<s>
The	O
scale-invariant	B-Algorithm
feature	I-Algorithm
transform	I-Algorithm
(	O
SIFT	O
)	O
is	O
a	O
computer	B-Application
vision	I-Application
algorithm	O
to	O
detect	O
,	O
describe	O
,	O
and	O
match	O
local	O
features	B-Algorithm
in	O
images	O
,	O
invented	O
by	O
David	O
Lowe	O
in	O
1999	O
.	O
</s>
<s>
Applications	O
include	O
object	O
recognition	O
,	O
robotic	O
mapping	O
and	O
navigation	O
,	O
image	B-Algorithm
stitching	I-Algorithm
,	O
3D	O
modeling	O
,	O
gesture	B-General_Concept
recognition	I-General_Concept
,	O
video	B-Operating_System
tracking	I-Operating_System
,	O
individual	O
identification	O
of	O
wildlife	O
and	O
match	O
moving	O
.	O
</s>
<s>
An	O
object	O
is	O
recognized	O
in	O
a	O
new	O
image	O
by	O
individually	O
comparing	O
each	O
feature	O
from	O
the	O
new	O
image	O
to	O
this	O
database	O
and	O
finding	O
candidate	O
matching	O
features	B-Algorithm
based	O
on	O
Euclidean	O
distance	O
of	O
their	O
feature	O
vectors	O
.	O
</s>
<s>
The	O
determination	O
of	O
consistent	O
clusters	O
is	O
performed	O
rapidly	O
by	O
using	O
an	O
efficient	O
hash	B-Algorithm
table	I-Algorithm
implementation	O
of	O
the	O
generalised	O
Hough	B-Algorithm
transform	I-Algorithm
.	O
</s>
<s>
Each	O
cluster	O
of	O
3	O
or	O
more	O
features	B-Algorithm
that	O
agree	O
on	O
an	O
object	O
and	O
its	O
pose	B-Architecture
is	O
then	O
subject	O
to	O
further	O
detailed	O
model	O
verification	O
and	O
subsequently	O
outliers	O
are	O
discarded	O
.	O
</s>
<s>
Finally	O
the	O
probability	O
that	O
a	O
particular	O
set	O
of	O
features	B-Algorithm
indicates	O
the	O
presence	O
of	O
an	O
object	O
is	O
computed	O
,	O
given	O
the	O
accuracy	O
of	O
fit	O
and	O
number	O
of	O
probable	O
false	O
matches	O
.	O
</s>
<s>
To	O
perform	O
reliable	O
recognition	O
,	O
it	O
is	O
important	O
that	O
the	O
features	B-Algorithm
extracted	O
from	O
the	O
training	O
image	O
be	O
detectable	O
even	O
under	O
changes	O
in	O
image	O
scale	O
,	O
noise	O
and	O
illumination	O
.	O
</s>
<s>
Another	O
important	O
characteristic	O
of	O
these	O
features	B-Algorithm
is	O
that	O
the	O
relative	O
positions	O
between	O
them	O
in	O
the	O
original	O
scene	O
should	O
n't	O
change	O
from	O
one	O
image	O
to	O
another	O
.	O
</s>
<s>
For	O
example	O
,	O
if	O
only	O
the	O
four	O
corners	O
of	O
a	O
door	O
were	O
used	O
as	O
features	B-Algorithm
,	O
they	O
would	O
work	O
regardless	O
of	O
the	O
door	O
's	O
position	O
;	O
but	O
if	O
points	O
in	O
the	O
frame	O
were	O
also	O
used	O
,	O
the	O
recognition	O
would	O
fail	O
if	O
the	O
door	O
is	O
opened	O
or	O
closed	O
.	O
</s>
<s>
Similarly	O
,	O
features	B-Algorithm
located	O
in	O
articulated	O
or	O
flexible	O
objects	O
would	O
typically	O
not	O
work	O
if	O
any	O
change	O
in	O
their	O
internal	O
geometry	O
happens	O
between	O
two	O
images	O
in	O
the	O
set	O
being	O
processed	O
.	O
</s>
<s>
However	O
,	O
in	O
practice	O
SIFT	O
detects	O
and	O
uses	O
a	O
much	O
larger	O
number	O
of	O
features	B-Algorithm
from	O
the	O
images	O
,	O
which	O
reduces	O
the	O
contribution	O
of	O
the	O
errors	O
caused	O
by	O
these	O
local	O
variations	O
in	O
the	O
average	O
error	O
of	O
all	O
feature	O
matching	O
errors	O
.	O
</s>
<s>
SIFT	O
can	O
robustly	O
identify	O
objects	O
even	O
among	O
clutter	O
and	O
under	O
partial	O
occlusion	O
,	O
because	O
the	O
SIFT	O
feature	O
descriptor	O
is	O
invariant	O
to	O
uniform	B-Algorithm
scaling	I-Algorithm
,	O
orientation	O
,	O
illumination	O
changes	O
,	O
and	O
partially	O
invariant	O
to	O
affine	B-Algorithm
distortion	I-Algorithm
.	O
</s>
<s>
The	O
detection	O
and	O
description	O
of	O
local	O
image	B-Algorithm
features	I-Algorithm
can	O
help	O
in	O
object	O
recognition	O
.	O
</s>
<s>
The	O
SIFT	O
features	B-Algorithm
are	O
local	O
and	O
based	O
on	O
the	O
appearance	O
of	O
the	O
object	O
at	O
particular	O
interest	O
points	O
,	O
and	O
are	O
invariant	O
to	O
image	O
scale	O
and	O
rotation	O
.	O
</s>
<s>
They	O
are	O
relatively	O
easy	O
to	O
match	O
against	O
a	O
(	O
large	O
)	O
database	O
of	O
local	O
features	B-Algorithm
but	O
,	O
however	O
,	O
the	O
high	O
dimensionality	O
can	O
be	O
an	O
issue	O
,	O
and	O
generally	O
probabilistic	O
algorithms	O
such	O
as	O
k-d	B-Data_Structure
trees	I-Data_Structure
with	O
best	B-Algorithm
bin	I-Algorithm
first	I-Algorithm
search	O
are	O
used	O
.	O
</s>
<s>
Object	O
description	O
by	O
set	O
of	O
SIFT	O
features	B-Algorithm
is	O
also	O
robust	O
to	O
partial	O
occlusion	O
;	O
as	O
few	O
as	O
3	O
SIFT	O
features	B-Algorithm
from	O
an	O
object	O
are	O
enough	O
to	O
compute	O
its	O
location	O
and	O
pose	B-Architecture
.	O
</s>
<s>
Lowe	O
's	O
method	O
for	O
image	B-Algorithm
feature	I-Algorithm
generation	O
transforms	O
an	O
image	O
into	O
a	O
large	O
collection	O
of	O
feature	O
vectors	O
,	O
each	O
of	O
which	O
is	O
invariant	O
to	O
image	O
translation	O
,	O
scaling	B-Algorithm
,	O
and	O
rotation	O
,	O
partially	O
invariant	O
to	O
illumination	O
changes	O
,	O
and	O
robust	O
to	O
local	O
geometric	O
distortion	O
.	O
</s>
<s>
These	O
features	B-Algorithm
share	O
similar	O
properties	O
with	O
neurons	O
in	O
the	O
primary	O
visual	O
cortex	O
that	O
encode	O
basic	O
forms	O
,	O
color	O
,	O
and	O
movement	O
for	O
object	O
detection	O
in	O
primate	O
vision	O
.	O
</s>
<s>
Key	O
locations	O
are	O
defined	O
as	O
maxima	O
and	O
minima	O
of	O
the	O
result	O
of	O
difference	B-Algorithm
of	I-Algorithm
Gaussians	I-Algorithm
function	O
applied	O
in	O
scale	B-Algorithm
space	I-Algorithm
to	O
a	O
series	O
of	O
smoothed	O
and	O
resampled	O
images	O
.	O
</s>
<s>
SIFT	O
descriptors	O
robust	O
to	O
local	O
affine	B-Algorithm
distortion	I-Algorithm
are	O
then	O
obtained	O
by	O
considering	O
pixels	O
around	O
a	O
radius	O
of	O
the	O
key	O
location	O
,	O
blurring	O
,	O
and	O
resampling	O
local	O
image	O
orientation	O
planes	O
.	O
</s>
<s>
Lowe	O
used	O
a	O
modification	O
of	O
the	O
k-d	B-Data_Structure
tree	I-Data_Structure
algorithm	O
called	O
the	O
best-bin-first	B-Algorithm
search	I-Algorithm
method	O
that	O
can	O
identify	O
the	O
nearest	B-Algorithm
neighbors	I-Algorithm
with	O
high	O
probability	O
using	O
only	O
a	O
limited	O
amount	O
of	O
computation	O
.	O
</s>
<s>
The	O
BBF	O
algorithm	O
uses	O
a	O
modified	O
search	O
ordering	O
for	O
the	O
k-d	B-Data_Structure
tree	I-Data_Structure
algorithm	O
so	O
that	O
bins	O
in	O
feature	O
space	O
are	O
searched	O
in	O
the	O
order	O
of	O
their	O
closest	O
distance	O
from	O
the	O
query	O
location	O
.	O
</s>
<s>
This	O
search	O
order	O
requires	O
the	O
use	O
of	O
a	O
heap-based	O
priority	B-Application
queue	I-Application
for	O
efficient	O
determination	O
of	O
the	O
search	O
order	O
.	O
</s>
<s>
The	O
best	O
candidate	O
match	O
for	O
each	O
keypoint	O
is	O
found	O
by	O
identifying	O
its	O
nearest	B-Algorithm
neighbor	I-Algorithm
in	O
the	O
database	O
of	O
keypoints	O
from	O
training	O
images	O
.	O
</s>
<s>
The	O
nearest	B-Algorithm
neighbors	I-Algorithm
are	O
defined	O
as	O
the	O
keypoints	O
with	O
minimum	O
Euclidean	O
distance	O
from	O
the	O
given	O
descriptor	O
vector	O
.	O
</s>
<s>
To	O
further	O
improve	O
the	O
efficiency	O
of	O
the	O
best-bin-first	B-Algorithm
algorithm	O
search	O
was	O
cut	O
off	O
after	O
checking	O
the	O
first	O
200	O
nearest	B-Algorithm
neighbor	I-Algorithm
candidates	O
.	O
</s>
<s>
For	O
a	O
database	O
of	O
100,000	O
keypoints	O
,	O
this	O
provides	O
a	O
speedup	O
over	O
exact	O
nearest	B-Algorithm
neighbor	I-Algorithm
search	I-Algorithm
by	O
about	O
2	O
orders	O
of	O
magnitude	O
,	O
yet	O
results	O
in	O
less	O
than	O
a	O
5%	O
loss	O
in	O
the	O
number	O
of	O
correct	O
matches	O
.	O
</s>
<s>
Hough	B-Algorithm
transform	I-Algorithm
is	O
used	O
to	O
cluster	O
reliable	O
model	O
hypotheses	O
to	O
search	O
for	O
keys	O
that	O
agree	O
upon	O
a	O
particular	O
model	O
pose	B-Architecture
.	O
</s>
<s>
Hough	B-Algorithm
transform	I-Algorithm
identifies	O
clusters	O
of	O
features	B-Algorithm
with	O
a	O
consistent	O
interpretation	O
by	O
using	O
each	O
feature	O
to	O
vote	O
for	O
all	O
object	O
poses	B-Architecture
that	O
are	O
consistent	O
with	O
the	O
feature	O
.	O
</s>
<s>
When	O
clusters	O
of	O
features	B-Algorithm
are	O
found	O
to	O
vote	O
for	O
the	O
same	O
pose	B-Architecture
of	O
an	O
object	O
,	O
the	O
probability	O
of	O
the	O
interpretation	O
being	O
correct	O
is	O
much	O
higher	O
than	O
for	O
any	O
single	O
feature	O
.	O
</s>
<s>
An	O
entry	O
in	O
a	O
hash	B-Algorithm
table	I-Algorithm
is	O
created	O
predicting	O
the	O
model	O
location	O
,	O
orientation	O
,	O
and	O
scale	O
from	O
the	O
match	O
hypothesis	O
.	O
</s>
<s>
The	O
hash	B-Algorithm
table	I-Algorithm
is	O
searched	O
to	O
identify	O
all	O
clusters	O
of	O
at	O
least	O
3	O
entries	O
in	O
a	O
bin	O
,	O
and	O
the	O
bins	O
are	O
sorted	O
into	O
decreasing	O
order	O
of	O
size	O
.	O
</s>
<s>
The	O
similarity	O
transform	O
implied	O
by	O
these	O
4	O
parameters	O
is	O
only	O
an	O
approximation	O
to	O
the	O
full	O
6	O
degree-of-freedom	O
pose	B-Architecture
space	O
for	O
a	O
3D	O
object	O
and	O
also	O
does	O
not	O
account	O
for	O
any	O
non-rigid	O
deformations	O
.	O
</s>
<s>
To	O
avoid	O
the	O
problem	O
of	O
boundary	O
effects	O
in	O
bin	O
assignment	O
,	O
each	O
keypoint	O
match	O
votes	O
for	O
the	O
2	O
closest	O
bins	O
in	O
each	O
dimension	O
,	O
giving	O
a	O
total	O
of	O
16	O
entries	O
for	O
each	O
hypothesis	O
and	O
further	O
broadening	O
the	O
pose	B-Architecture
range	O
.	O
</s>
<s>
Each	O
identified	O
cluster	O
is	O
then	O
subject	O
to	O
a	O
verification	O
procedure	O
in	O
which	O
a	O
linear	B-Algorithm
least	I-Algorithm
squares	I-Algorithm
solution	O
is	O
performed	O
for	O
the	O
parameters	O
of	O
the	O
affine	B-Algorithm
transformation	I-Algorithm
relating	O
the	O
model	O
to	O
the	O
image	O
.	O
</s>
<s>
This	O
equation	O
shows	O
a	O
single	O
match	O
,	O
but	O
any	O
number	O
of	O
further	O
matches	O
can	O
be	O
added	O
,	O
with	O
each	O
match	O
contributing	O
two	O
more	O
rows	O
to	O
the	O
first	O
and	O
last	O
matrix	B-Architecture
.	O
</s>
<s>
where	O
A	O
is	O
a	O
known	O
m-by-n	O
matrix	B-Architecture
(	O
usually	O
with	O
m	O
>	O
n	O
)	O
,	O
x	O
is	O
an	O
unknown	O
n-dimensional	O
parameter	O
vector	O
,	O
and	O
b	O
is	O
a	O
known	O
m-dimensional	O
measurement	O
vector	O
.	O
</s>
<s>
Outliers	O
can	O
now	O
be	O
removed	O
by	O
checking	O
for	O
agreement	O
between	O
each	O
image	B-Algorithm
feature	I-Algorithm
and	O
the	O
model	O
,	O
given	O
the	O
parameter	O
solution	O
.	O
</s>
<s>
Given	O
the	O
linear	B-Algorithm
least	I-Algorithm
squares	I-Algorithm
solution	O
,	O
each	O
match	O
is	O
required	O
to	O
agree	O
within	O
half	O
the	O
error	O
range	O
that	O
was	O
used	O
for	O
the	O
parameters	O
in	O
the	O
Hough	B-Algorithm
transform	I-Algorithm
bins	O
.	O
</s>
<s>
As	O
outliers	O
are	O
discarded	O
,	O
the	O
linear	B-Algorithm
least	I-Algorithm
squares	I-Algorithm
solution	O
is	O
re-solved	O
with	O
the	O
remaining	O
points	O
,	O
and	O
the	O
process	O
iterated	O
.	O
</s>
<s>
In	O
addition	O
,	O
a	O
top-down	O
matching	O
phase	O
is	O
used	O
to	O
add	O
any	O
further	O
matches	O
that	O
agree	O
with	O
the	O
projected	O
model	O
position	O
,	O
which	O
may	O
have	O
been	O
missed	O
from	O
the	O
Hough	B-Algorithm
transform	I-Algorithm
bin	O
due	O
to	O
the	O
similarity	O
transform	O
approximation	O
or	O
other	O
errors	O
.	O
</s>
<s>
This	O
method	O
first	O
computes	O
the	O
expected	O
number	O
of	O
false	O
matches	O
to	O
the	O
model	O
pose	B-Architecture
,	O
given	O
the	O
projected	O
size	O
of	O
the	O
model	O
,	O
the	O
number	O
of	O
features	B-Algorithm
within	O
the	O
region	O
,	O
and	O
the	O
accuracy	O
of	O
the	O
fit	O
.	O
</s>
<s>
A	O
Bayesian	O
probability	O
analysis	O
then	O
gives	O
the	O
probability	O
that	O
the	O
object	O
is	O
present	O
based	O
on	O
the	O
actual	O
number	O
of	O
matching	O
features	B-Algorithm
found	O
.	O
</s>
<s>
The	O
image	O
is	O
convolved	B-Language
with	O
Gaussian	O
filters	O
at	O
different	O
scales	O
,	O
and	O
then	O
the	O
difference	O
of	O
successive	O
Gaussian-blurred	B-Error_Name
images	O
are	O
taken	O
.	O
</s>
<s>
Keypoints	O
are	O
then	O
taken	O
as	O
maxima/minima	O
of	O
the	O
Difference	B-Algorithm
of	I-Algorithm
Gaussians	I-Algorithm
(	O
DoG	O
)	O
that	O
occur	O
at	O
multiple	O
scales	O
.	O
</s>
<s>
where	O
is	O
the	O
convolution	B-Language
of	O
the	O
original	O
image	O
with	O
the	O
Gaussian	B-Error_Name
blur	I-Error_Name
at	O
scale	O
,	O
i.e.	O
,	O
</s>
<s>
Hence	O
a	O
DoG	O
image	O
between	O
scales	O
and	O
is	O
just	O
the	O
difference	O
of	O
the	O
Gaussian-blurred	B-Error_Name
images	O
at	O
scales	O
and	O
.	O
</s>
<s>
For	O
scale	B-Algorithm
space	I-Algorithm
extrema	O
detection	O
in	O
the	O
SIFT	O
algorithm	O
,	O
the	O
image	O
is	O
first	O
convolved	B-Language
with	O
Gaussian-blurs	O
at	O
different	O
scales	O
.	O
</s>
<s>
The	O
convolved	B-Language
images	O
are	O
grouped	O
by	O
octave	O
(	O
an	O
octave	O
corresponds	O
to	O
doubling	O
the	O
value	O
of	O
)	O
,	O
and	O
the	O
value	O
of	O
is	O
selected	O
so	O
that	O
we	O
obtain	O
a	O
fixed	O
number	O
of	O
convolved	B-Language
images	O
per	O
octave	O
.	O
</s>
<s>
Then	O
the	O
Difference-of-Gaussian	O
images	O
are	O
taken	O
from	O
adjacent	O
Gaussian-blurred	B-Error_Name
images	O
per	O
octave	O
.	O
</s>
<s>
This	O
keypoint	O
detection	O
step	O
is	O
a	O
variation	O
of	O
one	O
of	O
the	O
blob	B-Algorithm
detection	I-Algorithm
methods	O
developed	O
by	O
Lindeberg	O
by	O
detecting	O
scale-space	B-Algorithm
extrema	O
of	O
the	O
scale	O
normalized	O
Laplacian	O
;	O
that	O
is	O
,	O
detecting	O
points	O
that	O
are	O
local	O
extrema	O
with	O
respect	O
to	O
both	O
space	O
and	O
scale	O
,	O
in	O
the	O
discrete	O
case	O
by	O
comparisons	O
with	O
the	O
nearest	O
26	O
neighbors	O
in	O
a	O
discretized	O
scale-space	B-Algorithm
volume	O
.	O
</s>
<s>
The	O
difference	B-Algorithm
of	I-Algorithm
Gaussians	I-Algorithm
operator	O
can	O
be	O
seen	O
as	O
an	O
approximation	O
to	O
the	O
Laplacian	O
,	O
with	O
the	O
implicit	O
normalization	O
in	O
the	O
pyramid	B-Algorithm
also	O
constituting	O
a	O
discrete	O
approximation	O
of	O
the	O
scale-normalized	O
Laplacian	O
.	O
</s>
<s>
Another	O
real-time	O
implementation	O
of	O
scale-space	B-Algorithm
extrema	O
of	O
the	O
Laplacian	O
operator	O
has	O
been	O
presented	O
by	O
Lindeberg	O
and	O
Bretzner	O
based	O
on	O
a	O
hybrid	O
pyramid	B-Algorithm
representation	O
,	O
which	O
was	O
used	O
for	O
human-computer	O
interaction	O
by	O
real-time	O
gesture	B-General_Concept
recognition	I-General_Concept
in	O
Bretzner	O
et	O
al	O
.	O
</s>
<s>
Scale-space	B-Algorithm
extrema	O
detection	O
produces	O
too	O
many	O
keypoint	O
candidates	O
,	O
some	O
of	O
which	O
are	O
unstable	O
.	O
</s>
<s>
The	O
interpolation	O
is	O
done	O
using	O
the	O
quadratic	O
Taylor	O
expansion	O
of	O
the	O
Difference-of-Gaussian	O
scale-space	B-Algorithm
function	O
,	O
with	O
the	O
candidate	O
keypoint	O
as	O
the	O
origin	O
.	O
</s>
<s>
A	O
similar	O
subpixel	O
determination	O
of	O
the	O
locations	O
of	O
scale-space	B-Algorithm
extrema	O
is	O
performed	O
in	O
the	O
real-time	O
implementation	O
based	O
on	O
hybrid	O
pyramids	B-Algorithm
developed	O
by	O
Lindeberg	O
and	O
his	O
co-workers	O
.	O
</s>
<s>
Otherwise	O
it	O
is	O
kept	O
,	O
with	O
final	O
scale-space	B-Algorithm
location	O
,	O
where	O
is	O
the	O
original	O
location	O
of	O
the	O
keypoint	O
.	O
</s>
<s>
Finding	O
these	O
principal	O
curvatures	O
amounts	O
to	O
solving	O
for	O
the	O
eigenvalues	O
of	O
the	O
second-order	O
Hessian	O
matrix	B-Architecture
,	O
H	O
:	O
</s>
<s>
This	O
processing	O
step	O
for	O
suppressing	O
responses	O
at	O
edges	O
is	O
a	O
transfer	O
of	O
a	O
corresponding	O
approach	O
in	O
the	O
Harris	B-General_Concept
operator	I-General_Concept
for	O
corner	O
detection	O
.	O
</s>
<s>
The	O
difference	O
is	O
that	O
the	O
measure	O
for	O
thresholding	O
is	O
computed	O
from	O
the	O
Hessian	O
matrix	B-Architecture
instead	O
of	O
a	O
second-moment	B-Algorithm
matrix	I-Algorithm
.	O
</s>
<s>
The	O
magnitude	O
and	O
direction	O
calculations	O
for	O
the	O
gradient	O
are	O
done	O
for	O
every	O
pixel	O
in	O
a	O
neighboring	O
region	O
around	O
the	O
keypoint	O
in	O
the	O
Gaussian-blurred	B-Error_Name
image	O
L	O
.	O
An	O
orientation	O
histogram	O
with	O
36	O
bins	O
is	O
formed	O
,	O
with	O
each	O
bin	O
covering	O
10	O
degrees	O
.	O
</s>
<s>
The	O
image	O
gradient	O
magnitudes	O
and	O
orientations	O
are	O
sampled	O
around	O
the	O
keypoint	O
location	O
,	O
using	O
the	O
scale	O
of	O
the	O
keypoint	O
to	O
select	O
the	O
level	O
of	O
Gaussian	B-Error_Name
blur	I-Error_Name
for	O
the	O
image	O
.	O
</s>
<s>
128	O
,	O
seems	O
high	O
,	O
descriptors	O
with	O
lower	O
dimension	O
than	O
this	O
do	O
n't	O
perform	O
as	O
well	O
across	O
the	O
range	O
of	O
matching	O
tasks	O
and	O
the	O
computational	O
cost	O
remains	O
low	O
due	O
to	O
the	O
approximate	O
BBF	O
(	O
see	O
below	O
)	O
method	O
used	O
for	O
finding	O
the	O
nearest	B-Algorithm
neighbor	I-Algorithm
.	O
</s>
<s>
To	O
test	O
the	O
distinctiveness	O
of	O
the	O
SIFT	O
descriptors	O
,	O
matching	O
accuracy	O
is	O
also	O
measured	O
against	O
varying	O
number	O
of	O
keypoints	O
in	O
the	O
testing	O
database	O
,	O
and	O
it	O
is	O
shown	O
that	O
matching	O
accuracy	O
decreases	O
only	O
very	O
slightly	O
for	O
very	O
large	O
database	O
sizes	O
,	O
thus	O
indicating	O
that	O
SIFT	O
features	B-Algorithm
are	O
highly	O
distinctive	O
.	O
</s>
<s>
SIFT	O
and	O
SIFT-like	O
GLOH	B-Algorithm
features	B-Algorithm
exhibit	O
the	O
highest	O
matching	O
accuracies	O
(	O
recall	O
rates	O
)	O
for	O
an	O
affine	B-Algorithm
transformation	I-Algorithm
of	O
50	O
degrees	O
.	O
</s>
<s>
Distinctiveness	O
of	O
descriptors	O
is	O
measured	O
by	O
summing	O
the	O
eigenvalues	O
of	O
the	O
descriptors	O
,	O
obtained	O
by	O
the	O
Principal	B-Application
components	I-Application
analysis	I-Application
of	O
the	O
descriptors	O
normalized	O
by	O
their	O
variance	O
.	O
</s>
<s>
PCA-SIFT	B-Algorithm
(	O
Principal	B-Application
Components	I-Application
Analysis	I-Application
applied	O
to	O
SIFT	O
descriptors	O
)	O
,	O
GLOH	B-Algorithm
and	O
SIFT	O
features	B-Algorithm
give	O
the	O
highest	O
values	O
.	O
</s>
<s>
Introduction	O
of	O
blur	O
affects	O
all	O
local	O
descriptors	O
,	O
especially	O
those	O
based	O
on	O
edges	O
,	O
like	O
shape	B-General_Concept
context	I-General_Concept
,	O
because	O
edges	O
disappear	O
in	O
the	O
case	O
of	O
a	O
strong	O
blur	O
.	O
</s>
<s>
But	O
GLOH	B-Algorithm
,	O
PCA-SIFT	B-Algorithm
and	O
SIFT	O
still	O
performed	O
better	O
than	O
the	O
others	O
.	O
</s>
<s>
However	O
,	O
most	O
recent	O
feature	O
descriptors	O
such	O
as	O
SURF	B-Algorithm
have	O
not	O
been	O
evaluated	O
in	O
this	O
study	O
.	O
</s>
<s>
SURF	B-Algorithm
has	O
later	O
been	O
shown	O
to	O
have	O
similar	O
performance	O
to	O
SIFT	O
,	O
while	O
at	O
the	O
same	O
time	O
being	O
much	O
faster	O
.	O
</s>
<s>
Other	O
studies	O
conclude	O
that	O
when	O
speed	O
is	O
not	O
critical	O
,	O
SIFT	O
outperforms	O
SURF	B-Algorithm
.	O
</s>
<s>
Specifically	O
,	O
disregarding	O
discretization	O
effects	O
the	O
pure	O
image	O
descriptor	O
in	O
SIFT	O
is	O
significantly	O
better	O
than	O
the	O
pure	O
image	O
descriptor	O
in	O
SURF	B-Algorithm
,	O
whereas	O
the	O
scale-space	B-Algorithm
extrema	O
of	O
the	O
determinant	O
of	O
the	O
Hessian	O
underlying	O
the	O
pure	O
interest	O
point	O
detector	O
in	O
SURF	B-Algorithm
constitute	O
significantly	O
better	O
interest	O
points	O
compared	O
to	O
the	O
scale-space	B-Algorithm
extrema	O
of	O
the	O
Laplacian	O
to	O
which	O
the	O
interest	O
point	O
detector	O
in	O
SIFT	O
constitutes	O
a	O
numerical	O
approximation	O
.	O
</s>
<s>
The	O
performance	O
of	O
image	O
matching	O
by	O
SIFT	O
descriptors	O
can	O
be	O
improved	O
in	O
the	O
sense	O
of	O
achieving	O
higher	O
efficiency	O
scores	O
and	O
lower	O
1-precision	O
scores	O
by	O
replacing	O
the	O
scale-space	B-Algorithm
extrema	O
of	O
the	O
difference-of-Gaussians	O
operator	O
in	O
original	O
SIFT	O
by	O
scale-space	B-Algorithm
extrema	O
of	O
the	O
determinant	O
of	O
the	O
Hessian	O
,	O
or	O
more	O
generally	O
considering	O
a	O
more	O
general	O
family	O
of	O
generalized	O
scale-space	B-Algorithm
interest	O
points	O
.	O
</s>
<s>
The	O
Euclidean	O
distance	O
between	O
SIFT-Rank	O
descriptors	O
is	O
invariant	O
to	O
arbitrary	O
monotonic	O
changes	O
in	O
histogram	O
bin	O
values	O
,	O
and	O
is	O
related	O
to	O
Spearman	B-General_Concept
's	I-General_Concept
rank	I-General_Concept
correlation	I-General_Concept
coefficient	I-General_Concept
.	O
</s>
<s>
Given	O
SIFT	O
's	O
ability	O
to	O
find	O
distinctive	O
keypoints	O
that	O
are	O
invariant	O
to	O
location	O
,	O
scale	O
and	O
rotation	O
,	O
and	O
robust	O
to	O
affine	B-Algorithm
transformations	I-Algorithm
(	O
changes	O
in	O
scale	O
,	O
rotation	O
,	O
shear	B-Algorithm
,	O
and	O
position	O
)	O
and	O
changes	O
in	O
illumination	O
,	O
they	O
are	O
usable	O
for	O
object	O
recognition	O
.	O
</s>
<s>
First	O
,	O
SIFT	O
features	B-Algorithm
are	O
obtained	O
from	O
the	O
input	O
image	O
using	O
the	O
algorithm	O
described	O
above	O
.	O
</s>
<s>
These	O
features	B-Algorithm
are	O
matched	O
to	O
the	O
SIFT	O
feature	O
database	O
obtained	O
from	O
the	O
training	O
images	O
.	O
</s>
<s>
This	O
feature	O
matching	O
is	O
done	O
through	O
a	O
Euclidean-distance	O
based	O
nearest	B-Algorithm
neighbor	I-Algorithm
approach	O
.	O
</s>
<s>
To	O
increase	O
robustness	O
,	O
matches	O
are	O
rejected	O
for	O
those	O
keypoints	O
for	O
which	O
the	O
ratio	O
of	O
the	O
nearest	B-Algorithm
neighbor	I-Algorithm
distance	O
to	O
the	O
second-nearest	O
neighbor	O
distance	O
is	O
greater	O
than	O
0.8	O
.	O
</s>
<s>
Finally	O
,	O
to	O
avoid	O
the	O
expensive	O
search	O
required	O
for	O
finding	O
the	O
Euclidean-distance-based	O
nearest	B-Algorithm
neighbor	I-Algorithm
,	O
an	O
approximate	O
algorithm	O
called	O
the	O
best-bin-first	B-Algorithm
algorithm	O
is	O
used	O
.	O
</s>
<s>
This	O
is	O
a	O
fast	B-Algorithm
method	O
for	O
returning	O
the	O
nearest	B-Algorithm
neighbor	I-Algorithm
with	O
high	O
probability	O
,	O
and	O
can	O
give	O
speedup	O
by	O
factor	O
of	O
1000	O
while	O
finding	O
nearest	B-Algorithm
neighbor	I-Algorithm
(	O
of	O
interest	O
)	O
95%	O
of	O
the	O
time	O
.	O
</s>
<s>
Therefore	O
,	O
to	O
increase	O
robustness	O
to	O
object	O
identification	O
,	O
we	O
want	O
to	O
cluster	O
those	O
features	B-Algorithm
that	O
belong	O
to	O
the	O
same	O
object	O
and	O
reject	O
the	O
matches	O
that	O
are	O
left	O
out	O
in	O
the	O
clustering	O
process	O
.	O
</s>
<s>
This	O
is	O
done	O
using	O
the	O
Hough	B-Algorithm
transform	I-Algorithm
.	O
</s>
<s>
This	O
will	O
identify	O
clusters	O
of	O
features	B-Algorithm
that	O
vote	O
for	O
the	O
same	O
object	O
pose	B-Architecture
.	O
</s>
<s>
When	O
clusters	O
of	O
features	B-Algorithm
are	O
found	O
to	O
vote	O
for	O
the	O
same	O
pose	B-Architecture
of	O
an	O
object	O
,	O
the	O
probability	O
of	O
the	O
interpretation	O
being	O
correct	O
is	O
much	O
higher	O
than	O
for	O
any	O
single	O
feature	O
.	O
</s>
<s>
Each	O
keypoint	O
votes	O
for	O
the	O
set	O
of	O
object	O
poses	B-Architecture
that	O
are	O
consistent	O
with	O
the	O
keypoint	O
's	O
location	O
,	O
scale	O
,	O
and	O
orientation	O
.	O
</s>
<s>
Bins	O
that	O
accumulate	O
at	O
least	O
3	O
votes	O
are	O
identified	O
as	O
candidate	O
object/pose	O
matches	O
.	O
</s>
<s>
If	O
the	O
projection	O
of	O
a	O
keypoint	O
through	O
these	O
parameters	O
lies	O
within	O
half	O
the	O
error	O
range	O
that	O
was	O
used	O
for	O
the	O
parameters	O
in	O
the	O
Hough	B-Algorithm
transform	I-Algorithm
bins	O
,	O
the	O
keypoint	O
match	O
is	O
kept	O
.	O
</s>
<s>
This	O
works	O
better	O
for	O
planar	O
surface	O
recognition	O
than	O
3D	B-Algorithm
object	I-Algorithm
recognition	I-Algorithm
since	O
the	O
affine	O
model	O
is	O
no	O
longer	O
accurate	O
for	O
3D	O
objects	O
.	O
</s>
<s>
SIFT	O
features	B-Algorithm
can	O
essentially	O
be	O
applied	O
to	O
any	O
task	O
that	O
requires	O
identification	O
of	O
matching	O
locations	O
between	O
images	O
.	O
</s>
<s>
motion	O
tracking	O
and	O
segmentation	O
,	O
robot	O
localization	O
,	O
image	O
panorama	B-Algorithm
stitching	I-Algorithm
and	O
epipolar	B-Algorithm
calibration	B-Algorithm
.	O
</s>
<s>
As	O
the	O
robot	O
moves	O
,	O
it	O
localizes	O
itself	O
using	O
feature	O
matches	O
to	O
the	O
existing	O
3D	O
map	O
,	O
and	O
then	O
incrementally	O
adds	O
features	B-Algorithm
to	O
the	O
map	O
while	O
updating	O
their	O
3D	O
positions	O
using	O
a	O
Kalman	O
filter	O
.	O
</s>
<s>
Recent	O
3D	O
solvers	O
leverage	O
the	O
use	O
of	O
keypoint	O
directions	O
to	O
solve	O
trinocular	O
geometry	O
from	O
three	O
keypoints	O
and	O
absolute	O
pose	B-Architecture
from	O
only	O
two	O
keypoints	O
,	O
an	O
often	O
disregarded	O
but	O
useful	O
measurement	O
available	O
in	O
SIFT	O
.	O
</s>
<s>
SIFT	O
feature	O
matching	O
can	O
be	O
used	O
in	O
image	B-Algorithm
stitching	I-Algorithm
for	O
fully	O
automated	O
panorama	B-Application
reconstruction	O
from	O
non-panoramic	O
images	O
.	O
</s>
<s>
The	O
SIFT	O
features	B-Algorithm
extracted	O
from	O
the	O
input	O
images	O
are	O
matched	O
against	O
each	O
other	O
to	O
find	O
k	O
nearest-neighbors	O
for	O
each	O
feature	O
.	O
</s>
<s>
Homographies	B-Algorithm
between	O
pairs	O
of	O
images	O
are	O
then	O
computed	O
using	O
RANSAC	B-Algorithm
and	O
a	O
probabilistic	O
model	O
is	O
used	O
for	O
verification	O
.	O
</s>
<s>
Because	O
there	O
is	O
no	O
restriction	O
on	O
the	O
input	O
images	O
,	O
graph	O
search	O
is	O
applied	O
to	O
find	O
connected	O
components	O
of	O
image	O
matches	O
such	O
that	O
each	O
connected	O
component	O
will	O
correspond	O
to	O
a	O
panorama	B-Application
.	O
</s>
<s>
Finally	O
for	O
each	O
connected	O
component	O
bundle	B-General_Concept
adjustment	I-General_Concept
is	O
performed	O
to	O
solve	O
for	O
joint	O
camera	O
parameters	O
,	O
and	O
the	O
panorama	B-Application
is	O
rendered	O
using	O
multi-band	O
blending	O
.	O
</s>
<s>
Because	O
of	O
the	O
SIFT-inspired	O
object	O
recognition	O
approach	O
to	O
panorama	B-Algorithm
stitching	I-Algorithm
,	O
the	O
resulting	O
system	O
is	O
insensitive	O
to	O
the	O
ordering	O
,	O
orientation	O
,	O
scale	O
and	O
illumination	O
of	O
the	O
images	O
.	O
</s>
<s>
The	O
input	O
images	O
can	O
contain	O
multiple	O
panoramas	B-Application
and	O
noise	O
images	O
(	O
some	O
of	O
which	O
may	O
not	O
even	O
be	O
part	O
of	O
the	O
composite	O
image	O
)	O
,	O
and	O
panoramic	O
sequences	O
are	O
recognized	O
and	O
rendered	O
as	O
output	O
.	O
</s>
<s>
This	O
application	O
uses	O
SIFT	O
features	B-Algorithm
for	O
3D	B-Algorithm
object	I-Algorithm
recognition	I-Algorithm
and	O
3D	O
modeling	O
in	O
context	O
of	O
augmented	B-General_Concept
reality	I-General_Concept
,	O
in	O
which	O
synthetic	O
objects	O
with	O
accurate	O
pose	B-Architecture
are	O
superimposed	O
on	O
real	O
images	O
.	O
</s>
<s>
This	O
is	O
used	O
with	O
bundle	B-General_Concept
adjustment	I-General_Concept
initialized	O
from	O
an	O
essential	B-General_Concept
matrix	I-General_Concept
or	O
trifocal	B-Algorithm
tensor	I-Algorithm
to	O
build	O
a	O
sparse	O
3D	O
model	O
of	O
the	O
viewed	O
scene	O
and	O
to	O
simultaneously	O
recover	O
camera	B-Architecture
poses	I-Architecture
and	O
calibration	B-Algorithm
parameters	O
.	O
</s>
<s>
For	O
online	O
match	O
moving	O
,	O
SIFT	O
features	B-Algorithm
again	O
are	O
extracted	O
from	O
the	O
current	O
video	O
frame	O
and	O
matched	O
to	O
the	O
features	B-Algorithm
already	O
computed	O
for	O
the	O
world	O
model	O
,	O
resulting	O
in	O
a	O
set	O
of	O
2D-to-3D	O
correspondences	O
.	O
</s>
<s>
These	O
correspondences	O
are	O
then	O
used	O
to	O
compute	O
the	O
current	O
camera	B-Architecture
pose	I-Architecture
for	O
the	O
virtual	O
projection	O
and	O
final	O
rendering	O
.	O
</s>
<s>
3D	O
extensions	O
of	O
SIFT	O
have	O
also	O
been	O
evaluated	O
for	O
true	O
3D	B-Algorithm
object	I-Algorithm
recognition	I-Algorithm
and	O
retrieval	O
.	O
</s>
<s>
Extensions	O
of	O
the	O
SIFT	O
descriptor	O
to	O
2+	O
1-dimensional	O
spatio-temporal	O
data	O
in	O
context	O
of	O
human	B-Application
action	I-Application
recognition	I-Application
in	O
video	O
sequences	O
have	O
been	O
studied	O
.	O
</s>
<s>
The	O
computation	O
of	O
local	O
position-dependent	O
histograms	O
in	O
the	O
2D	O
SIFT	O
algorithm	O
are	O
extended	O
from	O
two	O
to	O
three	O
dimensions	O
to	O
describe	O
SIFT	O
features	B-Algorithm
in	O
a	O
spatio-temporal	O
domain	O
.	O
</s>
<s>
For	O
application	O
to	O
human	B-Application
action	I-Application
recognition	I-Application
in	O
a	O
video	O
sequence	O
,	O
sampling	O
of	O
the	O
training	O
videos	O
is	O
carried	O
out	O
either	O
at	O
spatio-temporal	O
interest	O
points	O
or	O
at	O
randomly	O
determined	O
locations	O
,	O
times	O
and	O
scales	O
.	O
</s>
<s>
These	O
descriptors	O
are	O
then	O
clustered	O
to	O
form	O
a	O
spatio-temporal	O
Bag	B-General_Concept
of	I-General_Concept
words	I-General_Concept
model	I-General_Concept
.	O
</s>
<s>
The	O
Feature-based	O
Morphometry	O
(	O
FBM	O
)	O
technique	O
uses	O
extrema	O
in	O
a	O
difference	B-Algorithm
of	I-Algorithm
Gaussian	I-Algorithm
scale-space	B-Algorithm
to	O
analyze	O
and	O
classify	O
3D	O
magnetic	O
resonance	O
images	O
(	O
MRIs	O
)	O
of	O
the	O
human	O
brain	O
.	O
</s>
<s>
FBM	O
models	O
the	O
image	O
probabilistically	O
as	O
a	O
collage	O
of	O
independent	O
features	B-Algorithm
,	O
conditional	O
on	O
image	O
geometry	O
and	O
group	O
labels	O
,	O
e.g.	O
</s>
<s>
Features	B-Algorithm
are	O
first	O
extracted	O
in	O
individual	O
images	O
from	O
a	O
4D	O
difference	B-Algorithm
of	I-Algorithm
Gaussian	I-Algorithm
scale-space	B-Algorithm
,	O
then	O
modeled	O
in	O
terms	O
of	O
their	O
appearance	O
,	O
geometry	O
and	O
group	O
co-occurrence	O
statistics	O
across	O
a	O
set	O
of	O
images	O
.	O
</s>
<s>
This	O
normalization	O
scheme	O
termed	O
“	O
L1-sqrt	O
”	O
was	O
previously	O
introduced	O
for	O
the	O
block	O
normalization	O
of	O
HOG	B-Algorithm
features	B-Algorithm
whose	O
rectangular	O
block	O
arrangement	O
descriptor	O
variant	O
(	O
R-HOG	O
)	O
is	O
conceptually	O
similar	O
to	O
the	O
SIFT	O
descriptor	O
.	O
</s>
<s>
G-RIF	O
:	O
Generalized	O
Robust	O
Invariant	O
Feature	O
is	O
a	O
general	O
context	O
descriptor	O
which	O
encodes	O
edge	O
orientation	O
,	O
edge	O
density	O
and	O
hue	B-Language
information	O
in	O
a	O
unified	O
form	O
combining	O
perceptual	O
information	O
with	O
spatial	O
encoding	O
.	O
</s>
<s>
"	O
SURF	B-Algorithm
:	O
Speeded	B-Algorithm
Up	I-Algorithm
Robust	I-Algorithm
Features	I-Algorithm
"	O
is	O
a	O
high-performance	O
scale	O
-	O
and	O
rotation-invariant	O
interest	O
point	O
detector	O
/	O
descriptor	O
claimed	O
to	O
approximate	O
or	O
even	O
outperform	O
previously	O
proposed	O
schemes	O
with	O
respect	O
to	O
repeatability	O
,	O
distinctiveness	O
,	O
and	O
robustness	O
.	O
</s>
<s>
SURF	B-Algorithm
relies	O
on	O
integral	B-Algorithm
images	I-Algorithm
for	O
image	O
convolutions	B-Language
to	O
reduce	O
computation	O
time	O
,	O
builds	O
on	O
the	O
strengths	O
of	O
the	O
leading	O
existing	O
detectors	O
and	O
descriptors	O
(	O
using	O
a	O
fast	B-Algorithm
Hessian	O
matrix-based	O
measure	O
for	O
the	O
detector	O
and	O
a	O
distribution-based	O
descriptor	O
)	O
.	O
</s>
<s>
Integral	B-Algorithm
images	I-Algorithm
are	O
used	O
for	O
speed	O
and	O
only	O
64	O
dimensions	O
are	O
used	O
reducing	O
the	O
time	O
for	O
feature	O
computation	O
and	O
matching	O
.	O
</s>
<s>
PCA-SIFT	B-Algorithm
and	O
GLOH	B-Algorithm
are	O
variants	O
of	O
SIFT	O
.	O
</s>
<s>
PCA-SIFT	B-Algorithm
descriptor	O
is	O
a	O
vector	O
of	O
image	O
gradients	O
in	O
x	O
and	O
y	O
direction	O
computed	O
within	O
the	O
support	O
region	O
.	O
</s>
<s>
The	O
dimension	O
is	O
reduced	O
to	O
36	O
with	O
PCA	B-Application
.	O
</s>
<s>
Gradient	O
location-orientation	O
histogram	O
(	O
GLOH	B-Algorithm
)	O
is	O
an	O
extension	O
of	O
the	O
SIFT	O
descriptor	O
designed	O
to	O
increase	O
its	O
robustness	O
and	O
distinctiveness	O
.	O
</s>
<s>
The	O
size	O
of	O
this	O
descriptor	O
is	O
reduced	O
with	O
PCA	B-Application
.	O
</s>
<s>
The	O
covariance	O
matrix	B-Architecture
for	O
PCA	B-Application
is	O
estimated	O
on	O
image	O
patches	O
collected	O
from	O
various	O
images	O
.	O
</s>
<s>
Gauss-SIFT	O
is	O
a	O
pure	O
image	O
descriptor	O
defined	O
by	O
performing	O
all	O
image	O
measurements	O
underlying	O
the	O
pure	O
image	O
descriptor	O
in	O
SIFT	O
by	O
Gaussian	O
derivative	O
responses	O
as	O
opposed	O
to	O
derivative	O
approximations	O
in	O
an	O
image	B-Algorithm
pyramid	I-Algorithm
as	O
done	O
in	O
regular	O
SIFT	O
.	O
</s>
<s>
In	O
Lindeberg	O
(	O
2015	O
)	O
such	O
pure	O
Gauss-SIFT	O
image	O
descriptors	O
were	O
combined	O
with	O
a	O
set	O
of	O
generalized	O
scale-space	B-Algorithm
interest	O
points	O
comprising	O
the	O
Laplacian	O
of	O
the	O
Gaussian	O
,	O
the	O
determinant	O
of	O
the	O
Hessian	O
,	O
four	O
new	O
unsigned	O
or	O
signed	O
Hessian	O
feature	O
strength	O
measures	O
as	O
well	O
as	O
Harris-Laplace	O
and	O
Shi-and-Tomasi	O
interests	O
points	O
.	O
</s>
<s>
In	O
an	O
extensive	O
experimental	O
evaluation	O
on	O
a	O
poster	O
dataset	O
comprising	O
multiple	O
views	O
of	O
12	O
posters	O
over	O
scaling	B-Algorithm
transformations	O
up	O
to	O
a	O
factor	O
of	O
6	O
and	O
viewing	O
direction	O
variations	O
up	O
to	O
a	O
slant	O
angle	O
of	O
45	O
degrees	O
,	O
it	O
was	O
shown	O
that	O
substantial	O
increase	O
in	O
performance	O
of	O
image	O
matching	O
(	O
higher	O
efficiency	O
scores	O
and	O
lower	O
1-precision	O
scores	O
)	O
could	O
be	O
obtained	O
by	O
replacing	O
Laplacian	O
of	O
Gaussian	O
interest	O
points	O
by	O
determinant	O
of	O
the	O
Hessian	O
interest	O
points	O
.	O
</s>
<s>
A	O
quantitative	O
comparison	O
between	O
the	O
Gauss-SIFT	O
descriptor	O
and	O
a	O
corresponding	O
Gauss-SURF	O
descriptor	O
did	O
also	O
show	O
that	O
Gauss-SIFT	O
does	O
generally	O
perform	O
significantly	O
better	O
than	O
Gauss-SURF	O
for	O
a	O
large	O
number	O
of	O
different	O
scale-space	B-Algorithm
interest	O
point	O
detectors	O
.	O
</s>
<s>
This	O
study	O
therefore	O
shows	O
that	O
discregarding	O
discretization	O
effects	O
the	O
pure	O
image	O
descriptor	O
in	O
SIFT	O
is	O
significantly	O
better	O
than	O
the	O
pure	O
image	O
descriptor	O
in	O
SURF	B-Algorithm
,	O
whereas	O
the	O
underlying	O
interest	O
point	O
detector	O
in	O
SURF	B-Algorithm
,	O
which	O
can	O
be	O
seen	O
as	O
numerical	O
approximation	O
to	O
scale-space	B-Algorithm
extrema	O
of	O
the	O
determinant	O
of	O
the	O
Hessian	O
,	O
is	O
significantly	O
better	O
than	O
the	O
underlying	O
interest	O
point	O
detector	O
in	O
SIFT	O
.	O
</s>
<s>
use	O
the	O
FAST	B-Algorithm
corner	O
detector	O
for	O
feature	O
detection	O
.	O
</s>
<s>
The	O
algorithm	O
also	O
distinguishes	O
between	O
the	O
off-line	O
preparation	O
phase	O
where	O
features	B-Algorithm
are	O
created	O
at	O
different	O
scale	O
levels	O
and	O
the	O
on-line	O
phase	O
where	O
features	B-Algorithm
are	O
only	O
created	O
at	O
the	O
current	O
fixed	O
scale	O
level	O
of	O
the	O
phone	O
's	O
camera	O
image	O
.	O
</s>
<s>
In	O
addition	O
,	O
features	B-Algorithm
are	O
created	O
from	O
a	O
fixed	O
patch	O
size	O
of	O
15×15	O
pixels	O
and	O
form	O
a	O
SIFT	O
descriptor	O
with	O
only	O
36	O
dimensions	O
.	O
</s>
<s>
The	O
approach	O
is	O
mainly	O
restricted	O
by	O
the	O
amount	O
of	O
available	O
RAM	B-Architecture
.	O
</s>
<s>
KAZE	O
and	O
A-KAZE	O
(	O
KAZE	O
Features	B-Algorithm
and	O
Accelerated-Kaze	O
Features	B-Algorithm
)	O
is	O
a	O
new	O
2D	O
feature	O
detection	O
and	O
description	O
method	O
that	O
perform	O
better	O
compared	O
to	O
SIFT	O
and	O
SURF	B-Algorithm
.	O
</s>
