Sudowe, P. and Leibe, B. (2011) ‘Efficient use of geometric constraints for sliding window object
detection in video’, 8th International Conference on Computer Vision Systems.
Sun, M., Bao, S. and Savarese, S. (2012) ‘Object detection with geometrical context feedback
loop’, IJCV, Vol. 100, No. 2, pp.154–169.
Suna, Z., Bebisa, G. and Millerb, R. (2011) ‘Object detection using feature subset selection’,
18th IEEE International Conference on Image Processing, Vol. 37, pp.2165–2176.
Torralba, A., Murphy, K.P., Freeman, W.T and Rubin, M.A. (2003) ‘Context-based vision system
for place and object recognition’, Ninth IEEE Conference on Computer Vision.
Torrent, A., Lladó, X., Freixenet, J. and Torralba, A. (2013) ‘A boosting approach for the
simultaneous detection and segmentation of generic objects’, Journal of Pattern Recognition
Letters, Vol. 34, No. 13, pp.1490–1498.
Triesch, J. and Eckes, C. (2005) ‘Object recognition with deformable feature graphs: faces, hands,
and cluttered scenes’, Handbook of Pattern Recognition and Computer Vision, pp.461–480,
doi: 10.1142/9789812775320_0025.
Ugolotti, R., Nashed, Y., Mesejo, P., Ivekovič, S., Mussi, L. and Cagnoni, S. (2013) ‘Particle
swarm optimization and differential evolution for model based object detection’, Journal of
Applies Soft Computing, Vol. 13, No. 6, pp.3092–3105.
Vajda, P., Dufaux, F., Minh, T.H. and Ebrahimi, T. (2009) ‘Graph based approach for 3d object
duplicate detection’, Proceedings of the 10th International Workshop on Image Analysis for
Multimedia Interactive Services.
Verbeek, J. and Triggs, B. (2007) ‘Scene segmentation with conditional random fields learned from
partially labeled images’, Proceedings of Advances in Neural Information Processing Systems.
Wang, X., Bai, X., Liu, W. and Jan, L.L. (2011) ‘Feature context for image classification and
object detection’, IEEE Conference on Computer Vision and Pattern Recognition.
Wolf, L. and Bileschi, S. (2006) ‘A critical view of context’, International Journal of Computer
Vision, Vol. 69 No. 2, pp.251–261.
Yang, X., Liu, H. and Latecki, L.J. (2012) ‘Contour based object detection as dominant set
computation’, Journal on Pattern Recognition, Vol. 45 No. 5, pp.1927–1936.
Yanulevskaya, V., Uijlings, J. and Geusebroek, J.M. (2013) ‘Salient object detection: from pixels
to segments’, Image and Vision Computing, Vol. 31, No. 1, pp.31–42.
Yu, S.X., Gross, R. and Shi, J. (2002) ‘Concurrent object recognition and segmentation by graph
partitioning’, NIPS, MIT Press, pp.1383–1390.
 A review and an approach for object detection in images 237

Zang, H., Gao, W., Chen, X. and Zhao, D. (2006) ‘Object detection using spatial histogram
features’, Journal of Image and Vision Computing, Vol. 24, No. 4, pp.327–341.
Zhang, D.Q. and Chang, S.F. (2005) Learning Random Attributed Relational Graph for Part Based
Object Detection, Columbia University ADVENT Technical Report.
Zhao, X., Satoh, Y., Takauji, H., Kaneko, S., Iwata, K. and Ozaki, R. (2011) ‘Object detection
based on a robust and accurate statistical multi-point-pair model’, Journal of Pattern
Recognition, Vol. 44, No. 6, pp.1296–1311.
Zheng, W.S., Gong, S. and Xiang, T. (2009) ‘Quantifying contextual information for object
detection’, IEEE 12th International Conference, pp.932–939.
Zhu, Q., Wang, L., Wu, Y. and Shi, J. (2008) ‘Contour context selection for object detection:
a set-to-set contour matching approach’, Computer Vision – European Conference on
Computer Vision, pp.774–787.

View publication stats