I
nte
rna
t
io
na
l J
o
urna
l o
f
E
lect
rica
l a
nd
Co
m
p
ute
r
E
ng
in
ee
ring
(
I
J
E
CE
)
Vo
l.
9
,
No
.
4
,
A
u
g
u
s
t
201
9
,
p
p
.
2
5
4
8
~
2
5
5
5
I
SS
N:
2
0
8
8
-
8708
,
DOI
: 1
0
.
1
1
5
9
1
/
i
j
ec
e
.
v9
i
4
.
p
p
2
5
4
8
-
2555
2548
J
o
ur
na
l
ho
m
ep
a
g
e
:
h
ttp
:
//ia
e
s
co
r
e
.
co
m/
jo
u
r
n
a
ls
/in
d
ex
.
p
h
p
/
I
JE
C
E
Perf
o
r
m
a
nce
e
v
a
lua
tion o
f
f
in
e
-
t
un
ed
f
a
ster
R
-
CNN
o
n specific
M
S COCO
o
bjec
t
s
G
a
ri
m
a
Dev
na
ni
,
Ay
us
h J
a
is
w
a
l,
Ro
s
hn
i J
o
hn
,
Ra
j
a
t
Cha
ura
s
ia
,
Neha
T
irpu
de
De
p
a
rtme
n
t
o
f
C
o
m
p
u
ter
S
c
ien
c
e
a
n
d
E
n
g
in
e
e
rin
g
,
S
h
ri
Ra
m
d
e
o
b
a
b
a
Co
ll
e
g
e
o
f
En
g
in
e
e
rin
g
a
n
d
M
a
n
a
g
e
m
e
n
t
,
In
d
ia
Art
icle
I
nfo
AB
ST
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
J
u
n
4
,
2
0
1
8
R
ev
i
s
ed
J
an
4
,
20
19
A
cc
ep
ted
Mar
4
,
2
0
1
9
F
in
e
-
tu
n
in
g
o
f
a
m
o
d
e
l
is
a
m
e
t
h
o
d
t
h
a
t
is
m
o
st
o
f
ten
re
q
u
ired
to
c
a
ter
th
e
u
se
rs’
e
x
p
li
c
it
re
q
u
irem
e
n
t
s
.
Bu
t
th
e
q
u
e
sti
o
n
re
m
a
in
s
w
h
e
th
e
r
th
e
m
o
d
e
l
is
a
c
c
u
ra
te
e
n
o
u
g
h
to
b
e
u
se
d
f
o
r
a
c
e
rtain
a
p
p
li
c
a
ti
o
n
.
T
h
is
p
a
p
e
r
striv
e
s
to
p
re
se
n
t
th
e
m
e
tri
c
s
u
se
d
f
o
r
p
e
rf
o
r
m
a
n
c
e
e
v
a
lu
a
ti
o
n
o
f
a
Co
n
v
o
lu
ti
o
n
a
l
Ne
u
ra
l
Ne
t
w
o
rk
(
CNN
)
m
o
d
e
l.
T
h
e
e
v
a
lu
a
ti
o
n
is
b
a
se
d
o
n
t
h
e
train
in
g
p
ro
c
e
ss
w
h
ich
p
ro
v
i
d
e
s
u
s
w
it
h
i
n
term
e
d
iate
m
o
d
e
ls
a
f
ter
e
v
e
r
y
1
0
0
0
it
e
ra
ti
o
n
s.
W
h
il
e
1
0
0
0
it
e
ra
ti
o
n
s
is
n
o
t
su
b
sta
n
ti
a
l
e
n
o
u
g
h
o
v
e
r
th
e
ra
n
g
e
o
f
4
9
0
k
it
e
ra
ti
o
n
s,
th
e
g
ro
u
p
s
a
re
siz
e
d
w
it
h
1
0
0
k
it
e
ra
ti
o
n
s
e
a
c
h
.
No
w
,
th
e
in
ten
ti
o
n
w
a
s
to
c
o
m
p
a
re
th
e
re
c
o
rd
e
d
m
e
tri
c
s
to
e
v
a
lu
a
te
th
e
m
o
d
e
l
in
term
s
o
f
a
c
c
u
ra
c
y
.
T
h
e
train
in
g
m
o
d
e
l
u
se
d
t
h
e
se
t
o
f
sp
e
c
if
ic
c
a
teg
o
r
ies
c
h
o
se
n
f
ro
m
th
e
M
icro
so
f
t
Co
m
m
o
n
Ob
j
e
c
ts
in
Co
n
tex
t
(M
S
COCO
)
d
a
ta
se
t,
w
h
il
e
a
ll
o
w
in
g
u
se
r
s
to
u
se
th
e
ir
o
wn
e
x
tern
a
ll
y
a
v
a
il
a
b
le
i
m
a
g
e
s
t
o
tes
t
th
e
m
o
d
e
l’s
a
c
c
u
ra
c
y
.
Ou
r
train
e
d
m
o
d
e
l
e
n
s
u
re
d
th
a
t
a
ll
th
e
o
b
jec
ts
a
re
d
e
tec
ted
th
a
t
a
re
p
re
se
n
t
i
n
t
h
e
im
a
g
e
to
d
e
p
ict
th
e
e
f
f
e
c
t
o
f
p
re
c
isio
n
.
K
ey
w
o
r
d
s
:
Av
er
ag
e
p
r
ec
is
io
n
C
o
n
v
o
lu
tio
n
al
n
e
u
r
al
n
e
t
w
o
r
k
Dee
p
l
ea
r
n
in
g
Fin
e
-
t
u
n
in
g
Ob
j
ec
t
d
etec
tio
n
P
er
f
o
r
m
a
n
ce
e
v
al
u
atio
n
Co
p
y
rig
h
t
©
2
0
1
9
In
stit
u
te o
f
A
d
v
a
n
c
e
d
E
n
g
i
n
e
e
rin
g
a
n
d
S
c
ien
c
e
.
Al
l
rig
h
ts
re
se
rv
e
d
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
Gar
i
m
a
De
v
n
a
n
i,
Dep
ar
t
m
en
t o
f
C
o
m
p
u
ter
Scie
n
ce
an
d
E
n
g
i
n
ee
r
in
g
,
Sh
r
i
R
a
m
d
eo
b
ab
a
C
o
lleg
e
o
f
E
n
g
i
n
ee
r
i
n
g
a
n
d
Ma
n
a
g
e
m
en
t
,
R
a
m
d
eo
T
ek
d
i,
Gittik
h
ad
an
,
Kato
l Ro
ad
,
Nag
p
u
r
,
Ma
h
ar
as
h
tr
a,
I
n
d
ia.
E
m
ail:
g
ar
i
m
ad
e
v
n
a
n
i
@
g
m
a
il
.
co
m
1.
I
NT
RO
D
UCT
I
O
N
In
to
d
a
y
’
s
f
as
t
p
ac
ed
an
d
tech
n
o
lo
g
icall
y
s
a
v
v
y
w
o
r
ld
,
o
b
j
e
ct
d
etec
tio
n
h
a
s
cr
ea
ted
a
n
ic
h
e
f
o
r
its
el
f
.
T
h
e
ap
p
licatio
n
s
r
en
d
er
ed
f
o
r
th
e
s
a
m
e
f
o
c
u
s
o
n
k
n
o
w
led
g
e
ac
q
u
i
s
itio
n
,
a
n
o
m
a
l
y
d
etec
tio
n
.
So
m
e
ap
p
licatio
n
s
to
m
e
n
tio
n
ar
e
d
e
tectio
n
an
d
tr
ac
k
in
g
f
o
r
u
n
m
a
n
n
ed
ae
r
ial
v
eh
ic
les
as
p
r
esen
ted
b
y
Ka
m
ate
an
d
Y
il
m
az
er
[
1
]
;
an
d
h
u
m
an
-
r
o
b
o
t
in
ter
ac
tio
n
i
n
I
n
d
u
s
tr
y
4
.
0
f
ac
to
r
ies
as
e
x
p
lain
ed
b
y
So
n
n
tag
et
a
l.
[
2
].
T
h
e
ar
ch
itectu
r
e
o
f
t
h
ese
ap
p
licatio
n
s
w
o
r
k
s
o
n
l
y
w
it
h
u
s
ag
e
o
f
o
b
j
ec
t
d
etec
tio
n
alg
o
r
ith
m
s
.
A
l
s
o
k
n
o
w
n
a
s
p
ar
a
m
eter
f
i
n
e
-
tu
n
i
n
g
,
it
is
a
tr
an
s
f
er
lear
n
in
g
tech
n
iq
u
e,
w
h
er
ei
n
,
a
n
et
w
o
r
k
is
p
r
e
-
tr
ain
ed
o
n
a
d
if
f
er
en
t
d
ata
s
et
an
d
t
h
e
n
r
etr
ain
ed
o
n
th
e
tar
g
et
s
e
t
(
o
f
ca
te
g
o
r
i
es)
as
d
ef
i
n
ed
b
y
B
ec
h
er
er
et
al.
[
3
]
.
R
esear
ch
b
y
Y
o
s
in
s
k
i
et
al.
h
a
s
s
h
o
w
n
t
h
is
m
et
h
o
d
b
o
o
s
ts
t
h
e
p
er
f
o
r
m
an
ce
o
f
a
n
et
w
o
r
k
o
v
er
r
an
d
o
m
in
i
tializatio
n
[
4
]
.
T
h
is
tech
n
iq
u
e
is
co
n
s
id
er
ed
to
b
e
ad
v
an
tag
eo
u
s
b
ec
au
s
e
o
f
its
p
r
o
m
in
e
n
t
f
ea
t
u
r
e
o
f
allo
w
in
g
u
s
to
li
m
it
u
s
a
g
e
o
f
t
h
e
n
u
m
b
er
o
f
ca
te
g
o
r
ies to
a
s
p
ec
if
ic
s
et
t
h
at
h
elp
s
ten
d
to
th
e
n
ee
d
o
f
a
ce
r
tain
ap
p
licatio
n
.
Fin
e
-
tu
n
i
n
g
g
i
v
es
u
s
a
d
esira
b
le
o
u
tp
u
t
w
h
e
n
tr
y
i
n
g
to
o
v
er
co
m
e
p
r
o
b
le
m
s
l
ik
e:
-
W
h
en
w
e
u
s
e
R
-
C
NN
b
y
Gir
s
h
ic
k
et
al.
[
5
]
,
th
e
b
o
u
n
d
in
g
b
o
x
es
(
B
B
s
)
ar
e
g
en
er
ated
.
T
h
eses
b
o
x
es
cr
ea
ted
,
f
o
r
th
e
o
b
j
ec
ts
,
w
i
ll
h
av
e
s
o
m
e
p
ar
t o
f
th
e
b
ac
k
g
r
o
u
n
d
o
f
t
h
e
i
m
a
g
e
p
r
ese
n
t i
n
it.
T
h
is
b
ac
k
g
r
o
u
n
d
d
ep
en
d
en
ce
g
en
er
ated
ca
n
ea
s
il
y
m
is
lead
ca
te
g
o
r
izatio
n
an
d
d
etec
tio
n
as c
lai
m
ed
b
y
So
n
n
t
ag
et
al.
[
6
]
.
-
T
h
e
Fas
ter
R
-
C
NN
b
y
R
en
et
al.
[
7
]
w
o
r
k
s
r
ap
id
l
y
w
h
e
n
f
u
ll
i
m
ag
e
co
n
v
o
lu
t
io
n
al
f
ea
t
u
r
es
ar
e
s
h
ar
ed
w
i
t
h
th
e
d
etec
tio
n
n
et
w
o
r
k
u
s
ed
in
[
6
]
.
W
ith
th
e
al
g
o
r
ith
m
s
o
f
o
b
j
ec
t
d
etec
tio
n
h
a
v
in
g
u
n
d
er
g
o
n
e
m
u
l
tip
le
lev
el
s
o
f
ev
o
l
u
tio
n
o
v
er
th
e
y
ea
r
s
,
th
e
ac
c
u
r
ac
y
o
f
th
e
s
a
m
e
h
as
o
n
l
y
i
m
p
r
o
v
ed
.
B
u
t
th
e
i
m
p
le
m
e
n
tat
io
n
s
o
f
th
e
C
NN
s
s
ee
n
o
v
er
t
h
e
y
ea
r
s
d
o
n
o
t
d
is
cu
s
s
t
h
e
ac
c
u
r
ac
y
a
n
d
o
th
er
p
er
f
o
r
m
a
n
ce
m
etr
ics
w
h
e
n
w
e
ar
e
co
n
ce
r
n
ed
w
it
h
o
n
l
y
a
f
e
w
o
f
t
h
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2
0
8
8
-
8708
P
erfo
r
ma
n
ce
ev
a
lu
a
tio
n
o
f fin
e
-
tu
n
ed
fa
s
ter R
-
C
N
N
o
n
s
p
ec
ific MS
C
OC
O
o
b
jects (
Ga
r
im
a
Dev
n
a
n
i)
2549
ca
teg
o
r
ies
in
o
b
j
ec
t
d
etec
tio
n
.
A
l
s
o
,
th
e
n
u
m
b
er
o
f
iter
atio
n
s
p
er
f
o
r
m
ed
f
o
r
tr
ain
in
g
th
e
m
o
d
el
th
at
ca
n
ac
t
as
a
f
ac
to
r
in
th
e
r
e
f
i
n
e
m
e
n
t o
f
t
h
e
ac
cu
r
ac
y
o
f
t
h
e
m
o
d
el
h
a
v
e
n
’
t b
ee
n
co
n
s
id
er
ed
b
ef
o
r
e.
T
h
e
u
s
e
o
f
a
s
p
ec
i
f
ic
s
et
o
f
ca
teg
o
r
ies
f
o
r
tr
ain
i
n
g
o
f
an
o
b
j
ec
t
d
etec
tio
n
alg
o
r
it
h
m
is
w
h
a
t
is
ter
m
ed
as
f
in
e
-
tu
n
in
g
in
la
y
m
a
n
’
s
s
p
ee
ch
.
Gir
s
h
ic
k
et
al.
[
9
]
e
x
p
lain
ed
th
at
t
h
e
ap
p
licatio
n
o
f
th
e
f
in
e
-
tu
n
i
n
g
tech
n
iq
u
e
o
n
a
m
o
d
el
f
o
r
t
h
e
tar
g
et
d
ataset
s
h
o
w
s
i
m
p
r
o
v
e
m
en
t
in
t
h
e
m
ea
n
av
er
a
g
e
p
r
ec
is
io
n
.
T
h
e
w
o
r
k
i
n
g
o
f
f
in
e
-
tu
n
in
g
o
n
th
e
VGG_
C
NN_
M_
1
0
2
4
m
o
d
el
w
as
f
u
r
t
h
er
d
em
o
n
s
tr
ated
in
[
6
]
,
u
s
in
g
th
e
s
u
b
s
e
ts
o
f
th
e
d
ataset
o
f
t
h
e
Mic
r
o
s
o
f
t
C
o
m
m
o
n
Ob
j
ec
ts
in
C
o
n
te
x
t
(
MS
C
OC
O)
d
ata
s
et
as
d
ec
id
ed
b
y
th
e
u
s
er
as
d
o
n
e
b
y
L
i
n
et
al.
[
8
]
.
I
n
o
u
r
ap
p
r
o
ac
h
,
w
e
u
s
e
th
e
p
ath
ad
o
p
ted
in
[
6
]
,
b
r
an
ch
in
g
f
r
o
m
th
e
w
o
r
k
ac
co
m
p
li
s
h
ed
b
y
R
e
n
et
al.
[
7
]
w
h
ic
h
u
t
ilizes
ap
p
r
aisi
n
g
t
h
e
j
o
in
t
tr
ain
in
g
w
h
er
ein
t
h
e
r
eg
io
n
p
r
o
p
o
s
al
n
et
w
o
r
k
i
s
tr
ain
ed
j
o
in
tl
y
w
it
h
th
e
f
ast
R
-
C
NN
b
y
Gir
s
h
ick
[
9
]
,
g
i
v
in
g
th
e
r
esu
lts
w
i
th
s
i
g
n
if
ican
t
i
m
p
r
o
v
e
m
e
n
t
in
s
p
ee
d
.
T
h
is
d
e
m
o
n
s
tr
ates
f
i
n
e
-
tu
n
i
n
g
,
an
d
i
s
t
h
en
m
o
d
i
f
ied
f
o
r
e
n
s
u
r
in
g
t
h
at
al
l
t
h
e
ca
teg
o
r
ie
s
m
e
n
tio
n
ed
in
th
e
s
u
b
s
et
f
o
r
f
i
n
e
-
tu
n
i
n
g
ar
e
d
ete
cted
s
i
m
u
lta
n
eo
u
s
l
y
w
h
e
n
g
i
v
en
a
test
i
n
g
i
m
a
g
es
f
o
r
o
b
j
ec
t d
etec
tio
n
.
T
h
e
co
d
e
u
s
ed
is
f
o
r
k
ed
an
d
is
av
ailab
le
in
P
y
th
o
n
la
n
g
u
a
g
e
at
th
e
f
o
llo
w
i
n
g
lin
k
:
h
ttp
s
:/
/g
i
th
u
b
.
co
m
/a
y
u
s
h
j
ai
s
wal2
2
/S
y
n
ap
s
e
2.
I
M
P
L
E
M
E
NT
AT
I
O
N
Fo
r
o
u
r
p
u
r
p
o
s
es,
w
e
h
a
v
e
u
s
e
d
th
e
p
y
-
f
aster
-
r
cn
n
-
ft
(
h
ttp
s
://
g
it
h
u
b
.
co
m
/
DFKI
-
I
n
ter
ac
tiv
e
-
Ma
ch
i
n
e
-
L
ea
r
n
i
n
g
/p
y
-
f
aster
-
r
cn
n
-
ft
)
m
o
d
el
in
So
n
n
ta
g
et
al.
[
6
]
,
w
h
ic
h
i
s
t
h
e
f
o
r
k
o
f
py
-
f
aster
-
r
cn
n
(
h
ttp
s
://
g
it
h
u
b
.
co
m
/r
b
g
ir
s
h
ic
k
/
p
y
-
f
aster
-
r
cn
n
)
b
y
R
e
n
et
al.
[
7
]
allo
w
i
n
g
u
s
an
ap
p
r
o
p
r
iate
ap
p
licatio
n
o
f
th
e
f
i
n
e
-
tu
n
i
n
g
tech
n
iq
u
e
o
n
th
e
VGG_
C
NN_
M_
1
0
2
4
m
o
d
el
f
o
r
s
p
ec
if
ic
o
b
j
ec
t
ca
teg
o
r
ies
o
f
t
h
e
MS
C
OC
O
d
ataset
lis
ted
b
y
L
i
n
et
al.
[
8
]
.
2
.
1
.
Cho
o
s
ing
Ca
t
eg
o
ries f
r
o
m
M
S CO
CO
Da
t
a
s
et
Mic
r
o
s
o
f
t
C
O
C
O:
C
o
m
m
o
n
Ob
j
ec
t
i
n
C
o
n
te
x
t
is
a
lar
g
e
-
s
ca
le
d
ataset
t
h
at
ad
d
r
ess
es
t
h
r
ee
co
r
e
r
esear
ch
p
r
o
b
le
m
s
i
n
s
ce
n
e
u
n
d
er
s
tan
d
in
g
: d
etec
ti
n
g
n
o
n
-
ico
n
ic
v
ie
w
s
(
o
r
n
o
n
-
ca
n
o
n
ical
p
er
s
p
ec
tiv
es
[
1
0
]
)
o
f
o
b
j
ec
ts
,
co
n
tex
t
u
al
r
ea
s
o
n
in
g
b
et
w
ee
n
o
b
j
ec
ts
an
d
th
e
p
r
ec
i
s
e
2
D
lo
ca
lizat
io
n
o
f
o
b
j
ec
ts
as
ex
p
lai
n
ed
b
y
L
in
et
al
.
[
8
]
.
T
h
e
d
ataset
t
h
at
h
as
b
ee
n
u
s
ed
,
w
as
r
elea
s
ed
in
2
0
1
4
,
co
v
er
s
o
v
er
8
0
ca
teg
o
r
ies
o
f
o
b
j
e
cts.
T
h
e
d
ataset
co
n
s
is
ts
o
f
n
ea
r
l
y
1
6
4
k
i
m
ag
e
s
,
an
d
is
d
iv
id
ed
in
to
th
e
co
llectio
n
s
-
tr
a
in
i
n
g
,
v
alid
atio
n
a
n
d
test
i
n
g
-
s
eg
r
eg
a
tio
n
b
y
L
i
n
et
a
l
.
[
8
]
.
T
a
b
le
1
s
p
ec
if
ies n
u
m
b
e
r
o
f
i
m
ag
e
s
i
n
ea
ch
co
l
lectio
n.
T
ab
le
1
.
Nu
m
b
er
o
f
i
m
a
g
es i
n
MS
C
O
C
O
2
0
1
4
d
ataset
C
o
l
l
e
c
t
i
o
n
T
r
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
T
e
st
i
n
g
N
u
mb
e
r
o
f
i
m
a
g
e
s
8
2
,
7
8
3
4
0
,
5
0
4
4
0
,
7
7
5
T
h
er
e
ex
is
ts
n
ea
r
l
y
8
8
6
k
in
s
tan
ce
s
in
i
m
ag
e
s
o
f
th
e
tr
a
in
i
n
g
+
v
alid
atio
n
co
llecti
o
n
s
m
ad
e
b
y
L
i
n
et
al.
[
8
]
.
T
h
er
e
lib
er
ty
to
ch
o
o
s
e
an
y
s
u
b
s
et
o
f
ca
teg
o
r
ies
f
r
o
m
t
h
e
av
ailab
le
d
atas
et.
T
h
e
d
ataset
h
a
s
an
n
o
tatio
n
s
a
v
ailab
le
to
s
elec
t
ca
teg
o
r
ies
u
s
i
n
g
ca
teg
o
r
y
I
D
s
ass
i
g
n
ed
b
y
L
i
n
et
al.
[8
]
,
w
h
ich
ca
n
b
e
f
etc
h
ed
u
s
i
n
g
th
e
p
r
o
g
r
a
m
w
r
it
ten
in
[
6
]
.
Fo
r
o
u
r
p
er
f
o
r
m
a
n
ce
ev
al
u
atio
n
,
w
e
h
a
v
e
ch
o
s
e
n
th
e
f
o
ll
o
w
i
n
g
ca
te
g
o
r
ies
o
f
p
et
an
i
m
al
s
an
d
ca
ttle a
n
i
m
als
alo
n
g
w
it
h
t
h
e
h
u
m
an
s
: P
er
s
o
n
,
C
at
,
Do
g
, H
o
r
s
e
,
S
h
ee
p
,
an
d
C
o
w
.
2
.
2
.
T
ra
ini
ng
t
he
m
o
del
2
.
2
.
1
.
I
ns
t
a
lla
t
io
n a
nd
s
et
up
As
s
u
g
g
ested
i
n
[
6
]
,
f
o
r
tr
ain
in
g
t
h
e
m
o
d
el,
th
e
i
n
s
tallatio
n
o
f
th
e
Ub
u
n
t
u
1
7
.
0
4
6
4
-
b
it
o
p
er
atin
g
s
y
s
te
m
i
s
s
u
itab
le
b
ec
au
s
e
t
h
i
s
g
r
an
t
s
an
ea
s
y
a
n
d
q
u
ick
i
n
s
tallatio
n
p
r
o
v
id
in
g
m
i
n
i
m
al
co
m
p
ila
tio
n
p
r
o
ce
s
s
f
r
o
m
s
u
p
p
le
m
e
n
tar
y
s
o
u
r
ce
s
.
A
m
ac
h
i
n
e
w
it
h
at
least
o
n
e
GP
U
th
at
s
u
p
p
o
r
ts
C
UD
A
6
i
s
r
eq
u
ir
ed
f
o
r
th
e
f
u
n
ctio
n
i
n
g
o
f
py
-
faster
-
rcnn
-
ft
.
A
s
in
g
le
N
v
id
ia
G
T
X
1
0
8
0
w
a
s
d
ec
lar
ed
s
u
f
f
icie
n
t
f
o
r
th
e
e
x
ec
u
tio
n
us
i
n
g
th
e
VGG
C
NN
M
1
0
2
4
m
o
d
el
i
n
[
6
]
.
Fo
r
tr
ain
in
g
th
e
m
o
d
el,
th
er
e
w
as
u
s
a
g
e
o
f
Go
o
g
le
C
lo
u
d
P
latf
o
r
m
’
s
v
ir
tu
al
m
ac
h
in
e
w
it
h
th
e
Ub
u
n
t
u
1
7
.
0
4
6
4
-
b
it
o
p
er
atin
g
s
y
s
te
m
a
n
d
o
n
e
GP
U
w
ith
t
h
e
s
u
p
p
o
r
t
o
f
C
UD
A
8
.
A
s
i
n
g
le
Nv
id
ia
T
esla
K8
0
GP
U
p
r
o
v
ed
to
b
e
s
u
f
f
icie
n
t
f
o
r
th
e
tr
ain
in
g
p
u
r
p
o
s
e.
2
.
2
.
2
.
P
re
re
qu
is
it
es
T
h
e
f
in
e
-
tu
n
ed
m
o
d
el
to
b
e
g
en
er
ated
,
w
i
ll
b
e
attain
ed
u
s
i
n
g
th
e
VGG
-
1
6
m
o
d
el.
T
h
is
Fa
s
ter
R
-
C
NN
m
o
d
el
o
f
R
e
n
et
al.
i
s
p
r
e
-
tr
ain
ed
o
n
all
th
e
8
0
ca
te
g
o
r
ies
o
f
t
h
e
M
S
C
O
C
O
d
ata
s
et.
T
h
e
MS
C
O
C
O
d
ataset
n
ee
d
s
to
b
e
d
o
w
n
lo
ad
ed
s
ep
ar
atel
y
o
n
t
h
e
m
ac
h
in
e
to
b
e
u
s
ed
f
o
r
tr
ai
n
in
g
a
n
d
ex
ec
u
t
in
g
t
h
e
f
i
n
e
-
tu
n
in
g
o
f
a
n
alr
ea
d
y
p
r
e
-
tr
ain
e
d
m
o
d
el.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
9
,
No
.
4
,
A
u
g
u
s
t 2
0
1
9
:
2
5
4
8
-
2555
2550
2
.
2
.
3
.
I
nte
r
m
e
dia
t
e
s
na
ps
ho
t
s
W
h
en
tr
ai
n
i
n
g
t
h
e
m
o
d
el,
f
o
r
n
n
u
m
b
er
o
f
iter
atio
n
s
,
th
e
p
r
o
g
r
a
m
cr
e
ates
i
n
ter
m
ed
iate
s
n
ap
s
h
o
ts
o
f
th
e
m
o
d
el
f
o
r
ev
er
y
1
0
0
0
iter
a
tio
n
s
o
f
th
e
alg
o
r
it
h
m
.
T
h
ese
s
n
ap
s
h
o
t
s
w
ill b
e
te
s
ted
an
d
u
s
ed
to
d
eter
m
i
n
e
t
h
e
an
al
y
s
is
o
n
h
o
w
t
h
e
a
v
er
ag
e
p
r
ec
is
io
n
c
h
an
g
es
o
v
er
t
h
e
4
9
0
k
iter
atio
n
s
r
u
n
f
o
r
tr
ai
n
in
g
o
u
r
m
o
d
el.
W
e
w
il
l b
e
u
s
i
n
g
th
e
in
ter
m
ed
iate
m
o
d
el
s
g
en
er
ated
af
ter
ev
er
y
1
0
0
,
0
0
0
iter
atio
n
s
:
100k
,
200k
,
3
0
0
k
,
400k
,
an
d
4
9
0
k
.
Fig
u
r
e
1
s
h
o
w
ca
s
e
s
th
e
u
s
e
o
f
th
e
Fas
ter
R
-
C
NN
m
o
d
el
d
e
s
ig
n
ed
b
y
R
e
n
et
al.
[
7
]
,
to
ap
p
ly
t
h
e
f
i
n
e
-
tu
n
i
n
g
tech
n
iq
u
e
i
n
[
6
]
an
d
f
u
r
t
h
er
ev
alu
ate
t
h
e
m
o
d
el
f
o
r
p
er
f
o
r
m
a
n
ce
an
al
y
s
i
s.
Fig
u
r
e
1
.
B
asic f
lo
w
o
f
th
e
Fa
s
ter
R
-
C
NN
Mo
d
el
2
.
3
.
M
et
ho
do
lo
g
y
T
h
e
m
et
h
o
d
o
lo
g
y
h
as
b
ee
n
ad
o
p
ted
v
ia
th
e
u
s
e
o
f
t
h
e
VGG
-
1
6
Fas
ter
R
-
C
NN
m
o
d
elled
b
y
R
e
n
e
t
al.
[
7
]
a
n
d
is
i
n
co
r
p
o
r
ated
w
i
th
t
h
e
c
h
an
g
es
m
ad
e
to
ap
p
l
y
t
h
e
f
i
n
e
-
tu
n
i
n
g
t
ec
h
n
iq
u
e.
F
ig
u
r
e
2
d
ep
icts
th
e
b
asic
ar
c
h
itect
u
r
e
ap
p
lied
o
n
th
e
Fas
ter
R
-
C
NN
m
o
d
el
o
f
R
en
et
al
.
W
h
e
n
tr
ain
i
n
g
o
f
th
e
m
o
d
el
b
eg
in
s
,
th
e
f
ir
s
t
s
tep
is
to
f
etc
h
th
e
r
ele
v
an
t
i
m
a
g
es
f
r
o
m
t
h
e
MS
C
OC
O
tr
ain
i
n
g
d
ataset.
T
h
e
ter
m
r
elev
a
n
t
i
m
a
g
es
co
r
r
esp
o
n
d
s
t
h
e
i
m
ag
es
w
h
o
s
e
ca
te
g
o
r
y
I
Ds
h
av
e
b
ee
n
s
p
ec
i
f
ied
in
t
h
e
co
n
f
i
g
u
r
atio
n
f
ile.
T
h
e
ca
teg
o
r
y
I
Ds c
h
o
s
e
n
ar
e
u
s
er
s
p
ec
if
ic
a
n
d
ca
n
v
ar
y
f
r
o
m
p
er
s
o
n
to
p
er
s
o
n
.
Fig
u
r
e
2
.
Flo
w
o
f
t
h
e
Fi
n
e
-
t
u
n
in
g
a
n
d
an
al
y
s
is
o
f
t
h
e
Fa
s
ter
R
-
C
NN
m
o
d
el
2
.
3
.
1
.
Co
nv
o
lutio
na
l
net
w
o
r
k
la
y
er
T
h
is
s
p
ec
if
ic
ar
r
a
y
o
f
i
m
ag
e
s
i
s
th
e
n
p
ass
ed
o
n
to
t
h
e
co
n
v
o
l
u
tio
n
al
n
et
w
o
r
k
,
w
h
er
ei
n
th
e
i
m
ag
e
s
ar
e
p
r
o
ce
s
s
ed
to
o
b
tain
f
ea
tu
r
e
m
ap
s
af
ter
th
e
last
co
n
v
o
l
u
tio
n
al
lay
er
co
m
p
l
ete
s
its
ex
ec
u
ti
o
n
as
m
en
t
io
n
ed
b
y
R
en
et
al
.
[7
].
T
h
e
f
ea
tu
r
e
m
ap
s
attain
ed
ar
e
o
n
l
y
f
o
r
th
o
s
e
ca
teg
o
r
ies
w
h
ich
h
av
e
b
ee
n
s
p
ec
if
ied
b
y
t
h
e
u
s
e
r
.
2
.
3
.
2
.
Reg
io
n
pro
po
s
a
l net
wo
rk
T
h
e
f
ea
tu
r
e
m
ap
s
ac
q
u
ir
ed
ar
e
n
ex
t
p
r
o
ce
s
s
ed
th
r
o
u
g
h
th
e
s
lid
in
g
w
in
d
o
w
r
u
n
s
p
atiall
y
t
o
g
en
er
ate
p
r
o
p
o
s
als.
Fo
r
ea
ch
s
lid
in
g
w
i
n
d
o
w
o
f
t
h
e
s
ize
3×
3
,
a
g
r
o
u
p
o
f
9
an
ch
o
r
s
ar
e
g
en
er
ated
w
ith
t
h
e
s
a
m
e
ce
n
ter
(
,
)
.
T
h
ese
p
r
o
p
o
s
als
f
o
r
th
e
an
ch
o
r
s
ar
e
g
en
er
ated
w
it
h
3
d
if
f
er
en
t
asp
ec
t
r
atio
s
(
1
:1
,
1
:
2
,
an
d
2
:1
)
an
d
3
d
if
f
er
e
n
t
s
ca
le
s
(
1
2
8
,
2
5
6
,
an
d
5
1
2
)
s
im
ilar
to
w
o
r
k
o
f
R
en
et
al
.
T
h
is
en
tire
p
r
o
ce
s
s
o
cc
u
r
s
at
th
e
la
y
er
co
n
s
is
tin
g
o
f
t
h
e
R
eg
io
n
P
r
o
p
o
s
al
Net
w
o
r
k
(
R
P
N)
,
g
en
er
at
i
n
g
R
e
g
io
n
s
o
f
I
n
ter
e
s
t
(
R
o
I
)
.
T
h
is
w
as
d
e
v
elo
p
ed
to
s
u
b
s
tit
u
te
t
h
e
s
elec
ti
v
e
s
ea
r
ch
b
ein
g
ex
ec
u
ted
in
t
h
e
Fa
s
t
R
-
C
NN
o
f
Gir
s
h
ick
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2
0
8
8
-
8708
P
erfo
r
ma
n
ce
ev
a
lu
a
tio
n
o
f fin
e
-
tu
n
ed
fa
s
ter R
-
C
N
N
o
n
s
p
ec
ific MS
C
OC
O
o
b
jects (
Ga
r
im
a
Dev
n
a
n
i)
2551
T
h
e
in
tr
o
d
u
ctio
n
o
f
asp
ec
t
r
atio
s
an
d
s
ca
les
o
f
o
b
j
ec
ts
ca
m
e
in
to
p
ictu
r
e
d
u
r
in
g
th
e
co
n
s
tr
u
ctio
n
o
f
th
e
R
P
N
in
Fas
ter
R
-
C
NN.
T
h
is
w
as
i
n
cl
u
d
ed
to
h
a
n
d
le
t
h
e
d
if
f
er
en
t
s
ize
s
a
n
d
asp
ec
t
r
atio
s
o
f
t
h
e
o
b
j
ec
t
d
etec
ted
in
th
e
i
m
a
g
es.
Fo
r
ea
ch
o
f
th
e
s
e
an
ch
o
r
s
,
is
ev
alu
a
ted
to
in
d
icate
th
e
am
o
u
n
t
o
f
o
v
er
lap
w
it
h
t
h
e
g
r
o
u
n
d
-
tr
u
th
b
o
u
n
d
in
g
b
o
x
es,
*
1
0
.
7
1
0
.
3
0
if
I
o
U
p
if
I
o
U
o
th
e
r
w
is
e
(
1
)
w
h
er
e
,
I
o
U
is
in
ter
s
ec
t
io
n
o
v
e
r
u
n
io
n
a
n
d
is
e
x
p
r
ess
ed
as b
elo
w
:
A
n
c
h
o
r
G
r
o
u
n
d
t
r
u
t
h
B
o
x
I
o
U
A
n
c
h
o
r
G
r
o
u
n
d
t
r
u
t
h
B
o
x
(
2
)
Fin
all
y
t
h
e
3
×
3
s
p
atial
f
ea
t
u
r
es
ar
e
ex
tr
ac
ted
f
r
o
m
th
e
d
er
iv
ed
f
ea
t
u
r
e
m
ap
s
to
b
e
f
o
r
war
d
ed
to
a
s
m
al
ler
n
et
w
o
r
k
as o
f
R
e
n
et
a
l
.
[
7
]
.
2
.
3
.
3
.
M
ultila
y
er
ba
ck
pro
pa
g
a
t
io
n net
w
o
rk
T
h
is
s
m
aller
n
et
w
o
r
k
ac
co
m
p
lis
h
es
t
h
e
f
o
llo
w
i
n
g
t
w
o
task
s
to
o
b
tain
th
e
f
i
n
al
o
u
tp
u
t,
i.e
.
,
C
las
s
i
f
icati
o
n
(
cls)
a
n
d
R
eg
r
e
s
s
io
n
(
r
eg
)
,
as c
r
ea
ted
b
y
R
en
e
t a
l
.
[
7
]
.
T
h
e
task
o
f
r
eg
r
ess
io
n
g
e
n
e
r
ates
th
e
d
i
m
en
s
io
n
s
f
o
r
th
e
p
r
ed
icte
d
b
o
u
n
d
in
g
b
o
x
f
o
r
a
ce
r
tain
d
etec
ted
o
b
j
ec
t.
T
h
e
d
im
e
n
s
i
o
n
s
ar
e
g
i
v
e
n
as
(
,
,
,
ℎ
)
,
w
h
er
e
th
e
ea
ch
o
f
t
h
e
m
en
tio
n
ed
v
ar
ia
b
les
ar
e
d
ef
in
ed
as
f
o
llo
w
s
:
-
–
T
h
e
x
co
o
r
d
in
ate
o
f
th
e
ce
n
t
er
o
f
th
e
b
o
u
n
d
i
n
g
b
o
x
,
-
–
T
h
e
y
co
o
r
d
in
ate
o
f
th
e
ce
n
t
er
o
f
th
e
b
o
u
n
d
i
n
g
b
o
x
,
-
–
T
h
e
w
id
th
o
f
t
h
e
b
o
u
n
d
i
n
g
b
o
x
,
-
ℎ
–
T
h
e
h
eig
h
t o
f
th
e
b
o
u
n
d
in
g
b
o
x
.
As
R
en
et
al.
[
7
]
h
av
e
ex
p
lain
ed
,
t
h
e
task
o
f
class
i
f
icat
i
o
n
d
er
iv
es
th
e
p
r
o
b
a
b
ilit
y
in
d
icatin
g
w
h
et
h
er
th
e
b
o
x
o
b
tain
ed
f
r
o
m
t
h
e
r
eg
r
es
s
io
n
ta
s
k
co
n
tai
n
s
an
o
b
j
ec
t
o
r
d
o
es
it
h
o
ld
th
e
b
ac
k
g
r
o
u
n
d
(
0
f
o
r
n
o
o
b
j
ec
t)
.
T
h
is
la
y
er
co
n
s
i
s
ti
n
g
o
f
t
h
e
s
m
al
ler
n
et
w
o
r
k
i
s
wh
er
e
th
e
lo
s
s
is
ca
lcu
la
ted
o
v
er
th
e
r
u
n
o
f
t
h
e
iter
atio
n
s
,
b
ac
k
p
r
o
p
ag
atio
n
o
cc
u
r
r
in
g
af
ter
ev
er
y
2
0
0
iter
atio
n
s
.
T
h
e
lo
s
s
f
u
n
ctio
n
f
o
r
ev
er
y
a
n
ch
o
r
i
is
ex
p
r
ess
ed
as f
o
llo
w
s
:
**
11
(
{
}
,
{
}
)
(
,
)
(
,
)
i
i
c
l
s
i
i
r
e
g
i
i
c
l
s
r
e
g
L
p
t
L
p
p
L
t
t
NN
(
3
)
I
n
eq
u
atio
n
(
3
)
,
th
e
2
ter
m
s
in
ad
d
itio
n
ar
e
n
o
r
m
alize
d
b
y
an
d
an
d
w
e
ig
h
ted
b
y
a
b
alan
cin
g
p
ar
a
m
eter
.
Fo
r
o
u
r
i
m
p
le
m
en
ta
tio
n
a
s
tak
e
n
b
y
R
en
et
al.
[
7
]
,
th
e
ter
m
i
s
n
o
r
m
alize
d
b
y
t
h
e
m
i
n
i
-
b
atch
s
ize
o
f
t
h
e
v
al
u
e
=
256
an
d
th
e
ter
m
i
s
n
o
r
m
alize
d
b
y
t
h
e
n
u
m
b
er
o
f
an
c
h
o
r
lo
ca
ti
o
n
s
n
ea
r
l
y
~
2
,
400
.
T
h
e
d
ef
au
lt
v
alu
e
o
f
=
10
,
w
h
ich
s
ets
t
h
e
ter
m
s
an
d
as
r
o
u
g
h
l
y
eq
u
a
ll
y
w
ei
g
h
ted
,
as d
ec
id
ed
b
y
R
e
n
e
t a
l
.
[
7
]
.
T
h
e
v
alu
es
an
d
ar
e
c
alcu
lated
in
t
h
e
f
o
r
m
(
,
,
,
ℎ
)
,
as f
o
llo
w
s
:
,
,
l
og
,
l
og
aa
a
a
a
a
x
x
y
y
wh
t
w
w
w
h
(
4
)
**
**
*
,
,
l
og
,
l
og
aa
a
a
a
a
x
x
y
y
wh
t
w
w
w
h
(
5
)
w
h
er
e,
an
c
h
o
r
’
s
p
r
o
p
er
ties
,
d
e
f
i
n
ed
b
y
R
en
et
al.
[
7
]
,
ar
e
d
ep
icted
b
y
:
-
–
T
h
e
x
co
o
r
d
in
ate
o
f
th
e
ce
n
t
er
o
f
th
e
an
c
h
o
r
’
s
b
o
u
n
d
i
n
g
b
o
x
,
-
–
T
h
e
y
co
o
r
d
in
ate
o
f
t
h
e
ce
n
t
er
o
f
th
e
an
c
h
o
r
’
s
b
o
u
n
d
i
n
g
b
o
x
,
-
–
T
h
e
w
id
th
o
f
t
h
e
a
n
ch
o
r
’
s
b
o
u
n
d
i
n
g
b
o
x
,
-
ℎ
–
T
h
e
h
eig
h
t o
f
t
h
e
an
c
h
o
r
’
s
b
o
u
n
d
i
n
g
b
o
x
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
9
,
No
.
4
,
A
u
g
u
s
t 2
0
1
9
:
2
5
4
8
-
2555
2552
an
d
th
e
p
r
o
p
er
ties
o
f
g
r
o
u
n
d
-
tr
u
th
b
o
u
n
d
i
n
g
b
o
x
[
5
]
ar
e
ex
p
r
ess
ed
w
it
h
th
e
f
o
llo
w
in
g
:
-
-
T
h
e
x
co
o
r
d
in
ate
o
f
th
e
ce
n
te
r
o
f
th
e
b
o
u
n
d
i
n
g
b
o
x
,
-
-
T
h
e
y
co
o
r
d
in
ate
o
f
t
h
e
ce
n
te
r
o
f
th
e
b
o
u
n
d
i
n
g
b
o
x
,
-
-
T
h
e
w
id
th
o
f
th
e
b
o
u
n
d
in
g
b
o
x
,
-
ℎ
-
T
h
e
h
eig
h
t o
f
t
h
e
b
o
u
n
d
i
n
g
b
o
x
.
2
.
4
.
I
m
a
g
e
s
elec
t
io
n f
o
r
t
esti
ng
Fo
r
th
e
test
i
n
g
,
to
o
ls
/d
emo
2
.
p
y
is
th
e
p
r
o
g
r
a
m
th
at
h
as
b
ee
n
ex
ec
u
ted
.
I
n
[
6
]
,
a
v
er
s
io
n
o
f
th
e
p
r
o
ce
d
u
r
e
to
o
ls
/d
emo
.
p
y
w
as
u
s
ed
f
o
r
th
e
v
is
u
aliz
in
g
o
f
th
e
f
i
n
e
-
tu
n
ed
tr
ain
ed
m
o
d
el
.
I
n
t
h
eir
v
er
s
io
n
,
t
h
e
u
s
er
is
n
o
t
r
eq
u
ir
ed
to
m
a
n
u
a
l
l
y
s
ea
r
c
h
f
o
r
ap
p
licab
le
i
m
a
g
es
in
t
h
e
d
ataset
to
b
e
u
s
ed
f
o
r
th
e
test
i
n
g
o
n
t
h
e
m
o
d
el
i
n
[
6
]
.
I
n
o
u
r
v
er
s
io
n
,
th
e
i
m
a
g
es
t
h
e
u
s
er
w
is
h
ed
to
b
e
test
ed
ca
n
b
e
in
p
u
t
ex
ter
n
all
y
,
i.e
.
,
th
e
i
m
ag
e
s
n
ee
d
n
o
t
n
ec
es
s
ar
il
y
b
e
tak
e
n
f
r
o
m
th
e
d
ataset
b
u
t
ca
n
b
e
an
y
i
m
ag
e
t
h
e
y
o
w
n
o
r
h
a
v
e
d
o
w
n
lo
ad
ed
f
r
o
m
o
t
h
er
ex
ter
n
al
s
o
u
r
ce
s
.
R
u
n
n
i
n
g
th
e
to
o
ls
/d
emo
2
.
p
y
,
w
h
e
n
th
e
i
m
ag
e
is
g
iv
e
n
f
o
r
test
in
g
,
th
e
m
o
d
el
d
etec
ts
all
th
e
o
b
j
ec
ts
p
r
esen
t
in
th
e
i
m
ag
e
s
i
m
u
lta
n
eo
u
s
l
y
,
w
h
ic
h
m
ea
n
s
i
t
ac
co
m
p
li
s
h
e
s
a
co
m
p
r
eh
e
n
s
i
v
e
o
b
j
ec
t
d
etec
tio
n
f
o
r
o
u
r
s
p
ec
if
ic
s
et
o
f
ca
teg
o
r
i
es th
e
m
o
d
el
h
as b
ee
n
tr
ai
n
ed
o
n
.
3.
P
E
RF
O
RM
ANCE
E
VA
L
U
AT
I
O
N
I
n
th
e
MS
C
OC
O
d
ataset,
th
er
e
ar
e
n
ea
r
l
y
1
6
4
k
im
a
g
es
an
d
ea
ch
o
f
th
ese
i
m
ag
e
s
co
n
tai
n
in
s
tan
ce
s
o
f
m
u
l
tip
le
clas
s
es.
A
ls
o
,
t
h
e
m
o
d
el
u
s
es
co
n
v
o
lu
tio
n
to
d
etec
t
f
ea
t
u
r
es
w
h
ich
m
a
k
es
t
h
e
m
o
d
el
en
co
u
n
ter
th
o
u
s
an
d
s
o
f
n
e
g
at
i
v
e
i.e
.
,
b
ac
k
g
r
o
u
n
d
ex
a
m
p
les
f
o
r
ev
er
y
p
o
s
itiv
e
e
x
a
m
p
le.
I
n
o
b
j
ec
t
d
etec
tio
n
,
th
e
p
r
o
ce
s
s
is
b
ased
o
n
th
e
lo
ca
tio
n
o
f
t
h
e
o
b
j
ec
t
in
th
e
i
m
ag
e.
E
v
er
y
b
o
u
n
d
in
g
b
o
x
d
etec
ted
is
ass
ig
n
ed
to
t
h
e
g
r
o
u
n
d
tr
u
th
o
b
j
ec
ts
to
d
eter
m
i
n
e
w
h
et
h
er
it
is
a
tr
u
e
o
r
f
alse
p
o
s
itiv
e,
d
ep
en
d
i
n
g
o
n
t
h
e
a
m
o
u
n
t
o
f
o
v
er
lap
p
in
g
.
Fo
r
d
etec
tio
n
,
m
a
n
y
b
o
u
n
d
i
n
g
b
o
x
es
ar
e
cr
ea
te
d
b
y
th
e
m
o
d
el
w
it
h
a
p
r
o
b
a
b
ilit
y
o
r
co
n
f
id
e
n
ce
s
c
o
r
e
attac
h
ed
to
it.
I
f
th
er
e
ar
e
m
u
ltip
le
b
o
u
n
d
i
n
g
b
o
x
es
f
o
r
th
e
s
a
m
e
i
m
ag
e,
t
h
e
b
o
u
n
d
in
g
b
o
x
w
it
h
th
e
h
i
g
h
e
s
t
s
co
r
e
is
g
r
ee
d
il
y
s
elec
ted
[
1
1
]
.
T
h
e
d
etec
tio
n
is
co
n
s
id
er
ed
to
b
e
co
r
r
ec
t
w
h
en
t
h
e
ar
ea
o
f
o
v
er
lap
b
et
w
ee
n
d
etec
ted
b
o
u
n
d
in
g
b
o
x
B
p
an
d
th
e
g
r
o
u
n
d
tr
u
th
b
o
u
n
d
in
g
b
o
x
B
gt
ex
ce
ed
5
0
%,
as
cl
ai
m
ed
b
y
Gir
s
h
ic
k
[
9
]
,
also
ca
lled
as
I
n
ter
s
ec
tio
n
o
f
Un
io
n
(
I
o
U)
.
T
h
e
f
o
r
m
u
la
i
s
as f
o
llo
w
s
:
()
()
p
g
t
p
g
t
a
r
e
a
B
B
A
a
r
e
a
B
B
(
6
)
3
.
1
.
P
re
cisi
o
n
P
er
ce
n
tag
e
o
f
tr
u
e
p
o
s
iti
v
es
(
t
p
)
in
th
e
r
esu
lta
n
t
b
o
u
n
d
in
g
b
o
x
,
w
h
er
e
f
p
is
f
a
ls
e
p
o
s
iti
v
e
s
an
d
n
is
th
e
to
tal
n
u
m
b
er
o
f
b
o
u
n
d
in
g
b
o
x
es r
etr
iev
ed
.
t
p
t
p
P
r
e
c
i
s
i
o
n
t
p
f
p
n
(
7
)
3
.
2
.
Rec
a
ll
P
er
ce
n
tag
e
o
f
th
e
ac
t
u
al
o
b
j
ec
ts
d
etec
ted
w
h
er
e
f
n
is
f
alse
n
e
g
ati
v
es.
tp
R
e
c
a
ll
tp
fn
(
8
)
4.
RE
SU
L
T
S
A
ND
F
I
NDIN
G
4
.
1
.
M
ea
n
a
v
er
a
g
e
prec
is
io
n
T
ab
le
2
s
h
o
w
s
th
e
r
ec
o
r
d
ed
m
ea
n
av
er
a
g
e
p
r
ec
is
io
n
(
m
A
P
)
o
f
th
e
s
ix
ca
te
g
o
r
ies c
h
o
s
en
f
r
o
m
th
e
M
S
C
OC
O
d
ataset
f
o
r
t
h
e
i
n
ter
m
ed
iate
s
n
ap
s
h
o
ts
th
a
t
w
er
e
ev
alu
a
ted
.
T
h
e
v
al
u
es
d
ep
ict
th
at
o
v
er
t
h
e
4
9
0
k
iter
atio
n
s
,
t
h
e
m
o
d
el
s
tead
il
y
i
m
p
r
o
v
es it
s
elf
f
o
r
th
e
s
p
ec
i
f
ic
s
et
o
f
ca
teg
o
r
ies
.
T
h
e
n
u
m
b
er
s
i
n
T
ab
le
2
r
e
p
r
esen
t
t
h
at
ev
e
n
th
e
i
n
ter
m
ed
i
ate
s
n
ap
s
h
o
ts
o
r
m
o
d
els
cr
ea
t
ed
ca
n
b
e
u
s
ed
f
o
r
test
i
n
g
,
s
in
ce
t
h
e
y
h
a
v
e
a
r
aise
in
ac
cu
r
ac
y
co
m
p
ar
ed
to
th
e
last
in
ter
m
ed
iate
s
n
a
p
s
h
o
t.
T
h
e
n
u
m
b
er
s
h
elp
s
u
s
co
m
p
ar
e
th
e
f
i
n
al
r
es
u
lts
w
.
r
.
t.
t
h
e
en
tire
m
o
d
el.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2
0
8
8
-
8708
P
erfo
r
ma
n
ce
ev
a
lu
a
tio
n
o
f fin
e
-
tu
n
ed
fa
s
ter R
-
C
N
N
o
n
s
p
ec
ific MS
C
OC
O
o
b
jects (
Ga
r
im
a
Dev
n
a
n
i)
2553
T
ab
le
2
.
Me
an
av
er
ag
e
p
r
ec
is
i
o
n
(
m
A
P
)
@
I
o
U
=
[
0
.
5
0
,
0
.
9
5
]
o
v
er
th
e
in
ter
v
al
o
f
1
0
0
k
i
ter
atio
n
s
O
b
j
e
c
t
mA
P
(
M
e
a
n
A
v
e
r
a
g
e
P
r
e
c
i
si
o
n
)
1
0
0
k
I
t
e
r
a
t
i
o
n
s
2
0
0
k
I
t
e
r
a
t
i
o
n
s
3
0
0
k
I
t
e
r
a
t
i
o
n
s
4
0
0
k
I
t
e
r
a
t
i
o
n
s
4
9
0
k
I
t
e
r
a
t
i
o
n
s
En
t
i
r
e
M
o
d
e
l
1
5
.
5
1
7
.
5
1
9
.
3
2
5
.
3
2
6
.
2
P
e
r
so
n
2
2
.
0
2
2
.
7
2
4
.
8
2
8
.
9
2
9
.
3
Cat
2
1
.
5
2
8
.
9
2
9
.
0
3
7
.
3
3
8
.
9
D
o
g
1
3
.
5
1
3
.
0
1
6
.
7
2
4
.
8
2
4
.
5
H
o
r
se
1
6
.
5
1
7
.
0
1
8
.
7
2
5
.
3
2
7
.
0
S
h
e
e
p
1
1
.
3
1
1
.
4
1
4
.
3
1
9
.
9
2
0
.
4
C
o
w
7
.
9
1
2
.
0
1
2
.
2
1
5
.
9
1
7
.
1
4
.
2
.
L
o
s
s
f
u
nct
io
n
Usi
n
g
(
3
)
,
w
e
ca
lc
u
late
t
h
e
lo
s
s
af
ter
ev
er
y
1
0
0
k
iter
atio
n
s
.
T
h
e
r
esu
lts
s
h
o
w
th
a
t
th
e
lo
s
s
v
al
u
e
i
s
v
er
y
lo
w
at
t
h
e
e
n
d
o
f
th
e
tr
ain
i
n
g
,
w
h
i
le
e
x
p
er
ien
ci
n
g
s
o
m
e
f
l
u
ctu
a
tio
n
s
d
u
r
i
n
g
th
e
ite
r
atio
n
s
,
as
d
ep
icted
in
Fi
g
u
r
e
3
.
Fig
u
r
e
3
.
L
o
s
s
v
al
u
es v
s
.
n
u
m
b
er
o
f
iter
atio
n
s
o
n
th
e
M
S C
OC
O
2
0
1
4
tr
ain
s
et
4
.
3
.
P
r
ec
is
io
n
-
r
ec
a
ll c
urv
e
T
h
e
P
r
ec
is
io
n
-
R
ec
all
C
u
r
v
e,
p
lo
tted
u
s
i
n
g
t
h
e
ca
lc
u
l
atio
n
s
d
o
cu
m
e
n
ted
w
it
h
(
7
)
an
d
(
8
)
in
Fig
u
r
e
4
,
s
h
o
w
s
a
s
tead
y
g
r
o
w
t
h
o
v
er
th
e
s
p
an
th
e
tr
ai
n
i
n
g
o
f
th
e
m
o
d
el.
Fig
u
r
e
4
.
T
h
e
P
r
ec
is
io
n
-
R
ec
all
C
u
r
v
e
@
I
o
U
=
[
0
.
5
0
,
0
.
9
5
]
f
o
r
all
s
ized
o
b
j
ec
ts
o
v
er
4
9
0
k
iter
atio
n
s
4
.
4
.
P
r
ec
is
io
n o
f
I
o
U
p
r
edict
io
n
P
r
ec
is
io
n
o
f
I
o
U
is
a
m
ea
s
u
r
e
to
f
in
d
th
e
ti
g
h
tl
y
b
o
u
n
d
b
o
x
f
o
r
th
e
o
b
j
ec
ts
d
etec
ted
u
s
in
g
f
ea
t
u
r
e
m
ap
s
.
W
h
ile
t
h
e
“
all
”
attr
ib
u
t
e
d
ef
in
e
s
t
h
e
b
o
x
b
ei
n
g
p
lace
d
o
v
er
all
s
ized
o
b
j
ec
ts
in
t
h
e
i
m
a
g
e,
t
h
e
o
b
j
ec
ts
s
izes d
etec
te
d
ca
n
b
e
o
f
th
e
f
o
l
lo
w
i
n
g
t
y
p
es:
-
S
m
all
,
-
Me
d
iu
m
,
a
n
d
-
L
ar
g
e.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
9
,
No
.
4
,
A
u
g
u
s
t 2
0
1
9
:
2
5
4
8
-
2555
2554
Fig
u
r
e
5
d
ep
icts
th
e
v
ar
iatio
n
s
in
p
r
ec
is
io
n
w
h
en
co
n
s
id
er
i
n
g
th
e
s
izes
o
f
t
h
e
o
b
j
ec
ts
.
As
ex
p
ec
ted
th
e
o
b
j
ec
t
s
ized
as
la
r
g
e
h
av
e
th
e
h
i
g
h
e
s
t
p
r
ec
is
io
n
,
w
h
i
le
th
e
s
m
all
s
ized
ar
e
d
if
f
icu
lt
to
d
etec
t.
B
u
t
,
w
h
e
n
all
o
f
th
e
s
ize
s
ar
e
co
n
s
id
er
ed
to
g
eth
er
,
th
e
a
v
er
ag
e
p
r
ec
is
io
n
i
s
at
p
ar
w
ith
t
h
e
a
v
er
ag
e
p
r
ec
is
i
o
n
o
f
th
e
m
o
d
el.
T
h
e
o
th
er
p
a
r
am
eter
u
s
ed
w
.
r
.
t
.
m
ea
s
u
r
i
n
g
p
r
ec
is
io
n
o
f
I
o
U
is
th
e
ar
ea
p
er
ce
n
tag
es
w
h
e
n
all
s
izes
o
f
o
b
j
ec
ts
ar
e
co
n
s
id
er
ed
in
t
h
e
ar
ea
an
d
th
e
y
ar
e
as
f
o
llo
w
s
:
R
an
g
e
o
f
5
0
–
95
%
,
50%
,
an
d
75%
.
Fig
u
r
e
6
s
h
o
w
s
t
h
at
w
h
e
n
I
o
U
ar
ea
is
at
least 5
0
% th
e
p
r
ec
is
io
n
i
s
m
e
asu
r
ed
to
b
e
th
e
h
i
g
h
e
s
t.
Fig
u
r
e
5
.
A
v
er
ag
e
P
r
ec
is
io
n
v
s
A
r
ea
@
I
o
U
=
[
0
.
5
0
,
0
.
9
5
]
f
o
r
o
b
j
ec
ts
s
ized
as [
all,
s
m
al
l,
m
ed
i
u
m
,
lar
g
e]
Fig
u
r
e
6
.
A
v
er
a
g
e
P
r
ec
is
io
n
v
s
A
r
ea
@
I
o
U
f
o
r
all
s
ized
o
b
j
ec
ts
5.
CO
NCLU
SI
O
N
Fro
m
th
e
ab
o
v
e
d
er
iv
ed
r
esu
l
t
s
,
w
e
ca
n
co
n
cl
u
d
e
th
at
t
h
e
Fa
s
ter
R
-
C
NN
(
VGG
m
o
d
el)
m
a
in
tai
n
s
it
s
ac
cu
r
ac
y
w
h
e
n
tr
ai
n
ed
w
it
h
t
h
e
f
i
n
e
-
tu
n
i
n
g
tec
h
n
iq
u
e
a
s
w
el
l.
T
h
e
av
er
ag
e
p
r
ec
is
io
n
,
f
o
r
e
ac
h
cla
s
s
t
h
e
m
o
d
el
is
tr
ain
ed
o
n
,
d
ep
icts
a
s
tead
y
escalatio
n
,
e
n
s
u
r
in
g
co
r
r
ec
t d
etec
tio
n
o
f
o
b
j
ec
t b
y
th
e
m
o
d
el.
Fu
r
t
h
er
m
o
r
e,
o
v
er
th
e
co
u
r
s
e
o
f
tr
ain
in
g
th
e
m
o
d
el,
th
e
lo
s
s
v
alu
e
s
d
ep
ict
f
lu
ct
u
atio
n
s
b
u
t
b
y
t
h
e
en
d
o
f
th
e
p
r
o
ce
s
s
,
it
h
a
s
th
e
lo
w
e
s
t
p
o
s
s
ib
le
lo
s
s
i
n
cu
r
r
ed
,
r
ep
r
e
s
en
ti
n
g
th
e
m
o
d
el
’
s
b
etter
w
o
r
k
in
g
a
n
d
ac
cu
r
ac
y
.
T
h
e
g
r
o
w
t
h
o
f
th
e
P
r
ec
is
io
n
-
R
ec
all
cu
r
v
e
s
h
o
w
n
i
n
Fig
u
r
e
3
m
ea
n
s
t
h
at
th
e
m
o
d
el
d
etec
ts
th
e
ac
tu
al
o
b
j
ec
ts
co
r
r
ec
tly
,
i
n
s
id
e
t
h
e
b
o
u
n
d
in
g
b
o
x
.
T
h
is
i
m
p
lies
th
a
t t
h
e
ac
cu
r
ac
y
w
.
r
.
t
av
er
a
g
e
p
r
ec
is
io
n
an
d
r
ec
all,
in
s
id
e
t
h
e
b
o
u
n
d
in
g
b
o
x
f
o
r
o
b
j
ec
t d
etec
t
io
n
,
s
tead
il
y
r
i
s
e
o
v
er
th
e
co
u
r
s
e
o
f
th
e
4
9
0
k
iter
atio
n
s
r
u
n
to
tr
ain
th
e
m
o
d
el.
A
l
s
o
,
r
ea
d
in
g
Fig
u
r
e
4
,
w
e
ca
n
p
r
ec
is
el
y
co
n
cl
u
d
e
th
at
o
v
er
a
s
et
o
f
d
if
f
er
en
t
s
ized
o
b
jects,
lar
g
e
o
b
j
ec
ts
ar
e
ea
s
il
y
d
etec
ted
d
u
e
to
t
h
eir
lar
g
e
ar
ea
co
v
er
a
g
e,
h
e
n
ce
h
i
g
h
est
av
er
a
g
e
p
r
ec
is
io
n
.
T
h
e
s
m
all
o
b
j
ec
ts
f
o
llo
w
t
h
e
e
x
ac
t o
p
p
o
s
ite
p
atter
n
in
a
v
er
ag
e
p
r
ec
is
io
n
.
W
h
en
w
e
co
n
s
id
er
all
s
ized
o
b
j
ec
ts
f
o
r
p
r
ec
is
io
n
o
v
er
d
if
f
e
r
en
t
ar
ea
@
I
o
U
f
r
o
m
Fi
g
u
r
e
5
,
th
e
5
0
%
ar
ea
o
v
er
lap
g
i
v
es
t
h
e
m
a
x
i
m
u
m
p
r
ec
is
io
n
,
w
h
ile
7
5
%
ar
ea
o
v
er
lap
d
o
es
n
o
t,
as
o
n
e
w
o
u
ld
t
h
in
k
i
t
w
o
u
ld
.
T
h
e
av
er
ag
e
p
r
ec
is
io
n
o
f
en
t
ir
e
m
o
d
el
is
r
ec
o
r
d
ed
t
o
b
e
2
6
.
2
%
(
m
atch
i
n
g
w
ith
f
i
n
d
in
g
s
i
n
[
6
]
)
an
d
th
is
p
ar
am
eter
h
o
ld
s
v
a
lu
e
as
2
9
.
3
%
f
o
r
p
er
s
o
n
,
w
h
ic
h
is
0
.
1
%
less
th
a
n
th
e
r
ec
o
r
d
e
d
p
r
ec
is
io
n
in
[
6
]
.
A
ls
o
,
w
h
e
n
co
m
p
ar
i
n
g
th
e
v
al
u
e
o
f
t
h
e
av
er
ag
e
p
r
ec
is
io
n
f
o
r
ea
ch
ca
te
g
o
r
y
w
i
th
th
a
t
o
f
t
h
e
m
o
d
el,
n
e
ar
l
y
all
t
h
e
v
al
u
e
s
ar
e
in
r
an
g
e
o
f
p
r
ec
is
io
n
2
6
.
2
%
±
1
0
,
w
h
ic
h
d
r
a
w
s
th
e
co
n
clu
s
io
n
th
a
t
t
h
e
m
o
d
el
h
as
b
ec
o
m
e
s
p
ec
if
ica
ll
y
m
o
r
e
ac
c
u
r
ate
f
o
r
th
e
s
ix
ca
te
g
o
r
ies.
6.
SCO
P
E
Ho
w
e
v
er
,
th
e
r
esu
lt
s
d
o
b
u
ild
p
ath
s
to
d
is
co
v
er
.
T
h
e
f
u
r
th
e
r
w
o
r
k
th
at
ca
n
b
e
ac
co
m
p
lis
h
ed
is
b
y
r
ed
u
cin
g
t
h
e
n
u
m
b
er
o
f
iter
at
i
o
n
s
f
r
o
m
4
9
0
k
to
,
f
o
r
ex
a
m
p
l
e,
1
0
0
k
to
m
ea
s
u
r
e
w
h
eth
er
th
e
p
r
ec
is
io
n
r
e
m
a
in
s
s
a
m
e
o
r
n
o
t.
Als
o
,
t
h
e
e
f
f
ec
ti
v
e
ch
a
n
g
e
i
n
t
h
e
p
r
ec
is
io
n
ca
n
b
e
ch
ec
k
ed
b
y
c
h
a
n
g
in
g
t
h
e
n
u
m
b
er
o
f
o
b
j
ec
ts
o
f
th
e
s
u
b
s
et
u
s
ed
to
f
i
n
e
-
t
u
n
e
th
e
m
o
d
el,
w
h
ile
k
ee
p
i
n
g
t
h
e
iter
atio
n
s
as
4
9
0
k
.
An
o
t
h
e
r
d
ir
ec
tio
n
ca
n
b
e
u
n
co
v
er
ed
b
y
ch
ec
k
i
n
g
t
h
e
ex
ten
t
o
f
t
h
e
p
r
ec
is
io
n
ch
an
g
e
b
y
in
c
o
r
p
o
r
atin
g
b
o
t
h
o
f
th
e
ab
o
v
e
-
m
en
tio
n
ed
ch
a
n
g
es.
ACK
NO
WL
E
D
G
E
M
E
NT
S
T
h
e
au
th
o
r
s
w
i
s
h
to
th
a
n
k
Dr
.
M.
B
.
C
h
an
d
ak
,
P
r
o
f
ess
o
r
a
n
d
Hea
d
o
f
th
e
Dep
ar
tm
e
n
t,
C
o
m
p
u
ter
Scien
ce
a
n
d
E
n
g
i
n
ee
r
in
g
,
Sh
r
i
R
a
m
d
eo
b
ab
a
C
o
lle
g
e
o
f
E
n
g
i
n
ee
r
i
n
g
a
n
d
Ma
n
ag
e
m
e
n
t
,
Na
g
p
u
r
,
f
o
r
h
i
s
co
n
tin
u
o
u
s
en
co
u
r
a
g
e
m
e
n
t
to
p
u
r
s
u
e
t
h
is
r
esear
ch
.
A
l
s
o
,
to
ac
co
m
p
li
s
h
th
is
r
esear
c
h
,
th
e
au
th
o
r
s
i
m
m
en
s
el
y
ap
p
r
ec
iate
th
e
g
u
id
an
ce
a
n
d
co
n
s
ta
n
t
as
s
is
tan
ce
p
r
o
v
id
e
d
b
y
P
r
o
f
es
s
o
r
Neh
a
T
ir
p
u
d
e,
Su
p
er
v
i
s
o
r
an
d
Ass
i
s
tan
t
P
r
o
f
ess
o
r
,
C
o
m
p
u
t
er
Scien
ce
a
n
d
E
n
g
i
n
ee
r
in
g
,
Sh
r
i
R
a
m
d
eo
b
ab
a
C
o
lle
g
e
o
f
E
n
g
i
n
ee
r
in
g
a
n
d
Ma
n
ag
e
m
e
n
t
,
Na
g
p
u
r
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2
0
8
8
-
8708
P
erfo
r
ma
n
ce
ev
a
lu
a
tio
n
o
f fin
e
-
tu
n
ed
fa
s
ter R
-
C
N
N
o
n
s
p
ec
ific MS
C
OC
O
o
b
jects (
Ga
r
im
a
Dev
n
a
n
i)
2555
RE
F
E
R
E
NC
E
S
[1
]
S
.
Ka
m
a
te
a
n
d
N
.
Yilm
a
z
e
r
,
“
Ap
p
li
c
a
ti
o
n
o
f
Ob
jec
t
De
tec
ti
o
n
a
n
d
T
ra
c
k
in
g
T
e
c
h
n
iq
u
e
s
f
o
r
Un
m
a
n
n
e
d
A
e
rial
V
e
h
icle
s
,”
Pro
c
e
d
i
a
C
o
mp
u
ter
S
c
ien
c
e
,
v
ol
.
61
,
p
p
.
4
3
6
-
4
4
1
F
,
2
0
1
5
.
[2
]
D
.
S
o
n
n
tag
,
e
t
a
l
.
,
“
Ov
e
rv
ie
w
o
f
th
e
C
P
S
f
o
r
S
m
a
rt
F
a
c
to
ries
P
ro
jec
t:
De
e
p
L
e
a
rn
in
g
,
Kn
o
w
led
g
e
A
c
q
u
isit
io
n
,
A
n
o
m
a
l
y
De
te
c
ti
o
n
a
n
d
In
tell
ig
e
n
t
Us
e
r
In
terf
a
c
e
s
,”
S
p
rin
g
e
r In
ter
n
a
ti
o
n
a
l
P
u
b
li
sh
i
n
g
,
Ch
a
m
,
p
p
.
4
8
7
-
5
0
4
,
2
0
1
7
.
[3
]
N
.
Be
c
h
e
re
r,
e
t
a
l.
,
“
I
m
p
ro
v
in
g
o
p
ti
m
iza
ti
o
n
o
f
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
s
th
ro
u
g
h
p
a
ra
m
e
t
e
r
f
in
e
-
tu
n
i
n
g
,”
S
p
rin
g
e
r
I
n
tern
a
ti
o
n
a
l
P
u
b
l
ish
i
n
g
,
L
o
n
d
o
n
,
2
0
1
7
.
[4
]
Yo
sin
sk
i
J
.
,
e
t
a
l.
,
“
Ho
w
tran
s
fe
ra
b
le
a
re
fe
a
tu
re
s
in
d
e
e
p
n
e
u
ra
l
n
e
tw
o
rk
s
?
”
Ad
v
a
n
c
e
s
in
n
e
u
ra
l
in
f
o
rm
a
ti
o
n
p
ro
c
e
ss
in
g
sy
ste
ms
(
Pro
c
e
e
d
in
g
s
NIPS
)
,
v
ol
.
27
,
p
p
.
1
-
9
,
2
0
1
4
.
[5
]
R
.
G
irsh
ick
,
e
t
a
l.
,
“
Rich
f
e
a
tu
re
h
iera
rc
h
ies
f
o
r
a
c
c
u
ra
te
o
b
jec
t
d
e
te
c
ti
o
n
a
n
d
se
m
a
n
ti
c
se
g
m
e
n
tatio
n
,”
Pro
c
e
e
d
in
g
s
o
f
t
h
e
IEE
E
c
o
n
fer
e
n
c
e
o
n
c
o
mp
u
ter
v
isio
n
a
n
d
p
a
tt
e
rn
re
c
o
g
n
it
io
n
,
p
p
.
5
8
0
-
5
8
7
,
2
0
1
4
.
[6
]
D
.
S
o
n
n
tag
,
e
t
a
l.
,
“
F
in
e
-
t
u
n
i
n
g
d
e
e
p
CNN
m
o
d
e
ls
o
n
sp
e
c
if
ic
M
S
COCO
c
a
teg
o
ries
,”
Co
mp
u
ti
n
g
Res
e
a
rc
h
Rep
o
sit
o
ry
,
Co
mp
u
ter
Vi
si
o
n
a
n
d
Pa
tt
e
rn
Rec
o
g
n
it
io
n
,
e
p
rin
t
a
rX
iv
:
1
7
0
9
.
0
1
4
7
6
,
2
0
1
7
.
[7
]
S
.
Re
n
,
e
t
a
l.
,
“
F
a
ste
r
R
-
CNN
:
T
o
wa
rd
s
re
a
l
-
ti
m
e
o
b
jec
t
d
e
tec
ti
o
n
w
it
h
re
g
io
n
p
r
o
p
o
sa
l
n
e
tw
o
rk
s
,”
Ad
v
a
n
c
e
s
i
n
n
e
u
ra
l
in
f
o
rm
a
ti
o
n
p
ro
c
e
ss
in
g
sy
ste
ms
,
pp.
91
-
99
,
2
0
1
5
.
[8
]
T
.
Y
.
L
in
,
e
t
a
l.
,
“
M
icro
s
o
f
t
COCO
:
Co
m
m
o
n
o
b
jec
ts
in
c
o
n
t
e
x
t
,”
Eu
ro
p
e
a
n
c
o
n
fer
e
n
c
e
o
n
c
o
mp
u
ter
v
isio
n
,
pp.
7
4
0
-
755
,
2
0
1
4
.
[9
]
R.
G
irsh
ick
,
“
F
a
st R
-
CNN
,”
IEE
E
In
ter
n
a
ti
o
n
a
l
Co
n
fer
e
n
c
e
o
n
C
o
mp
u
ter
V
isio
n
(
ICCV)
,
2
0
1
5
.
[1
0
]
S
.
P
a
lm
e
r,
e
t
a
l.
,
“
Ca
n
o
n
ica
l
p
e
rsp
e
c
ti
v
e
a
n
d
th
e
p
e
rc
e
p
ti
o
n
o
f
o
b
je
c
ts
,”
At
ten
ti
o
n
a
n
d
p
e
rf
o
rm
a
n
c
e
I
X
,
v
ol
.
1
,
1
9
8
1
.
[1
1
]
D
.
De
sa
i,
Blo
g
;
h
tt
p
s://
d
e
sh
a
n
a
d
e
sa
i.
g
it
h
u
b
.
io
/
n
o
tes
/E
v
a
lu
a
ti
o
n
-
of
-
Re
su
lt
s
-
u
sin
g
-
M
e
a
n
-
A
v
g
-
P
re
c
isio
n
.
B
I
O
G
RAP
H
I
E
S
O
F
AUTH
O
RS
G
a
r
i
m
a
De
v
n
a
n
i
is
a
f
in
a
l
-
y
e
a
r
s
tu
d
e
n
t
in
t
h
e
De
p
a
rtm
e
n
t
o
f
Co
m
p
u
ter S
c
ien
c
e
a
n
d
E
n
g
in
e
e
rin
g
a
t
S
h
ri
Ra
m
d
e
o
b
a
b
a
Co
ll
e
g
e
o
f
En
g
in
e
e
rin
g
a
n
d
M
a
n
a
g
e
m
e
n
t.
S
h
e
w
il
l
b
e
re
c
e
iv
in
g
a
Ba
c
h
e
lo
r’s
d
e
g
re
e
in
Co
m
p
u
ter S
c
ien
c
e
a
n
d
En
g
in
e
e
rin
g
f
ro
m
Ra
sh
tras
a
n
t
T
u
k
d
o
ji
M
a
h
a
ra
j
Na
g
p
u
r
Un
iv
e
rsit
y
in
Na
g
p
u
r,
M
a
h
a
ra
sh
tra,
In
d
ia.
S
h
e
h
a
s
c
o
m
p
lete
d
p
ro
jec
ts
in
th
e
f
ield
s
o
f
W
e
b
a
n
d
A
n
d
r
o
id
A
p
p
De
v
e
lo
p
m
e
n
t,
Na
tu
ra
l
L
a
n
g
u
a
g
e
P
ro
c
e
ss
in
g
,
Da
ta
S
c
ien
c
e
,
a
n
d
M
a
c
h
in
e
L
e
a
rn
in
g
.
He
r
c
u
rre
n
t
re
se
a
rc
h
in
tere
sts in
c
lu
d
e
Da
ta S
c
ien
c
e
,
M
a
c
h
in
e
L
e
a
rn
in
g
,
a
n
d
Na
t
u
ra
l
L
a
n
g
u
a
g
e
P
ro
c
e
ss
in
g
.
Ay
u
sh
J
a
isw
a
l
is
a
f
in
a
l
-
y
e
a
r
st
u
d
e
n
t
o
f
th
e
De
p
a
rtme
n
t
o
f
Co
m
p
u
ter
S
c
ien
c
e
a
n
d
E
n
g
in
e
e
rin
g
a
t
S
h
ri
Ra
m
d
e
o
b
a
b
a
Co
l
leg
e
o
f
En
g
in
e
e
rin
g
a
n
d
M
a
n
a
g
e
m
e
n
t.
He
w
il
l
b
e
re
c
e
iv
in
g
a
B
a
c
h
e
lo
r’s
d
e
g
re
e
in
Co
m
p
u
ter S
c
ien
c
e
a
n
d
En
g
in
e
e
rin
g
f
ro
m
Ra
sh
tras
a
n
t
T
u
k
d
o
ji
M
a
h
a
ra
j
Na
g
p
u
r
Un
iv
e
rsit
y
in
Na
g
p
u
r,
M
a
h
a
ra
sh
tra,
In
d
ia.
He
h
a
s
w
o
rk
e
d
o
n
p
ro
jec
ts
in
th
e
f
ield
s
o
f
Da
ta
S
c
ien
c
e
,
a
n
d
M
a
c
h
in
e
L
e
a
rn
in
g
.
His
c
u
rre
n
t
re
se
a
rc
h
in
tere
sts
in
c
lu
d
e
A
rti
f
icia
l
In
telli
g
e
n
c
e
,
a
n
d
M
a
c
h
i
n
e
L
e
a
rn
in
g
.
He
p
lan
s to
p
u
rs
u
e
h
is
M
a
ste
r
o
f
T
e
c
h
n
o
lo
g
y
in
Co
m
p
u
t
e
r
S
c
ien
c
e
in
n
e
a
r
f
u
tu
re
.
Ro
s
h
n
i
J
o
h
n
is
a
f
in
a
l
-
y
e
a
r
stu
d
e
n
t
i
n
th
e
De
p
a
rtm
e
n
t
o
f
Co
m
p
u
ter
S
c
ien
c
e
a
n
d
E
n
g
in
e
e
rin
g
a
t
S
h
ri
Ra
m
d
e
o
b
a
b
a
Co
ll
e
g
e
o
f
En
g
in
e
e
rin
g
a
n
d
M
a
n
a
g
e
m
e
n
t.
S
h
e
w
il
l
b
e
re
c
e
iv
in
g
a
Ba
c
h
e
lo
r’s
d
e
g
re
e
in
Co
m
p
u
ter S
c
ien
c
e
a
n
d
En
g
in
e
e
rin
g
f
ro
m
Ra
sh
tras
a
n
t
T
u
k
d
o
ji
M
a
h
a
ra
j
Na
g
p
u
r
Un
iv
e
rsit
y
in
Na
g
p
u
r,
M
a
h
a
ra
sh
tra,
In
d
ia.
S
h
e
h
a
s
c
o
m
p
lete
d
p
ro
jec
ts
in
th
e
f
ield
s
o
f
Bu
sin
e
ss
In
t
e
ll
ig
e
n
c
e
,
Da
ta
S
c
ien
c
e
,
D
e
e
p
L
e
a
rn
in
g
a
n
d
Cy
b
e
r
S
e
c
u
rit
ies
.
He
r
c
u
rre
n
t
re
se
a
rc
h
in
tere
sts
in
c
lu
d
e
Da
ta
S
c
ien
c
e
,
De
e
p
L
e
a
rn
in
g
,
a
n
d
W
e
b
S
e
c
u
rit
ies
.
Ra
ja
t
Cha
u
r
a
sia
is
a
f
in
a
l
-
y
e
a
r
s
tu
d
e
n
t
in
th
e
De
p
a
rtm
e
n
t
o
f
Co
m
p
u
ter S
c
ien
c
e
a
n
d
En
g
in
e
e
rin
g
a
t
S
h
ri
Ra
m
d
e
o
b
a
b
a
Co
l
leg
e
o
f
En
g
in
e
e
rin
g
a
n
d
M
a
n
a
g
e
m
e
n
t.
He
w
il
l
b
e
re
c
e
iv
in
g
a
B
a
c
h
e
lo
r’s
d
e
g
re
e
in
Co
m
p
u
ter S
c
ien
c
e
a
n
d
En
g
in
e
e
rin
g
f
ro
m
Ra
sh
tras
a
n
t
T
u
k
d
o
ji
M
a
h
a
ra
j
Na
g
p
u
r
Un
iv
e
rsit
y
in
Na
g
p
u
r,
M
a
h
a
ra
sh
tra,
I
n
d
ia.
H
e
h
a
s
w
o
rk
e
d
o
n
n
u
m
e
ro
u
s
p
r
o
jec
ts
in
f
ield
s
o
f
W
e
b
De
v
e
lo
p
m
e
n
t
a
n
d
De
sig
n
,
E
n
terp
rise
A
p
p
li
c
a
t
io
n
De
v
e
lo
p
m
e
n
t,
Da
ta
A
n
a
l
y
ti
c
s,
a
n
d
M
a
c
h
in
e
L
e
a
rn
in
g
.
His
c
u
rre
n
t
re
se
a
rc
h
in
tere
sts
in
c
l
u
d
e
G
ra
p
h
ics
De
sig
n
in
g
,
a
n
d
M
a
c
h
i
n
e
L
e
a
rn
in
g
.
He
p
lan
s
t
o
p
u
rsu
e
h
is
M
a
ste
r’s d
e
g
re
e
in
n
e
a
r
f
u
tu
re
.
Ne
h
a
Tir
p
u
d
e
is
a
n
As
sista
n
t
P
r
o
f
e
ss
o
r
in
th
e
De
p
a
rtme
n
t
o
f
Co
m
p
u
ter
S
c
ien
c
e
a
n
d
En
g
in
e
e
rin
g
at
S
h
ri
Ra
m
d
e
o
b
a
b
a
Co
ll
e
g
e
o
f
En
g
in
e
e
rin
g
a
n
d
M
a
n
a
g
e
m
e
n
t.
S
h
e
re
c
e
iv
e
d
th
e
M
a
ste
r
o
f
T
e
c
h
n
o
lo
g
y
a
n
d
Ba
c
h
e
lo
r
o
f
En
g
i
n
e
e
rin
g
d
e
g
re
e
in
Co
m
p
u
ter
S
c
ie
n
c
e
a
n
d
E
n
g
in
e
e
rin
g
i
n
2
0
1
3
a
n
d
2
0
1
0
re
sp
e
c
ti
v
e
ly
.
H
e
r
c
u
rre
n
t
re
se
a
rc
h
a
n
d
p
u
b
li
c
a
ti
o
n
in
ter
e
sts
in
c
lu
d
e
b
i
o
-
m
e
d
ica
l
i
m
a
g
e
p
ro
c
e
ss
in
g
,
d
a
ta m
in
in
g
a
n
d
in
f
o
r
m
a
ti
o
n
se
c
u
rit
y
.
Evaluation Warning : The document was created with Spire.PDF for Python.