<s>
The	O
constellation	B-General_Concept
model	I-General_Concept
is	O
a	O
probabilistic	O
,	O
generative	O
model	O
for	O
category-level	O
object	O
recognition	O
in	O
computer	B-Application
vision	I-Application
.	O
</s>
<s>
Like	O
other	O
part-based	B-General_Concept
models	I-General_Concept
,	O
the	O
constellation	B-General_Concept
model	I-General_Concept
attempts	O
to	O
represent	O
an	O
object	O
class	O
by	O
a	O
set	O
of	O
N	O
parts	O
under	O
mutual	O
geometric	O
constraints	O
.	O
</s>
<s>
Because	O
it	O
considers	O
the	O
geometric	O
relationship	O
between	O
different	O
parts	O
,	O
the	O
constellation	B-General_Concept
model	I-General_Concept
differs	O
significantly	O
from	O
appearance-only	O
,	O
or	O
"	O
bag-of-words	B-General_Concept
"	O
representation	O
models	O
,	O
which	O
explicitly	O
disregard	O
the	O
location	O
of	O
image	B-Algorithm
features	I-Algorithm
.	O
</s>
<s>
The	O
Constellation	B-General_Concept
Model	I-General_Concept
takes	O
advantage	O
of	O
this	O
fact	O
by	O
explicitly	O
modeling	O
the	O
relative	O
location	O
,	O
relative	O
scale	O
,	O
and	O
appearance	O
of	O
these	O
parts	O
for	O
a	O
particular	O
object	O
category	O
.	O
</s>
<s>
Model	O
parameters	O
are	O
estimated	O
using	O
an	O
unsupervised	B-General_Concept
learning	I-General_Concept
algorithm	O
,	O
meaning	O
that	O
the	O
visual	O
concept	O
of	O
an	O
object	O
class	O
can	O
be	O
extracted	O
from	O
an	O
unlabeled	O
set	O
of	O
training	O
images	O
,	O
even	O
if	O
that	O
set	O
contains	O
"	O
junk	O
"	O
images	O
or	O
instances	O
of	O
objects	O
from	O
multiple	O
categories	O
.	O
</s>
<s>
The	O
Constellation	B-General_Concept
Model	I-General_Concept
,	O
as	O
introduced	O
by	O
Dr.	O
Perona	O
and	O
his	O
colleagues	O
,	O
was	O
a	O
probabilistic	O
adaptation	O
of	O
this	O
approach	O
.	O
</s>
<s>
made	O
the	O
significant	O
step	O
of	O
training	O
the	O
model	O
using	O
a	O
more	O
unsupervised	B-General_Concept
learning	I-General_Concept
process	O
,	O
which	O
precluded	O
the	O
necessity	O
for	O
tedious	O
hand-labeling	O
of	O
parts	O
.	O
</s>
<s>
In	O
the	O
first	O
step	O
,	O
a	O
standard	O
interest	O
point	O
detection	O
method	O
,	O
such	O
as	O
Harris	B-General_Concept
corner	B-Algorithm
detection	I-Algorithm
,	O
is	O
used	O
to	O
generate	O
interest	O
points	O
.	O
</s>
<s>
Image	B-Algorithm
features	I-Algorithm
generated	O
from	O
the	O
vicinity	O
of	O
these	O
points	O
are	O
then	O
clustered	O
using	O
k-means	B-Algorithm
or	O
another	O
appropriate	O
algorithm	O
.	O
</s>
<s>
In	O
this	O
process	O
of	O
vector	B-Algorithm
quantization	I-Algorithm
,	O
one	O
can	O
think	O
of	O
the	O
centroids	O
of	O
these	O
clusters	O
as	O
being	O
representative	O
of	O
the	O
appearance	O
of	O
distinctive	O
object	O
parts	O
.	O
</s>
<s>
are	O
estimated	O
using	O
expectation	B-Algorithm
maximization	I-Algorithm
.	O
</s>
<s>
These	O
are	O
then	O
reduced	O
to	O
10-15	O
dimensions	O
by	O
principal	B-Application
component	I-Application
analysis	I-Application
,	O
giving	O
the	O
appearance	O
information	O
.	O
</s>
<s>
The	O
task	O
of	O
learning	O
the	O
model	O
parameters	O
is	O
accomplished	O
by	O
expectation	B-Algorithm
maximization	I-Algorithm
.	O
</s>
<s>
The	O
Constellation	B-General_Concept
Model	I-General_Concept
as	O
conceived	O
by	O
Fergus	O
et	O
al	O
.	O
</s>
<s>
For	O
each	O
of	O
these	O
datasets	O
,	O
the	O
Constellation	B-General_Concept
Model	I-General_Concept
is	O
able	O
to	O
capture	O
the	O
"	O
essence	O
"	O
of	O
the	O
object	O
class	O
in	O
terms	O
of	O
appearance	O
and/or	O
shape	O
.	O
</s>
<s>
It	O
is	O
important	O
to	O
note	O
that	O
the	O
Constellation	B-General_Concept
Model	I-General_Concept
does	O
not	O
generally	O
account	O
for	O
significant	O
changes	O
in	O
orientation	O
.	O
</s>
<s>
In	O
terms	O
of	O
computational	O
complexity	O
,	O
the	O
Constellation	B-General_Concept
Model	I-General_Concept
is	O
very	O
expensive	O
.	O
</s>
<s>
Because	O
the	O
computation	O
of	O
sufficient	O
statistics	O
in	O
the	O
E-step	O
of	O
expectation	B-Algorithm
maximization	I-Algorithm
necessitates	O
evaluating	O
the	O
likelihood	O
for	O
every	O
hypothesis	O
,	O
learning	O
becomes	O
a	O
major	O
bottleneck	O
operation	O
.	O
</s>
<s>
This	O
allows	O
for	O
a	O
greater	O
number	O
of	O
model	O
parts	O
and	O
image	B-Algorithm
features	I-Algorithm
to	O
be	O
used	O
in	O
training	O
.	O
</s>
