<s>
The	O
ImageNet	B-General_Concept
project	O
is	O
a	O
large	O
visual	O
database	O
designed	O
for	O
use	O
in	O
visual	O
object	O
recognition	O
software	O
research	O
.	O
</s>
<s>
ImageNet	B-General_Concept
contains	O
more	O
than	O
20,000	O
categories	O
,	O
with	O
a	O
typical	O
category	O
,	O
such	O
as	O
"	O
balloon	O
"	O
or	O
"	O
strawberry	O
"	O
,	O
consisting	O
of	O
several	O
hundred	O
images	O
.	O
</s>
<s>
The	O
database	O
of	O
annotations	O
of	O
third-party	O
image	O
URLs	O
is	O
freely	O
available	O
directly	O
from	O
ImageNet	B-General_Concept
,	O
though	O
the	O
actual	O
images	O
are	O
not	O
owned	O
by	O
ImageNet	B-General_Concept
.	O
</s>
<s>
Since	O
2010	O
,	O
the	O
ImageNet	B-General_Concept
project	O
runs	O
an	O
annual	O
software	O
contest	O
,	O
the	O
ImageNet	B-General_Concept
Large	O
Scale	O
Visual	O
Recognition	O
Challenge	O
(	O
ILSVRC	O
)	O
,	O
where	O
software	O
programs	O
compete	O
to	O
correctly	O
classify	O
and	O
detect	O
objects	O
and	O
scenes	O
.	O
</s>
<s>
On	O
30	O
September	O
2012	O
,	O
a	O
convolutional	B-Architecture
neural	I-Architecture
network	I-Architecture
(	O
CNN	B-Architecture
)	O
called	O
AlexNet	B-Algorithm
achieved	O
a	O
top-5	O
error	O
of	O
15.3	O
%	O
in	O
the	O
ImageNet	B-General_Concept
2012	O
Challenge	O
,	O
more	O
than	O
10.8	O
percentage	O
points	O
lower	O
than	O
that	O
of	O
the	O
runner	O
up	O
.	O
</s>
<s>
This	O
was	O
made	O
feasible	O
due	O
to	O
the	O
use	O
of	O
graphics	B-Architecture
processing	I-Architecture
units	I-Architecture
(	O
GPUs	B-Architecture
)	O
during	O
training	O
,	O
an	O
essential	O
ingredient	O
of	O
the	O
deep	B-Algorithm
learning	I-Algorithm
revolution	O
.	O
</s>
<s>
In	O
2015	O
,	O
AlexNet	B-Algorithm
was	O
outperformed	O
by	O
Microsoft	O
's	O
very	B-Algorithm
deep	I-Algorithm
CNN	I-Algorithm
with	O
over	O
100	O
layers	O
,	O
which	O
won	O
the	O
ImageNet	B-General_Concept
2015	O
contest	O
.	O
</s>
<s>
AI	O
researcher	O
Fei-Fei	O
Li	O
began	O
working	O
on	O
the	O
idea	O
for	O
ImageNet	B-General_Concept
in	O
2006	O
.	O
</s>
<s>
In	O
2007	O
,	O
Li	O
met	O
with	O
Princeton	O
professor	O
Christiane	O
Fellbaum	O
,	O
one	O
of	O
the	O
creators	O
of	O
WordNet	B-General_Concept
,	O
to	O
discuss	O
the	O
project	O
.	O
</s>
<s>
As	O
a	O
result	O
of	O
this	O
meeting	O
,	O
Li	O
went	O
on	O
to	O
build	O
ImageNet	B-General_Concept
starting	O
from	O
the	O
word	O
database	O
of	O
WordNet	B-General_Concept
and	O
using	O
many	O
of	O
its	O
features	O
.	O
</s>
<s>
As	O
an	O
assistant	O
professor	O
at	O
Princeton	O
,	O
Li	O
assembled	O
a	O
team	O
of	O
researchers	O
to	O
work	O
on	O
the	O
ImageNet	B-General_Concept
project	O
.	O
</s>
<s>
They	O
presented	O
their	O
database	O
for	O
the	O
first	O
time	O
as	O
a	O
poster	O
at	O
the	O
2009	O
Conference	O
on	O
Computer	B-Application
Vision	I-Application
and	O
Pattern	O
Recognition	O
(	O
CVPR	O
)	O
in	O
Florida	O
.	O
</s>
<s>
ImageNet	B-General_Concept
crowdsources	B-Application
its	O
annotation	O
process	O
.	O
</s>
<s>
ImageNet	B-General_Concept
uses	O
a	O
variant	O
of	O
the	O
broad	O
WordNet	B-General_Concept
schema	O
to	O
categorize	O
objects	O
,	O
augmented	O
with	O
120	O
categories	O
of	O
dog	O
breeds	O
to	O
showcase	O
fine-grained	O
classification	O
.	O
</s>
<s>
One	O
downside	O
of	O
WordNet	B-General_Concept
use	O
is	O
the	O
categories	O
may	O
be	O
more	O
"	O
elevated	O
"	O
than	O
would	O
be	O
optimal	O
for	O
ImageNet	B-General_Concept
:	O
"	O
Most	O
people	O
are	O
more	O
interested	O
in	O
Lady	O
Gaga	O
or	O
the	O
iPod	O
Mini	O
than	O
in	O
this	O
rare	O
kind	O
of	O
diplodocus.	O
"	O
</s>
<s>
In	O
2012	O
ImageNet	B-General_Concept
was	O
the	O
world	O
's	O
largest	O
academic	O
user	O
of	O
Mechanical	O
Turk	O
.	O
</s>
<s>
There	O
are	O
various	O
subsets	O
of	O
the	O
ImageNet	B-General_Concept
dataset	O
used	O
in	O
various	O
context	O
.	O
</s>
<s>
One	O
of	O
the	O
most	O
highly	O
used	O
subset	O
of	O
ImageNet	B-General_Concept
is	O
the	O
"	O
ImageNet	B-General_Concept
Large	O
Scale	O
Visual	O
Recognition	O
Challenge	O
(	O
ILSVRC	O
)	O
2012-2017	O
image	O
classification	O
and	O
localization	O
dataset	O
"	O
.	O
</s>
<s>
This	O
is	O
also	O
referred	O
to	O
in	O
the	O
research	O
literature	O
as	O
ImageNet-1K	O
or	O
ILSVRC2017	O
,	O
reflecting	O
the	O
original	O
ILSVRC	O
challenge	O
that	O
involved	O
1,000	O
classes	O
.	O
</s>
<s>
ImageNet-1K	O
contains	O
1,281,167	O
training	O
images	O
,	O
50,000	O
validation	O
images	O
and	O
100,000	O
test	O
images	O
.	O
</s>
<s>
The	O
full	O
original	O
dataset	O
is	O
referred	O
to	O
as	O
ImageNet-21K	O
.	O
</s>
<s>
ImageNet-21k	O
contains	O
14,197,122	O
images	O
divided	O
into	O
21,841	O
classes	O
.	O
</s>
<s>
Some	O
papers	O
round	O
this	O
up	O
and	O
name	O
it	O
ImageNet-22k	O
.	O
</s>
<s>
The	O
resulting	O
annual	O
competition	O
is	O
now	O
known	O
as	O
the	O
ImageNet	B-General_Concept
Large	O
Scale	O
Visual	O
Recognition	O
Challenge	O
(	O
ILSVRC	O
)	O
.	O
</s>
<s>
The	O
ILSVRC	O
uses	O
a	O
"	O
trimmed	O
"	O
list	O
of	O
only	O
1000	O
image	O
categories	O
or	O
"	O
classes	O
"	O
,	O
including	O
90	O
of	O
the	O
120	O
dog	O
breeds	O
classified	O
by	O
the	O
full	O
ImageNet	B-General_Concept
schema	O
.	O
</s>
<s>
In	O
2012	O
,	O
a	O
deep	O
convolutional	B-Architecture
neural	I-Architecture
net	I-Architecture
called	O
AlexNet	B-Algorithm
achieved	O
16%	O
;	O
in	O
the	O
next	O
couple	O
of	O
years	O
,	O
top-5	O
error	O
rates	O
fell	O
to	O
a	O
few	O
percent	O
.	O
</s>
<s>
By	O
2015	O
,	O
researchers	O
at	O
Microsoft	O
reported	O
that	O
their	O
CNNs	B-Architecture
exceeded	O
human	O
ability	O
at	O
the	O
narrow	O
ILSVRC	O
tasks	O
.	O
</s>
<s>
In	O
2017	O
ImageNet	B-General_Concept
stated	O
it	O
would	O
roll	O
out	O
a	O
new	O
,	O
much	O
more	O
difficult	O
,	O
challenge	O
in	O
2018	O
that	O
involves	O
classifying	O
3D	O
objects	O
using	O
natural	O
language	O
.	O
</s>
<s>
The	O
applications	O
of	O
progress	O
in	O
this	O
area	O
would	O
range	O
from	O
robotic	O
navigation	O
to	O
augmented	B-General_Concept
reality	I-General_Concept
.	O
</s>
<s>
A	O
study	O
of	O
the	O
history	O
of	O
the	O
multiple	O
layers	O
(	O
taxonomy	O
,	O
object	O
classes	O
and	O
labeling	O
)	O
of	O
ImageNet	B-General_Concept
and	O
WordNet	B-General_Concept
in	O
2019	O
described	O
how	O
bias	O
is	O
deeply	O
embedded	O
in	O
most	O
classification	O
approaches	O
for	O
of	O
all	O
sorts	O
of	O
images	O
.	O
</s>
<s>
ImageNet	B-General_Concept
is	O
working	O
to	O
address	O
various	O
sources	O
of	O
bias	O
.	O
</s>
