<s>
Cascading	O
is	O
a	O
particular	O
case	O
of	O
ensemble	B-Algorithm
learning	I-Algorithm
based	O
on	O
the	O
concatenation	O
of	O
several	O
classifiers	B-General_Concept
,	O
using	O
all	O
information	O
collected	O
from	O
the	O
output	O
from	O
a	O
given	O
classifier	B-General_Concept
as	O
additional	O
information	O
for	O
the	O
next	O
classifier	B-General_Concept
in	O
the	O
cascade	O
.	O
</s>
<s>
Cascading	B-Algorithm
classifiers	I-Algorithm
are	O
trained	O
with	O
several	O
hundred	O
"	O
positive	O
"	O
sample	O
views	O
of	O
a	O
particular	O
object	O
and	O
arbitrary	O
"	O
negative	O
"	O
images	O
of	O
the	O
same	O
size	O
.	O
</s>
<s>
After	O
the	O
classifier	B-General_Concept
is	O
trained	O
it	O
can	O
be	O
applied	O
to	O
a	O
region	O
of	O
an	O
image	O
and	O
detect	O
the	O
object	O
in	O
question	O
.	O
</s>
<s>
To	O
search	O
for	O
the	O
object	O
in	O
the	O
entire	O
frame	O
,	O
the	O
search	O
window	O
can	O
be	O
moved	O
across	O
the	O
image	O
and	O
check	O
every	O
location	O
with	O
the	O
classifier	B-General_Concept
.	O
</s>
<s>
This	O
process	O
is	O
most	O
commonly	O
used	O
in	O
image	B-Algorithm
processing	I-Algorithm
for	O
object	O
detection	O
and	O
tracking	O
,	O
primarily	O
facial	B-General_Concept
detection	I-General_Concept
and	O
recognition	O
.	O
</s>
<s>
The	O
first	O
cascading	O
classifier	B-General_Concept
was	O
the	O
face	O
detector	O
of	O
Viola	O
and	O
Jones	O
(	O
2001	O
)	O
.	O
</s>
<s>
The	O
requirement	O
for	O
this	O
classifier	B-General_Concept
was	O
to	O
be	O
fast	O
in	O
order	O
to	O
be	O
implemented	O
on	O
low-power	O
CPUs	B-General_Concept
,	O
such	O
as	O
cameras	O
and	O
phones	O
.	O
</s>
<s>
It	O
can	O
be	O
seen	O
from	O
this	O
description	O
that	O
the	O
classifier	B-General_Concept
will	O
not	O
accept	O
faces	O
that	O
are	O
upside	O
down	O
(	O
the	O
eyebrows	O
are	O
not	O
in	O
a	O
correct	O
position	O
)	O
or	O
the	O
side	O
of	O
the	O
face	O
(	O
the	O
nose	O
is	O
no	O
longer	O
in	O
the	O
center	O
,	O
and	O
shadows	O
on	O
the	O
side	O
of	O
the	O
nose	O
might	O
be	O
missing	O
)	O
.	O
</s>
<s>
Separate	O
cascade	B-Algorithm
classifiers	I-Algorithm
have	O
to	O
be	O
trained	O
for	O
every	O
rotation	O
that	O
is	O
not	O
in	O
the	O
image	O
plane	O
(	O
side	O
of	O
face	O
)	O
and	O
will	O
have	O
to	O
be	O
retrained	O
or	O
run	O
on	O
rotated	O
features	O
for	O
every	O
rotation	O
that	O
is	O
in	O
the	O
image	O
plane	O
(	O
face	O
upside	O
down	O
or	O
tilted	O
to	O
the	O
side	O
)	O
.	O
</s>
<s>
The	O
training	O
procedure	O
for	O
one	O
stage	O
is	O
therefore	O
to	O
have	O
many	O
weak	B-Algorithm
learners	I-Algorithm
(	O
simple	O
pixel	O
difference	O
operators	O
)	O
,	O
train	O
them	O
as	O
a	O
group	O
(	O
raise	O
their	O
weight	O
if	O
they	O
give	O
correct	O
result	O
)	O
,	O
but	O
be	O
mindful	O
of	O
having	O
only	O
a	O
few	O
active	O
weak	B-Algorithm
learners	I-Algorithm
so	O
the	O
computation	O
time	O
remains	O
low	O
.	O
</s>
<s>
In	O
their	O
most	O
basic	O
versions	O
,	O
they	O
can	O
be	O
understood	O
as	O
choosing	O
,	O
at	O
each	O
step	O
,	O
between	O
adding	O
a	O
stage	O
or	O
adding	O
a	O
weak	B-Algorithm
learner	I-Algorithm
to	O
a	O
previous	O
stage	O
,	O
whichever	O
is	O
less	O
costly	O
,	O
until	O
the	O
desired	O
accuracy	O
has	O
been	O
reached	O
.	O
</s>
<s>
Every	O
stage	O
of	O
the	O
classifier	B-General_Concept
cannot	O
have	O
a	O
detection	O
rate	O
(	O
sensitivity	O
)	O
below	O
the	O
desired	O
rate	O
,	O
so	O
this	O
is	O
a	O
constrained	B-Application
optimization	I-Application
problem	I-Application
.	O
</s>
<s>
Cascade	B-Algorithm
classifiers	I-Algorithm
are	O
available	O
in	O
OpenCV	B-Language
,	O
with	O
pre-trained	O
cascades	O
for	O
frontal	O
faces	O
and	O
upper	O
body	O
.	O
</s>
<s>
Training	O
a	O
new	O
cascade	O
in	O
OpenCV	B-Language
is	O
also	O
possible	O
with	O
either	O
haar_training	O
or	O
train_cascades	O
methods	O
.	O
</s>
<s>
This	O
can	O
be	O
used	O
for	O
rapid	O
object	O
detection	O
of	O
more	O
specific	O
targets	O
,	O
including	O
non-human	O
objects	O
with	O
Haar-like	B-Algorithm
features	I-Algorithm
.	O
</s>
<s>
The	O
time	O
constraint	O
in	O
training	O
a	O
cascade	B-Algorithm
classifier	I-Algorithm
can	O
be	O
circumvented	O
using	O
cloud-computing	B-Architecture
methods	O
.	O
</s>
<s>
For	O
example	O
,	O
a	O
classifier	B-General_Concept
(	O
for	O
example	O
k-means	O
)	O
,	O
takes	O
a	O
vector	O
of	O
features	O
(	O
decision	O
variables	O
)	O
and	O
outputs	O
for	O
each	O
possible	O
classification	O
result	O
the	O
probability	O
that	O
the	O
vector	O
belongs	O
to	O
the	O
class	O
.	O
</s>
<s>
This	O
is	O
usually	O
used	O
to	O
take	O
a	O
decision	O
(	O
classify	O
into	O
the	O
class	O
with	O
highest	O
probability	O
)	O
,	O
but	O
cascading	B-Algorithm
classifiers	I-Algorithm
use	O
this	O
output	O
as	O
the	O
input	O
to	O
another	O
model	O
(	O
another	O
stage	O
)	O
.	O
</s>
<s>
Having	O
cascading	B-Algorithm
classifiers	I-Algorithm
enables	O
the	O
successive	O
stage	O
to	O
gradually	O
approximate	O
the	O
combinatorial	O
nature	O
of	O
the	O
classification	O
,	O
or	O
to	O
add	O
interaction	O
terms	O
in	O
classification	O
algorithms	O
that	O
cannot	O
express	O
them	O
in	O
one	O
stage	O
.	O
</s>
<s>
The	O
second	O
classifier	B-General_Concept
can	O
pick	O
up	O
this	O
higher	O
probability	O
and	O
make	O
a	O
decision	O
on	O
the	O
sign	O
of	O
feature3	O
.	O
</s>
<s>
In	O
a	O
bias-variance	B-General_Concept
decomposition	I-General_Concept
,	O
cascaded	O
models	O
are	O
usually	O
seen	O
as	O
lowering	O
bias	O
while	O
raising	O
variance	O
.	O
</s>
