I
AE
S In
t
er
na
t
io
na
l J
o
urna
l o
f
Art
if
icia
l In
t
ellig
ence
(
I
J
-
AI
)
Vo
l.
1
5
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
6
,
p
p
.
766
~
779
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
15
.i
1
.
p
p
7
6
6
-
7
7
9
766
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Im
pro
v
ing
ef
ficie
ncy
of autism
det
ection ba
sed o
n
f
a
cia
l ima
g
e
la
ndma
rks
Ng
uy
en
T
ro
ng
T
un
g
1,
2
,
Ng
o
Duc
Vinh
3
,
H
a
M
a
n
h T
o
a
n
4
,
Do
Na
ng
T
o
a
n
4
1
F
a
c
u
l
t
y
o
f
I
n
f
o
r
m
a
t
i
o
n
T
e
c
h
n
o
l
o
g
y
,
H
o
C
h
i
M
i
n
h
C
i
t
y
O
p
e
n
U
n
i
v
e
r
si
t
y
,
H
o
C
h
i
M
i
n
h
C
i
t
y
,
V
i
e
t
n
a
m
2
F
a
c
u
l
t
y
o
f
I
n
f
o
r
m
a
t
i
o
n
T
e
c
h
n
o
l
o
g
y
,
D
o
n
g
A
U
n
i
v
e
r
s
i
t
y
,
D
a
N
a
n
g
C
i
t
y
,
V
i
e
t
n
a
m
3
F
a
c
u
l
t
y
o
f
I
n
f
o
r
m
a
t
i
o
n
T
e
c
h
n
o
l
o
g
y
,
H
a
n
o
i
U
n
i
v
e
r
s
i
t
y
o
f
I
n
d
u
st
r
y
,
H
a
n
o
i
,
V
i
e
t
n
a
m
4
I
n
st
i
t
u
t
e
o
f
I
n
f
o
r
ma
t
i
o
n
Te
c
h
n
o
l
o
g
y
,
V
i
e
t
n
a
m
A
c
a
d
e
my
o
f
S
c
i
e
n
c
e
a
n
d
Te
c
h
n
o
l
o
g
y
,
H
a
n
o
i
,
V
i
e
t
n
a
m
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
J
u
n
1
1
,
2
0
2
5
R
ev
is
ed
No
v
4
,
2
0
2
5
Acc
ep
ted
J
an
1
0
,
2
0
2
6
Au
ti
sm
is
a
se
rio
u
s
m
e
n
tal
h
e
a
lt
h
p
r
o
b
lem
with
lo
n
g
-
term
e
ffe
c
ts
o
n
li
fe
.
Th
e
re
fo
re
,
e
a
rly
d
iag
n
o
sis
is
a
to
p
ica
l
iss
u
e
fo
r
e
ffe
c
ti
v
e
trea
tme
n
t.
T
h
is
stu
d
y
p
r
o
p
o
se
s
a
n
o
v
e
l
fa
c
i
a
l
lan
d
m
a
rk
tran
sf
o
rm
a
ti
o
n
-
b
a
se
d
d
a
ta
a
u
g
m
e
n
tatio
n
m
e
th
o
d
th
a
t
a
ll
o
ws
fo
r
th
e
g
e
n
e
ra
ti
o
n
o
f
g
e
o
m
e
tri
c
tran
sfo
rm
a
ti
o
n
s
re
late
d
to
fa
c
ial
g
e
o
m
e
try
.
Th
is
m
e
th
o
d
in
c
re
a
se
s
th
e
g
e
n
e
ra
li
z
a
b
il
it
y
a
n
d
p
r
o
v
i
d
e
s
a
p
e
rsp
e
c
ti
v
e
o
n
t
h
e
ro
le
o
f
fa
c
ial
r
e
g
io
n
s
in
a
u
ti
sm
d
e
tec
ti
o
n
.
T
h
e
p
r
o
p
o
se
d
a
u
g
m
e
n
tati
o
n
m
e
th
o
d
e
n
su
re
s
t
h
e
g
e
n
e
ra
ti
o
n
o
f
v
a
rian
ts
t
h
a
t
a
re
c
o
n
siste
n
t
wit
h
th
e
fa
c
ial
ima
g
e
stru
c
tu
re
a
n
d
t
h
e
n
a
tu
re
o
f
t
h
e
fa
c
ial
ima
g
e
.
Ne
x
t
,
c
o
n
d
u
c
t
a
c
o
m
p
re
h
e
n
siv
e
a
n
d
c
o
m
p
a
ra
ti
v
e
st
u
d
y
with
Eff
icie
n
tNe
t
-
B0
,
Eff
icie
n
tN
e
t
-
B4
,
Re
sN
e
t
-
1
8
,
Re
sN
e
t
-
5
0
,
Re
sN
e
t
-
1
0
1
,
M
o
b
i
leN
e
t
-
V2
,
De
n
se
Ne
t
-
1
2
1
a
n
d
De
n
se
Ne
t
-
2
0
1
.
A
lso
a
n
a
l
y
z
e
th
e
m
o
d
e
l'
s
a
tt
e
n
ti
o
n
o
v
e
r
t
h
e
m
a
in
re
g
io
n
s
o
f
t
h
e
fa
c
e
th
a
t
a
re
re
late
d
to
fa
c
ial
lan
d
m
a
rk
s.
T
h
e
re
su
lt
s
c
lea
rly
sh
o
w
th
a
t
t
h
e
m
o
d
e
ls
train
e
d
with
th
e
p
ro
p
o
se
d
m
e
th
o
d
o
u
t
p
e
rfo
rm
t
h
e
d
e
fa
u
lt
a
u
g
m
e
n
tati
o
n
m
e
th
o
d
.
S
p
e
c
ifi
c
a
ll
y
,
wh
e
n
a
v
e
ra
g
in
g
t
h
e
m
e
a
su
re
s
a
c
ro
ss
th
e
tes
ted
m
o
d
e
ls,
t
h
e
re
su
lt
s
a
re
0
.
9
0
5
4
1
7
fo
r
a
c
c
u
ra
c
y
,
0
.
9
6
2
1
3
3
fo
r
a
re
a
u
n
d
e
r
t
h
e
c
u
rv
e
(
AUC
)
,
0
.
9
1
9
8
fo
r
p
re
c
isio
n
,
0
.
8
8
8
3
3
3
fo
r
re
c
a
ll
,
a
n
d
0
.
9
0
3
6
7
8
f
o
r
F
1
-
sc
o
re
.
F
u
r
th
e
rm
o
re
,
wh
e
n
a
n
a
ly
z
i
n
g
t
h
e
g
ra
d
ien
t
-
w
e
ig
h
ted
c
las
s
a
c
ti
v
a
ti
o
n
m
a
p
p
i
n
g
(G
ra
d
-
CAM)
h
e
a
tma
p
s,
th
e
h
i
g
h
-
v
a
lu
e
r
e
g
io
n
s a
re
c
lea
rly
c
o
n
c
e
n
trate
d
o
n
th
e
m
a
in
a
re
a
s o
f
th
e
fa
c
e
.
S
o
u
rc
e
c
o
d
e
is
p
u
b
li
s
h
e
d
o
n
G
it
L
a
b
p
latfo
rm
.
K
ey
w
o
r
d
s
:
Au
tis
m
d
etec
tio
n
Dee
p
lear
n
in
g
Facial
au
g
m
en
tatio
n
Facial
ch
ild
im
ag
e
Facial
lan
d
m
ar
k
Hea
tm
ap
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
:
Do
Nan
g
T
o
a
n
I
n
s
titu
te
o
f
I
n
f
o
r
m
atio
n
T
ec
h
n
o
lo
g
y
,
Vietn
a
m
Aca
d
em
y
o
f
S
cien
ce
an
d
T
ec
h
n
o
lo
g
y
Han
o
i,
Vietn
am
E
m
ail:
d
n
to
an
@
io
it.a
c.
v
n
1.
I
NT
RO
D
UCT
I
O
N
On
e
o
f
th
e
im
p
o
r
tan
t
p
r
o
b
le
m
s
af
f
ec
tin
g
n
eu
r
o
d
ev
el
o
p
m
e
n
t
is
au
tis
m
s
p
ec
tr
u
m
d
is
o
r
d
er
(
ASD)
,
wh
ich
is
r
elate
d
to
ab
n
o
r
m
al
ities
in
th
e
d
ev
elo
p
m
en
t
o
f
t
h
e
b
r
ain
an
d
h
as
a
v
is
u
al
i
m
p
ac
t
o
n
p
h
y
s
ical
ch
ar
ac
ter
is
tics
an
d
f
ac
ial
ex
p
r
ess
io
n
s
[
1
]
.
Sig
n
s
o
f
au
tis
m
ca
n
ap
p
ea
r
i
n
m
an
y
d
if
f
e
r
en
t sit
u
atio
n
s
in
th
e
life
o
f
th
e
af
f
ec
ted
p
er
s
o
n
.
Au
tis
m
is
o
f
ten
ex
p
r
ess
ed
in
m
an
y
d
if
f
er
en
t
ch
ar
ac
ter
is
tics
,
s
u
ch
as
r
ep
etitiv
e
b
eh
av
io
r
s
an
d
a
lack
o
f
s
o
cial
co
m
m
u
n
i
ca
tio
n
[
2
]
.
T
h
is
p
h
e
n
o
m
e
n
o
n
,
wh
en
ap
p
ea
r
in
g
f
o
r
a
lo
n
g
tim
e,
will
h
av
e
s
er
io
u
s
im
p
ac
ts
o
n
th
e
liv
es o
f
au
tis
tic
p
eo
p
le
as we
ll a
s
th
o
s
e
ar
o
u
n
d
th
em
.
Su
lk
es
[
3
]
h
as
s
h
o
wn
th
at
th
e
r
is
k
o
f
h
av
in
g
a
n
o
th
er
c
h
ild
with
au
tis
m
is
ab
o
u
t
3
-
1
0
%
f
o
r
p
ar
en
t
s
wh
o
alr
ea
d
y
h
av
e
a
c
h
ild
wit
h
au
tis
m
.
I
n
th
e
Un
ited
States
,
it
is
esti
m
ated
t
h
at
1
in
5
9
c
h
ild
r
en
ag
ed
8
an
d
o
ld
er
h
as
au
tis
m
[
4
]
.
I
n
ad
d
i
tio
n
,
th
e
W
o
r
ld
Hea
lth
Or
g
a
n
izatio
n
h
as
p
u
b
lis
h
ed
alar
m
in
g
f
ig
u
r
es
o
n
th
e
p
r
ev
alen
ce
o
f
au
tis
m
.
Sp
ec
if
ically
,
ac
co
r
d
in
g
to
a
r
ep
o
r
t
[
5
]
,
wh
en
co
n
d
u
ctin
g
s
tatis
tics
o
n
1
6
%
o
f
c
h
ild
r
e
n
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
I
mp
r
o
vin
g
efficien
cy
o
f
a
u
tis
m
d
etec
tio
n
b
a
s
ed
o
n
f
a
cia
l ima
g
e
la
n
d
ma
r
ks
(
N
g
u
ye
n
Tr
o
n
g
Tu
n
g
)
767
g
lo
b
ally
,
0
.
6
7
%
o
f
th
e
m
h
a
v
e
au
tis
m
.
T
h
e
Vietn
am
ese
g
o
v
er
n
m
en
t
a
n
d
c
o
m
m
u
n
ity
h
a
v
e
b
ee
n
c
o
n
tr
ib
u
tin
g
s
ev
er
al
attem
p
ts
to
d
ea
l
with
au
tis
m
.
On
e
ca
s
e
is
th
e
ef
f
o
r
t
r
elatin
g
to
w
o
r
ld
a
u
tis
m
awa
r
en
ess
d
ay
o
f
t
h
e
Min
is
tr
y
o
f
L
ab
o
r
,
I
n
v
alid
s
an
d
So
cial
Af
f
air
s
[
6
]
.
T
h
is
ac
tiv
ity
h
elp
s
th
e
co
m
m
u
n
ity
u
n
d
er
s
tan
d
an
d
s
u
p
p
o
r
t
au
tis
tic
ch
ild
r
en
.
T
o
d
etec
t
a
u
tis
m
in
ch
ild
r
en
,
an
ef
f
ec
tiv
e
ap
p
r
o
ac
h
is
to
u
s
e
d
ee
p
lear
n
i
n
g
m
o
d
els.
On
t
h
e
o
n
e
h
a
n
d
,
tr
a
d
itio
n
al
d
iag
n
o
s
is
with
b
eh
av
io
r
al
ass
ess
m
en
t
is
tim
e
-
co
n
s
u
m
in
g
an
d
r
eq
u
i
r
es
ca
r
ef
u
l
o
b
s
er
v
atio
n
b
y
ex
p
e
r
ts
.
E
f
f
ec
t
iv
e
ass
es
s
m
en
ts
will
r
eq
u
ir
e
ex
p
er
ts
with
m
an
y
y
ea
r
s
o
f
ex
p
er
ien
ce
an
d
en
o
u
g
h
tim
e
to
o
b
s
er
v
e
ch
ild
r
e
n
,
n
o
t
t
o
m
en
ti
o
n
t
h
at
s
u
ch
ev
al
u
atio
n
s
will
b
e
s
u
b
jectiv
e.
O
n
th
e
o
th
er
h
an
d
,
th
e
u
s
e
o
f
d
ee
p
lear
n
in
g
ap
p
r
o
ac
h
es
h
as
b
ee
n
d
em
o
n
s
tr
ated
in
m
an
y
s
tu
d
ies
an
d
im
p
lem
e
n
tatio
n
s
f
o
r
im
ag
e
r
ec
o
g
n
itio
n
p
r
o
b
lem
s
in
th
e
m
ed
ical
f
ield
.
I
n
p
r
ac
tice,
th
e
a
p
p
licatio
n
o
f
d
ee
p
lear
n
in
g
m
et
h
o
d
s
will
b
e
a
way
to
s
cr
ee
n
q
u
ic
k
ly
,
o
b
jectiv
ely
,
an
d
ea
s
ily
ex
p
a
n
d
.
R
ec
o
g
n
izin
g
a
u
tis
m
f
r
o
m
f
ac
i
al
im
ag
es
is
a
co
m
p
lex
p
atter
n
r
ec
o
g
n
itio
n
task
in
v
o
lv
in
g
s
u
b
tle
f
ac
ial
cu
es.
T
h
at
is
b
ec
au
s
e
th
e
f
ac
i
al
ex
p
r
ess
io
n
s
o
f
au
tis
tic
ch
il
d
r
en
ar
e
n
o
t
alwa
y
s
o
b
v
io
u
s
.
T
h
ey
ca
n
ap
p
ea
r
in
s
m
all
ch
an
g
es in
ey
e
g
az
e,
f
ac
ial
ex
p
r
ess
io
n
,
o
r
th
e
co
r
r
elati
o
n
b
etwe
en
f
ac
ial
f
ea
tu
r
es.
T
h
is
r
eq
u
ir
es a
r
tific
ial
in
tellig
en
ce
m
o
d
els
to
b
e
ab
le
to
ex
p
lo
it
d
ee
p
f
ea
tu
r
es
in
s
tead
o
f
r
ely
in
g
s
o
lely
o
n
s
u
r
f
ac
e
v
is
u
al
cu
es.
I
n
ad
d
itio
n
,
ar
tific
ial
in
tellig
en
ce
m
o
d
els
n
ee
d
to
m
ee
t
s
ev
er
al
im
p
o
r
tan
t
r
eq
u
ir
em
e
n
ts
in
th
e
m
ed
ical
co
n
tex
t.
T
h
e
f
ir
s
t
is
g
e
n
er
aliza
b
ilit
y
.
Acc
o
r
d
in
g
ly
,
th
e
m
o
d
el
m
u
s
t
n
o
t
o
n
ly
p
er
f
o
r
m
well
o
n
tr
ain
in
g
d
ata
b
u
t
also
m
ain
tain
ac
ce
p
tab
le
p
er
f
o
r
m
an
ce
wh
en
ap
p
lied
to
n
ew
d
ata
f
r
o
m
d
if
f
er
e
n
t
s
o
u
r
ce
s
.
T
h
e
s
ec
o
n
d
i
s
in
ter
p
r
etab
ilit
y
.
Sp
ec
if
ically
,
p
r
ed
ictio
n
s
m
u
s
t
b
e
ac
co
m
p
a
n
ied
b
y
clea
r
e
x
p
lan
atio
n
s
,
s
u
ch
as
h
ea
t
m
ap
s
d
ep
ictin
g
atten
tio
n
to
h
el
p
d
o
cto
r
s
u
n
d
er
s
tan
d
wh
ich
r
eg
io
n
s
o
f
th
e
im
ag
e
th
e
m
o
d
el
f
o
c
u
s
es
o
n
.
T
h
e
th
ir
d
is
d
ata
-
ef
f
icien
cy
.
Sin
ce
m
ed
ical
d
ata,
esp
ec
ially
d
ata
o
n
ch
ild
r
en
with
au
tis
m
,
ar
e
o
f
ten
s
ca
r
ce
an
d
d
if
f
icu
lt
to
co
llect,
th
e
m
o
d
el
m
u
s
t
lear
n
well
ev
en
with
a
lim
ited
n
u
m
b
er
o
f
tr
ai
n
in
g
s
am
p
les.
T
h
is
ca
n
b
e
d
o
n
e
t
h
r
o
u
g
h
tr
an
s
f
er
lear
n
in
g
,
au
g
m
en
tatio
n
,
o
r
s
em
i
-
s
u
p
e
r
v
is
ed
lear
n
i
n
g
.
T
o
ex
tr
ac
t
r
ich
e
r
f
ea
t
u
r
es
f
r
o
m
lim
ited
d
ata,
o
t
h
er
a
p
p
r
o
ac
h
es
b
esid
es
tr
ad
itio
n
al
tr
an
s
f
e
r
lear
n
in
g
ca
n
also
h
elp
ex
p
lo
it
th
e
lim
ited
d
ata
m
o
r
e
e
f
f
ec
tiv
ely
in
m
ed
ical
ap
p
licatio
n
s
.
On
e
ex
am
p
le
is
m
u
lti
-
lay
e
r
f
in
e
-
tu
n
in
g
,
wh
ic
h
h
elp
s
c
o
n
v
o
lu
tio
n
al
n
e
u
r
al
n
etwo
r
k
(
C
NN)
m
o
d
els
lear
n
f
ea
tu
r
es
t
h
at
ar
e
m
o
r
e
s
u
ited
to
s
u
b
tle
f
ac
ial
cu
es.
An
o
th
er
e
x
am
p
le
is
m
u
lti
-
task
lear
n
in
g
,
wh
ich
allo
ws
th
e
co
m
b
i
n
atio
n
o
f
r
elate
d
task
s
,
s
u
ch
as
lan
d
m
ar
k
l
o
ca
lizatio
n
o
r
e
m
o
tio
n
p
r
ed
ictio
n
,
t
o
s
h
ar
e
a
co
m
m
o
n
r
ep
r
esen
tatio
n
.
I
n
a
d
d
itio
n
,
s
elf
-
s
u
p
er
v
is
ed
lear
n
in
g
lev
e
r
ag
es
u
n
lab
eled
d
ata
to
lear
n
f
ac
ial
s
tr
u
ctu
r
es
b
ef
o
r
e
cl
ass
if
icatio
n
.
T
h
ese
s
tr
ateg
ies p
r
o
m
is
e
to
p
r
o
v
id
e
m
o
r
e
ef
f
ec
tiv
e
f
ea
tu
r
es f
o
r
au
t
is
m
r
ec
o
g
n
itio
n
.
T
h
e
to
p
ic
o
f
a
u
tis
m
d
iag
n
o
s
is
h
as
attr
ac
ted
th
e
atten
tio
n
o
f
m
an
y
r
esear
ch
e
r
s
in
th
e
f
ield
o
f
ar
tific
ial
in
tellig
en
ce
.
I
n
2
0
2
2
,
a
s
tu
d
y
ev
alu
atin
g
ey
e
b
eh
av
io
r
s
was p
er
f
o
r
m
ed
in
[
7
]
f
o
r
au
tis
m
d
i
ag
n
o
s
is
.
T
h
e
s
tu
d
ies
wer
e
p
er
f
o
r
m
ed
with
th
e
co
n
s
tr
u
ctio
n
o
f
v
ar
io
u
s
task
s
,
with
p
ath
co
m
p
u
tatio
n
a
n
d
r
ec
o
g
n
itio
n
m
o
d
el
d
esig
n
.
Var
io
u
s
tech
n
iq
u
es
wer
e
test
ed
s
u
ch
as
R
es
N
et1
8
an
d
i
n
ce
p
tio
n
C
NN
as
well
as
im
ag
e
tr
an
s
f
o
r
m
atio
n
tech
n
iq
u
es
with
g
r
a
y
lev
el
co
-
o
cc
u
r
r
e
n
ce
m
atr
i
x
a
n
d
lo
ca
l
b
in
ar
y
p
atter
n
(
L
B
P)
.
A
s
tu
d
y
o
n
a
u
tis
m
d
etec
tio
n
on
m
a
g
n
etic
r
eso
n
an
ce
im
ag
i
n
g
(
MRI
)
d
ata
was
p
r
esen
ted
by
Hein
s
f
eld
et
a
l.
[
8
]
.
T
h
e
a
u
th
o
r
s
p
r
esen
ted
an
ar
ch
itectu
r
e
co
n
s
is
tin
g
o
f
two
co
n
v
o
lu
tio
n
al
e
n
co
d
e
r
s
an
d
t
ested
it
m
u
ltip
le
tim
es
u
s
in
g
cr
o
s
s
-
v
alid
atio
n
.
I
n
2
0
2
3
,
Far
o
o
q
et
a
l
.
[
9
]
p
u
b
lis
h
ed
th
eir
wo
r
k
ab
o
u
t
au
tis
tic
d
iag
n
o
s
is
with
f
ed
er
ated
lear
n
in
g
m
eth
o
d
.
T
h
e
y
co
m
b
in
ed
s
u
p
p
o
r
t
v
ec
to
r
m
a
ch
in
e
(
SVM)
an
d
lo
g
is
tic
r
e
g
r
ess
io
n
(
L
R
)
to
ex
p
er
im
en
t
o
n
tab
u
la
r
d
ata
an
d
r
ea
ch
ed
0
.
9
8
ac
c
u
r
ac
y
.
An
au
t
is
tic
class
if
ica
tio
n
s
tu
d
y
was
p
u
b
lis
h
ed
[
1
0
]
.
T
h
e
au
t
h
o
r
s
u
s
ed
a
f
r
am
ewo
r
k
f
o
r
ev
alu
atin
g
8
m
ac
h
in
e
lear
n
in
g
alg
o
r
ith
m
s
.
T
h
e
r
esu
lts
wer
e
p
er
f
o
r
m
ed
o
n
f
o
u
r
au
tis
tic
d
atasets
,
in
clu
d
in
g
to
d
d
ler
s
,
ad
o
lescen
ts
,
ch
ild
r
e
n
,
an
d
ad
u
lts
,
an
d
ev
alu
ate
d
with
v
ar
i
o
u
s
s
tatis
tical
ev
alu
atio
n
m
ea
s
u
r
es.
I
n
2
0
2
2
,
Kar
r
i
et
a
l
.
[
1
1
]
p
r
esen
ted
a
wo
r
k
u
s
in
g
f
ac
ial
im
a
g
es.
T
h
eir
wo
r
k
u
s
ed
Den
s
e
Net
f
o
r
id
en
tif
y
in
g
ASD
an
d
was
test
ed
o
n
a
f
ac
e
d
ata
s
et
o
n
th
e
Kag
g
le
p
latf
o
r
m
.
T
h
ey
also
b
u
ilt
a
s
im
p
le
web
to
o
l
to
s
u
p
p
o
r
t
t
h
e
m
ed
ical
f
ac
ilit
ies
.
I
n
2
0
2
3
,
G
h
az
al
et
a
l.
[
1
2
]
p
r
esen
ted
r
esear
ch
o
n
d
esig
n
in
g
a
C
NN
th
at
was
in
s
p
ir
ed
b
y
Alex
N
et
ar
ch
itectu
r
e.
T
h
ey
u
s
ed
in
p
u
t
as
f
ac
ial
im
ag
e
d
ata
a
n
d
tr
ied
to
ex
tr
ac
t
f
ac
ial
f
ea
tu
r
es
ef
f
ec
tiv
ely
.
T
h
e
au
th
o
r
s
ac
h
iev
e
d
8
7
.
6
% v
alid
a
tio
n
s
en
s
itiv
ity
,
8
7
.
6
% v
alid
ati
o
n
s
p
ec
if
icity
,
an
d
8
7
.
7
% v
ali
d
atio
n
ac
cu
r
ac
y
.
I
n
2
0
2
3
,
L
i
et
a
l
.
[
1
3
]
co
n
d
u
c
ted
a
s
tu
d
y
u
s
in
g
Mo
b
ileNetv
3
-
L
ar
g
e
an
d
Mo
b
ileNet
-
V2
to
d
iag
n
o
s
e
au
tis
m
b
ased
o
n
f
ac
ial
c
h
ild
im
ag
es.
T
h
e
a
u
th
o
r
s
d
esi
g
n
ed
a
f
r
am
ewo
r
k
u
s
in
g
t
r
a
n
s
f
er
lear
n
in
g
an
d
in
teg
r
atin
g
d
if
f
er
e
n
t
class
if
ier
s
.
I
n
r
esu
lts
,
th
eir
wo
r
k
ac
h
ie
v
ed
8
7
.
6
7
%
ac
cu
r
ac
y
e
v
alu
atin
g
Mo
b
ileNet
-
V3
-
L
ar
g
e
an
d
8
8
.
3
3
%
ac
c
u
r
ac
y
e
v
alu
atin
g
Mo
b
ileNet
-
V2
.
I
n
2
0
2
4
,
A
h
m
ad
et
a
l.
[
1
4
]
p
r
esen
ted
a
s
tu
d
y
t
o
d
etec
t
au
tis
m
f
r
o
m
f
ac
ial
im
a
g
es
u
s
in
g
m
an
y
m
o
d
els
as
v
is
u
al
g
eo
m
etr
y
g
r
o
u
p
(
VGG
)
1
6
,
VGG1
9
,
Mo
b
ileNet
-
V2
,
Alex
N
et,
R
esNe
t
-
3
4
,
an
d
R
esNet5
0
.
T
h
ey
u
s
ed
ap
p
r
o
x
im
at
ely
2
h
o
u
r
s
f
o
r
tr
ain
in
g
a
n
d
n
ea
r
ly
3
m
in
u
tes
f
o
r
test
in
g
.
T
h
ey
e
v
alu
ated
s
ev
er
a
l
r
eso
lu
tio
n
s
o
f
th
e
i
n
p
u
t
im
a
g
e
an
d
ac
h
iev
ed
th
e
h
ig
h
est
ac
c
u
r
ac
y
o
f
0
.
8
6
with
2
4
8
×2
4
8
.
I
n
2
0
2
4
,
R
ed
d
y
a
n
d
An
d
r
ew
[
1
5
]
c
o
n
d
u
cted
a
d
ee
p
lear
n
in
g
s
tu
d
y
to
class
if
y
au
tis
m
.
W
ith
a
tr
an
s
f
er
lear
n
in
g
ap
p
r
o
ac
h
,
th
r
ee
p
r
et
r
ain
ed
m
o
d
els,
in
clu
d
in
g
E
f
f
icien
t
N
etB
0
,
VGG1
6
,
an
d
VGG1
9
,
wer
e
ex
p
er
im
en
ted
with
.
I
n
th
e
r
esu
lts
,
th
e
au
th
o
r
s
r
ea
ch
e
d
th
e
h
ig
h
est
ac
cu
r
ac
y
is
0
.
8
7
9
f
o
r
th
e
E
f
f
icien
t
N
etB
0
m
o
d
el
i
n
th
eir
ex
p
er
im
en
t.
A
co
m
p
a
r
ativ
e
tab
le
s
u
m
m
a
r
izin
g
e
x
is
tin
g
m
eth
o
d
s
,
d
at
asets
,
au
g
m
en
tatio
n
s
tr
ateg
ies,
an
d
p
er
f
o
r
m
an
ce
m
etr
ics
is
p
r
esen
ted
in
T
ab
le
1
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J
Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
6
:
766
-
7
7
9
768
T
ab
le
1
.
T
h
e
co
m
p
ar
ativ
e
tab
l
e
s
u
m
m
ar
izin
g
e
x
is
tin
g
s
tu
d
ies
S
t
u
d
y
M
e
t
h
o
d
/
t
e
c
h
n
i
q
u
e
D
a
t
a
s
e
t
s
A
u
g
m
e
n
t
a
t
i
o
n
P
e
r
f
o
r
ma
n
c
e
me
t
r
i
c
s
A
h
me
d
e
t
a
l
.
[
7
]
,
2
0
2
2
LB
P
,
g
r
e
y
l
e
v
e
l
c
o
-
o
c
c
u
r
r
e
n
c
e
m
a
t
r
i
x
,
S
V
M
,
G
o
o
g
l
e
-
N
e
t
,
R
e
sN
e
t
-
18
F
i
g
s
h
a
r
e
d
a
t
a
(
e
y
e
-
t
r
a
c
k
i
n
g
sc
a
n
p
a
t
h
s)
F
l
i
p
p
i
n
g
,
mu
l
t
i
-
a
n
g
l
e
r
o
t
a
t
i
o
n
,
d
i
s
p
l
a
c
e
me
n
t
,
a
n
d
s
h
e
a
r
i
n
g
M
o
s
t
o
f
t
h
e
a
c
c
u
r
a
c
y
,
p
r
e
c
i
si
o
n
,
s
e
n
s
i
t
i
v
i
t
y
,
sp
e
c
i
f
i
c
i
t
y
,
a
n
d
A
U
C
a
r
e
mo
r
e
t
h
a
n
0
.
9
3
H
e
i
n
s
f
e
l
d
e
t
a
l
.
[
8
]
,
2
0
1
8
Tw
o
s
t
a
c
k
e
d
d
e
n
o
i
s
i
n
g
a
u
t
o
e
n
c
o
d
e
r
s
,
m
u
l
t
i
l
a
y
e
r
p
e
r
c
e
p
t
r
o
n
A
u
t
i
sm
b
r
a
i
n
i
ma
g
i
n
g
d
a
t
a
e
x
c
h
a
n
g
e
(
A
B
I
D
E
)
N
o
t
s
p
e
c
i
f
i
e
d
A
c
c
u
r
a
c
y
,
s
e
n
s
i
t
i
v
i
t
y
,
a
n
d
s
p
e
c
i
f
i
c
i
t
y
a
r
e
a
r
o
u
n
d
0
.
7
F
a
r
o
o
q
e
t
a
l
.
[
9
]
,
2
0
2
3
F
e
d
e
r
a
t
e
d
l
e
a
r
n
i
n
g
,
LR
,
S
V
M
F
o
u
r
d
a
t
a
s
e
t
s
a
c
c
o
r
d
i
n
g
t
o
t
h
e
q
u
a
n
t
i
t
a
t
i
v
e
c
h
e
c
k
l
i
s
t
f
o
r
a
u
t
i
sm
i
n
c
h
i
l
d
r
e
n
N
o
t
s
p
e
c
i
f
i
e
d
A
c
c
u
r
a
c
y
f
o
r
c
h
i
l
d
r
e
n
i
s
a
r
o
u
n
d
0
.
9
8
,
a
n
d
a
c
c
u
r
a
c
y
f
o
r
a
d
u
l
t
s
i
s
a
r
o
u
n
d
0
.
8
H
a
san
e
t
a
l
.
[
1
0
]
,
2
0
2
3
A
d
a
B
o
o
st
,
r
a
n
d
o
m
f
o
r
e
st
,
d
e
c
i
si
o
n
t
r
e
e
,
k
-
n
e
a
r
e
st
n
e
i
g
h
b
o
r
s,
g
a
u
ssi
a
n
n
a
i
v
e
B
a
y
e
s,
LR
,
S
V
M
,
a
n
d
l
i
n
e
a
r
d
i
s
c
r
i
mi
n
a
n
t
a
n
a
l
y
si
s
F
o
u
r
d
a
t
a
s
e
t
s
f
r
o
m
K
a
g
g
l
e
a
n
d
U
C
I
M
LA
N
o
t
s
p
e
c
i
f
i
e
d
M
o
s
t
o
f
t
h
e
a
c
c
u
r
a
c
y
,
p
r
e
c
i
s
i
o
n
,
r
e
c
a
l
l
,
F
1
-
s
c
o
r
e
,
a
n
d
A
U
C
a
r
e
m
o
r
e
t
h
a
n
0
.
9
6
Li
e
t
a
l
.
[
1
3
]
,
2
0
2
3
M
o
b
i
l
e
N
e
t
V
2
a
n
d
M
o
b
i
l
e
N
e
t
v
3
-
La
r
g
e
A
d
a
t
a
s
e
t
f
r
o
m
K
a
g
g
l
e
N
o
t
s
p
e
c
i
f
i
e
d
M
o
s
t
o
f
t
h
e
a
c
c
u
r
a
c
y
,
sen
s
i
t
i
v
i
t
y
,
sp
e
c
i
f
i
c
i
t
y
,
a
n
d
A
U
C
a
r
e
mo
r
e
t
h
a
n
0
.
9
0
A
h
m
a
d
e
t
a
l
.
[
1
4
]
,
2
0
2
4
R
e
sN
e
t
3
4
,
R
e
sN
e
t
5
0
,
A
l
e
x
N
e
t
,
M
o
b
i
l
e
N
e
t
V
2
,
V
G
G
1
6
,
a
n
d
V
G
G
1
9
A
d
a
t
a
s
e
t
f
r
o
m
K
a
g
g
l
e
F
l
i
p
p
i
n
g
B
e
st
a
c
c
u
r
a
c
y
i
s
0
.
9
2
i
n
t
h
e
c
a
se
o
f
R
e
sN
e
t
5
0
R
e
d
d
y
a
n
d
A
n
d
r
e
w
[
1
5
]
,
2
0
2
4
V
G
G
1
6
,
V
G
G
1
9
a
n
d
,
Ef
f
i
c
i
e
n
t
N
e
t
B
0
A
d
a
t
a
s
e
t
f
r
o
m
K
a
g
g
l
e
R
o
t
a
t
i
n
g
,
h
o
r
i
z
o
n
t
a
l
f
l
i
p
p
i
n
g
,
z
o
o
m
i
n
g
,
a
n
d
h
e
i
g
h
t
a
n
d
w
i
d
t
h
s
h
i
f
t
i
n
g
A
c
c
u
r
a
c
i
e
s
a
r
e
a
r
o
u
n
d
0
.
8
7
,
a
n
d
A
U
C
s
a
r
e
a
r
o
u
n
d
0
.
9
3
Ou
r
r
esear
ch
is
a
b
o
u
t
d
iag
n
o
s
in
g
au
tis
m
in
c
h
ild
r
en
with
a
d
ee
p
lear
n
in
g
ap
p
r
o
ac
h
.
T
h
is
wo
r
k
a
r
e
in
ter
ested
in
u
s
in
g
c
h
ild
r
en
'
s
f
ac
ial
im
ag
e
d
ata
b
y
ex
p
lo
itin
g
f
ac
ial
ex
p
r
ess
io
n
c
h
ar
ac
ter
is
tics
.
I
t
is
k
n
o
wn
t
h
at
o
n
e
o
f
th
e
ch
ar
ac
ter
is
tics
o
f
au
tis
m
is
ab
n
o
r
m
al
f
ac
ial
ex
p
r
ess
io
n
s
in
ch
ild
r
e
n
,
s
u
c
h
as
ab
n
o
r
m
al
s
ig
n
s
o
f
f
ac
ial
asy
m
m
etr
y
o
r
ab
n
o
r
m
a
l
f
ac
ial
d
e
v
elo
p
m
e
n
t
d
u
e
t
o
t
h
e
in
f
l
u
en
ce
o
f
n
eu
r
o
lo
g
ical
d
ev
elo
p
m
e
n
t.
B
ased
o
n
th
ese
o
b
s
er
v
atio
n
s
,
th
is
s
tu
d
y
f
o
cu
s
es
o
n
u
s
in
g
an
d
a
n
aly
zin
g
la
n
d
m
ar
k
s
o
n
ch
ild
r
en
'
s
f
ac
es
in
o
u
r
r
esear
ch
u
s
in
g
d
ee
p
lear
n
i
n
g
m
o
d
els f
o
r
d
iag
n
o
s
in
g
au
tis
m
.
M
o
r
e
s
p
e
ci
f
i
c
al
l
y
,
t
o
e
n
h
a
n
c
e
th
e
d
i
v
e
r
s
it
y
o
f
d
at
a
as
w
e
l
l a
s
th
e
p
e
r
f
o
r
m
a
n
c
e
o
f
d
e
e
p
l
e
a
r
n
i
n
g
m
o
d
e
l
s
,
a
n
a
u
g
m
e
n
t
a
t
i
o
n
t
e
c
h
n
i
q
u
e
u
s
i
n
g
f
a
c
i
a
l
la
n
d
m
a
r
k
s
is
p
r
o
p
o
s
e
d
.
O
u
r
a
u
g
m
e
n
t
a
t
i
o
n
d
i
f
f
er
s
f
r
o
m
g
e
o
m
e
t
r
i
c
t
r
a
n
s
f
o
r
m
a
t
i
o
n
s
i
n
s
e
m
a
n
t
i
cs
in
t
h
i
s
p
r
o
b
l
e
m
.
W
h
i
le
c
o
n
v
e
n
ti
o
n
a
l
g
e
o
m
e
t
r
ic
t
r
a
n
s
f
o
r
m
a
t
i
o
n
s
,
s
u
c
h
as
r
o
t
at
i
o
n
,
t
r
a
n
s
l
at
i
o
n
,
s
c
a
l
i
n
g
,
o
r
a
f
f
i
n
e
,
o
n
l
y
g
l
o
b
a
l
l
y
a
f
f
e
ct
t
h
e
e
n
t
ir
e
f
a
c
e
a
n
d
p
r
e
s
e
r
v
e
t
h
e
o
v
er
a
l
l
m
o
r
p
h
o
l
o
g
i
c
a
l
s
t
r
u
c
t
u
r
e
,
o
u
r
a
u
g
m
e
n
t
a
t
i
o
n
f
o
c
u
s
e
s
o
n
l
o
c
a
l
l
y
s
h
i
f
t
i
n
g
l
a
n
d
m
a
r
k
s
o
n
t
h
e
f
a
c
e
a
n
d
w
a
r
p
i
n
g
t
h
e
i
m
a
g
e
.
T
h
i
s
a
p
p
r
o
a
c
h
a
l
l
o
w
s
f
o
r
m
o
r
e
s
o
p
h
i
s
t
ic
a
t
e
d
d
e
f
o
r
m
a
t
i
o
n
s
t
h
a
t
a
cc
o
m
m
o
d
a
t
e
s
m
a
ll
c
h
a
n
g
e
s
i
n
f
a
c
i
a
l
e
x
p
r
es
s
i
o
n
a
n
d
g
e
o
m
e
t
r
i
c
s
t
r
u
c
t
u
r
e
.
As
a
r
es
u
l
t
,
t
h
e
m
o
d
e
l
ca
n
l
e
a
r
n
f
e
a
t
u
r
e
s
m
o
r
e
e
f
f
e
c
ti
v
e
l
y
t
h
an
g
l
o
b
a
l
g
e
o
m
e
t
r
i
c
t
r
a
n
s
f
o
r
m
a
t
i
o
n
s
i
n
t
h
e
p
r
o
b
l
e
m
o
f
a
u
t
i
s
m
r
e
c
o
g
n
i
t
i
o
n
f
r
o
m
f
a
c
i
a
l
i
m
a
g
es
.
I
n
p
r
e
v
i
o
u
s
s
t
u
d
i
es
,
f
a
c
i
a
l
l
a
n
d
m
a
r
k
s
h
a
v
e
a
l
s
o
b
e
e
n
a
p
p
l
i
e
d
t
o
e
x
p
r
e
s
s
i
o
n
a
n
a
l
y
s
is
o
r
t
o
s
u
p
p
o
r
t
m
o
r
p
h
o
l
o
g
i
c
a
l
m
o
d
e
l
i
n
g
,
s
u
c
h
a
s
i
n
[
1
6
]
f
o
r
d
i
s
e
n
t
a
n
g
li
n
g
e
x
p
r
e
s
s
i
o
n
a
n
d
id
e
n
t
i
t
y
,
o
r
[
1
7
]
f
o
r
d
i
a
g
n
o
s
i
n
g
m
a
n
d
i
b
u
l
a
r
d
e
f
o
r
m
i
t
y
.
H
o
w
e
v
e
r
,
i
n
t
e
r
m
s
o
f
i
m
a
g
e
a
u
g
m
e
n
t
a
t
i
o
n
,
c
o
n
v
e
n
t
i
o
n
a
l
i
m
a
g
e
a
u
g
m
e
n
t
a
t
i
o
n
m
et
h
o
d
s
s
u
ch
a
s
r
o
t
at
i
o
n
,
f
l
i
p
,
a
n
d
a
f
f
i
n
e
t
r
a
n
s
f
o
r
m
a
t
i
o
n
h
a
r
d
l
y
u
t
i
l
i
z
e
l
a
n
d
m
a
r
k
s
,
l
e
a
d
i
n
g
t
o
a
l
a
c
k
o
f
c
o
n
n
e
c
t
i
o
n
w
i
t
h
t
h
e
b
io
l
o
g
i
c
a
l
s
t
r
u
c
t
u
r
e
o
f
t
h
e
f
a
c
e
.
T
h
i
s
e
m
p
h
a
s
i
z
e
s
t
h
e
n
o
v
e
l
t
y
o
f
o
u
r
s
t
u
d
y
i
n
e
x
p
l
o
i
t
i
n
g
l
a
n
d
m
a
r
k
d
i
s
p
l
a
c
e
m
e
n
t
a
s
a
n
a
u
g
m
e
n
t
a
t
i
o
n
t
e
c
h
n
i
q
u
e
,
w
h
i
c
h
b
o
t
h
g
e
n
e
r
a
t
e
s
d
i
v
e
r
s
e
d
a
t
a
a
n
d
p
r
es
e
r
v
e
s
t
h
e
s
e
m
a
n
t
ic
s
o
f
f
a
c
ia
l
s
t
r
u
c
t
u
r
es
r
e
l
a
t
e
d
t
o
e
x
p
r
es
s
i
o
n
s
i
n
a
u
ti
s
ti
c
c
h
i
l
d
r
e
n
.
I
n
ad
d
itio
n
,
to
o
b
tain
an
o
b
j
ec
tiv
e
an
d
co
m
p
r
eh
e
n
s
iv
e
ev
alu
atio
n
,
co
m
p
ar
ativ
e
ex
p
er
i
m
en
ts
ar
e
co
n
d
u
cte
d
b
etwe
en
d
if
f
er
e
n
t w
ell
-
k
n
o
wn
m
o
d
els.
T
h
ese
ex
p
er
im
en
ts
an
aly
ze
th
e
ac
cu
r
ac
y
as we
ll a
s
ev
alu
ate
th
e
ab
ilit
y
to
d
ep
l
o
y
a
n
d
e
x
p
a
n
d
.
An
o
th
er
is
s
u
e
o
f
c
o
n
ce
r
n
i
s
to
ev
alu
ate
t
h
e
r
e
g
io
n
s
o
f
in
t
er
est
o
f
th
e
m
o
d
els
o
n
th
e
i
n
p
u
t
im
a
g
e
u
s
in
g
th
e
g
r
ad
ien
t
-
weig
h
ted
class
ac
tiv
atio
n
m
ap
p
i
n
g
(
Gr
ad
-
C
AM
)
te
ch
n
iq
u
e
i
n
r
elatio
n
to
m
ea
n
in
g
f
u
l
r
eg
i
o
n
s
o
n
th
e
f
ac
ial
im
ag
e.
T
h
is
way
ca
n
ex
p
lo
it
th
e
r
elatio
n
s
h
ip
b
etwe
en
th
e
m
o
d
el'
s
au
tis
m
r
ec
o
g
n
itio
n
an
d
th
e
f
ea
tu
r
e
lo
ca
tio
n
s
o
n
th
e
f
ac
ial
im
ag
e.
T
h
is
will
b
e
c
lear
ev
id
en
ce
o
f
th
e
r
o
le
o
f
f
ac
ial
ex
p
r
ess
io
n
f
ea
tu
r
es
in
au
tis
m
r
ec
o
g
n
itio
n
an
d
will
b
e
an
i
m
p
o
r
tan
t
b
asis
f
o
r
f
u
r
th
e
r
r
esear
ch
.
I
n
d
etail,
o
u
r
m
ain
co
n
tr
ib
u
tio
n
s
in
clu
d
e:
i)
Pro
p
o
s
e
a
n
o
v
el
f
ac
ial
im
ag
e
au
g
m
en
tatio
n
tec
h
n
iq
u
e
b
ased
o
n
d
is
p
lacin
g
f
ac
ial
lan
d
m
ar
k
s
to
im
p
r
o
v
e
th
e
p
er
f
o
r
m
an
ce
o
f
d
ee
p
lear
n
i
n
g
m
o
d
els.
ii)
C
o
m
p
r
eh
en
s
iv
ely
e
v
alu
ate
a
n
d
clar
if
y
o
u
r
h
y
p
o
th
esis
b
y
co
n
d
u
ctin
g
a
c
o
m
p
ar
ativ
e
s
tu
d
y
with
E
f
f
icien
tNet
-
B
0
,
E
f
f
icien
tNet
-
B
4
,
R
esNet
-
1
8
,
R
e
s
Net
-
5
0
,
R
esNet
-
1
0
1
,
Mo
b
ileNet
-
V2
,
Den
s
eNe
t
-
1
2
1
,
an
d
Den
s
eNe
t
-
2
0
1
.
iii)
An
aly
ze
th
e
in
ter
p
r
etab
ilit
y
o
f
m
o
d
els b
y
v
is
u
alizin
g
m
o
d
el
atten
tio
n
with
Gr
ad
-
C
AM
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
I
mp
r
o
vin
g
efficien
cy
o
f
a
u
tis
m
d
etec
tio
n
b
a
s
ed
o
n
f
a
cia
l ima
g
e
la
n
d
ma
r
ks
(
N
g
u
ye
n
Tr
o
n
g
Tu
n
g
)
769
2.
M
E
T
H
O
D
T
h
is
s
ec
tio
n
will
p
r
esen
t
th
e
s
p
ec
if
ic
co
n
te
n
ts
p
r
o
p
o
s
ed
in
o
u
r
r
esear
ch
.
B
asically
,
f
ac
ial
i
m
ag
e
d
ata
will
b
e
au
g
m
e
n
ted
with
a
f
o
cu
s
o
n
f
ac
ial
im
ag
e
au
g
m
en
tatio
n
tech
n
iq
u
es
b
ased
o
n
f
ac
ial
lan
d
m
ar
k
d
is
p
lace
m
en
t
an
d
f
ac
ial
im
a
g
e
war
p
i
n
g
.
B
esid
es,
th
e
f
a
m
o
u
s
d
ee
p
lear
n
in
g
m
o
d
els
ar
e
p
r
esen
ted
f
o
r
co
m
p
r
eh
e
n
s
iv
e
a
n
d
co
m
p
a
r
ativ
e
test
in
g
.
T
o
clar
if
y
th
eir
ef
f
ec
tiv
en
ess
,
ex
p
er
im
en
ts
w
ith
th
r
ee
d
if
f
e
r
en
t
au
g
m
en
tatio
n
s
tr
ateg
ies
ar
e
p
r
o
p
o
s
ed
,
a
n
d
th
e
p
r
ed
ictio
n
r
esu
lts
ar
e
an
aly
ze
d
in
r
elatio
n
to
f
ac
ial
r
eg
io
n
s
u
s
in
g
th
e
Gr
ad
-
C
AM
tech
n
iq
u
e
[
1
8
]
.
2
.
1
.
Aut
is
m
s
pect
rum
dis
o
rder
det
ec
t
io
n da
t
a
s
et
T
h
is
s
tu
d
y
co
n
d
u
cts
ex
p
er
im
e
n
ts
u
s
in
g
a
ch
ild
f
ac
e
im
a
g
e
d
ataset
p
u
b
lis
h
ed
o
n
th
e
Ka
g
g
l
e
p
latf
o
r
m
at
h
ttp
s
://www.
k
ag
g
le.
co
m
.
T
h
er
e
ar
e
to
tal
o
f
2
,
9
3
6
ch
ild
f
a
ce
im
ag
es in
th
is
d
ataset
an
d
th
ey
ar
e
d
iv
i
d
ed
in
to
two
g
r
o
u
p
s
,
in
clu
d
in
g
au
tis
tic
an
d
n
o
n
-
au
tis
tic.
Mo
r
e
s
p
ec
if
i
ca
lly
,
th
e
n
u
m
b
er
o
f
im
ag
es o
f
au
tis
tic
ch
ild
r
en
is
1
,
4
6
8
im
a
g
es
an
d
th
e
n
u
m
b
er
o
f
im
ag
es
o
f
n
o
n
-
au
tis
tic
ch
ild
r
en
is
1
,
4
6
8
im
a
g
es.
T
h
is
d
ataset
was
alr
ea
d
y
s
p
lit
in
to
th
r
ee
s
u
b
s
ets
a
s
th
e
tr
ain
s
et,
th
e
v
alid
atio
n
s
et,
an
d
th
e
test
s
et.
I
n
d
etail,
th
e
tr
ain
s
et
in
clu
d
es
1
,
2
6
8
im
a
g
es
o
f
n
o
n
-
au
tis
tic
ch
ild
r
en
an
d
1
,
2
6
8
im
ag
es
o
f
au
tis
tic
ch
ild
r
en
.
T
h
e
v
alid
atio
n
s
et
in
clu
d
es
5
0
im
ag
es
o
f
n
o
n
-
au
tis
tic
ch
ild
r
e
n
an
d
5
0
im
ag
es
o
f
au
tis
tic
ch
ild
r
en
.
T
h
e
test
s
et
in
clu
d
es
1
5
0
im
ag
es
o
f
n
o
n
-
au
tis
tic
ch
ild
r
en
an
d
1
5
0
im
ag
es o
f
au
tis
tic
ch
ild
r
en
.
2
.
2
.
P
r
o
po
s
ed
f
a
cia
l ima
g
e
a
ug
m
ent
a
t
io
n
Fo
r
tr
ain
in
g
d
ee
p
lear
n
in
g
m
o
d
els,
d
ata
au
g
m
en
tatio
n
p
lay
s
an
im
p
o
r
tan
t
r
o
le
in
d
ea
lin
g
with
d
at
a
s
ca
r
city
.
Def
au
lt
im
a
g
e
au
g
m
en
tatio
n
m
et
h
o
d
s
o
f
ten
in
c
lu
d
e
o
p
er
atio
n
s
s
u
c
h
as
f
lip
p
in
g
,
r
o
tatin
g
,
an
d
r
an
d
o
m
cr
o
p
p
i
n
g
.
T
h
ese
tech
n
iq
u
es
p
lay
an
im
p
o
r
tan
t
r
o
le
in
en
h
an
cin
g
th
e
g
en
e
r
aliza
tio
n
ab
ilit
y
o
f
d
ee
p
lear
n
in
g
m
o
d
els.
Ho
wev
e
r
,
f
o
r
f
ac
ial
im
a
g
e
d
ata,
d
ef
a
u
lt
m
eth
o
d
s
d
o
n
o
t
tak
e
a
d
v
an
ta
g
e
o
f
th
e
s
tr
u
ctu
r
al
f
ea
tu
r
es
o
f
th
e
f
ac
e.
T
h
is
s
tu
d
y
p
r
o
p
o
s
es
to
au
g
m
en
t
f
a
cial
im
ag
e
d
ata
b
ased
o
n
m
an
ip
u
latin
g
f
ac
ial
lan
d
m
ar
k
s
an
d
,
o
n
th
at
b
asis
,
g
en
er
ate
n
ew
f
ac
ial
im
ag
e
d
ata,
wh
ich
is
d
escr
ib
ed
in
Alg
o
r
ith
m
1
.
T
h
is
m
eth
o
d
h
elp
s
to
cr
ea
te
f
ac
ial
im
ag
e
v
ar
iatio
n
s
b
ased
o
n
th
e
s
tr
u
ctu
r
e
o
f
f
ac
ial
im
ag
e
d
ata.
T
h
u
s
,
it
cr
ea
tes
n
ew
d
at
a
s
am
p
les th
at
ar
e
co
n
s
is
ten
t w
ith
th
e
n
atu
r
e
o
f
f
ac
ial
im
ag
es.
Alg
o
r
ith
m
1
.
T
h
e
lan
d
m
ar
k
d
i
s
p
lace
m
en
t a
u
g
m
en
tatio
n
I
n
p
u
t: f
ac
e
im
ag
e
I
Ou
tp
u
t: a
u
g
m
e
n
ted
im
ag
e
I
'
Pro
ce
s
s
:
1
:
F=d
etec
t_
f
ac
e_
b
b
o
x
(
I
)
2:
L
=d
etec
t_
lan
d
m
ar
k
s
(
F,
I
)
3:
W
=c
alcu
late_
f
ac
e_
with
(
F)
4:
L
'
=d
is
p
lace
_
lan
d
m
ar
k
s
(
L
,
W
,
I
,
MA
X_
SHI
FT_
R
AT
I
O)
5:
tr
is
=d
elau
n
ay
_
tr
ian
g
u
latio
n
(
L
)
6
:
f
o
r
ea
ch
t in
tr
is
:
7:
p
1
=g
et_
v
er
tices(L
,
t)
8:
p
2
=g
et_
v
er
tices(L
'
,
t)
9:
T
=c
o
m
p
u
te_
t
r
an
s
f
o
r
m
(
p
1
,
p
2
)
10:
war
p
_
tr
ian
g
u
lar
_
r
e
g
io
n
(
I
,
I
'
,
p
1
,
p
2
,
T
)
T
h
e
p
r
o
ce
s
s
is
d
escr
ib
ed
in
Fig
u
r
e
1
.
W
ith
th
e
in
p
u
t
b
ein
g
a
f
ac
e
im
ag
e,
t
h
e
f
ir
s
t
s
tep
i
s
to
d
etec
t
f
ac
ial
lan
d
m
ar
k
s
.
T
h
ese
ar
e
th
e
p
o
in
ts
th
at
p
lay
an
im
p
o
r
tan
t
r
o
le
in
f
ac
ial
m
o
r
p
h
o
lo
g
y
s
u
ch
as
ey
e
co
r
n
er
s
an
d
n
o
s
e
p
o
in
ts
.
T
h
is
wo
r
k
u
s
es
a
s
et
o
f
6
8
f
ac
ial
p
o
in
ts
s
u
p
p
o
r
ted
in
t
h
e
Dlib
lib
r
ar
y
at
h
ttp
s
://d
lib
.
n
et/.
T
h
ese
lan
d
m
ar
k
s
a
r
e
th
e
b
asi
s
f
o
r
war
p
in
g
f
ac
ial
im
ag
es.
Fo
r
th
e
war
p
in
g
to
b
e
p
er
f
o
r
m
ed
,
th
e
s
et
o
f
f
ac
ial
lan
d
m
ar
k
s
is
tr
ian
g
u
lated
u
s
in
g
th
e
Dela
u
n
ay
tr
ian
g
u
latio
n
tech
n
iq
u
e
[
1
9
]
.
T
h
u
s
,
th
e
war
p
in
g
will
b
e
p
er
f
o
r
m
ed
b
y
in
ter
p
o
latin
g
t
h
e
p
ix
el
v
alu
es
in
ea
ch
s
u
b
-
tr
ian
g
le
o
f
th
e
r
esu
ltin
g
im
ag
e
b
ased
o
n
th
e
co
r
r
esp
o
n
d
in
g
p
o
s
itio
n
s
f
o
r
th
e
th
r
ee
v
er
tices
o
f
th
e
tr
ian
g
le.
T
h
is
ex
p
er
im
en
ts
also
h
av
e
th
e
o
p
tio
n
to
p
r
e
-
ca
lcu
late
th
e
tr
ian
g
u
latio
n
s
et
to
u
s
e
f
o
r
th
e
im
ag
es with
o
u
t h
av
in
g
to
r
ec
alcu
late
ea
ch
tim
e.
T
h
e
lan
d
m
ar
k
d
is
p
lace
m
en
t
a
u
g
m
en
tatio
n
m
eth
o
d
is
p
er
f
o
r
m
ed
b
y
d
etec
tin
g
6
8
f
ac
ial
lan
d
m
ar
k
s
u
s
in
g
th
e
Dlib
lib
r
ar
y
.
T
h
ese
p
o
in
ts
ar
e
t
h
en
r
an
d
o
m
ly
d
is
p
lace
d
with
in
a
lim
ited
r
an
g
e
b
ased
o
n
th
e
f
ac
e
wid
th
an
d
th
e
MA
X_
SHI
FT_
R
AT
I
O
s
ca
le
v
alu
e.
T
h
e
d
is
p
lace
m
en
t
r
atio
p
ar
a
m
eter
is
r
an
d
o
m
ly
g
e
n
er
ated
with
an
u
p
p
e
r
b
o
u
n
d
o
f
MA
X
_
SHI
FT_
R
AT
I
O,
an
d
s
p
ec
if
ic
ally
in
t
h
e
ex
p
er
im
en
t
MA
X_
SHI
FT_
R
AT
I
O
is
s
et
to
0
.
2
.
MA
X_
SHI
FT_
R
A
T
I
O
co
n
s
tr
ain
s
th
e
m
ax
im
u
m
d
is
p
lace
m
en
t
o
f
lan
d
m
ar
k
p
o
i
n
ts
o
n
th
e
f
ac
e.
Her
e
th
e
d
is
p
lace
m
en
t
o
f
th
e
lan
d
m
ar
k
is
ca
lc
u
lated
p
r
o
p
o
r
tio
n
ally
to
t
h
e
wid
th
o
f
th
e
f
ac
e,
to
e
n
s
u
r
e
t
h
at
th
e
d
ef
o
r
m
atio
n
is
alwa
y
s
p
r
o
p
o
r
t
io
n
al
to
th
e
s
ize
o
f
th
e
f
ac
e,
a
v
o
id
in
g
ab
n
o
r
m
al
d
ef
o
r
m
atio
n
o
f
th
e
f
ac
e
lead
in
g
to
u
n
r
ea
lis
tic
r
esu
lts
.
Nex
t,
t
h
e
f
ac
ial
r
eg
io
n
in
th
e
o
r
ig
i
n
al
im
ag
e
is
d
iv
id
ed
in
to
s
m
all
tr
ian
g
les
u
s
in
g
Dela
u
n
ay
tr
ian
g
u
latio
n
.
All
tr
ian
g
les
ar
e
iter
ated
th
r
o
u
g
h
an
d
ea
ch
tr
ian
g
le
is
co
m
p
u
ted
b
y
a
n
af
f
in
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J
Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
6
:
766
-
7
7
9
770
tr
an
s
f
o
r
m
atio
n
f
r
o
m
th
e
o
r
ig
i
n
al
lan
d
m
ar
k
lo
ca
tio
n
to
th
e
d
is
p
lace
d
lan
d
m
ar
k
.
B
ased
o
n
th
at,
th
e
r
esu
ltin
g
im
ag
e
is
g
en
er
ated
b
y
war
p
in
g
ea
ch
s
u
b
-
r
eg
io
n
in
t
h
e
o
r
ig
in
al
im
ag
e
ac
co
r
d
in
g
to
th
e
ca
lcu
lated
af
f
in
e
tr
an
s
f
o
r
m
atio
n
.
T
h
e
p
ar
a
m
et
er
s
in
o
u
r
a
u
g
m
en
tatio
n
p
r
o
ce
s
s
ar
e
cu
r
r
en
tly
s
et
b
ased
o
n
in
tu
itio
n
an
d
ex
p
er
ien
ce
.
I
n
th
e
f
u
tu
r
e,
als
o
co
n
s
id
er
ex
p
er
t
ev
alu
atio
n
.
I
n
th
is
way
,
ex
p
er
t
ev
al
u
atio
n
ca
n
v
alid
ate
th
e
r
ea
lis
m
o
f
au
g
m
e
n
ted
im
a
g
es f
o
r
th
e
a
u
tis
m
ass
ess
m
en
t ta
s
k
.
Fig
u
r
e
1
.
Pro
p
o
s
ed
lan
d
m
ar
k
d
is
p
lace
m
en
t
-
b
ased
au
g
m
en
ta
tio
n
p
r
o
ce
s
s
Du
r
in
g
th
e
d
ata
au
g
m
en
tatio
n
iter
atio
n
f
o
r
ea
ch
f
ac
e
im
ag
e,
th
e
lan
d
m
ar
k
s
ar
e
tr
an
s
f
o
r
m
ed
b
y
r
an
d
o
m
l
y
s
h
if
tin
g
th
em
in
a
n
eig
h
b
o
r
h
o
o
d
o
f
th
eir
o
r
ig
in
al
p
o
s
itio
n
s
.
T
h
e
d
is
p
lace
m
en
t
r
an
g
e
is
s
et
p
r
o
p
o
r
tio
n
al
t
o
th
e
h
o
r
iz
o
n
tal
len
g
th
o
f
th
e
f
ac
e
a
n
d
is
s
m
all
en
o
u
g
h
t
h
at
th
e
d
is
p
lace
m
en
t w
ill
n
o
t
d
is
r
u
p
t
t
h
e
s
tr
u
ctu
r
e
o
f
t
h
e
f
ac
e
in
th
e
im
ag
e.
On
ce
th
e
f
ac
ial
lan
d
m
a
r
k
s
h
av
e
b
ee
n
t
r
an
s
f
o
r
m
e
d
,
th
e
n
ew
f
ac
e
im
ag
e
is
war
p
ed
b
y
in
ter
p
o
latin
g
ea
ch
r
eg
io
n
o
f
th
e
id
en
tifie
d
tr
ian
g
l
es.
T
h
e
r
esu
ltin
g
im
ag
e
is
f
ed
in
to
th
e
d
ataset
f
o
r
m
o
d
el
tr
ain
in
g
.
So
m
e
e
x
am
p
le
r
esu
lts
wer
e
d
escr
ib
ed
in
Fig
u
r
e
2
.
Af
ter
th
e
tr
an
s
f
o
r
m
atio
n
u
s
in
g
lan
d
m
ar
k
s
d
is
p
lace
m
en
t,
th
e
im
ag
es
ar
e
f
u
r
th
er
d
iv
er
s
if
ie
d
u
s
in
g
th
e
d
ef
au
lt
au
g
m
e
n
tatio
n
m
eth
o
d
.
T
h
e
d
ef
au
lt
au
g
m
en
tatio
n
m
et
h
o
d
is
d
esig
n
ed
to
ac
co
m
m
o
d
ate
a
wid
e
r
an
g
e
o
f
in
p
u
t
im
ag
es,
n
o
t
j
u
s
t
f
ac
es.
First,
th
e
im
ag
es
ar
e
r
esized
t
o
a
ch
o
s
en
s
tan
d
ar
d
s
ize.
T
h
is
m
ak
es
im
ag
es
o
f
d
if
f
er
en
t sizes co
m
p
atib
le
with
d
ee
p
lear
n
in
g
m
o
d
els.
T
h
is
ca
s
e
will b
r
in
g
th
em
to
2
2
4
×2
2
4
.
Nex
t,
th
e
im
a
g
es
ar
e
r
an
d
o
m
ly
h
o
r
izo
n
tally
f
li
p
p
ed
with
a
s
p
ec
if
ied
p
r
o
b
ab
ilit
y
,
in
th
is
ca
s
e
5
0
%.
T
h
is
tr
an
s
f
o
r
m
atio
n
is
a
p
o
p
u
lar
ch
o
ice
to
h
elp
m
o
d
e
ls
lear
n
r
ef
lectio
n
v
a
r
iatio
n
s
,
an
d
it
is
also
s
u
itab
le
f
o
r
f
ac
es
b
ec
au
s
e
o
f
its
s
y
m
m
etr
y
.
Nex
t,
th
e
im
ag
es
ar
e
r
an
d
o
m
ly
r
o
tated
clo
c
k
wis
e
o
r
co
u
n
ter
clo
c
k
wis
e
with
i
n
a
s
p
ec
if
ied
an
g
le
r
an
g
e,
in
th
is
ca
s
e
1
0
d
e
g
r
ee
s
.
Nex
t,
r
an
d
o
m
ly
tr
an
s
f
o
r
m
th
e
p
ix
el
v
alu
es
in
te
r
m
s
o
f
b
r
ig
h
tn
ess
,
co
n
tr
ast,
s
atu
r
atio
n
,
an
d
h
u
e,
m
a
k
in
g
th
e
d
ataset
r
ich
er
in
ter
m
s
o
f
lig
h
tin
g
c
o
n
d
itio
n
s
.
Fin
a
lly
,
th
e
im
ag
es
ar
e
g
eo
m
etr
ically
tr
a
n
s
f
o
r
m
e
d
af
f
in
ely
with
s
m
all
o
f
f
s
ets.
I
n
th
e
im
p
lem
e
n
tatio
n
,
th
e
d
e
f
au
lt
au
g
m
en
tatio
n
tech
n
iq
u
es we
r
e
p
e
r
f
o
r
m
ed
wi
th
th
e
s
u
p
p
o
r
t o
f
th
e
Ko
r
n
ia
lib
r
ar
y
[
2
0
]
.
T
h
u
s
,
o
u
r
au
g
m
en
tatio
n
m
et
h
o
d
im
p
r
o
v
es
d
ata
ef
f
icie
n
c
y
,
wh
ich
is
a
m
ajo
r
A
I
ch
a
llen
g
e,
b
y
g
en
er
atin
g
m
o
r
e
r
ea
lis
tic
f
ac
ial
v
ar
ian
ts
f
r
o
m
th
e
o
r
ig
i
n
al
d
ata.
T
h
is
is
esp
ec
ially
tr
u
e
wh
en
wo
r
k
in
g
with
lim
ited
f
ac
ial
im
ag
e
d
atasets
.
I
n
s
tead
o
f
r
ely
in
g
o
n
s
im
p
le
g
eo
m
etr
ic
t
r
an
s
f
o
r
m
atio
n
s
,
t
h
is
m
eth
o
d
d
ir
ec
tly
ex
p
lo
its
th
e
f
ac
ial
s
tr
u
ctu
r
e,
th
er
eb
y
lear
n
in
g
m
o
r
e
s
em
an
t
ically
r
elev
an
t
f
ea
tu
r
es.
I
n
ad
d
itio
n
,
f
ac
ial
d
ata
o
f
ten
h
as
p
o
ten
tial
b
iases
in
ter
m
s
o
f
eth
n
icity
,
a
g
e,
o
r
g
en
d
e
r
.
Ou
r
lan
d
m
ar
k
d
is
p
lace
m
en
t m
eth
o
d
allo
ws
f
o
r
alg
o
r
ith
m
ically
c
o
n
s
is
ten
t
g
e
n
er
atio
n
o
f
m
o
d
els
with
r
a
n
d
o
m
d
is
p
lace
m
en
ts
b
ased
o
n
c
o
m
m
o
n
f
ac
ial
lan
d
m
ar
k
s
.
Du
e
to
th
ese
ch
ar
ac
ter
is
tics
,
o
u
r
m
eth
o
d
is
ab
le
to
co
n
tr
ib
u
te
to
r
e
d
u
cin
g
th
e
im
p
ac
t
o
f
p
o
ten
tial
b
iases
in
ter
m
s
o
f
id
ea
s
.
As a
r
esu
lt,
th
e
m
o
d
el
is
ex
p
ec
te
d
to
g
en
er
alize
b
etter
a
n
d
b
e
f
air
er
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
I
mp
r
o
vin
g
efficien
cy
o
f
a
u
tis
m
d
etec
tio
n
b
a
s
ed
o
n
f
a
cia
l ima
g
e
la
n
d
ma
r
ks
(
N
g
u
ye
n
Tr
o
n
g
Tu
n
g
)
771
Fig
u
r
e
2
.
So
m
e
r
esu
lts
o
f
th
e
p
r
o
p
o
s
ed
lan
d
m
a
r
k
d
is
p
lace
m
en
t
-
b
ased
au
g
m
en
tatio
n
p
r
o
ce
s
s
2
.
3
.
Dee
p lea
rning
mo
dels
E
f
f
i
ci
en
t
Ne
t
[
2
1
]
i
s
a
f
am
ily
o
f
C
N
N
ar
ch
i
te
ctu
r
e
s
th
at
s
u
p
p
o
r
t
s
t
h
e
a
b
i
li
ty
t
o
a
llo
w
t
h
e
m
o
d
e
l
t
o
s
c
al
e
ev
en
ly
ac
r
o
s
s
m
u
l
tip
l
e
d
im
en
s
io
n
s
s
u
ch
a
s
r
e
s
o
l
u
tio
n
,
wi
d
th
,
an
d
d
ep
th
.
T
h
e
o
p
ti
m
a
l
u
s
e
o
f
E
f
f
i
ci
en
t
Ne
t
ca
n
h
elp
th
e
p
r
o
g
r
am
s
t
il
l
ac
h
ie
v
e
h
ig
h
p
er
f
o
r
m
an
c
e
wh
il
e
b
e
in
g
ab
le
to
u
s
e
f
ew
er
r
e
s
o
u
r
ce
s
co
m
p
ar
ed
to
s
o
m
e
o
th
er
C
N
N
m
o
d
e
l
s
.
I
n
g
en
er
al
,
E
f
f
ic
ie
n
tN
et
ca
n
wo
r
k
ef
f
ec
t
iv
e
ly
w
ith
co
m
p
lex
im
ag
e
d
ata
an
d
p
r
o
v
id
e
s
g
en
er
a
li
za
t
io
n
ca
p
a
b
i
li
ti
e
s
.
T
h
i
s
ar
ch
i
te
c
tu
r
e
i
s
cu
s
t
o
m
i
ze
d
w
it
h
m
an
y
d
if
f
er
e
n
t
v
er
s
io
n
s
d
ep
en
d
in
g
o
n
th
e
s
ca
le
o
f
t
h
e
m
o
d
e
l.
T
h
i
s
s
tu
d
y
ch
o
s
e
t
wo
v
er
s
i
o
n
s
f
o
r
o
u
r
ex
p
e
r
im
en
t
as
E
f
f
icien
tNet
-
B
0
an
d
E
f
f
icien
tNet
-
B
4
.
R
esNet
[
2
2
]
is
a
well
-
k
n
o
wn
d
ee
p
n
etwo
r
k
ar
c
h
itectu
r
e
in
v
ar
io
u
s
d
ee
p
lear
n
in
g
p
r
o
b
le
m
s
with
a
r
esid
u
al
co
n
n
ec
tio
n
s
m
ec
h
an
is
m
th
at
h
elp
s
to
m
in
im
ize
th
e
p
h
en
o
m
e
n
o
n
o
f
g
r
a
d
ien
t v
an
is
h
in
g
as th
e
d
ep
th
o
f
th
e
n
etwo
r
k
in
cr
ea
s
es.
T
h
is
m
ec
h
an
is
m
is
s
p
ec
ial
in
t
h
at
it
allo
ws
r
esear
ch
er
s
to
tr
ai
n
v
er
y
d
ee
p
n
etwo
r
k
s
wh
ile
m
ain
tain
in
g
s
tab
ilit
y
.
T
h
an
k
s
to
th
at,
th
e
m
o
d
el
is
ca
p
ab
le
o
f
g
en
e
r
alizin
g
m
a
n
y
c
o
m
p
lex
f
ea
tu
r
es
in
im
ag
es.
Sp
ec
if
ic
v
er
s
io
n
s
o
f
R
esNet
ar
e
o
f
ten
n
am
ed
ac
co
r
d
in
g
to
th
e
d
e
p
th
o
f
th
e
a
r
ch
itectu
r
e.
T
h
is
s
tu
d
y
ch
o
o
s
es
3
v
er
s
io
n
s
: Res
Net
-
1
8
,
R
esNet
-
5
0
,
an
d
R
esNet
-
1
0
1
f
o
r
ev
alu
atin
g
th
e
p
r
o
p
o
s
ed
m
eth
o
d
.
Mo
b
ileNet
[
2
3
]
is
a
C
NN
ar
c
h
itectu
r
e
d
esig
n
e
d
f
o
r
u
s
e
in
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
s
ce
n
a
r
io
s
,
s
u
ch
as
m
o
b
ile
d
e
v
ices.
I
t
is
b
u
ilt
wit
h
d
ep
t
h
wis
e
s
ep
ar
ab
le
c
o
n
v
o
lu
tio
n
s
to
m
i
n
im
ize
th
e
n
u
m
b
e
r
o
f
p
ar
am
eter
s
a
n
d
th
e
co
m
p
u
tatio
n
al
b
u
r
d
en
w
h
ile
s
till
p
r
o
v
id
in
g
s
ig
n
if
ica
n
t
p
er
f
o
r
m
an
ce
g
ain
s
in
im
ag
e
class
if
icatio
n
p
r
o
b
lem
s
.
T
h
is
s
tu
d
y
u
s
e
s
Mo
b
ileNet
-
V2
f
o
r
e
x
p
er
im
e
n
ts
.
Similar
to
R
esNet,
Den
s
eNe
t
is
also
a
C
NN
ar
ch
itectu
r
e
d
esig
n
ed
t
o
im
p
r
o
v
e
th
e
p
r
o
p
ag
atio
n
o
f
g
r
ad
ien
t
s
ig
n
als.
I
n
De
n
s
eNe
t,
ea
ch
lay
e
r
is
co
n
n
ec
ted
to
all
p
r
ev
i
o
u
s
lay
er
s
.
T
h
is
n
o
t
o
n
ly
h
elp
s
in
th
e
p
r
o
p
a
g
atio
n
o
f
g
r
ad
ie
n
ts
b
u
t a
ls
o
co
n
tr
ib
u
tes to
th
e
r
eu
s
e
o
f
in
ter
m
ed
iate
f
ea
tu
r
es c
o
m
p
u
te
d
at
d
if
f
er
en
t le
v
els
o
f
ab
s
tr
ac
tio
n
.
T
h
is
s
tu
d
y
u
s
e
s
Den
s
eNe
t
-
1
2
1
an
d
Den
s
eNe
t
-
201.
C
o
m
p
ar
ed
t
o
n
ewe
r
ap
p
r
o
ac
h
es
s
u
ch
as
v
is
io
n
tr
a
n
s
f
o
r
m
er
s
,
s
win
tr
an
s
f
o
r
m
er
s
,
o
r
h
y
b
r
id
C
NN
-
r
ec
u
r
r
en
t
n
eu
r
al
n
etwo
r
k
(
R
NN
)
m
o
d
els,
v
is
io
n
t
r
an
s
f
o
r
m
er
s
h
av
e
th
e
ad
v
a
n
tag
e
o
f
m
o
d
elin
g
g
lo
b
al
r
elatio
n
s
h
ip
s
b
etwe
en
f
ac
ial
r
eg
io
n
s
,
b
u
t
o
f
ten
r
eq
u
ir
e
lar
g
e
d
atasets
f
o
r
ef
f
ec
tiv
e
tr
ain
in
g
,
wh
ich
is
d
if
f
icu
lt
to
m
ee
t
in
th
e
co
n
tex
t
o
f
lim
ited
p
ed
iatr
ic
au
tis
m
d
ata.
Swin
t
r
an
s
f
o
r
m
er
h
as
th
e
ad
v
an
ta
g
e
o
f
in
co
r
p
o
r
atin
g
a
h
ier
ar
ch
ical
s
tr
u
ctu
r
e
an
d
lo
c
al
atten
tio
n
,
b
u
t
co
m
es
with
a
h
ig
h
er
co
m
p
u
tatio
n
al
co
s
t.
I
n
ad
d
itio
n
,
h
y
b
r
id
C
NN
-
R
N
N
m
o
d
els
ar
e
m
ain
ly
s
u
itab
le
f
o
r
v
id
eo
d
ata
wh
en
an
aly
zin
g
f
ac
ial
m
o
tio
n
s
,
wh
ile
th
e
cu
r
r
en
t
s
tu
d
y
f
o
cu
s
es
o
n
s
till
im
ag
es.
T
h
e
r
ef
o
r
e,
t
h
is
wo
r
k
u
s
e
s
well
-
k
n
o
wn
C
NN
m
o
d
els
s
u
ch
a
s
E
f
f
icien
tNet
-
B
0
,
E
f
f
icien
tNet
-
B
4
,
R
esNet
-
1
8
,
R
esNet
-
5
0
,
R
es
Net
-
1
0
1
,
Mo
b
ileNet
-
V2
,
Den
s
eNe
t
-
121
,
a
n
d
Den
s
eNe
t
-
2
0
1
,
wh
ich
h
av
e
b
ee
n
p
r
o
v
en
ef
f
ec
tiv
e
in
ex
tr
ac
tin
g
lo
ca
l
f
ea
tu
r
e
s
f
r
o
m
f
ac
ial
im
ag
es.
T
h
is
is
co
n
s
is
ten
t
with
o
u
r
p
r
o
p
o
s
ed
co
n
tr
i
b
u
tio
n
r
eg
ar
d
in
g
th
e
u
tili
za
tio
n
o
f
f
ac
ial
lan
d
m
ar
k
s
.
T
o
tr
ain
m
o
d
els
ef
f
icien
tly
an
d
to
r
e
d
u
ce
o
v
er
f
itti
n
g
,
s
o
m
e
s
tr
ateg
ies
co
m
m
o
n
ly
u
s
ed
in
d
ee
p
lear
n
in
g
m
o
d
el
tr
ain
in
g
a
r
e
c
o
n
s
id
er
ed
.
O
n
e
e
x
am
p
le
is
ea
r
l
y
s
to
p
p
i
n
g
,
wh
er
e
th
e
tr
ain
in
g
p
r
o
ce
s
s
s
to
p
s
wh
en
th
e
ev
alu
atio
n
m
etr
ic
o
n
th
e
v
alid
atio
n
s
et
n
o
lo
n
g
er
im
p
r
o
v
es.
Oth
er
s
ar
e
d
r
o
p
o
u
t,
o
r
r
e
g
u
lar
izatio
n
,
w
h
ich
h
elp
im
p
r
o
v
e
th
e
g
en
er
aliza
tio
n
o
f
m
o
d
els.
2
.
4
.
Vis
ua
lizing
m
o
del a
t
t
ent
io
n wit
h G
ra
d
-
CAM
Gr
ad
-
C
AM
p
lay
s
an
im
p
o
r
tan
t
r
o
le
in
in
d
icatin
g
th
e
f
o
cu
s
o
f
C
NN
n
etwo
r
k
s
in
im
ag
e
r
e
co
g
n
itio
n
ap
p
licatio
n
s
.
C
o
n
v
o
l
u
tio
n
al
la
y
er
s
d
is
co
v
er
an
d
r
ec
o
r
d
s
p
ati
al
f
ea
tu
r
es
co
m
p
u
ted
f
r
o
m
an
in
p
u
t
im
ag
e.
Fo
r
a
tr
ain
ed
C
NN
m
o
d
el,
th
e
f
ir
s
t
co
n
v
o
l
u
tio
n
al
lay
er
s
will
p
lay
th
e
r
o
le
o
f
an
aly
zin
g
th
e
b
asic
f
ea
tu
r
es
o
f
th
e
im
ag
e
wh
ile
th
e
last
c
o
n
v
o
lu
tio
n
al
lay
e
r
s
will
m
o
d
el
th
e
s
e
m
an
tic
f
ea
tu
r
es.
T
h
u
s
,
th
e
last
lay
er
s
will
p
r
o
v
id
e
in
f
o
r
m
atio
n
th
at
ca
n
b
e
v
is
u
ally
m
ap
p
ed
to
th
e
lo
ca
tio
n
o
f
an
o
b
ject
in
th
e
i
n
p
u
t
im
ag
e.
I
n
t
h
is
s
tu
d
y
,
Gr
ad
-
C
AM
is
u
s
ed
to
in
ter
p
r
et
an
d
v
alid
ate
t
h
e
r
o
le
o
f
f
ac
ial
f
ea
tu
r
e
r
eg
i
o
n
s
in
au
tis
m
class
if
icatio
n
.
R
eg
io
n
s
in
f
ac
ial
im
ag
es
ar
e
in
tu
itiv
ely
u
n
d
er
s
to
o
d
as
r
e
g
io
n
s
ass
o
ciate
d
with
f
ac
ial
lan
d
m
ar
k
s
.
Giv
en
an
in
p
u
t
f
ac
ial
im
ag
e,
Gr
ad
-
C
AM
will
a
llo
w
th
e
g
en
er
atio
n
o
f
h
ea
tm
ap
s
with
d
if
f
e
r
en
t
tr
ain
ed
m
o
d
els.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J
Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
6
:
766
-
7
7
9
772
B
y
an
aly
zin
g
th
e
v
is
u
al
lo
ca
t
io
n
o
f
th
e
h
ea
tm
a
p
s
r
elativ
e
t
o
r
eg
io
n
s
in
th
e
f
ac
ial
im
a
g
e,
it
ca
n
b
e
a
n
aly
ze
wh
eth
er
th
e
m
o
d
els
p
ay
atten
tio
n
to
r
eg
i
o
n
s
im
p
o
r
tan
t
f
o
r
f
ac
ial
e
x
p
r
ess
io
n
.
T
h
is
wi
ll
also
n
ee
d
to
b
e
co
n
s
id
er
ed
i
n
r
elatio
n
to
th
e
ac
cu
r
ac
y
o
f
t
h
e
r
esu
ltin
g
m
o
d
els.
T
h
is
is
ev
id
en
ce
th
at
s
u
g
g
ests
a
co
r
r
elatio
n
b
etwe
en
im
p
o
r
ta
n
t f
ac
ial
lo
ca
t
io
n
s
an
d
au
tis
m
class
if
icatio
n
.
2
.
5
.
P
r
o
po
s
ed
t
ra
ini
ng
a
nd
ev
a
lua
t
io
n wo
r
k
f
lo
w
Fig
u
r
e
3
p
r
esen
ts
th
e
p
r
o
p
o
s
ed
wo
r
k
f
lo
w
in
o
u
r
s
tu
d
y
,
in
clu
d
in
g
b
o
t
h
th
e
tr
ain
in
g
p
h
a
s
e
an
d
th
e
ev
alu
atio
n
p
h
ase.
T
h
e
p
r
o
p
o
s
e
d
wo
r
k
f
l
o
w
was
d
esig
n
ed
to
a
n
aly
ze
th
e
ef
f
ec
t
o
f
f
ac
ial
lan
d
m
ar
k
s
o
n
t
h
e
way
d
ee
p
lear
n
i
n
g
m
o
d
els
class
if
y
im
ag
es
to
d
iag
n
o
s
e
a
u
tis
m
.
E
x
p
er
im
en
ts
wer
e
s
et
u
p
f
o
r
s
tr
ateg
ies
in
b
o
th
th
e
tr
ain
in
g
p
h
ase
an
d
th
e
ev
alu
ati
o
n
p
h
ase.
Fig
u
r
e
3
.
P
r
o
p
o
s
ed
tr
ain
i
n
g
an
d
ev
alu
atio
n
wo
r
k
f
lo
w
I
n
th
e
tr
ain
in
g
p
h
ase,
d
if
f
er
en
t
tr
ain
in
g
s
tr
ateg
ies
ar
e
s
et
u
p
f
o
r
all
th
e
d
ee
p
lear
n
in
g
m
o
d
els
in
th
e
ex
p
er
im
en
t.
Sp
ec
i
f
ically
,
th
e
r
e
ar
e
2
s
tr
ateg
ies:
d
ef
au
lt
au
g
m
en
tatio
n
an
d
th
e
p
r
o
p
o
s
ed
f
ac
ial
im
ag
e
au
g
m
en
tatio
n
m
eth
o
d
.
I
n
th
i
s
way
,
th
e
in
f
l
u
en
ce
o
f
f
ac
i
al
lan
d
m
ar
k
s
,
w
h
ich
r
e
p
r
ese
n
t
f
ac
ial
ex
p
r
ess
io
n
f
ea
tu
r
es,
o
n
th
e
tr
ain
i
n
g
p
er
f
o
r
m
an
ce
o
f
d
ee
p
lear
n
in
g
m
o
d
els.
T
h
u
s
,
in
th
e
ev
alu
atio
n
p
h
ase,
th
e
ex
p
ec
ted
r
esu
lt
is
th
at
th
e
p
r
o
p
o
s
ed
f
ac
ial
im
ag
e
au
g
m
en
tatio
n
s
tr
ateg
y
will
y
ield
th
e
h
ig
h
est
p
er
f
o
r
m
an
ce
,
an
d
th
e
r
eb
y
also
clar
if
y
th
e
ef
f
ec
ti
v
en
ess
o
f
th
is
tech
n
iq
u
e.
T
h
e
tr
ain
in
g
s
tr
ateg
ies
will
b
e
test
ed
with
v
ar
io
u
s
p
o
p
u
lar
C
NN
m
o
d
els,
n
am
ely
E
f
f
ici
en
tNet
-
B
0
,
E
f
f
icien
tNet
-
B
4
,
R
esNet
-
1
8
,
R
esNet
-
5
0
,
R
es
Net
-
1
0
1
,
Mo
b
ileNet
-
V2
,
Den
s
eNe
t
-
1
2
1
,
a
n
d
Den
s
eNe
t
-
2
0
1
.
T
h
e
co
m
p
r
eh
en
s
iv
e
a
n
d
co
m
p
ar
ativ
e
ev
alu
atio
n
with
v
ar
io
u
s
m
o
d
els
h
elp
s
to
co
n
f
ir
m
th
e
co
r
r
ec
tn
ess
o
f
th
e
p
r
o
p
o
s
ed
h
y
p
o
th
esis
.
T
h
eo
r
eti
ca
lly
,
th
e
p
r
o
p
o
s
ed
f
ac
ial
im
a
g
e
au
g
m
e
n
tatio
n
s
tr
ateg
y
will
ac
h
iev
e
th
e
h
i
g
h
est
p
er
f
o
r
m
an
ce
o
n
m
o
s
t
o
f
t
h
e
t
ested
C
NN
m
o
d
els.
T
h
is
e
x
p
er
im
en
t
also
p
r
o
v
id
es
a
p
er
s
p
ec
tiv
e
o
n
h
o
w
th
e
m
o
d
els co
m
p
a
r
e
with
ea
ch
o
th
er
in
ter
m
s
o
f
th
e
im
ag
e
d
ata
c
h
ar
ac
ter
is
tics
o
f
th
e
p
r
o
b
lem
.
I
n
ex
p
er
im
e
n
ts
,
im
ag
es
ar
e
n
o
r
m
alize
d
b
y
tr
a
n
s
f
o
r
m
i
n
g
th
e
p
ix
el
v
alu
es
b
ef
o
r
e
b
ein
g
f
ed
in
to
th
e
d
ee
p
lear
n
in
g
m
o
d
el.
T
h
is
is
im
p
o
r
tan
t
b
ec
au
s
e
it
h
elp
s
to
s
tab
ilize
th
e
r
an
g
e
o
f
weig
h
ts
o
f
d
ee
p
lear
n
in
g
m
o
d
els
d
u
r
in
g
tr
ain
i
n
g
.
First,
th
e
im
ag
es
ar
e
co
n
v
er
ted
to
3
2
-
b
it
f
lo
atin
g
p
o
in
t
f
o
r
m
at
with
th
e
p
ix
el
v
alu
e
r
an
g
e
o
f
[
0
,
1
]
.
T
h
en
,
th
e
p
ix
el
v
alu
es
ar
e
n
o
r
m
alize
d
b
ase
d
o
n
th
e
ex
p
ec
te
d
v
alu
e
an
d
s
tan
d
ar
d
d
ev
iatio
n
v
alu
e
ca
lcu
lated
o
n
th
e
I
m
ag
eNe
t
s
et.
T
h
is
h
elp
s
to
k
ee
p
th
e
d
is
tr
ib
u
tio
n
o
f
p
ix
el
v
alu
es
co
n
s
is
ten
t
b
ef
o
r
e
f
ee
d
in
g
in
t
o
th
e
d
ee
p
lear
n
i
n
g
m
o
d
el.
I
n
th
e
ev
alu
atio
n
p
h
ase,
th
e
tr
ain
ed
C
NN
m
o
d
els
will
b
e
ev
alu
ated
o
n
th
e
test
d
ata.
First,
class
if
icatio
n
s
co
r
es
wi
ll
b
e
c
alcu
lated
an
d
u
s
ed
as
a
cr
iter
io
n
f
o
r
c
o
m
p
ar
is
o
n
b
etwe
en
test
in
g
s
tr
ateg
ies
as
well
as b
etwe
en
s
p
ec
if
ic
m
o
d
els to
co
n
f
ir
m
th
e
h
y
p
o
th
eses
.
Seco
n
d
,
h
ea
tm
ap
r
esu
lts
will b
e
ca
lcu
lated
f
o
r
th
e
tr
ain
ed
m
o
d
els co
r
r
esp
o
n
d
in
g
to
th
e
in
p
u
t im
ag
es.
T
h
ese
h
ea
tm
ap
r
esu
lts
will
in
d
icate
th
e
atten
tio
n
r
eg
io
n
s
o
f
ea
ch
tr
ain
ed
m
o
d
el
f
o
r
a
s
p
ec
if
ic
in
p
u
t
im
ag
e.
T
h
ese
at
ten
tio
n
r
eg
io
n
s
will
b
e
d
is
cu
s
s
ed
b
ased
o
n
th
e
co
m
p
ar
is
o
n
with
th
e
l
o
ca
tio
n
s
o
f
im
p
o
r
tan
t
r
eg
io
n
s
in
t
h
e
im
ag
e,
s
p
ec
if
ically
th
e
f
a
cial
r
eg
io
n
s
ar
o
u
n
d
lan
d
m
ar
k
s
-
ar
ea
s
co
n
s
id
er
ed
to
r
ep
r
esen
t th
e
e
x
p
r
ess
iv
e
f
ea
tu
r
es o
f
h
u
m
an
f
ac
es.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
I
mp
r
o
vin
g
efficien
cy
o
f
a
u
tis
m
d
etec
tio
n
b
a
s
ed
o
n
f
a
cia
l ima
g
e
la
n
d
ma
r
ks
(
N
g
u
ye
n
Tr
o
n
g
Tu
n
g
)
773
3.
E
XP
E
R
I
M
E
N
T
AND
R
E
SU
L
T
S
T
h
is
s
ec
tio
n
p
r
esen
ts
th
e
d
etails
o
f
th
e
ex
p
er
im
en
t
as
well
as
th
e
a
n
aly
s
is
o
f
th
e
ex
p
e
r
im
en
t
al
r
esu
lts
.
First,
p
r
esen
t
th
e
ex
p
er
im
en
ta
l
s
etu
p
.
Seco
n
d
,
p
r
esen
t
th
e
im
p
ac
t
o
f
th
e
p
r
o
p
o
s
ed
m
et
h
o
d
.
T
h
ir
d
,
a
n
aly
ze
th
e
Gr
ad
-
C
AM
h
ea
tm
ap
s
in
r
elati
o
n
to
f
ac
ial
lan
d
m
ar
k
s
an
d
d
is
cu
s
s
th
e
r
elate
d
d
etails.
3
.
1
.
E
x
perim
ent
a
l set
up
I
n
th
e
e
x
p
er
im
e
n
ts
,
th
e
m
o
d
el
s
wo
u
ld
b
e
tr
ain
ed
u
s
in
g
th
e
Ad
am
alg
o
r
ith
m
[
2
4
]
.
B
y
u
s
in
g
a
tr
an
s
f
e
r
lear
n
in
g
ap
p
r
o
ac
h
,
th
o
s
e
m
o
d
els
wo
u
ld
b
e
f
in
e
-
tu
n
e
d
f
r
o
m
p
r
e
-
tr
ain
ed
m
o
d
els
p
r
o
v
id
ed
b
y
th
e
Py
T
o
r
ch
d
ee
p
lear
n
in
g
lib
r
a
r
y
at
h
ttp
s
://p
y
to
r
ch
.
o
r
g
.
T
o
en
s
u
r
e
th
e
p
er
f
o
r
m
an
ce
ac
h
iev
ed
f
o
r
t
r
ain
in
g
th
e
m
o
d
els,
th
e
ef
f
icien
t u
s
e
o
f
o
p
tim
izer
s
,
lea
r
n
in
g
r
ate
s
ch
ed
u
ler
s
,
an
d
b
atc
h
s
izes
will b
e
ex
p
lo
ited
in
th
e
ex
p
er
im
en
ts
.
T
h
is
will
ef
f
ec
tiv
ely
c
o
n
tr
o
l
th
e
u
p
d
ate
r
ate
o
f
th
e
m
o
d
el
weig
h
ts
,
th
e
tr
ain
in
g
tim
e,
an
d
th
e
g
e
n
er
aliza
tio
n
a
b
ilit
y
.
I
n
d
etail,
two
s
tep
s
o
f
ch
a
n
g
i
n
g
th
e
f
r
o
ze
n
s
tate
o
f
th
e
p
ar
am
eter
s
ar
e
p
er
f
o
r
m
e
d
to
o
p
ti
m
ize
th
e
u
p
d
ate
o
f
p
ar
ts
o
f
th
e
m
o
d
el
d
u
r
in
g
th
e
ex
p
er
im
en
t.
Fu
r
t
h
er
m
o
r
e,
th
e
cy
clica
l
lear
n
in
g
r
ates
m
ec
h
an
is
m
is
u
s
ed
[
2
5
]
to
h
elp
th
e
m
o
d
el
tr
ain
f
aster
w
h
ile
en
s
u
r
in
g
co
n
v
er
g
en
ce
b
y
allo
win
g
th
e
lear
n
in
g
r
ates
to
g
r
o
w.
B
esid
es,
th
e
b
atch
s
ize
v
alu
es
ar
e
also
ch
o
s
en
ap
p
r
o
p
r
iately
to
o
p
tim
ize
th
e
m
em
o
r
y
o
f
th
e
GPU
h
ar
d
war
e.
So
u
r
ce
co
d
e
is
p
u
b
lis
h
ed
o
n
Git
L
ab
p
latf
o
r
m
[
2
6
]
.
E
x
p
er
im
e
n
ts
wer
e
p
er
f
o
r
m
ed
o
n
th
e
Kag
g
le
p
latf
o
r
m
with
an
NVI
DI
A
T
esla
P1
0
0
GPU
with
1
6
GB
VR
AM
.
Kag
g
le
is
a
p
o
p
u
lar
p
latf
o
r
m
th
at
is
o
p
tim
ized
f
o
r
d
ee
p
lear
n
i
n
g
task
s
.
T
h
e
class
if
icatio
n
s
co
r
es u
s
ed
in
clu
d
e
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
e
ca
ll,
AUC,
F1
-
s
co
r
e.
3
.
2
.
T
he
im
pa
ct
o
f
t
he
pro
p
o
s
ed
a
ug
m
ent
a
t
io
n
m
et
ho
d
Fig
u
r
e
4
clea
r
ly
s
h
o
ws
th
at
th
e
m
o
d
els
tr
ain
ed
with
th
e
p
r
o
p
o
s
ed
f
ac
ial
im
ag
e
au
g
m
en
tati
o
n
m
eth
o
d
d
ata
h
av
e
s
ig
n
if
ican
tly
s
u
p
er
io
r
p
er
f
o
r
m
an
ce
o
n
all
m
etr
i
cs.
Sp
ec
if
ically
,
th
e
av
er
ag
e
ac
cu
r
ac
y
v
alu
e
is
0
.
9
0
5
4
1
7
,
th
e
av
er
a
g
e
AUC
v
alu
e
is
0
.
9
6
2
1
3
3
,
th
e
av
er
ag
e
p
r
ec
is
io
n
v
al
u
e
is
0
.
9
1
9
8
,
th
e
av
er
ag
e
r
ec
all
v
alu
e
is
0
.
8
8
8
3
3
3
,
a
n
d
th
e
a
v
er
ag
e
F1
-
s
co
r
e
v
alu
e
is
0
.
9
0
3
6
7
8
.
Ad
d
itio
n
ally
,
th
e
m
etr
ics
o
f
th
e
ca
s
e
u
s
in
g
d
ef
au
l
t
au
g
m
en
tatio
n
ar
e
lo
we
r
at
all
m
etr
ic
ty
p
es.
I
n
d
etail,
with
av
er
ag
e
p
r
ec
is
io
n
,
th
is
ca
s
e
ac
h
iev
es
0
.
8
7
1
2
5
,
wh
ich
is
0
.
0
3
4
1
6
7
lo
wer
th
a
n
th
e
p
r
o
p
o
s
ed
tech
n
iq
u
e.
Similar
ly
,
with
av
er
ag
e
AUC,
th
is
ca
s
e
ac
h
iev
es
0
.
9
5
0
5
6
1
an
d
is
0
.
0
1
1
5
7
2
l
o
wer
.
W
ith
av
e
r
ag
e
p
r
ec
is
io
n
,
th
i
s
ca
s
e
ac
h
iev
es
0
.
9
1
5
6
7
8
an
d
is
0
.
0
0
4
1
2
2
lo
wer
.
W
ith
av
er
ag
e
r
ec
all
,
th
is
ca
s
e
ac
h
iev
es
0
.
8
1
9
1
6
7
a
n
d
is
0
.
0
6
9
1
6
6
lo
wer
.
W
ith
av
er
ag
e
F1
-
s
co
r
e,
th
is
ca
s
e
ac
h
iev
es
0
.
8
6
3
2
0
9
an
d
is
0
.
0
4
0
4
6
9
lo
wer
.
T
h
er
ef
o
r
e,
t
h
ey
r
ef
lect
th
at
ap
p
l
y
in
g
t
h
e
p
r
o
p
o
s
ed
f
ac
ial
im
ag
e
au
g
m
en
tatio
n
m
eth
o
d
m
ak
es a
s
ig
n
if
ican
t d
if
f
er
e
n
ce
.
Fig
u
r
e
4
.
Av
e
r
ag
e
m
etr
ics f
o
r
th
e
d
ef
au
lt a
u
g
m
en
tatio
n
s
tr
ateg
y
an
d
th
e
p
r
o
p
o
s
ed
au
g
m
en
t
atio
n
s
tr
ateg
y
W
h
en
th
e
ex
p
er
im
en
tal
r
esu
lts
ar
e
ar
r
an
g
ed
b
y
th
e
ac
cu
r
ac
y
s
ca
le,
th
e
d
is
tr
ib
u
tio
n
o
f
th
e
lo
ca
tio
n
s
o
f
th
e
m
o
d
els
tr
ain
ed
with
th
e
th
r
ee
au
g
m
en
tatio
n
s
tr
ateg
ies
ca
n
b
e
s
ee
n
clea
r
ly
.
I
n
T
ab
le
2
,
th
e
s
ce
n
ar
io
s
tr
ain
ed
u
s
in
g
th
e
p
r
o
p
o
s
ed
f
a
cial
im
ag
e
au
g
m
en
tatio
n
m
et
h
o
d
d
em
o
n
s
tr
ated
h
ig
h
d
iag
n
o
s
tic
ac
cu
r
ac
y
.
Mo
r
e
s
p
ec
if
ically
,
am
o
n
g
th
e
8
h
i
g
h
est
-
ac
cu
r
ac
y
tr
ai
n
in
g
s
ce
n
a
r
io
s
,
6
wer
e
m
o
d
els
tr
ain
ed
u
s
in
g
th
e
p
r
o
p
o
s
ed
f
ac
ial
im
ag
e
au
g
m
en
tatio
n
m
e
th
o
d
,
ac
co
u
n
tin
g
f
o
r
7
5
%.
T
h
at
m
ea
n
s
th
e
s
ce
n
ar
io
s
wh
er
e
t
h
e
m
o
d
el
is
tr
ain
ed
with
d
ef
au
lt a
u
g
m
en
tatio
n
o
n
l
y
ac
co
u
n
t f
o
r
2
5
% o
f
th
e
8
b
est ca
s
es in
ter
m
s
o
f
ac
cu
r
ac
y
.
Mo
r
e
im
p
o
r
ta
n
tl
y
,
t
h
e
ca
s
es
t
h
at
d
o
n
o
t
u
s
e
t
h
e
p
r
o
p
o
s
e
d
tec
h
n
i
q
u
e
n
o
t
o
n
ly
ac
c
o
u
n
t
f
o
r
a
s
m
al
l
p
r
o
p
o
r
ti
o
n
b
u
t
als
o
r
an
k
v
e
r
y
l
o
w
i
n
th
e
to
p
8
ca
s
es
i
n
t
er
m
s
o
f
ac
c
u
r
a
cy
,
n
am
el
y
5
th
p
lac
e
wit
h
D
e
n
s
eN
et
-
201
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J
Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
6
:
766
-
7
7
9
774
an
d
7
th
p
l
ac
e
wi
th
E
f
f
i
ci
en
tNe
t
-
B
4
.
I
n
d
etai
l,
7
th
p
la
ce
an
d
6
th
p
la
ce
h
av
e
t
h
e
s
am
e
a
cc
u
r
ac
y
,
b
u
t
7
th
p
l
ac
e
h
as
a
cle
ar
l
y
s
m
a
lle
r
A
UC
.
T
h
is
c
lea
r
l
y
s
h
o
ws
t
h
e
d
is
a
d
v
a
n
ta
g
e
o
f
t
h
e
d
e
f
a
u
lt
a
u
g
m
e
n
t
ati
o
n
c
o
m
p
a
r
e
d
t
o
t
h
e
p
r
o
p
o
s
e
d
te
ch
n
i
q
u
e
.
T
ab
le
2
.
T
h
e
to
p
8
ca
s
es so
r
ted
in
ter
m
s
o
f
ac
cu
r
ac
y
S
t
r
a
t
e
g
y
M
o
d
e
l
A
c
c
u
r
a
c
y
AUC
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
sc
o
r
e
P
r
o
p
o
se
d
a
u
g
me
n
t
a
t
i
o
n
Ef
f
i
c
i
e
n
t
N
e
t
-
B4
0
.
9
2
6
6
6
7
0
.
9
6
7
9
3
3
0
.
9
2
6
6
6
7
0
.
9
2
6
6
6
7
0
.
9
2
6
6
6
7
P
r
o
p
o
se
d
a
u
g
me
n
t
a
t
i
o
n
R
e
sN
e
t
-
18
0
.
9
1
6
6
6
7
0
.
9
6
8
6
8
9
0
.
9
2
5
1
7
0
.
9
0
6
6
6
7
0
.
9
1
5
8
2
5
P
r
o
p
o
se
d
a
u
g
me
n
t
a
t
i
o
n
Ef
f
i
c
i
e
n
t
N
e
t
-
B0
0
.
9
1
6
6
6
7
0
.
9
7
4
0
8
9
0
.
9
3
1
0
3
4
0
.
9
0
.
9
1
5
2
5
4
P
r
o
p
o
se
d
a
u
g
me
n
t
a
t
i
o
n
M
o
b
i
l
e
N
e
t
-
V2
0
.
9
1
3
3
3
3
0
.
9
7
6
0
2
2
0
.
9
3
0
5
5
6
0
.
8
9
3
3
3
3
0
.
9
1
1
5
6
5
D
e
f
a
u
l
t
a
u
g
m
e
n
t
a
t
i
o
n
D
e
n
seN
e
t
-
2
0
1
0
.
9
0
.
9
7
3
3
3
3
0
.
8
8
9
6
1
0
0
.
9
1
3
3
3
3
0
.
9
0
1
3
1
6
P
r
o
p
o
se
d
a
u
g
me
n
t
a
t
i
o
n
D
e
n
seN
e
t
-
2
0
1
0
.
8
9
6
6
6
7
0
.
9
7
1
4
2
2
0
.
9
2
8
0
5
8
0
.
8
6
0
.
8
9
2
7
3
3
D
e
f
a
u
l
t
a
u
g
m
e
n
t
a
t
i
o
n
Ef
f
i
c
i
e
n
t
N
e
t
-
B4
0
.
8
9
6
6
6
7
0
.
9
6
5
1
1
1
0
,
9
2
8
0
5
8
0
.
8
6
0
.
8
9
2
7
3
4
P
r
o
p
o
se
d
a
u
g
me
n
t
a
t
i
o
n
D
e
n
seN
e
t
-
1
2
1
0
.
8
9
3
3
3
3
0
.
9
5
2
7
5
6
0
.
8
9
8
6
4
9
0
.
8
8
6
6
6
7
0
.
8
9
2
6
1
7
Fu
r
th
e
r
e
v
id
en
ce
is
s
h
o
w
n
in
T
a
b
le
3
w
it
h
t
h
e
8
wo
r
s
t
p
er
f
o
r
m
in
g
ca
s
es
s
o
r
te
d
b
y
ac
c
u
r
ac
y
m
e
asu
r
e.
Am
o
n
g
t
h
e
m
,
t
h
e
ca
s
e
u
s
i
n
g
t
h
e
p
r
o
p
o
s
ed
t
ec
h
n
iq
u
e
o
c
cu
p
i
es
o
n
ly
2
p
o
s
iti
o
n
s
,
e
q
u
i
v
al
e
n
t
to
2
5
%
o
f
t
h
e
t
o
tal
8
p
o
s
iti
o
n
s
.
T
h
ese
tw
o
p
o
s
iti
o
n
s
b
e
lo
n
g
t
o
t
h
e
R
esNe
t
-
5
0
m
o
d
el
wi
th
a
n
ac
cu
r
a
c
y
o
f
0
.
8
9
3
3
3
3
a
n
d
R
es
Net
-
1
0
1
wit
h
an
a
cc
u
r
ac
y
o
f
0
.
8
8
6
6
6
7
,
b
o
t
h
o
f
w
h
ic
h
a
r
e
as
y
m
p
to
tic
all
y
cl
o
s
e
t
o
0
.
9
.
Fu
r
t
h
er
m
o
r
e,
b
o
t
h
o
f
t
h
es
e
ca
s
es
ar
e
i
n
t
h
e
u
p
p
er
h
a
lf
o
f
t
h
e
t
a
b
le
,
w
h
i
ch
m
e
a
n
s
t
h
at
th
ey
a
r
e
in
th
e
h
i
g
h
-
a
cc
u
r
a
cy
r
e
g
i
o
n
o
f
th
e
t
ab
le
.
I
n
o
t
h
e
r
wo
r
d
s
,
t
h
e
ca
s
es
th
at
d
o
n
o
t
u
s
e
th
e
p
r
o
p
o
s
ed
tec
h
n
i
q
u
e
o
c
c
u
p
y
7
5
%
o
f
t
h
e
t
a
b
le
an
d
a
r
e
also
m
o
s
tl
y
i
n
t
h
e
lo
w
er
-
s
c
o
r
i
n
g
r
e
g
i
o
n
.
T
h
is
is
a
ls
o
a
cl
ea
r
d
em
o
n
s
t
r
at
io
n
o
f
o
u
r
h
y
p
o
t
h
esis
in
t
h
is
p
a
p
e
r
.
T
ab
le
3
.
T
h
e
b
o
tto
m
8
ca
s
es so
r
ted
in
ter
m
s
o
f
ac
cu
r
ac
y
S
t
r
a
t
e
g
y
M
o
d
e
l
A
c
c
u
r
a
c
y
AUC
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
sc
o
r
e
P
r
o
p
o
se
d
a
u
g
me
n
t
a
t
i
o
n
R
e
sN
e
t
-
50
0
.
8
9
3
3
3
3
0
.
9
4
1
9
1
1
0
.
9
1
5
4
9
3
0
.
8
6
6
6
6
7
0
.
8
9
0
4
1
1
D
e
f
a
u
l
t
a
u
g
m
e
n
t
a
t
i
o
n
D
e
n
seN
e
t
-
1
2
1
0
.
8
9
0
.
9
5
2
0
.
9
3
9
8
5
0
0
.
8
3
3
3
3
3
0
.
8
8
3
3
9
2
P
r
o
p
o
se
d
a
u
g
me
n
t
a
t
i
o
n
R
e
sN
e
t
-
101
0
.
8
8
6
6
6
7
0
.
9
4
4
2
4
4
0
.
9
0
2
7
7
8
0
.
8
6
6
6
6
7
0
.
8
8
4
3
5
4
D
e
f
a
u
l
t
a
u
g
m
e
n
t
a
t
i
o
n
M
o
b
i
l
e
N
e
t
-
V2
0
.
8
8
6
6
6
7
0
.
9
5
7
4
2
2
0
.
8
8
6
6
6
7
0
.
8
8
6
6
6
7
0
.
8
8
6
6
6
7
D
e
f
a
u
l
t
a
u
g
m
e
n
t
a
t
i
o
n
R
e
sN
e
t
-
18
0
.
8
6
3
3
3
3
0
.
9
5
9
2
4
4
0
.
9
5
0
4
1
3
0
.
7
6
6
6
6
7
0
.
8
4
8
7
0
8
D
e
f
a
u
l
t
a
u
g
m
e
n
t
a
t
i
o
n
Ef
f
i
c
i
e
n
t
N
e
t
-
B0
0
.
8
5
0
.
9
4
8
1
3
3
0
.
9
2
0
.
7
6
6
6
6
7
0
.
8
3
6
3
6
4
D
e
f
a
u
l
t
a
u
g
m
e
n
t
a
t
i
o
n
R
e
sN
e
t
-
50
0
.
8
4
6
6
6
7
0
.
9
3
1
4
2
2
0
.
9
1
9
3
5
5
0
.
7
6
0
.
8
3
2
1
1
7
D
e
f
a
u
l
t
a
u
g
m
e
n
t
a
t
i
o
n
R
e
sN
e
t
-
101
0
.
8
3
6
6
6
7
0
.
9
1
7
8
2
2
0
.
8
9
1
4
7
3
0
.
7
6
6
6
6
7
0
.
8
2
4
3
7
3
Nex
t,
s
tatis
tical
te
s
ts
u
s
in
g
a
p
air
ed
t
-
test
b
etwe
en
p
air
s
o
f
m
o
d
els
o
n
t
h
e
s
am
e
test
s
et
ar
e
also
in
clu
d
ed
t
o
v
alid
ate
im
p
r
o
v
e
m
en
ts
.
T
h
e
r
esu
lts
s
h
o
w
th
at
t
h
e
m
ajo
r
ity
o
f
m
o
d
els
s
h
o
w
c
lear
an
d
s
tatis
tically
s
ig
n
if
ican
t
d
if
f
er
e
n
ce
s
,
with
5
o
u
t
o
f
8
m
o
d
els
ac
h
iev
in
g
p
-
v
alu
e
less
th
an
0
.
0
5
.
Sp
ec
if
ically
,
R
esNet
-
18
r
esu
lts
with
a
t
-
s
tat
o
f
6
.
0
9
7
8
4
7
a
n
d
p
-
v
alu
e
less
th
an
0
.
0
0
0
0
0
1
,
in
d
icatin
g
th
at
th
e
r
e
is
a
s
tati
s
tical
ly
s
ig
n
if
ican
t d
if
f
er
en
ce
b
etwe
en
th
e
p
air
ed
g
r
o
u
p
s
.
Similar
ly
,
E
f
f
icien
tNet
-
B
0
h
as a
t
-
s
tat
o
f
-
3
.
2
7
4
0
0
6
,
p
-
v
alu
e
o
f
0
.
0
0
1
1
8
5
,
an
d
E
f
f
icien
tNet
-
B
4
h
as
a
t
-
s
tat
o
f
-
2
.
9
6
0
9
3
6
,
p
-
v
al
u
e
o
f
0
.
0
0
3
3
1
3
,
a
ls
o
d
em
o
n
s
tr
atin
g
s
ig
n
if
ican
t
im
p
r
o
v
em
en
ts
.
I
n
ad
d
itio
n
,
R
esNet
-
1
0
1
with
p
-
v
alu
e
o
f
0
.
0
0
4
8
5
7
a
n
d
De
n
s
eNe
t
-
2
0
1
with
p
-
v
alu
e
o
f
0
.
0
0
2
3
2
8
ar
e
also
s
tr
o
n
g
ev
id
en
ce
ag
ain
s
t
th
e
n
u
ll
h
y
p
o
th
esis
.
I
n
co
n
tr
ast,
Mo
b
ileNet
-
V2
m
o
d
el
with
p
-
v
alu
e
o
f
0
.
2
1
2
3
3
5
d
o
es
n
o
t
r
ea
ch
s
ig
n
if
ican
ce
.
T
h
ese
r
e
s
u
lts
co
n
f
ir
m
th
at
th
er
e
is
a
s
tatis
tically
r
eliab
le
im
p
r
o
v
em
e
n
t in
m
o
s
t o
f
th
e
te
s
t m
o
d
els.
T
h
is
also
p
r
o
v
es o
u
r
r
esear
ch
h
y
p
o
t
h
esis
f
o
r
th
e
p
r
o
p
o
s
ed
m
et
h
o
d
.
Ad
d
itio
n
ally
,
f
o
r
th
e
ca
s
es
ap
p
ly
in
g
th
e
d
ata
au
g
m
e
n
tatio
n
,
th
e
m
o
d
el
ca
lib
r
atio
n
is
ev
al
u
ated
u
s
in
g
th
e
ex
p
ec
ted
ca
lib
r
atio
n
er
r
o
r
(
E
C
E
)
i
n
d
ex
.
T
h
e
r
esu
lts
o
b
tain
ed
f
o
r
m
o
s
t
m
o
d
els
ar
e
r
elativ
ely
well
-
ca
lib
r
ated
with
q
u
ite
s
m
all
E
C
E
v
alu
es.
T
h
e
m
o
d
el
with
th
e
h
ig
h
est
E
C
E
is
R
esNe
t
-
1
0
1
with
E
C
E
o
f
0
.
1
0
0
2
9
5
.
T
h
e
r
em
ain
i
n
g
m
o
d
els
all
h
av
e
E
C
E
less
th
an
0
.
1
.
Acc
o
r
d
in
g
ly
,
th
e
R
esNet
-
5
0
m
o
d
el
ac
h
iev
ed
E
C
E
o
f
0
.
0
8
9
2
9
1
,
th
e
Den
s
e
Net
-
2
0
1
m
o
d
el
ac
h
iev
ed
E
C
E
o
f
0
.
0
8
4
4
7
0
,
th
e
Den
s
eNe
t
-
1
2
1
m
o
d
el
ac
h
ie
v
ed
E
C
E
o
f
0
.
0
8
1
2
3
0
,
th
e
E
f
f
icien
tNet
-
B
0
m
o
d
el
ac
h
iev
ed
E
C
E
o
f
0
.
0
7
5
3
1
3
,
th
e
Mo
b
ileNet
-
V
2
m
o
d
el
ac
h
iev
e
d
E
C
E
o
f
0
.
0
7
4
8
5
3
,
th
e
R
esNet
-
1
8
m
o
d
el
ac
h
iev
ed
E
C
E
o
f
0
.
0
6
9
9
4
3
,
th
e
E
f
f
icien
tNet
-
B
4
m
o
d
el
ac
h
iev
ed
E
C
E
o
f
0
.
0
6
4
4
0
9
.
T
h
is
also
co
n
tr
ib
u
tes
to
d
em
o
n
s
tr
atin
g
th
e
ef
f
ec
tiv
en
ess
o
f
d
ata
au
g
m
e
n
tatio
n
wh
en
in
th
e
ex
p
er
im
en
t,
th
e
ar
c
h
itectu
r
es
n
o
t
o
n
ly
ac
h
iev
ed
h
i
g
h
ac
cu
r
ac
y
b
u
t
also
p
r
o
v
id
ed
r
e
liab
le
p
r
o
b
ab
ilit
y
esti
m
ates,
wh
ich
is
im
p
o
r
tan
t in
th
e
m
ed
ical
c
o
n
tex
t.
I
n
ad
d
itio
n
,
s
o
m
e
in
f
o
r
m
atio
n
ab
o
u
t
th
e
c
o
m
p
u
tatio
n
al
ti
m
e,
m
em
o
r
y
u
s
ag
e,
an
d
tr
ain
in
g
tim
e
is
also
p
r
o
v
id
ed
.
T
h
e
ex
p
er
im
en
t
was
p
er
f
o
r
m
ed
with
th
e
Ka
g
g
le
s
er
v
er
,
s
o
th
e
m
em
o
r
y
is
a
llo
ca
ted
with
in
th
e
r
an
g
e
allo
wed
b
y
th
e
Ka
g
g
le
s
er
v
er
.
W
e
also
p
r
o
v
id
e
m
o
r
e
d
etails
ab
o
u
t
th
e
co
m
p
u
tatio
n
al
tim
e
an
d
t
r
ain
in
g
tim
e
f
o
r
ea
ch
ep
o
c
h
in
T
a
b
le
4
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
I
mp
r
o
vin
g
efficien
cy
o
f
a
u
tis
m
d
etec
tio
n
b
a
s
ed
o
n
f
a
cia
l ima
g
e
la
n
d
ma
r
ks
(
N
g
u
ye
n
Tr
o
n
g
Tu
n
g
)
775
T
ab
le
4
.
Av
e
r
ag
e
p
r
ed
icted
ti
m
es p
er
s
am
p
le
an
d
av
er
a
g
e
tr
ain
in
g
tim
es p
er
e
p
o
ch
M
o
d
e
l
A
v
e
r
a
g
e
p
r
e
d
i
c
t
e
d
t
i
me
p
e
r
sam
p
l
e
(
s
e
c
o
n
d
s)
A
v
e
r
a
g
e
t
r
a
i
n
i
n
g
t
i
me
p
e
r
e
p
o
c
h
(
sec
o
n
d
s)
Ef
f
i
c
i
e
n
t
N
e
t
-
B0
0
.
0
0
0
1
4
9
7
0
.
7
9
0
7
4
7
Ef
f
i
c
i
e
n
t
N
e
t
-
B4
0
.
0
0
0
6
1
0
1
1
6
.
3
7
9
6
8
6
R
e
sN
e
t
-
18
0
.
0
0
0
0
1
4
5
8
.
8
9
3
4
2
0
R
e
sN
e
t
-
50
0
.
0
0
0
1
1
0
8
8
.
7
9
8
2
0
5
R
e
sN
e
t
-
101
0
.
0
0
0
2
0
7
1
2
8
.
1
6
4
9
6
7
M
o
b
i
l
e
N
e
t
-
V2
0
.
0
0
0
1
0
5
6
9
.
0
1
1
8
7
5
D
e
n
seN
e
t
-
1
2
1
0
.
0
0
0
5
8
8
7
6
.
9
5
1
4
9
5
D
e
n
seN
e
t
-
2
0
1
0
.
0
0
1
0
1
8
1
1
1
.
6
7
9
6
5
4
3
.
3
.
Ana
ly
s
is
o
f
G
r
a
d
-
CAM
hea
t
m
a
ps
in re
la
t
io
n t
o
f
a
cia
l la
nd
m
a
rk
s
I
n
t
h
is
s
ec
t
io
n
,
t
h
e
r
ela
ti
o
n
s
h
i
p
b
et
we
en
t
h
e
m
o
d
el'
s
at
te
n
ti
o
n
,
r
ep
r
ese
n
te
d
b
y
G
r
a
d
-
C
AM
h
e
atm
a
p
s
,
an
d
k
e
y
r
eg
io
n
s
i
n
t
h
e
f
a
cia
l
i
m
a
g
e
s
a
n
a
ly
ze
d
.
T
h
is
wi
ll
c
o
n
t
r
i
b
u
te
t
o
r
e
v
ea
li
n
g
th
e
r
ela
ti
o
n
s
h
i
p
b
e
twe
en
t
h
e
q
u
ali
ty
o
f
au
tis
m
d
i
a
g
n
o
s
is
f
r
o
m
f
ac
i
al
i
m
a
g
es
a
n
d
t
h
e
i
m
ag
e
r
e
g
i
o
n
s
th
at
t
h
e
m
o
d
e
l
f
o
c
u
s
es
o
n
.
I
n
o
t
h
e
r
wo
r
d
s
,
d
o
es
th
e
m
o
d
el'
s
at
te
n
ti
o
n
t
o
k
ey
r
e
g
i
o
n
s
,
w
h
i
ch
ar
e
r
e
lat
e
d
to
f
a
cia
l
la
n
d
m
ar
k
s
,
af
f
e
ct
t
h
e
f
i
n
al
d
ia
g
n
o
s
ti
c
p
e
r
f
o
r
m
a
n
ce
?
W
e
s
ta
r
t
w
it
h
t
h
e
Gr
a
d
-
C
A
M
h
e
at
m
a
p
s
o
f
t
h
e
tw
o
b
est
-
p
e
r
f
o
r
m
i
n
g
ca
s
es
,
E
f
f
i
cie
n
tN
et
-
B
4
m
o
d
el
a
n
d
R
esNet
-
1
8
m
o
d
e
l,
b
o
t
h
o
f
w
h
i
c
h
h
a
v
e
b
ee
n
tr
ai
n
e
d
w
it
h
p
r
o
p
o
s
e
d
au
g
m
e
n
t
ati
o
n
tec
h
n
i
q
u
e.
T
h
e
s
p
e
cif
ic
r
es
u
lts
a
r
e
s
h
o
w
n
i
n
Fi
g
u
r
e
5
.
F
o
r
ea
ch
m
o
d
el
c
ase
,
t
h
e
r
es
u
lts
ar
e
p
l
o
tt
ed
in
tw
o
r
o
ws
.
T
h
e
f
i
r
s
t
r
o
w
s
h
o
ws
1
0
b
es
t
-
p
e
r
f
o
r
m
in
g
c
ases
,
an
d
s
ec
o
n
d
r
o
w
s
h
o
ws t
h
e
1
0
w
o
r
s
t
-
p
e
r
f
o
r
m
in
g
ca
s
es.
Fig
u
r
e
5
s
h
o
ws
a
s
tr
ik
in
g
d
em
o
n
s
tr
atio
n
o
f
o
u
r
p
r
e
d
ictio
n
h
y
p
o
t
h
esis
,
with
s
o
m
e
G
r
ad
-
C
AM
h
ea
tm
ap
s
o
f
t
h
e
E
f
f
icien
tNet
-
B
4
m
o
d
el
in
Fig
u
r
e
5
(
a)
a
n
d
th
e
R
esNet
-
1
8
m
o
d
el
in
Fi
g
u
r
e
5
(
b
)
u
s
in
g
th
e
p
r
o
p
o
s
ed
au
g
m
en
tatio
n
s
tr
ateg
y
.
I
n
t
h
e
f
i
r
s
t
r
o
w
o
f
two
ca
s
es,
it
is
ea
s
y
to
s
ee
th
at
th
e
m
o
d
els'
atten
tio
n
is
clea
r
ly
f
o
cu
s
ed
o
n
im
p
o
r
tan
t
f
ac
ial
r
eg
io
n
s
in
th
e
co
r
r
ec
t
p
r
ed
ictio
n
ca
s
es.
Sp
ec
if
ically
,
r
eg
io
n
s
s
u
ch
as
th
e
ey
es
an
d
n
o
s
e
ar
e
also
im
p
o
r
t
an
t
r
eg
io
n
s
m
ar
k
e
d
b
y
lan
d
m
a
r
k
s
co
m
m
o
n
ly
s
ee
n
in
f
ac
ial
e
x
p
r
ess
io
n
an
aly
s
is
.
T
h
is
r
ef
lects
th
at
th
e
m
o
d
els
h
av
e
ef
f
ec
tiv
el
y
lear
n
ed
f
ac
ial
ex
p
r
ess
io
n
f
ea
t
u
r
es
to
b
e
a
b
le
to
r
ec
o
g
n
ize
a
u
tis
m
well.
T
h
eo
r
etica
lly
,
au
tis
m
h
a
s
aty
p
ical
f
ac
ial
ex
p
r
ess
io
n
s
,
s
o
th
e
f
ac
t
th
at
Gr
a
d
-
C
AM
h
ea
tm
ap
s
ar
e
s
tr
o
n
g
ly
ass
o
ciate
d
with
r
eg
io
n
s
ass
o
ciate
d
with
f
ac
ial
lan
d
m
ar
k
s
s
u
g
g
ests
th
at
th
e
m
o
d
els
ca
p
tu
r
e
m
ea
n
in
g
f
u
l
f
ea
tu
r
es in
f
ac
ial
im
ag
es f
o
r
a
u
tis
m
d
iag
n
o
s
is
.
Op
p
o
s
itely
,
in
th
e
s
ec
o
n
d
r
o
w
o
f
two
ca
s
es,
th
e
im
ag
es
o
f
t
h
e
wr
o
n
g
r
esu
lts
s
h
o
w
th
at
t
h
e
h
ea
tm
ap
s
ar
e
o
f
ten
n
o
t
f
o
cu
s
ed
o
n
im
p
o
r
tan
t
ar
ea
s
o
f
th
e
f
ac
e,
wh
ich
ar
e
u
s
u
ally
d
eter
m
in
ed
b
y
lan
d
m
ar
k
s
.
T
h
e
r
esu
lts
ar
e
o
f
ten
s
ca
tter
ed
ac
r
o
s
s
m
an
y
ar
ea
s
o
f
t
h
e
f
ac
e,
e
v
en
o
u
ts
id
e
th
e
f
ac
e.
T
h
is
s
u
g
g
ests
th
at
th
e
m
o
d
els
ar
e
ca
p
tu
r
in
g
f
ea
tu
r
es
th
at
ar
e
ir
r
e
lev
an
t
o
r
h
a
v
e
litt
le
to
d
o
with
f
ac
ial
ex
p
r
ess
io
n
f
ea
tu
r
es
th
at
ar
e
im
p
o
r
tan
t
f
o
r
au
tis
m
.
T
h
is
co
u
ld
th
e
r
ef
o
r
e
b
e
an
ex
p
la
n
atio
n
f
o
r
wh
y
th
e
r
esu
lts
ar
e
p
o
o
r
o
r
u
n
s
tab
le.
I
n
Fig
u
r
e
6
,
Gr
ad
-
C
AM
h
ea
tm
ap
s
o
f
th
e
two
wo
r
s
t
-
p
er
f
o
r
m
in
g
ca
s
es
ar
e
p
r
esen
ted
,
th
e
R
esNe
t
-
50
m
o
d
el
in
Fig
u
r
e
6
(
a)
an
d
th
e
R
esNet
-
1
0
1
m
o
d
el
in
Fig
u
r
e
6
(
b
)
,
b
o
th
o
f
wh
ich
h
av
e
b
ee
n
tr
ain
ed
with
d
ef
au
lt
au
g
m
en
tatio
n
tec
h
n
iq
u
es.
Sim
ilar
ly
,
f
o
r
ea
ch
m
o
d
el
ca
s
e,
r
e
s
u
lts
ar
e
p
lo
tted
in
two
r
o
ws.
T
h
e
f
ir
s
t
r
o
w
s
h
o
ws
th
e
1
0
b
est
-
p
er
f
o
r
m
in
g
ca
s
es,
an
d
th
e
s
ec
o
n
d
r
o
w
s
h
o
ws th
e
1
0
wo
r
s
t
-
p
er
f
o
r
m
i
n
g
ca
s
es.
Fig
u
r
e
6
also
s
h
o
ws
ev
i
d
en
ce
f
o
r
o
u
r
p
r
ed
ictio
n
h
y
p
o
th
esis
.
I
n
ea
ch
ca
s
e,
t
h
e
f
ir
s
t
r
o
w
wit
h
th
e
b
est
r
esu
lts
co
n
s
is
ten
tly
ex
h
ib
its
a
s
tr
o
n
g
er
f
o
cu
s
o
n
im
p
o
r
tan
t
f
ac
ial
r
eg
io
n
s
th
an
th
e
s
ec
o
n
d
r
o
w
with
th
e
wo
r
s
t
r
esu
lts
.
T
h
is
also
h
elp
s
an
s
we
r
th
e
q
u
esti
o
n
o
f
th
e
r
elatio
n
s
h
ip
b
etwe
en
th
e
a
u
tis
m
d
iag
n
o
s
is
p
er
f
o
r
m
an
ce
o
f
th
e
d
ee
p
lear
n
in
g
m
o
d
el
an
d
th
e
m
o
d
el'
s
f
o
cu
s
o
n
im
p
o
r
tan
t
f
ac
ial
r
eg
io
n
s
,
wh
ich
ar
e
r
eg
io
n
s
th
at
ar
e
u
n
d
er
s
to
o
d
to
b
e
m
ar
k
e
d
b
y
la
n
d
m
ar
k
s
co
m
m
o
n
ly
u
s
ed
in
f
a
cial
ex
p
r
ess
io
n
s
tu
d
ies.
(
a)
(
b
)
Fig
u
r
e
5
.
So
m
e
Gr
ad
-
C
AM
h
e
atm
ap
s
u
s
in
g
th
e
p
r
o
p
o
s
ed
au
g
m
en
tatio
n
s
tr
ateg
y
f
o
r
(
a)
th
e
E
f
f
icien
tNet
-
B
4
m
o
d
el
an
d
(
b
)
th
e
R
esNet
-
1
8
m
o
d
el
Evaluation Warning : The document was created with Spire.PDF for Python.