I
nte
rna
t
io
na
l J
o
urna
l o
f
Adv
a
nces in Applie
d Science
s
(
I
J
AAS)
Vo
l.
1
3
,
No
.
4
,
Dec
em
b
er
2
0
2
4
,
p
p
.
9
9
1
~
1
0
0
3
I
SS
N:
2252
-
8
8
1
4
,
DOI
:
1
0
.
1
1
5
9
1
/ijaas
.
v
1
3
.
i
4
.
pp
991
-
1
0
0
3
991
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
a
s
.
ia
esco
r
e.
co
m
Rea
l
-
time
sm
o
ke
a
nd
fire
dete
ction
using
you
o
nly
lo
o
k
o
nce
v8
-
ba
sed
a
dv
a
nced
c
o
mputer
v
isio
n
a
nd
deep
lea
rning
Sh
a
k
ila
Ra
hm
a
n
1
,
Sy
ed
M
uh
a
m
m
a
d
H
a
s
na
t
J
a
m
ee
1
,
J
a
k
a
ria
K
ha
n
Ra
f
i
1
,
J
a
f
rin
Su
lt
a
na
J
uthi
1
,
Sa
j
ib
Abdu
l
Aziz
1
,
J
ia
Uddi
n
2
1
D
e
p
a
r
t
me
n
t
of
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
r
i
n
g
,
F
a
c
u
l
t
y
of
S
c
i
e
n
c
e
a
n
d
Te
c
h
n
o
l
o
g
y
,
A
meri
c
a
n
I
n
t
e
r
n
a
t
i
o
n
a
l
U
n
i
v
e
r
si
t
y
-
B
a
n
g
l
a
d
e
sh
,
D
h
a
k
a
,
B
a
n
g
l
a
d
e
s
h
2
D
e
p
a
r
t
me
n
t
o
f
A
r
t
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
a
n
d
B
i
g
D
a
t
a
,
C
o
l
l
e
g
e
o
f
E
n
d
i
c
o
t
t
,
W
o
o
s
o
n
g
U
n
i
v
e
r
si
t
y
,
D
a
e
j
e
o
n
,
R
e
p
u
b
l
i
c
of
K
o
r
e
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Feb
2
6
,
2
0
2
4
R
ev
is
ed
Au
g
2
9
,
2
0
2
4
Acc
ep
ted
Sep
4
,
2
0
2
4
F
ire
a
n
d
sm
o
k
e
p
o
se
se
v
e
re
t
h
r
e
a
ts,
c
a
u
sin
g
d
a
m
a
g
e
to
p
r
o
p
e
rt
y
a
n
d
t
h
e
e
n
v
iro
n
m
e
n
t
a
n
d
e
n
d
a
n
g
e
ri
n
g
l
iv
e
s.
Trad
it
io
n
a
l
f
ire
d
e
tec
ti
o
n
m
e
th
o
d
s
stru
g
g
le
with
a
c
c
u
ra
c
y
a
n
d
sp
e
e
d
,
h
in
d
e
rin
g
re
a
l
-
ti
m
e
d
e
tec
ti
o
n
.
Th
u
s,
th
i
s
stu
d
y
in
tro
d
u
c
e
s
an
imp
ro
v
e
d
fi
re
a
n
d
sm
o
k
e
d
e
tec
ti
o
n
a
p
p
ro
a
c
h
u
ti
li
z
in
g
th
e
y
o
u
o
n
l
y
l
o
o
k
o
n
c
e
(
YO
LO
)
v8
-
b
a
se
d
d
e
e
p
lea
rn
i
n
g
m
o
d
e
l.
Th
is
w
o
rk
a
ims
to
e
n
h
a
n
c
e
a
c
c
u
ra
c
y
a
n
d
sp
e
e
d
,
wh
ich
a
re
c
ru
c
ial
fo
r
e
a
rly
fire
d
e
tec
ti
o
n
.
T
h
e
m
e
th
o
d
o
l
o
g
y
i
n
v
o
lv
e
s
p
re
p
r
o
c
e
ss
in
g
a
larg
e
d
a
tas
e
t
c
o
n
tain
i
n
g
5
,
7
0
0
ima
g
e
s
d
e
p
ictin
g
fire
a
n
d
sm
o
k
e
sc
e
n
a
rio
s.
YO
LOv
8
h
a
s
b
e
e
n
train
e
d
a
n
d
v
a
li
d
a
ted
,
o
u
t
p
e
rfo
rm
in
g
s
o
m
e
b
a
se
li
n
e
m
o
d
e
ls
-
YO
LOv
7
,
YO
LOv
5
,
Re
sN
e
t
-
32,
a
n
d
M
o
b
il
e
Ne
t
-
v2
in
t
h
e
p
re
c
isio
n
,
re
c
a
ll
,
a
n
d
mean
a
v
e
ra
g
e
p
re
c
isio
n
(m
AP)
m
e
tri
c
s.
Th
e
p
ro
p
o
se
d
m
e
th
o
d
a
c
h
iev
e
s
6
8
.
3
%
p
re
c
isio
n
,
5
4
.
6
%
re
c
a
ll
,
6
0
.
7
%
F1
sc
o
re
,
a
n
d
5
7
.
3
%
m
AP.
I
n
teg
ra
ti
n
g
YO
LOv
8
in
fire
a
n
d
sm
o
k
e
d
e
t
e
c
ti
o
n
sy
ste
m
s
can
sig
n
ifi
c
a
n
tl
y
imp
ro
v
e
re
sp
o
n
se
ti
m
e
s,
e
n
h
a
n
c
e
th
e
a
b
il
it
y
to
m
it
ig
a
te
f
ire
o
u
tb
re
a
k
s,
a
n
d
p
o
ten
ti
a
ll
y
sa
v
e
li
v
e
s
a
n
d
p
ro
p
e
rty
.
Th
is
re
se
a
rc
h
a
d
v
a
n
c
e
s
fire
d
e
tec
ti
o
n
sy
s
tem
s
a
n
d
e
sta
b
li
sh
e
s
a
p
re
c
e
d
e
n
t
fo
r
a
p
p
l
y
in
g
d
e
e
p
lea
rn
i
n
g
tec
h
n
iq
u
e
s
to
c
rit
ica
l
sa
fe
ty
a
p
p
li
c
a
ti
o
n
s,
p
u
s
h
in
g
t
h
e
b
o
u
n
d
a
ries
of
in
n
o
v
a
ti
o
n
in
p
u
b
l
ic
sa
fe
ty
.
K
ey
w
o
r
d
s
:
C
o
m
p
u
ter
v
is
io
n
Fire
d
etec
tio
n
I
m
ag
e
p
r
o
ce
s
s
in
g
Ob
ject
d
etec
tio
n
m
o
d
el
Sm
o
k
e
d
etec
tio
n
Yo
u
o
n
ly
lo
o
k
o
n
ce
v8
T
h
is
is
an
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r
th
e
CC
BY
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
J
ia
Ud
d
in
Dep
ar
tm
en
t o
f
Ar
tific
ial
I
n
tellig
en
ce
an
d
B
ig
Data
,
C
o
lleg
e
o
f
E
n
d
ic
o
tt,
W
o
o
s
o
n
g
U
n
iv
er
s
ity
Dae
jeo
n
3
4
6
0
6
,
R
ep
u
b
lic
of
Ko
r
ea
E
m
ail:
jia.
u
d
d
in
@
wsu
.
ac
.
k
r
1.
I
NT
RO
D
UCT
I
O
N
T
h
e
m
ain
s
moke
d
etec
tio
n
p
l
ay
s
a
v
ital
r
o
le
in
m
itig
atin
g
t
h
e
r
is
k
s
ass
o
ciate
d
with
f
ir
es,
in
d
u
s
tr
ial
ac
cid
en
ts
,
an
d
n
atu
r
al
d
is
aster
s
.
T
r
ad
itio
n
al
m
eth
o
d
s
,
wh
ic
h
p
r
ed
o
m
in
an
tly
u
tili
ze
h
ea
t
or
o
p
tical
s
en
s
o
r
s
,
o
f
ten
s
u
f
f
er
f
r
o
m
d
ela
y
ed
r
esp
o
n
s
es
an
d
lim
itatio
n
s
in
ac
c
u
r
ac
y
,
p
ar
ticu
lar
l
y
in
ch
allen
g
in
g
e
n
v
ir
o
n
m
en
ts
wh
er
e
f
ac
to
r
s
s
u
ch
as
s
team
or
air
b
o
r
n
e
p
ar
ticles
can
lea
d
to
f
alse
alar
m
s
or
m
is
s
ed
d
etec
ti
o
n
s
.
T
h
e
ad
v
e
n
t
of
co
m
p
u
ter
v
is
io
n
a
n
d
m
ac
h
in
e
lear
n
in
g
(
ML
)
tech
n
o
lo
g
ies
h
as
p
av
ed
th
e
way
f
o
r
m
o
r
e
s
o
p
h
is
ticated
s
m
o
k
e
d
etec
tio
n
s
y
s
tem
s
.
Am
o
n
g
th
ese,
you
o
n
ly
lo
o
k
o
n
ce
(
YOL
O)
V8
[
1
]
m
o
d
el
s
tan
d
s
o
u
t
as
a
s
tate
-
of
-
th
e
-
ar
t
s
o
lu
tio
n
f
o
r
r
ea
l
-
tim
e
d
etec
tio
n
.
T
h
is
m
o
d
el,
p
ar
t
of
th
e
Y
OL
O
s
er
ies,
o
p
er
ates
on
a
s
in
g
le
-
s
tag
e
d
etec
tio
n
p
r
in
cip
le,
o
f
f
er
i
n
g
s
u
b
s
tan
tial
im
p
r
o
v
em
en
ts
in
s
p
ee
d
a
n
d
ac
cu
r
ac
y
o
v
e
r
its
p
r
e
d
ec
ess
o
r
s
an
d
two
-
s
tag
e
co
u
n
ter
p
a
r
ts
lik
e
f
aster
r
eg
io
n
-
b
ased
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
(
R
-
C
NN
)
[
2
]
,
[
3
]
,
Ma
s
k
R
-
C
NN
[
4
]
,
an
d
o
th
er
s
eg
m
e
n
tatio
n
-
b
ased
tech
n
iq
u
es
[
5
]
–
[
7
]
.
No
tab
ly
,
YOL
Ov
8
'
s
d
ee
p
ar
ch
itect
u
r
e
en
s
u
r
es
r
o
b
u
s
t
p
er
f
o
r
m
an
ce
ac
r
o
s
s
d
iv
e
r
s
e
c
o
n
d
itio
n
s
,
i
n
clu
d
in
g
poor
lig
h
tin
g
an
d
co
m
p
lex
b
ac
k
g
r
o
u
n
d
s
.
T
h
e
e
v
o
lu
tio
n
of
YOL
O
m
o
d
els
f
r
o
m
v
er
s
io
n
s
4
th
r
o
u
g
h
7
h
as
b
ee
n
m
a
r
k
ed
by
s
ig
n
if
ica
n
t
en
h
a
n
ce
m
en
ts
in
d
etec
tio
n
ca
p
ab
ilit
ies,
p
r
o
ce
s
s
in
g
s
p
ee
d
,
an
d
u
s
er
ac
ce
s
s
ib
ilit
y
.
Star
ti
n
g
with
YOL
Ov
4
[
8
]
,
w
h
ich
was
n
o
tab
le
f
o
r
its
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
8
1
4
I
n
t J Ad
v
Ap
p
l Sci
,
Vo
l.
1
3
,
No
.
4
,
Dec
em
b
er
2
0
2
4
:
991
-
1
0
0
3
992
ab
ilit
y
to
id
en
tify
a
wid
e
ar
r
ay
of
o
b
jects
ac
r
o
s
s
a
v
ast
d
ataset,
each
s
u
b
s
eq
u
en
t
v
er
s
io
n
h
as
in
tr
o
d
u
ce
d
im
p
r
o
v
em
e
n
ts
in
s
p
ee
d
,
ac
cu
r
ac
y
,
a
n
d
f
u
n
ctio
n
ality
,
cu
lm
i
n
atin
g
in
YOL
Ov
7
'
s
u
n
p
ar
all
eled
ef
f
icien
cy
[
9
]
an
d
YOL
Ov
8
'
s
ad
v
an
ce
d
f
ea
t
u
r
es
[
1
0
]
,
lik
e
m
u
lti
-
o
b
ject
tr
ac
k
in
g
an
d
o
p
tim
ized
p
er
f
o
r
m
a
n
ce
f
o
r
c
h
allen
g
in
g
d
etec
tio
n
s
.
R
esear
ch
lite
r
atu
r
e
u
n
d
er
s
co
r
es
th
e
ef
f
ec
tiv
en
ess
of
YOL
O
m
o
d
els
in
s
m
o
k
e
d
etec
tio
n
.
Fo
r
in
s
tan
ce
,
s
tu
d
ies
[
1
1
]
a
n
d
[
1
2
]
de
m
o
n
s
tr
ate
th
e
ap
p
licatio
n
of
YOL
Ov
5
an
d
YOL
Ov
6
,
r
esp
ec
tiv
ely
,
in
d
etec
tin
g
s
m
o
k
e
with
h
i
g
h
a
cc
u
r
ac
y
,
lev
e
r
ag
in
g
tem
p
o
r
al
in
f
o
r
m
atio
n
,
an
d
en
h
a
n
cin
g
s
p
atial
f
ea
tu
r
es
to
m
in
im
ize
f
alse
p
o
s
itiv
es.
T
h
is
wo
r
k
p
r
esen
ts
a
co
m
p
r
e
h
en
s
iv
e
ev
alu
atio
n
of
a
s
m
o
k
e
d
etec
tio
n
s
y
s
tem
b
ased
on
th
e
YOL
O
v
8
m
o
d
el.
By
tr
ain
in
g
th
e
m
o
d
el
on
a
cu
s
to
m
d
ataset
co
m
p
r
is
in
g
s
m
o
k
e
a
n
d
n
o
n
-
s
m
o
k
e
im
ag
es,
an
d
em
p
l
o
y
in
g
m
etr
ics
s
u
ch
as
p
r
ec
is
io
n
,
r
ec
all,
an
d
ac
cu
r
ac
y
,
th
e
s
tu
d
y
ai
m
s
to
o
f
f
er
an
ad
v
an
ce
d
s
o
lu
tio
n
f
o
r
f
ir
e
s
af
ety
.
T
h
e
s
y
s
tem
'
s
f
lex
ib
ilit
y
an
d
ef
f
icien
cy
m
ak
e
it
a
p
p
licab
le
in
r
ea
l
-
wo
r
ld
s
ce
n
a
r
io
s
,
in
clu
d
in
g
f
ir
e
d
etec
tio
n
,
in
d
u
s
tr
ial
s
af
ety
,
en
v
ir
o
n
m
en
tal
m
o
n
ito
r
in
g
,
an
d
p
o
ten
tial
ly
r
ev
o
lu
tio
n
izin
g
th
e
d
ev
el
o
p
m
en
t
of
s
m
o
k
e
d
etec
tio
n
tech
n
o
lo
g
ies.
T
h
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
is
u
n
q
u
esti
o
n
ab
ly
f
lex
ib
le
an
d
can
be
u
tili
ze
d
in
tr
u
e
cir
cu
m
s
tan
ce
s
,
s
u
ch
as
f
ir
e
d
etec
tio
n
s
y
s
tem
s
,
in
d
u
s
tr
ial
s
af
ety
s
y
s
tem
s
,
an
d
en
v
ir
o
n
m
en
tal
m
o
n
ito
r
in
g
s
y
s
tem
s
.
Ad
d
itio
n
ally
,
it
en
a
b
les
th
e
in
v
en
tio
n
of
s
er
v
ices
an
d
p
r
o
d
u
cts
s
u
ch
as
s
m
o
k
e
alar
m
s
an
d
s
m
o
k
e
d
etec
tio
n
ca
m
er
as.
YOL
O
v
8
is
a
p
o
wer
f
u
l
s
m
o
k
e
d
ete
ctio
n
m
o
d
el
th
at
is
well
-
s
u
i
ted
f
o
r
r
ea
l
-
wo
r
ld
ap
p
licatio
n
s
.
It
is
f
ast,
ac
cu
r
ate,
ef
f
icien
t,
an
d
v
e
r
s
atile.
YOL
O
v
8
is
f
ast
en
o
u
g
h
to
be
u
s
ed
f
o
r
r
ea
l
-
tim
e
s
m
o
k
e
d
etec
tio
n
;
it
can
d
etec
t
s
m
o
k
e
in
v
id
e
o
s
an
d
im
a
g
es
as
th
ey
ar
e
b
ein
g
ca
p
tu
r
e
d
.
T
h
is
is
im
p
o
r
tan
t
f
o
r
ap
p
licatio
n
s
s
u
ch
as
s
m
o
k
e
d
etec
tio
n
s
y
s
tem
s
in
b
u
ild
in
g
s
an
d
f
o
r
ests
.
YOL
O
v
8
is
als
o
ac
cu
r
ate,
ev
e
n
in
ch
allen
g
in
g
lig
h
tin
g
c
o
n
d
itio
n
s
an
d
co
m
p
lex
b
ac
k
g
r
o
u
n
d
s
,
in
clu
d
in
g
d
ar
k
r
o
o
m
s
,
f
o
g
g
y
f
o
r
ests
,
an
d
b
u
s
y
s
tr
ee
ts
.
YOL
O
v
8
is
ef
f
icien
t
a
n
d
r
e
q
u
ir
es
r
elativ
ely
f
ew
r
eso
u
r
ce
s
to
r
u
n
,
w
h
ich
m
a
k
es
it
s
u
itab
le
f
o
r
u
s
e
on
a
v
ar
iety
of
d
e
v
ices,
in
clu
d
in
g
s
m
ar
tp
h
o
n
es,
em
b
e
d
d
ed
s
y
s
tem
s
,
an
d
s
er
v
er
s
.
YOL
OV
v
8
is
v
er
s
atile
an
d
can
be
u
s
ed
to
d
etec
t
a
wid
e
v
ar
iet
y
of
s
m
o
k
e
ty
p
es
an
d
s
izes,
s
u
ch
as
f
ir
e
d
etec
tio
n
s
y
s
tem
s
,
in
d
u
s
tr
ial
s
af
ety
s
y
s
tem
s
,
an
d
en
v
ir
o
n
m
en
tal
m
o
n
ito
r
in
g
s
y
s
tem
s
.
T
h
e
p
r
o
p
o
s
ed
YOL
O
v
8
-
b
ased
s
m
o
k
e
d
etec
tio
n
tech
n
o
lo
g
y
h
as
th
e
p
o
ten
tial
to
s
ig
n
if
ica
n
t
ly
im
p
r
o
v
e
b
o
t
h
en
v
i
r
o
n
m
e
n
ta
l
an
d
p
u
b
lic
s
af
ety
.
T
h
e
m
ain
co
n
tr
ib
u
tio
n
of
th
e
p
r
o
p
o
s
ed
wo
r
k
is
,
i)
First,
a
la
r
g
e
d
ataset
co
n
tain
in
g
5
,
700
i
m
ag
es
h
as
b
ee
n
b
u
ilt.
T
h
e
n
th
e
co
llecte
d
im
ag
e
d
ata
is
p
r
e
-
p
r
o
ce
s
s
ed
u
s
in
g
v
ar
io
u
s
im
ag
e
p
r
o
ce
s
s
in
g
to
o
ls
s
u
ch
as
au
g
m
en
tatio
n
,
r
esizin
g
,
an
n
o
t
atin
g
,
or
lab
elin
g
;
ii)
Af
ter
t
h
at,
th
e
p
r
e
-
p
r
o
ce
s
s
ed
d
ata
is
p
r
o
ce
s
s
ed
wh
ich
p
r
ep
ar
es
t
h
e
d
ata
f
o
r
tr
ai
n
in
g
.
T
h
en
th
e
p
r
o
ce
s
s
ed
d
ata
is
t
r
ain
ed
u
s
in
g
th
e
YOL
Ov
8
-
b
ased
m
o
d
el
;
an
d
iii)
Fin
ally
,
th
e
r
esu
lts
of
th
e
p
r
o
p
o
s
ed
YOL
Ov
8
-
b
ased
m
o
d
el
p
er
f
o
r
m
an
ce
ar
e
c
o
m
p
ar
e
d
with
o
th
er
b
aselin
e
m
o
d
els,
s
u
ch
as
YOL
Ov
7
,
YOL
Ov
5
,
Mo
b
ile
N
et
v
2
,
a
n
d
R
es
N
et3
2
,
wh
er
e
YOL
Ov
8
o
u
t
p
er
f
o
r
m
s
th
em
in
all
ca
s
es.
2.
L
I
T
E
R
AT
U
RE
R
E
VI
E
W
Sm
o
k
e
an
d
f
ir
e
d
etec
tio
n
u
s
in
g
YOL
Ov
8
r
e
v
iews
a
d
i
v
er
s
e
r
an
g
e
of
s
tu
d
ies
s
h
o
w
ca
s
in
g
th
e
ev
o
lv
in
g
ca
p
a
b
ilit
ies
an
d
ap
p
licatio
n
s
of
th
e
YOL
O
alg
o
r
ith
m
in
s
m
o
k
e
an
d
f
ir
e
d
et
ec
tio
n
.
T
h
e
s
tu
d
ies
h
ig
h
lig
h
t
s
ig
n
if
ican
t
ad
v
an
c
em
en
ts
in
s
p
ee
d
,
an
d
ef
f
ici
en
cy
ac
r
o
s
s
v
ar
io
u
s
co
n
te
x
ts
,
in
clu
d
in
g
u
r
b
an
en
v
ir
o
n
m
en
ts
,
m
ar
i
n
e
e
n
g
in
e
r
o
o
m
s
,
an
d
f
o
r
est
ar
ea
s
.
Key
c
o
n
tr
ib
u
tio
n
s
in
v
o
lv
e
th
e
in
te
g
r
atio
n
of
lig
h
tweig
h
t
m
o
d
els
lik
e
L
ig
h
t
-
YOL
Ov
5
,
ac
co
m
p
lis
h
in
g
an
ad
ju
s
tm
en
t
b
etwe
en
ex
ac
t
n
ess
an
d
co
m
p
u
tatio
n
al
ef
f
ec
tiv
en
ess
f
o
r
r
ea
l
-
tim
e
ap
p
licatio
n
s
.
T
h
e
liter
atu
r
e
in
d
ic
ates
a
tr
en
d
to
war
d
more
s
o
p
h
is
ticated
,
ef
f
icien
t,
an
d
v
er
s
atile
YOL
O
-
b
ased
s
m
o
k
e
a
n
d
f
ir
e
d
etec
tio
n
s
y
s
tem
s
,
with
p
o
ten
tial
ap
p
licatio
n
s
in
em
er
g
in
g
tech
n
o
lo
g
ies
lik
e
th
e
in
ter
n
et
o
f
th
in
g
s
(
I
o
T
)
an
d
clo
u
d
c
o
m
p
u
tin
g
f
o
r
en
h
an
ce
d
em
er
g
en
cy
m
an
a
g
em
en
t
s
y
s
tem
s
.
W
h
ile
ad
v
an
ce
m
en
ts
in
YOL
O
-
b
ased
s
m
o
k
e
d
etec
tio
n
alg
o
r
ith
m
s
h
av
e
s
h
o
wn
s
ig
n
if
ican
t
p
r
o
g
r
ess
,
k
ey
ar
ea
s
lik
e
ad
a
p
tab
ilit
y
to
d
iv
er
s
e
an
d
c
h
allen
g
in
g
en
v
i
r
o
n
m
en
ts
,
co
m
p
u
tatio
n
al
e
f
f
ici
en
cy
f
o
r
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
s
ettin
g
s
,
an
d
s
p
ec
if
icity
to
v
ar
y
in
g
ty
p
es
of
s
m
o
k
e
an
d
f
ir
e
s
ce
n
ar
io
s
n
ee
d
f
u
r
th
er
r
esear
ch
an
d
d
ev
elo
p
m
e
n
t.
T
h
is
o
n
g
o
i
n
g
r
e
f
in
em
en
t
an
d
test
in
g
ar
e
cr
u
ci
al
f
o
r
th
e
p
r
ac
tical
an
d
wid
esp
r
ea
d
ap
p
licatio
n
of
th
ese
tech
n
o
lo
g
ies
in
r
ea
l
-
wo
r
ld
s
ce
n
ar
io
s
.
YOL
O
-
F
[
1
3
]
s
h
o
ws
h
ig
h
ac
cu
r
ac
y
u
s
in
g
YOL
Ov
3
an
d
YOL
Ov
4
f
o
r
f
lam
e
d
etec
tio
n
.
A
n
o
tab
l
e
s
h
o
r
tco
m
in
g
is
th
e
lac
k
of
test
in
g
in
d
iv
er
s
e
f
ir
e
s
ce
n
ar
i
o
s
,
p
ar
ticu
lar
ly
in
en
v
ir
o
n
m
en
ts
with
v
a
r
y
in
g
lig
h
t
an
d
s
m
o
k
e
in
ten
s
ities
.
T
h
is
r
aises
q
u
esti
o
n
s
ab
o
u
t
its
a
d
ap
tab
ilit
y
to
r
ea
l
-
wo
r
ld
,
d
iv
er
s
e
f
ir
e
s
itu
atio
n
s
.
In
r
esear
c
h
[
1
4
]
,
a
m
u
lti
-
task
l
ea
r
n
in
g
-
b
ased
f
o
r
est
f
ir
e
d
etec
tio
n
m
o
d
el
(
MTL
-
FF
DE
T
)
in
tr
o
d
u
ce
s
a
m
u
lti
-
ta
s
k
ap
p
r
o
ac
h
,
im
p
r
o
v
in
g
ac
cu
r
ac
y
an
d
r
ed
u
cin
g
f
alse
d
etec
ti
o
n
s
.
Desp
ite
th
ese
ad
v
an
ce
m
e
n
ts
,
th
e
m
o
d
el'
s
c
o
m
p
lex
ity
m
ig
h
t
h
in
d
er
its
d
ep
lo
y
m
en
t
in
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
e
n
v
ir
o
n
m
en
ts
,
an
d
its
p
e
r
f
o
r
m
an
ce
in
d
etec
tin
g
v
e
r
y
s
m
all
or
d
is
tan
t
f
ir
e
s
o
u
r
ce
s
h
as
not
b
ee
n
th
o
r
o
u
g
h
ly
e
v
alu
ated
.
In
r
esear
ch
[
1
5
]
,
co
m
m
o
n
wee
d
an
d
cr
o
p
d
ataset
(
CWC
)
-
YOL
Ov
5
s
u
s
ed
f
o
r
m
a
r
in
e
en
g
in
e
r
o
o
m
s
m
o
k
e
d
etec
tio
n
d
em
o
n
s
tr
ate
g
o
o
d
p
r
ec
is
io
n
an
d
r
ec
all
r
ates.
Ho
wev
er
,
th
e
m
o
d
el'
s
p
er
f
o
r
m
an
ce
in
en
v
ir
o
n
m
en
t
s
with
v
ar
y
in
g
s
m
o
k
e
d
e
n
s
ities
an
d
ty
p
es,
as
well
as
its
ef
f
ec
tiv
en
ess
in
d
etec
tin
g
s
m
o
k
e
at
th
e
ea
r
ly
s
tag
es
of
f
ir
e,
ar
e
ar
ea
s
th
at
n
ee
d
f
u
r
th
er
ex
p
lo
r
atio
n
.
I
m
p
r
o
v
ed
YO
L
Ov
7
[
1
6
]
is
ap
p
lied
f
o
r
co
al
m
in
e
d
etec
tio
n
th
at
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ad
v
Ap
p
l
Sci
I
SS
N:
2252
-
8
8
1
4
R
ea
l
-
time
s
mo
ke
a
n
d
fir
e
d
ete
ctio
n
u
s
in
g
yo
u
o
n
ly
lo
o
k
o
n
ce
v8
-
b
a
s
ed
a
d
va
n
ce
d
…
(
S
h
a
kila
R
a
h
ma
n
)
993
s
h
o
ws
ef
f
ec
tiv
en
ess
in
ch
allen
g
in
g
e
n
v
ir
o
n
m
en
ts
.
I
ts
ap
p
li
ca
tio
n
,
h
o
wev
e
r
,
is
lim
ited
to
s
p
ec
if
ic
s
ce
n
ar
io
s
(
co
al
m
in
es),
an
d
its
ad
ap
tab
i
lity
to
o
th
er
s
m
o
k
e
d
etec
tio
n
co
n
tex
ts
lik
e
f
o
r
est
f
ir
es
or
u
r
b
an
s
ettin
g
s
is
n
o
t
ad
d
r
ess
ed
.
In
r
esear
ch
[
1
7
]
,
a
s
m
ar
t
f
ir
e
d
etec
tio
n
s
y
s
tem
(
SF
DS
)
b
ased
on
YOL
Ov
8
is
ap
p
lied
f
o
r
s
m
ar
t
cities
th
at
in
teg
r
ate
with
I
o
T
an
d
clo
u
d
c
o
m
p
u
tin
g
,
o
f
f
e
r
in
g
h
ig
h
p
r
ec
is
io
n
.
Th
e
p
o
ten
ti
al
lim
itatio
n
is
th
e
d
ep
en
d
e
n
ce
on
r
o
b
u
s
t
n
etwo
r
k
in
f
r
astru
ctu
r
e
f
o
r
r
ea
l
-
tim
e
d
ata
p
r
o
ce
s
s
in
g
,
wh
ich
c
o
u
ld
be
a
ch
allen
g
e
in
less
d
ev
elo
p
e
d
ar
ea
s
.
In
r
esear
ch
[
1
8
]
,
E
n
h
an
ce
d
YOL
Ov
7
u
n
m
an
n
ed
ae
r
ial
v
eh
icles
(
UA
V
)
-
b
ased
f
o
r
est
f
ir
e
s
m
o
k
e
d
etec
tio
n
f
o
cu
s
es
on
a
er
ial
im
ag
er
y
.
T
h
e
m
o
d
el'
s
d
ep
en
d
en
c
y
on
th
e
q
u
ality
of
U
AV
im
ag
er
y
an
d
its
p
er
f
o
r
m
an
ce
in
d
if
f
er
e
n
t
wea
th
er
co
n
d
itio
n
s
ar
e
p
o
ten
tial
lim
itatio
n
s
.
In
r
esear
c
h
[
1
9
]
,
YOL
Ov
8
-
m
u
ltin
atio
n
al
co
r
p
o
r
atio
n
s
(
MN
C
)
is
u
s
ed
f
o
r
s
m
o
k
in
g
b
e
h
av
io
r
d
etec
tio
n
an
d
d
em
o
n
s
tr
ates
im
p
r
o
v
e
d
ac
cu
r
ac
y
.
T
h
e
alg
o
r
ith
m
'
s
s
p
ec
if
ic
f
o
cu
s
on
s
m
o
k
in
g
b
eh
av
io
r
lim
its
its
b
r
o
ad
e
r
ap
p
licab
ilit
y
to
g
en
er
al
f
ir
e
or
s
m
o
k
e
d
etec
tio
n
s
ce
n
ar
io
s
.
C
NN
-
b
ased
s
m
o
k
e
d
etec
tio
n
s
y
s
tem
[
2
0
]
in
co
r
p
o
r
ates
atten
tio
n
m
ec
h
an
is
m
s
an
d
im
p
r
o
v
ed
u
p
s
am
p
lin
g
.
T
h
e
s
y
s
tem
’
s
p
er
f
o
r
m
an
ce
in
d
etec
tin
g
s
m
o
k
e
f
r
o
m
d
if
f
er
en
t
s
o
u
r
ce
s
an
d
in
v
ar
io
u
s
lig
h
tin
g
c
o
n
d
itio
n
s
is
not
f
u
lly
ex
p
lo
r
e
d
.
In
r
esear
ch
[
2
1
]
,
a
ch
ar
ac
ter
is
tics
-
b
ased
f
ir
e
d
etec
tio
n
s
y
s
tem
ca
lled
Y
OL
O
-
v4
an
d
ViB
e
is
in
tr
o
d
u
ce
d
f
o
r
f
i
r
e
d
etec
tio
n
u
n
d
er
elec
tr
ic
f
ield
s
th
at
d
em
o
n
s
tr
ates
h
ig
h
ac
cu
r
ac
y
.
Ho
w
ev
er
,
th
e
alg
o
r
ith
m
’
s
p
er
f
o
r
m
an
ce
in
v
ar
ied
en
v
ir
o
n
m
en
tal
co
n
d
itio
n
s
b
ey
o
n
d
elec
tr
ic
f
ield
s
an
d
its
s
ca
l
ab
ilit
y
to
d
if
f
er
en
t
f
ir
e
d
etec
ti
o
n
s
ce
n
ar
io
s
r
em
ain
u
n
test
ed
.
In
r
esear
ch
[
2
2
]
,
YO
L
Ov
4
co
m
b
in
e
d
with
Mo
b
ileNetV3
is
ap
p
lied
f
o
r
f
ir
e
d
etec
tio
n
wh
ich
r
ed
u
ce
s
th
e
co
m
p
u
tatio
n
al
b
u
r
d
en
.
T
h
e
s
h
o
r
tco
m
in
g
lies
in
th
e
alg
o
r
ith
m
'
s
p
er
f
o
r
m
an
ce
on
less
p
o
wer
f
u
l
h
a
r
d
war
e
an
d
its
ad
a
p
tab
ilit
y
to
lar
g
er
,
more
c
o
m
p
lex
s
ce
n
es.
In
r
ese
ar
ch
[
1
2
]
,
YOL
Ov
6
-
b
ased
f
ir
e
d
etec
tio
n
in
s
m
ar
t
cities
d
em
o
n
s
tr
ates
YOL
Ov
6
’
s
ef
f
icien
cy
,
but
th
e
s
tu
d
y
lack
s
ex
ten
s
iv
e
ev
al
u
atio
n
in
d
if
f
e
r
en
t
u
r
b
an
s
ettin
g
s
,
esp
ec
ially
in
en
v
ir
o
n
m
en
ts
with
d
iv
er
s
e
lig
h
tin
g
an
d
o
cc
lu
s
io
n
ch
allen
g
es.
In
r
ese
ar
ch
er
s
[
2
3
]
–
[
2
5
]
,
im
p
r
o
v
e
d
YOL
Ov
5
is
p
r
o
p
o
s
ed
by
tar
g
etin
g
b
etter
ac
cu
r
a
cy
f
o
r
s
m
all
tar
g
ets.
A
p
o
ten
tial
lim
itatio
n
is
its
p
er
f
o
r
m
an
ce
in
d
etec
tin
g
f
ir
es
in
d
iv
er
s
e
ag
r
icu
ltu
r
al
e
n
v
ir
o
n
m
e
n
ts
an
d
v
ar
y
in
g
wea
th
er
co
n
d
itio
n
s
.
In
r
esear
ch
[
2
6
]
,
a
f
o
r
est
f
ir
e
class
if
icatio
n
an
d
d
etec
tio
n
m
o
d
el
(
FC
DM
)
is
p
r
o
p
o
s
ed
th
at
im
p
r
o
v
es
th
e
d
etec
tio
n
of
v
ar
io
u
s
f
o
r
est
f
ir
e
ty
p
es
b
u
t
m
ay
f
ac
e
ch
allen
g
es
in
d
if
f
er
en
tiatin
g
f
ir
es
in
d
e
n
s
e
f
o
r
est
ar
ea
s
an
d
u
n
d
er
v
ar
y
in
g
en
v
ir
o
n
m
e
n
tal
co
n
d
it
io
n
s
.
In
r
esear
c
h
[
2
7
]
,
th
e
y
c
o
m
p
ar
e
YOL
O
m
o
d
els
f
o
r
s
m
o
k
e
d
etec
tio
n
,
b
u
t
th
er
e
is
a
lack
of
in
-
d
ep
th
an
aly
s
is
of
each
m
o
d
el'
s
p
er
f
o
r
m
an
ce
in
r
ap
id
l
y
ch
an
g
in
g
en
v
ir
o
n
m
e
n
tal
co
n
d
itio
n
s
.
R
eST
iNet
[
2
8
]
s
h
o
wca
s
es
th
e
ap
p
licatio
n
of
T
i
n
y
-
YOL
O
-
b
ased
C
NN
ar
ch
itectu
r
e.
T
h
e
m
o
d
el'
s
ap
p
licab
ilit
y
to
f
ir
e
or
s
m
o
k
e
d
etec
tio
n
s
ce
n
ar
io
s
is
not
d
is
cu
s
s
ed
,
wh
ich
lim
its
its
r
elev
a
n
ce
to
th
is
s
p
ec
if
ic
r
ev
iew.
T
he
r
ev
iew
i
d
en
tifie
s
co
m
m
o
n
s
h
o
r
tco
m
in
g
s
ac
r
o
s
s
th
ese
s
tu
d
ies.
On
e
s
ig
n
if
ican
t
ch
alle
n
g
e
is
th
e
ad
ap
tab
ilit
y
of
th
ese
alg
o
r
ith
m
s
to
d
iv
er
s
e
an
d
d
y
n
am
ically
ch
an
g
in
g
en
v
ir
o
n
m
e
n
ts
,
s
u
ch
as
v
ar
y
in
g
wea
th
e
r
co
n
d
itio
n
s
an
d
lan
d
s
ca
p
es.
A
d
d
itio
n
ally
,
th
e
co
m
p
u
tatio
n
al
ef
f
icien
cy
of
more
co
m
p
lex
m
o
d
els
in
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
s
ettin
g
s
r
em
ain
s
a
co
n
ce
r
n
.
T
h
er
e
i
s
also
a
n
ee
d
f
o
r
f
u
r
th
e
r
r
esear
ch
to
i
m
p
r
o
v
e
th
e
m
o
d
els'
ef
f
ec
tiv
en
ess
in
d
etec
tin
g
d
i
f
f
er
en
t
t
y
p
es
an
d
s
tag
es
of
f
ir
es
an
d
s
m
o
k
e,
esp
ec
ially
in
ea
r
ly
d
etec
tio
n
s
ce
n
ar
io
s
.
Ov
er
all,
th
e
liter
atu
r
e
in
d
icat
es
a
s
tr
o
n
g
tr
en
d
to
war
d
s
m
o
r
e
s
o
p
h
is
ticated
,
ef
f
icien
t,
an
d
v
er
s
atile
YOL
O
-
b
ased
s
m
o
k
e
an
d
f
ir
e
d
etec
tio
n
s
y
s
tem
s
.
T
h
ese
s
y
s
t
em
s
s
h
o
w
p
r
o
m
is
e
in
n
o
t
o
n
l
y
en
h
an
ci
n
g
cu
r
r
en
t
f
ir
e
d
etec
tio
n
ca
p
ab
ilit
ies
.
T
h
ey
ar
e
also
in
teg
r
atin
g
with
em
er
g
in
g
tech
n
o
lo
g
ies
s
u
c
h
as
I
o
T
an
d
clo
u
d
co
m
p
u
tin
g
f
o
r
s
m
ar
ter
,
m
o
r
e
r
esp
o
n
s
iv
e
e
m
er
g
e
n
cy
m
a
n
ag
e
m
en
t sy
s
tem
s
.
3.
RE
S
E
ARCH
M
E
T
H
O
DO
L
O
G
Y
In
th
is
s
ec
tio
n
,
p
r
o
ce
d
u
r
es
f
o
r
ca
r
r
y
in
g
out
th
e
s
tu
d
y
ar
e
co
v
er
ed
.
A
f
r
a
m
ewo
r
k
f
o
r
th
e
d
et
ec
tio
n
an
d
lo
ca
lizatio
n
of
s
m
o
k
e
an
d
f
ir
e
zo
n
es
was
d
ev
elo
p
ed
u
s
in
g
th
e
YOL
Ov
8
m
o
d
e
l.
T
h
e
ar
ch
itectu
r
e
au
to
m
atica
lly
ex
tr
ac
te
d
m
a
n
y
s
m
o
k
e
an
d
f
ir
e
asp
ec
ts
f
r
o
m
an
in
p
u
t
im
a
g
e
in
o
r
d
er
to
id
e
n
tify
a
n
d
p
r
ec
is
ely
lo
ca
lize
th
e
p
o
ten
tial
lo
ca
tio
n
.
Fig
u
r
e
1
p
r
esen
ts
a
p
r
o
ce
s
s
f
lo
wch
ar
t
f
o
r
all
m
eth
o
d
s
u
s
ed
to
d
etec
t
s
m
o
k
e,
f
ir
e,
an
d
b
o
th
s
m
o
k
e
an
d
f
ir
e,
wh
ich
aid
s
in
a
d
ee
p
e
r
u
n
d
er
s
tan
d
i
n
g
of
th
e
p
ap
e
r
'
s
co
n
ce
p
t.
First
of
all,
co
llect
d
ata
f
r
o
m
d
if
f
er
en
t
s
o
u
r
ce
s
,
wh
ich
is
p
r
esen
ted
in
d
e
tail
in
s
u
b
s
ec
tio
n
3
.
1
.
T
h
en
th
e
co
llected
d
ata
is
p
r
e
-
p
r
o
ce
s
s
ed
wh
ich
is
ex
p
lain
ed
in
s
u
b
s
ec
tio
n
3
.
2
with
p
r
o
p
er
d
iag
r
am
s
an
d
ex
p
lan
atio
n
s
.
Af
ter
th
at,
th
e
p
r
e
-
p
r
o
ce
s
s
ed
d
ata
is
p
r
o
ce
s
s
ed
an
d
d
iv
id
ed
in
t
o
two
p
h
ases
.
In
p
h
ase
1,
th
e
tr
ain
i
n
g
p
ar
t
is
d
escr
ib
ed
.
Her
e,
a
l
ab
el
m
ap
is
cr
ea
ted
an
d
th
e
YOL
Ov
8
m
o
d
el
is
co
n
f
ig
u
r
e
d
.
T
h
en
th
e
d
ata
is
tr
ai
n
ed
with
th
e
p
r
o
p
o
s
ed
YOL
O
v
8
m
o
d
el.
Fin
ally
,
th
e
d
esire
d
tr
ain
in
g
g
r
a
p
h
s
f
r
o
m
th
e
tr
ain
in
g
p
r
o
ce
s
s
ar
e
ex
tr
ac
ted
.
In
p
h
ase
2,
th
e
test
in
g
p
ar
t
is
ex
p
lain
ed
.
Af
ter
co
m
p
letin
g
th
e
tr
ain
in
g
p
r
o
ce
s
s
,
th
e
d
ata
s
et
is
s
et
f
o
r
test
in
g
.
In
th
is
ca
s
e,
f
ir
s
t,
th
e
p
ar
am
eter
s
ar
e
ad
ju
s
ted
an
d
f
ee
d
d
ata
f
o
r
ac
c
u
r
ate
p
r
ed
ictio
n
s
.
T
h
en
,
we
m
u
s
t
go
th
r
o
u
g
h
s
o
m
e
s
tep
s
(
f
i
r
e
p
r
o
p
o
s
ed
r
etu
r
n
o
n
in
v
estme
n
t
(
R
o
I
)
p
o
llin
g
a
n
d
s
o
f
tm
a
x
)
.
Af
ter
th
at,
we
g
e
t
o
u
tp
u
t
with
a
b
o
u
n
d
in
g
box
wh
ile
ev
alu
atin
g
th
e
test
d
ata.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
8
1
4
I
n
t J Ad
v
Ap
p
l Sci
,
Vo
l.
1
3
,
No
.
4
,
Dec
em
b
er
2
0
2
4
:
991
-
1
0
0
3
994
Fig
u
r
e
1
.
W
o
r
k
f
lo
w
ar
ch
itectu
r
e
d
iag
r
am
f
o
r
th
e
wh
o
le
d
etec
tio
n
p
r
o
ce
s
s
3
.
1
.
I
ma
g
e
a
cquis
it
io
n
I
n
f
o
r
m
atio
n
co
llectio
n
is
a
v
ital
s
tep
th
at
in
clu
d
es
g
ath
er
in
g
d
ata
f
r
o
m
d
if
f
er
i
n
g
,
s
u
b
s
tan
tial
s
o
u
r
ce
s
.
W
ith
in
th
e
s
ettin
g
of
th
is
s
it
u
atio
n
,
in
th
is
p
o
n
d
er
,
th
e
in
f
o
r
m
atio
n
is
f
ir
s
t
co
llected
f
r
o
m
d
iv
er
s
e
o
n
lin
e
s
o
u
r
ce
s
,
wh
ich
ar
e
as
if
th
e
y
wer
e
p
ictu
r
es.
T
h
ese
o
n
lin
e
s
o
u
r
ce
s
can
be
web
s
ites
,
s
o
cial
m
ed
ia
p
latf
o
r
m
s
,
o
n
lin
e
d
atab
ases
,
or
a
n
y
o
t
h
er
s
ig
n
if
ican
t
o
n
lin
e
s
to
r
e.
To
s
u
p
p
lem
en
t
th
e
d
ataset,
o
th
er
p
h
o
to
s
f
r
o
m
th
e
web
ar
e
also
co
llected
in
t
h
is
d
ataset.
On
ce
th
e
v
ital
in
f
o
r
m
atio
n
an
d
p
ictu
r
es
ar
e
co
llected
f
r
o
m
o
n
lin
e
s
o
u
r
ce
s
,
a
cu
s
to
m
d
ataset
is
f
o
r
m
ed
.
T
h
i
s
co
m
p
ilatio
n
c
o
m
p
r
is
es
d
iv
er
s
e
im
ag
es
s
h
o
wca
s
in
g
d
a
y
a
n
d
n
ig
h
t
f
ir
es,
ae
r
ial
v
iews,
f
ix
ed
s
h
o
t
f
ir
es,
m
o
u
n
tain
,
s
u
r
f
ac
e,
tr
u
n
k
,
an
d
ca
n
o
p
y
f
ir
es,
as
well
as
n
atu
r
al
f
o
r
est
im
ag
es
with
d
is
tu
r
b
an
ce
s
.
T
ab
le
1
s
h
o
ws
th
e
m
ea
s
u
r
em
en
ts
of
th
e
p
ict
u
r
e
d
ataset,
wh
er
e
we
h
av
e
ca
teg
o
r
ized
p
ictu
r
e
d
atasets
in
to
th
r
ee
s
o
r
ts
(
f
ir
e
p
ic,
s
m
o
k
e
p
ic,
a
n
d
b
o
th
f
ir
e
an
d
s
m
o
k
e)
f
o
r
th
e
p
r
o
p
o
s
ed
f
r
a
m
ewo
r
k
.
T
ab
le
1.
Data
s
et
s
tatis
tic
s
C
a
t
e
g
o
r
y
Tr
a
i
n
Te
st
V
a
l
i
d
a
t
i
o
n
To
t
a
l
F
i
r
e
2
,
3
7
9
6
7
0
3
5
0
3
,
3
9
9
S
mo
k
e
6
6
5
1
9
0
95
9
5
0
F
i
r
e
a
n
d
S
m
o
k
e
9
4
6
2
7
0
1
3
5
1
,
3
5
1
To
t
a
l
3
,
9
9
0
1
,
1
3
0
5
8
0
5
,
7
0
0
A
d
ataset
s
er
v
es
as
one
of
th
e
m
o
s
t
p
r
o
f
itab
le
f
o
r
p
r
ep
ar
i
n
g
m
ac
h
i
n
e
lear
n
in
g
m
o
d
els,
c
o
n
d
u
ctin
g
in
v
esti
g
atio
n
s
,
or
p
e
r
f
o
r
m
in
g
ex
am
in
atio
n
s
.
T
h
e
c
u
s
to
m
d
at
aset
is
cu
r
ated
b
ased
on
th
e
n
e
ce
s
s
itie
s
an
d
tar
g
ets
of
th
e
ex
te
n
t,
g
u
ar
an
teein
g
th
at
it
in
clu
d
es
a
d
if
f
e
r
en
t
r
un
of
s
ig
n
if
ican
t
in
f
o
r
m
atio
n
to
s
u
p
p
o
r
t
th
e
p
lan
n
i
n
g
o
b
jectiv
es.
Fig
u
r
e
2
ap
p
ea
r
s
in
a
f
ew
s
am
p
le
p
ictu
r
e
d
atasets
.
In
T
ab
le
1,
it
is
ev
id
en
t
th
at
th
e
to
tal
n
u
m
b
er
of
f
ir
e
-
r
elate
d
im
ag
es
is
3
,
399.
T
h
ese
im
ag
es
ar
e
p
ar
titi
o
n
ed
f
o
r
d
if
f
er
e
n
t
p
u
r
p
o
s
es,
with
70%
allo
ca
ted
f
o
r
tr
ain
in
g
,
10%
f
o
r
v
alid
atio
n
,
an
d
th
e
r
em
ain
in
g
2
0
%
d
esig
n
ated
f
o
r
test
in
g
.
Similar
ly
,
we
u
s
ed
an
eq
u
al
p
er
ce
n
tag
e
of
im
ag
es
f
o
r
b
o
t
h
s
m
o
k
e
an
d
f
ir
e
.
T
h
e
n
u
m
b
e
r
of
p
ictu
r
es
is
950
an
d
1
,
3
5
1
,
r
esp
ec
tiv
ely
.
T
h
is
d
is
tr
ib
u
tio
n
en
s
u
r
es
a
co
m
p
r
e
h
en
s
iv
e
an
d
b
alan
ce
d
d
ataset
f
o
r
th
e
d
ev
elo
p
m
en
t
an
d
ev
alu
atio
n
of
m
o
d
els
r
elate
d
to
f
ir
e
d
etec
tio
n
or
an
a
ly
s
is
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ad
v
Ap
p
l
Sci
I
SS
N:
2252
-
8
8
1
4
R
ea
l
-
time
s
mo
ke
a
n
d
fir
e
d
ete
ctio
n
u
s
in
g
yo
u
o
n
ly
lo
o
k
o
n
ce
v8
-
b
a
s
ed
a
d
va
n
ce
d
…
(
S
h
a
kila
R
a
h
ma
n
)
995
Fig
u
r
e
2
.
Sam
p
les
of
r
aw
im
a
g
e
d
ata
3
.
2
.
Da
t
a
pre
-
pro
ce
s
s
ing
3
.
2
.
1
.
Aug
m
ent
a
t
io
n
Data
au
g
m
en
tatio
n
is
a
co
m
m
o
n
ly
em
p
lo
y
ed
tech
n
iq
u
e
aim
ed
at
en
h
an
cin
g
d
ataset
d
iv
er
s
ity
an
d
v
ar
iab
ilit
y
,
u
ltima
tely
en
h
an
cin
g
th
e
ex
ec
u
tio
n
an
d
f
lex
ib
ilit
y
of
ML
m
o
d
els.
We
h
av
e
c
o
llected
o
u
r
in
f
o
r
m
atio
n
f
r
o
m
d
i
f
f
er
en
t
s
o
u
r
ce
s
with
d
is
tin
ctiv
e
s
izes
an
d
r
eso
lu
tio
n
s
.
T
h
at'
s
wh
y
we
ap
p
ly
i
n
f
o
r
m
atio
n
en
lar
g
em
en
t
m
eth
o
d
s
in
a
f
ew
p
ictu
r
es
to
m
ak
e
th
e
in
f
o
r
m
atio
n
s
et
more
r
ea
s
o
n
ab
le.
We
h
av
e
co
n
n
ec
ted
d
is
tin
ctiv
e
in
f
o
r
m
atio
n
en
lar
g
em
en
t
p
r
o
ce
d
u
r
es
f
o
r
th
e
p
ictu
r
es,
as
tak
en
af
ter
.
In
ca
s
e
1,
we
ap
p
lied
cr
o
p
an
d
r
o
tatio
n
al
au
g
m
en
tatio
n
tech
n
iq
u
es.
Sin
ce
we
h
av
e
co
llecte
d
our
d
ata
s
et
f
r
o
m
d
if
f
er
en
t
s
o
u
r
ce
s
,
we
h
av
e
to
cr
o
p
s
o
m
e
d
ata
an
d
s
o
m
eti
m
es
ap
p
ly
r
o
tatio
n
to
o
u
r
d
ata
to
lab
el
th
e
s
h
ap
es.
T
h
i
s
m
eth
o
d
in
v
o
lv
es
au
g
m
en
tin
g
s
p
ec
if
ic
im
ag
es
by
r
an
d
o
m
ly
ap
p
ly
i
n
g
r
o
tatio
n
s
with
in
a
r
an
g
e
of
-
20
to
+2
0
d
eg
r
ee
s
.
T
h
is
au
g
m
en
tatio
n
tech
n
iq
u
e
aim
s
to
m
im
ic
r
ea
l
-
wo
r
ld
s
ce
n
ar
io
s
wh
er
e
o
b
jects
can
ap
p
ea
r
in
v
ar
io
u
s
o
r
ien
tatio
n
s
,
th
er
eb
y
m
ak
i
n
g
th
e
d
ataset
more
r
ep
r
esen
tativ
e.
E
x
p
o
s
in
g
th
e
m
o
d
el
to
r
o
tated
im
ag
es
d
u
r
in
g
tr
ain
in
g
im
p
r
o
v
es
its
ca
p
ac
ity
to
m
em
o
r
ize
an
d
g
en
er
alize
,
s
u
p
p
o
r
ti
n
g
s
u
p
er
io
r
ex
ec
u
tio
n
wh
en
e
x
p
er
ien
cin
g
p
iv
o
te
d
p
ictu
r
es
am
id
d
ed
u
ctio
n
or
t
esti
n
g
s
tag
es.
In
ca
s
e
2,
th
er
e
wer
e
s
o
m
e
im
ag
es
in
o
u
r
d
ata
s
et
th
at
lack
e
d
b
r
ig
h
tn
ess
,
wh
ich
is
wh
y
we
h
ad
to
m
ak
e
a
b
r
ig
h
tn
ess
ad
ju
s
tm
en
t.
We
u
s
ed
s
o
m
e
r
an
d
o
m
b
r
ig
h
tn
ess
ad
ju
s
tm
en
ts
in
th
e
d
ata
a
u
g
m
en
tatio
n
s
tep
.
T
h
e
p
er
ce
n
ta
g
e
of
th
e
a
d
ju
s
tm
en
t
was
b
etwe
en
+2
0
to
-
2
0
.
T
h
is
m
eth
o
d
p
r
o
v
e
d
in
s
tr
u
m
en
tal
in
im
p
r
o
v
in
g
th
e
v
is
ib
ilit
y
an
d
d
ef
in
itio
n
of
d
im
l
y
lit
im
ag
es,
r
en
d
er
in
g
th
em
more
c
o
n
d
u
civ
e
f
o
r
an
aly
s
is
a
n
d
f
ac
ilit
atin
g
th
eir
s
u
itab
ilit
y
f
o
r
m
o
d
el
tr
ain
in
g
p
u
r
p
o
s
es.
In
ca
s
e
3,
we
u
s
ed
th
e
f
lip
tech
n
iq
u
e.
Her
e,
we
s
elec
ted
s
o
m
e
r
an
d
o
m
d
ata
an
d
th
en
ju
s
t
f
lip
p
ed
it
in
a
d
if
f
er
en
t
p
o
s
itio
n
.
We
u
s
ed
th
is
tech
n
iq
u
e
to
en
h
a
n
ce
our
d
ata
s
et.
3
.
2
.
2
.
I
m
a
g
e
re
s
izing
T
h
e
p
r
o
ce
s
s
of
r
esizin
g
an
im
ag
e
in
v
o
lv
es
s
tan
d
ar
d
izin
g
th
e
d
im
en
s
io
n
s
of
all
im
ag
es
with
in
a
d
ataset,
en
s
u
r
in
g
u
n
if
o
r
m
ity
in
th
eir
s
h
ap
es.
In
th
is
co
n
tex
t,
th
e
en
tire
cu
s
to
m
d
ataset,
co
m
p
r
is
in
g
v
ar
io
u
s
im
ag
es,
u
n
d
er
wen
t
r
esizin
g
p
r
o
ce
d
u
r
es
to
ac
h
iev
e
d
im
e
n
s
io
n
s
of
6
4
0
×
6
4
0
p
ix
els.
R
esized
im
ag
es
ar
e
s
h
o
w
n
in
Fig
u
r
e
3.
T
h
en
,
th
e
r
esized
d
ata
is
a
s
et
of
an
n
o
tatio
n
s
.
Fig
u
r
e
3
.
Sam
p
les
of
r
esized
i
m
ag
e
d
ata
3
.
2
.
3
.
I
m
a
g
e
la
belin
g
or
a
nn
o
t
a
t
ing
First,
r
o
b
o
f
l
o
w
lab
elin
g
s
o
f
tw
ar
e
was
u
s
ed
to
c
h
o
o
s
e
th
e
n
e
ce
s
s
ar
y
s
ca
led
im
ag
e.
Fo
llo
win
g
th
at,
th
e
s
m
o
k
e
or
f
ir
e
r
eg
io
n
of
th
e
p
i
ctu
r
e
d
ata
was
h
ig
h
lig
h
ted
an
d
lab
eled
as
‘
s
m
o
k
e
’
an
d
‘
f
ir
e
’
u
s
in
g
th
e
o
p
tio
n
s
en
titl
ed
‘
b
o
u
n
d
in
g
b
o
x
to
o
l
’
an
d
‘
p
o
ly
g
o
n
to
o
l
’
.
T
h
e
a
n
n
o
tatio
n
g
ets
s
av
ed
au
to
m
atica
lly
,
an
d
th
e
r
e
is
an
ex
p
o
r
t
o
p
tio
n
th
at
c
r
ea
tes
th
e
T
XT
f
ile,
wh
ich
p
r
o
v
id
es
in
f
o
r
m
atio
n
in
-
d
ep
th
on
th
e
s
elec
ted
s
m
o
k
e
an
d
f
ir
e
r
eg
io
n
.
Fig
u
r
e
4
s
h
o
ws
a
f
l
o
wch
ar
t
f
o
r
lab
elin
g
th
e
r
esized
im
ag
e
d
ata
an
d
d
ata
p
r
e
-
p
r
o
ce
s
s
in
g
s
tep
s
.
Fig
u
r
e
4
(
a)
s
h
o
ws
th
e
f
lo
wch
a
r
t
f
o
r
lab
elin
g
th
e
im
a
g
es
.
T
h
e
ag
g
r
e
g
ate
of
th
e
in
f
o
r
m
atio
n
p
r
e
-
p
r
o
ce
s
s
in
g
s
tep
s
is
clea
r
ly
d
elin
ea
ted
in
Fig
u
r
e
4
(
b
)
an
d
lab
eled
p
o
r
tio
n
s
of
s
am
p
le
d
at
a
ar
e
s
h
o
wn
in
Fig
u
r
e
5.
Up
o
n
co
n
clu
d
in
g
th
ese
s
tep
s
,
th
e
p
r
ep
ar
ed
in
f
o
r
m
atio
n
ex
p
er
ien
ce
d
d
iv
is
io
n
,
ap
p
o
r
tio
n
in
g
70%
f
o
r
p
r
e
p
ar
atio
n
,
2
0
%
f
o
r
test
in
g
,
an
d
1
0
%
f
o
r
v
a
lid
atio
n
p
u
r
p
o
s
es
to
en
co
u
r
a
g
e
d
em
o
n
s
tr
atio
n
p
r
ep
ar
atio
n
.
T
h
e
n
th
e
p
r
e
-
p
r
o
ce
s
s
ed
d
ata
is
s
et
f
o
r
d
ata
p
r
o
ce
s
s
in
g
s
tep
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
8
1
4
I
n
t J Ad
v
Ap
p
l Sci
,
Vo
l.
1
3
,
No
.
4
,
Dec
em
b
er
2
0
2
4
:
991
-
1
0
0
3
996
(
a)
(
b
)
Fig
u
r
e
4
.
Flo
wch
ar
t
f
o
r
(
a)
lab
elin
g
th
e
r
esized
im
ag
e
d
ata
a
n
d
(
b
)
d
ata
p
r
e
-
p
r
o
ce
s
s
in
g
s
tep
s
Fig
u
r
e
5
.
Sam
p
les
of
lab
eled
i
m
ag
e
d
ata
3
.
3
.
Da
t
a
pro
ce
s
s
ing
In
th
is
d
ata
p
r
o
ce
s
s
in
g
co
n
tex
t,
o
n
ly
o
n
e
s
tep
is
co
n
s
id
er
ed
f
o
r
g
e
n
er
atin
g
a
T
XT
f
ile.
C
o
n
s
id
er
in
g
th
at,
a
f
ile
of
p
lain
tex
t
was
cr
ea
ted
f
o
r
ex
p
o
r
tin
g
d
ata
ea
s
ily
an
d
im
p
o
r
tin
g
in
a
s
tr
u
ctu
r
ed
m
an
n
er
.
T
h
e
n
th
e
p
r
o
ce
s
s
ed
d
ata
is
s
et
f
o
r
m
o
d
e
l
tr
ain
in
g
s
tep
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ad
v
Ap
p
l
Sci
I
SS
N:
2252
-
8
8
1
4
R
ea
l
-
time
s
mo
ke
a
n
d
fir
e
d
ete
ctio
n
u
s
in
g
yo
u
o
n
ly
lo
o
k
o
n
ce
v8
-
b
a
s
ed
a
d
va
n
ce
d
…
(
S
h
a
kila
R
a
h
ma
n
)
997
4.
RE
SU
L
T
S
AND
D
I
SCU
SS
I
O
N
Nu
m
er
o
u
s
p
iv
o
tal
tr
ain
in
g
c
o
n
f
ig
u
r
atio
n
s
a
n
d
h
y
p
e
r
p
ar
a
m
eter
s
ar
e
i
n
tr
icate
ly
wo
v
e
n
in
to
th
e
p
r
o
ce
s
s
of
tr
ain
in
g
th
e
YO
L
Ov
8
-
b
ased
m
o
d
el
f
o
r
d
ete
ctin
g
f
ir
e
an
d
s
m
o
k
e.
T
h
is
s
ec
tio
n
p
r
o
v
i
d
es
a
co
m
p
r
eh
e
n
s
iv
e
o
v
er
v
iew,
d
elv
in
g
in
to
th
e
h
y
p
er
p
ar
a
m
eter
s
m
eticu
lo
u
s
ly
u
tili
ze
d
th
r
o
u
g
h
o
u
t
th
e
tr
ain
in
g
r
eg
im
en
.
Du
r
in
g
t
h
e
tr
ain
in
g
p
h
ase,
an
ep
o
ch
c
o
u
n
t
of
1
0
0
is
s
et,
lev
er
ag
in
g
th
e
s
to
ch
asti
c
g
r
ad
ien
t
d
escen
t
(
SGD)
o
p
tim
izer
an
d
th
e
co
m
m
o
n
o
b
jects
in
co
n
tex
t
(
C
OC
O
)
p
r
e
-
tr
ai
n
ed
m
o
d
el.
To
c
u
r
t
ail
o
v
er
f
itti
n
g
an
d
s
tr
ea
m
lin
e
th
e
tr
ain
in
g
tr
ajec
to
r
y
,
an
ea
r
ly
s
to
p
p
in
g
m
ec
h
an
is
m
is
in
teg
r
ated
in
to
th
e
m
o
d
el.
T
h
is
in
ter
v
e
n
tio
n
h
alts
th
e
tr
ain
in
g
p
r
em
at
u
r
ely
wh
en
im
p
r
o
v
em
en
t
is
ab
s
en
t
o
v
er
th
e
last
50
e
p
o
ch
s
,
ef
f
e
ctiv
ely
p
r
ev
en
ti
n
g
u
n
n
ec
ess
ar
y
c
o
m
p
u
tatio
n
c
y
cl
es.
E
m
p
lo
y
i
n
g
an
ea
r
ly
s
to
p
p
i
n
g
tec
h
n
iq
u
e
with
a
p
atien
ce
v
alu
e
of
50
im
p
lies
th
at
if
no
d
is
ce
r
n
ib
le
en
h
an
ce
m
en
t
o
cc
u
r
s
f
o
r
50
co
n
s
e
cu
tiv
e
ep
o
ch
s
,
th
e
tr
ain
in
g
p
r
o
ce
s
s
will
ce
ase
au
to
m
atica
lly
.
It
i
s
wo
r
th
n
o
tin
g
th
at
f
u
tu
r
e
ex
p
e
r
im
en
ts
co
u
ld
e
x
p
lo
r
e
f
in
e
-
t
u
n
in
g
th
e
p
atien
ce
v
alu
e
to
ac
h
iev
e
ev
en
g
r
ea
ter
o
p
tim
izat
io
n
.
Mo
r
eo
v
er
,
m
eticu
lo
u
s
atten
tio
n
is
p
aid
to
o
th
er
p
iv
o
tal
p
ar
am
eter
s
.
Par
am
eter
s
s
u
ch
as
b
atch
s
ize,
lear
n
in
g
r
ate,
a
n
d
weig
h
t
d
ec
a
y
v
alu
es
ar
e
ca
lib
r
ated
at
1
6
,
0
.
0
1
,
an
d
0
.
0
0
1
,
r
esp
ec
tiv
ely
,
in
a
c
o
n
ce
r
te
d
ef
f
o
r
t
to
am
p
lify
m
o
d
el
o
p
tim
izatio
n
.
T
h
ese
n
u
a
n
ce
d
a
d
ju
s
tm
en
t
s
f
o
r
tify
th
e
m
o
d
el'
s
ab
ilit
y
to
d
is
ce
r
n
f
ir
e
a
n
d
s
m
o
k
e
in
s
tan
ce
s
with
h
eig
h
ten
ed
ac
cu
r
ac
y
an
d
r
o
b
u
s
tn
ess
,
war
r
an
tin
g
a
th
o
r
o
u
g
h
ex
p
lo
r
atio
n
of
p
o
ten
tial
r
ef
in
em
en
ts
in
f
o
r
th
co
m
in
g
e
x
p
er
im
en
ts
.
4.
1
.
M
o
del
ev
a
lua
t
i
o
n
T
h
e
ev
alu
atio
n
m
etr
ics
em
p
lo
y
ed
in
th
is
p
ap
er
to
ass
ess
th
e
m
o
d
el'
s
p
er
f
o
r
m
an
ce
in
clu
d
ed
p
r
ec
is
io
n
(
P),
r
ec
all
(
R
)
,
av
er
ag
e
p
r
ec
is
io
n
(
AP)
,
m
ea
n
av
er
a
g
e
p
r
ec
i
s
io
n
(
m
AP)
,
F1
s
co
r
e,
p
ar
am
eter
s
,
f
lo
atin
g
p
o
in
t
o
p
er
atio
n
s
(
FLOPs),
an
d
f
r
a
m
es
p
er
s
ec
o
n
d
(
FP
S).
AP
r
ep
r
esen
ts
th
e
ar
ea
u
n
d
er
th
e
p
r
ec
is
io
n
-
r
ec
all
(
PR
)
cu
r
v
e,
wh
ile
m
AP
s
ig
n
if
ies
th
e
av
er
ag
e
AP
ac
r
o
s
s
d
if
f
er
e
n
t
ca
teg
o
r
ies.
T
h
e
f
o
r
m
u
las
u
s
e
d
f
o
r
t
h
ese
m
etr
ics
ar
e
o
u
tlin
ed
as
(
1
)
-
(
3
)
.
Pre
cisi
o
n
=
+
(
1
)
R
ec
all
=
+
(
2
)
m
AP
=
∑
∫
(
)
−
(
3
)
T
r
u
e
p
o
s
itiv
e
(
T
P)
s
ig
n
if
ies
ac
cu
r
ate
class
if
icatio
n
of
a
s
am
p
le
as
p
o
s
itiv
e,
wh
ile
f
alse
p
o
s
itiv
e
(
FP
)
d
en
o
tes
in
c
o
r
r
ec
t
class
if
icatio
n
of
a
s
am
p
le
as
p
o
s
itiv
e.
Fals
e
n
eg
ativ
e
(
FN)
s
ig
n
if
ies
th
e
m
is
class
if
icatio
n
of
a
s
am
p
le
as
n
eg
ativ
e.
'
n
'
r
e
p
r
esen
ts
th
e
co
u
n
t
of
ca
teg
o
r
ies.
FLOPs
ar
e
a
m
ea
s
u
r
e
of
co
m
p
u
tatio
n
al
co
m
p
lex
ity
,
in
d
icatin
g
th
e
n
u
m
b
er
of
co
m
p
u
tatio
n
s
p
er
f
o
r
m
ed
by
a
m
o
d
el.
FPS
s
tan
d
s
f
o
r
f
r
am
es
p
er
s
ec
o
n
d
,
r
ep
r
esen
tin
g
th
e
r
ate
at
wh
ich
f
r
am
es
ar
e
tr
an
s
m
itted
.
T
h
e
YOL
Ov
8
m
o
d
el
was
ev
a
lu
ated
f
o
r
f
ir
e
a
n
d
s
m
o
k
e
d
et
ec
tio
n
with
in
a
d
ata
co
llectio
n
s
y
s
tem
.
T
h
e
ev
alu
atio
n
was
co
n
d
u
cte
d
u
s
in
g
Py
th
o
n
on
a
p
latf
o
r
m
eq
u
ip
p
e
d
with
C
UDA
1
2
.
0
an
d
NVI
DI
A
-
SMI
5
2
5
.
8
5
.
1
2
,
em
p
lo
y
in
g
a
GT
X
1650
GPU
with
4
GB
of
VR
AM
an
d
12
GB
of
R
AM
.
T
h
e
m
o
d
el
co
m
p
r
is
es
225
lay
er
s
an
d
1
1
1
3
8
3
0
9
p
ar
am
eter
s
,
s
h
o
wca
s
in
g
ef
f
icien
t
co
m
p
u
tatio
n
with
a
GFLO
P
v
alu
e
of
2
8
.
7
.
Var
io
u
s
m
etr
ics
wer
e
em
p
lo
y
e
d
to
g
a
u
g
e
th
e
m
o
d
el'
s
ef
f
icac
y
in
d
etec
tin
g
f
ir
e
an
d
s
m
o
k
e.
4.
2
.
Ana
ly
s
is
of
re
s
ults
Fig
u
r
e
6
s
h
o
ws
th
e
YOL
Ov
8
-
b
ased
r
ec
all
c
o
n
f
id
e
n
ce
c
u
r
v
e
an
d
tr
ai
n
in
g
g
r
a
p
h
with
8
5
ep
o
ch
s
.
T
h
e
YOL
Ov
8
-
b
ased
r
ec
all
co
n
f
id
e
n
ce
cu
r
v
e
is
s
h
o
wn
in
Fig
u
r
e
6
(
a)
f
o
r
in
d
iv
i
d
u
al
f
ir
e
an
d
s
m
o
k
e
class
es,
wh
er
e
all
class
v
alu
es
ar
e
ar
o
u
n
d
0
.
8
3
at
0
.
0
.
T
h
en
,
th
e
Y
OL
Ov
8
-
b
ased
tr
ai
n
in
g
g
r
ap
h
s
with
85
ep
o
ch
s
ar
e
d
e
p
icted
in
Fig
u
r
e
6
(
b
)
,
s
h
o
win
g
th
at
th
e
b
e
s
t
r
esu
lts
ar
e
o
b
tain
ed
at
tr
ain
in
g
s
tep
80
f
o
r
th
e
p
r
o
p
o
s
ed
s
ch
em
e.
T
h
u
s
,
th
e
d
ec
is
io
n
to
tr
ain
f
o
r
100
ep
o
ch
s
is
b
ased
on
th
e
o
b
s
er
v
ed
p
er
f
o
r
m
an
ce
.
T
h
e
ea
r
l
y
s
to
p
p
in
g
m
ec
h
a
n
is
m
was
not
ac
tiv
ated
b
ec
au
s
e
th
er
e
was
not
m
u
ch
d
if
f
e
r
en
ce
b
etwe
en
t
h
e
m
ax
im
u
m
ep
o
ch
an
d
th
e
b
est
r
esu
lt
s
tep
s
.
Fu
r
th
er
,
th
e
p
r
o
p
o
s
ed
YOL
Ov
8
ex
h
ib
its
a
r
ec
all
of
5
4
.
6
%,
a
p
r
ec
is
io
n
v
alu
e
of
6
8
.
3
%,
a
n
d
a
m
AP
of
5
7
.
3
%
wh
en
tr
ai
n
ed
with
85
ep
o
c
h
s
.
Fo
r
all
class
es,
th
e
r
ec
all
c
o
n
f
id
en
ce
th
r
esh
o
ld
v
alu
e
is
0
.
8
3
at
0
.
0
0
0
.
T
h
e
g
r
a
p
h
,
r
ec
all
v
s
.
c
o
n
f
id
e
n
ce
cu
r
v
e
m
er
g
ed
s
m
o
o
th
ly
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
8
1
4
I
n
t J Ad
v
Ap
p
l Sci
,
Vo
l.
1
3
,
No
.
4
,
Dec
em
b
er
2
0
2
4
:
991
-
1
0
0
3
998
(
a)
(
b
)
Fig
u
r
e
6
.
YOL
Ov
8
-
b
ased
(
a
)
r
ec
all
co
n
f
id
en
ce
cu
r
v
e
an
d
(
b
)
tr
ain
in
g
g
r
ap
h
with
8
5
e
p
o
ch
s
4
.
2
.
1
.
Co
nfusi
o
n
m
a
t
rix
A
co
n
f
u
s
io
n
m
atr
ix
is
a
f
u
n
d
am
en
tal
to
o
l
in
m
ac
h
in
e
lear
n
in
g
f
o
r
ev
alu
atin
g
th
e
p
er
f
o
r
m
an
ce
of
a
class
if
icatio
n
alg
o
r
ith
m
wh
ich
is
p
ar
ticu
lar
ly
u
s
ef
u
l
in
u
n
d
er
s
tan
d
in
g
th
e
s
tr
en
g
th
s
an
d
wea
k
n
ess
es
of
a
m
o
d
el,
ai
d
in
g
in
th
e
ass
ess
m
en
t
of
p
r
ec
is
io
n
,
r
ec
all,
a
n
d
o
v
er
all
ac
c
u
r
ac
y
.
T
h
e
co
n
f
u
s
io
n
m
at
r
ix
f
r
o
m
Fig
u
r
e
6
(
a
)
r
e
v
ea
ls
in
s
ig
h
tf
u
l
in
f
o
r
m
atio
n
a
b
o
u
t
th
e
m
o
d
el's
p
er
f
o
r
m
a
n
ce
.
It
d
ef
i
n
es
th
r
ee
class
es:
f
ir
e,
s
m
o
k
e,
an
d
b
ac
k
g
r
o
u
n
d
.
TP
i
n
d
icate
s
co
r
r
ec
tly
id
en
tifie
d
i
n
s
tan
ce
s
,
wh
ile
FP
r
ep
r
esen
t
s
m
is
class
if
icatio
n
s
.
Fo
r
f
ir
e,
th
e
m
o
d
el
c
o
r
r
ec
tly
id
en
tifie
s
70%
(
T
P)
as
f
ir
e
but
co
n
f
u
s
es
30%
(
FN)
wit
h
th
e
b
ac
k
g
r
o
u
n
d
.
Similar
ly
,
s
m
o
k
e
d
etec
tio
n
a
ch
iev
es
56%
(
T
P)
ac
cu
r
ac
y
,
but
m
is
class
if
ies
43%
(
FN)
as
b
ac
k
g
r
o
u
n
d
.
It
is
im
p
o
r
tan
t
to
n
o
te
th
at
th
e
v
al
u
es
with
in
th
e
m
atr
ix
r
an
g
e
f
r
o
m
0
.
0
1
to
0
.
7
0
,
o
f
f
er
in
g
f
u
r
th
er
in
s
ig
h
ts
in
to
s
p
ec
if
ic
co
n
f
id
en
ce
lev
els
f
o
r
each
class
if
icatio
n
.
T
h
is
an
a
ly
s
is
h
ig
h
lig
h
ts
ar
ea
s
f
o
r
im
p
r
o
v
em
en
t,
s
u
ch
as
p
o
ten
tially
r
ed
u
cin
g
f
alse
n
eg
ativ
es
to
en
h
an
ce
o
v
er
all
d
etec
tio
n
ac
cu
r
ac
y
.
An
aly
zin
g
th
e
tr
ain
in
g
c
u
r
v
e
in
Fig
u
r
e
6
(
b
)
,
we
can
s
ee
th
at
th
e
b
etter
r
esu
lt
is
ac
h
ie
v
ed
on
82
iter
atio
n
s
an
d
co
n
tin
u
es
up
to
85.
T
h
e
ev
alu
atio
n
of
th
e
p
r
o
p
o
s
ed
f
ir
e
-
r
elate
d
p
h
en
o
m
en
a
d
etec
tio
n
in
s
tan
ce
s
,
em
p
lo
y
in
g
th
e
YOL
Ov
8
m
o
d
el,
ex
ten
d
s
to
a
co
m
p
ar
ativ
e
an
aly
s
is
with
o
th
er
estab
lis
h
e
d
o
b
ject
d
etec
tio
n
m
o
d
els
s
u
ch
as
YOL
Ov
7
[
9
]
,
YOL
Ov
5
[
2
5
]
,
Mo
b
ilen
et
-
v2
[
2
9
]
,
a
n
d
R
esNet
-
32
[
3
0
]
.
T
h
e
p
e
r
f
o
r
m
a
n
ce
m
etr
ics
ar
e
m
eticu
lo
u
s
ly
ass
ess
ed
ac
r
o
s
s
all
cla
s
s
es
with
v
ar
y
in
g
iter
atio
n
s
tep
s
,
as
d
etai
led
in
T
ab
le
2.
T
h
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ad
v
Ap
p
l
Sci
I
SS
N:
2252
-
8
8
1
4
R
ea
l
-
time
s
mo
ke
a
n
d
fir
e
d
ete
ctio
n
u
s
in
g
yo
u
o
n
ly
lo
o
k
o
n
ce
v8
-
b
a
s
ed
a
d
va
n
ce
d
…
(
S
h
a
kila
R
a
h
ma
n
)
999
m
AP@
0
.
5
is
tr
ac
k
ed
d
u
r
in
g
t
h
e
tr
ain
in
g
p
h
ase,
in
d
icatin
g
t
h
e
m
o
d
el'
s
ab
ilit
y
to
lear
n
on
th
e
v
alid
atio
n
s
et,
with
a
h
ig
h
er
v
alu
e
d
e
n
o
tin
g
s
u
p
er
io
r
lear
n
in
g
.
Ad
d
itio
n
ally
,
th
e
F1
s
co
r
e
,
ca
lcu
lated
th
r
o
u
g
h
a
p
r
ec
is
e
f
o
r
m
u
la,
r
ev
e
als
th
at
th
e
YOL
Ov
8
-
b
ased
m
o
d
el
c
o
n
s
is
ten
tly
o
u
tp
e
r
f
o
r
m
s
o
th
er
s
,
b
o
asti
n
g
an
F1
s
co
r
e
an
d
m
AP@
0
.
5
v
alu
es
of
60
a
n
d
5
7
.
3
%,
r
esp
ec
tiv
ely
.
A
d
etailed
e
x
a
m
in
atio
n
of
m
o
d
el
co
m
p
lex
ities
in
T
ab
le
2
f
u
r
th
e
r
h
ig
h
lig
h
ts
YOL
Ov
8
'
s
s
u
p
er
io
r
ity
,
p
ar
ticu
lar
ly
o
v
er
YOL
Ov
7
,
wh
ic
h
ex
h
ib
its
th
e
h
ig
h
est
n
u
m
b
e
r
of
tr
ai
n
ab
le
p
a
r
am
eter
s
,
p
o
ten
tially
co
m
p
r
o
m
is
in
g
g
en
er
aliza
tio
n
ca
p
ac
ity
.
Fu
r
th
er
m
o
r
e
,
Fig
u
r
e
7
o
f
f
er
s
a
c
o
m
p
r
e
h
en
s
iv
e
v
iew
of
t
h
e
co
n
f
u
s
io
n
m
atr
ix
d
iag
r
am
at
th
e
8
5
th
iter
atio
n
,
s
h
o
wca
s
in
g
th
e
m
o
d
el's
p
er
f
o
r
m
an
ce
lim
itatio
n
s
with
an
ac
cu
r
ac
y
of
9
0
.
4
5
%.
T
h
is
m
atr
ix
v
is
u
ally
d
is
s
ec
ts
p
r
ed
icted
an
d
ac
tu
al
class
es,
with
n
o
tab
le
ac
cu
r
ac
y
p
ea
k
s
o
b
s
er
v
ed
f
o
r
th
e
f
ir
e
class
(
7
0
%)
an
d
a
co
r
r
esp
o
n
d
in
g
d
ip
f
o
r
th
e
s
moke
class
(
5
6
%).
T
h
e
YO
L
Ov
8
-
b
ased
m
o
d
el,
h
o
wev
er
,
r
ev
ea
ls
p
r
o
m
is
in
g
p
o
ten
tial
f
o
r
r
ea
l
-
tim
e
f
ir
e
a
n
d
s
m
o
k
e
d
etec
tio
n
in
d
iv
er
s
e
s
ce
n
ar
io
s
,
em
p
h
asizin
g
its
ef
f
ica
cy
in
s
till
im
ag
es,
v
id
eo
s
,
an
d
ca
m
er
a
f
ee
d
s
th
r
o
u
g
h
r
i
g
o
r
o
u
s
test
in
g
an
d
a
n
aly
s
is
.
Fig
u
r
e
7
.
C
o
n
f
u
s
io
n
m
atr
i
x
d
i
ag
r
am
f
o
r
85
ep
o
c
h
s
T
ab
le
2.
T
esti
n
g
e
x
ec
u
tio
n
of
YOL
Ov
8
with
YOL
Ov
7
,
YO
L
Ov
5
,
Mo
b
ileNet
V2
a
n
d
R
esNet
-
32
M
o
d
e
l
Ep
o
c
h
C
l
a
s
s
Tr
a
i
n
a
b
l
e
P
a
r
a
met
e
r
s
F1
S
c
o
r
e
mA
P
@
0
.
5
P
r
o
p
o
se
d
Y
O
LO
v8
50
A
l
l
1
1
.
1
3
M
0
.
5
8
7
0
.
5
5
5
P
r
o
p
o
se
d
Y
O
LO
v8
85
A
l
l
1
1
.
1
3
M
0
.
6
0
7
0
.
5
7
3
Y
O
LO
v7
[
9
]
50
A
l
l
3
7
.
2
M
0
.
4
3
0
0
.
3
7
2
Y
O
LO
v7
[
9
]
85
A
l
l
3
7
.
2
M
0
.
4
5
8
0
.
3
9
1
Y
O
LO
v5
[
2
5
]
50
A
l
l
7
.
2
M
0
.
4
8
7
0
.
4
2
6
Y
O
LO
v5
[
2
5
]
85
A
l
l
7
.
2
M
0
.
4
8
7
0
.
4
2
6
M
o
b
i
l
e
N
et
-
v2
[
2
9
]
50
A
l
l
3
.
4
M
0
.
3
5
9
0
.
3
2
4
M
o
b
i
l
e
N
et
-
v2
[
2
9
]
85
A
l
l
3
.
4
M
0
.
3
6
5
0
.
3
3
5
R
e
sN
e
t
-
32
[
3
0
]
50
A
l
l
0
.
4
7
M
0
.
2
6
7
0
.
2
4
5
R
e
sN
e
t
-
32
[
3
0
]
85
A
l
l
0
.
4
7
M
0
.
2
7
6
0
.
2
5
5
4.
3
.
Vis
ua
liza
t
io
n
T
h
e
m
o
d
el'
s
tr
ain
in
g
s
p
an
n
ed
100
ep
o
ch
s
,
s
ig
n
if
y
in
g
a
co
m
p
lete
iter
atio
n
th
r
o
u
g
h
th
e
e
n
tire
tr
ain
in
g
d
ataset
in
each
e
p
o
ch
.
T
h
e
tr
ain
in
g
p
r
o
ce
s
s
in
v
o
l
v
ed
iter
at
iv
e
u
p
d
ates
to
th
e
m
o
d
el'
s
p
ar
am
eter
s
b
ased
on
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
8
1
4
I
n
t J Ad
v
Ap
p
l Sci
,
Vo
l.
1
3
,
No
.
4
,
Dec
em
b
er
2
0
2
4
:
991
-
1
0
0
3
1000
ca
lcu
lated
lo
s
s
es
an
d
g
r
ad
ien
ts
.
T
h
e
d
ec
is
io
n
to
co
n
cl
u
d
e
th
e
tr
ain
in
g
at
85
ep
o
c
h
s
alig
n
s
with
th
e
co
n
f
ig
u
r
atio
n
s
ettin
g
s
,
with
o
p
tim
al
r
esu
lts
em
er
g
in
g
ar
o
u
n
d
th
e
80
th
s
tep
.
T
h
e
n
o
n
-
en
g
a
g
em
en
t
of
th
e
ea
r
l
y
s
to
p
p
in
g
m
ec
h
a
n
is
m
is
ju
s
tifi
ed
,
as
its
ac
t
iv
atio
n
cr
iter
ia
wer
en
'
t
m
et
w
ith
in
th
e
s
p
ec
if
ied
85
ep
o
ch
s
af
ter
ac
h
iev
in
g
t
h
e
b
est
r
esu
lt.
C
o
n
s
eq
u
en
tly
,
th
e
tr
ain
in
g
p
r
o
ce
s
s
h
alted
at
85
iter
atio
n
s
,
co
n
s
u
m
in
g
ap
p
r
o
x
im
ately
3
h
o
u
r
s
.
T
h
is
d
u
r
atio
n
,
s
u
b
ject
to
v
ar
iatio
n
s
b
ased
on
c
o
m
p
u
tatio
n
al
r
eso
u
r
ce
s
an
d
h
a
r
d
wa
r
e,
u
n
d
e
r
s
co
r
es
th
e
r
eso
u
r
ce
-
in
ten
s
iv
e
n
atu
r
e
of
th
e
tr
ain
in
g
.
T
h
e
p
e
r
f
o
r
m
an
ce
of
th
e
o
b
ject
d
etec
tio
n
m
o
d
el
p
ea
k
ed
at
85
ep
o
ch
s
an
d
th
en
s
tar
ted
to
d
ec
lin
e.
T
h
is
is
a
co
m
m
o
n
p
h
en
o
m
e
n
o
n
ca
lled
o
v
er
f
itti
n
g
,
wh
ich
o
cc
u
r
s
wh
en
th
e
m
o
d
el
m
em
o
r
izes
th
e
tr
ain
in
g
d
ata
to
o
well
an
d
lo
s
es
its
ab
ilit
y
to
g
en
er
alize
to
n
ew
d
ata
.
T
h
e
in
d
iv
id
u
al
d
etec
tio
n
ac
c
u
r
ac
ies
f
o
r
YOL
Ov
8
,
YOL
Ov
7
,
an
d
YOL
Ov
5
in
85
tr
ain
in
g
s
tep
s
ar
e
s
h
o
wn
in
Fig
u
r
e
8
,
co
n
s
id
er
in
g
all
th
e
class
e
s
.
In
th
e
ca
s
e
of
th
e
f
ir
s
t
r
o
w,
th
e
d
etec
tio
n
ac
cu
r
ac
ies
ar
e
90
an
d
51%
f
o
r
th
e
YOL
Ov
8
,
an
d
Y
OL
Ov
5
m
o
d
els,
but
we
s
ee
th
at
YOL
Ov
7
,
R
esNet
-
32
,
an
d
Mo
b
ileNet
-
v2
ca
n
n
o
t
d
etec
t
an
y
f
ir
e
p
o
r
tio
n
.
Similar
ly
,
f
o
r
th
e
YOL
Ov
8
m
o
d
el,
we
s
ee
in
th
e
s
ec
o
n
d
r
o
w’
s
f
ir
s
t
p
ictu
r
e
th
at
th
e
d
etec
tio
n
ac
cu
r
ac
y
f
o
r
s
m
o
k
e
is
64%
an
d
f
o
r
f
ir
e
is
6
6
%,
We
also
s
ee
th
at
f
o
r
YOL
Ov
7
an
d
R
esNet
-
32
th
er
e
was
no
d
etec
tio
n
f
o
r
th
is
p
ict
u
r
e,
b
u
t
f
o
r
th
e
s
am
e
p
ictu
r
e
,
th
e
YOL
Ov
5
a
n
d
Mo
b
ileNet
-
v2
m
o
d
els
can
d
etec
t
o
n
ly
th
e
f
ir
e
p
o
r
tio
n
,
wh
ich
is
53
an
d
5
2
%,
r
esp
ec
tiv
ely
.
Fo
r
th
e
YOL
Ov
8
m
o
d
el
a
n
d
M
o
b
ile
N
et
-
v2,
we
s
ee
in
th
e
t
h
ir
d
r
o
w’
s
f
i
r
s
t
an
d
last
p
ictu
r
es
th
at
th
e
d
etec
tio
n
ac
c
u
r
ac
y
f
o
r
s
m
o
k
e
is
75
an
d
5
9
%,
r
esp
ec
tiv
ely
,
but
f
o
r
th
e
s
am
e
p
ictu
r
e,
th
e
YOL
Ov
7
,
YOL
Ov
5
,
an
d
R
eNe
t
-
32
m
o
d
els
ca
n
n
o
t
d
etec
t
an
y
s
m
o
k
e
r
eg
io
n
.
L
astl
y
,
h
er
e
we
also
s
ee
th
e
d
etec
tio
n
ac
cu
r
ac
ies
ar
e
79,
60,
4
7
,
27,
an
d
4
9
%
f
o
r
YOL
Ov
8
,
YOL
Ov
7
,
YOL
Ov
5
,
R
esNet
-
32,
an
d
Mo
b
ileNet
-
v2
m
o
d
els
f
o
r
th
e
f
o
u
r
th
r
o
w.
T
h
u
s
,
f
o
r
all
class
es
in
m
o
s
t
ca
s
es,
th
e
d
etec
tio
n
ac
cu
r
ac
y
is
h
i
g
h
er
in
th
e
p
r
o
p
o
s
ed
YOL
Ov
8
-
b
ased
s
m
o
k
e
an
d
f
ir
e
d
etec
tio
n
mo
d
el
c
o
m
p
ar
ed
to
th
e
o
th
er
YOL
Ov
7
an
d
YOL
Ov
5
m
o
d
el
s
with
85
ep
o
ch
s
.
Am
o
n
g
all
t
h
e
m
o
d
els
th
at
ar
e
b
ein
g
tr
ai
n
ed
with
o
u
r
cu
s
to
m
d
ataset,
th
e
YOL
Ov
8
g
av
e
t
h
e
b
est
p
er
f
o
r
m
an
ce
m
etr
ics
al
o
n
g
with
d
etec
tio
n
ac
cu
r
ac
y
in
all
asp
ec
ts
.
T
h
e
m
o
d
el
is
tr
ain
ed
with
d
if
f
er
en
t
ep
o
ch
s
;
am
o
n
g
th
em
,
th
e
tr
a
in
in
g
cu
r
v
e
of
th
e
YOL
Ov
8
50
E
p
o
ch
is
s
lig
h
tly
b
etter
,
but
in
ter
m
s
of
all
o
t
h
er
ev
alu
atio
n
m
etr
ics,
th
e
Y
OL
Ov
8
85
E
p
o
ch
m
o
d
el
p
er
f
o
r
m
s
b
est,
an
d
th
e
p
er
f
o
r
m
an
ce
s
tar
ted
to
d
ec
r
ea
s
e
in
th
e
1
0
0
an
d
1
5
0
ep
o
ch
s
f
o
r
o
v
er
f
itti
n
g
.
Pro
p
o
s
ed
YOL
Ov
8
YOL
Ov
7
[
9
]
YOL
Ov
5
[
2
5
]
R
es
N
et
-
32
[
3
0
]
Mo
b
ile
N
et
-
v2
[
2
9
]
Fig
u
r
e
8
.
Sam
p
le
d
etec
ted
im
a
g
es
5.
CO
NCLU
SI
O
N
T
h
is
s
tu
d
y
p
r
esen
ts
a
s
ig
n
if
i
ca
n
t
leap
in
th
e
f
ield
of
s
m
o
k
e
an
d
f
ir
e
d
etec
tio
n
by
lev
e
r
ag
in
g
th
e
ca
p
ab
ilit
ies
of
th
e
YOL
Ov
8
d
ee
p
lear
n
in
g
m
o
d
el.
Dem
o
n
s
tr
atin
g
a
m
ar
k
ed
im
p
r
o
v
e
m
e
n
t
o
v
er
tr
a
d
itio
n
al
d
etec
tio
n
m
eth
o
d
s
an
d
p
r
e
v
i
o
u
s
iter
atio
n
s
of
th
e
YOL
O
m
o
d
el,
YOL
Ov
8
d
is
tin
g
u
is
h
ed
its
elf
th
r
o
u
g
h
Evaluation Warning : The document was created with Spire.PDF for Python.