<s>
In	O
computer	B-Application
vision	I-Application
,	O
the	O
problem	O
of	O
object	B-General_Concept
categorization	I-General_Concept
from	I-General_Concept
image	I-General_Concept
search	I-General_Concept
is	O
the	O
problem	O
of	O
training	O
a	O
classifier	B-General_Concept
to	O
recognize	O
categories	O
of	O
objects	O
,	O
using	O
only	O
the	O
images	O
retrieved	O
automatically	O
with	O
an	O
Internet	B-Application
search	I-Application
engine	I-Application
.	O
</s>
<s>
Ideally	O
,	O
automatic	O
image	O
collection	O
would	O
allow	O
classifiers	B-General_Concept
to	O
be	O
trained	O
with	O
nothing	O
but	O
the	O
category	O
names	O
as	O
input	O
.	O
</s>
<s>
This	O
problem	O
is	O
closely	O
related	O
to	O
that	O
of	O
content-based	B-Application
image	I-Application
retrieval	I-Application
(	O
CBIR	B-Application
)	O
,	O
where	O
the	O
goal	O
is	O
to	O
return	O
better	O
image	O
search	B-Application
results	I-Application
rather	O
than	O
training	O
a	O
classifier	B-General_Concept
for	O
image	O
recognition	O
.	O
</s>
<s>
Traditionally	O
,	O
classifiers	B-General_Concept
are	O
trained	O
using	O
sets	O
of	O
images	O
that	O
are	O
labeled	O
by	O
hand	O
.	O
</s>
<s>
The	O
use	O
of	O
Internet	B-Application
search	I-Application
engines	I-Application
to	O
automate	O
the	O
process	O
of	O
acquiring	O
large	O
sets	O
of	O
labeled	O
images	O
has	O
been	O
described	O
as	O
a	O
potential	O
way	O
of	O
greatly	O
facilitating	O
computer	B-Application
vision	I-Application
research	O
.	O
</s>
<s>
One	O
problem	O
with	O
using	O
Internet	O
image	O
search	B-Application
results	I-Application
as	O
a	O
training	O
set	O
for	O
a	O
classifier	B-General_Concept
is	O
the	O
high	O
percentage	O
of	O
unrelated	O
images	O
within	O
the	O
results	O
.	O
</s>
<s>
It	O
has	O
been	O
estimated	O
that	O
,	O
when	O
a	O
search	B-Application
engine	I-Application
such	O
as	O
Google	O
images	O
is	O
queried	O
with	O
the	O
name	O
of	O
an	O
object	O
category	O
(	O
such	O
as	O
airplane	O
?,	O
up	O
to	O
85%	O
of	O
the	O
returned	O
images	O
are	O
unrelated	O
to	O
the	O
category	O
.	O
</s>
<s>
Another	O
challenge	O
posed	O
by	O
using	O
Internet	O
image	O
search	B-Application
results	I-Application
as	O
training	O
sets	O
for	O
classifiers	B-General_Concept
is	O
that	O
there	O
is	O
a	O
high	O
amount	O
of	O
variability	O
within	O
object	O
categories	O
,	O
when	O
compared	O
with	O
categories	O
found	O
in	O
hand-labeled	O
datasets	O
such	O
as	O
Caltech	B-General_Concept
101	I-General_Concept
and	O
Pascal	B-Application
.	O
</s>
<s>
In	O
a	O
2005	O
paper	O
by	O
Fergus	O
et	O
al.	O
,	O
pLSA	B-General_Concept
(	O
probabilistic	B-General_Concept
latent	I-General_Concept
semantic	I-General_Concept
analysis	I-General_Concept
)	O
and	O
extensions	O
of	O
this	O
model	O
were	O
applied	O
to	O
the	O
problem	O
of	O
object	B-General_Concept
categorization	I-General_Concept
from	I-General_Concept
image	I-General_Concept
search	I-General_Concept
.	O
</s>
<s>
pLSA	B-General_Concept
was	O
originally	O
developed	O
for	O
document	B-Algorithm
classification	I-Algorithm
,	O
but	O
has	O
since	O
been	O
applied	O
to	O
computer	B-Application
vision	I-Application
.	O
</s>
<s>
It	O
makes	O
the	O
assumption	O
that	O
images	O
are	O
documents	O
that	O
fit	O
the	O
bag	B-General_Concept
of	I-General_Concept
words	I-General_Concept
model	I-General_Concept
.	O
</s>
<s>
pLSA	B-General_Concept
divides	O
documents	O
into	O
topics	O
as	O
well	O
.	O
</s>
<s>
The	O
pLSA	B-General_Concept
model	O
tells	O
us	O
the	O
probability	O
of	O
seeing	O
each	O
word	O
given	O
the	O
category	O
in	O
terms	O
of	O
topics	O
:	O
</s>
<s>
To	O
do	O
this	O
,	O
the	O
expectation	B-Algorithm
maximization	I-Algorithm
algorithm	I-Algorithm
is	O
used	O
,	O
with	O
the	O
following	O
objective	O
function	O
:	O
</s>
<s>
Absolute	O
position	O
pLSA	B-General_Concept
(	O
ABS-pLSA	O
)	O
attaches	O
location	O
information	O
to	O
each	O
visual	O
word	O
by	O
localizing	O
it	O
to	O
one	O
of	O
X	O
揵ins	O
?	O
in	O
the	O
image	O
.	O
</s>
<s>
Translation	O
and	O
scale	O
invariant	O
pLSA	B-General_Concept
(	O
TSI-pLSA	O
)	O
.	O
</s>
<s>
This	O
model	O
extends	O
pLSA	B-General_Concept
by	O
adding	O
another	O
latent	O
variable	O
,	O
which	O
describes	O
the	O
spatial	O
location	O
of	O
the	O
target	O
object	O
in	O
an	O
image	O
.	O
</s>
<s>
Again	O
,	O
the	O
parameters	O
and	O
can	O
be	O
solved	O
using	O
the	O
EM	B-Algorithm
algorithm	I-Algorithm
.	O
</s>
<s>
These	O
features	O
were	O
then	O
encoded	O
as	O
Scale-invariant	B-Algorithm
feature	I-Algorithm
transform	I-Algorithm
descriptors	O
,	O
and	O
vector	O
quantized	O
to	O
match	O
one	O
of	O
350	O
words	O
contained	O
in	O
a	O
codebook	O
.	O
</s>
<s>
One	O
important	O
question	O
in	O
the	O
TSI-pLSA	O
model	O
is	O
how	O
to	O
determine	O
the	O
values	O
that	O
the	O
random	O
variable	O
can	O
take	O
on	O
.	O
</s>
<s>
To	O
limit	O
the	O
number	O
of	O
possible	O
object	O
locations	O
to	O
a	O
reasonable	O
number	O
,	O
normal	O
pLSA	B-General_Concept
is	O
first	O
carried	O
out	O
on	O
the	O
set	O
of	O
images	O
,	O
and	O
for	O
each	O
topic	O
a	O
Gaussian	O
mixture	O
model	O
is	O
fit	O
over	O
the	O
visual	O
words	O
,	O
weighted	O
by	O
.	O
</s>
<s>
paper	O
compared	O
performance	O
of	O
the	O
three	O
pLSA	B-General_Concept
algorithms	O
(	O
pLSA	B-General_Concept
,	O
ABS-pLSA	O
,	O
and	O
TSI-pLSA	O
)	O
on	O
handpicked	O
datasets	O
and	O
images	O
returned	O
from	O
Google	O
searches	O
.	O
</s>
<s>
In	O
about	O
half	O
of	O
the	O
object	O
categories	O
tested	O
do	O
ABS-pLSA	O
and	O
TSI-pLSA	O
perform	O
significantly	O
better	O
than	O
regular	O
pLSA	B-General_Concept
,	O
and	O
in	O
only	O
2	O
categories	O
out	O
of	O
7	O
does	O
TSI-pLSA	O
perform	O
better	O
than	O
the	O
other	O
two	O
models	O
.	O
</s>
<s>
As	O
in	O
the	O
pLSA	B-General_Concept
approach	O
,	O
it	O
is	O
assumed	O
that	O
the	O
images	O
can	O
be	O
described	O
with	O
the	O
bag	B-General_Concept
of	I-General_Concept
words	I-General_Concept
model	I-General_Concept
.	O
</s>
<s>
The	O
distribution	O
of	O
topics	O
among	O
images	O
in	O
a	O
single	O
category	O
is	O
modeled	O
as	O
a	O
Dirichlet	B-General_Concept
process	I-General_Concept
(	O
a	O
type	O
of	O
non-parametric	B-General_Concept
probability	O
distribution	O
)	O
.	O
</s>
<s>
To	O
allow	O
the	O
sharing	O
of	O
topics	O
across	O
classes	O
,	O
each	O
of	O
these	O
Dirichlet	O
processes	O
is	O
modeled	O
as	O
a	O
sample	O
from	O
another	O
損arent	O
?	O
Dirichlet	B-General_Concept
process	I-General_Concept
.	O
</s>
<s>
These	O
can	O
be	O
gathered	O
automatically	O
,	O
using	O
the	O
first	O
page	O
or	O
so	O
of	O
images	O
returned	O
by	O
the	O
search	B-Application
engine	I-Application
(	O
which	O
tend	O
to	O
be	O
better	O
than	O
the	O
subsequent	O
images	O
)	O
.	O
</s>
<s>
To	O
learn	O
the	O
various	O
parameters	O
of	O
the	O
HDP	O
in	O
an	O
incremental	O
manner	O
,	O
Gibbs	B-Algorithm
sampling	I-Algorithm
is	O
used	O
over	O
the	O
latent	O
variables	O
.	O
</s>
<s>
Gibbs	B-Algorithm
sampling	I-Algorithm
involves	O
repeatedly	O
sampling	O
from	O
a	O
set	O
of	O
random	O
variables	O
in	O
order	O
to	O
approximate	O
their	O
distributions	O
.	O
</s>
<s>
At	O
each	O
iteration	O
,	O
and	O
can	O
be	O
obtained	O
from	O
model	O
learned	O
after	O
the	O
previous	O
round	O
of	O
Gibbs	B-Algorithm
sampling	I-Algorithm
,	O
where	O
is	O
a	O
topic	O
,	O
is	O
a	O
category	O
,	O
and	O
is	O
a	O
single	O
visual	O
word	O
.	O
</s>
<s>
The	O
size	O
of	O
the	O
OPTIMOL-retrieved	O
image	O
sets	O
surpass	O
that	O
of	O
large	O
human-labeled	O
image	O
sets	O
for	O
the	O
same	O
categories	O
,	O
such	O
as	O
those	O
found	O
in	O
Caltech	B-General_Concept
101	I-General_Concept
.	O
</s>
<s>
Classification	O
accuracy	O
:	O
Classification	O
accuracy	O
was	O
compared	O
to	O
the	O
accuracy	O
displayed	O
by	O
the	O
classifier	B-General_Concept
yielded	O
by	O
the	O
pLSA	B-General_Concept
methods	O
discussed	O
earlier	O
.	O
</s>
<s>
When	O
the	O
classifier	B-General_Concept
learns	O
incrementally	O
,	O
by	O
selecting	O
the	O
next	O
images	O
based	O
on	O
what	O
it	O
learned	O
from	O
the	O
previous	O
ones	O
,	O
three	O
important	O
results	O
are	O
observed	O
:	O
</s>
<s>
The	O
problem	O
of	O
content-based	B-Application
image	I-Application
retrieval	I-Application
is	O
that	O
of	O
improving	O
search	B-Application
results	I-Application
by	O
taking	O
into	O
account	O
visual	O
information	O
contained	O
in	O
the	O
images	O
themselves	O
.	O
</s>
<s>
Several	O
CBIR	B-Application
methods	O
make	O
use	O
of	O
classifiers	B-General_Concept
trained	O
on	O
image	O
search	B-Application
results	I-Application
,	O
to	O
refine	O
the	O
search	O
.	O
</s>
<s>
In	O
other	O
words	O
,	O
object	B-General_Concept
categorization	I-General_Concept
from	I-General_Concept
image	I-General_Concept
search	I-General_Concept
is	O
one	O
component	O
of	O
the	O
system	O
.	O
</s>
<s>
OPTIMOL	O
,	O
for	O
example	O
,	O
uses	O
a	O
classifier	B-General_Concept
trained	O
on	O
images	O
collected	O
during	O
previous	O
iterations	O
to	O
select	O
additional	O
images	O
for	O
the	O
returned	O
dataset	O
.	O
</s>
<s>
Examples	O
of	O
CBIR	B-Application
methods	O
that	O
model	O
object	O
categories	O
from	O
image	O
search	O
are	O
:	O
</s>
