I
AE
S
I
n
t
e
r
n
at
ion
al
Jou
r
n
al
of
Ar
t
if
icial
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
Vol.
14
,
No.
5
,
Oc
tober
2025
,
pp
.
3542
~
3553
I
S
S
N:
2252
-
8938
,
DO
I
:
10
.
11591/i
jai
.
v
14
.i
5
.
pp
35
42
-
3553
3542
Jou
r
n
al
h
omepage
:
ht
tp:
//
ij
ai
.
iaes
c
or
e
.
c
om
T
e
m
p
or
al
c
on
t
e
xt
o
f
l
ig
h
t
w
e
i
gh
t
n
e
t
w
o
r
k
m
od
e
l
f
or
d
e
t
e
c
t
i
n
g
b
oat
s ap
p
r
oa
c
h
i
n
g t
h
e
t
s
u
n
a
m
i
e
ar
ly
w
a
r
n
in
g syst
e
m
Wayan
Wira
Yogan
t
ar
a
1
,
3
,
S
u
p
r
i
j
an
t
o
2
,
A
n
ak
A
gu
n
g
N
gu
r
ah
A
n
an
d
a
Ku
s
u
m
a
4
,
Yuk
i
I
s
t
ian
t
o
5
1
M
a
s
te
r
'
s
P
r
ogr
a
m
of
I
ns
tr
ume
nt
a
ti
on a
nd C
ont
r
ol
, F
a
c
ul
ty
of
I
ndus
tr
ia
l
T
e
c
hnol
ogy, I
ns
ti
tu
t
T
e
knol
ogi
B
a
ndung,
B
a
ndung,
I
n
done
s
ia
2
I
n
s
t
r
u
m
e
nt
a
t
io
n
,
C
on
tr
o
l
a
n
d
A
u
t
om
a
t
io
n
R
e
s
e
a
r
c
h
G
r
o
u
p,
F
a
c
u
l
ty
o
f
I
nd
u
s
tr
i
a
l
T
e
c
h
n
o
lo
g
y
,
I
n
s
t
i
tu
t
T
e
k
no
l
og
i
B
a
nd
u
n
g,
B
a
n
du
n
g
,
I
n
do
n
e
s
i
a
3
R
e
s
e
a
r
c
h C
e
nt
e
r
f
or
E
le
c
tr
oni
c
s
, N
a
ti
ona
l
R
e
s
e
a
r
c
h
a
nd I
nnova
ti
on A
ge
nc
y, J
a
k
a
r
ta
, I
ndone
s
ia
4
R
e
s
e
a
r
c
h C
e
nt
e
r
f
or
T
e
le
c
omm
uni
c
a
ti
ons
, N
a
ti
ona
l
R
e
s
e
a
r
c
h a
nd I
nnova
ti
on A
ge
nc
y, J
a
ka
r
ta
, I
ndone
s
i
a
5
R
e
s
e
a
r
c
h C
e
nt
e
r
f
or
A
r
ti
f
ic
ia
l
I
nt
e
ll
ig
e
nc
e
a
nd C
ybe
r
S
e
c
ur
it
y,
N
a
ti
ona
l
R
e
s
e
a
r
c
h a
nd I
nnova
ti
on A
ge
nc
y,
J
a
ka
r
ta
, I
ndone
s
ia
Ar
t
icle
I
n
f
o
AB
S
T
RA
CT
A
r
ti
c
le
h
is
tor
y
:
R
e
c
e
ived
M
a
y
22,
2024
R
e
vis
e
d
J
ul
4,
2025
Ac
c
e
pted
Aug
6,
2025
T
h
e
t
s
u
n
ami
ear
l
y
w
arn
i
n
g
s
y
s
t
em
(
T
E
W
S)
i
s
a
d
ev
i
ce
t
h
a
t
d
et
ec
t
s
p
o
t
en
t
i
a
l
t
s
u
n
am
i
s
.
H
o
w
ev
er,
a
b
o
a
t
t
h
a
t
ap
p
ro
ac
h
es
T
E
W
S
i
s
a
s
o
u
rce
o
f
co
mmu
n
i
ca
t
i
o
n
d
i
s
t
u
r
b
an
ce.
A
co
n
v
o
l
u
t
i
o
n
al
n
e
u
ral
n
e
t
w
o
r
k
(CN
N
),
as
p
art
o
f
i
n
t
el
l
i
g
en
t
co
m
p
u
t
er
v
i
s
i
o
n
,
i
s
o
n
e
s
o
l
u
t
i
o
n
fo
r
d
et
ec
t
i
n
g
b
o
at
s
an
d
p
ro
v
i
d
i
n
g
a
w
arn
i
n
g
t
o
mo
v
e
aw
ay
fro
m
t
h
e
T
E
W
S
area.
W
at
er
s
eg
me
n
t
a
t
i
o
n
an
d
refi
n
emen
t
-
t
em
p
o
ra
l
(
W
aSR
-
T
)
,
as
t
h
e
cu
rren
t
ad
v
an
ce
d
CN
N
n
e
t
w
o
rk
,
e
x
h
i
b
i
t
s
i
m
p
res
s
i
v
e
p
erf
o
rman
ce
i
n
d
et
ec
t
i
n
g
o
b
j
ec
t
o
b
s
t
acl
es
i
n
t
h
e
mari
n
e
d
o
ma
i
n
,
al
t
h
o
u
g
h
i
t
req
u
i
re
s
a
p
o
w
erfu
l
c
o
mp
u
t
a
t
i
o
n
a
l
d
ev
i
ce.
In
t
h
e
p
ap
er,
w
e
p
r
o
p
o
s
e
a
mo
d
i
fi
ca
t
i
o
n
o
f
W
aSR
-
T
,
rep
l
aci
n
g
t
h
e
mo
s
t
co
mp
u
t
a
t
i
o
n
a
l
l
y
i
n
t
e
n
s
i
v
e
s
t
a
g
es
w
i
t
h
a
l
i
g
h
t
w
e
i
g
h
t
v
er
s
i
o
n
ca
l
l
e
d
l
i
g
h
t
w
e
i
g
h
t
W
aSR
-
T
.
O
n
t
h
e
p
r
o
p
o
s
e
d
l
i
g
h
t
w
ei
g
h
t
W
aSR
-
T
,
t
h
e
p
rev
i
o
u
s
en
co
d
er
o
f
W
aSR
-
T
w
as
rep
l
aced
w
i
t
h
Mo
b
i
l
e
N
e
t
V
3
,
an
d
s
o
m
e
feat
u
re
l
ay
er
ma
p
s
w
ere
re
d
u
ce
d
as
i
n
p
u
t
t
o
t
h
e
d
ec
o
d
er.
F
o
r
t
r
ai
n
i
n
g
an
d
v
a
l
i
d
at
i
n
g
t
h
e
l
i
g
h
t
w
ei
g
h
t
W
aSR
-
T
,
t
h
e
i
mag
e
d
at
a
s
et
re
p
res
e
n
t
i
n
g
t
h
e
o
p
en
s
ea
an
d
o
u
r
e
x
t
e
n
d
e
d
d
a
t
as
e
t
fro
m
I
n
d
o
n
e
s
i
a's
o
cea
n
reg
i
o
n
w
er
e
u
s
e
d
.
Bas
ed
o
n
t
h
e
q
u
a
n
t
i
t
a
t
i
v
e
res
u
l
t
s
an
d
e
v
al
u
at
i
o
n
o
f
t
h
e
co
mp
u
t
a
t
i
o
n
a
l
l
o
ad
,
t
h
e
s
en
s
i
t
i
v
i
t
y
t
o
d
et
ec
t
a
b
o
at
fo
r
W
aSR
-
T
a
n
d
l
i
g
h
t
w
ei
g
h
t
W
aS
R
-
T
i
s
9
5
.
7
1
%
a
n
d
9
0
.
0
0
%
,
res
p
ec
t
i
v
el
y
.
T
h
e
l
i
g
h
t
w
e
i
g
h
t
W
aSR
-
T
re
q
u
i
red
l
e
s
s
memo
r
y
at
3
2
.
5
7
%
,
res
u
l
t
i
n
g
i
n
a
0
.
0
7
6
1
%
red
u
ct
i
o
n
i
n
t
o
t
al
p
ro
ce
s
s
i
n
g
t
i
me
c
o
mp
are
d
t
o
t
h
e
o
ri
g
i
n
al
W
aSR
-
T
.
T
h
erefo
re,
o
u
r
p
ro
p
o
s
ed
l
i
g
h
t
w
ei
g
h
t
W
aSR
-
T
i
s
p
ro
m
i
s
i
n
g
fo
r
u
s
e
as
t
h
e
cen
t
ral
p
ar
t
o
f
an
i
n
t
el
l
i
g
en
t
mari
t
i
me
co
mp
u
t
e
r
v
i
s
i
o
n
s
y
s
t
em
i
n
T
E
W
S.
K
e
y
w
o
r
d
s
:
L
ight
we
ight
W
a
S
R
-
T
M
a
r
it
im
e
c
omput
e
r
vis
ion
M
obil
e
Ne
tV3
N
e
t
wo
r
k
ma
r
in
e
o
bj
e
c
t
de
te
c
t
io
n
T
s
una
mi
e
a
r
ly
wa
r
ning
s
ys
tem
Th
i
s
i
s
a
n
o
p
en
a
c
ces
s
a
r
t
i
c
l
e
u
n
d
e
r
t
h
e
CC
B
Y
-
SA
l
i
ce
n
s
e.
C
or
r
e
s
pon
din
g
A
u
th
or
:
S
upr
ij
a
nto
I
ns
tr
umenta
ti
on,
C
ont
r
ol
a
nd
Auto
mation
R
e
s
e
a
r
c
h
Gr
oup,
F
a
c
ult
y
o
f
I
ndus
tr
ial
T
e
c
hnology
I
ns
ti
tut
T
e
knologi
B
a
ndung
B
a
ndung,
I
ndone
s
ia
E
mail:
s
upr
i89@i
tb
.
a
c
.
id
1.
I
NT
RODU
C
T
I
O
N
A
t
s
una
mi
e
a
r
ly
wa
r
ning
s
ys
tem
(
T
E
W
S
)
is
c
r
uc
ial
f
or
mi
nim
izing
the
potential
da
mage
a
nd
los
s
of
li
f
e
c
a
us
e
d
by
na
tu
r
a
l
dis
a
s
ter
s
.
T
E
W
S
ba
s
e
d
on
ts
una
mi
buoys
a
r
e
de
ployed
in
the
de
e
p
oc
e
a
n
c
l
os
e
to
a
pos
s
ibl
e
loca
ti
on
of
s
e
a
e
a
r
thquake
s
a
nd
e
quipped
with
pr
e
s
s
ur
e
s
e
ns
or
s
to
de
tec
t
c
ha
nge
s
in
wa
ter
l
e
ve
l
[
1]
.
W
he
n
a
ts
una
mi
pa
s
s
e
s
ove
r
the
buoy
,
it
r
e
gis
te
r
s
the
pr
e
s
s
ur
e
c
ha
nge
a
nd
s
e
nds
thi
s
da
ta
to
m
onit
or
ing
s
tations
,
whic
h
c
a
n
then
a
s
s
e
s
s
the
potential
thr
e
a
t
f
r
om
the
ts
una
mi
.
T
he
qua
li
ty
of
da
ta
tr
a
ns
f
e
r
us
ing
a
c
ous
ti
c
li
nk
c
omm
unica
ti
on
de
pe
nds
on
the
dis
tanc
e
r
a
nge
be
twe
e
n
a
n
oc
e
a
n
bott
om
unit
(
OB
U)
a
nd
a
s
ur
f
a
c
e
buoy
,
a
s
we
ll
a
s
nois
e
int
e
r
f
e
r
e
nc
e
due
to
unde
s
ir
e
d
a
c
ous
ti
c
wa
ve
s
.
T
he
pr
e
s
e
nc
e
of
una
uthor
ize
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
T
e
mpor
al
c
ontex
t
of
li
ghtw
e
ight
ne
tw
or
k
mode
l
for
de
tec
ti
ng
boats
appr
oac
hing
…
(
W
ay
an
W
ir
a
Y
og
antar
a
)
3543
boa
ts
in
the
ins
tallation
a
r
e
a
of
the
ts
una
mi
buoy
s
ys
tem
is
a
s
our
c
e
of
dis
tur
ba
nc
e
that
mus
t
be
a
voided.
A
de
vice
that
de
tec
ts
una
uthor
ize
d
boa
ts
a
ppr
oa
c
hi
ng
the
s
ur
f
a
c
e
buoy
is
r
e
quir
e
d
to
pr
e
ve
nt
dis
r
upti
on
to
the
f
unc
ti
on
of
th
is
ts
una
mi
de
tec
ti
on
s
ys
tem.
T
he
de
v
ice
ha
s
not
ye
t
be
e
n
ins
talled
on
the
e
xis
ti
ng
T
E
W
S
s
ys
tem
in
the
I
ndone
s
ian
ope
n
s
e
a
.
Va
r
ious
s
e
ns
or
s
,
s
uc
h
a
s
mar
ine
r
a
da
r
a
nd
opti
c
a
l
s
e
ns
or
-
ba
s
e
d
c
a
mer
a
s
,
a
r
e
c
omm
only
us
e
d
f
or
objec
t
de
tec
ti
on
on
the
s
e
a
s
ur
f
a
c
e
[
2]
.
A
de
vice
with
a
n
opti
c
a
l
c
a
mer
a
is
ga
ini
ng
p
r
omi
ne
nc
e
a
s
a
lea
ding
objec
t
de
tec
ti
on
method
w
he
n
s
uppor
ted
by
pr
o
pe
r
c
omput
e
r
vis
ion
methods
[
3]
–
[
5
]
.
P
r
e
vious
c
las
s
ica
l
c
omput
e
r
vis
ion
methods
f
o
r
objec
t
de
tec
ti
on
r
e
ly
on
s
im
ple
c
lus
ter
ing
of
f
e
a
tur
e
s
r
e
late
d
to
s
pe
c
if
i
c
gr
oups
of
objec
ts
,
whic
h
a
r
e
not
e
xpr
e
s
s
ive
e
nough
f
or
a
c
c
ur
a
te
de
tec
ti
on
in
ob
jec
t
de
tec
ti
on
e
nvir
on
m
e
nts
[
6]
.
A
c
ur
r
e
nt
int
e
ll
igent
c
omput
e
r
vis
ion
method
f
or
objec
t
de
tec
ti
on
is
ba
s
e
d
on
c
las
s
if
ying
e
ve
r
y
pix
e
l
in
a
n
im
a
ge
us
ing
a
c
onvolut
ional
ne
ur
a
l
ne
twor
k
(
C
N
N)
to
pr
ovide
va
luable
in
f
or
mation
f
or
s
c
e
ne
unde
r
s
tanding
a
nd
objec
t
de
tec
ti
on
[
7]
–
[
9]
.
T
he
s
tate
-
of
-
the
-
a
r
t
(
S
OT
A)
C
NN
ne
twor
k
f
or
int
e
ll
igent
c
omput
e
r
vis
i
on
objec
t
de
tec
ti
o
n
ha
s
be
c
ome
a
n
e
s
tablis
he
d
a
ppr
oa
c
h
in
a
utonom
ous
gr
ound
ve
hicle
s
[
10]
–
[
14]
.
H
owe
ve
r
,
the
e
xis
ti
ng
S
OT
A
C
NN
ne
twor
k
,
pr
im
a
r
il
y
de
v
e
loped
in
ter
r
e
s
tr
ial
gr
ound
s
c
e
ne
s
,
is
inade
qua
t
e
in
the
mar
it
im
e
domain
.
T
he
wa
ter
s
e
gmenta
ti
on
a
nd
r
e
f
ineme
nt
(
W
a
S
R
)
[
15]
–
[
17]
is
one
o
f
the
S
O
T
A
C
NN
ne
twor
ks
that
ha
s
good
pe
r
f
or
manc
e
f
o
r
unmanne
d
s
ur
f
a
c
e
ve
hi
c
les
(
USVs
)
.
T
he
W
a
S
R
ne
twor
k
c
ons
is
ts
of
a
c
o
ntr
a
c
ti
ng
pa
th
(
e
nc
ode
r
)
a
nd
a
n
e
xpa
ns
ive
pa
th
(
de
c
ode
r
)
c
ompr
is
ing
s
e
ve
r
a
l
inf
or
mation
f
us
ion
a
nd
f
e
a
tur
e
s
c
a
li
ng
blocks
.
T
he
W
a
S
R
de
c
ode
r
wa
s
de
s
igned
to
f
us
e
iner
ti
a
l
mea
s
ur
e
ment
unit
(
I
M
U)
a
nd
vis
ua
l
im
a
ge
inf
or
mation
to
incr
e
a
s
e
the
a
c
c
ur
a
c
y
of
pos
it
ive
objec
t
de
tec
ti
ons
.
T
he
a
r
c
hit
e
c
tur
e
upgr
a
de
of
th
e
W
a
S
R
ne
twor
k
wa
s
p
r
opos
e
d
to
e
xtr
a
c
t
the
tempor
a
l
c
ont
e
xt
f
r
om
a
s
e
que
nc
e
of
f
r
a
mes
to
di
f
f
e
r
e
nti
a
te
obje
c
ts
f
r
om
r
e
f
lec
ti
ons
,
a
nd
it
is
c
a
ll
e
d
W
a
S
R
-
T
[
18]
.
W
a
S
R
-
T
ha
s
be
e
n
r
e
por
ted
to
r
e
duc
e
the
number
of
f
a
ls
e
pos
it
ive
(
F
P
)
de
tec
ti
ons
c
ompar
e
d
with
W
a
S
R
.
W
e
r
e
quir
e
d
a
S
O
T
A
C
NN
ne
twor
k
to
de
tec
t
una
u
thor
ize
d
boa
ts
a
ppr
oa
c
hing
ts
una
mi
buoys
a
s
pa
r
t
of
a
n
int
e
ll
igent
c
omput
e
r
vis
ion
s
ys
tem.
W
a
S
R
-
T
is
one
of
the
be
s
t
-
pe
r
f
or
mi
ng
mar
it
im
e
obs
tac
les
f
or
USVs
[
17]
.
T
he
r
e
f
o
r
e
,
W
a
S
R
-
T
is
one
of
the
c
a
ndidate
s
f
or
the
una
uthor
ize
d
s
hip
de
tec
ti
on
s
ys
tem.
W
hil
e
a
c
hieving
im
pr
e
s
s
ive
objec
t
de
tec
ti
on
r
e
s
ult
s
f
or
USVs
,
the
W
a
S
R
-
T
ne
twor
k
r
e
qui
r
e
s
a
powe
r
f
ul
c
omput
a
ti
ona
l
de
vice
[
6]
.
I
n
thi
s
wor
k,
we
pr
opos
e
d
a
modi
f
ica
ti
on
of
W
a
S
R
-
T
with
r
e
plac
e
ments
f
or
the
mos
t
c
omput
a
ti
ona
ll
y
int
e
ns
ive
s
tage
s
,
c
a
ll
e
d
the
li
ghtwe
ight
W
a
S
R
-
T
.
T
he
objec
ti
ve
of
li
ghtwe
ight
W
a
S
R
-
T
is
to
c
las
s
if
y
a
nd
labe
l
e
a
c
h
pixel
of
a
r
e
c
or
de
d
im
a
g
e
a
s
a
n
una
uthor
ize
d
boa
t
wi
th
a
ba
c
kgr
ound
s
e
a
or
a
r
e
a
int
e
r
f
a
c
e
be
twe
e
n
the
s
e
a
a
nd
the
s
ky
that
a
ppr
oa
c
he
s
T
E
W
S
with
inexpe
ns
ive
c
omput
a
ti
ona
l
r
e
s
our
c
e
s
a
nd
go
od
pe
r
f
o
r
manc
e
.
F
or
the
tr
a
ini
ng
a
nd
va
li
da
ti
on
of
the
li
ghtwe
ight
W
a
S
R
-
T
,
a
unique
da
tas
e
t
wa
s
uti
li
z
e
d
that
a
c
c
ur
a
tely
r
e
pr
e
s
e
nts
the
ope
n
s
e
a
.
T
his
da
ta
s
e
t
wa
s
s
e
lec
ted
f
r
om
the
ope
n
da
tas
e
t
mar
it
i
me
s
e
mantic
s
e
gmenta
ti
on
tr
a
ini
ng
[
18]
,
[
19]
,
a
nd
f
r
om
da
ta
s
e
ts
a
va
il
a
ble
on
a
n
op
e
n
we
bs
it
e
.
W
e
a
l
s
o
c
r
e
a
ted
a
n
e
xtens
ive
da
tas
e
t
f
r
om
the
I
ndone
s
ian
T
E
W
S
a
r
e
a
,
whic
h
is
ins
talled
in
the
ope
n
s
e
a
.
T
he
n,
the
a
bil
it
y
of
W
a
S
R
-
T
a
nd
li
ghtwe
ight
W
a
S
R
-
T
to
de
tec
t
a
n
una
uthor
ize
d
boa
t
us
ing
thes
e
da
tas
e
ts
wa
s
pe
r
f
or
med.
T
his
pa
pe
r
is
s
tr
uc
tur
e
d
a
s
f
o
ll
ows
:
s
e
c
ti
on
2
int
r
oduc
e
s
the
mate
r
ial
a
nd
methods
,
including
the
c
onc
e
pt
of
una
uthor
ize
d
boa
t
de
tec
ti
on
f
or
T
E
W
S
,
the
da
tas
e
ts
us
e
d
f
or
de
ve
lopi
ng
the
S
O
T
A
C
NN
ne
twor
k,
a
nd
a
r
e
view
of
the
e
xis
ti
ng
a
r
c
hit
e
c
tur
e
,
W
a
S
R
-
T
,
a
nd
li
ghtwe
ight
W
a
S
R
-
T
.
S
e
c
ti
on
3
de
s
c
r
ibes
the
de
tails
of
the
e
xpe
r
im
e
nt
pr
oc
e
dur
e
s
.
T
he
r
e
s
ult
s
a
nd
dis
c
us
s
ion
a
r
e
de
s
c
r
ibed
in
s
e
c
ti
on
4.
F
inal
ly,
in
s
e
c
ti
on
5,
we
dr
a
w
our
c
onc
lus
ions
.
2.
M
E
T
HO
D
2.
1
.
T
h
e
c
on
c
e
p
t
of
u
n
au
t
h
or
ize
d
b
oa
t
d
e
t
e
c
t
io
n
f
or
t
h
e
t
s
u
n
am
i
b
u
oy
s
ys
t
e
m
T
he
typi
c
a
l
ts
una
mi
buoy
s
ys
tem
c
ons
is
ts
of
two
main
pa
r
ts
:
a
n
OB
U
that
mea
s
ur
e
s
c
ha
nge
s
in
s
e
a
leve
l
he
ight
a
t
the
s
e
a
be
d
a
nd
a
s
ur
f
a
c
e
buoy
tha
t
tr
a
ns
mi
ts
mea
s
ur
e
ment
da
ta
to
a
ts
una
mi
da
ta
c
e
nter
a
s
s
hown
in
F
igur
e
1.
T
he
OB
U
a
nd
s
ur
f
a
c
e
buoy
e
xc
ha
nge
inf
or
mation
us
ing
a
c
ous
ti
c
c
omm
unica
ti
on
modems
that
a
r
e
ve
r
y
s
us
c
e
pti
ble
to
nois
e
int
e
r
f
e
r
e
nc
e
,
f
or
e
xa
mpl
e
,
f
r
om
the
s
hip's
pr
ope
ll
e
r
s
.
I
n
s
ome
c
a
s
e
s
,
a
s
ur
f
a
c
e
buoy
is
us
e
d
a
s
a
boa
t
moo
r
ing,
whic
h
c
a
n
c
r
e
a
te
pr
oblems
in
a
c
ous
ti
c
li
nk
c
omm
unica
ti
on
d
ue
to
the
s
hif
ti
ng
pos
it
ion
o
f
the
buoy
s
ys
tem.
I
t
may
a
ls
o
im
pa
c
t
the
c
omm
unica
ti
on
c
ha
nne
l
be
twe
e
n
the
OB
U
a
nd
the
s
ur
f
a
c
e
buoy
[
20]
.
An
una
uthor
ize
d
boa
t
a
ppr
oa
c
hing
the
T
E
W
S
ins
tallation
a
r
e
a
is
a
s
our
c
e
of
dis
tur
ba
nc
e
.
B
oa
ts
or
other
objec
ts
a
ppr
oa
c
hing
t
he
ins
tallation
a
r
e
a
mus
t
be
wa
r
ne
d
to
lea
ve
im
media
tely
a
nd
noti
f
y
the
da
ta
c
e
nter
of
f
ice
o
f
the
r
e
s
ult
ing
di
s
tur
ba
nc
e
s
.
An
in
telli
ge
nt
c
omput
e
r
vis
ion
s
ys
tem
is
one
o
f
the
s
olut
ions
that
may
be
ins
talled
on
the
highes
t
s
ur
f
a
c
e
of
the
boys
a
s
s
hown
in
F
igur
e
1
.
T
he
S
OT
A
C
NN
n
e
t
wor
k,
a
s
th
e
c
e
ntr
a
l
p
a
r
t
of
a
n
i
nt
e
ll
ig
e
nt
c
o
mp
ut
e
r
vi
s
io
n
s
y
s
te
m,
m
u
s
t
b
e
d
e
ve
lo
pe
d
ba
s
e
d
o
n
t
he
c
on
c
e
pt
of
s
e
m
a
n
ti
c
s
e
g
m
e
nt
a
ti
o
n
[
21]
–
[
23]
.
S
e
ma
nt
ic
s
e
gm
e
n
ta
ti
on
i
s
u
s
e
d
t
o
c
l
a
s
s
if
y
a
nd
l
a
be
l
e
a
c
h
p
ix
e
l
of
a
r
e
c
or
d
e
d
i
mag
e
a
s
e
it
h
e
r
a
n
un
a
u
th
or
i
z
e
d
b
o
a
t,
a
ba
c
kgr
ou
nd
w
a
t
e
r
s
u
r
f
a
c
e
,
or
a
n
a
r
e
a
i
nt
e
r
f
a
c
e
b
e
t
we
e
n
t
h
e
s
e
a
a
nd
s
k
y.
T
hi
s
S
OT
A
C
N
N
ne
tw
or
k
m
u
s
t
b
e
a
bl
e
to
d
e
te
c
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
5
,
Oc
tober
2025
:
35
42
-
3553
3544
s
t
a
t
i
c
a
n
d
dy
na
mi
c
un
a
uth
or
i
z
e
d
b
o
a
t
s
of
v
a
r
io
u
s
s
h
a
p
e
s
a
n
d
s
iz
e
s
,
in
c
lu
din
g
th
o
s
e
n
ot
s
e
e
n
d
ur
in
g
t
r
a
in
in
g.
Ad
di
ti
o
n
a
ll
y,
th
e
S
OT
A
C
N
N
n
e
t
wor
k
mu
s
t
b
e
hi
gh
ly
a
d
a
pt
a
bl
e
to
c
h
a
ll
e
ng
in
g
a
n
d
d
y
na
mi
c
w
a
te
r
f
e
a
tur
e
s
,
a
ll
o
win
g
T
E
W
S
to
op
e
r
a
t
e
e
f
f
e
c
ti
v
e
ly.
D
ue
to
e
nv
ir
o
nm
e
n
t
s
a
nd
c
on
dit
io
n
s
,
a
n
int
e
l
li
ge
nt
c
o
mp
ut
e
r
v
i
s
io
n
a
n
d
S
OT
A
C
N
N
n
e
t
w
or
k
mu
s
t
w
or
k
f
or
r
e
a
l
-
wor
ld
s
ma
ll
-
s
iz
e
d
e
ne
r
gy
-
c
on
s
tr
a
i
ne
d
T
E
W
S
.
T
he
r
e
f
or
e
,
a
li
gh
tw
e
i
gh
t
n
e
t
wo
r
k
th
a
t
c
a
n
be
r
u
n
o
n
a
d
e
v
ic
e
wit
h
l
im
i
t
e
d
me
mor
y
a
nd
a
s
m
a
l
l
a
r
c
hi
t
e
c
tur
a
l
d
e
vi
c
e
i
s
r
e
qui
r
e
d.
F
igur
e
1.
C
onf
igur
a
ti
on
of
the
I
ndone
s
ian
T
E
W
S
2.
2
.
Op
e
n
s
e
a
im
age
d
at
as
e
t
s
f
or
t
h
e
d
e
ve
lop
m
e
n
t
of
S
OT
A
CNN
n
e
t
wor
k
D
e
te
c
t
in
g
ob
je
c
ts
o
n
ma
r
it
im
e
op
e
n
s
e
a
s
w
it
h
C
N
N
s
pr
e
s
e
nt
s
s
e
v
e
r
a
l
c
ha
ll
e
n
ge
s
,
l
ik
e
th
e
c
o
mpl
e
xit
y
of
o
bj
e
c
t
s
,
e
nv
ir
o
nm
e
n
t
a
l
c
on
di
ti
o
ns
,
d
a
t
a
a
va
il
a
b
il
it
y,
c
l
a
s
s
im
ba
la
n
c
e
,
a
n
om
a
l
y
de
t
e
c
ti
on,
s
m
a
l
l
ob
je
c
t
d
e
t
e
c
ti
on,
a
n
d
m
od
e
l
r
o
bu
s
tn
e
s
s
.
L
a
r
ge
d
a
t
a
s
e
t
s
r
e
le
va
nt
t
o
t
h
e
o
p
e
n
s
e
a
do
ma
in
a
r
e
r
e
q
uir
e
d
t
o
s
u
pp
or
t
t
he
d
e
v
e
l
op
me
nt
of
t
h
e
S
OT
A
C
N
N
n
e
t
w
or
k
a
s
t
h
e
c
e
ntr
a
l
p
a
r
t
of
a
n
in
te
ll
ig
e
n
t
c
om
pu
te
r
v
i
s
i
on
s
y
s
t
e
m
f
or
T
E
W
S
.
W
e
c
a
n
no
t
r
e
l
y
s
o
le
ly
on
c
ur
r
e
n
tl
y
a
v
a
il
a
bl
e
m
a
r
i
ti
me
d
a
t
a
s
e
t
s
,
s
u
c
h
a
s
th
e
m
a
r
it
i
me
s
e
m
a
n
ti
c
s
e
gm
e
nt
a
ti
on
tr
a
i
ni
ng
d
a
t
a
s
e
t
[
19]
,
b
e
c
a
u
s
e
m
o
s
t
e
xi
s
t
in
g
i
ma
g
e
s
s
h
ow
s
e
a
c
o
nd
it
i
on
s
i
n
c
a
lm
a
r
e
a
s
s
u
c
h
a
s
b
a
y
s
,
h
a
r
bor
s
,
r
i
ve
r
s
,
o
r
e
s
t
ua
r
y
a
r
e
a
s
.
I
n
c
o
ntr
a
s
t,
th
e
a
va
il
a
b
le
d
a
ta
s
e
t
i
s
r
e
la
ti
ve
ly
l
im
i
t
e
d
f
or
o
pe
n
s
e
a
c
on
dit
io
n
s
.
T
he
pr
opos
e
d
C
NN
ne
twor
k
mus
t
be
tr
a
ined
with
a
n
im
a
ge
da
tas
e
t
r
e
leva
nt
to
ope
n
-
s
e
a
e
nvir
onments
.
T
he
r
e
f
or
e
,
in
the
s
tudy,
ope
n
-
s
e
a
domain
da
ta
s
e
ts
we
r
e
r
e
c
or
de
d
dir
e
c
tl
y
f
r
o
m
the
lo
c
a
ti
on
in
the
ope
n
s
e
a
of
the
I
ndone
s
ia
n
Oc
e
a
n,
whe
r
e
T
E
W
S
wa
s
c
ur
r
e
ntl
y
ins
talled,
to
obtain
a
c
tual
s
e
a
c
ondit
ions
.
Additi
ona
l
da
tas
e
ts
of
mar
ine
domains
r
e
leva
nt
t
o
c
ondit
ions
on
the
ope
n
s
e
a
we
r
e
a
ls
o
c
oll
e
c
ted
f
r
o
m
a
n
ope
n
we
bs
it
e
.
2.
2.
1.
Op
e
n
d
at
as
e
t
s
A
da
tas
e
t
c
ontaining
im
a
ge
f
r
a
mes
typ
ica
l
of
c
on
dit
ions
on
the
ope
n
s
e
a
is
r
e
quir
e
d
to
de
ve
lop
the
S
OT
A
-
C
NN
a
lgor
it
hm
f
or
una
utho
r
ize
d
boa
t
de
tec
ti
on
on
T
E
W
S
.
Ope
nly
a
c
c
e
s
s
ibl
e
im
a
ge
tr
a
ini
ng
da
tas
e
ts
f
or
the
mar
ine
domain
ha
ve
be
e
n
a
va
il
a
ble
f
or
tail
or
ing
to
de
ve
lop
obs
tac
le
de
tec
ti
on
method
s
in
s
m
a
ll
-
s
ize
d,
USVs
.
F
or
e
xa
mpl
e
,
M
a
S
T
r
1325
is
the
da
tas
e
t
that
c
ontains
1325
diver
s
e
im
a
ge
s
c
a
ptur
e
d
ove
r
two
ye
a
r
s
with
r
e
a
l
USV
,
c
ove
r
ing
a
r
a
nge
o
f
r
e
a
li
s
ti
c
c
ondit
ions
e
nc
ounter
e
d
in
a
c
oa
s
tal
s
ur
ve
il
lanc
e
tas
k
[
19]
.
T
he
e
xtende
d
da
tas
e
t
of
M
a
S
T
r
13
25
a
dde
d
153
im
a
ge
s
(
including
thei
r
p
r
e
c
e
ding
f
r
a
mes
)
a
nd
us
e
d
the
c
ode
na
me
M
a
S
T
r
1478
f
or
thi
s
a
ddit
ional
da
tas
e
t.
M
a
S
T
r
1
478
tr
a
ini
ng
im
a
ge
s
r
e
pr
e
s
e
nt
c
ha
ll
e
nging
objec
t
s
due
to
mi
r
r
or
ing
,
r
e
f
lec
ti
ons
,
a
nd
s
un
gli
tt
e
r
s
[
18]
.
T
he
e
nti
r
e
da
tas
e
t
wa
s
e
xpli
c
i
tl
y
c
r
e
a
ted
f
o
r
tr
a
ini
ng
t
he
USV;
L
a
tely,
the
pr
im
a
r
y
pu
r
pos
e
of
thi
s
USV
is
to
s
ur
v
e
y
ve
s
s
e
l
s
in
a
c
oa
s
tal
domain,
s
o
that
s
e
a
c
ondit
io
ns
,
s
uc
h
a
s
thos
e
in
ba
ys
,
r
iver
e
s
tuar
ies
,
a
nd
ha
r
bo
r
s
,
dom
i
na
te
the
main
f
r
a
me
im
a
ge
s
.
Additi
ona
l
da
tas
e
ts
of
mar
ine
domains
r
e
leva
nt
t
o
c
ondit
ions
in
the
ope
n
s
e
a
we
r
e
a
ls
o
c
oll
e
c
ted
f
r
om
a
n
ope
n
-
a
c
c
e
s
s
we
bs
it
e
[
24]
.
W
e
s
e
lec
ted
1
50
im
a
ge
s
that
c
los
e
ly
r
e
s
e
mbl
e
the
typi
c
a
l
c
ondit
ions
a
nd
objec
ts
c
a
ptur
e
d
by
the
c
a
mer
a
on
the
ope
n
s
e
a
.
T
he
s
e
lec
ted
im
a
ge
s
r
e
pr
e
s
e
nt
va
r
io
us
s
ha
pe
s
a
nd
s
ize
s
of
boa
ts
c
a
ptur
e
d
withi
n
the
s
pe
c
if
ic
f
ield
of
view
of
the
c
a
mer
a
,
unde
r
dif
f
e
r
e
nt
we
a
ther
c
ondit
ions
.
E
xa
mpl
e
s
of
s
e
lec
ti
ng
ope
n
-
s
e
a
f
r
a
me
im
a
ge
s
f
r
om
the
M
a
S
T
r
1325
a
nd
M
a
S
T
r
1478
da
tas
e
ts
,
a
s
we
ll
a
s
f
r
om
the
ope
n
we
bs
it
e
,
a
r
e
s
hown
in
F
igur
e
2
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
T
e
mpor
al
c
ontex
t
of
li
ghtw
e
ight
ne
tw
or
k
mode
l
for
de
tec
ti
ng
boats
appr
oac
hing
…
(
W
ay
an
W
ir
a
Y
og
antar
a
)
3545
F
igur
e
2.
An
e
xa
mpl
e
o
f
s
e
lec
ti
ng
ope
n
s
e
a
f
r
a
me
i
mage
s
f
r
om
the
M
a
S
T
r
1325,
M
a
S
T
r
1478
da
tas
e
ts
,
a
nd
the
ope
n
we
bs
it
e
2.
2.
2.
Re
c
or
d
in
g
i
m
age
d
a
t
as
e
t
s
f
r
om
t
h
e
op
e
n
s
e
a
in
I
n
d
on
e
s
ian
t
e
r
r
it
or
y
T
E
W
S
pr
im
a
r
il
y
ope
r
a
tes
in
the
ope
n
s
e
a
o
f
I
nd
one
s
ian
ter
r
it
or
y
,
whic
h
is
r
e
latively
f
a
r
f
r
om
the
c
oa
s
t.
Ne
w
da
tas
e
ts
we
r
e
r
e
c
or
de
d
f
r
om
loca
ti
ons
c
los
e
to
the
ope
r
a
ti
ona
l
a
r
e
a
of
T
E
W
S
.
T
he
I
n
done
s
ian
T
E
W
S
a
r
e
p
r
im
a
r
il
y
ope
r
a
ted
in
the
ope
n
s
e
a
of
t
he
I
ndi
a
n
Oc
e
a
n,
in
the
s
outher
n
r
e
gion
o
f
J
a
va
I
s
land
,
a
nd
on
the
we
s
t
c
oa
s
t
of
S
umatr
a
I
s
land.
T
he
f
ive
e
xis
ti
ng
T
E
W
S
loca
ti
ons
a
r
e
il
lus
tr
a
ted
in
F
igur
e
3(
a
)
a
nd
the
ts
una
mi
buoy
de
ploym
e
nt
pr
oc
e
s
s
a
t
s
e
a
is
s
hown
i
n
F
igur
e
3
(
b)
.
T
h
e
ne
w
d
a
t
a
s
e
t
w
a
s
r
e
c
o
r
d
e
d
f
r
o
m
the
c
a
me
r
a
ins
ta
l
led
o
n
th
e
B
a
r
un
a
J
a
ya
r
e
s
e
a
r
c
h
ve
s
s
e
l
,
w
hi
c
h
is
r
e
s
po
ns
i
bl
e
f
o
r
r
e
c
o
ve
r
in
g
a
nd
ma
in
ta
in
in
g
th
e
T
E
W
S
s
ys
t
e
m
.
W
e
s
ys
tem
a
t
ic
a
l
ly
r
e
c
o
r
de
d
i
ma
ge
s
r
e
p
r
e
s
e
nt
ing
v
a
r
i
ous
c
on
di
t
io
ns
on
t
he
o
pe
n
s
e
a
ne
a
r
th
e
T
E
W
S
s
ys
tem
,
inc
l
ud
in
g
mo
r
ni
ng
,
a
f
t
e
r
n
oo
n
w
it
h
pe
a
k
s
u
nl
ig
h
t,
a
nd
e
ve
ni
ng
a
s
t
he
s
un
s
e
ts
.
A
n
e
xa
mp
le
of
r
e
c
o
r
de
d
i
ma
ge
s
us
e
d
f
o
r
t
r
a
in
i
ng
a
nd
tes
ti
ng
ou
r
p
r
o
pos
e
d
S
O
T
A
C
NN
ne
tw
or
k
is
s
ho
wn
i
n
F
i
gu
r
e
4
.
T
h
e
s
e
i
mag
e
s
r
e
p
r
e
s
e
n
t
va
r
i
ous
s
h
a
pe
s
a
nd
s
iz
e
s
o
f
u
na
ut
ho
r
ize
d
boa
ts
c
a
p
tu
r
e
d
i
n
the
s
p
e
c
i
f
ic
f
i
e
l
d
o
f
vi
e
w
o
f
th
e
c
a
me
r
a
in
s
ta
l
led
on
t
he
B
a
r
un
a
J
a
y
a
r
e
s
e
a
r
c
h
ve
s
s
e
l
.
(
a
)
(
b)
F
igur
e
3.
T
he
f
ive
e
xis
ti
ng
T
E
W
S
loca
ti
ons
a
r
e
mar
ke
d
with
(
a
)
B
KG
,
S
UN
,
C
XP,
M
L
G,
a
nd
DPS
,
a
nd
(
b)
a
n
e
xa
mpl
e
of
the
t
s
una
mi
boys
a
s
pa
r
t
of
T
E
W
S
on
loca
ti
on
S
UN
F
igur
e
4.
E
xa
mpl
e
of
im
a
ge
s
us
e
d
f
or
tr
a
ini
ng
a
nd
tes
ti
ng
our
pr
opos
e
d
ne
twor
k
f
r
om
the
c
a
mer
a
ins
t
a
ll
e
d
on
the
B
a
r
una
J
a
ya
r
e
s
e
a
r
c
h
ve
s
s
e
l
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
5
,
Oc
tober
2025
:
35
42
-
3553
3546
2.
3.
S
OT
A
CN
N
n
e
t
wor
k
f
or
b
oat
d
e
t
e
c
t
io
n
in
t
h
e
op
e
n
s
e
a
M
ode
r
n
C
NN
a
r
c
hit
e
c
tur
e
s
a
t
the
f
or
e
f
r
on
t
of
ma
r
it
im
e
objec
t
de
tec
ti
on
typi
c
a
ll
y
e
mpl
oy
s
e
mantic
s
e
gmenta
ti
on
methods
.
T
he
s
e
a
ppr
oa
c
he
s
pr
oc
e
s
s
vis
ua
l
input
a
t
the
pixel
leve
l,
c
a
tegor
izing
e
a
c
h
pixel
int
o
dis
ti
nc
t
c
las
s
e
s
—
c
omm
only
s
e
a
,
s
ky,
or
obs
tr
uc
ti
ve
e
leme
nts
s
u
c
h
a
s
ve
s
s
e
l
s
,
buoys
,
or
f
loating
d
e
br
is
[
6]
.
T
his
gr
a
nular
c
las
s
if
ica
ti
on
a
ll
ows
f
or
a
c
c
ur
a
te
identif
ica
ti
on
a
nd
loca
li
z
a
ti
on
of
objec
ts
withi
n
the
c
ompl
e
x
a
nd
of
ten
unpr
e
dicta
ble
mar
it
im
e
e
nvi
r
onment
.
T
o
pe
r
f
or
m
e
f
f
e
c
ti
ve
ly
in
r
e
a
l
-
wor
ld
c
ondi
ti
on
s
,
thes
e
a
dva
nc
e
d
a
lgo
r
it
hms
mus
t
be
highl
y
a
da
ptable
to
the
dyna
mi
c
a
nd
v
is
ua
ll
y
c
ha
ll
e
nging
na
tur
e
of
oc
e
a
n
s
ur
f
a
c
e
s
.
Va
r
iations
in
li
ghti
ng
,
r
e
f
lec
ti
ons
f
r
om
s
unli
ght,
s
hif
ti
ng
wa
ve
pa
tt
e
r
ns
,
a
nd
a
t
mos
phe
r
ic
dis
tur
ba
nc
e
s
c
a
n
dr
a
s
ti
c
a
ll
y
a
lt
e
r
the
vis
ua
l
c
ha
r
a
c
ter
is
ti
c
s
of
the
s
e
a
a
nd
s
ky,
c
ompl
ica
ti
ng
ob
jec
t
r
e
c
ognit
ion.
T
he
r
e
f
or
e
,
the
models
mus
t
be
r
e
s
il
ient
to
s
uc
h
f
luctua
ti
ons
a
nd
c
a
pa
ble
of
ge
ne
r
a
li
z
ing
a
c
r
os
s
diver
s
e
s
c
e
na
r
ios
not
e
nc
ounter
e
d
dur
ing
t
r
a
ini
ng
.
I
n
a
d
di
ti
on,
d
e
t
e
c
ti
on
s
y
s
te
m
s
m
u
s
t
b
e
pr
of
i
c
ie
n
t
i
n
r
e
c
o
gn
i
z
in
g
a
w
id
e
a
r
r
a
y
o
f
ob
s
t
r
u
c
t
iv
e
o
r
ha
z
a
r
d
ou
s
o
bj
e
c
t
s
,
wh
ic
h
ma
y
dif
f
e
r
gr
e
a
tl
y
in
f
or
m,
s
c
a
l
e
,
t
e
xtu
r
e
,
a
nd
or
i
e
nt
a
t
io
n.
T
hi
s
i
n
c
lu
d
e
s
b
ot
h
f
a
m
il
iar
m
a
r
i
ti
me
h
a
z
a
r
d
s
a
n
d
no
ve
l
o
r
p
r
e
vi
ou
s
ly
un
s
e
e
n
it
e
m
s
a
b
s
e
nt
f
r
o
m
th
e
t
r
a
in
in
g
da
t
a
.
A
c
o
mpr
e
he
n
s
i
ve
a
n
d
di
ve
r
s
e
d
a
t
a
s
e
t
th
a
t
c
a
ptu
r
e
s
e
xt
r
e
me
a
n
d
va
r
i
e
d
s
e
a
c
on
di
ti
on
s
i
s
e
s
s
e
n
ti
a
l
t
o
s
u
pp
or
t
th
i
s
c
a
pa
bi
li
t
y.
A
c
hi
e
v
in
g
r
o
bu
s
t
g
e
n
e
r
a
li
z
a
t
io
n
d
e
m
a
n
d
s
a
dv
a
n
c
e
d
f
e
a
tu
r
e
e
x
tr
a
c
ti
o
n
a
nd
r
e
pr
e
s
e
nt
a
ti
on
le
a
r
ni
ng
w
it
hi
n
C
N
N
f
r
a
me
wo
r
k
s
,
of
te
n
e
nh
a
nc
e
d
thr
ou
g
h
te
c
h
ni
qu
e
s
s
u
c
h
a
s
d
a
t
a
a
ugm
e
nt
a
t
io
n
a
n
d
tr
a
n
s
f
e
r
l
e
a
r
n
in
g.
T
he
o
ve
r
a
r
c
hi
ng
ob
je
c
ti
v
e
i
s
to
e
n
s
ur
e
c
on
s
i
s
te
nt
a
nd
a
c
c
ur
a
te
d
e
t
e
c
ti
on
p
e
r
f
or
m
a
n
c
e
a
c
r
o
s
s
a
b
r
o
a
d
s
p
e
c
tr
um
of
op
e
r
a
ti
on
a
l
e
n
v
ir
onm
e
nt
s
.
T
he
r
e
b
y
c
o
ntr
i
bu
ti
ng
to
s
a
f
e
r
n
a
v
ig
a
t
io
n
a
n
d
m
or
e
e
f
f
e
c
ti
ve
m
a
r
it
i
me
s
ur
ve
il
l
a
nc
e
.
2.
3.
1.
Ar
c
h
it
e
c
t
u
r
e
a
n
alys
is
WaS
R
-
T
an
d
l
igh
t
w
e
igh
t
WaS
R
-
T
T
he
W
a
S
R
[
15
]
–
[
17
]
is
one
o
f
th
e
S
O
T
A
C
NN
ne
t
wor
ks
t
ha
t
pe
r
f
or
m
w
e
ll
f
o
r
objec
t
de
tec
ti
o
n
in
t
he
mar
i
ti
m
e
do
main
.
T
he
W
a
S
R
de
c
ode
r
wa
s
de
s
i
gne
d
to
f
us
e
I
M
U
a
nd
vis
ua
l
i
mage
inf
or
mati
on
t
o
i
nc
r
e
a
s
e
the
a
c
c
ur
a
c
y
o
f
pos
i
ti
ve
o
bjec
t
d
e
tec
t
ions
.
How
e
ve
r
,
W
a
S
R
s
t
il
l
ne
e
ds
to
im
pr
o
ve
it
s
pe
r
f
or
m
a
nc
e
in
c
a
s
e
s
whe
r
e
a
s
ingl
e
i
mage
is
qu
it
e
c
ha
ll
e
ngin
g
to
d
is
ti
n
gui
s
h
th
e
r
e
f
lec
ti
ve
pr
ope
r
ti
e
s
o
f
the
w
a
te
r
s
u
r
f
a
c
e
a
nd
o
bj
e
c
ts
that
of
te
n
oc
c
u
r
on
the
ope
n
s
e
a
.
T
he
modi
f
ie
d
a
r
c
hit
e
c
t
ur
e
of
the
W
a
S
R
ne
two
r
k
,
c
a
l
led
W
a
S
R
-
T
,
is
a
modi
f
ica
ti
on
of
the
e
nc
ode
r
to
a
c
c
omm
oda
te
m
ult
ipl
e
s
e
que
nc
e
im
a
ge
f
r
a
mes
a
s
input
,
whic
h
ha
s
be
e
n
pr
opos
e
d
to
e
xtr
a
c
t
s
pa
ti
o
-
tempor
a
l
textur
e
a
nd
c
o
pe
with
r
e
f
lec
ti
ons
[
18]
.
T
he
W
a
S
R
-
T
ne
twor
k
a
r
c
hit
e
c
tur
e
is
s
hown
in
F
igur
e
5
(
a
)
.
T
he
e
nc
ode
r
ba
c
kbone
of
W
a
S
R
-
T
uti
li
z
e
s
a
R
e
s
Ne
t
-
101
ba
c
kbone
f
e
a
tur
in
g
a
tr
ous
c
onvolut
ions
.
Dur
ing
ne
twor
k
r
e
tr
a
ini
n
g,
a
ll
laye
r
s
a
r
e
e
xpa
nde
d,
a
ll
owing
the
r
e
s
idual
pa
r
ts
of
the
ne
twor
k
to
de
lve
de
e
pe
r
int
o
the
f
e
a
tur
e
s
pa
c
e
of
the
i
nput
im
a
ge
.
I
n
the
c
ontext
of
a
n
e
nc
ode
r
f
o
r
W
a
S
R
-
T,
R
e
s
Ne
t
-
101
c
ompr
is
e
s
f
our
r
e
s
idual
c
onvolu
ti
ona
l
blocks
(
R
e
s
2,
R
e
s
3,
R
e
s
4,
a
nd
R
e
s
5)
.
T
his
a
r
c
hit
e
c
tur
e
of
W
a
S
R
-
T
ha
s
be
e
n
uti
li
z
e
d
to
e
nc
ode
the
va
r
ied
a
p
pe
a
r
a
nc
e
of
ope
n
s
e
a
s
c
e
ne
s
,
including
e
leme
nts
li
ke
boa
ts
,
wa
ter
,
a
nd
s
ky,
a
nd
to
c
las
s
if
y
e
a
c
h
r
e
gion
withi
n
the
ta
r
ge
t
im
a
ge
f
r
a
me
X
∈
R
^
(
3×
H×
W
)
.
T
o
im
pr
ove
the
pr
e
diction
a
c
c
ur
a
c
y,
th
e
r
ole
o
f
R
e
s
N
et
-
101
ha
s
be
e
n
e
xtende
d
to
e
nc
ode
dis
c
r
im
inative
tempor
a
l
inf
or
mation
a
bout
loca
l
f
e
a
tur
e
a
ppe
a
r
a
nc
e
c
ha
nge
of
the
r
e
gion
of
the
ta
r
ge
t
im
a
ge
ove
r
T
p
r
e
c
e
ding
c
ontext
f
r
a
mes
c
a
ppe
d
by
M
e
leme
nts
of
double
-
s
tr
u
c
k
c
a
p
R
,
M
∈
R
^
(
T
×
3×
H×
W
)
a
s
s
hown
in
F
igur
e
5
(
a
)
.
T
he
im
a
ge
input
(
X)
a
nd
c
ontext
f
r
a
me
(
M
)
a
r
e
f
ir
s
t
e
nc
ode
d
with
a
R
e
s
ne
t
-
101
e
nc
ode
r
ne
twor
k
that
pr
oduc
e
s
pe
r
-
f
r
a
me
f
e
a
tur
e
maps
f
r
a
me
X_
F
∈
R
^
(
N×
H×
W
)
a
nd
M
_F
∈
R
^
(
N×
H×
W
)
,
whe
r
e
N
is
the
number
of
c
ha
nne
ls
of
f
e
a
tur
e
maps
.
T
he
tempor
a
l
c
ont
e
xt
modul
e
(
T
C
M
)
e
xtr
a
c
ts
tempor
a
l
inf
or
mation
f
r
om
the
e
mbeddings
of
the
c
ontext
a
nd
tar
ge
t
f
r
a
mes
.
T
o
maintain
the
s
tr
uc
tur
e
a
nd
qua
nti
ty
of
input
c
ha
nne
ls
to
the
de
c
ode
r
,
T
C
M
ini
ti
a
ll
y
de
c
r
e
a
s
e
s
the
dim
e
ns
ionalit
y
of
pe
r
-
f
r
a
me
f
e
a
tur
e
maps
XF
a
nd
M
F
int
o
N/2
-
dim
e
ns
ional
pe
r
-
f
r
a
me
r
e
pr
e
s
e
ntations
.
F
inally,
the
output
f
r
om
T
C
M
is
f
e
d
to
the
f
i
r
s
t
f
us
ion
block
c
a
ll
e
d
the
a
tt
e
nti
on
r
e
f
ineme
nt
modul
e
(
AR
M
1
a
nd
AR
M
2
in
F
igu
r
e
5
(
a
)
)
.
AR
M
1
is
us
e
d
to
a
djus
t
the
we
ight
s
of
input
f
e
a
tur
e
c
ha
nne
ls
ba
s
e
d
on
th
e
ir
c
ontent
withi
n
the
c
ha
nne
ls
.
T
he
indi
vidual
we
ight
s
f
or
e
a
c
h
c
ha
nne
l
a
r
e
de
ter
mi
ne
d
by
taking
the
a
ve
r
a
ge
of
the
input
f
e
a
tur
e
s
a
c
r
os
s
s
pa
ti
a
l
dim
e
ns
ions
.
T
his
pr
oc
e
s
s
yields
a
1×
1
f
e
a
tur
e
ve
c
tor
,
whic
h
then
unde
r
goe
s
a
1×
1
c
onvolut
ion
a
nd
pa
s
s
e
s
thr
ou
gh
a
s
igm
oid
a
c
ti
va
ti
on
f
unc
ti
on.
T
he
s
e
c
ond
f
us
ion
block
is
de
noted
a
s
a
f
e
a
tur
e
f
us
ion
modul
e
(
F
F
M
)
.
I
t
c
ombi
ne
s
f
e
a
tu
r
e
s
f
r
om
va
r
ious
ne
twor
k
br
a
nc
he
s
by
c
onc
a
tena
ti
ng
them,
f
ol
lowe
d
by
a
3×
3
c
onvolut
ion.
T
he
thi
r
d
major
block
is
r
e
f
e
r
r
e
d
to
a
s
a
tr
ous
s
pa
ti
a
l
pyr
a
mi
d
pooli
ng
(
ASP
P
)
a
nd
S
o
f
tM
a
x
.
ASP
P
s
im
ult
a
ne
ous
ly
uti
li
z
e
s
c
onvolut
ions
with
va
r
ying
dil
a
ti
on
r
a
tes
a
nd
c
ombi
ne
s
the
ge
ne
r
a
ted
r
e
pr
e
s
e
ntations
to
e
f
f
e
c
ti
ve
ly
c
a
ptur
e
the
objec
t
a
nd
im
a
ge
c
ontext
a
t
va
r
ious
s
c
a
les
a
s
s
how
n
in
F
igu
r
e
5(
a
)
.
One
of
the
potential
pr
oblems
in
us
ing
W
a
S
R
-
T
t
o
de
tec
t
una
uthor
ize
d
boa
ts
a
ppr
oa
c
hing
T
E
W
S
is
it
s
r
e
quir
e
ment
f
o
r
a
powe
r
f
ul
c
omput
a
ti
ona
l
mac
hine.
T
he
e
nc
ode
r
of
W
a
S
R
-
T
is
pr
im
a
r
il
y
r
e
s
pons
ibl
e
f
or
memor
y
c
ons
umpt
ion
due
to
it
s
uti
li
z
a
ti
on
of
the
R
e
s
Ne
t
-
101
a
s
the
b
a
c
kbone
.
One
s
tr
a
tegy
to
a
d
dr
e
s
s
the
is
s
ue
is
to
r
e
plac
e
the
e
nc
ode
r
with
a
li
gh
twe
ight
ba
c
kbone
that
c
a
n
ope
r
a
te
on
low
-
powe
r
de
vice
s
.
M
obil
e
Ne
ts
is
on
e
of
a
c
las
s
of
li
ghtwe
ight
a
r
c
hit
e
c
tur
e
s
c
ur
r
e
ntl
y
us
e
d
in
va
r
ious
a
ppli
c
a
ti
ons
[
24]
–
[
26]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
T
e
mpor
al
c
ontex
t
of
li
ghtw
e
ight
ne
tw
or
k
mode
l
for
de
tec
ti
ng
boats
appr
oac
hing
…
(
W
ay
an
W
ir
a
Y
og
antar
a
)
3547
M
obil
e
Ne
t
is
pr
e
f
e
r
r
e
d
a
s
a
n
e
nc
ode
r
ba
c
kbone
f
or
i
mage
s
e
gmenta
ti
on
ne
two
r
ks
f
o
r
va
r
ious
a
ppl
ica
ti
ons
[
27]
–
[
29]
.
C
ompar
e
d
with
R
e
s
Ne
t
-
101,
M
obil
e
Ne
t
ha
s
f
e
we
r
pa
r
a
mete
r
s
a
nd
r
e
quir
e
s
les
s
me
mor
y
a
nd
s
tor
a
ge
,
whic
h
c
a
n
be
a
dva
ntage
ous
,
e
s
pe
c
ially
in
de
ploym
e
nt
s
c
e
na
r
ios
.
How
e
ve
r
,
li
mi
ted
r
e
s
e
a
r
c
h
ha
s
be
e
n
r
e
por
ted
on
the
a
ppli
c
a
ti
ons
of
M
obil
e
Ne
t
in
th
e
mar
ine
domain
.
T
he
W
a
S
R
-
T
with
a
n
e
nc
od
e
r
us
ing
M
obil
e
Ne
t
is
r
e
f
e
r
r
e
d
to
a
s
the
l
igh
twe
ight
W
a
S
R
-
T
,
wi
th
it
s
ne
twor
k
a
r
c
hit
e
c
tur
e
il
lus
tr
a
ted
in
F
igur
e
5(
b)
.
M
o
bi
le
N
e
t
s
[
24]
,
[
30]
,
[
31]
f
a
c
t
or
i
z
e
c
o
nv
ol
ut
ion
s
in
to
d
e
e
p
a
n
d
pr
e
c
i
s
e
c
o
nv
ol
ut
io
n
s
,
pr
op
os
in
g
h
a
r
d
-
s
wi
s
h
a
c
ti
va
ti
on
f
u
n
c
ti
on
s
,
s
q
u
e
e
z
e
-
a
n
d
-
e
xc
it
e
b
lo
c
k
s
[
32]
,
a
n
d
f
ut
ur
e
ne
ur
a
l
a
r
c
hi
te
c
t
ur
e
s
e
a
r
c
h
(
NA
S
)
to
f
i
nd
th
e
b
e
s
t
a
r
c
hit
e
c
t
ur
e
f
or
m
ob
il
e
C
P
U
s
.
M
ob
il
e
N
e
t
s
f
a
c
tor
iz
e
s
c
on
vo
lu
ti
o
n
s
in
to
d
e
p
th
wi
s
e
c
o
nv
ol
uti
on
s
a
nd
p
r
o
po
s
e
s
a
h
a
r
d
s
wi
s
h
a
c
t
iva
ti
on
f
u
nc
ti
on
(
h
-
s
w
i
s
h)
.
T
h
e
h
-
s
wi
s
h
n
on
li
ne
a
r
i
ty
is
e
m
plo
y
e
d
to
m
in
im
iz
e
t
he
nu
mb
e
r
of
tr
a
i
ni
ng
p
a
r
a
m
e
ter
s
a
nd
r
e
d
u
c
e
th
e
m
od
e
l
c
o
mp
le
xi
ty
a
n
d
s
i
z
e
.
U
n
li
k
e
t
he
s
ta
nd
a
r
d
u
s
e
of
c
on
vo
lu
ti
on
i
n
R
e
s
Ne
t,
de
pt
hw
i
s
e
s
e
p
a
r
a
b
le
c
o
nv
ol
ut
io
n
s
pl
it
s
t
he
c
o
mp
ut
a
t
io
n
i
nt
o
tw
o
s
te
p
s
:
a
de
pt
hw
i
s
e
c
o
nv
olu
ti
on
a
pp
li
e
s
a
s
i
ng
le
c
on
vo
lu
ti
on
a
l
f
i
lt
e
r
to
e
a
c
h
inp
ut
c
h
a
nn
e
l,
a
n
d
p
oi
nt
wi
s
e
c
on
vo
lut
io
n
i
s
u
s
e
d
to
c
r
e
a
t
e
a
l
in
e
a
r
c
o
mbi
n
a
ti
on
o
f
th
e
ou
tp
ut
of
t
he
d
e
p
th
wi
s
e
c
onv
ol
ut
io
n.
T
he
de
pt
hw
i
s
e
c
o
nv
ol
uti
on
a
l
ke
r
n
e
l
i
s
a
l
e
a
r
n
a
b
l
e
p
a
r
a
me
t
e
r
a
pp
li
e
d
t
o
e
a
c
h
i
np
ut
c
h
a
n
n
e
l
s
e
p
a
r
a
t
e
ly,
i
n
c
r
e
a
s
i
ng
mo
d
e
l
e
f
f
i
c
i
e
nc
y
a
nd
r
e
du
c
i
ng
c
omp
ut
a
t
io
n
c
o
s
t
s
in
th
e
M
o
bi
le
N
e
t
s
n
e
t
w
or
k.
I
t
i
s
a
l
s
o
s
h
a
r
e
d
a
c
r
o
s
s
a
ll
in
pu
t
c
h
a
n
ne
l
s
.
F
ur
ther
mor
e
,
to
s
e
a
r
c
h
f
or
the
be
s
t
ke
r
ne
l
s
ize
in
t
he
de
pthwis
e
c
onvolut
ion,
NA
S
wa
s
e
mpl
oye
d
to
f
ind
the
be
s
t
a
r
c
hit
e
c
tur
e
s
to
f
ulf
il
l
the
low
-
r
e
s
our
c
e
d
ha
r
dwa
r
e
platf
or
ms
in
ter
ms
of
s
ize
,
pe
r
f
or
ma
nc
e
,
a
nd
late
nc
y
[
6]
,
[
30]
,
[
31
]
.
On
the
li
ghtwe
ight
W
a
S
R
-
T
,
the
e
nc
ode
r
be
gins
with
a
s
tem
block
that
pr
oc
e
s
s
e
s
the
input
im
a
ge
to
e
xtr
a
c
t
f
unda
menta
l
f
e
a
tur
e
s
.
T
he
s
ha
pe
of
the
r
e
s
ult
ing
f
e
a
tur
e
map
is
inf
luenc
e
d
by
downs
a
mpl
ing
ope
r
a
ti
ons
,
typi
c
a
ll
y
i
nvo
lvi
ng
s
t
r
ide
-
2
c
onvolut
ions
.
F
or
a
n
input
im
a
ge
of
51
2×
384×
3
(
width×he
ight
×
c
ha
nne
ls
)
,
thi
s
downs
a
mpl
ing
wo
uld
yield
a
f
e
a
tur
e
map
s
ize
of
256×
192×
3
a
t
th
e
output
s
tage
1.
W
e
uti
li
z
e
the
s
kip
c
onne
c
ti
on
f
r
om
the
f
i
r
s
t
a
nd
s
e
c
ond
r
e
s
idual
blocks
of
M
obil
e
Ne
tV3,
s
im
il
a
r
to
the
W
a
S
R
-
T
a
r
c
hit
e
c
tur
e
,
whe
r
e
the
s
kip
c
onne
c
ti
on
is
loca
ted
a
t
s
tage
s
2
a
nd
3.
Unlike
the
or
igi
na
l
W
a
S
R
-
T
,
the
late
nt
f
e
a
tur
e
dim
e
ns
ion
of
the
F
F
M
wa
s
r
e
duc
e
d
f
r
om
1024
to
128,
a
nd
the
F
F
M
o
utput
wa
s
96×
128
to
r
e
duc
e
the
c
omput
a
ti
ona
l
load
in
the
de
c
ode
r
pa
r
ts
.
(
a
)
(
b)
F
igu
r
e
5
.
T
w
o
d
i
f
f
e
r
e
nt
W
a
S
R
-
T
ne
t
wor
k
a
r
c
h
it
e
c
t
ur
e
(
a
)
W
a
S
R
-
T
a
n
d
(
b
)
li
ghtwe
igh
t
W
a
S
R
-
T
ne
tw
or
k
a
r
c
hi
tec
tu
r
e
3.
E
XP
E
RI
M
E
NT
P
ROCE
DU
RE
S
3.
1
.
I
m
age
d
a
t
as
e
t
f
or
t
r
ain
in
g
an
d
e
valu
at
ion
T
he
da
tas
e
ts
us
e
d
f
or
tr
a
ini
ng
the
S
OT
A
C
NN
ne
twor
k
we
r
e
s
e
lec
ted
f
r
om
126
im
a
ge
s
(
including
their
pr
e
c
e
ding
f
r
a
mes
)
c
hos
e
n
f
r
om
M
a
S
T
r
1325
,
M
a
S
T
r
1478,
a
nd
s
uppleme
nted
with
164
im
a
ge
s
(
i
nc
ludi
ng
their
p
r
e
c
e
ding
f
r
a
mes
)
f
r
om
the
e
xtende
d
i
mage
da
tas
e
t
of
th
e
ope
n
s
e
a
in
I
ndone
s
ia's
ter
r
it
or
y.
T
he
a
nnotation
im
a
ge
f
or
t
r
a
ini
ng
the
pr
opos
e
d
S
O
T
A
C
NN
ne
twor
k
wa
s
a
va
il
a
ble
f
r
om
the
da
t
a
s
e
ts
of
M
a
S
T
r
1325
a
nd
M
a
S
T
r
1478
.
T
he
r
e
f
o
r
e
,
the
e
xtende
d
im
a
ge
da
tas
e
ts
f
r
om
the
ope
n
we
bs
it
e
a
nd
the
e
xtende
d
im
a
ge
da
tas
e
t
we
r
e
manua
ll
y
a
nnotate
d
pe
r
pixel
f
or
th
r
e
e
s
e
mantic
c
omponents
:
s
e
a
,
s
ky,
a
nd
boa
t.
T
he
labe
li
ng
f
o
r
the
e
xtende
d
da
tas
e
t
is
c
onduc
ted
us
i
ng
ove
r
164
s
e
lec
ted
im
a
ge
s
f
or
the
tr
a
ini
ng
pr
oc
e
s
s
of
the
li
ghtwe
ight
W
a
S
R
-
T
ne
twor
k.
T
he
a
nnotation
pr
oc
e
s
s
wa
s
c
a
r
r
i
e
d
out
us
ing
the
L
a
be
l
M
e
tool
s
[
33]
.
An
e
xa
mpl
e
of
a
n
a
nnotation
pr
oc
e
s
s
us
ing
L
a
be
l
M
e
is
s
hown
in
F
igur
e
6.
L
a
be
ls
in
gr
ound
-
tr
uth
a
n
notation
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
5
,
Oc
tober
2025
:
35
42
-
3553
3548
mas
ks
c
or
r
e
s
pond
to
the
f
oll
owing
va
lues
,
i
.
e
.
,
boa
t,
s
e
a
,
a
nd
s
ky
a
r
e
given
labe
ls
1
(
F
igu
r
e
6(
a
)
)
,
F
ig
ur
e
6
(
b
)
gives
2
(
va
lue
two)
,
a
nd
F
igur
e
6(
c
)
gives
3
(
va
lue
thr
e
e
)
,
s
uc
c
e
s
s
ively.
(
a
)
(
b)
(
c
)
F
igur
e
6.
E
xa
mpl
e
of
a
n
a
nnotation
p
r
oc
e
s
s
us
ing
L
a
be
lM
e
:
(
a
)
t
he
pr
oc
e
s
s
of
a
manua
l
a
nnotation
o
f
im
a
ge
,
(
b)
manua
ll
y
a
nnotate
d
pe
r
p
ixel
f
o
r
th
r
e
e
s
e
mantic
c
omponents
:
s
e
a
,
s
ky,
a
nd
boa
t
,
a
nd
(
c
)
i
mage
mas
king
that
r
e
pr
e
s
e
nts
the
labe
l
c
la
s
s
3.
2
.
T
r
ain
in
g
s
e
t
u
p
li
g
h
t
we
igh
t
WaS
R
-
T
T
he
p
r
opos
e
d
li
ghtwe
ight
W
a
S
R
-
T
wa
s
tr
a
ined
us
ing
290
im
a
ge
da
tas
e
ts
,
including
their
pr
e
c
e
ding
f
r
a
mes
with
T
=
3,
a
s
we
ll
a
s
the
c
or
r
e
s
ponding
a
n
notation
im
a
ge
s
.
T
he
s
ize
of
input
im
a
ge
s
f
o
r
t
r
a
i
ning
wa
s
512
×
384
×
3.
T
he
da
tas
e
t
wa
s
divi
de
d
int
o
mi
ni
-
ba
tche
s
of
10
im
a
ge
s
to
im
pr
ove
the
e
f
f
icie
nc
y
of
the
tr
a
ini
ng
pr
oc
e
s
s
.
T
he
a
da
pti
ve
mo
ment
e
s
ti
mation
(
Ada
m
)
opti
mi
z
e
r
uti
li
z
e
s
a
s
qua
r
e
g
r
a
dient
to
a
da
pt
lea
r
ni
ng
r
a
tes
.
Hype
r
pa
r
a
mete
r
tr
a
ini
ng
wa
s
c
onduc
ted
thr
ough
va
r
ious
c
onf
igur
a
ti
ons
to
a
c
hieve
the
be
s
t
pos
s
ibl
e
r
e
s
ult
s
,
whic
h
a
r
e
s
umm
a
r
ize
d
in
T
a
ble
1
.
T
a
ble
1.
T
he
hype
r
pa
r
a
mete
r
t
r
a
ini
ng
o
f
the
Ada
m
-
opti
mi
z
e
r
P
a
r
a
me
te
r
V
a
lu
e
L
e
a
r
ni
ng r
a
te
10
−
0
.
6
L
e
a
r
ni
ng r
a
te
de
c
a
y
0.9
W
e
ig
ht
s
de
c
a
y
10
−
0
.
6
E
poc
h
500
B
a
tc
h s
iz
e
6
M
ome
nt
um va
lu
e
0.9
P
a
ti
e
nc
e
50
T
he
tr
a
ini
ng
of
the
li
gh
twe
ight
W
a
S
R
-
T
wa
s
pe
r
f
or
med
us
ing
a
f
a
c
il
it
y
in
the
labor
a
tor
y
of
high
-
pe
r
f
or
manc
e
c
omput
ing
a
t
the
I
ndone
s
ian
Na
ti
ona
l
R
e
s
e
a
r
c
h
a
nd
I
nnova
ti
on
Age
nc
y,
e
quipped
with
a
n
NV
I
DI
A
DG
X1.
T
he
opti
mi
z
e
d
ne
twor
k
pa
r
a
m
e
ter
s
of
li
gh
twe
ight
W
a
S
R
-
T
,
ba
s
e
d
on
the
c
r
it
e
r
ion
of
mi
nim
um
tr
a
in/
los
s
va
lue,
a
r
e
0
.
0009.
A
va
l/
a
c
c
ur
a
c
y
of
0
.
995
c
a
n
be
a
c
hieve
d
us
ing
the
hype
r
p
a
r
a
mete
r
tr
a
ini
ng
of
Ada
m
-
o
pti
mi
z
e
r
with
290
e
poc
hs
.
Af
te
r
the
tr
a
ini
ng
pr
oc
e
s
s
,
the
l
ight
we
ight
W
a
S
R
-
T
wa
s
tes
ted
to
de
tec
t
una
utho
r
ize
d
boa
ts
with
va
r
ious
s
ha
pe
s
a
nd
s
ize
s
that
may
ha
ve
be
e
n
unc
ove
r
e
d
in
tr
a
ini
ng
da
tas
e
ts
.
T
he
tes
ti
ng
da
tas
e
ts
c
ons
is
t
of
140
i
mage
s
s
e
lec
t
e
d
f
r
om
a
n
ope
n
we
bs
it
e
a
nd
da
tas
e
ts
r
e
c
or
de
d
f
r
om
the
B
a
r
una
J
a
ya
r
e
s
e
a
r
c
h
ve
s
s
e
l
in
I
ndone
s
ia's
ter
r
it
or
y,
in
a
ddit
ion
to
the
da
tas
e
ts
us
e
d
f
or
da
ta
t
r
a
ini
ng.
As
a
r
e
f
e
r
e
nc
e
f
or
e
va
luating
de
tec
ti
on
a
c
c
ur
a
c
y,
the
p
r
e
-
tr
a
ini
ng
or
igi
na
l
W
a
S
R
-
T
models
a
r
e
publi
c
ly
a
va
il
a
ble
on
GitHub
[
34]
.
T
he
pr
e
-
tr
a
ined
or
igi
na
l
W
a
S
R
-
T
wa
s
a
ls
o
a
s
s
e
s
s
e
d
to
de
tec
t
boa
ts
us
ing
s
im
il
a
r
da
tas
e
ts
to
thos
e
us
e
d
f
or
l
ight
we
ight
W
a
S
R
-
T.
As
mentioned
in
the
a
nnotation
p
r
oc
e
s
s
f
or
the
im
a
ge
da
tas
e
t,
the
ne
twor
k's
tar
ge
t
output
c
ons
is
ts
of
thr
e
e
c
las
s
labe
ls
:
boa
t,
s
e
a
,
a
nd
s
ky.
T
he
objec
ti
v
e
of
the
pr
opos
e
d
ne
twor
k
f
or
de
tec
ti
ng
a
boa
t
a
pp
r
oa
c
hing
the
T
E
W
S
is
to
identif
y
the
a
r
e
a
of
pixels
labe
led
by
the
output
ne
twor
k
that
c
or
r
e
s
ponds
to
a
c
las
s
labe
l
o
f
the
boa
t's
g
r
ound
tr
uth
,
with
the
ba
c
kgr
ound
im
a
g
e
s
e
r
ving
a
s
the
labe
l
c
las
s
:
s
e
a
or
the
int
e
r
f
a
c
e
be
t
we
e
n
s
e
a
a
nd
s
ky.
T
he
pe
r
f
or
manc
e
of
l
ight
we
ight
W
a
S
R
-
T
in
c
ompar
is
on
with
the
o
r
igi
na
l
W
a
S
R
-
T
wa
s
e
va
luate
d
us
ing
vis
ua
l
qua
li
ty
a
s
s
e
s
s
ment
s
[
35]
,
[
36]
with
c
r
i
ter
ia
a
s
f
oll
ows
:
−
I
f
the
output
of
the
ne
tw
or
k
p
r
oduc
e
s
a
c
las
s
labe
l
of
a
boa
t
that
pe
r
f
e
c
tl
y
ove
r
laps
with
the
loca
ti
on
a
nd
a
n
a
r
e
a
of
the
pixels
labe
l
of
a
boa
t
is
a
li
gne
d
to
t
he
boa
t
gr
ound
tr
uth
,
it
is
a
s
ubjec
ti
ve
a
s
s
e
s
s
ment
a
s
a
tr
ue
pos
it
ive
(
T
P
)
.
−
I
f
the
output
of
the
ne
twor
k
p
r
oduc
e
s
a
c
la
s
s
labe
l
of
a
boa
t
wi
th
ins
uf
f
icie
nt
ove
r
lap
with
the
loca
ti
on,
a
nd
the
a
r
e
a
of
the
pixels
labe
led
a
s
a
boa
t
is
s
l
ight
ly
s
pr
e
a
d
r
e
lative
to
the
boa
t
g
r
ound
t
r
uth,
it
is
a
s
ubjec
ti
ve
a
s
s
e
s
s
m
e
nt
as
a
F
P
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
T
e
mpor
al
c
ontex
t
of
li
ghtw
e
ight
ne
tw
or
k
mode
l
for
de
tec
ti
ng
boats
appr
oac
hing
…
(
W
ay
an
W
ir
a
Y
og
antar
a
)
3549
−
I
f
the
output
of
the
ne
twor
k
pr
oduc
e
s
a
c
las
s
labe
l
of
a
boa
t
that
li
e
s
outs
ide
the
loca
ti
on
a
nd
a
n
a
r
e
a
of
the
pixels
labe
l
of
a
boa
t
is
s
pr
e
a
d
r
e
lative
to
the
b
oa
t
gr
ound
t
r
uth,
it
is
a
s
ubjec
ti
ve
a
s
s
e
s
s
m
e
nt
a
s
a
f
a
ls
e
ne
ga
ti
ve
(
F
N)
.
Ove
r
a
ll
metr
ics
to
mea
s
ur
e
model
ne
twor
k
pe
r
f
o
r
manc
e
be
twe
e
n
l
ight
we
ight
W
a
S
R
-
T
a
nd
or
igi
na
l
W
a
S
R
-
T
we
r
e
e
va
luate
d
us
ing
=
+
a
nd
=
+
ba
s
e
d
on
ove
r
a
ll
tes
ti
ng
of
i
mage
da
tas
e
ts
.
4.
RE
S
UL
T
S
AN
D
DI
S
CU
S
S
I
ON
U
n
li
ke
p
r
e
v
i
ous
a
p
p
r
o
a
c
h
e
s
th
a
t
f
oc
us
on
e
n
ha
nc
ing
e
xis
t
in
g
ob
jec
t
de
tec
t
io
n
met
ho
ds
by
i
n
tr
od
uc
in
g
n
e
w
a
lg
o
r
i
th
ms
w
hi
le
us
i
ng
the
s
a
me
da
tas
e
t
,
t
his
s
tu
d
y
ta
ke
s
a
d
i
f
f
e
r
e
nt
di
r
e
c
t
io
n
.
W
e
e
va
lua
te
th
e
pe
r
f
o
r
ma
nc
e
o
f
a
p
r
e
-
e
s
ta
bl
is
h
e
d
ob
je
c
t
de
te
c
t
io
n
mo
de
l
by
a
p
ply
i
ng
i
t
to
a
n
e
n
t
ir
e
l
y
ne
w
d
a
tas
e
t
d
is
t
i
nc
t
f
r
o
m
the
o
ne
us
e
d
dur
i
ng
it
s
o
r
ig
in
a
l
d
e
ve
lo
p
men
t
.
T
he
o
nl
y
mo
d
if
ica
t
io
n
mad
e
to
t
he
m
od
e
l
is
the
r
e
pla
c
e
me
nt
o
f
i
ts
b
a
c
kb
one
e
nc
ode
r
:
R
e
s
N
e
t
-
10
1
is
s
u
bs
t
it
u
ted
w
i
th
M
o
bi
leN
e
t
V3
.
T
h
is
c
h
a
n
ge
is
i
nt
e
n
de
d
to
a
s
s
e
s
s
th
e
im
pa
c
t
o
n
d
e
t
e
c
ti
on
s
pe
e
d
,
pa
r
ti
c
u
la
r
l
y
in
th
e
c
on
te
xt
o
f
d
e
p
lo
y
men
t
on
r
e
s
o
ur
c
e
-
c
ons
t
r
a
ine
d
de
v
ic
e
s
.
P
r
i
or
r
e
s
e
a
r
c
h
ha
s
d
e
m
ons
t
r
a
te
d
th
a
t
us
i
ng
a
l
ig
ht
e
r
ba
c
kb
on
e
e
nc
od
e
r
c
a
n
s
ig
ni
f
ica
nt
l
y
a
c
c
e
le
r
a
te
the
d
e
t
e
c
ti
on
p
r
oc
e
s
s
[
6
]
,
a
l
th
ou
gh
th
is
o
f
te
n
c
o
mes
a
t
th
e
c
o
s
t
of
r
e
d
uc
e
d
a
c
c
u
r
a
c
y
wh
e
n
ope
r
a
t
in
g
o
n
l
ow
-
po
we
r
h
a
r
dw
a
r
e
.
I
n
li
ne
wi
th
the
methodology
p
r
e
s
e
nted
in
[
6
]
,
we
c
onduc
ted
a
c
ompar
a
ti
ve
a
na
lys
is
us
ing
M
obil
e
Ne
t
V3
a
mor
e
li
ghtwe
ight
e
nc
ode
r
ba
c
kbone
than
R
e
s
Ne
t
-
101
a
nd
e
va
luate
d
both
a
r
c
hit
e
c
tur
e
s
on
a
ne
wly
int
r
oduc
e
d
ope
n
oc
e
a
n
da
tas
e
t
that
ha
d
not
pr
e
vio
us
ly
be
e
n
a
va
il
a
ble.
B
oth
ne
twor
ks
'
pe
r
f
o
r
manc
e
to
de
tec
t
boa
ts
wa
s
e
va
luat
e
d
whe
n
tes
ti
ng
140
im
a
ge
da
tas
e
ts
r
un
us
ing
laptop
c
omput
e
r
s
with
a
n
I
ntel
C
or
e
i7
-
10510U
(
qua
d
-
c
or
e
HT
1
.
8
GH
z
,
tur
bo
4.
9
GH
z
)
a
nd
16
GB
R
AM
.
W
hil
e
r
unning
the
pr
og
r
a
m,
we
e
va
luate
s
ome
c
omput
a
ti
ona
l
load
pa
r
a
mete
r
s
.
i.
e
.
,
pe
r
c
e
ntage
of
C
P
U
r
e
s
our
c
e
s
us
e
d
by
a
p
r
oc
e
s
s
(
%
C
P
U)
,
pe
r
c
e
ntage
of
phys
ica
l
memor
y
us
e
d
by
a
p
r
oc
e
s
s
(
%
memor
y)
,
tot
a
l
p
r
oc
e
s
s
ing
ti
me
a
nd
r
a
te
s
e
gmenta
ti
ons
pe
r
it
e
r
a
ti
on
(
s
/i
t
)
in
the
c
ontext
pr
oduc
e
s
a
c
las
s
labe
l
of
boa
t
to
tes
ti
ng
a
ll
im
a
ge
da
tas
e
ts
.
T
he
s
umm
a
r
y
of
c
omput
a
ti
ona
l
load
pa
r
a
mete
r
s
is
s
hown
in
T
a
ble
2
.
T
a
ble
2.
S
umm
a
r
y
of
c
omput
a
ti
ona
l
load
pa
r
a
mete
r
s
be
twe
e
n
li
ghtwe
ight
(
L
)
-
W
a
S
R
-
T
a
nd
W
a
S
R
-
T
M
ode
l
T
e
s
ti
ng i
ma
ge
s
C
P
U
(
%
)
M
e
mor
y (
%
)
T
ot
a
l
pr
oc
e
s
s
in
g t
im
e
R
a
te
(
s
/i
t)
W
a
S
R
-
T
140
190
13.2
1:
13:
56
20.07
L
-
W
a
S
R
-
T
140
160
4.3
0:
02:
45
1.33
F
ur
ther
mor
e
,
the
e
xa
mpl
e
qua
nti
tative
r
e
s
ult
s
of
li
ghtwe
ight
W
a
S
R
-
T
a
nd
or
igi
na
l
W
a
S
R
-
T
f
r
om
tr
a
ini
ng
im
a
ge
da
tas
e
ts
of
the
ope
n
s
e
a
in
I
ndone
s
i
a
's
ter
r
it
or
y
a
nd
a
n
ope
n
we
bs
it
e
a
r
e
s
e
que
nti
a
ll
y
s
hown
in
F
igur
e
7
.
T
he
f
ou
r
tar
ge
t
f
r
a
me
im
a
ge
s
f
r
om
e
a
c
h
type
o
f
da
tas
e
t
we
r
e
e
va
luate
d
us
ing
v
is
ua
l
qua
li
ty
a
s
s
e
s
s
ment
to
de
ter
mi
ne
the
T
P
,
F
P
,
a
nd
F
N
r
a
t
e
s
.
T
he
s
umm
a
r
y
o
f
the
vis
ua
l
qua
li
ty
a
s
s
e
s
s
me
nt
of
the
or
igi
na
l
W
a
S
R
-
T
a
nd
li
ghtwe
ight
W
a
S
R
-
T
,
whic
h
de
tec
ts
the
c
las
s
labe
l
of
a
boa
t
f
r
om
e
ight
tes
ti
ng
im
a
ge
da
tas
e
ts
a
s
s
hown
in
T
a
ble
3.
T
he
vis
ua
l
qua
li
ty
a
s
s
e
s
s
ment
of
both
ne
twor
ks
to
de
tec
t
the
c
las
s
l
a
be
l
of
a
boa
t
f
r
om
e
ight
tes
ti
ng
im
a
ge
da
tas
e
ts
a
s
s
hown
i
n
F
igur
e
7
wa
s
tabula
ted
in
T
a
ble
3
,
whic
h
de
s
c
r
ibes
the
output
of
both
ne
twor
ks
ba
s
e
d
on
c
r
it
e
r
ia
T
P
,
F
P
,
a
nd
F
N
.
T
he
s
umm
a
r
y
of
qua
li
tative
r
e
s
ult
s
,
i
nc
ludi
ng
li
ghtwe
ight
W
a
S
R
-
T
a
nd
W
a
S
R
-
T
f
or
a
ll
t
r
a
ini
ng
i
mage
da
tas
e
ts
,
is
s
hown
in
T
a
ble
4
.
F
igur
e
7.
E
xa
mpl
e
of
q
ua
li
tative
r
e
s
ult
s
of
li
ghtwe
i
ght
W
a
S
R
-
T
a
nd
or
igi
na
l
W
a
S
R
-
T
f
r
om
a
tes
ti
ng
im
a
ge
da
tas
e
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
5
,
Oc
tober
2025
:
35
42
-
3553
3550
B
a
s
e
d
on
the
qua
li
tative
r
e
s
ult
s
in
T
a
bles
3
a
nd
4
,
if
the
vis
ua
l
qua
li
ty
a
s
s
e
s
s
ment
is
de
ter
mi
ne
d
a
s
T
P
,
li
ghtwe
ight
W
a
S
R
-
T
p
r
oduc
e
s
a
c
las
s
labe
l
f
or
a
boa
t
,
s
im
il
a
r
to
the
or
ig
inal
W
a
S
R
-
T
.
T
he
a
r
e
a
of
the
pixel
labe
l
of
the
boa
t
is
c
ompl
e
tely
de
tec
ted.
I
f
t
he
ne
twor
k
output
r
e
s
ult
s
in
a
vis
ua
l
qua
li
ty
a
s
s
e
s
s
ment
of
FP
,
li
ghtwe
ight
W
a
S
R
-
T
pr
oduc
e
s
a
n
a
r
e
a
of
pixel
labe
led
a
s
a
boa
t
that
is
s
li
ghtl
y
s
pr
e
a
d
or
s
maller
r
e
lative
to
the
gr
ound
t
r
uth
o
f
the
boa
t
labe
l
.
I
n
the
vis
ua
l
qua
li
ty
a
s
s
e
s
s
ment
a
s
F
P
,
the
number
of
f
a
ls
e
de
tec
ti
ons
pr
oduc
e
d
by
the
W
a
S
R
-
T
is
s
maller
tha
n
that
of
t
he
li
ghtwe
ight
W
a
S
R
-
T
.
T
he
mos
t
c
omm
on
s
our
c
e
of
f
a
ls
e
de
tec
ti
ons
is
due
to
the
r
e
f
lec
ti
on
of
the
wa
ter
on
t
he
s
e
a
or
the
int
e
r
f
a
c
e
be
twe
e
n
the
s
e
a
a
nd
the
s
ky.
I
n
thi
s
c
a
s
e
,
the
a
r
e
a
of
the
pixel
labe
l
o
f
the
boa
t
is
s
ti
ll
de
tec
ted.
T
he
r
e
f
or
e
,
i
f
the
c
las
s
labe
l
of
the
boa
t
c
a
n
be
de
tec
ted,
thi
s
inf
or
mation
c
a
n
s
ti
ll
be
us
e
d
to
ge
ne
r
a
te
a
n
a
lar
m.
He
nc
e
,
the
qua
nti
tative
r
e
s
ult
s
of
li
g
htwe
ight
W
a
S
R
-
T
,
a
long
with
s
ubjec
ti
ve
a
s
s
e
s
s
ment
in
ter
ms
of
T
P
a
nd
F
P
,
a
r
e
us
e
f
ul
f
or
int
e
ll
igent
c
ompu
ter
vis
i
on
in
T
E
W
S
.
B
a
s
e
d
on
th
is
c
on
di
ti
on
,
we
a
ls
o
p
r
o
pos
e
d
a
n
a
d
d
it
i
on
a
l
e
va
lu
a
t
io
n
c
r
i
te
r
io
n
t
o
m
e
a
s
ur
e
w
he
t
he
r
the
n
e
t
wo
r
k
c
ou
ld
ge
ne
r
a
te
a
n
a
l
a
r
m
du
e
to
un
a
u
th
o
r
i
z
e
d
boa
ts
a
p
pr
oa
c
h
in
g
T
E
W
S
.
T
he
s
e
ns
it
iv
i
ty
o
f
g
e
ne
r
a
t
in
g
a
T
E
W
S
a
l
a
r
m
is
f
o
r
m
ul
a
t
e
d
a
s
t
he
r
a
t
i
o
be
tw
e
e
n
th
e
s
um
o
f
T
P
a
nd
F
P
a
n
d
the
t
o
ta
l
tr
a
i
ni
ng
i
ma
ge
d
a
ta
s
e
t
.
T
h
e
s
e
ns
i
ti
vi
ty
of
ge
n
e
r
a
ti
ng
T
E
W
S
a
l
a
r
ms
f
o
r
W
a
S
R
-
T
a
nd
li
gh
tw
e
i
gh
t
W
a
S
R
-
T
pe
r
f
o
r
ma
nc
e
r
e
s
ul
ts
a
r
e
9
5
.
71
%
a
nd
90
.
00
%
,
r
e
s
pe
c
t
iv
e
l
y
.
A
lt
ho
ug
h
th
e
s
e
ns
it
iv
it
y
of
ge
ne
r
a
ti
ng
t
he
T
E
W
S
a
la
r
m
o
f
W
a
S
R
-
T
is
s
li
gh
tl
y
b
e
t
te
r
tha
n
tha
t
o
f
li
gh
tw
e
i
gh
t
W
a
S
R
-
T
,
the
c
om
pu
ta
ti
on
a
l
lo
a
d
o
f
l
ig
ht
we
ig
ht
W
a
S
R
-
T
is
s
ig
ni
f
ic
a
n
tl
y
l
o
we
r
t
ha
n
tha
t
of
W
a
S
R
T
.
B
a
s
e
d
o
n
t
he
tes
t
in
g
r
e
s
u
l
ts
tab
ul
a
t
e
d
in
T
a
b
le
2
,
li
gh
tw
e
i
gh
t
W
a
S
R
-
T
r
e
q
ui
r
e
d
les
s
m
e
mo
r
y
,
a
t
32
.
57
%
,
a
n
d
t
he
t
o
ta
l
p
r
o
c
e
s
s
in
g
t
i
me
wa
s
r
e
d
uc
e
d
t
o
0
.
07
61
%
c
o
mpa
r
e
d
to
th
e
o
r
i
g
ina
l
W
a
S
R
-
T
.
T
a
ble
3
.
T
he
s
umm
a
r
y
o
f
vis
ua
l
qua
l
it
y
a
s
s
e
s
s
ment
o
f
or
igi
na
l
W
a
S
R
-
T
a
n
d
li
g
htwe
ight
W
a
S
R
-
T
T
a
r
ge
t
f
r
a
me
A
s
s
e
s
s
me
nt
r
e
s
ul
ts
W
a
S
R
T
A
s
s
e
s
s
me
nt
r
e
s
ul
ts
L
-
W
a
S
R
T
R
e
ma
r
ks
A
TP
FN
L
ig
ht
w
e
ig
ht
W
a
S
R
-
T
pr
oduc
e
s
th
e
f
a
ls
e
c
la
s
s
la
be
l
of
a
boa
t
th
a
t
is
a
r
e
f
le
c
ti
on
of
l
ig
ht
on t
he
s
e
a
. (
ma
r
ke
d w
it
h t
he
r
e
d c
ir
c
le
)
.
B
TP
FP
L
ig
ht
w
e
ig
ht
W
a
S
R
-
T
de
te
c
te
d
th
e
la
be
l
a
r
e
a
of
th
e
boa
t
(
ma
r
ke
d
w
it
h
th
e
dot
a
r
r
ow
l
in
e
)
. I
n
t
hi
s
c
a
s
e
,
l
ig
ht
w
e
ig
ht
W
a
S
R
-
T
de
te
c
te
d t
he
c
la
s
s
l
a
be
l
of
t
he
s
ky
w
it
h t
he
s
e
a
on t
he
i
nt
e
r
f
a
c
e
be
tw
e
e
n s
e
a
a
nd
s
ky (
ma
r
ke
d w
it
h t
he
r
e
d c
ir
c
le
)
.
C
TP
FP
L
ig
ht
w
e
ig
ht
W
a
S
R
-
T
de
te
c
te
d
th
e
la
be
l
a
r
e
a
of
th
e
boa
t
(
ma
r
ke
d
w
it
h
th
e
dot
a
r
r
ow
li
ne
)
.
H
ow
e
ve
r
,
th
e
pi
xe
l
la
be
l
of
a
boa
t
is
s
li
ght
ly
s
pr
e
a
d
r
e
la
ti
ve
to
th
e
gr
ound
tr
ue
la
be
l
of
th
e
boa
t.
I
n
th
i
s
c
a
s
e
,
a
r
e
f
le
c
ti
on
of
li
g
ht
on
th
e
s
e
a
is
de
te
c
te
d a
s
t
h
e
p
ix
e
l
la
be
l
of
a
boa
t
(
ma
r
ke
d w
it
h t
he
r
e
d c
ir
c
le
)
.
D
TP
FP
L
ig
ht
w
e
ig
ht
W
a
S
R
-
T
de
te
c
te
d
th
e
la
be
l
a
r
e
a
of
th
e
boa
t
(
ma
r
ke
d
w
it
h
th
e
dot
a
r
r
ow
li
ne
)
.
H
ow
e
ve
r
,
th
e
pi
xe
l
la
be
l
of
a
boa
t
is
s
li
ght
ly
s
pr
e
a
d
r
e
la
ti
ve
to
th
e
gr
ound
tr
ue
la
be
l
of
th
e
boa
t.
I
n
th
i
s
c
a
s
e
,
a
r
e
f
le
c
ti
on
of
li
g
ht
on
th
e
s
e
a
is
de
te
c
te
d a
s
t
h
e
pi
xe
l
la
be
l
of
a
boa
t
(
ma
r
ke
d w
it
h t
he
r
e
d c
ir
c
le
)
.
T
a
ble
4.
Qua
li
tative
r
e
s
ult
s
l
ight
we
ight
W
a
S
R
-
T
a
nd
W
a
S
R
-
T
f
or
a
ll
tr
a
ini
ng
im
a
ge
da
tas
e
ts
M
ode
l
TP
FP
FN
P
r
e
c
is
io
n (
%
)
R
e
c
a
ll
(
%
)
W
a
S
R
-
T
94
40
15
70.15
86.24
L
ig
ht
w
e
ig
ht
W
a
S
R
-
T
77
49
14
61.11
86.14
5.
CONC
L
USI
ON
T
he
de
ve
lopm
e
nt
a
nd
im
pleme
ntation
of
the
p
r
opos
e
d
li
ghtwe
ight
W
a
S
R
-
T
ne
twor
ks
to
de
tec
t
una
uthor
ize
d
boa
ts
a
ppr
oa
c
hing
T
E
W
S
a
s
a
n
int
e
g
r
a
l
pa
r
t
of
a
n
int
e
ll
igent
c
omput
e
r
vis
ion
s
ys
tem
in
a
n
ope
n
s
e
a
domain
ha
ve
be
e
n
dis
c
u
s
s
e
d.
B
a
s
e
d
on
the
qu
a
nti
tative
r
e
s
ult
s
a
nd
e
va
luation
of
the
c
omput
a
ti
ona
l
load,
l
ight
we
ight
W
a
S
R
T
,
de
s
igned
a
s
the
c
e
ntr
a
l
pa
r
t
o
f
a
n
int
e
ll
igent
c
ompu
ter
vis
ion
s
ys
tem
in
T
E
W
S
,
s
howe
d
pr
omi
s
e
f
or
f
ur
the
r
im
pleme
ntation
in
c
omput
a
ti
on
a
l
de
vice
s
with
a
s
mall
a
r
c
hit
e
c
tur
a
l
f
ootpr
in
t.
F
utu
r
e
wor
k
will
f
oc
us
on
r
e
a
l
-
wor
ld
tes
ti
ng
of
the
l
i
ghtwe
ight
W
a
S
R
-
T
ne
twor
k
on
buoy
platf
or
ms
in
diver
s
e
wa
ter
e
nvir
onments
.
Additi
ona
l
e
xpe
r
im
e
nts
unde
r
e
xtr
e
me
we
a
ther
c
ondit
ions
s
hould
be
c
onduc
ted
to
e
n
s
ur
e
the
ne
twor
k's
r
obus
tnes
s
.
F
ur
the
r
mor
e
,
to
im
p
r
ove
li
ghtwe
ight
W
a
S
R
-
T
pe
r
f
or
manc
e
,
we
e
nha
nc
e
da
t
a
s
e
t
a
ugmenta
ti
on,
pr
e
pr
oc
e
s
s
ing
tec
hniques
,
a
nd
the
im
pleme
ntation
of
li
ghtwe
ight
W
a
S
R
-
T
,
whic
h
c
ur
r
e
ntl
y
uti
li
z
e
s
a
lar
ge
M
obil
e
Ne
tV3
with
pa
r
a
ll
e
l
pr
oc
e
s
s
ing
on
a
s
mall
GPU
a
r
c
hit
e
c
tur
e
,
s
uc
h
a
s
the
NV
I
DI
A
J
e
ts
on
Na
no.
AC
KNOWL
E
DGM
E
N
T
S
T
he
a
uthor
s
wou
ld
li
ke
to
e
xpr
e
s
s
their
s
ince
r
e
gr
a
ti
tude
to
M
r
.
Andi
Kur
n
ianto
f
or
his
invalua
ble
a
s
s
is
t
a
nc
e
in
da
ta
pr
e
p
r
oc
e
s
s
ing
a
nd
a
nnotation
du
r
ing
the
e
a
r
ly
s
tage
s
of
th
is
s
tudy.
W
e
a
ls
o
thank
M
r
.
Ar
ie
f
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
T
e
mpor
al
c
ontex
t
of
li
ghtw
e
ight
ne
tw
or
k
mode
l
for
de
tec
ti
ng
boats
appr
oac
hing
…
(
W
ay
an
W
ir
a
Y
og
antar
a
)
3551
R
uf
iyanto
f
or
his
s
uppor
t
in
pr
oo
f
r
e
a
ding
the
manus
c
r
ipt
a
nd
pr
ovidi
ng
he
lpf
ul
s
ugge
s
ti
ons
to
im
pr
o
ve
c
lar
it
y
a
nd
c
ohe
r
e
nc
e
.
S
pe
c
ial
thanks
to
M
r
.
De
dy
I
r
a
wa
n
f
or
his
ins
ight
f
ul
d
is
c
us
s
ions
on
model
a
r
c
hit
e
c
tur
e
s
e
lec
ti
on,
whic
h
gr
e
a
tl
y
in
f
luenc
e
d
the
di
r
e
c
ti
on
of
our
e
xpe
r
i
ments
.
F
UN
DI
NG
I
NF
ORM
AT
I
ON
Author
s
s
tate
no
f
unding
invol
ve
d.
AU
T
HO
R
CONT
RI
B
U
T
I
ONS
S
T
AT
E
M
E
N
T
T
his
jour
na
l
us
e
s
the
C
ontr
ibut
o
r
R
oles
T
a
xo
nomy
(
C
R
e
diT
)
to
r
e
c
ognize
indi
vidual
a
uthor
c
ontr
ibut
ions
,
r
e
duc
e
a
utho
r
s
hip
dis
putes
,
a
nd
f
a
c
il
it
a
te
c
oll
a
bor
a
ti
on.
Nam
e
of
Au
t
h
or
C
M
So
V
a
Fo
I
R
D
O
E
Vi
Su
P
Fu
W
a
ya
n
W
i
r
a
Y
og
a
n
ta
r
a
✓
✓
✓
✓
✓
✓
✓
✓
S
upr
ij
a
nto
✓
✓
✓
✓
✓
✓
Ana
k
Agung
Ngur
a
h
Ana
nda
Kus
uma
✓
✓
✓
✓
✓
Yuki
I
s
ti
a
nto
✓
✓
✓
✓
✓
✓
C
:
C
onc
e
pt
ua
li
z
a
ti
on
M
:
M
e
th
odol
ogy
So
:
So
f
twa
r
e
Va
:
Va
li
da
ti
on
Fo
:
Fo
r
ma
l
a
na
ly
s
is
I
:
I
nve
s
ti
ga
ti
on
R
:
R
e
s
our
c
e
s
D
:
D
a
ta
C
ur
a
ti
on
O
:
W
r
it
in
g
-
O
r
ig
in
a
l
D
r
a
f
t
E
:
W
r
it
in
g
-
R
e
vi
e
w
&
E
di
ti
ng
Vi
:
Vi
s
ua
li
z
a
ti
on
Su
:
Su
pe
r
vi
s
io
n
P
:
P
r
oj
e
c
t
a
dmi
ni
s
tr
a
ti
on
Fu
:
Fu
ndi
ng a
c
qui
s
it
io
n
CONF
L
I
CT
OF
I
NT
E
RE
S
T
S
T
AT
E
M
E
N
T
T
he
a
uthor
s
s
tate
no
c
on
f
li
c
t
o
f
int
e
r
e
s
t.
I
NF
ORM
E
D
CONSE
NT
W
e
ha
ve
obtaine
d
inf
or
med
c
ons
e
nt
f
r
om
a
ll
ind
ivi
dua
ls
include
d
in
thi
s
s
tudy.
E
T
HI
CA
L
AP
P
ROVA
L
T
his
pa
pe
r
doe
s
not
invol
ve
pe
ople
or
a
nim
a
ls
;
no
inves
ti
ga
ti
on
h
a
s
invol
ve
d
human
s
ubjec
ts
.
T
he
r
e
f
or
e
,
the
a
utho
r
s
did
not
s
e
e
k
a
ppr
ova
l
f
r
om
a
ny
ins
ti
tut
ional
r
e
view
boa
r
d.
DA
T
A
AV
AI
L
A
B
I
L
I
T
Y
T
he
da
ta
that
s
uppor
t
the
f
indi
ngs
o
f
thi
s
s
tudy
a
r
e
a
va
il
a
ble
f
r
om
the
c
or
r
e
s
ponding
a
uthor
,
[
S
]
,
upon
r
e
a
s
ona
ble
r
e
que
s
t.
RE
F
E
RE
NC
E
S
[
1]
L
.
Z
ha
o,
F
.
Y
u,
J
.
H
ou,
P
.
W
a
ng,
a
nd
T
.
F
a
n,
“
T
he
r
ol
e
of
ts
una
mi
buoy
pl
a
ye
d
in
ts
una
mi
w
a
r
ni
ng
a
nd
it
s
a
ppl
ic
a
ti
on
in
S
out
h
C
hi
na
S
e
a
,”
T
he
or
e
ti
c
al
and A
ppl
ie
d M
e
c
hani
c
s
L
e
tt
e
r
s
, vol
. 3,
no. 3, 2013, doi:
10.1063/2.1303202
.
[
2]
B
.
B
ovc
on
a
nd
M
.
K
r
is
t
a
n,
“
W
a
S
R
-
A
w
a
te
r
s
e
gme
nt
a
ti
o
n
a
nd
r
e
f
in
e
me
nt
ma
r
it
im
e
obs
ta
c
le
de
te
c
ti
on
ne
t
w
or
k,”
I
E
E
E
T
r
ans
ac
ti
ons
on C
y
be
r
n
e
ti
c
s
, vol
. 52, no. 12, pp. 12661
–
12674,
D
e
c
. 2022, doi:
10.1109/T
C
Y
B
.2021.3085856.
[
3]
F
.
E
.
T
.
S
c
höl
le
r
,
M
.
B
la
nke
,
M
.
K
.
P
le
ng
e
-
F
e
id
e
nha
ns
’
l,
a
nd
L
.
N
a
lp
a
nt
id
is
,
“
V
is
io
n
-
ba
s
e
d
obj
e
c
t
tr
a
c
ki
ng
in
ma
r
in
e
e
nvi
r
onme
nt
s
us
in
g
f
e
a
tu
r
e
s
f
r
om
ne
ur
a
l
ne
twor
k
de
te
c
ti
ons
,
”
I
F
A
C
-
P
ape
r
s
O
nL
in
e
,
vol
.
53,
no.
2,
pp.
14517
–
14523,
2020,
doi
:
10.1016/j
.i
f
a
c
ol
.2020.12.1455.
[
4]
D
.
K
.
P
r
a
s
a
d,
D
.
R
a
ja
n,
L
.
R
a
c
hma
w
a
ti
,
E
.
R
a
ja
b
a
ll
y,
a
nd
C
.
Q
ue
k,
“
V
id
e
o
pr
oc
e
s
s
in
g
f
r
om
e
le
c
tr
o
-
opt
ic
a
l
s
e
ns
or
s
f
or
ob
je
c
t
de
te
c
ti
on
a
nd
tr
a
c
ki
ng
in
a
ma
r
it
im
e
e
nvi
r
onme
nt
:
a
s
ur
v
e
y,”
I
E
E
E
T
r
ans
ac
ti
ons
on
I
nt
e
ll
ig
e
nt
T
r
ans
po
r
ta
ti
on
Sy
s
te
m
s
,
vol
.
18,
no. 8
, pp. 1993
–
2016, Aug. 2017, do
i:
10.1109/T
I
T
S
.2016.2634580.
[
5]
D
.
Q
ia
o,
G
.
L
iu
,
T
.
L
v,
W
.
L
i,
a
nd
J
.
Z
ha
ng,
“
M
a
r
in
e
vi
s
io
n
-
ba
s
e
d
s
it
ua
ti
ona
l
a
w
a
r
e
ne
s
s
us
in
g
di
s
c
r
im
in
a
ti
ve
de
e
p
le
a
r
ni
ng:
A
s
ur
ve
y,”
J
our
nal
of
M
ar
in
e
S
c
ie
nc
e
and E
ngi
ne
e
r
in
g
, vol
. 9, no. 4,
A
pr
. 2021, doi:
10.3390/j
ms
e
9040397.
[
6]
M
.
T
e
r
š
e
k,
L
.
Ž
us
t,
a
nd
M
.
K
r
is
ta
n,
“
e
W
a
S
R
—
a
n
e
mbe
dde
d
-
c
omput
e
-
r
e
a
dy
ma
r
it
im
e
obs
ta
c
le
de
te
c
ti
on
n
e
tw
or
k,”
Se
n
s
or
s
,
vol
. 23, no. 12, J
un. 2023, doi:
10.3390/s
23125386.
[
7]
A
.
F
.
A
bba
s
,
U
.
U
.
S
he
ik
h,
F
.
T
.
A
l
-
D
hi
e
f
,
a
nd
M
.
N
.
H
.
M
o
hd,
“
A
c
ompr
e
he
ns
iv
e
r
e
vi
e
w
of
ve
hi
c
le
de
te
c
ti
on
us
in
g
c
omp
ut
e
r
vi
s
io
n,”
T
e
lk
om
ni
k
a
(
T
e
le
c
om
m
uni
c
at
io
n
C
om
put
in
g
E
le
c
tr
o
ni
c
s
and
C
ont
r
ol
)
,
vol
.
19,
no.
3,
pp.
838
–
850,
J
un.
2021,
doi
:
10.12928/T
E
L
K
O
M
N
I
K
A
.v19i3.12880.
Evaluation Warning : The document was created with Spire.PDF for Python.