I
AE
S In
t
er
na
t
i
o
na
l J
o
urna
l o
f
Art
if
icia
l In
t
ellig
ence
(
I
J
-
AI
)
Vo
l.
14
,
No
.
4
,
A
u
g
u
s
t
20
25
,
p
p
.
2
6
3
4
~
2
6
4
5
I
SS
N:
2
2
5
2
-
8938
,
DOI
: 1
0
.
1
1
5
9
1
/i
j
ai.
v
14
.i
4
.
p
p
2
6
3
4
-
2
6
4
5
2634
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Enha
ncing
f
a
ce
ma
sk
de
tec
tion per
f
o
r
m
a
nce w
ith
co
m
pre
hensiv
e d
a
tas
et
a
nd Y
O
L
O
v
8
T
ro
ng
T
hu
a
H
uy
nh
,
H
o
a
ng
T
ha
nh
Ng
uy
en
I
n
f
o
r
mat
i
o
n
S
e
c
u
r
i
t
y
T
e
c
h
n
o
l
o
g
y
L
a
b
,
F
a
c
u
l
t
y
o
f
I
n
f
o
r
mat
i
o
n
T
e
c
h
n
o
l
o
g
y
,
P
o
st
s a
n
d
T
e
l
e
c
o
mm
u
n
i
c
a
t
i
o
n
s
I
n
st
i
t
u
t
e
o
f
T
e
c
h
n
o
l
o
g
y
,
H
o
C
h
i
M
i
n
h
C
i
t
y
,
V
i
e
t
n
a
m
Art
icle
I
nfo
AB
ST
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Au
g
5
,
2
0
2
4
R
ev
i
s
ed
Ma
r
2
6
,
2
0
2
5
A
cc
ep
ted
J
u
n
8
,
2
0
2
5
In
th
e
c
o
n
tex
t
o
f
th
e
COV
ID
-
1
9
p
a
n
d
e
m
ic
a
n
d
th
e
risk
o
f
si
m
il
a
r
in
f
e
c
ti
o
u
s
d
ise
a
se
s,
m
o
n
it
o
ri
n
g
a
n
d
p
ro
m
o
ti
n
g
p
u
b
li
c
h
e
a
lt
h
m
e
a
su
re
s
li
k
e
w
e
a
rin
g
f
a
c
e
m
a
sk
s
h
a
v
e
b
e
c
o
m
e
c
ru
c
ial
in
c
o
n
tro
ll
in
g
v
iru
s
tra
n
sm
issio
n
.
De
e
p
lea
rn
i
n
g
-
b
a
se
d
m
a
sk
re
c
o
g
n
it
io
n
sy
st
e
m
s
p
la
y
a
n
im
p
o
rtan
t
ro
le,
b
u
t
t
h
e
ir
e
ffe
c
ti
v
e
n
e
ss
d
e
p
e
n
d
s
o
n
t
h
e
q
u
a
li
ty
a
n
d
d
iv
e
rsit
y
o
f
train
in
g
d
a
tas
e
ts.
T
h
is
stu
d
y
p
ro
p
o
se
s
th
e
d
iv
e
rse
a
n
d
ro
b
u
st
d
a
tas
e
t
f
o
r
f
a
c
e
m
a
sk
d
e
tec
ti
o
n
(
DRFM
D),
d
e
sig
n
e
d
to
a
d
d
re
ss
li
m
it
a
ti
o
n
s
o
f
e
x
isti
n
g
d
a
tas
e
t
s
a
n
d
e
n
h
a
n
c
e
m
a
sk
re
c
o
g
n
it
io
n
m
o
d
e
ls'
p
e
r
fo
rm
a
n
c
e
.
DRFM
D
in
teg
ra
tes
d
a
ta
f
ro
m
so
u
rc
e
s
su
c
h
a
s
A
IZOO,
f
a
c
e
m
a
sk
d
e
tec
to
r
b
y
Ka
ra
n
-
M
a
li
k
(
KFM
D
)
,
m
a
sk
e
d
f
a
c
e
s
(M
A
F
A
)
,
M
OX
A
3
K,
p
ro
p
e
rly
w
e
a
rin
g
m
a
s
k
e
d
f
a
c
e
d
e
tec
ti
o
n
d
a
tas
e
t
(P
W
M
F
D)
,
a
n
d
th
e
Zal
o
A
I
c
h
a
ll
e
n
g
e
2
0
2
2
,
c
o
m
p
risin
g
1
4
,
7
2
7
im
a
g
e
s
w
it
h
2
9
,
8
4
6
in
sta
n
c
e
s,
d
i
v
id
e
d
in
t
o
trai
n
in
g
,
v
a
li
d
a
ti
o
n
,
a
n
d
tes
ti
n
g
se
ts.
T
h
e
d
a
tas
e
t
'
s
s
c
a
le
a
n
d
d
iv
e
rsit
y
e
n
su
r
e
h
ig
h
e
r
a
c
c
u
ra
c
y
a
n
d
b
e
tt
e
r
g
e
n
e
ra
li
z
a
ti
o
n
f
o
r
m
a
sk
re
c
o
g
n
it
io
n
m
o
d
e
ls.
Ex
p
e
rim
e
n
ts
w
it
h
v
a
r
iatio
n
s
o
f
th
e
YO
L
O
v
8
m
o
d
e
l
(n
,
s,
m
,
l,
x
),
a
n
a
d
v
a
n
c
e
d
o
b
jec
t
d
e
tec
ti
o
n
a
lg
o
rit
h
m
,
o
n
th
e
DRFM
D
d
a
tas
e
t,
d
e
m
o
n
stra
te
su
p
e
rio
r
p
e
rf
o
r
m
a
n
c
e
th
ro
u
g
h
m
e
tri
c
s
li
k
e
p
r
e
c
isio
n
,
re
c
a
ll
,
a
n
d
m
A
P
@5
0
.
A
d
d
it
io
n
a
ll
y
,
c
o
m
p
a
ri
so
n
s
w
it
h
p
re
v
io
u
s
d
a
tas
e
t
li
k
e
F
M
M
D
s
h
o
w
th
a
t
m
o
d
e
ls
train
e
d
o
n
DRF
M
D
m
a
in
tain
stro
n
g
g
e
n
e
ra
li
z
a
ti
o
n
c
a
p
a
b
i
li
ti
e
s
a
n
d
h
ig
h
e
r
p
e
rf
o
rm
a
n
c
e
.
Th
is
stu
d
y
sig
n
if
ica
n
tl
y
c
o
n
tri
b
u
tes
to
im
p
ro
v
in
g
a
c
c
u
ra
c
y
o
f
p
u
b
li
c
h
e
a
lt
h
m
o
n
it
o
rin
g
s
y
ste
m
s,
a
id
in
g
in
th
e
p
re
v
e
n
ti
o
n
o
f
h
a
z
a
rd
s
f
ro
m
in
f
e
c
ti
o
u
s
d
ise
a
se
s
a
n
d
a
ir
p
o
ll
u
ti
o
n
.
K
ey
w
o
r
d
s
:
Data
s
et
Dee
p
lear
n
in
g
Face
m
as
k
d
etec
tio
n
Ob
j
ec
t d
etec
tio
n
YOL
O
v
8
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
r
o
n
g
T
h
u
a
Hu
y
n
h
I
n
f
o
r
m
a
tio
n
Sec
u
r
it
y
T
ec
h
n
o
l
o
g
y
L
ab
,
Facu
lt
y
o
f
I
n
f
o
r
m
ati
o
n
T
ec
h
n
o
lo
g
y
P
o
s
ts
an
d
T
elec
o
m
m
u
n
icatio
n
s
I
n
s
ti
tu
te
o
f
T
ec
h
n
o
lo
g
y
1
1
Ng
u
y
e
n
Di
n
h
C
h
ieu
Stre
et,
Dis
tr
ict
1
,
Ho
C
h
i M
i
n
h
C
it
y
,
Vietn
a
m
E
m
ail:
t
h
u
ah
t
@
p
tit.ed
u
.
v
n
1.
I
NT
RO
D
UCT
I
O
N
I
n
a
n
er
a
m
ar
k
ed
b
y
g
lo
b
al
h
ea
lt
h
c
h
alle
n
g
e
s
,
tec
h
n
o
lo
g
y
p
la
y
s
a
cr
u
cial
r
o
le
in
p
u
b
l
ic
s
a
f
et
y
,
p
ar
ticu
lar
l
y
t
h
r
o
u
g
h
f
ac
e
m
as
k
r
ec
o
g
n
itio
n
tec
h
n
o
lo
g
y
.
T
h
i
s
in
n
o
v
at
io
n
u
s
es
ad
v
a
n
ce
d
alg
o
r
ith
m
s
a
n
d
d
ee
p
lear
n
in
g
to
ac
cu
r
atel
y
id
e
n
ti
f
y
in
d
iv
id
u
a
ls
w
ea
r
in
g
m
as
k
s
,
e
n
s
u
r
in
g
co
m
p
lia
n
ce
w
it
h
p
u
b
li
c
h
ea
lth
g
u
id
eli
n
es.
T
h
e
u
r
g
en
t
n
ee
d
f
o
r
ef
f
ec
ti
v
e
p
r
ev
en
tiv
e
m
ea
s
u
r
es
d
u
r
in
g
th
e
C
OVI
D
-
1
9
p
an
d
e
m
i
c
h
ig
h
li
g
h
ted
th
e
li
m
ita
tio
n
s
o
f
m
a
n
u
al
e
n
f
o
r
ce
m
en
t,
p
r
o
m
p
ti
n
g
t
h
e
d
ev
el
o
p
m
e
n
t
o
f
au
to
m
ated
s
o
l
u
tio
n
s
li
k
e
f
ac
e
m
as
k
r
ec
o
g
n
itio
n
.
T
h
is
tec
h
n
o
lo
g
y
r
eliab
l
y
m
o
n
ito
r
s
co
m
p
lia
n
ce
in
cr
o
w
d
ed
p
u
b
lic
s
p
ac
es
s
u
ch
as
air
p
o
r
ts
an
d
s
h
o
p
p
in
g
ce
n
ter
s
,
r
ed
u
c
in
g
v
ir
u
s
tr
an
s
m
i
s
s
io
n
r
i
s
k
an
d
en
h
a
n
ci
n
g
p
u
b
lic
h
ea
lth
s
a
f
et
y
[
1
]
,
[
2
]
.
B
ey
o
n
d
m
o
n
ito
r
i
n
g
,
it
aid
s
d
ata
co
llectio
n
an
d
an
al
y
s
i
s
,
o
f
f
er
i
n
g
i
n
s
i
g
h
ts
i
n
to
co
m
p
lia
n
ce
r
ates
an
d
th
e
s
u
cc
es
s
o
f
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
A
r
ti
f
I
n
tell
I
SS
N:
2252
-
8938
E
n
h
a
n
ci
n
g
fa
ce
m
a
s
k
d
etec
tio
n
p
erfo
r
ma
n
ce
w
ith
co
mp
r
eh
e
n
s
ive
d
a
ta
s
et
a
n
d
YOLOv8
(
T
r
o
n
g
Th
u
a
Hu
yn
h
)
2635
h
ea
lt
h
ca
m
p
aig
n
s
.
A
d
d
itio
n
all
y
,
it
s
i
n
te
g
r
atio
n
w
it
h
o
t
h
er
s
u
r
v
eilla
n
ce
s
y
s
te
m
s
f
o
r
m
s
a
co
m
p
r
e
h
e
n
s
i
v
e
p
u
b
lic
s
af
et
y
n
e
t
w
o
r
k
,
b
o
o
s
tin
g
p
r
ep
ar
ed
n
ess
f
o
r
f
u
t
u
r
e
h
ea
l
th
cr
is
es.
Dee
p
lear
n
in
g
ap
p
r
o
ac
h
es
u
s
ed
f
o
r
m
as
k
r
ec
o
g
n
itio
n
f
o
cu
s
o
n
id
en
tify
i
n
g
an
d
ca
teg
o
r
izin
g
in
d
iv
id
u
als
ac
co
r
d
in
g
to
t
h
eir
ad
h
er
en
ce
to
m
as
k
-
w
ea
r
in
g
p
r
o
to
co
ls
.
Am
o
n
g
t
h
ese
m
e
th
o
d
s
,
co
n
v
o
l
u
tio
n
a
l
n
eu
r
al
n
e
t
w
o
r
k
s
(
C
N
Ns)
s
ta
n
d
o
u
t
d
u
e
to
th
eir
ex
ce
p
tio
n
al
ca
p
ab
ilit
ies
in
i
m
ag
e
a
n
a
l
y
s
i
s
,
m
a
k
i
n
g
t
h
e
m
p
ar
ticu
lar
l
y
e
f
f
ec
ti
v
e
f
o
r
m
as
k
d
etec
tio
n
ap
p
licatio
n
s
[
3
]
.
Fa
n
et
a
l
.
[
4
]
p
r
o
p
o
s
ed
a
lig
h
t
w
e
ig
h
t
m
as
k
d
etec
tio
n
s
u
ite
b
ased
o
n
d
ee
p
lear
n
i
n
g
,
co
m
b
in
in
g
t
h
e
r
esid
u
al
c
o
n
tex
t
atte
n
tio
n
m
o
d
u
le
(
R
C
A
M)
a
n
d
Gau
s
s
ia
n
h
ea
t
m
ap
r
eg
r
ess
io
n
(
SGHR
)
t
o
im
p
r
o
v
e
f
ea
tu
r
e
ex
tr
ac
tio
n
ca
p
ab
ilit
ies.
Usi
n
g
Mo
b
ileNe
t
as
th
e
b
ac
k
b
o
n
e,
th
is
m
o
d
el
i
s
s
u
itab
le
f
o
r
e
m
b
ed
d
ed
s
y
s
te
m
s
an
d
ac
h
ie
v
es
h
i
g
h
p
er
f
o
r
m
a
n
ce
.
R
C
AM
h
elp
s
f
o
cu
s
o
n
i
m
p
o
r
tan
t
m
a
s
k
-
r
elate
d
ar
ea
s
,
w
h
ile
SGH
R
lear
n
s
th
e
d
is
ti
n
g
u
i
s
h
in
g
f
ea
t
u
r
es
b
et
w
ee
n
m
a
s
k
ed
a
n
d
u
n
m
as
k
ed
f
ac
es.
T
ests
o
n
t
h
e
A
I
Z
OO
a
n
d
Mo
x
a3
K
d
ataset
s
s
h
o
w
ed
t
h
e
m
o
d
el
ac
h
iev
ed
h
i
g
h
er
m
ea
n
a
v
er
ag
e
p
r
ec
is
io
n
(
m
A
P
)
th
a
n
YO
L
Ov
3
-
ti
n
y
b
y
1
.
7
%
a
n
d
1
0
.
4
7
%,
r
esp
ec
tiv
el
y
.
J
o
o
d
i
et
a
l.
[
5
]
in
tr
o
d
u
ce
s
a
n
o
v
el
d
ee
p
lear
n
in
g
m
o
d
el
f
o
r
m
a
s
k
d
etec
tio
n
,
s
tr
u
ct
u
r
ed
in
t
w
o
s
ta
g
e
s
:
f
ac
e
d
etec
tio
n
u
s
i
n
g
t
h
e
Ha
ar
ca
s
ca
d
e
d
etec
to
r
an
d
class
i
f
icat
io
n
w
it
h
a
C
NN
m
o
d
el
b
u
ilt
f
r
o
m
s
cr
atc
h
.
E
x
p
er
i
m
e
n
t
s
co
n
d
u
cted
u
s
in
g
t
h
e
b
en
c
h
m
ar
k
m
as
k
ed
f
ac
es
(
M
AF
A
)
d
ata
s
et
ac
h
iev
ed
r
elati
v
el
y
h
i
g
h
ac
cu
r
ac
y
ac
r
o
s
s
d
i
f
f
er
e
n
t
l
ea
r
n
in
g
r
ates
w
h
ile
m
ai
n
tai
n
in
g
lo
w
co
m
p
u
tat
io
n
al
c
o
m
p
le
x
it
y
.
Si
m
ilar
l
y
,
t
h
e
p
r
o
p
o
s
ed
m
o
d
el
in
[
6
]
en
ab
les
r
ea
l
-
ti
m
e
m
a
s
k
-
w
ea
r
i
n
g
r
ec
o
g
n
itio
n
b
ased
o
n
th
e
Mo
b
ileNetV2
ar
ch
itectu
r
e,
ap
p
licab
le
to
em
b
ed
d
ed
d
e
v
ices
s
u
c
h
as
th
e
NVI
DI
A
J
etso
n
Na
n
o
.
E
x
p
er
im
en
tal
r
es
u
lts
in
d
icate
a
v
er
y
h
i
g
h
ac
c
u
r
ac
y
r
ate
in
b
o
th
tr
ain
i
n
g
a
n
d
test
in
g
.
T
h
e
m
o
d
el
i
s
also
d
esi
g
n
ed
to
b
e
lig
h
t
w
eig
h
t
a
n
d
ef
f
ic
ien
t,
s
u
p
p
o
r
tin
g
m
u
lti
-
m
as
k
d
etec
tio
n
,
w
h
ic
h
i
s
b
en
ef
icia
l in
cr
o
w
d
ed
en
v
ir
o
n
m
en
ts
w
h
er
e
m
u
ltip
le
i
n
d
iv
id
u
als n
ee
d
to
b
e
m
o
n
ito
r
ed
s
i
m
u
ltan
eo
u
s
l
y
.
Sev
er
al
s
tu
d
ie
s
h
av
e
e
x
p
l
o
r
ed
d
ee
p
lear
n
in
g
m
o
d
els
f
o
r
m
as
k
d
etec
tio
n
.
R
e
s
ea
r
ch
b
y
Kh
o
r
a
m
d
el
et
a
l.
[
7
]
,
th
r
ee
m
o
d
els
(
SS
D,
YO
L
O
v
4
-
t
in
y
,
a
n
d
YOL
O
v
4
-
tin
y
-
3
l)
w
er
e
test
e
d
o
n
1
,
5
3
1
im
a
g
es,
w
it
h
YO
L
O
v
4
-
ti
n
y
ac
h
iev
in
g
th
e
h
i
g
h
e
s
t
m
A
P
(
8
5
.
3
1
%)
an
d
5
0
.
6
6
FP
S,
s
u
itab
le
f
o
r
r
ea
l
-
ti
m
e
u
s
e
.
Al
-
D
m
o
u
r
et
a
l.
[
8
]
p
r
o
p
o
s
ed
a
C
NN
-
b
ased
s
y
s
te
m
to
r
ec
o
g
n
ize
co
v
er
ed
f
ac
es,
ac
h
ie
v
i
n
g
h
i
g
h
ac
cu
r
ac
y
i
n
d
is
tin
g
u
is
h
i
n
g
m
as
k
ed
f
r
o
m
u
n
m
a
s
k
ed
f
ac
e
s
.
A
d
d
itio
n
all
y
,
t
h
e
DB
-
YO
L
O
m
a
s
k
d
ete
ctio
n
al
g
o
r
ith
m
[
9
]
,
in
te
g
r
ated
in
to
an
An
d
r
o
id
ap
p
,
d
em
o
n
s
tr
ated
h
ig
h
p
r
ec
is
io
n
an
d
a
d
etec
tio
n
s
p
ee
d
o
f
3
3
FP
S
u
s
in
g
a
lig
h
t
w
ei
g
h
t
ar
ch
itect
u
r
e
b
ase
d
o
n
YO
L
O
v
5
,
o
p
ti
m
ized
f
o
r
m
o
b
ile
d
ev
ices.
A
b
u
r
ae
d
et
a
l.
[
1
0
]
co
m
p
ar
ed
YOL
O
v
5
a
n
d
YO
L
Ov
6
f
o
r
d
etec
tin
g
i
m
p
ac
t
cr
ater
s
o
n
m
ar
s
a
n
d
th
e
m
o
o
n
.
T
h
e
r
e
s
u
lt
s
i
n
d
icate
t
h
at
YOL
O
v
6
o
u
tp
er
f
o
r
m
ed
YOL
Ov
5
in
s
p
ee
d
an
d
ac
cu
r
ac
y
w
i
th
A
d
a
m
o
p
ti
m
izer
.
T
h
e
YOL
O
v
8
alg
o
r
it
h
m
[
1
1
]
r
ep
r
esen
ts
an
ad
v
an
ce
d
o
b
j
ec
t
d
etec
tio
n
f
r
a
m
e
w
o
r
k
,
r
en
o
w
n
ed
f
o
r
its
ex
ce
lle
n
t
ac
c
u
r
ac
y
,
r
ea
l
-
ti
m
e
p
r
o
ce
s
s
in
g
ca
p
ab
ilit
ies,
a
n
d
r
o
b
u
s
t
p
er
f
o
r
m
an
ce
.
As
a
n
ev
o
l
u
tio
n
o
f
t
h
e
YO
L
O
s
er
ies,
YO
L
O
v
8
m
ai
n
tai
n
s
th
e
f
u
n
d
a
m
e
n
tal
p
r
in
c
ip
l
e
o
f
p
er
f
o
r
m
i
n
g
o
b
j
ec
t
d
etec
tio
n
in
a
s
in
g
le
p
a
s
s
t
h
r
o
u
g
h
a
n
e
u
r
al
n
e
t
w
o
r
k
[
1
2
]
,
[
1
3
]
,
m
ak
in
g
i
t
h
ig
h
l
y
e
f
f
icie
n
t
a
n
d
s
u
itab
le
f
o
r
r
ea
l
-
ti
m
e
ap
-
p
l
icatio
n
s
.
YO
L
Ov
8
e
m
p
lo
y
s
a
d
ee
p
co
n
v
o
l
u
tio
n
al
n
e
u
r
al
n
e
t
w
o
r
k
(
D
C
NN)
w
it
h
m
u
ltip
le
co
n
v
o
lu
tio
n
al
la
y
e
r
s
,
d
o
w
n
-
s
a
m
p
li
n
g
,
an
d
u
p
-
s
a
m
p
li
n
g
o
p
er
atio
n
s
.
T
h
is
ar
ch
itectu
r
e
allo
w
s
f
o
r
ca
p
tu
r
in
g
f
ea
t
u
r
es
at
v
ar
io
u
s
s
c
ales
an
d
p
r
eser
v
in
g
cr
u
cial
s
p
atial
i
n
f
o
r
m
atio
n
f
o
r
p
r
ec
is
e
o
b
j
ec
t
id
en
tif
icatio
n
an
d
lo
ca
lizatio
n
.
YO
L
O
v
8
p
r
o
ce
s
s
es
th
e
i
n
p
u
t
i
m
a
g
e
b
y
s
e
g
m
e
n
ti
n
g
it
in
to
a
g
r
id
,
w
h
er
e
ea
c
h
ce
l
l
i
s
r
esp
o
n
s
ib
le
f
o
r
p
r
ed
ictin
g
b
o
u
n
d
i
n
g
b
o
x
e
s
alo
n
g
w
it
h
th
eir
co
r
r
esp
o
n
d
in
g
cla
s
s
p
r
o
b
ab
ilit
ies.
T
h
is
s
tr
u
ct
u
r
ed
g
r
i
d
-
b
ased
tech
n
iq
u
e
allo
w
s
YO
L
O
v
8
to
e
f
f
ec
tiv
el
y
id
en
ti
f
y
m
u
lt
ip
le
o
b
j
ec
ts
w
i
th
i
n
a
n
i
m
a
g
e,
ac
co
m
m
o
d
atin
g
v
ar
iatio
n
s
i
n
s
ize
a
n
d
a
s
p
ec
t
r
at
io
[
1
0
]
.
Mo
r
e
o
v
er
,
YOL
O
v
8
i
n
teg
r
ate
s
ad
v
a
n
ce
d
tech
n
iq
u
es
s
u
c
h
as
b
atch
n
o
r
m
aliza
t
io
n
,
d
r
o
p
o
u
t,
an
d
co
m
p
le
x
ac
ti
v
atio
n
f
u
n
ctio
n
s
,
i
m
p
r
o
v
i
n
g
ac
c
u
r
ac
y
a
n
d
r
ec
all
r
ate
s
co
m
p
ar
ed
t
o
p
r
ev
io
u
s
v
er
s
io
n
s
.
T
h
ese
e
n
h
an
ce
m
e
n
ts
r
ed
u
ce
er
r
o
r
s
an
d
im
p
r
o
v
e
r
eliab
ilit
y
i
n
d
etec
tin
g
o
b
j
ec
ts
ac
r
o
s
s
d
if
f
er
en
t scen
ar
io
s
.
De
w
i
et
a
l.
[
1
4
]
in
tr
o
d
u
ce
d
an
d
u
t
ilized
a
d
atase
t
w
e
r
e
f
er
to
as
f
ac
e
a
n
d
m
ed
ical
m
ask
d
ata
s
et
(
FMM
D
)
,
w
h
ich
i
s
a
co
m
b
in
a
tio
n
o
f
th
e
f
ac
e
m
a
s
k
d
ataset
(
FMD)
[
1
5
]
an
d
th
e
m
ed
ical
m
ask
d
ataset
(
MM
D)
[
1
6
]
,
in
tr
ain
in
g
a
n
d
ev
al
u
ati
n
g
m
as
k
r
ec
o
g
n
itio
n
.
Ho
w
e
v
er
,
FMM
D
s
till
h
as
s
o
m
e
l
i
m
itat
io
n
s
,
s
u
c
h
as
s
m
a
ll
s
ca
le,
lack
o
f
d
ata
s
o
u
r
ce
d
iv
er
s
it
y
,
an
d
in
s
u
f
f
icie
n
t
lab
el
q
u
alit
y
an
d
d
etail,
w
ith
o
n
l
y
1
,
0
6
7
im
ag
e
s
an
d
5
,
7
9
6
in
s
tan
ce
s
.
T
o
ad
d
r
ess
th
ese
li
m
i
tatio
n
s
an
d
e
n
h
a
n
ce
th
e
e
f
f
ec
t
iv
e
n
es
s
o
f
m
as
k
r
ec
o
g
n
itio
n
m
o
d
el
s
,
w
e
p
r
o
p
o
s
e
a
n
e
w
,
m
o
r
e
d
iv
er
s
e,
an
d
r
o
b
u
s
t
d
ataset
ca
lled
d
iv
er
s
e
an
d
r
o
b
u
s
t
d
ataset
f
o
r
f
a
ce
m
as
k
d
etec
tio
n
(
DR
FMD)
.
A
d
d
itio
n
all
y
,
th
is
w
o
r
k
e
x
p
lo
r
es a
n
d
an
al
y
ze
s
t
h
e
h
u
m
a
n
in
t
h
e
lo
o
p
(
HI
T
L
)
-
MM
D
[
1
7
]
,
an
o
p
en
-
ac
ce
s
s
d
ataset
d
esig
n
ed
to
co
n
tr
ib
u
te
to
th
e
g
lo
b
al
f
i
g
h
t
a
g
ai
n
s
t
C
OVI
D
-
1
9
.
HI
T
L
-
MM
D
p
r
o
v
id
es
a
r
ich
an
d
d
iv
er
s
e
d
ata
s
o
u
r
ce
,
s
u
p
p
le
m
e
n
ti
n
g
a
n
d
i
m
p
r
o
v
i
n
g
u
p
o
n
e
x
is
tin
g
r
esear
ch
.
T
h
e
m
ai
n
co
n
tr
ib
u
tio
n
s
o
f
t
h
is
w
o
r
k
ar
e:
i
)
co
n
s
tr
u
cti
n
g
a
lar
g
e
-
s
ca
le
a
n
d
d
i
v
er
s
e
d
ataset
w
i
t
h
1
0
,
3
0
4
im
ag
e
s
an
d
2
0
,
6
0
3
in
s
tan
ce
s
to
i
m
p
r
o
v
e
r
ec
o
g
n
itio
n
p
er
f
o
r
m
a
n
ce
co
m
p
ar
ed
to
p
r
ev
io
u
s
d
atasets
.
ii
)
i
m
p
le
m
en
t
in
g
a
d
ee
p
lear
n
i
n
g
-
b
ased
o
b
j
ec
t
d
etec
tio
n
m
o
d
el
th
at
ca
n
a
u
to
m
a
ti
ca
ll
y
id
en
t
if
y
an
d
lo
ca
te
f
ac
e
s
w
it
h
o
u
t
m
as
k
s
,
f
ac
es
w
it
h
m
ask
s
,
an
d
i
m
p
r
o
p
er
ly
w
o
r
n
m
as
k
s
i
n
i
m
a
g
es.
iii
)
a
n
al
y
zi
n
g
a
n
d
co
m
p
ar
in
g
YOL
O
v
8
n
,
YO
L
O
v
8
s
,
YO
L
Ov
8
m
,
YO
L
O
v
8
l,
an
d
YO
L
Ov
8
x
m
o
d
els
to
id
en
t
if
y
a
n
d
ev
alu
a
te
t
h
e
b
en
e
f
it
s
an
d
li
m
itatio
n
s
r
elate
d
to
u
s
i
n
g
YO
L
O
v
8
in
m
as
k
r
ec
o
g
n
itio
n
s
y
s
te
m
s
.
B
esid
es
t
h
e
i
n
t
r
o
d
u
ctio
n
,
w
e
w
ill
p
r
esen
t th
e
d
atase
t c
o
n
s
tr
u
ctio
n
m
e
th
o
d
o
lo
g
y
i
n
s
ec
t
io
n
2
.
S
ec
tio
n
3
illu
s
tr
ates tr
ai
n
in
g
r
es
u
lts
,
e
v
al
u
atio
n
a
n
d
d
is
cu
s
s
io
n
.
Fi
n
all
y
,
co
n
cl
u
s
io
n
w
ill b
e
p
r
o
v
id
ed
in
s
ec
tio
n
4
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8938
I
n
t J
A
r
ti
f
I
n
tell
,
Vo
l.
14
,
No
.
4
,
A
u
g
u
s
t
2025
:
2
6
3
4
-
2645
2636
2.
M
E
T
H
O
D
I
n
th
i
s
s
ec
tio
n
,
w
e
b
r
ie
f
l
y
in
tr
o
d
u
ce
t
w
o
d
atase
ts
u
s
ed
i
n
p
r
ev
io
u
s
s
t
u
d
ies,
n
a
m
el
y
FMM
D
[
1
4
]
an
d
HI
T
L
-
MM
D
[
1
7
]
.
T
h
ese
d
atasets
p
la
y
a
cr
u
cia
l
r
o
le
in
r
esear
ch
o
n
f
ac
e
r
ec
o
g
n
itio
n
w
i
th
an
d
w
i
th
o
u
t
m
a
s
k
s
,
p
r
o
v
id
in
g
a
r
eliab
le
r
e
f
er
en
c
e
s
o
u
r
ce
f
o
r
co
m
p
u
ter
v
i
s
io
n
-
b
ased
r
ec
o
g
n
i
tio
n
s
y
s
te
m
s
.
Nex
t,
w
e
p
r
ese
n
t
a
d
etailed
ex
p
lan
atio
n
o
f
th
e
m
e
th
o
d
o
lo
g
y
u
s
ed
to
co
n
s
tr
u
ct
th
e
p
r
o
p
o
s
ed
d
ataset,
ca
lled
DR
FMD
.
T
h
is
s
ec
tio
n
w
il
l
co
v
er
th
e
d
ata
co
llectio
n
p
r
o
ce
s
s
,
th
e
cr
iter
ia
f
o
r
i
m
ag
e
s
elec
tio
n
,
p
r
ep
r
o
ce
s
s
in
g
s
tep
s
,
an
d
k
e
y
f
ea
t
u
r
es
th
at
m
a
k
e
D
R
FMD
a
d
iv
er
s
e
an
d
r
o
b
u
s
t
d
ataset.
T
h
ese
ch
ar
ac
ter
is
tics
en
s
u
r
e
it
s
ef
f
ec
ti
v
en
es
s
in
s
u
p
p
o
r
tin
g
m
o
d
el
s
f
o
r
f
ac
e
m
as
k
d
etec
tio
n
.
2
.
1
.
F
a
ce
m
a
s
k
da
t
a
s
et
a
nd
m
e
dica
l
m
a
s
k
da
t
a
s
et
De
w
i
et
a
l.
[
1
4
]
in
tr
o
d
u
ce
d
a
n
d
u
s
ed
t
h
e
FMM
D
d
ata
s
et,
w
h
ic
h
is
a
co
m
b
i
n
atio
n
o
f
F
MD
[
1
5
]
an
d
MM
D
[
1
6
]
,
f
o
r
tr
ain
in
g
a
n
d
ev
alu
ati
n
g
m
a
s
k
r
ec
o
g
n
itio
n
w
it
h
an
in
p
u
t
r
eso
lu
t
io
n
(
im
ag
e
s
ize)
o
f
4
1
6
.
T
h
e
FMD
is
a
p
u
b
licl
y
av
a
il
ab
le
d
ataset
o
f
M
A
F
A
,
w
h
ic
h
in
c
lu
d
es
8
5
3
i
m
a
g
es,
s
to
r
ed
in
P
A
S
C
AL
VO
C
f
o
r
m
at.
T
h
e
MM
D
i
n
cl
u
d
es 6
8
2
i
m
ag
e
s
,
w
it
h
o
v
er
3
,
0
0
0
MA
F
A
w
ea
r
i
n
g
m
ed
ical
m
as
k
s
.
T
h
e
co
m
b
i
n
atio
n
o
f
th
ese
t
w
o
d
ataset
s
r
esu
lted
i
n
a
d
is
tin
ct
an
d
m
o
r
e
ex
te
n
s
iv
e
d
ataset,
w
it
h
a
to
tal
o
f
1
,
4
1
5
co
llected
im
a
g
es
u
n
d
er
g
o
i
n
g
a
r
ig
o
r
o
u
s
s
elec
ti
o
n
p
r
o
ce
s
s
.
L
o
w
-
q
u
alit
y
o
r
d
u
p
licate
i
m
a
g
es
f
r
o
m
t
h
e
o
r
ig
in
a
l
d
atase
t
w
er
e
r
e
m
o
v
ed
to
en
s
u
r
e
th
e
q
u
alit
y
an
d
co
n
s
is
te
n
c
y
o
f
th
e
f
in
a
l
d
ataset.
W
e
w
ill
u
s
e
th
i
s
d
ataset
to
ev
alu
ate
a
n
d
co
m
p
ar
e
it
w
it
h
o
u
r
p
r
o
p
o
s
ed
d
ataset
(
DR
FMD)
i
n
th
e
n
e
x
t
s
ec
tio
n
.
2
.
2
.
H
um
a
n i
n
t
he
lo
o
p
m
e
d
ica
l
m
a
s
k da
t
a
s
et
H
I
T
L
[
1
7
]
p
r
o
v
i
d
e
s
a
n
o
p
en
-
a
cc
e
s
s
d
a
t
as
e
t
d
es
ig
n
e
d
t
o
s
u
p
p
o
r
t
g
l
o
b
a
l
ef
f
o
r
t
s
in
c
o
m
b
a
t
in
g
C
O
V
I
D
-
1
9
.
T
h
is
d
ataset
co
m
p
r
is
es
6
,
0
0
0
p
u
b
licl
y
av
ai
lab
le
i
m
a
g
es,
c
ar
ef
u
ll
y
c
u
r
ated
to
en
s
u
r
e
d
iv
er
s
it
y
b
y
i
n
cl
u
d
in
g
in
d
iv
id
u
als
f
r
o
m
v
ar
io
u
s
et
h
n
ic
b
ac
k
g
r
o
u
n
d
s
,
ag
e
g
r
o
u
p
s
,
a
n
d
g
eo
g
r
ap
h
ic
r
eg
io
n
s
.
F
u
r
th
er
m
o
r
e,
i
t
in
co
r
p
o
r
ates
2
0
d
is
tin
ct
t
y
p
es
o
f
ac
ce
s
s
o
r
ies
a
n
d
ca
teg
o
r
izes
f
ac
ial
i
m
ag
e
s
in
to
th
r
ee
g
r
o
u
p
s
:
w
ea
r
i
n
g
m
a
s
k
s
co
r
r
ec
tly
,
n
o
t
w
ea
r
i
n
g
m
as
k
s
,
an
d
i
m
p
r
o
p
er
m
a
s
k
u
s
a
g
e.
T
h
e
d
ataset
w
a
s
co
m
p
i
led
an
d
a
n
n
o
tated
b
y
r
e
f
u
g
ee
w
o
r
k
er
s
a
f
f
iliated
w
it
h
HI
T
L
in
B
u
l
g
ar
ia.
T
o
p
r
o
m
o
te
ac
ce
s
s
ib
ilit
y
an
d
b
r
o
ad
er
u
s
ag
e,
t
h
is
MMD
h
as
b
ee
n
r
elea
s
ed
in
to
th
e
p
u
b
lic
d
o
m
a
in
u
n
d
er
th
e
C
C
0
1
.
0
licen
s
e.
I
n
t
h
i
s
s
tu
d
y
,
w
e
u
s
e
t
h
e
HI
T
L
-
MM
D
d
ata
s
et
to
ev
alu
a
te
th
e
m
o
d
el
tr
ain
ed
o
n
th
e
p
r
ev
io
u
s
FMM
D
d
at
aset
an
d
o
u
r
p
r
o
p
o
s
ed
d
ataset.
Fo
r
th
e
ev
al
u
atio
n
,
w
e
u
s
ed
th
e
L
ab
elI
m
g
to
o
l
[
1
8
]
to
an
n
o
tate
d
ata
f
o
r
1
,
3
1
1
im
ag
es
w
it
h
a
to
tal
o
f
1
,
5
9
8
in
s
tan
ce
s
,
w
h
er
e
th
e
lab
els
w
ith
o
u
t
m
as
k
(
0
)
,
w
ith
m
as
k
(
1
)
,
an
d
w
ea
r
m
a
s
k
i
n
co
r
r
ec
t
(
2
)
ar
e
4
6
2
,
1
,
0
3
0
,
an
d
1
0
6
r
esp
ec
tiv
el
y
.
2
.
3
.
Div
er
s
e
a
nd
ro
bu
s
t
da
t
a
s
et
f
o
r
f
a
ce
m
a
s
k
det
ec
t
io
n
T
h
e
d
ataset
w
e
p
r
o
p
o
s
e
in
t
h
is
s
tu
d
y
is
ca
l
led
DR
F
MD
,
w
h
ic
h
is
co
llected
f
r
o
m
v
ar
io
u
s
s
o
u
r
ce
s
w
it
h
p
ar
tial
o
r
co
m
p
lete
d
ata
f
r
o
m
A
I
Z
OO
[
1
9
]
,
f
ac
e
m
as
k
d
et
ec
to
r
b
y
Kar
a
n
-
Ma
li
k
(
KFM
D
)
[
2
0
]
,
MO
XA
3
k
[
2
1
]
,
MA
F
A
[
2
2
]
,
an
d
p
r
o
p
er
ly
w
ea
r
i
n
g
m
as
k
ed
f
ac
e
d
etec
tio
n
d
ataset
(
P
W
MFD)
[
2
3
]
.
T
h
e
d
ataset
co
n
s
tr
u
ct
io
n
m
et
h
o
d
is
d
escr
ib
ed
in
Fig
u
r
e
1
.
T
h
e
DR
F
MD
d
ataset
is
b
u
i
lt
b
y
a
g
g
r
eg
ati
n
g
v
ar
io
u
s
d
ata
s
o
u
r
ce
s
to
e
n
r
ic
h
t
h
e
f
in
a
l
d
at
aset.
E
ac
h
in
p
u
t
d
ataset
is
p
ar
tiall
y
o
r
f
u
ll
y
u
s
ed
,
t
h
en
r
e
v
ie
w
ed
a
n
d
r
ef
i
n
ed
as
n
ec
es
s
ar
y
,
an
d
f
i
n
all
y
co
n
v
er
t
ed
to
YOL
O
lab
elin
g
s
ta
n
d
ar
d
s
.
A
d
d
itio
n
all
y
,
w
e
co
llected
an
d
lab
eled
s
o
m
e
d
ata
f
r
o
m
th
e
Z
alo
A
I
c
h
alle
n
g
e
2
0
2
2
d
ataset
[
2
4
]
,
an
d
ad
j
u
s
ted
s
o
m
e
lab
els o
f
t
h
e
a
f
o
r
e
m
en
tio
n
ed
d
ataset
s
.
Fig
u
r
e
1
.
T
h
e
p
r
o
ce
s
s
o
f
cr
ea
tin
g
t
h
e
D
R
FMD
d
ataset
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
A
r
ti
f
I
n
tell
I
SS
N:
2252
-
8938
E
n
h
a
n
ci
n
g
fa
ce
m
a
s
k
d
etec
tio
n
p
erfo
r
ma
n
ce
w
ith
co
mp
r
eh
e
n
s
ive
d
a
ta
s
et
a
n
d
YOLOv8
(
T
r
o
n
g
Th
u
a
Hu
yn
h
)
2637
T
h
e
DR
FMD
d
ataset
i
n
cl
u
d
es
1
4
,
7
2
7
im
a
g
es
w
i
th
2
9
,
8
4
6
in
s
tan
ce
s
,
co
m
p
r
is
i
n
g
1
0
,
3
0
4
in
s
tan
ce
s
f
o
r
th
e
tr
ai
n
in
g
s
e
t,
1
,
4
7
4
in
s
tan
c
es
f
o
r
th
e
v
alid
atio
n
s
et,
a
n
d
2
,
9
4
9
in
s
ta
n
ce
s
f
o
r
th
e
tes
t
s
e
t.
W
e
n
o
ticed
th
at
FMD
o
f
te
n
lack
th
e
n
u
m
b
er
o
f
in
s
tan
ce
s
o
f
i
m
p
r
o
p
e
r
l
y
w
o
r
n
m
a
s
k
s
.
T
h
er
ef
o
r
e,
w
e
f
o
cu
s
ed
o
n
ex
tr
ac
tin
g
al
l
i
m
a
g
es
w
i
th
i
m
p
r
o
p
er
l
y
w
o
r
n
m
a
s
k
s
f
r
o
m
t
h
e
M
A
F
A
a
n
d
KFMD
d
atase
ts
.
T
h
e
KF
MD
d
ataset
[
2
0
]
,
w
h
ic
h
o
r
ig
in
a
ll
y
o
n
l
y
h
ad
lab
els
f
o
r
w
ea
r
i
n
g
m
a
s
k
s
an
d
n
o
t
w
ea
r
i
n
g
m
as
k
s
,
w
a
s
an
n
o
tat
ed
w
it
h
i
m
p
r
o
p
er
l
y
w
o
r
n
m
as
k
lab
els
u
s
i
n
g
t
h
e
L
ab
elI
m
g
to
o
l.
W
e
th
en
ad
j
u
s
te
d
th
e
lab
els
to
en
s
u
r
e
t
h
e
ac
cu
r
ac
y
o
f
ea
ch
lab
el.
Fo
r
th
e
A
I
Z
O
O
-
T
ec
h
an
d
M
OX
A
3
K
d
atasets
,
w
h
ic
h
o
n
l
y
h
ad
lab
els
f
o
r
w
ea
r
in
g
m
as
k
s
an
d
n
o
t
w
ea
r
i
n
g
m
as
k
s
,
w
e
ad
d
ed
i
m
p
r
o
p
er
l
y
w
o
r
n
m
as
k
lab
el
s
a
n
d
ex
tr
ac
ted
p
ar
t
o
f
t
h
e
d
ata.
T
h
e
P
W
F
MD
d
ataset
is
q
u
it
e
lar
g
e,
w
ith
9
,
2
0
5
i
m
a
g
es
o
f
v
a
r
y
in
g
s
ize
s
,
f
r
o
m
s
m
all
to
lar
g
e,
s
o
w
e
e
x
tr
ac
ted
a
s
u
b
s
et
o
f
4
,
7
0
3
im
a
g
e
s
.
T
h
e
p
ar
am
eter
s
o
f
t
h
ese
d
ata
s
o
u
r
ce
s
ar
e
d
etailed
in
T
ab
le
1
,
w
ith
all
i
m
a
g
es
i
n
th
e
d
ata
s
ets
b
ein
g
r
ea
l
i
m
ag
e
s
.
Mo
s
t
d
ataset
s
ar
e
lab
eled
f
o
r
all
i
m
ag
e
s
ize
s
(
s
m
all,
m
ed
i
u
m
,
lar
g
e)
,
ex
ce
p
t
f
o
r
t
h
e
M
AF
A
d
ataset,
w
h
ich
is
lab
eled
o
n
l
y
f
o
r
m
ed
i
u
m
a
n
d
l
ar
g
e
s
izes.
T
ab
le
1
.
Su
m
m
ar
y
o
f
s
o
m
e
p
ar
a
m
et
er
s
o
f
o
p
en
d
ataset
s
f
o
r
f
ac
e
d
etec
tio
n
to
b
u
ild
th
e
DR
F
MD
[
2
5
]
D
a
t
a
se
t
n
a
me
K
e
y
f
e
a
t
u
r
e
s
N
u
mb
e
r
o
f
i
mag
e
s
C
l
a
ssi
f
i
c
a
t
i
o
n
M
a
s
k
n
u
mb
e
r
H
e
a
d
p
o
si
t
i
o
n
S
c
e
n
c
e
A
I
ZO
O
-
Te
c
h
T
h
e
d
a
t
a
se
t
i
s
c
r
e
a
t
e
d
b
y
mo
d
i
f
y
i
n
g
i
n
c
o
r
r
e
c
t
a
n
n
o
t
a
t
i
o
n
s
f
r
o
m
t
h
e
W
I
D
ER
F
a
c
e
a
n
d
M
A
F
A
d
a
t
a
se
t
s
T
r
a
i
n
:
6
,
1
3
0
V
a
l
i
d
:
1
,
8
3
9
Tw
o
1
2
,
6
2
0
f
a
c
e
s
w
i
t
h
o
u
t
mask
s
;
4
,
0
3
4
M
A
F
A
D
i
v
e
r
si
t
y
M
e
d
i
u
m
K
F
M
D
[
2
0
]
I
mag
e
s c
r
e
a
t
e
d
b
y
K
a
r
a
n
1
,
5
0
8
Tw
o
7
5
3
f
a
c
e
s
w
i
t
h
o
u
t
mas
k
;
7
5
5
f
a
c
e
w
i
t
h
mask
e
d
;
D
i
v
e
r
si
t
y
M
e
d
i
u
m
M
A
F
A
T
h
e
i
mag
e
s
a
r
e
f
r
o
m
t
h
e
I
n
t
e
r
n
e
t
,
a
n
n
o
t
a
t
e
d
w
i
t
h
si
x
a
t
t
r
i
b
u
t
e
s
f
o
r
e
a
c
h
f
a
c
e
r
e
g
i
o
n
,
si
m
i
l
a
r
t
o
t
h
e
mo
r
e
o
c
c
l
u
d
e
d
f
a
c
e
d
a
t
a
se
t
s
3
0
,
8
1
1
M
a
n
y
t
y
p
e
s
o
f
mask
s
3
5
,
8
0
6
M
A
F
A
D
i
v
e
r
si
t
y
C
o
mp
l
e
x
M
O
X
A
3
K
T
h
e
i
mag
e
s
c
o
me
f
r
o
m
K
a
g
g
l
e
d
a
t
a
se
t
s
f
e
a
t
u
r
i
n
g
d
a
t
a
f
r
o
m
R
u
ss
i
a
,
I
t
a
l
y
,
C
h
i
n
a
,
a
n
d
I
n
d
i
a
d
u
r
i
n
g
t
h
e
o
n
g
o
i
n
g
p
a
n
d
e
mi
c
3
,
0
0
0
Tw
o
9
,
1
6
1
f
a
c
e
s
w
i
t
h
o
u
t
mask
s
;
3
0
1
5
M
A
F
A
D
i
v
e
r
si
t
y
C
o
mp
l
e
x
P
W
M
F
D
M
o
r
e
t
h
a
n
h
a
l
f
o
f
t
h
e
i
mag
e
s
a
r
e
c
o
l
l
e
c
t
e
d
f
r
o
m
W
I
D
ER
F
a
c
e
,
M
A
F
A
,
a
n
d
R
W
M
F
D
.
T
h
e
“
W
i
t
h
M
a
s
k
”
c
l
a
ss
r
e
q
u
i
r
e
s
c
o
v
e
r
i
n
g
b
o
t
h
t
h
e
f
a
c
e
a
n
d
n
o
se
9
,
2
0
5
T
h
r
e
e
1
0
,
4
7
1
f
a
c
e
s
w
i
t
h
o
u
t
mask
s
;
7
,
6
9
5
c
o
r
r
e
c
t
M
A
F
A
;
3
6
6
i
n
c
o
r
r
e
c
t
M
A
F
A
F
r
o
n
t
a
l
t
o
P
r
o
f
i
l
e
M
e
d
i
u
m
T
h
e
liv
e
n
ess
d
etec
tio
n
d
atase
t
-
Z
alo
A
I
c
h
alle
n
g
e
2
0
2
2
is
u
s
ed
to
tr
ai
n
a
n
d
test
m
o
d
el
s
f
o
r
f
ac
e
liv
e
n
ess
d
etec
tio
n
,
ai
m
i
n
g
to
d
is
tin
g
u
i
s
h
b
et
w
ee
n
r
ea
l
an
d
f
ak
e
f
ac
es.
I
t
i
n
cl
u
d
es
m
u
ltip
le
v
id
eo
s
u
n
d
er
v
ar
io
u
s
li
g
h
tin
g
co
n
d
itio
n
s
a
n
d
co
n
te
x
ts
.
Fo
r
t
h
is
d
ataset,
w
e
e
x
tr
ac
ted
f
r
a
m
e
s
f
r
o
m
m
u
ltip
le
v
id
eo
s
a
n
d
m
an
u
all
y
lab
eled
th
e
m
u
s
in
g
L
ab
elI
m
g
[
1
8
]
,
a
s
o
f
t
w
ar
e
to
o
l
d
esig
n
ed
f
o
r
v
is
u
all
y
a
n
n
o
tat
in
g
a
n
d
id
en
ti
f
y
in
g
o
b
j
ec
ts
w
it
h
i
n
i
m
ag
e
s
.
Fin
a
ll
y
,
all
i
m
a
g
es
a
n
d
lab
els
f
r
o
m
th
e
af
o
r
e
m
e
n
tio
n
ed
d
atasets
w
er
e
co
n
v
er
ted
to
YOL
O
s
ta
n
d
ar
d
lab
elin
g
f
o
r
m
at.
C
o
n
s
eq
u
en
tl
y
,
o
u
r
DR
F
M
D
d
ataset
is
m
o
r
e
d
iv
er
s
e
i
n
t
er
m
s
o
f
th
e
n
u
m
b
er
o
f
i
m
a
g
es,
m
as
k
t
y
p
es,
d
is
tr
ib
u
tio
n
o
f
m
a
s
k
-
w
ea
r
i
n
g
,
n
o
n
-
m
as
k
-
w
ea
r
in
g
,
an
d
i
m
p
r
o
p
er
m
as
k
-
w
ea
r
in
g
,
h
ea
d
p
o
s
itio
n
s
,
s
u
r
r
o
u
n
d
in
g
co
n
te
x
ts
,
a
n
d
i
m
ag
e
d
i
m
e
n
s
io
n
s
.
T
h
e
DR
FMD
d
atase
t
co
n
s
is
ts
o
f
a
to
tal
o
f
1
4
,
7
2
7
i
m
a
g
es,
w
h
ic
h
ar
e
ca
teg
o
r
ized
in
to
t
h
r
ee
s
u
b
s
ets:
i
)
a
tr
ain
i
n
g
s
et
co
n
tai
n
i
n
g
1
0
,
3
0
4
s
a
m
p
l
es,
ii
)
a
v
alid
atio
n
s
et
co
m
p
r
is
i
n
g
1
,
4
7
4
s
a
m
p
le
s
,
an
d
ii
i
)
a
test
s
et
w
i
th
2
,
9
4
9
s
am
p
le
s
.
F
u
r
th
er
d
etail
s
ab
o
u
t
th
e
d
ataset
p
ar
am
eter
s
ca
n
b
e
f
o
u
n
d
in
T
ab
le
2
.
T
ab
le
2
.
Sy
n
th
e
tic
p
ar
a
m
eter
s
o
f
th
e
D
R
FMD
d
at
ase
t
D
a
t
a
se
t
T
r
a
i
n
V
a
l
i
d
T
e
st
T
o
t
a
l
A
I
ZO
O
2
,
6
8
6
3
8
3
7
8
2
3
,
8
5
1
K
F
M
D
[
1
8
]
4
9
5
64
1
2
6
6
8
5
M
A
F
A
1
,
0
0
6
1
4
1
2
8
9
1
,
4
3
6
M
O
X
A
3
k
9
9
9
1
4
4
2
9
6
1
,
4
3
9
P
W
M
F
D
3
,
2
8
3
4
8
2
9
3
8
4
,
7
0
3
Z
a
l
o
A
I
C
h
a
l
e
n
g
e
2
0
2
2
1
,
8
3
5
2
6
0
5
1
8
2
,
6
1
3
D
R
F
M
D
1
0
,
3
0
4
1
,
4
7
4
2
,
9
4
9
1
4
,
7
2
7
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8938
I
n
t J
A
r
ti
f
I
n
tell
,
Vo
l.
14
,
No
.
4
,
A
u
g
u
s
t
2025
:
2
6
3
4
-
2645
2638
A
cc
o
r
d
in
g
to
T
ab
le
3
(
DR
F
MD
d
ataset
co
lu
m
n
)
,
th
e
d
a
ta
lab
els
w
er
e
r
ea
n
n
o
tated
t
o
YOL
O
s
tan
d
ar
d
s
w
it
h
t
h
r
ee
ca
teg
o
r
ie
s
:
w
it
h
o
u
t
m
a
s
k
(
0
)
w
it
h
1
4
,
1
5
7
lab
els,
w
it
h
m
as
k
(
1
)
w
i
th
1
1
,
5
9
0
lab
els,
an
d
w
ea
r
m
a
s
k
in
co
r
r
ec
t
(
2
)
w
it
h
2
,
8
6
6
la
b
els.
C
lear
l
y
,
w
h
e
n
co
m
p
ar
i
n
g
all
p
ar
a
m
eter
s
(
i
m
ag
e
s
,
i
n
s
ta
n
ce
s
w
it
h
o
u
t
m
as
k
,
i
n
s
ta
n
ce
s
w
it
h
m
as
k
,
i
n
s
tan
ce
s
w
ea
r
m
as
k
in
co
r
r
ec
t)
ac
r
o
s
s
d
if
f
er
en
t
d
atasets
(
tr
ai
n
,
v
alid
,
test
)
,
th
e
D
R
FM
D
d
ataset
d
em
o
n
s
tr
ate
s
a
s
i
g
n
i
f
ican
t
a
d
v
an
ta
g
e
i
n
q
u
a
n
tit
y
.
Sp
ec
i
f
ical
l
y
,
it
co
n
tai
n
s
ap
p
r
o
x
im
a
tel
y
1
0
ti
m
es
m
o
r
e
i
m
a
g
es
an
d
ar
o
u
n
d
4
ti
m
es
m
o
r
e
in
s
tan
ce
s
th
a
n
t
h
e
FMM
D
d
ataset.
A
d
d
itio
n
al
l
y
,
i
n
D
R
FMD,
t
h
e
d
is
tr
ib
u
tio
n
o
f
I
n
s
tan
ce
s
with
m
as
k
a
n
d
I
n
s
ta
n
ce
s
w
it
h
o
u
t
m
a
s
k
is
m
o
r
e
b
alan
ce
d
co
m
p
ar
ed
to
th
e
FM
MD
d
ataset.
T
ab
le
3
.
Key
p
ar
a
m
eter
s
o
f
t
h
e
DR
FMD
d
ataset
v
s
FMM
D
d
ataset
P
a
r
a
me
t
e
r
s
D
R
F
M
D
d
a
t
a
se
t
F
M
M
D
d
a
t
a
se
t
T
r
a
i
n
V
a
l
i
d
T
e
st
T
o
t
a
l
T
r
a
i
n
V
a
l
i
d
T
e
st
I
mag
e
s
1
0
,
3
0
4
1
,
4
7
4
2
,
9
4
9
1
4
,
7
2
7
1
,
0
6
7
4
5
6
5
0
7
I
n
st
a
n
c
e
s
2
0
,
6
0
3
3
,
0
5
2
6
,
1
9
1
2
9
,
8
4
6
5
,
7
9
6
2
,
1
5
6
2
,
6
6
3
W
i
t
h
o
u
t
M
a
s
k
(
0
)
9,
6
8
3
1
,
3
9
2
3
,
0
8
2
1
4
,
1
5
7
1
,
0
3
0
3
5
2
4
4
9
W
i
t
h
M
a
s
k
(
1
)
8
,
9
2
6
1
,
3
7
0
2
,
5
2
7
1
1
,
5
9
0
4
,
5
8
9
1
,
7
2
8
2
,
1
2
2
W
e
a
r
M
a
s
k
I
n
c
o
r
r
e
c
t
(
2
)
1
,
9
9
4
2
9
0
5
8
2
2
8
6
6
1
7
7
76
92
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
3
.
1
.
T
ra
ini
ng
re
s
ult
T
o
ev
alu
ate
th
e
e
f
f
ec
tiv
e
n
es
s
o
f
o
b
j
ec
t
d
etec
tio
n
m
o
d
els,
a
v
er
ag
e
p
r
ec
is
io
n
(
A
P
)
is
co
m
m
o
n
l
y
u
s
ed
,
in
co
r
p
o
r
atin
g
k
e
y
m
etr
ics
s
u
c
h
as
in
ter
s
ec
tio
n
o
v
er
u
n
io
n
(
I
o
U)
,
p
r
ec
is
io
n
,
an
d
r
ec
all
.
A
s
m
en
tio
n
ed
in
[
2
6
]
,
th
ese
m
etr
ics
ar
e
m
at
h
e
m
atic
all
y
d
ef
i
n
ed
i
n
(
1
)
to
(
3
)
.
I
o
U
m
ea
s
u
r
es
t
h
e
o
v
er
lap
b
et
w
ee
n
t
h
e
p
r
ed
icte
d
b
o
u
n
d
in
g
b
o
x
(
p
r
ed
)
an
d
t
h
e
ac
tu
al
g
r
o
u
n
d
tr
u
th
b
o
x
(
g
t)
.
P
r
ec
is
io
n
as
s
ess
e
s
t
h
e
ac
c
u
r
ac
y
o
f
t
h
e
m
o
d
el
’
s
o
u
tp
u
ts
,
w
h
er
ea
s
R
ec
all
ev
al
u
ates its
ab
ilit
y
to
d
etec
t a
ll g
t i
n
s
ta
n
ce
s
.
=
ℎ
=
∩
∪
(
1
)
=
+
=
(
2
)
=
+
(
3
)
w
h
er
e
T
P
r
ep
r
esen
ts
tr
u
e
p
o
s
it
i
v
e,
FP
r
ep
r
esen
ts
f
alse p
o
s
iti
v
e,
FN r
ep
r
esen
ts
f
alse
n
eg
a
tiv
e,
an
d
N
r
ep
r
esen
ts
th
e
to
tal
n
u
m
b
er
o
f
r
ec
o
v
er
ed
o
b
j
ec
ts
,
in
clu
d
i
n
g
tr
u
e
p
o
s
iti
v
es a
n
d
f
al
s
e
p
o
s
itiv
e
s
.
I
n
o
b
j
ec
t
d
etec
tio
n
,
m
u
ltip
le
o
b
j
ec
t
class
es
n
ee
d
to
b
e
id
en
ti
f
ied
.
T
h
e
m
A
P
in
d
ex
i
s
u
s
ed
t
o
co
m
p
u
te
th
e
A
P
f
o
r
ea
ch
cla
s
s
a
n
d
t
h
e
n
d
er
iv
e
th
e
o
v
er
all
a
v
er
ag
e.
T
h
is
m
etr
ic
p
r
o
v
id
es
a
co
m
p
r
eh
en
s
i
v
e
ev
a
lu
at
io
n
o
f
th
e
m
o
d
el
’
s
p
er
f
o
r
m
an
ce
ac
r
o
s
s
all
ca
teg
o
r
ies,
co
n
s
i
d
er
in
g
t
h
e
v
ar
y
in
g
d
if
f
ic
u
lt
y
lev
els
i
n
d
etec
ti
n
g
d
if
f
er
e
n
t
o
b
j
ec
ts
.
A
cc
o
r
d
in
g
t
o
D
e
w
i
et
a
l.
[
1
4
]
,
m
A
P
is
d
ef
i
n
ed
in
(
4
)
,
w
h
er
e
th
e
v
ar
ia
b
le
p
(
o
)
r
ep
r
esen
ts
d
etec
tio
n
ac
cu
r
ac
y
.
=
∫
(
)
1
0
(
4
)
Data
au
g
m
e
n
tat
io
n
i
s
a
w
id
el
y
ad
o
p
ted
tech
n
iq
u
e
in
d
ee
p
lear
n
in
g
,
ai
m
ed
at
en
h
a
n
cin
g
t
h
e
v
ar
iab
ilit
y
o
f
a
tr
ain
in
g
d
atas
et
b
y
ap
p
l
y
i
n
g
d
i
f
f
er
e
n
t
tr
an
s
f
o
r
m
atio
n
s
to
th
e
o
r
ig
i
n
al
d
ata.
T
h
r
o
u
g
h
o
u
t
th
e
tr
ain
i
n
g
p
r
o
ce
s
s
,
v
ar
io
u
s
au
g
m
en
tatio
n
m
et
h
o
d
s
,
in
cl
u
d
in
g
p
ad
d
in
g
,
cr
o
p
p
in
g
,
an
d
h
o
r
izo
n
tal
f
lip
p
in
g
,
a
m
o
n
g
o
th
er
s
,
ar
e
u
tili
ze
d
.
T
h
ese
tech
n
iq
u
es
p
la
y
a
cr
u
c
ial
r
o
le
in
th
e
d
ev
elo
p
m
e
n
t
o
f
lar
g
e
-
s
ca
le
n
eu
r
al
n
et
w
o
r
k
s
d
u
e
to
th
eir
ef
f
ec
ti
v
e
n
es
s
in
i
m
p
r
o
v
in
g
m
o
d
el
g
e
n
er
aliza
ti
o
n
.
I
n
o
u
r
ex
p
er
i
m
e
n
t,
w
e
tr
a
in
ed
th
e
m
o
d
el
f
o
r
1
0
0
e
p
o
ch
s
w
it
h
a
w
ei
g
h
t
d
ec
a
y
o
f
0
.
0
0
0
5
,
an
in
itial
lear
n
in
g
r
ate
o
f
0
.
0
1
,
a
f
in
al
lear
n
in
g
r
ate
o
f
0
.
0
1
,
a
b
atch
s
ize
o
f
1
6
,
an
in
p
u
t
i
m
a
g
e
s
ize
o
f
6
4
0
,
an
d
a
n
I
o
U
th
r
esh
o
ld
o
f
0
.
7
.
F
u
r
th
er
m
o
r
e,
we
ap
p
lied
a
Mo
s
aic
co
n
f
i
g
u
r
atio
n
o
f
1
.
0
f
o
r
th
e
f
ir
s
t
9
0
ep
o
ch
s
,
s
et
clo
s
e_
m
o
s
ai
c
to
1
0
,
an
d
u
s
ed
m
i
x
u
p
at
0
.
2
4
3
.
A
d
d
itio
n
al
d
ata
au
g
m
e
n
tatio
n
p
ar
am
e
ter
s
in
cl
u
d
ed
h
s
v
_
h
at
0
.
0
1
3
8
,
h
s
v
_
s
at
0
.
6
6
4
,
h
s
v
_
v
at
0
.
4
6
4
,
tr
an
s
late
at
0
.
1
,
s
ca
le
at
0
.
8
9
8
,
an
d
s
h
ea
r
at
0
.
6
0
2
.
A
ll
t
h
e
a
b
o
v
e
co
n
f
i
g
u
r
atio
n
s
w
il
l
b
e
u
s
ed
to
tr
ain
all
f
i
v
e
YOL
O
v
8
m
o
d
els
o
n
b
o
th
th
e
FMM
D
an
d
DR
FMD
d
ata
s
ets.
W
e
th
e
n
e
v
alu
a
te
th
e
r
e
s
u
lt
s
f
o
r
all
tr
ain
ed
m
o
d
els
o
n
t
h
e
te
s
ts
et
o
f
th
e
t
h
r
ee
d
atasets
FMM
D,
D
R
FMD,
an
d
HI
T
L
-
MM
D.
T
h
e
tr
ain
i
n
g
e
n
v
ir
o
n
m
e
n
t
f
o
r
t
h
e
m
o
d
el
s
w
as
ca
r
r
ied
o
u
t
o
n
a
Dell
R
7
3
0
s
er
v
er
,
w
h
ich
in
c
lu
d
es
t
w
o
N
v
id
ia
T
esla
P
4
0
GP
U
ac
ce
ler
ato
r
s
w
it
h
2
4
GB
o
f
R
A
M
ea
c
h
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
A
r
ti
f
I
n
tell
I
SS
N:
2252
-
8938
E
n
h
a
n
ci
n
g
fa
ce
m
a
s
k
d
etec
tio
n
p
erfo
r
ma
n
ce
w
ith
co
mp
r
eh
e
n
s
ive
d
a
ta
s
et
a
n
d
YOLOv8
(
T
r
o
n
g
Th
u
a
Hu
yn
h
)
2639
t
w
o
Xeo
n
E
2
6
8
0
v
4
ce
n
tr
al
p
r
o
ce
s
s
o
r
s
,
an
d
1
2
8
GB
o
f
DDR4
2
4
0
0
b
u
s
m
e
m
o
r
y
.
T
h
e
YOL
O
v
8
tr
ain
i
n
g
p
r
o
ce
s
s
w
as p
er
f
o
r
m
ed
o
n
t
w
o
GP
Us to
ac
h
iev
e
r
ea
l
-
ti
m
e
d
etec
tio
n
ca
p
ab
ilit
ies.
T
h
e
tr
ain
in
g
r
esu
lts
o
f
t
h
e
Y
OL
O
v
8
m
o
d
els
(
n
,
s
,
m
,
l,
x
)
o
n
th
e
FMM
D
d
ataset
(
4
5
6
im
a
g
es
,
2
1
5
6
in
s
tan
ce
s
)
a
n
d
th
e
DR
F
MD
d
ataset
(
1
,
4
7
4
im
a
g
es,
3
,
0
5
2
in
s
tan
ce
s
)
ar
e
s
h
o
w
n
i
n
T
ab
le
4
.
I
n
th
i
s
tab
le,
w
e
o
n
l
y
p
r
esen
t
t
h
e
o
v
er
all
tr
ain
i
n
g
r
es
u
lt
s
f
o
r
all
i
m
a
g
es.
Deta
iled
r
esu
lt
s
f
o
r
th
e
ca
s
e
s
w
it
h
o
u
t
m
a
s
k
(
0
)
,
w
it
h
m
a
s
k
(
1
)
,
an
d
w
ea
r
m
as
k
i
n
co
r
r
ec
t
(
2
)
ar
e
p
u
b
lis
h
ed
i
n
[
2
7
]
.
T
h
e
tr
ain
i
n
g
p
er
f
o
r
m
a
n
ce
i
s
e
v
al
u
ated
o
n
th
e
m
etr
ics
p
r
ec
is
io
n
(
P
)
,
r
ec
all
(
R
)
,
an
d
m
AP
@
5
0
(
r
ep
r
e
s
en
t
s
t
h
e
m
A
P
v
a
lu
e
w
h
e
n
t
h
e
I
o
U
is
0
.
5
)
.
T
h
e
s
p
ec
if
ic
r
esu
lts
ar
e:
i)
YO
L
Ov
8
n
(
n
a
n
o
)
ac
h
ie
v
ed
P
=0
.
8
3
8
,
R
=0
.
7
9
6
,
an
d
m
A
P
@
5
0
=0
.
8
4
9
.
Desp
ite
b
ein
g
th
e
s
m
al
lest
v
er
s
io
n
,
t
h
i
s
m
o
d
el
s
h
o
w
s
r
elat
iv
el
y
h
i
g
h
ac
c
u
r
ac
y
an
d
r
ec
o
g
n
i
tio
n
ca
p
ab
ilit
y
;
ii)
YO
L
O
v
8
s
(
s
m
all
)
s
ig
n
i
f
ica
n
tl
y
i
m
p
r
o
v
ed
w
i
th
P
=0
.
8
4
5
,
R
=0
.
8
2
0
,
an
d
m
AP
@
5
0
=0
.
8
7
2
,
in
d
icatin
g
en
h
an
ce
d
d
etec
tio
n
o
f
h
ar
d
er
o
b
j
ec
ts
;
iii)
YO
L
O
v
8
m
(
m
ed
i
u
m
)
f
u
r
th
er
e
n
h
an
ce
d
p
er
f
o
r
m
a
n
ce
w
i
th
P
=0
.
8
6
2
,
R
=0
.
8
2
3
,
an
d
m
A
P
@
5
0
=0
.
8
8
8
,
d
em
o
n
s
tr
ati
n
g
s
tr
o
n
g
er
o
b
j
ec
t
r
ec
o
g
n
itio
n
ca
p
ab
ilit
ies
;
i
v
)
YO
L
O
v
8
l
(
lar
g
e)
m
ai
n
tai
n
ed
h
ig
h
ac
cu
r
ac
y
w
i
th
P
=0
.
8
5
9
,
R
=0
.
8
4
4
,
an
d
m
A
P
@
5
0
=0
.
8
8
9
,
in
d
icatin
g
s
tab
le
p
er
f
o
r
m
an
ce
in
o
b
j
ec
t
d
etec
tio
n
an
d
c
lass
if
ica
tio
n
;
v
)
YO
L
O
v
8
x
(
ex
tr
a
-
lar
g
e)
a
ch
iev
ed
th
e
h
i
g
h
est
p
er
f
o
r
m
a
n
ce
w
it
h
P
=0
.
8
5
9
,
R
=0
.
8
3
8
,
an
d
m
A
P
@
5
0
=0
.
8
9
5
,
s
h
o
w
ca
s
i
n
g
s
u
p
er
io
r
o
b
j
ec
t
r
ec
o
g
n
itio
n
an
d
clas
s
if
icatio
n
ca
p
ab
ilit
ies.
T
ab
le
4
.
T
r
ain
in
g
p
er
f
o
r
m
a
n
ce
f
o
r
5
m
o
d
els YO
L
O
v
8
o
n
FM
MD
an
d
DR
F
MD
d
atasets
M
o
d
e
l
F
M
M
D
d
a
t
a
se
t
(
4
5
6
i
mag
e
s,
2
,
1
5
6
i
n
st
a
n
c
e
s)
D
R
F
M
D
d
a
t
a
se
t
(
1
,
4
7
4
i
m
a
g
e
s,
3
,
0
5
2
i
n
st
a
n
c
e
s)
P
R
mA
P
@
5
0
P
R
mA
P
@
5
0
Y
O
L
O
v
8
n
0
.
9
0
4
0
.
7
7
7
0
.
8
6
8
0
.
8
3
8
0
.
7
9
6
0
.
8
4
9
Y
O
L
O
v
8
s
0
.
9
4
5
0
.
8
6
3
0
.
9
2
4
0
.
8
4
5
0
.
8
2
0
0
.
8
7
2
Y
O
L
O
v
8
m
0
.
9
6
6
0
.
8
7
9
0
.
9
4
8
0
.
8
6
2
0
.
8
2
3
0
.
8
8
8
Y
O
L
O
v
8
l
0
.
9
7
9
0
.
8
9
8
0
.
9
5
6
0
.
8
5
9
0
.
8
4
4
0
.
8
8
9
Y
O
L
O
v
8
x
0
.
9
5
5
0
.
9
2
4
0
.
9
6
3
0
.
8
5
9
0
.
8
3
8
0
.
8
9
5
Fig
u
r
e
2
i
llu
s
t
r
at
es
th
e
tr
ai
n
in
g
ch
ar
ts
o
f
YOL
Ov
8
n
,
YOL
Ov
8
s
,
YOL
Ov
8
m
,
YO
L
Ov
8
l,
an
d
YOL
Ov
8
x
o
v
er
1
0
0
e
p
o
ch
s
o
n
th
e
DR
F
MD
d
atas
et.
T
h
e
l
o
s
s
v
alu
es f
o
r
YOL
Ov
8
n
as
s
h
o
w
n
in
Fig
u
r
e
2
(
a)
a
r
e
b
o
x
_
l
o
s
s
=
1
.
1
8
8
,
cls
_
l
o
s
s
=
0
.
6
8
3
1
7
,
an
d
d
f
l_
l
o
s
s
=1
.
2
2
1
6
.
Sim
ilar
ly
,
YOL
Ov
8
s
as
s
h
o
w
n
in
Fig
u
r
e
2
(
b
)
h
as
b
o
x
_
l
o
s
s
=
1
.
1
0
2
4
,
cls
_
l
o
s
s
=
0
.
5
7
8
3
,
an
d
d
f
l_
lo
s
s
=1
.
1
6
3
9
,
w
h
ile
YOL
Ov
8
m
as
s
h
o
w
n
in
Fig
u
r
e
2
(
c
)
r
ec
o
r
d
s
b
o
x
_
l
o
s
s
=
1
.
0
1
6
7
,
cls
_
l
o
s
s
=
0
.
5
0
4
,
an
d
d
f
l_
l
o
s
s
=1
.
1
5
3
1
.
F
o
r
YOL
Ov
8
l
as
s
h
o
w
n
in
Fig
u
r
e
2
(
d
)
,
th
e
v
alu
es
a
r
e
b
o
x
_
l
o
s
s
=
1
.
0
0
3
6
,
cls
_
l
o
s
s
=
0
.
4
8
4
5
9
,
an
d
d
f
l_
lo
s
s
=1
.
1
9
3
8
,
w
h
er
ea
s
YOL
Ov
8
x
as
s
h
o
w
n
in
Fig
u
r
e
2
(
e)
ex
h
ib
it
s
b
o
x
_
lo
s
s
=0
.
9
7
0
3
2
,
cls_
lo
s
s
=0
.
4
5
9
3
6
,
an
d
d
f
l_
lo
s
s
=1
.
1
9
9
2
.
T
h
ese
r
esu
lts
s
u
g
g
es
t
th
at
lar
g
er
m
o
d
els
(
YOL
O
v
8
m
,
YO
L
O
v
8
l,
YOL
Ov
8
x
)
o
u
tp
er
f
o
r
m
s
m
al
ler
o
n
es
(
YOL
O
v
8
n
,
YO
L
O
v
8
s
)
ac
r
o
s
s
all
t
y
p
es
o
f
lo
s
s
f
u
n
ctio
n
s
.
A
s
m
o
d
els
b
ec
o
m
e
m
o
r
e
co
m
p
le
x
,
b
o
x
_
lo
s
s
an
d
cls_
lo
s
s
d
ec
r
ea
s
e
s
ig
n
i
f
ican
tl
y
,
h
ig
h
li
g
h
ti
n
g
th
eir
i
m
p
r
o
v
ed
o
b
j
ec
t
lo
ca
lizatio
n
an
d
class
i
f
icatio
n
ca
p
ab
ilit
i
es.
Ho
w
e
v
er
,
d
f
l_
lo
s
s
r
e
m
ai
n
s
r
elati
v
e
l
y
s
tab
le
ac
r
o
s
s
d
if
f
er
en
t
m
o
d
els,
w
i
th
o
n
l
y
m
i
n
o
r
f
lu
ct
u
atio
n
s
,
i
n
d
ic
atin
g
t
h
at
t
h
eir
ab
ilit
y
to
lear
n
d
is
tr
ib
u
tio
n
w
eig
h
t
s
d
o
es n
o
t v
ar
y
s
i
g
n
i
f
ican
tl
y
.
3
.
2
.
E
v
a
lua
t
i
o
n r
esu
lt
s
a
nd
dis
cu
s
s
io
n
T
h
is
ev
al
u
atio
n
s
ec
tio
n
is
c
o
n
d
u
cted
b
ased
o
n
t
h
e
tr
ain
i
n
g
r
e
s
u
l
ts
u
s
i
n
g
b
o
th
t
h
e
F
MM
D
an
d
DR
FMD
d
ataset
s
as
p
r
esen
te
d
in
s
ec
tio
n
3
.
1
.
Fo
r
co
n
v
e
n
i
en
ce
i
n
d
is
c
u
s
s
i
n
g
t
h
e
r
es
u
lt
s
,
w
e
d
is
ti
n
g
u
i
s
h
as
f
o
llo
w
s
:
i)
t
h
e
tr
ai
n
i
n
g
r
es
u
l
ts
w
ith
t
h
e
FMM
D
d
ataset
f
o
r
all
f
iv
e
YO
L
O
v
8
m
o
d
els
ar
e
r
ef
er
r
ed
to
as
YOL
O
v
8
m
o
d
el
s
w
it
h
F
M
M
D
d
ataset
(
Y
w
FMM
D)
;
an
d
i
i)
t
h
e
tr
ai
n
in
g
r
es
u
lt
s
w
it
h
t
h
e
p
r
o
p
o
s
ed
DR
FMD
d
ataset
f
o
r
all
f
i
v
e
YO
L
Ov
8
m
o
d
el
s
ar
e
r
ef
er
r
ed
to
as
YO
L
O
v
8
m
o
d
els
w
it
h
DR
FMD
d
ataset
(
Y
w
DR
FMD)
.
I
n
th
e
n
e
x
t
s
ec
tio
n
s
,
to
av
o
id
co
n
f
u
s
io
n
w
ith
to
o
m
an
y
n
u
m
b
er
s
,
w
e
f
ilter
ed
o
u
t
th
e
d
etailed
d
ata
in
clu
d
in
g
w
it
h
o
u
t
m
a
s
k
(
0
)
,
w
i
th
m
as
k
(
1
)
,
an
d
w
ea
r
m
as
k
i
n
co
r
r
ec
tl
y
(
2
)
.
T
h
e
d
etails ar
e
p
r
o
v
id
ed
i
n
[
2
7
]
.
3
.
2
.
1
.
E
v
a
lua
t
e
t
he
Y
w
F
M
M
D
m
o
de
ls
v
ia
v
a
rio
us
t
estset
s
I
n
t
h
is
s
ec
tio
n
,
w
e
u
s
e
th
r
ee
t
ests
et
s
:
FM
MD
,
HI
T
L
-
MM
D,
an
d
D
R
FMD,
a
s
d
escr
ib
ed
i
n
s
ec
tio
n
2
,
to
ev
al
u
ate
Y
w
F
MM
D.
T
ab
le
5
p
r
esen
ts
th
e
e
v
al
u
atio
n
r
es
u
lts
o
f
t
h
e
Y
O
L
Ov
8
m
o
d
els
(
n
,
s
,
m
,
l,
x
)
o
n
th
e
FMM
D
d
ataset
w
it
h
i)
t
h
e
te
s
ts
et
e
x
tr
ac
ted
f
r
o
m
t
h
e
FM
MD
d
ataset
its
el
f
(
5
0
7
i
m
ag
e
s
,
2
,
6
6
3
in
s
t
an
ce
s
)
,
ii)
th
e
test
s
et
ex
tr
ac
ted
f
r
o
m
t
h
e
DR
FMD
d
atase
t (
2
,
9
4
9
im
a
g
es,
6
,
1
9
1
in
s
tan
ce
s
)
,
an
d
iii)
t
h
e
test
s
et
ex
tr
ac
ted
f
r
o
m
t
h
e
HI
T
L
-
MM
D
d
ataset
(
1
,
3
1
1
im
a
g
es,
2
,
9
6
4
in
s
ta
n
c
es).
R
es
u
lt
s
i
n
co
l
u
m
n
(
I
)
s
h
o
w
t
h
at
th
e
m
o
d
el
s
ac
h
iev
e
v
er
y
h
i
g
h
p
er
f
o
r
m
a
n
c
e
o
n
th
e
d
ataset
th
e
y
w
er
e
tr
ai
n
ed
o
n
(
th
e
FMM
D
d
ataset
it
s
elf
)
,
w
ith
p
r
ec
is
io
n
,
r
ec
all,
an
d
m
A
P
m
e
tr
ics
all
a
b
o
v
e
av
er
ag
e.
T
h
is
i
n
d
icate
s
th
at
t
h
e
m
o
d
el
s
ar
e
w
ell
-
tr
ain
ed
an
d
ca
p
ab
le
o
f
g
o
o
d
r
ec
o
g
n
itio
n
o
n
th
e
s
ee
n
d
ataset.
R
esu
l
ts
in
co
l
u
m
n
s
(
I
I
)
an
d
(
I
I
I
)
s
h
o
w
t
h
at
th
e
p
er
f
o
r
m
an
ce
o
f
t
h
e
YOL
O
v
8
m
o
d
els
w
i
th
t
h
e
F
MM
D
d
ataset
d
ec
r
ea
s
e
s
,
i
n
d
icatin
g
t
h
at
th
e
g
e
n
er
aliza
tio
n
ca
p
ab
ilit
y
o
f
t
h
e
FMM
D
d
ataset
is
n
o
t d
iv
er
s
e
en
o
u
g
h
,
lead
i
n
g
to
u
n
d
er
f
i
tti
n
g
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8938
I
n
t J
A
r
ti
f
I
n
tell
,
Vo
l.
14
,
No
.
4
,
A
u
g
u
s
t
2025
:
2
6
3
4
-
2645
2640
(
a)
(
b
)
(
c)
(
d
)
(
e)
Fig
u
r
e
2
.
T
r
ain
in
g
p
er
f
o
r
m
a
n
c
e
u
s
i
n
g
(
a)
YO
L
O
v
8
n
,
(
b
)
YOL
O
v
8
s
,
(
c)
YOL
O
v
8
m
,
(
d
)
YOL
O
v
8
l,
an
d
(
e)
YOL
Ov
8
x
w
it
h
DR
FMD
d
ataset
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
A
r
ti
f
I
n
tell
I
SS
N:
2252
-
8938
E
n
h
a
n
ci
n
g
fa
ce
m
a
s
k
d
etec
tio
n
p
erfo
r
ma
n
ce
w
ith
co
mp
r
eh
e
n
s
ive
d
a
ta
s
et
a
n
d
YOLOv8
(
T
r
o
n
g
Th
u
a
Hu
yn
h
)
2641
T
ab
le
5
.
T
esti
n
g
r
es
u
lt o
f
Y
wFMM
D
m
o
d
els o
n
FMM
D,
D
R
FMD,
a
n
d
HI
T
L
-
MM
D
t
es
ts
ets
M
o
d
e
l
F
M
M
D
T
e
st
se
t
(
I
)
D
R
F
M
D
T
e
st
se
t
(
I
I
)
H
I
TL
-
M
M
D
T
e
st
se
t
(
I
I
I
)
P
R
mA
P
@
5
0
P
R
mA
P
@
5
0
P
R
mA
P
@
5
0
Y
O
L
O
v
8
n
0
.
9
0
4
0
.
8
1
8
0
.
8
7
6
0
.
7
0
2
0
.
5
7
7
0
.
6
0
4
0
.
7
7
2
0
.
5
9
6
0
.
6
6
1
Y
O
L
O
v
8
s
0
.
9
2
7
0
.
8
7
3
0
.
9
2
0
0
.
7
1
2
0
.
5
9
5
0
.
6
1
5
0
.
7
4
7
0
.
6
2
6
0
.
6
8
4
Y
O
L
O
v
8
m
0
.
9
6
8
0
.
8
8
3
0
.
9
3
7
0
.
7
2
0
0
.
6
3
2
0
.
6
5
5
0
.
7
1
5
0
.
6
5
3
0
.
6
8
2
Y
O
L
O
v
8
l
0
.
9
7
3
0
.
9
2
2
0
.
9
6
7
0
.
7
1
9
0
.
6
3
4
0
.
6
5
8
0.
7
1
0
0
.
6
3
5
0
.
6
8
3
Y
O
L
O
v
8
x
0
.
9
6
6
0
.
9
1
7
0
.
9
4
8
0
.
6
9
1
0
.
6
3
3
0
.
6
3
9
0
.
7
2
9
0
.
6
5
0
0
.
6
7
5
B
ased
o
n
T
a
b
le
5
,
w
e
s
ee
t
h
at
w
h
e
n
tr
ain
i
n
g
an
d
test
in
g
t
h
e
Y
w
FMM
D
m
o
d
el
o
n
th
e
FM
MD
d
ataset
its
el
f
,
t
h
e
r
es
u
lts
ar
e
g
o
o
d
in
m
o
s
t
v
er
s
io
n
s
o
f
t
h
e
Y
w
F
M
MD
m
o
d
el.
Ho
w
e
v
er
,
w
h
e
n
e
v
alu
a
tin
g
o
n
t
h
e
n
e
w
d
atasets
D
R
FMD
an
d
HI
T
L
-
MM
D,
th
e
P
r
ec
is
io
n
,
R
ec
all,
an
d
m
A
P
@
5
0
m
etr
ic
s
all
d
ec
lin
e
s
ig
n
i
f
ica
n
tl
y
.
Sp
ec
if
icall
y
,
f
o
r
YO
L
Ov
8
m
,
th
e
p
r
ec
is
io
n
w
h
e
n
tr
ain
i
n
g
an
d
test
i
n
g
w
i
th
t
h
e
FM
MD
,
DR
FMD,
an
d
HI
T
L
-
MM
D
d
atasets
ar
e
as
f
o
llo
w
s
:
0
.
9
6
6
,
0
.
9
6
8
(
↑
0
.
0
2
)
,
0
.
7
2
0
(
↓
0
.
2
4
6
)
,
an
d
0
.
7
1
5
(
↓
0
.
2
5
1
)
.
Fo
r
r
ec
all,
th
e
r
esu
lt
s
ar
e
0
.
8
7
9
,
0
.
8
8
3
(
↑
0
.
0
4
)
,
0
.
6
3
2
(
↓
0
.
2
4
7
)
,
an
d
0
.
6
5
3
(
↓
0
.
2
2
6
)
.
Fo
r
m
A
P
@
5
0
,
th
e
r
esu
lt
s
ar
e
0
.
9
4
8
,
0
.
9
3
7
(
↓
0
.
0
4
3
)
,
0
.
6
5
5
(
↓
0
.
2
9
3
)
,
an
d
0
.
6
8
2
(
↓
0
.
2
6
6
)
.
T
h
is
in
d
icate
s
th
at
tr
ain
i
n
g
o
n
t
h
e
F
MM
D
d
ataset
r
es
u
lts
in
an
o
v
er
f
itted
m
o
d
el.
I
n
ter
m
s
o
f
d
atase
t
s
ize,
DR
F
MD
(
1
4
,
7
2
7
im
a
g
es
a
n
d
2
9
,
8
4
6
in
s
ta
n
ce
s
)
i
s
m
u
ch
lar
g
er
t
h
an
FMM
D
(
2
,
0
3
0
i
m
ag
e
s
a
n
d
1
0
,
6
1
5
in
s
tan
ce
s
)
,
as
s
h
o
w
n
i
n
T
ab
le
3
.
I
n
ter
m
s
o
f
i
n
s
ta
n
ce
d
is
tr
i
b
u
tio
n
,
D
R
FM
D
i
s
r
elativ
el
y
m
o
r
e
b
alan
ce
d
co
m
p
ar
ed
to
FMM
D.
No
tab
l
y
,
th
e
p
r
o
p
o
r
tio
n
o
f
i
m
p
r
o
p
er
l
y
w
o
r
n
m
as
k
s
i
n
DR
FMD
i
s
9
.
6
%
(
2
,
8
6
6
in
s
tan
ce
s
)
co
m
p
ar
ed
to
3
.
2
%
(
3
4
5
in
s
ta
n
ce
s
)
i
n
FMM
D.
Du
e
to
th
is
,
w
h
e
n
test
i
n
g
w
it
h
th
e
test
s
ets
o
f
DR
FMD
an
d
HI
T
L
-
MM
D,
t
h
e
r
es
u
lt
s
s
i
g
n
i
f
ican
tl
y
d
ec
lin
e,
d
e
m
o
n
s
tr
ati
n
g
t
h
at
t
h
e
FMM
D
d
ataset
is
q
u
ite
li
m
ite
d
in
th
e
n
u
m
b
er
o
f
in
s
ta
n
ce
s
,
e
s
p
ec
iall
y
f
o
r
i
m
p
r
o
p
er
ly
w
o
r
n
m
as
k
s
.
3
.
2
.
2
.
E
v
a
lua
t
e
t
he
Y
w
DRF
M
D
m
o
del
s
v
ia
v
a
rio
us
t
estset
s
I
n
t
h
is
s
ec
tio
n
,
w
e
u
s
e
th
r
ee
te
s
ts
et
s
: F
MM
D,
HI
T
L
-
MM
D,
an
d
DR
FMD
a
s
d
escr
ib
e
d
i
n
s
ec
tio
n
2
to
ev
alu
a
te
Y
w
D
R
FMD
to
s
ee
th
e
co
n
tr
ib
u
tio
n
o
f
t
h
e
p
r
o
p
o
s
ed
DR
FMD
d
ataset.
T
a
b
le
6
p
r
esen
ts
th
e
ev
alu
a
tio
n
r
es
u
lt
s
o
f
YO
L
O
v
8
m
o
d
els
(
n
,
s
,
m
,
l,
x
)
o
n
t
h
e
DR
FMD
d
ataset
w
i
th
i)
th
e
te
s
ts
et
e
x
tr
ac
ted
f
r
o
m
th
e
D
R
F
MD
d
atase
t
it
s
el
f
(
2
,
9
4
9
im
a
g
e
s
,
6
,
1
9
1
in
s
ta
n
ce
s
)
,
ii)
t
h
e
test
s
et
e
x
tr
ac
ted
f
r
o
m
th
e
FMM
D
d
ataset
(
5
0
7
im
a
g
es,
2
,
6
6
3
in
s
tan
ce
s
)
,
an
d
iii)
th
e
test
s
et
ex
tr
ac
te
d
f
r
o
m
t
h
e
HI
T
L
-
MM
D
d
ata
s
et
(
1
,
3
1
1
im
a
g
es,
2
,
9
6
4
in
s
tan
ce
s
)
.
R
es
u
lt
s
i
n
co
lu
m
n
(
I
)
s
h
o
w
t
h
at
t
h
e
Y
OL
O
v
8
m
o
d
el
s
tr
ai
n
ed
o
n
t
h
e
D
R
FMD
d
ataset
p
er
f
o
r
m
w
ell
o
n
t
h
e
test
s
et
o
f
th
is
d
ata
s
et
i
ts
el
f
(
D
R
FM
D)
.
E
as
y
to
s
ee
t
h
at,
t
h
e
p
r
ec
is
io
n
an
d
r
ec
all
m
etr
ics
ar
e
h
ig
h
f
o
r
all
v
er
s
io
n
s
o
f
Y
OL
O
v
8
.
E
s
p
ec
iall
y
,
t
h
e
lar
g
er
m
o
d
els
lik
e
Y
OL
Ov
8
l
a
n
d
YOL
O
v
8
x
h
av
e
t
h
e
h
ig
h
e
s
t
m
A
P
@
5
0
(
0
.
8
6
7
an
d
0
.
5
6
0
r
esp
ec
tiv
el
y
)
.
T
h
is
d
em
o
n
s
tr
ates
th
a
t
th
e
DR
FMD
d
ataset
p
r
o
v
id
es
a
s
o
lid
f
o
u
n
d
atio
n
f
o
r
tr
ain
i
n
g
r
ec
o
g
n
itio
n
m
o
d
els
w
it
h
h
i
g
h
ac
cu
r
ac
y
an
d
r
ec
o
g
n
itio
n
ca
p
ab
ilit
y
.
R
es
u
lts
i
n
co
lu
m
n
s
(
I
I
)
an
d
(
I
I
I
)
s
h
o
w
th
at
th
e
Y
w
D
R
FMD
m
o
d
e
l,
w
h
e
n
te
s
ted
o
n
o
t
h
er
d
at
asets
(
F
MM
D
a
n
d
HI
T
L
-
MM
D)
,
h
as r
ed
u
ce
d
p
er
f
o
r
m
a
n
ce
co
m
p
ar
ed
to
w
h
e
n
t
ested
o
n
th
e
D
R
FMD
d
atase
t.
Ho
w
e
v
er
,
th
e
p
r
ec
is
io
n
an
d
r
ec
all
m
etr
ics
r
e
m
ai
n
r
elati
v
el
y
h
i
g
h
,
r
an
g
i
n
g
f
r
o
m
0
.
7
7
1
t
o
0
.
8
3
0
an
d
0
.
7
7
8
to
0
.
7
9
5
f
o
r
p
r
ec
is
io
n
,
an
d
f
r
o
m
0
.
6
8
6
to
0
.
7
5
8
an
d
0
.
6
4
6
t
o
0
.
7
1
6
f
o
r
r
ec
all
o
n
t
h
e
FMM
D
an
d
HI
T
L
-
MM
D
tes
ts
et
s
,
r
esp
ec
ti
v
el
y
.
No
tab
l
y
,
m
A
P
@
5
0
m
ai
n
tai
n
s
ac
ce
p
tab
le
v
al
u
e
s
,
w
ith
0
.
7
3
7
to
0
.
8
0
0
f
o
r
FMM
D
a
n
d
0
.
7
0
5
to
0
.
7
7
9
f
o
r
HI
T
L
-
MM
D.
T
h
ese
r
es
u
l
ts
s
u
g
g
e
s
t
t
h
at
th
e
m
o
d
el
tr
ain
ed
o
n
DR
FMD
d
em
o
n
s
tr
ate
s
s
tr
o
n
g
g
en
er
a
liz
atio
n
ca
p
ab
ilit
y
,
allo
w
i
n
g
it
to
p
er
f
o
r
m
e
f
f
ec
ti
v
el
y
o
n
o
th
er
d
atasets
d
esp
ite
n
o
t
ac
h
iev
in
g
th
e
h
ig
h
est
p
er
f
o
r
m
an
ce
.
Ov
er
all,
t
h
e
YO
L
Ov
8
m
o
d
el
tr
ain
ed
o
n
th
e
DR
FMD
d
ata
s
et
(
Y
w
D
R
FM
D)
d
e
m
o
n
s
tr
ates
g
o
o
d
r
ec
o
g
n
itio
n
a
n
d
class
if
ica
tio
n
ca
p
ab
ilit
ies
o
n
t
h
is
d
ata
s
et
its
el
f
,
w
h
ile
al
s
o
s
h
o
w
i
n
g
t
h
e
ab
ilit
y
to
g
e
n
er
alize
an
d
ap
p
l
y
to
o
th
er
d
at
asets
w
ith
g
o
o
d
p
er
f
o
r
m
a
n
ce
.
T
h
is
af
f
ir
m
s
th
e
d
iv
er
s
it
y
a
n
d
r
o
b
u
s
tn
e
s
s
o
f
th
e
DR
FMD
d
ataset,
m
a
k
i
n
g
a
s
i
g
n
if
ica
n
t
co
n
tr
ib
u
tio
n
to
im
p
r
o
v
in
g
t
h
e
p
er
f
o
r
m
a
n
ce
o
f
o
b
j
ec
t
r
ec
o
g
n
i
tio
n
m
o
d
els.
B
ased
o
n
T
ab
le
6
,
w
e
o
b
s
er
v
e
t
h
at
w
h
e
n
tr
ai
n
in
g
a
n
d
test
i
n
g
t
h
e
Y
w
DR
F
MD
m
o
d
el
o
n
t
h
e
DR
FMD
d
atase
t
its
el
f
,
it
y
i
eld
s
g
o
o
d
r
esu
lts
ac
r
o
s
s
m
o
s
t
v
er
s
io
n
s
o
f
t
h
e
Y
w
DR
F
MD
m
o
d
el.
W
e
also
ev
alu
a
ted
th
e
m
o
d
el
o
n
o
th
er
d
atasets
s
u
ch
as
FMM
D
an
d
HI
T
L
-
MM
D,
w
h
er
e
th
e
p
ar
a
m
eter
s
P
r
ec
is
io
n
,
R
ec
all,
an
d
m
A
P
@
5
0
s
h
o
w
ed
a
s
lig
h
t
d
ec
r
ea
s
e.
Sp
ec
i
f
icall
y
,
f
o
r
YOL
Ov
8
m
,
th
e
p
r
ec
is
io
n
w
h
en
tr
ain
ed
a
n
d
t
ested
w
it
h
t
h
e
D
R
FMD,
FM
MD
,
an
d
HI
T
L
-
MM
D
d
atase
ts
ar
e
0
.
8
6
2
,
0
.
8
5
8
(
↓
0
.
0
0
4
)
,
0
.
8
0
7
(
↓
0
.
0
5
5
)
,
an
d
0
.
7
9
5
(
↓
0
.
0
6
7
)
,
r
esp
ec
tiv
el
y
;
f
o
r
r
ec
all,
th
e
r
es
u
lts
ar
e
0
.
8
2
3
,
0
.
8
0
0
(
↓
0
.
0
2
3
)
,
0
.
7
5
0
(
↓
0
.
0
7
3
)
,
an
d
0
.
7
1
6
(
↓
0
.
1
0
7
)
,
r
esp
ec
tiv
el
y
;
f
o
r
m
A
P
@
5
0
,
th
e
r
esu
lts
ar
e
0
.
8
8
8
,
0
.
8
5
6
(
↓
0
.
0
3
2
)
,
0
.
7
8
3
(
↓
0
.
1
0
5
)
,
an
d
0
.
7
7
9
(
↓
0
.
1
0
9
)
,
r
esp
ec
tiv
el
y
.
T
h
is
in
d
icate
s
th
at
tr
ain
i
n
g
o
n
th
e
DR
FMD
d
ataset
r
esu
lt
s
in
a
m
o
d
el
w
i
th
h
i
g
h
g
en
er
al
izatio
n
ca
p
ab
ilit
ies.
I
n
c
l
u
d
i
n
g
t
h
e
H
I
T
L
-
M
M
D
d
a
t
a
s
e
t
t
o
e
v
a
l
u
a
t
e
t
h
e
a
c
c
u
r
a
c
y
o
f
th
e
m
o
d
e
l
s
(
Y
w
F
M
M
D
a
n
d
Yw
D
R
F
M
D
)
is
ai
m
ed
at
te
s
ti
n
g
th
e
g
e
n
er
aliza
tio
n
a
n
d
o
b
j
ec
tiv
it
y
o
f
t
h
ese
m
o
d
els.
Sp
ec
i
f
icall
y
,
HI
T
L
-
MM
D
i
n
cl
u
d
es
ca
s
es
o
f
i
m
p
r
o
p
er
l
y
w
o
r
n
m
as
k
s
,
p
ar
tiall
y
co
v
er
ed
f
ac
es,
o
r
o
th
er
o
b
s
tr
u
ctio
n
s
,
p
r
o
v
id
in
g
a
m
o
r
e
co
m
p
r
e
h
en
s
iv
e
as
s
ess
m
e
n
t
o
f
th
e
m
o
d
el
'
s
ca
p
ab
ilit
ie
s
.
W
h
ile
d
atase
ts
l
ik
e
FMM
D
a
n
d
DR
FMD
o
f
f
er
a
s
o
lid
f
o
u
n
d
atio
n
f
o
r
m
o
d
el
t
r
ain
in
g
a
n
d
te
s
ti
n
g
,
t
h
e
d
i
v
er
s
it
y
an
d
co
m
p
lex
i
t
y
o
f
HI
T
L
-
MM
D
h
i
g
h
li
g
h
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8938
I
n
t J
A
r
ti
f
I
n
tell
,
Vo
l.
14
,
No
.
4
,
A
u
g
u
s
t
2025
:
2
6
3
4
-
2645
2642
asp
ec
ts
t
h
at
o
t
h
er
d
ataset
s
m
i
g
h
t
m
is
s
.
T
h
e
ab
s
e
n
ce
o
f
HI
T
L
-
MM
D
co
u
ld
r
es
u
lt
in
t
h
e
m
o
d
el
b
ein
g
l
i
m
i
ted
to
th
e
s
ee
n
d
ata
s
co
p
e
an
d
r
ed
u
ce
its
p
r
ac
tical
ap
p
licatio
n
.
Fig
u
r
e
3
s
h
o
w
s
th
e
e
v
al
u
ati
o
n
r
esu
lt
s
b
et
w
ee
n
th
e
Y
w
FMM
D
an
d
Y
w
D
R
F
MD
m
o
d
els
o
n
th
e
HI
T
L
-
M
MD
test
s
et.
I
t
is
clea
r
t
h
at
t
h
e
Y
w
D
R
FMD
m
o
d
el,
tr
ain
ed
o
n
th
e
DR
FMD
d
ataset,
p
er
f
o
r
m
s
s
i
g
n
if
ican
t
l
y
b
etter
th
an
th
e
Y
w
F
MM
D
m
o
d
el,
w
h
ich
w
a
s
tr
ain
ed
o
n
t
h
e
FMM
D
d
ata
s
et
.
T
h
is
d
e
m
o
n
s
tr
ates
th
at
t
h
e
YOL
O
v
8
m
o
d
els
tr
ai
n
ed
o
n
t
h
e
D
R
FMD
d
ataset
h
av
e
b
e
tter
g
e
n
er
aliza
tio
n
ca
p
ab
ilit
ies.
T
h
is
co
n
f
ir
m
s
t
h
e
d
iv
er
s
it
y
a
n
d
r
o
b
u
s
tn
e
s
s
o
f
t
h
e
p
r
o
p
o
s
ed
DR
FMD
d
ataset.
T
ab
le
6
.
T
esti
n
g
r
es
u
lt o
f
Y
wDRF
MD
m
o
d
els o
n
D
R
FMD,
FMM
D,
an
d
HI
T
L
-
MM
D
t
e
s
t
s
ets
M
o
d
e
l
D
R
F
M
D
T
e
st
se
t
(
I
)
F
M
M
D
T
e
st
se
t
(
I
I
)
H
I
TL
-
M
M
D
T
e
st
se
t
(
I
I
I
)
P
R
mA
P
@
5
0
P
R
mA
P
@
5
0
P
R
mA
P
@
5
0
Y
O
L
O
v
8
n
0
.
8
4
0
0
.
7
5
6
0
.
8
1
4
0
.
8
1
0
0
.
6
8
6
0
.
7
3
7
0
.
7
7
8
0
.
6
4
6
0
.
7
0
5
Y
O
L
O
v
8
s
0
.
8
3
9
0
.
7
9
7
0
.
8
4
4
0
.
8
1
4
0
.
7
1
5
0
.
7
7
2
0
.
7
7
8
0
.
6
9
8
0
.
7
3
1
Y
O
L
O
v
8
m
0
.
8
5
8
0
.
8
0
0
0
.
8
5
6
0
.
8
0
7
0
.
7
5
0
0
.
7
8
3
0
.
7
9
5
0
.
7
1
6
0
.
7
7
9
Y
O
L
O
v
8
l
0
.
8
6
6
0
.
8
1
1
0
.
8
6
7
0
.
7
7
1
0
.
7
5
8
0
.
7
7
6
0
.
7
8
7
0
.
7
1
2
0
.
7
6
9
Y
O
L
O
v
8
x
0
.
8
6
0
0
.
8
2
7
0
.
8
7
4
0
.
8
3
0
0
.
7
5
3
0
.
8
0
0
0
.
7
8
6
0
.
7
1
5
0
.
7
7
0
Fig
u
r
e
3
.
C
o
m
p
ar
in
g
th
e
p
er
f
o
r
m
an
ce
o
f
Y
w
DR
FMD
v
s
Y
wFMM
D
m
o
d
els o
n
t
h
e
s
a
m
e
te
s
ts
et
HI
T
L
-
MMD
3
.
2
.
3
.
Dis
cus
s
io
n o
n t
he
ef
f
ec
t
iv
eness
o
f
t
he
Yw
DRF
M
D
c
o
m
pa
re
d
t
o
o
t
her
s
t
ud
ie
s
I
n
th
i
s
d
is
cu
s
s
io
n
,
w
e
s
elec
t
tw
o
s
t
u
d
ies
[
5
]
,
[
6
]
,
to
c
o
m
p
ar
e
w
it
h
o
u
r
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
.
Sin
ce
th
e
d
atasets
u
s
ed
i
n
t
h
ese
s
t
u
d
ie
s
ar
e
d
if
f
er
en
t,
co
m
p
ar
in
g
ac
cu
r
ac
y
m
etr
ic
s
o
r
ev
al
u
atio
n
m
ea
s
u
r
es
s
u
ch
as
p
r
ec
is
io
n
,
r
ec
all
,
F1
s
co
r
e,
o
r
m
A
P
@
is
n
o
t
ap
p
r
o
p
r
iate.
T
h
er
ef
o
r
e,
ev
en
t
h
o
u
g
h
t
h
e
m
o
d
el
s
p
r
o
p
o
s
ed
in
[
5
]
,
[
6
]
ac
h
iev
e
s
ig
n
i
f
ica
n
tl
y
h
ig
h
er
ac
cu
r
ac
y
th
a
n
o
u
r
s
o
l
u
tio
n
,
t
h
e
d
atasets
u
s
ed
in
t
h
ese
p
r
o
p
o
s
als
h
av
e
n
o
tab
le
d
if
f
er
e
n
ce
s
.
A
cc
o
r
d
in
g
to
J
o
o
d
i
et
a
l.
[
5
]
,
th
e
M
AF
A
d
atase
t
p
r
i
m
ar
il
y
f
o
cu
s
es
o
n
f
ac
es
w
it
h
m
as
k
s
,
w
h
ic
h
m
a
y
li
m
it
t
h
e
d
iv
er
s
it
y
o
f
s
ce
n
ar
io
s
an
d
co
n
te
x
t
s
i
n
w
h
ic
h
f
ac
es
a
r
e
ca
p
tu
r
ed
.
T
h
is
co
u
ld
a
f
f
ec
t
th
e
m
o
d
el’
s
ab
ilit
y
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
A
r
ti
f
I
n
tell
I
SS
N:
2252
-
8938
E
n
h
a
n
ci
n
g
fa
ce
m
a
s
k
d
etec
tio
n
p
erfo
r
ma
n
ce
w
ith
co
mp
r
eh
e
n
s
ive
d
a
ta
s
et
a
n
d
YOLOv8
(
T
r
o
n
g
Th
u
a
Hu
yn
h
)
2643
to
g
en
er
alize
to
r
ea
l
-
w
o
r
ld
s
it
u
atio
n
s
w
h
er
e
li
g
h
tin
g
co
n
d
iti
o
n
s
,
a
n
g
le
s
,
a
n
d
b
ac
k
g
r
o
u
n
d
s
v
ar
y
s
i
g
n
if
ican
t
l
y
.
A
lt
h
o
u
g
h
t
h
e
M
AF
A
d
ataset
co
n
tai
n
s
a
s
u
b
s
ta
n
tial
n
u
m
b
er
o
f
i
m
a
g
es
(
3
5
,
8
0
6
MA
F
A
)
,
s
elec
ti
n
g
5
,
9
0
2
i
m
a
g
es
f
o
r
f
ac
e
an
al
y
s
i
s
m
a
y
n
o
t
en
co
m
p
a
s
s
all
p
o
s
s
ib
le
v
a
r
iatio
n
s
i
n
m
a
s
k
s
,
s
u
c
h
as
d
i
f
f
er
en
t
t
y
p
es,
co
lo
r
s
,
an
d
s
t
y
le
s
.
T
h
is
co
u
ld
li
m
it
th
e
m
o
d
el’
s
r
o
b
u
s
tn
e
s
s
in
d
etec
tin
g
m
as
k
s
u
n
d
er
d
iv
er
s
e
co
n
d
i
tio
n
s
.
A
d
d
itio
n
all
y
,
th
e
d
atase
t
s
e
lectio
n
f
o
r
t
h
e
s
t
u
d
y
f
o
c
u
s
es
o
n
f
r
o
n
tal
f
ac
es,
w
h
ic
h
m
a
y
n
o
t
f
u
l
l
y
r
ep
r
esen
t
th
e
ch
a
llen
g
e
s
o
f
d
etec
tin
g
m
a
s
k
s
o
n
f
ac
e
s
ca
p
t
u
r
ed
at
d
if
f
er
en
t
an
g
les
o
r
in
m
o
tio
n
.
T
h
is
co
u
ld
b
e
a
li
m
ita
tio
n
w
h
en
ap
p
l
y
in
g
th
e
m
o
d
el
in
d
y
n
a
m
ic
e
n
v
ir
o
n
m
en
ts
.
W
it
h
a
s
p
ec
if
ic
f
o
c
u
s
o
n
M
A
F
A
,
th
er
e
i
s
a
r
i
s
k
t
h
at
m
o
d
els
tr
ai
n
ed
o
n
th
e
M
AF
A
d
ata
s
et
m
a
y
o
v
er
f
i
t
to
its
ch
ar
ac
ter
i
s
tics
,
p
o
ten
ti
all
y
r
ed
u
c
in
g
t
h
eir
e
f
f
ec
ti
v
en
e
s
s
o
n
o
th
er
d
ataset
s
o
r
in
r
ea
l
-
w
o
r
ld
ap
p
licatio
n
s
w
h
er
e
f
ac
e
s
ar
e
ei
th
er
u
n
m
ask
ed
o
r
o
n
l
y
p
ar
tiall
y
co
v
er
ed
.
Me
an
w
h
ile,
t
h
e
co
m
p
r
e
h
en
s
iv
e
d
ataset
p
r
o
p
o
s
ed
in
o
u
r
s
t
u
d
y
ad
d
r
ess
e
s
t
h
ese
i
s
s
u
es
w
h
ile
m
ain
tain
in
g
h
ig
h
p
er
f
o
r
m
an
ce
w
h
e
n
ev
al
u
ated
o
n
d
if
f
er
en
t d
atasets
,
as
s
h
o
w
n
in
t
h
e
p
r
ev
io
u
s
s
ec
tio
n
(
3
.
2
.
2
)
.
R
esear
ch
b
y
Ha
s
s
a
n
et
a
l.
[
6
]
,
th
e
d
ataset
co
n
s
i
s
ts
o
f
2
,
1
6
5
i
m
ag
e
s
o
f
M
A
F
A
a
n
d
1
,
9
3
0
i
m
a
g
es
o
f
u
n
m
a
s
k
ed
f
ac
es.
W
h
ile
t
h
i
s
m
a
y
b
e
s
u
f
f
icie
n
t
f
o
r
in
itial
m
o
d
el
tr
ain
i
n
g
,
it
m
a
y
n
o
t
b
e
lar
g
e
en
o
u
g
h
to
ca
p
tu
r
e
th
e
f
u
ll
v
ar
iab
ilit
y
o
f
r
ea
l
-
w
o
r
ld
s
ce
n
ar
io
s
,
p
o
ten
tiall
y
a
f
f
e
ctin
g
th
e
m
o
d
el's
g
e
n
er
aliza
ti
o
n
ca
p
ab
ilit
y
.
T
h
e
d
ataset
is
ca
te
g
o
r
ized
in
to
o
n
l
y
t
w
o
clas
s
es
m
a
s
k
ed
a
n
d
u
n
m
as
k
ed
f
ac
e
s
.
T
h
is
b
i
n
ar
y
cla
s
s
i
f
icatio
n
d
o
es
n
o
t
ac
co
u
n
t
f
o
r
p
ar
tiall
y
w
o
r
n
m
a
s
k
s
o
r
i
m
p
r
o
p
er
l
y
u
s
ed
m
as
k
s
,
w
h
ic
h
ar
e
co
m
m
o
n
in
r
ea
l
-
wo
r
ld
s
itu
atio
n
s
a
n
d
m
a
y
lead
to
m
is
cla
s
s
i
f
icat
io
n
.
T
h
e
im
a
g
e
s
in
t
h
e
d
ataset
w
er
e
cr
o
p
p
e
d
to
f
o
cu
s
s
o
lel
y
o
n
t
h
e
f
ac
e
s
.
W
h
ile
t
h
is
s
i
m
p
li
f
ie
s
th
e
m
o
d
el’
s
tas
k
,
i
t
m
a
y
n
o
t
r
ef
lect
r
ea
l
-
w
o
r
ld
co
n
d
itio
n
s
w
h
er
e
f
ac
e
s
ar
e
n
o
t
al
w
a
y
s
p
er
f
ec
t
l
y
alig
n
ed
o
r
f
u
ll
y
v
i
s
ib
le,
p
o
ten
tiall
y
i
m
p
ac
ti
n
g
th
e
m
o
d
el’
s
p
er
f
o
r
m
a
n
ce
in
p
r
ac
tical
ap
p
licatio
n
s
.
A
d
d
itio
n
al
l
y
,
w
it
h
a
r
elat
iv
el
y
s
m
all
d
ata
s
et,
t
h
e
r
ep
o
r
ted
ac
cu
r
ac
y
o
f
9
9
%
d
u
r
in
g
tr
ai
n
i
n
g
an
d
1
0
0
%
d
u
r
i
n
g
test
i
n
g
r
aise
s
co
n
ce
r
n
s
ab
o
u
t
o
v
er
f
itti
n
g
,
w
h
er
e
th
e
m
o
d
e
l
lear
n
s
th
e
tr
ai
n
i
n
g
d
ata
to
o
w
ell
b
u
t
f
ails
to
p
er
f
o
r
m
e
f
f
ec
ti
v
el
y
o
n
u
n
s
ee
n
d
ata.
T
h
is
is
s
u
e
i
s
p
ar
ticu
la
r
l
y
co
n
ce
r
n
i
n
g
i
f
t
h
e
d
ataset
d
o
es
n
o
t
in
cl
u
d
e
a
w
id
e
r
a
n
g
e
o
f
v
ar
iatio
n
s
in
f
a
cial
ap
p
ea
r
an
ce
s
a
n
d
m
as
k
t
y
p
es.
Me
an
w
h
ile,
th
e
DR
FMD
d
ataset
w
e
p
r
o
p
o
s
e
o
f
f
er
s
g
r
ea
ter
d
iv
er
s
it
y
i
n
ter
m
s
o
f
d
e
m
o
g
r
ap
h
ics,
li
g
h
tin
g
co
n
d
itio
n
s
,
an
d
m
u
ltip
le
i
m
ag
e
an
g
les.
T
h
is
is
p
ar
ticu
lar
l
y
u
s
e
f
u
l
f
o
r
ef
f
ec
t
iv
el
y
d
etec
tin
g
v
ar
io
u
s
ca
s
e
s
o
f
b
o
th
p
r
o
p
er
an
d
i
m
p
r
o
p
er
m
a
s
k
-
w
ea
r
i
n
g
,
en
s
u
r
in
g
a
m
o
r
e
r
o
b
u
s
t a
n
d
g
e
n
er
aliza
b
le
m
o
d
el
f
o
r
r
ea
l
-
w
o
r
ld
ap
p
licatio
n
s
.
4.
CO
NCLU
SI
O
N
T
h
is
s
t
u
d
y
f
o
cu
s
es
o
n
p
r
o
p
o
s
in
g
t
h
e
DR
FMD
d
ata
s
et
a
n
d
ap
p
ly
i
n
g
YO
L
O
v
8
m
o
d
el
s
to
i
m
p
r
o
v
e
m
as
k
r
ec
o
g
n
itio
n
p
er
f
o
r
m
a
n
c
e
o
n
f
ac
e
s
ac
r
o
s
s
v
ar
io
u
s
i
n
p
u
t
i
m
a
g
e
t
y
p
e
s
.
T
h
e
r
esu
lt
s
o
b
tain
ed
f
r
o
m
tr
ain
i
n
g
an
d
test
i
n
g
o
n
th
e
D
R
FMD
d
ataset
s
h
o
w
t
h
at
YO
L
Ov
8
ca
n
ac
cu
r
atel
y
d
etec
t
an
d
clas
s
i
f
y
ca
s
es
o
f
w
ea
r
i
n
g
m
as
k
s
,
n
o
t
w
ea
r
i
n
g
m
as
k
s
,
a
n
d
w
ea
r
in
g
m
a
s
k
s
i
m
p
r
o
p
er
l
y
w
it
h
h
ig
h
ac
cu
r
ac
y
.
E
x
p
er
i
m
e
n
ts
d
e
m
o
n
s
tr
ate
t
h
at
th
e
Y
OL
Ov
8
m
o
d
el
tr
ai
n
ed
o
n
D
R
FMD
o
u
tp
er
f
o
r
m
s
YO
L
Ov
8
m
o
d
el
s
tr
ai
n
ed
o
n
o
t
h
er
d
atasets
li
k
e
FMM
D,
p
r
o
v
in
g
i
ts
b
r
o
ad
ap
p
licab
ilit
y
in
p
u
b
lic
h
ea
lth
m
o
n
ito
r
in
g
an
d
d
is
ea
s
e
p
r
ev
e
n
tio
n
.
T
h
is
p
r
o
p
o
s
ed
d
ataset
is
co
m
p
iled
f
r
o
m
r
ep
u
tab
le
s
o
u
r
ce
s
s
u
ch
as
A
I
Z
O
O,
KFM
D,
MA
F
A
,
MO
X
A
3
K,
an
d
t
h
e
Z
alo
A
I
c
h
alle
n
g
e,
en
s
u
r
in
g
g
r
ea
ter
d
iv
er
s
it
y
a
n
d
g
en
er
aliza
t
io
n
ca
p
ab
ilit
y
f
o
r
t
h
e
m
o
d
el.
A
d
d
itio
n
all
y
,
u
s
i
n
g
d
ata
au
g
m
e
n
tatio
n
tech
n
iq
u
es
s
u
ch
a
s
p
ad
d
in
g
,
cr
o
p
p
in
g
,
an
d
h
o
r
izo
n
tal
f
l
ip
p
in
g
h
as
e
n
h
an
ce
d
t
h
e
m
o
d
el'
s
p
er
f
o
r
m
a
n
ce
,
en
ab
lin
g
it
to
b
etter
h
an
d
le
d
i
v
er
s
e
r
ea
l
-
w
o
r
ld
s
itu
at
io
n
s
.
T
h
is
r
esear
c
h
s
i
g
n
i
f
ican
tl
y
co
n
t
r
ib
u
tes
to
i
m
p
r
o
v
in
g
th
e
e
f
f
ec
ti
v
en
e
s
s
an
d
ac
c
u
r
ac
y
o
f
m
a
s
k
r
ec
o
g
n
itio
n
s
y
s
te
m
s
,
esp
ec
iall
y
in
t
h
e
co
n
tex
t
o
f
cu
r
r
en
t
p
u
b
lic
h
ea
lt
h
is
s
u
es,
an
d
o
p
en
s
u
p
n
e
w
d
ir
ec
tio
n
s
f
o
r
d
ev
elo
p
in
g
r
ich
d
atasets
a
n
d
ad
v
an
ce
d
d
ee
p
lear
n
in
g
tech
n
iq
u
es.
Fu
r
t
h
er
m
o
r
e,
b
ased
o
n
t
h
e
p
r
o
p
o
s
ed
DR
FMD
d
ata
s
et,
o
u
r
f
u
t
u
r
e
r
esear
ch
ai
m
s
to
e
n
h
an
ce
t
h
e
Y
O
L
Ov
8
m
o
d
el
a
n
d
s
u
b
s
eq
u
e
n
t Y
O
L
O
v
er
s
io
n
s
to
r
ed
u
ce
tr
ain
i
n
g
t
i
m
e
an
d
f
u
r
t
h
er
i
m
p
r
o
v
e
ac
c
u
r
ac
y
.
ACK
NO
WL
E
D
G
M
E
NT
S
T
h
e
au
th
o
r
s
e
x
te
n
d
th
e
ir
ap
p
r
ec
iatio
n
to
t
h
e
P
o
s
ts
a
n
d
T
elec
o
m
m
u
n
icatio
n
s
I
n
s
tit
u
te
o
f
T
ec
h
n
o
lo
g
y
(
PT
I
T
)
in
Vietn
a
m
f
o
r
f
o
r
f
in
a
n
cial
s
u
p
p
o
r
tin
g
a
n
d
f
ac
il
itati
n
g
th
i
s
r
esear
ch
.
F
UNDIN
G
I
NF
O
RM
AT
I
O
N
T
h
is
r
esear
ch
r
ec
eiv
ed
n
o
s
p
ec
if
ic
g
r
a
n
t
f
r
o
m
a
n
y
f
u
n
d
i
n
g
a
g
en
c
y
in
t
h
e
p
u
b
lic,
co
m
m
er
ci
al,
o
r
n
o
t
-
f
o
r
-
p
r
o
f
it secto
r
s
.
AUTHO
R
CO
NT
RIB
UT
I
O
NS ST
A
T
E
M
E
NT
T
h
is
j
o
u
r
n
al
u
s
e
s
th
e
C
o
n
tr
ib
u
to
r
R
o
les
T
ax
o
n
o
m
y
(
C
R
ed
iT
)
to
r
ec
o
g
n
ize
in
d
i
v
id
u
al
au
th
o
r
co
n
tr
ib
u
tio
n
s
,
r
ed
u
ce
au
t
h
o
r
s
h
ip
d
is
p
u
tes,
an
d
f
ac
ilit
ate
co
lla
b
o
r
atio
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.