I
AE
S In
t
er
na
t
io
na
l J
o
urna
l o
f
Art
if
icia
l In
t
ellig
ence
(
I
J
-
AI
)
Vo
l.
1
5
,
No
.
3
,
J
u
n
e
2
0
2
6
,
p
p
.
2774
~
2
7
8
5
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
15
.i
3
.
p
p
2
7
7
4
-
2
7
8
5
2774
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
RG
B
-
D
sa
lient
ob
ject dete
ction
wit
h
lo
ca
l f
ea
tur
e an
d sema
ntic
seg
menta
tion
Z
ha
ng
Wa
ng
1
,
K
im
O
n
Chin
1
,
Ra
y
ner
Alf
re
d
1
,
J
un
y
i C
ha
i
2
,
Rund
o
ng
Z
ha
ng
2
,
So
o
See
Cha
i
3
1
F
a
c
u
l
t
y
o
f
C
o
mp
u
t
i
n
g
a
n
d
I
n
f
o
r
mat
i
c
s
,
U
n
i
v
e
r
si
t
i
M
a
l
a
y
s
i
a
S
a
b
a
h
,
K
o
t
a
K
i
n
a
b
a
l
u
,
M
a
l
a
y
si
a
2
C
o
l
l
e
g
e
o
f
S
c
i
e
n
c
e
a
n
d
T
e
c
h
n
o
l
o
g
y
,
N
i
n
g
b
o
U
n
i
v
e
r
s
i
t
y
,
N
i
n
g
b
o
,
C
h
i
n
a
3
F
a
c
u
l
t
y
o
f
C
o
mp
u
t
e
r
S
c
i
e
n
c
e
a
n
d
I
n
f
o
r
mat
i
o
n
Te
c
h
n
o
l
o
g
y
,
U
n
i
v
e
r
si
t
i
M
a
l
a
y
s
i
a
S
a
r
a
w
a
k
,
S
a
r
a
w
a
k
,
M
a
l
a
y
s
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
J
u
n
1
4
,
2
0
2
5
R
ev
is
ed
Ma
r
2
3
,
2
0
2
6
Acc
ep
ted
Ma
y
7
,
2
0
2
6
Re
d
,
g
re
e
n
,
b
lu
e
–
d
e
p
t
h
(
RG
B
-
D
)
sa
li
e
n
t
o
b
jec
t
d
e
tec
ti
o
n
(S
OD
)
f
o
c
u
se
s
o
n
id
e
n
ti
f
y
in
g
v
isu
a
ll
y
p
r
o
m
in
e
n
t
o
b
jec
ts
b
y
sim
u
lati
n
g
h
u
m
a
n
v
isu
a
l
p
e
rc
e
p
ti
o
n
.
Wh
il
e
e
x
isti
n
g
RG
B
-
D
S
OD
m
e
th
o
d
s
h
a
v
e
d
e
m
o
n
stra
ted
re
su
lt
s,
th
e
re
re
m
a
in
c
h
a
ll
e
n
g
e
s
i
n
e
f
fe
c
ti
v
e
ly
lev
e
ra
g
i
n
g
e
x
tri
n
sic
c
u
e
s
a
n
d
e
n
h
a
n
c
in
g
fe
a
t
u
re
re
p
re
se
n
tati
o
n
.
To
a
d
d
re
ss
t
h
e
se
l
i
m
it
a
ti
o
n
s,
n
o
v
e
l
RG
B
-
D
S
OD
m
o
d
e
l
wit
h
lo
c
a
l
fe
a
tu
re
e
x
trac
ti
o
n
a
n
d
se
m
a
n
ti
c
se
g
m
e
n
tati
o
n
(L
F
S
S
)
is
in
tr
o
d
u
c
e
d
,
wh
ic
h
is
b
u
il
t
o
n
a
n
e
n
c
o
d
e
r
-
d
e
c
o
d
e
r
a
rc
h
it
e
c
tu
re
.
T
h
e
e
n
c
o
d
e
r
p
re
p
ro
c
e
ss
e
s
th
e
in
p
u
t
ima
g
e
s
b
y
m
e
rg
in
g
RG
B
a
n
d
d
e
p
th
d
a
ta
th
ro
u
g
h
a
c
h
a
n
n
e
l
a
n
d
sp
a
ti
a
l
a
tt
e
n
ti
o
n
(CS
A)
m
o
d
u
le.
A
lo
c
a
l
fe
a
tu
re
e
x
trac
ti
o
n
m
o
d
u
le
f
u
rth
e
r
re
fin
e
s
th
is
fu
si
o
n
.
T
h
e
d
e
c
o
d
e
r
c
o
n
sist
s
o
f
th
re
e
k
e
y
m
o
d
u
les
:
i
)
t
h
e
m
u
lt
i
-
fe
a
tu
r
e
e
x
trac
ti
o
n
(M
F
E)
m
o
d
u
le
e
n
h
a
n
c
e
s
b
a
se
fe
a
tu
re
s
th
ro
u
g
h
d
iv
e
rse
c
o
n
v
o
l
u
ti
o
n
a
l
o
p
e
ra
ti
o
n
s;
ii
)
th
e
se
m
a
n
ti
c
se
g
m
e
n
tatio
n
e
n
h
a
n
c
e
m
e
n
t
(S
S
E)
m
o
d
u
le
o
p
ti
m
ize
s
fe
a
tu
re
s
v
ia
sp
a
ti
a
l
p
y
ra
m
id
p
o
o
li
n
g
a
n
d
a
tro
u
s
c
o
n
v
o
lu
ti
o
n
;
a
n
d
ii
i
)
t
h
e
lo
c
a
l/
g
lo
b
a
l
a
g
re
e
m
e
n
t
a
n
d
e
d
g
e
d
e
tec
ti
o
n
(LG
E)
m
o
d
u
l
e
th
a
t
e
n
a
b
les
m
u
lt
i
-
lev
e
l
fe
a
tu
re
in
tera
c
ti
o
n
a
n
d
e
d
g
e
d
e
tec
ti
o
n
.
Th
e
se
m
o
d
u
l
e
s
wo
rk
se
q
u
e
n
ti
a
ll
y
t
o
e
n
h
a
n
c
e
a
n
d
e
x
trac
t
sa
li
e
n
t
o
b
jec
ts.
L
F
S
S
is
e
v
a
lu
a
t
e
d
o
n
si
x
sta
n
d
a
rd
RG
B
-
D
S
O
D
d
a
tas
e
ts
(NJ
U2
K
,
NLP
R
,
S
TE
RE
,
L
F
S
D
,
SSD
,
S
IP
)
by
f
o
u
r
m
e
tri
c
s,
o
u
tp
e
rfo
rm
in
g
th
e
c
o
m
p
a
riso
n
m
o
d
e
ls
wit
h
u
p
to
1
.
2
%
F
-
m
e
a
su
re
imp
ro
v
e
m
e
n
t
.
LF
S
S
is
fo
u
n
d
t
o
b
e
a
v
e
rsa
ti
le m
o
d
e
l
,
o
ff
e
rin
g
v
a
lu
a
b
le a
p
p
li
c
a
ti
o
n
s i
n
e
n
g
in
e
e
rin
g
.
K
ey
w
o
r
d
s
:
Atten
tio
n
m
ec
h
an
is
m
I
m
ag
e
p
r
o
ce
s
s
in
g
L
o
ca
l f
ea
tu
r
e
e
x
tr
ac
tio
n
Mu
lti
-
s
ca
le
f
u
s
io
n
R
GB
-
D
s
alien
t o
b
ject
d
etec
tio
n
Sem
an
tic
s
eg
m
en
tatio
n
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
:
Kim
On
C
h
in
Facu
lty
o
f
C
o
m
p
u
tin
g
an
d
I
n
f
o
r
m
atics,
Un
iv
er
s
iti Ma
lay
s
ia
Sab
ah
Ko
ta
Kin
ab
alu
,
Ma
lay
s
ia
E
m
ail:
k
im
o
n
ch
i
n
@
u
m
s
.
ed
u
.
m
y
1.
I
NT
RO
D
UCT
I
O
N
Salien
t
o
b
ject
d
etec
tio
n
(
SO
D)
s
ee
k
s
to
s
im
u
late
b
io
lo
g
ic
al
v
is
io
n
to
ca
p
tu
r
e
th
e
m
o
s
t
s
ig
n
if
ican
t
r
eg
io
n
s
o
f
an
im
ag
e.
I
t
h
as
b
e
en
wid
ely
ap
p
lie
d
in
v
a
r
io
u
s
d
o
wn
s
tr
ea
m
task
s
,
i
n
clu
d
in
g
o
b
ject
d
etec
tio
n
[
1
]
,
im
ag
e
class
if
icatio
n
[
2
]
,
s
em
an
tic
s
eg
m
en
tatio
n
[
3
]
,
co
-
s
alie
n
cy
d
etec
tio
n
[
4
]
,
a
n
d
o
b
ject
tr
ac
k
in
g
.
T
h
e
SOD
m
o
d
els g
en
er
ally
ac
co
m
p
lis
h
two
task
s
:
i
)
d
etec
tin
g
th
e
m
o
s
t
p
r
o
m
in
en
t o
b
jects a
n
d
ii
)
ac
cu
r
ately
s
eg
m
en
tin
g
th
e
r
eg
io
n
s
o
f
th
ese
o
b
jects.
T
o
im
p
r
o
v
e
th
e
ef
f
ec
tiv
en
ess
o
f
th
ese
task
s
,
b
o
th
in
tr
i
n
s
ic
an
d
ex
tr
in
s
ic
cu
es
ar
e
ty
p
ically
u
tili
ze
d
in
SOD.
Dep
th
m
ap
s
s
er
v
e
as
v
alu
ab
le
ex
tr
in
s
ic
c
u
es
th
at
ca
n
e
n
h
an
ce
th
e
p
er
f
o
r
m
an
ce
o
f
s
alien
c
y
m
o
d
els.
T
h
e
ad
v
en
t
o
f
d
e
p
th
ca
m
er
as
h
as
m
ad
e
it
ea
s
ier
to
ac
ce
s
s
d
ep
th
in
f
o
r
m
atio
n
f
r
o
m
a
s
ce
n
e
.
B
y
ex
tr
ac
tin
g
r
ed
,
g
r
ee
n
,
b
l
u
e
(
R
GB
)
,
an
d
d
ep
t
h
in
to
a
u
n
if
ie
d
r
ep
r
esen
tati
o
n
,
f
u
s
io
n
-
b
ased
ap
p
r
o
ac
h
es
em
p
o
wer
s
alien
cy
m
o
d
els
to
u
tili
ze
d
ep
th
cu
es,
m
u
lti
-
s
ca
le
f
ea
tu
r
e
f
u
s
io
n
,
in
t
eg
r
ity
,
an
d
co
n
tex
tu
al
in
f
o
r
m
atio
n
to
ac
cu
r
ately
id
en
tify
s
alien
t
o
b
jects.
Desp
ite
th
ese
ad
v
a
n
ce
m
en
ts
,
e
x
is
tin
g
r
e
d
,
g
r
ee
n
,
b
lu
e
–
d
ep
t
h
(
R
GB
-
D
)
m
o
d
els
[
5
]
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
R
GB
-
D
s
a
lien
t o
b
ject
d
etec
tio
n
w
ith
lo
ca
l fe
a
tu
r
e
a
n
d
s
ema
n
tic
s
eg
men
ta
tio
n
(
Zh
a
n
g
Wa
n
g
)
2775
th
at
in
co
r
p
o
r
ate
ex
tr
in
s
ic
cu
e
s
an
d
s
alien
cy
ex
tr
ac
tio
n
m
et
h
o
d
s
s
till
f
ac
e
s
ev
er
al
ch
allen
g
es.
T
h
e
f
u
s
io
n
o
f
d
ep
th
m
ap
s
with
R
GB
im
ag
es
r
em
ain
s
a
cr
itical
ar
ea
o
f
ex
p
lo
r
atio
n
.
E
a
r
ly
R
GB
-
D
SOD
m
o
d
els
s
im
p
l
y
co
n
ca
ten
ated
th
e
d
e
p
th
m
ap
with
th
e
R
GB
ch
an
n
els,
r
esu
ltin
g
in
a
cr
u
d
e
a
n
d
s
u
b
o
p
tim
al
ap
p
r
o
ac
h
.
R
ec
en
t
m
o
d
els
u
s
e
m
u
lti
-
s
ca
le,
m
u
lti
-
m
o
d
al
f
u
s
io
n
m
o
d
u
les
[
6
]
–
[
9
]
,
to
ac
h
iev
e
b
etter
in
te
g
r
atio
n
o
f
R
GB
an
d
d
ep
th
in
f
o
r
m
atio
n
.
Ho
wev
er
,
th
ese
m
eth
o
d
s
o
f
ten
f
ail
to
f
u
lly
e
x
p
lo
it
lo
ca
l
f
ea
tu
r
e
d
etails,
in
d
icatin
g
r
o
o
m
f
o
r
im
p
r
o
v
em
e
n
t
in
ef
f
ec
tiv
el
y
u
tili
zin
g
ex
tr
in
s
ic
c
u
es.
Fu
r
t
h
er
m
o
r
e
,
cu
r
r
en
t
SOD
m
eth
o
d
s
b
ased
o
n
d
ee
p
lear
n
in
g
ty
p
ically
ad
o
p
t
c
r
o
s
s
-
m
o
d
al
an
d
m
u
lti
-
s
ca
le
f
e
atu
r
e
ex
tr
ac
tio
n
,
d
ee
p
b
o
u
n
d
ar
y
-
b
ased
lear
n
in
g
m
ec
h
an
is
m
s
,
to
p
-
d
o
wn
m
o
d
el
in
g
,
a
n
d
c
o
n
tex
t
-
b
ased
m
o
d
el
in
g
.
Alth
o
u
g
h
th
ese
a
p
p
r
o
ac
h
es
ex
tr
ac
t
s
alien
cy
f
r
o
m
v
ar
io
u
s
p
e
r
s
p
ec
tiv
es,
th
e
y
o
f
ten
f
all
s
h
o
r
t
in
b
asic
f
ea
t
u
r
e
e
x
tr
ac
tio
n
,
s
em
an
tic
s
eg
m
en
tatio
n
,
a
n
d
ed
g
e
d
etec
tio
n
.
T
h
e
r
ef
o
r
e
,
th
er
e
is
s
till
s
ig
n
if
ican
t p
o
ten
tial f
o
r
e
n
h
an
cin
g
s
alien
cy
e
x
tr
ac
tio
n
m
eth
o
d
s
.
A
n
R
GB
-
D
SOD
b
ased
o
n
lo
ca
l
f
ea
tu
r
e
e
x
tr
ac
tio
n
an
d
s
em
an
tic
s
eg
m
en
tatio
n
(
L
FS
S)
is
p
r
o
p
o
s
ed
with
an
e
n
co
d
e
r
-
d
ec
o
d
er
s
tr
u
c
tu
r
e
t
o
tack
le
t
h
ese
is
s
u
es.
T
h
e
en
co
d
er
p
ar
t
u
tili
ze
s
p
y
r
am
i
d
v
is
io
n
tr
an
s
f
o
r
m
er
v
2
(
PVTv
2
)
[
1
0
]
to
s
er
v
e
as
th
e
b
ac
k
b
o
n
e
n
etwo
r
k
a
n
d
co
m
b
in
es
a
ch
an
n
el/s
p
atial
atten
tio
n
m
o
d
u
le
with
a
lo
ca
l f
ea
tu
r
e
ex
tr
ac
tio
n
m
o
d
u
l
e
to
en
h
an
ce
th
e
in
teg
r
atio
n
o
f
R
G
B
an
d
d
ep
th
in
f
o
r
m
atio
n
.
I
n
th
e
d
ec
o
d
e
r
p
ar
t
,
th
e
ex
tr
ac
tio
n
o
f
b
asic f
ea
tu
r
es
is
o
p
tim
ized
m
ain
ly
th
r
o
u
g
h
t
r
ip
le
co
n
v
o
lu
tio
n
.
T
h
e
o
u
tp
u
t
f
ea
tu
r
es a
r
e
f
u
r
th
er
en
h
an
ce
d
b
y
s
em
an
tic
s
eg
m
e
n
tatio
n
,
an
d
th
e
ef
f
ec
tiv
e
n
ess
o
f
l
o
ca
l/g
lo
b
al
a
g
r
ee
m
en
t
d
e
tectio
n
is
im
p
r
o
v
ed
th
r
o
u
g
h
ed
g
e
d
etec
tio
n
.
T
h
e
m
ain
co
n
tr
ib
u
tio
n
s
o
f
th
i
s
wo
r
k
ar
e
as
f
o
llo
ws:
f
ir
s
t
ly
,
an
en
co
d
e
r
is
d
esig
n
ed
th
at
in
co
r
p
o
r
ates
an
im
ag
e
p
r
e
p
r
o
ce
s
s
in
g
m
o
d
u
le
b
ased
o
n
t
h
e
PVTv
2
b
a
ck
b
o
n
e
,
in
teg
r
at
in
g
with
a
c
h
an
n
el
a
n
d
s
p
atial
atten
tio
n
(
C
SA)
m
ec
h
an
is
m
,
a
co
n
v
o
l
u
tio
n
n
e
x
t
ex
tr
em
e
(
C
o
n
v
NeX
t
)
-
b
ased
[
1
1
]
lo
ca
l
f
ea
tu
r
e
ex
tr
ac
tio
n
m
o
d
u
le,
an
d
a
b
s
o
lu
te
p
o
s
itio
n
e
n
co
d
in
g
(
APE)
to
en
ab
le
ef
f
ec
tiv
e
cr
o
s
s
-
m
o
d
al
f
u
s
io
n
o
f
R
GB
an
d
d
e
p
th
f
ea
tu
r
es.
T
h
en
,
a
s
em
an
tic
s
eg
m
en
tatio
n
e
n
h
an
ce
m
en
t
(
SS
E
)
m
o
d
u
le
is
d
ev
elo
p
ed
with
in
th
e
d
ec
o
d
er
th
at
s
p
atial
p
y
r
am
id
p
o
o
lin
g
[
1
2
]
is
o
p
tim
ized
b
y
atr
o
u
s
co
n
v
o
l
u
tio
n
to
e
n
h
an
ce
m
u
lti
-
s
ca
le
s
em
an
tic
u
n
d
er
s
tan
d
i
n
g
a
n
d
b
o
u
n
d
ar
y
r
ec
o
g
n
itio
n
.
Fin
ally
,
s
y
s
tem
ati
c
ex
p
er
im
e
n
ts
ar
e
c
o
n
d
u
cted
o
n
eig
h
t
s
tate
-
of
-
th
e
-
ar
t
(
SOTA
)
m
o
d
els
ac
r
o
s
s
s
ix
b
en
c
h
m
ar
k
d
atasets
,
p
r
o
v
in
g
th
at
th
e
p
r
o
p
o
s
ed
L
FS
S
f
r
am
ewo
r
k
ac
h
ie
v
es
ex
ce
llen
t p
er
f
o
r
m
an
ce
in
th
e
m
u
lti
-
s
ca
le
an
d
m
u
lti
-
m
o
d
al
f
u
s
io
n
o
f
th
e
R
GB
-
D
SOD
ta
s
k
s
.
2.
RE
L
AT
E
D
WO
RK
S
2
.
1
.
RG
B
-
D
s
a
lient
o
bje
ct
d
et
ec
t
io
n
m
o
del
T
r
ad
itio
n
al
R
GB
-
D
SOD
m
o
d
els
g
en
er
ally
u
tili
ze
d
ca
lcu
latio
n
s
b
ased
o
n
co
l
o
r
,
b
o
u
n
d
ar
y
,
co
n
tr
ast,
tex
tu
r
e,
an
d
p
r
io
r
k
n
o
wled
g
e.
Fo
r
in
s
tan
ce
,
Z
h
o
u
et
a
l.
[
1
3
]
d
eliv
er
ed
a
th
o
r
o
u
g
h
s
u
r
v
e
y
o
f
th
ese
tr
ad
itio
n
al
m
eth
o
d
s
,
d
etailin
g
r
eg
i
o
n
-
b
ased
,
co
n
tr
ast
-
b
ased
,
an
d
b
o
u
n
d
ar
y
-
b
ased
s
tr
ateg
ies
.
J
i
et
a
l
.
[
1
4
]
in
tr
o
d
u
ce
d
a
ca
lib
r
ated
R
GB
-
D
s
alien
cy
d
etec
tio
n
ap
p
r
o
ac
h
to
a
d
d
r
ess
in
ter
-
m
o
d
al
in
co
n
s
is
ten
cies
u
s
in
g
s
p
atial
p
r
io
r
s
.
C
h
en
et
a
l.
[
1
5
]
tr
ea
ted
d
ep
th
as
a
f
o
u
r
th
c
h
an
n
el
v
ia
a
3
D
c
o
n
v
o
lu
ti
o
n
al
n
e
u
r
al
n
etwo
r
k
(
3D
-
C
NN
)
,
en
ab
lin
g
p
r
e
-
f
u
s
io
n
o
f
R
GB
an
d
d
ep
th
cr
o
s
s
-
m
o
d
al
f
ea
tu
r
es
in
th
e
en
co
d
er
s
tag
e.
Su
n
et
a
l
.
[
1
6
]
p
r
o
p
o
s
ed
a
d
ep
th
-
s
en
s
itiv
e
d
etec
tio
n
m
o
d
u
le
co
m
b
in
ed
with
au
t
o
m
atic
m
u
lti
-
m
o
d
al
f
u
s
io
n
to
g
en
er
ate
r
ef
in
ed
s
alien
cy
m
ap
s
.
Alth
o
u
g
h
th
ese
ea
r
ly
m
eth
o
d
s
ac
h
iev
ed
in
itial su
cc
ess
,
th
eir
p
r
ec
is
io
n
a
n
d
r
o
b
u
s
tn
ess
r
em
ain
ed
lim
ited
.
E
ar
ly
R
GB
-
D
SOD
m
o
d
els
w
ith
d
ee
p
lear
n
in
g
m
o
v
ed
b
e
y
o
n
d
h
an
d
c
r
af
ted
f
ea
tu
r
es.
C
o
n
g
et
a
l.
[
1
7
]
p
r
esen
ted
C
I
R
-
Net,
wh
ich
in
teg
r
ates
cr
o
s
s
-
m
o
d
ality
in
ter
a
ctiv
e
m
o
d
u
les
f
o
r
a
d
ap
tiv
e
f
u
s
io
n
o
f
R
GB
an
d
d
ep
th
cu
es.
Z
h
a
n
g
et
a
l.
[
1
8
]
p
r
o
p
o
s
ed
B
T
S
-
Net,
a
b
i
-
d
ir
e
ctio
n
al
n
etwo
r
k
with
tr
an
s
f
e
r
an
d
s
elec
tio
n
th
at
p
u
r
if
ies
n
o
is
y
d
ep
th
in
f
o
r
m
a
tio
n
th
r
o
u
g
h
p
r
o
g
r
ess
iv
e
f
u
s
io
n
an
d
s
elec
tio
n
m
ec
h
an
is
m
s
.
J
ia
et
a
l.
[
1
9
]
in
tr
o
d
u
ce
d
SiaTr
an
s
,
a
Siam
ese
tr
an
s
f
o
r
m
er
n
etwo
r
k
in
c
o
r
p
o
r
ati
n
g
d
ep
t
h
-
im
ag
e
q
u
ali
ty
class
if
icatio
n
to
g
u
id
e
d
y
n
am
ic
c
r
o
s
s
-
m
o
d
al
f
u
s
io
n
.
Mo
s
t
r
ec
en
tly
,
Yi
et
a
l.
[
2
0
]
p
r
o
p
o
s
ed
GL
-
DM
Net,
wh
ich
u
s
es
a
d
u
al
m
u
tu
al
lear
n
in
g
f
r
am
ewo
r
k
w
ith
g
lo
b
al
-
p
a
r
tial
awa
r
en
ess
v
ia
p
o
s
itio
n
an
d
ch
a
n
n
el
m
u
tu
al
f
u
s
io
n
m
o
d
u
les,
ac
h
iev
in
g
SOTA
p
er
f
o
r
m
an
c
e.
Desp
ite
th
ese
ad
v
an
ce
s
,
k
ey
ch
allen
g
es
r
em
ain
u
n
r
eso
lv
ed
,
s
u
ch
as
n
o
is
y
d
ep
th
m
a
p
s
,
r
ea
l
-
tim
e
p
r
o
ce
s
s
in
g
d
em
a
n
d
s
,
an
d
p
o
o
r
g
e
n
er
al
izatio
n
ac
r
o
s
s
d
iv
er
s
e
s
ce
n
es.
2
.
2
.
RG
B
-
D
f
us
io
n m
et
ho
d
a
nd
lo
ca
l f
ea
t
ure
ex
t
ra
ct
io
n
R
GB
-
D
f
u
s
io
n
s
ch
em
es
ar
e
ab
le
to
b
e
g
en
e
r
ally
class
if
ied
in
to
ea
r
ly
f
u
s
io
n
,
late
f
u
s
io
n
,
an
d
m
u
lti
-
s
ca
le
f
u
s
io
n
.
E
ar
ly
f
u
s
io
n
co
n
ca
ten
ates
th
e
d
ep
th
ch
an
n
el
to
th
e
R
GB
ch
an
n
els
as
an
ad
d
itio
n
al
ch
an
n
el.
L
ate
f
u
s
io
n
u
s
es
a
m
u
lti
-
s
ca
le
m
o
d
el
b
ased
o
n
two
s
tr
ea
m
s
to
f
u
s
e
R
G
B
an
d
d
ep
th
.
Mu
lti
-
s
ca
le
f
u
s
io
n
m
o
d
els
ad
o
p
t
a
m
u
lti
-
l
ev
el
ap
p
r
o
ac
h
wh
e
r
e
ea
ch
le
v
el
o
u
tp
u
ts
a
f
ea
tu
r
e
m
ap
af
ter
f
u
s
io
n
,
an
d
th
e
f
in
al
s
alien
cy
is
g
en
er
ated
th
r
o
u
g
h
th
e
in
ter
ac
tio
n
ac
r
o
s
s
m
u
ltip
le
s
ca
les.
Du
e
to
th
e
lim
i
tatio
n
s
o
f
th
e
f
ir
s
t
two
f
u
s
io
n
m
eth
o
d
s
,
r
ec
en
t
m
o
d
els
ty
p
ically
u
s
e
v
a
r
i
o
u
s
m
u
lti
-
s
ca
le
f
u
s
io
n
tech
n
iq
u
es
to
im
p
r
o
v
e
s
alien
cy
ex
tr
ac
tio
n
.
I
n
ad
d
itio
n
to
th
e
f
u
s
io
n
m
e
th
o
d
,
t
h
e
C
o
n
v
NeX
t
lo
ca
l
f
e
atu
r
e
ex
tr
ac
tio
n
m
o
d
u
le
c
o
u
l
d
also
b
e
ad
ap
ted
t
o
f
u
r
th
er
im
p
r
o
v
e
th
e
f
u
s
io
n
ef
f
ec
t
o
f
R
GB
-
D
m
o
d
els.
C
o
n
v
NeX
t
ap
p
lies
th
e
lo
c
al
r
ec
ep
tiv
e
f
ield
t
o
s
elf
-
atten
tio
n
,
en
a
b
lin
g
t
h
e
m
o
d
el
to
b
etter
c
o
n
ce
n
tr
ate
o
n
l
o
ca
l
d
etails
wh
ile
s
till
co
n
s
id
e
r
in
g
g
lo
b
al
c
o
n
tex
t.
C
o
n
v
NeX
t
p
r
o
v
id
es
b
o
th
th
e
g
lo
b
al
f
ea
tu
r
e
ex
tr
ac
tio
n
c
ap
ab
ilit
y
o
f
T
r
an
s
f
o
r
m
er
an
d
th
e
lo
ca
l
f
ea
tu
r
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
3
,
J
u
n
e
2
0
2
6
:
2
7
7
4
-
2
7
8
5
2776
ca
p
tu
r
e
ca
p
ab
ilit
y
o
f
th
e
C
NN
.
I
ts
h
ier
ar
c
h
ical
s
tr
u
ctu
r
e
also
allo
ws
th
e
m
o
d
el
to
o
b
t
ain
in
f
o
r
m
atio
n
at
d
if
f
er
en
t le
v
els.
2
.
3
.
E
nh
a
ncem
ent
m
e
t
ho
d ba
s
ed
o
n m
ulti
-
s
ca
le
s
em
a
ntic
s
eg
m
ent
a
t
io
n
I
n
th
e
R
GB
-
D
SOD,
an
en
h
an
ce
m
en
t
s
ch
em
e
is
u
s
u
ally
ap
p
lied
af
ter
th
e
im
ag
e
p
r
e
p
r
o
ce
s
s
in
g
lay
er
to
f
u
r
th
er
e
x
p
lo
r
e
th
e
r
elatio
n
s
h
ip
s
b
etwe
en
f
ea
tu
r
es.
Sp
atial
p
y
r
am
id
p
o
o
lin
g
(
Dee
p
L
a
b
v
3
+)
b
ased
o
n
atr
o
u
s
co
n
v
o
l
u
tio
n
[
2
1
]
is
o
n
e
s
u
ch
en
h
an
ce
m
en
t
s
ch
em
e
u
ti
lized
in
s
em
an
tic
s
eg
m
en
ta
tio
n
.
I
n
s
em
an
tic
s
eg
m
en
tatio
n
task
s
,
p
y
r
a
m
id
p
o
o
li
n
g
an
d
en
c
o
d
er
-
d
ec
o
d
er
ar
c
h
itectu
r
es
ar
e
f
r
eq
u
en
tly
u
s
ed
,
a
n
d
Dee
p
L
ab
v
3
+
c
o
m
b
in
es
th
e
a
d
v
an
tag
es
o
f
b
o
th
.
Sin
ce
p
y
r
a
m
id
p
o
o
lin
g
ca
n
ca
u
s
e
th
e
lo
s
s
o
f
o
b
ject
b
o
u
n
d
a
r
y
d
etails,
atr
o
u
s
co
n
v
o
lu
ti
o
n
is
u
s
ed
to
co
m
p
e
n
s
ate
f
o
r
th
is
s
h
o
r
tco
m
in
g
.
T
h
e
e
f
f
icien
cy
a
n
d
ac
cu
r
ac
y
o
f
th
e
m
o
d
el
ar
e
f
u
r
th
e
r
im
p
r
o
v
ed
b
y
u
s
in
g
d
ep
t
h
wis
e
s
ep
ar
ab
le
co
n
v
o
lu
tio
n
.
Dee
p
L
ab
v
3
+
is
wid
ely
a
p
p
lied
in
m
an
y
m
o
d
els
d
u
e
to
its
ef
f
ec
tiv
en
ess
.
Fo
r
ex
am
p
le,
C
h
en
et
a
l.
[
2
2
]
p
r
esen
ted
a
d
en
s
e
r
esid
u
al
n
etwo
r
k
(
DR
Net)
b
ased
o
n
Dee
p
L
ab
v
3
+,
wh
ich
in
teg
r
ates
a
d
en
s
ely
co
n
n
ec
ted
C
NN
(
DC
NN)
an
d
a
r
esid
u
al
n
etwo
r
k
(
R
esNet)
f
o
r
co
n
s
tr
u
ctio
n
ex
t
r
ac
tio
n
f
r
o
m
r
em
o
te
s
en
s
in
g
im
ag
er
y
(
R
SI)
.
Mo
r
e
o
v
er
,
Du
et
a
l.
[
2
3
]
p
r
esen
ted
a
m
o
d
el
b
ased
o
n
Dee
p
L
ab
v
3
+
an
d
o
b
ject
-
b
ased
im
ag
e
an
aly
s
is
(
OB
I
A)
f
o
r
r
e
m
o
te
s
en
s
in
g
im
ag
es’
s
em
an
ti
c
s
eg
m
en
tatio
n
.
3.
M
E
T
H
O
D
3
.
1
.
O
v
er
v
iew
o
f
lo
ca
l
f
ea
t
u
re
ex
t
ra
ct
i
o
n a
nd
s
em
a
ntic
s
eg
m
ent
a
t
i
o
n
T
h
e
L
FS
S
is
p
r
o
p
o
s
ed
t
o
en
h
an
ce
th
e
ef
f
ec
ti
v
en
ess
o
f
th
e
R
GB
-
D
s
alien
cy
task
.
As
illu
s
tr
ated
in
Fig
u
r
e
1
,
L
FS
S
ad
o
p
ts
an
en
co
d
er
-
d
ec
o
d
er
f
r
am
ewo
r
k
.
T
h
e
en
co
d
er
p
r
e
p
r
o
ce
s
s
es
th
e
R
G
B
-
D
in
p
u
t
im
ag
es
u
s
in
g
a
d
u
al
PVTv
2
b
ac
k
b
o
n
e
n
etwo
r
k
with
APE.
I
t
u
s
es a
co
m
b
in
atio
n
o
f
a
C
SA m
o
d
u
le
with
a
lo
ca
l f
ea
tu
r
e
ex
tr
ac
tio
n
m
o
d
u
le
,
C
o
n
v
Nex
t
,
to
ac
h
iev
e
b
etter
R
GB
an
d
d
e
p
th
f
u
s
io
n
ef
f
ec
t
,
g
en
er
atin
g
m
u
lti
-
s
ca
le
R
GB
-
D
f
ea
tu
r
e
r
ep
r
esen
tatio
n
s
ac
r
o
s
s
f
o
u
r
le
v
els.
Fig
u
r
e
1
.
T
h
e
p
r
o
p
o
s
ed
L
FS
S f
r
am
ewo
r
k
is
b
ased
o
n
a
n
en
c
o
d
er
-
d
ec
o
d
er
m
o
d
el
T
h
e
d
ec
o
d
e
r
is
d
iv
id
ed
in
t
o
th
r
ee
m
o
d
u
les.
Firstl
y
,
th
e
m
u
lti
-
f
ea
tu
r
e
ex
tr
ac
tio
n
(
MFE
)
m
o
d
u
le
f
u
s
es
f
ea
tu
r
e
f
r
o
m
th
r
ee
d
is
tin
ct
co
n
v
o
lu
tio
n
al
k
e
r
n
els
to
ca
p
tu
r
e
d
iv
er
s
e
l
o
ca
l
p
atter
n
s
.
Nex
t,
th
e
SS
E
m
o
d
u
l
e
in
co
r
p
o
r
ates
atr
o
u
s
co
n
v
o
lu
ti
o
n
s
an
d
s
p
atial
p
y
r
am
i
d
p
o
o
l
in
g
to
r
ein
f
o
r
ce
g
lo
b
al
co
n
te
x
tu
al
u
n
d
e
r
s
tan
d
in
g
an
d
im
p
r
o
v
e
b
o
u
n
d
ar
y
r
ec
o
g
n
itio
n
.
Fin
ally
,
th
e
lo
ca
l/g
lo
b
al
ag
r
ee
m
en
t
with
th
e
ed
g
e
d
etec
tio
n
(
L
GE
)
m
o
d
u
le
p
r
o
m
o
tes
cr
o
s
s
-
s
ca
le
f
ea
tu
r
e
co
n
s
is
ten
cy
an
d
e
n
h
an
ce
s
co
n
to
u
r
p
r
ec
is
io
n
th
r
o
u
g
h
i
n
teg
r
ated
ed
g
e
d
etec
tio
n
m
ec
h
a
n
is
m
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
R
GB
-
D
s
a
lien
t o
b
ject
d
etec
tio
n
w
ith
lo
ca
l fe
a
tu
r
e
a
n
d
s
ema
n
tic
s
eg
men
ta
tio
n
(
Zh
a
n
g
Wa
n
g
)
2777
3
.
2
.
I
ma
g
e
prepro
ce
s
s
ing
T
h
e
im
ag
e
p
r
e
p
r
o
ce
s
s
in
g
m
o
d
u
le
u
s
es
a
d
u
al
PVTv
2
b
ac
k
b
o
n
e
n
etwo
r
k
with
p
r
e
-
tr
ain
ed
APE
weig
h
ts
f
o
r
b
ase
in
f
o
r
m
atio
n
o
n
th
e
R
GB
-
D
ch
an
n
els.
T
h
e
en
tire
f
u
s
io
n
p
r
o
ce
s
s
is
p
er
f
o
r
m
ed
ac
r
o
s
s
f
o
u
r
lev
el
s
.
T
h
e
e
x
p
er
im
e
n
ts
in
d
ic
ate
th
at
ad
d
in
g
an
ad
d
itio
n
al
C
o
n
v
NeX
t
m
o
d
u
le
f
o
r
lo
ca
l
f
ea
tu
r
e
ex
tr
ac
tio
n
im
p
r
o
v
es
R
GB
-
D
f
u
s
io
n
ac
c
u
r
ac
y
f
o
r
d
ep
th
m
a
p
s
.
Ho
w
ev
er
,
th
e
R
GB
m
ap
s
ar
e
n
o
t
p
r
o
ce
s
s
ed
b
y
th
e
C
o
n
v
NeX
t
m
o
d
u
le.
Dep
th
f
ea
tu
r
es,
p
r
o
ce
s
s
ed
b
y
PVTv
2
an
d
C
o
n
v
NeX
t
with
APE,
ar
e
p
ass
ed
in
to
th
e
C
SA
m
o
d
u
le,
an
d
th
e
r
esu
ltin
g
f
ea
tu
r
es
ar
e
co
m
b
in
ed
with
R
GB
f
ea
tu
r
es
p
r
o
ce
s
s
ed
b
y
PVTv
2
with
APE.
Fin
ally
,
th
e
co
m
b
i
n
ed
r
esu
lt
p
ass
es
th
r
o
u
g
h
th
e
C
o
n
v
NeX
t
m
o
d
u
le
f
o
r
an
o
th
e
r
r
o
u
n
d
o
f
lo
ca
l
f
e
atu
r
e
ex
tr
ac
tio
n
to
co
m
p
lete
th
e
R
GB
-
D
f
u
s
io
n
.
T
h
e
f
o
r
m
u
la
f
o
r
th
e
f
u
s
io
n
is
s
h
o
wn
as
(
1
)
.
=
(
(
+
(
)
)
)
(
1
)
W
h
er
e
an
d
ar
e
th
e
PVTv
2
+
APE
–
d
er
iv
ed
f
ea
t
u
r
es
f
o
r
th
e
R
GB
an
d
d
ep
th
s
tr
ea
m
s
(
th
e
latter
in
clu
d
in
g
a
n
in
itial C
o
n
v
NeX
t r
ef
in
em
en
t)
.
T
h
e
d
ep
th
‐
g
u
id
e
d
atten
tio
n
ter
m
is
(
2
)
an
d
(
3
)
.
(
)
=
(
(
)
)
(
2
)
(
)
=
(
(
(
)
)
)
⊙
，
(
)
=
(
(
(
)
)
)
⊙
(
3
)
W
h
er
e
an
d
ar
e
ch
an
n
el‐
wis
e
an
d
s
p
atial
m
a
x
‐
p
o
o
lin
g
,
(
·
)
is
a
s
h
ar
ed
m
u
lti‐lay
e
r
p
e
r
c
ep
tr
o
n
,
(
·
)
d
en
o
tes
a
3
×3
c
o
n
v
o
lu
tio
n
,
s
tan
d
s
th
e
s
ig
m
o
id
,
an
d
⊙
i
n
d
icate
s
elem
en
t‐
wis
e
m
u
ltip
licatio
n
.
C
o
llectin
g
{
}
=
1
4
y
ield
s
th
e
m
u
lti‐scale
f
u
s
ed
r
ep
r
esen
tatio
n
.
−
C
SA:
in
tr
ad
itio
n
al
R
GB
s
alie
n
cy
m
eth
o
d
s
,
a
well
-
ch
o
s
en
f
u
s
io
n
m
eth
o
d
ca
n
en
h
a
n
ce
r
el
ev
an
t
f
ea
tu
r
es
an
d
s
u
p
p
r
ess
n
o
is
e
wh
en
m
er
g
in
g
two
m
o
d
alities
.
As
s
h
o
wn
in
Fig
u
r
e
2
,
th
e
C
SA
m
o
d
u
l
e
ex
em
p
lifie
s
an
ef
f
ec
tiv
e
f
u
s
io
n
s
ch
em
e
b
y
u
s
in
g
b
o
th
C
SA
m
ec
h
an
is
m
s
.
T
h
e
C
SA
m
o
d
u
le
ac
h
iev
es
ef
f
icien
t
an
d
lig
h
tweig
h
t
R
GB
-
D
f
u
s
io
n
th
r
o
u
g
h
s
im
p
le
m
ax
-
p
o
o
lin
g
-
b
a
s
ed
ch
an
n
el
atten
tio
n
an
d
s
p
a
tial
atten
tio
n
,
co
m
b
in
ed
with
a
s
h
ar
e
d
M
L
P.
T
h
is
ap
p
r
o
ac
h
e
n
h
an
ce
s
th
e
f
u
s
io
n
o
f
m
u
lti
-
m
o
d
alit
y
in
f
o
r
m
atio
n
,
lead
in
g
to
im
p
r
o
v
e
d
s
alien
cy
d
etec
tio
n
.
−
C
o
n
v
NeX
t:
th
e
r
o
le
o
f
th
e
C
o
n
v
NeX
t
m
o
d
u
le
is
to
p
er
f
o
r
m
lo
ca
l
f
ea
tu
r
e
ex
tr
ac
tio
n
i
n
en
h
an
cin
g
th
e
R
GB
-
D
f
u
s
io
n
.
I
t
is
in
ef
f
ec
tiv
e
to
ap
p
ly
th
e
C
o
n
v
NeX
t
m
o
d
u
le
to
b
o
t
h
R
GB
an
d
d
ep
th
ch
an
n
els
s
im
u
ltan
eo
u
s
ly
.
T
h
er
e
f
o
r
e,
t
h
e
C
o
n
v
NeX
t
o
p
er
atio
n
is
ap
p
lied
o
n
ly
to
t
h
e
d
ep
th
m
ap
.
Af
ter
war
d
,
th
e
R
GB
-
D
d
u
al
-
s
tr
ea
m
f
ea
tu
r
es
ar
e
f
u
s
ed
by
th
e
C
SA
m
o
d
u
le
.
T
h
e
R
GB
f
ea
tu
r
e
m
ap
,
p
r
o
c
ess
ed
th
r
o
u
g
h
C
o
n
v
NeX
t,
is
th
en
co
n
ca
ten
ated
with
th
e
f
u
s
ed
f
ea
tu
r
es.
T
h
is
ap
p
r
o
ac
h
h
as
b
ee
n
d
e
m
o
n
s
tr
ated
t
o
im
p
r
o
v
e
t
h
e
f
in
al
s
alien
cy
ac
c
u
r
ac
y
.
−
APE:
in
th
e
PVTv
2
en
co
d
er
b
ac
k
b
o
n
e
n
etwo
r
k
,
th
e
im
ag
e
is
p
ar
titi
o
n
ed
in
to
s
ev
er
al
p
a
tch
es.
PVTv
2
in
h
er
en
tly
co
n
tain
s
r
elativ
e
p
o
s
itio
n
en
co
d
in
g
f
o
r
th
e
p
ix
el
s
with
in
ea
ch
im
ag
e
p
atch
.
Ho
wev
er
,
it
d
o
es
n
o
t
s
u
p
p
o
r
t
APE
cr
o
s
s
-
p
atch
es
at
all
f
o
u
r
le
v
els.
T
o
ad
d
r
ess
th
is
,
a
tr
u
n
ca
ted
n
o
r
m
al
d
is
tr
ib
u
tio
n
i
s
ad
o
p
ted
f
o
r
p
atch
p
o
s
itio
n
as APE
.
T
h
is
m
eth
o
d
s
lig
h
tly
im
p
r
o
v
es th
e
m
o
d
el’
s
ac
cu
r
ac
y
.
Fig
u
r
e
2
.
T
h
e
ch
ar
t
o
f
th
e
C
SA RGB
-
D
f
u
s
io
n
m
o
d
u
le
3
.
3
.
M
ulti
-
f
ea
t
ure
e
x
t
ra
ct
i
o
n
A
s
s
h
o
w
n
i
n
Fi
g
u
r
e
3
,
a
n
M
FE
m
o
d
u
l
e
c
o
n
t
a
i
n
i
n
g
t
h
r
e
e
d
i
f
f
e
r
e
n
t
t
y
p
e
s
a
n
d
s
i
z
es
o
f
k
e
r
n
e
ls
is
u
s
e
d
to
e
x
t
r
a
c
t
d
i
v
e
r
s
e
f
e
a
t
u
r
es
,
e
n
r
i
c
h
in
g
t
h
e
l
a
y
e
r
s
o
f
i
n
f
o
r
m
a
t
i
o
n
.
T
h
e
f
o
r
m
u
l
a
f
o
r
t
h
e
M
F
E
m
o
d
u
le
i
s
s
h
o
w
n
a
s
(
4
)
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
3
,
J
u
n
e
2
0
2
6
:
2
7
7
4
-
2
7
8
5
2778
=
(
(
(
)
,
(
)
,
(
)
)
)
(
4
)
W
h
er
e
p
r
esen
ts
th
e
MFE
m
o
d
u
le
o
u
tp
u
t
f
ea
tu
r
e
at
th
e
lev
el
,
(
∈
{
1
,
2
,
3
,
4
}
)
,
r
esp
ec
tiv
ely
,
in
d
icate
s
th
e
y
ield
ed
f
ea
tu
r
es
f
r
o
m
t
h
e
R
GB
-
D
f
u
s
io
n
m
o
d
u
le,
(
·
)
d
en
o
tes
o
r
ig
i
n
al
co
n
v
o
l
u
tio
n
,
(
·
)
d
en
o
tes
asy
m
m
etr
ic
co
n
v
o
l
u
tio
n
with
th
r
ee
d
if
f
er
en
t
k
er
n
els,
(
·
)
d
en
o
tes
d
y
n
am
ic
co
n
v
o
l
u
tio
n
with
a
weig
h
t c
o
m
p
u
tatio
n
k
er
n
el
b
a
s
ed
o
n
an
atten
tio
n
m
ec
h
a
n
is
m
.
T
h
e
f
o
r
m
u
la
f
o
r
(
·
)
is
s
h
o
wn
as
(
5
)
.
(
)
=
(
∗
3
×
3
)
⊕
(
∗
1
×
3
)
⊕
(
∗
3
×
1
)
(
5
)
W
h
er
e
d
en
o
tes
th
e
in
p
u
t
f
ea
tu
r
e,
∗
d
en
o
tes
co
n
v
o
lu
tio
n
,
n
×
m
d
en
o
tes
n
×
m
k
er
n
el
s
h
ar
e
d
in
th
e
s
am
e
win
d
o
w,
⊕
d
en
o
tes ele
m
en
t
-
wis
e
ad
d
itio
n
.
T
h
e
MFE
m
o
d
u
le
o
u
tp
u
ts
=
{
1
,
2
,
3
,
4
}
.
A
d
y
n
am
ic
co
n
v
o
lu
tio
n
m
o
d
u
le
was
p
r
o
p
o
s
e
d
to
en
ab
le
th
e
weig
h
ts
o
f
ea
ch
p
ix
el
to
b
e
d
y
n
am
ically
ad
ju
s
ted
d
u
r
in
g
th
e
co
n
v
o
lu
ti
o
n
p
r
o
ce
s
s
f
o
r
l
o
ca
l
f
ea
t
u
r
e
e
x
tr
ac
tio
n
.
T
h
is
m
o
d
u
le
en
h
a
n
ce
s
th
e
co
n
v
o
l
u
tio
n
o
p
er
atio
n
'
s
ab
ilit
y
to
p
er
ce
iv
e
lo
ca
l
f
ea
tu
r
es
b
y
in
co
r
p
o
r
atin
g
an
atten
tio
n
m
ec
h
an
is
m
.
T
h
is
m
ec
h
an
is
m
im
p
r
o
v
es
th
e
r
o
b
u
s
tn
ess
o
f
th
e
b
ase
f
ea
tu
r
e
ex
tr
ac
tio
n
,
a
n
d
ex
p
er
im
e
n
ts
h
av
e
d
em
o
n
s
tr
ated
its
ef
f
ec
tiv
en
ess
in
en
r
ich
in
g
th
e
ag
g
r
eg
ate
d
f
e
atu
r
es f
o
r
s
alien
cy
d
etec
tio
n
.
Fig
u
r
e
3
.
T
h
e
ch
ar
t
o
f
th
e
MFE
m
o
d
u
le
3
.
4
.
Se
m
a
ntic
s
eg
m
ent
a
t
io
n
enha
ncem
ent
As
s
h
o
wn
in
Fig
u
r
e
4
,
an
up
-
d
o
wn
-
s
am
p
lin
g
(
UDS)
m
o
d
u
le
is
u
s
ed
to
i
n
teg
r
ate
r
e
latio
n
s
h
ip
s
b
etwe
en
u
p
p
er
a
n
d
lo
wer
s
ca
les
to
m
in
e
in
f
o
r
m
atio
n
ac
r
o
s
s
m
u
ltip
le
s
ca
les.
Fo
r
th
e
cu
r
r
en
t
lev
el,
th
e
n
e
x
t
lev
el
is
u
p
-
s
am
p
led
a
n
d
t
h
e
p
r
ev
io
u
s
lev
el
is
d
o
wn
-
s
am
p
l
ed
s
o
th
at
th
e
th
r
ee
n
eig
h
b
o
r
i
n
g
lev
els
r
ea
ch
th
e
s
am
e
im
ag
e
s
ize.
T
h
e
f
ea
t
u
r
e
s
f
r
o
m
th
e
t
h
r
ee
lay
e
r
s
ar
e
co
n
ca
ten
ated
f
ir
s
t
an
d
th
en
o
u
tp
u
t
.
T
h
e
f
o
r
m
u
la
f
o
r
th
e
s
am
p
lin
g
m
o
d
u
le
is
s
h
o
wn
(
6
)
.
W
h
er
e
p
r
esen
ts
th
e
r
esu
lt
o
f
th
e
u
p
-
d
o
wn
s
am
p
lin
g
m
o
d
u
le
o
f
th
e
lev
el
,
(
∈
{
1
,
2
,
3
,
4
}
)
,
r
esp
ec
tiv
ely
,
s
ig
n
if
ies
th
e
o
u
tp
u
t f
ea
tu
r
es f
r
o
m
th
e
M
FE
m
o
d
u
le.
=
(
−
1
,
,
+
1
)
(
6
)
Fig
u
r
e
4
.
T
h
e
s
tr
u
ctu
r
e
o
f
th
e
SS
E
m
o
d
u
le
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
R
GB
-
D
s
a
lien
t o
b
ject
d
etec
tio
n
w
ith
lo
ca
l fe
a
tu
r
e
a
n
d
s
ema
n
tic
s
eg
men
ta
tio
n
(
Zh
a
n
g
Wa
n
g
)
2779
Nex
t,
th
r
ee
Dee
p
L
ab
v
3
+
m
o
d
u
les
with
d
if
f
er
en
t
d
ilatio
n
r
ates
o
f
3
,
6
,
a
n
d
9
a
r
e
u
s
ed
t
o
en
h
an
ce
s
alien
t
f
ea
tu
r
es
th
r
o
u
g
h
im
ag
e
s
eg
m
en
tatio
n
.
T
h
e
en
h
a
n
ce
m
en
t
o
f
th
e
m
o
d
u
le
b
y
th
e
SS
E
m
ad
e
th
e
ed
g
es
o
f
th
e
o
v
er
all
s
alien
t e
d
g
e
clea
r
er
an
d
m
o
r
e
d
e
f
in
ed
.
T
h
e
en
h
an
ce
m
en
t f
o
r
m
u
la
is
s
h
o
wn
as
(
7
)
.
=
(
(
(
3
+
(
(
(
(
)
)
)
)
)
)
)
(
7
)
W
h
er
e
p
r
esen
ts
th
e
o
u
tco
m
e
o
f
th
e
SS
E
m
o
d
u
le,
d
em
o
n
s
tr
ate
th
e
o
u
tp
u
t
f
ea
tu
r
e
o
f
th
e
u
p
-
d
o
w
n
s
am
p
lin
g
m
o
d
u
le,
3
+
d
en
o
tes
en
h
an
ce
m
o
d
u
le
b
ased
o
n
Dee
p
L
ab
v
3
+
with
a
d
ilatio
n
r
ate
o
f
,
(
∈
{
3
,
6
,
9
}
)
,
r
esp
ec
tiv
ely
.
T
h
e
SS
E
m
o
d
u
l
e
o
u
tp
u
ts
=
{
1
,
2
,
3
,
4
}
.
3
.
5
.
L
o
ca
l
a
nd
g
lo
ba
l a
g
re
e
m
ent
a
nd
edg
e
det
ec
t
io
n
Nex
t,
a
ca
p
s
u
le
n
etwo
r
k
is
a
d
o
p
ted
t
o
ad
d
r
ess
th
e
lo
ca
l
a
n
d
g
lo
b
al
ag
r
ee
m
en
t
o
f
s
alien
t
o
b
jects.
Ad
d
itio
n
ally
,
a
d
o
u
b
le
f
u
s
io
n
m
o
d
u
le
b
ased
o
n
ed
g
e
d
e
tectio
n
g
u
i
d
an
ce
is
in
t
r
o
d
u
c
ed
to
e
n
h
an
ce
th
e
ef
f
ec
tiv
en
ess
o
f
th
e
L
GE
m
o
d
u
le.
T
h
e
p
o
s
e
v
ec
to
r
s
o
u
tp
u
t
f
r
o
m
th
e
ca
p
s
u
le
n
etwo
r
k
ar
e
co
n
ca
ten
ated
with
th
e
ed
g
e
in
f
o
r
m
atio
n
p
r
o
ce
s
s
ed
b
y
d
o
u
b
le
f
u
s
io
n
t
o
f
o
r
m
th
e
f
in
al
o
u
tp
u
t p
o
s
e
v
ec
to
r
s
.
As
s
h
o
wn
in
Fig
u
r
e
5
,
th
e
s
o
u
r
ce
o
f
th
e
L
GE
is
th
e
o
u
tp
u
t
f
r
o
m
th
e
f
o
r
m
er
SS
E
m
o
d
u
le
at
ea
ch
lev
el.
T
h
e
p
r
im
ar
y
ca
p
s
u
le
an
d
th
e
L
GE
ca
p
s
u
le
ar
e
f
ir
s
t
b
u
ilt
to
o
b
tain
lo
ca
l
an
d
g
lo
b
al
ag
r
ee
m
en
t
r
esu
lts
.
Nex
t,
th
e
d
o
u
b
le
f
u
s
io
n
m
o
d
u
le
is
u
tili
ze
d
to
ex
tr
ac
t
im
ag
e
ed
g
e
in
f
o
r
m
atio
n
,
e
n
h
an
cin
g
t
h
e
p
o
s
e
esti
m
atio
n
o
f
th
e
ca
p
s
u
le
n
etwo
r
k
o
u
t
p
u
t.
T
h
en
,
a
Sm
is
h
ac
tiv
atio
n
f
u
n
ctio
n
is
ca
r
ef
u
lly
d
esig
n
ed
to
en
h
an
ce
o
p
tim
izatio
n
d
u
r
i
n
g
tr
ain
in
g
.
Fin
ally
,
th
e
p
o
s
e
o
f
th
e
ca
p
s
u
le
n
etwo
r
k
is
co
n
ca
ten
ated
with
th
e
o
u
tp
u
t
o
f
d
o
u
b
le
f
u
s
io
n
to
p
r
o
d
u
ce
th
e
o
u
tp
u
t
o
f
th
e
L
GE
m
o
d
u
le.
A
n
d
,
p
r
esen
ts
th
e
o
u
tp
u
t
f
ea
tu
r
e
o
f
th
e
d
o
u
b
le
fu
s
io
n
m
o
d
u
le,
⊕
d
en
o
tes
elem
en
t
-
wis
e
ad
d
itio
n
,
F
S
mis
h
(
·
)
and
S
mis
h
(
·
)
d
en
o
te
two
e
f
f
ec
tiv
e
ac
tiv
atio
n
f
u
n
ctio
n
s
,
1
,
2
,
3
d
en
o
te
th
e
o
u
tp
u
t
f
ea
tu
r
es
f
r
o
m
th
e
L
GE
m
o
d
u
le,
(
·
)
d
en
o
tes
d
o
wn
-
s
am
p
lin
g
,
(
·
)
d
en
o
tes
up
-
s
am
p
lin
g
.
T
h
e
f
o
r
m
u
la
f
o
r
th
e
L
GE
p
o
s
e
is
s
h
o
wn
as
(
8
)
.
_
=
(
(
(
(
2
)
,
3
,
(
4
)
)
)
,
)
(
8
)
W
h
er
e
_
p
r
esen
ts
th
e
o
u
tp
u
t
p
o
s
e
o
f
th
e
L
GE
m
o
d
u
le,
2
,
3
,
4
d
en
o
te
o
u
tp
u
t
f
ea
tu
r
es
f
r
o
m
SS
E
m
o
d
u
le,
(
·
)
d
en
o
tes ca
p
s
u
le
n
etwo
r
k
.
T
h
e
f
o
r
m
u
la
f
o
r
1
,
2
,
3
is
s
h
o
wn
in
Fig
u
r
e
1
.
T
h
e
L
GE
m
o
d
u
le
o
u
tp
u
ts
=
{
1
,
2
,
3
,
_
}
.
Fig
u
r
e
5
.
T
h
e
s
tr
u
ctu
r
e
o
f
th
e
L
GE
m
o
d
u
le
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
3
,
J
u
n
e
2
0
2
6
:
2
7
7
4
-
2
7
8
5
2780
3.
6
.
L
o
s
s
f
un
ct
io
ns
T
h
e
m
o
d
el
f
i
n
ally
o
u
tp
u
ts
=
{
1
⊗
1
,
1
,
2
,
3
,
_
}
,
wh
er
e
1
p
r
esen
ts
o
u
tp
u
t
f
ea
tu
r
e
1
o
f
th
e
SS
E
m
o
d
u
le
,
⊗
d
en
o
tes
elem
en
t
-
wis
e
m
u
ltip
licatio
n
,
r
ef
er
s
to
th
e
o
u
tp
u
t
f
ea
tu
r
e
o
f
th
e
L
GE
m
o
d
u
le.
B
in
ar
y
cr
o
s
s
-
en
tr
o
p
y
(
BCE
)
an
d
in
ter
s
e
ctio
n
o
v
er
u
n
io
n
(
I
o
U
)
l
o
s
s
f
u
n
ctio
n
s
ar
e
ad
o
p
ted
as su
p
er
v
is
io
n
s
tr
ateg
ies.
T
h
e
f
o
r
m
u
la
is
s
h
o
wn
as
(
9
)
.
F
iv
e
f
ea
tu
r
es in
l
f
s
s
ar
e
o
u
tp
u
t
as th
e
f
in
al
p
r
ed
ictio
n
s
u
n
d
er
th
e
co
n
s
tr
ain
t
o
f
th
e
l
f
s
s
lo
s
s
a
s
th
e
s
u
p
er
v
is
io
n
s
tr
ateg
y
,
an
d
th
e
s
u
m
m
atio
n
o
f
th
ese
f
iv
e
lo
s
s
es
i
s
th
e
f
in
al
lo
s
s
.
=
+
(
9
)
4.
E
XP
E
R
I
M
E
N
T
A
L
RE
SUL
T
S AN
D
A
NALY
SI
S
4
.
1
.
E
x
perim
ent
a
l set
t
ing
s
S
-
m
ea
s
u
r
e
(
)
,
m
ax
im
u
m
E
-
m
ea
s
u
r
e
(
)
,
m
ax
im
u
m
F
-
m
ea
s
u
r
e
(
)
,
an
d
m
ea
n
ab
s
o
lu
te
e
r
r
o
r
(
MA
E
)
is
u
s
ed
f
o
r
R
GB
-
D
SOD’
s
p
er
f
o
r
m
an
ce
co
m
p
a
r
i
s
o
n
.
T
h
en
,
th
e
m
o
d
els
ar
e
e
v
alu
ate
d
b
ased
o
n
s
ix
d
atasets
:
NJ
U2
K
[
1
]
,
NL
PR
[
2
4
]
,
STE
R
E
[
2
5
]
,
L
FS
D
[
2
6
]
,
SS
D
[
2
7
]
,
an
d
SIP
[
2
8
]
.
1
,
4
8
5
im
ag
es
f
r
o
m
NJU2
K
an
d
7
0
0
im
ag
es
f
r
o
m
NL
PR
ar
e
s
elec
ted
a
s
th
e
tr
ain
in
g
d
ataset,
an
d
th
e
o
th
e
r
im
ag
es
in
th
e
two
d
atasets
ar
e
s
er
v
ed
as
th
e
test
d
atasets
in
th
e
tr
ain
in
g
.
All
s
ix
d
atasets
ar
e
s
u
b
s
eq
u
en
tly
u
s
ed
as
test
d
atasets
f
o
r
th
e
f
in
al
ev
alu
atio
n
.
E
ig
h
t
d
ee
p
-
lear
n
in
g
m
o
d
els
ar
e
ad
o
p
ted
f
o
r
co
m
p
ar
is
o
n
p
u
r
p
o
s
es:
th
e
DF
[
2
4
]
,
B
AS
Net
[
2
9
]
,
DSA2
F
[
1
6
]
,
AFNet
[
3
0
]
,
DM
R
A
[
3
1
]
,
PiC
ANe
t
[
3
2
]
,
DC
MF
[
3
3
]
,
an
d
B
B
S
-
Net
[
3
4
]
.
T
h
e
wh
o
l
e
ex
p
er
im
en
t
is
co
n
d
u
cted
o
n
a
s
er
v
er
eq
u
ip
p
e
d
with
an
I
n
tel
C
o
r
e
i7
-
1
2
7
0
0
K
3
.
6
GHz
C
PU,
an
d
an
NVI
DI
A
GeFo
r
ce
R
T
X
3
0
9
0
GPU
(
2
4
GB
v
id
eo
m
em
o
r
y
)
.
T
h
e
s
o
f
twar
e
en
v
ir
o
n
m
en
t
u
tili
ze
s
Py
th
o
n
3
.
7
.
2
o
n
W
in
d
o
ws
1
1
,
with
Py
T
o
r
ch
1
.
1
3
.
0
+c
u
1
1
7
s
er
v
i
n
g
as
th
e
d
ee
p
lear
n
i
n
g
f
r
am
ewo
r
k
.
T
h
e
in
p
u
t
im
a
g
es
ar
e
u
n
if
o
r
m
ly
r
esized
to
3
8
4
×3
8
4
p
ix
els.
T
h
e
SGD
o
p
tim
izer
is
u
s
ed
with
an
in
itial
lear
n
in
g
r
ate
o
f
0
.
0
5
,
a
tr
ian
g
u
lar
lear
n
in
g
r
ate
s
ch
ed
u
le
(
0
.
0
0
0
5
–
0
.
0
5
)
,
weig
h
t
d
ec
ay
o
f
0
.
0
0
0
1
,
a
n
d
a
b
atch
s
ize
o
f
4
.
T
h
e
m
o
d
el
is
tr
ain
ed
f
o
r
2
0
0
e
p
o
ch
s
.
T
h
e
d
ata
an
d
s
o
u
r
ce
co
d
e
ar
e
o
p
en
l
y
av
ailab
le
in
GitHu
b
at
h
ttp
s
://g
ith
u
b
.
co
m
/wz
-
zr
d
-
ce
n
tin
/LFSS.
4
.
2
.
Q
ua
ntit
a
t
iv
e
e
v
a
lua
t
io
n
Qu
an
titativ
e
r
esu
lts
ar
e
illu
s
tr
ated
in
T
a
b
le
1;
th
e
r
esu
lts
d
em
o
n
s
tr
ate
th
at
th
e
p
r
o
p
o
s
ed
m
o
d
el
o
u
tp
er
f
o
r
m
s
all
co
m
p
ar
is
o
n
m
o
d
els
ac
r
o
s
s
f
o
u
r
e
v
alu
atio
n
m
etr
ics
o
n
s
ix
b
en
c
h
m
ar
k
d
a
tasets
.
Am
o
n
g
d
ee
p
lear
n
in
g
-
b
ased
m
o
d
els,
o
u
r
L
FS
S
m
eth
o
d
ac
h
iev
es
th
e
b
est
o
v
er
all
p
er
f
o
r
m
a
n
ce
.
Sp
e
cif
ically
,
th
e
L
FS
S
s
u
r
p
ass
es th
e
B
B
S
-
Net
[
3
4
]
,
with
im
p
r
o
v
em
e
n
ts
o
f
u
p
to
1
.
2
0
%,
1
.
4
1
%,
2
.
6
8
%,
an
d
0
.
0
9
,
r
esp
ec
tiv
ely
.
T
ab
le
1
.
Qu
a
n
titativ
e
co
m
p
ar
i
s
o
n
o
f
eig
h
t d
ee
p
lear
n
in
g
m
o
d
els
D
a
t
a
s
e
t
M
e
t
r
i
c
D
e
e
p
l
e
a
r
n
i
n
g
mo
d
e
l
s
Th
i
s
m
o
d
e
l
DF
B
A
S
N
e
t
D
S
A
2
F
A
F
N
e
t
D
M
R
A
P
i
C
A
N
e
t
D
C
M
F
BBS
-
N
e
t
LFS
S
N
JU
2
K
[
1
]
α
↑
.
7
6
3
.
8
6
5
.
8
7
5
.
8
8
4
.
8
7
1
.
8
8
0
.
9
1
2
.
9
2
1
.
9
2
8
β
↑
.
8
0
4
.
8
6
1
.
8
7
1
.
8
8
5
.
8
6
5
.
8
6
8
.
9
0
7
.
9
2
0
.
9
3
3
ε
↑
.
8
6
4
.
9
0
9
.
9
1
7
.
9
2
5
.
9
1
6
.
9
2
2
.
9
4
5
.
9
4
9
.
9
5
9
M↓
.
1
4
1
.
0
5
8
.
0
5
2
.
0
5
3
.
0
5
7
.
0
6
1
.
0
3
9
.
0
3
5
.
0
2
8
N
LPR
[2
4
]
α
↑
.
8
0
2
.
9
0
4
.
8
8
0
.
7
5
9
.
8
7
9
.
9
0
1
.
9
1
8
.
9
3
0
.
9
3
4
β
↑
.
7
7
8
.
8
8
1
.
8
5
4
.
6
9
9
.
8
5
3
.
8
7
5
.
9
0
3
.
9
1
8
.
9
2
5
ε
↑
.
8
8
0
.
9
3
8
.
9
3
0
.
8
3
6
.
9
3
3
.
9
4
4
.
9
5
4
.
9
6
1
.
9
6
5
M↓
.
0
8
5
.
0
3
3
.
0
3
8
.
0
7
0
.
0
3
9
.
0
3
8
.
0
2
7
.
0
2
3
.
0
1
9
S
TER
E
[
2
5
]
α
↑
.
7
5
7
.
8
8
3
.
8
7
8
.
8
2
0
.
8
4
4
.
8
9
0
.
9
0
4
.
9
0
8
.
9
1
9
β
↑
.
7
5
7
.
8
7
5
.
8
7
5
.
8
1
2
.
8
4
1
.
8
7
6
.
8
9
5
.
9
0
3
.
9
1
6
ε
↑
.
8
4
7
.
9
2
8
.
9
3
0
.
8
8
9
.
9
0
9
.
9
3
1
.
9
4
5
.
9
4
2
.
9
5
4
M↓
.
1
4
1
.
0
4
8
.
0
4
7
.
0
7
4
.
0
6
5
.
0
5
6
.
0
4
0
.
0
4
1
.
0
3
2
LFS
D
[
2
6
]
α
↑
.
7
8
3
.
8
0
2
.
8
2
8
.
8
1
7
.
8
0
7
.
8
2
7
.
8
5
8
.
8
6
4
.
8
7
4
β
↑
.
8
1
3
.
7
9
1
.
8
3
3
.
8
1
9
.
8
1
1
.
8
1
6
.
8
5
0
.
8
5
8
.
8
8
1
ε
↑
.
8
5
7
.
8
5
9
.
8
9
0
.
8
6
2
.
8
7
8
.
8
6
9
.
8
9
3
.
9
0
1
.
9
1
1
M↓
.
1
4
5
.
0
9
7
.
0
8
3
.
0
9
2
.
0
9
5
.
0
9
8
.
0
7
4
.
0
7
2
.
0
6
6
SSD
[
2
7
]
α
↑
.
7
4
7
.
8
1
1
.
8
2
6
.
8
1
1
.
8
2
6
.
8
3
7
.
8
6
4
.
8
8
2
.
8
8
5
β
↑
.
7
3
5
.
7
9
6
.
8
0
2
.
7
8
5
.
7
9
9
.
7
9
8
.
8
4
1
.
8
5
9
.
8
7
7
ε
↑
.
8
2
8
.
8
8
2
.
8
7
4
.
8
6
6
.
8
7
4
.
8
8
2
.
9
1
9
.
9
1
9
.
9
3
2
M↓
.
1
4
2
.
0
7
4
.
0
6
4
.
0
7
8
.
0
7
8
.
0
7
1
.
0
5
6
.
0
4
4
.
0
4
1
S
I
P
[
2
8
]
α
↑
.
6
5
3
.
8
1
5
.
8
4
6
.
6
2
8
.
8
1
9
.
8
5
0
.
8
5
3
.
8
7
9
.
8
7
9
β
↑
.
6
5
7
.
8
1
8
.
8
5
5
.
5
8
2
.
8
2
0
.
8
4
9
.
8
5
6
.
8
8
3
.
8
9
1
ε
↑
.
7
5
9
.
8
8
2
.
9
0
2
.
7
4
4
.
8
7
7
.
9
1
0
.
9
0
4
.
9
2
2
.
9
2
7
M↓
.
1
8
5
.
0
8
1
.
0
6
4
.
1
6
1
.
0
8
3
.
0
7
3
.
0
6
5
.
0
5
5
.
0
4
9
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
R
GB
-
D
s
a
lien
t o
b
ject
d
etec
tio
n
w
ith
lo
ca
l fe
a
tu
r
e
a
n
d
s
ema
n
tic
s
eg
men
ta
tio
n
(
Zh
a
n
g
Wa
n
g
)
2781
Fu
r
th
er
m
o
r
e
,
as
s
h
o
wn
in
Fig
u
r
e
6
,
th
e
r
ed
lin
es
p
r
esen
tin
g
o
u
r
L
FS
S
in
d
icate
s
u
p
er
io
r
p
er
f
o
r
m
a
n
ce
o
v
er
o
th
er
m
o
d
els.
As
s
h
o
w
n
in
Fig
u
r
e
7
,
th
e
m
o
d
el
is
e
v
a
lu
ated
in
d
if
f
er
e
n
t
s
ce
n
ar
io
s
.
I
n
th
e
f
ir
s
t
lin
e,
th
e
s
alien
t
o
b
ject
is
p
ar
ticu
lar
ly
p
r
o
m
in
en
t.
I
n
th
e
s
ec
o
n
d
lin
e,
t
h
er
e
ar
e
m
an
if
o
ld
s
alien
t
o
b
ject
s
.
I
n
th
e
th
ir
d
lin
e,
th
e
s
alien
t
tar
g
et
is
ex
tr
em
ely
elo
n
g
ated
.
I
n
th
e
f
o
u
r
th
lin
e,
th
e
s
alien
t
tar
g
et
clo
s
ely
r
esem
b
les
th
e
b
ac
k
g
r
o
u
n
d
.
I
n
th
e
f
if
th
lin
e,
th
e
d
is
tin
ctio
n
b
etwe
en
th
e
f
o
r
eg
r
o
u
n
d
an
d
th
e
b
ac
k
g
r
o
u
n
d
is
v
er
y
lo
w.
I
n
th
e
s
ix
th
lin
e,
th
e
tar
g
et
h
as
a
co
m
p
lex
s
h
ap
e.
T
h
ese
r
esu
lts
d
em
o
n
s
tr
ated
th
at
o
u
r
L
FS
S
m
o
d
el
ac
h
iev
es
th
e
h
ig
h
est ac
cu
r
ac
y
am
o
n
g
th
e
c
o
m
p
ar
is
o
n
m
o
d
els.
F
i
g
u
r
e
6
.
P
r
e
c
is
i
o
n
-
r
e
c
al
l
a
n
d
F
-
m
e
a
s
u
r
e
c
u
r
v
es
o
f
p
r
o
p
o
s
e
d
m
e
t
h
o
d
a
n
d
o
t
h
e
r
S
O
T
A
w
i
t
h
s
i
x
p
o
p
u
l
a
r
d
a
t
a
s
e
ts
Fig
u
r
e
7
.
T
h
e
v
is
u
al
co
m
p
ar
is
o
n
o
f
o
u
r
m
eth
o
d
an
d
o
th
er
SOTA
s
with
s
ix
im
ag
es f
r
o
m
p
o
p
u
lar
d
atasets
an
d
t
h
e
im
ag
es so
r
ted
f
r
o
m
b
est to
wo
r
s
t
4
.
3
.
Co
m
plex
it
y
a
na
ly
s
is
As
s
h
o
wn
in
T
ab
le
2
,
a
co
m
p
l
ex
ity
an
aly
s
is
o
f
ea
ch
f
u
n
ctio
n
al
m
o
d
u
le
in
L
FS
S
is
p
r
esen
ted
,
to
talin
g
1
6
1
.
4
m
illi
o
n
p
a
r
am
eter
s
an
d
5
7
.
3
g
ig
a
f
lo
atin
g
-
p
o
in
t
o
p
e
r
atio
n
s
(
GFLO
Ps
)
.
T
h
e
p
ar
a
m
eter
co
u
n
t
ac
r
o
s
s
m
o
d
u
les
ex
h
i
b
its
a
p
y
r
am
id
-
s
h
ap
ed
d
is
tr
ib
u
tio
n
,
with
th
e
b
o
tto
m
-
lev
el
f
ea
tu
r
e
e
x
tr
ac
tio
n
PVTv
2
m
o
d
u
le
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
3
,
J
u
n
e
2
0
2
6
:
2
7
7
4
-
2
7
8
5
2782
h
av
in
g
th
e
lar
g
est
p
ar
am
eter
co
u
n
t
(
1
2
4
.
0
m
illi
o
n
)
,
th
e
m
id
d
le
-
lev
el
C
SA
+
C
o
n
v
NeX
t
f
u
s
io
n
m
o
d
u
le
f
o
llo
win
g
n
ex
t
(
3
0
.
6
m
illi
o
n
)
,
an
d
th
e
h
ig
h
-
lev
el
s
em
an
tic
m
o
d
u
les
MFE
,
SS
E
,
an
d
L
GE
h
av
in
g
th
e
s
m
allest
p
ar
am
eter
c
o
u
n
ts
(
3
.
8
m
illi
o
n
,
1
.
6
m
illi
o
n
,
an
d
1
.
4
m
illi
o
n
r
esp
ec
tiv
ely
)
,
wh
ic
h
in
lin
e
w
ith
th
e
h
ier
a
r
ch
ical
r
ep
r
esen
tatio
n
p
r
in
ci
p
le
o
f
d
e
ep
lear
n
i
n
g
.
On
t
h
e
R
T
X
3
0
9
0
with
a
t
h
eo
r
etica
l
p
ea
k
FP
3
2
th
r
o
u
g
h
p
u
t
o
f
3
5
.
6
ter
a
f
lo
atin
g
-
p
o
in
t
o
p
er
a
tio
n
s
p
er
s
ec
o
n
d
(
T
FLOPS
)
,
an
d
ac
co
u
n
tin
g
f
o
r
o
p
er
atio
n
al
ef
f
icien
cy
lo
s
s
es,
L
FS
S
ca
n
th
eo
r
etica
lly
ac
h
ie
v
e
2
0
-
4
0
f
r
am
es
p
er
s
ec
o
n
d
(
FPS
)
.
I
n
r
ea
l
test
s
,
h
o
wev
er
,
th
e
I
n
tel
R
ea
lSen
s
e
d
ep
th
ca
m
er
a
(
3
0
FP
S)
an
d
o
t
h
er
I
/O
co
n
s
tr
ain
ts
lead
to
th
e
p
ip
elin
e
b
ein
g
m
ai
n
tain
ed
at
1
0
FP
S.
T
h
is
v
alu
e
f
alls
with
in
th
e
n
o
r
m
al
r
an
g
e
o
f
8
-
2
0
FP
S,
wh
ich
is
ty
p
ical
f
o
r
t
r
an
s
f
o
r
m
er
-
i
n
ten
s
iv
e
R
GB
-
D
SOD
m
o
d
els.
C
o
n
s
eq
u
en
tly
,
L
FS
S
o
p
er
ates
well
with
in
a
m
an
a
g
ea
b
le
co
m
p
u
tatio
n
al
b
u
d
g
et
a
n
d
is
r
ea
d
ily
d
e
p
lo
y
ab
le
in
en
g
in
ee
r
in
g
ap
p
licatio
n
s
.
T
ab
le
2
.
T
h
e
p
ar
am
ete
r
co
u
n
ts
an
d
FLOPs (
FP
3
2
)
u
s
ed
b
y
ea
ch
m
o
d
u
le
o
f
L
FS
S
M
o
d
u
l
e
s
P
a
r
a
me
t
e
r
c
o
u
n
t
(
F
P
3
2
)
F
LO
P
s
(
F
P
3
2
)
F
u
l
l
m
o
d
e
l
(
LFS
S
)
1
6
1
.
4
m
i
l
l
i
o
n
5
7
.
3
G
P
V
Tv
2
1
2
4
.
0
m
i
l
l
i
o
n
3
4
.
0
G
C
S
A
+
C
o
n
v
N
e
X
t
3
0
.
6
mi
l
l
i
o
n
2
2
.
2
G
M
F
E
3
.8
m
i
l
l
i
o
n
0
.
5
G
SSE
1
.6
m
i
l
l
i
o
n
0
.
3
G
LG
E
1
.
4
m
i
l
l
i
o
n
0
.
3
G
5.
CO
NCLU
SI
O
N
L
F
S
S
i
s
p
r
o
p
o
s
ed
,
an
R
G
B
-
D
S
OD
m
o
d
e
l
in
t
eg
r
a
ti
n
g
L
F
S
S
.
T
h
e
m
o
d
e
l
e
m
p
lo
y
s
en
co
d
er
-
d
e
co
d
er
ar
ch
it
ec
tu
r
e
.
T
h
e
en
co
d
er
u
ti
l
ize
s
a
PV
T
v
2
b
ac
k
b
o
n
e
w
i
th
AP
E
,
en
h
an
c
ed
b
y
C
SA
an
d
C
o
n
v
Ne
Xt
m
o
d
u
l
es
to
f
u
s
e
R
G
B
-
D
d
a
ta
wh
i
le
p
r
es
er
v
i
n
g
lo
ca
l
f
e
atu
r
e
s
.
T
h
e
d
ec
o
d
e
r
co
m
b
in
e
s
d
y
n
am
ic
co
n
v
o
lu
t
io
n
,
s
em
an
t
ic
s
eg
m
en
ta
tio
n
,
an
d
ed
g
e
d
et
ec
tio
n
th
r
o
u
g
h
MF
E
,
S
SE,
L
G
E
,
an
d
o
t
h
er
m
o
d
u
le
s
to
ex
tr
ac
t
s
a
li
en
t
o
b
j
ec
t
s
.
T
h
e
p
er
f
o
r
m
an
ce
o
f
th
e
n
et
w
o
r
k
is
te
s
t
ed
u
s
in
g
s
ix
d
a
ta
s
e
ts
a
n
d
f
o
u
r
m
e
tr
i
c
s
.
Af
ter
co
m
p
ar
i
n
g
it
to
eig
h
t
d
ee
p
l
ea
r
n
in
g
m
o
d
el
s
,
th
e
r
e
s
u
lt
s
s
h
o
w
th
a
t
L
F
S
S
ac
h
iev
e
s
th
e
h
ig
h
e
s
t
p
r
e
ci
s
io
n
.
T
h
e
m
o
d
el
c
an
f
u
n
c
t
io
n
a
s
a
s
tan
d
alo
n
e
m
o
d
u
le
an
d
c
an
al
s
o
b
e
ap
p
l
ied
to
o
th
er
f
ie
ld
s
o
f
m
ac
h
in
e
v
i
s
io
n
t
o
d
el
iv
er
ex
ce
ll
en
t
p
er
f
o
r
m
an
c
e.
I
n
th
e
f
iel
d
o
f
en
g
in
ee
r
in
g
ap
p
l
ic
at
io
n
s
,
e
v
en
w
i
th
lo
w
-
q
u
al
i
ty
R
G
B
an
d
d
ep
th
d
at
a
an
d
ex
tr
em
e
ly
co
m
p
lex
b
a
ck
g
r
o
u
n
d
s
,
L
F
SS
e
f
f
e
ct
iv
e
ly
ex
t
r
ac
t
s
s
al
ien
t
o
b
j
ec
ts
.
At
th
e
m
a
c
r
o
l
ev
e
l,
th
e
L
F
S
S
m
o
d
el
ac
cu
r
a
te
ly
id
en
tif
ie
s
t
h
e
co
r
r
ec
t
n
u
m
b
er
o
f
s
a
li
en
t
o
b
je
ct
s
,
wh
i
le
at
th
e
m
ic
r
o
lev
e
l,
it
ca
p
tu
r
e
s
d
eta
i
led
o
b
j
ec
t
ed
g
e
s
.
Ad
d
i
t
io
n
a
lly
,
L
F
S
S
i
s
an
i
n
tu
it
iv
e
an
d
g
en
er
a
l
izab
le
m
o
d
el
t
h
at
ca
n
b
e
r
e
ad
i
ly
in
t
eg
r
a
ted
in
to
o
th
e
r
v
i
s
io
n
s
y
s
t
em
s
.
F
UNDING
I
NF
O
R
M
A
T
I
O
N
T
h
is
p
ap
er
is
s
u
p
p
o
r
ted
b
y
t
h
e
Gen
er
al
R
esear
ch
Pro
jects
o
f
Z
h
ejian
g
Pr
o
v
in
cial
Dep
ar
tm
en
t
o
f
E
d
u
ca
tio
n
(
Gr
an
t N
o
.
Y
2
0
2
4
5
6
1
0
1
)
.
AUTHO
R
CO
NT
RI
B
UT
I
O
NS ST
A
T
E
M
E
N
T
T
h
is
jo
u
r
n
al
u
s
es
th
e
C
o
n
tr
ib
u
to
r
R
o
les
T
ax
o
n
o
m
y
(
C
R
ed
iT)
to
r
ec
o
g
n
ize
in
d
iv
id
u
al
au
th
o
r
co
n
tr
ib
u
tio
n
s
,
r
ed
u
ce
au
th
o
r
s
h
ip
d
is
p
u
tes,
an
d
f
ac
ilit
ate
co
llab
o
r
atio
n
.
Na
m
e
o
f
Aut
ho
r
C
M
So
Va
Fo
I
R
D
O
E
Vi
Su
P
Fu
Z
h
an
g
W
an
g
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
Kim
On
C
h
in
✓
✓
✓
✓
✓
✓
✓
R
ay
n
er
Alf
r
ed
✓
✓
✓
✓
✓
✓
✓
J
u
n
y
i Ch
a
i
✓
✓
✓
✓
✓
✓
✓
✓
R
u
n
d
o
n
g
Z
h
an
g
✓
✓
✓
✓
✓
✓
So
o
See
C
h
ai
✓
✓
✓
✓
✓
✓
✓
C
:
C
o
n
c
e
p
t
u
a
l
i
z
a
t
i
o
n
M
:
M
e
t
h
o
d
o
l
o
g
y
So
:
So
f
t
w
a
r
e
Va
:
Va
l
i
d
a
t
i
o
n
Fo
:
Fo
r
mal
a
n
a
l
y
s
i
s
I
:
I
n
v
e
s
t
i
g
a
t
i
o
n
R
:
R
e
so
u
r
c
e
s
D
:
D
a
t
a
C
u
r
a
t
i
o
n
O
:
W
r
i
t
i
n
g
-
O
r
i
g
i
n
a
l
D
r
a
f
t
E
:
W
r
i
t
i
n
g
-
R
e
v
i
e
w
&
E
d
i
t
i
n
g
Vi
:
Vi
su
a
l
i
z
a
t
i
o
n
Su
:
Su
p
e
r
v
i
s
i
o
n
P
:
P
r
o
j
e
c
t
a
d
mi
n
i
st
r
a
t
i
o
n
Fu
:
Fu
n
d
i
n
g
a
c
q
u
i
si
t
i
o
n
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
R
GB
-
D
s
a
lien
t o
b
ject
d
etec
tio
n
w
ith
lo
ca
l fe
a
tu
r
e
a
n
d
s
ema
n
tic
s
eg
men
ta
tio
n
(
Zh
a
n
g
Wa
n
g
)
2783
CO
NF
L
I
C
T
O
F
I
N
T
E
R
E
S
T
ST
A
T
E
M
E
NT
T
h
e
au
th
o
r
d
ec
lar
es
th
at
th
er
e
ar
e
n
o
k
n
o
wn
co
n
f
licts
o
f
in
ter
est
ass
o
ciate
d
wi
th
th
is
p
u
b
licatio
n
.
T
o
th
e
b
est o
f
th
e
a
u
th
o
r
’
s
k
n
o
wl
ed
g
e,
n
o
f
in
a
n
cial,
p
er
s
o
n
al,
p
r
o
f
ess
io
n
al,
o
r
i
n
s
titu
tio
n
al
r
elatio
n
s
h
ip
s
ex
is
t th
at
co
u
ld
h
av
e
i
n
f
lu
en
ce
d
th
e
d
esig
n
,
co
n
d
u
ct,
i
n
ter
p
r
etatio
n
,
o
r
r
ep
o
r
tin
g
o
f
t
h
e
wo
r
k
i
n
an
y
in
ap
p
r
o
p
r
iate
m
an
n
er
.
T
h
e
au
th
o
r
also
co
n
f
ir
m
s
th
at
th
is
m
a
n
u
s
cr
ip
t
was
p
r
e
p
ar
ed
in
d
e
p
en
d
e
n
tly
a
n
d
o
b
jectiv
ely
,
with
o
u
t
an
y
ex
ter
n
al
p
r
ess
u
r
e
o
r
in
ter
e
s
t th
at
m
ig
h
t c
o
m
p
r
o
m
is
e
its
in
teg
r
ity
.
DATA AV
AI
L
AB
I
L
I
T
Y
T
h
e
d
ata
a
n
d
s
o
u
r
ce
co
d
e
h
av
e
b
ee
n
m
a
d
e
p
u
b
licly
av
ailab
le
f
r
o
m
GitHu
b
r
e
p
o
s
ito
r
y
at
h
ttp
s
://g
ith
u
b
.
co
m
/wz
-
zr
d
-
ce
n
tin
/LFSS
.
RE
F
E
R
E
NC
E
S
[
1
]
A
.
B
o
r
j
i
,
M
.
M
.
C
h
e
n
g
,
Q
.
H
o
u
,
H
.
Ji
a
n
g
,
a
n
d
J.
Li
,
“
S
a
l
i
e
n
t
o
b
j
e
c
t
d
e
t
e
c
t
i
o
n
:
a
s
u
r
v
e
y
,
”
C
o
m
p
u
t
a
t
i
o
n
a
l
V
i
su
a
l
M
e
d
i
a
,
v
o
l
.
5
,
n
o
.
2
,
p
p
.
1
1
7
–
1
5
0
,
2
0
1
9
,
d
o
i
:
1
0
.
1
0
0
7
/
s
4
1
0
9
5
-
0
1
9
-
0
1
4
9
-
9.
[
2
]
M
.
M
a
sa
n
a
,
X
.
L
i
u
,
B
.
Tw
a
r
d
o
w
s
k
i
,
M
.
M
e
n
t
a
,
A
.
D
.
B
a
g
d
a
n
o
v
,
a
n
d
J.
V
.
D
.
W
e
i
j
e
r
,
“
C
l
a
ss
-
i
n
c
r
e
me
n
t
a
l
l
e
a
r
n
i
n
g
:
s
u
r
v
e
y
a
n
d
p
e
r
f
o
r
m
a
n
c
e
e
v
a
l
u
a
t
i
o
n
o
n
i
m
a
g
e
c
l
a
s
si
f
i
c
a
t
i
o
n
,
”
I
EEE
T
r
a
n
s
a
c
t
i
o
n
s o
n
P
a
t
t
e
r
n
An
a
l
y
si
s a
n
d
Ma
c
h
i
n
e
I
n
t
e
l
l
i
g
e
n
c
e
,
v
o
l
.
4
5
,
n
o
.
5
,
p
p
.
5
5
1
3
–
5
5
3
3
,
M
a
y
2
0
2
3
,
d
o
i
:
1
0
.
1
1
0
9
/
TPA
M
I
.
2
0
2
2
.
3
2
1
3
4
7
3
.
[
3
]
I
.
U
l
k
u
a
n
d
E
.
A
k
a
g
ü
n
d
ü
z
,
“
A
su
r
v
e
y
o
n
d
e
e
p
l
e
a
r
n
i
n
g
-
b
a
s
e
d
a
r
c
h
i
t
e
c
t
u
r
e
s
f
o
r
sem
a
n
t
i
c
se
g
me
n
t
a
t
i
o
n
o
n
2
D
i
ma
g
e
s,
”
A
p
p
l
i
e
d
Art
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
,
v
o
l
.
3
6
,
n
o
.
1
,
2
0
2
2
,
d
o
i
:
1
0
.
1
0
8
0
/
0
8
8
3
9
5
1
4
.
2
0
2
2
.
2
0
3
2
9
2
4
.
[
4
]
W
.
W
a
n
g
,
Q
.
L
a
i
,
H
.
F
u
,
J
.
S
h
e
n
,
H
.
Li
n
g
,
a
n
d
R
.
Y
a
n
g
,
“
S
a
l
i
e
n
t
o
b
j
e
c
t
d
e
t
e
c
t
i
o
n
i
n
t
h
e
d
e
e
p
l
e
a
r
n
i
n
g
e
r
a
:
a
n
i
n
-
d
e
p
t
h
su
r
v
e
y
,
”
I
EEE
T
ra
n
s
a
c
t
i
o
n
s
o
n
P
a
t
t
e
r
n
An
a
l
y
si
s
a
n
d
M
a
c
h
i
n
e
I
n
t
e
l
l
i
g
e
n
c
e
,
v
o
l
.
4
4
,
n
o
.
6
,
p
p
.
3
2
3
9
–
3
2
5
9
,
2
0
2
2
,
d
o
i
:
1
0
.
1
1
0
9
/
TPA
M
I
.
2
0
2
1
.
3
0
5
1
0
9
9
.
[
5
]
A
.
C
h
e
n
,
X
.
Li
,
T.
H
e
,
J
.
Zh
o
u
,
a
n
d
D
.
C
h
e
n
,
“
A
d
v
a
n
c
i
n
g
i
n
R
G
B
-
D
sa
l
i
e
n
t
o
b
j
e
c
t
d
e
t
e
c
t
i
o
n
:
a
s
u
r
v
e
y
,
”
A
p
p
l
i
e
d
S
c
i
e
n
c
e
s
,
v
o
l
.
1
4
,
n
o
.
1
7
,
2
0
2
4
,
d
o
i
:
1
0
.
3
3
9
0
/
a
p
p
1
4
1
7
8
0
7
8
.
[
6
]
Y
.
P
e
n
g
,
Z
.
Z
h
a
i
,
a
n
d
M
.
F
e
n
g
,
“
S
L
M
S
F
-
N
e
t
:
a
se
ma
n
t
i
c
l
o
c
a
l
i
z
a
t
i
o
n
a
n
d
mu
l
t
i
-
sc
a
l
e
f
u
si
o
n
n
e
t
w
o
r
k
f
o
r
R
G
B
-
D
sal
i
e
n
t
o
b
j
e
c
t
d
e
t
e
c
t
i
o
n
,
”
S
e
n
so
rs
,
v
o
l
.
2
4
,
n
o
.
4
,
2
0
2
4
,
d
o
i
:
1
0
.
3
3
9
0
/
s
2
4
0
4
1
1
1
7
.
[
7
]
M
.
Z
h
o
n
g
,
J
.
S
u
n
,
P
.
R
e
n
,
F
.
W
a
n
g
,
a
n
d
F
.
S
u
n
,
“
M
A
G
N
e
t
:
m
u
l
t
i
-
sc
a
l
e
a
w
a
r
e
n
e
ss
a
n
d
g
l
o
b
a
l
f
u
s
i
o
n
n
e
t
w
o
r
k
f
o
r
R
G
B
-
D
s
a
l
i
e
n
t
o
b
j
e
c
t
d
e
t
e
c
t
i
o
n
,
”
K
n
o
w
l
e
d
g
e
-
B
a
se
d
S
y
st
e
m
s
,
v
o
l
.
2
9
9
,
2
0
2
4
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
k
n
o
s
y
s.
2
0
2
4
.
1
1
2
1
2
6
.
[
8
]
C
.
S
u
n
,
Q
.
Z
h
a
n
g
,
C
.
Z
h
u
a
n
g
,
a
n
d
M
.
Z
h
a
n
g
,
“
B
M
F
N
e
t
:
b
i
f
u
r
c
a
t
e
d
m
u
l
t
i
-
m
o
d
a
l
f
u
s
i
o
n
n
e
t
w
o
r
k
f
o
r
R
G
B
-
D
sal
i
e
n
t
o
b
j
e
c
t
d
e
t
e
c
t
i
o
n
,
”
I
m
a
g
e
a
n
d
V
i
si
o
n
C
o
m
p
u
t
i
n
g
,
v
o
l
.
1
4
7
,
2
0
2
4
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
i
m
a
v
i
s.
2
0
2
4
.
1
0
5
0
4
8
.
[
9
]
Y
.
P
e
n
g
,
M
.
F
e
n
g
,
a
n
d
Z.
Z
h
e
n
g
,
“
R
G
B
-
D
sa
l
i
e
n
t
o
b
j
e
c
t
d
e
t
e
c
t
i
o
n
me
t
h
o
d
b
a
se
d
o
n
m
u
l
t
i
-
mo
d
a
l
f
u
si
o
n
a
n
d
c
o
n
t
o
u
r
g
u
i
d
a
n
c
e
,
”
I
EEE
A
c
c
e
ss
,
v
o
l
.
1
1
,
p
p
.
1
4
5
2
1
7
–
1
4
5
2
3
0
,
2
0
2
3
,
d
o
i
:
1
0
.
1
1
0
9
/
A
C
C
ESS
.
2
0
2
3
.
3
3
4
4
6
4
4
.
[
1
0
]
W
.
W
a
n
g
e
t
a
l
.
,
“
P
V
T
v
2
:
i
m
p
r
o
v
e
d
b
a
sel
i
n
e
s
w
i
t
h
p
y
r
a
m
i
d
v
i
si
o
n
t
r
a
n
s
f
o
r
mer,”
C
o
m
p
u
t
a
t
i
o
n
a
l
V
i
s
u
a
l
Me
d
i
a
,
v
o
l
.
8
,
n
o
.
3
,
p
p
.
4
1
5
–
4
2
4
,
2
0
2
2
,
d
o
i
:
1
0
.
1
0
0
7
/
s
4
1
0
9
5
-
0
2
2
-
0
2
7
4
-
8.
[
1
1
]
Z.
L
i
u
,
H
.
M
a
o
,
C
.
Y
.
W
u
,
C
.
F
e
i
c
h
t
e
n
h
o
f
e
r
,
T.
D
a
r
r
e
l
l
,
a
n
d
S
.
X
i
e
,
“
A
C
o
n
v
N
e
t
f
o
r
t
h
e
2
0
2
0
s,”
i
n
2
0
2
2
I
EEE
/
C
V
F
C
o
n
f
e
r
e
n
c
e
o
n
C
o
m
p
u
t
e
r
Vi
s
i
o
n
a
n
d
P
a
t
t
e
r
n
Re
c
o
g
n
i
t
i
o
n
(
C
VPR)
,
N
e
w
O
r
l
e
a
n
s
,
U
n
i
t
e
d
S
t
a
t
e
s
,
2
0
2
2
,
p
p
.
1
1
9
6
6
–
1
1
9
7
6
,
d
o
i
:
1
0
.
1
1
0
9
/
C
V
P
R
5
2
6
8
8
.
2
0
2
2
.
0
1
1
6
7
.
[
1
2
]
K
.
H
e
,
X
.
Z
h
a
n
g
,
S
.
R
e
n
,
a
n
d
J.
S
u
n
,
“
S
p
a
t
i
a
l
p
y
r
a
mi
d
p
o
o
l
i
n
g
i
n
d
e
e
p
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
t
w
o
r
k
s
f
o
r
v
i
s
u
a
l
r
e
c
o
g
n
i
t
i
o
n
,
”
I
EEE
T
ra
n
s
a
c
t
i
o
n
s
o
n
P
a
t
t
e
rn
A
n
a
l
y
si
s
a
n
d
Ma
c
h
i
n
e
I
n
t
e
l
l
i
g
e
n
c
e
,
v
o
l
.
3
7
,
n
o
.
9
,
p
p
.
1
9
0
4
–
1
9
1
6
,
2
0
1
5
,
d
o
i
:
1
0
.
1
1
0
9
/
TPA
M
I
.
2
0
1
5
.
2
3
8
9
8
2
4
.
[
1
3
]
T.
Z
h
o
u
,
D
.
P
.
F
a
n
,
M
.
M
.
C
h
e
n
g
,
J.
S
h
e
n
,
a
n
d
L
.
S
h
a
o
,
“
R
G
B
-
D
s
a
l
i
e
n
t
o
b
j
e
c
t
d
e
t
e
c
t
i
o
n
:
a
s
u
r
v
e
y
,
”
C
o
m
p
u
t
a
t
i
o
n
a
l
V
i
s
u
a
l
Me
d
i
a
,
v
o
l
.
7
,
n
o
.
1
,
p
p
.
3
7
–
6
9
,
2
0
2
1
,
d
o
i
:
1
0
.
1
0
0
7
/
s
4
1
0
9
5
-
020
-
0
1
9
9
-
z.
[
1
4
]
W
.
J
i
e
t
a
l
.
,
“
C
a
l
i
b
r
a
t
e
d
R
G
B
-
D
sa
l
i
e
n
t
o
b
j
e
c
t
d
e
t
e
c
t
i
o
n
,
”
i
n
2
0
2
1
I
E
E
E/
C
VF
C
o
n
f
e
r
e
n
c
e
o
n
C
o
m
p
u
t
e
r
Vi
si
o
n
a
n
d
P
a
t
t
e
r
n
Re
c
o
g
n
i
t
i
o
n
(
C
VPR)
,
2
0
2
1
,
p
p
.
9
4
6
6
–
9
4
7
6
,
d
o
i
:
1
0
.
1
1
0
9
/
C
V
P
R
4
6
4
3
7
.
2
0
2
1
.
0
0
9
3
5
.
[
1
5
]
Q
.
C
h
e
n
,
Z.
Li
u
,
Y
.
Zh
a
n
g
,
K
.
F
u
,
Q
.
Zh
a
o
,
a
n
d
H
.
D
u
,
“
R
G
B
-
D
sa
l
i
e
n
t
o
b
j
e
c
t
d
e
t
e
c
t
i
o
n
v
i
a
3
D
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
s
,
”
T
h
e
3
5
t
h
AA
A
I
C
o
n
f
e
re
n
c
e
o
n
A
rt
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
(
A
AA
I
-
21)
,
p
p
.
1
0
6
3
–
1
0
7
1
,
M
a
y
2
0
2
1
,
d
o
i
:
1
0
.
1
6
0
9
/
a
a
a
i
.
v
3
5
i
2
.
1
6
1
9
1
.
[
1
6
]
P
.
S
u
n
,
W
.
Zh
a
n
g
,
H
.
W
a
n
g
,
S
.
Li
,
a
n
d
X
.
Li
,
“
D
e
e
p
R
G
B
-
D
sa
l
i
e
n
c
y
d
e
t
e
c
t
i
o
n
w
i
t
h
d
e
p
t
h
-
se
n
si
t
i
v
e
a
t
t
e
n
t
i
o
n
a
n
d
a
u
t
o
m
a
t
i
c
mu
l
t
i
-
m
o
d
a
l
f
u
s
i
o
n
,
”
i
n
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
I
EEE
C
o
m
p
u
t
e
r
S
o
c
i
e
t
y
C
o
n
f
e
re
n
c
e
o
n
C
o
m
p
u
t
e
r
V
i
si
o
n
a
n
d
P
a
t
t
e
rn
Re
c
o
g
n
i
t
i
o
n
,
2
0
2
1
,
p
p
.
1
4
0
7
–
1
4
1
7
,
d
o
i
:
1
0
.
1
1
0
9
/
C
V
P
R
4
6
4
3
7
.
2
0
2
1
.
0
0
1
4
6
.
[
1
7
]
R
.
C
o
n
g
e
t
a
l
.
,
“
C
I
R
-
N
e
t
:
c
r
o
ss
-
m
o
d
a
l
i
t
y
i
n
t
e
r
a
c
t
i
o
n
a
n
d
r
e
f
i
n
e
m
e
n
t
f
o
r
R
G
B
-
D
sal
i
e
n
t
o
b
j
e
c
t
d
e
t
e
c
t
i
o
n
,
”
I
E
EE
T
ra
n
sa
c
t
i
o
n
s
o
n
I
m
a
g
e
Pr
o
c
e
ssi
n
g
,
v
o
l
.
3
1
,
p
p
.
6
8
0
0
–
6
8
1
5
,
2
0
2
2
,
d
o
i
:
1
0
.
1
1
0
9
/
TI
P
.
2
0
2
2
.
3
2
1
6
1
9
8
.
[
1
8
]
W
.
Z
h
a
n
g
,
Y
.
Ji
a
n
g
,
K
.
F
u
,
a
n
d
Q
.
Zh
a
o
,
“
B
TS
-
N
e
t
:
bi
-
d
i
r
e
c
t
i
o
n
a
l
t
r
a
n
sf
e
r
-
a
n
d
-
sel
e
c
t
i
o
n
n
e
t
w
o
r
k
f
o
r
R
G
B
-
D
sal
i
e
n
t
o
b
j
e
c
t
d
e
t
e
c
t
i
o
n
,
”
i
n
I
2
0
2
1
I
EEE
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
re
n
c
e
o
n
Mu
l
t
i
m
e
d
i
a
a
n
d
Ex
p
o
(
I
C
M
E)
,
S
h
e
n
z
h
e
n
,
C
h
i
n
a
,
2
0
2
1
,
p
p
.
1
–
6
,
d
o
i
:
1
0
.
1
1
0
9
/
I
C
M
E5
1
2
0
7
.
2
0
2
1
.
9
4
2
8
2
6
3
.
[
1
9
]
X
.
Z.
J
i
a
,
C
.
L.
D
o
n
g
Y
e
,
a
n
d
Y
.
J
.
P
e
n
g
,
“
S
i
a
Tr
a
n
s:
S
i
a
mes
e
t
r
a
n
sf
o
r
m
e
r
n
e
t
w
o
r
k
f
o
r
R
G
B
-
D
sa
l
i
e
n
t
o
b
j
e
c
t
d
e
t
e
c
t
i
o
n
w
i
t
h
d
e
p
t
h
i
ma
g
e
c
l
a
ssi
f
i
c
a
t
i
o
n
,
”
I
m
a
g
e
a
n
d
V
i
si
o
n
C
o
m
p
u
t
i
n
g
,
v
o
l
.
1
2
7
,
2
0
2
2
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
i
ma
v
i
s
.
2
0
2
2
.
1
0
4
5
4
9
.
[
2
0
]
K
.
Y
i
,
Y
.
Li
,
J
.
X
u
,
a
n
d
J
.
Zh
a
n
g
,
“
D
u
a
l
mu
t
u
a
l
l
e
a
r
n
i
n
g
n
e
t
w
o
r
k
w
i
t
h
g
l
o
b
a
l
-
l
o
c
a
l
a
w
a
r
e
n
e
s
s
f
o
r
R
G
B
-
D
sa
l
i
e
n
t
o
b
j
e
c
t
d
e
t
e
c
t
i
o
n
,
”
C
i
r
c
u
i
t
s
,
S
y
s
t
e
m
s,
a
n
d
S
i
g
n
a
l
P
ro
c
e
s
s
i
n
g
,
v
o
l
.
4
4
,
n
o
.
1
0
,
p
p
.
7
5
4
9
–
7
5
7
6
,
Ju
n
.
2
0
2
5
,
d
o
i
:
1
0
.
1
0
0
7
/
s0
0
0
3
4
-
025
-
0
3
1
4
3
-
4.
[
2
1
]
L.
C
.
C
h
e
n
,
Y
.
Z
h
u
,
G
.
P
a
p
a
n
d
r
e
o
u
,
F
.
S
c
h
r
o
f
f
,
a
n
d
H
.
A
d
a
m
,
“
E
n
c
o
d
e
r
-
d
e
c
o
d
e
r
w
i
t
h
a
t
r
o
u
s
se
p
a
r
a
b
l
e
c
o
n
v
o
l
u
t
i
o
n
f
o
r
sema
n
t
i
c
i
ma
g
e
s
e
g
me
n
t
a
t
i
o
n
,
”
C
o
m
p
u
t
e
r
V
i
si
o
n
–
EC
C
V
2
0
1
8
(
E
C
C
V
2
0
1
8
)
:
1
5
t
h
Eu
r
o
p
e
a
n
C
o
n
f
e
r
e
n
c
e
,
p
p
.
8
3
3
–
8
5
1
,
2
0
1
8
,
d
o
i
:
1
0
.
1
0
0
7
/
9
7
8
-
3
-
030
-
0
1
2
3
4
-
2
_
4
9
.
[
2
2
]
M
.
C
h
e
n
e
t
a
l
.
,
“
D
R
‐
N
e
t
:
a
n
i
m
p
r
o
v
e
d
n
e
t
w
o
r
k
f
o
r
b
u
i
l
d
i
n
g
e
x
t
r
a
c
t
i
o
n
f
r
o
m
h
i
g
h
r
e
so
l
u
t
i
o
n
r
e
m
o
t
e
s
e
n
s
i
n
g
i
m
a
g
e
,
”
Re
m
o
t
e
S
e
n
si
n
g
,
v
o
l
.
1
3
,
n
o
.
2
,
p
p
.
1
–
1
9
,
2
0
2
1
,
d
o
i
:
1
0
.
3
3
9
0
/
r
s1
3
0
2
0
2
9
4
.
[
2
3
]
S
.
D
u
,
S
.
D
u
,
B
.
Li
u
,
a
n
d
X
.
Z
h
a
n
g
,
“
I
n
c
o
r
p
o
r
a
t
i
n
g
D
e
e
p
La
b
v
3
+
a
n
d
o
b
j
e
c
t
-
b
a
se
d
i
m
a
g
e
a
n
a
l
y
si
s
f
o
r
se
ma
n
t
i
c
seg
m
e
n
t
a
t
i
o
n
o
f
v
e
r
y
h
i
g
h
r
e
s
o
l
u
t
i
o
n
r
e
mo
t
e
se
n
si
n
g
i
ma
g
e
s,
”
I
n
t
e
rn
a
t
i
o
n
a
l
J
o
u
r
n
a
l
o
f
D
i
g
i
t
a
l
E
a
rt
h
,
v
o
l
.
1
4
,
n
o
.
3
,
p
p
.
3
5
7
–
3
7
8
,
2
0
2
1
,
d
o
i
:
1
0
.
1
0
8
0
/
1
7
5
3
8
9
4
7
.
2
0
2
0
.
1
8
3
1
0
8
7
.
Evaluation Warning : The document was created with Spire.PDF for Python.