<s>
The	O
Viola	B-General_Concept
–	I-General_Concept
Jones	I-General_Concept
object	I-General_Concept
detection	I-General_Concept
framework	I-General_Concept
is	O
a	O
machine	O
learning	O
object	B-General_Concept
detection	I-General_Concept
framework	O
proposed	O
in	O
2001	O
by	O
Paul	O
Viola	O
and	O
Michael	O
Jones	O
.	O
</s>
<s>
It	O
was	O
motivated	O
primarily	O
by	O
the	O
problem	O
of	O
face	B-General_Concept
detection	I-General_Concept
,	O
although	O
it	O
can	O
be	O
adapted	O
to	O
the	O
detection	O
of	O
other	O
object	O
classes	O
.	O
</s>
<s>
The	O
algorithm	O
is	O
efficient	O
for	O
its	O
time	O
,	O
able	O
to	O
detect	O
faces	O
in	O
384	O
by	O
288	O
pixel	O
images	O
at	O
15	O
frames	O
per	O
second	O
on	O
a	O
conventional	O
700MHz	O
Intel	B-General_Concept
Pentium	I-General_Concept
III	I-General_Concept
.	O
</s>
<s>
While	O
it	O
has	O
lower	O
accuracy	O
than	O
more	O
modern	O
methods	O
such	O
as	O
convolutional	B-Architecture
neural	I-Architecture
network	I-Architecture
,	O
its	O
efficiency	O
and	O
compact	O
size	O
(	O
only	O
around	O
50k	O
parameters	O
,	O
compared	O
to	O
millions	O
of	O
parameters	O
for	O
typical	O
CNN	B-Architecture
like	O
DeepFace	O
)	O
means	O
it	O
is	O
still	O
used	O
in	O
cases	O
with	O
limited	O
computational	O
power	O
.	O
</s>
<s>
For	O
example	O
,	O
in	O
the	O
original	O
paper	O
,	O
they	O
reported	O
that	O
this	O
face	O
detector	O
could	O
run	O
on	O
the	O
Compaq	B-Application
iPAQ	I-Application
at	O
2	O
fps	O
(	O
this	O
device	O
has	O
a	O
low	O
power	O
StrongARM	B-Device
without	O
floating	O
point	O
hardware	O
)	O
.	O
</s>
<s>
Face	B-General_Concept
detection	I-General_Concept
is	O
a	O
binary	O
classification	O
problem	O
combined	O
with	O
a	O
localization	O
problem	O
:	O
given	O
a	O
picture	O
,	O
decide	O
whether	O
it	O
contains	O
faces	O
,	O
and	O
construct	O
bounding	B-Algorithm
boxes	I-Algorithm
for	O
the	O
faces	O
.	O
</s>
<s>
for	O
a	O
general	O
picture	O
with	O
a	O
face	O
of	O
unknown	O
size	O
and	O
orientation	O
,	O
one	O
can	O
perform	O
blob	B-Algorithm
detection	I-Algorithm
to	O
discover	O
potential	O
faces	O
,	O
then	O
scale	O
and	O
rotate	O
them	O
into	O
the	O
upright	O
,	O
full-sized	O
position	O
.	O
</s>
<s>
the	O
brightness	O
of	O
the	O
image	O
can	O
be	O
corrected	O
by	O
white	B-Algorithm
balancing	I-Algorithm
.	O
</s>
<s>
the	O
bounding	B-Algorithm
boxes	I-Algorithm
can	O
be	O
found	O
by	O
sliding	O
a	O
window	O
across	O
the	O
entire	O
picture	O
,	O
and	O
marking	O
down	O
every	O
window	O
that	O
contains	O
a	O
face	O
.	O
</s>
<s>
Viola	O
–	O
Jones	O
is	O
essentially	O
a	O
boosted	B-Algorithm
feature	B-General_Concept
learning	I-General_Concept
algorithm	O
,	O
trained	O
by	O
running	O
a	O
modified	O
AdaBoost	B-Algorithm
algorithm	O
on	O
Haar	B-Algorithm
feature	I-Algorithm
classifiers	O
to	O
find	O
a	O
sequence	O
of	O
classifiers	O
.	O
</s>
<s>
Haar	B-Algorithm
feature	I-Algorithm
classifiers	O
are	O
crude	O
,	O
but	O
allows	O
very	O
fast	O
computation	O
,	O
and	O
the	O
modified	O
AdaBoost	B-Algorithm
constructs	O
a	O
strong	O
classifier	O
out	O
of	O
many	O
weak	O
ones	O
.	O
</s>
<s>
Consider	O
a	O
perceptron	B-Algorithm
defined	O
by	O
two	O
variables	O
.	O
</s>
<s>
A	O
Haar	B-Algorithm
feature	I-Algorithm
classifier	O
is	O
a	O
perceptron	B-Algorithm
with	O
a	O
very	O
special	O
kind	O
of	O
that	O
makes	O
it	O
extremely	O
cheap	O
to	O
calculate	O
.	O
</s>
<s>
The	O
Haar	O
features	O
used	O
in	O
the	O
Viola-Jones	B-General_Concept
algorithm	I-General_Concept
are	O
a	O
subset	O
of	O
the	O
more	O
general	O
Haar	B-Algorithm
basis	I-Algorithm
functions	I-Algorithm
,	O
which	O
have	O
been	O
used	O
previously	O
in	O
the	O
realm	O
of	O
image-based	O
object	B-General_Concept
detection	I-General_Concept
.	O
</s>
<s>
While	O
crude	O
compared	O
to	O
alternatives	O
such	O
as	O
steerable	B-Algorithm
filters	I-Algorithm
,	O
Haar	O
features	O
are	O
sufficiently	O
complex	O
to	O
match	O
features	O
of	O
typical	O
human	O
faces	O
.	O
</s>
<s>
Further	O
,	O
the	O
design	O
of	O
Haar	O
features	O
allows	O
for	O
efficient	O
computation	O
of	O
using	O
only	O
constant	O
number	O
of	O
additions	O
and	O
subtractions	O
,	O
regardless	O
of	O
the	O
size	O
of	O
the	O
rectangular	O
features	O
,	O
using	O
the	O
summed-area	B-Algorithm
table	I-Algorithm
.	O
</s>
<s>
Perform	O
a	O
certain	O
modified	O
AdaBoost	B-Algorithm
training	O
on	O
the	O
set	O
of	O
all	O
Haar	B-Algorithm
feature	I-Algorithm
classifiers	O
of	O
dimension	O
,	O
until	O
a	O
desired	O
level	O
of	O
precision	O
and	O
recall	O
is	O
reached	O
.	O
</s>
<s>
The	O
modified	O
AdaBoost	B-Algorithm
algorithm	O
would	O
output	O
a	O
sequence	O
of	O
Haar	B-Algorithm
feature	I-Algorithm
classifiers	O
.	O
</s>
<s>
The	O
details	O
of	O
the	O
modified	O
AdaBoost	B-Algorithm
algorithm	O
is	O
detailed	O
below	O
.	O
</s>
<s>
Thus	O
,	O
the	O
object	B-General_Concept
detection	I-General_Concept
framework	O
employs	O
a	O
variant	O
of	O
the	O
learning	O
algorithm	O
AdaBoost	B-Algorithm
to	O
both	O
select	O
the	O
best	O
features	O
and	O
to	O
train	O
classifiers	O
that	O
use	O
them	O
.	O
</s>
<s>
Each	O
weak	B-Algorithm
classifier	I-Algorithm
is	O
a	O
threshold	O
function	O
based	O
on	O
the	O
feature	O
.	O
</s>
<s>
A	O
simple	O
2-feature	O
classifier	O
can	O
achieve	O
almost	O
100%	O
detection	O
rate	O
with	O
50%	O
FP	B-General_Concept
rate	I-General_Concept
.	O
</s>
<s>
In	O
the	O
case	O
of	O
faces	O
,	O
the	O
first	O
classifier	O
in	O
the	O
cascade	O
–	O
called	O
the	O
attentional	O
operator	O
–	O
uses	O
only	O
two	O
features	O
to	O
achieve	O
a	O
false	O
negative	O
rate	O
of	O
approximately	O
0%	O
and	O
a	O
false	B-General_Concept
positive	I-General_Concept
rate	I-General_Concept
of	O
40%	O
.	O
</s>
<s>
f	O
=	O
the	O
maximum	O
acceptable	O
false	B-General_Concept
positive	I-General_Concept
rate	I-General_Concept
per	O
layer	O
.	O
</s>
<s>
Ftarget	O
=	O
target	O
overall	O
false	B-General_Concept
positive	I-General_Concept
rate	I-General_Concept
.	O
</s>
<s>
Because	O
the	O
activation	O
of	O
each	O
classifier	O
depends	O
entirely	O
on	O
the	O
behavior	O
of	O
its	O
predecessor	O
,	O
the	O
false	B-General_Concept
positive	I-General_Concept
rate	I-General_Concept
for	O
an	O
entire	O
cascade	O
is	O
:	O
</s>
<s>
Thus	O
,	O
to	O
match	O
the	O
false	B-General_Concept
positive	I-General_Concept
rates	I-General_Concept
typically	O
achieved	O
by	O
other	O
detectors	O
,	O
each	O
classifier	O
can	O
get	O
away	O
with	O
having	O
surprisingly	O
poor	O
performance	O
.	O
</s>
<s>
For	O
example	O
,	O
for	O
a	O
32-stage	O
cascade	O
to	O
achieve	O
a	O
false	B-General_Concept
positive	I-General_Concept
rate	I-General_Concept
of	O
,	O
each	O
classifier	O
need	O
only	O
achieve	O
a	O
false	B-General_Concept
positive	I-General_Concept
rate	I-General_Concept
of	O
about	O
65%	O
.	O
</s>
<s>
In	O
videos	O
of	O
moving	O
objects	O
,	O
one	O
need	O
not	O
apply	O
object	B-General_Concept
detection	I-General_Concept
to	O
each	O
frame	O
.	O
</s>
<s>
Instead	O
,	O
one	O
can	O
use	O
tracking	O
algorithms	O
like	O
the	O
KLT	O
algorithm	O
to	O
detect	O
salient	O
features	O
within	O
the	O
detection	O
bounding	B-Algorithm
boxes	I-Algorithm
and	O
track	O
their	O
movement	O
between	O
frames	O
.	O
</s>
