<s>
In	O
computer	B-Application
vision	I-Application
,	O
maximally	B-Algorithm
stable	I-Algorithm
extremal	I-Algorithm
regions	I-Algorithm
(	O
MSER	O
)	O
are	O
used	O
as	O
a	O
method	O
of	O
blob	B-Algorithm
detection	I-Algorithm
in	O
images	O
.	O
</s>
<s>
to	O
find	O
correspondences	B-Algorithm
between	O
image	O
elements	O
from	O
two	O
images	O
with	O
different	O
viewpoints	O
.	O
</s>
<s>
Maximally	B-Algorithm
stable	I-Algorithm
extremal	I-Algorithm
region	I-Algorithm
Let	O
an	O
extremal	O
region	O
such	O
as	O
all	O
points	O
on	O
it	O
have	O
an	O
intensity	O
smaller	O
than	O
.	O
</s>
<s>
The	O
concept	O
more	O
simply	O
can	O
be	O
explained	O
by	O
thresholding	B-Algorithm
.	O
</s>
<s>
A	O
maximally	B-Algorithm
stable	I-Algorithm
extremal	I-Algorithm
region	I-Algorithm
is	O
found	O
when	O
size	O
of	O
one	O
of	O
these	O
black	O
areas	O
is	O
the	O
same	O
(	O
or	O
near	O
the	O
same	O
)	O
than	O
in	O
previous	O
image	O
.	O
</s>
<s>
Multi-scale	O
detection	O
without	O
any	O
smoothing	O
involved	O
,	O
both	O
fine	O
and	O
large	O
structure	O
is	O
detected.Note	O
,	O
however	O
,	O
that	O
detection	O
of	O
MSERs	O
in	O
a	O
scale	O
pyramid	O
improves	O
repeatability	O
,	O
and	O
number	O
of	O
correspondences	B-Algorithm
across	O
scale	O
changes	O
.	O
</s>
<s>
In	O
Mikolajczyk	O
et	O
al.	O
,	O
six	O
region	O
detectors	O
are	O
studied	O
(	O
Harris-affine	B-Algorithm
,	O
Hessian-affine	B-Algorithm
,	O
MSER	O
,	O
edge-based	O
regions	O
,	O
intensity	O
extrema	O
,	O
and	O
salient	B-Algorithm
regions	I-Algorithm
)	O
.	O
</s>
<s>
Scale	O
change	O
–	O
Following	O
Hessian-affine	B-Algorithm
detector	O
,	O
MSER	O
comes	O
in	O
second	O
under	O
a	O
scale	O
change	O
and	O
in-plane	O
rotation	O
.	O
</s>
<s>
After	O
sorting	O
,	O
pixels	O
are	O
marked	O
in	O
the	O
image	O
,	O
and	O
the	O
list	O
of	O
growing	O
and	O
merging	O
connected	O
components	O
and	O
their	O
areas	O
is	O
maintained	O
using	O
the	O
union-find	B-Algorithm
algorithm	I-Algorithm
.	O
</s>
<s>
By	O
applying	O
RANSAC	B-Algorithm
to	O
the	O
centers	O
of	O
gravity	O
of	O
the	O
regions	O
,	O
a	O
rough	O
epipolar	B-Algorithm
geometry	I-Algorithm
can	O
be	O
computed	O
.	O
</s>
<s>
An	O
affine	B-Algorithm
transformation	I-Algorithm
between	O
pairs	O
of	O
potentially	O
corresponding	O
regions	O
is	O
computed	O
,	O
and	O
correspondences	B-Algorithm
define	O
it	O
up	O
to	O
a	O
rotation	O
,	O
which	O
is	O
then	O
determined	O
by	O
epipolar	B-Algorithm
lines	O
.	O
</s>
<s>
RANSAC	B-Algorithm
is	O
applied	O
again	O
with	O
a	O
more	O
narrow	O
threshold	O
,	O
and	O
the	O
final	O
epipolar	B-Algorithm
geometry	I-Algorithm
is	O
estimated	O
by	O
the	O
eight-point	B-General_Concept
algorithm	I-General_Concept
.	O
</s>
<s>
This	O
algorithm	O
can	O
be	O
tested	O
here	O
(	O
Epipolar	B-Algorithm
or	O
homography	O
geometry	O
constrained	O
matches	O
)	O
:	O
</s>
<s>
The	O
MSER	O
algorithm	O
has	O
been	O
used	O
in	O
text	O
detection	O
by	O
Chen	O
by	O
combining	O
MSER	O
with	O
Canny	B-Algorithm
edges	I-Algorithm
.	O
</s>
<s>
Canny	B-Algorithm
edges	I-Algorithm
are	O
used	O
to	O
help	O
cope	O
with	O
the	O
weakness	O
of	O
MSER	O
to	O
blur	O
.	O
</s>
<s>
To	O
enhance	O
the	O
MSER	O
regions	O
any	O
pixels	O
outside	O
the	O
boundaries	O
formed	O
by	O
Canny	B-Algorithm
edges	I-Algorithm
are	O
removed	O
.	O
</s>
<s>
When	O
these	O
are	O
minimized	O
the	O
graph	O
is	O
then	O
cut	B-Algorithm
to	O
separate	O
the	O
text	O
nodes	O
from	O
the	O
non-text	O
nodes	O
.	O
</s>
<s>
The	O
MSER	O
algorithm	O
has	O
been	O
adapted	O
to	O
colour	O
images	O
,	O
by	O
replacing	O
thresholding	B-Algorithm
of	O
the	O
intensity	O
function	O
with	O
agglomerative	B-Algorithm
clustering	I-Algorithm
,	O
based	O
on	O
colour	O
gradients	O
.	O
</s>
