I
AE
S
I
nte
rna
t
io
na
l J
o
urna
l o
f
Ro
bo
t
ics a
nd
Aut
o
m
a
t
io
n
(
I
J
RA)
Vo
l.
1
4
,
No
.
2
,
J
u
n
e
20
2
5
,
p
p
.
2
60
~
2
72
I
SS
N:
2722
-
2
5
8
6
,
DOI
:
1
0
.
1
1
5
9
1
/i
jr
a
.
v
14
i
2
.
pp
2
60
-
2
72
260
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
r
a
.
ia
esco
r
e.
co
m
CP_SDUN
et:
ro
a
d extra
ction
using
SDUN
et
a
nd
cen
t
erline
preserv
ing
dice
los
s
B
a
y
u Sa
t
ria
P
er
s
a
da
,
M
uh
a
m
m
a
d Rif
qi P
riy
o
Su
s
a
nto
,
L
a
k
s
m
it
a
Ra
ha
dia
nti,
Ania
t
i
M
urni
Ary
m
urt
hy
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
a
s
I
n
d
o
n
e
si
a
,
J
a
k
a
r
t
a
,
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
No
v
6
,
2
0
2
4
R
ev
is
ed
J
an
3
1
,
2
0
2
5
Acc
ep
ted
Ma
r
5
,
2
0
2
5
Ex
isti
n
g
a
u
to
m
a
ti
c
ro
a
d
m
a
p
e
x
trac
ti
o
n
a
p
p
ro
a
c
h
e
s
o
n
re
m
o
te
se
n
sin
g
ima
g
e
s
o
ften
fa
il
b
e
c
a
u
se
th
e
y
c
a
n
n
o
t
u
n
d
e
rsta
n
d
th
e
s
p
a
ti
a
l
c
o
n
tex
t
o
f
a
n
ima
g
e
.
M
a
in
l
y
b
e
c
a
u
se
th
e
y
c
o
u
l
d
n
o
t
lea
rn
th
e
sp
a
ti
a
l
c
o
n
te
x
t
o
f
th
e
ima
g
e
a
n
d
o
n
l
y
k
n
e
w
th
e
str
u
c
tu
re
o
r
tex
tu
re
o
f
th
e
ima
g
e
.
T
h
e
se
a
p
p
r
o
a
c
h
e
s
o
n
l
y
fo
c
u
s
o
n
re
g
i
o
n
a
l
a
c
c
u
ra
c
y
i
n
ste
a
d
o
f
c
o
n
n
e
c
ti
v
it
y
.
Th
e
re
f
o
re
,
m
o
st
a
p
p
ro
a
c
h
e
s
p
r
o
d
u
c
e
d
isc
o
n
ti
n
u
o
u
s
o
u
t
p
u
ts
c
a
u
se
d
b
y
b
u
il
d
in
g
s,
sh
a
d
o
ws
,
a
n
d
sim
il
a
rit
y
to
ri
v
e
rs.
Th
is
stu
d
y
a
d
d
re
ss
e
s
th
e
c
h
a
ll
e
n
g
e
o
f
a
u
t
o
m
a
ti
c
ro
a
d
e
x
trac
ti
o
n
,
f
o
c
u
sin
g
o
n
e
n
h
a
n
c
in
g
r
o
a
d
c
o
n
n
e
c
ti
v
it
y
a
n
d
se
g
m
e
n
tatio
n
a
c
c
u
ra
c
y
b
y
p
ro
p
o
si
n
g
a
n
e
two
r
k
-
b
a
se
d
r
o
a
d
e
x
trac
ti
o
n
th
a
t
u
se
s
a
sp
a
ti
a
l
in
ten
sifier
m
o
d
u
le
(DU
LR)
a
n
d
d
e
n
se
ly
c
o
n
n
e
c
ted
U
-
Ne
t
a
rc
h
it
e
c
tu
re
(S
DU
Ne
t)
with
a
c
o
n
n
e
c
ti
v
it
y
-
p
r
e
se
rv
in
g
l
o
ss
fu
n
c
ti
o
n
(CP
_
c
lDice
)
c
a
ll
e
d
CP
_
S
DU
Ne
t.
Th
is
st
u
d
y
a
n
a
ly
z
e
s
th
e
CP
_
c
lDice
lo
ss
fu
n
c
ti
o
n
f
o
r
th
e
ro
a
d
e
x
trac
ti
o
n
tas
k
c
o
m
p
a
re
d
to
t
h
e
BCE
Lo
ss
f
u
n
c
ti
o
n
t
o
train
th
e
S
DU
Ne
t
m
o
d
e
l.
Th
e
re
su
lt
sh
o
ws
th
a
t
CP
_
S
DU
Ne
t,
p
e
rfo
rm
s
b
e
st
u
sin
g
a
n
ima
g
e
siz
e
o
f
1
2
8
×
1
2
8
p
i
x
e
ls
a
n
d
t
ra
in
e
d
wit
h
th
e
w
h
o
le
d
a
tas
e
t
with
a
c
o
m
b
in
a
ti
o
n
o
f
2
0
%
c
l
Dic
e
a
n
d
8
0
%
d
ice
lo
ss
.
Th
e
p
r
o
p
o
se
d
m
e
th
o
d
o
b
tai
n
s
a
c
lDice
sc
o
re
o
f
0
.
8
5
a
n
d
a
n
In
tere
st
o
v
e
r
Un
i
o
n
(I
o
U)
sc
o
re
o
f
0
.
6
5
f
o
r
th
e
tes
ti
n
g
d
a
ta.
T
h
e
se
fin
d
i
n
g
s
d
e
m
o
n
stra
te
th
e
p
o
te
n
ti
a
l
o
f
CP
_
S
DU
Ne
t
fo
r
re
li
a
b
le ro
a
d
e
x
trac
ti
o
n
.
K
ey
w
o
r
d
s
:
C
o
n
n
ec
tiv
ity
p
r
eser
v
i
n
g
C
P_
clDice
C
P_
SDU
Net
Dice
lo
s
s
R
o
ad
ex
tr
ac
tio
n
SDUNet
Th
is
is
a
n
o
p
en
a
cc
ess
a
r
ticle
u
n
d
er th
e
C
C
B
Y
-
SA
licen
s
e.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
B
ay
u
Satr
ia
Per
s
ad
a
Facu
lty
o
f
C
o
m
p
u
ter
Scien
ce
,
Un
iv
er
s
itas
I
n
d
o
n
esia
So
u
th
J
ak
ar
ta,
J
ak
ar
ta,
J
av
a,
I
n
d
o
n
esia
E
m
ail: b
ay
u
s
atr
iap
er
s
ad
a@
g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
Au
to
m
atic
r
o
ad
m
ap
g
e
n
er
ati
o
n
o
r
r
o
ad
ex
tr
ac
tio
n
f
r
o
m
im
ag
es
b
en
ef
its
m
an
y
a
p
p
licatio
n
s
,
s
u
ch
as
au
to
n
o
m
o
u
s
v
eh
icles,
u
r
b
an
p
l
an
n
in
g
,
r
o
ad
n
etwo
r
k
p
la
n
n
in
g
,
an
d
lan
d
in
v
esti
g
atio
n
.
B
ased
o
n
th
e
d
ata
u
s
ed
,
r
o
ad
ex
tr
ac
tio
n
task
s
ca
n
b
e
d
iv
id
ed
in
to
th
r
ee
ty
p
es,
i.e
.
,
r
o
a
d
ex
tr
ac
tio
n
u
s
in
g
a
2
D
d
ataset,
a
3
D
d
ataset,
o
r
b
o
th
ty
p
es
s
im
u
ltan
eo
u
s
ly
.
R
o
ad
ex
t
r
ac
tio
n
u
s
in
g
a
2
D
d
at
aset
ca
n
b
e
d
o
n
e
u
s
in
g
o
p
tical
im
ag
es
[
1
]
,
[
2
]
o
r
SAR
im
ag
es
[
3
]
,
[
4
]
,
wh
ich
ea
ch
h
av
e
th
eir
c
h
ar
ac
ter
is
tics
.
T
h
is
p
ap
er
f
o
cu
s
es
o
n
u
s
in
g
2
D
d
atasets
,
s
p
ec
if
ically
o
p
tical
im
ag
es,
to
p
er
f
o
r
m
au
to
m
atic
r
o
ad
e
x
tr
ac
tio
n
.
T
h
e
m
ain
p
r
o
b
le
m
in
r
o
ad
e
x
tr
ac
tio
n
f
r
o
m
s
atellite
im
ag
es
is
to
lab
el
th
e
r
o
ad
as
f
o
r
eg
r
o
u
n
d
an
d
o
th
er
th
in
g
s
as
b
ac
k
g
r
o
u
n
d
,
cr
ea
tin
g
a
b
in
a
r
y
s
eg
m
en
tatio
n
p
r
o
b
lem
.
C
o
m
p
ar
e
d
with
o
th
er
im
ag
e
s
eg
m
en
tatio
n
p
r
o
b
lem
s
,
m
an
y
f
ac
to
r
s
af
f
ec
t r
o
ad
e
x
tr
ac
tio
n
.
T
h
e
f
ir
s
t is th
at
r
o
ad
s
ar
e
th
in
an
d
lo
n
g
,
s
p
an
n
in
g
a
lar
g
e
ar
ea
ev
en
t
h
o
u
g
h
th
ei
r
ar
ea
(
p
ix
el
c
o
u
n
t)
is
s
m
all.
Nex
t,
th
e
g
eo
m
etr
ic
f
ea
t
u
r
es
o
f
r
o
a
d
s
ar
e
v
er
y
s
im
ilar
to
r
iv
er
s
an
d
r
ailway
s
.
T
h
ese
p
r
o
b
lem
s
m
ak
e
it
tr
ick
y
ev
en
f
o
r
p
r
o
f
ess
io
n
als
to
d
if
f
er
en
tiate
b
etwe
en
th
em
.
Ad
d
itio
n
ally
,
r
o
a
d
s
ca
n
o
v
er
lap
in
a
c
o
m
p
lex
in
ter
s
ec
t
io
n
,
a
n
d
f
in
ally
,
in
s
atellite
im
ag
es,
r
o
a
d
s
ca
n
b
e
o
b
s
cu
r
ed
b
y
tr
ee
s
,
th
eir
s
h
ad
o
ws,
an
d
b
u
ild
in
g
s
'
s
h
ad
o
ws.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
AE
S
I
n
t
J
R
o
b
&
A
u
to
m
I
SS
N:
2722
-
2
5
8
6
C
P
_
S
DUNet:
r
o
a
d
ex
tr
a
ctio
n
u
s
in
g
S
DUNet a
n
d
ce
n
terl
in
e
p
r
eser
vin
g
d
ice
lo
s
s
(
B
a
yu
S
a
t
r
ia
P
ers
a
d
a
)
261
I
n
ea
r
ly
s
tu
d
ies,
r
esear
ch
e
r
s
d
esig
n
ed
h
an
d
-
cr
af
te
d
f
ea
t
u
r
es
f
o
r
r
o
a
d
e
x
tr
ac
tio
n
[
5
]
–
[
9
]
.
T
h
is
co
n
v
en
tio
n
al
tech
n
iq
u
e
u
s
es
tex
tu
r
e
s
tr
u
ctu
r
e,
s
u
p
er
-
p
i
x
el
p
ar
titi
o
n
in
g
[
1
0
]
,
r
eg
io
n
-
g
r
o
win
g
alg
o
r
ith
m
[
1
1
]
an
d
o
th
er
m
eth
o
d
s
f
o
r
r
o
ad
e
x
tr
ac
tio
n
.
Ho
wev
er
,
s
atellite
im
ag
er
y
h
as
g
r
o
w
n
m
ass
iv
ely
in
r
ec
en
t
y
ea
r
s
an
d
p
r
o
d
u
ce
d
v
er
y
h
ig
h
-
r
eso
lu
tio
n
im
ag
es.
T
h
ese
im
ag
es
b
r
i
n
g
n
ew
c
h
allen
g
es
f
o
r
r
o
ad
e
x
tr
ac
tio
n
d
u
e
to
th
e
co
m
p
lex
s
tr
u
ctu
r
e
o
f
th
e
r
o
ad
an
d
d
iv
er
s
e
b
ac
k
g
r
o
u
n
d
d
is
tr
ib
u
tio
n
.
W
ith
th
ese
n
ew
ch
alle
n
g
es,
co
n
v
e
n
tio
n
al
m
eth
o
d
s
s
tar
t
to
u
n
d
er
p
e
r
f
o
r
m
.
I
n
r
ec
e
n
t
y
ea
r
s
,
th
e
d
ev
el
o
p
m
en
t
o
f
n
eu
r
al
n
etwo
r
k
s
s
u
ch
as
c
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NN)
h
as
i
m
p
r
o
v
e
d
i
n
lo
ts
o
f
s
em
an
tic
s
eg
m
en
tatio
n
task
s
,
wh
ich
in
cl
u
d
es
r
o
a
d
e
x
tr
ac
tio
n
[
1
2
]
–
[
1
5
]
.
R
ec
en
tly
,
th
e
U
-
Net
[
1
6
]
ar
c
h
itectu
r
e
h
as
b
ee
n
u
s
ed
ex
ten
s
iv
ely
f
o
r
s
eg
m
e
n
tatio
n
task
s
.
Alth
o
u
g
h
C
NN
-
b
ased
m
eth
o
d
s
an
d
U
-
Net
ar
e
u
b
iq
u
ito
u
s
in
s
eg
m
en
tatio
n
task
s
,
th
ey
a
r
e
o
f
te
n
in
ad
eq
u
ate
f
o
r
r
em
o
te
s
en
s
in
g
im
ag
es.
T
o
a
d
d
r
ess
th
e
is
s
u
e,
Yan
g
et
a
l.
[
1
7
]
d
esig
n
ed
a
f
r
a
m
ewo
r
k
b
ased
o
n
s
p
atial
-
en
h
an
ce
d
an
d
d
en
s
ely
co
n
n
ec
ted
U
-
Net
to
a
d
d
r
ess
th
e
p
r
o
b
lem
.
Ad
d
itio
n
ally
,
th
e
tr
ain
in
g
s
tr
ateg
ies
ca
n
also
b
e
m
o
d
if
ied
t
o
o
p
tim
ize
m
o
d
els
f
o
r
ex
tr
ac
t
io
n
task
s
.
T
h
e
m
o
d
el'
s
lo
s
s
f
u
n
ctio
n
p
la
y
s
an
im
p
o
r
ta
n
t
r
o
le
as
it
g
u
i
d
es
th
e
m
o
d
el
in
ac
h
iev
in
g
t
h
e
b
est
r
esu
lt
b
ased
o
n
th
e
d
if
f
er
en
ce
b
etwe
en
th
e
p
r
ed
icted
v
al
u
e
a
n
d
its
g
r
o
u
n
d
tr
u
th
.
Fo
r
a
task
th
at
u
s
es
s
e
g
m
en
tatio
n
,
in
th
is
ca
s
e,
r
o
ad
ex
tr
ac
tio
n
,
ch
o
o
s
in
g
th
e
lo
s
s
f
u
n
ctio
n
is
im
p
o
r
tan
t.
Sev
er
al
s
tu
d
ies
h
av
e
p
r
o
p
o
s
ed
alter
n
ate
lo
s
s
f
u
n
ctio
n
s
th
at
co
u
ld
in
c
r
ea
s
e
th
e
o
u
tco
m
e
o
f
r
o
ad
e
x
tr
ac
tio
n
[
1
8
]
.
On
e
o
f
th
em
u
s
es
s
tr
u
ctu
r
al
s
im
ilar
ity
as
a
lo
s
s
f
u
n
ctio
n
,
wh
ic
h
is
a
jo
in
t
lo
s
s
o
f
cr
o
s
s
en
tr
o
p
y
an
d
Dice
lo
s
s
ca
lled
cr
o
s
s
-
en
tr
o
p
y
-
d
ice
-
lo
s
s
(
C
E
DL
)
[
1
9
]
f
u
n
ctio
n
.
Ho
wev
er
,
n
o
n
e
o
f
th
em
m
ain
tain
th
e
r
o
ad
c
o
n
n
ec
tiv
ity
.
Fo
r
th
at
r
ea
s
o
n
,
Ab
d
o
llah
i
et
a
l.
[
2
0
]
p
r
o
p
o
s
ed
u
s
in
g
C
P_
clDice
.
T
h
is
n
ew
m
ea
s
u
r
e
co
m
p
ar
es
th
e
in
ter
s
ec
tio
n
o
f
p
ix
els
an
d
t
h
eir
m
o
r
p
h
o
lo
g
ical
s
k
eleto
n
to
p
r
eser
v
e
th
e
r
o
ad
'
s
co
n
n
ec
tiv
ity
a
n
d
o
b
tain
a
b
etter
p
er
f
o
r
m
an
ce
o
f
th
e
r
o
a
d
ex
t
r
ac
tio
n
m
o
d
el.
I
n
th
is
p
ap
er
,
we
p
r
o
p
o
s
ed
a
n
o
v
el
a
p
p
r
o
ac
h
to
a
u
to
m
at
ic
r
o
ad
ex
t
r
ac
tio
n
.
W
e
in
tr
o
d
u
ce
u
s
in
g
s
p
atial
-
en
h
an
ce
d
an
d
d
en
s
ely
co
n
n
ec
ted
U
-
Net
to
ca
p
tu
r
e
s
p
atial
co
n
tex
t
f
r
o
m
a
n
im
ag
e
ca
lled
C
P_
SDU
Net.
C
P_
SDU
Net
ca
n
p
r
o
d
u
ce
a
g
en
er
ally
b
etter
ac
cu
r
ac
y
o
n
r
o
ad
s
eg
m
en
tatio
n
u
s
in
g
C
P_
clDice
as
its
lo
s
s
f
u
n
ctio
n
to
en
s
u
r
e
r
o
ad
co
n
n
ec
tiv
ity
is
p
r
eser
v
ed
,
wh
ich
u
s
es
a
s
p
atial
in
ten
s
if
ier
(
D
UL
R
m
o
d
u
le)
an
d
d
en
s
ely
co
n
n
ec
te
d
U
-
Net
with
co
n
n
ec
tiv
ity
p
r
eser
v
in
g
lo
s
s
f
u
n
ctio
n
.
B
y
en
h
an
ci
n
g
s
p
a
tial
co
n
tex
t
ca
p
tu
r
e
an
d
e
x
p
licitly
m
ain
tain
i
n
g
r
o
ad
c
o
n
n
ec
tiv
ity
,
th
is
m
eth
o
d
aim
s
t
o
o
v
er
co
m
e
lim
it
atio
n
s
in
e
x
is
tin
g
tech
n
iq
u
es.
O
u
r
w
o
r
k
im
p
r
o
v
es
s
eg
m
en
tatio
n
ac
cu
r
ac
y
an
d
en
s
u
r
es
p
r
ac
tical
u
s
ab
ilit
y
f
o
r
ap
p
licatio
n
s
th
at
d
em
an
d
r
eliab
le
r
o
ad
n
etwo
r
k
ex
tr
ac
tio
n
,
s
u
ch
as
u
r
b
a
n
i
n
f
r
astru
ctu
r
e
p
lan
n
i
n
g
a
n
d
a
u
to
n
o
m
o
u
s
v
e
h
icle
s
y
s
tem
s
.
2.
M
E
T
H
O
D
S
T
h
is
wo
r
k
p
r
o
p
o
s
es
a
n
ew
r
o
ad
ex
tr
ac
tio
n
m
eth
o
d
ca
lled
c
o
n
n
ec
tiv
ity
p
r
eser
v
in
g
s
p
atial
en
h
an
ce
d
an
d
d
en
s
ely
U
-
Net
(
C
P_
SDUNet)
.
T
h
e
p
r
o
p
o
s
ed
tech
n
iq
u
e
is
a
m
o
d
if
icatio
n
o
f
t
h
e
o
r
ig
in
al
d
ee
p
lear
n
in
g
m
o
d
el
SDUNet
[
1
7
]
,
wh
ich
is
b
ased
o
n
U
-
Net
ar
ch
itectu
r
e
b
u
t
with
b
etter
p
er
f
o
r
m
a
n
ce
b
y
en
h
an
cin
g
s
p
atial
co
n
tex
t.
Ad
d
itio
n
ally
,
it
co
n
s
is
ts
o
f
a
co
n
n
ec
tiv
ity
-
awa
r
e
s
im
ilar
ity
m
ea
s
u
r
e
b
ased
o
n
th
e
in
ter
s
ec
tio
n
o
f
th
e
s
k
eleto
n
an
d
its
m
ask
(
C
P_
clDice
)
[
2
0
]
t
o
p
r
eser
v
e
co
n
n
ec
t
iv
ity
o
f
th
e
r
o
ad
th
at
h
as b
ee
n
ex
tr
ac
ted
.
2
.
1
.
SDUNet
SDUNet
[
1
7
]
is
a
d
ee
p
n
etwo
r
k
ar
ch
itectu
r
e
th
at
is
b
ased
o
n
U
-
Net
[
1
6
]
an
d
R
esNet
[
2
1
]
.
SDUNet
co
n
s
is
ts
o
f
th
r
ee
m
ajo
r
p
ar
ts
:
d
ec
o
d
er
,
b
o
ttlen
ec
k
,
an
d
d
e
co
d
er
.
T
h
e
en
c
o
d
er
is
f
illed
with
co
n
v
o
lu
tio
n
a
l
b
lo
ck
an
d
R
eL
U
Activ
atio
n
,
wh
ich
later
will
b
e
d
o
wn
s
am
p
led
u
s
in
g
Ma
x
p
o
o
lin
g
,
w
h
ile
th
e
d
ec
o
d
er
is
u
p
-
s
am
p
led
with
tr
an
s
p
o
s
e
co
n
v
o
lu
tio
n
.
T
h
e
p
r
o
m
in
en
t f
ea
t
u
r
e
o
f
SDUNet
is
th
e
u
s
e
o
f
a
s
k
ip
co
n
n
ec
tio
n
,
wh
ic
h
was
in
s
p
ir
ed
b
y
R
esNet
[
2
1
]
.
Sk
ip
co
n
n
ec
tio
n
allo
ws
th
e
s
en
d
in
g
o
f
in
f
o
r
m
atio
n
with
o
u
t
th
e
n
ee
d
to
g
o
th
r
o
u
g
h
lay
er
s
o
f
co
n
v
o
lu
tio
n
.
Hen
ce
,
p
r
e
v
en
ts
v
an
is
h
in
g
g
r
ad
ien
t p
r
o
b
lem
s
.
SDUNet
is
also
in
s
p
ir
ed
b
y
D
en
s
eNe
t
wh
er
e
ea
ch
lay
er
is
c
o
n
n
ec
ted
to
t
h
e
o
t
h
er
s
o
th
at
i
t
is
ab
le
to
im
p
r
o
v
e
f
ea
t
u
r
e
f
lo
ws.
T
h
e
en
co
d
er
co
n
s
is
ts
o
f
4
d
en
s
e
b
lo
c
k
s
,
th
e
b
o
ttlen
ec
k
co
n
s
is
ts
o
f
1
d
en
s
e
b
lo
ck
,
a
n
d
last
ly
,
th
e
d
ec
o
d
er
co
n
s
is
ts
o
f
4
d
en
s
e
b
lo
c
k
s
[
1
7
]
.
2
.
1
.
1
.
Dense
blo
ck
SDUNet
i
s
a
s
er
ies
o
f
d
en
s
e
b
lo
ck
s
th
at
m
ain
ly
co
n
s
is
t
o
f
m
u
ltip
le
lay
er
s
wh
ich
ex
tr
ac
t
in
f
o
r
m
atio
n
th
at
is
s
h
ap
ed
lik
e
th
e
letter
U
an
d
ar
e
co
n
n
ec
ted
to
ea
ch
o
th
er
,
wh
ich
was
in
s
p
ir
ed
b
y
De
n
s
eNe
t
ar
ch
itectu
r
e
[
2
2
]
.
T
h
e
b
lo
ck
s
a
r
e
co
n
n
ec
ted
to
p
r
eser
v
e
th
e
n
atu
r
e
o
f
f
o
r
wa
r
d
f
ee
d
in
g
,
wh
er
e
ea
ch
lay
er
will
ad
d
ad
d
itio
n
al
in
f
o
r
m
atio
n
f
r
o
m
al
l
p
r
ec
ed
i
n
g
la
y
er
s
an
d
p
ass
its
f
ea
tu
r
e
m
a
p
to
s
u
b
s
eq
u
en
t
lay
er
s
.
First,
th
e
in
p
u
t
ten
s
o
r
p
ass
es
th
r
o
u
g
h
a
b
atch
n
o
r
m
aliza
tio
n
lay
er
t
o
r
e
-
s
ca
le
an
d
s
tan
d
ar
d
ize
its
v
alu
e.
T
h
e
n
,
it
is
ac
tiv
ated
to
a
r
ec
tifie
d
lin
ea
r
u
n
it
ac
tiv
ati
o
n
f
u
n
ctio
n
,
f
o
llo
wed
b
y
a
3
×3
co
n
v
o
l
u
tio
n
lay
er
.
SDUNet
u
s
es
a
m
o
d
if
ied
v
er
s
io
n
o
f
De
n
s
eNe
t,
wh
er
e
a
f
ter
g
o
in
g
th
r
o
u
g
h
a
3
×3
co
n
v
o
lu
tio
n
lay
er
,
th
e
ten
s
o
r
m
o
v
e
s
to
a
d
r
o
p
o
u
t
lay
er
to
d
r
o
p
s
o
m
e
n
e
u
r
o
n
s
r
an
d
o
m
l
y
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
2
5
8
6
I
AE
S
I
n
t
J
R
o
b
&
A
u
to
m
,
Vo
l
.
14
,
No
.
2
,
J
u
n
e
20
2
5
:
2
60
-
2
72
262
Fig
u
r
e
1
illu
s
tr
ates
d
en
s
e
b
l
o
ck
s
u
s
ed
in
SDUNet.
T
h
e
p
u
r
p
o
s
e
o
f
u
s
in
g
d
e
n
s
e
b
lo
c
k
s
was
to
en
h
an
ce
in
f
o
r
m
atio
n
f
lo
w
b
e
twee
n
lay
er
s
.
C
o
n
s
eq
u
e
n
tly
,
th
e
ℎ
lay
er
r
ec
ei
v
es
th
e
f
ea
tu
r
e
m
ap
s
o
f
all
p
r
ec
ed
in
g
lay
er
s
,
0
,
…
,
−
1
as in
p
u
t.
T
h
e
f
o
r
m
u
la
is
r
ep
r
esen
ted
in
(
1
)
.
=
(
[
0
,
…
,
−
1
]
)
(
1
)
wh
er
e
[
0
,
…
,
−
1
]
r
ep
r
esen
t
th
e
co
n
ca
ten
atio
n
o
f
th
e
f
ea
tu
r
e
m
ap
s
g
en
e
r
ated
f
r
o
m
all
lay
er
s
b
ef
o
r
eh
a
n
d
an
d
(
.
)
r
ef
e
r
s
to
a
n
o
n
lin
ea
r
m
a
p
p
i
n
g
f
u
n
ctio
n
.
Fig
u
r
e
1
.
Den
s
e
b
l
o
ck
in
SDU
Net
ar
ch
itectu
r
e
2
.
1
.
2
.
T
ra
ns
it
io
n
la
y
er
T
h
e
tr
a
n
s
itio
n
lay
er
co
n
s
is
ts
o
f
two
p
ar
ts
:
a
d
o
wn
-
tr
a
n
s
itio
n
lay
er
a
n
d
an
u
p
-
tr
an
s
itio
n
lay
er
.
T
h
e
d
o
wn
-
tr
an
s
itio
n
lay
er
is
ap
p
lie
d
to
d
o
wn
s
am
p
le
th
e
f
ea
tu
r
e
m
ap
,
wh
ich
th
e
d
en
s
e
b
lo
ck
c
o
u
ld
n
o
t
d
o
.
So
,
th
e
tr
an
s
itio
n
lay
er
co
n
s
is
ts
o
f
b
atch
n
o
r
m
aliza
tio
n
to
s
tan
d
ar
d
iz
e,
f
o
llo
wed
b
y
a
1
×1
co
n
v
o
l
u
t
io
n
lay
er
an
d
a
2
×2
p
o
o
lin
g
lay
er
to
d
o
wn
s
am
p
lin
g
.
C
o
n
v
e
r
s
ely
,
th
e
u
p
-
tr
an
s
itio
n
lay
e
r
will
p
er
f
o
r
m
th
e
u
p
-
s
am
p
lin
g
u
s
in
g
th
e
s
am
e
ap
p
r
o
ac
h
as
th
e
d
o
wn
-
t
r
an
s
itio
n
lay
er
,
b
u
t
it
u
s
es
a
t
r
an
s
p
o
s
ed
co
n
v
o
lu
tio
n
in
s
tea
d
o
f
a
2
×2
p
o
o
lin
g
lay
er
.
2
.
1
.
3
.
Do
wn
up
lef
t
rig
ht
co
n
v
o
lutio
n m
o
du
le
(
DULR)
SDUNet
al
s
o
p
r
o
p
o
s
ed
a
n
o
th
e
r
way
to
s
p
atially
en
h
an
ce
th
e
s
p
atial
co
n
tex
t
f
ea
tu
r
e
b
y
d
o
i
n
g
wh
at
is
ca
lled
a
DUL
R
m
o
d
u
le
[
2
3
]
.
Fig
u
r
e
2
illu
s
tr
ates
th
e
d
etail
s
o
f
th
e
DUL
R
m
o
d
u
le
ar
ch
itectu
r
e,
wh
ich
co
n
s
is
ts
o
f
u
p
c
o
n
v
o
lu
tio
n
,
d
o
wn
c
o
n
v
o
lu
tio
n
,
lef
t
co
n
v
o
l
u
tio
n
,
a
n
d
r
ig
h
t
co
n
v
o
lu
tio
n
.
As
s
h
o
wn
i
n
th
e
f
ig
u
r
e,
if
we
tak
e
d
o
wn
co
n
v
o
lu
tio
n
an
d
lef
t
co
n
v
o
lu
tio
n
as
an
ex
am
p
le,
th
e
lef
t
co
n
v
o
lu
tio
n
will
r
ed
u
ce
th
e
f
ea
tu
r
e
m
ap
f
r
o
m
a
s
ize
o
f
C
×
W
×
H
to
C
×
1
×
H
s
o
th
at
it
will
co
n
v
o
lv
e
a
lo
n
g
th
o
s
e
H.
Me
an
wh
ile,
Do
wn
co
n
v
o
lu
tio
n
will
r
ed
u
ce
a
C
×
W
×
H
f
ea
tu
r
e
m
ap
to
b
ec
o
m
e
C
×
W
×
1
s
o
th
at
it
co
n
v
o
l
v
es
f
o
r
th
e
W
ten
s
o
r
o
n
ly
.
I
t
will
co
n
v
o
lv
e
ev
er
y
C
×
W
×
1
f
o
r
ev
er
y
s
lice
u
n
til
it
b
ec
o
m
es
th
e
H
ag
ain
s
o
th
at
th
e
f
ea
tu
r
e
m
ap
af
ter
it
g
ets
co
n
v
o
lu
ted
is
b
ac
k
to
C
×
W
×
H.
E
v
er
y
co
n
v
o
lu
tio
n
will
b
e
ac
tiv
ated
u
s
in
g
a
n
o
n
lin
ea
r
f
u
n
ctio
n
t
h
at
is
a
r
ec
tifie
d
lin
ea
r
u
n
it
(
R
eL
U)
.
Similar
ly
,
th
e
o
th
er
c
o
n
v
o
lu
ti
o
n
h
as th
e
s
am
e
f
u
n
cti
o
n
b
u
t
g
o
es in
a
d
if
f
er
e
n
t d
ir
ec
tio
n
.
T
h
e
in
ten
tio
n
o
f
u
s
in
g
DUL
R
is
s
o
th
at
th
r
o
u
g
h
DUL
R
th
e
f
ea
tu
r
e
m
ap
p
o
te
n
tially
co
n
tain
s
p
er
s
p
ec
tiv
e
ch
an
g
e
d
u
e
to
th
e
d
if
f
e
r
en
t
c
o
m
p
u
tin
g
o
r
d
e
r
s
,
t
h
e
ef
f
ec
t
o
f
ag
g
r
eg
atio
n
f
o
r
e
ac
h
r
o
w
is
d
if
f
er
en
t,
th
u
s
m
ak
in
g
it lik
e
a
ch
an
g
e
in
p
er
s
p
ec
tiv
e
[
2
3
]
.
2
.
2
.
Co
nn
ec
t
iv
it
y
preserv
ing
ce
nte
rline
dice
co
ef
f
icient
(
CP
_
clDice
)
.
T
h
e
ce
n
ter
lin
e
d
ice
co
ef
f
icien
t
was
in
tr
o
d
u
ce
d
b
y
Sh
it
et
a
l.
[
2
4
]
with
th
e
in
ten
tio
n
o
f
f
in
d
in
g
wh
at
is
a
g
o
o
d
p
i
x
els
-
wis
e
m
ea
s
u
r
e
to
b
e
n
ch
m
a
r
k
t
u
b
u
lar
s
eg
m
en
tatio
n
wh
ile
g
u
ar
an
teein
g
th
e
p
r
eser
v
atio
n
o
f
th
e
n
etwo
r
k
to
p
o
lo
g
y
s
in
ce
th
ey
ar
e
u
s
in
g
b
io
m
ed
ical
im
ag
es
t
h
at
is
b
r
ain
v
ascu
lar
d
ataset.
T
h
ey
th
u
s
p
r
o
p
o
s
ed
th
e
ce
n
ter
lin
e
d
ice
c
o
ef
f
icien
t
(
clDice
)
,
as sh
o
wn
in
(
2
)
an
d
(
3
)
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
AE
S
I
n
t
J
R
o
b
&
A
u
to
m
I
SS
N:
2722
-
2
5
8
6
C
P
_
S
DUNet:
r
o
a
d
ex
tr
a
ctio
n
u
s
in
g
S
DUNet a
n
d
ce
n
terl
in
e
p
r
eser
vin
g
d
ice
lo
s
s
(
B
a
yu
S
a
t
r
ia
P
ers
a
d
a
)
263
Fig
u
r
e
2
.
DUL
R
m
o
d
u
le
ar
ch
i
tectu
r
e
(
,
)
=
|
∩
|
(
2
)
(
,
)
=
|
∩
|
(
3
)
clDice
will
cr
ea
te
s
o
f
t
s
k
eleto
n
izatio
n
f
o
r
t
h
e
r
ea
l
g
r
o
u
n
d
tr
u
th
m
ask
(
)
an
d
th
e
p
r
ed
icted
m
ask
f
r
o
m
th
e
n
etwo
r
k
(
)
.
T
h
e
s
k
eleto
n
izatio
n
will
b
e
ca
lled
Sl
f
o
r
th
e
s
k
eleto
n
izatio
n
o
f
th
e
g
r
o
u
n
d
tr
u
t
h
m
ask
an
d
f
o
r
th
e
s
k
eleto
n
izatio
n
o
f
th
e
p
r
ed
icted
m
ask
.
T
h
e
id
ea
o
f
clDice
is
th
e
co
m
p
u
tatio
n
o
f
s
o
f
t
-
Dice
,
th
at
i
s
to
co
m
p
u
te
th
e
f
r
ac
tio
n
o
f
th
at
lies
with
in
,
wh
ich
we
ca
ll
th
e
to
p
o
lo
g
y
p
r
ec
is
io
n
(
)
an
d
th
e
f
r
ac
tio
n
o
f
th
at
lies
with
in
,
wh
ic
h
we
ca
ll
th
e
to
p
o
lo
g
y
s
en
s
itiv
ity
(
)
.
W
h
en
an
d
h
av
e
b
ee
n
ca
lc
u
lated
,
it
co
u
ld
b
e
co
n
s
id
er
e
d
th
at
is
s
u
s
ce
p
tib
le
to
f
alse
p
o
s
itiv
es
an
d
T
s
en
s
to
f
alse
n
eg
ativ
e.
Sin
ce
we
n
ee
d
to
m
ax
im
ize
b
o
t
h
p
r
ec
is
io
n
an
d
s
en
s
itiv
ity
,
we
d
ef
in
e
clDice
to
b
e
h
ar
m
o
n
ic
m
ea
n
also
k
n
o
wn
as
F
-
1
o
r
D
ice,
th
e
f
o
r
m
u
la
is
s
h
o
wn
in
(
4
)
[
2
0
]
.
(
,
)
=
2
+
(
4
)
wh
er
e
is
th
e
to
p
o
lo
g
y
p
r
ec
is
i
o
n
an
d
th
e
to
p
o
l
o
g
y
s
en
s
itiv
ity
.
E
x
tr
ac
tin
g
a
n
ac
cu
r
ate
s
k
el
eto
n
is
ess
en
tial to
th
e
C
P_
clDice
m
et
h
o
d
.
Fig
u
r
e
3
illu
s
tr
ates h
o
w
S
o
f
t
-
clDice
wo
r
k
s
.
Fig
u
r
e
3
.
C
o
m
p
u
tatio
n
o
f
So
f
t
-
clDice
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
2
5
8
6
I
AE
S
I
n
t
J
R
o
b
&
A
u
to
m
,
Vo
l
.
14
,
No
.
2
,
J
u
n
e
20
2
5
:
2
60
-
2
72
264
Ho
wev
er
,
s
k
eleto
n
izatio
n
u
s
in
g
er
o
s
io
n
an
d
d
ilatio
n
is
n
o
t
f
u
lly
d
if
f
e
r
en
tiab
le
a
n
d
t
h
er
ef
o
r
e
ca
n
n
o
t
b
e
ca
p
tu
r
ed
b
y
a
lo
s
s
f
u
n
ctio
n
.
I
n
s
tead
,
d
o
in
g
m
o
r
p
h
o
lo
g
i
ca
l
th
in
n
in
g
,
we
u
s
e
an
alter
n
ativ
e
th
at
is
u
s
in
g
m
ax
-
p
o
o
lin
g
an
d
m
in
-
p
o
o
lin
g
.
T
h
u
s
,
a
s
o
f
t
-
s
k
eleto
n
izatio
n
tech
n
iq
u
e
was
p
r
o
p
o
s
ed
to
b
e
th
e
alter
n
ativ
e,
wh
er
e
an
iter
ativ
e
m
in
-
an
d
m
ax
-
p
o
o
lin
g
is
ap
p
lied
t
o
ap
p
r
o
x
im
ately
s
im
u
late
er
o
s
io
n
an
d
d
ilatio
n
.
Fig
u
r
e
4
illu
s
tr
ates h
o
w
m
in
-
p
o
o
lin
g
,
m
ax
-
p
o
o
lin
g
,
a
n
d
th
e
u
s
e
o
f
R
eL
U
ac
tiv
atio
n
f
u
n
ctio
n
co
u
ld
b
e
an
alter
n
ativ
e
.
Fig
u
r
e
4
.
Mo
r
p
h
o
lo
g
ical
th
in
n
in
g
u
s
in
g
m
i
n
an
d
m
ax
p
o
o
lin
g
2
.
3
.
P
r
o
po
s
ed
m
et
ho
d
(
CP
_
SDUNet
)
Ou
r
p
r
o
p
o
s
ed
m
eth
o
d
co
m
b
i
n
ed
SDUNet
with
th
e
C
P_
cl
Dice
lo
s
s
f
u
n
ctio
n
.
Pre
v
io
u
s
s
tu
d
ies
[
2
0
]
p
r
o
v
e
d
th
at
clDice
as
a
lo
s
s
f
u
n
ctio
n
im
p
r
o
v
ed
th
e
s
eg
m
en
tatio
n
r
esu
lt.
T
h
e
r
ef
o
r
e
,
we
u
s
e
th
e
s
p
atially
en
h
an
ce
d
ex
tr
ac
t
o
r
i
n
SDU
Net
to
o
b
tain
t
h
e
b
est
r
o
a
d
ex
tr
ac
tio
n
n
etwo
r
k
an
d
p
r
e
s
er
v
e
its
n
etwo
r
k
’
s
to
p
o
lo
g
y
.
Sin
ce
th
e
o
b
jectiv
e
i
s
to
m
ain
tain
to
p
o
lo
g
y
wh
ile
a
ch
iev
in
g
ac
cu
r
ate
s
eg
m
e
n
tatio
n
an
d
n
o
t le
ar
n
th
e
s
k
eleto
n
izatio
n
,
it
is
p
r
o
p
o
s
ed
th
at
clDice
co
m
b
i
n
e
with
s
o
f
t
-
Dice
to
b
e
u
s
ed
as
a
lo
s
s
f
u
n
ct
io
n
r
e
p
r
esen
ted
i
n
(
5
)
.
=
(
1
−
)
(
1
−
)
+
(
1
−
)
(
5
)
w
h
er
e
α
is
th
e
c
o
ef
f
icien
t
th
at
d
iv
id
es
th
e
wo
r
k
f
o
r
ea
ch
l
o
s
s
f
u
n
ctio
n
an
d
a
∈
[
0
,
0
,
5
]
,
is
th
e
C
P_
clDice
L
o
s
s
,
m
ak
in
g
th
is
lo
s
s
f
u
n
c
tio
n
a
co
n
n
ec
tiv
ity
p
r
eser
v
i
n
g
th
e
C
en
ter
lin
e
d
ice
lo
s
s
f
u
n
ctio
n
.
Fig
u
r
e
5
illu
s
tr
ates w
h
at
th
e
p
r
o
p
o
s
ed
n
etwo
r
k
ar
ch
itectu
r
e
lo
o
k
s
lik
e.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
AE
S
I
n
t
J
R
o
b
&
A
u
to
m
I
SS
N:
2722
-
2
5
8
6
C
P
_
S
DUNet:
r
o
a
d
ex
tr
a
ctio
n
u
s
in
g
S
DUNet a
n
d
ce
n
terl
in
e
p
r
eser
vin
g
d
ice
lo
s
s
(
B
a
yu
S
a
t
r
ia
P
ers
a
d
a
)
265
Fig
u
r
e
5
.
C
P_
SDUNet
ar
ch
itectu
r
e
2
.
4
.
E
x
perim
ent
s
T
h
e
ex
p
er
im
en
t
will
b
e
d
iv
id
e
d
in
to
two
p
ar
ts
.
T
h
e
f
ir
s
t
is
t
h
e
C
P_
SDUNet
p
er
f
o
r
m
an
ce
e
x
p
er
im
en
t
an
d
th
e
jo
in
t
l
o
s
s
co
n
f
ig
u
r
at
io
n
ex
p
er
im
en
t.
All
th
e
e
x
p
e
r
im
en
ts
u
s
ed
NVI
DI
A
DGX
A
-
1
0
0
GPU
with
T
en
s
o
r
Flo
w
as its
f
r
am
ewo
r
k
.
2
.
4
.
1
.
E
x
perim
ent
des
ig
n
T
h
e
ex
p
e
r
im
en
tal
d
esig
n
u
s
ed
f
o
r
th
is
r
esear
ch
is
illu
s
tr
ated
in
Fig
u
r
e
6
.
First,
f
o
r
th
e
d
at
a
p
r
ep
r
o
ce
s
s
in
g
s
tag
e,
th
e
in
p
u
t f
r
o
m
th
e
d
ataset
will
b
e
tak
e
n
b
y
its
o
r
ig
in
al
s
ize,
t
h
en
it
is
c
r
o
p
p
e
d
an
d
r
esized
to
2
5
6
×
2
5
6
p
ix
els
o
r
1
2
8
×
1
2
8
p
ix
els.
E
ac
h
s
ize
is
u
s
ed
in
d
if
f
e
r
en
t
ex
p
e
r
im
en
ts
.
Af
ter
th
e
im
ag
e
is
p
r
ep
r
o
ce
s
s
ed
,
it
g
o
es
to
an
a
u
g
m
en
tatio
n
wh
er
e
th
e
d
ata
will
b
e
in
cr
ea
s
ed
v
ia
au
g
m
e
n
tatio
n
,
th
en
to
th
e
m
o
d
e
l
f
o
r
t
r
a
i
n
i
n
g
.
A
f
t
e
r
t
h
at
,
th
e
r
e
s
u
l
t
o
f
e
a
c
h
m
o
d
e
l
w
i
ll
b
e
e
v
a
l
u
a
t
e
d
u
s
i
n
g
t
h
e
c
l
Di
c
e
m
e
tr
i
c
a
n
d
I
o
U
s
c
o
r
e
.
Fig
u
r
e
6
.
Me
th
o
d
s
u
s
ed
in
t
h
is
r
esear
ch
2
.
4
.
2
.
Da
t
a
s
et
a
nd
prepro
ce
s
s
ing
I
n
th
ese
ex
p
er
im
en
ts
,
we
u
s
ed
th
e
Dee
p
Glo
b
e
r
o
a
d
ex
tr
ac
t
io
n
d
ataset
[
2
5
]
.
T
h
e
Dee
p
Glo
b
e
R
o
ad
E
x
tr
ac
tio
n
d
ataset
co
n
s
is
ts
o
f
6
,
2
2
6
ae
r
ial
im
ag
es
f
o
r
tr
ai
n
in
g
,
1
,
2
4
3
v
alid
atio
n
ae
r
ial
im
ag
es,
an
d
1
,
1
0
1
ae
r
ial
test
im
ag
es.
T
h
e
r
eso
l
u
tio
n
o
f
ea
ch
im
ag
e
is
1
0
2
4
×
1
0
2
4
p
ix
els.
T
h
e
Dee
p
Glo
b
e
R
o
ad
E
x
tr
ac
tio
n
d
ataset
co
v
er
s
im
ag
es
ca
p
tu
r
ed
o
v
er
T
h
ailan
d
,
I
n
d
o
n
esia,
an
d
I
n
d
ia.
T
h
e
im
ag
e
o
f
th
e
Dee
p
Glo
b
e
R
o
ad
E
x
tr
ac
tio
n
d
ataset
ca
n
b
e
s
ee
n
in
Fig
u
r
e
7
,
with
t
h
e
ex
am
p
l
e
o
f
th
e
r
ea
l
d
ataset
in
Fig
u
r
e
7
(
a)
an
d
th
e
im
ag
e
m
ask
in
Fig
u
r
e
7
(
b
)
.
B
ef
o
r
e
th
e
d
ataset
ca
n
b
e
u
s
e
d
,
th
e
d
ataset
will
b
e
p
r
ep
r
o
c
ess
ed
.
T
h
e
im
ag
es
will
b
e
f
ir
s
t
cr
o
p
p
ed
an
d
r
e
d
u
ce
d
in
to
f
o
u
r
p
ar
ts
.
E
ac
h
p
a
r
t
is
th
e
s
am
e
s
ize,
5
1
2
×5
1
2
p
i
x
els.
Af
ter
t
h
at,
th
e
g
r
o
u
n
d
tr
u
t
h
/m
ask
im
ag
es
will
b
e
n
o
r
m
alize
d
t
o
a
b
in
ar
y
im
ag
e.
An
e
x
am
p
le
o
f
th
e
p
r
ep
r
o
ce
s
s
ed
im
ag
es
is
illu
s
tr
ated
in
Fig
u
r
e
8
.
T
h
e
im
ag
es
th
at
h
av
e
b
ee
n
cr
o
p
p
ed
a
n
d
r
esized
th
en
n
ee
d
to
b
e
p
r
o
ce
s
s
ed
f
u
r
th
er
with
d
ata
b
alan
cin
g
.
T
h
e
d
ata
b
alan
cin
g
is
in
ten
d
ed
to
r
em
o
v
e
a
n
y
i
n
s
ig
n
if
ican
t
im
a
g
e.
T
h
e
e
x
am
p
le
is
illu
s
tr
ated
in
Fig
u
r
e
9
with
th
e
m
ask
im
a
g
e
th
at
h
as
s
ig
n
if
ican
t
in
f
o
r
m
atio
n
ex
p
lain
ed
in
Fig
u
r
e
9
(
a)
an
d
th
e
n
o
n
-
s
ig
n
if
ican
t im
ag
e
in
Fig
u
r
e
9
(
b
)
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
2
5
8
6
I
AE
S
I
n
t
J
R
o
b
&
A
u
to
m
,
Vo
l
.
14
,
No
.
2
,
J
u
n
e
20
2
5
:
2
60
-
2
72
266
(
a)
(
b
)
Fig
u
r
e
7
.
E
x
am
p
le
im
a
g
e
o
f
D
ee
p
Glo
b
e
d
ataset
(
a)
r
ea
l im
ag
e
an
d
(
b
)
m
ask
im
a
g
e
Fig
u
r
e
8
.
Pre
p
r
o
ce
s
s
ed
im
ag
es
T
h
e
f
ir
s
t
im
ag
e
is
s
ig
n
if
ican
tl
y
u
s
ef
u
l
b
ec
au
s
e
it
c
o
n
tain
s
th
e
r
o
ad
th
at
th
e
m
o
d
el
n
ee
d
s
to
lear
n
,
a
n
d
th
e
later
im
ag
e
d
o
es
n
o
t
h
av
e
an
y
;
th
u
s
,
it
n
ee
d
s
to
b
e
r
em
o
v
ed
.
No
t
ev
e
r
y
th
in
g
b
u
t
o
n
ly
p
ar
ts
o
f
it;
th
u
s
,
we
d
o
d
ata
b
alan
cin
g
b
y
r
em
o
v
in
g
im
ag
es
th
at
o
n
ly
h
av
e
0
.
0
1
%
r
o
ad
p
i
x
els
an
d
b
elo
w.
T
h
e
d
is
tr
ib
u
tio
n
o
f
th
e
d
ata
b
ef
o
r
e
an
d
af
te
r
b
alan
ci
n
g
ca
n
b
e
illu
s
tr
ated
in
Fig
u
r
e
1
0
,
wh
ich
s
h
o
ws
th
e
im
ag
e
d
is
tr
ib
u
tio
n
b
ef
o
r
e
d
ata
b
alan
cin
g
in
Fig
u
r
e
1
0
(
a
)
an
d
af
ter
d
ata
b
alan
ci
n
g
in
Fi
g
u
r
e
1
0
(
b
)
.
(
a)
(
b
)
Fig
u
r
e
9
.
E
x
am
p
le
k
in
d
s
o
f
im
ag
e
with
(
a)
in
f
o
r
m
atio
n
(
s
ig
n
i
f
ican
t)
an
d
(
b
)
n
o
i
n
f
o
r
m
atio
n
(
in
s
ig
n
if
ican
t)
(
a)
(
b
)
Fig
u
r
e
1
0
.
I
m
ag
e
d
is
tr
ib
u
tio
n
(
a)
b
ef
o
r
e
an
d
(
b
)
af
ter
d
ata
b
a
lan
cin
g
Evaluation Warning : The document was created with Spire.PDF for Python.
I
AE
S
I
n
t
J
R
o
b
&
A
u
to
m
I
SS
N:
2722
-
2
5
8
6
C
P
_
S
DUNet:
r
o
a
d
ex
tr
a
ctio
n
u
s
in
g
S
DUNet a
n
d
ce
n
terl
in
e
p
r
eser
vin
g
d
ice
lo
s
s
(
B
a
yu
S
a
t
r
ia
P
ers
a
d
a
)
267
T
h
e
n
ex
t step
in
v
o
lv
es d
ata
au
g
m
en
tatio
n
to
in
cr
ea
s
e
th
e
d
at
a
v
o
lu
m
e.
T
h
e
au
g
m
en
tatio
n
t
h
at
will b
e
u
s
ed
is
r
an
d
o
m
r
o
tatio
n
an
d
f
lip
p
in
g
.
Af
ter
t
h
at,
th
e
d
ataset
co
u
ld
b
e
u
s
ed
f
o
r
t
r
ain
in
g
b
u
t
will
b
e
u
s
ed
d
if
f
er
en
tly
f
o
r
ea
ch
e
x
p
er
im
e
n
t.
T
h
e
d
ataset
will
also
v
ar
y
b
etwe
en
h
alf
an
d
t
h
e
wh
o
le
o
f
th
e
r
ea
l
d
ataset.
Ad
d
itio
n
ally
,
th
e
d
ataset
will b
e
d
iv
id
e
d
in
to
tr
ain
i
n
g
,
v
alid
atio
n
,
an
d
test
in
g
with
an
8
0
:1
0
:1
0
p
r
o
p
o
r
tio
n
.
2
.
4
.
2
.
E
x
perim
ent
a
l
s
et
t
ing
s
a
nd
ev
a
lua
t
io
n
W
e
p
er
f
o
r
m
th
e
p
r
o
p
o
s
ed
m
o
d
el
u
s
in
g
th
e
T
en
s
o
r
Flo
w
f
r
am
ewo
r
k
.
T
h
e
h
y
p
er
p
ar
am
et
er
s
f
o
r
th
e
ex
p
er
im
en
t
a
r
e
as
f
o
llo
ws:
T
h
e
lear
n
in
g
r
ate
f
o
r
tr
ain
in
g
is
0
.
0
0
1
,
an
d
we
also
im
p
lem
en
ted
a
lear
n
in
g
r
ate
s
ch
ed
u
ler
u
s
in
g
R
ed
u
ce
L
R
On
Plateau
with
p
atien
ce
o
f
4
an
d
a
r
ed
u
ce
d
f
ac
t
o
r
o
f
0
.
1
.
T
h
e
m
o
d
el
was
tr
ain
e
d
f
o
r
4
0
e
p
o
ch
s
,
u
s
in
g
Ad
am
as
its
o
p
tim
izer
with
o
u
t
weig
h
t
d
ec
ay
.
T
h
e
p
r
o
p
o
r
tio
n
f
o
r
t
h
e
C
P_
clDice
lo
s
s
th
at
was
u
s
ed
was
9
0
%
s
o
f
t
-
d
ice
an
d
1
0
%
clDice
lo
s
s
.
Fo
r
all
ex
p
er
im
en
ts
,
m
etr
ics
th
at
ar
e
b
ein
g
u
s
ed
to
m
o
n
ito
r
in
th
e
tr
ain
in
g
p
h
ase
ar
e
as
in
(
6
)
to
(
8
)
.
=
2
×
(
+
)
+
(
+
)
(
6
)
c
=
2
×
(
,
)
×
(
,
)
(
,
)
+
(
,
)
(
7
)
=
2
×
+
+
(
8
)
T
P,
FP
,
T
N,
a
n
d
FN
ar
e
s
eq
u
en
tially
th
e
tr
u
e
p
o
s
itiv
es,
f
alse
p
o
s
itiv
es,
tr
u
e
n
eg
ativ
es,
an
d
f
alse
n
e
g
ativ
es
b
ased
o
n
th
e
m
o
d
el
p
r
ed
ictio
n
an
d
th
e
g
r
o
u
n
d
t
r
u
th
/m
ask
.
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
3
.
1
.
CP
_
SDUNet
perf
o
rm
a
n
ce
T
h
e
m
ain
ex
p
er
im
e
n
t
is
in
ten
d
ed
to
ev
alu
ate
SDUNet
tr
ain
ed
b
y
v
ar
i
o
u
s
lo
s
s
f
u
n
ctio
n
s
f
o
r
r
o
ad
ex
tr
ac
tio
n
,
o
n
e
o
f
wh
ich
is
th
e
p
r
o
p
o
s
ed
m
et
h
o
d
C
P_
SDUNet.
T
h
is
ex
p
er
im
en
t
is
im
p
lem
en
ted
u
s
in
g
256
×
2
5
6
p
ix
els im
ag
es a
n
d
5
0
% o
f
th
e
ac
tu
al
d
ata
f
o
r
tr
ain
in
g
th
e
m
o
d
el.
Du
e
to
th
e
s
ize
o
f
th
e
d
ata,
we
u
s
ed
h
alf
o
f
t
h
e
ac
tu
al
d
ata
in
th
is
ex
p
er
im
en
t.
T
h
e
r
esu
lt f
r
o
m
th
e
ex
p
er
im
en
t is sh
o
wn
in
T
ab
le
1
.
T
ab
le
1
.
T
h
e
tr
ain
in
g
,
v
alid
ati
o
n
,
an
d
test
in
g
r
esu
lt
o
f
SDUNet
f
o
r
its
lo
s
s
f
u
n
ctio
n
v
a
r
iatio
n
s
V
a
l
u
e
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
Te
st
i
n
g
S
D
U
N
e
t
u
si
n
g
B
C
E
Lo
ss
S
D
U
N
e
t
u
s
i
n
g
C
P
_
c
l
D
i
c
e
Lo
s
s
(
C
P
_
S
D
U
N
e
t
)
S
D
U
N
e
t
u
si
n
g
B
C
E
Lo
ss
S
D
U
N
e
t
u
s
i
n
g
C
P
_
c
l
D
i
c
e
Lo
s
s
(
C
P
_
S
D
U
N
e
t
)
S
D
U
N
e
t
u
si
n
g
B
C
E
Lo
ss
S
D
U
N
e
t
u
s
i
n
g
C
P
_
c
l
D
i
c
e
Lo
s
s
(
C
P
_
S
D
U
N
e
t
)
I
o
U
0
.
5
3
0
3
0
.
6
1
0
.
4
4
2
2
0
.
5
7
8
1
0
.
5
2
0
.
5
9
3
c
l
D
i
c
e
0
.
7
7
5
0
0
.
8
2
9
3
0
.
6
9
6
4
0
.
8
0
9
6
0
.
3
0
8
0
.
8
8
Lo
ss
0
.
0
5
8
1
0
.
2
2
8
3
0
.
1
2
2
5
0
.
2
5
2
2
-
T
h
e
r
esu
lt
g
iv
en
i
n
T
ab
le
1
s
h
o
ws
th
at
SDUNet
u
s
in
g
B
C
E
lo
s
s
p
er
f
o
r
m
ed
wo
r
s
e
co
m
p
ar
ed
to
SDUNet
u
s
in
g
C
P_
clDice
lo
s
s
f
o
r
its
tr
ain
in
g
,
v
alid
atio
n
,
a
n
d
test
in
g
r
esu
lts
.
W
ith
th
e
C
P_
clDice
lo
s
s
,
th
e
m
o
d
el
co
u
ld
ac
h
iev
e
b
etter
p
er
f
o
r
m
a
n
ce
with
an
in
ter
est
o
v
er
u
n
io
n
(
I
OU)
s
co
r
e
o
f
0
.
6
1
co
m
p
a
r
ed
t
o
B
C
E
L
o
s
s
with
an
I
o
U
s
co
r
e
o
f
0
.
5
3
.
A
0
.
0
8
d
if
f
er
en
ce
r
esu
lted
i
n
a
v
er
y
d
is
tin
ct
o
u
tco
m
e.
T
h
e
v
alid
atio
n
p
h
ase
also
h
ad
a
b
i
g
d
if
f
er
en
ce
,
with
th
e
C
P_
clDice
lo
s
s
ac
h
iev
in
g
an
I
o
U
s
co
r
e
o
f
0
.
5
7
8
1
,
an
d
th
e
B
C
E
lo
s
s
o
b
t
ain
in
g
a
n
I
o
U
s
co
r
e
o
f
0
.
4
4
.
A
wh
o
le
0
.
1
3
d
if
f
er
en
ce
b
etw
ee
n
th
e
two
.
Hen
ce
,
th
e
test
in
g
r
esu
lt
th
at
was
s
h
o
wn
in
T
ab
le
1
f
u
r
th
er
co
n
f
ir
m
ed
th
at
SDUNet
u
s
in
g
C
P_
cl
Dice
(
C
P_
SDU
Net)
lo
s
s
was
ab
le
to
o
u
tp
er
f
o
r
m
S
DUNe
t
u
s
in
g
B
C
E
L
o
s
s
with
an
I
o
U
s
co
r
e
o
f
0
.
5
9
f
o
r
C
P_
clDice
lo
s
s
an
d
a
n
I
o
U
s
co
r
e
o
f
0
.
5
2
f
o
r
B
C
E
lo
s
s
,
a
d
is
tin
ct
0
.
7
d
if
f
e
r
en
ce
.
I
n
ter
esti
n
g
ly
,
ev
en
th
o
u
g
h
B
C
E
lo
s
s
g
en
er
ally
p
r
o
d
u
ce
s
a
g
o
o
d
I
o
U
Sco
r
e,
S
DUNe
t
with
B
C
E
lo
s
s
o
b
tain
e
d
a
v
er
y
lo
w
clDice
s
co
r
e,
wh
ich
m
ea
n
s
th
at
th
e
m
o
d
el
co
u
ld
c
r
ea
te
ex
tr
ac
ti
o
n
co
r
r
ec
tly
b
u
t
d
id
n
o
t
g
iv
e
a
v
er
y
g
o
o
d
ce
n
ter
lin
e
f
o
r
its
r
o
a
d
.
T
h
e
r
esu
lt
o
f
th
e
two
m
o
d
els
is
s
h
o
wn
in
Fig
u
r
e
1
1
,
wh
ich
s
h
o
ws
th
e
r
esu
lt
o
f
th
e
p
r
ed
icted
r
o
a
d
wh
en
SDUNet
is
tr
ain
ed
u
s
in
g
clDice
lo
s
s
in
Fig
u
r
e
1
1
(
a)
an
d
t
h
e
r
esu
lt
o
f
th
e
p
r
ed
ic
ted
r
o
a
d
wh
e
n
SDUn
et
is
t
r
ain
ed
u
s
in
g
B
C
E
L
o
s
s
in
Fig
u
r
e
1
1
(
b
)
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
2
5
8
6
I
AE
S
I
n
t
J
R
o
b
&
A
u
to
m
,
Vo
l
.
14
,
No
.
2
,
J
u
n
e
20
2
5
:
2
60
-
2
72
268
(
a)
(
b
)
Fig
u
r
e
1
1
.
R
esu
lt o
f
th
e
p
r
e
d
ic
ted
r
o
ad
f
r
o
m
test
in
g
d
ata
(
a)
SDUNet
u
s
in
g
clDice
an
d
(
b
)
SDUNet
u
s
in
g
B
C
E
L
o
s
s
Nex
t,
tr
ain
in
g
with
o
n
ly
h
alf
o
f
th
e
d
ataset
an
d
a
wh
o
le
d
ataset
wo
u
ld
r
esu
lt
in
a
d
if
f
er
en
t
o
u
tco
m
e.
Ho
wev
er
,
th
e
p
r
o
b
lem
with
u
s
in
g
th
e
wh
o
le
d
ataset
is
th
at
th
e
d
ata
n
ee
d
s
to
b
e
tr
an
s
f
o
r
m
ed
in
to
a
n
I
m
a
g
e
with
a
s
ize
o
f
1
2
8
×
1
2
8
p
i
x
els
s
o
th
at
it
d
o
es
n
o
t
e
x
ce
ed
th
e
m
em
o
r
y
lim
it,
wh
ich
m
ay
le
ad
to
th
e
lo
s
s
o
f
its
d
etails.
Hen
ce
,
we
co
n
d
u
cted
an
ex
p
er
im
e
n
t u
s
in
g
1
2
8
×
1
2
8
p
ix
els in
p
u
t w
ith
th
e
wh
o
le
d
a
taset a
s
it
s
in
p
u
t.
T
h
e
ex
p
e
r
im
en
t
was
in
ten
d
e
d
to
test
C
P_
SDUNet
u
s
in
g
th
e
wh
o
le
o
f
th
e
Dee
p
Glo
b
e
R
o
ad
E
x
tr
ac
tio
n
d
ataset
b
y
r
ed
u
cin
g
th
e
in
p
u
t
im
ag
e
s
ize
to
1
2
8
×1
2
8
p
ix
els.
T
h
u
s
,
th
e
r
esu
lt
i
s
s
h
o
wn
in
T
ab
le
2
an
d
Fig
u
r
e
1
2
.
T
ab
le
2
.
T
h
e
tr
ain
in
g
,
v
alid
ati
o
n
an
d
test
in
g
r
esu
lt
o
f
C
P_
SDUNet
u
s
in
g
im
ag
e
s
ize
o
f
1
2
8
×
128
p
ix
els
V
a
l
u
e
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
Te
st
i
n
g
c
l
D
i
c
e
0
.
8
0
4
9
0
.
7
7
6
7
0
.
7
2
9
I
o
U
0
.
6
0
4
1
0
.
5
5
0
2
0
.
5
7
8
Lo
ss
0
.
2
4
3
3
0
.
2
8
4
1
-
Fig
u
r
e
1
2
.
T
esti
n
g
d
ata
r
esu
lt
u
s
in
g
SDUNet
with
im
ag
e
s
ize
o
f
1
2
8
×
1
2
8
p
i
x
els
Evaluation Warning : The document was created with Spire.PDF for Python.
I
AE
S
I
n
t
J
R
o
b
&
A
u
to
m
I
SS
N:
2722
-
2
5
8
6
C
P
_
S
DUNet:
r
o
a
d
ex
tr
a
ctio
n
u
s
in
g
S
DUNet a
n
d
ce
n
terl
in
e
p
r
eser
vin
g
d
ice
lo
s
s
(
B
a
yu
S
a
t
r
ia
P
ers
a
d
a
)
269
Fro
m
wh
at
we
ac
h
iev
ed
,
t
h
e
r
esu
lt
s
h
o
wed
th
at
u
s
in
g
th
e
w
h
o
le
d
ataset
d
o
es
g
iv
e
a
b
etter
r
esu
lt,
b
u
t
n
o
t
th
at
s
ig
n
if
ican
t.
T
h
e
test
in
g
r
esu
lt
s
co
r
e
f
o
r
clDice
d
e
cr
ea
s
es
f
r
o
m
0
.
7
3
8
to
0
.
7
2
9
wh
ile
th
e
I
o
U
s
co
r
e
in
cr
ea
s
es
to
0
.
5
7
8
.
Fro
m
t
h
e
t
esti
n
g
r
esu
lts
,
we
co
u
ld
also
s
ee
th
at
th
e
m
o
d
el
h
as
s
u
cc
ess
f
u
lly
ex
tr
ac
ted
th
e
r
o
ad
s
,
b
u
t
it
is
s
till
co
n
f
u
s
in
g
s
in
ce
th
e
im
ag
e
b
ec
am
e
m
u
ch
s
m
aller
.
Ma
n
y
p
i
x
els
th
at
wer
e
s
u
p
p
o
s
ed
t
o
s
u
p
p
o
r
t
th
e
lear
n
in
g
wer
e
eit
h
er
b
lu
r
r
ed
o
r
r
em
o
v
ed
b
ec
au
s
e
o
f
th
e
r
esizin
g
.
T
h
u
s
,
u
s
in
g
th
e
wh
o
le
d
ataset
wh
ile
lo
wer
in
g
th
e
r
eso
lu
tio
n
l
ea
d
s
to
s
u
b
o
p
tim
al
c
h
an
g
es.
3
.
2
.
J
o
int
lo
s
s
v
a
ria
t
i
o
n per
f
o
rm
a
nce
W
e
h
av
e
test
ed
C
P_
SD
UNe
t
u
s
in
g
th
e
co
m
b
in
atio
n
o
f
9
0
%
Dice
lo
s
s
an
d
1
0
% c
lDice
lo
s
s
as
its
lo
s
s
f
u
n
ctio
n
co
n
f
ig
u
r
atio
n
.
Her
e,
we
wan
t
to
ex
p
er
im
en
t
with
o
th
er
co
m
b
i
n
atio
n
s
to
f
in
d
th
e
b
est
co
m
b
in
atio
n
s
f
o
r
r
o
ad
ex
t
r
ac
tio
n
u
s
in
g
C
P_
clDice
.
3
.
2
.
1
.
8
0
%
Dice
L
o
s
s
a
nd
2
0
%
clDice
lo
s
s
T
h
e
in
p
u
t
im
ag
e
is
s
et
to
b
e
1
2
8
×
1
2
8
p
ix
els.
T
h
e
b
atch
s
ize
an
d
o
th
er
h
y
p
er
p
ar
am
eter
s
ar
e
th
e
s
am
e
co
n
f
ig
u
r
atio
n
as th
e
p
r
ev
io
u
s
ex
p
er
im
en
ts
.
T
h
e
r
esu
lt is
s
h
o
wn
in
T
ab
le
3
an
d
Fig
u
r
e
1
3
.
T
ab
le
3
.
T
r
ai
n
in
g
,
v
alid
atio
n
a
n
d
test
in
g
r
esu
lt
f
o
r
SDUNet
u
s
in
g
8
0
%
d
ice
lo
s
s
an
d
2
0
% c
l
Dice
lo
s
s
V
a
l
u
e
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
Te
st
i
n
g
c
l
D
i
c
e
0
.
8
0
6
8
0
.
7
9
4
3
0
.
8
5
I
o
U
0
.
5
9
2
3
0
.
5
5
8
5
0
.
6
5
Lo
ss
0
.
2
4
3
9
0
.
2
6
8
3
-
Fig
u
r
e
1
3
.
T
esti
n
g
d
ata
r
esu
lt
f
o
r
C
P_
SDUNet
u
s
in
g
th
e
co
m
b
in
atio
n
o
f
8
0
%
d
ice
lo
s
s
an
d
2
0
% c
lDice
lo
s
s
T
h
e
r
esu
lts
wer
e
b
etter
th
a
n
u
s
in
g
th
e
1
0
%
co
m
b
i
n
atio
n
wh
er
e
th
e
test
in
g
s
co
r
e
o
f
cl
Dice
is
0
.
8
5
,
an
d
th
e
I
o
U
is
0
.
6
5
,
wh
ic
h
i
s
th
e
b
est
I
o
U
s
o
f
a
r
,
m
ea
n
in
g
th
e
m
o
d
el
co
u
ld
p
r
ed
ict
r
o
ad
s
well.
Fro
m
th
e
test
in
g
r
esu
lt
im
ag
es,
we
co
u
ld
also
s
ee
th
at
it
is
alm
o
s
t
p
er
f
ec
t
o
n
th
e
s
ec
o
n
d
im
ag
e,
wh
ile
f
o
r
t
h
e
f
i
r
s
t
im
ag
e,
th
e
m
o
d
el
s
till
co
u
ld
n
o
t p
er
f
ec
tly
e
x
tr
ac
t r
o
a
d
s
.
T
h
is
is
d
u
e
to
th
e
s
id
e
ef
f
ec
t o
f
r
esizin
g
in
p
u
t im
ag
es.
3
.
2
.
2
.
7
0
%
Dice
L
o
s
s
a
nd
3
0
%
clDice
lo
s
s
Kn
o
win
g
th
at
th
e
co
m
b
in
atio
n
o
f
2
0
%
clDice
lo
s
s
an
d
8
0
%
d
ice
lo
s
s
g
av
e
a
b
etter
r
esu
lt,
we
tr
ied
to
ev
en
f
u
r
th
er
in
cr
ea
s
e
th
e
p
r
o
p
o
r
tio
n
f
o
r
clDice
lo
s
s
,
with
3
0
%
an
d
7
0
%
b
ein
g
d
ice
lo
s
s
.
T
h
e
r
esu
lt
o
f
th
is
ex
p
er
im
en
t is d
escr
ib
e
d
in
T
a
b
le
4
an
d
Fig
u
r
e
1
4
.
T
h
e
r
esu
lt
in
d
icate
s
th
at
in
cr
e
asin
g
clDice
lo
s
s
p
r
o
p
o
r
tio
n
t
o
3
0
%
le
d
to
a
p
o
o
r
p
er
f
o
r
m
an
ce
.
T
h
e
m
o
d
el
s
tar
ted
to
lear
n
o
n
ly
o
n
e
p
ar
t:
th
at
is
th
e
r
o
ad
.
T
h
e
r
esu
lt
m
ay
b
e
d
u
e
to
th
e
d
if
f
ic
u
lty
o
f
lear
n
in
g
th
e
r
o
ad
’
s
lo
ca
tio
n
b
ec
au
s
e
o
f
clD
ice
lo
s
s
.
clDice
lo
s
s
u
s
es
m
o
r
p
h
o
lo
g
y
to
m
a
k
e
s
k
eleto
n
izatio
n
,
an
d
b
ec
au
s
e
th
e
f
ac
to
r
h
as
b
ec
o
m
e
g
r
ea
ter
,
it
i
s
h
ar
d
er
f
o
r
th
e
m
o
d
el
to
d
is
ce
r
n
wh
ich
p
a
r
t
is
th
e
r
o
ad
.
T
h
u
s
,
th
e
ex
p
er
im
e
n
t
s
h
o
wed
th
at
in
cr
ea
s
in
g
alp
h
a
f
o
r
clDice
is
n
o
t
alwa
y
s
g
o
o
d
f
o
r
th
e
m
o
d
el,
a
n
d
th
e
b
est
f
it
f
o
r
SDUNet
is
th
e
co
m
b
in
atio
n
o
f
2
0
% c
lDice
lo
s
s
an
d
0
.
8
d
ice
l
o
s
s
.
T
ab
le
4
.
T
r
ai
n
in
g
,
v
alid
atio
n
a
n
d
test
in
g
r
esu
lt
f
o
r
SDUNet
u
s
in
g
7
0
%
d
ice
lo
s
s
an
d
3
0
% c
l
Dice
lo
s
s
V
a
l
u
e
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
Te
st
i
n
g
c
l
D
i
c
e
0
.
0
6
8
1
0
.
0
6
9
0
0
.
0
8
I
o
U
0
.
0
3
5
5
0
.
0
3
5
7
0
.
0
3
Lo
ss
0
.
9
3
1
7
0
.
9
3
1
2
-
Evaluation Warning : The document was created with Spire.PDF for Python.