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
.
1
0
9
7
~
1
1
1
4
I
SS
N:
2252
-
8
7
7
6
,
DOI
:
1
0
.
1
1
5
9
1
/iji
ct
.
v15
i
3
.
pp
1
0
9
7
-
1
1
1
4
1097
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ict.
ia
esco
r
e.
co
m
Exploring
mul
ti
-
a
nswer v
isua
l ques
tion a
nswering
wi
th
o
bje
ct
detec
tion: a sy
ste
ma
tic
revi
ew
Nida
ul
H
a
s
a
na
t
i
1,
2
,
T
a
ufik
D
j
a
t
na
3
,
I
ma
s
Su
k
a
esih
Sita
ng
g
a
ng
4
,
Arif
I
m
a
m
Su
ro
s
o
5
1
D
o
c
t
o
r
a
l
P
r
o
g
r
a
m i
n
C
o
mp
u
t
e
r
S
c
i
e
n
c
e
,
S
c
h
o
o
l
o
f
D
a
t
a
S
c
i
e
n
c
e
,
M
a
t
h
e
ma
t
i
c
s,
a
n
d
I
n
f
o
r
mat
i
c
s
,
I
P
B
U
n
i
v
e
r
si
t
y
,
B
o
g
o
r
,
I
n
d
o
n
e
si
a
2
D
e
p
a
r
t
me
n
t
o
f
I
n
f
o
r
mat
i
o
n
S
y
s
t
e
ms
,
S
y
a
r
i
f
H
i
d
a
y
a
t
u
l
l
a
h
S
t
a
t
e
I
sl
a
m
i
c
U
n
i
v
e
r
si
t
y
(
U
I
N
)
,
Jak
a
r
t
a
,
I
n
d
o
n
e
s
i
a
3
D
i
v
i
si
o
n
o
f
A
g
r
o
i
n
d
u
s
t
r
i
a
l
E
n
g
i
n
e
e
r
i
n
g
a
n
d
Te
c
h
n
o
l
o
g
y
,
F
a
c
u
l
t
y
o
f
E
n
g
i
n
e
e
r
i
n
g
a
n
d
T
e
c
h
n
o
l
o
g
y
,
I
P
B
U
n
i
v
e
r
si
t
y
,
B
o
g
o
r
,
I
n
d
o
n
e
s
i
a
4
S
c
h
o
o
l
o
f
D
a
t
a
S
c
i
e
n
c
e
,
M
a
t
h
e
ma
t
i
c
s
,
a
n
d
I
n
f
o
r
mat
i
c
s
,
I
P
B
U
n
i
v
e
r
si
t
y
,
B
o
g
o
r
,
I
n
d
o
n
e
si
a
5
S
c
h
o
o
l
o
f
B
u
s
i
n
e
ss,
I
P
B
U
n
i
v
e
r
s
i
t
y
,
B
o
g
o
r
,
I
n
d
o
n
e
si
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
J
an
2
6
,
2
0
2
6
R
ev
is
ed
J
u
n
2
,
2
0
2
6
Acc
ep
ted
J
u
l 5
,
2
0
2
6
Visu
a
l
q
u
e
stio
n
a
n
sw
e
rin
g
(VQ
A)
is
a
c
h
a
ll
e
n
g
in
g
re
se
a
rc
h
a
re
a
th
a
t
e
n
a
b
les
m
a
c
h
in
e
s
to
a
n
sw
e
r
n
a
tu
ra
l
lan
g
u
a
g
e
q
u
e
sti
o
n
s
b
a
se
d
o
n
v
isu
a
l
c
o
n
ten
t
b
y
jo
in
t
ly
u
n
d
e
rsta
n
d
i
n
g
ima
g
e
s
a
n
d
tex
t.
Co
n
v
e
n
ti
o
n
a
l
VQ
A
sy
ste
m
s
ty
p
ica
ll
y
p
ro
d
u
c
e
a
sin
g
le
a
n
sw
e
r
fo
r
e
a
c
h
ima
g
e
–
q
u
e
sti
o
n
p
a
ir.
H
o
we
v
e
r,
m
a
n
y
re
a
l
-
wo
rld
v
isu
a
l
q
u
e
stio
n
s
a
re
a
m
b
i
g
u
o
u
s
o
r
c
o
m
p
lex
,
a
ll
o
wi
n
g
m
u
l
ti
p
le
v
a
li
d
a
n
sw
e
rs
to
e
x
ist.
Th
is
sy
ste
m
a
ti
c
li
tera
tu
re
re
v
iew
(S
LR)
fo
c
u
se
s
o
n
m
u
lt
i
-
a
n
sw
e
r
VQ
A
sy
ste
m
s
a
n
d
th
e
u
se
o
f
o
b
jec
t
d
e
tec
ti
o
n
,
fo
l
lo
win
g
th
e
P
RIS
M
A
2
0
2
0
g
u
id
e
li
n
e
s.
We
a
n
a
ly
z
e
d
5
8
p
e
e
r
-
re
v
iew
e
d
jo
u
rn
a
l
a
rti
c
les
re
tri
e
v
e
d
fr
o
m
t
h
e
S
c
o
p
u
s
d
a
tab
a
se
p
u
b
li
sh
e
d
b
e
twe
e
n
2
0
2
0
a
n
d
2
0
2
5
.
Te
n
o
f
th
e
se
stu
d
ies
c
lea
rly
sta
ted
t
h
a
t
g
e
n
e
ra
ti
n
g
m
u
lt
ip
le
a
n
sw
e
rs
wa
s
th
e
ir
m
a
in
g
o
a
l.
F
o
rt
y
-
e
ig
h
t
o
th
e
rs
i
n
d
irec
tl
y
su
p
p
o
rte
d
a
n
sw
e
r
v
a
ri
a
b
il
it
y
b
y
u
sin
g
o
b
jec
t
-
b
a
se
d
o
r
m
u
lt
i
-
in
st
a
n
c
e
re
a
so
n
in
g
.
T
h
ro
u
g
h
th
is
re
v
iew
,
we
e
x
a
m
in
e
th
e
c
u
rre
n
t
m
e
th
o
d
o
l
o
g
i
e
s
fo
r
s
u
p
p
o
rti
n
g
m
u
lt
i
-
a
n
sw
e
r
g
e
n
e
ra
ti
o
n
,
in
c
lu
d
in
g
m
o
d
e
l
a
rc
h
i
tec
tu
re
,
d
a
t
a
se
ts,
a
n
d
e
v
a
lu
a
ti
o
n
m
e
tri
c
s.
M
o
st
m
u
l
ti
-
a
n
sw
e
r
g
e
n
e
ra
ti
o
n
a
p
p
r
o
a
c
h
e
s
u
ti
li
z
e
a
tt
e
n
ti
o
n
m
e
c
h
a
n
ism
s,
g
ra
p
h
n
e
u
ra
l
n
e
two
rk
s,
a
n
d
tran
sf
o
rm
e
r
-
b
a
se
d
m
o
d
e
ls.
A
d
d
it
i
o
n
a
l
ly
,
we
p
ro
p
o
se
a
ta
x
o
n
o
m
y
o
f
m
u
lt
i
-
a
n
sw
e
r
VQ
A
o
rg
a
n
ize
d
a
lo
n
g
f
o
u
r
d
i
m
e
n
sio
n
s.
Li
m
it
a
ti
o
n
s
a
re
id
e
n
t
ifi
e
d
in
d
a
t
a
se
ts
a
n
d
e
v
a
lu
a
ti
o
n
m
e
tri
c
s
(i.
e
.
,
a
n
sw
e
r
a
m
b
ig
u
it
y
/su
b
jec
ti
v
i
ty
).
F
u
tu
re
r
e
se
a
rc
h
sh
o
u
ld
fo
c
u
s
o
n
imp
r
o
v
i
n
g
m
o
d
e
l
in
terp
re
tab
il
it
y
a
n
d
d
e
sig
n
in
g
a
n
e
v
a
l
u
a
ti
o
n
fr
a
m
e
wo
rk
t
h
a
t
i
n
c
o
rp
o
ra
tes
su
b
jec
ti
v
e
a
n
d
c
o
n
tex
t
-
se
n
si
ti
v
e
r
e
sp
o
n
se
s.
K
ey
w
o
r
d
s
:
Mu
lti
-
an
s
wer
Mu
lti
-
lab
el
Mu
ltip
le
an
s
wer
s
Ob
ject
d
etec
tio
n
Sy
s
tem
atic
r
ev
iew
Vis
u
al
q
u
esti
o
n
an
s
wer
in
g
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
:
T
au
f
ik
Djatn
a
Div
is
io
n
o
f
Ag
r
o
in
d
u
s
tr
ial
E
n
g
in
ee
r
in
g
a
n
d
T
ec
h
n
o
l
o
g
y
,
Facu
lty
o
f
E
n
g
in
ee
r
i
n
g
an
d
T
ec
h
n
o
lo
g
y
I
PB
Un
iv
er
s
ity
B
o
g
o
r
,
I
n
d
o
n
esia
E
m
ail: ta
u
f
ik
d
jatn
a@
ap
p
s
.
ip
b
.
ac
.
id
1.
I
NT
RO
D
UCT
I
O
N
Vis
u
al
q
u
esti
o
n
an
s
wer
in
g
(
VQA)
r
ep
r
esen
ts
a
m
u
ltid
is
cip
lin
ar
y
r
esear
ch
d
o
m
ai
n
th
at
co
n
v
er
g
es
co
m
p
u
ter
v
is
io
n
a
n
d
n
atu
r
al
l
an
g
u
ag
e
p
r
o
ce
s
s
in
g
to
en
ab
le
m
ac
h
in
es
to
co
m
p
r
eh
e
n
d
an
d
r
esp
o
n
d
to
n
atu
r
al
lan
g
u
ag
e
q
u
er
ies
b
ased
o
n
v
is
u
al
co
n
ten
t
[
1
]
.
T
h
is
r
ap
id
ly
ev
o
lv
i
n
g
f
ield
h
as
witn
ess
ed
s
u
b
s
tan
tial
ad
v
an
ce
m
e
n
t
th
r
o
u
g
h
t
h
e
in
t
eg
r
atio
n
o
f
d
ee
p
lear
n
in
g
tec
h
n
o
lo
g
ies,
th
e
g
r
o
wth
o
f
la
r
g
e
-
s
ca
le
m
u
ltimo
d
al
d
atasets
,
an
d
th
e
d
e
v
elo
p
m
e
n
t
o
f
s
o
p
h
is
ticated
v
is
io
n
-
la
n
g
u
ag
e
ap
p
r
o
ac
h
es
ac
r
o
s
s
d
iv
er
s
e
ap
p
licatio
n
d
o
m
ain
s
,
in
clu
d
in
g
h
ea
lth
ca
r
e,
r
o
b
o
tics
,
r
em
o
te
s
en
s
in
g
,
an
d
ar
tific
ial
in
tellig
en
ce
s
y
s
tem
s
[
2
]
,
[
3
]
.
Desp
ite
th
ese
s
ig
n
if
ican
t
d
ev
elo
p
m
en
ts
,
co
n
v
en
tio
n
al
VQ
A
s
y
s
tem
s
p
r
ed
o
m
in
an
tly
o
p
er
ate
u
n
d
e
r
th
e
r
estrictiv
e
ass
u
m
p
tio
n
o
f
s
in
g
le
-
an
s
wer
g
en
er
atio
n
,
wh
er
ein
ea
ch
v
is
u
al
q
u
esti
o
n
is
m
ap
p
ed
to
p
r
ec
is
ely
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
:
1
0
9
7
-
1
1
1
4
1098
o
n
e
co
r
r
e
ct
r
esp
o
n
s
e.
T
h
e
s
in
g
le
-
an
s
wer
p
ar
ad
i
g
m
p
r
o
v
es
f
u
n
d
am
en
tally
in
a
d
eq
u
ate
wh
en
f
ac
ed
with
am
b
ig
u
o
u
s
v
is
u
al
r
ea
liti
es
,
wh
er
e
a
s
in
g
le
im
ag
e
o
f
ten
h
o
ld
s
m
u
ltip
le
lay
er
s
o
f
tr
u
th
th
at
ca
n
n
o
t
b
e
s
u
m
m
ar
ized
in
a
s
in
g
le
lab
el
[
4
]
.
R
ec
en
t
r
esear
ch
s
h
o
ws
th
at
VQA
s
y
s
tem
s
s
h
o
u
ld
b
e
ab
le
to
p
r
o
v
id
e
m
u
ltip
le
co
r
r
ec
t
a
n
s
wer
s
f
o
r
m
u
ltip
le
o
b
jects
in
a
co
m
p
lex
v
is
u
al
s
ce
n
e,
am
b
ig
u
o
u
s
q
u
esti
o
n
s
,
an
d
/o
r
q
u
esti
o
n
s
h
av
in
g
s
u
b
je
ctiv
e
an
s
wer
s
[
5
]
.
T
h
e
em
er
g
en
ce
o
f
m
u
lt
i
-
an
s
wer
VQA
s
y
s
tem
s
ad
d
r
ess
es
s
ev
er
al
f
u
n
d
am
en
tal
lim
itatio
n
s
in
tr
ad
itio
n
al
ap
p
r
o
ac
h
es.
T
o
d
a
y
’
s
v
is
u
al
co
n
ten
t
is
co
m
p
o
s
ed
o
f
co
m
p
lex
s
ce
n
es
with
m
u
ltip
le
o
b
jects,
v
ar
y
in
g
attr
ib
u
tes,
an
d
co
m
p
le
x
r
elatio
n
s
h
ip
s
b
etwe
en
th
e
m
,
all
o
f
w
h
ich
m
ay
cr
ea
te
s
itu
at
io
n
s
w
h
er
e
m
u
ltip
le
c
o
r
r
ec
t
an
s
wer
s
m
ay
v
alid
ly
ap
p
ly
to
a
s
in
g
le
q
u
esti
o
n
.
Fo
r
ex
am
p
l
e,
if
an
im
ag
e
co
n
tain
s
m
u
ltip
le
s
h
ad
es
o
f
b
an
an
a
co
lo
r
,
s
ev
er
al
r
esp
o
n
s
es
co
u
l
d
in
d
icate
wh
at
co
lo
r
b
an
a
n
a
is
p
r
esen
t
in
th
e
im
ag
e.
Ad
d
itio
n
ally
,
q
u
esti
o
n
s
r
eg
ar
d
in
g
h
u
m
an
em
o
tio
n
s
o
r
h
u
m
an
ac
tiv
ity
c
o
u
ld
h
av
e
d
if
f
er
en
t
in
ter
p
r
etatio
n
s
o
f
w
h
at
is
o
cc
u
r
r
in
g
b
ased
o
n
a
u
s
er
’
s
cu
ltu
r
e,
th
eir
s
o
cial
co
n
tex
t,
an
d
th
eir
in
d
iv
id
u
al
ex
p
er
ien
ce
s
[
6
]
.
As
VQA
s
y
s
tem
s
b
ec
o
m
e
in
cr
ea
s
in
g
ly
in
ter
ac
tiv
e,
co
m
p
r
eh
en
s
iv
e
a
n
s
wer
s
co
v
er
in
g
m
u
ltip
le
v
alid
in
ter
p
r
etatio
n
s
o
f
v
is
u
al
co
n
te
n
t
b
ec
o
m
e
ess
en
tial.
Ob
ject
d
etec
tio
n
is
an
ess
en
tial
tech
n
o
lo
g
y
f
o
r
th
e
VQA
s
y
s
tem
to
g
en
er
ate
m
u
ltip
le
an
s
wer
s
b
ec
au
s
e
it
h
elp
s
cr
ea
te
a
b
as
e
v
is
u
al
u
n
d
er
s
tan
d
in
g
o
f
th
e
s
ce
n
e
an
d
id
en
tif
y
,
lo
ca
te,
a
n
d
d
e
f
in
e
m
u
ltip
le
en
titi
es
o
r
o
b
jects
in
a
c
o
m
p
lex
e
n
v
ir
o
n
m
en
t.
W
h
en
u
s
ed
with
o
b
ject
-
b
ased
r
ep
r
esen
tatio
n
m
o
d
els,
VQA
s
y
s
tem
s
ca
n
s
y
s
tem
atic
ally
an
aly
ze
th
e
v
is
u
al
co
n
ten
t,
co
r
r
elate
th
e
id
en
tifie
d
o
b
jects,
an
d
p
r
o
d
u
c
e
m
u
ltip
le
an
s
wer
s
f
o
r
a
g
iv
e
n
q
u
esti
o
n
b
ased
o
n
r
ic
h
v
is
u
al
d
ata
[
7
]
.
W
ith
o
u
t
p
r
ec
is
e
o
b
je
ct
d
etec
tio
n
,
a
VQA
s
y
s
tem
o
p
er
ates
with
o
u
t
ex
p
licit
v
is
u
al
g
r
o
u
n
d
in
g
.
Ob
ject
d
etec
tio
n
in
teg
r
atio
n
s
er
v
es
as
a
v
is
u
al
an
ch
o
r
th
at
allo
ws
th
e
s
y
s
tem
to
g
r
an
u
lar
l
y
d
is
s
ec
t
ea
ch
en
tity
,
h
elp
clar
if
y
am
b
ig
u
o
u
s
q
u
esti
o
n
s
,
an
d
en
ab
le
th
e
cr
ea
tio
n
o
f
a
co
m
p
lete
s
et
o
f
an
s
wer
s
t
h
at
f
u
lly
r
e
p
r
esen
t th
e
c
o
m
p
le
x
ity
o
f
v
is
u
al
co
n
ten
t.
Fro
m
an
in
f
o
r
m
atio
n
an
d
co
m
m
u
n
icatio
n
tec
h
n
o
lo
g
y
(
I
C
T
)
p
er
s
p
ec
tiv
e,
m
u
lti
-
an
s
wer
VQA
s
y
s
tem
s
in
tr
o
d
u
ce
s
ev
e
r
al
co
m
p
u
tatio
n
al
an
d
co
m
m
u
n
icatio
n
ch
all
en
g
es.
Mo
d
er
n
VQA
p
ip
elin
es
ty
p
ically
r
ely
o
n
clo
u
d
-
b
ased
p
r
o
ce
s
s
in
g
o
f
h
ig
h
-
r
eso
lu
tio
n
im
a
g
es
co
m
b
in
ed
with
m
u
ltimo
d
al
d
ee
p
lear
n
in
g
m
o
d
els,
wh
ich
r
eq
u
i
r
e
s
u
b
s
tan
tial
co
m
p
u
t
atio
n
al
p
o
wer
an
d
d
ata
t
r
an
s
m
is
s
io
n
.
I
n
r
ea
l
-
wo
r
ld
d
e
p
lo
y
m
en
t
s
ce
n
a
r
io
s
,
ed
g
e
d
ev
ices
m
u
s
t
tr
an
s
m
it
v
is
u
al
d
ata
to
a
ce
n
tr
al
s
er
v
er
f
o
r
in
f
er
en
ce
,
r
esu
ltin
g
in
s
ig
n
if
ican
t
laten
cy
an
d
b
an
d
wid
th
ch
allen
g
es
f
o
r
tim
e
-
cr
itical
ap
p
licatio
n
s
s
u
ch
a
s
m
ed
ical
d
iag
n
o
s
tic
s
y
s
tem
s
,
s
m
ar
t
ag
r
ic
u
ltu
r
e
m
o
n
ito
r
in
g
,
a
n
d
i
n
tellig
en
t
s
u
r
v
eillan
ce
s
y
s
tem
s
.
T
h
er
e
f
o
r
e,
d
ev
elo
p
in
g
VQA
ar
ch
itectu
r
e
s
th
at
o
p
tim
ize
t
h
e
b
alan
ce
b
etwe
en
in
f
er
en
c
e
ac
cu
r
ac
y
an
d
c
o
m
p
u
tatio
n
al
o
r
co
m
m
u
n
icatio
n
o
v
er
h
ea
d
is
cr
itical
f
o
r
p
r
ac
tical
I
C
T
d
ep
lo
y
m
e
n
ts
[
8
]
,
[
9
]
.
His
to
r
ically
,
VQA
r
esear
ch
h
a
s
p
r
o
g
r
ess
ed
r
ap
i
d
ly
d
u
e
to
a
d
v
an
ce
s
in
d
ee
p
lear
n
i
n
g
an
d
m
u
ltimo
d
al
lear
n
in
g
.
E
ar
ly
s
tu
d
ies
p
r
im
ar
ily
f
o
cu
s
ed
o
n
s
in
g
le
-
an
s
wer
p
r
ed
ictio
n
,
wh
er
e
ea
c
h
im
ag
e
–
q
u
esti
o
n
p
air
was
ass
u
m
ed
to
h
a
v
e
o
n
e
co
r
r
ec
t
r
esp
o
n
s
e.
T
h
e
f
o
u
n
d
atio
n
al
VQA
d
ataset
was
in
tr
o
d
u
ce
d
alo
n
g
s
id
e
C
NN
-
L
STM
m
o
d
els
as
th
e
b
aselin
e
f
r
a
m
ewo
r
k
f
o
r
jo
in
tly
p
r
o
ce
s
s
in
g
v
is
u
al
an
d
tex
t
u
al
in
f
o
r
m
at
io
n
[
1
]
.
Ho
wev
er
,
r
esear
ch
er
s
s
o
o
n
r
ec
o
g
n
ized
t
h
at
ex
p
ec
tin
g
o
n
e
co
r
r
ec
t
an
s
wer
to
a
v
is
u
al
q
u
esti
o
n
was
o
f
ten
u
n
r
ea
lis
tic,
as
h
u
m
an
an
n
o
tato
r
s
co
u
ld
p
r
o
v
id
e
d
if
f
er
en
t
y
et
e
q
u
ally
v
a
lid
r
esp
o
n
s
es
to
th
e
s
am
e
q
u
esti
o
n
.
I
t
h
as
b
ee
n
d
em
o
n
s
tr
ated
th
a
t
d
is
ag
r
ee
m
e
n
t
am
o
n
g
h
u
m
an
an
n
o
tato
r
s
is
f
r
eq
u
en
t
in
VQA
an
n
o
tatio
n
s
,
co
n
f
ir
m
in
g
th
at
m
u
ltip
le
v
alid
an
s
wer
s
m
ay
e
x
is
t
an
d
th
at
m
o
d
els
m
u
s
t
b
e
ab
le
to
h
a
n
d
le
th
is
v
ar
iab
ilit
y
[
6
]
.
Fu
r
th
er
m
o
r
e,
th
r
ee
p
r
im
ar
y
f
ac
to
r
s
co
n
tr
ib
u
tin
g
to
an
s
wer
m
u
ltip
licity
h
av
e
b
ee
n
id
e
n
tifie
d
:
v
is
u
al
co
m
p
lex
ity
,
s
em
an
tic
am
b
ig
u
ity
,
a
n
d
s
u
b
jectiv
e
i
n
te
r
p
r
etatio
n
[
4
]
.
Su
b
s
eq
u
en
t
ar
ch
itectu
r
al
in
n
o
v
atio
n
s
h
av
e
s
u
b
s
tan
tially
im
p
r
o
v
ed
v
is
u
al
r
ea
s
o
n
i
n
g
in
VQ
A
s
y
s
tem
s
.
On
e
o
f
th
e
f
ir
s
t
ex
p
licit
f
r
a
m
ewo
r
k
s
f
o
r
m
u
lti
-
an
s
wer
VQ
A
was
p
r
o
p
o
s
ed
b
y
in
t
r
o
d
u
cin
g
m
ec
h
a
n
is
m
s
to
g
en
er
ate
m
o
r
e
th
a
n
o
n
e
an
s
we
r
f
o
r
a
s
in
g
le
v
is
u
al
q
u
esti
o
n
[
8
]
.
I
n
p
a
r
allel,
o
b
ject
d
etec
tio
n
h
as e
m
er
g
ed
as a
n
ess
en
tial
tech
n
o
lo
g
y
f
o
r
VQA.
I
t
h
as
b
ee
n
d
em
o
n
s
tr
ated
th
a
t
ap
p
ly
i
n
g
Fas
ter
R
-
C
NN
f
o
r
o
b
ject
-
b
ased
v
is
u
al
r
ep
r
esen
tatio
n
s
ig
n
if
ican
tly
im
p
r
o
v
es
v
is
u
al
g
r
o
u
n
d
in
g
a
n
d
r
ea
s
o
n
in
g
p
e
r
f
o
r
m
an
ce
[
1
0
]
.
R
ec
en
t
lar
g
e
v
is
io
n
–
lan
g
u
a
g
e
m
o
d
els,
in
cl
u
d
in
g
B
L
I
P
-
2
[
1
1
]
an
d
Flam
in
g
o
[
1
2
]
,
h
av
e
f
u
r
th
e
r
ad
v
an
ce
d
m
u
ltimo
d
a
l
r
ea
s
o
n
in
g
ca
p
ab
ilit
ies
b
y
i
n
teg
r
atin
g
lar
g
e
lan
g
u
ag
e
m
o
d
els
with
r
ich
v
is
u
al
r
ep
r
esen
tatio
n
s
.
Do
m
ain
-
s
p
ec
if
ic
VQA
ap
p
licatio
n
s
(
in
clu
d
in
g
m
ed
ical
im
ag
in
g
,
a
g
r
icu
ltu
r
e,
an
d
r
em
o
te
s
en
s
in
g
)
h
av
e
f
u
r
t
h
er
em
p
h
asized
th
e
n
ec
ess
ity
o
f
m
u
lti
-
an
s
wer
r
ea
s
o
n
in
g
,
as
q
u
esti
o
n
s
in
th
ese
d
o
m
ain
s
o
f
ten
r
eq
u
ir
e
m
u
ltip
le
attr
ib
u
tes
o
r
o
b
s
er
v
atio
n
s
to
b
e
d
escr
ib
ed
s
im
u
ltan
eo
u
s
ly
[
1
3
]
,
[
1
4
]
.
Desp
ite
s
ig
n
if
ican
t
p
r
o
g
r
ess
a
cr
o
s
s
th
ese
d
ir
ec
tio
n
s
,
th
e
lite
r
atu
r
e
s
till
lack
s
a
s
y
s
tem
atic
s
y
n
th
esi
s
s
p
ec
if
ically
f
o
cu
s
ed
o
n
m
u
lti
-
an
s
wer
VQA
s
y
s
tem
s
o
r
th
e
r
o
le
o
f
o
b
ject
-
lev
el
v
is
u
al
r
ep
r
esen
tatio
n
s
in
s
u
p
p
o
r
tin
g
a
n
s
wer
m
u
ltip
licity
.
Mo
s
t
ex
is
tin
g
s
u
r
v
ey
s
an
al
y
ze
VQA
ar
ch
itectu
r
es
f
r
o
m
th
e
p
er
s
p
ec
tiv
e
o
f
m
o
d
el
ev
o
lu
tio
n
o
r
m
u
ltimo
d
al
f
u
s
io
n
s
tr
ateg
ies,
in
clu
d
in
g
tr
an
s
f
o
r
m
er
-
b
ased
m
o
d
els
[
1
5
]
,
g
r
ap
h
r
ea
s
o
n
in
g
f
r
am
ewo
r
k
s
[
1
6
]
,
[
1
7
]
,
an
d
d
o
m
ain
-
s
p
ec
if
ic
im
p
lem
en
tati
o
n
s
[
1
8
]
–
[
2
0
]
,
with
o
u
t
ex
am
in
in
g
h
o
w
cu
r
r
en
t
s
y
s
tem
s
h
an
d
le
s
itu
atio
n
s
in
wh
ich
m
u
ltip
le
v
alid
an
s
wer
s
m
ay
ex
is
t.
T
h
is
g
ap
m
o
tiv
ates
th
e
p
r
esen
t
s
y
s
tem
atic
r
ev
iew.
T
o
th
e
b
est
o
f
o
u
r
k
n
o
wled
g
e,
th
is
r
e
v
iew
is
o
n
e
o
f
th
e
f
ir
s
t
to
s
y
s
tem
atica
lly
an
aly
ze
t
h
e
r
elatio
n
s
h
ip
b
etwe
en
o
b
ject
-
le
v
el
v
is
u
al
g
r
o
u
n
d
i
n
g
an
d
an
s
w
er
m
u
ltip
licity
in
m
u
lti
-
an
s
wer
VQA
s
y
s
tem
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
E
xp
lo
r
in
g
mu
lti
-
a
n
s
w
er v
is
u
a
l q
u
esti
o
n
a
n
s
w
erin
g
w
ith
o
b
ject
d
etec
tio
n
:
a
s
ystema
tic
… (
N
id
a
u
l H
a
s
a
n
a
ti
)
1099
T
h
e
m
ain
r
esear
c
h
q
u
esti
o
n
s
g
u
id
in
g
th
is
s
y
s
tem
atic
r
ev
iew
ar
e
as
f
o
llo
ws.
R
Q1
:
W
h
at
ex
is
tin
g
m
eth
o
d
o
l
o
g
ical
ap
p
r
o
ac
h
es
en
ab
le
m
u
ltip
le
an
s
wer
g
e
n
er
atio
n
in
co
n
tem
p
o
r
ar
y
V
QA
s
y
s
tem
s
?
R
Q2
:
Ho
w
d
o
es
o
b
ject
d
etec
tio
n
co
n
tr
ib
u
te
to
r
ea
s
o
n
in
g
ca
p
ab
ilit
y
an
d
a
n
s
wer
m
u
ltip
licity
in
VQA
m
o
d
els?
R
Q3
:
W
h
at
ar
e
th
e
d
ataset
s
a
n
d
ev
alu
atio
n
m
etr
ics
co
m
m
o
n
ly
u
s
ed
in
s
tu
d
ies
th
at
co
m
b
in
e
o
b
ject
d
etec
tio
n
an
d
m
u
lti
-
an
s
wer
VQA?
R
Q4
:
W
h
at
ar
e
th
e
p
r
in
cip
al
lim
itatio
n
s
,
p
er
s
is
ten
t
ch
allen
g
es,
an
d
p
r
o
m
is
in
g
f
u
tu
r
e
r
esear
ch
d
ir
ec
tio
n
s
id
e
n
tifie
d
i
n
th
e
liter
atu
r
e
r
eg
ar
d
i
n
g
o
b
je
ct
-
b
ased
m
u
lti
-
an
s
wer
VQA
s
y
s
tem
s
?
T
h
e
co
n
ce
p
t
o
f
m
u
lti
-
an
s
wer
VQA
r
em
ain
s
in
co
n
s
is
ten
tly
d
ef
in
e
d
ac
r
o
s
s
s
tu
d
ies.
T
ab
le
1
s
u
m
m
ar
izes
co
m
m
o
n
f
o
r
m
u
la
tio
n
s
o
f
s
in
g
le
-
an
s
wer
a
n
d
m
u
lti
-
an
s
wer
VQA
task
s
,
wh
ich
d
if
f
er
in
h
o
w
th
e
m
o
d
el
r
ep
r
esen
ts
th
e
an
s
wer
s
p
ac
e
an
d
h
an
d
l
es
m
u
ltip
le
v
alid
an
s
wer
s
.
L
et
a
n
im
a
g
e
b
e
d
en
o
ted
as
I
,
a
q
u
esti
o
n
as Q,
an
a
n
s
wer
as a
,
an
d
a
n
an
s
wer
s
p
ac
e
as
.
T
ab
le
1
.
Fo
r
m
al
f
o
r
m
u
latio
n
s
o
f
VQA
task
s
r
elate
d
to
m
u
lti
-
an
s
wer
r
ea
s
o
n
in
g
F
o
r
mu
l
a
t
i
o
n
M
a
t
h
e
ma
t
i
c
a
l
e
x
p
r
e
ssi
o
n
D
e
scri
p
t
i
o
n
S
i
n
g
l
e
-
a
n
sw
e
r
VQA
:
(
,
)
→
,
w
h
e
r
e
∈
Ea
c
h
i
ma
g
e
–
q
u
e
st
i
o
n
p
a
i
r
i
s
ma
p
p
e
d
t
o
o
n
e
d
o
mi
n
a
n
t
a
n
sw
e
r
(
S
o
f
t
M
ax
o
u
t
p
u
t
l
a
y
e
r
s
)
.
M
u
l
t
i
-
l
a
b
e
l
V
Q
A
:
(
,
)
→
{
1
,
2
,
.
.
.
,
}
M
u
l
t
i
p
l
e
a
n
sw
e
r
s
p
r
e
d
i
c
t
e
d
si
m
u
l
t
a
n
e
o
u
s
l
y
u
si
n
g
m
u
l
t
i
-
l
a
b
e
l
c
l
a
ss
i
f
i
c
a
t
i
o
n
(s
i
g
mo
i
d
o
u
t
p
u
t
l
a
y
e
r
s)
.
G
e
n
e
r
a
t
i
v
e
m
u
l
t
i
-
a
n
sw
e
r
:
(
,
)
→
seq
u
e
n
c
e
(
1
,
2
,
.
.
.
,
)
Th
e
mo
d
e
l
g
e
n
e
r
a
t
e
s a
se
q
u
e
n
c
e
o
f
p
l
a
u
s
i
b
l
e
a
n
sw
e
r
s
b
a
s
e
d
o
n
c
o
n
d
i
t
i
o
n
a
l
p
r
o
b
a
b
i
l
i
t
y
d
i
s
t
r
i
b
u
t
i
o
n
s.
To
p
-
k
c
a
n
d
i
d
a
t
e
r
a
n
k
i
n
g
:
(
,
)
→
To
p
-
(
)
Th
e
mo
d
e
l
r
a
n
k
s
c
a
n
d
i
d
a
t
e
a
n
sw
e
r
s a
n
d
r
e
t
u
r
n
s se
v
e
r
a
l
t
o
p
p
r
e
d
i
c
t
i
o
n
s fr
o
m
a
p
r
e
d
e
f
i
n
e
d
a
n
sw
e
r
sp
a
c
e
(
m
u
l
t
i
p
l
e
-
c
h
o
i
c
e
VQA).
I
mp
l
i
c
i
t
m
u
l
t
i
-
a
n
sw
e
r
r
e
a
s
o
n
i
n
g
:
(
,
,
1
,
…
,
)
→
M
o
d
e
l
r
e
a
s
o
n
s
o
v
e
r
mu
l
t
i
p
l
e
d
e
t
e
c
t
e
d
o
b
j
e
c
t
s
b
u
t
o
u
t
p
u
t
s
a
s
i
n
g
l
e
a
n
sw
e
r
(
o
b
j
e
c
t
-
b
a
s
e
d
VQA).
B
ased
o
n
th
ese
f
o
r
m
u
latio
n
s
,
th
is
r
ev
iew
r
eg
ar
d
s
a
s
tu
d
y
as
s
u
p
p
o
r
tin
g
m
u
lti
-
an
s
wer
VQA
wh
en
at
least
o
n
e
o
f
th
e
f
o
llo
win
g
c
o
n
d
itio
n
s
is
s
atis
f
ied
:
(
i
)
th
e
m
o
d
el
ex
p
licitly
p
r
e
d
icts
m
u
ltip
le
an
s
wer
s
,
s
u
ch
as
th
r
o
u
g
h
m
u
lti
-
lab
el
o
r
an
s
wer
-
s
et
o
u
tp
u
ts
;
(
i
)
th
e
m
o
d
el
g
en
er
ates
s
ev
er
al
s
em
an
tically
d
i
s
tin
ct
r
esp
o
n
s
es
f
o
r
th
e
s
am
e
im
ag
e
–
q
u
esti
o
n
p
air
;
o
r
(
iii
)
th
e
m
o
d
el
em
p
lo
y
s
o
b
ject
-
lev
el
o
r
r
eg
io
n
-
b
ased
r
ea
s
o
n
in
g
m
ec
h
a
n
is
m
s
th
at
en
ab
le
in
ter
p
r
etatio
n
o
f
m
u
ltip
le
v
is
u
al
en
titi
es.
T
h
e
m
ain
g
o
als
o
f
th
is
s
y
s
t
em
atic
r
ev
iew
ar
e
to
:
(
i
)
p
r
o
v
id
e
a
n
o
r
g
an
ize
d
s
u
m
m
a
r
y
o
f
r
ec
en
t
ad
v
an
ce
s
in
m
u
lti
-
an
s
wer
VQA
s
y
s
tem
s
in
co
r
p
o
r
atin
g
o
b
ject
d
etec
tio
n
;
(
ii
)
ca
teg
o
r
ize
an
d
co
m
p
a
r
e
s
tate
-
of
-
th
e
-
ar
t
d
ee
p
lear
n
in
g
m
eth
o
d
s
b
ased
o
n
ar
c
h
itectu
r
al
d
esig
n
an
d
m
u
ltimo
d
al
f
u
s
io
n
s
tr
ateg
ies;
(
iii
)
o
v
er
v
iew
co
m
m
o
n
ly
u
s
ed
d
atasets
an
d
ev
alu
atio
n
m
etr
ics
r
elev
an
t
to
m
u
lti
-
an
s
wer
o
r
o
b
ject
-
awa
r
e
VQA
s
y
s
tem
s
;
an
d
(
iv
)
id
en
tif
y
k
e
y
ch
allen
g
es
an
d
f
u
t
u
r
e
r
esea
r
ch
d
ir
ec
tio
n
s
f
o
r
m
o
r
e
r
o
b
u
s
t
an
d
in
ter
p
r
etab
le
m
u
lti
-
an
s
wer
VQA
m
o
d
els.
T
o
s
u
p
p
o
r
t
th
ese
g
o
als,
th
is
r
ev
iew
f
u
r
th
er
c
o
n
tr
i
b
u
tes
a
s
tr
u
ctu
r
ed
f
o
u
r
-
d
im
en
s
io
n
al
ta
x
o
n
o
m
y
o
f
m
u
lti
-
an
s
wer
VQA,
o
r
g
a
n
iz
in
g
ex
is
tin
g
ap
p
r
o
ac
h
es
alo
n
g
task
f
o
r
m
u
latio
n
,
v
is
u
al
r
ep
r
esen
tatio
n
s
tr
ateg
y
,
s
o
u
r
ce
s
o
f
a
n
s
wer
m
u
lt
ip
licity
,
an
d
ev
alu
atio
n
s
tr
ateg
y
d
im
en
s
io
n
s
,
with
p
ar
ticu
lar
em
p
h
asis
o
n
h
o
w
o
b
ject
-
lev
el
v
is
u
al
g
r
o
u
n
d
i
n
g
en
ab
les
an
s
wer
m
u
ltip
licity
.
T
h
e
r
em
ain
d
e
r
o
f
th
is
p
ap
er
is
o
r
g
an
ize
d
as
f
o
llo
ws.
Sectio
n
2
d
escr
ib
es
th
e
r
esear
ch
m
eth
o
d
o
lo
g
y
.
Sect
io
n
3
p
r
esen
ts
th
e
r
esu
lts
an
d
d
is
cu
s
s
io
n
.
Sectio
n
4
co
n
cl
u
d
es th
e
p
a
p
er
.
2.
RE
S
E
ARCH
M
E
T
H
O
D
T
h
is
s
y
s
tem
atic
r
ev
iew
em
p
lo
y
ed
th
e
Pre
f
er
r
ed
R
ep
o
r
ti
n
g
I
tem
s
f
o
r
Sy
s
tem
atic
R
ev
iews
an
d
Me
ta
-
An
aly
s
es
(
PR
I
SMA)
2
0
2
0
g
u
id
elin
es
t
o
en
s
u
r
e
a
co
m
p
r
eh
en
s
iv
e,
tr
an
s
p
ar
e
n
t,
an
d
r
ep
r
o
d
u
ci
b
le
in
v
esti
g
atio
n
[
2
1
]
.
T
h
e
co
m
p
leted
PR
I
SMA
2
0
2
0
ch
ec
k
lis
t is p
r
o
v
id
ed
in
Su
p
p
lem
en
tar
y
Ma
ter
ial,
T
ab
le
S1
.
2
.
1
.
Sea
rc
h
s
t
ra
t
eg
y
a
nd
info
rm
a
t
i
o
n r
eso
urce
s
T
h
e
s
ea
r
ch
s
tr
ateg
y
co
m
b
in
ed
th
r
ee
k
ey
wo
r
d
g
r
o
u
p
s
:
th
e
VQA
task
(
“
v
is
u
al
q
u
esti
o
n
an
s
wer
in
g
”
)
,
m
u
lti
-
an
s
wer
co
n
ce
p
ts
(
“
m
u
ltip
le
an
s
wer
”
,
“
m
u
lti
-
an
s
wer
”
,
“
m
u
lti
-
lab
el
”
)
,
a
n
d
o
b
ject
-
lev
el
v
is
u
al
r
ep
r
esen
tatio
n
s
(
“
o
b
ject
d
etec
tio
n
”
,
“
r
eg
io
n
”
,
“
b
o
u
n
d
in
g
b
o
x
”
)
.
T
h
ese
ter
m
s
wer
e
s
elec
ted
to
ca
p
tu
r
e
b
o
th
ex
p
licit
m
u
lti
-
an
s
wer
ap
p
r
o
ac
h
es
an
d
s
tu
d
ies
th
at
im
p
licitly
s
u
p
p
o
r
t
an
s
wer
v
ar
iab
ilit
y
t
h
r
o
u
g
h
o
b
ject
-
lev
el
r
ea
s
o
n
in
g
.
T
h
e
k
e
y
wo
r
d
c
ateg
o
r
ies an
d
B
o
o
lean
s
ea
r
c
h
s
tr
in
g
s
u
s
ed
in
th
e
q
u
er
y
ar
e
s
u
m
m
ar
ized
in
T
ab
le
2
.
T
h
e
s
ea
r
ch
was
co
n
d
u
cted
in
J
an
u
ar
y
2
0
2
6
u
s
in
g
th
e
Sco
p
u
s
d
atab
ase,
wh
ich
was
s
elec
ted
f
o
r
its
ex
ten
s
iv
e
co
v
er
a
g
e
o
f
p
ee
r
-
r
ev
iewe
d
jo
u
r
n
als
in
th
e
f
ield
s
o
f
co
m
p
u
te
r
v
is
io
n
an
d
ar
t
if
icial
in
tellig
en
ce
,
in
clu
d
in
g
p
u
b
licatio
n
s
f
r
o
m
I
E
E
E
,
AC
M,
E
ls
ev
ier
,
a
n
d
Sp
r
in
g
er
.
Fil
ter
s
wer
e
ap
p
lied
to
lim
it
r
esu
lts
to
E
n
g
lis
h
-
lan
g
u
a
g
e
jo
u
r
n
al
ar
tic
les p
u
b
lis
h
ed
b
etwe
en
J
an
u
a
r
y
2
0
2
0
an
d
Dec
e
m
b
er
2
0
2
5
.
T
ec
h
n
ical
ar
ch
itectu
r
al
ter
m
s
s
u
ch
as
“
o
b
ject
d
etec
tio
n
,
”
“
r
eg
io
n
,
”
an
d
“
b
o
u
n
d
in
g
b
o
x
”
wer
e
p
r
io
r
itized
o
v
e
r
ab
s
tr
ac
t
s
em
an
tic
ter
m
s
to
en
h
an
ce
r
ec
all
o
f
s
tu
d
ies
th
at
im
p
licit
ly
ad
d
r
ess
an
s
wer
v
ar
iab
ilit
y
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
:
1
0
9
7
-
1
1
1
4
1100
th
r
o
u
g
h
o
b
ject
-
lev
el
r
ea
s
o
n
in
g
with
o
u
t
e
x
p
licitly
u
s
in
g
am
b
ig
u
ity
ter
m
i
n
o
lo
g
y
.
T
h
is
m
eth
o
d
o
lo
g
ical
d
ec
is
io
n
is
b
ec
au
s
e
in
th
e
VQA
d
o
m
ain
,
q
u
esti
o
n
am
b
ig
u
ity
is
o
f
ten
co
n
s
id
er
e
d
a
co
n
ce
p
tu
al
p
h
en
o
m
en
o
n
,
wh
ile
tech
n
ical
s
o
lu
tio
n
s
ar
e
im
p
lem
en
ted
th
r
o
u
g
h
v
is
u
aliza
tio
n
an
d
en
tity
lo
ca
lizatio
n
m
ec
h
an
is
m
s
.
T
ab
le
2
.
Key
wo
r
d
ca
teg
o
r
ies a
n
d
s
ea
r
ch
ter
m
s
,
a
n
d
B
o
o
lean
s
ea
r
ch
s
tr
in
g
s
K
e
y
w
o
r
d
c
a
t
e
g
o
r
y
S
e
a
r
c
h
t
e
r
ms
R
a
t
i
o
n
a
l
e
C
o
r
e
t
a
sk
“
V
i
su
a
l
q
u
e
s
t
i
o
n
a
n
sw
e
r
i
n
g
”
U
sed
i
n
f
u
l
l
f
o
r
m
t
o
a
v
o
i
d
a
m
b
i
g
u
i
t
y
w
i
t
h
o
t
h
e
r
m
e
a
n
i
n
g
s
o
f
t
h
e
a
c
r
o
n
y
m
V
Q
A
,
su
c
h
a
s
v
i
d
e
o
q
u
a
l
i
t
y
a
ss
u
r
a
n
c
e
.
M
u
l
t
i
-
a
n
sw
e
r
c
o
n
c
e
p
t
“
M
u
l
t
i
p
l
e
a
n
sw
e
r
”
,
“
m
u
l
t
i
-
a
n
sw
e
r
”
,
“
mu
l
t
i
a
n
sw
e
r
”
,
“
mu
l
t
i
-
l
a
b
e
l
”
To
i
d
e
n
t
i
f
y
V
Q
A
s
t
u
d
i
e
s
t
h
a
t
a
l
l
o
w
mo
r
e
t
h
a
n
o
n
e
v
a
l
i
d
a
n
sw
e
r
f
o
r
a
si
n
g
l
e
i
ma
g
e
–
q
u
e
st
i
o
n
p
a
i
r
.
O
b
j
e
c
t
-
l
e
v
e
l
v
i
s
u
a
l
r
e
p
r
e
s
e
n
t
a
t
i
o
n
“
O
b
j
e
c
t
d
e
t
e
c
t
i
o
n
”
,
“
r
e
g
i
o
n
”
,
“
b
o
u
n
d
i
n
g
b
o
x
”
To
r
e
t
r
i
e
v
e
s
t
u
d
i
e
s
t
h
a
t
su
p
p
o
r
t
m
u
l
t
i
-
o
b
j
e
c
t
a
n
d
mu
l
t
i
-
a
n
sw
e
r
r
e
a
so
n
i
n
g
t
h
r
o
u
g
h
o
b
j
e
c
t
-
l
e
v
e
l
v
i
su
a
l
g
r
o
u
n
d
i
n
g
,
e
v
e
n
w
h
e
n
t
h
e
t
e
r
m
“
m
u
l
t
i
p
l
e
a
n
sw
e
r
”
i
s
n
o
t
e
x
p
l
i
c
i
t
l
y
u
s
e
d
.
S
c
o
p
u
s
TI
TLE
-
A
B
S
-
K
EY
(
“
v
i
su
a
l
q
u
e
s
t
i
o
n
a
n
sw
e
r
i
n
g
”
)
A
N
D
((
“
mu
l
t
i
p
l
e
a
n
sw
e
r
”
O
R
“
m
u
l
ti
-
a
n
sw
e
r
”
O
R
“
mu
l
t
i
a
n
sw
e
r
”
O
R
“
m
u
l
t
i
-
l
a
b
e
l
”
)
O
R
(
“
o
b
j
e
c
t
d
e
t
e
c
t
i
o
n
”
O
R
“
r
e
g
i
o
n
”
O
R
“
b
o
u
n
d
i
n
g
b
o
x
”
).
Jo
u
r
n
a
l
s,
En
g
l
i
s
h
,
2
0
2
0
-
2
0
2
5
2
.
2
.
E
lig
ibi
lity
cr
it
er
ia
Stu
d
ies
wer
e
in
clu
d
ed
i
f
th
ey
wer
e
p
u
b
lis
h
ed
in
a
p
ee
r
-
r
e
v
iewe
d
jo
u
r
n
al
in
d
ex
e
d
in
Sco
p
u
s
,
wr
itten
in
E
n
g
lis
h
,
p
u
b
lis
h
ed
b
etwe
en
J
an
u
ar
y
2
0
2
0
an
d
Dec
em
b
e
r
2
0
2
5
,
an
d
d
ir
ec
tly
a
d
d
r
ess
ed
VQA
task
s
wh
ile
eith
er
ex
p
licitly
en
d
o
r
s
in
g
m
u
lti
-
an
s
wer
f
o
r
m
u
latio
n
s
o
r
im
p
licitly
f
ac
ilit
atin
g
m
u
ltip
le
v
alid
an
s
wer
s
th
r
o
u
g
h
o
b
ject
-
b
ased
,
r
e
g
io
n
-
b
ased
,
o
r
m
u
lti
-
in
s
tan
ce
r
ea
s
o
n
in
g
m
ec
h
an
is
m
s
.
T
h
e
r
estrictio
n
to
p
e
er
-
r
ev
iewe
d
jo
u
r
n
al
ar
ticles
en
s
u
r
es
a
m
i
n
im
u
m
q
u
ality
th
r
esh
o
l
d
c
o
n
s
is
ten
t
with
estab
lis
h
ed
SLR
p
r
ac
tices
in
a
p
p
lied
c
o
m
p
u
te
r
s
cien
ce
[
2
2
]
.
Stu
d
ies
wer
e
ex
clu
d
ed
if
th
ey
lack
ed
ex
p
e
r
im
en
t
al
v
alid
atio
n
o
r
r
eliab
le
ev
alu
atio
n
p
r
o
ce
d
u
r
es,
d
id
n
o
t
f
o
cu
s
o
n
VQA
task
s
,
o
r
d
id
n
o
t
em
p
lo
y
o
b
ject
-
b
ased
o
r
r
e
g
io
n
-
b
ased
r
ea
s
o
n
in
g
to
s
u
p
p
o
r
t
m
u
ltip
le
v
alid
an
s
wer
s
.
Su
r
v
ey
p
ap
e
r
s
,
r
ev
iew
ar
ticles,
an
d
o
th
er
s
ec
o
n
d
ar
y
s
tu
d
ies
wer
e
also
ex
clu
d
ed
.
Fu
tu
r
e
r
ev
iews m
ay
co
n
s
id
er
in
clu
d
in
g
h
ig
h
-
im
p
ac
t c
o
n
f
er
en
ce
p
r
o
c
ee
d
in
g
s
to
ex
p
an
d
s
co
p
e.
2
.
3
.
Select
io
n
pro
ce
s
s
T
h
e
s
elec
tio
n
p
r
o
ce
s
s
f
o
llo
we
d
f
o
u
r
s
eq
u
e
n
tial
PR
I
SMA
p
h
ases
:
(
i
)
I
d
en
tific
atio
n
,
in
wh
i
ch
r
ec
o
r
d
s
wer
e
r
etr
iev
ed
f
r
o
m
Sco
p
u
s
u
s
in
g
th
e
p
r
ed
ef
in
e
d
s
ea
r
c
h
s
tr
in
g
;
(
ii
)
Scr
ee
n
in
g
,
in
wh
ich
two
au
th
o
r
s
in
d
ep
en
d
en
tly
s
cr
ee
n
ed
titl
e
s
an
d
a
b
s
tr
ac
ts
ag
ain
s
t
in
c
lu
s
io
n
an
d
ex
cl
u
s
io
n
cr
iter
i
a;
(
iii
)
E
lig
ib
ilit
y
ass
es
s
m
en
t,
in
wh
ich
f
u
ll
-
tex
t
ar
ticles
wer
e
ev
alu
ated
in
d
ep
en
d
e
n
tly
b
y
th
e
s
am
e
r
ev
iewe
r
s
,
with
d
is
ag
r
ee
m
en
ts
r
eso
lv
ed
th
r
o
u
g
h
d
is
cu
s
s
io
n
an
d
ad
d
itio
n
al
au
th
o
r
co
n
s
u
ltatio
n
wh
en
n
ee
d
e
d
;
an
d
(
iv
)
Fin
al
in
clu
s
io
n
,
in
wh
ich
s
tu
d
ies
s
ati
s
f
y
in
g
all
cr
iter
ia
wer
e
in
clu
d
ed
in
th
e
f
in
al
s
y
n
t
h
esis
.
T
h
e
co
m
p
lete
p
r
o
ce
s
s
is
s
u
m
m
ar
ized
in
th
e
PR
I
SMA
f
lo
w
d
iag
r
am
a
s
s
h
o
wn
in
Fig
u
r
e
1
.
Alth
o
u
g
h
a
f
o
r
m
al
i
n
ter
-
r
ater
r
eliab
ilit
y
co
ef
f
icien
t
(
e
.
g
.
,
C
o
h
en
’
s
Kap
p
a)
was
n
o
t
ca
lc
u
lated
p
r
io
r
to
c
o
n
s
en
s
u
s
r
eso
lu
tio
n
,
th
e
u
s
e
o
f
clea
r
ly
d
ef
in
ed
elig
ib
ilit
y
c
r
iter
ia
an
d
in
d
ep
en
d
en
t
m
u
l
ti
-
s
tag
e
s
cr
ee
n
in
g
was
in
te
n
d
ed
to
m
in
im
iz
e
s
u
b
jectiv
ity
in
t
h
e
s
elec
tio
n
p
r
o
ce
s
s
,
co
n
s
is
ten
t
with
estab
li
s
h
ed
SLR
p
r
ac
tices
in
ar
tific
i
al
in
tellig
en
ce
a
n
d
co
m
p
u
ter
v
is
io
n
r
esear
ch
[
2
2
]
.
2
.
4
.
Da
t
a
e
x
t
ra
ct
i
o
n a
nd
qu
a
lity
a
s
s
ess
m
ent
Data
ex
tr
ac
tio
n
ca
p
tu
r
ed
d
escr
ip
tiv
e
ch
ar
ac
ter
is
tics
(
ap
p
licatio
n
d
o
m
ai
n
,
r
esear
ch
o
b
jectiv
es,
d
ataset
u
tili
za
tio
n
)
,
m
eth
o
d
o
lo
g
ical
d
etails
(
ev
alu
atio
n
m
etr
ics,
o
b
j
ec
t
-
lev
el
v
is
u
al
r
ep
r
esen
tatio
n
s
tr
ateg
ies,
ex
p
licit
o
r
im
p
licit
m
u
lti
-
an
s
wer
s
u
p
p
o
r
t)
,
an
d
s
tu
d
y
-
lev
el
o
b
s
er
v
atio
n
s
(
co
n
tr
ib
u
tio
n
s
,
lim
itatio
n
s
,
f
u
tu
r
e
d
ir
ec
tio
n
s
)
.
Qu
ality
ass
ess
m
en
t
was
ad
a
p
ted
f
o
r
AI
-
b
ased
VQA
r
esear
ch
u
s
in
g
f
iv
e
c
r
iter
ia:
(
i
)
clar
ity
o
f
p
r
o
b
lem
f
o
r
m
u
latio
n
,
(
ii
)
tr
an
s
p
ar
en
c
y
o
f
d
ataset
d
escr
ip
tio
n
s
,
(
iii
)
ap
p
r
o
p
r
iaten
ess
o
f
o
b
ject
d
etec
tio
n
in
teg
r
atio
n
,
(
iv
)
clar
ity
o
f
ev
alu
a
tio
n
p
r
o
to
co
ls
,
an
d
(
v
)
co
m
p
leten
ess
o
f
r
ep
o
r
ted
r
esu
lts
.
E
ac
h
c
r
i
te
r
i
o
n
was
i
n
d
e
p
e
n
d
e
n
tl
y
r
at
ed
as
l
o
w,
m
o
d
er
at
e,
o
r
h
i
g
h
r
is
k
o
f
b
ias
b
y
tw
o
r
e
v
ie
wer
s
,
wi
th
d
is
a
g
r
e
em
en
ts
r
eso
lv
ed
th
r
o
u
g
h
co
n
s
en
s
u
s
.
Mo
s
t
s
t
u
d
ies
r
ec
e
iv
ed
lo
w
r
is
k
r
ati
n
g
s
ac
r
o
s
s
al
l
f
i
v
e
cr
ite
r
i
a,
p
a
r
ti
c
u
la
r
l
y
t
h
o
s
e
t
h
at
u
s
ed
well
-
k
n
o
w
n
p
u
b
lic
d
at
asets
a
n
d
r
e
p
o
r
t
ed
t
h
ei
r
ev
al
u
a
ti
o
n
m
et
h
o
d
s
cl
ea
r
l
y
an
d
co
m
p
let
el
y
.
St
u
d
ies
t
h
at
r
el
i
ed
o
n
c
u
s
t
o
m
d
at
asets
wit
h
o
u
t
p
u
b
l
ic
ac
c
ess
o
r
t
h
at
p
r
o
v
i
d
e
d
i
n
c
o
m
p
let
e
d
es
cr
ip
ti
o
n
s
o
f
t
h
eir
ev
al
u
at
i
o
n
p
r
o
ce
d
u
r
es
we
r
e
ass
i
g
n
e
d
m
o
d
er
ate
r
is
k
r
a
ti
n
g
s
.
Hi
g
h
r
i
s
k
o
f
b
ias
was
n
o
t
id
e
n
t
if
ie
d
in
a
n
y
o
f
t
h
e
i
n
c
lu
d
e
d
s
t
u
d
i
es,
as
al
l
h
a
d
u
n
d
e
r
g
o
n
e
p
ee
r
-
r
e
v
i
ew
v
ali
d
a
ti
o
n
p
r
io
r
t
o
i
n
cl
u
s
i
o
n
.
Qu
ali
ty
r
at
in
g
s
we
r
e
u
s
e
d
t
o
co
n
t
ex
tu
a
lize
f
i
n
d
in
g
s
d
u
r
i
n
g
n
a
r
r
at
iv
e
s
y
n
t
h
esis
r
at
h
e
r
t
h
a
n
as
g
r
o
u
n
d
s
f
o
r
ex
cl
u
s
i
o
n
.
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
E
xp
lo
r
in
g
mu
lti
-
a
n
s
w
er v
is
u
a
l q
u
esti
o
n
a
n
s
w
erin
g
w
ith
o
b
ject
d
etec
tio
n
:
a
s
ystema
tic
… (
N
id
a
u
l H
a
s
a
n
a
ti
)
1101
2
.
5
.
Sy
nthesis
m
et
ho
d
T
h
e
s
y
n
th
esis
f
o
llo
wed
a
n
ar
r
ativ
e
an
d
th
em
atic
ap
p
r
o
ac
h
in
ac
co
r
d
an
ce
with
th
e
PR
I
SMA
2
0
2
0
g
u
id
elin
es.
Nar
r
ativ
e
s
y
n
th
esis
was
u
s
ed
to
q
u
alitativ
ely
s
u
m
m
ar
ize
an
d
in
ter
p
r
et
f
in
d
in
g
s
f
r
o
m
th
e
i
n
cl
u
d
ed
s
tu
d
ies,
wh
ile
th
em
atic
s
y
n
th
esis
o
r
g
an
ize
d
th
e
s
tu
d
i
es
in
to
c
o
n
ce
p
tu
al
ca
teg
o
r
i
es
b
ased
o
n
k
ey
m
eth
o
d
o
l
o
g
ical
ch
ar
ac
ter
is
tics
r
elev
an
t to
m
u
lti
-
an
s
wer
VQA
r
esear
ch
.
Stu
d
ies
wer
e
g
r
o
u
p
ed
b
ased
o
n
m
eth
o
d
o
lo
g
ical
ch
a
r
ac
ter
i
s
tics
r
elev
an
t
to
th
e
r
esear
ch
o
b
jectiv
es,
in
clu
d
in
g
task
f
o
r
m
u
latio
n
f
o
r
m
u
lti
-
an
s
wer
VQA,
o
b
ject
-
l
ev
el
v
is
u
al
r
ep
r
esen
tatio
n
s
tr
a
teg
ies,
an
d
d
ataset
u
s
ag
e
an
d
ev
alu
atio
n
p
r
o
to
c
o
ls
.
T
h
ese
g
r
o
u
p
in
g
s
wer
e
d
ef
in
ed
p
r
io
r
to
th
e
s
y
n
th
esis
p
r
o
ce
s
s
to
en
s
u
r
e
co
n
s
is
ten
cy
ac
r
o
s
s
th
e
in
clu
d
e
d
s
tu
d
ies.
E
x
tr
ac
ted
d
ata
wer
e
s
u
b
s
eq
u
en
tly
s
tan
d
ar
d
ized
to
ad
d
r
ess
v
ar
iatio
n
s
in
ter
m
in
o
lo
g
y
an
d
r
e
p
o
r
tin
g
s
ty
les.
T
h
e
ch
a
r
ac
ter
is
tics
o
f
in
d
iv
i
d
u
al
s
tu
d
ies
wer
e
s
u
m
m
ar
iz
ed
in
th
e
s
tu
d
y
c
h
ar
ac
ter
is
tics
s
ec
tio
n
.
A
n
ar
r
ativ
e
s
y
n
th
esis
was
th
e
n
co
n
d
u
cted
to
i
d
en
tify
r
ec
u
r
r
in
g
m
eth
o
d
o
l
o
g
ical
p
atter
n
s
a
cr
o
s
s
th
e
r
e
v
iewe
d
s
tu
d
ies.
T
h
is
ap
p
r
o
ac
h
was
s
elec
ted
b
ec
au
s
e
s
u
b
s
tan
tial
h
eter
o
g
en
eity
in
task
d
ef
in
iti
o
n
s
,
d
atasets
,
an
d
ev
alu
atio
n
m
etr
ics p
r
ec
lu
d
ed
th
e
u
s
e
o
f
q
u
an
titativ
e
m
eta
-
a
n
aly
s
is
.
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
3
.
1
.
Study
s
elec
t
io
n pro
ce
s
s
T
h
e
d
atab
ase
s
ea
r
ch
id
en
tifie
d
6
0
0
r
ec
o
r
d
s
f
r
o
m
Sco
p
u
s
.
Af
ter
ap
p
ly
in
g
f
ilter
s
f
o
r
p
u
b
lica
tio
n
y
ea
r
,
lan
g
u
ag
e,
an
d
d
o
cu
m
e
n
t
ty
p
e,
4
0
6
r
ec
o
r
d
s
wer
e
r
em
o
v
e
d
,
leav
in
g
1
9
4
f
o
r
titl
e
a
n
d
ab
s
tr
ac
t
s
cr
ee
n
in
g
.
Of
th
ese,
1
3
5
wer
e
e
x
clu
d
e
d
b
ased
o
n
p
r
ed
ef
in
e
d
cr
iter
ia
,
r
etain
in
g
5
9
ar
ticles
f
o
r
f
u
l
l
-
tex
t
ass
es
s
m
en
t,
as
s
u
m
m
ar
ized
in
T
a
b
le
3
.
Fo
llo
win
g
f
u
ll
-
tex
t
ev
alu
atio
n
,
o
n
e
ad
d
itio
n
al
s
tu
d
y
was
ex
clu
d
ed
,
y
ield
in
g
a
f
i
n
al
co
r
p
u
s
o
f
5
8
s
tu
d
ies.
Am
o
n
g
th
ese,
1
0
ex
p
licitly
f
o
r
m
u
late
d
m
u
lti
-
an
s
wer
VQA
as
th
eir
p
r
im
ar
y
o
b
jectiv
e,
wh
ile
4
8
im
p
licitly
s
u
p
p
o
r
ted
m
u
lti
-
an
s
wer
r
ea
s
o
n
in
g
th
r
o
u
g
h
o
b
ject
-
b
ased
,
r
e
g
io
n
-
b
ased
,
o
r
m
u
lti
-
in
s
tan
ce
ap
p
r
o
ac
h
es,
as sh
o
wn
in
T
a
b
le
4
an
d
Fig
u
r
e
1
.
T
ab
le
3
.
R
ea
s
o
n
s
f
o
r
ex
clu
s
io
n
d
u
r
i
n
g
ab
s
tr
ac
t scr
ee
n
in
g
R
e
a
s
o
n
f
o
r
e
x
c
l
u
s
i
o
n
N
u
mb
e
r
o
f
p
a
p
e
r
s
D
u
p
l
i
c
a
t
e
r
e
c
o
r
d
s
i
d
e
n
t
i
f
i
e
d
d
u
r
i
n
g
s
c
r
e
e
n
i
n
g
4
S
u
r
v
e
y
/
r
e
v
i
e
w
a
r
t
i
c
l
e
s
4
I
n
su
f
f
i
c
i
e
n
t
v
a
l
i
d
a
t
i
o
n
o
r
V
Q
A
n
o
t
p
r
i
mary
f
o
c
u
s
2
N
o
t
V
Q
A
t
a
s
k
34
V
Q
A
h
a
s
g
o
o
d
o
b
j
e
c
t
d
e
t
e
c
t
i
o
n
b
u
t
a
si
n
g
l
e
a
n
sw
e
r
o
n
l
y
62
S
i
n
g
l
e
-
a
n
sw
e
r
V
Q
A
w
i
t
h
o
u
t
o
b
j
e
c
t
-
b
a
se
d
o
r
mu
l
t
i
-
i
n
st
a
n
c
e
r
e
a
s
o
n
i
n
g
29
To
t
a
l
e
x
c
l
u
d
e
d
a
b
s
t
r
a
c
t
1
3
5
T
ab
le
4
.
Fin
al
s
tu
d
y
class
if
icat
io
n
S
c
r
e
e
n
i
n
g
o
u
t
c
o
me
N
u
mb
e
r
o
f
s
t
u
d
i
e
s
Ex
p
l
i
c
i
t
m
u
l
t
i
-
a
n
sw
e
r
10
I
mp
l
i
c
i
t
m
u
l
t
i
-
a
n
sw
e
r
(
o
b
j
e
c
t
-
b
a
se
d
,
r
e
g
i
o
n
-
b
a
se
d
,
o
r
m
u
l
t
i
-
i
n
st
a
n
c
e
r
e
a
s
o
n
i
n
g
)
4
8
S
t
u
d
i
e
s t
h
a
t
m
e
t
a
l
l
e
l
i
g
i
b
i
l
i
t
y
c
r
i
t
e
r
i
a
5
8
Fig
u
r
e
1
.
P
R
I
SMA
f
lo
w
d
iag
r
am
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
:
1
0
9
7
-
1
1
1
4
1102
3
.
2
.
Study
cha
ra
c
t
er
is
t
ics
T
h
is
s
ec
tio
n
d
escr
ib
es
th
e
m
ain
ch
ar
a
cter
is
tics
o
f
th
e
5
8
in
clu
d
e
d
s
tu
d
ies
ac
r
o
s
s
f
o
u
r
k
ey
d
im
en
s
io
n
s
:
1
)
g
en
er
al
-
p
u
r
p
o
s
e
v
er
s
u
s
d
o
m
ain
-
s
p
ec
if
i
c
ca
teg
o
r
izatio
n
,
2
)
ex
p
lic
it
v
er
s
u
s
im
p
licit
m
u
lti
-
an
s
wer
f
o
r
m
u
latio
n
,
3
)
d
ataset
u
tili
za
tio
n
,
an
d
4
)
o
b
je
ct
-
ba
s
ed
v
is
u
al
r
e
p
r
esen
tatio
n
s
tr
ateg
ies.
Deta
iled
ch
ar
ac
ter
is
tics
f
o
r
all
5
8
in
cl
u
d
ed
s
tu
d
ies
ar
e
p
r
o
v
id
ed
i
n
T
ab
le
S2
(
n
=2
8
d
o
m
ain
-
s
p
e
cif
ic
s
tu
d
ies)
an
d
T
ab
le
S3
(
n
=
3
0
g
e
n
er
al
-
p
u
r
p
o
s
e
s
tu
d
ies)
in
Su
p
p
lem
en
tar
y
Ma
ter
ial
.
3
.
2
.
1
.
G
ener
a
l
-
pu
rpo
s
e
v
s
.
do
m
a
in
-
sp
ec
if
ic
VQ
A
Fro
m
th
e
5
8
in
cl
u
d
ed
s
tu
d
ie
s
,
3
0
f
o
cu
s
o
n
g
en
er
al
-
p
u
r
p
o
s
e
VQA,
aim
in
g
to
an
s
wer
o
p
en
-
en
d
ed
q
u
esti
o
n
s
ac
r
o
s
s
d
iv
er
s
e
ev
er
y
d
ay
im
ag
es,
wh
ile
th
e
r
em
ain
in
g
2
8
a
d
d
r
ess
d
o
m
ai
n
-
s
p
ec
if
ic
ap
p
licatio
n
s
,
in
clu
d
in
g
m
e
d
ical
im
ag
in
g
,
r
e
m
o
te
s
en
s
in
g
,
ag
r
icu
lt
u
r
e,
an
d
d
o
cu
m
en
t
an
aly
s
is
.
Fig
u
r
e
2
p
r
esen
ts
th
e
o
v
er
all
d
is
tr
ib
u
tio
n
o
f
th
e
in
cl
u
d
ed
s
tu
d
ies.
Sp
ec
if
ically
,
Fig
u
r
e
2
(
a)
co
m
p
ar
es
th
e
p
r
o
p
o
r
tio
n
s
o
f
g
en
er
al
-
p
u
r
p
o
s
e
an
d
d
o
m
ain
-
s
p
ec
if
ic
VQA
s
tu
d
ies,
wh
er
ea
s
Fig
u
r
e
2
(
b
)
illu
s
tr
ates
th
e
d
is
tr
ib
u
tio
n
o
f
d
o
m
ai
n
-
s
p
ec
if
ic
ap
p
licatio
n
ar
ea
s
.
Gen
er
al
-
p
u
r
p
o
s
e
s
tu
d
ies
ar
e
ca
teg
o
r
ized
b
y
task
ty
p
e
(
e.
g
.
,
k
n
o
wled
g
e
-
b
ased
,
r
elatio
n
al,
o
r
o
p
en
-
e
n
d
ed
r
ea
s
o
n
in
g
)
,
wh
ile
d
o
m
ain
-
s
p
ec
if
ic
s
tu
d
ies
ar
e
ca
teg
o
r
ized
b
y
ap
p
licatio
n
d
o
m
ain
(
e.
g
.
,
m
ed
ical
im
ag
in
g
,
r
e
m
o
te
s
en
s
in
g
,
o
r
a
g
r
icu
ltu
r
e)
,
as su
m
m
ar
ized
in
T
ab
le
5
.
(
a)
(
b
)
Fig
u
r
e
2
.
Ov
e
r
all
s
tu
d
y
d
is
tr
ib
u
tio
n
o
f
m
u
lti
-
an
s
wer
VQA
(
n
=5
8
)
(
a
)
g
en
e
r
al
-
p
u
r
p
o
s
e
v
s
.
d
o
m
ain
-
s
p
ec
if
ic
VQA
an
d
(
b
)
d
is
tr
ib
u
tio
n
o
f
d
o
m
ain
-
s
p
ec
if
ic
VQA
ap
p
licati
o
n
s
(
n
=
2
8
)
T
ab
le
5
.
Dis
tr
ib
u
tio
n
o
f
in
cl
u
d
ed
s
tu
d
ies
D
o
ma
i
n
V
Q
A
N
u
m.
o
f
s
t
u
d
i
e
s
S
t
u
d
i
e
s
D
o
m
a
i
n
-
s
p
e
c
i
f
i
c
VQ
A
(
n
=
2
8
)
M
e
d
i
c
a
l
9
[
2
3
]
–
[
3
1
]
R
e
m
o
t
e
se
n
s
i
n
g
13
[3
2
]
–
[4
4
]
A
g
r
i
c
u
l
t
u
r
e
2
[
1
4
]
,
[
4
5
]
D
o
c
u
me
n
t
/
Le
g
a
l
3
[
4
6
]
–
[
4
8
]
O
t
h
e
r
(
F
a
c
i
a
l
a
n
a
l
y
s
i
s)
1
[
4
9
]
G
e
n
e
r
a
l
p
u
r
p
o
ses
VQ
A
(
n
=
3
0
)
G
e
n
e
r
a
l
o
p
e
n
-
e
n
d
e
d
r
e
a
s
o
n
i
n
g
13
[5
0
]
–
[
62
]
R
e
l
a
t
i
o
n
a
l
a
n
d
s
p
a
t
i
a
l
r
e
a
s
o
n
i
n
g
7
[
6
3
]
–
[
6
9
]
K
n
o
w
l
e
d
g
e
-
b
a
se
d
V
Q
A
(
K
B
-
VQA)
5
[
7
0
]
–
[
7
4
]
S
p
e
c
i
a
l
i
z
e
d
t
a
s
k
s (c
o
u
n
t
i
n
g
a
n
d
d
e
b
i
a
si
n
g
,
e
x
p
l
a
i
n
a
b
i
l
i
t
y
)
3
[
7
5
]
–
[
7
7
]
M
u
l
t
i
-
a
n
sw
e
r
a
n
d
m
u
l
t
i
-
c
h
o
i
c
e
d
e
si
g
n
2
[
8
]
,
[
7
8
]
3
.
2
.
2
.
E
x
pli
cit
v
s
.
im
pli
cit
mu
lt
i
-
a
ns
wer
f
o
rm
ula
t
io
ns
E
x
p
licit
m
u
lti
-
an
s
wer
s
tu
d
ies
d
ir
ec
tly
d
esig
n
m
o
d
els
to
g
en
er
ate
m
u
ltip
le
an
s
wer
s
th
r
o
u
g
h
m
u
lti
-
lab
el
p
r
ed
ictio
n
,
s
tr
u
ct
u
r
ed
o
u
tp
u
ts
,
o
r
g
en
er
ativ
e
m
ec
h
an
is
m
s
.
I
n
co
n
tr
ast,
m
o
s
t
o
f
th
e
r
ev
iewe
d
liter
atu
r
e
(
8
3
%,
n
=4
8
)
s
u
p
p
o
r
ts
im
p
licit
m
u
lti
-
an
s
wer
r
ea
s
o
n
i
n
g
,
i
n
wh
ich
m
o
d
els
in
co
r
p
o
r
ate
o
b
ject
-
b
ased
,
r
eg
io
n
-
b
ased
,
o
r
m
u
lti
-
in
s
tan
ce
m
ec
h
an
is
m
s
ev
en
th
o
u
g
h
th
e
f
in
al
p
r
e
d
ictio
n
r
em
ain
s
a
s
in
g
le
an
s
wer
.
I
n
th
ese
m
o
d
els,
o
b
ject
d
etec
tio
n
en
ab
les
d
if
f
e
r
en
tiatio
n
a
n
d
co
m
p
ar
is
o
n
o
f
m
u
ltip
le
v
i
s
u
al
en
titi
es
b
ef
o
r
e
g
en
er
atin
g
a
r
esp
o
n
s
e,
estab
lis
h
in
g
th
e
a
r
ch
itectu
r
al
f
o
u
n
d
atio
n
f
o
r
an
s
wer
v
a
r
iab
ilit
y
.
A
cr
itical
d
is
tin
ctio
n
ex
is
ts
b
et
wee
n
o
b
ject
d
etec
tio
n
u
s
ed
f
o
r
m
u
lti
-
en
tity
r
ea
s
o
n
in
g
v
er
s
u
s
d
etec
tio
n
u
s
ed
m
er
ely
to
e
n
h
an
ce
s
in
g
l
e
-
an
s
wer
ac
cu
r
ac
y
as
s
h
o
wn
in
Fig
u
r
e
3
.
Am
o
n
g
th
e
1
3
5
ex
clu
d
e
d
s
tu
d
ies,
6
2
em
p
lo
y
e
d
o
b
ject
d
etec
tio
n
o
n
ly
f
o
r
s
in
g
le
-
an
s
wer
en
h
a
n
c
em
en
t
an
d
wer
e
th
er
e
f
o
r
e
ex
cl
u
d
ed
.
As
s
h
o
wn
in
Fig
u
r
e
3
(
a)
,
o
n
ly
1
0
s
t
u
d
i
e
s
(
1
7
%
)
e
x
p
l
i
c
it
l
y
f
o
r
m
u
l
a
t
e
d
m
u
l
t
i
-
a
n
s
w
e
r
VQ
A
,
w
h
i
l
e
t
h
e
r
e
m
a
i
n
i
n
g
4
8
(
8
3
%
)
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
E
xp
lo
r
in
g
mu
lti
-
a
n
s
w
er v
is
u
a
l q
u
esti
o
n
a
n
s
w
erin
g
w
ith
o
b
ject
d
etec
tio
n
:
a
s
ystema
tic
… (
N
id
a
u
l H
a
s
a
n
a
ti
)
1103
a
d
d
r
e
s
s
e
d
a
n
s
w
e
r
v
a
r
ia
b
i
l
it
y
i
m
p
l
i
c
it
l
y
t
h
r
o
u
g
h
o
b
j
e
c
t
-
b
a
s
e
d
r
e
as
o
n
i
n
g
m
e
c
h
a
n
i
s
m
s
.
A
l
t
h
o
u
g
h
i
m
p
l
i
ci
t
a
p
p
r
o
a
c
h
e
s
d
o
n
o
t
e
x
p
l
i
c
it
l
y
g
e
n
e
r
a
t
e
m
u
l
t
i
p
l
e
o
u
t
p
u
t
s
,
t
h
ey
a
r
e
i
n
c
l
u
d
e
d
b
e
c
a
u
s
e
o
b
je
ct
-
c
e
n
t
r
i
c
r
e
as
o
n
i
n
g
p
r
o
v
i
d
e
s
t
h
e
a
r
c
h
i
t
e
ct
u
r
a
l
f
o
u
n
d
a
t
i
o
n
r
e
q
u
i
r
e
d
f
o
r
f
u
t
u
r
e
m
u
l
t
i
-
a
n
s
w
e
r
e
x
t
e
n
s
i
o
n
s
.
F
i
g
u
r
e
3
(
b
)
r
e
v
e
a
l
s
a
n
u
p
w
a
r
d
p
u
b
l
i
c
a
t
i
o
n
t
r
e
n
d
,
w
it
h
2
0
2
4
–
2
0
2
5
a
c
c
o
u
n
t
i
n
g
f
o
r
6
3
.
8
%
o
f
i
n
c
l
u
d
e
d
s
t
u
d
i
es
(
n
=
3
7
)
.
N
o
t
a
b
l
y
,
8
o
f
t
h
e
1
0
e
x
p
l
i
c
i
t
m
u
l
t
i
-
a
n
s
w
e
r
s
t
u
d
i
e
s
w
e
r
e
p
u
b
l
is
h
e
d
i
n
t
h
is
p
e
r
i
o
d
,
r
e
f
l
e
c
t
i
n
g
g
r
o
wi
n
g
a
w
a
r
e
n
ess
o
f
i
m
p
l
i
c
it
a
p
p
r
o
a
c
h
l
i
m
i
t
at
i
o
n
s
.
T
e
n
s
t
u
d
i
es
e
x
p
l
i
cit
l
y
f
o
r
m
u
l
a
t
e
d
m
u
lt
i
-
a
n
s
w
e
r
g
en
e
r
a
t
i
o
n
a
s
t
h
e
i
r
p
r
i
m
a
r
y
o
b
j
e
ct
i
v
e
.
T
h
es
e
i
n
cl
u
d
e
d
o
m
a
i
n
-
s
p
e
c
i
f
i
c
a
p
p
l
i
c
at
i
o
n
s
i
n
p
e
s
t
m
a
n
a
g
e
m
e
n
t
,
d
is
as
t
e
r
i
n
t
er
p
r
e
t
a
t
i
o
n
,
f
a
c
ia
l
e
x
p
r
es
s
i
o
n
a
n
a
l
y
s
is
,
a
n
d
m
e
d
i
c
a
l
V
Q
A
,
as
w
e
l
l
as
g
e
n
e
r
a
l
-
p
u
r
p
o
s
e
a
p
p
r
o
a
c
h
e
s
a
d
d
r
e
s
s
i
n
g
m
u
l
t
i
p
l
e
-
c
h
o
i
c
e
V
QA
d
i
r
ec
t
m
u
l
t
i
p
l
e
-
a
n
s
w
e
r
g
e
n
e
r
a
t
i
o
n
,
c
a
u
s
al
i
t
y
-
g
u
i
d
e
d
r
ea
s
o
n
i
n
g
,
a
n
d
k
n
o
w
l
e
d
g
e
-
a
u
g
m
e
n
t
e
d
VQA
a
s
d
e
ta
i
l
e
d
i
n
T
a
b
le
6
.
(
a)
(
b
)
Fig
u
r
e
3
.
Mu
lti
-
a
n
s
wer
VQA
f
o
r
m
u
latio
n
d
is
tr
ib
u
tio
n
(
n
=5
8
)
(
a)
n
u
m
b
er
o
f
e
x
p
licit v
s
im
p
l
icit stu
d
ies
an
d
(
b
)
p
u
b
licatio
n
tr
en
d
s
b
y
y
ea
r
an
d
m
u
lti
-
an
s
wer
ty
p
e
T
ab
le
6
.
E
x
p
licit
m
u
lti
-
an
s
wer
VQA
f
o
r
m
u
latio
n
s
ID
S
t
u
d
y
VQA
t
y
p
e
O
b
j
e
c
t
d
e
t
e
c
t
i
o
n
D
a
t
a
s
e
t
K
e
y
c
o
n
t
r
i
b
u
t
i
o
n
O
u
t
p
u
t
mec
h
a
n
i
sm
1
H
o
sse
i
n
a
b
a
d
e
t
a
l
.
E
M
Q
A
[
8
]
O
p
e
n
-
e
n
d
e
d
S
l
i
d
i
n
g
w
i
n
d
o
w
I
c
o
n
Q
u
e
st
i
o
n
A
n
sw
e
r
(
I
C
Q
A
)
,
D
A
Q
U
A
R
9
4
%
l
e
ss
o
p
e
r
a
t
i
o
n
s,
r
e
a
l
-
t
i
me
su
i
t
a
b
l
e
S
o
f
t
M
a
x
U
n
i
o
n
a
g
g
r
e
g
a
t
i
o
n
2
R
a
h
n
e
mo
o
n
f
a
r
e
t
a
l
.
R
e
sc
u
e
A
D
I
[
4
4
]
R
e
m
o
t
e
s
e
n
s
i
n
g
b
o
u
n
d
i
n
g
b
o
x
e
s
R
e
sc
u
e
A
d
a
p
t
i
v
e
D
i
sast
e
r
I
n
t
e
r
p
r
e
t
a
t
i
o
n
(
A
D
I
)
A
c
c
u
r
a
c
y
9
%
h
i
g
h
e
r
t
h
a
n
e
x
i
s
t
i
n
g
V
Q
A
met
h
o
d
s
S
t
r
u
c
t
u
r
e
d
o
u
t
p
u
t
t
u
p
l
e
s
(
,
,
)
3
L
iu
e
t
a
l
.
S
C
A
G
[
7
0
]
K
n
o
w
l
e
d
g
e
-
b
a
s
e
d
F
a
st
e
r
R
-
C
N
N
V
Q
A
v
2
,
O
K
-
V
Q
A
,
A
K
(
A
mb
i
g
u
o
u
s
K
n
o
w
l
e
d
g
e
)
A
c
c
u
r
a
c
y
:
V
Q
A
v
2
=
7
1
.
0
,
O
K
-
V
Q
A
=
5
4
.
3
6
,
A
K
=
5
0
.
4
2
;
4
.
6
2
%
i
mp
r
o
v
e
me
n
t
M
u
l
t
i
-
t
a
r
g
e
t
r
e
st
r
i
c
t
i
o
n
l
o
ss
4
Li
e
t
a
l
.
H
o
r
t
i
V
Q
A
-
PP
[
4
5
]
H
o
r
t
i
c
u
l
t
u
r
e
(
P
e
st
m
a
n
a
g
e
me
n
t
)
S
e
ma
n
t
i
c
seg
m
e
n
t
a
t
i
o
n
,
p
e
s
t
-
p
r
e
d
a
t
o
r
c
o
-
o
c
c
u
r
r
e
n
c
e
d
e
t
e
c
t
i
o
n
H
o
r
t
i
V
Q
A
-
PP
(
c
u
st
o
m
mu
l
t
i
m
o
d
a
l
d
a
t
a
se
t
(
3
0
p
e
st
+
1
0
p
r
e
d
a
t
o
r
c
a
t
e
g
o
r
i
e
s)
)
M
u
l
t
i
-
l
a
b
e
l
a
c
c
u
r
a
c
y
,
P
r
e
c
i
s
i
o
n
,
R
e
c
a
l
l
,
F1
-
sco
r
e
,
I
o
U
,
H
a
m
m
i
n
g
Lo
ss,
M
a
c
r
o
-
F
1
,
B
LEU
-
4
,
R
O
U
G
E
-
L,
M
ET
EO
R
,
E
M
,
G
P
T
sco
r
e
S
i
g
m
o
i
d
a
c
t
i
v
a
t
i
o
n
h
e
a
d
s,
mu
l
t
i
-
l
a
b
e
l
mat
c
h
i
n
g
5
Zh
a
n
g
e
t
a
l
.
H
i
C
A
-
VQA
[
2
4
]
M
e
d
i
c
a
l
V
Q
A
R
e
g
i
o
n
-
b
a
s
e
d
w
i
t
h
H
i
e
r
a
r
c
h
i
c
a
l
M
o
d
e
l
i
n
g
+
C
r
o
ss
-
A
t
t
e
n
t
i
o
n
R
a
d
-
R
e
s
t
r
u
c
t
F1
-
sc
o
r
e
1
8
%
i
mp
r
o
v
e
me
n
t
H
i
e
r
a
r
c
h
i
c
a
l
A
n
sw
e
r
D
e
c
o
d
e
r
(
S
i
g
m
o
i
d
)
6
U
p
a
d
h
y
a
y
a
n
d
Tr
i
p
a
t
h
y
[
7
3
]
M
u
l
t
i
p
l
e
c
h
o
i
c
e
V
Q
A
I
mag
e
r
e
g
i
o
n
s
w
i
t
h
a
t
t
e
n
t
i
o
n
V
i
su
a
l
7
W
,
V
Q
A
v
2
C
o
m
p
e
t
i
t
i
v
e
p
e
r
f
o
r
m
a
n
c
e
S
i
n
g
l
e
-
b
e
st
(
T
o
p
-
4
r
e
f
i
n
e
me
n
t
)
7
M
i
a
o
e
t
a
l
.
G
A
T2
R
[
7
4
]
C
G
C
N
(
c
a
u
s
a
l
i
t
y
-
g
u
i
d
e
d
)
O
b
j
e
c
t
+
c
a
u
sa
l
g
r
a
p
h
V
Q
A
2
.
0
,
V
Q
A
v
1
i
mp
r
o
v
e
d
p
e
r
f
o
r
ma
n
c
e
S
i
n
g
l
e
-
b
e
st
sel
e
c
t
i
o
n
(
S
o
f
t
M
a
x
)
8
Li
u
e
t
a
l
.
M
2
TV
Q
A
[
6
3
]
K
n
o
w
l
e
d
g
e
-
a
u
g
me
n
t
e
d
VQA
M
u
l
t
i
-
g
r
a
n
u
l
a
r
i
t
y
f
e
a
t
u
r
e
s/
m
u
l
t
i
-
r
e
g
i
o
n
se
l
e
c
t
i
o
n
OK
-
V
Q
A
,
K
R
V
Q
A
Ef
f
e
c
t
i
v
e
v
e
r
i
f
i
c
a
t
i
o
n
Tr
i
p
l
e
t
r
e
a
so
n
i
n
g
9
Li
u
e
t
a
l
.
I
C
Q
-
Tr
a
n
sE
[
7
8
]
K
n
o
w
l
e
d
g
e
-
b
a
s
e
d
V
Q
A
To
p
-
K
r
e
g
i
o
n
s
(
D
y
n
a
mi
c
v
i
s
u
a
l
r
e
g
i
o
n
se
l
e
c
t
i
o
n
)
OK
-
V
Q
A
,
K
R
V
Q
A
LLM
-
e
n
h
a
n
c
e
d
c
a
p
t
i
o
n
b
r
i
d
g
e
K
n
o
w
l
e
d
g
e
g
r
a
p
h
c
o
mp
l
e
t
i
o
n
10
Y
u
a
n
e
t
a
l
.
E
x
p
-
VQA
[
4
9
]
F
a
c
i
a
l
e
x
p
r
e
ss
i
o
n
a
n
a
l
y
si
s
I
n
st
r
u
c
t
B
LI
P
-
b
a
s
e
d
v
i
su
a
l
f
e
a
t
u
r
e
s,
f
a
c
i
a
l
r
e
g
i
o
n
s a
n
a
l
y
si
s
S
y
n
t
h
e
si
z
e
d
V
Q
A
p
a
i
r
s (em
o
t
i
o
n
c
l
a
ss
i
f
i
c
a
t
i
o
n
+
A
c
t
i
o
n
U
n
i
t
(
A
U
)
d
e
t
e
c
t
i
o
n
)
S
t
a
t
e
-
of
-
t
h
e
-
a
r
t
z
e
r
o
-
sh
o
t
A
U
d
e
t
e
c
t
i
o
n
G
e
n
e
r
a
t
i
v
e
n
a
r
r
a
t
i
v
e
c
a
p
t
i
o
n
i
n
g
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
:
1
0
9
7
-
1
1
1
4
1104
3
.
2
.
3
.
Da
t
a
s
et
utiliza
t
i
o
n
A
v
ar
iety
o
f
p
u
b
lic
an
d
d
o
m
ain
-
s
p
ec
if
ic
b
e
n
ch
m
a
r
k
d
atasets
ar
e
em
p
lo
y
ed
ac
r
o
s
s
ex
p
licit
m
u
lti
-
an
s
wer
VQA
s
tu
d
ies.
As
s
u
m
m
ar
ized
in
T
ab
le
7
,
co
m
m
o
n
l
y
u
s
ed
g
en
er
al
-
p
u
r
p
o
s
e
VQA
d
ataset
s
in
clu
d
e
VQA
v
2
,
GQA,
O
K
-
VQA,
an
d
Vis
u
al7
W
,
wh
ich
ar
e
wid
ely
ad
o
p
ted
f
o
r
o
p
en
-
d
o
m
ain
v
is
u
al
r
ea
s
o
n
in
g
task
s
.
I
n
ad
d
itio
n
,
s
ev
er
al
d
o
m
ain
-
s
p
ec
if
ic
d
atasets
h
av
e
em
er
g
e
d
to
s
u
p
p
o
r
t
s
p
ec
ialized
ap
p
licatio
n
s
,
s
u
ch
as
R
escu
e
ADI
f
o
r
r
em
o
te
s
en
s
in
g
d
is
a
s
ter
in
ter
p
r
etatio
n
,
R
ad
-
R
eStr
u
ct
f
o
r
s
tr
u
ctu
r
e
d
m
ed
ical
r
ep
o
r
t
g
en
e
r
atio
n
,
an
d
Ho
r
tiVQA
-
PP
f
o
r
ag
r
icu
ltu
r
al
p
est
m
an
ag
em
en
t.
Fu
r
th
er
m
o
r
e,
f
o
r
f
in
e
-
g
r
ain
e
d
f
ac
ial
ex
p
r
ess
i
o
n
an
aly
s
is
,
s
tu
d
ies
lik
e
E
x
p
-
VQA
lev
er
ag
e
s
y
n
th
esized
VQA
p
air
s
d
er
iv
ed
f
r
o
m
estab
lis
h
ed
em
o
tio
n
an
d
Acti
o
n
Un
it
(
AU)
d
atasets
,
in
clu
d
in
g
Af
f
ec
tNet,
B
P4
D,
DI
SF
A,
an
d
R
AF
-
AU,
to
en
ab
le
m
u
lti
-
s
ca
le
r
ea
s
o
n
in
g
f
r
o
m
m
ac
r
o
to
m
icr
o
ex
p
r
ess
io
n
s
.
On
ly
a
lim
ited
n
u
m
b
e
r
o
f
d
atasets
ex
p
licitly
s
u
p
p
o
r
t
m
u
lti
-
an
s
wer
r
ea
s
o
n
in
g
.
Data
s
ets
s
u
ch
as
DAQU
A
R
an
d
I
C
QA
allo
w
m
u
ltip
le
v
ali
d
an
s
wer
s
o
r
o
b
ject
-
lev
el
e
n
tity
d
escr
i
p
tio
n
s
f
o
r
a
s
in
g
le
im
ag
e
–
q
u
esti
o
n
p
air
,
wh
er
ea
s
m
o
s
t
wid
ely
u
s
ed
b
en
ch
m
ar
k
s
,
in
clu
d
in
g
VQA
v
2
an
d
OK
-
VQA,
im
p
licitly
co
n
tain
an
s
wer
v
a
r
iab
ilit
y
b
u
t
ar
e
f
r
eq
u
en
tly
ev
al
u
ated
u
s
in
g
a
s
in
g
le
-
an
s
wer
p
r
e
d
ictio
n
p
a
r
ad
ig
m
.
Data
s
ets
s
u
ch
as
Vis
u
al7
W
an
d
VQA
-
Me
d
p
r
o
v
i
d
e
lim
ite
d
an
s
wer
v
a
r
iab
ilit
y
co
m
p
ar
e
d
to
m
o
r
e
r
ec
en
t
s
tr
u
ctu
r
ed
b
en
c
h
m
ar
k
s
lik
e
KR
VQA,
wh
ich
f
o
c
u
s
es
o
n
m
u
lti
-
s
tep
r
ea
s
o
n
in
g
(
tha
t
is
c
r
uc
ia
l
for
the
M
2TV
QA
a
n
d
IC
Q
-
Tr
a
n
s
E
m
ode
l
)
.
C
o
n
s
eq
u
en
tly
,
th
e
ch
o
i
ce
o
f
d
ataset
r
ef
lects
b
o
t
h
th
e
tar
g
et
ap
p
licatio
n
d
o
m
ain
a
n
d
th
e
e
x
ten
t to
wh
ic
h
it su
p
p
o
r
ts
r
ea
s
o
n
in
g
o
v
er
m
u
ltip
le
v
is
u
al
en
titi
es.
I
n
ter
m
s
o
f
a
n
s
wer
m
u
ltip
licity
an
d
a
m
b
ig
u
ity
lev
el,
d
atasets
s
u
ch
as
VQA
v
2
,
OK
-
VQA,
an
d
I
C
QA
p
r
o
v
id
e
m
u
ltip
le
h
u
m
an
s
o
r
p
r
o
g
r
am
m
atica
lly
g
en
er
ate
d
an
s
wer
s
p
er
q
u
esti
o
n
,
wh
ic
h
in
tr
o
d
u
ce
s
a
h
i
g
h
lev
e
l
o
f
an
s
wer
am
b
ig
u
ity
an
d
r
e
f
le
cts
th
e
d
iv
er
s
ity
o
f
in
ter
p
r
etatio
n
s
.
I
n
co
n
tr
ast,
s
tr
u
ctu
r
ed
d
at
asets
s
u
ch
as
GQ
A
an
d
R
SVQA
ty
p
ically
p
r
o
v
id
e
a
s
in
g
le
d
o
m
in
an
t
an
s
wer
d
er
i
v
ed
f
r
o
m
s
tr
u
ctu
r
e
d
a
n
n
o
tatio
n
s
o
r
s
ce
n
e
g
r
ap
h
s
,
r
esu
ltin
g
in
lo
wer
am
b
ig
u
ity
b
u
t m
o
r
e
d
eter
m
in
is
tic
r
ea
s
o
n
in
g
task
s
.
T
ab
le
7
.
Data
s
ets em
p
lo
y
ed
ac
r
o
s
s
th
e
ex
p
licit m
u
lti
-
an
s
wer
VQA
s
tu
d
ies
D
a
t
a
s
e
t
#
I
mag
e
s
A
v
g
.
A
n
sw
e
r
s
A
n
n
o
t
a
t
i
o
n
s
t
y
l
e
A
mb
i
g
u
i
t
y
l
e
v
e
l
O
b
j
e
c
t
-
l
e
v
e
l
a
n
n
o
t
a
t
i
o
ns
M
u
l
t
i
-
a
n
sw
e
r
c
a
p
a
b
i
l
i
t
y
K
e
y
c
h
a
r
a
c
t
e
r
i
st
i
c
s
V
Q
A
v
2
.
0
~
2
0
4
k
10
1
0
h
u
m
a
n
a
n
sw
e
r
s
H
i
g
h
N
o
n
e
I
mp
l
i
c
i
t
C
o
n
se
n
s
u
s
-
b
a
se
d
e
v
a
l
u
a
t
i
o
n
GQA
~
1
1
3
k
1
S
c
e
n
e
g
r
a
p
h
Lo
w
S
c
e
n
e
g
r
a
p
h
s
I
mp
l
i
c
i
t
R
e
l
a
t
i
o
n
a
l
r
e
a
so
n
i
n
g
f
o
c
u
s
OK
-
VQA
~
1
4
k
10
M
u
l
t
i
p
l
e
h
u
m
a
n
a
n
sw
e
r
s
H
i
g
h
R
e
g
i
o
n
f
e
a
t
u
r
e
s
I
mp
l
i
c
i
t
O
u
t
s
i
d
e
k
n
o
w
l
e
d
g
e
r
e
q
u
i
r
e
d
K
R
V
Q
A
3
2
,
9
1
0
~
4
.
8
K
n
o
w
l
e
d
g
e
r
o
u
t
e
d
Lo
w
1
9
3
,
4
4
9
t
r
i
p
l
e
t
s
I
mp
l
i
c
i
t
K
n
o
w
l
e
d
g
e
-
r
o
u
t
e
d
r
e
a
so
n
i
n
g
V
i
su
a
l
7
W
~
4
7
k
4
M
u
l
t
i
p
l
e
c
h
o
i
c
e
M
o
d
e
r
a
t
e
B
o
u
n
d
i
n
g
b
o
x
e
s
(
B
B
s)
Li
mi
t
e
d
V
i
su
a
l
g
r
o
u
n
d
i
n
g
f
o
c
u
s
DAQUAR
~
1
.
4
k
1
–
3
H
u
ma
n
a
n
n
o
t
a
t
i
o
n
s
M
o
d
e
r
a
t
e
O
b
j
e
c
t
l
a
b
e
l
s
Ex
p
l
i
c
i
t
O
b
j
e
c
t
p
r
e
se
n
c
e
a
n
d
c
o
u
n
t
i
n
g
I
C
Q
A
~
4
2
k
M
u
l
t
i
p
l
e
P
r
o
g
r
a
m
mat
i
c
a
l
l
y
g
e
n
e
r
a
t
e
d
H
i
g
h
I
c
o
n
o
b
j
e
c
t
l
a
b
e
l
s
Ex
p
l
i
c
i
t
S
p
e
c
i
f
i
c
a
l
l
y
f
o
r
mu
l
t
i
-
a
n
sw
e
r
R
e
sc
u
e
A
D
I
4
,
0
4
4
~
3
.
3
S
t
r
u
c
t
u
r
e
d
t
u
p
l
e
s
(
P
,
R
,
A
)
M
o
d
e
r
a
t
e
1
6
,
9
4
9
m
a
sk
s
,
1
4
,
4
8
3
B
B
s
Ex
p
l
i
c
i
t
D
i
sast
e
r
i
n
t
e
r
p
r
e
t
a
t
i
o
n
t
u
p
l
e
s
R
S
V
Q
A
~
1
M
1
A
u
t
o
ma
t
i
c
a
l
l
y
g
e
n
e
r
a
t
e
d
Lo
w
B
o
u
n
d
i
n
g
b
o
x
e
s
I
mp
l
i
c
i
t
U
r
b
a
n
r
e
m
o
t
e
sen
s
i
n
g
d
a
t
a
R
a
d
-
R
e
S
t
r
u
c
t
3
,
7
2
0
~
4
8
3
-
t
i
e
r
h
i
e
r
a
r
c
h
y
M
o
d
e
r
a
t
e
1
7
8
p
a
t
h
o
l
.
v
o
c
a
b
Ex
p
l
i
c
i
t
H
i
e
r
a
r
c
h
i
c
a
l
c
l
i
n
i
c
a
l
r
e
p
o
r
t
s
VQA
-
M
e
d
~
4
k
1
Ex
p
e
r
t
a
g
e
n
e
r
a
t
e
d
n
n
o
t
a
t
i
o
n
s
Lo
w
-
M
o
d
e
r
a
t
e
N
o
n
e
Li
mi
t
e
d
C
l
i
n
i
c
a
l
Q
A
b
e
n
c
h
mar
k
s
H
o
r
t
i
V
Q
A
-
PP
1
4
,
4
6
0
6
.
3
P
e
st
-
p
r
e
d
a
t
o
r
mat
c
h
i
n
g
M
o
d
e
r
a
t
e
M
u
l
t
i
-
l
a
b
e
l
c
a
t
e
g
o
r
i
e
s
Ex
p
l
i
c
i
t
P
e
st
-
p
r
e
d
a
t
o
r
c
o
-
o
c
c
u
r
r
e
n
c
e
Ex
p
-
VQA
~
5
.
8
M
M
u
l
t
i
-
sca
l
e
M
u
l
t
i
-
sca
l
e
c
a
p
t
i
o
n
i
n
g
M
o
d
e
r
a
t
e
9
f
a
c
i
a
l
r
e
g
i
o
n
s
Ex
p
l
i
c
i
t
M
a
c
r
o
t
o
mi
c
r
o
-
e
x
p
r
e
ss
i
o
n
s
2
.
3
.
4
.
O
bje
ct
-
ba
s
ed
v
is
ua
l r
epre
s
ent
a
t
io
ns
T
h
ese
o
b
ject
-
lev
el
r
ep
r
esen
tat
io
n
s
s
er
v
e
as
th
e
p
r
im
ar
y
v
is
u
al
in
p
u
ts
f
o
r
s
u
b
s
eq
u
en
t
r
ea
s
o
n
in
g
an
d
an
s
wer
p
r
ed
ictio
n
,
f
o
r
m
i
n
g
a
co
n
s
is
ten
t
v
is
u
al
f
o
u
n
d
atio
n
a
cr
o
s
s
b
o
th
g
en
e
r
al
-
p
u
r
p
o
s
e
an
d
d
o
m
ain
-
s
p
ec
if
ic
VQA
s
tu
d
ies
as
s
h
o
wn
in
Fig
u
r
e
4
.
W
h
ile
Fig
u
r
e
4
(
a
)
s
h
o
ws
th
at
r
eg
io
n
-
b
ased
f
ea
t
u
r
es
ar
e
th
e
m
o
s
t
co
m
m
o
n
l
y
ad
o
p
te
d
ap
p
r
o
ac
h
ac
r
o
s
s
all
5
8
s
tu
d
ies
(
3
9
.
7
%),
Fig
u
r
e
4
(
b
)
r
ev
ea
ls
a
d
if
f
er
en
t
p
atter
n
am
o
n
g
th
e
1
0
ex
p
licit m
u
l
ti
-
an
s
wer
s
tu
d
i
es:
atten
tio
n
-
b
ased
r
eg
io
n
f
ea
tu
r
es e
m
er
g
e
as th
e
d
o
m
i
n
an
t stra
teg
y
(
5
0
%,
n
=5
)
,
f
o
llo
wed
b
y
Fas
ter
R
-
C
NN
-
b
ased
d
etec
tio
n
(
2
0
%,
n
=2
)
a
n
d
s
lid
in
g
-
win
d
o
w
d
etec
tio
n
,
tr
an
s
f
o
r
m
er
-
b
ased
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
E
xp
lo
r
in
g
mu
lti
-
a
n
s
w
er v
is
u
a
l q
u
esti
o
n
a
n
s
w
erin
g
w
ith
o
b
ject
d
etec
tio
n
:
a
s
ystema
tic
… (
N
id
a
u
l H
a
s
a
n
a
ti
)
1105
o
b
ject
d
etec
tio
n
,
an
d
s
em
an
tic
s
eg
m
en
tatio
n
,
ea
c
h
r
ep
r
e
s
en
tin
g
1
0
.
0
%
(
n
=
1
)
.
T
h
is
s
h
if
t
s
u
g
g
ests
th
at
r
esear
ch
er
s
d
esig
n
in
g
ex
p
licit
m
u
lti
-
an
s
wer
s
y
s
tem
s
ten
d
to
f
av
o
r
atten
tio
n
m
ec
h
an
i
s
m
s
o
v
er
g
en
er
al
r
eg
io
n
-
b
ased
ap
p
r
o
ac
h
es,
lik
ely
b
ec
au
s
e
atten
tio
n
allo
ws
m
o
d
els
to
s
elec
tiv
ely
f
o
cu
s
o
n
m
u
ltip
le
v
is
u
al
r
eg
io
n
s
s
im
u
ltan
eo
u
s
ly
.
T
h
is
o
b
s
er
v
atio
n
f
u
r
t
h
er
h
ig
h
lig
h
ts
th
e
im
p
o
r
tan
ce
o
f
o
b
ject
-
le
v
e
l
r
ep
r
esen
tatio
n
s
as
a
f
u
n
d
a
m
en
tal
co
m
p
o
n
e
n
t in
e
n
ab
lin
g
m
u
lti
-
en
tity
r
ea
s
o
n
i
n
g
in
m
u
lti
-
an
s
wer
VQA
s
y
s
tem
s
.
(
a)
(
b
)
Fig
u
r
e
4
.
A
c
o
n
s
is
ten
t v
is
u
al
f
o
u
n
d
atio
n
ac
r
o
s
s
b
o
t
h
g
en
e
r
al
-
p
u
r
p
o
s
e
an
d
d
o
m
ain
-
s
p
ec
if
ic
VQA
(
a)
o
b
ject
d
etec
tio
n
ap
p
r
o
ac
h
e
s
(
All Stu
d
ies,
n
=5
8
)
an
d
(
b
)
d
is
tr
ib
u
tio
n
o
f
v
is
u
al
r
ep
r
esen
ta
tio
n
m
eth
o
d
s
am
o
n
g
e
x
p
licit m
u
lti
-
an
s
wer
VQA
m
o
d
els (
n
=1
0
)
3
.
3
.
Na
rr
a
t
iv
e
s
y
nthesis
o
f
t
he
inclu
ded stu
dies
T
o
s
y
n
th
esize
th
e
f
in
d
in
g
s
o
f
th
e
r
ev
iewe
d
s
tu
d
ies,
th
e
lite
r
atu
r
e
is
an
aly
ze
d
ac
co
r
d
in
g
t
o
th
e
f
o
u
r
an
aly
tical
d
im
en
s
io
n
s
illu
s
tr
a
ted
in
Fig
u
r
e
5
,
n
am
el
y
task
f
o
r
m
u
latio
n
,
v
is
u
al
r
e
p
r
es
en
tatio
n
s
tr
ateg
ies,
s
o
u
r
c
es o
f
an
s
wer
m
u
ltip
licity
,
an
d
e
v
alu
atio
n
s
tr
ateg
ies.
Fig
u
r
e
5
.
T
a
x
o
n
o
m
y
m
u
lti
-
an
s
wer
VQA
3
.
3
.
1
.
M
ulti
-
a
ns
wer
v
s
.
im
pli
cit
m
ulti
-
a
ns
wer
f
o
rm
ula
t
i
o
n (
t
a
s
k
f
o
rm
ula
t
io
n
)
:
RQ
1
T
h
e
f
ir
s
t
d
im
en
s
io
n
o
f
th
e
tax
o
n
o
m
y
co
n
ce
r
n
s
h
o
w
th
e
VQA
task
is
f
o
r
m
ally
d
ef
in
ed
t
o
ac
co
m
m
o
d
ate
m
u
ltip
le
v
alid
a
n
s
wer
s
.
T
h
r
ee
m
ain
task
f
o
r
m
u
latio
n
s
em
er
g
e
f
r
o
m
liter
atu
r
e.
T
h
e
m
o
s
t e
x
p
licit
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
:
1
0
9
7
-
1
1
1
4
1106
ap
p
r
o
ac
h
m
o
d
els
VQA
as
a
m
u
lti
-
lab
el
class
if
icatio
n
p
r
o
b
lem
,
wh
e
r
e
m
u
ltip
le
an
s
wer
s
ca
n
b
e
p
r
ed
icted
s
im
u
ltan
eo
u
s
ly
f
o
r
a
s
in
g
le
im
ag
e
–
q
u
esti
o
n
p
air
.
T
h
is
f
o
r
m
u
latio
n
is
p
ar
ticu
lar
ly
p
r
ev
alen
t
in
d
o
m
ain
-
s
p
ec
if
ic
ap
p
licatio
n
s
,
s
u
ch
as
m
ed
ical
im
ag
in
g
an
d
ag
r
icu
ltu
r
e,
w
h
er
e
q
u
esti
o
n
s
n
atu
r
ally
co
r
r
esp
o
n
d
t
o
m
u
ltip
le
attr
ib
u
tes,
co
n
d
itio
n
s
,
o
r
f
in
d
i
n
g
s
.
A
s
ec
o
n
d
ap
p
r
o
ac
h
r
elies
o
n
c
o
n
s
en
s
u
s
-
b
ased
m
o
d
elin
g
,
in
wh
ich
m
u
ltip
le
a
n
s
wer
s
ar
e
d
e
r
iv
ed
f
r
o
m
an
n
o
tato
r
ag
r
ee
m
en
t
o
r
p
r
o
b
a
b
ilis
tic
co
n
f
id
en
ce
d
is
tr
ib
u
tio
n
s
.
T
h
is
f
o
r
m
u
latio
n
is
co
m
m
o
n
ly
ass
o
ciate
d
with
g
en
er
al
-
p
u
r
p
o
s
e
VQ
A
b
en
ch
m
ar
k
s
th
at
p
r
o
v
id
e
m
u
ltip
le
h
u
m
an
an
n
o
tatio
n
s
p
er
q
u
esti
o
n
,
allo
win
g
s
y
s
tem
s
to
r
ef
lect
an
s
wer
v
ar
ia
b
ilit
y
with
o
u
t e
x
p
licitly
r
ed
ef
in
in
g
th
e
o
u
tp
u
t sp
ac
e.
T
h
e
th
ir
d
f
o
r
m
u
latio
n
ad
o
p
ts
g
en
er
ativ
e
m
u
lti
-
a
n
s
wer
m
o
d
elin
g
,
w
h
er
e
m
o
d
els
ar
e
d
e
s
ig
n
ed
to
pr
o
d
u
ce
a
s
et
o
r
s
eq
u
en
ce
o
f
p
lau
s
ib
le
an
s
wer
s
r
ath
e
r
th
a
n
a
s
in
g
le
p
r
ed
ictio
n
.
W
h
ile
t
h
is
ap
p
r
o
ac
h
alig
n
s
well
with
r
ec
en
t
ad
v
an
ce
s
in
lar
g
e
v
is
io
n
–
lan
g
u
a
g
e
an
d
m
u
ltimo
d
al
lan
g
u
a
g
e
m
o
d
els,
it
r
em
ain
s
u
n
d
er
e
x
p
lo
r
e
d
in
th
e
r
ev
iewe
d
liter
atu
r
e
d
u
e
t
o
ch
allen
g
e
s
in
co
n
tr
o
llab
ilit
y
a
n
d
ev
al
u
atio
n
.
R
esear
ch
o
n
m
u
lti
-
a
n
s
wer
VQA
d
em
o
n
s
tr
ates
f
u
n
d
am
en
ta
l
d
iv
er
g
e
n
ce
s
in
ar
c
h
itectu
r
al
s
tr
ateg
ies
f
o
r
h
an
d
lin
g
an
s
wer
m
u
ltip
lic
ity
.
B
ased
o
n
n
ar
r
ati
v
e
s
y
n
th
e
s
is
,
a
cr
itical
ap
p
r
aisal
o
f
th
e
1
0
ex
p
licit
s
tu
d
ies
(
1
7
%),
as
d
et
ailed
in
T
ab
le
6
,
r
ev
ea
ls
two
d
is
tin
ct
ar
ch
itectu
r
al
co
h
o
r
ts
:
a)
T
r
u
ly
ex
p
licit
m
o
d
els
(
n
=
3
)
,
wh
ich
in
tr
o
d
u
ce
ar
ch
itectu
r
al
m
o
d
if
icatio
n
s
d
esig
n
ed
to
g
en
er
ate
an
s
wer
s
ets
th
r
o
u
g
h
m
ec
h
an
is
m
s
s
u
ch
as
m
u
lti
-
lab
el
p
r
e
d
ictio
n
o
r
r
eg
io
n
ag
g
r
eg
atio
n
.
T
h
ese
m
o
d
els
i
n
tr
o
d
u
ce
d
ir
ec
t
ar
ch
itectu
r
al
m
o
d
if
icatio
n
s
to
th
e
o
u
tp
u
t la
y
er
th
at
e
n
ab
le
th
e
s
im
u
ltan
eo
u
s
g
en
er
atio
n
o
f
m
u
ltip
le
an
s
wer
s
[
8
]
,
[
2
4
]
,
[
4
5
]
.
b)
Ag
g
r
eg
atio
n
a
n
d
lo
s
s
in
n
o
v
a
tio
n
s
:
So
f
tMa
x
Un
io
n
was
in
t
r
o
d
u
ce
d
in
th
e
E
MQ
A
s
y
s
te
m
to
ag
g
r
eg
ate
p
r
o
b
a
b
ilit
ies
f
r
o
m
im
ag
e
g
lim
p
s
es,
g
en
er
atin
g
a
n
s
wer
s
s
im
u
ltan
eo
u
s
ly
v
ia
s
p
atial
o
u
tp
u
t
m
ap
v
ec
to
r
s
[
8
]
.
SC
AG
em
p
lo
y
ed
s
ig
m
o
id
r
estrictio
n
lo
s
s
o
r
m
u
lti
-
tar
g
e
t
r
estrictio
n
lo
s
s
to
ac
co
m
m
o
d
ate
m
u
ltip
le
g
r
o
u
n
d
-
tr
u
th
lab
els,
y
et
in
f
er
e
n
ce
r
etu
r
n
s
th
e
h
ig
h
est
-
s
co
r
in
g
an
s
wer
[
7
0
]
.
c)
Stru
ctu
r
ed
an
d
m
u
lti
-
lab
el
o
u
tp
u
ts
:
Ho
r
tiVQA
-
PP
an
d
Hi
C
A
-
VQA
u
ti
lize
s
ig
m
o
id
ac
tiv
atio
n
h
ea
d
s
to
p
r
ed
ict
m
u
ltip
le
p
est
-
p
r
ed
ato
r
r
elatio
n
s
h
ip
s
o
r
p
at
h
o
lo
g
ical
a
ttrib
u
tes
co
n
cu
r
r
en
tly
[
2
4
]
,
[
4
5
]
.
R
escu
eAD
I
p
r
o
d
u
ce
s
s
tr
u
ctu
r
ed
o
u
tp
u
t
tu
p
les
(
P,
R
,
A)
to
r
ec
o
r
d
p
la
n
s
an
d
i
n
ter
m
ed
iate
r
esu
lts
;
h
o
wev
er
,
th
e
o
u
tp
u
t
el
em
en
t A
r
em
ain
s
a
s
in
g
le
an
s
wer
in
n
atu
r
al
lan
g
u
ag
e
[
3
6
]
.
d)
E
x
p
licit
T
ask
Mo
d
els
(
n
=7
)
:
Up
ad
h
y
ay
an
d
T
r
i
p
ath
y
[
4
9
]
,
GAT
2
R
[
6
3
]
,
I
C
Q
-
T
r
a
n
s
E
[
7
3
]
,
M2
T
VQA
[
7
4
]
,
E
x
p
-
VQA
[
7
8
]
,
SC
AG
[
7
0
]
,
a
n
d
R
escu
eAD
I
[
4
4
]
.
T
h
ese
m
o
d
els
ad
d
r
ess
m
u
lti
-
en
tity
o
r
co
m
p
lex
task
r
ea
s
o
n
in
g
b
u
t
r
em
ain
co
n
s
tr
ain
ed
to
s
in
g
le
-
an
s
wer
o
u
tp
u
ts
at
in
f
er
en
ce
(
e.
g
.
,
SC
AG
ap
p
lies
m
u
lti
-
lab
el
lo
s
s
o
n
ly
d
u
r
in
g
t
r
ain
in
g
,
wh
ile
R
escu
eAD
I
o
u
tp
u
ts
s
tr
u
ctu
r
ed
task
tu
p
les
with
a
s
in
g
le
n
atu
r
al
lan
g
u
ag
e
a
n
s
wer
)
.
e)
R
elatio
n
al
an
d
s
ca
led
r
ea
s
o
n
in
g
:
g
r
a
p
h
atten
tio
n
n
etwo
r
k
s
(
GAT
)
wer
e
u
s
ed
f
o
r
ex
p
licit
r
elatio
n
al
m
o
d
elin
g
[
7
4
]
,
an
d
E
x
p
-
VQA
p
r
o
v
i
d
es
s
ca
le
-
s
p
ec
if
ic
ca
p
tio
n
in
g
(
m
ac
r
o
,
m
eso
,
m
icr
o
)
to
d
escr
ib
e
f
ac
ial
ac
tio
n
s
in
a
s
in
g
le
n
ar
r
ativ
e
r
e
s
p
o
n
s
e
[
4
9
]
.
f)
R
ef
in
em
en
t
an
d
tr
ip
lets
:
a
tw
o
-
s
tag
e
r
ef
in
em
en
t
o
f
T
o
p
-
4
c
an
d
id
ate
s
ets
h
av
e
b
ee
n
em
p
l
o
y
ed
[
7
3
]
,
wh
ile
tr
ip
let
r
ea
s
o
n
in
g
to
p
r
ed
ict
a
s
p
ec
if
ic
tail
en
tity
b
ased
o
n
m
u
lti
-
r
eg
io
n
s
elec
tio
n
h
as
b
ee
n
ad
o
p
ted
i
n
k
n
o
wled
g
e
-
b
ased
VQA
f
r
am
e
wo
r
k
s
,
M2
T
VQA
an
d
I
C
Q
-
T
r
an
s
E
[
6
3
]
,
[
7
8
]
.
T
h
is
tax
o
n
o
m
y
co
n
f
ir
m
s
a
s
ig
n
if
ican
t
g
ap
i
n
th
e
f
ield
:
wh
ile
co
n
tem
p
o
r
ar
y
VQA
m
o
d
els
h
av
e
d
ev
elo
p
e
d
r
o
b
u
s
t
p
er
ce
p
tu
al
ca
p
ab
ilit
y
o
f
id
en
tif
y
in
g
an
d
g
r
o
u
n
d
in
g
m
u
ltip
le
en
titi
es,
s
u
p
p
o
r
ted
b
y
8
3
%
(
n
=4
8
)
o
f
th
e
r
e
v
iewe
d
s
tu
d
ies,
th
e
p
r
im
ar
y
b
o
ttlen
ec
k
r
em
ain
s
in
h
o
w
th
ey
‘
s
p
ea
k
’
(
th
e
o
u
tp
u
t
m
ec
h
a
n
is
m
)
.
Ou
r
an
aly
s
is
r
ev
ea
ls
th
at
o
n
ly
1
7
%
(
n
=1
0
)
o
f
th
e
s
tu
d
ies
ex
p
licitly
f
o
r
m
u
late
th
e
task
to
ad
d
r
ess
an
s
wer
m
u
ltip
licity
.
3
.
3
.
2
.
Vis
ua
l R
epre
s
ent
a
t
io
n
Str
a
t
eg
y
:
RQ
2
T
h
e
s
ec
o
n
d
d
im
e
n
s
io
n
o
f
t
h
is
tax
o
n
o
m
y
is
th
at
v
is
u
al
r
e
p
r
es
en
tatio
n
s
tr
ateg
ies
in
m
u
lti
-
an
s
wer
VQA
h
av
e
ev
o
lv
e
d
f
r
o
m
h
o
lis
tic
-
s
ce
n
e
en
co
d
in
g
to
m
o
r
e
g
r
a
n
u
lar
,
en
tity
-
ce
n
tr
ic
r
e
p
r
esen
ta
tio
n
s
.
B
ased
o
n
th
e
r
ev
iewe
d
s
tu
d
ies,
th
ese
s
tr
ateg
ies
ca
n
b
e
b
r
o
ad
ly
ca
teg
o
r
ize
d
in
to
th
r
ee
g
r
o
u
p
s
:
g
lo
b
al
im
ag
e
r
ep
r
esen
tatio
n
s
,
o
b
ject
-
lev
el
o
r
r
eg
io
n
-
b
ased
r
e
p
r
esen
tatio
n
s
,
an
d
h
y
b
r
id
s
tr
at
eg
ies th
at
co
m
b
in
e
b
o
th
.
Glo
b
al
im
ag
e
f
ea
tu
r
es
ca
p
tu
r
e
h
o
lis
tic
v
is
u
al
in
f
o
r
m
atio
n
f
r
o
m
th
e
en
tire
s
ce
n
e.
Ho
w
ev
er
,
s
u
ch
r
ep
r
esen
tatio
n
s
ar
e
g
en
er
ally
in
s
u
f
f
icien
t
f
o
r
m
u
lti
-
an
s
wer
r
ea
s
o
n
in
g
b
ec
au
s
e
th
ey
la
ck
th
e
s
p
atial
an
d
in
s
tan
ce
-
lev
el
g
r
an
u
lar
ity
r
eq
u
ir
ed
to
d
is
tin
g
u
is
h
m
u
ltip
le
r
el
ev
an
t
e
n
titi
es
with
in
co
m
p
lex
im
ag
es.
As
a
r
esu
lt,
o
b
ject
-
lev
el
o
r
r
eg
io
n
-
b
ased
r
ep
r
esen
tatio
n
s
h
av
e
e
m
er
g
ed
as
th
e
d
o
m
in
an
t
s
tr
at
eg
y
in
m
u
lti
-
a
n
s
wer
VQA.
T
h
ese
r
ep
r
esen
tatio
n
s
ar
e
ty
p
ically
d
er
iv
ed
f
r
o
m
o
b
ject
d
etec
tio
n
o
r
r
e
g
io
n
p
r
o
p
o
s
al
m
ec
h
an
is
m
s
,
en
ab
lin
g
m
o
d
els
to
lo
ca
lize
m
u
ltip
le
o
b
jects
an
d
ex
tr
ac
t
th
eir
attr
ib
u
tes
s
o
th
at
d
if
f
er
e
n
t
v
i
s
u
al
en
titi
es
ca
n
b
e
ass
o
ciate
d
with
d
if
f
er
en
t
p
lau
s
ib
le
an
s
wer
s
.
An
o
th
er
em
er
g
in
g
d
ir
ec
ti
o
n
in
v
o
lv
es
h
y
b
r
id
v
is
u
al
r
ep
r
esen
tatio
n
s
,
wh
ich
c
o
m
b
in
e
g
lo
b
al
co
n
tex
tu
al
i
n
f
o
r
m
atio
n
with
lo
ca
lized
o
b
ject
-
lev
el
f
ea
tu
r
es.
Evaluation Warning : The document was created with Spire.PDF for Python.