I
nte
rna
t
io
na
l J
o
urna
l o
f
I
nfo
rm
a
t
ics a
nd
Co
m
m
un
ica
t
io
n T
ec
hn
o
lo
g
y
(
I
J
-
I
CT
)
Vo
l.
15
,
No
.
3
,
Sep
tem
b
er
20
26
,
p
p
.
9
3
5
~
9
4
3
I
SS
N:
2252
-
8
7
7
6
,
DOI
:
1
0
.
1
1
5
9
1
/iji
ct
.
v15
i
3
.
pp
935
-
9
4
3
935
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ict.
ia
esco
r
e.
co
m
Cro
ss
-
mo
da
l atte
ntion fusio
n
using
vis
io
n t
ra
nsfo
rm
ers
for
ro
bust stud
ent
att
entivenes
s es
tima
tion
Ra
j
a
s
ek
a
r
a
n
M
a
riswa
m
y
,
P
ra
v
ee
n Sun
da
r
D
e
p
a
r
t
me
n
t
o
f
C
o
mp
u
t
e
r
S
c
i
e
n
c
e
,
A
d
h
i
p
a
r
a
s
a
k
t
h
i
C
o
l
l
e
g
e
o
f
A
r
t
s
a
n
d
S
c
i
e
n
c
e
,
T
h
i
r
u
v
a
l
l
u
v
a
r
U
n
i
v
e
r
si
t
y
,
V
e
l
l
o
r
e
,
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
J
u
n
2
1
,
2
0
2
4
R
ev
is
ed
Ap
r
2
2
,
2
0
2
6
Acc
ep
ted
Ma
y
7
,
2
0
2
6
Au
to
m
a
ted
st
u
d
e
n
t
a
tt
e
n
t
iv
e
n
e
ss
e
stim
a
ti
o
n
is
a
fu
n
d
a
m
e
n
tal
c
o
m
p
o
n
e
n
t
o
f
in
telli
g
e
n
t
e
-
lea
rn
i
n
g
sy
ste
m
s
a
n
d
a
d
a
p
ti
v
e
c
las
sro
o
m
a
n
a
l
y
ti
c
s.
Trad
it
io
n
a
l
c
o
n
v
o
lu
ti
o
n
a
l
a
n
d
re
c
u
rre
n
t
a
rc
h
i
tec
tu
re
s
o
ften
str
u
g
g
le
to
m
o
d
e
l
l
o
n
g
-
ra
n
g
e
tem
p
o
ra
l
d
e
p
e
n
d
e
n
c
ies
a
n
d
c
o
m
p
lex
in
ter
-
m
o
d
a
l
re
latio
n
sh
i
p
s
i
n
h
e
re
n
t
in
e
n
g
a
g
e
m
e
n
t
b
e
h
a
v
io
r
.
T
o
a
d
d
re
s
s
th
e
se
li
m
it
a
ti
o
n
s,
th
is
p
a
p
e
r
p
ro
p
o
se
s
a
c
ro
ss
-
m
o
d
a
l
a
tt
e
n
ti
o
n
fu
si
o
n
fra
m
e
wo
rk
b
u
il
t
u
p
o
n
a
v
isi
o
n
tr
a
n
sfo
rm
e
r
(ViT)
b
a
c
k
b
o
n
e
fo
r
ro
b
u
st
st
u
d
e
n
t
a
tt
e
n
ti
v
e
n
e
ss
e
stim
a
ti
o
n
.
T
h
e
p
ro
p
o
se
d
a
rc
h
it
e
c
tu
re
lev
e
ra
g
e
s
p
a
tch
-
b
a
se
d
v
isu
a
l
e
n
c
o
d
i
n
g
th
ro
u
g
h
a
ViT
to
c
a
p
tu
re
g
lo
b
a
l
sp
a
ti
a
l
d
e
p
e
n
d
e
n
c
ies
,
wh
i
le
b
e
h
a
v
io
ra
l
c
u
e
s
su
c
h
a
s
g
a
z
e
d
irec
ti
o
n
,
h
e
a
d
p
o
se
,
a
n
d
b
li
n
k
d
y
n
a
m
ics
a
re
e
m
b
e
d
d
e
d
i
n
to
a
sh
a
r
e
d
late
n
t
re
p
re
se
n
tatio
n
s
p
a
c
e
.
A
c
ro
ss
-
m
o
d
a
l
m
u
lt
i
-
h
e
a
d
a
tt
e
n
ti
o
n
m
e
c
h
a
n
ism
is
in
tro
d
u
c
e
d
to
d
y
n
a
m
ica
ll
y
lea
rn
in
tera
c
ti
o
n
s
b
e
twe
e
n
v
isu
a
l
a
n
d
b
e
h
a
v
i
o
ra
l
m
o
d
a
li
ti
e
s,
re
p
lac
in
g
sta
ti
c
we
ig
h
ted
fu
si
o
n
stra
teg
ies
.
Tem
p
o
ra
l
d
y
n
a
m
ics
a
re
m
o
d
e
led
u
si
n
g
a
Tran
sfo
rm
e
r
e
n
c
o
d
e
r,
e
n
a
b
li
n
g
e
ffe
c
ti
v
e
l
o
n
g
-
ra
n
g
e
se
q
u
e
n
c
e
m
o
d
e
li
n
g
wi
th
o
u
t
re
c
u
r
re
n
t
d
e
p
e
n
d
e
n
c
ies
.
E
x
p
e
rime
n
tal
e
v
a
lu
a
ti
o
n
o
n
a
b
e
n
c
h
m
a
rk
a
tt
e
n
t
iv
e
n
e
ss
d
a
tas
e
t
d
e
m
o
n
stra
tes
su
p
e
rio
r
p
e
rfo
rm
a
n
c
e
c
o
m
p
a
re
d
to
CNN
–
LS
T
M
-
b
a
se
d
m
o
d
e
ls,
a
c
h
iev
i
n
g
imp
ro
v
e
d
a
c
c
u
ra
c
y
,
F
1
-
sc
o
re
,
a
n
d
ro
b
u
st
n
e
ss
u
n
d
e
r
c
h
a
l
len
g
i
n
g
li
g
h
t
i
n
g
a
n
d
o
c
c
lu
sio
n
c
o
n
d
it
io
n
s
.
Ab
latio
n
stu
d
ies
v
a
li
d
a
te
th
e
c
o
n
tri
b
u
t
io
n
o
f
c
ro
ss
-
m
o
d
a
l
a
tt
e
n
ti
o
n
a
n
d
tran
sfo
rm
e
r
-
b
a
se
d
tem
p
o
ra
l
m
o
d
e
li
n
g
.
Th
e
p
r
o
p
o
se
d
fra
m
e
wo
rk
m
a
in
tain
s
re
a
l
-
ti
m
e
fe
a
sib
il
it
y
wh
i
le sig
n
ifi
c
a
n
tl
y
e
n
h
a
n
c
i
n
g
d
isc
rimin
a
ti
v
e
c
a
p
a
b
il
it
y
.
K
ey
w
o
rd
s
:
C
ar
d
io
v
ascu
lar
d
is
ea
s
e
Dee
p
lear
n
in
g
E
ch
o
ca
r
d
i
o
g
r
am
im
a
g
in
g
Io
T
Pre
d
ictio
n
VGNet
-
C
NN
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
ajasek
ar
an
Ma
r
is
wam
y
Dep
ar
tm
en
t o
f
C
o
m
p
u
ter
Scie
n
ce
,
Ad
h
ip
a
r
asak
th
i Co
lleg
e
o
f
Ar
ts
an
d
Scien
ce
,
T
h
ir
u
v
allu
v
ar
Un
iv
er
s
ity
Vello
r
e
,
I
n
d
ia
E
m
ail:
m
ar
ajasek
ar
an
@
g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
C
ar
d
io
v
ascu
lar
d
is
ea
s
es
(
C
VDs)
r
em
ain
th
e
lea
d
in
g
ca
u
s
e
o
f
d
ea
th
g
l
o
b
ally
,
ac
c
o
u
n
ti
n
g
f
o
r
an
esti
m
ated
.
T
h
e
r
ap
id
ex
p
an
s
io
n
o
f
o
n
lin
e
ed
u
ca
tio
n
an
d
v
i
r
tu
al
class
r
o
o
m
s
h
as f
u
n
d
am
en
t
ally
tr
an
s
f
o
r
m
ed
th
e
lan
d
s
ca
p
e
o
f
teac
h
in
g
an
d
lear
n
in
g
.
Ma
s
s
iv
e
o
p
en
o
n
lin
e
c
o
u
r
s
es
(
MO
OC
s
)
,
h
y
b
r
id
i
n
s
tr
u
ctio
n
al
m
o
d
els,
a
n
d
f
u
lly
r
em
o
te
a
ca
d
em
ic
p
r
o
g
r
am
s
h
av
e
en
a
b
led
s
ca
lab
le
ac
ce
s
s
to
ed
u
ca
tio
n
ac
r
o
s
s
g
eo
g
r
a
p
h
ical
an
d
s
o
cio
-
ec
o
n
o
m
ic
b
o
u
n
d
a
r
ies.
Ho
wev
er
,
th
is
d
ig
ital
tr
an
s
f
o
r
m
atio
n
h
as
also
in
tr
o
d
u
ce
d
a
cr
i
tical
ch
allen
g
e:
th
e
d
if
f
icu
lty
o
f
m
o
n
ito
r
in
g
s
tu
d
en
t
atten
tiv
en
ess
in
t
h
e
ab
s
e
n
ce
o
f
p
h
y
s
ical
class
r
o
o
m
s
u
p
er
v
is
io
n
[
1
]
-
[
3
]
.
Un
lik
e
tr
ad
itio
n
al
in
-
p
e
r
s
o
n
en
v
ir
o
n
m
en
ts
wh
er
e
in
s
tr
u
ct
o
r
s
ca
n
o
b
s
er
v
e
f
ac
ial
e
x
p
r
e
s
s
io
n
s
,
p
o
s
tu
r
e,
an
d
b
eh
av
io
r
al
cu
es
to
g
au
g
e
en
g
ag
em
en
t,
o
n
lin
e
p
latf
o
r
m
s
o
f
ten
lack
r
eliab
le
m
ec
h
an
is
m
s
to
ass
ess
wh
eth
er
lear
n
er
s
ar
e
atten
tiv
e,
d
is
tr
ac
ted
,
o
r
d
is
e
n
g
a
g
ed
.
C
o
n
s
eq
u
en
tly
,
au
to
m
ated
atten
tiv
en
ess
m
o
n
ito
r
in
g
h
as
em
er
g
ed
as
a
c
r
u
cial
c
o
m
p
o
n
en
t
o
f
in
tellig
en
t
e
-
lea
r
n
in
g
s
y
s
tem
s
,
en
ab
lin
g
a
d
ap
tiv
e
co
n
ten
t
d
eliv
er
y
,
p
er
s
o
n
alize
d
f
ee
d
b
ac
k
,
an
d
d
a
ta
-
d
r
iv
en
p
ed
ag
o
g
ical
in
ter
v
en
tio
n
s
[
4
]
-
[
6
]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
15
,
No
.
3
,
Sep
tem
b
er
20
26
:
935
-
9
4
3
936
Stu
d
en
t
atten
tiv
en
e
s
s
is
in
h
er
en
tly
a
m
u
lti
-
d
im
en
s
io
n
al
co
n
s
tr
u
ct
th
at
m
an
if
ests
th
r
o
u
g
h
v
is
u
al
an
d
b
eh
av
io
r
al
s
ig
n
als
s
u
ch
as
g
a
ze
d
ir
ec
tio
n
,
h
ea
d
o
r
ien
tatio
n
,
f
ac
ial
m
icr
o
-
e
x
p
r
ess
io
n
s
,
b
li
n
k
r
ate,
a
n
d
p
o
s
tu
r
e
d
y
n
am
ics.
T
h
ese
s
ig
n
als
ev
o
l
v
e
o
v
e
r
tim
e
a
n
d
i
n
ter
ac
t
in
c
o
m
p
lex
wa
y
s
.
R
ec
en
t
ad
v
a
n
ce
s
in
co
m
p
u
ter
v
is
io
n
an
d
d
ee
p
lear
n
in
g
h
a
v
e
f
ac
ili
tated
th
e
d
ev
el
o
p
m
en
t
o
f
a
u
t
o
m
ated
en
g
ag
em
en
t
d
etec
tio
n
s
y
s
tem
s
,
p
r
im
ar
ily
lev
er
ag
in
g
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NNs)
f
o
r
s
p
atial
f
ea
tu
r
e
e
x
tr
ac
tio
n
a
n
d
r
ec
u
r
r
en
t
n
eu
r
al
n
etwo
r
k
s
(
R
NNs),
p
ar
ticu
lar
ly
lo
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
L
STM
)
m
o
d
els,
f
o
r
tem
p
o
r
al
m
o
d
elin
g
.
W
h
ile
th
ese
ar
ch
itectu
r
es
h
av
e
ac
h
iev
ed
m
o
d
er
ate
s
u
cc
ess
,
th
ey
ex
h
ib
it
n
o
tab
le
lim
itatio
n
s
wh
en
ap
p
lied
to
th
e
n
u
an
ce
d
task
o
f
atten
tiv
en
ess
esti
m
atio
n
in
u
n
co
n
s
tr
ain
ed
o
n
lin
e
e
n
v
i
r
o
n
m
en
ts
[
7
]
-
[
9
]
.
C
NN
-
b
ased
ap
p
r
o
ac
h
es
ex
c
el
at
ca
p
tu
r
in
g
lo
ca
l
s
p
atial
p
atter
n
s
th
r
o
u
g
h
h
ier
ar
c
h
ic
al
f
ea
tu
r
e
ex
tr
ac
tio
n
.
Ho
wev
e
r
,
th
eir
in
h
er
en
t
in
d
u
ctiv
e
b
ias
to
war
d
lo
ca
lity
r
estricts
th
eir
ab
ilit
y
to
m
o
d
el
g
lo
b
a
l
co
n
tex
tu
al
r
elatio
n
s
h
ip
s
with
in
an
im
ag
e.
E
n
g
a
g
em
en
t
-
r
ela
ted
cu
es
o
f
ten
r
eq
u
ir
e
h
o
lis
tic
in
ter
p
r
etatio
n
;
f
o
r
in
s
tan
ce
,
s
u
b
tle
co
m
b
in
atio
n
s
o
f
ey
e
g
az
e
d
ev
iatio
n
an
d
h
ea
d
p
o
s
e
s
h
if
ts
m
ay
in
d
icate
co
g
n
itiv
e
d
is
en
g
ag
em
en
t.
Stan
d
ar
d
C
NN
ar
ch
itectu
r
es
r
ely
o
n
p
r
o
g
r
e
s
s
iv
ely
s
tack
ed
co
n
v
o
lu
tio
n
al
f
ilter
s
to
en
lar
g
e
th
e
r
ec
ep
tiv
e
f
ield
,
b
u
t
th
is
p
r
o
c
ess
r
em
ain
s
co
n
s
tr
ain
ed
b
y
s
p
atial
lo
ca
lity
an
d
d
ep
th
-
d
ep
en
d
en
t
ag
g
r
eg
atio
n
.
C
o
n
s
eq
u
en
tly
,
C
NN
-
b
ased
s
y
s
tem
s
m
ay
s
tr
u
g
g
le
to
ca
p
tu
r
e
l
o
n
g
-
r
a
n
g
e
s
p
atial
d
e
p
en
d
e
n
cies
th
at
ar
e
ess
en
tial
f
o
r
r
o
b
u
s
t a
tten
ti
v
en
ess
in
f
er
e
n
ce
[
1
0
]
-
[
1
2
]
.
T
o
ad
d
r
ess
tem
p
o
r
al
d
y
n
am
i
cs,
m
an
y
ex
is
tin
g
w
o
r
k
s
in
c
o
r
p
o
r
ate
L
STM
o
r
g
ated
r
ec
u
r
r
en
t
u
n
it
(
GR
U)
n
etwo
r
k
s
.
R
ec
u
r
r
en
t
ar
ch
itectu
r
es
ar
e
d
esig
n
ed
to
p
r
o
ce
s
s
s
eq
u
en
tial
d
ata
b
y
m
ain
tain
in
g
h
id
d
en
s
tates
th
at
p
r
o
p
a
g
ate
in
f
o
r
m
atio
n
a
cr
o
s
s
tim
e
s
tep
s
.
Alt
h
o
u
g
h
L
STM
s
m
itig
ate
th
e
v
an
is
h
in
g
g
r
ad
ie
n
t
p
r
o
b
lem
r
elativ
e
to
v
a
n
illa
R
NNs,
th
ey
s
till
ex
h
ib
it
lim
itatio
n
s
in
m
o
d
elin
g
lo
n
g
-
r
a
n
g
e
d
e
p
en
d
e
n
cies,
p
ar
ticu
lar
ly
in
ex
ten
d
ed
v
i
d
e
o
s
eq
u
en
ce
s
.
T
h
e
s
eq
u
e
n
tial
n
atu
r
e
o
f
r
ec
u
r
r
e
n
ce
in
tr
o
d
u
ce
s
co
m
p
u
tatio
n
al
in
ef
f
icien
cies
an
d
r
estricts
p
ar
alleliza
tio
n
d
u
r
in
g
tr
ain
in
g
[
1
3
]
-
[
1
7
]
.
Mo
r
e
o
v
er
,
th
e
m
em
o
r
y
m
ec
h
an
is
m
o
f
L
STM
s
m
ay
n
o
t
ef
f
ec
tiv
ely
c
ap
tu
r
e
co
m
p
lex
tem
p
o
r
al
in
ter
ac
tio
n
s
s
p
an
n
in
g
lo
n
g
in
ter
v
al
s
,
esp
ec
ially
wh
en
atten
tiv
en
ess
p
atter
n
s
f
lu
ctu
at
e
g
r
ad
u
ally
o
v
er
tim
e.
As
a
r
esu
lt,
r
ec
u
r
r
en
t
m
o
d
els
ca
n
p
r
o
d
u
ce
u
n
s
tab
le
o
r
d
elay
ed
r
esp
o
n
s
es to
s
u
b
tle
en
g
ag
em
en
t sh
if
ts
.
T
wo
co
r
e
ch
allen
g
es
th
er
ef
o
r
e
em
er
g
e
in
atten
tiv
e
n
ess
esti
m
atio
n
:
(
1
)
e
f
f
ec
tiv
e
m
o
d
elin
g
o
f
lo
n
g
-
r
an
g
e
tem
p
o
r
al
d
e
p
en
d
e
n
cies,
an
d
(
2
)
d
y
n
am
ic
in
ter
-
m
o
d
al
f
ea
tu
r
e
in
ter
ac
tio
n
.
First,
atten
tiv
en
ess
is
n
o
t
s
o
lely
d
eter
m
in
e
d
b
y
in
s
tan
tan
eo
u
s
b
eh
av
io
r
b
u
t
r
ath
er
b
y
s
u
s
tain
ed
p
atter
n
s
ac
r
o
s
s
ex
ten
d
ed
tem
p
o
r
a
l
win
d
o
ws.
C
ap
tu
r
in
g
th
ese
p
atter
n
s
r
eq
u
ir
es
ar
ch
itectu
r
es
ca
p
ab
le
o
f
g
lo
b
al
tem
p
o
r
al
r
ea
s
o
n
in
g
with
o
u
t
s
eq
u
en
tial
b
o
ttlen
ec
k
s
.
Seco
n
d
,
v
is
u
al
an
d
b
eh
av
io
r
al
s
ig
n
als
ar
e
n
o
t
in
d
ep
en
d
en
t;
th
ey
i
n
ter
ac
t
in
n
o
n
-
lin
ea
r
an
d
co
n
tex
t
-
s
en
s
itiv
e
way
s
.
A
r
o
b
u
s
t
en
g
ag
em
en
t
esti
m
atio
n
s
y
s
tem
m
u
s
t
th
er
ef
o
r
e
lear
n
cr
o
s
s
-
m
o
d
al
r
elatio
n
s
h
ip
s
r
ath
er
th
a
n
t
r
ea
ti
n
g
m
o
d
alities
as is
o
lated
f
ea
tu
r
e
s
tr
ea
m
s
[
1
8
]
-
[
2
1
]
.
T
r
an
s
f
o
r
m
e
r
-
b
ased
ar
ch
itect
u
r
es
o
f
f
er
a
p
r
i
n
cip
led
s
o
l
u
tio
n
to
th
ese
ch
allen
g
es.
Or
ig
in
ally
in
tr
o
d
u
ce
d
in
n
atu
r
al
lan
g
u
ag
e
p
r
o
ce
s
s
in
g
,
th
e
T
r
a
n
s
f
o
r
m
er
r
e
p
lace
s
r
ec
u
r
r
en
ce
with
s
elf
-
atten
tio
n
m
ec
h
an
i
s
m
s
th
at
d
ir
ec
tly
m
o
d
el
p
air
wis
e
in
ter
ac
tio
n
s
ac
r
o
s
s
s
eq
u
en
ce
elem
e
n
ts
.
Self
-
atten
tio
n
en
ab
les
g
lo
b
al
d
ep
en
d
e
n
cy
m
o
d
elin
g
b
y
co
m
p
u
tin
g
r
elev
a
n
ce
s
co
r
es
b
etwe
en
all
p
o
s
itio
n
s
in
a
s
eq
u
en
ce
s
im
u
ltan
eo
u
s
ly
.
T
h
is
d
esig
n
elim
in
ates
s
eq
u
en
tial
co
n
s
tr
ain
ts
,
im
p
r
o
v
es
p
ar
a
llelizatio
n
,
an
d
en
h
an
ce
s
lo
n
g
-
r
an
g
e
d
e
p
en
d
e
n
cy
ca
p
tu
r
e.
I
n
co
m
p
u
ter
v
is
io
n
,
th
e
v
is
io
n
tr
an
s
f
o
r
m
er
(
ViT
)
ex
ten
d
s
th
is
p
ar
ad
i
g
m
b
y
d
iv
id
in
g
im
ag
es
in
to
p
atch
es
an
d
p
r
o
ce
s
s
in
g
t
h
em
as
to
k
en
s
eq
u
en
ce
s
,
th
e
r
eb
y
en
ab
lin
g
g
lo
b
al
s
p
atial
atten
tio
n
.
Un
lik
e
C
NNs,
ViT
s
d
o
n
o
t
r
ely
o
n
lo
ca
lity
-
b
iased
co
n
v
o
l
u
tio
n
al
f
ilter
s
;
in
s
tead
,
th
ey
lear
n
co
n
tex
tu
al
r
elatio
n
s
h
ip
s
ac
r
o
s
s
th
e
en
tire
im
ag
e
f
r
o
m
th
e
o
u
ts
et
[
2
2
]
-
[
2
5
].
Mo
tiv
ated
b
y
th
ese
ad
v
an
ta
g
e
s
,
th
is
p
ap
er
p
r
o
p
o
s
es
a
cr
o
s
s
-
m
o
d
al
atten
tio
n
f
u
s
io
n
f
r
am
e
wo
r
k
b
u
ilt
upon
ViT
f
o
r
r
o
b
u
s
t
s
tu
d
en
t
atten
tiv
en
ess
es
tim
atio
n
.
T
h
e
p
r
o
p
o
s
ed
ar
ch
itectu
r
e
r
ep
lace
s
co
n
v
en
tio
n
al
co
n
v
o
l
u
tio
n
al
an
d
r
ec
u
r
r
en
t
m
o
d
u
les
with
tr
an
s
f
o
r
m
er
-
b
a
s
ed
s
p
atial
an
d
tem
p
o
r
al
m
o
d
elin
g
co
m
p
o
n
en
ts
.
Vis
u
al
f
r
am
es
ar
e
p
r
o
ce
s
s
ed
th
r
o
u
g
h
a
ViT
b
ac
k
b
o
n
e
to
c
ap
tu
r
e
g
lo
b
al
s
p
atial
d
ep
en
d
e
n
cies.
B
eh
av
io
r
al
f
ea
tu
r
es,
in
clu
d
in
g
g
az
e
d
ir
ec
t
io
n
,
h
ea
d
p
o
s
e,
an
d
b
lin
k
d
y
n
a
m
ics,
ar
e
em
b
ed
d
e
d
in
to
a
s
h
a
r
ed
laten
t
s
p
ac
e.
A
cr
o
s
s
-
m
o
d
al
m
u
lti
-
h
ea
d
atten
t
io
n
m
ec
h
an
is
m
d
y
n
am
ically
l
ea
r
n
s
in
ter
ac
tio
n
s
b
etwe
en
v
is
u
al
an
d
b
e
h
av
io
r
al
m
o
d
alities
,
en
ab
lin
g
ad
a
p
tiv
e
f
ea
tu
r
e
weig
h
tin
g
b
ased
o
n
c
o
n
tex
tu
al
r
elev
an
ce
.
T
em
p
o
r
a
l
d
ep
en
d
en
cies
ar
e
s
u
b
s
eq
u
en
tly
m
o
d
eled
u
s
in
g
a
T
r
an
s
f
o
r
m
er
e
n
co
d
er
,
allo
win
g
co
m
p
r
eh
en
s
iv
e
s
eq
u
e
n
ce
-
lev
el
r
ea
s
o
n
in
g
with
o
u
t r
ec
u
r
r
en
ce
.
2.
L
I
T
E
R
AT
U
RE
R
E
VI
E
W
Au
to
m
ated
s
tu
d
en
t
atten
tiv
en
ess
d
etec
tio
n
h
as
ev
o
lv
e
d
s
ig
n
if
ican
tly
with
th
e
g
r
o
wth
o
f
o
n
lin
e
an
d
h
y
b
r
id
lear
n
i
n
g
e
n
v
ir
o
n
m
en
ts
.
E
ar
ly
r
esear
ch
p
r
im
ar
ily
f
o
cu
s
ed
o
n
v
is
u
al
o
b
s
er
v
atio
n
an
d
b
eh
av
io
r
al
an
aly
s
is
with
in
s
tr
u
ctu
r
ed
class
r
o
o
m
s
ettin
g
s
.
Neg
r
o
n
an
d
Gr
a
v
es
[
1
]
in
tr
o
d
u
ce
d
th
e
class
r
o
o
m
atten
tiv
en
ess
class
if
icatio
n
to
o
l
(
C
lass
AC
T
)
,
a
s
y
s
tem
d
esig
n
ed
to
ca
teg
o
r
ize
atten
tiv
en
ess
lev
els
u
s
in
g
v
is
u
al
m
o
n
ito
r
in
g
tech
n
iq
u
es
.
T
h
eir
f
r
am
ewo
r
k
d
em
o
n
s
tr
ated
th
e
f
ea
s
ib
ilit
y
o
f
au
to
m
ate
d
e
n
g
ag
em
e
n
t
class
if
icatio
n
;
h
o
wev
e
r
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
C
r
o
s
s
-
mo
d
a
l a
tten
tio
n
f
u
s
io
n
u
s
in
g
visi
o
n
tr
a
n
s
fo
r
mers
fo
r
r
o
b
u
s
t
s
tu
d
en
t
… (
R
a
ja
s
ek
a
r
a
n
Ma
r
is
w
a
my
)
937
it
r
elied
o
n
h
a
n
d
cr
a
f
ted
f
ea
tu
r
es
an
d
r
u
le
-
b
ased
d
ec
is
io
n
m
e
ch
an
is
m
s
,
lim
itin
g
ad
a
p
tab
ilit
y
in
u
n
c
o
n
s
tr
ain
ed
o
n
lin
e
lear
n
in
g
s
ce
n
ar
io
s
.
W
ith
th
e
in
cr
ea
s
in
g
ad
o
p
tio
n
o
f
e
-
lea
r
n
in
g
p
latf
o
r
m
s
,
r
ese
ar
ch
atten
tio
n
s
h
if
ted
to
war
d
b
eh
av
io
r
al
m
o
n
ito
r
in
g
i
n
v
ir
tu
al
en
v
ir
o
n
m
en
ts
.
R
ev
ad
ek
ar
et
a
l.
[
2
]
p
r
o
p
o
s
ed
a
s
y
s
tem
f
o
r
g
au
g
i
n
g
s
tu
d
en
t
atten
tio
n
u
s
in
g
f
ac
ial
d
etec
tio
n
an
d
ac
ti
v
ity
t
r
ac
k
in
g
i
n
o
n
lin
e
class
r
o
o
m
s
.
Alth
o
u
g
h
th
eir
wo
r
k
em
p
h
asized
co
n
tin
u
o
u
s
en
g
ag
em
e
n
t
m
o
n
ito
r
in
g
,
it
lack
ed
ad
v
an
ce
d
tem
p
o
r
al
m
o
d
elin
g
s
tr
ateg
ies.
Sh
ah
et
a
l.
[
3
]
f
u
r
th
e
r
ex
ten
d
e
d
b
eh
av
io
r
al
a
n
aly
s
is
b
y
in
c
o
r
p
o
r
atin
g
g
az
e
d
ir
ec
tio
n
,
h
ea
d
o
r
ien
tatio
n
,
an
d
f
ac
ial
ex
p
r
ess
io
n
s
to
ass
ess
att
en
tiv
en
es
s
d
u
r
in
g
e
-
lear
n
in
g
s
ess
io
n
s
.
W
h
ile
ef
f
ec
tiv
e
in
d
em
o
n
s
tr
atin
g
th
e
im
p
o
r
tan
ce
o
f
b
eh
av
i
o
r
al
cu
es,
th
e
s
tu
d
y
em
p
lo
y
ed
c
o
n
v
e
n
tio
n
al
m
ac
h
in
e
lear
n
in
g
cla
s
s
if
ier
s
with
o
u
t
m
o
d
elin
g
lo
n
g
-
r
an
g
e
tem
p
o
r
al
d
ep
en
d
e
n
cies.
Su
b
s
eq
u
en
t
s
tu
d
ies
e
m
p
h
asize
d
f
ac
e
-
ba
s
ed
a
n
d
o
cu
lar
f
ea
tu
r
e
ex
tr
ac
tio
n
f
o
r
atten
tiv
e
n
ess
d
etec
tio
n
.
Pan
d
ey
et
a
l.
[
4
]
i
n
tr
o
d
u
c
ed
a
f
ac
e
d
etec
tio
n
-
b
ased
atten
tiv
en
ess
m
ea
s
u
r
e
tailo
r
ed
f
o
r
class
r
o
o
m
en
v
ir
o
n
m
en
ts
.
T
h
ei
r
ap
p
r
o
ac
h
f
o
cu
s
ed
o
n
id
e
n
tify
in
g
f
ac
ial
p
r
esen
ce
an
d
o
r
ien
tatio
n
b
u
t
r
em
ain
ed
lim
ited
to
f
r
am
e
-
lev
el
in
f
er
e
n
ce
.
Pai
et
a
l.
[
5
]
u
tili
ze
d
C
NNs
f
o
r
r
ea
l
-
tim
e
ey
e
m
o
n
ito
r
in
g
to
d
ete
ct
d
r
o
wsi
n
ess
an
d
atten
tiv
en
ess
.
Alth
o
u
g
h
C
NNs
im
p
r
o
v
ed
s
p
atial
f
ea
tu
r
e
lea
r
n
in
g
,
t
h
eir
lo
ca
lity
-
b
iased
a
r
ch
itectu
r
e
r
estricte
d
g
lo
b
al
co
n
tex
tu
al
m
o
d
elin
g
.
Hea
d
p
o
s
e
esti
m
atio
n
h
as
also
b
ee
n
ex
p
lo
r
ed
as
a
s
tr
o
n
g
in
d
icato
r
o
f
en
g
ag
em
en
t.
Pin
zo
n
-
Go
n
za
lez
an
d
B
ar
b
a
-
G
u
am
an
[
6
]
em
p
lo
y
ed
h
ea
d
p
o
s
itio
n
esti
m
atio
n
to
d
eter
m
in
e
atten
tio
n
lev
els
in
r
em
o
t
e
class
r
o
o
m
s
.
T
h
eir
f
in
d
in
g
s
v
a
lid
ated
h
ea
d
o
r
ien
tatio
n
as
a
r
elia
b
le
b
eh
av
io
r
al
cu
e;
h
o
we
v
er
,
th
e
ap
p
r
o
ac
h
tr
ea
ted
m
o
d
alities
in
d
ep
en
d
e
n
tly
an
d
d
id
n
o
t
m
o
d
el
in
ter
-
m
o
d
al
in
ter
ac
tio
n
s
.
Mo
r
e
co
m
p
r
eh
en
s
iv
e
m
o
d
elin
g
s
tr
ateg
ies
wer
e
later
p
r
o
p
o
s
ed
b
y
E
lb
awa
b
an
d
Hen
r
iq
u
es
[
7
]
,
wh
o
c
o
m
b
in
e
d
em
o
tio
n
al
an
d
n
o
n
-
em
o
tio
n
al
f
ea
tu
r
e
s
u
s
in
g
m
ac
h
i
n
e
lear
n
in
g
tech
n
iq
u
es
to
en
h
an
ce
atten
t
iv
en
ess
p
r
ed
ictio
n
.
W
h
ile
in
c
o
r
p
o
r
atin
g
af
f
ec
tiv
e
s
ig
n
als
im
p
r
o
v
ed
p
e
r
f
o
r
m
an
c
e,
th
e
r
elian
ce
o
n
tr
ad
itio
n
al
lear
n
in
g
p
ip
elin
es
lim
ited
th
e
ab
ilit
y
to
ca
p
tu
r
e
co
m
p
lex
tem
p
o
r
al
an
d
cr
o
s
s
-
m
o
d
al
d
e
p
en
d
e
n
cies
.
Fro
m
an
ev
alu
atio
n
s
tan
d
p
o
in
t,
r
eliab
ilit
y
ass
e
s
s
m
en
t
r
em
ain
s
ess
en
tial
in
at
ten
tiv
en
ess
m
o
n
ito
r
in
g
s
y
s
tem
s
.
T
h
e
in
tr
ac
lass
co
r
r
el
atio
n
co
ef
f
icien
t
(
I
C
C
)
f
r
am
e
wo
r
k
p
r
o
p
o
s
ed
b
y
B
i
an
d
K
u
esten
[
8
]
p
r
o
v
i
d
es
a
s
tatis
t
ical
f
o
u
n
d
atio
n
f
o
r
ass
es
s
in
g
in
ter
-
r
ater
co
n
s
i
s
ten
cy
an
d
m
o
d
el
r
eliab
ilit
y
.
Su
ch
ev
al
u
atio
n
s
tr
ateg
ies ar
e
r
elev
an
t w
h
en
v
alid
atin
g
e
n
g
a
g
em
en
t d
etec
tio
n
m
o
d
els ag
ai
n
s
t a
n
n
o
tated
g
r
o
u
n
d
tr
u
th
lab
els.
3.
MET
H
O
D
T
h
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
is
d
esig
n
ed
to
esti
m
ate
s
tu
d
en
t
atten
tiv
en
ess
f
r
o
m
v
id
eo
s
tr
ea
m
s
in
b
o
th
r
ea
l
-
tim
e
an
d
o
f
f
lin
e
m
o
d
es.
T
h
e
s
y
s
tem
p
r
o
c
ess
es
in
p
u
t
v
i
d
eo
f
r
am
es
to
ex
t
r
ac
t
v
is
u
al
a
n
d
b
e
h
av
io
r
al
cu
es,
f
u
s
es
th
ese
m
o
d
alities
u
s
in
g
atten
tio
n
-
d
r
iv
e
n
m
ec
h
an
is
m
s
,
m
o
d
els
tem
p
o
r
al
d
ep
e
n
d
en
cies
th
r
o
u
g
h
a
tr
an
s
f
o
r
m
er
e
n
co
d
e
r
,
an
d
o
u
tp
u
ts
a
p
r
o
b
a
b
ilis
tic
atten
tiv
en
ess
s
co
r
e.
T
h
e
o
v
e
r
all
p
ip
elin
e
co
n
s
is
ts
o
f
th
e
f
o
llo
win
g
s
tag
es:
a)
Fra
m
e
ac
q
u
is
itio
n
an
d
p
r
ep
r
o
c
ess
in
g
b)
ViT
-
b
ased
s
p
atial
f
ea
tu
r
e
e
x
tr
a
ctio
n
c)
B
eh
av
io
r
al
f
ea
tu
r
e
e
m
b
ed
d
in
g
(
g
az
e
an
d
h
ea
d
p
o
s
e)
d)
C
r
o
s
s
-
m
o
d
al
m
u
lti
-
h
ea
d
atten
t
io
n
f
u
s
io
n
e)
T
em
p
o
r
al
tr
a
n
s
f
o
r
m
er
en
co
d
in
g
f)
Atten
tiv
en
ess
class
if
icatio
n
T
h
e
d
ataset
is
th
e
d
ataset
f
o
r
af
f
ec
tiv
e
s
tates
in
e
-
en
v
ir
o
n
m
en
ts
(
DAiSEE
)
,
a
p
u
b
lic
b
en
ch
m
ar
k
co
n
tain
in
g
o
v
er
9
,
0
0
0
v
id
e
o
clip
s
o
f
s
tu
d
en
ts
in
o
n
lin
e
lea
r
n
in
g
en
v
ir
o
n
m
en
ts
.
E
ac
h
cli
p
is
an
n
o
tated
wit
h
f
o
u
r
a
f
f
ec
tiv
e
s
tates
—
en
g
ag
em
en
t,
b
o
r
ed
o
m
,
co
n
f
u
s
io
n
,
an
d
f
r
u
s
tr
atio
n
—
at
f
o
u
r
in
ten
s
ity
lev
els
(
lo
w,
m
ed
iu
m
,
h
ig
h
,
v
er
y
h
i
g
h
)
.
T
h
e
DAiSEE
d
ataset
p
r
o
v
id
es
ex
ter
n
al
v
alid
atio
n
an
d
e
n
s
u
r
es
th
e
m
o
d
el
g
en
er
alize
s
ac
r
o
s
s
d
iv
er
s
e
lear
n
in
g
co
n
tex
ts
.
T
h
e
p
r
o
p
o
s
ed
ar
c
h
itectu
r
e
f
o
llo
ws
a
d
u
al
-
s
tr
ea
m
tr
an
s
f
o
r
m
er
-
b
ased
d
esig
n
th
at
e
x
p
lici
tly
m
o
d
els
co
m
p
lem
en
tar
y
v
is
u
al
an
d
b
eh
av
io
r
al
cu
es
b
e
f
o
r
e
in
teg
r
a
tin
g
th
em
th
r
o
u
g
h
atten
tio
n
-
d
r
iv
en
f
u
s
io
n
.
T
h
e
f
r
am
ewo
r
k
is
s
tr
u
ctu
r
ed
to
ca
p
tu
r
e
s
p
atial
co
n
tex
t,
b
eh
av
io
r
al
d
y
n
a
m
ics,
an
d
lo
n
g
-
r
an
g
e
tem
p
o
r
al
d
ep
en
d
e
n
cies in
a
u
n
i
f
ied
en
d
-
to
-
en
d
tr
ai
n
ab
le
s
y
s
tem
.
3
.
1
.
Vis
ua
l
s
t
re
a
m
T
h
e
v
is
u
al
s
tr
ea
m
is
r
esp
o
n
s
i
b
le
f
o
r
e
x
tr
ac
tin
g
r
ich
s
p
atial
r
ep
r
esen
tatio
n
s
f
r
o
m
r
aw
R
GB
f
r
am
es
ca
p
tu
r
ed
d
u
r
in
g
o
n
lin
e
lear
n
i
n
g
s
ess
io
n
s
.
E
ac
h
i
n
p
u
t
f
r
am
e
is
f
ir
s
t
r
esized
an
d
n
o
r
m
aliz
ed
,
th
en
p
ar
titi
o
n
ed
in
to
f
ix
e
d
-
s
ize
n
o
n
-
o
v
er
lap
p
in
g
p
atch
es.
I
f
th
e
f
r
am
e
r
eso
lu
t
io
n
is
H×
W
,
an
d
th
e
p
atch
s
ize
is
P×P,
th
e
im
a
g
e
is
d
ec
o
m
p
o
s
ed
in
to
=
2
p
atch
es.
E
ac
h
p
atch
is
f
latten
ed
a
n
d
li
n
ea
r
ly
p
r
o
jecte
d
in
to
a
h
ig
h
-
d
im
en
s
io
n
al
em
b
ed
d
in
g
s
p
ac
e
th
r
o
u
g
h
a
le
ar
n
ab
le
p
r
o
jectio
n
m
at
r
ix
,
f
o
r
m
in
g
a
s
eq
u
e
n
ce
o
f
p
atch
to
k
e
n
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
15
,
No
.
3
,
Sep
tem
b
er
20
26
:
935
-
9
4
3
938
T
o
r
etain
s
p
atial
s
tr
u
ctu
r
e,
p
o
s
itio
n
al
en
co
d
i
n
g
s
ar
e
ad
d
e
d
to
ea
ch
to
k
en
em
b
ed
d
i
n
g
.
T
h
ese
en
r
ich
ed
to
k
en
s
ar
e
th
en
p
ass
ed
th
r
o
u
g
h
a
s
tack
o
f
ViT
en
co
d
e
r
b
lo
ck
s
.
E
ac
h
b
lo
c
k
co
n
s
is
ts
o
f
m
u
lti
-
h
ea
d
s
elf
-
atten
tio
n
(
MH
SA)
an
d
f
ee
d
-
f
o
r
war
d
lay
er
s
with
r
esid
u
al
co
n
n
ec
tio
n
s
a
n
d
lay
er
n
o
r
m
a
lizatio
n
.
T
h
e
s
elf
-
atten
tio
n
m
ec
h
a
n
is
m
allo
ws
ea
ch
p
atch
to
k
e
n
to
atten
d
t
o
all
o
t
h
er
p
atch
es
with
in
t
h
e
f
r
a
m
e,
th
e
r
eb
y
m
o
d
elin
g
g
lo
b
al
s
p
atial
d
ep
en
d
en
cies.
T
h
is
is
p
ar
ticu
lar
ly
i
m
p
o
r
tan
t f
o
r
atten
tiv
en
ess
esti
m
atio
n
,
wh
er
e
f
ac
ial
cu
es,
ey
e
r
eg
i
o
n
s
,
an
d
co
n
tex
t
u
al
p
o
s
tu
r
e
in
f
o
r
m
atio
n
m
ay
in
ter
ac
t n
o
n
-
lo
ca
lly
.
T
h
e
o
u
tp
u
t
o
f
th
e
ViT
b
ac
k
b
o
n
e
is
a
r
ef
in
ed
v
is
u
al
to
k
en
r
e
p
r
e
s
en
tatio
n
th
at
ca
p
tu
r
es
h
o
li
s
tic
s
p
atia
l
co
n
tex
t.
A
s
p
ec
ial
class
if
icatio
n
to
k
en
o
r
p
o
o
led
r
ep
r
esen
tatio
n
is
ex
tr
ac
te
d
to
s
u
m
m
a
r
ize
th
e
f
r
am
e
-
le
v
el
v
is
u
al
in
f
o
r
m
atio
n
.
T
h
is
r
ep
r
e
s
en
tatio
n
en
co
d
es
f
ac
ial
o
r
ien
t
atio
n
,
ey
e
r
eg
i
o
n
p
atter
n
s
,
p
o
s
tu
r
e
in
d
icato
r
s
,
an
d
o
th
er
ap
p
ea
r
an
ce
-
b
ased
cu
es r
elev
an
t to
atten
tiv
en
ess
.
3
.
2
.
B
eha
v
io
r
a
l
s
t
re
a
m
T
h
e
b
eh
a
v
io
r
al
s
tr
ea
m
o
p
er
ate
s
in
p
ar
allel
to
e
x
p
licitly
m
o
d
el
s
tr
u
ctu
r
ed
b
eh
av
io
r
al
f
ea
tu
r
es
d
er
iv
ed
f
r
o
m
th
e
s
am
e
v
id
e
o
f
r
am
e
.
I
n
s
tead
o
f
r
ely
in
g
s
o
lely
o
n
im
p
licit
lear
n
in
g
f
r
o
m
p
ix
els,
th
is
s
tr
ea
m
in
co
r
p
o
r
ates
in
ter
p
r
etab
le
b
eh
av
io
r
al
in
d
icato
r
s
s
u
ch
as
g
az
e
d
ir
ec
tio
n
v
ec
to
r
s
an
d
h
ea
d
p
o
s
e
an
g
les
(
y
aw,
p
itch
,
an
d
r
o
ll).
T
h
e
g
az
e
v
e
cto
r
,
ty
p
ically
r
e
p
r
esen
ted
in
3
D
s
p
ac
e,
ca
p
tu
r
es
wh
er
e
th
e
s
tu
d
en
t
is
lo
o
k
in
g
r
elativ
e
to
th
e
ca
m
er
a
o
r
s
cr
ee
n
.
Hea
d
p
o
s
e
p
ar
am
eter
s
q
u
an
tify
o
r
ien
tatio
n
d
ev
iatio
n
s
th
at
m
ay
in
d
icate
d
is
tr
ac
tio
n
.
T
h
ese
f
ea
tu
r
es a
r
e
co
n
ca
ten
ated
to
f
o
r
m
a
b
e
h
av
i
o
r
al
f
ea
tu
r
e
v
ec
to
r
at
ea
c
h
tim
e
s
tep
.
T
o
alig
n
th
e
b
eh
a
v
io
r
al
f
ea
tu
r
es
with
th
e
v
is
u
al
em
b
e
d
d
in
g
s
p
a
ce
,
th
e
c
o
n
ca
ten
ated
v
ec
t
o
r
is
p
ass
ed
th
r
o
u
g
h
lin
ea
r
p
r
o
jectio
n
lay
er
s
.
T
h
ese
f
u
lly
co
n
n
ec
ted
lay
er
s
tr
an
s
f
o
r
m
lo
w
-
d
im
en
s
io
n
a
l
b
eh
av
io
r
al
in
p
u
ts
in
to
a
h
ig
h
er
-
d
im
en
s
io
n
al
b
e
h
av
io
r
al
em
b
ed
d
i
n
g
c
o
m
p
ati
b
le
with
tr
an
s
f
o
r
m
er
atten
tio
n
o
p
e
r
atio
n
s
.
No
n
-
lin
ea
r
ac
t
iv
atio
n
f
u
n
ctio
n
s
an
d
n
o
r
m
aliza
tio
n
m
ay
b
e
ap
p
lied
to
en
h
an
ce
r
ep
r
esen
tatio
n
ca
p
ac
ity
an
d
s
tab
ilize
tr
ain
in
g
.
T
h
e
r
esu
ltin
g
b
e
h
a
v
io
r
al
em
b
ed
d
in
g
p
r
o
v
i
d
es
co
m
p
ac
t,
s
tr
u
ctu
r
ed
i
n
f
o
r
m
atio
n
ab
o
u
t
s
tu
d
e
n
t
en
g
ag
em
e
n
t th
at
co
m
p
lem
en
ts
th
e
h
ig
h
-
d
im
en
s
io
n
al
v
is
u
al
to
k
en
r
e
p
r
esen
tatio
n
.
3
.
3
.
Cro
s
s
-
mo
da
l f
us
io
n a
nd
t
em
po
ra
l
m
o
delin
g
Af
ter
in
d
ep
en
d
en
t
en
co
d
in
g
,
th
e
v
is
u
al
to
k
en
r
ep
r
esen
tatio
n
an
d
b
e
h
av
io
r
al
em
b
ed
d
i
n
g
ar
e
f
u
s
ed
u
s
in
g
cr
o
s
s
-
m
o
d
al
m
u
lti
-
h
ea
d
atten
tio
n
.
I
n
t
h
is
m
ec
h
an
is
m
,
o
n
e
m
o
d
ality
(
e.
g
.
,
v
is
u
a
l
to
k
en
s
)
s
er
v
es
as
q
u
er
ies,
wh
ile
th
e
o
t
h
er
m
o
d
a
lity
(
b
eh
av
io
r
al
em
b
e
d
d
in
g
s
)
p
r
o
v
id
es
k
e
y
s
an
d
v
alu
es.
T
h
i
s
en
ab
les
ad
ap
tiv
e
weig
h
tin
g
o
f
b
e
h
av
io
r
al
c
u
es c
o
n
d
itio
n
e
d
o
n
v
is
u
al
co
n
tex
t,
allo
win
g
th
e
m
o
d
el
t
o
d
y
n
am
ically
d
eter
m
in
e
th
e
r
elativ
e
im
p
o
r
ta
n
ce
o
f
g
az
e
a
n
d
h
ea
d
p
o
s
e
in
f
o
r
m
atio
n
f
o
r
e
ac
h
f
r
am
e
.
T
h
e
f
u
s
ed
f
ea
tu
r
es
ar
e
th
en
o
r
g
a
n
ized
in
to
tem
p
o
r
al
s
eq
u
en
ce
s
an
d
p
ass
ed
to
a
t
r
an
s
f
o
r
m
e
r
-
b
ased
t
em
p
o
r
al
en
co
d
er
.
Un
lik
e
r
ec
u
r
r
en
t
a
r
ch
itectu
r
es,
th
e
tem
p
o
r
al
tr
an
s
f
o
r
m
er
m
o
d
els
lo
n
g
-
r
an
g
e
d
ep
en
d
en
cies
u
s
in
g
s
elf
-
at
ten
tio
n
ac
r
o
s
s
tim
e
s
tep
s
,
en
ab
lin
g
th
e
s
y
s
tem
to
ca
p
tu
r
e
s
u
s
tain
ed
e
n
g
ag
em
e
n
t
p
atter
n
s
an
d
g
r
a
d
u
al
s
h
if
ts
in
atten
tiv
en
ess
.
Fin
a
lly
,
th
e
tem
p
o
r
ally
en
co
d
ed
r
e
p
r
esen
tatio
n
is
p
r
o
ce
s
s
ed
th
r
o
u
g
h
a
class
if
icatio
n
h
ea
d
with
a
So
f
tm
ax
lay
er
to
p
r
o
d
u
ce
atten
tiv
en
ess
p
r
o
b
ab
ilit
ies.
T
h
is
h
ier
ar
ch
ical
d
esig
n
en
s
u
r
es
r
o
b
u
s
t
s
p
atial
m
o
d
elin
g
,
ex
p
licit
b
eh
av
io
r
al
in
teg
r
atio
n
,
a
n
d
ef
f
icien
t seq
u
en
ce
lear
n
in
g
with
in
a
f
u
lly
atten
tio
n
-
d
r
iv
en
f
r
am
ewo
r
k
.
3
.
4
.
ViT
ba
c
k
bo
ne
T
o
o
v
er
c
o
m
e
lo
ca
lity
co
n
s
tr
ai
n
ts
o
f
co
n
v
o
l
u
tio
n
al
n
etwo
r
k
s
,
th
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
em
p
lo
y
s
a
ViT
b
ac
k
b
o
n
e
f
o
r
s
p
atial
f
ea
tu
r
e
ex
tr
ac
tio
n
.
E
ac
h
in
p
u
t
f
r
a
m
e
∈
×
×
is
d
iv
i
d
ed
i
n
to
n
o
n
-
o
v
er
l
ap
p
in
g
p
atch
es o
f
s
ize
P×P.
E
ac
h
p
atc
h
is
f
latten
ed
an
d
lin
ea
r
ly
p
r
o
j
ec
ted
in
to
a
lat
en
t e
m
b
ed
d
in
g
s
p
ac
e.
=
(
ℎ
)
+
is
th
e
lear
n
ab
le
p
r
o
jectio
n
m
a
tr
ix
r
ep
r
esen
ts
p
o
s
itio
n
al
en
c
o
d
in
g
d
en
o
tes p
atch
e
m
b
ed
d
in
g
s
T
h
is
tr
an
s
f
o
r
m
s
th
e
im
ag
e
in
to
a
s
eq
u
en
ce
o
f
to
k
en
s
s
u
itab
le
f
o
r
tr
an
s
f
o
r
m
e
r
p
r
o
ce
s
s
in
g
.
E
ac
h
tr
an
s
f
o
r
m
er
en
c
o
d
er
b
lo
c
k
co
m
p
u
tes
s
elf
-
atten
tio
n
ac
r
o
s
s
all
p
atch
to
k
en
s
,
en
ab
li
n
g
g
lo
b
al
s
p
atial
d
ep
en
d
e
n
cy
m
o
d
elin
g
.
T
h
e
atten
tio
n
m
ec
h
an
is
m
is
d
e
f
in
ed
as:
(
,
,
)
=
(
√
)
wh
er
e:
Q
,
K,
V
ar
e
q
u
e
r
y
,
k
ey
,
a
n
d
v
a
lu
e
m
atr
ices
is
th
e
k
ey
d
im
e
n
s
io
n
T
h
is
allo
ws ea
ch
p
atch
to
atte
n
d
to
e
v
er
y
o
th
er
p
atch
,
ca
p
t
u
r
in
g
h
o
lis
tic
co
n
tex
tu
al
r
elatio
n
s
h
ip
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
C
r
o
s
s
-
mo
d
a
l a
tten
tio
n
f
u
s
io
n
u
s
in
g
visi
o
n
tr
a
n
s
fo
r
mers
fo
r
r
o
b
u
s
t
s
tu
d
en
t
… (
R
a
ja
s
ek
a
r
a
n
Ma
r
is
w
a
my
)
939
3
.
5
.
B
eha
v
io
r
a
l
f
ea
t
ure
e
m
bedd
ing
B
eh
av
io
r
al
cu
es
p
r
o
v
id
e
c
o
m
p
lem
en
tar
y
in
f
o
r
m
atio
n
to
v
i
s
u
al
ap
p
ea
r
an
ce
f
ea
tu
r
es.
T
h
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
em
b
ed
s
g
az
e
v
ec
to
r
s
an
d
h
ea
d
p
o
s
e
p
a
r
am
eter
s
i
n
to
a
u
n
if
ied
laten
t sp
ac
e.
L
et
th
e
b
eh
av
i
o
r
al
f
ea
tu
r
e
v
ec
t
o
r
at
tim
e
t
b
e:
=
[
,
]
wh
er
e:
=
g
az
e
d
ir
ec
tio
n
v
ec
to
r
=
h
ea
d
p
o
s
e
an
g
les (
y
aw,
p
itc
h
,
r
o
ll)
3
.
6
.
G
a
ze
v
ec
t
o
r
enco
din
g
T
h
e
g
az
e
d
i
r
ec
tio
n
is
r
ep
r
esen
ted
as a
3
D
v
ec
to
r
an
d
p
r
o
jecte
d
in
to
th
e
e
m
b
ed
d
in
g
d
im
en
s
io
n
:
′
=
+
3
.
7
.
H
ea
d
po
s
e
em
bedd
ing
Hea
d
p
o
s
e
an
g
les ar
e
s
im
ilar
ly
p
r
o
jecte
d
:
′
=
ℎ
+
ℎ
t
h
e
f
in
al
b
eh
a
v
io
r
al
em
b
ed
d
i
n
g
is
:
′
=
[
′
,
′
]
T
h
is
em
b
ed
d
in
g
alig
n
s
d
im
e
n
s
io
n
ality
with
v
is
u
al
to
k
en
s
f
o
r
f
u
s
io
n
.
3
.
7
.
Cro
s
s
-
mo
da
l a
t
t
ent
io
n f
us
io
n
I
n
s
t
e
a
d
o
f
s
t
a
ti
c
c
o
n
c
a
t
e
n
a
ti
o
n
o
r
w
e
i
g
h
t
e
d
s
u
m
m
a
t
i
o
n
,
t
h
e
p
r
o
p
o
s
e
d
f
r
a
m
e
w
o
r
k
e
m
p
l
o
y
s
c
r
o
s
s
-
m
o
d
al
m
u
l
t
i
-
h
e
a
d
a
t
te
n
t
i
o
n
t
o
d
y
n
a
m
ic
a
l
l
y
m
o
d
e
l
i
n
te
r
d
e
p
e
n
d
e
n
c
i
e
s
b
e
t
w
e
e
n
v
is
u
a
l
a
n
d
b
e
h
a
v
i
o
r
a
l
m
o
d
a
l
i
t
i
es
.
Giv
en
v
is
u
al
to
k
en
s
an
d
b
eh
a
v
io
r
al
em
b
e
d
d
in
g
′
,
cr
o
s
s
-
atten
tio
n
is
co
m
p
u
ted
as:
(
,
,
)
=
(
√
)
wh
er
e
q
u
er
ies
ar
e
d
er
iv
e
d
f
r
o
m
v
is
u
al
f
ea
tu
r
es
an
d
k
ey
s
/v
a
lu
es
f
r
o
m
b
e
h
av
io
r
al
em
b
ed
d
i
n
g
s
(
o
r
v
ice
v
er
s
a
d
ep
en
d
i
n
g
o
n
co
n
f
ig
u
r
atio
n
)
.
3
.
8
.
M
ulti
-
hea
d a
t
t
ent
io
n f
o
rm
ula
t
io
n
(
,
,
)
=
(
ℎ
1
,
.
.
.
,
ℎ
ℎ
)
E
ac
h
h
ea
d
lea
r
n
s
d
is
tin
ct
in
ter
-
m
o
d
al
r
elatio
n
s
h
ip
s
,
en
ab
lin
g
:
−
Ad
ap
tiv
e
m
o
d
ality
weig
h
tin
g
−
C
o
n
tex
t
-
awa
r
e
f
ea
tu
r
e
i
n
ter
ac
t
io
n
−
R
o
b
u
s
tn
ess
to
n
o
is
e
in
in
d
iv
id
u
al
m
o
d
alities
T
h
e
f
u
s
ed
r
ep
r
esen
tatio
n
is
d
e
n
o
ted
as:
3
.
9
.
T
em
po
r
a
l
t
ra
ns
f
o
r
m
er
enco
der
Atten
tiv
en
ess
is
in
h
er
en
tly
s
eq
u
en
tial.
I
n
s
tead
o
f
r
ec
u
r
r
e
n
t
m
o
d
elin
g
,
t
h
e
p
r
o
p
o
s
ed
f
r
am
ew
o
r
k
u
s
es
a
tr
an
s
f
o
r
m
er
e
n
co
d
e
r
to
ca
p
tu
r
e
lo
n
g
-
r
an
g
e
tem
p
o
r
al
d
ep
en
d
en
cies.
Giv
en
th
e
f
u
s
ed
s
e
q
u
en
ce
:
=
{
1
,
2
,
.
.
.
,
}
T
em
p
o
r
al
e
n
co
d
in
g
is
p
er
f
o
r
m
ed
as:
′
=
(
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
15
,
No
.
3
,
Sep
tem
b
er
20
26
:
935
-
9
4
3
940
T
h
is
m
ec
h
an
is
m
:
−
E
n
ab
les g
lo
b
al
tem
p
o
r
al
atten
t
io
n
−
E
lim
in
ates r
ec
u
r
r
en
ce
c
o
n
s
tr
ai
n
ts
−
I
m
p
r
o
v
es p
ar
alleliza
tio
n
−
C
ap
tu
r
es lo
n
g
-
ter
m
e
n
g
ag
e
m
e
n
t p
atter
n
s
Po
s
itio
n
al
en
co
d
in
g
s
ar
e
a
d
d
e
d
to
p
r
eser
v
e
tem
p
o
r
al
o
r
d
er
in
g
.
3
.
1
0
.
At
t
ent
iv
eness
cla
s
s
if
ic
a
t
io
n hea
d
T
h
e
tem
p
o
r
ally
en
co
d
ed
r
ep
r
e
s
en
tatio
n
Z
′ is
p
ass
ed
th
r
o
u
g
h
a
f
u
lly
co
n
n
ec
ted
class
if
icatio
n
h
ea
d
.
Fo
r
ea
ch
tim
e
s
tep
:
=
(
′
+
)
wh
er
e:
:
is
th
e
o
u
tp
u
t p
r
o
jectio
n
m
atr
i
x
:
is
b
ias
:
is
th
e
atten
tiv
en
ess
p
r
o
b
ab
ilit
y
d
is
tr
ib
u
tio
n
T
h
e
f
in
al
atten
tiv
en
ess
s
co
r
e
c
an
b
e
co
m
p
u
ted
as:
−
Per
-
f
r
am
e
p
r
o
b
a
b
ilit
y
−
Sli
d
in
g
win
d
o
w
av
e
r
ag
e
−
Seq
u
en
ce
-
lev
el
ag
g
r
eg
atio
n
C
o
m
p
ar
ed
to
c
o
n
v
e
n
tio
n
al
C
NN
–
L
STM
p
ip
elin
es,
th
e
p
r
o
p
o
s
ed
tr
an
s
f
o
r
m
er
-
b
ased
f
r
am
ewo
r
k
in
tr
o
d
u
ce
s
s
ev
er
al
ar
ch
itectu
r
al
ad
v
an
tag
es
th
at
s
ig
n
if
ican
tl
y
en
h
a
n
ce
atten
tiv
en
ess
esti
m
atio
n
p
er
f
o
r
m
a
n
ce
.
First,
b
y
em
p
lo
y
in
g
a
Vi
T
b
ac
k
b
o
n
e
i
n
s
tead
o
f
co
n
v
o
lu
ti
o
n
al
lay
er
s
,
th
e
m
o
d
el
ca
p
tu
r
es
g
lo
b
al
s
p
atial
d
ep
en
d
e
n
cies
ac
r
o
s
s
th
e
en
tire
f
r
am
e
th
r
o
u
g
h
s
elf
-
atten
tio
n
,
r
ath
er
th
a
n
r
ely
in
g
o
n
th
e
lim
i
ted
r
ec
ep
tiv
e
f
ield
s
o
f
co
n
v
o
lu
tio
n
al
k
er
n
els.
4.
RE
SU
L
T
S AN
D
P
E
RF
O
RM
ANCE
E
VA
L
UAT
I
O
N
T
h
is
s
ec
tio
n
p
r
esen
ts
th
e
q
u
an
titativ
e
an
d
c
o
m
p
ar
ati
v
e
e
v
alu
atio
n
o
f
th
e
p
r
o
p
o
s
ed
T
r
an
s
f
o
r
m
er
-
b
ased
cr
o
s
s
-
m
o
d
al
atten
tiv
en
e
s
s
est
im
atio
n
f
r
am
ewo
r
k
.
T
h
e
m
o
d
el
is
ev
alu
ated
u
s
in
g
s
ta
n
d
ar
d
class
if
icatio
n
m
etr
ics
in
clu
d
in
g
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all
,
F1
-
s
co
r
e,
an
d
a
r
ea
u
n
d
er
th
e
R
OC
cu
r
v
e
(
AU
C
)
.
Per
f
o
r
m
a
n
ce
is
co
m
p
ar
ed
ag
ain
s
t
b
aselin
e
d
e
ep
lear
n
in
g
m
o
d
els
an
d
s
elec
t
ed
m
eth
o
d
s
f
r
o
m
p
r
i
o
r
liter
atu
r
e.
T
h
e
d
ataset
was
d
iv
id
ed
in
t
o
tr
ain
in
g
(
7
0
%),
v
alid
atio
n
(
1
5
%),
a
n
d
test
in
g
(
1
5
%)
s
u
b
s
ets
u
s
in
g
s
tr
atif
ied
s
am
p
lin
g
.
All
ex
p
er
im
en
ts
wer
e
co
n
d
u
cted
u
s
in
g
id
en
tical
p
r
ep
r
o
ce
s
s
in
g
an
d
ev
alu
atio
n
p
r
o
to
c
o
ls
to
en
s
u
r
e
f
air
n
ess
.
T
h
e
d
ataset
was
d
iv
id
ed
in
to
tr
ain
in
g
(
7
0
%),
v
alid
atio
n
(
1
5
%),
an
d
test
in
g
(
1
5
%)
s
u
b
s
ets
u
s
in
g
s
tr
atif
ied
s
am
p
lin
g
.
All
ex
p
er
im
en
ts
wer
e
co
n
d
u
cte
d
u
s
i
n
g
id
en
tical
p
r
ep
r
o
ce
s
s
in
g
an
d
ev
alu
atio
n
p
r
o
to
co
ls
to
en
s
u
r
e
f
air
n
ess
.
4
.
1
.
O
v
er
a
ll
cla
s
s
if
ica
t
io
n pe
rf
o
rma
nce
T
o
ev
alu
ate
th
e
co
m
p
lete
f
r
am
ewo
r
k
,
all
p
r
im
ar
y
clas
s
if
icatio
n
m
etr
ics
o
n
th
e
test
s
et
ar
e
s
u
m
m
ar
ized
in
T
ab
le
1.
T
h
e
h
ig
h
r
ec
all
in
d
icate
s
s
tr
o
n
g
d
etec
tio
n
ca
p
ab
ilit
y
f
o
r
atten
tiv
e
s
tu
d
en
ts
,
wh
ile
h
i
g
h
s
p
ec
if
icity
co
n
f
ir
m
s
r
eliab
le
i
d
en
tific
atio
n
o
f
in
att
en
tiv
e
s
t
ates.
T
h
e
b
ala
n
ce
d
F1
-
s
co
r
e
d
em
o
n
s
tr
ates
s
tab
le
p
er
f
o
r
m
an
ce
ac
r
o
s
s
b
o
th
class
es.
T
ab
le
1
.
Ov
e
r
all
p
er
f
o
r
m
an
ce
m
etr
ics
M
e
t
r
i
c
P
r
o
p
o
se
d
f
r
a
m
e
w
o
r
k
A
c
c
u
r
a
c
y
9
4
.
8
7
%
P
r
e
c
i
s
i
o
n
9
3
.
9
5
%
R
e
c
a
l
l
(
S
e
n
si
t
i
v
i
t
y
)
9
5
.
4
2
%
S
p
e
c
i
f
i
c
i
t
y
9
3
.
8
1
%
F1
-
S
c
o
r
e
9
4
.
6
8
%
AUC
0
.
9
7
2
4
.
2
.
Abla
t
io
n
s
t
ud
y
T
o
q
u
a
n
tify
th
e
im
p
a
ct
o
f
ar
ch
itectu
r
al
co
m
p
o
n
en
ts
,
a
d
et
ailed
ab
latio
n
s
tu
d
y
was
co
n
d
u
cted
as
s
h
o
wn
in
T
a
b
le
2
.
All
ev
al
u
atio
n
m
etr
ics
ar
e
r
ep
o
r
ted
f
o
r
f
ai
r
n
ess
.
T
h
e
r
esu
lts
d
em
o
n
s
tr
ate
th
at
ea
ch
ar
ch
itectu
r
al
en
h
a
n
ce
m
en
t c
o
n
tr
ib
u
tes co
n
s
is
ten
tly
ac
r
o
s
s
all
ev
alu
atio
n
m
etr
ics.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
C
r
o
s
s
-
mo
d
a
l a
tten
tio
n
f
u
s
io
n
u
s
in
g
visi
o
n
tr
a
n
s
fo
r
mers
fo
r
r
o
b
u
s
t
s
tu
d
en
t
… (
R
a
ja
s
ek
a
r
a
n
Ma
r
is
w
a
my
)
941
T
ab
le
2
.
Ab
latio
n
s
tu
d
y
with
m
u
ltip
le
m
etr
ics
M
o
d
e
l
v
a
r
i
a
n
t
A
c
c
u
r
a
c
y
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
S
c
o
r
e
AUC
C
N
N
+
LSTM
(
B
a
sel
i
n
e
)
8
6
.
3
2
%
8
5
.
7
4
%
8
6
.
1
0
%
8
5
.
9
1
%
0
.
8
8
4
V
i
si
o
n
t
r
a
n
sf
o
r
m
e
r
o
n
l
y
9
0
.
1
4
%
8
9
.
6
2
%
9
0
.
3
1
%
8
9
.
7
2
%
0
.
9
2
1
V
i
T
+
B
e
h
a
v
i
o
r
a
l
(
C
o
n
c
a
t
e
n
a
t
i
o
n
)
9
2
.
0
8
%
9
1
.
7
3
%
9
1
.
5
8
%
9
1
.
6
5
%
0
.
9
4
3
V
i
T
+
C
r
o
ss
-
M
o
d
a
l
A
t
t
e
n
t
i
o
n
9
3
.
1
1
%
9
2
.
8
5
%
9
2
.
5
8
%
9
2
.
7
0
%
0
.
9
5
6
F
u
l
l
p
r
o
p
o
s
e
d
f
r
a
mew
o
r
k
9
4
.
8
7
%
9
3
.
9
5
%
9
5
.
4
2
%
9
4
.
6
8
%
0
.
9
7
2
4
.
3
.
Co
m
pa
riso
n wit
h
ex
is
t
ing
m
et
ho
ds
T
o
p
o
s
itio
n
th
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
a
g
ain
s
t
p
r
io
r
liter
atu
r
e,
m
u
ltip
le
p
e
r
f
o
r
m
an
ce
m
etr
ics
ar
e
co
m
p
ar
ed
with
r
ep
r
esen
tativ
e
ap
p
r
o
ac
h
es
as
in
T
ab
le
3.
T
h
e
p
r
o
p
o
s
ed
m
et
h
o
d
c
o
n
s
is
ten
tly
o
u
tp
er
f
o
r
m
s
ex
is
tin
g
tech
n
iq
u
es
ac
r
o
s
s
all
ev
alu
atio
n
m
etr
ics,
n
o
t
o
n
ly
ac
cu
r
ac
y
.
T
h
e
im
p
r
o
v
em
e
n
t
in
r
ec
all
in
d
icate
s
b
etter
d
etec
tio
n
o
f
en
g
ag
ed
s
tu
d
en
ts
,
wh
ile
h
ig
h
e
r
p
r
ec
is
io
n
r
ed
u
ce
s
f
alse e
n
g
ag
em
e
n
t p
r
e
d
i
ctio
n
s
.
T
ab
le
3
.
C
o
m
p
a
r
ativ
e
p
er
f
o
r
m
an
ce
with
ex
is
tin
g
s
tu
d
ies
M
e
t
h
o
d
A
c
c
u
r
a
c
y
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
S
c
o
r
e
El
b
a
w
a
b
a
n
d
H
e
n
r
i
q
u
e
s
[
7
]
8
9
.
3
0
%
8
8
.
9
2
%
8
8
.
6
0
%
8
8
.
7
5
%
S
h
a
h
e
t
a
l
.
[
3
]
8
5
.
6
0
%
8
4
.
8
8
%
8
5
.
1
0
%
8
4
.
9
2
%
P
a
n
d
e
y
e
t
a
l
.
[
4
]
8
3
.
4
0
%
8
2
.
9
5
%
8
2
.
8
0
%
8
2
.
8
8
%
P
r
o
p
o
se
d
f
r
a
m
e
w
o
r
k
9
4
.
8
7
%
9
3
.
9
5
%
9
5
.
4
2
%
9
4
.
6
8
%
4
.
4
.
Co
m
pu
t
a
t
io
na
l
e
f
f
iciency
T
o
en
s
u
r
e
p
r
ac
tical
f
ea
s
ib
ilit
y
,
co
m
p
u
tatio
n
al
m
etr
ics
ar
e
also
an
aly
ze
d
as
s
h
o
wn
in
T
a
b
le
4.
Desp
ite
a
h
ig
h
e
r
p
ar
am
eter
co
u
n
t,
th
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
m
ain
t
ain
s
co
m
p
etitiv
e
i
n
f
er
en
ce
s
p
ee
d
d
u
e
to
p
ar
allel
atten
tio
n
co
m
p
u
tatio
n
.
T
h
e
r
e
s
u
lts
co
n
f
ir
m
th
e
ef
f
ec
tiv
en
ess
o
f
th
e
p
r
o
p
o
s
ed
T
r
an
s
f
o
r
m
er
-
b
ased
cr
o
s
s
-
m
o
d
al
f
r
am
ewo
r
k
f
o
r
s
tu
d
en
t
atten
tiv
en
ess
esti
m
atio
n
.
T
h
e
m
o
d
el
ac
h
ie
v
es
s
u
p
er
i
o
r
ac
c
u
r
ac
y
a
n
d
F1
-
s
co
r
e
co
m
p
ar
ed
to
b
aselin
e
an
d
e
x
is
tin
g
ap
p
r
o
ac
h
es,
in
d
icatin
g
s
tr
o
n
g
o
v
er
all
class
if
icatio
n
ca
p
a
b
ilit
y
an
d
b
alan
ce
d
er
r
o
r
d
is
tr
ib
u
tio
n
.
I
n
ad
d
itio
n
to
ac
c
u
r
ac
y
,
th
e
f
r
am
ewo
r
k
m
ain
tain
s
s
tab
le
p
r
ec
is
io
n
an
d
r
ec
all
v
alu
es,
d
em
o
n
s
tr
atin
g
its
ab
ilit
y
to
co
r
r
ec
tly
id
en
tif
y
b
o
th
atten
tiv
e
an
d
in
atten
tiv
e
s
tu
d
en
ts
with
o
u
t
b
ias
to
war
d
a
s
p
ec
if
ic
class
.
T
h
e
h
ig
h
AU
C
f
u
r
th
er
r
ef
lects
s
tr
o
n
g
class
s
ep
ar
ab
ilit
y
an
d
r
eliab
le
d
i
s
cr
im
in
atio
n
ac
r
o
s
s
d
ec
is
io
n
th
r
esh
o
ld
s
.
T
h
e
i
n
clu
s
io
n
o
f
tr
an
s
f
o
r
m
er
-
b
ased
tem
p
o
r
al
en
c
o
d
in
g
s
ig
n
if
ican
tly
en
h
an
ce
s
p
er
f
o
r
m
an
ce
b
y
ca
p
tu
r
in
g
lo
n
g
-
r
an
g
e
en
g
ag
e
m
en
t
p
atter
n
s
th
at
s
tatic
f
r
am
e
-
lev
el
m
o
d
els
ca
n
n
o
t
r
ep
r
esen
t.
I
m
p
o
r
ta
n
tly
,
th
ese
p
e
r
f
o
r
m
a
n
ce
g
ain
s
d
o
n
o
t
co
m
p
r
o
m
i
s
e
co
m
p
u
tatio
n
al
f
ea
s
ib
ilit
y
.
Desp
ite
in
cr
ea
s
ed
ar
ch
itectu
r
al
co
m
p
lex
ity
,
th
e
m
o
d
el
p
r
eser
v
es
p
r
ac
tical
in
f
er
en
ce
ef
f
icien
cy
th
r
o
u
g
h
p
ar
allel
atten
tio
n
m
ec
h
an
is
m
s
,
m
ak
in
g
it su
itab
le
f
o
r
r
ea
l
-
tim
e
o
r
n
ea
r
r
ea
l
-
tim
e
d
ep
lo
y
m
en
t in
o
n
lin
e
lear
n
in
g
en
v
ir
o
n
m
en
ts
.
T
ab
le
4
.
Mo
d
el
co
m
p
lex
ity
an
d
in
f
er
e
n
ce
p
er
f
o
r
m
a
n
ce
M
o
d
e
l
P
a
r
a
me
t
e
r
s (
M
)
I
n
f
e
r
e
n
c
e
t
i
me
(
ms/fr
a
m
e
)
Th
r
o
u
g
h
p
u
t
(
F
P
S
)
C
N
N
+
LSTM
1
2
.
4
1
8
.
7
53
V
i
si
o
n
t
r
a
n
sf
o
r
m
e
r
2
1
.
8
1
6
.
2
61
P
r
o
p
o
se
d
f
r
a
m
e
w
o
r
k
2
4
.
6
1
7
.
4
57
5.
CO
NCLU
SI
O
N
T
h
is
p
ap
er
p
r
esen
ted
a
T
r
an
s
f
o
r
m
er
-
b
ased
cr
o
s
s
-
m
o
d
al
f
r
a
m
ewo
r
k
f
o
r
r
o
b
u
s
t
s
tu
d
en
t
atten
tiv
en
ess
esti
m
atio
n
in
o
n
lin
e
lear
n
in
g
en
v
ir
o
n
m
en
ts
.
T
h
e
p
r
o
p
o
s
ed
ar
ch
itectu
r
e
in
teg
r
ates
a
ViT
b
ac
k
b
o
n
e
f
o
r
g
lo
b
al
s
p
atial
d
ep
en
d
en
cy
m
o
d
elin
g
,
b
eh
av
io
r
al
f
ea
t
u
r
e
em
b
e
d
d
in
g
f
o
r
g
az
e
a
n
d
h
ea
d
p
o
s
e
r
ep
r
esen
tatio
n
,
cr
o
s
s
-
m
o
d
al
m
u
lti
-
h
ea
d
atten
tio
n
f
o
r
ad
ap
tiv
e
f
ea
tu
r
e
f
u
s
io
n
,
a
n
d
a
tr
an
s
f
o
r
m
er
-
b
ased
tem
p
o
r
a
l
en
co
d
er
f
o
r
lo
n
g
-
r
an
g
e
s
eq
u
en
ce
m
o
d
elin
g
.
Un
lik
e
co
n
v
e
n
tio
n
al
C
NN
–
L
ST
M
ar
ch
itectu
r
es,
th
e
f
r
am
ewo
r
k
is
f
u
lly
atten
tio
n
-
d
r
iv
en
,
en
a
b
lin
g
d
y
n
am
ic
m
u
ltimo
d
al
in
ter
ac
tio
n
with
o
u
t
r
ec
u
r
r
en
t
b
o
ttlen
ec
k
s
.
Qu
an
titativ
e
ev
alu
atio
n
d
em
o
n
s
tr
ates
th
e
ef
f
ec
tiv
en
es
s
o
f
th
e
p
r
o
p
o
s
ed
m
eth
o
d
.
T
h
e
m
o
d
el
ac
h
iev
ed
9
4
.
8
7
%
ac
cu
r
ac
y
,
9
3
.
9
5
%
p
r
ec
is
io
n
,
9
5
.
4
2
%
r
ec
all,
9
4
.
6
8
%
F1
-
s
co
r
e,
an
d
an
AUC
o
f
0
.
9
7
2
o
n
th
e
test
s
et.
C
o
m
p
ar
ed
to
th
e
C
NN
–
L
STM
b
aselin
e
(
8
6
.
3
2
%
ac
cu
r
ac
y
,
8
5
.
9
1
%
F1
-
s
co
r
e)
,
th
e
p
r
o
p
o
s
ed
a
p
p
r
o
ac
h
im
p
r
o
v
e
d
p
e
r
f
o
r
m
a
n
ce
b
y
m
o
r
e
th
an
8
%
in
ac
cu
r
a
cy
an
d
n
ea
r
l
y
9
%
in
F1
-
s
co
r
e.
T
h
e
ab
latio
n
s
tu
d
y
f
u
r
th
er
c
o
n
f
ir
m
e
d
in
cr
em
en
tal
g
ain
s
f
r
o
m
c
r
o
s
s
-
m
o
d
al
atten
tio
n
an
d
te
m
p
o
r
al
tr
an
s
f
o
r
m
er
en
co
d
in
g
.
Ad
d
itio
n
ally
,
th
e
m
o
d
el
m
ain
tain
ed
p
r
ac
tical
in
f
er
en
ce
ef
f
icien
cy
(
1
7
.
4
m
s
p
er
f
r
am
e
)
,
s
u
p
p
o
r
tin
g
n
ea
r
r
e
al
-
tim
e
d
ep
lo
y
m
en
t.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
15
,
No
.
3
,
Sep
tem
b
er
20
26
:
935
-
9
4
3
942
F
UNDING
I
NF
O
R
M
A
T
I
O
N
Au
th
o
r
s
s
tate
n
o
f
u
n
d
in
g
in
v
o
lv
ed
.
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
R
ajasek
ar
an
Ma
r
is
wam
y
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
Pra
v
ee
n
Su
n
d
a
r
✓
✓
✓
✓
✓
✓
✓
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
CO
NF
L
I
C
T
O
F
I
N
T
E
R
E
S
T
ST
A
T
E
M
E
NT
Au
th
o
r
s
s
tate
n
o
co
n
f
lict o
f
in
t
er
est.
DATA AV
AI
L
AB
I
L
I
T
Y
Data
av
ailab
ilit
y
is
n
o
t
ap
p
li
ca
b
le
to
th
is
p
ap
er
as
n
o
n
e
w
d
ata
wer
e
cr
ea
ted
o
r
an
aly
ze
d
in
th
is
s
tu
d
y
.
RE
F
E
R
E
NC
E
S
[
1
]
T.
P
.
N
e
g
r
o
n
a
n
d
C
.
A
.
G
r
a
v
e
s,
“
C
l
a
s
sr
o
o
m
a
t
t
e
n
t
i
v
e
n
e
ss
c
l
a
ss
i
f
i
c
a
t
i
o
n
t
o
o
l
(
C
l
a
ssA
C
T)
:
T
h
e
sy
s
t
e
m
i
n
t
r
o
d
u
c
t
i
o
n
,
”
i
n
2
0
1
7
I
E
EE
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
Pe
r
v
a
si
v
e
C
o
m
p
u
t
i
n
g
a
n
d
C
o
m
m
u
n
i
c
a
t
i
o
n
s W
o
rks
h
o
p
s
(
Pe
r
C
o
m
W
o
rk
sh
o
p
s)
,
M
a
r
.
2
0
1
7
,
p
p
.
2
6
–
2
9
,
d
o
i
:
1
0
.
1
1
0
9
/
p
e
r
c
o
mw
.
2
0
1
7
.
7
9
1
7
5
1
3
.
[
2
]
A
.
R
e
v
a
d
e
k
a
r
,
S
.
O
a
k
,
A
.
G
a
d
e
k
a
r
,
a
n
d
P
.
B
i
d
e
,
“
G
a
u
g
i
n
g
a
t
t
e
n
t
i
o
n
o
f
st
u
d
e
n
t
s
i
n
a
n
e
-
l
e
a
r
n
i
n
g
e
n
v
i
r
o
n
me
n
t
,
”
i
n
2
0
2
0
I
EE
E
4
t
h
C
o
n
f
e
re
n
c
e
o
n
I
n
f
o
rm
a
t
i
o
n
&
a
m
p
;
C
o
m
m
u
n
i
c
a
t
i
o
n
T
e
c
h
n
o
l
o
g
y
(
C
I
C
T
)
,
D
e
c
.
2
0
2
0
,
p
p
.
1
–
6
,
d
o
i
:
1
0
.
1
1
0
9
/
c
i
c
t
5
1
6
0
4
.
2
0
2
0
.
9
3
1
2
0
4
8
.
[
3
]
N.
A
.
S
h
a
h
,
K
.
M
e
e
n
a
k
sh
i
,
A
.
A
g
a
r
w
a
l
,
a
n
d
S
.
S
i
v
a
s
u
b
r
a
ma
n
i
a
n
,
“
A
ssessme
n
t
o
f
st
u
d
e
n
t
a
t
t
e
n
t
i
v
e
n
e
ss
t
o
e
-
l
e
a
r
n
i
n
g
b
y
mo
n
i
t
o
r
i
n
g
b
e
h
a
v
i
o
u
r
a
l
e
l
e
m
e
n
t
s,
”
i
n
2
0
2
1
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
C
o
m
p
u
t
e
r C
o
m
m
u
n
i
c
a
t
i
o
n
a
n
d
I
n
f
o
rm
a
t
i
c
s
(
I
C
C
C
I
)
,
Ja
n
.
2
0
2
1
,
p
p
.
1
–
7
,
d
o
i
:
1
0
.
1
1
0
9
/
i
c
c
c
i
5
0
8
2
6
.
2
0
2
1
.
9
4
0
2
2
8
3
.
[
4
]
R
.
K
.
P
a
n
d
e
y
,
A
.
A
.
F
a
r
i
d
i
,
a
n
d
G
.
S
h
r
i
v
a
s
t
a
v
a
,
“
A
t
t
e
n
t
i
v
e
n
e
ss
mea
s
u
r
e
i
n
c
l
a
ssr
o
o
m
e
n
v
i
r
o
n
me
n
t
u
si
n
g
f
a
c
e
d
e
t
e
c
t
i
o
n
,
”
i
n
2
0
2
1
6
t
h
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
I
n
v
e
n
t
i
v
e
C
o
m
p
u
t
a
t
i
o
n
T
e
c
h
n
o
l
o
g
i
e
s
(
I
C
I
C
T
)
,
Ja
n
.
2
0
2
1
,
p
p
.
1
0
5
3
–
1
0
5
8
,
d
o
i
:
1
0
.
1
1
0
9
/
i
c
i
c
t
5
0
8
1
6
.
2
0
2
1
.
9
3
5
8
6
0
0
.
[
5
]
R
.
P
a
i
,
A
.
D
u
b
e
y
,
a
n
d
N
.
M
a
n
g
a
o
n
k
a
r
,
“
R
e
a
l
t
i
m
e
e
y
e
m
o
n
i
t
o
r
i
n
g
s
y
s
t
e
m
u
s
i
n
g
C
N
N
f
o
r
d
r
o
w
s
i
n
e
s
s
a
n
d
a
t
t
e
n
t
i
v
e
n
e
s
s
s
y
s
t
e
m
,
”
i
n
2
0
2
1
A
s
i
a
n
C
o
n
f
e
r
e
n
c
e
o
n
I
n
n
o
v
a
t
i
o
n
i
n
T
e
c
h
n
o
l
o
g
y
(
A
S
I
A
N
C
O
N
)
,
A
u
g
.
2
0
2
1
,
p
p
.
1
–
4
,
d
o
i
:
1
0
.
1
1
0
9
/
a
s
i
a
n
c
o
n
5
1
3
4
6
.
2
0
2
1
.
9
5
4
4
6
2
4
.
[
6
]
J.
G
.
P
i
n
z
o
n
-
G
o
n
z
a
l
e
z
a
n
d
L.
B
a
r
b
a
-
G
u
a
ma
n
,
“
U
se
o
f
h
e
a
d
p
o
s
i
t
i
o
n
e
s
t
i
m
a
t
i
o
n
f
o
r
a
t
t
e
n
t
i
o
n
l
e
v
e
l
d
e
t
e
c
t
i
o
n
i
n
r
e
mo
t
e
c
l
a
ssr
o
o
ms,”
i
n
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
F
u
t
u
re
T
e
c
h
n
o
l
o
g
i
e
s
C
o
n
f
e
r
e
n
c
e
(
FT
C
)
2
0
2
1
,
V
o
l
u
m
e
1
,
S
p
r
i
n
g
e
r
I
n
t
e
r
n
a
t
i
o
n
a
l
P
u
b
l
i
s
h
i
n
g
,
2
0
2
1
,
p
p
.
2
7
5
–
2
9
3
.
[
7
]
M
.
El
b
a
w
a
b
a
n
d
R
.
H
e
n
r
i
q
u
e
s,
“
M
a
c
h
i
n
e
l
e
a
r
n
i
n
g
a
p
p
l
i
e
d
t
o
s
t
u
d
e
n
t
a
t
t
e
n
t
i
v
e
n
e
s
s
d
e
t
e
c
t
i
o
n
:
u
si
n
g
e
m
o
t
i
o
n
a
l
a
n
d
n
o
n
-
e
m
o
t
i
o
n
a
l
mea
su
r
e
s
,
”
Ed
u
c
a
t
i
o
n
a
n
d
I
n
f
o
rm
a
t
i
o
n
T
e
c
h
n
o
l
o
g
i
e
s
,
v
o
l
.
2
8
,
n
o
.
1
2
,
p
p
.
1
5
7
1
7
–
1
5
7
3
7
,
M
a
y
2
0
2
3
,
d
o
i
:
1
0
.
1
0
0
7
/
s
1
0
6
3
9
-
023
-
1
1
8
1
4
-
5.
[
8
]
J.
B
i
a
n
d
C
.
K
u
e
st
e
n
,
“
I
n
t
r
a
c
l
a
ss
c
o
r
r
e
l
a
t
i
o
n
c
o
e
f
f
i
c
i
e
n
t
(
<
s
c
p
>
I
C
C
<
/
scp
>
)
:
A
f
r
a
mew
o
r
k
f
o
r
m
o
n
i
t
o
r
i
n
g
a
n
d
a
ssess
i
n
g
p
e
r
f
o
r
m
a
n
c
e
o
f
t
r
a
i
n
e
d
s
e
n
s
o
r
y
p
a
n
e
l
s
a
n
d
p
a
n
e
l
i
st
s,”
J
o
u
r
n
a
l
o
f
S
e
n
so
r
y
S
t
u
d
i
e
s
,
v
o
l
.
2
7
,
n
o
.
5
,
p
p
.
3
5
2
–
3
6
4
,
2
0
1
2
,
d
o
i
:
1
0
.
1
1
1
1
/
j
.
1
7
4
5
-
4
5
9
x
.
2
0
1
2
.
0
0
3
9
9
.
x
.
[
9
]
S
.
T,
A
.
B
.
J
,
V
.
V
,
B
.
A
.
J,
G
.
V
.
A
.
S
,
a
n
d
D
.
K
,
“
A
ssessm
e
n
t
o
f
s
t
u
d
e
n
t
a
t
t
e
n
t
i
v
e
n
e
s
s
i
n
c
l
a
ssr
o
o
m
e
n
v
i
r
o
n
me
n
t
u
s
i
n
g
d
e
e
p
l
e
a
r
n
i
n
g
,
”
i
n
2
0
2
3
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
re
n
c
e
o
n
C
i
r
c
u
i
t
P
o
w
e
r
a
n
d
C
o
m
p
u
t
i
n
g
T
e
c
h
n
o
l
o
g
i
e
s
(
I
C
C
PC
T
)
,
A
u
g
.
2
0
2
3
,
p
p
.
2
6
–
3
0
,
d
o
i
:
1
0
.
1
1
0
9
/
i
c
c
p
c
t
5
8
3
1
3
.
2
0
2
3
.
1
0
2
4
5
1
9
4
.
[
1
0
]
S
.
G
u
p
t
a
a
n
d
P
.
K
u
mar,
“
A
t
t
e
n
t
i
o
n
r
e
c
o
g
n
i
t
i
o
n
s
y
s
t
e
m i
n
o
n
l
i
n
e
l
e
a
r
n
i
n
g
p
l
a
t
f
o
r
m u
s
i
n
g
E
EG
si
g
n
a
l
s
,
”
i
n
Em
e
r
g
i
n
g
T
e
c
h
n
o
l
o
g
i
e
s
f
o
r
S
m
a
rt
C
i
t
i
e
s
,
S
p
r
i
n
g
e
r
S
i
n
g
a
p
o
r
e
,
2
0
2
1
,
p
p
.
1
3
9
–
1
5
2
.
[1
1
]
B
.
S
u
n
e
t
a
l
.
,
“
S
t
u
d
e
n
t
c
l
a
ss
b
e
h
a
v
i
o
r
d
a
t
a
se
t
:
a
v
i
d
e
o
d
a
t
a
set
f
o
r
r
e
c
o
g
n
i
z
i
n
g
,
d
e
t
e
c
t
i
n
g
,
a
n
d
c
a
p
t
i
o
n
i
n
g
st
u
d
e
n
t
s
’
b
e
h
a
v
i
o
r
s
i
n
c
l
a
ssr
o
o
m
sce
n
e
s,”
N
e
u
ra
l
C
o
m
p
u
t
i
n
g
a
n
d
Ap
p
l
i
c
a
t
i
o
n
s
,
v
o
l
.
3
3
,
n
o
.
1
4
,
p
p
.
8
3
3
5
–
8
3
5
4
,
Ja
n
.
2
0
2
1
,
d
o
i
:
1
0
.
1
0
0
7
/
s
0
0
5
2
1
-
020
-
0
5
5
8
7
-
y.
[1
2
]
M
.
S
a
r
i
,
A
.
M
o
u
ssa
o
u
i
,
a
n
d
A
.
H
a
d
i
d
,
“
A
si
m
p
l
e
y
e
t
e
f
f
e
c
t
i
v
e
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
m
o
d
e
l
t
o
c
l
a
ss
i
f
y
f
a
c
i
a
l
e
x
p
r
e
ssi
o
n
s
,
”
i
n
M
o
d
e
l
l
i
n
g
a
n
d
I
m
p
l
e
m
e
n
t
a
t
i
o
n
o
f
C
o
m
p
l
e
x
S
y
s
t
e
m
s
,
S
p
r
i
n
g
e
r
I
n
t
e
r
n
a
t
i
o
n
a
l
P
u
b
l
i
s
h
i
n
g
,
2
0
2
0
,
p
p
.
1
8
8
–
2
0
2
.
[1
3
]
A
.
A
.
R
a
v
i
n
d
r
a
n
,
“
I
n
t
e
r
n
e
t
-
of
-
t
h
i
n
g
s
e
d
g
e
c
o
m
p
u
t
i
n
g
s
y
st
e
ms
f
o
r
s
t
r
e
a
mi
n
g
v
i
d
e
o
a
n
a
l
y
t
i
c
s:
t
r
a
i
l
s
b
e
h
i
n
d
a
n
d
t
h
e
p
a
t
h
s
a
h
e
a
d
,
”
I
o
T
,
v
o
l
.
4
,
n
o
.
4
,
p
p
.
4
8
6
–
5
1
3
,
O
c
t
.
2
0
2
3
,
d
o
i
:
1
0
.
3
3
9
0
/
i
o
t
4
0
4
0
0
2
1
.
[1
4
]
A
.
I
.
W
a
n
g
a
n
d
R
.
Ta
h
i
r
,
“
Th
e
e
f
f
e
c
t
o
f
u
si
n
g
K
a
h
o
o
t
!
f
o
r
l
e
a
r
n
i
n
g
–
A
l
i
t
e
r
a
t
u
r
e
r
e
v
i
e
w
,
”
C
o
m
p
u
t
e
rs &a
m
p
;
Ed
u
c
a
t
i
o
n
,
v
o
l
.
1
4
9
,
p
.
1
0
3
8
1
8
,
M
a
y
2
0
2
0
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
c
o
m
p
e
d
u
.
2
0
2
0
.
1
0
3
8
1
8
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
C
r
o
s
s
-
mo
d
a
l a
tten
tio
n
f
u
s
io
n
u
s
in
g
visi
o
n
tr
a
n
s
fo
r
mers
fo
r
r
o
b
u
s
t
s
tu
d
en
t
… (
R
a
ja
s
ek
a
r
a
n
Ma
r
is
w
a
my
)
943
[1
5
]
F.
-
C
.
Li
n
,
H
.
-
H
.
N
g
o
,
C
.
-
R
.
D
o
w
,
K
.
-
H
.
L
a
m,
a
n
d
H
.
L.
L
e
,
“
S
t
u
d
e
n
t
b
e
h
a
v
i
o
r
r
e
c
o
g
n
i
t
i
o
n
sy
s
t
e
m
f
o
r
t
h
e
c
l
a
ssr
o
o
m
e
n
v
i
r
o
n
me
n
t
b
a
s
e
d
o
n
s
k
e
l
e
t
o
n
p
o
se
e
st
i
ma
t
i
o
n
a
n
d
p
e
r
s
o
n
d
e
t
e
c
t
i
o
n
,
”
S
e
n
s
o
rs
,
v
o
l
.
2
1
,
n
o
.
1
6
,
p
.
5
3
1
4
,
A
u
g
.
2
0
2
1
,
d
o
i
:
1
0
.
3
3
9
0
/
s
2
1
1
6
5
3
1
4
.
[1
6
]
J.
S
h
e
n
,
H
.
Y
a
n
g
,
J.
L
i
,
a
n
d
Z
.
C
h
e
n
g
,
“
A
sse
ssi
n
g
l
e
a
r
n
i
n
g
e
n
g
a
g
e
me
n
t
b
a
s
e
d
o
n
f
a
c
i
a
l
e
x
p
r
e
s
si
o
n
r
e
c
o
g
n
i
t
i
o
n
i
n
M
O
O
C
’
s
sce
n
a
r
i
o
,
”
M
u
l
t
i
m
e
d
i
a
S
y
s
t
e
m
s
,
v
o
l
.
2
8
,
n
o
.
2
,
p
p
.
4
6
9
–
4
7
8
,
O
c
t
.
2
0
2
1
,
d
o
i
:
1
0
.
1
0
0
7
/
s
0
0
5
3
0
-
0
2
1
-
0
0
8
5
4
-
x.
[1
7
]
M.
-
E.
O
t
g
o
n
b
o
l
d
e
t
a
l
.
,
“
S
H
EL5
K
:
a
n
e
x
t
e
n
d
e
d
d
a
t
a
set
a
n
d
b
e
n
c
h
mar
k
i
n
g
f
o
r
safe
t
y
h
e
l
m
e
t
d
e
t
e
c
t
i
o
n
,
”
S
e
n
so
rs
,
v
o
l
.
2
2
,
n
o
.
6
,
p
.
2
3
1
5
,
M
a
r
.
2
0
2
2
,
d
o
i
:
1
0
.
3
3
9
0
/
s2
2
0
6
2
3
1
5
.
[
18
]
S
.
W
a
n
g
,
C
.
Zh
a
n
g
,
Y
.
S
h
u
,
a
n
d
Y
.
Li
u
,
“
L
i
v
e
v
i
d
e
o
a
n
a
l
y
t
i
c
s
w
i
t
h
F
P
G
A
-
b
a
se
d
smar
t
c
a
mer
a
s,”
i
n
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
2
0
1
9
Wo
r
k
sh
o
p
o
n
H
o
t
T
o
p
i
c
s
i
n
V
i
d
e
o
A
n
a
l
y
t
i
c
s
a
n
d
I
n
t
e
l
l
i
g
e
n
t
E
d
g
e
s
,
O
c
t
.
2
0
1
9
,
p
p
.
9
–
1
4
,
d
o
i
:
1
0
.
1
1
4
5
/
3
3
4
9
6
1
4
.
3
3
5
6
0
2
7
.
[
19
]
S
.
W
i
l
l
e
r
m
a
r
k
a
n
d
M
.
G
e
l
l
e
r
st
e
d
t
,
“
F
a
c
i
n
g
r
a
d
i
c
a
l
d
i
g
i
t
a
l
i
z
a
t
i
o
n
:
c
a
p
t
u
r
i
n
g
t
e
a
c
h
e
r
s’
t
r
a
n
s
i
t
i
o
n
t
o
v
i
r
t
u
a
l
c
l
a
ssr
o
o
ms
t
h
r
o
u
g
h
i
d
e
a
l
t
y
p
e
e
x
p
e
r
i
e
n
c
e
s
,
”
J
o
u
r
n
a
l
o
f
Ed
u
c
a
t
i
o
n
a
l
C
o
m
p
u
t
i
n
g
Re
se
a
r
c
h
,
v
o
l
.
6
0
,
n
o
.
6
,
p
p
.
1
3
5
1
–
1
3
7
2
,
F
e
b
.
2
0
2
2
,
d
o
i
:
1
0
.
1
1
7
7
/
0
7
3
5
6
3
3
1
2
1
1
0
6
9
4
2
4
.
[2
0
]
X
.
W
a
n
g
,
T
.
L
i
u
,
J.
W
a
n
g
,
a
n
d
J.
Ti
a
n
,
“
U
n
d
e
r
s
t
a
n
d
i
n
g
l
e
a
r
n
e
r
c
o
n
t
i
n
u
a
n
c
e
i
n
t
e
n
t
i
o
n
:
a
c
o
mp
a
r
i
s
o
n
o
f
l
i
v
e
v
i
d
e
o
l
e
a
r
n
i
n
g
,
p
r
e
-
r
e
c
o
r
d
e
d
v
i
d
e
o
l
e
a
r
n
i
n
g
a
n
d
h
y
b
r
i
d
v
i
d
e
o
l
e
a
r
n
i
n
g
i
n
C
O
V
I
D
-
1
9
p
a
n
d
e
mi
c
,
”
I
n
t
e
r
n
a
t
i
o
n
a
l
J
o
u
rn
a
l
o
f
H
u
m
a
n
–
C
o
m
p
u
t
e
r
I
n
t
e
r
a
c
t
i
o
n
,
v
o
l
.
3
8
,
n
o
.
3
,
p
p
.
2
6
3
–
2
8
1
,
2
0
2
1
,
d
o
i
:
1
0
.
1
0
8
0
/
1
0
4
4
7
3
1
8
.
2
0
2
1
.
1
9
3
8
3
8
9
.
[2
1
]
Y
.
W
u
,
L
.
Z
h
a
n
g
,
G
.
C
h
e
n
,
a
n
d
P
.
N
.
M
i
c
h
e
l
i
n
i
,
“
U
n
c
o
n
s
t
r
a
i
n
e
d
f
a
c
i
a
l
e
x
p
r
e
ssi
o
n
r
e
c
o
g
n
i
t
o
n
b
a
se
d
o
n
c
a
sc
a
d
e
d
e
c
i
s
i
o
n
a
n
d
g
a
b
o
r
f
i
l
t
e
r
s,”
i
n
2
0
2
0
2
5
t
h
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
P
a
t
t
e
r
n
Re
c
o
g
n
i
t
i
o
n
(
I
C
P
R)
,
J
a
n
.
2
0
2
1
,
p
p
.
3
3
3
6
–
3
3
4
1
,
d
o
i
:
1
0
.
1
1
0
9
/
i
c
p
r
4
8
8
0
6
.
2
0
2
1
.
9
4
1
1
9
8
3
.
[
2
2
]
R
.
P
e
r
i
y
a
sam
y
,
M
.
G
o
v
i
n
d
a
r
a
j
i
,
I
.
N
a
su
r
u
l
l
a
,
V
.
S
r
i
n
i
v
a
s
a
n
,
a
n
d
K
.
R
a
ma
D
e
v
i
,
“
O
p
t
i
m
i
z
i
n
g
c
a
l
l
c
e
n
t
e
r
a
g
e
n
t
e
f
f
i
c
i
e
n
c
y
t
h
r
o
u
g
h
d
e
e
p
l
e
a
r
n
i
n
g
-
b
a
s
e
d
c
l
a
ss
i
f
i
c
a
t
i
o
n
s
u
s
i
n
g
S
M
F
C
C
A
E
,
”
I
n
t
e
r
n
a
t
i
o
n
a
l
J
o
u
r
n
a
l
o
f
Re
c
o
n
f
i
g
u
r
a
b
l
e
a
n
d
Em
b
e
d
d
e
d
S
y
s
t
e
m
s
(
I
J
RE
S
)
,
v
o
l
.
1
5
,
n
o
.
1
,
2
0
2
6
[2
3
]
Z.
Tr
a
b
e
l
s
i
,
F
.
A
l
n
a
j
j
a
r
,
M
.
M
.
A
.
P
a
r
a
mb
i
l
,
M
.
G
o
c
h
o
o
,
a
n
d
L.
A
l
i
,
“
R
e
a
l
-
t
i
me
a
t
t
e
n
t
i
o
n
m
o
n
i
t
o
r
i
n
g
s
y
st
e
m
f
o
r
c
l
a
s
sr
o
o
m:
a
d
e
e
p
l
e
a
r
n
i
n
g
a
p
p
r
o
a
c
h
f
o
r
st
u
d
e
n
t
’
s
b
e
h
a
v
i
o
r
r
e
c
o
g
n
i
t
i
o
n
,
”
Bi
g
D
a
t
a
a
n
d
C
o
g
n
i
t
i
v
e
C
o
m
p
u
t
i
n
g
,
v
o
l
.
7
,
n
o
.
1
,
p
.
4
8
,
M
a
r
.
2
0
2
3
,
d
o
i
:
1
0
.
3
3
9
0
/
b
d
c
c
7
0
1
0
0
4
8
.
[2
4
]
Z.
W
a
n
g
,
F
.
Ze
n
g
,
S
.
L
i
u
,
a
n
d
B
.
Ze
n
g
,
“
O
A
EN
e
t
:
o
r
i
e
n
t
e
d
a
t
t
e
n
t
i
o
n
e
n
s
e
mb
l
e
f
o
r
a
c
c
u
r
a
t
e
f
a
c
i
a
l
e
x
p
r
e
ss
i
o
n
r
e
c
o
g
n
i
t
i
o
n
,
”
Pa
t
t
e
r
n
Re
c
o
g
n
i
t
i
o
n
,
v
o
l
.
1
1
2
,
p
.
1
0
7
6
9
4
,
A
p
r
.
2
0
2
1
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
p
a
t
c
o
g
.
2
0
2
0
.
1
0
7
6
9
4
.
[2
5
]
S
.
H
u
ss
a
i
n
a
n
d
J
.
J
e
y
a
c
h
i
d
r
a
,
“
A
st
a
c
k
e
d
c
l
a
ss
i
f
i
e
r
mo
d
e
l
f
o
r
e
n
h
a
n
c
e
d
st
u
d
e
n
t
p
e
r
f
o
r
ma
n
c
e
p
r
e
d
i
c
t
i
o
n
i
n
e
-
l
e
a
r
n
i
n
g
e
n
v
i
r
o
n
m
e
n
t
s
,
”
I
n
d
o
n
e
s
i
a
n
J
o
u
r
n
a
l
o
f
El
e
c
t
r
i
c
a
l
E
n
g
i
n
e
e
r
i
n
g
a
n
d
I
n
f
o
rm
a
t
i
c
s (I
J
EEI)
,
v
o
l
.
1
4
,
n
o
.
1
,
2
0
2
6
.
B
I
O
G
RAP
H
I
E
S O
F
AUTH
O
RS
Ra
ja
se
k
a
r
a
n
Ma
r
iswa
m
y
is
a
P
h
.
D.
sc
h
o
lar
a
t
Ad
h
ip
a
ra
sa
k
t
h
i
Co
l
leg
e
o
f
Arts
a
n
d
S
c
ien
c
e
,
sp
e
c
ializin
g
i
n
Artif
icia
l
In
telli
g
e
n
c
e
a
n
d
M
a
c
h
i
n
e
Lea
rn
in
g
.
His
re
se
a
rc
h
fo
c
u
se
s
o
n
E
-
lea
rn
in
g
a
n
a
l
y
ti
c
s,
p
a
rti
c
u
larly
e
x
a
m
in
i
n
g
stu
d
e
n
t
a
t
t
e
n
ti
v
e
n
e
ss
a
n
d
a
ss
e
ss
m
e
n
t
m
e
th
o
d
o
lo
g
ies
i
n
li
v
e
lec
tu
re
s
c
o
m
p
a
re
d
t
o
p
re
-
re
c
o
rd
e
d
v
i
d
e
o
-
b
a
se
d
in
str
u
c
ti
o
n
.
His
wo
r
k
a
ims
to
d
e
v
e
lo
p
in
tell
ig
e
n
t
,
d
a
ta
-
d
riv
e
n
fra
m
e
wo
rk
s
t
o
e
n
h
a
n
c
e
e
n
g
a
g
e
m
e
n
t
m
o
n
i
to
ri
n
g
a
n
d
lea
rn
in
g
o
u
tco
m
e
s
i
n
d
ig
it
a
l
e
d
u
c
a
ti
o
n
e
n
v
ir
o
n
m
e
n
ts.
He
se
rv
e
s
a
s
a
Lab
o
ra
to
r
y
Tec
h
n
icia
n
in
th
e
De
p
a
rtme
n
t
o
f
Co
m
p
u
ter
S
c
ien
c
e
a
t
Th
iru
v
a
ll
u
v
a
r
Un
i
v
e
rsity
.
His
a
c
a
d
e
m
ic
in
tere
sts
in
c
lu
d
e
d
e
e
p
lea
rn
in
g
,
e
d
u
c
a
ti
o
n
a
l
d
a
ta
m
in
in
g
,
lea
rn
in
g
a
n
a
l
y
ti
c
s,
a
n
d
AI
-
d
ri
v
e
n
p
e
rfo
rm
a
n
c
e
e
v
a
lu
a
ti
o
n
sy
ste
m
s.
He
is
c
o
m
m
it
ted
to
a
d
v
a
n
c
in
g
tec
h
n
o
l
o
g
y
-
e
n
a
b
led
e
d
u
c
a
ti
o
n
th
r
o
u
g
h
in
n
o
v
a
ti
v
e
a
n
d
sc
a
lab
le AI
so
lu
ti
o
n
s.
He
c
a
n
b
e
c
o
n
tac
ted
a
t
e
m
a
il
:
m
a
ra
jas
e
k
a
ra
n
@g
m
a
il
.
c
o
m
.
Dr
.
P
ra
v
ee
n
Su
nd
a
r
M
CA,
M
.
P
h
il
.
,
P
h
.
D.
is
a
n
I
n
d
ian
a
c
a
d
e
m
icia
n
a
n
d
re
se
a
rc
h
e
r
c
u
rre
n
tl
y
se
rv
i
n
g
a
s
As
so
c
iate
P
ro
fe
ss
o
r
&
He
a
d
o
f
t
h
e
De
p
a
rtme
n
t
o
f
C
o
m
p
u
ter
S
c
ien
c
e
a
n
d
Ap
p
li
c
a
ti
o
n
s
a
t
Ad
h
i
p
a
ra
sa
k
th
i
Co
ll
e
g
e
o
f
Arts
a
n
d
S
c
ien
c
e
,
G
.
B.
Na
g
a
r,
Ka
lav
a
i
–
6
3
2
5
0
6
,
Tam
il
Na
d
u
,
In
d
i
a
.
He
lea
d
s
th
e
d
e
p
a
rtme
n
t
in
b
o
t
h
tea
c
h
in
g
a
n
d
re
se
a
rc
h
,
with
re
sp
o
n
si
b
il
it
ies
sp
a
n
n
i
n
g
c
u
r
ricu
lu
m
d
e
v
e
lo
p
m
e
n
t,
re
se
a
rc
h
su
p
e
rv
isi
o
n
,
a
n
d
a
c
a
d
e
m
ic
a
d
m
in
istratio
n
.
Dr.
S
u
n
d
a
r
b
e
g
a
n
h
is
a
c
a
d
e
m
ic
jo
u
r
n
e
y
wit
h
a
Ba
c
h
e
lo
r
o
f
C
o
m
p
u
ter
Ap
p
li
c
a
ti
o
n
s
(BCA),
fo
ll
o
w
e
d
b
y
a
M
a
ste
r
o
f
Co
m
p
u
ter
Ap
p
li
c
a
ti
o
n
s
(
M
CA).
He
we
n
t
o
n
to
e
a
rn
a
n
M
.
P
h
il
.
in
C
o
m
p
u
ter
S
c
ien
c
e
a
n
d
late
r
c
o
m
p
lete
d
h
is
P
h
.
D.,
wit
h
d
o
c
to
ra
l
re
se
a
rc
h
fo
c
u
se
d
o
n
e
d
u
c
a
ti
o
n
a
l
d
a
ta
m
in
in
g
a
n
d
st
u
d
e
n
t
e
n
g
a
g
e
m
e
n
t.
Be
fo
re
jo
in
i
n
g
h
is
c
u
rre
n
t
in
stit
u
ti
o
n
,
h
e
g
a
i
n
e
d
e
x
p
e
rie
n
c
e
in
b
o
th
in
d
u
str
y
a
n
d
a
c
a
d
e
m
ia,
in
c
l
u
d
i
n
g
ro
les
i
n
s
o
ftwa
re
d
e
v
e
lo
p
m
e
n
t
a
n
d
a
s
He
a
d
o
f
t
h
e
M
CA
De
p
a
rtme
n
t
a
t
a
n
o
th
e
r
c
o
ll
e
g
e
.
His
re
se
a
rc
h
in
tere
sts
in
c
lu
d
e
m
a
c
h
in
e
lea
rn
i
n
g
,
d
a
ta
m
in
in
g
,
e
-
lea
rn
in
g
,
d
a
ta
re
tri
e
v
a
l,
a
n
d
s
u
p
e
r
v
ise
d
/u
n
su
p
e
rv
ise
d
lea
rn
in
g
,
a
n
d
h
e
h
a
s
p
u
b
l
ish
e
d
m
u
lt
ip
le
p
a
p
e
rs
in
i
n
tern
a
ti
o
n
a
l
j
o
u
r
n
a
ls
a
n
d
c
o
n
fe
re
n
c
e
s.
Dr.
S
u
n
d
a
r
h
a
s
a
lso
fil
e
d
a
p
a
te
n
t
re
l
a
ted
to
a
n
“
I
n
telli
g
e
n
t
Dr
u
g
Ab
u
se
Asc
e
rta
in
S
y
ste
m
.
He
c
a
n
b
e
c
o
n
tac
ted
a
t
e
m
a
il
:
p
ra
v
e
e
n
su
n
d
a
rp
v
@
g
m
a
il
.
c
o
m
.
Evaluation Warning : The document was created with Spire.PDF for Python.