I
nd
o
ne
s
ia
n J
o
urna
l o
f
E
lect
rica
l En
g
ineering
a
nd
Co
m
pu
t
er
Science
Vo
l.
40
,
No
.
2
,
N
o
v
em
b
e
r
2
0
2
5
,
p
p
.
7
1
9
~
7
3
4
I
SS
N:
2
5
0
2
-
4
7
5
2
,
DOI
: 1
0
.
1
1
5
9
1
/ijeecs.v
40
.i
2
.
p
p
719
-
7
3
4
719
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ee
cs.ia
esco
r
e.
co
m
End
-
to
-
en
d sy
ste
m for t
ra
nsla
ting
b
a
ha
sa
isya
r
a
t Ind
o
nesia
sig
n
la
ng
ua
g
e ges
tures int
o
I
ndo
nesia
n
t
ex
t
Sa
t
ria
P
utr
a
,
E
rdef
i R
a
k
un
D
e
p
a
r
t
me
n
t
o
f
C
o
mp
u
t
e
r
S
c
i
e
n
c
e
,
F
a
c
u
l
t
y
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
,
U
n
i
v
e
r
si
t
y
o
f
I
n
d
o
n
e
si
a
,
D
e
p
o
k
,
I
n
d
o
n
e
s
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Dec
10
,
2
0
2
4
R
ev
is
ed
J
u
l
7
,
2
0
2
5
Acc
ep
ted
Oct
14
,
2
0
2
5
Th
is
stu
d
y
a
d
d
re
ss
e
s
c
rit
ica
l
c
h
a
l
len
g
e
s
in
d
e
v
e
l
o
p
i
n
g
a
n
e
n
d
-
to
-
e
n
d
b
a
h
a
s
a
isy
a
ra
t
In
d
o
n
e
sia
(
BIS
IND
O
)
S
LT
b
y
in
te
g
ra
ti
n
g
a
d
v
a
n
c
e
d
d
e
e
p
lea
rn
in
g
tec
h
n
iq
u
e
s
to
o
v
e
rc
o
m
e
c
o
m
p
l
e
x
b
a
c
k
g
ro
u
n
d
i
n
terfe
re
n
c
e
,
tr
a
n
siti
o
n
a
l
g
e
stu
re
re
c
o
g
n
it
i
o
n
,
a
n
d
li
m
it
a
ti
o
n
s
in
d
a
tas
e
t
a
v
a
il
a
b
il
it
y
.
Wh
il
e
e
x
isti
n
g
S
LT
sy
ste
m
s
stru
g
g
le
with
i
so
late
d
wo
rd
re
c
o
g
n
it
io
n
a
n
d
m
a
n
u
a
l
p
re
p
ro
c
e
ss
in
g
,
o
u
r
wo
r
k
i
n
tro
d
u
c
e
s
th
re
e
k
e
y
i
n
n
o
v
a
ti
o
n
s:
(
1
)
imp
lem
e
n
tatio
n
o
f
YO
LOv
8
fo
r
o
p
ti
m
ize
d
o
b
jec
t
d
e
tec
ti
o
n
,
a
c
h
iev
in
g
8
8
%
m
AP
a
n
d
re
d
u
c
i
n
g
WE
R
to
1
1
.
4
0
%
,
o
u
tp
e
rfo
rm
i
n
g
YO
LOv
5
/v
7
i
n
h
a
n
d
li
n
g
c
o
m
p
lex
b
a
c
k
g
ro
u
n
d
s;
(2
)
a
u
t
o
m
a
ted
re
m
o
v
a
l
o
f
tran
siti
o
n
a
l
g
e
stu
re
s
u
sin
g
Th
re
sh
o
l
d
c
o
n
d
it
io
n
a
l
ra
n
d
o
m
field
s
(
TCRF
)
,
wh
ic
h
a
tt
a
in
e
d
9
5
.
6
8
%
a
c
c
u
ra
c
y
,
sig
n
ifi
c
a
n
tl
y
imp
r
o
v
i
n
g
u
p
o
n
M
o
b
i
leN
e
tV2
’s
p
e
rfo
rm
a
n
c
e
(W
ER:
6
.
8
9
%
v
s.
9
3
.
5
3
%
);
a
n
d
(3
)
e
n
d
-
to
-
e
n
d
BIS
IND
O
S
LT
b
y
e
x
p
a
n
s
io
n
o
f
th
e
BIS
IND
O
d
a
tas
e
t
t
o
4
3
5
wo
rd
lab
e
ls,
e
n
a
b
li
n
g
c
o
m
p
re
h
e
n
si
v
e
se
n
ten
c
e
-
lev
e
l
tran
sla
ti
o
n
.
Ex
p
e
rime
n
tal
re
su
lt
s
d
e
m
o
n
stra
te
th
e
s
y
ste
m
’s
ro
b
u
stn
e
ss
,
with
8
.
3
1
%
o
f
WE
R,
8
4
.
1
3
%
o
f
S
Ac
c
,
a
n
d
8
7
.
0
8
%
o
f
S
a
c
re
BLE
U
a
fter
d
a
tas
e
t
e
x
p
a
n
si
o
n
a
n
d
re
d
u
n
d
a
n
c
y
e
li
m
in
a
ti
o
n
th
r
o
u
g
h
g
r
o
u
p
in
g
m
e
th
o
d
s.
Th
e
p
ro
p
o
se
d
fra
m
e
wo
rk
o
p
e
ra
t
e
s
with
o
u
t
m
a
n
u
a
l
in
ter
v
e
n
ti
o
n
,
m
a
rk
i
n
g
a
su
b
sta
n
ti
a
l
a
d
v
a
n
c
e
m
e
n
t
t
o
wa
rd
r
e
a
l
-
wo
rld
a
p
p
li
c
a
b
il
it
y
.
K
ey
w
o
r
d
s
:
B
a
h
a
s
a
is
ya
r
a
t
I
n
d
o
n
esia
Ob
ject
d
etec
tio
n
Sig
n
lan
g
u
a
g
e
tr
an
s
latio
n
T
h
r
esh
o
ld
c
o
n
d
itio
n
al
r
an
d
o
m
f
ield
s
YOL
Ov
7
YOL
Ov
8
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
E
r
d
ef
i Rak
u
n
Dep
ar
tm
en
t o
f
C
o
m
p
u
ter
Scie
n
ce
,
Facu
lty
o
f
C
o
m
p
u
ter
Scie
n
ce
,
Un
iv
er
s
ity
o
f
I
n
d
o
n
esia
1
6
4
2
4
Dep
o
k
,
I
n
d
o
n
esia
E
m
ail: e
f
i@
cs.u
i.a
c.
id
1.
I
NT
RO
D
UCT
I
O
N
Dev
elo
p
in
g
in
clu
s
iv
e
c
o
m
m
u
n
icatio
n
to
o
ls
f
o
r
th
e
d
ea
f
c
o
m
m
u
n
ity
r
ep
r
esen
ts
an
u
r
g
en
t
n
ec
ess
ity
,
as
th
ey
f
ac
e
c
o
m
p
lex
ch
allen
g
es
in
d
aily
in
ter
ac
tio
n
s
.
Hea
r
in
g
im
p
ai
r
m
en
t,
w
h
eth
er
c
o
n
g
en
ital
o
r
ac
q
u
ir
e
d
[
1
]
,
o
f
ten
f
o
r
ce
s
d
ea
f
in
d
iv
i
d
u
als
to
r
ely
p
r
im
ar
ily
o
n
v
is
u
al
co
m
m
u
n
icatio
n
m
eth
o
d
s
,
s
u
ch
as
s
ig
n
lan
g
u
ag
e,
f
o
r
ef
f
ec
tiv
e
co
m
m
u
n
icatio
n
[
2
]
.
Ho
wev
e
r
,
th
ese
m
eth
o
d
s
en
co
u
n
ter
s
ig
n
if
ica
n
t
o
b
s
tacle
s
,
in
clu
d
in
g
lim
ited
p
u
b
lic
u
n
d
er
s
tan
d
in
g
an
d
a
la
ck
o
f
p
r
ac
tical
co
m
m
u
n
icatio
n
aid
s
.
C
u
r
r
en
t
s
ig
n
lan
g
u
ag
e
tr
an
s
latio
n
(
SLT
)
s
o
lu
tio
n
s
b
ased
o
n
C
NNs
[
3
]
–
[
8
]
,
r
e
m
ain
lim
ited
to
r
e
co
g
n
iz
in
g
in
d
iv
i
d
u
al
letter
s
/wo
r
d
s
with
o
u
t f
u
ll sen
ten
ce
tr
an
s
latio
n
ca
p
ab
ilit
ies.
T
h
ese
lim
itatio
n
s
u
n
d
er
s
co
r
e
th
e
cr
u
cial
n
ee
d
f
o
r
an
e
n
d
-
to
-
e
n
d
s
y
s
tem
th
at
ca
n
tr
an
s
late
co
m
p
lete
s
en
ten
ce
s
in
r
ea
l
-
tim
e
ac
r
o
s
s
v
a
r
io
u
s
c
o
n
d
itio
n
s
with
o
u
t
r
eq
u
ir
i
n
g
h
u
m
an
i
n
ter
v
en
tio
n
.
I
m
p
lem
en
tin
g
s
u
c
h
a
s
o
lu
tio
n
wo
u
ld
b
e
tr
an
s
f
o
r
m
ativ
e
f
o
r
t
h
e
d
ea
f
co
m
m
u
n
ity
,
e
n
ab
lin
g
im
p
r
o
v
e
d
ac
ce
s
s
to
ed
u
ca
ti
o
n
an
d
em
p
lo
y
m
en
t,
f
ac
ilit
atin
g
s
o
cial
in
ter
ac
tio
n
s
,
an
d
p
r
o
m
o
tin
g
g
r
ea
te
r
s
o
cieta
l
in
clu
s
io
n
.
T
h
is
ad
v
an
ce
d
s
y
s
tem
wo
u
l
d
n
o
t
o
n
ly
o
v
er
co
m
e
ex
is
tin
g
s
o
lu
tio
n
s
h
o
r
tc
o
m
in
g
s
b
u
t
also
estab
lis
h
a
m
o
r
e
n
atu
r
al
co
m
m
u
n
icatio
n
b
r
id
g
e
b
etwe
e
n
th
e
d
ea
f
co
m
m
u
n
ity
an
d
th
e
h
ea
r
in
g
s
o
ciety
.
T
wo
m
ain
s
ig
n
lan
g
u
ag
es
ar
e
u
s
ed
in
I
n
d
o
n
esia:
S
is
tem
is
ya
r
a
t
B
a
h
a
s
a
I
n
d
o
n
esia
(
SIBI
)
an
d
b
a
h
a
s
a
isy
a
ra
t
In
d
o
n
e
sia
(
B
I
SIN
DO
)
.
SIBI,
th
e
o
f
f
icial
s
ig
n
lan
g
u
a
g
e
r
ec
o
g
n
ized
b
y
law,
f
o
llo
ws
th
e
g
r
am
m
atica
l
s
tr
u
ctu
r
e
o
f
s
p
o
k
en
I
n
d
o
n
esi
an
.
I
n
co
n
tr
ast,
B
I
SIN
DO
h
a
s
ev
o
lv
ed
o
r
g
a
n
ically
with
in
d
ea
f
co
m
m
u
n
ities
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
40
,
No
.
2
,
No
v
em
b
er
20
25
:
7
1
9
-
7
3
4
720
f
ea
tu
r
in
g
a
m
o
r
e
in
tu
itiv
e
g
r
am
m
ar
a
n
d
r
eg
io
n
al
v
ar
iatio
n
s
[
9
]
.
Desp
ite
its
wid
esp
r
ea
d
u
s
e,
B
I
SIN
DO
r
em
ain
s
u
n
d
er
-
r
esear
ch
ed
c
o
m
p
ar
ed
to
SIBI,
h
ig
h
lig
h
tin
g
a
cr
itical
g
ap
in
th
e
liter
atu
r
e.
Hen
ce
,
th
is
r
esear
ch
aim
s
to
d
ev
elo
p
an
en
d
-
to
-
e
n
d
ap
p
licatio
n
to
tr
a
n
s
late
B
I
SI
NDO
s
ig
n
lan
g
u
ag
e
in
to
a
s
en
ten
ce
in
r
ea
l
-
wo
r
ld
co
n
d
itio
n
s
.
SLT
p
r
esen
ts
th
r
ee
s
ig
n
if
ican
t
ch
allen
g
es.
First,
tr
an
s
itio
n
al
g
estu
r
es,
wh
ich
ar
e
wo
r
d
s
in
a
s
en
ten
ce
th
at
ca
r
r
y
n
o
s
p
ec
if
ic
m
ea
n
i
n
g
,
ca
n
b
e
ch
allen
g
i
n
g
to
r
e
co
g
n
ize
d
u
e
to
th
eir
v
a
r
iab
il
ity
,
wh
ich
r
ed
u
ce
s
tr
an
s
latio
n
ac
cu
r
ac
y
.
Fo
r
e
x
a
m
p
le,
in
th
e
B
I
SIN
DO
s
en
ten
ce
"
Dia
La
r
i
K
a
r
en
a
Ta
ku
t
"
(
"He
r
u
n
s
b
ec
au
s
e
h
e
is
s
ca
r
ed
")
,
th
ese
m
ea
n
in
g
less
tr
an
s
itio
n
s
o
cc
u
r
at
th
e
s
tar
t,
b
etwe
en
wo
r
d
s
,
an
d
at
th
e
en
d
(
Fig
u
r
e
1
)
.
Sin
ce
th
ey
ac
t
as
ir
r
elev
an
t
in
p
u
ts
in
co
n
tin
u
o
u
s
s
ig
n
in
g
,
au
to
m
ated
r
em
o
v
al
o
f
s
u
ch
g
estu
r
e
s
ca
n
en
h
an
ce
th
e
s
y
s
tem
’
s
wo
r
d
r
ec
o
g
n
itio
n
p
e
r
f
o
r
m
an
ce
.
Fig
u
r
e
1
.
T
r
an
s
itio
n
al
g
estu
r
es
Ad
d
r
ess
in
g
th
e
is
s
u
e
o
f
tr
a
n
s
itio
n
al
g
estu
r
es,
r
esear
ch
b
y
[
1
0
]
u
s
ed
th
e
Au
t
o
m
ated
E
p
en
th
esis
R
em
o
v
al
m
eth
o
d
with
T
C
R
F
in
SIBI,
s
ig
n
if
ican
tly
r
ed
u
ci
n
g
th
e
wo
r
d
e
r
r
o
r
r
ate
(
W
E
R
)
to
3
.
3
6
7
%
an
d
im
p
r
o
v
in
g
s
en
ten
ce
ac
c
u
r
ac
y
(
SAcc)
to
8
0
.
1
8
%.
H
o
wev
er
,
th
is
m
eth
o
d
h
as
n
o
t
y
et
b
ee
n
ap
p
lied
to
B
I
SIN
DO.
T
h
e
s
ec
o
n
d
ch
allen
g
e
f
o
r
SLT
is
th
e
p
r
esen
ce
o
f
co
m
p
lex
b
ac
k
g
r
o
u
n
d
s
.
I
n
r
ea
l
-
wo
r
ld
s
ce
n
ar
io
s
,
s
ig
n
lan
g
u
ag
e
is
o
f
ten
p
er
f
o
r
m
ed
i
n
en
v
ir
o
n
m
en
ts
with
v
ar
y
i
n
g
b
ac
k
g
r
o
u
n
d
s
,
s
u
ch
as
s
h
o
p
p
in
g
m
alls
,
s
ch
o
o
ls
,
o
r
p
u
b
lic
s
p
ac
es.
T
h
ese
co
m
p
l
ex
b
ac
k
g
r
o
u
n
d
s
ca
n
s
ig
n
if
ic
an
tly
in
ter
f
er
e
with
ac
c
u
r
ate
ly
s
eg
m
en
tin
g
k
e
y
f
ea
tu
r
es,
s
u
ch
as
f
ac
ial
e
x
p
r
es
s
io
n
s
an
d
h
an
d
m
o
v
em
en
ts
,
w
h
ich
ar
e
c
r
itical
f
o
r
in
te
r
p
r
etin
g
s
ig
n
lan
g
u
ag
e.
I
n
co
n
tr
ast,
a
p
lain
g
r
ee
n
b
ac
k
g
r
o
u
n
d
p
r
o
v
id
es
clea
r
an
d
d
is
tin
ct
f
ea
tu
r
es
f
o
r
s
eg
m
en
tatio
n
,
an
d
co
m
p
lex
en
v
ir
o
n
m
en
ts
in
tr
o
d
u
ce
n
o
is
e
an
d
v
is
u
al
d
is
tr
ac
tio
n
s
th
at
r
ed
u
ce
th
e
s
y
s
tem
'
s
ac
cu
r
ac
y
(
Fig
u
r
e
2
)
.
T
h
is
n
o
is
e
ca
n
lead
to
m
is
in
ter
p
r
etatio
n
s
o
r
f
ailu
r
es in
r
ec
o
g
n
izin
g
s
ig
n
s
,
u
ltima
tely
h
in
d
er
in
g
th
e
s
y
s
tem
'
s
p
er
f
o
r
m
an
ce
.
Fig
u
r
e
2
.
C
o
m
p
le
x
b
ac
k
g
r
o
u
n
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
E
n
d
-
to
-
e
n
d
s
ystem
fo
r
tr
a
n
s
la
tin
g
b
a
h
a
s
a
is
ya
r
a
t I
n
d
o
n
esia
s
ig
n
la
n
g
u
a
g
e
g
estu
r
es
…
(
S
a
tr
ia
P
u
tr
a
)
721
R
esear
ch
by
[
1
1
]
ad
d
r
ess
ed
th
e
ch
allen
g
e
o
f
co
m
p
le
x
b
ac
k
g
r
o
u
n
d
s
in
B
I
SIN
DO
u
s
in
g
t
h
e
Fas
ter
R
-
C
NN
[
1
2
]
an
d
YOL
Ov
5
[
1
3
]
o
b
ject
d
etec
tio
n
m
eth
o
d
s
.
E
ac
h
m
eth
o
d
ac
h
iev
e
d
a
m
ea
n
av
er
ag
e
p
r
ec
is
io
n
(
m
AP)
s
co
r
e
o
f
7
1
.
3
%
an
d
7
4
.
1
%,
t
h
o
u
g
h
YOL
Ov
5
o
u
t
p
er
f
o
r
m
ed
Fas
ter
R
-
C
NN
in
W
E
R
at
1
6
.
4
2
%
,
Sacr
eBLE
U
at
6
7
.
7
7
%
,
an
d
S
Acc
at
4
9
.
2
9
%.
I
n
th
is
s
tu
d
y
,
th
e
tr
an
s
itio
n
g
estu
r
es
wer
e
m
a
n
u
ally
r
em
o
v
e
d
.
A
s
tu
d
y
o
n
SIBI
b
y
[
1
4
]
also
e
m
p
lo
y
ed
R
etin
aNe
t
[
1
5
]
f
o
r
o
b
ject
d
etec
tio
n
,
ac
h
iev
in
g
a
W
E
R
o
f
4
.
1
%
an
d
a
n
SAcc
o
f
7
8
.
9
9
%.
W
h
ile
th
ese
r
esu
lts
d
em
o
n
s
tr
ate
p
r
o
g
r
ess
in
h
a
n
d
lin
g
co
m
p
le
x
b
ac
k
g
r
o
u
n
d
s
,
t
h
er
e
is
s
till
r
o
o
m
f
o
r
im
p
r
o
v
e
m
en
t.
T
h
e
p
er
f
o
r
m
an
ce
m
etr
ics,
s
u
ch
as
m
AP,
W
E
R
,
an
d
SAcc,
c
an
p
o
ten
tially
b
e
en
h
an
ce
d
b
y
lev
er
a
g
in
g
m
o
r
e
ad
v
an
ce
d
o
b
ject
d
etec
tio
n
m
eth
o
d
s
lik
e
YOL
Ov
7
an
d
YO
L
Ov
8
.
T
h
ese
latest
m
o
d
els
o
f
f
e
r
im
p
r
o
v
ed
ac
cu
r
ac
y
,
s
p
ee
d
,
a
n
d
r
o
b
u
s
tn
ess
in
d
etec
tin
g
a
n
d
s
eg
m
e
n
tin
g
o
b
jects
in
co
m
p
le
x
en
v
ir
o
n
m
en
ts
.
T
h
e
th
ir
d
c
h
allen
g
e
SLT
f
ac
es
,
esp
ec
ially
in
B
I
SIN
DO,
is
th
e
lack
o
f
d
atasets
.
T
h
e
latest
r
esear
ch
b
y
[
1
1
]
o
n
l
y
u
s
ed
4
0
s
en
ten
ce
s
with
1
5
2
wo
r
d
s
,
wh
ile
o
th
er
s
tu
d
ies,
s
u
ch
as
[
1
6
]
–
[
1
8
]
,
s
ti
ll
u
tili
ze
d
as
letter
d
atasets
,
n
o
t
as
s
en
ten
ce
s
.
T
h
is
d
ataset
lim
itatio
n
f
u
n
d
am
en
tally
co
n
s
tr
ain
s
s
y
s
tem
d
ev
elo
p
m
en
t
a
n
d
p
er
f
o
r
m
an
ce
v
alid
ati
o
n
,
n
ec
ess
itatin
g
th
e
ex
p
an
s
io
n
o
f
th
e
d
ataset
to
cr
ea
te
an
en
d
-
to
-
en
d
B
I
SIN
DO
SLT
.
Ad
d
r
ess
in
g
th
e
id
en
tifie
d
ch
al
len
g
es
in
SLT
s
y
s
tem
s
,
p
ar
ticu
lar
ly
f
o
r
B
I
SIN
DO,
th
is
s
tu
d
y
d
ev
elo
p
s
an
en
d
-
to
-
e
n
d
tr
a
n
s
latio
n
ap
p
l
icatio
n
th
at
tack
les
t
h
r
ee
c
r
itical
ch
allen
g
es:
(
1
)
Data
s
et
lim
i
tatio
n
s
-
ex
p
an
d
i
n
g
b
ey
o
n
d
p
r
ev
io
u
s
co
n
s
tr
ain
ed
d
atasets
to
cr
ea
te
a
m
o
r
e
c
o
m
p
r
eh
en
s
iv
e
B
I
SIN
DO
r
ep
r
e
s
en
tatio
n
;
(
2
)
Ob
ject
d
etec
tio
n
o
p
tim
izatio
n
-
i
m
p
lem
en
tin
g
c
u
ttin
g
-
ed
g
e
YOL
Ov
7
[
1
9
]
an
d
YOL
Ov
8
[
2
0
]
ar
ch
itectu
r
es
to
en
h
an
ce
m
AP
an
d
o
v
er
c
o
m
e
co
m
p
lex
b
ac
k
g
r
o
u
n
d
in
ter
f
e
r
en
ce
;
an
d
(
3
)
Au
to
m
ated
ep
en
th
esis
r
em
o
v
al
-
ad
ap
tin
g
th
e
T
C
R
F
m
eth
o
d
[
2
1
]
to
B
I
SIN
DO
f
o
r
elim
in
atin
g
tr
an
s
itio
n
al
g
estu
r
es,
th
e
r
eb
y
im
p
r
o
v
in
g
m
o
d
el
tr
ain
in
g
an
d
s
y
s
tem
ac
cu
r
ac
y
.
T
h
ese
in
teg
r
ated
s
o
lu
tio
n
s
en
ab
le
f
u
lly
au
to
m
ate
d
o
p
e
r
a
tio
n
wh
ile
m
ak
in
g
th
r
ee
k
e
y
c
o
n
tr
ib
u
tio
n
s
:
f
ir
s
t,
estab
lis
h
in
g
a
r
o
b
u
s
t
B
I
SIN
DO
d
ataset;
s
ec
o
n
d
,
a
d
v
an
cin
g
d
etec
tio
n
ca
p
ab
ilit
ies
th
r
o
u
g
h
s
tate
-
of
-
th
e
-
ar
t
YOL
O
m
o
d
els;
an
d
th
i
r
d
,
p
io
n
ee
r
in
g
tr
a
n
s
itio
n
al
g
e
s
tu
r
e
p
r
o
ce
s
s
in
g
in
B
I
SIN
DO
th
r
o
u
g
h
au
t
o
m
ated
ep
en
th
esis
r
em
o
v
al.
2.
RE
L
AT
E
D
WO
RK
S
T
ab
le
1
p
r
esen
ts
s
ev
er
al
s
tu
d
i
es
r
elate
d
to
SLT
s
y
s
tem
s
.
in
th
e
c
o
n
tex
t
o
f
I
n
d
o
n
esian
s
ig
n
lan
g
u
ag
e
an
d
f
o
r
eig
n
s
ig
n
lan
g
u
ag
es,
wh
ich
h
av
e
p
r
o
v
id
ed
v
alu
ab
l
e
th
eo
r
etica
l
f
o
u
n
d
ati
o
n
s
to
s
u
p
p
o
r
t
th
e
u
p
co
m
i
n
g
r
esear
ch
.
T
ab
le
1
.
R
elate
d
wo
r
k
s
in
th
e
f
ield
o
f
SLT
s
y
s
tem
s
A
u
t
h
o
r
S
u
mm
a
r
y
M
e
t
h
o
d
D
a
t
a
s
e
t
R
e
s
u
l
t
s
(
%)
[
2
2
]
D
e
e
p
C
N
N
f
o
r
f
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
o
n
S
I
B
I
R
e
s
n
e
t
5
0
a
n
d
M
o
b
i
l
e
N
e
t
v
2
S
I
B
I
S
e
n
t
e
n
c
e
s
A
c
c
:
9
9
.
8
9
[
2
3
]
B
I
S
I
N
D
O
l
e
t
t
e
r
l
e
v
e
l
u
si
n
g
Y
O
LO
v
3
Y
O
LO
v
3
B
I
S
I
N
D
O
S
e
n
t
e
n
c
e
s
I
mag
e
:
1
0
0
f
1
-
sc
o
r
e
;
V
i
d
e
o
:
8
4
.
3
8
f1
-
sc
o
r
e
[
2
4
]
M
o
d
i
f
i
e
d
s
k
e
l
e
t
o
n
,
d
e
t
e
c
t
i
o
n
h
e
a
d
,
a
n
d
l
o
ss f
u
n
c
t
i
o
n
o
f
Y
O
LO
v
5
Y
O
LO
v
5
c
o
mb
i
n
e
d
w
i
t
h
t
h
e
LST
M
n
e
t
w
o
r
k
a
n
d
O
p
e
n
P
o
s
e
M
a
n
d
a
r
i
n
W
o
r
d
s
A
c
c
:
9
8
.
8
7
[
2
5
]
U
si
n
g
G
o
o
g
l
e
M
e
d
i
a
P
i
p
e
t
o
e
x
t
r
a
c
t
k
e
y
p
o
i
n
t
s fr
o
m
t
h
e
p
a
l
m
LSTM
B
a
n
g
l
a
S
e
n
t
e
n
c
e
s
A
c
c
:
9
2
.
0
7
[
2
6
]
S
i
g
n
L
a
n
g
u
a
g
e
D
e
t
e
c
t
i
o
n
U
si
n
g
C
N
N
-
Y
O
LO
v
8
C
N
N
-
Y
O
LO
v
8
I
n
d
i
a
n
W
o
r
d
A
c
c
:
9
8
.
9
[
2
7
]
M
a
l
a
y
si
a
n
S
i
g
n
La
n
g
u
a
g
e
M
o
b
i
l
e
A
p
p
l
i
c
a
t
i
o
n
U
s
i
n
g
D
e
e
p
Le
a
r
n
i
n
g
Y
O
LO
v
5
M
a
l
a
y
si
a
n
S
i
g
n
La
n
g
u
a
g
e
mA
P
:
8
7
.
2
3
T
C
R
F
[
2
8
]
a
f
r
am
ewo
r
k
th
at
co
n
s
id
er
s
th
r
esh
o
ld
s
b
y
a
d
d
in
g
lab
els
f
o
r
n
o
n
-
s
ig
n
G
p
atter
n
s
in
th
e
o
r
ig
in
al
C
R
F
an
d
u
tili
zin
g
t
h
e
weig
h
ts
f
r
o
m
th
e
tr
an
s
itio
n
f
u
n
ctio
n
s
an
d
ex
is
tin
g
f
ea
tu
r
es
in
th
e
C
R
F.
T
h
er
ef
o
r
e,
T
C
R
F
in
clu
d
es
lab
els
=
{
1
,
…,
,
},
wh
er
e
(
1
,
.
.
,
l)
r
e
f
er
s
to
th
e
n
u
m
b
er
o
f
C
R
F
lab
els.
T
h
e
weig
h
t o
f
th
e
f
ea
tu
r
e
f
u
n
cti
o
n
f
o
r
th
e
n
o
n
-
s
ig
n
lab
el
G,
,
i
s
ca
lcu
lated
u
s
in
g
E
q
u
atio
n
1
.
T
h
u
s
,
d
eter
m
in
in
g
th
is
th
r
esh
o
ld
v
alu
e
is
cr
u
cial
in
o
p
tim
izin
g
s
y
s
tem
p
er
f
o
r
m
a
n
ce
.
(
)
=
̅
+
√
(
1
)
th
e
v
alu
e
̅
=
∑
1
=
1
(
)
,
wh
er
e
k
is
th
e
n
u
m
b
er
o
f
C
R
F
lab
els,
an
d
is
t
h
e
v
ar
ian
ce
o
f
weig
h
t
m
.
Sev
er
al
s
tu
d
ies
th
at
im
p
le
m
en
t
T
C
R
F,
s
u
ch
as
[
2
9
]
to
c
ateg
o
r
ize
h
u
m
an
ac
tio
n
s
,
[
3
0
]
to
r
ec
o
g
n
ize
f
ac
ial
ex
p
r
ess
io
n
s
,
an
d
[
3
1
]
to
in
ter
p
r
et
h
u
m
an
b
o
d
y
m
o
v
em
en
ts
f
o
r
r
o
b
o
t c
o
m
m
a
n
d
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
40
,
No
.
2
,
No
v
em
b
er
20
25
:
7
1
9
-
7
3
4
722
3.
M
E
T
H
O
D
Dev
elo
p
in
g
an
en
d
-
to
-
en
d
a
p
p
licatio
n
f
o
r
t
r
an
s
latin
g
B
I
SI
NDO
s
ig
n
lan
g
u
ag
e
in
to
tex
t
r
eq
u
ir
es
a
co
m
p
r
eh
e
n
s
iv
e
ap
p
r
o
ac
h
en
c
o
m
p
ass
in
g
s
ev
er
al
cr
u
cia
l
s
t
ag
es.
T
h
ese
s
tag
es
in
clu
d
e
d
ataset
p
r
ep
ar
atio
n
,
ex
p
er
im
en
tal
d
esig
n
,
an
d
r
esu
l
t e
v
alu
atio
n
,
as illu
s
tr
ated
in
F
ig
u
r
e
3
.
Fig
u
r
e
3
.
Me
th
o
d
o
lo
g
y
I
n
th
e
d
ataset
p
r
ep
ar
atio
n
s
tag
e,
we
will
ex
p
lain
h
o
w
t
h
e
d
at
aset
f
o
r
th
is
r
esear
ch
was
co
ll
ec
ted
,
h
o
w
th
e
d
ataset
was
p
r
ep
r
o
ce
s
s
ed
,
an
d
wh
at
an
n
o
tatio
n
s
wer
e
in
v
o
lv
ed
.
I
n
th
e
ex
p
e
r
im
en
t
s
tag
e,
we
will
ex
p
lain
th
e
s
tep
-
by
-
s
tep
p
r
o
ce
s
s
o
f
ad
d
r
ess
in
g
th
e
p
r
o
b
lem
s
tar
g
ete
d
in
th
is
r
esear
ch
.
Fin
ally
,
t
h
e
ev
alu
atio
n
m
etr
ics
u
s
ed
to
m
ea
s
u
r
e
t
h
e
p
e
r
f
o
r
m
a
n
ce
o
f
th
e
en
d
-
to
-
en
d
a
p
p
licatio
n
f
o
r
tr
an
s
latin
g
B
I
SIN
DO
S
L
T
in
to
tex
t
will
b
e
ex
p
lain
ed
.
A
d
etailed
d
is
cu
s
s
io
n
o
f
ea
ch
s
tag
e
will b
e
p
r
esen
ted
in
th
e
f
o
llo
win
g
s
ec
tio
n
s
.
3
.
1
.
Da
t
a
s
et
prepa
ra
t
io
n
3
.
1
.
1
.
Da
t
a
s
et
c
o
llect
io
n
We
ex
ten
d
ed
th
e
d
ataset
in
itially
em
p
lo
y
ed
b
y
[
1
1
]
,
wh
ich
co
m
p
r
is
ed
4
0
B
I
SIN
DO
s
en
ten
ce
s
s
p
an
n
in
g
1
5
2
wo
r
d
class
es,
in
co
r
p
o
r
atin
g
an
ad
d
itio
n
al
1
5
0
s
en
ten
ce
s
to
im
p
r
o
v
e
t
h
e
s
y
s
tem
's
lex
ical
co
v
er
ag
e
an
d
en
h
a
n
ce
its
g
en
er
aliza
tio
n
ca
p
ab
ilit
y
,
in
cr
ea
s
in
g
th
e
t
o
tal
n
u
m
b
er
o
f
wo
r
d
class
es
to
3
5
2
.
T
h
e
in
teg
r
atio
n
o
f
th
e
o
r
i
g
in
al
an
d
s
u
p
p
lem
en
tar
y
d
atasets
y
ield
ed
a
co
m
p
r
eh
en
s
iv
e
c
o
r
p
u
s
o
f
1
9
0
B
I
SIN
DO
s
en
ten
ce
s
ac
r
o
s
s
4
3
5
class
es,
with
r
ep
r
esen
tativ
e
ex
a
m
p
les
p
r
esen
ted
in
T
ab
le
2
.
Fo
r
th
e
ad
d
itio
n
al
d
ataset,
we
co
l
lecte
d
p
r
im
a
r
y
d
ata
in
c
o
llab
o
r
atio
n
with
th
e
L
an
g
u
ag
e
R
esear
ch
L
ab
o
r
at
o
r
y
team
a
t
th
e
Fak
u
ltas
I
lm
u
B
u
d
ay
a,
Un
iv
er
s
itas
I
n
d
o
n
esi
a
(
L
R
B
I
FIB
UI
)
.
T
wo
d
ea
f
s
ig
n
er
s
d
em
o
n
s
tr
ated
th
e
1
5
0
B
I
SIN
DO
s
en
ten
ce
s
th
r
ee
tim
es,
r
esu
ltin
g
i
n
9
0
0
v
id
eo
s
,
wh
ile
two
i
n
ter
p
r
eter
s
tr
an
s
lated
th
eir
g
estu
r
es
(
Fig
u
r
e
4
)
.
Fig
u
r
es
4
(
a
)
an
d
4
(
b
)
illu
s
tr
ate
th
e
d
ataset
co
llectio
n
p
r
o
ce
s
s
,
wh
ile
Fig
u
r
es 4
(
c
)
an
d
4
(
d
)
d
e
p
ict
th
e
d
e
af
s
ig
n
er
s
in
v
o
lv
ed
.
T
ab
le
2
.
E
x
am
p
le
o
f
1
5
0
B
I
SI
NDO
s
en
ten
ce
s
N
o
.
S
e
n
t
e
n
c
e
s
B
IS
IN
D
O Gl
o
ss
1
D
i
a
S
u
k
a
Me
n
g
e
n
a
k
a
n
B
a
j
u
Me
r
a
h
M
u
d
a
H
e
/
S
h
e
l
i
k
e
s
t
o
w
e
a
r
p
i
n
k
c
l
o
t
h
e
s.
D
i
a
-
b
a
j
u
-
m
e
ra
h
m
u
d
a
-
d
i
a
-
s
u
k
a
-
p
a
k
a
i
.
H
e
/
S
h
e
-
S
h
i
r
t
-
P
i
n
k
-
H
e
/
S
h
e
-
Li
k
e
s
-
W
e
a
r
.
2
H
a
ri
I
n
i
A
d
a
l
a
h
H
a
ri
M
i
n
g
g
u
To
d
a
y
i
s Su
n
d
a
y
.
H
a
ri
–
i
n
i
-
m
i
n
g
g
u
.
D
a
y
–
Th
i
s
–
S
u
n
d
a
y
.
3
S
a
y
a
B
u
t
u
h
Pe
n
g
g
a
r
i
s U
n
t
u
k
Me
n
g
u
k
u
r P
a
n
j
a
n
g
n
y
a
.
I
n
e
e
d
a
r
u
l
e
r
t
o
m
e
a
s
u
r
e
i
t
s l
e
n
g
t
h
.
Pe
n
g
g
a
r
i
s
–
i
n
i
–
s
a
y
a
-
b
u
t
u
h
–
u
n
t
u
k
-
a
p
a
-
p
a
n
j
a
n
g
-
u
k
u
r.
R
u
l
e
r
-
Th
i
s
-
I
-
N
e
e
d
-
F
o
r
-
W
h
a
t
-
Le
n
g
t
h
-
M
e
a
s
u
r
e
.
4
S
a
y
a
M
e
m
i
l
i
k
i
S
a
u
d
a
r
a
K
a
n
d
u
n
g
Y
a
n
g
B
a
i
k
.
I
h
a
v
e
a
k
i
n
d
s
i
b
l
i
n
g
.
S
a
y
a
-
p
u
n
y
a
-
s
a
u
d
a
r
a
k
a
n
d
u
n
g
-
d
i
a
-
b
a
i
k
.
I
-
H
a
v
e
-
S
i
b
l
i
n
g
-
H
e
/
S
h
e
-
K
i
n
d
.
5
J
a
n
g
a
n
L
u
p
a
Ma
k
a
n
J
a
m
b
u
S
a
a
t
S
a
r
a
p
a
n
.
D
o
n
'
t
f
o
r
g
e
t
t
o
e
a
t
g
u
a
v
a
a
t
b
r
e
a
k
f
a
st
.
Pa
g
i
-
j
a
m
b
u
-
m
a
k
a
n
-
l
u
p
a
-
j
a
n
g
a
n
.
M
o
r
n
i
n
g
-
G
u
a
v
a
-
Ea
t
-
F
o
r
g
e
t
-
D
o
n
'
t
.
(
a)
(b
)
(
c)
(
d
)
Fig
u
r
e
4
.
Data
s
et
co
llectio
n
p
r
o
ce
s
s
,
(
a)
d
ataset
co
llectio
n
d
o
cu
m
en
tatio
n
f
o
r
s
ig
n
e
r
o
n
e
,
(
b
)
d
ataset
co
llectio
n
d
o
c
u
m
en
tatio
n
f
o
r
s
ig
n
er
two
,
(
c)
s
ig
n
er
o
n
e,
a
n
d
(
d
)
s
ig
n
er
two
T
h
e
s
elec
tio
n
o
f
wo
r
d
s
f
o
r
th
e
1
5
0
-
s
en
ten
ce
B
I
SIN
DO
d
ataset
was
b
ased
o
n
th
e
B
I
SIN
DO
d
ictio
n
ar
y
,
with
th
e
p
r
im
ar
y
o
b
jectiv
e
o
f
th
is
s
tu
d
y
b
ein
g
to
in
co
r
p
o
r
ate
all
v
o
ca
b
u
la
r
y
en
tr
ies
f
r
o
m
th
e
d
ictio
n
ar
y
to
en
s
u
r
e
co
m
p
r
e
h
en
s
iv
e
lin
g
u
is
tic
r
ep
r
esen
tatio
n
.
A
co
m
p
ar
ativ
e
an
aly
s
is
o
f
th
e
d
atasets
u
s
ed
in
S
t
ar
t
D
at
as
e
t
P
r
epar
at
i
on
E
xp
er
i
m
ent
s
R
es
ul
t
E
val
ua
t
i
on
F
i
ni
sh
3
.
1
3
.
2
3
.
3
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
E
n
d
-
to
-
e
n
d
s
ystem
fo
r
tr
a
n
s
la
tin
g
b
a
h
a
s
a
is
ya
r
a
t I
n
d
o
n
esia
s
ig
n
la
n
g
u
a
g
e
g
estu
r
es
…
(
S
a
tr
ia
P
u
tr
a
)
723
th
is
s
tu
d
y
is
p
r
esen
ted
in
T
a
b
le
3
.
W
e
v
is
u
alize
d
th
e
r
ela
tio
n
s
h
ip
b
etwe
en
th
e
two
d
at
asets
u
s
in
g
a
Ven
n
d
iag
r
am
(
Fig
u
r
e
5
)
,
r
e
v
ea
lin
g
6
9
o
v
e
r
lap
p
in
g
wo
r
d
class
es.
T
ab
le
3
.
Deta
il d
ataset
A
sp
e
c
t
40
-
B
I
S
I
N
D
O
sen
t
e
n
c
e
s
1
5
0
-
B
I
S
I
N
D
O
sen
t
e
n
c
e
s
S
i
g
n
e
r
4
2
S
e
n
t
e
n
c
e
s
40
1
5
0
N
u
mb
e
r
o
f
v
i
d
e
o
s
4
2
0
9
0
0
W
o
r
l
d
c
l
a
ss
e
s
1
5
2
3
5
2
D
i
c
t
i
o
n
a
r
y
w
o
r
d
c
o
v
e
r
(
%)
1
3
.
0
2
3
0
.
1
6
To
t
a
l
f
r
a
mes
9
2
,
2
4
6
1
2
4
,
0
2
1
Fig
u
r
e
5
.
Ven
n
d
iag
r
am
o
f
th
e
d
ataset
3
.
1
.
2
.
Da
t
a
s
et
pre
-
pro
ce
s
s
in
g
T
h
e
B
I
SIN
DO
g
estu
r
es
wer
e
r
ec
o
r
d
ed
u
s
in
g
a
s
m
ar
tp
h
o
n
e
at
a
r
eso
lu
tio
n
o
f
1
9
2
0
×
1
0
8
0
an
d
a
f
r
am
e
r
ate
o
f
3
0
FP
S,
r
esu
ltin
g
in
1
2
4
,
0
2
1
f
r
am
es.
E
ac
h
f
r
am
e
w
as
r
esized
to
2
2
4
×
2
2
4
f
o
r
ef
f
icien
t
tr
ain
in
g
an
d
p
r
o
ce
s
s
in
g
.
T
o
s
im
u
late
r
ea
l
-
wo
r
ld
co
n
d
itio
n
s
,
b
ac
k
g
r
o
u
n
d
s
wer
e
m
o
d
if
ied
to
r
e
p
r
esen
t
n
in
e
en
v
ir
o
n
m
e
n
ts
:
b
an
k
,
o
f
f
ice,
class
r
o
o
m
,
s
h
o
p
p
in
g
m
all,
m
ar
k
et,
r
estau
r
an
t,
h
o
s
p
ital,
s
ch
o
o
l,
an
d
s
tr
ee
t.
Fig
u
r
e
6
o
u
tlin
es
th
e
d
ataset
p
r
ep
ar
atio
n
p
r
o
ce
s
s
.
3
.
1
.
3
.
Da
t
a
s
et
a
nn
o
t
a
t
io
n
T
h
e
d
ata
a
n
n
o
tatio
n
in
v
o
lv
ed
ap
p
ly
in
g
two
ty
p
es
o
f
an
n
o
tatio
n
s
.
First,
ea
ch
f
r
am
e
was
la
b
eled
with
th
e
co
r
r
esp
o
n
d
in
g
wo
r
d
b
ase
d
o
n
th
e
g
estu
r
e,
in
clu
d
i
n
g
t
r
an
s
itio
n
al
lab
els.
Fig
u
r
e
7
illu
s
tr
a
tes
th
e
wo
r
d
an
n
o
tatio
n
p
r
o
ce
s
s
,
wh
er
e
th
e
an
n
o
tated
s
en
ten
ce
is
"
Dia
La
r
i
K
a
r
en
a
Ta
ku
t
"
(
"He
/
Sh
e
R
u
n
s
B
ec
au
s
e
Scar
ed
")
.
Ho
wev
er
,
th
e
o
r
ig
in
al
s
en
ten
ce
o
r
d
er
is
"
K
a
r
en
a
Ta
ku
t,
Dia
La
r
i
"
(
"He
r
u
n
s
b
ec
au
s
e
h
e
is
s
ca
r
ed
")
.
T
h
e
lab
els
f
o
r
th
e
s
en
te
n
ce
ar
e
0
-
63
-
0
-
77
-
0
-
65
-
0
-
1
3
3
,
with
lab
el
0
in
d
icatin
g
tr
an
s
itio
n
g
estu
r
es
at
th
e
b
eg
in
n
in
g
,
b
etwe
en
wo
r
d
s
,
an
d
at
th
e
en
d
o
f
g
estu
r
es.
Fiv
e
an
n
o
tato
r
s
p
ar
ticip
ated
in
th
is
an
n
o
tatio
n
p
r
o
ce
s
s
to
m
in
im
ize
b
ias.
Fig
u
r
e
6
.
Data
s
et
p
r
ep
a
r
atio
n
s
tep
s
T
h
e
v
i
d
e
o
o
f
BIS
IN
D
O
s
e
n
t
a
n
c
e
T
h
e
s
e
ri
e
s
o
f
BIS
I
N
D
O
s
e
n
t
e
n
c
e
fra
m
e
s
w
i
t
h
a
g
re
e
n
b
a
c
k
g
r
o
u
n
d
T
h
e
s
e
ri
e
s
o
f
BIS
I
N
D
O
s
e
n
t
e
n
c
e
fra
m
e
s
w
i
t
h
c
o
m
p
l
e
x
b
a
c
k
g
ro
u
n
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
40
,
No
.
2
,
No
v
em
b
er
20
25
:
7
1
9
-
7
3
4
724
Fig
u
r
e
7
.
W
o
r
d
lab
elin
g
a
n
n
o
t
atio
n
T
h
e
s
u
b
s
eq
u
en
t
p
r
o
ce
s
s
in
v
o
lv
ed
ass
ig
n
in
g
b
o
u
n
d
in
g
b
o
x
lab
els
f
o
r
o
b
ject
d
etec
tio
n
t
ask
s
an
d
estab
lis
h
in
g
g
r
o
u
n
d
tr
u
t
h
f
o
r
th
r
ee
ca
teg
o
r
ies:
f
ac
e,
r
ig
h
t
h
a
n
d
,
an
d
lef
t
h
an
d
.
Nin
e
an
n
o
t
ato
r
s
p
er
f
o
r
m
ed
th
e
b
o
u
n
d
in
g
b
o
x
lab
elin
g
.
Giv
e
n
th
e
en
o
r
m
o
u
s
v
o
lu
m
e
o
f
f
r
a
m
es
r
eq
u
ir
in
g
b
o
u
n
d
in
g
b
o
x
e
s
,
th
e
m
ea
n
s
q
u
ar
e
d
er
r
o
r
(
MSE
)
[
3
2
]
an
d
s
tr
u
ctu
r
al
s
im
ilar
ity
in
d
ex
(
SS
I
M)
[
3
3
]
m
eth
o
d
s
wer
e
em
p
lo
y
e
d
with
a
9
4
%
th
r
esh
o
ld
.
T
h
is
s
y
s
t
em
atic
ap
p
r
o
ac
h
b
o
u
n
d
b
o
x
lab
els
f
o
r
3
8
,
7
3
6
o
u
t
o
f
1
2
4
,
0
2
1
f
r
am
es.
W
e
u
s
e
th
e
o
p
en
-
s
o
u
r
ce
ap
p
licatio
n
lab
elI
m
g
f
o
r
th
is
task
.
3
.
2
.
E
x
perim
ent
s
T
h
is
r
esear
ch
co
n
s
is
ts
o
f
th
r
ee
ex
p
er
im
en
ts
,
as
illu
s
tr
ated
in
Fig
u
r
e
8
.
E
x
p
er
im
en
t
1
aim
s
t
o
im
p
r
o
v
e
th
e
o
b
jec
t
d
etec
tio
n
m
et
h
o
d
u
s
ed
in
th
e
p
r
e
v
io
u
s
s
tu
d
y
b
y
[
1
1
]
with
t
h
e
latest
o
b
ject
d
etec
tio
n
m
et
h
o
d
.
E
x
p
er
im
en
t
2
aim
s
to
d
eter
m
in
e
th
e
m
o
s
t
e
f
f
ec
tiv
e
m
eth
o
d
f
o
r
r
ec
o
g
n
izin
g
tr
an
s
itio
n
g
estu
r
es.
E
x
p
er
im
en
t
3
aim
s
to
d
ev
elo
p
t
h
e
o
p
tim
al
m
o
d
el
f
o
r
en
d
-
to
-
en
d
B
I
SIN
DO
SLT
.
Fig
u
r
e
8
.
E
x
p
er
im
e
n
tal
f
lo
w
3
.
2
.
1
.
B
a
s
eline
A
s
tu
d
y
co
n
d
u
cted
b
y
[
1
1
]
ai
m
to
d
ev
elo
p
an
a
u
to
m
ated
B
is
in
d
o
r
ec
o
g
n
itio
n
s
y
s
tem
in
a
co
m
p
lex
b
ac
k
g
r
o
u
n
d
,
u
tili
zin
g
4
0
B
I
SIN
DO
with
s
tag
es
o
f
o
b
ject
d
etec
tio
n
,
s
k
in
co
lo
r
s
eg
m
en
tatio
n
,
f
ea
tu
r
e
ex
tr
ac
tio
n
,
an
d
class
if
icatio
n
(
Fig
u
r
e
9
)
,
wh
ich
will
s
er
v
e
as
a
b
aselin
e.
T
h
e
r
esu
lts
o
f
Fas
ter
R
-
C
NN
an
d
YOL
Ov
5
s
h
o
wed
m
AP
s
co
r
es
o
f
7
1
.
3
%
an
d
7
4
.
1
%,
r
esp
e
ctiv
ely
,
an
d
W
E
R
,
SAc
c,
an
d
Sacr
eBLE
U
wer
e
2
6
.
8
1
%,
3
5
.
0
0
%,
4
5
.
8
6
%,
a
n
d
1
6
.
4
2
%,
4
9
.
2
9
%,
6
7
.
7
7
%,
r
esp
ec
tiv
ely
,
wh
ic
h
in
d
icate
s
th
at
YOL
Ov
5
o
u
tp
er
f
o
r
m
s
Fas
ter
R
-
C
N
N,
it
s
h
o
u
ld
b
e
n
o
ted
th
at
in
th
i
s
s
tu
d
y
,
r
em
o
v
in
g
tr
a
n
s
itio
n
a
l
g
estu
r
es
is
d
o
n
e
m
an
u
ally
in
th
e
d
ata
p
r
ep
r
o
ce
s
s
in
g
p
r
o
ce
s
s
(
Fig
u
r
e
9
,
p
r
o
ce
s
s
1
)
.
B
ased
o
n
th
ese
r
esu
lts
,
th
er
e
a
r
e
s
till
g
ap
s
th
at
ca
n
b
e
im
p
r
o
v
ed
,
s
u
ch
as in
o
b
ject
d
etec
tio
n
,
wh
ich
will im
p
ac
t th
e
p
er
f
o
r
m
an
ce
o
f
th
e
tr
an
s
latio
n
r
esu
lt
s
an
d
th
e
im
p
lem
e
n
tatio
n
o
f
a
u
to
m
atic
tr
an
s
itio
n
g
estu
r
e
r
em
o
v
al,
s
o
t
h
at
th
e
s
y
s
tem
b
u
ilt
r
ef
lects
r
ea
l
-
wo
r
ld
co
n
d
itio
n
s
.
3
.
2
.
2
.
E
x
perim
ent
1
(
I
m
pro
v
ed
o
bje
ct
det
ec
t
io
n)
T
h
is
ex
p
er
im
en
t
aim
s
to
en
h
a
n
ce
s
y
s
tem
p
er
f
o
r
m
an
ce
in
o
b
ject
d
etec
tio
n
(
Fig
u
r
e
9
,
Pro
ce
s
s
2
)
b
y
u
tili
zin
g
th
e
YOL
Ov
7
x
an
d
YOL
Ov
8
x
m
o
d
els,
as
s
h
o
wn
in
Fig
u
r
e
1
0
.
B
o
th
m
o
d
els
o
f
f
er
d
is
tin
ct
ad
v
an
tag
es:
YOL
Ov
7
f
o
cu
s
e
s
o
n
o
p
tim
izin
g
th
e
E
-
E
L
A
N
ar
ch
itectu
r
e
an
d
s
ca
lin
g
f
o
r
ef
f
icien
cy
,
wh
ile
I
m
pr
oved
O
bj
e
c
t
D
et
ec
t
i
on
S
t
ar
t
D
a
t
a
s
e
t
40
S
e
nt
e
nc
e
s
Im
rpove
d
T
ra
ns
i
t
i
on
G
e
s
t
ure
s
Re
c
ogn
i
t
i
on
D
a
t
a
s
e
t
150
S
e
nt
e
nc
e
s
E
va
l
ua
t
i
on
(
m
A
P
, A
cc
,
WE
R
, S
A
cc
,
S
ac
r
e
B
L
E
U
)
Fi
ni
s
h
E
val
ua
t
i
on
(
m
A
P, A
cc
,
WE
R
, S
A
cc
,
Sac
r
e
B
L
E
U
)
T
he
B
e
s
t
O
be
j
c
t
D
e
t
e
c
t
i
on
M
e
t
hod
T
he B
es
t
Model
E
x
p
er
eim
ent
1
E
x
p
er
eim
ent
2
T
he
Be
s
t
T
ra
ns
i
t
i
on
G
e
s
t
ur
e
s
Re
c
ogni
t
i
on
M
e
t
hod
E
nd
-
To
-
E
nd
BIS
IN
D
O
S
L
T
E
val
ua
t
i
on
(
m
A
P
, A
cc
,
WE
R
, S
A
cc
,
S
ac
r
e
B
L
E
U
)
E
x
p
er
eim
en
t 3
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
E
n
d
-
to
-
e
n
d
s
ystem
fo
r
tr
a
n
s
la
tin
g
b
a
h
a
s
a
is
ya
r
a
t I
n
d
o
n
esia
s
ig
n
la
n
g
u
a
g
e
g
estu
r
es
…
(
S
a
tr
ia
P
u
tr
a
)
725
YOL
Ov
8
em
p
h
asizes
m
u
lti
-
lev
el
f
ea
tu
r
e
ex
t
r
ac
tio
n
th
r
o
u
g
h
a
co
m
b
in
atio
n
o
f
C
SP
Dar
k
n
e
t,
SP
PF
,
an
d
PAN.
Alth
o
u
g
h
YOL
Ov
8
is
a
n
ewe
r
v
er
s
io
n
o
f
YOL
Ov
7
,
th
e
o
r
ig
in
al
YOL
Ov
8
p
ap
er
d
o
es
n
o
t
ex
p
licitly
co
m
p
ar
e
th
e
p
er
f
o
r
m
a
n
ce
o
f
th
e
two
Y
OL
O
v
er
s
io
n
s
.
T
h
e
ex
p
er
im
en
t
u
tili
ze
s
4
0
B
I
SIN
DO
s
en
ten
ce
s
,
with
th
e
o
b
ject
d
etec
tio
n
m
eth
o
d
as
th
e
in
d
ep
en
d
en
t
v
ar
iab
le.
I
ts
s
u
cc
ess
is
m
ea
s
u
r
ed
b
y
th
e
m
AP,
W
E
R
,
SAcc,
an
d
Sacr
eBLE
U
s
co
r
es.
Fig
u
r
e
9
.
B
aselin
e
[
1
1
]
Fig
u
r
e
1
0
.
I
m
p
r
o
v
e
d
d
etec
tio
n
o
b
jects
3
.
2
.
3
.
E
x
perim
ent
2
(
I
m
pro
v
ed
t
ra
ns
it
io
n g
esture
s
re
co
g
nitio
n
)
T
h
is
ex
p
er
im
en
t
ad
d
r
ess
es
th
e
p
r
o
b
lem
o
f
r
em
o
v
in
g
tr
a
n
s
itio
n
al
g
estu
r
es.
I
n
th
e
s
tu
d
y
[
1
1
]
,
th
e
r
em
o
v
al
o
f
tr
an
s
itio
n
al
g
estu
r
es
was
ca
r
r
ied
o
u
t
at
th
e
d
ata
p
r
e
p
r
o
ce
s
s
in
g
s
t
ag
e
(
Fig
u
r
e
9
,
p
r
o
ce
s
s
1
)
,
s
o
in
th
is
s
tu
d
y
,
a
s
tag
e
will
b
e
ad
d
ed
b
etwe
en
s
tag
es
4
an
d
5
(
Fig
u
r
e
9
,
p
r
o
ce
s
s
es
4
an
d
5
)
,
tr
an
s
itio
n
al
g
estu
r
es
r
ec
o
g
n
itio
n
s
o
th
at
th
e
s
y
s
tem
ca
n
r
em
o
v
e
th
e
m
au
to
m
ati
ca
lly
.
T
h
e
u
ltima
te
g
o
al
is
to
b
u
ild
an
en
d
-
to
-
en
d
B
I
SIN
DO
SLT
.
I
n
ad
d
itio
n
,
we
also
ad
d
ed
th
e
s
an
d
wich
m
ajo
r
ity
v
o
tin
g
an
d
s
h
o
r
t
wo
r
d
-
f
r
a
m
e
s
eq
u
en
c
e
r
elab
elin
g
s
tag
es
af
ter
th
e
tr
an
s
itio
n
al
g
estu
r
es
s
tag
e
an
d
b
ef
o
r
e
t
h
e
au
to
m
atic
r
em
o
v
al
o
f
tr
an
s
itio
n
al
g
estu
r
es,
s
o
th
at
th
e
r
esu
ltin
g
d
ata
ar
e
clea
n
,
as seen
in
Fig
u
r
e
1
1
.
I
n
s
tead
o
f
u
s
in
g
o
r
i
g
in
a
l la
b
els to
r
ec
o
g
n
ize
tr
an
s
itio
n
g
estu
r
es,
we
em
p
l
o
y
two
ap
p
r
o
ac
h
es:
f
ir
s
t,
M
o
b
ileNet
-
p
r
ed
icted
la
b
els,
an
d
s
ec
o
n
d
,
T
C
R
F
-
p
r
ed
icted
la
b
els.
T
h
e
d
ataset
u
s
ed
co
n
s
is
ts
o
f
4
0
B
I
SIN
DO
s
en
ten
ce
s
,
with
th
e
tr
an
s
itio
n
g
estu
r
e
r
ec
o
g
n
itio
n
m
eth
o
d
as th
e
in
d
ep
en
d
en
t v
a
r
iab
le.
Fig
u
r
e
1
1
.
I
m
p
r
o
v
e
d
tr
an
s
itio
n
al
g
estu
r
e
r
em
o
v
al
3
.
2
.
4
.
E
x
perim
ent
3
(
P
r
o
po
s
ed
end
-
to
-
end
B
I
SI
NDO
S
L
T
)
T
h
is
s
tu
d
y
p
r
o
p
o
s
es
an
en
d
-
to
-
en
d
B
I
SIN
DO
SLT
ap
p
licatio
n
,
d
ev
elo
p
e
d
th
r
o
u
g
h
E
x
p
e
r
im
en
ts
1
an
d
2
.
T
h
e
s
y
s
tem
p
r
o
ce
s
s
es
B
I
SIN
DO
v
id
eo
f
r
am
es
as
in
p
u
t
an
d
g
e
n
er
ates
co
r
r
es
p
o
n
d
in
g
B
I
SIN
DO
s
en
ten
ce
s
as
o
u
tp
u
t,
o
p
er
ati
n
g
f
u
lly
au
to
m
atica
lly
with
o
u
t
h
u
m
a
n
in
ter
v
e
n
tio
n
i
n
r
e
al
-
wo
r
ld
co
n
d
itio
n
s
(
Fig
u
r
e
1
2
)
.
T
h
e
s
tag
es
in
v
o
l
v
ed
ar
e
(
1
)
o
b
ject
d
etec
tio
n
,
(
2
)
s
k
in
co
lo
r
s
eg
m
e
n
tatio
n
,
(
3
)
f
ea
tu
r
e
ex
tr
ac
tio
n
,
(
4
)
tr
a
n
s
itio
n
g
estu
r
es
r
ec
o
g
n
i
tio
n
,
(
5
)
s
an
d
wich
m
aj
o
r
ity
v
o
tin
g
a
n
d
s
h
o
r
t
wo
r
d
-
f
r
am
e
s
eq
u
en
ce
r
elab
elin
g
,
(
6
)
tr
an
s
itio
n
r
em
o
v
in
g
,
(
7
)
f
r
am
e
eq
u
aliza
tio
n
,
an
d
(
8
)
class
if
icatio
n
.
E
v
alu
atio
n
em
p
lo
y
s
th
r
ee
d
at
aset
co
n
f
ig
u
r
atio
n
s
:
4
0
s
en
ten
ce
s
(
1
5
2
lab
els),
1
5
0
s
en
ten
ce
s
(
3
5
2
lab
els),
an
d
th
eir
c
o
m
b
in
ed
1
9
0
-
s
en
ten
ce
d
ataset
(
4
3
5
lab
els),
with
s
en
t
en
ce
an
d
lab
el
q
u
a
n
tity
as
th
e
in
d
ep
en
d
en
t
v
ar
iab
les
f
o
r
ass
ess
in
g
th
e
im
p
ac
t
o
f
s
y
s
tem
p
er
f
o
r
m
a
n
ce
.
D
at
a P
re
parat
i
on
T
r
ans
i
t
i
on R
e
m
ovi
ng
F
r
am
e E
qual
i
za
t
i
on
O
bj
ec
t
D
et
ec
t
i
on
S
ki
n C
ol
or
S
egm
e
nt
at
i
on
F
ea
t
ur
e E
xt
r
ac
t
i
on
C
l
as
s
i
f
i
ca
t
i
on
E
val
ua
t
i
on
1
2
3
4
5
6
B
I
S
IN
D
O
F
r
am
es
F
as
t
er
R
-
C
N
N
Y
O
L
O
v5
Y
O
L
O
v7
Y
O
L
O
v8
B
I
S
IN
D
O
F
r
am
e
s
wit
h
B
B
X
D
e
tec
ti
o
n
O
b
j
ec
t (
m
A
P)
L
S
T
M
C
l
as
s
i
f
i
ca
t
i
on
Mobi
l
eN
et
V
2 F
eat
ur
e
E
xt
r
ac
t
i
on
Mobi
l
eN
et
pr
edi
ct
ed
l
a
bel
s
S
a
nd
w
i
c
h
M
a
j
or
i
t
y
V
ot
i
ng
,
S
hort
w
ord
-
fra
m
e
s
e
que
nc
e
re
l
a
be
l
i
ng
,
A
ut
om
a
t
i
c
T
r
ans
i
t
i
on
r
em
ovi
ng
T
C
R
F
pr
edi
ct
e
d l
abel
s
4
5
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
40
,
No
.
2
,
No
v
em
b
er
20
25
:
7
1
9
-
7
3
4
726
Fig
u
r
e
1
2
.
E
n
d
-
to
-
en
d
o
f
B
I
SIN
DO
SLT
Ob
ject
d
etec
tio
n
is
th
e
in
itial
s
tag
e
o
f
id
en
tify
in
g
th
e
s
ig
n
er
'
s
f
ac
e,
r
ig
h
t
h
an
d
,
an
d
lef
t
h
an
d
in
in
p
u
t
im
ag
es
with
co
m
p
lex
b
ac
k
g
r
o
u
n
d
s
,
p
r
o
d
u
cin
g
is
o
lated
r
eg
io
n
s
.
T
h
e
aim
is
to
e
x
tr
ac
t
k
e
y
f
ea
tu
r
es
f
o
r
f
u
r
th
e
r
pr
o
ce
s
s
in
g
.
T
h
e
m
o
d
els
wer
e
tr
ain
ed
with
a
b
atch
s
ize
o
f
1
6
f
o
r
1
5
0
e
p
o
ch
s
.
Per
f
o
r
m
an
c
e
is
m
ea
s
u
r
ed
u
s
in
g
m
AP
o
n
th
e
test
s
et,
with
th
e
b
aselin
e
s
et
as
th
e
r
esu
lts
f
r
o
m
[
1
1
]
'
s
s
tu
d
y
,
wh
ich
s
er
v
es
a
s
th
e
g
r
o
u
n
d
tr
u
t
h
to
en
s
u
r
e
r
eliab
le
d
etec
tio
n
q
u
ali
ty
.
T
h
e
s
ec
o
n
d
s
tag
e
is
s
k
in
co
lo
r
s
eg
m
en
tatio
n
.
I
m
ag
es
f
r
o
m
t
h
e
p
r
ev
io
u
s
s
tag
e
s
till
co
n
tain
b
o
u
n
d
in
g
b
o
x
es,
r
eq
u
ir
i
n
g
an
ad
d
itio
n
a
l
s
tep
to
is
o
late
an
d
ex
t
r
ac
t
o
n
ly
t
h
e
f
ac
e
a
n
d
h
an
d
s
.
T
h
i
s
s
tag
e
ap
p
lies
th
e
m
eth
o
d
p
r
o
p
o
s
ed
b
y
[
3
4
]
,
b
y
c
o
m
b
in
in
g
n
o
r
m
ali
ze
d
R
GB
,
HSV,
an
d
YC
b
C
r
co
lo
r
s
p
ac
es.
T
h
e
th
ir
d
s
tag
e
p
er
f
o
r
m
s
f
ea
t
u
r
e
ex
tr
ac
tio
n
u
s
in
g
Mo
b
ileN
etV2
[
3
5
]
to
tr
a
n
s
f
o
r
m
s
eg
m
e
n
ted
h
an
d
an
d
f
ac
e
r
eg
io
n
s
in
t
o
m
ea
n
in
g
f
u
l
g
estu
r
e
r
ep
r
esen
tatio
n
s
.
T
h
is
s
tag
e
p
r
o
ce
s
s
es
s
eg
m
en
ted
f
r
am
es
an
d
th
ei
r
B
I
SIN
DO
lab
el
s
th
r
o
u
g
h
a
m
o
d
el
tr
ain
ed
with
a
b
atch
s
ize
o
f
1
6
f
o
r
1
0
0
ep
o
c
h
s
,
g
en
er
atin
g
1
2
8
0
-
d
im
en
s
io
n
al
f
ea
tu
r
e
v
ec
to
r
s
(
s
to
r
ed
as
Nu
m
Py
a
r
r
ay
s
)
f
o
r
ea
ch
f
r
am
e.
T
h
e
p
u
r
p
o
s
e
is
to
cr
ea
te
d
is
cr
im
in
ativ
e
f
ea
tu
r
es
th
at
ca
p
tu
r
e
ess
en
tial
g
estu
r
e
ch
ar
ac
ter
is
tics
wh
ile
r
ed
u
cin
g
co
m
p
u
tatio
n
al
co
m
p
lex
ity
.
Pre
d
ictin
g
ac
cu
r
ac
y
m
ea
s
u
r
es
p
er
f
o
r
m
a
n
ce
,
wh
ich
v
alid
ates
h
o
w
well
th
e
ex
tr
ac
ted
f
ea
tu
r
es
r
ep
r
esen
t
B
I
SIN
DO
g
estu
r
es.
T
h
e
f
o
u
r
t
h
s
tag
e
f
o
cu
s
es
o
n
tr
an
s
itio
n
g
estu
r
e
r
ec
o
g
n
itio
n
u
s
in
g
Mo
b
ileNetV2
an
d
T
C
R
F
p
r
ed
ictio
n
lab
els.
T
h
ese
two
m
eth
o
d
s
a
im
to
d
eter
m
in
e
th
e
b
est
r
o
b
u
s
t
m
eth
o
d
f
o
r
r
ec
o
g
n
izin
g
tr
an
s
itio
n
g
estu
r
es.
Mo
b
ilen
et
p
lay
s
a
d
u
al
r
o
le;
in
ad
d
itio
n
to
b
ein
g
an
e
x
t
r
ac
tio
n
f
ea
tu
r
e
,
it
is
also
task
ed
with
m
ak
i
n
g
p
r
ed
ictio
n
s
b
ased
o
n
th
e
r
esu
lt
in
g
lab
els.
T
h
e
o
u
tp
u
t
o
f
th
is
s
tag
e
is
a
s
eq
u
e
n
ce
o
f
f
ea
t
u
r
es f
o
r
ea
c
h
B
I
SIN
DO
g
estu
r
e,
as sh
o
wn
in
Fig
u
r
e
1
3
.
Fig
u
r
e
1
3
.
I
llu
s
tr
atio
n
o
f
tr
an
s
itio
n
g
estu
r
e
r
ec
o
g
n
itio
n
T
C
R
F
is
tr
ain
ed
with
a
b
atch
s
ize
o
f
1
6
an
d
1
0
0
ep
o
c
h
s
,
an
d
th
e
ac
cu
r
ac
y
is
m
ea
s
u
r
ed
f
o
r
ea
c
h
th
r
esh
o
ld
v
alu
e.
As
an
illu
s
tr
atio
n
,
Fig
u
r
e
1
4
p
r
esen
ts
th
e
p
r
o
b
ab
ilit
y
g
r
ap
h
f
o
r
ea
c
h
class
(
tr
an
s
itio
n
o
r
wo
r
d
g
estu
r
es)
b
ased
o
n
th
e
B
I
SIN
DO
s
en
ten
ce
"
Ma
u
P
erg
i
B
a
n
d
u
n
g
B
eso
k
"
(
"Wan
t
-
Go
-
B
a
n
d
u
n
g
-
T
o
m
o
r
r
o
w")
.
T
h
e
b
est
-
p
e
r
f
o
r
m
in
g
M
o
b
ileN
etV2
an
d
T
C
R
F
m
o
d
els
g
en
er
ated
th
is
g
r
a
p
h
.
W
o
r
d
g
estu
r
e
s
ap
p
ea
r
i
n
f
r
am
es
88
–
9
2
,
1
0
6
–
1
1
4
,
1
2
6
–
1
3
7
,
a
n
d
1
5
1
–
1
5
6
,
wh
ile
all
o
t
h
er
f
r
a
m
es r
ep
r
esen
t tr
an
s
itio
n
al
g
est
u
r
es.
Fig
u
r
e
1
4
.
Pr
o
b
ab
ilit
y
g
r
ap
h
f
o
r
“
Ma
u
-
P
erg
i
-
B
a
n
d
u
n
g
-
B
eso
k
”
(
"Wan
t
-
Go
-
B
an
d
u
n
g
-
T
o
m
o
r
r
o
w")
F
r
am
e E
qual
i
za
t
i
on
O
bj
ec
t
D
et
ec
t
i
on
(
m
A
P)
Ski
n C
ol
or
Segm
ent
at
i
on
Mul
t
i
C
ol
or
S
pace
F
ea
t
ur
e E
xt
r
ac
t
i
on
M
obi
l
ene
t
V
2
(
A
cc
ur
ac
y)
Fra
me
wi
th
bounding box
Segment
ed fr
ame
T
r
ans
i
t
i
on G
e
s
t
ur
es
R
ec
ogni
t
i
on
T
r
ans
i
t
i
on r
em
ovi
ng
Fea
ture
, 12
80
C
l
as
s
i
f
i
ca
t
i
on
L
S
T
M
(
A
cc
ur
ac
y)
Fea
ture
, 12
80
1
B
I
S
IN
D
O
F
r
a
m
e
s
4
0
,
1
5
0
d
a
n
1
9
0
-
sen
t
en
c
es
2
3
6
7
Fea
ture
, 12
80
Featur
e, 128
0
S
a
ndw
i
c
h M
a
j
ori
t
y V
ot
i
ng,
S
hort
w
ord
-
fra
m
e
s
e
que
nc
e
re
l
a
be
l
i
ng,
Fea
ture
, 12
80
5
8
B
IS
I
N
D
O
Sen
ten
c
es
4
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
E
n
d
-
to
-
e
n
d
s
ystem
fo
r
tr
a
n
s
la
tin
g
b
a
h
a
s
a
is
ya
r
a
t I
n
d
o
n
esia
s
ig
n
la
n
g
u
a
g
e
g
estu
r
es
…
(
S
a
tr
ia
P
u
tr
a
)
727
T
h
e
f
if
th
s
tag
e
c
o
m
p
r
is
es
t
wo
s
eq
u
en
tial
s
tep
s
:
San
d
wi
ch
m
in
o
r
ity
v
o
tin
g
.
T
h
is
s
te
p
r
ep
lace
s
tr
an
s
itio
n
lab
els
s
an
d
wich
ed
b
etwe
en
wo
r
d
lab
els
with
th
e
co
r
r
esp
o
n
d
in
g
wo
r
d
lab
els.
Sh
o
r
t
wo
r
d
-
f
r
a
m
e
s
eq
u
en
ce
r
elab
elin
g
:
W
o
r
d
lab
els
f
lan
k
ed
b
y
tr
an
s
itio
n
lab
els
ar
e
r
ec
lass
if
ied
as
tr
an
s
itio
n
lab
els.
I
n
alig
n
m
en
t
with
[
1
0
]
,
th
is
s
tu
d
y
ap
p
lies
a
co
n
s
tr
ain
t
wh
er
eb
y
a
m
ax
im
u
m
o
f
two
co
n
s
ec
u
tiv
e
lab
els
m
ay
b
e
m
o
d
if
ied
in
ea
ch
s
tep
,
as illu
s
tr
a
ted
in
Fig
u
r
e
1
5
.
Fig
u
r
e
1
5
.
Po
s
t
-
tr
an
s
itio
n
g
estu
r
e
r
ec
o
g
n
itio
n
T
h
e
s
ix
th
s
tag
e
in
v
o
lv
es
r
em
o
v
in
g
tr
an
s
itio
n
al
g
estu
r
es,
wh
er
e
tr
an
s
itio
n
al
g
estu
r
es
ar
e
id
en
tifie
d
an
d
elim
in
ated
u
s
in
g
Mo
b
ile
NetV2
an
d
T
C
R
F
-
p
r
ed
icted
lab
els.
Fra
m
es
ass
o
ciate
d
with
a
lab
el
v
alu
e
o
f
ze
r
o
an
d
th
eir
co
r
r
esp
o
n
d
i
n
g
lab
e
ls
ar
e
d
is
ca
r
d
ed
,
as
d
ep
icted
in
Fig
u
r
e
1
6
.
T
h
is
p
r
o
ce
s
s
en
s
u
r
es
th
at
o
n
ly
m
ea
n
in
g
f
u
l g
estu
r
e
s
eg
m
e
n
ts
ar
e
p
r
eser
v
e
d
f
o
r
s
u
b
s
eq
u
en
t
wo
r
d
p
r
e
d
ictio
n
.
Fig
u
r
e
1
6
.
T
r
an
s
itio
n
r
e
m
o
v
in
g
Th
e
s
ev
en
th
s
tag
e
in
v
o
lv
es
f
r
am
e
eq
u
aliza
tio
n
,
as
s
h
o
w
n
in
Fig
u
r
e
1
7
,
e
n
s
u
r
in
g
u
n
if
o
r
m
in
p
u
t
s
eq
u
en
ce
len
g
th
s
r
eq
u
ir
e
d
b
y
th
e
L
STM
-
b
ased
s
eq
2
s
eq
ar
ch
itectu
r
e.
T
h
is
s
tan
d
ar
d
izatio
n
p
r
o
ce
s
s
b
eg
in
s
b
y
ca
lcu
latin
g
th
e
av
e
r
ag
e
f
r
am
e
len
g
th
af
ter
r
em
o
v
in
g
t
h
e
t
r
an
s
itio
n
.
Fo
r
s
eq
u
e
n
ce
s
ex
ce
ed
in
g
th
is
av
er
a
g
e
len
g
th
,
th
e
s
y
s
tem
ca
lcu
lates
th
e
MSE
b
etwe
en
a
d
jace
n
t
f
r
am
e
p
air
s
,
r
a
n
k
s
th
em
ac
co
r
d
in
g
ly
,
a
n
d
r
e
m
o
v
e
s
th
e
last
n
f
r
am
es
(
wh
er
e
n
eq
u
als
th
e
d
if
f
er
en
ce
b
etwe
en
cu
r
r
en
t
a
n
d
av
er
a
g
e
len
g
th
s
)
as
s
h
o
wn
in
Fig
u
r
e
1
7
(
a
)
.
C
o
n
v
er
s
ely
,
s
h
o
r
ter
s
eq
u
en
ce
s
a
r
e
p
a
d
d
ed
b
y
d
u
p
licatin
g
t
h
eir
f
i
n
al
f
r
am
e
u
n
til
th
ey
r
ea
c
h
th
e
tar
g
et
len
g
th
(
Fig
u
r
e
1
7
b
)
.
T
h
e
tar
g
et
len
g
th
o
f
1
5
f
r
am
es
p
er
wo
r
d
was
d
eter
m
in
e
d
b
y
d
iv
id
in
g
t
h
e
to
tal
f
r
am
e
c
o
u
n
t
ac
r
o
s
s
all
wo
r
d
s
b
y
th
e
n
u
m
b
er
o
f
wo
r
d
c
lass
es.
T
h
is
eq
u
aliza
tio
n
p
r
o
ce
d
u
r
e
is
ap
p
lied
u
n
if
o
r
m
ly
to
b
o
th
f
ea
t
u
r
e
f
r
am
es a
n
d
th
eir
co
r
r
esp
o
n
d
in
g
la
b
els.
(
a)
(
b
)
Fig
u
r
e
1
7
.
Fra
m
e
eq
u
aliza
tio
n
(
a)
Fra
m
e
m
o
r
e
th
a
n
1
5
,
(
b
)
Fra
m
e
less
th
an
1
5
0
0
0
0
0
0
0
0
160
160
0
160
160
0
0
0
20
20
0
0
0
0
412
412
412
412
412
0
0
0
0
0
0
0
0
227
227
0
0
227
227
227
0
0
0
0
3
0
0
0
0
0
0
0
0
0
160
160
160
160
160
0
0
0
20
20
0
0
0
0
412
412
412
412
412
0
0
0
0
0
0
0
0
227
227
227
227
227
227
227
0
0
0
0
3
0
0
0
0
0
0
0
0
0
160
160
160
160
160
0
0
0
0
0
0
0
0
0
412
412
412
412
412
0
0
0
0
0
0
0
0
227
227
227
227
227
227
227
0
0
0
0
0
0
W
o
r
d
L
ab
els
S
an
d
w
i
c
h
M
aj
o
r
i
t
y Vo
t
i
n
g
S
h
o
r
t
W
o
r
d
-
F
r
am
e S
equ
enc
e Re
l
ab
eli
n
g
L
a
b
el
s
0
0
0
0
0
0
0
0
160
160
160
160
160
0
0
0
0
0
0
0
0
0
412
412
412
412
412
0
0
0
0
0
0
0
0
227
227
227
227
227
227
227
0
0
0
0
0
0
F
r
a
m
es
L
a
b
el
s
160
160
160
160
160
412
412
412
412
412
227
227
227
227
227
227
227
F
r
a
m
es
T
r
a
n
s
i
t
i
o
n
Rem
o
vi
n
g
. . . . . .
1
2
3
12
13
. . . . . .
1
2
3
12
13
14
15
. . . . . .
1
2
3
12
13
14
15
16
17
. . . . . .
1
2
3
12
13
14
15
F
r
a
m
e
s
l
e
s
s
t
ha
n 15
E
qua
li
z
e
d f
r
a
m
e
s
t
o
15
F
r
a
m
e
s
m
o
r
e
t
ha
n 15
E
qua
li
z
e
d f
r
a
m
e
s
t
o
15
C
o
p
i
e
d
F
r
a
m
e
MS
E
C
a
l
c
u
l
a
t
i
o
n
. . . . . .
1
2
3
12
13
. . . . . .
1
2
3
12
13
14
15
. . . . . .
F
r
a
m
e
s
l
e
s
s
t
ha
n 1
5
E
quali
z
e
d f
r
a
m
e
s
t
o
15
F
r
a
m
e
s
m
o
r
e
t
ha
n 1
5
C
o
p
i
e
d
F
r
a
m
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
40
,
No
.
2
,
No
v
em
b
er
20
25
:
7
1
9
-
7
3
4
728
T
h
e
f
in
al
s
tag
e
o
f
th
e
p
r
o
p
o
s
e
d
m
eth
o
d
o
l
o
g
y
is
class
if
icatio
n
,
u
tili
zin
g
a
1
-
lay
er
B
id
ir
ec
tio
n
al
lo
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
L
STM
)
n
etwo
r
k
[
3
6
]
.
T
h
is
m
o
d
el
was
ch
o
s
en
b
ased
o
n
p
r
io
r
r
esear
c
h
[
1
1
]
.
C
o
n
f
ig
u
r
ed
with
2
5
6
h
id
d
e
n
u
n
its
,
a
b
atch
s
ize
o
f
1
6
,
an
d
tr
ain
ed
f
o
r
3
0
0
ep
o
c
h
s
,
th
e
m
o
d
el
en
s
u
r
e
s
ef
f
ec
tiv
e
lear
n
in
g
an
d
p
r
ec
is
e
class
if
icatio
n
o
f
in
p
u
t d
ata,
o
p
tim
izin
g
p
er
f
o
r
m
a
n
ce
f
o
r
B
I
SIN
DO
g
estu
r
e
r
ec
o
g
n
itio
n
.
All
p
r
o
ce
s
s
es
wer
e
co
n
d
u
cte
d
u
s
in
g
th
e
s
am
e
m
ac
h
in
e
as
i
n
th
e
s
tu
d
y
b
y
[
1
1
]
,
a
n
NVI
D
I
A
DGX
-
1
s
er
v
er
with
L
in
u
x
OS (
#
4
8
-
U
b
u
n
tu
SMP)
,
an
8
0
-
c
o
r
e
x
8
6
_
6
4
C
PU,
5
1
2
GB
D
DR
4
m
em
o
r
y
,
an
d
a
n
NVI
DI
A
T
esla V
1
0
0
-
SXM2
-
3
2
GB
GPU.
T
h
is
allo
ws th
e
r
esu
lts
to
b
e
co
m
p
ar
e
d
with
p
r
e
v
io
u
s
r
ese
ar
ch
.
3.
3
.
E
v
a
lua
t
i
o
n
Fiv
e
m
etr
ics
wer
e
u
s
ed
:
m
AP,
Acc
u
r
ac
y
,
W
E
R
,
Sacr
eB
L
E
U,
an
d
SAcc.
m
AP
ass
es
s
es
o
b
ject
d
etec
tio
n
b
y
c
o
m
p
ar
i
n
g
p
r
ed
ictio
n
s
to
g
r
o
u
n
d
tr
u
th
u
s
in
g
th
e
in
ter
s
ec
tio
n
o
v
e
r
u
n
io
n
(
I
o
U)
m
etr
ic.
I
t
is
ca
lcu
lated
u
s
in
g
E
q
u
atio
n
2
.
=
1
∑
=
1
(
2
)
W
h
er
e
N
is
th
e
n
u
m
b
er
o
f
clas
s
es,
an
d
APi is
th
e
AP
f
o
r
th
e
i
-
th
class
.
Acc
u
r
ac
y
m
ea
s
u
r
es
th
e
p
er
ce
n
tag
e
o
f
co
r
r
ec
tly
p
r
ed
icte
d
d
ata,
as
d
e
f
in
ed
in
(
3
)
.
I
t
is
cr
u
cial
f
o
r
ev
alu
atin
g
p
e
r
f
o
r
m
an
ce
in
f
ea
t
u
r
e
ex
tr
ac
tio
n
,
tr
an
s
itio
n
g
estu
r
e
r
ec
o
g
n
itio
n
,
a
n
d
clas
s
if
icatio
n
s
tag
es.
=
+
+
+
+
(
3
)
W
h
er
e
tr
u
e
p
o
s
itiv
e
(
TP
)
is
th
e
n
u
m
b
er
o
f
p
o
s
itiv
e
ca
s
es
co
r
r
ec
tly
p
r
ed
icted
.
T
r
u
e
n
eg
ati
v
e
(
TN
)
is
th
e
n
u
m
b
er
o
f
n
eg
ativ
e
ca
s
es
co
r
r
ec
tly
p
r
ed
icted
.
Fals
e
p
o
s
itiv
e
(
FP
)
is
th
e
n
u
m
b
e
r
o
f
n
eg
ativ
e
ca
s
es
in
co
r
r
ec
tly
p
r
ed
icted
as
p
o
s
itiv
e.
F
alse
n
eg
ativ
e
(
FN
)
is
th
e
n
u
m
b
er
o
f
p
o
s
itiv
e
ca
s
es
in
co
r
r
ec
tly
p
r
e
d
icted
as
n
eg
ativ
e
b
y
th
e
m
o
d
el.
W
E
R
m
ea
s
u
r
es
t
h
e
r
atio
o
f
th
e
n
u
m
b
e
r
o
f
in
c
o
r
r
ec
t
wo
r
d
s
g
en
er
ated
b
y
th
e
s
y
s
tem
to
th
e
to
tal
n
u
m
b
er
o
f
wo
r
d
s
in
th
e
g
iv
en
d
o
cu
m
e
n
t o
r
d
ataset.
In
(
4
)
ca
l
cu
lates WER
.
=
+
+
=
+
+
(
4
)
W
h
er
e
S
is
th
e
n
u
m
b
er
o
f
s
u
b
s
titu
tio
n
er
r
o
r
s
,
D
is
th
e
n
u
m
b
er
o
f
d
eleti
o
n
er
r
o
r
s
,
I
is
th
e
n
u
m
b
er
o
f
in
s
er
tio
n
er
r
o
r
s
,
an
d
N
is
th
e
t
o
tal
n
u
m
b
e
r
o
f
wo
r
d
s
in
th
e
an
aly
ze
d
s
en
ten
ce
o
r
d
o
c
u
m
en
t.
Sacr
eBLE
U
ass
e
s
s
e
s
th
e
q
u
al
ity
o
f
m
ac
h
in
e
tr
a
n
s
latio
n
s
b
e
twee
n
lan
g
u
ag
es.
I
ts
g
o
al
is
t
o
ev
alu
ate
h
o
w
ef
f
ec
tiv
ely
m
ac
h
in
e
tr
an
s
latio
n
s
co
m
p
ar
e
to
h
u
m
a
n
t
r
an
s
latio
n
s
.
Sacr
eBLE
U
u
s
es
a
f
o
r
m
u
la
m
o
d
if
ie
d
b
ased
o
n
n
-
g
r
a
m
p
r
ec
is
io
n
t
o
ev
alu
ate
th
is
.
T
h
e
eq
u
atio
n
ca
n
b
e
ex
p
lain
ed
as sh
o
wn
i
n
(
5
)
.
=
×
(
5
)
W
h
er
e
B
L
E
U
is
th
e
s
co
r
e
ca
lcu
lated
u
s
in
g
n
-
g
r
am
p
r
ec
is
i
o
n
-
m
o
d
if
ied
,
B
P
is
th
e
b
r
ev
it
y
p
e
n
alty
f
ac
to
r
,
wh
ich
ad
d
r
ess
es
th
e
is
s
u
e
wh
en
m
ac
h
in
e
tr
a
n
s
latio
n
o
u
t
p
u
ts
ar
e
s
h
o
r
ter
th
an
h
u
m
an
r
ef
er
e
n
ce
tr
an
s
latio
n
s
.
SAcc
m
ea
s
u
r
es
th
e
co
r
r
ec
tn
ess
o
f
an
en
tire
s
en
ten
ce
o
r
s
eq
u
en
ce
o
f
p
r
ed
ictio
n
s
b
y
c
o
m
p
ar
in
g
it
t
o
a
r
ef
er
en
ce
s
en
ten
ce
.
SAcc
en
s
u
r
es
th
at
th
e
g
e
n
er
ated
s
en
te
n
ce
m
atch
es
th
e
r
e
f
er
en
ce
w
ith
o
u
t
er
r
o
r
s
.
I
t
is
ca
lcu
lated
u
s
in
g
(
6
)
.
(
)
=
(
6
)
4.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
4
.
1
.
Resul
t
s
o
f
ex
perim
ent
1
(
I
m
pro
v
ed
o
bje
ct
det
ec
t
i
o
n
)
E
x
p
er
im
en
t
1
aim
s
to
im
p
r
o
v
e
o
b
ject
d
etec
tio
n
p
e
r
f
o
r
m
an
ce
u
s
in
g
YOL
Ov
7
an
d
YOL
Ov
8
,
th
e
latest
o
b
ject
d
etec
tio
n
m
eth
o
d
s
.
T
h
e
s
u
cc
ess
o
f
th
is
ex
p
er
im
en
t
is
m
ea
s
u
r
ed
b
y
th
e
in
cr
ea
s
e
in
m
AP
o
n
o
b
ject
d
etec
tio
n
,
th
e
im
p
r
o
v
em
en
t
in
ac
cu
r
ac
y
o
n
M
o
b
ileNet
ex
tr
ac
tio
n
f
ea
tu
r
e
s
,
th
e
en
h
an
ce
m
en
t
in
L
STM
class
if
icatio
n
,
an
d
th
e
r
ed
u
ctio
n
in
co
m
p
u
tatio
n
tim
e
at
ea
c
h
s
tag
e
(
T
ab
le
4
)
.
As
s
h
o
wn
in
T
ab
le
4
,
YOL
Ov
8
s
ig
n
if
ican
tly
im
p
r
o
v
es
o
b
jec
t
d
etec
tio
n
p
er
f
o
r
m
a
n
ce
,
as
in
d
icate
d
b
y
its
in
cr
ea
s
in
g
m
AP
s
co
r
es.
T
h
is
ad
v
an
ce
m
en
t
is
c
o
m
p
le
m
en
ted
b
y
en
h
a
n
ce
d
ac
c
u
r
ac
y
in
Mo
b
ileNet's
f
ea
tu
r
e
ex
tr
ac
tio
n
ca
p
ab
ilit
y
,
with
f
r
am
e
lab
el
class
if
icatio
n
s
h
o
win
g
s
u
b
s
tan
tial
im
p
r
o
v
em
en
t
wh
en
co
m
p
ar
e
d
to
Evaluation Warning : The document was created with Spire.PDF for Python.