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.
15
,
No
.
2
,
A
p
r
il
2
0
2
6
,
p
p
.
1
4
7
3
~
1
4
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
2
.
p
p
1
4
7
3
-
1
4
8
5
1473
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Sca
ler enhance
d
deforma
ble at
tent
io
n wit
h
g
ra
ph
ne
ura
l
network
for v
ideo
com
press
io
n
Rev
a
t
hi K
a
s
ina
t
ha
perum
a
l
1
,
H
o
s
a
nn
a
P
rincy
e
P
er
ia
pa
nd
i
2
1
D
e
p
a
r
t
me
n
t
o
f
El
e
c
t
r
o
n
i
c
s a
n
d
C
o
m
mu
n
i
c
a
t
i
o
n
En
g
i
n
e
e
r
i
n
g
,
S
.
E.
A
.
C
o
l
l
e
g
e
o
f
E
n
g
i
n
e
e
r
i
n
g
a
n
d
Te
c
h
n
o
l
o
g
y
,
V
i
sv
e
sv
a
r
a
y
a
T
e
c
h
n
o
l
o
g
i
c
a
l
U
n
i
v
e
r
s
i
t
y
,
B
e
l
a
g
a
v
i
,
I
n
d
i
a
2
D
e
p
a
r
t
me
n
t
o
f
El
e
c
t
r
o
n
i
c
s a
n
d
C
o
m
mu
n
i
c
a
t
i
o
n
En
g
i
n
e
e
r
i
n
g
,
S
r
i
S
a
i
r
a
m
C
o
l
l
e
g
e
o
f
E
n
g
i
n
e
e
r
i
n
g
,
V
i
sv
e
sv
a
r
a
y
a
T
e
c
h
n
o
l
o
g
i
c
a
l
U
n
i
v
e
r
s
i
t
y
,
B
e
l
a
g
a
v
i
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Au
g
1
,
2
0
2
5
R
ev
is
ed
J
an
13
,
2
0
2
6
Acc
ep
ted
J
an
25
,
2
0
2
6
Vid
e
o
c
o
m
p
re
ss
io
n
is
wi
d
e
ly
u
se
d
t
o
re
d
u
c
e
b
a
n
d
wi
d
th
a
n
d
st
o
ra
g
e
re
q
u
irem
e
n
ts
wh
e
n
sto
ri
n
g
a
n
d
tran
sm
it
ti
n
g
v
id
e
o
s,
m
o
st
e
x
isti
n
g
n
e
u
ra
l
v
id
e
o
c
o
m
p
re
ss
io
n
a
p
p
ro
a
c
h
e
s
a
d
o
p
t
t
h
e
p
re
d
ictiv
e
re
si
d
u
e
-
c
o
d
i
n
g
fra
m
e
wo
rk
,
wh
ic
h
is
su
b
o
p
ti
m
a
l
fo
r
re
m
o
v
i
n
g
re
d
u
n
d
a
n
c
y
a
c
ro
s
s
fra
m
e
s.
Ad
d
it
i
o
n
a
ll
y
,
m
in
imiz
in
g
o
n
l
y
t
h
e
p
ix
e
l
-
wise
d
iffere
n
c
e
s
b
e
twe
e
n
th
e
ra
w
a
n
d
d
e
c
o
m
p
re
ss
e
d
fra
m
e
s
is
in
e
ffe
c
ti
v
e
in
imp
ro
v
i
n
g
t
h
e
p
e
rc
e
p
t
u
a
l
q
u
a
li
ty
o
f
th
e
v
id
e
o
s,
b
l
o
c
k
in
g
a
rti
fa
c
ts
d
e
g
ra
d
e
th
e
v
isu
a
l
q
u
a
li
t
y
,
e
sp
e
c
ially
n
e
a
r
e
d
g
e
s
a
n
d
tex
t
u
re
a
re
a
s.
He
n
c
e
,
to
so
lv
e
th
e
se
p
r
o
b
lem
s,
t
h
is
re
se
a
rc
h
p
ro
p
o
se
s
a
sc
a
ler
e
n
h
a
n
c
e
d
d
e
fo
rm
a
b
le
a
tt
e
n
ti
o
n
g
ra
p
h
n
e
u
ra
l
n
e
two
rk
(S
EDA
-
G
NN
)
to
u
ti
li
z
e
d
f
o
r
re
d
u
c
e
in
ter
-
fra
m
e
re
d
u
n
d
a
n
c
y
b
y
e
m
p
lo
y
i
n
g
a
d
e
fo
rm
a
b
le
a
tt
e
n
ti
o
n
m
e
c
h
a
n
ism
th
a
t
e
fficie
n
tl
y
c
a
p
tu
re
s
m
o
ti
o
n
a
n
d
stru
c
tu
ra
l
c
h
a
n
g
e
s,
t
h
e
re
b
y
m
in
imiz
in
g
re
d
u
n
d
a
n
c
y
.
M
o
d
e
ll
i
n
g
c
o
m
p
le
x
tem
p
o
ra
l
d
y
n
a
m
ics
with
g
ra
p
h
n
e
u
ra
l
n
e
two
r
k
s
(
G
NN
s
)
c
a
p
tu
re
s
d
e
p
e
n
d
e
n
c
ies
b
e
twe
e
n
fra
m
e
s,
th
e
re
b
y
fa
c
il
it
a
ti
n
g
h
i
g
h
l
y
e
fficie
n
t
v
id
e
o
e
n
c
o
d
i
n
g
,
t
h
e
n
c
o
n
stra
in
e
d
d
irec
ti
o
n
a
l
e
n
h
a
n
c
e
m
e
n
t
fil
ter
(CDEF
)
e
ffe
c
ti
v
e
ly
re
d
u
c
e
s
b
lo
c
k
in
g
a
rti
f
a
c
ts
wh
il
e
p
re
se
rv
in
g
s
h
a
rp
e
d
g
e
s
th
ro
u
g
h
d
irec
ti
o
n
a
l
a
n
d
c
o
n
stra
i
n
e
d
f
il
te
rin
g
,
t
h
e
re
b
y
imp
r
o
v
i
n
g
v
isu
a
l
q
u
a
li
t
y
in
c
o
m
p
re
ss
e
d
v
i
d
e
o
.
T
h
e
S
EDA
-
G
NN
a
p
p
r
o
a
c
h
a
c
h
iev
e
d
a
b
j
o
n
teg
a
a
rd
d
e
lt
a
b
it
ra
te
(
BD
-
BR
)
re
d
u
c
ti
o
n
o
f
2
.
3
7
2
%
o
n
t
h
e
jo
i
n
t
c
o
l
lab
o
ra
ti
v
e
tea
m
o
n
v
id
e
o
c
o
d
in
g
(JCT
-
VC)
d
a
tab
a
se
a
n
d
3
.
2
3
0
%
o
f
BD
-
BR
o
n
t
h
e
u
lt
ra
v
i
d
e
o
g
ro
u
p
(UV
G
)
d
a
tas
e
t,
d
e
m
o
n
stra
ti
n
g
sig
n
ifi
c
a
n
t
p
e
rfo
rm
a
n
c
e
wh
e
n
c
o
m
p
a
re
d
to
in
v
e
rti
b
le
n
e
u
ra
l
n
e
two
rk
s
(INN
s).
K
ey
w
o
r
d
s
:
C
o
n
s
tr
ain
ed
d
ir
ec
tio
n
al
en
h
an
ce
m
e
n
t f
ilter
Gr
ap
h
n
e
u
r
al
n
etwo
r
k
Hig
h
ef
f
icien
cy
v
id
eo
en
co
d
in
g
Scaler
en
h
an
ce
d
d
ef
o
r
m
ab
le
atten
tio
n
Vid
eo
co
m
p
r
ess
io
n
Vid
eo
q
u
ality
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
:
R
ev
ath
i K
asin
ath
ap
er
u
m
al
Dep
ar
tm
e
n
t
o
f
E
l
ec
t
r
o
n
ics
a
n
d
C
o
m
m
u
n
i
ca
ti
o
n
E
n
g
i
n
ee
r
i
n
g
,
S.E
.
A
.
C
o
ll
eg
e
o
f
E
n
g
in
ee
r
i
n
g
a
n
d
T
ec
h
n
o
l
o
g
y
Vis
v
esv
ar
ay
a
T
ec
h
n
o
l
o
g
ical
Un
iv
er
s
ity
B
elag
av
i,
Kar
n
atak
a,
I
n
d
ia
E
m
ail: r
ev
ath
is
elv
ar
aj1
2
2
9
@
g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
Vid
eo
en
co
d
in
g
,
also
k
n
o
wn
as
v
id
eo
co
m
p
r
ess
io
n
,
e
n
ab
les
th
e
s
ig
n
if
ican
t
tr
an
s
m
is
s
io
n
an
d
s
to
r
ag
e
o
f
v
i
d
eo
c
o
n
ten
t.
Stan
d
ar
d
o
f
v
i
d
eo
c
o
d
in
g
f
o
llo
w
a
s
tr
ateg
y
wh
ich
u
tili
ze
s
ad
v
a
n
ce
d
m
eth
o
d
s
lik
e
in
tr
a
-
f
r
am
e
p
r
ed
ictio
n
,
v
ar
iab
le
b
lo
ck
s
izes,
p
r
ed
ictio
n
o
f
in
ter
-
f
r
am
e,
q
u
an
tizatio
n
,
en
t
r
o
p
y
,
a
n
d
tr
a
n
s
f
er
co
d
in
g
to
ef
f
ec
tiv
ely
en
h
a
n
ce
co
m
p
r
ess
io
n
ef
f
icien
c
y
[
1
]
.
T
h
e
h
ig
h
-
e
f
f
icien
cy
v
i
d
eo
co
d
i
n
g
(
HE
VC
)
s
tan
d
ar
d
was
lau
n
ch
ed
b
y
th
e
jo
in
t
c
o
llab
o
r
ativ
e
team
o
n
v
id
eo
c
o
d
in
g
(
J
C
T
-
VC
)
to
d
if
f
e
r
en
tiat
e
class
es
b
ased
o
n
v
id
eo
q
u
ality
.
Vid
eo
c
o
m
p
r
ess
io
n
s
tan
d
ar
d
s
h
a
v
e
ad
v
a
n
ce
d
,
i
n
clu
d
in
g
a
d
v
an
ce
d
v
id
e
o
co
d
in
g
(
AVC),
HE
VC
,
v
er
s
atile
v
id
eo
co
d
in
g
(
VVC),
an
d
th
e
en
h
an
ce
d
c
o
m
p
r
ess
io
n
m
o
d
el
(
E
C
M)
,
to
d
eliv
er
h
ig
h
er
-
q
u
ality
v
id
eo
with
less
b
it
co
n
s
u
m
p
tio
n
[
2
]
,
[
3
]
.
W
ith
th
e
r
is
e
o
f
m
eta
v
er
s
e
-
en
ab
led
o
n
lin
e
c
o
m
m
u
n
icatio
n
,
in
ter
ac
tiv
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
15
,
No
.
2
,
Ap
r
il
20
26
:
1
4
7
3
-
1
4
8
5
1474
v
id
eo
co
d
in
g
h
as
m
ad
e
s
u
s
tain
ab
le
p
r
o
g
r
ess
ac
h
iev
in
g
lo
w
laten
cy
Ho
wev
er
,
a
n
u
ltra
lo
w
d
elay
is
ca
u
s
ed
b
y
h
ig
h
-
v
o
lu
m
e
d
ata
i
n
v
i
d
eo
co
m
m
u
n
icatio
n
with
r
aw
s
ig
n
al
s
[
4
]
.
HE
VC
/H.
2
6
5
h
as
in
cr
e
asin
g
ly
s
tr
u
g
g
le
d
to
m
ee
t
th
e
d
em
an
d
s
o
f
s
tr
ea
m
in
g
p
latf
o
r
m
s
as
a
wid
ely
ad
o
p
ted
d
ig
ital
m
ed
iu
m
v
id
eo
is
h
ig
h
ly
v
alu
ed
f
o
r
d
eliv
er
in
g
d
y
n
am
ic
an
d
im
m
er
s
iv
e
v
is
u
al
ex
p
e
r
ien
ce
s
[
5
]
.
Ho
wev
er
,
p
r
o
d
u
cin
g
a
v
id
eo
th
at
ef
f
ec
tiv
ely
co
n
v
ey
s
th
e
id
ea
o
f
ae
s
th
etic
q
u
ality
is
o
f
ten
m
o
r
e
ex
p
e
n
s
iv
e
an
d
co
m
p
le
x
,
th
e
v
id
eo
c
o
m
p
r
ess
io
n
m
eth
o
d
u
s
es
an
au
to
en
co
d
e
r
-
s
ty
le
n
etwo
r
k
f
o
r
en
co
d
in
g
an
d
d
e
co
d
in
g
.
Alth
o
u
g
h
an
au
t
o
en
co
d
er
ca
n
ca
p
tu
r
e
im
p
o
r
tan
t
in
f
o
r
m
atio
n
f
o
r
r
ec
o
n
s
tr
u
ctio
n
,
a
n
y
n
eg
lecte
d
in
f
o
r
m
atio
n
lo
s
t
d
u
r
in
g
e
n
co
d
in
g
ca
n
n
o
t
b
e
r
ec
o
v
er
e
d
d
u
r
in
g
d
ec
o
d
in
g
,
wh
ich
lead
s
to
ir
r
ev
er
s
ib
le
d
eg
r
ad
atio
n
[
6
]
,
[
7
]
.
T
h
e
d
ee
p
co
n
tex
tu
al
v
id
e
o
co
m
p
r
e
s
s
io
n
(
DC
VC
)
ap
p
r
o
ac
h
is
p
r
esen
ted
f
o
r
co
n
d
itio
n
al
co
d
in
g
-
b
ased
d
ee
p
v
id
e
o
co
m
p
r
ess
io
n
,
in
DC
VC
,
a
v
alu
ab
le
co
n
tex
t
is
ex
tr
ac
te
d
as
a
co
n
d
itio
n
to
co
m
p
r
ess
th
e
cu
r
r
en
t
f
r
a
m
e
an
d
im
p
r
o
v
e
th
e
co
m
p
r
ess
io
n
ef
f
icien
cy
[
8
]
.
T
h
e
p
ar
titi
o
n
s
tr
u
ctu
r
e
an
d
in
tr
a
p
r
ed
ictio
n
ar
e
p
er
f
o
r
m
ed
u
s
in
g
th
e
m
ax
i
m
u
m
tim
e
in
VVC
in
tr
a
co
d
in
g
an
d
b
r
u
te
-
f
o
r
ce
r
ec
u
r
s
iv
e
(
r
ate/d
is
to
r
tio
n
o
p
tim
izatio
n
o
r
R
DO)
s
ea
r
ch
,
th
at
d
eter
m
in
es
th
e
o
p
tim
al
p
ar
titi
o
n
an
d
c
h
o
o
s
e
th
e
b
est
in
tr
a
-
p
r
ed
ictio
n
m
o
d
el
[
9
]
.
I
n
a
s
em
an
tic
co
m
m
u
n
icatio
n
s
y
s
tem
,
r
aw
im
ag
es
ar
e
c
o
n
v
er
te
d
in
to
s
e
m
an
tic
f
ea
tu
r
es
at
th
e
t
r
an
s
m
itter
an
d
th
en
d
ec
o
d
ed
at
th
e
r
e
ce
iv
er
to
r
ec
o
n
s
tr
u
ct
th
e
d
ec
o
m
p
r
ess
ed
im
a
g
e.
Fo
r
task
-
e
x
ec
u
tio
n
ap
p
licatio
n
s
,
o
n
l
y
task
-
s
p
ec
if
ic
s
em
an
tic
in
f
o
r
m
atio
n
is
ex
tr
ac
ted
an
d
e
n
co
d
e
d
at
th
e
tr
an
s
m
itter
[
1
0
]
.
I
n
all
in
tr
a
-
co
n
f
i
g
u
r
atio
n
s
,
ev
er
y
f
r
am
e
is
in
d
ep
en
d
en
tly
en
co
d
ed
b
y
co
n
s
id
er
in
g
an
im
ag
e
co
m
p
r
ess
io
n
ap
p
r
o
ac
h
,
wh
ich
tr
ea
ts
ea
ch
im
ag
e
as
a
s
tan
d
alo
n
e
en
tity
to
ex
p
lo
it
s
p
atial
r
ed
u
n
d
an
c
y
with
in
a
s
in
g
le
f
r
am
e
th
at
is
d
ec
o
d
ed
with
o
u
t
r
ef
e
r
en
ce
s
to
o
th
er
s
,
en
h
an
cin
g
r
o
b
u
s
tn
ess
[
1
1
]
.
T
r
ad
itio
n
al
m
eth
o
d
s
f
ail
to
e
x
p
lo
it
tem
p
o
r
al
r
e
d
u
n
d
an
cy
,
s
im
ilar
ity
a
m
o
n
g
co
n
s
ec
u
tiv
e
f
r
am
es,
an
d
k
ey
f
ac
t
o
r
s
in
v
id
eo
co
d
in
g
ef
f
icien
c
y
,
b
y
i
g
n
o
r
in
g
th
e
in
ter
f
r
am
e
c
o
r
r
elatio
n
s
th
at
av
o
id
s
ig
n
if
ican
t
o
p
p
o
r
tu
n
ities
to
d
e
cr
ea
s
e
b
itra
te
an
d
en
h
an
ce
e
f
f
ec
tiv
en
ess
in
co
m
p
r
ess
io
n
[
1
2
]
.
T
h
e
e
x
is
ti
n
g
d
ee
p
n
e
u
r
al
n
et
w
o
r
k
-
b
ase
d
a
u
t
o
e
n
co
d
e
r
a
p
p
r
o
ac
h
es
a
r
e
ap
p
l
ie
d
t
o
v
i
d
eo
co
m
p
r
ess
io
n
,
wh
e
r
e
o
p
tim
izi
n
g
r
a
te
d
is
to
r
t
i
o
n
t
r
a
d
e
o
f
f
s
t
o
r
e
d
u
c
e
b
it
r
at
e
w
h
il
e
p
r
ese
r
v
i
n
g
r
ec
o
n
s
t
r
u
c
ted
f
r
am
e
q
u
ali
ty
.
T
h
ese
ap
p
r
o
ac
h
ed
u
s
ed
f
o
r
n
o
n
-
l
in
ea
r
,
a
d
a
p
ti
v
e
r
e
p
r
ese
n
t
a
tio
n
w
h
i
c
h
is
m
o
r
e
f
l
e
x
i
b
le
t
o
c
ap
tu
r
e
c
o
m
p
l
e
x
s
p
at
ial
v
a
r
i
ati
o
n
s
ac
r
o
s
s
f
r
a
m
e
s
[
1
3
]
,
[
1
4
]
.
T
h
e
t
r
a
d
it
io
n
al
e
n
co
d
i
n
g
te
ch
n
i
q
u
es
h
e
lp
s
t
o
p
r
e
d
ic
t
s
p
ati
o
t
em
p
o
r
al
r
el
ati
o
n
s
h
i
p
am
o
n
g
b
l
o
c
k
r
e
g
i
o
n
s
i
n
c
o
n
s
e
cu
ti
v
e
f
r
a
m
es
t
o
im
p
r
o
v
e
c
o
m
p
r
ess
io
n
e
f
f
ic
ie
n
cy
a
n
d
ac
cu
r
ac
y
o
f
th
e
m
o
d
el
p
e
r
f
o
r
m
a
n
c
e
t
o
r
ed
u
c
e
m
in
im
i
ze
r
es
id
u
a
l
b
lo
ck
er
r
o
r
[
1
5
]
.
T
h
e
h
i
g
h
f
r
e
q
u
e
n
c
y
e
m
b
e
d
d
in
g
ap
p
r
o
ac
h
es
a
r
e
u
ti
liz
ed
f
o
r
f
r
a
m
e
r
ec
o
n
s
t
r
u
cti
o
n
,
a
n
d
als
o
v
a
r
i
o
u
s
te
ch
n
i
q
u
es
li
k
e
a
d
ap
t
iv
e
q
u
a
n
t
iza
ti
o
n
f
a
ct
o
r
s
.
M
ac
h
i
n
e
le
ar
n
i
n
g
(
ML
)
a
n
d
d
ee
p
l
ea
r
n
in
g
(
DL
)
t
ec
h
n
iq
u
es
h
el
p
s
t
o
i
m
p
r
o
v
e
th
e
e
f
f
ici
en
c
y
a
n
d
co
m
p
r
ess
i
o
n
v
i
d
e
o
q
u
a
lit
y
o
f
th
e
d
at
a
[
1
6
]
,
[
1
7
]
.
T
h
ese
m
et
h
o
d
s
s
o
l
v
e
t
h
e
c
h
al
le
n
g
es
li
k
e
la
r
g
e
a
n
d
co
m
p
le
x
m
o
ti
o
n
s
i
n
t
h
e
v
i
d
e
o
s
an
d
i
n
v
id
e
o
p
u
s
h
i
n
g
l
ea
r
n
e
d
v
a
r
i
et
y
o
f
d
i
f
f
er
e
n
t
m
o
ti
o
n
p
at
te
r
n
s
w
h
ich
is
f
u
r
t
h
e
r
a
wa
y
f
r
o
m
p
r
ac
t
ica
l
u
s
a
b
il
it
y
i
n
c
o
n
s
u
m
e
r
d
e
v
i
ce
s
.
T
h
e
u
s
i
n
g
o
f
g
e
n
e
r
i
c
f
r
am
ew
o
r
k
to
c
o
n
tr
o
l
th
e
s
c
al
e
m
o
t
io
n
v
e
ct
o
r
s
o
n
p
e
r
f
r
am
e
b
asis
to
ap
p
r
o
x
i
m
a
tel
y
m
a
tc
h
t
h
e
r
a
n
g
e
o
f
m
o
t
io
n
in
th
e
t
r
a
in
in
g
v
i
d
e
o
[
1
8
]
,
[
1
9
]
.
A
s
ev
er
al
r
ese
ar
c
h
es
f
a
ci
n
g
ch
al
len
g
es
i
n
p
r
e
d
i
cti
n
g
c
o
d
i
n
g
t
o
el
im
in
ate
th
e
s
h
o
r
t
-
te
r
m
i
n
t
er
-
f
r
am
e
a
n
d
in
tr
a
-
f
r
am
e
r
ed
u
n
d
a
n
c
ies
t
h
a
t
lea
d
t
o
p
o
o
r
e
f
f
i
ci
en
c
y
in
co
m
p
r
ess
i
o
n
,
ex
is
ti
n
g
m
et
h
o
d
s
a
r
e
n
o
t
e
f
f
ec
ti
v
e
t
o
g
e
n
e
r
a
ti
n
g
l
o
n
g
te
r
m
b
a
ck
g
r
o
u
n
d
f
r
o
m
t
h
e
n
o
is
e
[
2
0
]
.
D
u
et
a
l
.
[
2
1
]
i
m
p
l
em
en
te
d
co
n
t
ex
tu
al
g
e
n
e
r
ati
v
e
v
id
e
o
co
m
p
r
ess
i
o
n
wi
th
t
r
a
n
s
f
o
r
m
e
r
s
(
C
G
VC
-
T
)
t
h
a
t
ad
o
p
te
d
a
g
en
er
ati
v
e
a
d
v
e
r
s
a
r
i
al
n
etw
o
r
k
(
GAN
)
t
o
im
p
r
o
v
e
co
d
i
n
g
e
f
f
ici
en
c
y
b
y
u
s
i
n
g
co
n
te
x
t
u
al
c
o
d
i
n
g
an
d
en
h
a
n
ce
p
e
r
c
ep
tu
al
q
u
ali
ty
.
GA
Ns
r
ec
o
n
s
tr
u
ct
f
in
er
d
etai
ls
an
d
m
an
ag
e
h
i
g
h
v
is
u
al
q
u
al
ity
a
t
l
o
w
er
b
it
r
a
tes
b
y
e
n
a
b
lin
g
h
ig
h
er
c
o
m
p
r
ess
io
n
e
f
f
ic
ien
c
y
an
d
r
e
d
u
ce
d
b
a
n
d
wi
d
t
h
u
s
a
g
e
.
H
o
we
v
e
r
,
t
h
e
GAN
m
o
d
e
l
was
t
r
ai
n
e
d
b
y
m
i
n
im
izi
n
g
t
h
e
m
e
an
s
q
u
a
r
e
e
r
r
o
r
(
MSE
)
,
wh
ic
h
af
f
e
cte
d
t
h
e
y
iel
d
i
n
th
e
s
m
o
o
t
h
e
d
f
r
a
m
es
a
n
d
le
d
t
o
u
n
s
atis
f
ac
to
r
y
p
er
ce
p
t
u
a
l
q
u
al
it
y
.
Sh
e
n
g
e
t
a
l
.
[
2
2
]
d
e
v
e
lo
p
ed
a
te
m
p
o
r
al
co
n
t
e
x
t
m
i
n
i
n
g
(
T
C
M)
a
p
p
r
o
ac
h
t
h
a
t
p
r
o
p
a
g
a
te
d
f
e
at
u
r
es
b
e
f
o
r
e
r
e
co
n
s
t
r
u
ct
in
g
t
h
e
f
r
am
e,
a
n
d
t
h
ese
f
e
at
u
r
es
we
r
e
s
to
r
e
d
i
n
a
g
e
n
e
r
al
ize
d
d
e
co
d
e
d
p
ic
tu
r
e
b
u
f
f
er
.
M
u
lt
i
-
s
ca
l
e
te
m
p
o
r
a
l
c
o
n
te
x
ts
h
el
p
ed
t
o
e
x
t
r
a
ct
m
u
lti
-
s
c
ale
t
em
p
o
r
al
c
o
n
te
x
ts
f
r
o
m
th
e
p
r
o
p
a
g
a
te
d
f
ea
t
u
r
es
,
w
h
ic
h
e
n
h
a
n
ce
d
th
e
ac
c
u
r
ac
y
o
f
th
e
te
m
p
o
r
al
in
f
o
r
m
a
ti
o
n
.
T
h
e
t
em
p
o
r
al
c
o
n
t
ex
t
r
ef
i
lli
n
g
(
T
C
R
)
m
e
th
o
d
le
ar
n
s
t
h
e
te
m
p
o
r
al
co
n
t
e
n
t
w
h
ic
h
is
co
m
b
i
n
e
d
i
n
t
o
t
h
e
c
o
m
p
r
ess
i
o
n
s
c
h
e
m
e
.
T
h
is
i
n
v
o
l
v
es
a
c
o
n
te
x
t
u
a
l
e
n
c
o
d
e
r
-
d
ec
o
d
e
r
,
f
r
am
e
g
en
e
r
at
o
r
,
a
n
d
tem
p
o
r
al
co
n
t
e
x
t
e
n
co
d
er
.
T
h
e
p
a
r
a
lle
liz
ati
o
n
-
f
r
ie
n
d
l
y
e
n
co
d
in
g
a
p
p
r
o
ac
h
e
lim
in
at
es
u
s
e
o
f
a
n
au
to
r
e
g
r
ess
i
v
e
en
t
r
o
p
y
m
o
d
e
l
t
o
ac
h
i
e
v
e
p
r
a
c
tica
l
en
co
d
i
n
g
a
n
d
d
ec
o
d
i
n
g
t
i
m
es
.
H
o
w
ev
e
r
,
t
h
e
T
C
M
m
et
h
o
d
f
ac
e
d
a
s
t
r
u
g
g
l
e
d
e
p
e
n
d
i
n
g
o
n
ac
cu
r
a
te
f
ea
t
u
r
e
p
r
o
p
a
g
ati
o
n
,
w
h
ic
h
c
au
s
ed
e
r
r
o
r
s
w
h
e
n
t
h
e
m
et
h
o
d
was
n
o
t
p
r
o
p
e
r
l
y
h
a
n
d
le
d
.
Gu
o
et
a
l.
[
2
3
]
im
p
lem
en
ted
an
in
v
e
r
tib
le
e
n
co
d
in
g
-
d
ec
o
d
i
n
g
n
etwo
r
k
f
o
r
d
ee
p
-
v
i
d
eo
co
m
p
r
ess
io
n
.
I
n
v
er
tib
le
n
eu
r
al
n
etwo
r
k
s
(
I
NNs)
u
tili
ze
d
to
p
r
eser
v
e
b
o
th
s
p
atial
a
n
d
tem
p
o
r
al
i
n
f
o
r
m
atio
n
.
I
NN
is
p
r
im
ar
ily
ap
p
lied
in
r
esid
u
al
n
o
n
lin
ea
r
tr
an
s
f
o
r
m
atio
n
s
to
en
h
an
ce
th
e
r
ep
r
esen
tatio
n
p
o
wer
o
f
a
n
etwo
r
k
.
I
n
v
er
tib
le
e
n
co
d
in
g
h
el
p
ed
t
o
m
an
ag
e
m
o
tio
n
in
f
o
r
m
ati
o
n
an
d
r
ed
u
ce
ar
tifa
cts
in
co
m
p
r
ess
ed
v
id
e
o
s
.
Ho
wev
er
,
I
NN
f
ac
ed
d
if
f
icu
lt
y
in
f
u
lly
o
p
tim
izin
g
c
o
n
tex
tu
al
en
co
d
in
g
;
wh
ile
m
o
tio
n
c
o
m
p
r
ess
io
n
in
cr
ea
s
ed
,
th
e
ex
p
l
o
r
atio
n
o
f
co
n
tex
tu
al
en
co
d
in
g
lag
g
ed
.
T
h
a
n
g
a
n
d
B
an
g
[
2
4
]
d
ev
elo
p
ed
a
h
ier
ar
c
h
ical
r
an
d
o
m
ac
ce
s
s
co
d
in
g
m
eth
o
d
wh
ic
h
r
e
d
u
ce
d
b
id
ir
ec
tio
n
al
tem
p
o
r
al
r
ed
u
n
d
an
cy
to
en
h
a
n
ce
th
e
s
ig
n
if
i
ca
n
t
co
d
i
n
g
o
f
th
e
tr
ad
itio
n
al
d
ee
p
n
e
u
r
al
v
id
eo
co
m
p
r
ess
io
n
ap
p
r
o
ac
h
.
T
h
e
m
eth
o
d
ap
p
lied
a
n
in
ter
p
o
latio
n
n
etwo
r
k
v
i
d
e
o
f
r
am
es
to
en
h
a
n
ce
t
h
e
p
r
ed
ictio
n
o
f
th
e
in
ter
-
f
r
am
e.
T
h
e
p
r
o
p
o
s
ed
m
et
h
o
d
im
p
r
o
v
ed
in
ter
-
p
r
ed
ictio
n
p
r
ed
ictio
n
,
th
er
e
b
y
g
en
er
atin
g
in
ter
m
e
d
iate
f
r
a
m
es
th
at
r
ed
u
ce
d
r
esid
u
al
e
r
r
o
r
s
.
I
n
th
e
o
p
tim
ized
b
itra
te
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
S
ca
ler en
h
a
n
ce
d
d
efo
r
ma
b
le
a
tten
tio
n
w
ith
g
r
a
p
h
n
eu
r
a
l n
e
tw
o
r
k
fo
r
…
(
R
ev
a
th
i Ka
s
in
a
th
a
p
eru
ma
l
)
1475
allo
ca
tio
n
,
f
r
am
es
at
h
ig
h
er
le
v
els
in
th
e
g
r
o
u
p
o
f
p
ict
u
r
es
(
Go
P)
wer
e
ass
ig
n
ed
lo
wer
b
it
r
ates
b
ec
au
s
e
th
ey
wer
e
n
o
t u
s
ed
as
r
ef
er
e
n
ce
f
r
a
m
es.
Ho
wev
er
,
th
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
m
o
s
tly
f
o
cu
s
ed
o
n
t
h
e
co
n
f
ig
u
r
atio
n
o
f
d
elay
,
wh
er
e
th
e
o
r
d
e
r
o
f
f
r
a
m
es
m
atch
es
th
eir
en
co
d
in
g
o
r
d
er
,
wh
ich
a
f
f
ec
ts
th
e
co
m
p
r
ess
io
n
ef
f
icien
cy
o
f
th
e
v
id
eo
s
eq
u
en
ce
s
.
L
u
et
a
l.
[
2
5
]
im
p
lem
en
te
d
an
en
d
-
to
-
en
d
d
ee
p
v
id
e
o
co
m
p
r
ess
io
n
f
r
am
ewo
r
k
,
wh
ic
h
was
u
tili
ze
d
f
o
r
p
ix
el
-
wis
e
m
o
tio
n
in
f
o
r
m
atio
n
lear
n
ed
f
r
o
m
an
o
p
tical
f
l
o
w
n
etwo
r
k
a
n
d
f
u
r
th
er
co
m
p
r
ess
ed
b
y
an
au
to
-
en
co
d
er
n
etwo
r
k
to
s
av
e
b
its
.
I
n
r
ate
-
d
is
to
r
tio
n
o
p
tim
izatio
n
,
all
m
o
d
u
les
wer
e
jo
in
tly
tr
ain
ed
u
s
in
g
a
s
in
g
le
lo
s
s
f
u
n
ctio
n
to
b
alan
ce
co
m
p
r
ess
io
n
ef
f
icie
n
cy
an
d
v
id
e
o
q
u
ality
.
An
ad
ap
tiv
e
q
u
an
tizatio
n
lay
er
r
ed
u
ce
d
th
e
n
u
m
b
er
o
f
p
ar
am
eter
s
r
eq
u
ir
ed
f
o
r
v
a
r
iab
l
e
b
itra
te
co
d
in
g
.
R
esid
u
al
co
m
p
r
ess
io
n
u
s
es
d
ee
p
n
etwo
r
k
s
to
tr
an
s
f
o
r
m
an
d
en
co
d
e
r
esid
u
al
i
n
f
o
r
m
atio
n
ef
f
icien
tly
.
Ho
wev
e
r
,
t
h
e
b
lo
ck
-
b
ased
m
eth
o
d
in
tr
o
d
u
ce
s
in
ac
c
u
r
ate
m
o
tio
n
i
n
f
o
r
m
atio
n
an
d
th
u
s
d
e
g
r
ad
es
th
e
co
m
p
r
ess
io
n
p
er
f
o
r
m
an
ce
.
T
h
is
r
esear
ch
h
as
v
ar
io
u
s
ex
is
tin
g
ch
allen
g
es,
s
u
ch
as
tr
ad
itio
n
al
ap
p
r
o
ac
h
es
th
at
f
ac
e
m
an
y
p
r
o
b
lem
s
,
s
u
ch
as
p
o
o
r
p
er
ce
p
t
u
al
q
u
ality
,
wh
ich
is
ca
u
s
ed
b
y
s
m
o
o
th
ed
f
r
am
es.
T
h
e
m
eth
o
d
h
as
d
if
f
icu
lty
p
e
r
f
o
r
m
in
g
f
ea
tu
r
e
p
r
o
p
ag
atio
n
ac
cu
r
ately
,
wh
ich
af
f
ec
ts
th
e
o
v
er
all
m
o
d
el
p
er
f
o
r
m
an
c
e.
I
m
p
r
o
p
er
m
o
tio
n
in
f
o
r
m
atio
n
r
ed
u
ce
s
th
e
co
m
p
r
ess
io
n
p
er
f
o
r
m
an
ce
.
W
h
en
th
e
m
o
tio
n
co
m
p
r
ess
io
n
in
cr
ea
s
ed
,
th
e
m
o
d
el
f
ac
ed
p
r
o
b
lem
s
an
d
d
eg
r
ad
e
d
th
e
s
y
s
tem
p
er
f
o
r
m
an
ce
,
th
e
c
o
n
tr
ib
u
tio
n
s
o
f
th
is
wo
r
k
ar
e
a
s
f
o
llo
ws:
−
T
h
e
s
ca
ler
en
h
an
ce
d
d
e
f
o
r
m
a
b
le
atten
tio
n
g
r
ap
h
n
eu
r
al
n
et
wo
r
k
(
SEDA
-
GNN)
co
n
s
is
ts
o
f
d
ef
o
r
m
ab
l
e
atten
tio
n
m
ec
h
an
is
m
s
to
d
y
n
a
m
ically
in
teg
r
ate
an
d
alig
n
in
f
o
r
m
atio
n
ac
r
o
s
s
f
r
am
es,
wh
ich
ef
f
icien
tly
ca
p
tu
r
es
m
o
tio
n
an
d
s
tr
u
ctu
r
al
ch
an
g
es
to
m
in
im
ize
in
ter
-
f
r
am
e
r
e
d
u
n
d
an
cy
,
atten
tio
n
m
o
d
u
les
th
a
t
f
o
cu
s
o
n
p
r
eser
v
in
g
v
is
u
all
y
s
ig
n
if
ican
t
f
ea
tu
r
es,
SED
A
-
GNN,
im
p
r
o
v
e
th
e
v
is
u
a
l
f
id
elity
o
f
co
m
p
r
ess
ed
v
id
e
o
s
an
d
ad
d
r
es
s
th
e
lim
itatio
n
s
o
f
p
ix
el
-
wis
e
lo
s
s
f
u
n
ctio
n
s
.
−
T
h
e
co
n
s
tr
ain
ed
d
ir
ec
tio
n
al
en
h
an
ce
m
e
n
t
f
ilter
(
C
DE
F)
ef
f
ec
tiv
ely
r
ed
u
ce
s
b
lo
ck
i
n
g
ar
tifa
cts
wh
ile
p
r
eser
v
in
g
s
h
ar
p
ed
g
es,
an
d
it
s
d
ir
ec
tio
n
al
an
d
c
o
n
s
tr
ain
ed
f
ilter
in
g
ap
p
r
o
ac
h
p
r
im
ar
ily
t
ar
g
ets
r
in
g
in
g
ar
tifa
cts ac
r
o
s
s
d
etec
ted
ed
g
e
d
ir
ec
tio
n
s
,
th
er
eb
y
im
p
r
o
v
in
g
th
e
v
is
u
al
q
u
ality
in
c
o
m
p
r
ess
ed
v
id
eo
.
−
SEDA
-
GNN
m
o
d
els
ca
p
tu
r
e
co
m
p
lex
tem
p
o
r
al
d
e
p
en
d
e
n
cies
am
o
n
g
f
r
am
es
to
p
r
o
v
id
e
m
o
r
e
ef
f
icien
t
an
d
ac
cu
r
ate
v
id
e
o
co
m
p
r
ess
io
n
.
B
y
d
y
n
am
ically
m
o
d
elli
n
g
n
o
n
-
lo
ca
l
s
p
atio
-
tem
p
o
r
al
r
elatio
n
s
h
ip
s
b
ased
o
n
m
o
tio
n
c
u
es,
it e
f
f
ici
en
tly
im
p
r
o
v
es th
e
v
id
e
o
q
u
ality
.
T
h
is
r
esear
ch
is
o
r
g
an
ized
as
f
o
llo
ws
.
S
ec
tio
n
2
d
escr
ib
es
t
h
e
r
esear
ch
m
eth
o
d
.
Sectio
n
3
r
ep
r
esen
ts
th
e
r
esu
lts
an
d
d
is
cu
s
s
io
n
.
T
h
e
s
ec
tio
n
4
co
n
cl
u
d
es th
e
r
esear
ch
.
2.
RE
S
E
ARCH
M
E
T
H
O
D
I
n
th
is
r
esear
ch
,
th
e
SEDA
-
GNN
m
eth
o
d
m
in
im
izes
th
e
r
ed
u
n
d
a
n
cy
an
d
p
ix
el
-
wis
e
d
if
f
er
en
ce
s
am
o
n
g
th
e
r
aw
f
r
am
es
an
d
en
h
an
ce
s
th
e
v
id
eo
s
p
er
ce
p
t
u
al
q
u
ality
an
d
co
m
p
r
ess
io
n
p
r
o
c
ess
to
im
p
r
o
v
e
th
e
co
m
p
r
ess
io
n
v
id
e
o
q
u
ality
.
T
h
e
p
r
o
p
o
s
ed
SEDA
-
GNN
ap
p
r
o
ac
h
h
elp
s
to
d
eter
m
in
e
th
e
f
in
al
co
d
i
n
g
u
n
it
(
CU
)
s
ize
an
d
r
e
d
u
ce
d
th
e
C
U
s
id
e
th
er
eb
y
im
p
r
o
v
in
g
o
v
er
all
m
o
d
el
p
er
f
o
r
m
an
ce
.
T
h
e
C
DE
F
m
eth
o
d
s
ig
n
if
ican
tly
d
ec
r
ea
s
es
th
e
b
l
o
ck
in
g
ar
tifa
cts
b
y
p
r
eser
v
in
g
s
h
ar
p
ed
g
es
an
d
te
x
tu
r
e,
Fig
u
r
e
1
r
e
p
r
esen
t
th
e
f
lo
wch
ar
t o
f
th
e
p
r
o
p
o
s
ed
m
et
h
o
d
.
Fig
u
r
e
1
.
T
h
e
f
lo
w
c
h
ar
t o
f
p
r
o
p
o
s
ed
SEDA
-
GNN
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
15
,
No
.
2
,
Ap
r
il
20
26
:
1
4
7
3
-
1
4
8
5
1476
2
.
1
.
Da
t
a
s
et
des
cr
iptio
n
T
h
e
u
ltra
v
id
e
o
g
r
o
u
p
(
UVG)
[
2
6
]
d
ataset
with
(
1
9
2
0
×1
0
8
0
)
HD
v
id
eo
s
eq
u
en
ce
s
h
ig
h
,
th
e
h
ig
h
ef
f
icien
cy
v
i
d
eo
q
u
ality
(
HE
VC
)
class
B
with
f
u
ll
-
HD
v
id
eo
s
is
5
with
r
eso
lu
tio
n
o
f
v
id
eo
s
eq
u
en
ce
s
(
1
9
2
0
×1
0
8
0
)
.
HE
VC
class
C
with
4
v
id
eo
s
eq
u
en
ce
s
h
av
in
g
8
3
2
×
4
8
0
r
eso
lu
tio
n
s
.
HE
VC
class
D
h
as
4
v
id
eo
s
eq
u
en
ce
s
with
a
r
eso
lu
tio
n
o
f
4
1
6
×2
4
0
r
eso
lu
tio
n
,
with
a
Y
UV4
2
0
r
aw
f
o
r
m
at
av
ailab
le
with
d
if
f
er
e
n
t f
r
am
e
r
ates
f
r
o
m
2
4
f
r
a
m
es
p
er
s
e
co
n
d
to
1
2
0
f
p
s
.
W
h
er
e
Y
r
ep
r
esen
ts
th
e
lu
m
a
co
m
p
o
n
e
n
t
(
b
r
ig
h
tn
ess
)
an
d
U
an
d
V
ar
e
th
e
ch
r
o
m
a
co
m
p
o
n
en
ts
(
co
lo
r
in
f
o
r
m
atio
n
)
.
T
h
e
J
C
T
-
VC
[
2
7
]
d
ataset
is
a
g
r
o
u
p
o
f
v
id
e
o
c
o
d
in
g
e
x
p
er
ts
f
r
o
m
th
e
I
n
ter
n
atio
n
al
T
elec
o
m
m
u
n
icatio
n
Un
io
n
-
T
e
leco
m
m
u
n
icatio
n
Stan
d
ar
d
izat
io
n
Secto
r
(
I
T
U
-
T
)
s
tu
d
y
g
r
o
u
p
1
6
v
id
eo
co
d
in
g
ex
p
er
ts
g
r
o
u
p
(
VC
E
G)
,
an
d
th
e
I
n
ter
n
atio
n
al
Or
g
an
iz
atio
n
f
o
r
Stan
d
ar
d
izatio
n
(
I
SO)
/
I
n
ter
n
atio
n
al
E
lectr
o
tech
n
ical
C
o
m
m
is
s
io
n
(
I
E
C
)
/J
o
in
t
T
ec
h
n
ical
C
o
m
m
ittee
1
(
J
T
C
)
1
/Su
b
co
m
m
ittee
(
SC
)
2
9
/
W
o
r
k
in
g
Gr
o
u
p
(
W
G)
1
1
.
T
h
e
m
o
v
in
g
p
ictu
r
e
ex
p
er
ts
’
g
r
o
u
p
(
MPE
G)
i
s
a
n
ew
-
g
e
n
er
atio
n
s
tan
d
ar
d
f
o
r
c
r
ea
tin
g
v
id
eo
co
d
in
g
in
2
0
1
0
,
f
o
r
h
i
g
h
-
q
u
ali
ty
v
id
eo
c
o
d
in
g
m
in
im
izes
th
e
am
o
u
n
t
o
f
d
ata
r
eq
u
ir
e
d
b
y
a
p
p
r
o
x
im
ately
5
0
%.
T
h
e
d
atab
ase
co
n
s
is
ts
o
f
3
4
v
id
eo
s
,
wh
ich
ar
e
d
iv
id
e
d
in
to
tr
ain
in
g
1
4
v
id
eo
s
,
6
f
o
r
v
ali
d
atio
n
,
an
d
test
in
g
1
4
v
id
eo
s
.
T
r
ain
i
n
g
an
d
test
in
g
o
f
r
an
d
o
m
ly
s
elec
ted
v
id
eo
s
o
f
d
if
f
er
e
n
t r
eso
lu
tio
n
s
to
im
p
r
o
v
e
ef
f
icien
c
y
an
d
ass
es
s
p
er
f
o
r
m
an
ce
.
2
.
2
.
P
re
pro
ce
s
s
ing
I
n
th
is
p
h
ase,
a
C
DE
F
is
u
tili
z
ed
to
elim
i
n
ate
ar
tifa
cts
s
u
ch
as
r
in
g
in
g
ar
o
u
n
d
h
ar
d
e
d
g
es.
T
h
er
ef
o
r
e,
ap
p
ly
in
g
two
d
ir
ec
tio
n
al
f
ilte
r
s
it
is
ac
h
iev
ed
(
4
5
°
o
f
f
)
f
o
r
ev
er
y
p
i
x
el,
p
r
o
p
er
f
ilter
s
el
ec
tio
n
is
p
er
f
o
r
m
e
d
ac
co
r
d
in
g
to
th
e
o
p
tim
izatio
n
an
d
is
b
ased
o
n
th
e
m
in
im
izatio
n
o
f
(
1
)
.
W
h
er
e
th
e
g
r
o
u
p
o
f
p
ix
els
in
a
s
elec
ted
d
ir
ec
tio
n
is
an
d
th
e
m
ea
n
o
f
th
e
g
r
o
u
p
is
,
C
DE
F
is
an
in
-
lo
o
p
r
esto
r
atio
n
f
ilter
th
at
is
ap
p
lied
to
lo
o
p
r
esto
r
atio
n
u
n
its
(
L
R
U
)
with
6
4
×6
4
,
1
2
8
×
1
2
8
,
o
r
2
5
6
×2
5
6
-
p
ix
el
b
lo
ck
s
,
ev
e
r
y
L
R
U
in
d
ep
en
d
en
tly
c
h
o
o
s
es
th
e
r
esto
r
atio
n
o
p
tio
n
f
r
o
m
o
n
e
o
f
th
e
th
r
ee
p
o
s
s
ib
ilit
ies.
2
=
∑
∑
(
−
,
)
2
,
(
1
)
2
.
3
.
CU
pa
rt
it
i
o
nin
g
T
h
e
in
p
u
ts
g
ath
e
r
ed
an
d
f
ed
in
to
a
s
tr
u
ct
u
r
ed
n
eu
r
al
n
etwo
r
k
,
wh
ich
tr
ain
ed
u
s
in
g
a
s
p
ec
if
ic
lo
s
s
f
u
n
ctio
n
,
f
in
ally
HE
VC
is
in
teg
r
ated
in
to
t
h
e
tr
ain
ed
m
o
d
el,
an
d
t
h
e
o
r
i
g
in
al
tr
av
e
r
s
al
s
ea
r
ch
p
r
o
ce
s
s
is
r
ep
lace
d
.
T
h
e
d
e
v
elo
p
e
d
GNN
n
ee
d
s
to
g
o
th
r
o
u
g
h
th
e
m
etr
i
cs
o
f
th
e
th
r
ee
lay
er
s
,
wh
ich
ar
e
co
r
r
elate
d
to
th
e
g
lo
b
al
m
ea
n
s
q
u
a
r
e
r
esid
u
al
(
GM
SR
)
,
th
e
o
v
e
r
all
tex
tu
r
e
c
o
m
p
lex
ity
is
in
itially
ev
al
u
ate
d
b
y
c
o
m
p
u
tin
g
t
h
e
R
MSE
,
th
e
co
m
p
lex
ity
o
f
th
e
lo
ca
l
tex
tu
r
e
is
m
ea
s
u
r
ed
as
R
M
SE
≤
.
I
f
GM
SR
≤
,
th
e
C
U
is
s
m
o
o
th
,
an
d
in
d
ec
is
io
n
-
m
ak
in
g
,
n
o
s
m
all
s
ize
is
p
ar
titi
o
n
ed
,
GNN
is
u
s
ed
to
d
eter
m
i
n
e
wh
eth
er
f
u
r
th
e
r
p
ar
titi
o
n
in
g
is
r
eq
u
ir
e
d
to
d
et
er
m
in
es
C
U
tex
tu
r
e
co
m
p
lex
it
y
d
u
r
in
g
f
ast
p
ar
titi
o
n
in
g
a
n
d
s
elec
ts
a
b
etter
C
U
p
ar
titi
o
n
b
y
co
n
s
id
er
in
g
th
r
ee
m
etr
ics.
2
.
4
.
G
ra
ph
neura
l net
wo
rk
T
h
e
tr
ad
itio
n
al
m
eth
o
d
s
f
ac
ed
d
if
f
icu
lties
with
non
-
E
u
c
lid
ea
n
d
ata,
an
GNN
is
p
r
o
p
o
s
ed
f
o
r
g
r
ap
h
-
s
tr
u
ct
u
r
ed
d
ata
with
a
b
etter
s
co
p
e
f
o
r
v
id
eo
co
m
p
r
ess
io
n
.
A
GNN
co
n
s
is
t
s
o
f
a
g
r
ap
h
co
n
v
o
lu
tio
n
al
n
etwo
r
k
(
GC
N)
,
g
r
a
p
h
atten
tio
n
n
etwo
r
k
s
(
GAN)
,
an
d
g
r
ap
h
au
to
-
en
c
o
d
er
s
(
GAE
)
f
o
r
co
m
p
r
ess
io
n
.
I
n
v
id
eo
s
u
m
m
ar
izatio
n
,
ev
er
y
v
id
eo
is
p
r
esen
ted
as
a
g
r
ap
h
wh
er
e
f
ea
tu
r
es
ar
e
f
r
am
es
th
at
a
r
e
tr
ea
ted
as
g
r
ap
h
n
o
d
es
an
d
t
h
e
r
elatio
n
s
h
ip
s
am
o
n
g
f
r
am
es
ar
e
e
n
co
d
e
d
a
s
ed
g
e
weig
h
ts
wh
ich
ef
f
ec
ti
v
ely
ca
p
t
u
r
in
g
th
e
tem
p
o
r
al
an
d
s
p
atial
d
ep
e
n
d
e
n
cies
b
etwe
en
th
e
v
i
d
eo
f
r
am
es,
Fig
u
r
e
2
r
ep
r
esen
ts
th
e
ar
c
h
itectu
r
e
d
iag
r
am
o
f
SEDA
-
GNN
ap
p
r
o
ac
h
.
Fig
u
r
e
2
.
Ar
c
h
itectu
r
e
d
iag
r
a
m
o
f
SEDA
-
GNN
m
eth
o
d
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
S
ca
ler en
h
a
n
ce
d
d
efo
r
ma
b
le
a
tten
tio
n
w
ith
g
r
a
p
h
n
eu
r
a
l n
e
tw
o
r
k
fo
r
…
(
R
ev
a
th
i Ka
s
in
a
th
a
p
eru
ma
l
)
1477
T
h
e
in
tr
o
d
u
ctio
n
o
f
GNN
h
as
b
ee
n
s
u
cc
ess
f
u
l
an
d
ef
f
icien
t
in
th
e
d
o
m
ain
o
f
v
id
e
o
s
u
m
m
ar
izatio
n
,
Fig
u
r
e
2
th
e
ar
ch
itectu
r
e
d
iag
r
am
o
f
th
e
p
r
o
p
o
s
ed
m
eth
o
d
,
a
v
id
eo
f
r
am
e
in
p
u
t
is
g
iv
en
as
in
p
u
t
g
r
ap
h
an
d
th
eir
tem
p
o
r
al
co
n
n
ec
tio
n
s
.
A
g
r
ap
h
is
p
r
o
ce
s
s
ed
v
ia
two
g
r
a
p
h
co
n
v
o
lu
tio
n
lay
e
r
s
,
ev
er
y
l
ay
er
is
f
o
llo
wed
b
y
r
ec
tifie
d
lin
ea
r
u
n
it
(
R
eL
U
)
ac
tiv
atio
n
f
u
n
ctio
n
to
ex
tr
ac
t
s
p
atio
-
tem
p
o
r
al
f
ea
t
u
r
e
le
v
els,
s
u
b
s
eq
u
en
tly
th
e
f
ea
tu
r
es
f
r
o
m
all
f
r
am
es
a
r
e
c
o
n
ca
ten
ated
an
d
f
o
r
war
d
ed
th
r
o
u
g
h
a
f
u
lly
co
n
n
ec
ted
lay
er
(
FC
N)
.
Fin
ally
,
it
p
r
o
v
id
es
Y
as
an
o
u
tco
m
e
th
at
r
ep
r
esen
ts
co
m
p
r
ess
ed
v
i
d
eo
in
f
o
r
m
atio
n
.
Sp
ec
tr
al
g
r
ap
h
co
n
v
o
lu
tio
n
is
ad
o
p
ted
b
y
th
e
o
r
ig
in
al
GC
N
,
wh
ich
is
d
ef
in
ed
i
n
(
2
)
.
∗
=
(
2
)
W
h
er
e
th
e
n
o
r
m
alize
d
L
ap
laci
an
m
atr
ix
o
f
th
e
ei
g
en
v
ec
to
r
g
r
ap
h
m
atr
i
x
is
; th
u
s
,
=
,
an
d
is
th
e
in
p
u
t
d
ata,
tr
ain
ab
le
c
o
n
v
o
lu
tio
n
al
k
er
n
el
p
ar
am
eter
is
=
(
)
.
B
ec
au
s
e
o
f
th
e
lar
g
e
n
u
m
b
e
r
o
f
p
ar
am
eter
s
an
d
c
o
m
p
u
tatio
n
al
co
m
p
lex
ity
,
an
ad
v
an
ce
d
GC
N
was c
alcu
lated
u
s
in
g
(
3
)
.
∗
=
∑
−
1
=
0
(
3
)
T
o
ad
d
r
ess
th
e
is
s
u
e
o
f
h
ig
h
co
m
p
u
tatio
n
al
co
s
ts
,
two
ap
p
r
o
x
im
atio
n
ap
p
r
o
ac
h
es
ar
e
u
tili
ze
d
,
C
h
eb
y
s
h
ev
co
n
v
o
l
u
tio
n
al
k
er
n
el:
s
o
lv
in
g
h
ig
h
m
u
ltip
licativ
e
co
m
p
lex
it
y
p
r
o
b
lem
o
f
wid
e
d
en
s
e
m
atr
ices,
ap
p
r
o
x
im
ate
ar
e
p
r
o
v
id
ed
u
s
in
g
C
h
eb
y
s
h
e
v
p
o
ly
n
o
m
ials
is
(
⋀
)
.
T
h
e
f
o
r
m
u
la
f
o
r
th
is
ap
p
r
o
x
im
atio
n
is
g
iv
e
n
b
y
(
4
)
:
′
(
⋀
)
≈
∑
′
=
0
(
⋀
̃
)
(
4
)
W
h
e
r
e
̃
=
2
−
,
th
is
f
o
r
m
u
la
is
(
|
|
)
a
c
o
m
p
u
t
ati
o
n
al
ly
c
o
m
p
le
x
it
y
a
n
d
d
ep
e
n
d
s
o
n
t
h
e
n
u
m
b
er
o
f
ed
g
es
i
n
t
h
e
g
r
a
p
h
ce
n
t
r
a
l
n
o
d
e
f
i
r
s
t
-
o
r
d
er
n
ei
g
h
b
o
r
s
:
p
r
o
b
le
m
s
ar
e
s
im
p
l
if
y
i
n
g
a
n
d
alle
v
i
at
e
o
v
e
r
f
itti
n
g
b
e
ca
u
s
e
o
f
l
o
ca
l
n
o
d
e
i
n
la
r
g
e
s
ize
.
A
s
et
=
1
,
m
ea
n
i
n
g
w
h
i
ch
o
n
l
y
ce
n
t
r
al
n
o
d
e
a
n
d
its
f
i
r
s
t
-
o
r
d
e
r
p
r
o
x
i
m
i
ty
ar
e
s
u
b
s
t
it
u
te
d
,
a
cc
o
r
d
in
g
l
y
,
0
(
)
=
1
,
1
(
̃
)
=
̃
an
d
to
2
,
t
h
e
c
o
n
v
o
l
u
ti
o
n
wit
h
t
h
e
n
ew
ap
p
r
o
x
im
ate
f
o
r
m
u
la
is
g
iv
e
n
i
n
(
5
)
.
W
h
er
e
t
h
e
ei
g
e
n
v
al
u
e
r
an
g
e
o
f
+
−
1
2
1
2
is
[
0
,
2
]
,
g
r
a
d
i
en
t
d
is
a
p
p
e
ar
an
ce
an
d
m
u
ltil
a
y
e
r
c
o
n
v
o
l
u
t
i
o
n
le
ad
to
t
h
e
g
r
ad
ie
n
t
ex
p
l
o
s
i
o
n
g
i
v
e
n
i
n
(
6
)
,
w
h
e
r
e
̃
=
+
,
̃
=
∑
̃
,
th
e
eig
e
n
v
al
u
e
r
a
n
g
e
is
n
o
r
m
a
liz
e
d
am
o
n
g
[
0
,
1
]
,
s
u
b
s
ti
tu
ti
n
g
(
6
)
in
to
(
7
)
.
∗
≈
0
′
+
1
′
(
−
)
=
(
+
−
1
2
1
2
)
(
5
)
+
−
1
2
−
1
2
→
̃
−
1
2
̃
̃
−
1
2
(
6
)
∗
=
̃
−
1
2
̃
̃
−
1
2
(
7
)
2
.
5
.
Sca
le
-
enha
nced
def
o
rma
ble a
t
t
ent
io
n
Scale
-
en
h
an
ce
d
d
ef
o
r
m
ab
le
atten
tio
n
(
SEDA
)
h
as
b
ee
n
p
r
o
p
o
s
ed
,
wh
ic
h
co
m
b
in
es
s
ca
le
ag
g
r
eg
atio
n
,
d
ilated
s
am
p
lin
g
,
p
o
s
itio
n
u
p
d
atin
g
,
atten
tio
n
ca
lcu
latio
n
,
an
d
f
ea
tu
r
e
in
te
g
r
atio
n
m
ec
h
an
is
m
th
at
h
elp
s
t
o
im
p
r
o
v
e
in
s
tan
ce
awa
r
en
ess
ac
r
o
s
s
d
iv
er
s
e
s
ca
les
an
d
s
h
ap
es.
Scale
a
g
g
r
eg
atio
n
co
llects
m
u
lti
-
r
eso
lu
tio
n
f
ea
tu
r
es;
d
ila
ted
s
am
p
lin
g
u
s
es
lear
n
ed
o
f
f
s
ets
to
s
am
p
le
r
elev
an
t
p
o
in
ts
at
d
if
f
er
en
t
s
ca
les
.
Po
s
itio
n
u
p
d
atin
g
a
d
ju
s
ts
r
ef
e
r
en
ce
p
o
in
ts
d
y
n
am
ically
,
atte
n
tio
n
ca
lcu
latio
n
c
o
m
p
u
tes
w
eig
h
ts
o
v
er
s
am
p
led
p
o
s
itio
n
s
,
an
d
f
ea
t
u
r
e
in
teg
r
atio
n
f
u
s
es th
e
r
esu
lts
in
to
r
ic
h
er
r
ep
r
esen
tatio
n
s
.
S
c
a
l
e
a
g
g
r
e
g
a
ti
o
n
:
l
e
t
i
s
a
n
in
p
u
t
f
e
a
t
u
r
e
o
f
t
h
e
S
E
D
A
,
ac
c
o
r
d
i
n
g
t
o
t
h
e
c
o
n
v
o
l
u
t
i
o
n
a
l
m
u
l
t
i
-
h
e
a
d
s
e
l
f
-
a
tt
e
n
t
i
o
n
(
M
H
SA
)
,
t
h
e
v
a
lu
e
i
s
g
e
n
e
r
a
t
e
d
=
[
1
,
2
,
…
]
×
×
×
f
r
o
m
X
u
s
i
n
g
l
i
n
e
a
r
p
r
o
je
c
t
i
o
n
,
W
h
e
r
e
,
×
×
×
i
s
t
h
e
−
ℎ
(
=
1
,
2
,
…
,
)
h
e
a
d
.
I
n
t
h
e
co
n
v
o
l
u
t
i
o
n
a
l
M
HS
A
f
r
a
m
ew
o
r
k
,
a
l
l
t
h
e
a
t
t
e
n
t
i
o
n
h
e
a
d
s
o
p
e
r
a
t
e
o
n
th
e
s
a
m
e
s
c
a
l
e
.
C
o
n
s
e
q
u
e
n
t
l
y
,
t
h
e
i
n
t
e
r
-
l
e
v
e
l
m
u
l
ti
s
ca
l
e
p
r
o
p
e
r
t
y
c
a
n
n
o
t
b
e
e
f
f
e
c
t
i
v
e
l
y
e
x
p
l
o
i
te
d
,
w
h
i
c
h
le
a
d
s
t
o
a
l
i
m
it
e
d
m
u
l
t
is
c
a
le
r
e
p
r
e
s
e
n
t
a
ti
o
n
f
o
r
e
a
c
h
f
e
a
t
u
r
e
m
a
p
.
T
o
s
o
l
v
e
t
h
i
s
i
s
s
u
e
,
t
h
e
p
r
i
n
c
i
p
l
e
o
f
a
d
i
l
at
e
d
t
r
a
n
s
f
o
r
m
e
r
i
s
a
d
o
p
t
e
d
i
n
w
h
ic
h
e
a
c
h
h
e
a
d
s
a
m
p
l
e
f
e
a
t
u
r
e
p
o
i
n
t
u
s
e
s
a
d
i
f
f
e
r
e
n
t
d
i
l
a
t
i
o
n
r
a
t
e
,
h
o
w
e
v
e
r
t
h
e
o
r
i
g
in
a
l
d
i
l
a
t
e
d
t
r
a
n
s
f
o
r
m
e
r
f
a
il
e
d
to
f
a
c
i
l
i
ta
t
e
i
n
te
r
a
c
t
i
o
n
s
b
e
t
we
en
d
i
f
f
e
r
e
n
t
s
c
a
le
s
.
T
h
er
ef
o
r
e,
d
ev
elo
p
in
g
an
ag
g
r
eg
atio
n
ap
p
r
o
ac
h
th
at
co
n
s
o
li
d
ates
M
h
ea
d
s
in
to
R
(
R
<
M)
s
ca
le
h
ea
d
s
th
r
o
u
g
h
d
y
n
am
ic
s
elec
tio
n
w
eig
h
ts
{
}
=
1
is
p
r
e
d
icted
f
r
o
m
t
h
e
v
alu
e
t
h
r
o
u
g
h
g
lo
b
al
av
er
a
g
e
p
o
o
lin
g
(
GAP)
,
a
lin
ea
r
lay
er
,
an
d
So
f
t
Ma
x
ac
tiv
atio
n
.
T
h
e
co
m
p
u
tat
io
n
s
o
f
{
}
=
1
ar
e
s
u
m
m
ar
ized
as in
(
8
)
an
d
let
̃
∈
×
×
×
b
e
th
e
v
alu
e
f
o
r
th
e
−
ℎ
s
c
ale
-
h
ea
d
;
th
en
,
̃
(
=
1
,
2
,
…
.
,
)
is
pr
oduc
e
d
by
(
9
)
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
15
,
No
.
2
,
Ap
r
il
20
26
:
1
4
7
3
-
1
4
8
5
1478
W
h
er
e
,
is
th
e
−
ℎ
elem
en
t
o
f
.
Su
b
s
eq
u
en
tly
,
SEDA
is
ap
p
lie
d
to
th
e
s
ca
le
h
ea
d
s
{
̃
}
=
1
r
esp
ec
tiv
ely
,
th
at
s
h
o
win
g
th
e
d
y
n
am
ic
p
r
esen
tatio
n
o
f
ac
tiv
atio
n
f
u
n
ctio
n
s
.
{
}
=
(
(
(
(
)
)
)
)
(
8
)
̃
=
⨂
=
[
,
1
,
,
2
2
,
…
.
,
,
,
]
(
9
)
Dilated
s
am
p
lin
g
:
f
o
r
d
if
f
er
e
n
t
-
s
ca
le
h
ea
d
s
,
d
if
f
er
en
t
d
ilatio
n
r
ates
ar
e
u
s
ed
to
s
am
p
le
th
e
en
tr
ies
o
f
th
e
v
alu
e
at
th
e
tar
g
et
q
u
er
y
p
o
in
t,
with
o
u
t
lo
s
s
o
f
g
e
n
er
ality
,
th
e
d
ilatio
n
r
ate
is
ass
ig
n
ed
to
th
e
−
ℎ
ℎ
(
=
1
,
2
,
)
,
g
iv
en
t
h
e
q
u
er
y
p
o
s
itio
n
∈
2
o
f
th
e
−
ℎ
q
u
e
r
y
.
T
h
e
v
alu
e
s
am
p
lin
g
s
tr
ateg
ies
f
o
r
d
if
f
er
e
n
t
s
ca
le
h
ea
d
s
ar
e
d
ef
in
ed
as
f
o
llo
ws:
f
o
r
=
1
,
th
e
v
alu
e
p
o
i
n
ts
ar
e
s
am
p
led
with
in
a
s
tan
d
ar
d
K
×K
win
d
o
w
ce
n
ter
ed
at
,
f
o
r
≥
2
,
th
e
s
am
p
lin
g
p
o
s
itio
n
s
ar
e
d
eter
m
i
n
ed
b
y
ap
p
ly
i
n
g
d
ilatio
n
r
ate
to
th
e
s
am
p
lin
g
g
r
id
.
I
n
s
h
o
r
t
d
ilated
s
am
p
lin
g
s
et
Γ
,
f
o
r
q
u
er
y
p
o
in
t
,
at
th
e
−
ℎ
s
ca
le
-
h
ea
d
i
s
g
iv
en
b
y
(
1
0
)
,
a
d
d
itio
n
ally
,
at
th
e
b
o
u
n
d
ar
y
o
f
f
ea
tu
r
e
m
ap
,
z
er
o
-
p
a
d
d
in
g
is
u
s
ed
to
ass
is
t d
ilated
s
am
p
lin
g
.
,
=
{
,
|
,
=
+
.
,
[
−
2
,
2
]
2
}
(
1
0
)
Po
s
itio
n
u
p
d
ate:
th
e
s
tr
ateg
y
i
n
(
1
0
)
s
am
p
les
f
ea
tu
r
e
p
o
i
n
ts
with
in
a
r
eg
u
lar
×
win
d
o
w
f
o
r
t
h
e
−
ℎ
h
ea
d
w
h
er
e
=
+
(
−
1
)
(
−
1
)
,
alth
o
u
g
h
b
asic
d
ilated
s
am
p
lin
g
en
lar
g
es
th
e
r
ec
ep
tiv
e
f
ield
it
lack
s
th
e
f
lex
ib
ilit
y
to
ac
cu
r
at
ely
m
o
d
el
th
e
i
n
d
iv
id
u
al
in
s
tan
ce
s
.
T
o
ad
d
r
ess
th
is
lim
itatio
n
,
a
s
et
o
f
lear
n
a
b
le
p
o
s
itio
n
o
f
f
s
ets
is
in
tr
o
d
u
ce
d
to
u
p
d
ate
th
e
s
am
p
led
p
o
s
itio
n
s
ad
ap
tiv
el
y
th
u
s
th
e
s
am
p
led
p
o
s
itio
n
s
ar
e
u
p
d
ated
as
(
1
1
)
,
wh
er
e
∆
,
∈
2
d
en
o
tes
th
e
o
f
f
s
et
f
o
r
p
o
s
itio
n
s
,
,
d
en
o
tes
b
ilin
ea
r
in
ter
p
o
latio
n
,
f
o
llo
win
g
DAT
.
T
h
e
lo
ca
tio
n
s
,
ar
e
also
n
o
r
m
alize
d
to
th
e
r
a
n
g
e
[
0
,
1
]
t
h
e
o
f
f
s
et
∆
,
∈
[
0
,
1
]
is
p
r
ed
icted
f
r
o
m
th
e
−
ℎ
q
u
er
y
u
s
in
g
a
lin
e
ar
lay
er
th
at
is
{
∆
,
}
=
(
,
)
h
en
ce
f
o
r
t
h
e
−
ℎ
q
u
er
y
at
lo
ca
tio
n
,
,
th
e
v
alu
e
p
o
in
ts
ar
e
s
elec
ted
f
r
o
m
th
e
s
et
Γ
,
to
ca
lcu
late
th
e
atten
tio
n
at
th
e
−
ℎ
s
ca
le
h
ea
d
.
̃
,
=
{
(
,
+
∆
,
)
,
,
∈
,
(
1
1
)
Atten
tio
n
ca
lcu
latio
n
:
b
a
s
ed
o
n
(
8
)
to
(
1
0
)
,
th
e
−
ℎ
v
alu
e
p
o
in
t
f
o
r
th
e
−
ℎ
q
u
er
y
at
th
e
−
ℎ
s
ca
le
-
h
ea
d
,
wh
er
e
̃
=
(
(
,
+
∆
,
)
)
,
Γ
,
,
in
s
tead
o
f
u
s
in
g
d
o
t
p
r
o
d
u
ct
s
im
ilar
i
ty
,
as
in
ty
p
ical
s
elf
-
atten
tio
n
.
T
h
e
atten
tio
n
weig
h
ts
ar
e
p
r
e
d
icted
d
ir
ec
tly
f
r
o
m
th
e
q
u
er
y
u
s
in
g
a
lin
ea
r
lay
er
a
n
d
So
f
tMa
x
f
u
n
ctio
n
f
o
r
ca
lcu
latio
n
.
L
et
,
,
b
e
th
e
atten
tio
n
weig
h
t
b
etwe
en
th
e
−
ℎ
q
u
er
y
an
d
t
h
e
−
ℎ
k
ey
at
th
e
−
ℎ
s
ca
le
h
ea
d
{
,
,
}
|
Γ
,
|
=
1
is
o
b
tain
ed
b
y
(
1
2
)
,
wer
e
|
Γ
,
|
co
m
p
o
n
e
n
ts
th
e
p
o
wer
o
f
s
et
Γ
,
th
en
th
e
atten
tio
n
o
f
th
e
−
ℎ
s
ca
le
-
h
ea
d
is
ca
lcu
lated
as (
1
3
)
.
{
,
,
}
=
1
|
,
|
=
(
(
,
)
)
(
1
2
)
,
=
∑
,
,
|
,
|
=
1
̃
(
(
,
+
∆
,
)
)
(
1
3
)
W
h
er
e
,
d
en
o
tes
th
e
−
ℎ
elem
en
t
in
th
e
o
u
tp
u
t
o
f
th
e
−
ℎ
s
ca
le
-
h
ea
d
s
s
s
h
o
wn
in
(
1
2
)
,
ev
er
y
s
ca
le
h
ea
d
in
SEDA
o
p
er
ates
o
n
th
e
co
m
p
lete
in
p
u
t
f
ea
t
u
r
e,
.
T
h
is
o
p
er
atio
n
is
d
if
f
e
r
en
t
f
r
o
m
a
t
y
p
ical
MH
SA,
in
wh
ich
ea
ch
h
ea
d
o
p
er
ates
o
n
a
p
ar
titi
o
n
o
f
th
e
in
p
u
t
f
ea
t
u
r
e
to
ca
lc
u
late
s
ca
le
h
ea
d
.
T
h
e
r
ef
o
r
e,
th
e
d
ilatio
n
r
ate
d
ef
in
e
d
in
(
1
3
)
ca
n
b
e
s
et
to
an
y
v
alu
e
with
o
u
t
th
e
d
im
en
s
io
n
al
m
is
m
atch
is
s
u
e
th
at
m
ay
o
cc
u
r
in
a
ty
p
ical
MH
SA.
Featu
r
e
in
teg
r
atio
n
:
th
e
f
in
al
o
u
tp
u
t
o
f
SEDA
is
p
r
o
d
u
ce
d
th
r
o
u
g
h
th
e
f
ea
tu
r
e
in
teg
r
atio
n
o
f
all
s
ca
le
h
ea
d
s
s
p
ec
if
ically
,
it
is
o
p
er
at
ed
th
r
o
u
g
h
th
e
co
n
ca
ten
atio
n
o
p
er
atio
n
f
o
llo
wed
b
y
a
lin
ea
r
lay
er
W
O,
wh
ich
is
f
o
r
m
u
lated
as
(
1
4
)
.
W
h
er
e
l
h
as
th
e
s
am
e
d
ef
in
it
io
n
as
in
(
1
5
)
an
d
th
e
atten
tio
n
weig
h
ts
(
∑
∑
,
,
|
Γ
,
|
=
1
̃
(
(
1
(
,
+
∆
,
,
)
.
=
1
=
1
r
ec
allin
g
th
e
o
v
er
all
ar
ch
itect
u
r
e,
th
e
SEDA
is
eq
u
ip
p
ed
with
th
e
s
elf
-
atten
tio
n
o
f
th
e
en
co
d
er
a
n
d
cr
o
s
s
-
atten
tio
n
o
f
th
e
d
ec
o
d
er
wh
e
r
e
=
[
1
,
2
,
…
.
,
]
i
s
th
e
s
p
lit
o
f
ac
co
r
d
in
g
to
t
h
e
d
im
en
s
io
n
s
o
f
{
,
}
=
1
.
T
h
e
d
if
f
er
en
ce
b
etwe
en
th
ese
two
S
E
D
A
ar
r
an
g
em
e
n
ts
is
th
at
th
e
q
u
er
ies
ar
e
d
if
f
er
en
t
s
p
ec
if
ically
q
u
er
ies
in
th
e
s
elf
-
atten
tio
n
o
f
th
e
en
co
d
er
ar
e
f
r
o
m
im
ag
e
f
ea
tu
r
es,
wh
e
r
ea
s
q
u
er
i
es in
th
e
d
ec
o
d
e
r
cr
o
s
s
-
atten
ti
o
n
o
r
i
g
in
ate
f
r
o
m
o
b
ject
q
u
er
i
es.
=
(
1
,
,
2
,
…
.
,
,
)
.
=
∑
,
=
1
.
(
1
4
)
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
S
ca
ler en
h
a
n
ce
d
d
efo
r
ma
b
le
a
tten
tio
n
w
ith
g
r
a
p
h
n
eu
r
a
l n
e
tw
o
r
k
fo
r
…
(
R
ev
a
th
i Ka
s
in
a
th
a
p
eru
ma
l
)
1479
=
∑
(
∑
∑
,
,
|
,
|
=
1
̃
(
(
1
(
,
+
∆
,
,
)
)
=
1
=
1
.
(
1
5
)
2
.
5
.
1
.
Alg
o
rit
hm
f
o
r
pro
po
s
e
d SE
DA
-
G
NN
I
n
p
u
t:
GNN
m
o
d
el
with
p
ar
am
eter
an
d
th
e
tr
ain
in
g
d
ataset
is
UVG
an
d
J
C
T
-
V
C
an
d
r
ate
-
d
is
to
r
tio
n
an
d
p
e
r
ce
p
tio
n
lo
s
s
ar
e
co
n
s
id
er
ed
,
w
h
ich
ar
e
r
ep
r
esen
ted
as
ℒ
(
.
)
,
1
0
0
ep
o
ch
s
wer
e
in
clu
d
ed
,
wh
ich
is
s
h
o
wn
as
,
o
u
tp
u
t:
tr
ain
ed
m
o
d
el
p
ar
am
eter
.
T
h
e
tr
ain
in
g
lo
o
p
u
p
d
atin
g
iter
ativ
ely
o
v
er
1
0
0
ep
o
ch
es
th
er
e
b
y
m
in
im
izin
g
th
e
lo
s
s
tr
ad
eo
f
f
am
o
n
g
c
o
m
p
r
ess
io
n
r
ate,
p
er
ce
p
tu
al
q
u
a
lity
an
d
d
is
to
r
tio
n
q
u
ality
.
T
h
e
f
in
al
tr
ai
n
ed
m
o
d
el
p
ar
a
m
eter
s
wh
ich
b
alan
ce
s
b
est
wh
en
co
m
p
etein
g
lo
s
s
es
ac
r
o
s
s
b
o
th
d
ataset
s
as sh
o
wn
in
Alg
o
r
ith
m
1
.
T
h
e
p
r
o
p
o
s
ed
GNN
ap
p
r
o
ac
h
u
s
ed
SEDA
b
y
co
n
s
id
er
in
g
(
4
)
ad
ap
tiv
el
y
s
am
p
les
ac
r
o
s
s
d
ilated
r
ates
an
d
m
u
ltip
le
s
ca
les,
d
y
n
am
ic
ally
u
p
d
ate
th
eir
p
o
s
itio
n
ac
c
o
r
d
in
g
to
lear
n
ed
o
f
f
s
ets.
T
h
i
s
p
r
o
v
i
d
es
m
o
d
el
to
atten
d
ef
f
ec
tiv
ely
ac
r
o
s
s
tim
e
an
d
s
p
ac
e
f
o
r
s
alien
t
v
id
eo
f
ea
tu
r
es
ef
f
ec
tiv
ely
.
B
y
(
7
)
q
u
an
tifie
s
th
e
lo
s
s
f
u
n
ctio
n
q
u
ality
o
f
v
id
eo
c
o
m
p
r
ess
io
n
to
ca
lcu
late
r
ate
o
f
d
is
to
r
tio
n
.
T
h
e
GNN
m
o
d
el
h
elp
s
to
p
r
e
d
ict
C
U
p
ar
titi
o
n
in
g
with
lo
we
r
co
m
p
lex
ity
an
d
it lea
d
s
to
e
f
f
icien
t e
n
co
d
in
g
b
y
en
h
a
n
cin
g
q
u
ality
o
f
v
id
e
o
.
Alg
o
r
ith
m
1
.
GNN
tr
ain
in
g
p
r
o
ce
d
u
r
e
1
: p
r
o
ce
d
u
r
e
T
r
ain
GNN
(
f
,
d
a
ta
tr
ain
in
g
,
ℒ
,
o
p
tim
izer
,
T
)
2
:
f
o
r
ep
o
ch
=
1
to
T
d
o
3
:
f
o
r
ea
c
h
b
atch
in
_
do
4
:
(
X,
A,
Y)
←
E
x
tr
ac
tFeatu
r
eAn
d
L
ab
els (
)
#
n
o
d
e
f
ea
tu
r
es X,
ad
jace
n
cy
A,
la
b
els Y
5
:
Op
tim
izer
.
ze
r
o
_
g
r
ad
(
)
#
clea
r
p
r
ev
i
o
u
s
g
r
ad
ien
ts
6
:
̂
←
f
(
x
,
A)
u
s
in
g
(
4
)
#
f
o
r
wa
r
d
p
ass
th
r
o
u
g
h
th
e
GNN
7
:
L
o
s
s
←
ℒ
(
̂
,
Y)
u
s
in
g
(
7
)
#
co
m
p
u
te
lo
s
s
8
:
L
o
s
s
.
b
ac
k
war
d
(
)
#
b
ac
k
p
r
o
p
ag
ate
g
r
a
d
ien
ts
9
:
Op
tim
izer
.
s
tep
(
)
#
u
p
d
ate
m
o
d
el
p
a
r
a
m
eter
s
1
0
:
en
d
f
o
r
1
1
:
e
n
d
f
o
r
1
2
:
r
et
u
r
n
1
3
: e
n
d
p
r
o
ce
d
u
r
e
2
.
6
.
No
t
a
t
io
n
t
a
ble
T
h
e
n
o
tatio
n
tab
le
s
u
m
m
ar
ize
s
th
e
all
v
ar
ia
b
le
an
d
th
eir
s
y
m
b
o
ls
u
s
ed
f
o
r
th
e
v
id
eo
co
m
p
r
ess
io
n
;
ea
ch
n
o
tatio
n
d
ef
i
n
es
th
e
p
a
r
a
m
eter
s
r
elate
d
to
th
e
tr
an
s
f
er
co
ef
f
icien
t
an
d
v
i
d
eo
f
r
am
es
m
o
tio
n
esti
m
atio
n
.
T
h
e
en
co
d
in
g
p
r
o
ce
s
s
lik
e
p
r
e
-
p
r
o
ce
s
s
in
g
,
C
U
p
ar
titi
o
n
in
g
,
p
r
ed
ictio
n
an
d
en
t
r
o
p
y
co
d
in
g
r
ep
r
esen
ted
b
y
its
r
elate
d
s
y
m
b
o
ls
.
T
h
e
v
id
eo
c
o
m
p
r
ess
io
n
m
ath
em
atica
l
r
e
p
r
esen
tatio
n
s
ar
e
p
r
o
v
id
ed
in
n
o
tatio
n
tab
le
wh
ich
p
r
o
v
id
es
th
e
clar
ity
an
d
co
n
s
is
ten
cy
to
u
n
d
er
s
tan
d
in
T
ab
l
e
1
.
No
tatio
n
an
d
th
eir
d
escr
ip
tio
n
s
u
s
ed
in
th
e
v
id
eo
co
m
p
r
ess
io
n
.
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
h
e
SEDA
-
GNN
ap
p
r
o
ac
h
is
s
im
u
lated
with
MA
T
L
AB
2
0
2
0
a
en
v
i
r
o
n
m
e
n
t,
an
d
,
win
d
o
ws
1
0
OS,
i5
p
r
o
ce
s
s
o
r
,
an
d
1
6
GB
R
A
M
ar
e
th
e
r
eq
u
ir
ed
s
y
s
tem
co
n
f
ig
u
r
atio
n
s
.
T
h
e
SEDA
-
GNN
p
er
f
o
r
m
an
ce
m
etr
ics
u
s
ed
f
o
r
ev
al
u
atin
g
m
o
d
el
p
er
f
o
r
m
an
ce
a
r
e
b
j
o
n
teg
aa
r
d
d
elta
p
ea
k
-
to
-
s
ig
n
al
n
o
is
e
r
atio
(
B
D
-
PS
NR
)
,
m
u
lti
-
s
ca
le
–
s
tr
u
ctu
r
al
s
im
ilar
ity
in
d
e
x
m
ea
s
u
r
e
(
MS
-
SS
I
M)
an
d
b
j
o
n
teg
aa
r
d
d
elta
b
it
r
ate
(
B
D
-
B
R
)
with
p
er
ce
n
tag
e
o
f
co
d
in
g
tim
e
(
∆
)
,
th
e
m
ath
em
atica
l
ex
p
r
ess
io
n
s
ar
e
p
r
o
v
id
ed
i
n
(
1
6
)
to
(
1
8
)
.
W
h
er
e
an
d
d
en
o
te
th
e
en
co
d
in
g
tim
e,
a
n
d
ar
e
th
e
PS
NR
r
atio
,
an
d
,
d
ef
in
e
th
e
b
itra
te
in
th
e
HM
o
u
tp
u
t
o
f
th
e
p
r
o
p
o
s
ed
m
eth
o
d
,
th
e
r
ate
d
if
f
er
en
ce
is
d
ef
in
ed
as
∆
=
ℎ−
,
wh
er
e
ℎ
an
d
p
r
esen
ts
th
e
lo
w
a
n
d
h
ig
h
p
o
in
ts
o
f
th
e
c
u
r
v
e
r
an
g
e
o
f
R
D,
as
p
r
esen
ted
in
also
,
th
e
b
it
r
ates
ar
e
an
d
o
f
SEDA
-
GNN
an
d
th
e
a
ctu
al
m
eth
o
d
.
△
=
1
4
∑
(
)
−
(
)
(
)
×
100
4
=
1
(
1
6
)
−
=
−
(
1
7
)
−
=
−
×
100
(
1
8
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
15
,
No
.
2
,
Ap
r
il
20
26
:
1
4
7
3
-
1
4
8
5
1480
3
.
1
.
P
er
f
o
r
m
a
nce
a
na
ly
s
is
T
ab
le
1
r
e
p
r
esen
ts
th
e
UVG
an
d
J
C
T
-
VC
d
ataset
clas
s
d
is
tr
ib
u
tio
n
s
with
th
e
r
eso
lu
tio
n
s
izes
an
d
d
if
f
er
en
t
s
eq
u
en
ce
s
.
T
h
e
J
C
T
-
VC
d
atab
ase
v
id
eo
s
eq
u
en
ce
s
ar
e
d
iv
id
ed
in
to
v
ar
io
u
s
r
eso
lu
tio
n
s
,
with
ea
ch
class
h
av
in
g
a
d
if
f
er
en
t
r
eso
lu
tio
n
.
T
ab
le
2
p
r
o
v
id
es
a
d
etail
ed
ex
p
lan
atio
n
o
f
th
e
r
eso
l
u
tio
n
d
is
tr
ib
u
tio
n
s
an
d
s
eq
u
en
ce
s
an
d
p
r
esen
ts
th
e
r
eso
lu
tio
n
o
f
th
e
d
ataset
class
es.
T
ab
le
1
.
Su
m
m
a
r
izes th
e
s
y
m
b
o
ls
an
d
v
a
r
iab
les u
s
ed
in
th
e
v
id
eo
co
m
p
r
ess
io
n
p
r
o
ce
s
s
S
y
mb
o
l
D
e
scri
p
t
i
o
n
2
D
i
r
e
c
t
i
o
n
a
l
e
r
r
o
r
m
e
t
r
i
c
s
f
o
r
f
i
l
t
e
r
se
l
e
c
t
i
o
n
i
n
C
D
EF
I
n
d
e
x
o
v
e
r
d
i
r
e
c
t
i
o
n
a
l
f
i
l
t
e
r
s
o
r
g
r
o
u
p
s
,
P
i
x
e
l
p
b
e
l
o
n
g
i
n
g
t
o
t
h
e
p
i
x
e
l
g
r
o
u
p
P
i
n
d
i
r
e
c
t
i
o
n
d
,
g
r
o
u
p
k
.
ₚ
I
n
t
e
n
s
i
t
y
o
r
v
a
l
u
e
o
f
p
i
x
e
l
p
.
,
M
e
a
n
i
n
t
e
n
s
i
t
y
o
f
p
i
x
e
l
s
i
n
g
r
o
u
p
,
C
o
n
st
r
a
i
n
e
d
d
i
r
e
c
t
i
o
n
a
l
e
n
h
a
n
c
e
m
e
n
t
f
i
l
t
e
r
.
Lo
o
p
r
e
st
o
r
a
t
i
o
n
u
n
i
t
—
p
r
o
c
e
ss
e
s
b
l
o
c
k
s (e
.
g
.
,
6
4
×
6
4
,
1
2
8
×
1
2
8
,
2
5
6
×
2
5
6
)
.
C
o
d
i
n
g
u
n
i
t
(
i
n
H
EV
C
p
a
r
t
i
t
i
o
n
i
n
g
)
.
R
o
o
t
me
a
n
s
sq
u
a
r
e
e
r
r
o
r
—
g
l
o
b
a
l
t
e
x
t
u
r
e
c
o
m
p
l
e
x
i
t
y
me
t
r
i
c
.
Lo
c
a
l
t
e
x
t
u
r
e
c
o
m
p
l
e
x
i
t
y
m
e
a
s
u
r
e
.
ₐ
,
ᵦ
Th
r
e
s
h
o
l
d
s A
a
n
d
B
f
o
r
t
e
x
t
u
r
e
c
o
m
p
l
e
x
i
t
y
d
e
c
i
s
i
o
n
-
ma
k
i
n
g
.
G
N
N
,
G
C
N
,
G
A
N
,
G
A
E
G
r
a
p
h
n
e
u
r
a
l
n
e
t
w
o
r
k
;
g
r
a
p
h
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
t
w
o
r
k
;
g
r
a
p
h
a
t
t
e
n
t
i
o
n
n
e
t
w
o
r
k
;
g
r
a
p
h
a
u
t
o
-
e
n
c
o
d
e
r
.
,
,
,
,
L
–
n
o
r
ma
l
i
z
e
d
L
a
p
l
a
c
i
a
n
ma
t
r
i
x
U
–
e
i
g
e
n
v
e
c
t
o
r
s
o
f
L
λ
–
e
i
g
e
n
v
a
l
u
e
s
–
t
r
a
i
n
a
b
l
e
f
i
l
t
e
r
k
e
r
n
e
l
x
–
n
o
d
e
f
e
a
t
u
r
e
/
d
a
t
a
v
e
c
t
o
r
ₖ
(
⋀
)
C
h
e
b
y
s
h
e
v
p
o
l
y
n
o
mi
a
l
o
f
o
r
d
e
r
k
a
p
p
l
i
e
d
t
o
sc
a
l
e
d
La
p
l
a
c
i
a
n
.
ₘ
ₐ
ₓ
M
a
x
i
m
u
m
e
i
g
e
n
v
a
l
u
e
o
f
L
.
=
(
2
/
ₘ
ₐ
ₓ
)
−
ₙ
S
c
a
l
e
d
La
p
l
a
c
i
a
n
.
′
(
⋀
)
C
h
e
b
y
s
h
e
v
-
b
a
se
d
a
p
p
r
o
x
i
mat
i
o
n
o
f
s
p
e
c
t
r
a
l
c
o
n
v
o
l
u
t
i
o
n
.
,
,
Â
,
Ḋ
A
–
a
d
j
a
c
e
n
c
y
m
a
t
r
i
x
D
–
d
e
g
r
e
e
ma
t
r
i
x
Â
=
A
+
I
ₙ,
Ḋ
=
∑
ⱼ
Â
ᵢⱼ
(
d
e
g
r
e
e
o
f
Â
)
.
ₙ
I
d
e
n
t
i
t
y
m
a
t
r
i
x
o
f
s
i
z
e
n
.
S
ED
A
S
c
a
l
e
-
e
n
h
a
n
c
e
d
d
e
f
o
r
ma
b
l
e
a
t
t
e
n
t
i
o
n
mec
h
a
n
i
sm
.
N
u
mb
e
r
o
f
a
t
t
e
n
t
i
o
n
h
e
a
d
s
i
n
M
H
S
A
.
,
,
F
e
a
t
u
r
e
d
i
me
n
si
o
n
s:
c
h
a
n
n
e
l
s
,
h
e
i
g
h
t
,
w
i
d
t
h
.
=
[
₁
,
₂
,
…
,
]
V
a
l
u
e
v
e
c
t
o
r
s fo
r
e
a
c
h
a
t
t
e
n
t
i
o
n
h
e
a
d
.
(
<
)
N
u
mb
e
r
o
f
a
g
g
r
e
g
a
t
e
d
sca
l
e
-
h
e
a
d
s
a
f
t
e
r
S
ED
A
w
e
i
g
h
t
i
n
g
.
W
e
i
g
h
t
f
o
r
t
h
e
−
th
sca
l
e
-
h
e
a
d
(
a
f
t
e
r
S
o
f
t
M
a
x
)
.
̃
A
g
g
r
e
g
a
t
e
d
v
a
l
u
e
t
e
n
s
o
r
f
o
r
t
h
e
−
th
sca
l
e
-
h
e
a
d
.
,
,
P
ᵢ
–
Q
u
e
r
y
p
o
s
i
t
i
o
n
f
o
r
−
th
q
u
e
r
y
;
–
,
,
v
a
l
u
e
sam
p
l
i
n
g
p
o
si
t
i
o
n
f
o
r
−
th
p
o
i
n
t
i
n
h
e
a
d
r
.
∆
,
Le
a
r
n
a
b
l
e
o
f
f
set
a
p
p
l
i
e
d
t
o
sa
mp
l
i
n
g
p
o
s
i
t
i
o
n
,
,
P
o
si
t
i
o
n
ma
p
p
i
n
g
a
n
d
b
i
l
i
n
e
a
r
i
n
t
e
r
p
o
l
a
t
i
o
n
o
p
e
r
a
t
o
r
.
,
Q
u
e
r
y
v
e
c
t
o
r
a
t
p
o
s
i
t
i
o
n
i
n
h
e
a
d
.
,
,
A
t
t
e
n
t
i
o
n
w
e
i
g
h
t
f
r
o
m
q
u
e
r
y
t
o
k
e
y
a
t
h
e
a
d
,
O
u
t
p
u
t
v
a
l
u
e
a
t
q
u
e
r
y
f
o
r
h
e
a
d
,
a
f
t
e
r
w
e
i
g
h
t
e
d
s
u
mm
a
t
i
o
n
1
,
,
2
,
…
.
,
,
S
e
t
o
f
o
u
t
p
u
t
s fr
o
m
e
a
c
h
sc
a
l
e
-
h
e
a
d
a
t
p
o
si
t
i
o
n
.
Li
n
e
a
r
p
r
o
j
e
c
t
i
o
n
w
e
i
g
h
t
s f
o
r
t
h
e
r
-
ℎ
sc
a
l
e
-
h
e
a
d
.
ᵢ
F
i
n
a
l
i
n
t
e
g
r
a
t
e
d
o
u
t
p
u
t
a
t
p
o
si
t
i
o
n
a
f
t
e
r
c
o
n
c
a
t
e
n
a
t
i
o
n
a
n
d
p
r
o
j
e
c
t
i
o
n
Γ
,
S
e
t
o
f
sa
mp
l
e
d
p
o
s
i
t
i
o
n
s
a
t
sc
a
l
e
-
h
e
a
d
r
f
o
r
q
u
e
r
y
i
T
ab
le
2
.
C
lass
d
is
tr
ib
u
tio
n
o
f
UVG
d
ataset
D
a
t
a
s
e
t
N
u
mb
e
r
o
f
v
i
d
e
o
se
q
u
e
n
c
e
s
(VS)
S
e
q
u
e
n
c
e
R
e
s
o
l
u
t
i
o
n
H
EV
C
-
B
5
C
a
c
t
u
s
B
Q
t
e
r
r
a
c
e
P
a
r
k
s
c
e
n
e
K
i
m
o
n
o
B
a
s
k
e
t
b
a
l
l
d
r
i
v
e
1
9
2
0
×
1
0
8
0
H
EV
C
-
C
4
B
Q
ma
l
l
P
a
r
t
y
sce
n
e
R
a
c
e
h
o
r
s
e
B
a
s
k
e
t
b
a
l
l
d
r
i
l
l
8
3
2
×
4
8
0
H
EV
C
-
D
4
B
a
s
k
e
t
b
a
l
l
p
a
ss
B
Q
sq
u
a
r
e
R
a
c
e
h
o
r
s
e
s
B
l
o
w
i
n
g
b
u
b
b
l
e
s
4
1
6
×
2
4
0
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
S
ca
ler en
h
a
n
ce
d
d
efo
r
ma
b
le
a
tten
tio
n
w
ith
g
r
a
p
h
n
eu
r
a
l n
e
tw
o
r
k
fo
r
…
(
R
ev
a
th
i Ka
s
in
a
th
a
p
eru
ma
l
)
1481
T
a
b
l
e
3
p
r
e
s
e
n
t
s
t
h
e
p
er
f
o
r
m
a
n
c
e
o
f
S
E
D
A
-
G
N
N
ap
p
r
o
a
ch
o
f
v
a
r
io
u
s
f
r
o
m
B
t
o
D
c
l
as
s
e
s
w
i
t
h
m
e
t
r
i
c
s
n
a
m
e
ly
B
D
-
B
R
,
M
S
-
S
S
I
M
,
B
D
-
P
S
N
R
,
a
n
d
∆
,
f
o
r
av
e
r
a
g
e
p
er
f
o
r
m
a
n
c
e
i
n
d
if
f
e
r
e
n
t
c
l
a
s
s
e
s
a
r
e
e
v
a
l
u
a
t
e
d
,
t
h
e
S
E
D
A
-
G
N
N
a
p
p
r
o
a
ch
a
t
ta
i
n
s
an
B
D
-
B
R
a
v
er
a
g
e
o
f
3
.
3
7
2
%
,
B
D
-
P
S
N
R
o
f
-
8
7
.
5
d
B
,
∆
o
f
–
1
5
3
.
0
%
,
a
v
er
a
g
e
a
n
d
-
6
9
.
2
%
o
f
M
S
-
S
S
I
M
.
T
h
e
B
c
l
a
s
s
c
o
n
s
i
s
t
s
o
f
5
f
u
l
l
HD
s
e
q
u
en
c
e
s
w
i
t
h
a
r
e
s
o
l
u
t
io
n
o
f
1
9
2
0
×1
0
8
0
,
w
h
i
c
h
ar
e
u
s
e
d
to
ev
a
l
u
a
t
e
th
e
p
e
r
f
o
r
m
a
n
c
e
o
f
h
i
g
h
-
r
e
s
o
lu
t
io
n
co
d
in
g
i
n
b
o
t
h
l
o
w
-
d
e
l
a
y
co
n
f
i
g
u
r
a
t
io
n
s
a
n
d
r
a
n
d
o
m
a
c
c
e
s
s
,
c
l
a
s
s
C
co
n
t
a
in
s
8
3
2
×
4
8
0
m
id
-
r
e
s
o
lu
ti
o
n
s
e
q
u
e
n
c
e
s
th
a
t
r
e
p
r
e
s
e
n
t
m
o
d
er
a
t
e
-
r
e
s
o
lu
t
i
o
n
co
n
t
en
t
t
a
r
g
e
t
s
f
o
r
b
r
o
a
d
c
a
s
t
a
n
d
s
tr
e
a
m
in
g
.
C
l
a
s
s
D
h
av
i
n
g
4
lo
w
e
r
-
r
e
s
o
l
u
t
io
n
4
1
6
×
2
4
0
s
e
q
u
e
n
ce
s
u
s
e
d
to
co
d
e
c
b
e
h
a
v
io
r
u
n
d
e
r
c
o
n
d
i
t
io
n
s
o
f
co
m
p
a
ct
v
id
e
o
-
l
i
k
e
v
i
d
eo
c
o
n
f
e
r
en
c
in
g
,
t
h
e
p
r
o
p
o
s
e
d
S
E
D
A
-
G
N
N
ap
p
r
o
a
ch
i
s
co
m
p
a
r
e
d
b
a
s
ed
o
n
c
l
a
s
s
e
s
B
,
C
,
D
,
an
d
E
w
i
t
h
r
e
s
p
e
c
t
t
o
th
e
J
C
T
-
V
C
d
a
t
a
s
et
.
T
h
e
d
ev
elo
p
ed
SEDA
-
GNN
m
o
d
el
h
elp
s
in
r
ed
u
cin
g
th
e
co
m
p
u
tatio
n
al
c
o
m
p
lex
ity
with
h
ig
h
f
id
elity
an
d
m
ain
tain
s
a
s
u
p
er
io
r
r
ate
-
d
is
to
r
tio
n
tr
ad
e
-
o
f
f
.
T
h
e
SEDA
-
GNN
ex
h
ib
ited
s
u
p
er
io
r
co
m
p
r
ess
io
n
ef
f
icien
cy
with
s
u
b
s
tan
tial
∆T
r
ed
u
ctio
n
an
d
s
ig
n
if
ican
t
MS
-
SS
I
M
im
p
r
o
v
em
en
t.
T
ab
le
4
co
m
p
ar
is
o
n
b
etwe
en
th
e
p
r
o
p
o
s
ed
m
eth
o
d
an
d
th
e
p
r
o
p
o
s
ed
SEDA
-
GNN
with
o
th
er
ex
is
tin
g
DL
a
p
p
r
o
ac
h
es r
esp
ec
tiv
ely
.
T
ab
le
3
.
C
lass
-
wis
e
p
er
f
o
r
m
an
ce
co
m
p
ar
is
o
n
o
f
th
e
SEDA
-
GNN
ap
p
r
o
ac
h
o
n
J
C
T
-
VC
d
ataset
C
l
a
s
ses
BD
-
B
R
(
%)
BD
-
P
S
N
R
(
%)
∆T
(
%)
MS
-
S
S
I
M
(
%)
B
2
.
4
6
1
-
0
.
0
9
6
-
9
4
.
7
6
0
-
5
3
.
7
1
C
2
.
9
8
3
-
0
.
1
0
4
-
9
4
.
1
3
6
-
5
3
.
3
8
D
3
.
2
0
9
-
0
.
1
3
8
-
9
3
.
8
4
5
-
5
2
.
1
6
E
3
.
1
2
3
-
0
.
1
9
2
-
9
4
.
7
8
6
-
5
3
.
8
1
A
v
e
r
a
g
e
3
.
3
7
2
-
8
7
.
5
-
1
5
3
.
0
-
6
9
.
2
T
ab
le
4
.
Per
f
o
r
m
an
ce
an
aly
s
is
o
f
d
if
f
er
en
t D
L
tech
n
iq
u
es wi
th
SEDA
-
GNN
M
e
t
h
o
d
BD
-
B
R
(
%)
BD
-
P
S
N
R
(
d
B
)
∆T
(
%)
MS
-
S
S
I
M
(
%)
C
N
N
1
.
4
6
1
-
0
.
9
0
6
-
9
5
.
7
6
0
-
5
1
.
7
1
R
N
N
3
.
8
9
3
-
0
.
4
0
1
-
9
3
.
1
4
8
-
5
2
.
4
9
LSTM
2
.
0
2
9
-
0
.
3
8
1
-
9
4
.
8
3
6
-
5
3
.
1
8
V
i
T
-
r
e
t
3
.
4
8
3
-
0
.
2
3
4
-
9
2
.
6
5
4
-
5
4
.
5
4
S
ED
A
-
GNN
2
.
3
7
2
-
8
7
.
5
-
1
5
3
.
0
-
6
9
.
2
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
b
ala
n
ce
s
v
is
u
al
q
u
ality
an
d
co
m
p
r
e
s
s
io
n
g
ain
,
lead
in
g
to
h
ig
h
p
er
ce
p
tu
al
q
u
ality
an
d
l
o
wer
b
it
r
ates.
C
o
m
p
ar
ed
to
C
NN,
R
NN,
L
STM
,
an
d
ViT
-
r
et,
SEDA
-
GNN
ac
h
iev
ed
th
e
b
est
tr
ad
e
-
o
f
f
ac
r
o
s
s
all
k
ey
m
etr
ics
it
d
eliv
er
s
en
h
an
ce
d
v
is
u
al
q
u
ality
an
d
co
m
p
r
ess
io
n
ef
f
icien
cy
,
th
er
eb
y
v
alid
atin
g
its
r
o
b
u
s
t
ar
ch
itect
u
r
e.
T
ab
le
5
r
ep
r
esen
ts
th
e
co
m
p
ar
is
o
n
b
etwe
en
th
e
e
x
is
tin
g
m
eth
o
d
a
n
d
with
p
r
o
p
o
s
ed
SEDA
-
GNN
to
attain
s
b
etter
MS
-
SS
I
M
r
ate.
T
h
e
SEDA
-
GNN
ap
p
r
o
ac
h
h
elp
s
in
m
ess
ag
e
p
ass
in
g
b
y
ca
p
tu
r
in
g
b
o
th
tem
p
o
r
al
a
n
d
s
p
atial
d
ep
en
d
e
n
cies
wh
en
co
m
p
ar
ed
to
o
th
er
tr
ad
itio
n
al
m
et
h
o
d
s
.
SEDA
-
GNN
m
ain
tain
s
co
n
s
is
ten
t
p
er
f
o
r
m
a
n
ce
ac
r
o
s
s
v
ar
io
u
s
co
n
te
n
t
class
es,
with
an
av
er
ag
e
B
D
-
B
R
o
f
o
n
ly
3
.
3
2
%,
th
e
h
ig
h
MS
-
SS
I
M
s
co
r
es
ac
r
o
s
s
s
ce
n
es
f
u
r
th
er
co
n
f
ir
m
ed
th
eir
ad
ap
tab
ilit
y
to
d
if
f
e
r
en
t
v
id
eo
ch
ar
ac
ter
is
tics
.
T
ab
le
6
r
ep
r
e
s
en
ts
a
co
m
p
ar
is
o
n
b
etwe
en
th
e
tr
ad
itio
n
al
m
eth
o
d
with
p
r
o
p
o
s
ed
SEDA
-
GNN
ap
p
r
o
ac
h
h
a
n
d
les
B
D
-
B
R
ac
r
o
s
s
all
UVG
class
es
an
d
th
e
a
b
ilit
y
to
g
en
e
r
alize
c
o
m
p
r
ess
io
n
v
is
u
al
f
id
elity
a
n
d
ef
f
icien
cy
b
ased
o
n
s
ce
n
e
co
m
p
lex
ity
.
T
ab
le
5
.
Per
f
o
r
m
an
ce
an
aly
s
i
s
o
f
SEDA
-
GNN
with
ex
is
tin
g
m
eth
o
d
s
M
e
t
h
o
d
BD
-
B
R
(
%)
BD
-
P
S
N
R
(
d
B
)
∆T
(
%)
MS
-
S
S
I
M
(
%)
C
N
N
1
.
4
6
1
-
0
.
9
0
6
-
9
5
.
7
6
0
-
5
1
.
7
1
R
N
N
3
.
8
9
3
-
0
.
4
0
1
-
9
3
.
1
4
8
-
5
2
.
4
9
LSTM
2
.
0
2
9
-
0
.
3
8
1
-
9
4
.
8
3
6
-
5
3
.
1
8
V
i
T
-
r
e
t
3
.
4
8
3
-
0
.
2
3
4
-
9
2
.
6
5
4
-
5
4
.
5
4
S
ED
A
-
GNN
2
.
3
7
2
-
8
7
.
5
-
1
5
3
.
0
-
6
9
.
2
T
ab
le
6
.
Per
f
o
r
m
an
ce
o
f
SED
A
-
GNN
ap
p
r
o
ac
h
f
o
r
class
es u
s
in
g
UVG
d
ataset
C
l
a
s
ses
BD
-
B
R
(
%)
BD
-
P
S
N
R
(
d
B
)
∆T
(
%)
MS
-
S
S
I
M
(
%)
B
e
a
u
t
y
3
.
8
1
0
-
0
.
8
4
3
-
8
4
.
3
2
4
-
4
3
.
2
6
Jo
c
k
e
y
3
.
4
7
7
-
0
.
2
6
3
-
8
5
.
5
4
9
-
4
4
.
8
8
C
i
t
y
A
l
l
e
y
2
.
8
2
4
-
0
.
9
2
9
-
8
5
.
7
0
4
-
4
2
.
7
4
F
l
o
w
e
r
F
o
c
u
s
3
.
9
8
9
-
0
.
8
6
2
-
8
6
.
3
2
6
-
4
5
.
3
9
S
u
n
B
a
t
h
3
.
7
3
6
-
0
.
4
3
2
-
8
6
.
8
1
7
-
4
5
.
6
8
A
v
e
r
a
g
e
3
.
3
2
0
-
0
.
2
0
3
-
8
5
.
6
7
3
-
4
4
.
7
9
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
15
,
No
.
2
,
Ap
r
il
20
26
:
1
4
7
3
-
1
4
8
5
1482
3
.
2
.
Co
m
pa
ra
t
iv
e
a
na
ly
s
is
I
n
t
h
is
s
ec
t
io
n
,
t
h
e
SE
DA
-
GN
N
a
p
p
r
o
a
ch
is
c
o
m
p
a
r
e
d
wit
h
t
r
a
d
it
io
n
al
m
et
h
o
d
s
,
s
u
ch
as
C
R
L
VC
[
2
1
]
an
d
T
C
M
[
2
2
]
,
o
n
t
h
e
UVG
d
ataset,
in
T
ab
le
6
th
e
SEDA
-
GNN
ap
p
r
o
ac
h
was
ev
alu
ated
u
s
in
g
th
e
d
atab
ase
JCT
-
VC
,
wh
er
ea
s
T
ab
le
7
s
h
o
ws
th
e
r
esu
lts
f
o
r
th
e
J
C
T
-
V
C
d
ataset.
I
n
T
ab
le
s
7
an
d
8
,
f
o
r
v
id
eo
s
eq
u
e
n
ce
s
in
th
e
J
C
T
-
VC
d
ataset,
SEDA
-
GNN
attain
s
less
B
D
-
B
R
o
f
5
.
8
5
%
o
n
a
v
er
ag
e
,
an
d
f
o
r
th
e
tr
af
f
ic
s
eq
u
en
ce
s
,
attain
s
4
.
3
2
% r
esp
ec
tiv
ely
,
b
y
u
tili
zin
g
a
g
r
ap
h
-
b
ased
co
n
t
ex
t.
SEDA
-
GNN
h
elp
s
to
ac
h
iev
e
b
o
th
s
u
s
tain
ab
le
MS
-
SS
I
M
an
d
B
D
-
r
ate,
th
er
eb
y
p
r
o
v
id
in
g
its
ca
p
ab
ilit
y
to
co
m
p
r
ess
th
e
d
ata
m
o
r
e
ef
f
icien
t
ly
.
T
ab
le
7
.
C
o
m
p
a
r
ativ
e
an
aly
s
is
o
f
th
e
e
x
is
tin
g
m
eth
o
d
s
with
SEDA
-
GNN
in
ter
m
s
o
f
th
e
UVG
d
ataset
D
a
t
a
s
e
t
s
M
e
t
h
o
d
s
BD
-
r
a
t
e
(
%)
MS
-
S
S
I
M
(
%)
F
I
D
K
I
D
LPI
P
S
UVG
C
R
LV
C
[
2
1
]
-
4
3
.
1
-
3
1
.
6
-
4
9
.
9
2
1
.
8
8
5
.
2
3
0
.
0
5
7
.
1
2
5
.
2
-
1
7
.
7
-
2
0
.
6
1
5
.
8
-
3
2
.
5
5
.
7
-
5
7
.
2
1
8
.
5
-
1
8
.
5
-
5
4
.
8
-
6
3
.
4
2
5
.
7
-
2
5
.
6
-
4
8
.
5
-
4
7
.
8
-
3
0
.
5
-
4
5
.
2
-
1
1
.
7
-
3
1
.
8
-
3
2
.
5
-
6
5
.
3
S
ED
A
-
GNN
3
.
2
3
0
-
8
4
.
5
-
4
3
.
6
-
7
6
.
2
JC
T
-
VC
C
R
LV
C
-
1
0
.
5
-
6
5
.
3
-
1
2
.
7
-
1
2
.
4
-
1
2
.
8
-
1
5
.
9
-
3
2
.
6
-
3
2
.
5
-
1
5
.
2
-
2
5
.
2
-
3
2
.
8
-
2
4
.
6
-
3
2
.
5
-
3
6
.
7
-
6
5
.
2
-
5
0
.
3
S
ED
A
-
GNN
-
4
7
.
8
-
4
7
.
8
-
7
7
.
1
+
7
2
.
4
I
n
T
ab
le
8
,
t
h
e
p
r
o
p
o
s
ed
S
E
DA
-
GNN
m
eth
o
d
ac
h
iev
es
ef
f
icien
t
B
D
r
ed
u
ctio
n
an
d
MS
-
SS
I
M
o
u
tp
er
f
o
r
m
in
g
th
e
T
C
M
[
2
2
]
m
eth
o
d
wh
ich
h
ig
h
lig
h
ts
its
e
f
f
icien
cy
in
co
m
p
r
ess
in
g
HE
VC
-
R
G
B
d
ata
wh
ile
p
r
eser
v
in
g
h
i
g
h
p
er
ce
p
tu
al
an
d
r
ec
o
n
s
tr
u
ctio
n
q
u
ality
.
I
n
th
e
tab
le
,
th
e
p
r
o
p
o
s
ed
m
eth
o
d
co
m
p
ar
is
o
n
is
g
iv
en
b
etwe
en
PS
NR
v
alu
e
an
d
MS
-
SS
I
M
r
ate
with
th
e
tr
ad
itio
n
al
m
eth
o
d
with
r
esp
ec
t
to
th
e
class
e
s
.
W
h
il
e
m
an
ag
in
g
t
h
e
PS
NR
an
d
B
D
-
r
ate,
th
e
p
r
o
p
o
s
ed
SEDA
-
GN
N
m
eth
o
d
attain
s
ef
f
icien
c
y
i
n
p
er
ce
p
t
u
al
q
u
ality
b
y
o
p
tim
izin
g
th
e
g
r
ap
h
-
b
ased
r
ep
r
esen
tatio
n
ac
r
o
s
s
tem
p
o
r
a
l a
n
d
s
p
atial
d
o
m
ain
s
.
T
ab
le
8
.
Per
f
o
r
m
an
ce
o
f
th
e
SEDA
-
GNN
m
eth
o
d
f
o
r
th
e
J
C
T
-
VC
d
ataset
in
ter
m
s
o
f
class
es
D
a
t
a
s
e
t
M
e
t
h
o
d
P
S
N
R
(
ΔB
D
-
r
a
t
e
)
MS
-
S
S
I
M
(
ΔB
D
-
r
a
t
e
)
H
EV
C
-
R
G
B
TC
M
[
2
2
]
-
1
5
.
2
-
4
8
.
5
+
4
.
7
-
4
7
.
2
-
1
0
.
2
-
5
5
.
1
-
6
.
4
-
5
3
.
9
-
2
3
.
0
-
5
1
.
4
-
2
1
.
0
-
3
5
.
0
O
v
e
r
a
l
l
(
P
S
N
R
)
S
ED
A
-
GNN
-
1
4
.
4
-
3
7
.
7
I
n
T
ab
le
9
,
th
e
SEDA
-
GNN
ap
p
r
o
ac
h
is
co
m
p
ar
ed
with
o
t
h
er
ex
is
tin
g
d
ee
p
lear
n
in
g
m
et
h
o
d
s
u
s
in
g
d
if
f
er
en
t p
e
r
f
o
r
m
an
ce
m
etr
ics wh
ich
h
av
e
th
e
h
ig
h
est M
S
-
S
SIM
g
ain
s
an
d
th
e
lo
west B
D
-
B
R
an
d
△
T
v
alu
es.
T
h
is
co
m
p
ar
is
o
n
s
h
o
wed
b
ett
er
v
is
u
al
q
u
ality
an
d
s
u
p
e
r
io
r
r
ate
-
d
is
to
r
tio
n
ef
f
icien
c
y
f
o
r
th
e
J
C
T
-
VC
an
d
UVG
d
atasets
with
r
esp
ec
t
to
an
o
t
h
er
ex
is
tin
g
alg
o
r
it
h
m
.
T
h
e
s
u
p
er
io
r
r
ate
-
d
is
to
r
tio
n
ef
f
icien
c
y
o
f
co
m
p
r
ess
ed
v
id
e
o
s
is
m
o
r
e
ef
f
ec
tiv
e
wh
en
m
a
n
ag
in
g
an
d
e
n
h
an
cin
g
th
e
q
u
ality
b
it
r
ate
wh
en
co
m
p
ar
ed
with
o
th
er
ex
is
tin
g
m
et
h
o
d
s
.
T
ab
le
9
.
Per
f
o
r
m
an
ce
o
f
th
e
SEDA
-
GNN
m
eth
o
d
with
o
th
e
r
d
ee
p
lear
n
i
n
g
alg
o
r
ith
m
s
M
e
t
h
o
d
BD
-
B
R
(
%)
BD
-
P
S
N
R
(
d
B
)
△
T
P
S
N
R
(
%)
MS
-
S
S
I
M
(
%)
V
M
A
F
A
l
e
x
N
e
t
8
.
4
3
-
0
.
5
6
-
6
8
.
9
6
-
2
9
.
8
2
-
2
9
.
6
7
7
3
.
4
0
R
e
sN
e
t
7
.
0
8
-
0
.
4
9
-
7
1
.
3
1
-
2
7
.
9
3
-
2
6
.
4
5
8
2
.
3
8
C
N
N
5
.
6
8
-
0
.
4
8
-
6
5
.
5
3
-
2
1
.
9
2
-
3
1
.
3
4
7
3
.
4
8
V
i
T
8
.
8
5
-
0
.
4
7
-
6
3
.
2
3
-
2
9
.
0
2
-
2
3
.
8
7
6
9
.
0
3
S
w
i
n
t
r
a
n
sf
o
r
mer
1
2
.
3
2
-
0
.
3
9
-
6
2
.
2
0
-
2
8
.
0
8
-
2
3
.
2
6
7
2
.
8
2
H
y
b
r
i
d
C
N
N
-
V
i
T
1
2
.
6
7
-
0
.
4
5
-
6
5
.
4
3
-
5
2
.
1
8
-
4
7
.
7
6
8
7
.
6
7
S
W
D
A
-
C
N
N
7
.
3
0
-
0
.
3
8
-
8
5
.
8
7
-
4
2
.
1
9
-
3
7
.
7
6
8
7
.
6
7
HDVC
6
.
4
9
+
0
.
3
5
-
8
9
.
7
6
+
3
2
.
9
3
+
4
0
.
9
2
-
P
r
o
p
o
se
d
S
ED
A
-
G
N
N
f
o
r
U
V
G
5
.
8
5
+
0
.
2
5
-
9
1
.
8
8
+
3
4
.
1
2
+
4
2
.
3
0
9
0
.
8
9
P
r
o
p
o
se
d
S
ED
A
-
G
N
N
f
o
r
J
C
T
-
VC
4
.
3
2
0
.
2
4
-
9
0
.
7
7
+
2
4
.
2
1
-
3
2
.
0
3
8
9
.
0
8
Evaluation Warning : The document was created with Spire.PDF for Python.