Computational Models of Machine Vision

Goal, Role and Success

Tayyaba Azim and Mahesan Niranjan

Communications, Signals, Processing and Control (CSPC) Group,

School of Electronics and Computer Science, University of Southampton, Southampton, U.K.

Keywords:

Object Detection, Feature Learning, Deep Models, Support Vector Machines, Fisher Kernel.

Abstract:

This paper surveys the learning algorithms of visual features representation and the computational modelling

approaches proposed with the aim of developing better artiﬁcial object recognition systems. It turns out that

most of the learning theories and schemas have been developed either in the spirit of understanding biological

facts of vision or designing machines that provide better or competitive perception power than humans. In this

study, we discuss and analyse the impact of notable statistical approaches that map the cognitive neural activity

at macro level formally, as well as those that work independently without any biological inspiration towards

the goal of developing better classiﬁers. With the ultimate objective of classiﬁcation in hand, the dimensions

of research in computer vision and AI in general, have expanded so much so that it has become important to

understand if our goals and diagnostics of the visual input learning are correct or not. We ﬁrst highlight the

mainstream approaches that have been proposed to solve the classiﬁcation task ever since the advent of the

ﬁeld, and then suggest some criterion of success that can guide the direction of the future research.

1 INTRODUCTION

Artiﬁcial object recognition has for long remained an

important problem in computer vision because of its

wide applications and the persistent gap in the perfor-

mance between the humans and artiﬁcial scene recog-

nition systems. The capability of human visual sys-

tem supersedes machine vision in many respects. For

example the very large number of object categories

that humans learn and recognize accurately despite

the variability in their position, size, viewpoint, il-

lumination, clutter and distractions, motivates us to

understand the functionality of human perception and

emulate it in artiﬁcial systems. However, it is not easy

to accomplish this task because of the limited neuro-

physiological ﬁndings on how brain learns and orga-

nizes the visual input. Fortunately, the advances in

both vision algorithms and hardware have made prac-

tical visual object recognition within reach, as can be

seen in systems deployed on airports and highways

for security and risk assessments. However, a com-

prehensive solution to this problem where machines

are as good as humans in terms of accuracy and speed

of recognition, still evades the reach of even the best

researchers with only partial solutions and limited

success in constrained environment being the state of

the art (Aggarwal et al., 1996).

One of the straightforward ways of refuting a the-

ory of how humans learn is to show that the machine

predictions do not match the given data. This is why

many recognition algorithms are assessed based on

their prediction accuracy and precision on benchmark

data sets. Another approach to deal with this issue is

to check whether the internal representations of vi-

sual input match that of the brain. This line of thought

has lead to various biologically inspired feature de-

tectors, graphical models and learning algorithms that

learn features resembling physiological signals of the

brain. In this paper, we ﬁrst of all provide a survey of

the mainstream approaches of object recognition al-

gorithms and then suggest some cues that might help

in guiding the progress of object recognition systems

in future.

2 EARLIER COMPUTATIONAL

MODELS OF OBJECT

RECOGNITION-ERA 1930-2005

Some of the early ideas on computational mod-

elling of visual object recognition can be traced back

to Gestalt’s work in the 1930s (Wertheimer, 1938),

(Fulcher, 2003) that describes how visual objects can

179

Azim T..

Computational Models of Object Recognition - Goal, Role and Success.

DOI: 10.5220/0004737601790186

In Proceedings of the 9th International Conference on Computer Vision Theory and Applications (VISAPP-2014), pages 179-186

ISBN: 978-989-758-003-1

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

be separated from each other and from the back-

ground. The Gestalt psychologists maintained that

humans constantly search for a ‘good ﬁt’ between

the visual image and the stored memories of visual

objects that are naturally organized in the brain as

patterns based on their continuity, similarity, closure,

proximity and symmetry. These deﬁned principles of

perception that assist grouping of stimuli and were

minimally effected by an individual’s past experience

are known as the Laws of Pragnanz. Gestalt theory

laid much of the groundwork for the study of object

recognition, however it was criticized heavily because

of being more descriptive than explanatory enough to

clarify the functioning of human vision. More tan-

gible work that draws ideas from the study of neural

processing in computational form, comes from Hubel

and Wiesel (Hubel and Wiesel, 1962), (Hubel and

Wiesel, 1965), who by observing the cat’s visual cor-

tex introduced the concept of hierarchical visual in-

formation processing in the receptive ﬁelds that are

regions where the action of light causes reﬂex in the

neurons. According to their ﬁndings, the receptive

ﬁelds of cells at one level of the visual system are

formed from input by cells at a lower level of the vi-

sual system. In this way, small, simple receptiveﬁelds

could be combined to form large, complex receptive

ﬁelds thus accounting for a progressiveincrease in the

complexity of physiological receptive ﬁelds of cells.

It was this discovery of receptive ﬁelds and feed for-

ward architecture that later on lead to the develop-

ment of many different hierarchical models of ma-

chine vision (Fukushima, 1988), (Wallis and Rolls,

1996), (Riesenhuber and Poggio, 1999), (Deco and

Rolls, 2006).

One of the most inﬂuential work on understand-

ing visual scene analysis after Hubel and Wiesel is

that of David Marr, who proposed hierarchical mod-

elling of the visual system from simple to complex

at three independent levels of abstraction: computa-

tional, algorithmic and implementational levels. In

computer analogy, these can be roughly understood

as task, software and hardware levels. According to

Marr, separating the three levels allow those inter-

ested in cognition to focus on the level they are most

interested in, while simultaneously allowing those not

specially interested in cognition (like computer scien-

tists) to provide valuable insight from their speciﬁc

point of view. The tri-level hypothesis is not with-

out any objections, but it remained a valuable tool

to aid in the study of cognitive science (and cogni-

tion in general). Apart from the tri-level hypothesis,

Marr also proposed an intermediate stage of informa-

tion representation - the 2-1/2D sketch - between the

2D image on the retina and a 3D description of the

world in cortex (Marr and Poggio, 1979), (Grimson,

1981). The idea of a primal sketch is similar to a

pencil drawing by an artist in which different areas

of a scene are shaded to give depth to it. This bot-

tom up hierarchical processing insight although sem-

inal, has now been modiﬁed by the recent research

(Serre et al., 2007b), (Rolls et al., 2009). However,

it highly inﬂuenced the state of the art object recog-

nition systems circa 1965-1980, giving birth to object

centered/shape based models that focused on ﬁnding

the correct representation for visual primitives, and

represented objects hierarchically in terms of their

structural properties. Marr and Nisihara’s idea of part

based structural representation was based on hierar-

chically stored three dimensional volumes of general-

ized cones (or cylinders) and their spatial relationship

to one another (Marr and Nishihara, 1978). This ap-

proach of using geometric primitives was an attempt

to reconstruct the shape of objects, in a similar vein to

how some other inspired approaches in parallel line

of work (Nevatia and Binford, 1973) were trying to

reconstruct the scene, however, they did not provide

any empirical support for the proposed model. Com-

pared to this initially proposed model by Marr, the

most well received structural descriptive model was

the recognition by components (RBC) model by Bie-

derman (Biederman, 1987) who reﬁned Marr’s model

of object recognition in important ways and provided

empirical support for the proposed theory : First and

foremost improvement was the psychophysical sup-

port of the RBC model (Biederman, 1986), (Bieder-

man and Cooper, 1991). Another deﬁning factor of

the recognition by components (RBC) theory was its

ability to recognize the objects regardless of the view-

ing angle, known as viewpoint invariance. Although

this model made sensible assumptions of how human

visual system may parse a scene, it was not without

caveats in practice. The main disadvantages of such

shape-based methods are: the dependency on reliable

extraction of geometric primitives(lines, circles, etc.),

the ambiguity in interpretation of the detected prim-

itives (presence of primitives that are not modelled),

the restricted modelling capability owned by a class of

objects which are composed of few easily detectable

elements, and the need to create the models manually

(Matas and Obdrzalek, 2004).

In contrast to the view point independent method

proposed by Biederman, the decade of 1990s saw

the evolution of appearance/view based models in

which the objects are represented with respect to their

viewpoint, thus entailing multiple representations that

place higher demands on memory capacity; however

it does potentially reduce the degree of computation

necessary for deriving higher-level object represen-

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

180

tations in object centered models. Based on the de-

rived features, these methods can be sub-divided into

two main classes, i.e., local and global approaches.

A local approach grabs a feature from a small re-

gion of an image (object) which is ideally a distinc-

tive property of the object’s view/projection to the

camera. Examples of local features of an object are

the color, mean gradient or mean gray value of pixels

from small region. In contrast, the global approaches

grab features that cover the information content of

the whole image. This varies from simple statisti-

cal measures (e.g., mean values or histograms of fea-

tures) to more sophisticated dimensionality reduction

techniques, i.e., subspace methods, such as princi-

ple component analysis (PCA), independent compo-

nent analysis (ICA), or non negative matrix factoriza-

tion (NMF). Some of the popular methods that come

into this category of models were proposed by (Turk

and Pentland, 1991), (Linsker, 1992), (Lades et al.,

1993), (Ojala et al., 1994), (Murase and Nayar, 1995),

(Bell and Sejnowski, 1997), (Lowe, 1999). The ques-

tion of whether the human visual system uses a view

based or an object centered representation has been

a subject of much controversy (for reviews see ref-

erences (Logothetis and Sheinberg, 1996) and (Tarr

and Blthoff, 1998)). We will just mention here the

fact that the psychophysical and physiological data

from humans and monkeys actually supports a view

based approach. View/appearance based approach is

attractive since it does not require image features or

geometric primitives to be detected and matched. But

their limitations, i.e. the necessity of dense sampling

of training views and the low robustness to occlu-

sion and cluttered background, make them suitable

mainly for certain applications with limited or con-

trolled variations in the image formation conditions,

e.g. for industrial inspection (Matas and Obdrzalek,

2004).

In order to address the issues faced by object cen-

tered and appearance based models, feature based

methods were proposed next, in which the objects

are represented by a set of view independent local

features which are automatically computed from the

training images and stored as a database for probing

the class of the test images later. Putting local fea-

tures into correspondence is an approach that is ro-

bust to object occlusion and cluttered background in

principle. Thus, when part of an object is occluded by

other objects in the scene, only features of that part

are missed and as long as there are enough features

detected in the unoccluded part, the object can be rec-

ognized. Examples of such features that have been

widely used for object recognition are scale invari-

ant feature transform (SIFT) (Lowe, 2004), histogram

of gradients (HoG) (Dalal and Triggs, 2005), haar

wavelet feature set (Viola and Jones, 2001), etc. Such

local patch based methods hold biological plausibil-

ity and tend to show beneﬁts over global approaches

when supported by mathematical models and neural

network frameworks in object categorization (Leibe

and Schiele, 2003).

One of the most popular ways of transforming a

set of low level features extracted from an image into

a high level image representation is the bag of vi-

sual words (BoW) inspired by the traditional bag of

words technique for text analysis. The BoW algo-

rithm constructs a codebook analogous to a dictionary

from the collection of orderless patch based features,

where each codeword in the codebook is a represen-

tative of several similar patches attained through the

clustering process; consequently the test image can

be represented by the histogram of the codewords.

Several state-of-the-art visual object recognition sys-

tems (Csurka et al., 2004), (Zhang et al., 2007), (Li

et al., 2008), (Wu and Rehg, 2011) ﬁt into this gen-

eral framework of codebook based object recognition

models. After the image features are represented in

the codebook of bag of words model, learning and

recognition can be done in a generative or discrimi-

native way. One of the greatest challenges in building

up a codebook based model is the computation time

required for clustering million of feature data points.

(Ramanan and Niranjan, 2010) proposed a solution to

this problem by presenting a sequential one-pass al-

gorithm that creates the codebook in a drastically re-

duced time.

From a computational scientist’s point of view,

it goes without saying that this continuous research

effort of developing bio-inspired architecture, learn-

ing algorithms and features was only taking place be-

cause the machines had not achieved human compat-

ible speed and accuracy of detecting scenes. In this

respect, the ﬁrst researcher to quantify the timing of

the visual scene understanding in humans was Simon

Thorpe (Thorpe et al., 1996), who explained through

event related potentials (ERP) analysis, the amount of

time it takes to categorize the visual scene in cortex,

which is 150ms. Progress towards understanding ob-

ject recognition was driven by exploring and linking

phenomenon at different levels of abstraction. At one

end, where hierarchical generative models and learn-

ing algorithms inspired from the cortex were being

improved, statistical methods independent of the biol-

ogy of the visual system were also being developed in

parallel. One popular paradigm which gathered a lot

of attention since mid 1990s was the Vapnik theory

of support vector machines (SVM) which showed im-

pressive classiﬁcation performance on many bench-

ComputationalModelsofMachineVision-Goal,RoleandSuccess

181

mark data sets. SVMs utilize a principle called kernel

trick that computes dot products in high dimensional

feature spaces using simple functions called kernels

deﬁned on pairs of input patterns. This trick enables

us to get a linearly separable hyperplane for the data

which is otherwise nonlinearly separable in the input

space. Not only did the SVM classiﬁer work suc-

cessfully with the state of the art BoW feature space

(derived from the BoW model discussed just above),

but also with Fisher kernels (Jaakkola and Haussler,

1998) that combined the beneﬁts of generative and

discriminative approaches to pattern classiﬁcation by

deriving a kernel from a generative model of the data.

Kernel classiﬁers like SVM proved their signiﬁcance

in various applications but they require a large amount

of labelled training data as well as aprior deﬁnition

of a suitable similarity metric/feature space in which

naive similarity metrics sufﬁce the classiﬁcation to

perform well. This requirement invites criticism by

the researchers who are of the view that arranging a

large amount of labelled data for many objects is ex-

pensive/impractical.

Although most of the proposed object recognition

systems are inspired from the hierarchical nature of

the primate cortex, it is worth mentioning that the

neural connectivity and learning algorithms of these

models have evolved with time. Earlier, most of the

computational efforts were focussed on feed forward

processing of information but since these feed for-

ward connections just constitute a small fraction of

the total connectivity in cortex, researchers shifted

their attention towards the development of systems

that made use of the back projection feedback too

(Rumelhart et al., 1986). Feedback using back pro-

jection provides the opportunity of using previous

knowledge, memory and task dependent expectations

in a system (Kreiman, 2008), (Karklin and Lewicki,

2005). This change in neural connectivity revolution-

ized the learning algorithms used in undirected graph-

ical models (Rumelhart et al., 1986), directed graph-

ical models (Hinton et al., 1995) and non graphical

models (Rumelhart et al., 1986). Although these the-

ories failed to answer the scientiﬁc question of how

the brain learns visual features, they produced two

neat tricks: one for learning directed graphical models

(Thulasiraman and Swamy, 1992) and the other one

for undirected models (Karklin and Lewicki, 2009),

(Hinton et al., 2006). Another inﬂuential fact that was

established was that individual neurons are not sufﬁ-

cient for discriminating between objects; rather popu-

lation of neurons should be analysed - a neuronal be-

havior also pointed out by the Wilson-Cowan model

(Wilson and Cowan, 1972) in early 1970s and later

addressed in many computational neuroscience prob-

lems (Sejnowsky, 1976), (Amit and Brunel, 1997),

(Brunel, 2000), (Hertz et al., 2004).

3 RECENT COMPUTATIONAL

MODELS OF IMAGE

UNDERSTANDING-ERA

2006-PRESENT

Much of the progress experienced in the last decade

has produced an overwhelming body of object recog-

nition results without explaining anything signiﬁcant

about the perception and vision phenomenon in hu-

man visual system. The success of these artiﬁcial sys-

tems is determined by the overall recognition accu-

racy and the time they take to categorize these images.

In order to cater the speed of recognition, it is worth

mentioning the spectacular success of Poggio and

Serre (Serre et al., 2007a) for developing an imme-

diate recognition system, which is the fastest known

form of computer object recognition against humans.

In this system, the parallel processing paradigm was

implemented rather than the conventional serial pro-

cessing machine learning. When analysed with ani-

mal presence/absence test with humans, the computer

did as well as the humans, and thus better than the

best machine vision programs available so far. Im-

mediate object recognition laid a new foundation of

overall visual recognition and extending this theory

to solve harder perception problem requires recruiting

higher levels of brain function which would take more

time and computational complexity for implementa-

tion. This extension has already began to spread in

the Neuroscience community; an example being Stan

Bileschi who applied this model to scene recognition

(Bileschi, 2006), which is a derivation of higher order

judgements, like it is a farm, a barn, a forest, etc.

As far as the goal of gaining better accuracy is

concerned, deep learning and representation has been

the subject of much recent research ever since the pro-

posed breakthrough in feature learning by Hinton in

2006 (Hinton et al., 2006). The central idea of his

greedy layer wise pre-training procedure is based on

training each layer of the graphical model indepen-

dently in an unsupervisedway and then taking the fea-

tures learnt at the previous layer as input to the next

level. The features learnt by the deep model can ei-

ther be used as an input to a standard supervised ma-

chine learning predictor such as support vector ma-

chines or as an initialization for a deep supervised

neural networks like multi-layer perceptron (MLP).

This idea of greedy layer wise unsupervised train-

ing was followed up quickly by the rest (Hinton and

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

182

Nair, 2009), (Taylor and Hinton, 2009), (Krizhevsky

et al., 2012) as deep architectures showed potential

of progressively learning more abstract features at

higher levels of representation, yielding better classi-

ﬁcation error (Larochelle and Bengio, 2008), (Erhan

et al., 2010) quality of samples generated by the prob-

ability distributions (Hinton and Salakhutdinov,2009)

and the invariance of properties learnt by the classi-

ﬁer . The recent work of (Krizhevsky et al., 2012)

also shows that with proper initialization of param-

eters and choice of non linearity, it is not necessary

to do unsupervised pretraining of the model as re-

quired by other deep networks.This ﬁnding reinforces

the hypothesis that the unsupervised pretraining acts

as a prior that brings little/no improvement over pure

supervised learning from scratch when training data

is large. The deep learning algorithms ﬁrst proved

their dominance over the MNIST digits data set by

breaking the SVMs classiﬁcation supremacy, and then

moved on to object recognition in natural images. The

latest breakthrough has been achieved on the Image

net data set, bringing the error rate of the state of

the art algorithms from 26.1% to 15.3% (Krizhevsky

et al., 2012) on 10K classes of objects. While deep

learning algorithms are making inﬂuential progress,

another impressive approach making its mark in par-

allel is that of Fisher kernels with SVMs (Jaakkola

and Haussler, 1998). The Fisher kernels made a

successful come back by ﬁrst showing its classiﬁca-

tion advantage over the state of the art bag of words

approach (Perronnin and Dance, 2007), (Perronnin

et al., 2010), and then showing its successful use with

large scale data sets like PASCAL VOC 2007 (Csurka

and Perronnin, 2011), CALTECH-256 (Sanchez and

Perronnin, 2011) and ImageNet-10K (Sanchez et al.,

2013). Currently, the second best performance after

deep convolution network (Krizhevsky et al., 2012)

achieved on the Image Net 10K classiﬁcation task is

shown by the Fisher kernels (Sanchez et al., 2013)

derived from a gaussian mixture model designed for

SIFT, local binary pattern (LBP) and GIST descrip-

tors of the data.

4 SOME GUIDING PRINCIPLES

FOR THE PROGRESS OF

OBJECT RECOGNITION

In this section, we point out some misleading prac-

tices by the research community in the ﬁeld of com-

puter vision and introduce some novel aspects of re-

search that have either been ignored completely or

given less attention so far. We maintain that taking

care of these aspects might improve the progress of

artiﬁcial object recognition systems in the future.

• Evaluation of the Benchmark Data Sets

In order to evaluate the strength of the learn-

ing algorithms and performance of classiﬁers, the

experiments are usually conducted on standard

benchmark data sets for comparison. Pinto (Pinto

et al., 2008) argued that publicly available data

sets such as Caltech-101 and PASCAL VOC im-

age sets lack in several aspects that can actually

mislead the progress in the long-term interest of

being able to achieve near human levels of recog-

nition. To prove this claim, he carried out the ex-

periments on a V1 like model which was based on

the known properties of simple cells of primate

visual cortex. The model was a population of

locally normalized, thresholded Gabor functions

spanning a range of orientations and frequencies.

This model contained no explicit mechanism to

tolerate variation in object position, size/pose and

shape. A standard one-versus-all approach was

used to generate the multi-class SVM classiﬁer

from the training images. It was found that this

V1 like model performed remarkably well on the

Caltech-101 data set but when tested on a care-

fully controlled object recognition task that just

consisted of two classes, the problem proved sub-

stantially harder for the V1 like model, exactly

as one would expect for an incomplete model of

object recognition. This proved that the V1 like

model performed well previously not because of

it being a good model of object recognition but

because the natural image sets were inadequate.

Ponce et al. (Ponce et al., 2006) also pointed out

some of the issues present in the current standard

data sets (i.e. UIUC, Caltech-4 and Caltech-101)

used for judging the performance of developed

object recognition systems. The most commonly

observed problems in all these data sets were the

limited range of variability in viewpoint, orienta-

tion of different instances in each category, no oc-

clusion and background clutter. We have not seen

any work that objects these claims about the in-

adequacy of these standard data sets or provides a

counter solution to this problem. We suggest that

there should be a formal mechanism of assigning

a challenging score to each of the benchmark data

sets in practice; based on this measure, the ones

that are too simple should be discarded for ex-

perimentation in the future. Such an initiative is

important to provide a uniform test bed to all the

competing algorithms on fair basis of evaluation

deﬁned explicitly through the challenging score.

ComputationalModelsofMachineVision-Goal,RoleandSuccess

183

• Impact of Learning Algorithms, Features and

Amount of Training Data

The object/scene classiﬁcation approaches often

focus on one of the three aspects of the recogni-

tion problem: the amount of training data, the ef-

ﬁciency of learning algorithm and the quality of

feature representations. It is important to know

which of these factors are responsible for humans

superior classiﬁcation performance. The answer

to this question was investigated by (Parikh and

Zitnick, 2010), who compared the human and ma-

chine responses on similar problems to evaluate

which of the three factors: learning algorithm,

amount of training data and features, are respon-

sible for better performance. They found no evi-

dence that human pattern matching algorithms are

better than standard machine learning algorithms.

Also humans do not take advantage of increased

amount of data, thus the main factor impacting the

accuracies is the choice of features. we maintain

that these observations should be further investi-

gated and not ignored in order to focus the efforts

in the right direction.

• Integration between Physiological Recordings

and Empirical Results of Object Recognition

Learning systems inspired by the biology and

evaluated by their classiﬁcation performance have

become much more sophisticated in the last few

decades. However, there is a need to directly

verify the empirical results of machine recogni-

tion algorithms with the physiological recordings.

Physiological data may offer an avenue for rec-

ognizing aspects of recognition that may be less

obvious for humans but more suitable for com-

puters. Such recognized cues could be integrated

within a machine’s control architecture to make

it more capable of responding to visual signals in

real time.

• Addition of Time Dynamics

Most of the well known computational models re-

viewed here do not take into explicit account the

fact that the retinal input usually has a time com-

ponent associated to it. It is important to con-

sider the time dynamics of the neural circuit as

the objects in our surroundings move and the eyes

show movement as well. Thus, measured neu-

ronal responses are functions of time and even

for an image presented in a ﬂash, different types

of information is carried out over time (Perrett

and Oram, 1993), (Sugase et al., 1999) or in the

time structure of the neuronal response. Incorpo-

rating the time dimension in neuronal models of

recognition is a challenge that began in the last

decade and is now actively being persued (Re-

ichert et al., 2011b), (Reichert et al., 2011a). One

of the interesting work in this regard is of (Nishi-

moto et al., 2011) who experimented on recon-

structing the visual brain activity elicited by natu-

ral scene movies in humans. The time dynamics

of the system is captured through a motion-energy

model that describes how spatial and temporal

information are represented in voxels through-

out the visual cortex and then uses a Bayesian

approach to combine estimated encoding models

with a sampled natural movie prior for movie re-

construction. Under the Bayesian framework, the

probability that the image s evoked response r, is

given by the posterior distribution, p(s|r) as fol-

lows: p(s|r) ∝ p(s)

∏

|s). To produce a re-

construction, p(s|r) is evaluated for a large num-

ber of images. The image with the highest pos-

terior probability, p(s|r) is selected as the recon-

struction, a method commonly known as maxi-

mum a posteriori(MAP) estimate. Much of the

excitement surrounding this work is motivated by

the ultimate objective of directly picturing subjec-

tive mental phenomenon such as visual imagery

(Thirion et al., 2006) or dreams. We argue that

time is an interesting dimension of the data, which

if added to the existing models, can assist in mak-

ing interesting discoveries about the human vision

that could be deployed in the artiﬁcial systems.

5 CONCLUSIONS

The huge successes of the Annual Meeting on Object

Perception, Attention, and Memory (in its 21st year),

Proceedings of the Neural Information Processing

Systems (in its 26th year) and the Annual Meeting of

Vision Sciences Society (in its 13th year) serve as a

measure of how far the ﬁeld has come and what con-

sistently stays unexplored. With many of the theo-

ries already presented by various researchers, it is now

high time to integrate all the efforts from various com-

munities of machine learning, neuroscience and phys-

iology to reassure the objectives of learning efﬁcient

visual representations in machines. With joint efforts,

one may discover the explanation for the underlying

factors of image understanding in the visual cortex,

as well as initiate signiﬁcant debate on those areas

where theories from different disciplines mismatch.

It is anticipated that future advances in brain signal

measurement and the development of more sophisti-

cated encoding models of visual information will lead

to a better apprehension of the complete neural model

of human visual system, thus paving a way for human

competitive object recognition systems.

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

184

REFERENCES

Aggarwal, J., Ghosh, J., Nair, D., and Taha, I. (1996).

A comparative study of three paradigms for ob-

ject recognition - bayesian statistics, neural networks

and expert systems. In Image Understanding: A

Festschrift for Azriel Rosenfeld, pages 241–262. So-

ciety Press.

Amit, D. and Brunel, N. (1997). Model of global spon-

taneous activity and local structured activity during

delay periods in the cerebral corte. Cereb. Cortex,

7:237–252.

Bell, A. and Sejnowski, T. (1997). The ‘Independent Com-

ponents’ of natural scenes are edge ﬁlters. Vision Re-

search, 37:3327–3338.

Biederman, I. (1986). Human image understanding: recent

research and a theory. In Second workshop on Human

and Machine Vision II, pages 13–57.

Biederman, I. (1987). Recognition-by-components: A the-

ory of human image understanding. Psychological Re-

view, 94(2):115–147.

Biederman, I. and Cooper, E. (1991). Priming contour-

deleted images: Evidence for intermediate represen-

tations in visual object recognition. Cognitive Psy-

chology, 23(3):393 – 419.

Bileschi, S. (2006). Streetscenes: Towards Scene Under-

standing in Still Images. PhD thesis, MIT.

Brunel, N. (2000). Dynamics of sparsely connected net-

works of excitatory and inhibitory spiking neurons.

Comput. Neurosci., 8:183–208.

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

C. (2004). Visual categorization with bags of key-

points. In Workshop on Statistical Learning in Com-

puter Vision, (ECCV), pages 1–22.

Csurka, G. and Perronnin, F. (2011). Fisher vectors: Be-

yond bag-of-visual-words image representations. In

Computer Vision, Imaging and Computer Graphics.

Theory and Applications, volume 229, pages 28–42.

Dalal, N. and Triggs, B. (2005). Histograms of oriented

gradients for human detection. In CVPR, pages 886–

893.

Deco, G. and Rolls, E. (2006). Decision-making and webers

law: a neurophysiological model. European Journal

of Neuroscience, 24:901–916.

Erhan, D., Bengio, Y., Courville, A., Manzagol, P., Vincent,

P., and Bengio, S. (2010). Why does unsupervised

pre-training help deep learning? JMLR, 11:625–660.

Fukushima, K. (1988). Neocognitron: a hierarchical neural

network capable of visual pattern recognition. Neural

Networks, 1(2):119–13.

Fulcher, E. (2003). Cognitive Psychology. NY, 1st edition.

Grimson, W. (1981). A computer implementation of a the-

ory of human stereo vision. Philosophical Transac-

tions of the Royal Society of London. Series B, Bio-

logical Sciences, 292(1058):217–253.

Hertz, J., Lerchner, A., and Ahmadi, M. (2004). Mean ﬁeld

methods for cortical network dynamics. pages 71–89.

Springer-Verlag.

Hinton, G., Dayan, P., Frey, B., and Neal, R. M. (1995).

The wake-sleep algorithm for self-organizing neural

networks. Science.

Hinton, G. and Nair, V. (2009). 3D object recognition with

Deep Belief Nets. In NIPS.

Hinton, G., Osindero, S., and Teh, Y. (2006). A fast learning

algorithm for Deep Belief Nets. Neural Computation,

18:1527–1554.

Hinton, G. and Salakhutdinov, R. (2009). Semantic hashing.

Approximate Reasoning, 50(7):969–978.

Hubel, D. and Wiesel, T. (1962). Receptive ﬁelds, binocu-

lar interaction and functional architecture in the cat’s

visual cortex. Physiology, 160:106–154.

Hubel, D. and Wiesel, T. (1965). Receptive ﬁelds and func-

tional architecture in two non-striate visual areas (18

and 19) of the cat. Neurophysiology, 18:229–289.

Jaakkola, T. and Haussler, D. (1998). Exploiting generative

models in discriminative classiﬁers. In NIPS, pages

487–493.

Karklin, Y. and Lewicki, M. (2005). A hierarchical

Bayesian model for learning non-linear statistical reg-

ularities in non-stationary natural signals. Neural

Computation, 17(2):397–423.

Karklin, Y. and Lewicki, M. (2009). Emergence of com-

plex cell properties by learning to generalize in natural

scenes. Nature, 457(2):83–86.

Kreiman, G. (2008). Biological object recognition. Schol-

arpedia, 3(6):26–67.

Krizhevsky, A., Sutskever, I., and Hinton, G. (2012). Im-

agenet classiﬁcation with deep convolutional neural

networks. NIPS.

Lades, M., Vorbruggen, J., Buhmann, J., Lange, J., Mals-

burg, C., Wurtz, R., and Konen, W. (1993). Distortion

invariant object recognition in the dynamic link archi-

tecture. Computers, 42(3).

Larochelle, H. and Bengio, Y. (2008). Classiﬁcation us-

ing discriminative restricted boltzmann machines. In

ICML.

Leibe, B. and Schiele, B. (2003). Interleaved object catego-

rization and segmentation. In BMVC, pages 759–768.

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

compact codebook for efﬁcient object categorization.

In IEEE Workshop on ACV, pages 1–6.

Linsker, R. (1992). Local synaptic learning rules sufﬁce

to maximise mutual information in a linear network.

Neural Computation, 4:691–702.

Logothetis, N. and Sheinberg, D. (1996). Visual object

recognition. Annual Review Neuroscience, 19:577–

621.

Lowe, D. (1999). Object recognition from local scale-

invariant features. In ICCV, volume 2, pages 1150–

1157.

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

Invariant keypoints. IJCV, 60(2):91–110.

Marr, D. and Nishihara, H. (1978). Representation

and recognition of the spatial organization of three-

dimensional shapes. 200(1140):269–294.

Marr, D. and Poggio, T. (1979). A computational the-

ory of human stereo vision. Proceedings of the

Royal Society of London. Series B, Biological Sci-

ences, 204(1156):301–328.

Matas, J. and Obdrzalek, S. (2004). Object recognition

methods based on transformation covariant features.

In EUSIPCO.

ComputationalModelsofMachineVision-Goal,RoleandSuccess

185

Murase, H. and Nayar, S. (1995). Visual learning and recog-

nition of 3-d objects from appearance. Computer Vi-

sion, 14(1):5–24.

Nevatia, K. and Binford, T. (1973). Structured descriptions

of complex objects. In IJCAI, pages 641–647.

Nishimoto, S., Vu, A., Naselaris, T., Benjamini, Y., Yu, B.,

and Gallant, J. (2011). Reconstructing visual experi-

ences from brain activity evoked by natural movies.

Current Biology, 21(19):1641–1646.

Ojala, T., Pietikainen, M., and Harwood, D. (1994). Perfor-

mance evaluation of texture measures with classiﬁca-

tion based on kullback discrimination of distributions.

In ICPR, volume 1, pages 582–585.

Parikh, D. and Zitnick, C. (2010). The role of features,

algorithms and data in visual recognition. In CVPR,

pages 2328 –2335.

Perrett, D. and Oram, M. (1993). Neurophysiology of shape

processing. IVC, 11:317–333.

Perronnin, F. and Dance, C. (2007). Fisher kernels on vi-

sual vocabularies for image categorization. In CVPR,

pages 1–8.

Perronnin, F., Snchez, J., and Mensink, T. (2010). Improv-

ing the ﬁsher kernel for large-scale image classiﬁca-

tion. In ECCV.

Pinto, N., Cox, D., and DiCarlo, J. (2008). Why is real-

world visual object recognition hard? PLoS Compu-

tational Biology, 41(1).

Ponce, J., Berg, T., Everingham, M., Forsyth, D., Hebert,

M., Lazebnik, S., Marszalek, M., Schmid, C., Russell,

B., Torralba, A., Williams, C., Zhang, J., and Zisser-

man, A. (2006). Dataset issues in object recognition.

volume 4170, pages 29–48. Springer Verlag.

Ramanan, A. and Niranjan, M. (2010). A One-pass

Resource-Allocating Codebook for patch-based visual

object recognition. In MLSP.

Reichert, D., Series, P., and Storkey, A. (2011a). Halluci-

nations in charles bonnet syndrome induced by home-

ostasis: a deep boltzmann machine model. In NIPS,

volume 23, pages 2020–2028.

Reichert, D., Series, P., and Storkey, A. (2011b). A hierar-

chical generative model of recurrent object-based at-

tention in the visual cortex. In Artiﬁcial neural net-

works (ANN), ICANN, pages 18–25.

Riesenhuber, M. and Poggio, T. (1999). Hierarchical mod-

els of object recognition in cortex. Nature Neuro-

science, 2:1019–1025.

Rolls, E., Loh, M., Deco, G., and Winterer, G. (2008-

2009). Computational models of schizophrenia and

dopamine modulation in the prefrontal cortex. Nature

Rev. Neurosci., 9(9):696–709.

Rumelhart, D., Hinton, G., and Williams, R. (1986). Learn-

ing representations by back-propagating errors. Na-

ture, 323(6088):533–536.

Sanchez, J. and Perronnin, F. (2011). High-dimensional

signature compression for large-scale image classiﬁ-

cation. In CVPR, pages 1665 –1672.

Sanchez, J., Perronnin, F., Mensink, T., and Verbeek, J.

(2013). Image Classiﬁcation with the FV: Theory &

Practice. Technical Report RR-8209, INRIA.

Sejnowsky, T. (1976). On global properties of neuronal in-

teraction. Biol. Cybern, 22:85–95.

Serre, T., Kreiman, G., Kouh, M., Cadieu, C., Knoblich,

U., and Poggio, T. (2007a). A quantitative theory of

immediate visual recognition. Progress in Brain Re-

search, 165:33–56.

Serre, T., Wolf, L., Bileschi, S., Riesenhuber, M., and

Poggio, T. (2007b). Robust object recognition with

cortex-like mechanisms. PAMI, 29:411–426.

Sugase, Y., Yamane, S., Ueno, S., and Kawano, K. (1999).

Global and ﬁne information coded by single neurons

in the temporal visual cortex. Nature, pages 869–873.

Tarr, M. and Blthoff, H. (1998). Image-based object recog-

nition in man, monkey and machine. Cognition, 67(1-

2):1–20.

Taylor, G. and Hinton, G. (2009). Factored conditional

Restricted Boltzmann Machine for modeling motion

style. In ICML.

Thirion, B., Duchesnay, E., Hubbard, E., Dubois, J., Po-

line, J., Lebihan, D., and Dehaene, S. (2006). Inverse

retinotopy: inferring the visual content of images from

brain activation patterns. Neuroimg., 33(4):1104–

1116.

Thorpe, S., Fize, D., and Marlot, C. (1996). Speed of pro-

cessing in the human visual system. Nature, 381:520–

522.

Thulasiraman, K. and Swamy, M. (1992). Graphs: Theory

and Algorithms. John Wiley & Sons, Inc., NY.

Turk, M. and Pentland, A. (1991). Eigenfaces for recogni-

tion. Cognitive Neuroscience, 3(1):71–86.

Viola, P. and Jones, M. (2001). Rapid object detection us-

ing a boosted cascade of simple features. In CVPR,

volume 1, pages I–511 – I–518.

Wallis, G. and Rolls, E. (1996). A model of invariant object

recognition in the visual system. Prog. Neurobiol.,

51:167–194.

Wertheimer, M. (1938). Laws of organization in perceptual

forms. W. Ellis, W (Ed. & Trans.), London: Routledge

& Kegan Paul(Original work published in 1923).

Wilson, H. and Cowan, J. (1972). Excitatory and inhibitory

interactions in localized populations of model neu-

rons. Biophysical Journal, 12(1):1–24.

Wu, J. and Rehg, J. (2011). Centrist: A visual descriptor for

scene categorization. PAMI, 33(8):1489–1501.

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

tures and kernels for classiﬁcation of texture and ob-

ject categories: a comprehensive study. Computer Vi-

sion, 73:213–238.

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

186