<s>
In	O
computer	B-Application
vision	I-Application
and	O
image	B-Algorithm
processing	I-Algorithm
,	O
a	O
feature	B-Algorithm
is	O
a	O
piece	O
of	O
information	O
about	O
the	O
content	O
of	O
an	O
image	B-Algorithm
;	O
typically	O
about	O
whether	O
a	O
certain	O
region	O
of	O
the	O
image	B-Algorithm
has	O
certain	O
properties	O
.	O
</s>
<s>
Features	O
may	O
be	O
specific	O
structures	O
in	O
the	O
image	B-Algorithm
such	O
as	O
points	O
,	O
edges	O
or	O
objects	O
.	O
</s>
<s>
Features	O
may	O
also	O
be	O
the	O
result	O
of	O
a	O
general	O
neighborhood	B-Algorithm
operation	I-Algorithm
or	O
feature	B-Algorithm
detection	O
applied	O
to	O
the	O
image	B-Algorithm
.	O
</s>
<s>
Other	O
examples	O
of	O
features	O
are	O
related	O
to	O
motion	O
in	O
image	B-Algorithm
sequences	O
,	O
or	O
to	O
shapes	O
defined	O
in	O
terms	O
of	O
curves	O
or	O
boundaries	O
between	O
different	O
image	B-Algorithm
regions	O
.	O
</s>
<s>
More	O
broadly	O
a	O
feature	B-Algorithm
is	O
any	O
piece	O
of	O
information	O
which	O
is	O
relevant	O
for	O
solving	O
the	O
computational	O
task	O
related	O
to	O
a	O
certain	O
application	O
.	O
</s>
<s>
This	O
is	O
the	O
same	O
sense	O
as	O
feature	B-Algorithm
in	O
machine	O
learning	O
and	O
pattern	O
recognition	O
generally	O
,	O
though	O
image	B-Algorithm
processing	I-Algorithm
has	O
a	O
very	O
sophisticated	O
collection	O
of	O
features	O
.	O
</s>
<s>
The	O
feature	B-Algorithm
concept	O
is	O
very	O
general	O
and	O
the	O
choice	O
of	O
features	O
in	O
a	O
particular	O
computer	B-Application
vision	I-Application
system	O
may	O
be	O
highly	O
dependent	O
on	O
the	O
specific	O
problem	O
at	O
hand	O
.	O
</s>
<s>
There	O
is	O
no	O
universal	O
or	O
exact	O
definition	O
of	O
what	O
constitutes	O
a	O
feature	B-Algorithm
,	O
and	O
the	O
exact	O
definition	O
often	O
depends	O
on	O
the	O
problem	O
or	O
the	O
type	O
of	O
application	O
.	O
</s>
<s>
Nevertheless	O
,	O
a	O
feature	B-Algorithm
is	O
typically	O
defined	O
as	O
an	O
"	O
interesting	O
"	O
part	O
of	O
an	O
image	B-Algorithm
,	O
and	O
features	O
are	O
used	O
as	O
a	O
starting	O
point	O
for	O
many	O
computer	B-Application
vision	I-Application
algorithms	O
.	O
</s>
<s>
Since	O
features	O
are	O
used	O
as	O
the	O
starting	O
point	O
and	O
main	O
primitives	O
for	O
subsequent	O
algorithms	O
,	O
the	O
overall	O
algorithm	O
will	O
often	O
only	O
be	O
as	O
good	O
as	O
its	O
feature	B-Algorithm
detector	O
.	O
</s>
<s>
Consequently	O
,	O
the	O
desirable	O
property	O
for	O
a	O
feature	B-Algorithm
detector	O
is	O
repeatability	O
:	O
whether	O
or	O
not	O
the	O
same	O
feature	B-Algorithm
will	O
be	O
detected	O
in	O
two	O
or	O
more	O
different	O
images	O
of	O
the	O
same	O
scene	O
.	O
</s>
<s>
Feature	B-Algorithm
detection	O
is	O
a	O
low-level	O
image	B-Algorithm
processing	I-Algorithm
operation	O
.	O
</s>
<s>
That	O
is	O
,	O
it	O
is	O
usually	O
performed	O
as	O
the	O
first	O
operation	O
on	O
an	O
image	B-Algorithm
,	O
and	O
examines	O
every	O
pixel	B-Algorithm
to	O
see	O
if	O
there	O
is	O
a	O
feature	B-Algorithm
present	O
at	O
that	O
pixel	B-Algorithm
.	O
</s>
<s>
If	O
this	O
is	O
part	O
of	O
a	O
larger	O
algorithm	O
,	O
then	O
the	O
algorithm	O
will	O
typically	O
only	O
examine	O
the	O
image	B-Algorithm
in	O
the	O
region	O
of	O
the	O
features	O
.	O
</s>
<s>
As	O
a	O
built-in	O
pre-requisite	O
to	O
feature	B-Algorithm
detection	O
,	O
the	O
input	O
image	B-Algorithm
is	O
usually	O
smoothed	O
by	O
a	O
Gaussian	B-Error_Name
kernel	O
in	O
a	O
scale-space	B-Algorithm
representation	I-Algorithm
and	O
one	O
or	O
several	O
feature	B-Algorithm
images	O
are	O
computed	O
,	O
often	O
expressed	O
in	O
terms	O
of	O
local	O
image	B-Algorithm
derivative	I-Algorithm
operations	O
.	O
</s>
<s>
Occasionally	O
,	O
when	O
feature	B-Algorithm
detection	O
is	O
computationally	B-General_Concept
expensive	I-General_Concept
and	O
there	O
are	O
time	O
constraints	O
,	O
a	O
higher	O
level	O
algorithm	O
may	O
be	O
used	O
to	O
guide	O
the	O
feature	B-Algorithm
detection	O
stage	O
,	O
so	O
that	O
only	O
certain	O
parts	O
of	O
the	O
image	B-Algorithm
are	O
searched	O
for	O
features	O
.	O
</s>
<s>
There	O
are	O
many	O
computer	B-Application
vision	I-Application
algorithms	O
that	O
use	O
feature	B-Algorithm
detection	O
as	O
the	O
initial	O
step	O
,	O
so	O
as	O
a	O
result	O
,	O
a	O
very	O
large	O
number	O
of	O
feature	B-Algorithm
detectors	O
have	O
been	O
developed	O
.	O
</s>
<s>
These	O
vary	O
widely	O
in	O
the	O
kinds	O
of	O
feature	B-Algorithm
detected	O
,	O
the	O
computational	O
complexity	O
and	O
the	O
repeatability	O
.	O
</s>
<s>
When	O
features	O
are	O
defined	O
in	O
terms	O
of	O
local	O
neighborhood	B-Algorithm
operations	I-Algorithm
applied	O
to	O
an	O
image	B-Algorithm
,	O
a	O
procedure	O
commonly	O
referred	O
to	O
as	O
feature	B-Algorithm
extraction	O
,	O
one	O
can	O
distinguish	O
between	O
feature	B-Algorithm
detection	O
approaches	O
that	O
produce	O
local	O
decisions	O
whether	O
there	O
is	O
a	O
feature	B-Algorithm
of	O
a	O
given	O
type	O
at	O
a	O
given	O
image	B-Algorithm
point	O
or	O
not	O
,	O
and	O
those	O
who	O
produce	O
non-binary	O
data	O
as	O
result	O
.	O
</s>
<s>
Although	O
local	O
decisions	O
are	O
made	O
,	O
the	O
output	O
from	O
a	O
feature	B-Algorithm
detection	O
step	O
does	O
not	O
need	O
to	O
be	O
a	O
binary	O
image	B-Algorithm
.	O
</s>
<s>
The	O
result	O
is	O
often	O
represented	O
in	O
terms	O
of	O
sets	O
of	O
(	O
connected	O
or	O
unconnected	O
)	O
coordinates	O
of	O
the	O
image	B-Algorithm
points	O
where	O
features	O
have	O
been	O
detected	O
,	O
sometimes	O
with	O
subpixel	O
accuracy	O
.	O
</s>
<s>
When	O
feature	B-Algorithm
extraction	O
is	O
done	O
without	O
local	O
decision	O
making	O
,	O
the	O
result	O
is	O
often	O
referred	O
to	O
as	O
a	O
feature	B-Algorithm
image	B-Algorithm
.	O
</s>
<s>
Consequently	O
,	O
a	O
feature	B-Algorithm
image	B-Algorithm
can	O
be	O
seen	O
as	O
an	O
image	B-Algorithm
in	O
the	O
sense	O
that	O
it	O
is	O
a	O
function	O
of	O
the	O
same	O
spatial	O
(	O
or	O
temporal	O
)	O
variables	O
as	O
the	O
original	O
image	B-Algorithm
,	O
but	O
where	O
the	O
pixel	B-Algorithm
values	O
hold	O
information	O
about	O
image	B-Algorithm
features	I-Algorithm
instead	O
of	O
intensity	O
or	O
color	O
.	O
</s>
<s>
This	O
means	O
that	O
a	O
feature	B-Algorithm
image	B-Algorithm
can	O
be	O
processed	O
in	O
a	O
similar	O
way	O
as	O
an	O
ordinary	O
image	B-Algorithm
generated	O
by	O
an	O
image	B-Algorithm
sensor	O
.	O
</s>
<s>
Feature	B-Algorithm
images	O
are	O
also	O
often	O
computed	O
as	O
integrated	O
step	O
in	O
algorithms	O
for	O
feature	B-Algorithm
detection	O
.	O
</s>
<s>
In	O
some	O
applications	O
,	O
it	O
is	O
not	O
sufficient	O
to	O
extract	O
only	O
one	O
type	O
of	O
feature	B-Algorithm
to	O
obtain	O
the	O
relevant	O
information	O
from	O
the	O
image	B-Algorithm
data	O
.	O
</s>
<s>
Instead	O
two	O
or	O
more	O
different	O
features	O
are	O
extracted	O
,	O
resulting	O
in	O
two	O
or	O
more	O
feature	B-Algorithm
descriptors	O
at	O
each	O
image	B-Algorithm
point	O
.	O
</s>
<s>
A	O
common	O
practice	O
is	O
to	O
organize	O
the	O
information	O
provided	O
by	O
all	O
these	O
descriptors	O
as	O
the	O
elements	O
of	O
one	O
single	O
vector	O
,	O
commonly	O
referred	O
to	O
as	O
a	O
feature	B-Algorithm
vector	I-Algorithm
.	O
</s>
<s>
The	O
set	O
of	O
all	O
possible	O
feature	B-Algorithm
vectors	I-Algorithm
constitutes	O
a	O
feature	B-Algorithm
space	O
.	O
</s>
<s>
A	O
common	O
example	O
of	O
feature	B-Algorithm
vectors	I-Algorithm
appears	O
when	O
each	O
image	B-Algorithm
point	O
is	O
to	O
be	O
classified	O
as	O
belonging	O
to	O
a	O
specific	O
class	O
.	O
</s>
<s>
Assuming	O
that	O
each	O
image	B-Algorithm
point	O
has	O
a	O
corresponding	O
feature	B-Algorithm
vector	I-Algorithm
based	O
on	O
a	O
suitable	O
set	O
of	O
features	O
,	O
meaning	O
that	O
each	O
class	O
is	O
well	O
separated	O
in	O
the	O
corresponding	O
feature	B-Algorithm
space	O
,	O
the	O
classification	B-General_Concept
of	O
each	O
image	B-Algorithm
point	O
can	O
be	O
done	O
using	O
standard	O
classification	B-General_Concept
method	O
.	O
</s>
<s>
The	O
input	O
data	O
fed	O
to	O
the	O
neural	B-Architecture
network	I-Architecture
is	O
often	O
given	O
in	O
terms	O
of	O
a	O
feature	B-Algorithm
vector	I-Algorithm
from	O
each	O
image	B-Algorithm
point	O
,	O
where	O
the	O
vector	O
is	O
constructed	O
from	O
several	O
different	O
features	O
extracted	O
from	O
the	O
image	B-Algorithm
data	O
.	O
</s>
<s>
Edges	O
are	O
points	O
where	O
there	O
is	O
a	O
boundary	O
(	O
or	O
an	O
edge	B-Algorithm
)	O
between	O
two	O
image	B-Algorithm
regions	O
.	O
</s>
<s>
In	O
general	O
,	O
an	O
edge	B-Algorithm
can	O
be	O
of	O
almost	O
arbitrary	O
shape	O
,	O
and	O
may	O
include	O
junctions	O
.	O
</s>
<s>
In	O
practice	O
,	O
edges	O
are	O
usually	O
defined	O
as	O
sets	O
of	O
points	O
in	O
the	O
image	B-Algorithm
which	O
have	O
a	O
strong	O
gradient	O
magnitude	O
.	O
</s>
<s>
Furthermore	O
,	O
some	O
common	O
algorithms	O
will	O
then	O
chain	O
high	O
gradient	O
points	O
together	O
to	O
form	O
a	O
more	O
complete	O
description	O
of	O
an	O
edge	B-Algorithm
.	O
</s>
<s>
These	O
algorithms	O
usually	O
place	O
some	O
constraints	O
on	O
the	O
properties	O
of	O
an	O
edge	B-Algorithm
,	O
such	O
as	O
shape	O
,	O
smoothness	O
,	O
and	O
gradient	O
value	O
.	O
</s>
<s>
The	O
terms	O
corners	O
and	O
interest	O
points	O
are	O
used	O
somewhat	O
interchangeably	O
and	O
refer	O
to	O
point-like	O
features	O
in	O
an	O
image	B-Algorithm
,	O
which	O
have	O
a	O
local	O
two	O
dimensional	O
structure	O
.	O
</s>
<s>
The	O
name	O
"	O
Corner	B-Algorithm
"	O
arose	O
since	O
early	O
algorithms	O
first	O
performed	O
edge	B-Algorithm
detection	I-Algorithm
,	O
and	O
then	O
analysed	O
the	O
edges	O
to	O
find	O
rapid	O
changes	O
in	O
direction	O
(	O
corners	O
)	O
.	O
</s>
<s>
These	O
algorithms	O
were	O
then	O
developed	O
so	O
that	O
explicit	O
edge	B-Algorithm
detection	I-Algorithm
was	O
no	O
longer	O
required	O
,	O
for	O
instance	O
by	O
looking	O
for	O
high	O
levels	O
of	O
curvature	O
in	O
the	O
image	B-Algorithm
gradient	I-Algorithm
.	O
</s>
<s>
It	O
was	O
then	O
noticed	O
that	O
the	O
so-called	O
corners	O
were	O
also	O
being	O
detected	O
on	O
parts	O
of	O
the	O
image	B-Algorithm
which	O
were	O
not	O
corners	O
in	O
the	O
traditional	O
sense	O
(	O
for	O
instance	O
a	O
small	O
bright	O
spot	O
on	O
a	O
dark	O
background	O
may	O
be	O
detected	O
)	O
.	O
</s>
<s>
These	O
points	O
are	O
frequently	O
known	O
as	O
interest	O
points	O
,	O
but	O
the	O
term	O
"	O
corner	B-Algorithm
"	O
is	O
used	O
by	O
tradition	O
.	O
</s>
<s>
Blobs	O
provide	O
a	O
complementary	O
description	O
of	O
image	B-Algorithm
structures	O
in	O
terms	O
of	O
regions	O
,	O
as	O
opposed	O
to	O
corners	O
that	O
are	O
more	O
point-like	O
.	O
</s>
<s>
Nevertheless	O
,	O
blob	B-Algorithm
descriptors	O
may	O
often	O
contain	O
a	O
preferred	O
point	O
(	O
a	O
local	O
maximum	O
of	O
an	O
operator	O
response	O
or	O
a	O
center	O
of	O
gravity	O
)	O
which	O
means	O
that	O
many	O
blob	B-Algorithm
detectors	I-Algorithm
may	O
also	O
be	O
regarded	O
as	O
interest	O
point	O
operators	O
.	O
</s>
<s>
Blob	B-Algorithm
detectors	I-Algorithm
can	O
detect	O
areas	O
in	O
an	O
image	B-Algorithm
which	O
are	O
too	O
smooth	O
to	O
be	O
detected	O
by	O
a	O
corner	B-Algorithm
detector	I-Algorithm
.	O
</s>
<s>
Consider	O
shrinking	O
an	O
image	B-Algorithm
and	O
then	O
performing	O
corner	B-Algorithm
detection	I-Algorithm
.	O
</s>
<s>
The	O
detector	O
will	O
respond	O
to	O
points	O
which	O
are	O
sharp	O
in	O
the	O
shrunk	O
image	B-Algorithm
,	O
but	O
may	O
be	O
smooth	O
in	O
the	O
original	O
image	B-Algorithm
.	O
</s>
<s>
It	O
is	O
at	O
this	O
point	O
that	O
the	O
difference	O
between	O
a	O
corner	B-Algorithm
detector	I-Algorithm
and	O
a	O
blob	B-Algorithm
detector	I-Algorithm
becomes	O
somewhat	O
vague	O
.	O
</s>
<s>
Nevertheless	O
,	O
due	O
to	O
their	O
response	O
properties	O
to	O
different	O
types	O
of	O
image	B-Algorithm
structures	O
at	O
different	O
scales	O
,	O
the	O
LoG	O
and	O
DoH	O
blob	B-Algorithm
detectors	I-Algorithm
are	O
also	O
mentioned	O
in	O
the	O
article	O
on	O
corner	B-Algorithm
detection	I-Algorithm
.	O
</s>
<s>
A	O
ridge	O
descriptor	O
computed	O
from	O
a	O
grey-level	O
image	B-Algorithm
can	O
be	O
seen	O
as	O
a	O
generalization	O
of	O
a	O
medial	O
axis	O
.	O
</s>
<s>
Unfortunately	O
,	O
however	O
,	O
it	O
is	O
algorithmically	O
harder	O
to	O
extract	O
ridge	O
features	O
from	O
general	O
classes	O
of	O
grey-level	O
images	O
than	O
edge-	O
,	O
corner	B-Algorithm
-	O
or	O
blob	B-Algorithm
features	O
.	O
</s>
<s>
Feature	B-Algorithm
detection	O
includes	O
methods	O
for	O
computing	O
abstractions	O
of	O
image	B-Algorithm
information	O
and	O
making	O
local	O
decisions	O
at	O
every	O
image	B-Algorithm
point	O
whether	O
there	O
is	O
an	O
image	B-Algorithm
feature	I-Algorithm
of	O
a	O
given	O
type	O
at	O
that	O
point	O
or	O
not	O
.	O
</s>
<s>
The	O
resulting	O
features	O
will	O
be	O
subsets	O
of	O
the	O
image	B-Algorithm
domain	O
,	O
often	O
in	O
the	O
form	O
of	O
isolated	O
points	O
,	O
continuous	O
curves	O
or	O
connected	O
regions	O
.	O
</s>
<s>
One	O
of	O
these	O
methods	O
is	O
the	O
scale-invariant	B-Algorithm
feature	I-Algorithm
transform	I-Algorithm
(	O
SIFT	O
)	O
.	O
</s>
<s>
+	O
Common	O
feature	B-Algorithm
detectors	O
and	O
their	O
classification:Feature	O
detectorEdgeCornerBlobRidge	O
Canny	B-Algorithm
Sobel	B-Algorithm
Harris	B-General_Concept
&	I-General_Concept
Stephens	I-General_Concept
/	I-General_Concept
Plessey	I-General_Concept
SUSAN	O
Shi	O
&	O
Tomasi	O
Level	O
curve	O
curvature	O
FAST	B-Algorithm
Laplacian	O
of	O
Gaussian	B-Error_Name
Difference	O
of	O
GaussiansJ	O
.	O
</s>
<s>
Determinant	O
of	O
Hessian	O
Hessian	O
strength	O
feature	B-Algorithm
measuresT	O
.	O
</s>
<s>
Lindeberg	O
"	O
Scale	O
selection	O
properties	O
of	O
generalized	O
scale-space	B-Algorithm
interest	O
point	O
detectors	O
"	O
,	O
Journal	O
of	O
Mathematical	O
Imaging	O
and	O
Vision	O
,	O
Volume	O
46	O
,	O
Issue	O
2	O
,	O
pages	O
177-210	O
,	O
2013.T	O
.	O
</s>
<s>
Lindeberg	O
``Image	O
matching	O
using	O
generalized	O
scale-space	B-Algorithm
interest	O
points	O
"	O
,	O
Journal	O
of	O
Mathematical	O
Imaging	O
and	O
Vision	O
,	O
volume	O
52	O
,	O
number	O
1	O
,	O
pages	O
3-36	O
,	O
2015	O
.	O
</s>
<s>
MSER	B-Algorithm
Principal	O
curvature	O
ridgesR	O
.	O
</s>
<s>
Haralick	O
,	O
"	O
Ridges	O
and	O
Valleys	O
on	O
Digital	B-Algorithm
Images	I-Algorithm
"	O
,	O
Computer	B-Application
Vision	I-Application
,	O
Graphics	O
,	O
and	O
Image	B-Algorithm
Processing	I-Algorithm
vol	O
.	O
</s>
<s>
D	O
.	O
Eberly	O
,	O
R	O
.	O
Gardner	O
,	O
B	O
.	O
Morse	O
,	O
S	O
.	O
Pizer	O
,	O
C	O
.	O
Scharlach	O
,	O
Ridges	O
for	O
image	B-Algorithm
analysis	O
,	O
Journal	O
of	O
Mathematical	O
Imaging	O
and	O
Vision	O
,	O
v	O
.	O
4	O
n	O
.	O
4	O
,	O
pp	O
.	O
</s>
<s>
Once	O
features	O
have	O
been	O
detected	O
,	O
a	O
local	O
image	B-Algorithm
patch	O
around	O
the	O
feature	B-Algorithm
can	O
be	O
extracted	O
.	O
</s>
<s>
This	O
extraction	O
may	O
involve	O
quite	O
considerable	O
amounts	O
of	O
image	B-Algorithm
processing	I-Algorithm
.	O
</s>
<s>
The	O
result	O
is	O
known	O
as	O
a	O
feature	B-Algorithm
descriptor	O
or	O
feature	B-Algorithm
vector	I-Algorithm
.	O
</s>
<s>
Among	O
the	O
approaches	O
that	O
are	O
used	O
to	O
feature	B-Algorithm
description	O
,	O
one	O
can	O
mention	O
N-jets	B-Algorithm
and	O
local	O
histograms	O
(	O
see	O
scale-invariant	B-Algorithm
feature	I-Algorithm
transform	I-Algorithm
for	O
one	O
example	O
of	O
a	O
local	O
histogram	O
descriptor	O
)	O
.	O
</s>
<s>
In	O
addition	O
to	O
such	O
attribute	O
information	O
,	O
the	O
feature	B-Algorithm
detection	O
step	O
by	O
itself	O
may	O
also	O
provide	O
complementary	O
attributes	O
,	O
such	O
as	O
the	O
edge	B-Algorithm
orientation	O
and	O
gradient	O
magnitude	O
in	O
edge	B-Algorithm
detection	I-Algorithm
and	O
the	O
polarity	O
and	O
the	O
strength	O
of	O
the	O
blob	B-Algorithm
in	O
blob	B-Algorithm
detection	I-Algorithm
.	O
</s>
<s>
Edge	B-Algorithm
direction	O
,	O
changing	O
intensity	O
,	O
autocorrelation	O
.	O
</s>
<s>
Works	O
with	O
any	O
parameterizable	O
feature	B-Algorithm
(	O
class	O
variables	O
,	O
cluster	O
detection	O
,	O
etc	O
..	O
)	O
</s>
<s>
A	O
specific	O
image	B-Algorithm
feature	I-Algorithm
,	O
defined	O
in	O
terms	O
of	O
a	O
specific	O
structure	O
in	O
the	O
image	B-Algorithm
data	O
,	O
can	O
often	O
be	O
represented	O
in	O
different	O
ways	O
.	O
</s>
<s>
For	O
example	O
,	O
an	O
edge	B-Algorithm
can	O
be	O
represented	O
as	O
a	O
boolean	O
variable	O
in	O
each	O
image	B-Algorithm
point	O
that	O
describes	O
whether	O
an	O
edge	B-Algorithm
is	O
present	O
at	O
that	O
point	O
.	O
</s>
<s>
Alternatively	O
,	O
we	O
can	O
instead	O
use	O
a	O
representation	O
which	O
provides	O
a	O
certainty	O
measure	O
instead	O
of	O
a	O
boolean	O
statement	O
of	O
the	O
edge	B-Algorithm
's	O
existence	O
and	O
combine	O
this	O
with	O
information	O
about	O
the	O
orientation	O
of	O
the	O
edge	B-Algorithm
.	O
</s>
<s>
Similarly	O
,	O
the	O
color	O
of	O
a	O
specific	O
region	O
can	O
either	O
be	O
represented	O
in	O
terms	O
of	O
the	O
average	O
color	O
(	O
three	O
scalars	O
)	O
or	O
a	O
color	B-General_Concept
histogram	I-General_Concept
(	O
three	O
functions	O
)	O
.	O
</s>
<s>
When	O
a	O
computer	B-Application
vision	I-Application
system	O
or	O
computer	B-Application
vision	I-Application
algorithm	O
is	O
designed	O
the	O
choice	O
of	O
feature	B-Algorithm
representation	O
can	O
be	O
a	O
critical	O
issue	O
.	O
</s>
<s>
In	O
some	O
cases	O
,	O
a	O
higher	O
level	O
of	O
detail	O
in	O
the	O
description	O
of	O
a	O
feature	B-Algorithm
may	O
be	O
necessary	O
for	O
solving	O
the	O
problem	O
,	O
but	O
this	O
comes	O
at	O
the	O
cost	O
of	O
having	O
to	O
deal	O
with	O
more	O
data	O
and	O
more	O
demanding	O
processing	O
.	O
</s>
<s>
In	O
this	O
discussion	O
,	O
an	O
instance	O
of	O
a	O
feature	B-Algorithm
representation	O
is	O
referred	O
to	O
as	O
a	O
,	O
or	O
simply	O
descriptor	O
.	O
</s>
<s>
Two	O
examples	O
of	O
image	B-Algorithm
features	I-Algorithm
are	O
local	O
edge	B-Algorithm
orientation	O
and	O
local	O
velocity	O
in	O
an	O
image	B-Algorithm
sequence	O
.	O
</s>
<s>
In	O
the	O
case	O
of	O
orientation	O
,	O
the	O
value	O
of	O
this	O
feature	B-Algorithm
may	O
be	O
more	O
or	O
less	O
undefined	O
if	O
more	O
than	O
one	O
edge	B-Algorithm
are	O
present	O
in	O
the	O
corresponding	O
neighborhood	O
.	O
</s>
<s>
Local	O
velocity	O
is	O
undefined	O
if	O
the	O
corresponding	O
image	B-Algorithm
region	O
does	O
not	O
contain	O
any	O
spatial	O
variation	O
.	O
</s>
<s>
As	O
a	O
consequence	O
of	O
this	O
observation	O
,	O
it	O
may	O
be	O
relevant	O
to	O
use	O
a	O
feature	B-Algorithm
representation	O
which	O
includes	O
a	O
measure	O
of	O
certainty	O
or	O
confidence	O
related	O
to	O
the	O
statement	O
about	O
the	O
feature	B-Algorithm
value	O
.	O
</s>
<s>
Otherwise	O
,	O
it	O
is	O
a	O
typical	O
situation	O
that	O
the	O
same	O
descriptor	O
is	O
used	O
to	O
represent	O
feature	B-Algorithm
values	O
of	O
low	O
certainty	O
and	O
feature	B-Algorithm
values	O
close	O
to	O
zero	O
,	O
with	O
a	O
resulting	O
ambiguity	O
in	O
the	O
interpretation	O
of	O
this	O
descriptor	O
.	O
</s>
<s>
In	O
particular	O
,	O
if	O
a	O
featured	O
image	B-Algorithm
will	O
be	O
used	O
in	O
subsequent	O
processing	O
,	O
it	O
may	O
be	O
a	O
good	O
idea	O
to	O
employ	O
a	O
feature	B-Algorithm
representation	O
that	O
includes	O
information	O
about	O
certainty	O
or	O
confidence	O
.	O
</s>
<s>
This	O
enables	O
a	O
new	O
feature	B-Algorithm
descriptor	O
to	O
be	O
computed	O
from	O
several	O
descriptors	O
,	O
for	O
example	O
computed	O
at	O
the	O
same	O
image	B-Algorithm
point	O
but	O
at	O
different	O
scales	O
,	O
or	O
from	O
different	O
but	O
neighboring	O
points	O
,	O
in	O
terms	O
of	O
a	O
weighted	O
average	O
where	O
the	O
weights	O
are	O
derived	O
from	O
the	O
corresponding	O
certainties	O
.	O
</s>
<s>
In	O
the	O
simplest	O
case	O
,	O
the	O
corresponding	O
computation	O
can	O
be	O
implemented	O
as	O
a	O
low-pass	O
filtering	O
of	O
the	O
featured	O
image	B-Algorithm
.	O
</s>
<s>
The	O
resulting	O
feature	B-Algorithm
image	B-Algorithm
will	O
,	O
in	O
general	O
,	O
be	O
more	O
stable	O
to	O
noise	O
.	O
</s>
<s>
In	O
addition	O
to	O
having	O
certainty	O
measures	O
included	O
in	O
the	O
representation	O
,	O
the	O
representation	O
of	O
the	O
corresponding	O
feature	B-Algorithm
values	O
may	O
itself	O
be	O
suitable	O
for	O
an	O
averaging	O
operation	O
or	O
not	O
.	O
</s>
<s>
Most	O
feature	B-Algorithm
representations	O
can	O
be	O
averaged	O
in	O
practice	O
,	O
but	O
only	O
in	O
certain	O
cases	O
can	O
the	O
resulting	O
descriptor	O
be	O
given	O
a	O
correct	O
interpretation	O
in	O
terms	O
of	O
a	O
feature	B-Algorithm
value	O
.	O
</s>
<s>
For	O
example	O
,	O
if	O
the	O
orientation	O
of	O
an	O
edge	B-Algorithm
is	O
represented	O
in	O
terms	O
of	O
an	O
angle	O
,	O
this	O
representation	O
must	O
have	O
a	O
discontinuity	O
where	O
the	O
angle	O
wraps	O
from	O
its	O
maximal	O
value	O
to	O
its	O
minimal	O
value	O
.	O
</s>
<s>
There	O
are	O
other	O
representations	O
of	O
edge	B-Algorithm
orientation	O
,	O
such	O
as	O
the	O
structure	B-Algorithm
tensor	I-Algorithm
,	O
which	O
are	O
averageable	O
.	O
</s>
<s>
Another	O
example	O
relates	O
to	O
motion	O
,	O
where	O
in	O
some	O
cases	O
only	O
the	O
normal	O
velocity	O
relative	O
to	O
some	O
edge	B-Algorithm
can	O
be	O
extracted	O
.	O
</s>
<s>
Features	O
detected	O
in	O
each	O
image	B-Algorithm
can	O
be	O
matched	O
across	O
multiple	O
images	O
to	O
establish	O
corresponding	O
features	O
such	O
as	O
corresponding	O
points	O
.	O
</s>
<s>
The	O
algorithm	O
is	O
based	O
on	O
comparing	O
and	O
analyzing	O
point	O
correspondences	O
between	O
the	O
reference	O
image	B-Algorithm
and	O
the	O
target	O
image	B-Algorithm
.	O
</s>
<s>
If	O
any	O
part	O
of	O
the	O
cluttered	O
scene	O
shares	O
correspondences	O
greater	O
than	O
the	O
threshold	O
,	O
that	O
part	O
of	O
the	O
cluttered	O
scene	O
image	B-Algorithm
is	O
targeted	O
and	O
considered	O
to	O
include	O
the	O
reference	O
object	O
there	O
.	O
</s>
