<s>
The	O
histogram	B-Algorithm
of	I-Algorithm
oriented	I-Algorithm
gradients	I-Algorithm
(	O
HOG	O
)	O
is	O
a	O
feature	B-General_Concept
descriptor	I-General_Concept
used	O
in	O
computer	B-Application
vision	I-Application
and	O
image	B-Algorithm
processing	I-Algorithm
for	O
the	O
purpose	O
of	O
object	B-General_Concept
detection	I-General_Concept
.	O
</s>
<s>
This	O
method	O
is	O
similar	O
to	O
that	O
of	O
edge	O
orientation	O
histograms	O
,	O
scale-invariant	B-Algorithm
feature	I-Algorithm
transform	I-Algorithm
descriptors	O
,	O
and	O
shape	B-General_Concept
contexts	I-General_Concept
,	O
but	O
differs	O
in	O
that	O
it	O
is	O
computed	O
on	O
a	O
dense	O
grid	O
of	O
uniformly	O
spaced	O
cells	O
and	O
uses	O
overlapping	O
local	O
contrast	O
normalization	O
for	O
improved	O
accuracy	O
.	O
</s>
<s>
However	O
,	O
usage	O
only	O
became	O
widespread	O
in	O
2005	O
when	O
Navneet	O
Dalal	O
and	O
Bill	O
Triggs	O
,	O
researchers	O
for	O
the	O
French	O
National	O
Institute	O
for	O
Research	O
in	O
Computer	O
Science	O
and	O
Automation	O
(	O
INRIA	O
)	O
,	O
presented	O
their	O
supplementary	O
work	O
on	O
HOG	O
descriptors	O
at	O
the	O
Conference	O
on	O
Computer	B-Application
Vision	I-Application
and	O
Pattern	O
Recognition	O
(	O
CVPR	O
)	O
.	O
</s>
<s>
In	O
this	O
work	O
they	O
focused	O
on	O
pedestrian	B-General_Concept
detection	I-General_Concept
in	O
static	O
images	O
,	O
although	O
since	O
then	O
they	O
expanded	O
their	O
tests	O
to	O
include	O
human	O
detection	O
in	O
videos	O
,	O
as	O
well	O
as	O
to	O
a	O
variety	O
of	O
common	O
animals	O
and	O
vehicles	O
in	O
static	O
imagery	O
.	O
</s>
<s>
The	O
essential	O
thought	O
behind	O
the	O
histogram	B-Algorithm
of	I-Algorithm
oriented	I-Algorithm
gradients	I-Algorithm
descriptor	O
is	O
that	O
local	O
object	O
appearance	O
and	O
shape	O
within	O
an	O
image	O
can	O
be	O
described	O
by	O
the	O
distribution	O
of	O
intensity	O
gradients	O
or	O
edge	O
directions	O
.	O
</s>
<s>
The	O
most	O
common	O
method	O
is	O
to	O
apply	O
the	O
1-D	O
centered	O
,	O
point	O
discrete	O
derivative	B-Algorithm
mask	I-Algorithm
in	O
one	O
or	O
both	O
of	O
the	O
horizontal	O
and	O
vertical	O
directions	O
.	O
</s>
<s>
Dalal	O
and	O
Triggs	O
tested	O
other	O
,	O
more	O
complex	O
masks	O
,	O
such	O
as	O
the	O
3x3	O
Sobel	B-Algorithm
mask	I-Algorithm
or	O
diagonal	O
masks	O
,	O
but	O
these	O
masks	O
generally	O
performed	O
more	O
poorly	O
in	O
detecting	O
humans	O
in	O
images	O
.	O
</s>
<s>
They	O
also	O
experimented	O
with	O
Gaussian	B-Error_Name
smoothing	I-Error_Name
before	O
applying	O
the	O
derivative	B-Algorithm
mask	I-Algorithm
,	O
but	O
similarly	O
found	O
that	O
omission	O
of	O
any	O
smoothing	O
performed	O
better	O
in	O
practice	O
.	O
</s>
<s>
The	O
R-HOG	O
blocks	O
appear	O
quite	O
similar	O
to	O
the	O
scale-invariant	B-Algorithm
feature	I-Algorithm
transform	I-Algorithm
(	O
SIFT	B-Algorithm
)	O
descriptors	O
;	O
however	O
,	O
despite	O
their	O
similar	O
formation	O
,	O
R-HOG	O
blocks	O
are	O
computed	O
in	O
dense	O
grids	O
at	O
some	O
single	O
scale	O
without	O
orientation	O
alignment	O
,	O
whereas	O
SIFT	B-Algorithm
descriptors	O
are	O
usually	O
computed	O
at	O
sparse	O
,	O
scale-invariant	O
key	O
image	O
points	O
and	O
are	O
rotated	O
to	O
align	O
orientation	O
.	O
</s>
<s>
In	O
addition	O
,	O
the	O
R-HOG	O
blocks	O
are	O
used	O
in	O
conjunction	O
to	O
encode	O
spatial	O
form	O
information	O
,	O
while	O
SIFT	B-Algorithm
descriptors	O
are	O
used	O
singly	O
.	O
</s>
<s>
C-HOG	O
blocks	O
appear	O
similar	O
to	O
shape	B-General_Concept
context	I-General_Concept
descriptors	O
,	O
but	O
differ	O
strongly	O
in	O
that	O
C-HOG	O
blocks	O
contain	O
cells	O
with	O
several	O
orientation	O
channels	O
,	O
while	O
shape	B-General_Concept
contexts	I-General_Concept
only	O
make	O
use	O
of	O
a	O
single	O
edge	O
presence	O
count	O
in	O
their	O
formulation	O
.	O
</s>
<s>
Dalal	O
and	O
Triggs	O
used	O
HOG	O
descriptors	O
as	O
features	O
in	O
a	O
support	B-Algorithm
vector	I-Algorithm
machine	I-Algorithm
(	O
SVM	B-Algorithm
)	O
;	O
however	O
,	O
HOG	O
descriptors	O
are	O
not	O
tied	O
to	O
a	O
specific	O
machine	O
learning	O
algorithm	O
.	O
</s>
<s>
In	O
their	O
original	O
human	O
detection	O
experiment	O
,	O
Dalal	O
and	O
Triggs	O
compared	O
their	O
R-HOG	O
and	O
C-HOG	O
descriptor	O
blocks	O
against	O
generalized	O
Haar	O
wavelets	O
,	O
PCA-SIFT	B-Algorithm
descriptors	O
,	O
and	O
shape	B-General_Concept
context	I-General_Concept
descriptors	O
.	O
</s>
<s>
Generalized	O
Haar	O
wavelets	O
are	O
oriented	O
Haar	O
wavelets	O
,	O
and	O
were	O
used	O
in	O
2001	O
by	O
Mohan	O
,	O
Papageorgiou	O
,	O
and	O
Poggio	O
in	O
their	O
own	O
object	B-General_Concept
detection	I-General_Concept
experiments	O
.	O
</s>
<s>
PCA-SIFT	B-Algorithm
descriptors	O
are	O
similar	O
to	O
SIFT	B-Algorithm
descriptors	O
,	O
but	O
differ	O
in	O
that	O
principal	B-Application
component	I-Application
analysis	I-Application
is	O
applied	O
to	O
the	O
normalized	O
gradient	O
patches	O
.	O
</s>
<s>
PCA-SIFT	B-Algorithm
descriptors	O
were	O
first	O
used	O
in	O
2004	O
by	O
Ke	O
and	O
Sukthankar	O
and	O
were	O
claimed	O
to	O
outperform	O
regular	O
SIFT	B-Algorithm
descriptors	O
.	O
</s>
<s>
Finally	O
,	O
shape	B-General_Concept
contexts	I-General_Concept
use	O
circular	O
bins	O
,	O
similar	O
to	O
those	O
used	O
in	O
C-HOG	O
blocks	O
,	O
but	O
only	O
tabulate	O
votes	O
on	O
the	O
basis	O
of	O
edge	O
presence	O
,	O
making	O
no	O
distinction	O
with	O
regards	O
to	O
orientation	O
.	O
</s>
<s>
Shape	B-General_Concept
contexts	I-General_Concept
were	O
originally	O
used	O
in	O
2001	O
by	O
Belongie	O
,	O
Malik	O
,	O
and	O
Puzicha	O
.	O
</s>
<s>
As	O
for	O
the	O
results	O
,	O
the	O
C-HOG	O
and	O
R-HOG	O
block	O
descriptors	O
perform	O
comparably	O
,	O
with	O
the	O
C-HOG	O
descriptors	O
maintaining	O
a	O
slight	O
advantage	O
in	O
the	O
detection	O
miss	O
rate	O
at	O
fixed	O
false	B-General_Concept
positive	I-General_Concept
rates	I-General_Concept
across	O
both	O
data	O
sets	O
.	O
</s>
<s>
On	O
the	O
MIT	O
set	O
,	O
the	O
C-HOG	O
and	O
R-HOG	O
descriptors	O
produced	O
a	O
detection	O
miss	O
rate	O
of	O
essentially	O
zero	O
at	O
a	O
10−4	O
false	B-General_Concept
positive	I-General_Concept
rate	I-General_Concept
.	O
</s>
<s>
On	O
the	O
INRIA	O
set	O
,	O
the	O
C-HOG	O
and	O
R-HOG	O
descriptors	O
produced	O
a	O
detection	O
miss	O
rate	O
of	O
roughly	O
0.1	O
at	O
a	O
10−4	O
false	B-General_Concept
positive	I-General_Concept
rate	I-General_Concept
.	O
</s>
<s>
The	O
generalized	O
Haar	O
wavelets	O
represent	O
the	O
next	O
highest	O
performing	O
approach	O
:	O
they	O
produced	O
roughly	O
a	O
0.01	O
miss	O
rate	O
at	O
a	O
10−4	O
false	B-General_Concept
positive	I-General_Concept
rate	I-General_Concept
on	O
the	O
MIT	O
set	O
,	O
and	O
roughly	O
a	O
0.3	O
miss	O
rate	O
on	O
the	O
INRIA	O
set	O
.	O
</s>
<s>
The	O
PCA-SIFT	B-Algorithm
descriptors	O
and	O
shape	B-General_Concept
context	I-General_Concept
descriptors	O
both	O
performed	O
fairly	O
poorly	O
on	O
both	O
data	O
sets	O
.	O
</s>
<s>
Both	O
methods	O
produced	O
a	O
miss	O
rate	O
of	O
0.1	O
at	O
a	O
10−4	O
false	B-General_Concept
positive	I-General_Concept
rate	I-General_Concept
on	O
the	O
MIT	O
set	O
and	O
nearly	O
a	O
miss	O
rate	O
of	O
0.5	O
at	O
a	O
10−4	O
false	B-General_Concept
positive	I-General_Concept
rate	I-General_Concept
on	O
the	O
INRIA	O
set	O
.	O
</s>
<s>
As	O
part	O
of	O
the	O
Pascal	O
Visual	O
Object	O
Classes	O
2006	O
Workshop	O
,	O
Dalal	O
and	O
Triggs	O
presented	O
results	O
on	O
applying	O
histogram	B-Algorithm
of	I-Algorithm
oriented	I-Algorithm
gradients	I-Algorithm
descriptors	O
to	O
image	O
objects	O
other	O
than	O
humans	O
,	O
such	O
as	O
cars	O
,	O
buses	O
,	O
and	O
bicycles	O
,	O
as	O
well	O
as	O
common	O
animals	O
such	O
as	O
dogs	O
,	O
cats	O
,	O
and	O
cows	O
.	O
</s>
<s>
As	O
part	O
of	O
the	O
2006	O
European	O
Conference	O
on	O
Computer	B-Application
Vision	I-Application
(	O
ECCV	O
)	O
,	O
Dalal	O
and	O
Triggs	O
teamed	O
up	O
with	O
Cordelia	O
Schmid	O
to	O
apply	O
HOG	O
detectors	O
to	O
the	O
problem	O
of	O
human	O
detection	O
in	O
films	O
and	O
videos	O
.	O
</s>
<s>
When	O
testing	O
on	O
two	O
large	O
datasets	O
taken	O
from	O
several	O
movies	O
,	O
the	O
combined	O
HOG-IMH	O
method	O
yielded	O
a	O
miss	O
rate	O
of	O
approximately	O
0.1	O
at	O
a	O
false	B-General_Concept
positive	I-General_Concept
rate	I-General_Concept
.	O
</s>
<s>
At	O
the	O
Intelligent	O
Vehicles	O
Symposium	O
in	O
2006	O
,	O
F	O
.	O
Suard	O
,	O
A	O
.	O
Rakotomamonjy	O
,	O
and	O
A	O
.	O
Bensrhair	O
introduced	O
a	O
complete	O
system	O
for	O
pedestrian	B-General_Concept
detection	I-General_Concept
based	O
on	O
HOG	O
descriptors	O
.	O
</s>
<s>
Then	O
support	B-Algorithm
vector	I-Algorithm
machine	I-Algorithm
classifiers	O
operate	O
on	O
the	O
HOG	O
descriptors	O
taken	O
from	O
these	O
smaller	O
positions	O
of	O
interest	O
to	O
formulate	O
a	O
decision	O
regarding	O
the	O
presence	O
of	O
a	O
pedestrian	O
.	O
</s>
<s>
At	O
the	O
IEEE	O
Conference	O
on	O
Computer	B-Application
Vision	I-Application
and	O
Pattern	O
Recognition	O
in	O
2006	O
,	O
Qiang	O
Zhu	O
,	O
Shai	O
Avidan	O
,	O
Mei-Chen	O
Yeh	O
,	O
and	O
Kwang-Ting	O
Cheng	O
presented	O
an	O
algorithm	O
to	O
significantly	O
speed	O
up	O
human	O
detection	O
using	O
HOG	O
descriptor	O
methods	O
.	O
</s>
<s>
Their	O
method	O
uses	O
HOG	O
descriptors	O
in	O
combination	O
with	O
the	O
cascading	B-Algorithm
classifiers	I-Algorithm
algorithm	O
normally	O
applied	O
with	O
great	O
success	O
to	O
face	O
detection	O
.	O
</s>
<s>
In	O
order	O
to	O
isolate	O
the	O
blocks	O
best	O
suited	O
for	O
human	O
detection	O
,	O
they	O
applied	O
the	O
AdaBoost	B-Algorithm
algorithm	O
to	O
select	O
those	O
blocks	O
to	O
be	O
included	O
in	O
the	O
cascade	O
.	O
</s>
<s>
At	O
the	O
IEEE	O
International	O
Conference	O
on	O
Image	B-Algorithm
Processing	I-Algorithm
in	O
2010	O
,	O
Rui	O
Hu	O
,	O
Mark	O
Banard	O
,	O
and	O
John	O
Collomosse	O
extended	O
the	O
HOG	O
descriptor	O
for	O
use	O
in	O
sketch	O
based	O
image	O
retrieval	O
(	O
SBIR	O
)	O
.	O
</s>
<s>
A	O
dense	O
orientation	O
field	O
was	O
extrapolated	O
from	O
dominant	O
responses	O
in	O
the	O
Canny	B-Algorithm
edge	I-Algorithm
detector	I-Algorithm
under	O
a	O
Laplacian	O
smoothness	O
constraint	O
,	O
and	O
HOG	O
computed	O
over	O
this	O
field	O
.	O
</s>
<s>
This	O
enabled	O
the	O
descriptor	O
to	O
be	O
used	O
within	O
a	O
content-based	B-Application
image	I-Application
retrieval	I-Application
system	O
searchable	O
by	O
free-hand	O
sketched	O
shapes	O
.	O
</s>
<s>
The	O
GF-HOG	O
adaptation	O
was	O
shown	O
to	O
outperform	O
existing	O
gradient	O
histogram	O
descriptors	O
such	O
as	O
SIFT	B-Algorithm
,	O
SURF	B-Algorithm
,	O
and	O
HOG	O
by	O
around	O
15	O
percent	O
at	O
the	O
task	O
of	O
SBIR	O
.	O
</s>
<s>
His	O
histogram	O
of	O
oriented	O
residuals	O
descriptor	O
(	O
HOR	O
)	O
was	O
successfully	O
used	O
in	O
object	B-General_Concept
detection	I-General_Concept
tasks	O
of	O
3d	O
pointclouds	O
.	O
</s>
