<s>
Region-based	O
Convolutional	O
Neural	O
Networks	O
(	O
R-CNN	B-Algorithm
)	O
are	O
a	O
family	O
of	O
machine	O
learning	O
models	O
for	O
computer	B-Application
vision	I-Application
and	O
specifically	O
object	B-General_Concept
detection	I-General_Concept
.	O
</s>
<s>
The	O
original	O
goal	O
of	O
R-CNN	B-Algorithm
was	O
to	O
take	O
an	O
input	O
image	O
and	O
produce	O
a	O
set	O
of	O
bounding	O
boxes	O
as	O
output	O
,	O
where	O
each	O
bounding	O
box	O
contains	O
an	O
object	O
and	O
also	O
the	O
category	O
(	O
e.g.	O
</s>
<s>
More	O
recently	O
,	O
R-CNN	B-Algorithm
has	O
been	O
extended	O
to	O
perform	O
other	O
computer	B-Application
vision	I-Application
tasks	O
.	O
</s>
<s>
The	O
following	O
covers	O
some	O
of	O
the	O
versions	O
of	O
R-CNN	B-Algorithm
that	O
have	O
been	O
developed	O
.	O
</s>
<s>
November	O
2013	O
:	O
R-CNN	B-Algorithm
.	O
</s>
<s>
Given	O
an	O
input	O
image	O
,	O
R-CNN	B-Algorithm
begins	O
by	O
applying	O
a	O
mechanism	O
called	O
Selective	O
Search	O
to	O
extract	O
regions	B-Algorithm
of	I-Algorithm
interest	I-Algorithm
(	O
ROI	O
)	O
,	O
where	O
each	O
ROI	O
is	O
a	O
rectangle	O
that	O
may	O
represent	O
the	O
boundary	O
of	O
an	O
object	O
in	O
image	O
.	O
</s>
<s>
For	O
each	O
ROI	O
's	O
output	O
features	O
,	O
a	O
collection	O
of	O
support-vector	B-Algorithm
machine	I-Algorithm
classifiers	O
is	O
used	O
to	O
determine	O
what	O
type	O
of	O
object	O
(	O
if	O
any	O
)	O
is	O
contained	O
within	O
the	O
ROI	O
.	O
</s>
<s>
April	O
2015	O
:	O
Fast	O
R-CNN	B-Algorithm
.	O
</s>
<s>
While	O
the	O
original	O
R-CNN	B-Algorithm
independently	O
computed	O
the	O
neural	O
network	O
features	O
on	O
each	O
of	O
as	O
many	O
as	O
two	O
thousand	O
regions	B-Algorithm
of	I-Algorithm
interest	I-Algorithm
,	O
Fast	O
R-CNN	B-Algorithm
runs	O
the	O
neural	O
network	O
once	O
on	O
the	O
whole	O
image	O
.	O
</s>
<s>
As	O
in	O
the	O
original	O
R-CNN	B-Algorithm
,	O
the	O
Fast	O
R-CNN	B-Algorithm
uses	O
Selective	O
Search	O
to	O
generate	O
its	O
region	O
proposals	O
.	O
</s>
<s>
June	O
2015	O
:	O
Faster	O
R-CNN	B-Algorithm
.	O
</s>
<s>
While	O
Fast	O
R-CNN	B-Algorithm
used	O
Selective	O
Search	O
to	O
generate	O
ROIs	O
,	O
Faster	O
R-CNN	B-Algorithm
integrates	O
the	O
ROI	O
generation	O
into	O
the	O
neural	O
network	O
itself	O
.	O
</s>
<s>
March	O
2017	O
:	O
Mask	O
R-CNN	B-Algorithm
.	O
</s>
<s>
While	O
previous	O
versions	O
of	O
R-CNN	B-Algorithm
focused	O
on	O
object	B-General_Concept
detection	I-General_Concept
,	O
Mask	O
R-CNN	B-Algorithm
adds	O
instance	O
segmentation	O
.	O
</s>
<s>
Mask	O
R-CNN	B-Algorithm
also	O
replaced	O
ROIPooling	O
with	O
a	O
new	O
method	O
called	O
ROIAlign	O
,	O
which	O
can	O
represent	O
fractions	O
of	O
a	O
pixel	O
.	O
</s>
<s>
June	O
2019	O
:	O
Mesh	O
R-CNN	B-Algorithm
adds	O
the	O
ability	O
to	O
generate	O
a	O
3D	O
mesh	O
from	O
a	O
2D	O
image	O
.	O
</s>
<s>
Region-based	O
convolutional	O
neural	O
networks	O
have	O
been	O
used	O
for	O
tracking	O
objects	O
from	O
a	O
drone-mounted	O
camera	O
,	O
locating	O
text	O
in	O
an	O
image	O
,	O
and	O
enabling	O
object	B-General_Concept
detection	I-General_Concept
in	O
Google	B-General_Concept
Lens	I-General_Concept
.	O
</s>
<s>
Mask	O
R-CNN	B-Algorithm
serves	O
as	O
one	O
of	O
seven	O
tasks	O
in	O
the	O
MLPerf	O
Training	O
Benchmark	O
,	O
which	O
is	O
a	O
competition	O
to	O
speed	O
up	O
the	O
training	O
of	O
neural	O
networks	O
.	O
</s>
