<s>
In	O
computer	B-Application
vision	I-Application
,	O
3D	B-Algorithm
object	I-Algorithm
recognition	I-Algorithm
involves	O
recognizing	O
and	O
determining	O
3D	O
information	O
,	O
such	O
as	O
the	O
pose	B-Architecture
,	O
volume	O
,	O
or	O
shape	O
,	O
of	O
user-chosen	O
3D	O
objects	O
in	O
a	O
photograph	O
or	O
range	B-Algorithm
scan	I-Algorithm
.	O
</s>
<s>
Typically	O
,	O
an	O
example	O
of	O
the	O
object	O
to	O
be	O
recognized	O
is	O
presented	O
to	O
a	O
vision	O
system	O
in	O
a	O
controlled	O
environment	O
,	O
and	O
then	O
for	O
an	O
arbitrary	O
input	O
such	O
as	O
a	O
video	B-Application
stream	I-Application
,	O
the	O
system	O
locates	O
the	O
previously	O
presented	O
object	O
.	O
</s>
<s>
This	O
can	O
be	O
done	O
either	O
off-line	O
,	O
or	O
in	O
real-time	B-General_Concept
.	O
</s>
<s>
Due	O
to	O
the	O
low	O
cost	O
and	O
ease	O
of	O
acquiring	O
photographs	O
,	O
a	O
significant	O
amount	O
of	O
research	O
has	O
been	O
devoted	O
to	O
3D	B-Algorithm
object	I-Algorithm
recognition	I-Algorithm
in	O
photographs	O
.	O
</s>
<s>
For	O
simplicity	O
,	O
many	O
existing	O
algorithms	O
have	O
focused	O
on	O
recognizing	O
rigid	O
objects	O
consisting	O
of	O
a	O
single	O
part	O
,	O
that	O
is	O
,	O
objects	O
whose	O
spatial	O
transformation	O
is	O
a	O
Euclidean	B-Algorithm
motion	I-Algorithm
.	O
</s>
<s>
Feature-based	O
approaches	O
work	O
well	O
for	O
objects	O
which	O
have	O
distinctive	O
features	B-Algorithm
.	O
</s>
<s>
Thus	O
far	O
,	O
objects	O
which	O
have	O
good	O
edge	O
features	B-Algorithm
or	O
blob	B-Algorithm
features	B-Algorithm
have	O
been	O
successfully	O
recognized	O
;	O
for	O
example	O
detection	O
algorithms	O
,	O
see	O
Harris	B-Algorithm
affine	I-Algorithm
region	I-Algorithm
detector	I-Algorithm
and	O
SIFT	B-Algorithm
,	O
respectively	O
.	O
</s>
<s>
Feature-based	O
object	O
recognizers	O
generally	O
work	O
by	O
pre-capturing	O
a	O
number	O
of	O
fixed	O
views	O
of	O
the	O
object	O
to	O
be	O
recognized	O
,	O
extracting	O
features	B-Algorithm
from	O
these	O
views	O
,	O
and	O
then	O
in	O
the	O
recognition	O
process	O
,	O
matching	O
these	O
features	B-Algorithm
to	O
the	O
scene	O
and	O
enforcing	O
geometric	O
constraints	O
.	O
</s>
<s>
The	O
method	O
starts	O
by	O
assuming	O
that	O
objects	O
undergo	O
globally	O
rigid	B-Algorithm
transformations	I-Algorithm
.	O
</s>
<s>
Because	O
smooth	O
surfaces	O
are	O
locally	O
planar	O
,	O
affine	O
invariant	O
features	B-Algorithm
are	O
appropriate	O
for	O
matching	O
:	O
the	O
paper	O
detects	O
ellipse-shaped	O
regions	O
of	O
interest	O
using	O
both	O
edge-like	O
and	O
blob-like	O
features	B-Algorithm
,	O
and	O
as	O
per	O
[	O
Lowe	O
2004 ]	O
,	O
finds	O
the	O
dominant	O
gradient	O
direction	O
of	O
the	O
ellipse	O
,	O
converts	O
the	O
ellipse	O
into	O
a	O
parallelogram	O
,	O
and	O
takes	O
a	O
SIFT	B-Algorithm
descriptor	O
on	O
the	O
resulting	O
parallelogram	O
.	O
</s>
<s>
Color	O
information	O
is	O
used	O
also	O
to	O
improve	O
discrimination	O
over	O
SIFT	B-Algorithm
features	B-Algorithm
alone	O
.	O
</s>
<s>
The	O
center	O
points	O
of	O
such	O
matching	O
features	B-Algorithm
correspond	O
,	O
and	O
detected	O
features	B-Algorithm
are	O
aligned	O
along	O
the	O
dominant	O
gradient	O
direction	O
,	O
so	O
the	O
points	O
at	O
(	O
1	O
,	O
0	O
)	O
in	O
the	O
local	O
coordinate	O
system	O
of	O
the	O
feature	O
parallelogram	O
also	O
correspond	O
,	O
as	O
do	O
the	O
points	O
(	O
0	O
,	O
1	O
)	O
in	O
the	O
parallelogram	O
's	O
local	O
coordinates	O
.	O
</s>
<s>
Thus	O
for	O
every	O
pair	O
of	O
matching	O
features	B-Algorithm
in	O
nearby	O
views	O
,	O
three	O
point	O
pair	O
correspondences	O
are	O
known	O
.	O
</s>
<s>
Given	O
at	O
least	O
two	O
matching	O
features	B-Algorithm
,	O
a	O
multi-view	O
affine	O
structure	B-Algorithm
from	I-Algorithm
motion	I-Algorithm
algorithm	O
(	O
see	O
[	O
Tomasi	O
and	O
Kanade	O
1992 ]	O
)	O
can	O
be	O
used	O
to	O
construct	O
an	O
estimate	O
of	O
points	O
positions	O
(	O
up	O
to	O
an	O
arbitrary	O
affine	O
transformation	O
)	O
.	O
</s>
<s>
therefore	O
selects	O
two	O
adjacent	O
views	O
,	O
uses	O
a	O
RANSAC-like	O
method	O
to	O
select	O
two	O
corresponding	O
pairs	O
of	O
features	B-Algorithm
,	O
and	O
adds	O
new	O
features	B-Algorithm
to	O
the	O
partial	O
model	O
built	O
by	O
RANSAC	B-Algorithm
so	O
long	O
as	O
they	O
are	O
under	O
an	O
error	O
term	O
.	O
</s>
<s>
Thus	O
for	O
any	O
given	O
pair	O
of	O
adjacent	O
views	O
,	O
the	O
algorithm	O
creates	O
a	O
partial	O
model	O
of	O
all	O
features	B-Algorithm
visible	O
in	O
both	O
views	O
.	O
</s>
<s>
Global	O
minimization	O
is	O
used	O
to	O
reduce	O
the	O
error	O
,	O
then	O
a	O
Euclidean	O
upgrade	O
is	O
used	O
to	O
change	O
the	O
model	O
's	O
feature	O
positions	O
from	O
3D	O
coordinates	O
unique	O
up	O
to	O
affine	O
transformation	O
to	O
3D	O
coordinates	O
that	O
are	O
unique	O
up	O
to	O
Euclidean	B-Algorithm
motion	I-Algorithm
.	O
</s>
<s>
At	O
the	O
end	O
of	O
this	O
step	O
,	O
one	O
has	O
a	O
model	O
of	O
the	O
target	O
object	O
,	O
consisting	O
of	O
features	B-Algorithm
projected	O
into	O
a	O
common	O
3D	O
space	O
.	O
</s>
<s>
To	O
recognize	O
an	O
object	O
in	O
an	O
arbitrary	O
input	O
image	O
,	O
the	O
paper	O
detects	O
features	B-Algorithm
,	O
and	O
then	O
uses	O
RANSAC	B-Algorithm
to	O
find	O
the	O
affine	O
projection	O
matrix	O
which	O
best	O
fits	O
the	O
unified	O
object	O
model	O
to	O
the	O
2D	O
scene	O
.	O
</s>
<s>
If	O
this	O
RANSAC	B-Algorithm
approach	O
has	O
sufficiently	O
low	O
error	O
,	O
then	O
on	O
success	O
,	O
the	O
algorithm	O
both	O
recognizes	O
the	O
object	O
and	O
gives	O
the	O
object	O
's	O
pose	B-Architecture
in	O
terms	O
of	O
an	O
affine	O
projection	O
.	O
</s>
