I
nte
rna
t
io
na
l J
o
urna
l o
f
E
lect
rica
l a
nd
Co
m
pu
t
er
E
ng
ineering
(
I
J
E
CE
)
Vo
l.
15
,
No
.
5
,
Octo
b
er
20
25
,
p
p
.
5
0
3
1
~
5
0
4
4
I
SS
N:
2088
-
8
7
0
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijece.
v
15
i
5
.
pp
5
0
3
1
-
5
0
4
4
5031
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ec
e.
ia
esco
r
e.
co
m
Efficien
t
fall det
e
ction using
light
w
eig
ht
ne
twork
to
enha
nce
sm
a
rt
inte
rnet
of
things
P
inro
lin
v
ic
D.
K
.
M
a
nem
bu
,
J
a
ne
I
v
o
nn
e
L
it
o
uw
,
F
eisy
Dia
ne
K
a
m
bey
,
Abdu
l H
a
ris J
un
us
O
nto
wirj
o
,
Vec
k
y
C.
P
o
eko
el,
M
uh
a
m
a
d Dw
is
na
nto
P
utr
o
D
e
p
a
r
t
me
n
t
o
f
El
e
c
t
r
i
c
a
l
En
g
i
n
e
e
r
i
n
g
,
F
a
c
u
l
t
y
o
f
En
g
i
n
e
e
r
i
n
g
,
S
a
m
R
a
t
u
l
a
n
g
i
U
n
i
v
e
r
si
t
y
,
M
a
n
a
d
o
,
I
n
d
o
n
e
si
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Au
g
1
5
,
2
0
2
4
R
ev
is
ed
Ma
y
2
6
,
2
0
2
5
Acc
ep
ted
J
u
n
3
0
,
2
0
2
5
F
a
ll
d
e
tec
ti
o
n
a
u
t
o
m
a
ti
c
a
ll
y
re
c
o
g
n
ize
s
h
u
m
a
n
fa
ll
s,
m
a
in
l
y
t
o
m
o
n
it
o
r
a
n
d
p
re
v
e
n
t
se
v
e
re
i
n
ju
r
y
a
n
d
p
o
ten
t
ial
fa
talit
ies
.
It
c
a
n
b
e
d
e
v
e
l
o
p
e
d
b
y
a
p
p
l
y
in
g
d
e
e
p
lea
rn
i
n
g
m
e
th
o
d
s
t
o
re
c
o
g
n
i
z
e
h
u
m
a
n
s
u
b
jec
ts
d
u
ri
n
g
fa
ll
in
c
i
d
e
n
ts
a
n
d
imp
lem
e
n
ted
i
n
t
h
e
in
tern
e
t
o
f
t
h
in
g
s
(Io
T)
t
o
m
o
n
it
o
r
p
a
ti
e
n
t
a
n
d
e
l
d
e
rl
y
in
d
i
v
id
u
a
ls’
a
c
ti
v
it
y
.
T
h
e
d
e
v
e
lo
p
m
e
n
t
o
f
o
b
jec
t
d
e
tec
ti
o
n
p
re
se
n
ts
y
o
u
o
n
ly
lo
o
k
o
n
c
e
v
8
(YO
LOv
8
)
a
s
a
n
in
f
lu
e
n
ti
a
l
n
e
two
r
k
,
b
u
t
it
s
e
fficie
n
c
y
n
e
e
d
s
t
o
b
e
imp
ro
v
e
d
.
A
m
o
d
if
ied
YO
LO
v
8
a
rc
h
i
tec
tu
re
is
p
r
o
p
o
se
d
t
o
i
n
tro
d
u
c
e
a
n
o
v
e
l
l
ig
h
twe
ig
h
t
n
e
two
r
k
v
e
rsi
o
n
c
a
ll
e
d
YO
LOv
8
-
H
y
p
e
rn
a
n
o
(Y
OLOv
8
h
)
th
a
t
re
c
o
g
n
ize
s
fa
ll
e
v
e
n
ts.
T
h
e
b
a
c
k
b
o
n
e
i
n
c
o
r
p
o
ra
tes
a
c
o
m
b
in
e
d
sp
a
ti
a
l
a
n
d
c
h
a
n
n
e
l
a
tt
e
n
t
io
n
m
o
d
u
le,
wh
ich
e
n
h
a
n
c
e
s
fo
c
u
s
o
n
h
u
m
a
n
su
b
jec
ts
b
y
c
o
n
c
e
n
tratin
g
o
n
m
o
v
e
m
e
n
t
p
a
tt
e
rn
s
to
d
e
tec
t
fa
ll
s
m
o
re
a
c
c
u
ra
tely
.
Th
is
wo
rk
a
lso
o
ffe
rs
a
c
o
n
se
c
u
ti
v
e
se
lec
ti
v
e
e
n
h
a
n
c
e
m
e
n
t
(CS
E)
m
o
d
u
le
t
o
imp
ro
v
e
e
fficie
n
c
y
a
n
d
e
ffe
c
ti
v
e
n
e
ss
in
fe
a
tu
re
e
x
trac
ti
o
n
wh
il
e
re
d
u
c
in
g
c
o
m
p
u
tati
o
n
a
l
c
o
sts.
Th
e
n
e
c
k
str
u
c
tu
re
is
m
o
d
ifi
e
d
b
y
a
d
d
i
n
g
a
li
g
h
twe
ig
h
t
b
o
tt
le
n
e
c
k
n
e
two
rk
.
Th
e
p
ro
p
o
s
e
d
n
e
tw
o
rk
re
c
o
n
str
u
c
ts
fe
a
tu
re
m
a
p
s
i
n
d
e
p
th
,
p
a
y
i
n
g
m
o
re
a
tt
e
n
ti
o
n
to
a
c
c
u
ra
te
h
u
m
a
n
m
o
v
e
m
e
n
t
p
a
t
tern
s
a
n
d
e
n
h
a
n
c
in
g
e
fficie
n
c
y
a
n
d
e
ffe
c
ti
v
e
n
e
ss
in
fe
a
tu
re
e
x
trac
ti
o
n
.
Ex
p
e
rime
n
tal
re
su
lt
s
o
f
YO
LOv
8
h
wit
h
th
e
l
ig
h
t
b
o
tt
len
e
c
k
a
n
d
c
o
n
se
c
u
ti
v
e
se
lec
ti
v
e
e
n
h
a
n
c
e
m
e
n
t
m
o
d
u
les
sh
o
w
g
i
g
a
flo
a
ti
n
g
-
p
o
i
n
t
o
p
e
ra
ti
o
n
s
p
e
r
se
c
o
n
d
s
(
G
F
LOP
S
)
o
f
5
.
6
a
n
d
1
,
1
9
4
,
4
4
0
p
a
ra
m
e
ters
.
Th
e
m
o
d
e
l
p
e
rfo
rm
a
n
c
e
is
c
a
lcu
late
d
in
m
e
a
n
a
v
e
ra
g
e
p
re
c
is
io
n
,
a
c
h
iev
in
g
0
.
6
0
3
a
n
d
0
.
7
3
2
o
n
t
h
e
Le
2
i
a
n
d
F
a
ll
e
n
d
a
tas
e
ts,
re
sp
e
c
ti
v
e
ly
.
T
h
e
se
re
su
lt
s
d
e
m
o
n
stra
te
th
a
t
t
h
e
o
p
ti
m
ize
d
n
e
two
r
k
im
p
ro
v
e
s
a
c
c
u
ra
c
y
p
e
rfo
rm
a
n
c
e
wh
il
e
m
a
in
tain
i
n
g
li
g
h
twe
i
g
h
t
c
o
m
p
u
ti
n
g
re
q
u
irem
e
n
ts
th
a
t
c
a
n
r
u
n
sm
o
o
t
h
ly
o
n
Io
T
d
e
v
ice
s
,
a
c
h
iev
in
g
c
o
m
p
a
ra
b
le
sp
e
e
d
a
n
d
e
fficie
n
c
y
su
it
a
b
le
fo
r
o
p
e
ra
ti
o
n
o
n
lo
w
-
c
o
st co
m
p
u
t
in
g
d
e
v
ice
s.
K
ey
w
o
r
d
s
:
E
d
g
e
d
e
v
ice
E
f
f
icien
t c
o
m
p
u
tatio
n
Fall d
etec
tio
n
L
ig
h
tweig
h
t m
o
d
u
le
Sm
ar
t
in
ter
n
et
o
f
t
h
in
g
s
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
:
Pin
r
o
lin
v
ic
D.
K.
Ma
n
em
b
u
Dep
ar
tm
en
t o
f
E
lectr
ical
E
n
g
i
n
ee
r
in
g
,
Facu
lty
o
f
E
n
g
in
ee
r
in
g
,
Sam
R
atu
lan
g
i U
n
iv
er
s
ity
B
ah
u
C
am
p
u
s
,
Ma
n
ad
o
9
5
1
1
5
,
I
n
d
o
n
esia
E
m
ail: p
m
an
em
b
u
@
u
n
s
r
at.
ac
.
id
1.
I
NT
RO
D
UCT
I
O
N
Vis
io
n
-
b
ased
f
all
d
etec
tio
n
is
a
r
ewa
r
d
in
g
task
th
at
a
n
aly
z
es
an
d
p
r
ed
icts
f
all
ev
e
n
ts
,
wh
ich
ca
n
ca
u
s
e
s
er
io
u
s
in
ju
r
ies s
u
ch
as
d
is
ab
ilit
y
,
p
ar
aly
s
is
,
o
r
d
ea
th
.
I
t is esp
ec
ial
ly
cr
u
cial
f
o
r
eld
er
ly
p
atien
ts
,
as f
alls
ca
n
b
e
f
atal
[
1
]
.
T
h
e
r
is
k
o
f
f
a
lls
in
cr
ea
s
es
d
u
e
to
p
h
y
s
ical
an
d
m
en
tal
d
ec
lin
e
in
f
lu
en
ce
d
b
y
ag
in
g
.
I
t
m
a
k
es
f
all
d
etec
tio
n
s
y
s
tem
s
ess
en
tial
in
h
ea
lth
ca
r
e
to
im
p
r
o
v
e
its
q
u
ality
.
T
h
e
r
ef
o
r
e
,
f
all
d
ete
ctio
n
s
y
s
tem
s
ar
e
a
p
r
o
m
is
in
g
s
o
lu
tio
n
f
o
r
r
e
d
u
cin
g
th
e
r
is
k
o
f
f
alls
an
d
t
h
eir
h
ea
lth
co
n
s
eq
u
en
ce
s
[
2
]
.
Me
an
wh
ile,
it
is
im
p
o
s
s
ib
le
to
p
r
ev
en
t
f
alls
,
b
u
t
p
h
y
s
ical
ex
er
cise
an
d
tech
n
o
lo
g
ical
s
o
lu
tio
n
s
ca
n
co
m
p
letely
h
elp
r
ed
u
ce
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
15
,
No
.
5
,
Octo
b
e
r
20
25
:
5
0
3
1
-
5
0
4
4
5032
th
eir
f
r
e
q
u
en
c
y
.
I
n
ter
n
et
o
f
th
in
g
s
(
I
o
T
)
u
tili
ze
s
th
is
d
et
ec
tio
n
s
y
s
tem
to
r
ed
u
ce
r
is
k
an
d
im
p
r
o
v
e
t
h
e
p
er
f
o
r
m
an
ce
o
f
m
o
n
ito
r
in
g
th
e
ac
tiv
ities
o
f
p
atien
ts
an
d
th
e
eld
er
ly
[
3
]
,
[
4
]
.
B
esid
es,
f
all
d
etec
tio
n
ca
n
en
h
an
ce
I
o
T
ab
ilit
y
b
y
en
a
b
lin
g
co
n
tin
u
o
u
s
r
em
o
te
m
o
n
ito
r
in
g
o
f
th
e
en
v
ir
o
n
m
en
t
[
5
]
.
T
h
is
s
y
s
tem
co
m
m
o
n
l
y
r
e
q
u
ir
es
e
d
g
e
d
ev
ices
with
lo
w
co
m
p
u
tatio
n
c
o
s
t,
th
u
s
d
em
an
d
i
n
g
a
lig
h
tw
eig
h
t
an
d
ef
f
ec
tiv
e
alg
o
r
ith
m
.
T
h
e
s
m
ar
t
I
o
T
ca
n
im
p
lem
en
t
m
o
r
e
u
s
ef
u
l
p
r
ev
en
tiv
e
m
ea
s
u
r
es
b
y
u
tili
zin
g
th
ese
d
etec
tio
n
s
y
s
tem
s
[
6
]
–
[
8
]
.
I
t c
an
d
ec
r
ea
s
e
th
e
im
p
ac
t o
f
f
all
in
ju
r
ies an
d
p
r
o
m
o
te
h
ea
lth
y
an
d
ac
tiv
e
l
if
esty
les.
T
h
e
d
ee
p
lear
n
in
g
a
p
p
r
o
ac
h
h
as
p
lay
ed
a
cr
u
cial
r
o
le
i
n
ex
tr
ac
tin
g
c
o
m
p
lex
in
f
o
r
m
atio
n
an
d
ac
cu
r
ately
p
r
e
d
ictin
g
o
b
jects
an
d
b
e
h
av
io
r
s
[
9
]
,
[
1
0
]
.
T
h
es
e
n
etwo
r
k
s
ar
e
e
f
f
ec
tiv
e
d
u
e
to
th
eir
n
o
n
-
lin
ea
r
o
p
er
atio
n
s
an
d
d
ee
p
lay
er
s
,
wh
ich
s
im
u
ltan
eo
u
s
ly
p
r
o
ce
s
s
f
ea
tu
r
e
m
ap
s
.
T
h
e
n
ee
d
s
an
d
ch
allen
g
es
in
co
m
p
u
ter
v
is
io
n
em
p
h
asize
co
n
v
o
lu
ti
o
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NN)
as
m
o
d
er
n
m
eth
o
d
s
t
h
at
o
p
tim
ally
f
ilter
o
u
t
im
p
o
r
tan
t
f
ea
t
u
r
es
[
1
1
]
–
[
1
4
]
.
C
NNs
h
av
e
p
r
o
v
en
v
er
y
ef
f
ec
tiv
e
in
c
o
m
p
u
ter
v
is
io
n
task
s
s
u
ch
as
o
b
ject
r
ec
o
g
n
itio
n
an
d
im
ag
e
class
if
icatio
n
.
T
h
e
g
e
n
er
al
c
o
n
v
o
lu
tio
n
lay
er
is
f
o
llo
wed
b
y
an
ac
tiv
atio
n
lay
er
to
o
f
f
e
r
a
n
o
n
-
lin
ea
r
ity
f
u
n
ctio
n
.
C
NNs
f
o
r
f
all
d
etec
tio
n
h
a
v
e
p
a
v
ed
th
e
wa
y
f
o
r
p
o
wer
f
u
l
al
g
o
r
ith
m
s
s
u
ch
as
y
o
u
o
n
ly
lo
o
k
o
n
ce
(
YOL
O)
.
T
h
i
s
n
etwo
r
k
o
f
f
e
r
s
an
ef
f
icien
t
s
o
lu
tio
n
b
y
d
etec
tin
g
o
b
jects
in
a
s
in
g
le
p
r
o
ce
s
s
,
en
ab
lin
g
r
ea
l
-
tim
e
o
b
ject
d
ete
ctio
n
with
o
u
t
s
ac
r
if
icin
g
ac
cu
r
ac
y
[
1
5
]
,
[
1
6
]
.
YOL
Ov
8
,
th
e
latest
v
er
s
io
n
o
f
t
h
e
YOL
O
f
am
ily
,
s
h
o
ws
s
ig
n
if
ic
an
t
ad
v
a
n
ce
m
en
ts
o
v
er
its
p
r
e
d
ec
ess
o
r
s
.
I
t
ac
h
iev
es
h
ig
h
er
ac
cu
r
ac
y
i
n
o
b
ject
d
etec
tio
n
,
m
ak
in
g
it
h
ig
h
ly
r
eliab
le
f
o
r
ap
p
licatio
n
s
r
eq
u
i
r
in
g
r
o
b
u
s
t
r
ea
l
-
tim
e
d
etec
tio
n
[
1
7
]
.
Ho
wev
er
,
d
esp
ite
its
ac
cu
r
ac
y
,
it
r
e
q
u
ir
es
ex
ten
s
iv
e
en
e
r
g
y
r
eso
u
r
ce
s
.
YOL
Ov
8
-
n
an
o
h
as
b
ee
n
p
r
esen
ted
as
a
m
o
r
e
ef
f
icien
t
alg
o
r
ith
m
b
u
t
s
till
r
u
n
s
s
lo
wly
o
n
d
e
v
ices
with
lim
ited
co
m
p
u
tin
g
ca
p
ac
it
y
[
1
8
]
.
T
h
e
r
ef
o
r
e
,
im
p
r
o
v
in
g
ef
f
icie
n
cy
in
its
d
ev
elo
p
m
en
t
is
im
p
e
r
ativ
e
to
c
r
ea
te
a
lig
h
ter
ar
c
h
itectu
r
e
wit
h
o
u
t
co
m
p
r
o
m
is
in
g
d
etec
tio
n
p
er
f
o
r
m
a
n
ce
in
r
ea
l
-
wo
r
ld
ap
p
licatio
n
s
.
Hu
m
a
n
f
all
d
etec
tio
n
h
as
b
ee
n
ex
te
n
s
iv
ely
s
tu
d
ied
,
with
m
o
d
els d
esig
n
ed
to
q
u
ick
ly
r
e
d
u
ce
r
escu
e
tim
es a
n
d
s
ig
n
if
ic
an
tly
id
en
tify
h
u
m
an
b
o
d
y
m
o
v
em
en
ts
.
Ho
wev
er
,
d
ev
elo
p
in
g
a
s
u
itab
le
ar
c
h
itectu
r
e
p
r
esen
ts
s
ev
er
al
ch
alle
n
g
es,
p
ar
ticu
lar
ly
in
ca
p
tu
r
in
g
g
lo
b
al
a
n
d
lo
ca
l
in
f
o
r
m
atio
n
wh
ile
m
ain
tain
in
g
d
etec
tio
n
ac
c
u
r
ac
y
.
T
h
e
GL
-
YOL
O
-
L
ite
m
o
d
el,
d
e
v
elo
p
e
d
b
y
[
1
9
]
,
in
teg
r
ates
a
tr
an
s
f
o
r
m
e
r
b
lo
c
k
an
d
an
atten
tio
n
m
o
d
u
le
in
to
th
e
Y
OL
Ov
5
ar
ch
itectu
r
e
to
a
d
d
r
e
s
s
th
ese
is
s
u
es.
T
h
e
o
v
er
lap
p
i
n
g
c
h
allen
g
e
in
co
m
p
lex
en
v
i
r
o
n
m
e
n
ts
is
ad
d
r
ess
ed
b
y
th
e
ef
f
icien
t
d
iv
e
r
s
e
b
r
an
ch
b
lo
c
k
-
YOL
O
(ED
-
YOL
O)
m
o
d
el,
w
h
ich
u
s
es
YOL
Ov
5
s
as
its
b
ac
k
b
o
n
e
[
2
0
]
.
T
h
is
r
esear
ch
p
r
o
d
u
ce
s
a
r
ea
l
-
tim
e
f
ea
tu
r
e
ex
tr
ac
tio
n
th
at
en
co
u
r
ag
es
th
e
n
etwo
r
k
to
wo
r
k
o
p
tim
ally
.
An
o
th
er
s
tu
d
y
[
8
]
p
r
o
p
o
s
ed
a
v
is
io
n
-
b
ased
f
all
d
etec
tio
n
s
y
s
tem
th
at
em
p
lo
y
s
o
b
ject
tr
ac
k
in
g
a
n
d
im
ag
e
en
h
an
ce
m
e
n
t
tech
n
iq
u
es.
Pra
ctica
l
ap
p
licatio
n
s
d
r
iv
e
n
ew
r
esear
c
h
f
o
cu
s
ed
o
n
p
r
esen
tin
g
lig
h
tweig
h
t
alg
o
r
it
h
m
s
.
A
s
tu
d
y
[
2
1
]
in
tr
o
d
u
ce
d
a
lig
h
tweig
h
t
C
NN
ar
ch
itectu
r
e
u
s
in
g
YOL
Ov
5
,
wh
ich
r
ep
lace
s
th
e
en
tire
b
ac
k
b
o
n
e
with
Sh
u
f
f
leNe
tV2
.
A
s
tu
d
y
[
2
2
]
p
r
esen
ted
a
m
eth
o
d
th
at
in
te
g
r
ates
co
n
v
o
lu
tio
n
an
d
i
n
f
o
r
m
atio
n
s
u
p
p
r
ess
io
n
lay
er
s
to
r
ed
u
ce
co
m
p
u
tatio
n
al
o
v
er
h
ea
d
wh
ile
m
ain
tain
in
g
o
p
tim
al
d
etec
tio
n
p
er
f
o
r
m
an
ce
.
T
h
e
p
r
o
p
o
s
ed
s
tu
d
y
p
r
esen
ts
an
ef
f
icien
t
s
o
lu
tio
n
f
o
r
f
all
d
etec
tio
n
b
y
in
tr
o
d
u
cin
g
a
r
e
s
o
u
r
ce
-
ef
f
icien
t
ap
p
r
o
ac
h
,
en
ab
lin
g
im
p
lem
en
tatio
n
ac
r
o
s
s
m
u
ltip
le
p
latf
o
r
m
s
,
in
clu
d
in
g
I
o
T
-
b
ased
h
ar
d
war
e
,
p
ar
ticu
lar
ly
e
d
g
e
d
ev
ices.
E
n
h
an
ce
m
e
n
t
m
o
d
u
les
h
av
e
b
ee
n
wid
el
y
u
s
ed
to
im
p
r
o
v
e
th
e
p
r
ec
is
io
n
o
f
o
b
ject
lo
ca
lizatio
n
[
2
3
]
–
[
2
5
]
.
T
h
e
m
o
d
u
le
ad
o
p
te
d
an
atten
tio
n
m
ec
h
a
n
is
m
d
esi
g
n
ed
to
o
p
tim
ize
ac
c
u
r
ac
y
an
d
ef
f
icie
n
cy
i
n
th
is
m
o
d
el,
en
h
an
cin
g
its
ab
ilit
y
to
d
etec
t
f
alls
.
T
h
is
m
o
d
u
le
h
elp
s
th
e
ex
tr
ac
tio
n
f
ea
tu
r
e
f
o
cu
s
o
n
th
e
p
e
r
s
o
n
'
s
b
o
d
y
,
h
ig
h
lig
h
tin
g
s
p
ec
if
ic
attr
ib
u
tes
o
f
m
o
v
em
e
n
t
p
atter
n
s
an
d
b
o
d
y
p
o
s
itio
n
s
th
at
in
d
icate
f
allin
g
ac
tiv
ity
.
T
h
is
wo
r
k
f
o
cu
s
es
o
n
in
teg
r
atin
g
th
e
e
n
h
an
ce
m
e
n
t
m
o
d
u
le
in
to
th
e
n
etwo
r
k
'
s
b
ac
k
b
o
n
e
to
im
p
r
o
v
e
th
e
p
r
ec
is
io
n
o
f
f
all
d
etec
tio
n
.
T
h
e
n
etwo
r
k
g
ain
s
ad
d
itio
n
al
ca
p
ab
ilit
y
in
ex
tr
ac
tin
g
ess
en
tial
in
f
o
r
m
atio
n
b
y
ef
f
ec
tiv
ely
s
ep
ar
atin
g
elem
e
n
ts
f
r
o
m
th
e
b
ac
k
g
r
o
u
n
d
.
T
h
is
ad
v
an
tag
e
is
ac
h
iev
ed
with
o
u
t
ad
d
in
g
s
ig
n
if
ican
t
co
m
p
u
tatio
n
o
r
in
c
r
ea
s
in
g
th
e
n
u
m
b
er
o
f
p
ar
am
eter
s
.
T
h
e
s
u
m
m
ar
y
o
f
t
h
e
p
o
te
n
tial im
p
ac
t a
n
d
c
o
n
tr
ib
u
ti
o
n
s
o
f
th
is
s
tu
d
y
is
as f
o
llo
ws:
a.
An
ef
f
icien
t
f
all
d
etec
tio
n
s
y
s
tem
is
d
ev
elo
p
ed
as
an
I
o
T
-
b
ased
m
o
n
ito
r
in
g
s
y
s
tem
th
at
o
p
er
ates
o
n
lo
w
-
co
s
t c
o
m
p
u
tin
g
d
ev
ices.
b.
T
h
is
s
tu
d
y
p
r
o
p
o
s
es
a
co
n
s
e
cu
tiv
e
s
elec
tiv
e
en
h
an
ce
m
e
n
t
(
C
SE)
m
o
d
u
le
th
at
m
o
d
if
ie
s
th
e
s
tr
u
ctu
r
al
en
h
an
ce
m
e
n
t
o
f
YOL
Ov
8
-
n
a
n
o
to
im
p
r
o
v
e
f
all
d
etec
tio
n
p
er
f
o
r
m
an
ce
.
T
h
is
m
o
d
i
f
icatio
n
r
ef
in
es
th
e
tar
g
et
f
ea
tu
r
es o
f
th
e
h
u
m
an
b
o
d
y
s
p
ec
if
ic
to
f
all
ev
en
ts
.
c.
E
x
ten
s
iv
e
ev
alu
atio
n
is
co
n
d
u
cted
to
m
ea
s
u
r
e
th
e
p
er
f
o
r
m
a
n
ce
o
f
th
e
p
r
o
p
o
s
ed
d
etec
to
r
co
m
p
ar
ed
to
o
th
er
lig
h
tweig
h
t
n
etwo
r
k
d
etec
to
r
s
.
Ad
d
itio
n
ally
,
th
e
s
tu
d
y
an
al
y
ze
s
th
e
m
o
d
el'
s
ef
f
icien
cy
b
y
ex
am
i
n
in
g
th
e
p
r
o
p
o
s
ed
m
o
d
el'
s
n
u
m
b
er
o
f
p
ar
am
eter
s
,
co
m
p
u
tatio
n
al
co
m
p
lex
ity
,
an
d
in
f
er
e
n
ce
tim
e.
2.
M
E
T
H
O
D
2
.
1
.
B
a
ck
bo
ne
I
n
c
o
m
p
u
ter
v
is
io
n
,
th
e
ter
m
"b
ac
k
b
o
n
e"
is
an
alo
g
o
u
s
t
o
th
e
h
u
m
an
b
ac
k
b
o
n
e
th
at
s
u
p
p
o
r
ts
th
e
b
o
d
y
.
Similar
ly
,
in
YOL
O,
t
h
e
b
ac
k
b
o
n
e
is
th
e
p
r
im
ar
y
f
o
u
n
d
atio
n
f
o
r
C
NN
ar
ch
itectu
r
e
f
o
r
ex
t
r
ac
tin
g
in
f
o
r
m
atio
n
f
r
o
m
i
n
p
u
t
im
a
g
es.
I
n
YOL
Ov
8
,
th
e
C
3
m
o
d
u
le
f
r
o
m
YOL
Ov
5
h
as
b
ee
n
u
p
d
ate
d
t
o
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
E
fficien
t fa
ll d
etec
tio
n
u
s
in
g
li
g
h
tw
eig
h
t n
etw
o
r
k
to
en
h
a
n
ce
s
ma
r
t
…
(
P
in
r
o
lin
vic
D.
K
.
M
a
n
emb
u
)
5033
co
n
v
o
l
u
tio
n
al
two
f
aster
(
C
2
F).
T
h
is
u
p
d
ate
im
p
r
o
v
es
f
e
atu
r
e
ex
t
r
ac
tio
n
b
y
r
etain
in
g
in
f
o
r
m
atio
n
m
o
r
e
q
u
ick
ly
a
n
d
ef
f
icien
tly
.
Af
ter
th
e
C
2
F
s
tag
e,
th
e
p
r
o
ce
s
s
ed
o
u
tp
u
t
g
o
es
to
th
e
SP
PF
s
ta
g
e.
T
h
is
s
tag
e
a
d
d
s
v
ar
iatio
n
s
to
th
e
in
f
o
r
m
atio
n
b
ef
o
r
e
it
m
o
v
es
to
th
e
n
ec
k
o
f
t
h
e
YOL
O
ar
ch
itectu
r
e.
B
ef
o
r
e
en
ter
in
g
th
e
n
ec
k
,
a
n
ew
lay
er
ca
lled
C
2
F
-
C
SE
i
s
ad
d
ed
af
ter
SP
PF
.
I
t
en
s
u
r
es
th
at
th
e
v
ar
ied
in
f
o
r
m
atio
n
en
h
an
ce
s
th
e
m
o
d
el'
s
ab
ilit
y
to
h
ig
h
lig
h
t
im
p
o
r
tan
t
v
er
tical
an
d
h
o
r
izo
n
tal
in
f
o
r
m
atio
n
s
ep
ar
ately
f
r
o
m
t
h
e
two
s
p
atial
d
im
en
s
io
n
s
.
T
h
is
ap
p
r
o
ac
h
m
ak
es
th
e
m
o
d
el
m
o
r
e
f
o
cu
s
ed
o
n
ca
p
t
u
r
in
g
ess
en
tial
f
ea
tu
r
es
in
th
e
im
ag
e.
T
h
e
n
u
m
b
e
r
o
f
ch
an
n
el
lay
er
s
is
m
o
d
if
ied
t
o
r
ed
u
ce
c
o
m
p
u
tatio
n
in
th
e
p
r
o
p
o
s
ed
n
etwo
r
k
b
ac
k
b
o
n
e
.
L
im
itin
g
ch
an
n
el
ass
ig
n
m
en
t
en
co
u
r
ag
es
th
e
n
e
two
r
k
to
e
x
tr
ac
t
f
ea
tu
r
es
f
aster
d
u
r
i
n
g
th
e
tr
ain
in
g
an
d
in
f
e
r
en
ce
p
r
o
ce
s
s
es.
A
n
ew
YOL
Ov
8
s
ize
v
ar
ian
t
ca
l
led
YOL
Ov
8
h
(
YOL
Ov
8
-
Hy
p
er
n
an
o
)
lim
its
th
e
m
ax
im
u
m
n
u
m
b
er
o
f
c
h
an
n
el
s
in
th
e
n
etwo
r
k
lay
er
to
1
2
8
,
a
s
p
r
esen
ted
in
Fig
u
r
e
1
.
I
t
s
ig
n
if
ican
tly
r
e
d
u
ce
s
th
e
n
u
m
b
er
o
f
p
ar
am
eter
s
a
n
d
co
m
p
u
tatio
n
al
c
o
m
p
lex
ity
co
m
p
ar
ed
to
th
e
n
a
n
o
v
e
r
s
io
n
.
Fig
u
r
e
1
.
T
h
e
p
r
o
p
o
s
ed
a
r
ch
it
ec
tu
r
e
is
im
p
r
o
v
e
d
f
r
o
m
th
e
Y
OL
Ov
8
n
an
o
v
er
s
io
n
.
I
t c
o
n
s
is
ts
o
f
a
b
ac
k
b
o
n
e
as
th
e
m
ain
ex
tr
ac
t
o
r
f
ea
tu
r
e,
a
n
ec
k
to
r
elate
in
f
o
r
m
atio
n
in
d
if
f
er
en
t f
r
e
q
u
en
cies,
an
d
th
e
h
ea
d
is
r
esp
o
n
s
ib
le
f
o
r
p
r
ed
ictin
g
th
e
lo
ca
tio
n
a
n
d
d
im
en
s
io
n
o
f
an
o
b
ject
2
.
1
.
1
.
C2
F
T
h
e
co
n
v
o
lu
tio
n
al
two
f
aster
(
C
2
F)
m
o
d
u
le
in
YOL
Ov
8
is
in
s
p
ir
ed
b
y
th
e
p
r
e
v
io
u
s
v
er
s
io
n
o
f
th
e
co
n
v
o
l
u
tio
n
al
th
r
ee
(
C
3
)
b
lo
c
k
.
T
h
is
m
o
d
u
le
is
d
esig
n
ed
to
im
p
r
o
v
e
m
o
d
el
p
er
f
o
r
m
an
c
e
an
d
ef
f
icien
cy
in
YOL
Ov
8
.
T
h
e
C
2
F
m
o
d
u
le
c
o
m
p
r
is
es
two
co
n
v
o
lu
tio
n
o
p
er
atio
n
s
at
th
e
n
etwo
r
k
'
s
b
eg
in
n
in
g
a
n
d
e
n
d
.
T
h
e
in
p
u
t
in
f
o
r
m
atio
n
is
s
p
lit
in
to
two
p
ar
ts
af
ter
th
e
f
i
r
s
t
co
n
v
o
l
u
tio
n
.
T
h
e
f
ir
s
t
p
ar
t
p
ass
es
th
e
in
p
u
t
in
f
o
r
m
atio
n
b
y
p
er
f
o
r
m
in
g
a
r
esid
u
al
o
p
er
atio
n
.
I
n
co
n
tr
ast,
th
e
s
ec
o
n
d
p
ar
t
ap
p
lies
a
b
o
ttlen
ec
k
th
at
u
tili
ze
s
co
n
v
o
lu
tio
n
s
with
d
if
f
er
e
n
t
k
er
n
el
s
izes
to
ac
h
iev
e
o
p
tim
al
ef
f
icien
c
y
an
d
ef
f
ec
tiv
e
n
ess
.
Fu
r
th
er
m
o
r
e,
th
e
m
o
d
el
co
m
b
in
es
b
o
th
f
ea
tu
r
es
to
en
r
ich
t
h
e
d
if
f
er
en
t
in
f
o
r
m
ati
o
n
.
At
th
e
en
d
o
f
th
e
C
2
F
m
o
d
u
le,
t
h
e
m
o
d
el's
p
er
f
o
r
m
an
ce
is
en
h
an
ce
d
m
o
r
e
e
f
f
icien
tly
b
y
em
p
lo
y
in
g
a
1
×
1
co
n
v
o
lu
tio
n
o
p
er
atio
n
t
o
co
n
s
o
lid
ate
t
h
e
in
f
o
r
m
atio
n
.
2
.
1
.
2
.
C2
F
-
CSE
T
h
e
C
2
F
-
C
SE
m
o
d
u
le
m
o
d
if
ies
a
b
asic
m
o
d
u
le
o
f
C
2
F
b
y
ad
d
in
g
a
c
o
n
s
ec
u
tiv
e
s
elec
tiv
e
en
h
an
ce
m
e
n
t
(
C
SE)
m
o
d
u
le.
As
illu
s
tr
ated
in
Fig
u
r
e
2
,
tw
o
atten
tio
n
m
o
d
u
les
ar
e
d
e
v
elo
p
ed
t
o
im
p
r
o
v
e
t
h
e
m
o
d
el'
s
ca
p
ab
ilit
y
in
ca
p
tu
r
i
n
g
an
d
u
tili
zin
g
s
p
atial
an
d
ch
an
n
el
in
f
o
r
m
atio
n
,
r
esp
ec
ti
v
ely
.
T
h
e
p
r
o
p
o
s
ed
m
o
d
u
le
ca
n
in
c
r
ea
s
e
th
e
ab
ilit
y
o
f
ex
tr
ac
to
r
f
ea
t
u
r
es
to
d
is
cr
im
in
ate
b
etwe
en
v
ital
in
f
o
r
m
atio
n
an
d
tr
iv
ial
f
ea
tu
r
es.
I
ts
o
b
jectiv
e
is
to
f
o
cu
s
m
o
r
e
o
n
v
alu
ab
le
f
ea
tu
r
es
in
th
e
f
ea
tu
r
e
in
p
u
t.
B
esid
es,
it
p
ay
s
atten
tio
n
to
th
e
cr
itical
co
n
te
x
t
o
f
th
e
im
a
g
e
o
b
ject.
T
h
e
im
p
r
o
v
em
en
t
a
ls
o
aim
s
to
en
h
an
ce
t
h
e
m
o
d
el
'
s
ab
ilit
y
to
ca
p
tu
r
e
f
ea
tu
r
es o
f
in
ter
est wh
ile
o
p
tim
izin
g
th
e
ef
f
icien
cy
an
d
p
e
r
f
o
r
m
an
ce
o
f
th
e
m
o
d
el.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
15
,
No
.
5
,
Octo
b
e
r
20
25
:
5
0
3
1
-
5
0
4
4
5034
Fig
u
r
e
2
.
T
h
e
p
r
o
p
o
s
ed
C
SE
m
o
d
u
le
c
o
m
b
in
es c
h
a
n
n
el
an
d
s
p
atial
r
ep
r
esen
tatio
n
.
I
t im
p
l
em
en
ts
a
s
q
u
ee
ze
an
d
ex
citatio
n
m
o
d
u
le
at
th
e
b
eg
in
n
in
g
o
f
th
e
m
o
d
u
le
to
h
i
g
h
lig
h
t v
ital in
f
o
r
m
atio
n
al
o
n
g
th
e
ch
an
n
el
m
ap
an
d
s
p
atial
en
h
an
ce
m
en
t to
ca
p
tu
r
e
th
e
v
alu
ab
le
f
ea
t
u
r
es in
a
lar
g
er
s
p
atial
ar
ea
2
.
1
.
3
.
CSE
m
o
du
le
T
h
e
p
r
o
p
o
s
ed
n
etwo
r
k
u
tili
ze
s
th
e
s
q
u
ee
ze
-
an
d
-
ex
citatio
n
(
SE)
[
2
3
]
an
d
s
p
atial
atten
tio
n
b
lo
c
k
s
'
f
ea
tu
r
e
-
s
elec
tiv
e
ab
ilit
y
.
T
h
is
m
o
d
u
le
is
b
eliev
ed
to
im
p
r
o
v
e
th
e
p
r
ec
is
io
n
o
f
t
h
e
d
et
ec
tio
n
n
etwo
r
k
b
y
s
u
m
m
ar
izin
g
th
e
r
e
p
r
esen
tatio
n
f
ea
tu
r
e
a
n
d
g
e
n
er
atin
g
we
ig
h
ted
s
ca
lin
g
.
A
SE
b
lo
ck
i
s
th
e
f
ir
s
t
atten
tio
n
m
o
d
u
le
d
esig
n
e
d
to
im
p
r
o
v
e
n
etwo
r
k
p
er
f
o
r
m
a
n
ce
b
y
ex
p
l
icitly
m
o
d
elin
g
th
e
r
elatio
n
s
h
ip
b
etwe
en
f
ea
tu
r
e
ch
an
n
els
th
r
o
u
g
h
a
f
ea
tu
r
e
r
ec
alib
r
atio
n
p
r
o
ce
s
s
.
T
h
is
p
r
o
ce
s
s
as
s
ig
n
s
weig
h
ts
to
ea
c
h
f
ea
tu
r
e
ch
an
n
el.
Featu
r
e
r
ec
alib
r
atio
n
ap
p
lies
a
s
q
u
ee
ze
tech
n
iq
u
e
in
co
r
p
o
r
atin
g
s
p
atial
in
f
o
r
m
atio
n
in
to
th
e
c
h
an
n
el
d
escr
ip
to
r
s
,
an
d
th
e
ex
citatio
n
p
r
o
ce
s
s
lear
n
s
th
e
ch
a
n
n
el'
s
ac
tiv
atio
n
co
r
r
esp
o
n
d
i
n
g
to
t
h
e
in
p
u
t.
T
h
is
m
o
d
u
le
ca
n
f
o
r
m
u
late
as (
1
)
:
(
)
=
(
2
(
1
)
)
⨂
,
(
1
)
wh
er
e
=
1
×
∑
∑
,
,
.
=
1
=
1
(
2
)
I
n
th
e
in
itial
p
r
o
ce
s
s
,
th
e
in
p
u
t
in
f
o
r
m
atio
n
(
)
is
m
o
d
eled
b
y
ca
p
tu
r
in
g
th
e
g
lo
b
al
av
e
r
ag
e
r
e
g
io
n
o
f
th
e
f
ea
tu
r
e
m
a
p
ac
r
o
s
s
ea
ch
c
h
an
n
el
t
h
r
o
u
g
h
th
e
o
p
e
r
atio
n
o
f
.
T
h
e
r
e
p
r
esen
ted
f
ea
tu
r
e
i
s
th
en
p
r
o
ce
s
s
ed
th
r
o
u
g
h
two
f
u
lly
c
o
n
n
ec
te
d
l
ay
er
s
to
m
o
d
el
ch
a
n
n
el
d
ep
en
d
en
cies.
T
h
is
p
r
o
ce
s
s
is
f
o
llo
wed
b
y
ap
p
ly
i
n
g
th
e
r
ec
tifie
d
lin
ea
r
u
n
it
(
R
eL
U)
f
u
n
ctio
n
.
T
h
is
ac
tiv
atio
n
elim
i
n
ates
n
eg
ativ
e
in
p
u
t
v
alu
es,
t
h
er
eb
y
p
r
ev
e
n
tin
g
ir
r
elev
an
t
o
r
d
etr
im
en
tal
in
f
o
r
m
atio
n
p
r
o
p
ag
atio
n
i
n
s
u
b
s
eq
u
en
t
co
m
p
u
tatio
n
s
.
I
t
en
s
u
r
es
th
at
cr
itical
n
eu
r
o
n
s
ar
e
n
o
t
h
in
d
er
e
d
b
y
lo
w
o
r
n
e
g
ativ
e
s
co
r
es,
en
ab
lin
g
th
e
m
o
d
el
to
f
o
cu
s
o
n
v
alu
a
b
le
f
ea
tu
r
es
ef
f
ec
tiv
ely
.
T
h
e
weig
h
ts
o
f
th
e
two
f
u
lly
co
n
n
ec
ted
lay
er
s
ar
e
d
en
o
te
d
as
1
an
d
2
,
an
d
a
s
ig
m
o
id
f
u
n
ctio
n
(
)
is
em
p
lo
y
ed
to
g
en
er
ate
weig
h
ted
p
r
o
b
ab
ilit
y
s
co
r
es.
Su
b
s
eq
u
en
tly
,
th
e
o
u
tp
u
t
v
ec
to
r
f
r
o
m
t
h
e
s
ig
m
o
id
ac
tiv
atio
n
i
s
m
u
ltip
lied
with
th
e
o
r
ig
in
al
in
p
u
t
f
ea
tu
r
e
m
a
p
to
r
ef
i
n
e
th
e
in
itial
in
f
o
r
m
atio
n
b
ase
d
o
n
ch
a
n
n
el
-
wis
e
r
ep
r
esen
tatio
n
.
T
h
e
SE
n
etwo
r
k
p
r
o
v
id
es
o
n
ly
ch
a
n
n
el
-
s
p
ec
if
ic
atten
tio
n
i
n
f
ea
tu
r
e
e
x
tr
ac
tio
n
a
n
d
lack
s
en
h
an
ce
d
s
p
atial
r
ep
r
esen
tatio
n
.
T
h
er
e
f
o
r
e,
th
is
s
tu
d
y
i
n
co
r
p
o
r
ates
a
s
p
atial
atten
tio
n
m
o
d
u
le
to
im
p
r
o
v
e
th
e
en
h
an
ce
m
e
n
t
o
f
f
ea
t
u
r
es.
T
h
is
ad
d
itio
n
allo
ws
th
e
n
etwo
r
k
t
o
f
in
d
i
n
ter
esti
n
g
in
f
o
r
m
atio
n
in
s
p
atial
co
v
er
ag
e,
en
ab
lin
g
it
to
r
ec
o
g
n
ize
s
p
ec
if
ic
d
im
en
s
io
n
p
atter
n
s
to
in
d
icate
f
all
f
ea
tu
r
es.
T
h
e
ch
an
n
el
an
d
s
p
atial
in
f
o
r
m
atio
n
co
m
b
in
atio
n
ca
n
ac
cu
r
ately
u
n
d
er
s
tan
d
u
n
b
ala
n
ce
d
b
o
d
y
p
o
s
itio
n
s
as
in
d
ica
tio
n
s
o
f
a
f
all
wh
ile
ig
n
o
r
in
g
ir
r
elev
a
n
t a
r
ea
s
.
I
n
d
etail,
th
e
f
u
s
io
n
m
o
d
u
le
ca
n
illu
s
tr
ate
as (
3
)
:
(
)
=
(
7
7
[
(
(
)
)
,
(
(
)
)
]
)
⊗
(
)
.
(
3
)
T
h
e
SE
n
etwo
r
k
b
o
o
s
ts
th
e
q
u
ality
o
f
in
p
u
t
f
ea
tu
r
es,
wh
ic
h
ca
n
im
p
r
o
v
e
p
er
f
o
r
m
a
n
ce
b
y
ass
ig
n
in
g
m
o
r
e
r
elev
an
t
weig
h
t
to
im
p
o
r
tan
t
f
ea
tu
r
e
ch
an
n
els.
T
h
e
o
u
tp
u
t
th
en
ad
ap
tiv
ely
r
ewe
ig
h
t
s
es
s
en
tial
f
ea
tu
r
es
in
th
e
s
p
atial
d
im
en
s
io
n
,
en
h
a
n
cin
g
th
e
m
o
d
el'
s
p
er
f
o
r
m
an
c
e
o
n
s
p
atially
m
ap
p
e
d
f
ea
tu
r
e
s
.
Av
er
ag
e
p
o
o
lin
g
(
A
vg
P
o
o
l
)
an
d
m
ax
p
o
o
lin
g
(
Ma
xP
o
o
l
)
ar
e
u
s
ed
in
p
a
r
allel
b
lo
ck
s
to
o
b
tain
f
ea
tu
r
e
s
u
m
m
ar
ies.
T
h
e
two
s
p
atial
f
ea
tu
r
es
ar
e
f
u
s
ed
u
s
in
g
th
e
co
n
ca
ten
ate
o
p
e
r
atio
n
(
[]
)
,
an
d
th
en
a
7
×
7
co
n
v
o
lu
ti
o
n
f
ilter
(
7
7
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
E
fficien
t fa
ll d
etec
tio
n
u
s
in
g
li
g
h
tw
eig
h
t n
etw
o
r
k
to
en
h
a
n
ce
s
ma
r
t
…
(
P
in
r
o
lin
vic
D.
K
.
M
a
n
emb
u
)
5035
ex
tr
ac
ts
f
ea
tu
r
e
m
ap
in
f
o
r
m
ati
o
n
to
co
v
er
a
wid
er
r
ec
e
p
tiv
e
f
ield
an
d
d
etec
t
m
o
r
e
c
o
m
p
le
x
p
atter
n
s
.
Fin
ally
,
s
ig
m
o
id
ac
tiv
atio
n
(
)
em
p
h
asizes
an
d
h
ig
h
lig
h
ts
th
e
s
p
atial
atten
tio
n
m
ap
.
T
h
e
p
r
o
p
o
s
ed
r
esear
ch
in
teg
r
ates
th
e
m
o
d
if
ied
SE
n
etwo
r
k
in
to
th
e
b
ac
k
b
o
n
e
s
tr
u
ctu
r
e
o
f
th
e
YOL
Ov
8
ar
ch
itectu
r
e
to
ex
t
r
ac
t
d
ee
p
er
f
ea
tu
r
es
b
y
em
p
h
asizin
g
th
e
lear
n
ed
weig
h
t
r
ep
r
esen
ted
in
s
p
atial
an
d
ch
an
n
el
m
ap
s
.
T
h
e
co
m
b
in
ed
en
h
a
n
ce
m
en
t
m
o
d
u
le
im
p
r
o
v
es
th
e
n
etwo
r
k
'
s
ac
cu
r
ac
y
in
r
ec
o
g
n
izi
n
g
an
d
em
p
h
asizin
g
ess
en
tial
f
ea
tu
r
es.
T
h
e
m
o
d
el
also
f
o
cu
s
es
o
n
ef
f
icien
tly
o
p
e
r
atin
g
in
a
r
ea
lis
tic
ap
p
licatio
n
s
y
s
tem
.
Mo
r
e
o
v
er
,
th
e
atten
tio
n
m
o
d
u
le
ca
n
en
h
an
ce
th
e
o
p
tim
izatio
n
o
f
th
e
f
ea
tu
r
e
lear
n
in
g
p
r
o
ce
s
s
.
2
.
1
.
4
.
SPP
F
A
s
p
atial
p
y
r
am
id
p
o
o
lin
g
f
a
s
ter
(
SPPF
)
i
s
an
im
p
o
r
tan
t
c
o
m
p
o
n
en
t
in
th
e
b
ac
k
b
o
n
e
th
at
aim
s
to
in
cr
ea
s
e
o
b
ject
d
etec
tio
n
ca
p
a
b
ilit
y
with
h
ig
h
ef
f
icien
cy
o
n
d
iv
er
s
e
in
p
u
t
f
ea
tu
r
es.
T
h
is
m
o
d
u
le
o
p
tim
izes
th
e
o
r
ig
in
al
v
e
r
s
io
n
o
f
s
p
atial
p
y
r
am
id
p
o
o
lin
g
(
SPP
)
b
y
p
o
o
lin
g
f
ea
tu
r
es
u
s
in
g
v
ar
y
in
g
k
er
n
el
s
izes
(
e.
g
.
,
5
×
5
,
9
×
9
,
1
3
×
1
3
)
.
T
h
en
,
th
e
r
esu
lts
ar
e
co
m
b
in
e
d
to
cr
ea
te
a
r
ich
er
an
d
d
iv
er
s
e
f
ea
tu
r
e
r
ep
r
esen
tatio
n
.
T
h
is
p
r
o
ce
s
s
h
elp
s
th
e
m
o
d
el
ca
p
tu
r
e
in
f
o
r
m
atio
n
at
d
if
f
er
en
t
s
ca
les,
m
ak
in
g
d
etec
tin
g
o
b
jects
o
f
v
ar
i
o
u
s
s
izes
in
im
ag
es
ea
s
ier
.
Ad
d
itio
n
ally
,
th
is
m
o
d
u
le
im
p
r
o
v
es
co
m
p
u
t
atio
n
al
ef
f
icien
cy
.
I
t
ca
n
s
p
ee
d
u
p
th
e
in
f
er
e
n
ce
p
r
o
ce
s
s
an
d
s
tr
en
g
th
en
th
e
d
et
ec
tio
n
p
r
ec
is
io
n
,
m
ak
i
n
g
it
an
ex
citin
g
co
m
p
o
n
en
t
in
th
e
YO
L
Ov
8
ar
c
h
itectu
r
e,
esp
ec
ially
in
th
e
b
ac
k
b
o
n
e
p
ar
t.
2
.
2
.
Nec
k
T
h
e
n
ec
k
m
o
d
u
le
aim
s
to
r
ec
eiv
e
an
d
co
m
b
in
e
f
ea
t
u
r
es
f
r
o
m
v
ar
io
u
s
r
eso
lu
tio
n
lev
els
p
r
o
d
u
ce
d
in
th
e
b
ac
k
b
o
n
e
an
d
th
en
co
n
n
e
ct
th
e
in
f
o
r
m
atio
n
f
r
o
m
th
e
b
ac
k
b
o
n
e
t
o
th
e
h
ea
d
.
I
t
h
elp
s
im
p
r
o
v
e
t
h
e
f
ea
tu
r
e
r
ep
r
esen
tatio
n
b
ef
o
r
e
p
ass
in
g
it
to
th
e
f
in
al
p
r
ed
ictio
n
.
T
h
e
PANet
m
o
d
u
le
is
ad
o
p
te
d
to
e
n
ab
le
f
ea
tu
r
e
ex
tr
ac
tio
n
f
r
o
m
d
if
f
er
en
t
lev
el
s
o
f
r
eso
lu
tio
n
o
n
th
e
m
a
p
b
y
en
h
an
cin
g
th
e
m
o
d
el
to
r
ec
o
g
n
ize
d
if
f
er
e
n
t
-
s
ized
in
f
o
r
m
atio
n
.
PANet
u
tili
ze
s
r
ap
id
f
ea
tu
r
e
f
u
s
io
n
b
y
ex
tr
ac
tin
g
m
o
r
e
c
o
m
p
r
e
h
en
s
iv
e
in
f
o
r
m
atio
n
.
A
lig
h
t
b
o
ttlen
ec
k
m
o
d
u
le
is
o
f
f
er
ed
i
n
th
e
C
2
F
m
o
d
u
le
at
ea
ch
p
r
e
d
ictio
n
lay
er
.
T
h
is
b
o
ttlen
ec
k
s
tr
u
ctu
r
e
v
ar
iatio
n
en
h
an
ce
s
f
ea
tu
r
e
ex
tr
ac
tio
n
ef
f
ec
tiv
en
ess
wh
ile
r
ed
u
cin
g
co
m
p
u
tatio
n
al
co
s
t.
I
n
o
r
d
e
r
to
im
p
r
o
v
e
th
e
ef
f
icien
cy
o
f
th
e
n
etwo
r
k
,
it
p
r
o
p
o
s
es
C
2
F
-
Nex
t.
T
h
is
m
o
d
u
le
is
in
s
p
ir
ed
b
y
th
e
b
asic
m
o
d
u
le
o
f
c
o
n
v
o
lu
tio
n
al
two
Fas
ter
,
b
u
t
th
e
b
o
ttle
n
ec
k
p
ar
t
is
m
o
d
if
ied
u
s
in
g
th
e
lig
h
t
b
o
ttlen
ec
k
.
T
h
e
s
tr
u
ctu
r
e
o
f
th
e
l
ig
h
t
b
o
ttl
en
ec
k
ap
p
lies
th
e
s
tan
d
ar
d
b
o
t
tlen
ec
k
d
esig
n
[
2
6
]
,
as p
r
esen
ted
in
Fig
u
r
e
3
.
T
h
e
m
o
d
u
le
ca
n
r
ed
u
ce
c
o
m
p
u
tati
o
n
al
co
s
t
wh
ile
m
ain
tain
in
g
ex
tr
ac
tio
n
ab
ilit
y
with
o
u
t
s
ig
n
if
ican
tly
d
ec
lin
in
g
p
r
ec
is
io
n
.
Dep
th
wis
e
co
n
v
o
l
u
tio
n
ap
p
lies
a
s
in
g
le
ch
an
n
el
o
f
f
ilter
o
p
er
atio
n
t
h
at
c
o
m
p
r
o
m
is
es
m
ix
ed
in
f
o
r
m
atio
n
o
f
ea
ch
c
h
an
n
el
i
n
p
u
t.
T
h
is
p
r
o
ce
s
s
ca
n
s
av
e
m
an
y
p
ar
a
m
eter
s
an
d
a
r
a
p
id
ex
tr
ac
tio
n
p
r
o
ce
s
s
.
T
h
e
b
o
ttlen
ec
k
m
o
d
u
le
s
tr
u
c
tu
r
e
in
co
r
p
o
r
ates
s
ev
er
al
b
lo
ck
s
tr
u
ctu
r
es,
wh
ic
h
ad
o
p
t
d
ep
th
wis
e
o
p
er
atio
n
(
DW
)
at
in
p
u
t
in
f
o
r
m
atio
n
(
X
)
u
s
in
g
a
5
×
5
k
e
r
n
el,
as
s
h
o
w
n
in
Fig
u
r
e
4
.
T
h
e
lar
g
e
f
ilter
ca
p
tu
r
es
a
s
izab
le
s
p
atial
ar
ea
f
r
o
m
in
p
u
t
f
ea
t
u
r
es
an
d
h
elp
s
t
h
e
n
etwo
r
k
to
in
c
r
ea
s
e
th
e
v
ar
iety
o
f
th
e
elem
en
t
o
b
ject
r
elatio
n
s
h
ip
.
Fu
r
th
er
m
o
r
e,
it
u
tili
ze
s
L
ay
er
No
r
m
(
L
N)
to
p
r
o
ce
s
s
ea
ch
f
ea
tu
r
e
with
in
a
la
y
er
b
y
n
o
r
m
alizin
g
ea
ch
s
am
p
le'
s
in
f
o
r
m
atio
n
.
A
l
ig
h
t b
o
ttlen
ec
k
is
f
o
r
m
u
late
d
a
s
(
4
)
:
(
)
=
(
(
(
)
)
)
,
(
4
)
wh
er
e
(
)
=
(
(
(
)
)
)
.
(
5
)
T
h
is
m
o
d
u
le
ap
p
lies
p
o
in
twis
e
co
n
v
o
lu
tio
n
(
PW
)
,
wh
ich
em
p
lo
y
s
a
1
×
1
co
n
v
o
lu
tio
n
al
b
lo
ck
to
in
teg
r
ate
in
f
o
r
m
atio
n
f
r
o
m
v
ar
io
u
s
ch
an
n
els
wh
ile
p
r
eser
v
in
g
th
e
s
p
atial
d
im
en
s
io
n
s
o
f
th
e
f
ea
tu
r
es.
T
h
e
Gau
s
s
ian
er
r
o
r
lin
ea
r
u
n
it
(
GE
L
U
)
ac
tiv
atio
n
f
u
n
ctio
n
a
ls
o
h
elp
s
o
p
tim
ize
th
e
m
o
d
e
l's
p
er
f
o
r
m
a
n
ce
b
y
im
p
lem
en
tin
g
a
Gau
s
s
ian
ap
p
r
o
ac
h
to
g
en
er
ate
s
m
all
n
eg
ativ
e
in
p
u
ts
.
Su
b
s
eq
u
e
n
tly
,
g
lo
b
al
r
esp
o
n
s
e
n
o
r
m
aliza
tio
n
(
GR
)
n
o
r
m
alize
s
th
e
ac
tiv
atio
n
o
u
tp
u
t
ac
r
o
s
s
all
s
p
atial
f
ea
tu
r
es
with
in
a
lay
er
,
t
h
u
s
en
h
an
cin
g
th
e
tr
ain
in
g
p
r
o
ce
s
s
an
d
im
p
r
o
v
in
g
m
o
d
el
p
er
f
o
r
m
a
n
ce
.
Fin
ally
,
p
o
in
twis
e
co
n
v
o
lu
tio
n
i
s
ap
p
lied
in
th
e
last
m
o
d
u
le
to
r
ec
o
n
s
tr
u
ct
th
e
o
u
t
p
u
t
ch
an
n
el
an
d
r
e
f
in
e
s
in
g
le
s
p
atial
f
ea
tu
r
es,
r
esu
ltin
g
in
a
m
o
r
e
p
r
ec
is
e
m
o
d
el.
T
h
e
p
r
o
p
o
s
ed
m
o
d
u
le
f
o
cu
s
es o
n
ef
f
icien
tly
ex
tr
ac
tin
g
co
m
b
in
atio
n
f
ea
tu
r
es
in
th
e
n
ec
k
s
tag
e.
I
t
u
tili
ze
s
lig
h
t
o
p
er
atio
n
with
a
lin
ea
r
p
r
o
ce
s
s
with
o
u
t
in
v
o
l
v
in
g
m
u
lti
-
ch
a
n
n
el
m
ix
in
g
.
I
t
en
s
u
r
es
th
at
t
h
e
s
elec
ted
f
ea
tu
r
e
wo
r
k
s
ef
f
ec
tiv
ely
,
co
m
p
r
o
m
is
in
g
th
e
c
o
m
p
u
tatio
n
al
lo
ad
.
2
.
3
.
H
ea
d
I
n
its
f
in
al
s
tag
e,
th
e
m
o
d
el
em
p
lo
y
s
th
r
ee
h
ea
d
s
th
at
c
o
n
s
titu
te
a
n
eu
r
al
n
etwo
r
k
r
esp
o
n
s
ib
le
f
o
r
p
r
ed
ictin
g
th
e
lo
ca
tio
n
s
a
n
d
c
lass
es
o
f
o
b
jects.
T
h
ese
h
ea
d
s
d
eter
m
in
e
th
e
co
r
r
esp
o
n
d
in
g
class
es'
b
o
u
n
d
in
g
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
15
,
No
.
5
,
Octo
b
e
r
20
25
:
5
0
3
1
-
5
0
4
4
5036
b
o
x
l
o
ca
tio
n
s
an
d
d
im
e
n
s
io
n
s
with
o
u
t
ad
d
i
n
g
a
d
d
itio
n
al
c
o
m
p
u
tatio
n
al
o
v
e
r
h
ea
d
.
T
h
es
e
o
u
tp
u
t
h
ea
d
s
ar
e
tr
ain
ed
to
g
en
e
r
ate
o
f
f
s
et
s
co
r
es
attac
h
ed
to
th
e
f
in
al
lay
er
s
.
T
h
e
ty
p
es
o
f
o
u
tp
u
t
h
ea
d
s
ca
n
v
ar
y
d
ep
en
d
in
g
o
n
th
e
o
b
ject
d
etec
tio
n
alg
o
r
ith
m
an
d
task
r
eq
u
ir
em
en
ts
.
I
n
s
tead
o
f
u
s
in
g
an
ch
o
r
s
f
o
r
p
r
ed
ict
ed
b
o
x
es,
YOL
Ov
8
ad
o
p
ts
an
a
n
ch
o
r
-
f
r
ee
d
etec
tio
n
m
eth
o
d
th
at
d
ir
ec
tly
p
r
ed
i
cts
an
o
b
ject'
s
ce
n
ter
r
ath
er
t
h
an
its
o
f
f
s
et
f
r
o
m
a
p
r
ed
ef
in
e
d
b
o
x
.
T
h
is
ap
p
r
o
a
ch
r
ed
u
ce
s
th
e
n
u
m
b
er
o
f
b
o
x
p
r
ed
ictio
n
s
,
s
p
ee
d
s
u
p
p
o
s
t
-
p
r
o
ce
s
s
in
g
,
a
n
d
s
im
p
lifie
s
th
e
n
etwo
r
k
,
m
ak
i
n
g
it
f
aster
an
d
im
p
r
o
v
i
n
g
th
e
s
u
itab
ilit
y
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
with
lo
w
-
co
s
t
h
ar
d
war
e
co
n
f
ig
u
r
atio
n
s
.
Ho
wev
er
,
o
b
ject
d
etec
tio
n
o
f
te
n
en
co
u
n
ter
s
in
ac
cu
r
ac
ies
o
r
m
is
s
ed
d
etec
tio
n
s
,
lead
in
g
to
er
r
o
r
s
.
T
h
e
in
ter
s
e
ctio
n
o
v
er
u
n
io
n
(
I
o
U)
m
etr
ic
ad
d
r
ess
es
th
is
b
y
ca
lcu
lati
n
g
th
e
r
atio
o
f
th
e
o
v
er
lap
ar
ea
o
f
b
o
u
n
d
in
g
b
o
x
es
to
th
eir
u
n
io
n
ar
ea
.
I
t
ca
n
f
u
r
th
er
b
e
u
s
ed
to
co
m
p
u
te
th
e
co
m
p
lete
I
o
U
(
C
I
o
U)
,
wh
ich
in
co
r
p
o
r
ates
f
ac
to
r
s
s
u
ch
as
th
e
d
is
tan
ce
b
et
wee
n
b
o
u
n
d
in
g
b
o
x
ce
n
ter
s
an
d
an
asp
ec
t
r
atio
f
o
r
s
ca
le.
Dis
tr
ib
u
tio
n
f
o
ca
l
lo
s
s
(
DFL)
an
d
b
i
n
ar
y
cr
o
s
s
-
en
tr
o
p
y
(
B
C
E
)
lo
s
s
f
u
n
ctio
n
s
ar
e
e
m
p
lo
y
ed
to
e
v
alu
ate
th
e
b
o
u
n
d
in
g
b
o
x
es
r
e
g
r
ess
io
n
an
d
class
if
icatio
n
ac
cu
r
ac
y
o
f
d
etec
ted
o
b
jects,
r
esp
ec
tiv
ely
[
2
7
]
.
T
h
ese
cr
itical
lo
s
s
f
u
n
ctio
n
s
p
lay
a
p
iv
o
tal
r
o
le
i
n
tr
ain
i
n
g
t
h
e
m
o
d
el,
en
a
b
lin
g
e
n
h
an
ce
m
en
ts
in
p
r
ed
ictiv
e
p
er
f
o
r
m
an
ce
ac
r
o
s
s
s
u
cc
ess
iv
e
iter
atio
n
s
.
Fig
u
r
e
3
.
Mo
d
if
ied
C
2
f
with
li
g
h
t b
o
ttlen
ec
k
ap
p
lies
ef
f
icien
t o
p
er
atio
n
.
I
t o
n
ly
p
r
o
v
id
es a
n
ex
ten
s
iv
e
co
m
p
u
tatio
n
al
e
f
f
o
r
t
o
n
h
alf
o
f
th
e
p
ar
ts
b
u
t d
o
es n
o
t ig
n
o
r
e
th
e
f
ea
tu
r
es o
f
th
e
r
est
Fig
u
r
e
4
.
T
h
e
p
r
o
p
o
s
ed
lig
h
t b
o
ttlen
ec
k
m
o
d
u
le
ap
p
lies
a
lar
g
e
k
er
n
el
o
p
er
atio
n
.
I
t
u
tili
ze
s
d
ep
th
wis
e
co
n
v
o
l
u
tio
n
with
a
lar
g
e
k
er
n
e
l size
to
ca
p
tu
r
e
a
wid
er
s
p
atia
l a
r
ea
3.
DATAS
E
T
AND
I
M
P
L
E
M
E
NT
A
T
I
O
N
S
E
T
UP
3
.
1
.
Da
t
a
s
et
T
h
e
f
all
d
etec
tio
n
d
ataset
u
s
es
L
e2
i
[
2
2
]
an
d
c
o
n
s
is
ts
o
f
2
class
es
:
s
tan
d
in
g
an
d
f
allin
g
h
u
m
a
n
s
u
b
jects.
T
h
ese
im
ag
es
ar
e
s
o
u
r
ce
d
f
r
o
m
th
e
cr
e
d
ib
le
web
s
ite
R
o
b
o
f
lo
w,
w
h
ich
f
o
cu
s
es
o
n
two
m
ain
o
b
jectiv
es:
s
tan
d
in
g
d
etec
tio
n
an
d
f
allin
g
d
etec
tio
n
.
T
h
is
d
ataset
co
m
p
r
is
es
3
0
1
0
im
ag
es,
d
iv
id
ed
in
t
o
tr
ain
in
g
,
test
in
g
,
an
d
v
alid
atio
n
s
u
b
s
ets,
with
a
d
ataset
s
p
lit
o
f
7
0
%
f
o
r
tr
ain
in
g
,
2
0
%
f
o
r
t
esti
n
g
,
an
d
1
0
%
f
o
r
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
E
fficien
t fa
ll d
etec
tio
n
u
s
in
g
li
g
h
tw
eig
h
t n
etw
o
r
k
to
en
h
a
n
ce
s
ma
r
t
…
(
P
in
r
o
lin
vic
D.
K
.
M
a
n
emb
u
)
5037
v
alid
atio
n
.
T
h
e
s
ec
o
n
d
f
allen
d
ataset
[
2
8
]
c
o
n
s
is
ts
o
f
3
class
es:
f
allen
,
s
itti
n
g
,
an
d
s
ta
n
d
in
g
.
T
h
e
d
ataset
in
clu
d
es 3
,
2
9
0
im
ag
es,
d
iv
id
e
d
in
to
th
r
ee
p
ar
ts
: 7
4
% tr
ain
in
g
,
1
5
% v
alid
atio
n
,
an
d
1
0
% testi
n
g
.
3
.
2
.
I
m
ple
m
ent
a
t
io
n set
up
T
h
e
p
r
o
p
o
s
ed
m
eth
o
d
u
tili
ze
s
th
e
Py
T
o
r
ch
f
r
am
ewo
r
k
an
d
r
eq
u
ir
es
s
p
ec
ialized
h
ar
d
war
e,
in
clu
d
in
g
an
AM
D
R
y
ze
n
5
4
5
0
0
6
-
c
o
r
e
C
PU
@
4
.
2
GHz
p
r
o
ce
s
s
,
3
2
GB
R
AM
,
an
d
an
R
T
X
4
0
6
0
T
I
g
r
a
p
h
ics
ca
r
d
,
to
in
cr
ea
s
e
th
e
tr
ain
in
g
s
p
ee
d
.
T
h
e
s
y
s
tem
u
tili
ze
s
th
e
f
all
d
etec
tio
n
d
ataset
f
o
r
tr
ai
n
in
g
a
n
d
ev
alu
atio
n
.
T
h
e
ev
alu
atio
n
p
r
o
ce
s
s
r
elies
o
n
t
h
e
v
alu
es
o
f
av
er
a
g
e
p
r
ec
is
io
n
(
AP)
a
n
d
m
ea
n
av
er
a
g
e
p
r
e
cisi
o
n
(
m
AP)
,
u
s
in
g
an
I
o
U
th
r
esh
o
ld
o
f
0
.
5
.
T
h
e
t
r
ain
in
g
was
p
er
f
o
r
m
e
d
o
v
er
2
0
0
ep
o
ch
s
with
a
b
atch
s
ize
o
f
1
6
.
T
h
e
o
p
tim
izer
u
s
ed
was
Sto
ch
asti
c
g
r
ad
ien
t
d
escen
t
(
SGD
)
,
with
a
m
o
m
e
n
tu
m
o
f
0
.
9
3
7
an
d
a
lea
r
n
in
g
r
ate
o
f
0
.
0
1
.
T
h
is
wo
r
k
im
p
lem
e
n
ts
m
o
s
aic
au
g
m
en
tatio
n
th
at
u
tili
ze
s
cr
o
p
,
f
lip
,
zo
o
m
,
an
d
s
h
if
t
g
e
o
m
etr
y
a
p
p
r
o
ac
h
es
to
en
r
ich
th
e
d
ata
ch
allen
g
es.
T
h
e
m
o
s
aic
o
n
ly
was
co
n
d
u
cted
o
n
1
9
0
ep
o
c
h
s
,
an
d
th
e
r
em
ain
d
e
r
u
s
ed
n
o
r
m
al
m
o
d
e.
T
h
e
im
ag
e
i
n
th
e
o
v
er
all
e
x
p
er
im
en
t
is
g
en
er
alize
d
with
t
h
e
in
p
u
t
s
ize
d
im
en
s
io
n
o
f
6
4
0
×
6
4
0
.
I
n
test
in
g
s
p
ee
d
,
th
e
m
o
d
el
em
b
ed
d
ed
i
n
a
J
etso
n
Nan
o
4
GB
r
ep
r
esen
tin
g
an
I
o
T
d
e
v
ice
th
at
d
ir
ec
tly
co
n
n
ec
ts
with
a
web
ca
m
E
y
esec
4
K
in
liv
e
s
tr
ea
m
m
o
d
e.
4.
E
XP
E
R
I
M
E
N
T
AND
R
E
SU
L
T
S
T
h
is
s
ec
tio
n
in
v
esti
g
ates
th
e
p
r
o
p
o
s
ed
m
o
d
el'
s
ev
alu
atio
n
r
e
s
u
lts
with
two
f
all
d
etec
tio
n
d
atasets
th
at
m
ea
s
u
r
e
co
n
f
o
r
m
an
ce
to
in
ter
s
ec
tio
n
o
v
er
u
n
io
n
.
T
h
is
ex
p
e
r
im
en
t
also
co
m
p
ar
es
th
e
m
ea
n
av
er
ag
e
p
r
ec
is
io
n
(
m
AP)
p
er
f
o
r
m
a
n
ce
with
o
th
er
lig
h
tweig
h
t
d
etec
tio
n
m
o
d
els
with
in
th
e
s
co
p
e
o
f
th
e
YOL
O
f
am
ily
.
I
n
ad
d
itio
n
,
ef
f
icien
c
y
co
m
p
ar
is
o
n
s
ar
e
co
n
d
u
cted
b
y
m
ea
s
u
r
in
g
th
e
n
u
m
b
e
r
o
f
p
a
r
am
eter
s
,
e
f
f
icien
cy
,
an
d
d
ata
p
r
o
ce
s
s
in
g
s
p
ee
d
.
Fu
r
t
h
er
m
o
r
e,
a
c
o
m
p
r
e
h
en
s
iv
e
an
aly
s
is
o
f
th
e
m
o
d
el
is
p
r
esen
ted
in
th
i
s
s
tu
d
y
,
wh
ic
h
f
i
n
d
s
th
e
u
s
ag
e
im
p
ac
t o
f
th
e
p
r
o
p
o
s
ed
m
o
d
u
les.
4
.
1
.
Abla
t
io
n study
T
h
e
ab
latio
n
s
tu
d
y
p
r
esen
ts
th
e
p
r
o
p
o
s
ed
m
o
d
u
le
in
v
esti
g
ati
o
n
th
at
im
p
r
o
v
es
th
e
p
er
f
o
r
m
a
n
ce
o
f
th
e
YOL
Ov
8
n
an
o
v
er
s
io
n
.
T
h
e
i
n
ten
d
ed
n
etwo
r
k
YOL
Ov
8
h
-
LB
-
C
SE
was
co
m
p
ar
ed
at
ea
ch
s
tep
with
m
o
d
if
ie
d
b
lo
ck
s
tr
u
ctu
r
es
to
s
ee
th
e
im
p
ac
t
o
f
m
o
d
if
icatio
n
s
.
As
s
h
o
wn
in
T
ab
le
1
,
th
e
YOL
Ov
8
h
-
LB
-
C
SE
m
o
d
u
le
in
to
th
e
o
r
ig
in
al
YOL
Ov
8
n
h
as
a
n
o
tab
le
d
ec
r
ea
s
e
in
Pa
r
am
eter
s
b
y
6
0
.
3
4
%
an
d
FL
OP
b
y
2
2
.
2
2
%.
I
n
ad
d
itio
n
,
th
e
r
e
is
an
im
p
r
o
v
e
m
en
t
in
th
e
p
er
f
o
r
m
a
n
ce
m
o
d
el
b
y
1
.
1
7
%
an
d
0
.
6
9
%
o
n
th
e
L
e2
i
an
d
Fallen
d
atasets
,
r
esp
ec
tiv
ely
.
T
h
e
r
esear
ch
er
s
also
r
ec
o
n
s
tr
u
cted
th
e
ch
an
n
el
d
im
e
n
s
io
n
s
o
f
th
e
o
r
ig
in
al
YOL
OV8
n
an
o
,
wh
ich
was
lim
ited
to
a
m
ax
im
u
m
o
f
1
2
8
c
h
an
n
els.
T
h
is
m
o
d
if
ied
m
o
d
el
is
ca
lled
YOL
OV8
h
(
Hy
p
er
n
a
n
o
)
b
ec
a
u
s
e
th
er
e
is
a
s
ig
n
if
ican
t
d
ec
r
ea
s
e
in
lear
n
ab
le
p
ar
am
eter
s
b
y
6
9
.
8
4
%
an
d
th
e
n
u
m
b
er
o
f
o
p
er
atio
n
s
b
y
2
6
.
3
9
%.
Mo
r
e
o
v
er
,
it
ad
d
e
d
a
lig
h
t
b
o
ttlen
ec
k
m
o
d
u
le
to
th
e
YOL
OV8
-
H
y
p
er
n
a
n
o
m
o
d
u
le
s
tr
u
ctu
r
e,
wh
ich
h
elp
s
im
p
r
o
v
e
ac
cu
r
ac
y
with
o
u
t
s
ig
n
if
ican
tly
s
ac
r
if
icin
g
co
m
p
u
tati
o
n
co
s
t.
T
h
e
lig
h
t
b
o
ttlen
ec
k
m
o
d
u
le
s
tr
u
ctu
r
e
i
s
d
esig
n
ed
to
m
ain
tain
ef
f
icien
cy
an
d
f
ea
tu
r
e
ex
t
r
ac
tio
n
c
ap
ab
ilit
ies,
an
d
th
e
co
m
b
in
ed
m
o
d
el
is
ca
lled
YO
L
Ov
8
h
-
L
B
.
T
ab
le
1
.
Ab
latio
n
ex
p
e
r
im
en
ts
with
d
if
f
er
en
t im
p
r
o
v
em
en
t st
r
ateg
ies.
I
t a
d
d
s
th
e
p
r
o
p
o
s
ed
m
o
d
u
les
u
n
til th
ey
r
ea
c
h
th
e
e
n
tire
p
r
o
p
o
s
ed
n
etwo
r
k
M
o
d
e
l
s
G
F
LO
P
S
P
a
r
a
me
t
e
r
s
mA
P
@
0
.
5
:
0
.
9
5
o
n
Le
2
i
d
a
t
a
s
e
t
mA
P
@
0
.
5
o
n
F
a
l
l
e
n
d
a
t
a
s
e
t
Y
O
LO
v
8
n
7
.
2
3
,
0
1
1
,
2
3
8
0
.
5
9
6
0
.
7
2
7
Y
O
LO
v
8
h
5
.
3
9
0
7
,
2
3
8
0
.
5
7
3
0
.
7
0
8
Y
O
LO
v
8
h
-
LB
5
.
8
1
,
1
8
5
,
5
9
8
0
.
5
7
1
0
.
7
1
2
Y
OLOv
8
h
-
LB
-
C
S
E
5
.
6
1
,
1
9
4
,
4
4
0
0
.
6
0
3
0
.
7
3
2
Fu
r
th
er
m
o
r
e
,
th
e
YOL
Ov
8
h
with
L
B
an
d
C
SE
co
m
b
in
es
a
lig
h
t
b
o
ttlen
ec
k
with
t
h
e
p
r
o
p
o
s
ed
en
h
an
ce
m
e
n
t
m
o
d
u
le,
d
esig
n
ed
to
im
p
r
o
v
e
n
etwo
r
k
p
er
f
o
r
m
an
ce
b
y
r
ec
alib
r
atin
g
th
e
in
p
u
t
f
ea
tu
r
es.
T
h
ese
f
in
d
in
g
s
p
r
o
v
e
th
at
th
e
en
h
a
n
ce
d
YOL
Ov
8
p
r
o
v
id
es
s
u
p
er
io
r
d
etec
tio
n
ef
f
icac
y
in
f
a
ll
d
etec
tio
n
.
I
t
also
b
en
ef
its
f
r
o
m
th
e
lig
h
tweig
h
t
in
co
r
p
o
r
atio
n
o
f
m
o
d
u
les,
lea
d
in
g
to
r
ed
u
ce
d
m
o
d
el
c
o
m
p
l
ex
ity
.
I
n
cl
u
d
in
g
th
e
lig
h
t
b
o
ttlen
ec
k
an
d
Sq
u
ee
ze
-
an
d
-
E
x
citatio
n
m
o
d
u
les
allo
ws
th
e
m
o
d
el
to
ca
p
tu
r
e
ess
en
tial
in
f
o
r
m
atio
n
a
n
d
r
ec
alib
r
ate
f
ea
tu
r
es
ef
f
ec
tiv
e
ly
,
im
p
r
o
v
in
g
o
v
er
all
ac
cu
r
ac
y
with
o
u
t
ad
d
in
g
s
ig
n
if
ic
an
t
co
m
p
u
tatio
n
al
o
v
er
h
ea
d
.
T
h
is
en
h
an
ce
m
e
n
t
en
co
u
r
a
g
es
YOL
Ov
8
to
b
e
p
ar
ticu
lar
ly
s
u
itab
le
f
o
r
d
e
p
lo
y
m
en
t
in
r
ea
l
-
w
o
r
ld
ap
p
licatio
n
s
with
lim
ited
c
o
m
p
u
tatio
n
al
r
eso
u
r
ce
s
.
R
ed
u
cin
g
th
e
n
u
m
b
er
o
f
p
a
r
am
eter
s
an
d
FLOPs
m
ak
es
th
e
m
o
d
el
m
o
r
e
ef
f
icien
t
an
d
f
aster
,
wh
ich
is
cr
u
cial
f
o
r
r
ea
l
-
ti
m
e
f
all
d
etec
tio
n
s
y
s
tem
s
.
Fu
r
th
er
m
o
r
e,
ca
r
ef
u
lly
r
ec
o
n
s
tr
u
ctin
g
th
e
ch
a
n
n
el
d
i
m
en
s
io
n
s
en
s
u
r
es
th
at
th
e
m
o
d
el
r
em
ain
s
co
m
p
ac
t
wh
ile
m
ain
tain
in
g
h
ig
h
p
er
f
o
r
m
an
ce
,
m
a
k
in
g
it a
n
id
e
al
s
o
lu
tio
n
f
o
r
ed
g
e
c
o
m
p
u
tin
g
d
ev
ices.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
15
,
No
.
5
,
Octo
b
e
r
20
25
:
5
0
3
1
-
5
0
4
4
5038
4
.
2
.
E
v
a
lua
t
i
o
n o
n da
t
a
s
et
s
T
h
is
s
tu
d
y
co
n
d
u
cted
a
v
is
u
al
an
aly
s
is
to
illu
s
tr
ate
th
e
d
etec
tio
n
p
er
f
o
r
m
a
n
ce
o
f
th
e
m
o
d
if
ied
YOL
Ov
8
u
n
d
er
v
ar
io
u
s
c
o
n
d
it
io
n
s
,
as
s
h
o
wn
in
Fig
u
r
e
s
5
(
a)
an
d
5
(
b
)
.
E
ac
h
s
et
o
f
test
im
ag
es
co
n
s
is
ts
o
f
two
co
m
p
o
n
en
ts
:
f
allin
g
an
d
s
tan
d
in
g
ca
teg
o
r
ies.
T
h
e
lef
t
p
ar
t
p
r
esen
ts
th
e
o
r
ig
in
al
p
h
o
to
,
wh
ile
th
e
r
ig
h
t
p
ar
t
illu
s
tr
ates th
e
h
ea
tm
ap
r
esu
lts
o
f
th
e
m
o
d
if
ied
YOL
Ov
8
alg
o
r
ith
m
.
T
h
is
v
is
u
aliza
tio
n
u
tili
ze
s
th
e
E
ig
en
-
C
AM
ap
p
r
o
ac
h
,
h
ig
h
lig
h
tin
g
th
e
m
o
s
t
im
p
o
r
tan
t
f
ea
tu
r
es
in
r
ed
p
ix
els.
T
h
e
h
ea
tm
ap
o
f
th
e
f
allin
g
ca
teg
o
r
y
is
s
h
o
wn
in
Fig
u
r
e
5
(
a)
,
d
e
m
o
n
s
tr
atin
g
th
at
th
e
p
r
o
p
o
s
ed
m
o
d
el
em
p
h
asizes
v
alu
ab
le
in
f
o
r
m
atio
n
o
n
th
e
b
o
d
y
p
ar
t o
f
th
e
f
allen
o
b
ject.
As a
r
esu
lt,
th
is
m
o
d
el
ca
n
ef
f
ec
tiv
el
y
r
ec
o
g
n
ize
th
e
f
allin
g
p
o
s
itio
n
.
Fo
r
th
e
s
tan
d
in
g
ca
teg
o
r
y
,
th
e
h
ea
tm
ap
in
d
icat
es
th
at
th
e
m
o
d
el
f
o
cu
s
es
o
n
th
e
co
r
r
ec
t
p
r
e
d
ictio
n
,
s
h
o
win
g
th
at
th
e
h
ea
tm
ap
ar
ea
is
in
clin
ed
v
er
tically
.
I
t
also
h
ig
h
lig
h
ts
th
e
s
h
o
u
ld
er
s
an
d
f
ee
t
as
th
e
m
ain
in
d
icato
r
s
o
f
th
e
s
tan
d
in
g
p
o
s
itio
n
.
(
a)
(
b
)
Fig
u
r
e
5
.
Hea
tm
ap
o
b
s
er
v
atio
n
o
f
th
e
p
r
o
p
o
s
ed
d
etec
to
r
.
I
t t
ests
o
n
(
a)
L
e2
i a
n
d
(
b
)
Fallen
d
atasets
.
T
h
e
tar
g
et
o
b
ject
is
d
etec
ted
t
h
r
o
u
g
h
g
r
ee
n
b
o
x
es
T
h
is
s
tu
d
y
i
n
v
esti
g
ates
th
e
m
ea
n
av
e
r
ag
e
p
r
ec
is
io
n
o
f
ea
ch
p
r
ed
ictio
n
a
g
ain
s
t
ea
ch
class
lab
el.
T
h
e
co
n
f
u
s
io
n
m
at
r
ix
o
n
t
h
e
L
e2
i
d
ataset
is
p
r
esen
ted
in
Fig
u
r
e
6
(
a)
.
T
h
e
f
all
class
ex
h
ib
its
th
e
h
ig
h
est
ac
cu
r
ac
y
,
with
0
.
9
6
in
s
tan
ce
s
co
r
r
ec
tly
class
if
ied
.
Ho
wev
er
,
th
er
e
is
a
0
.
0
4
m
is
class
if
icatio
n
r
ate,
wh
er
e
in
s
tan
ce
s
o
f
f
alls
ar
e
in
co
r
r
ec
tly
class
if
ied
as
"stan
d
in
g
.
"
C
o
n
v
e
r
s
ely
,
th
e
s
tan
d
in
g
class
h
as
a
co
r
r
ec
t
cl
ass
if
icatio
n
r
ate
o
f
0
.
8
9
b
u
t
o
b
tain
s
a
0
.
1
1
m
is
c
lass
if
icatio
n
r
ate,
with
in
s
tan
ce
s
th
at
s
h
o
u
ld
b
e
class
if
ied
as
s
tan
d
in
g
b
ein
g
in
co
r
r
ec
tly
id
e
n
tifie
d
as
f
all.
T
h
is
an
aly
s
is
h
ig
h
lig
h
ts
th
e
m
o
d
el'
s
s
tr
en
g
th
s
an
d
ar
ea
s
f
o
r
im
p
r
o
v
em
e
n
t
in
d
is
tin
g
u
is
h
in
g
b
etwe
en
f
all
an
d
s
tan
d
in
g
e
v
en
ts
.
Fig
u
r
e
6
(
b
)
illu
s
tr
ates
th
e
co
n
f
u
s
io
n
m
atr
ix
f
o
r
th
e
Fallen
d
ataset,
wh
ich
in
clu
d
es
th
r
ee
class
e
s
:
f
allen
,
s
itti
n
g
,
an
d
s
tan
d
i
n
g
.
T
h
e
s
tan
d
i
n
g
class
ex
h
ib
its
th
e
h
i
g
h
est
co
r
r
ec
t
class
if
icatio
n
r
ate
at
0
.
8
1
.
Ho
wev
er
,
th
e
r
e
ar
e
m
is
class
if
icatio
n
s
co
r
es
o
f
0
.
0
2
in
s
tan
d
in
g
in
s
tan
ce
s
,
in
c
o
r
r
ec
tly
class
if
y
in
g
it
as
s
itti
n
g
,
an
d
b
ac
k
g
r
o
u
n
d
m
is
class
if
icat
io
n
o
f
0
.
1
6
.
Fo
r
th
e
f
allen
class
,
th
e
co
r
r
ec
t
p
r
ed
ictio
n
r
ate
is
0
.
6
6
,
b
u
t
th
er
e
a
r
e
f
allen
m
is
class
if
icatio
n
s
o
f
0
.
0
6
p
r
ed
icted
as
s
itti
n
g
an
d
b
a
ck
g
r
o
u
n
d
m
is
class
if
icatio
n
s
o
f
0
.
2
8
.
T
h
e
s
itti
n
g
class
o
b
tain
s
th
e
lo
wes
t
co
r
r
ec
t
p
r
ed
ictio
n
r
ate
o
f
0
.
6
0
,
w
ith
m
is
clas
s
if
icatio
n
s
o
f
0
.
0
7
,
0
.
1
0
,
an
d
0
.
2
3
as
f
allen
,
s
tan
d
in
g
,
an
d
b
ac
k
g
r
o
u
n
d
,
r
esp
ec
tiv
el
y
.
T
h
is
an
aly
s
is
in
d
icate
s
ar
ea
s
wh
er
e
th
e
m
o
d
el'
s
ac
cu
r
ac
y
ca
n
im
p
r
o
v
e,
p
ar
ticu
lar
ly
d
is
tin
g
u
i
s
h
in
g
b
etwe
en
s
im
ilar
p
o
s
tu
r
e
s
.
T
h
is
wo
r
k
co
m
p
ar
es
th
e
p
er
f
o
r
m
an
ce
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
with
th
e
ef
f
icien
t
YOL
O
f
am
ilies
.
I
t
s
h
o
ws
th
at
o
u
r
d
etec
to
r
is
s
u
p
er
io
r
to
co
m
p
etito
r
s
,
s
u
ch
as
YOL
Ov
3
tin
y
,
YOL
Ov
5
n
,
Y
OL
Ov
6
n
,
YOL
Ov
7
tin
y
,
YOL
Ov
8
n
,
an
d
YOL
Ov
1
0
n
.
Per
f
o
r
m
a
n
ce
ev
alu
atio
n
s
h
o
ws
th
at
YOL
Ov
8
h
-
LB
-
C
SE
ac
h
iev
es
th
e
b
est
p
r
ec
is
io
n
,
m
ea
s
u
r
ed
with
m
AP
o
f
0
.
6
0
3
in
0
.
5
:
9
5
I
o
U.
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
o
u
tp
e
r
f
o
r
m
s
th
e
o
r
ig
in
al
YOL
Ov
8
n
,
d
if
f
e
r
in
g
b
y
a
m
AP
o
f
5
.
6
1
%.
E
v
en
it
is
s
u
p
er
io
r
to
th
e
n
ew
v
er
s
io
n
o
f
th
e
lig
h
tweig
h
t
YOL
Ov
1
0
n
.
T
h
is
o
b
s
er
v
atio
n
also
co
m
p
ar
es
th
e
p
er
f
o
r
m
an
ce
o
f
YOL
Ov
8
h
-
L
B
with
s
ev
er
al
atten
tio
n
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
E
fficien
t fa
ll d
etec
tio
n
u
s
in
g
li
g
h
tw
eig
h
t n
etw
o
r
k
to
en
h
a
n
ce
s
ma
r
t
…
(
P
in
r
o
lin
vic
D.
K
.
M
a
n
emb
u
)
5039
m
o
d
u
les,
a
n
d
t
h
e
m
aj
o
r
ity
p
r
esen
ts
im
p
r
o
v
in
g
p
r
ec
is
io
n
.
A
n
en
h
an
ce
m
en
t
b
lo
c
k
is
ass
ig
n
ed
to
in
cr
ea
s
e
t
h
e
ab
ilit
y
o
f
ex
t
r
ac
tio
n
f
ea
t
u
r
es,
alth
o
u
g
h
it
o
n
l
y
d
o
es
n
o
t
s
ig
n
if
ican
tly
ad
d
to
t
h
e
co
m
p
u
tatio
n
co
s
t.
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
with
C
SE
o
b
tain
s
a
m
AP
o
f
5
.
9
8
%
h
ig
h
er
th
an
YOL
Ov
8
h
-
LB
-
SE,
wh
ich
u
s
es
th
e
Sq
u
ee
ze
E
x
citatio
n
atten
tio
n
m
o
d
u
le.
T
h
is
s
tr
u
ctu
r
e
o
n
ly
f
o
cu
s
es
o
n
ch
an
n
el
-
wis
e
en
h
an
ce
m
en
t.
Fu
r
th
er
m
o
r
e
,
th
e
p
r
o
p
o
s
ed
m
o
d
el
also
p
er
f
o
r
m
s
b
etter
th
an
C
B
A
M
atten
t
io
n
,
wh
ich
d
if
f
e
r
s
b
y
2
%
m
AP.
C
B
AM
u
s
e
s
a
co
n
f
ig
u
r
atio
n
s
im
ilar
to
C
SE
b
u
t
with
o
v
er
-
in
-
ch
a
n
n
el
ex
tr
ac
tio
n
.
C
o
m
p
ar
ed
t
o
th
e
C
AN
an
d
E
L
A
atten
tio
n
m
o
d
u
les,
o
u
r
n
etwo
r
k
s
h
o
ws
h
ig
h
er
m
AP.
Alth
o
u
g
h
b
o
th
o
f
th
ese
atten
tio
n
s
in
v
o
lv
e
co
n
t
ex
t
d
is
co
v
er
y
f
r
o
m
th
e
s
p
atial
r
e
g
io
n
s
o
f
t
h
e
m
a
p
,
th
e
y
a
r
e
n
o
t
r
o
b
u
s
t
en
o
u
g
h
to
d
is
cr
im
in
ate
b
etwe
en
f
a
llin
g
an
d
s
tan
d
in
g
f
ea
tu
r
es.
GC
NE
T
an
d
DAN
p
er
f
o
r
m
e
d
well
in
th
e
o
b
ject
d
e
tectio
n
task
b
u
t c
o
u
ld
n
o
t o
u
tp
er
f
o
r
m
C
SE.
(
a)
(
b
)
Fig
u
r
e
6
.
C
o
n
f
u
s
io
n
m
atr
ices
o
f
m
o
d
el
p
r
e
d
ictio
n
.
I
t e
v
alu
at
es o
n
(
a)
L
e
2
i a
n
d
(
b
)
Fallen
d
atasets
On
th
e
o
th
er
h
an
d
,
T
ab
le
2
s
h
o
ws
a
co
m
p
ar
is
o
n
o
f
YOL
Ov
8
h
-
LB
-
C
SE
p
er
f
o
r
m
a
n
ce
with
o
th
er
d
etec
to
r
s
o
n
th
e
Fallen
d
ataset.
T
h
e
p
r
o
p
o
s
ed
n
etwo
r
k
ac
h
ie
v
es m
AP@
0
.
5
o
f
0
.
7
3
2
,
i
n
d
ic
atin
g
th
at
o
u
r
m
o
d
el
o
u
tp
er
f
o
r
m
s
YOL
Ov
8
n
b
y
0
.
6
8
8
%
m
AP
an
d
s
u
r
p
ass
es
YOL
Ov
7
tin
y
b
y
3
.
3
9
%.
O
n
th
is
d
ataset,
th
e
p
r
o
p
o
s
ed
m
o
d
el
also
co
m
p
ar
es
th
e
p
r
ec
is
io
n
with
o
th
er
YOL
O
lig
h
tweig
h
t
m
o
d
els.
Alth
o
u
g
h
YOL
Ov
6
n
o
u
tp
er
f
o
r
m
s
o
u
r
d
etec
to
r
,
o
th
er
lig
h
tweig
h
t
d
etec
to
r
s
u
n
d
e
r
p
er
f
o
r
m
.
YOL
Ov
6
n
ac
h
ie
v
e
s
h
ig
h
er
p
r
ec
is
io
n
th
an
th
e
p
r
o
p
o
s
ed
m
o
d
el
b
y
0
.
0
0
3
,
b
u
t th
e
m
o
d
el
g
en
er
ates
m
o
r
e
p
ar
a
m
eter
s
an
d
co
m
p
u
ta
tio
n
.
Mo
r
eo
v
er
,
th
is
s
atis
f
ac
to
r
y
p
er
f
o
r
m
an
ce
also
o
u
tp
er
f
o
r
m
s
th
e
m
AP
o
f
YOL
Ov
5
n
b
y
0
.
0
2
1
.
A
co
m
p
a
r
is
o
n
with
th
e
s
tate
-
of
-
th
e
-
ar
t
n
etwo
r
k
,
YOL
Ov
1
0
n
,
s
h
o
ws
th
at
o
u
r
d
etec
to
r
is
s
u
p
er
io
r
b
y
0
.
2
%.
T
h
e
H
y
p
er
n
a
n
o
v
er
s
io
n
s
h
o
ws
a
lo
wer
p
er
f
o
r
m
an
ce
th
an
th
e
f
u
ll
p
r
o
p
o
s
ed
n
etwo
r
k
.
T
h
is
r
e
s
u
lt
r
ep
r
esen
ts
th
at
th
e
p
r
o
p
o
s
ed
g
ain
m
o
d
u
le
ca
n
im
p
r
o
v
e
p
er
f
o
r
m
an
ce
in
r
ec
o
g
n
izin
g
f
allin
g
,
s
itti
n
g
,
an
d
s
tan
d
in
g
ac
tiv
ities
,
th
er
eb
y
d
em
o
n
s
tr
atin
g
th
e
p
r
ac
tical
im
p
licatio
n
s
o
f
o
u
r
w
o
r
k
.
4
.
3
.
E
v
a
lua
t
i
o
n o
f
m
o
del e
f
f
i
ciency
T
h
e
d
esig
n
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
co
n
s
id
er
s
th
e
ad
v
a
n
tag
es
th
at
h
elp
t
h
e
ap
p
licatio
n
s
ce
n
ar
io
.
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
is
d
esig
n
ed
with
atten
tio
n
to
s
ev
e
r
al
im
p
o
r
tan
t
asp
ec
ts
,
s
u
ch
as
a
lo
w
n
u
m
b
e
r
o
f
g
ig
a
f
lo
atin
g
-
p
o
in
t
o
p
er
atio
n
s
p
er
s
ec
o
n
d
s
(
GFLO
P
S
)
an
d
a
m
i
n
im
al
n
u
m
b
er
o
f
p
ar
am
ete
r
s
co
m
p
ar
ed
to
o
th
er
m
o
d
els.
T
h
is
r
esear
ch
p
r
io
r
iti
ze
s
ef
f
icien
cy
;
a
m
o
r
e
ef
f
icien
t
m
o
d
el
ca
n
p
e
r
f
o
r
m
m
an
y
c
o
m
p
u
tatio
n
al
task
s
with
f
ewe
r
o
p
e
r
atio
n
s
an
d
f
e
wer
tr
ain
ab
le
weig
h
ts
.
T
h
is
is
s
u
e
is
d
ir
ec
tly
r
elate
d
t
o
th
e
n
u
m
b
er
o
f
GFLO
P
S
an
d
th
e
to
tal
n
u
m
b
e
r
o
f
p
ar
a
m
eter
s
.
An
aly
s
is
o
f
th
e
co
m
p
ar
ativ
e
ex
p
er
im
e
n
ts
YOL
Ov
8
h
-
LB
-
C
SE
is
th
e
ch
ea
p
est
m
o
d
el.
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
is
v
er
y
lig
h
tweig
h
t
co
m
p
ar
ed
to
th
e
ef
f
icien
t
YOL
O
d
etec
to
r
s
,
as
s
h
o
wn
in
T
ab
le
s
2
an
d
3
.
T
h
e
o
r
ig
in
al
v
er
s
io
n
o
f
YOL
Ov
8
n
g
e
n
er
ates
2
.
5
tim
es
lar
g
er
th
an
o
u
r
p
r
o
p
o
s
ed
d
etec
to
r
.
Ou
r
d
etec
to
r
u
s
es
o
n
ly
1
.
4
6
tim
es
less
o
p
er
atio
n
u
s
ag
e.
Fu
r
th
er
m
o
r
e,
th
e
p
ar
am
eter
s
an
d
n
u
m
b
er
o
f
o
p
er
atio
n
s
a
r
e
wid
el
y
u
s
ed
b
y
YOL
Ov
3
tin
y
an
d
YOL
Ov
7
tin
y
.
T
h
u
s
,
t
h
is
also
r
e
q
u
ir
es
a
s
ig
n
if
ican
t
p
r
o
ce
s
s
in
g
d
ev
ice
m
em
o
r
y
wh
ile
wea
k
en
in
g
th
e
d
ata
p
r
o
ce
s
s
in
g
s
p
ee
d
.
T
h
e
s
p
ee
d
o
f
d
ata
p
r
o
ce
s
s
in
g
d
eter
m
in
es
th
e
r
eliab
ilit
y
o
f
a
m
eth
o
d
wh
e
n
im
p
lem
e
n
ted
o
n
a
d
ev
ice.
I
n
o
r
d
er
to
s
u
p
p
o
r
t
th
e
s
m
ar
t
I
o
T
s
y
s
tem
,
th
is
wo
r
k
ev
al
u
ates
th
e
p
r
o
p
o
s
ed
m
o
d
el
s
p
e
ed
o
n
an
NVI
DI
A
J
etso
n
Nan
o
with
a
VR
AM
o
f
4
GB
.
T
h
is
d
ev
ice
is
co
m
m
o
n
ly
u
s
ed
as
an
ed
g
e
d
ev
ice
f
o
r
i
n
tellig
en
t
s
y
s
tem
s
.
B
ased
o
n
th
e
g
r
ap
h
in
Fig
u
r
e
7
,
th
e
p
r
o
p
o
s
ed
m
o
d
el
with
l
ig
h
t
b
o
ttlen
ec
k
an
d
C
SE
ac
h
iev
ed
a
s
p
ee
d
o
f
1
0
.
6
4
FP
S
wh
en
test
ed
o
n
th
e
d
ev
ic
e.
C
o
m
p
ar
ed
to
th
e
li
g
h
tweig
h
t
m
o
d
el
YOL
Ov
8
h
,
it
is
r
ed
u
ce
d
b
y
2
5
.
4
%.
On
th
e
o
th
er
h
an
d
,
YOL
Ov
8
h
-
LB
-
C
SE
ac
h
iev
es
4
0
.
1
%
f
aster
co
m
p
ar
ed
to
YOL
Ov
3
-
tin
y
.
An
o
th
er
co
m
p
ar
is
o
n
is
th
at
th
e
p
r
o
p
o
s
ed
m
o
d
el
is
8
.
5
%
an
d
9
.
8
%
s
lo
wer
th
an
YOL
Ov
8
n
a
n
d
YOL
Ov
6
n
,
r
esp
ec
tiv
ely
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
15
,
No
.
5
,
Octo
b
e
r
20
25
:
5
0
3
1
-
5
0
4
4
5040
A
co
m
p
ar
is
o
n
with
th
e
m
o
s
t
p
o
p
u
lar
YOL
O
v
er
s
io
n
s
,
s
u
ch
a
s
th
e
YOL
Ov
5
n
,
s
h
o
ws
th
at
o
u
r
m
o
d
el
is
s
lo
wer
th
an
th
is
co
m
p
etito
r
'
s
,
wh
ich
is
o
u
r
m
o
d
el'
s
wea
k
n
ess
.
T
h
e
ef
f
icien
cy
o
f
th
e
n
u
m
b
er
o
f
p
ar
am
eter
s
an
d
co
m
p
u
tatio
n
al
co
m
p
lex
ity
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
o
u
tp
e
r
f
o
r
m
s
th
at
o
f
th
e
ef
f
icien
t
YOL
O
d
etec
to
r
s
.
Ho
wev
er
,
it
r
eq
u
ir
es
m
o
r
e
p
r
o
ce
s
s
in
g
m
em
o
r
y
.
I
t
is
d
u
e
to
d
ep
th
-
wis
e
o
p
er
atio
n
s
th
at
ap
p
ly
b
r
an
ch
in
g
o
p
er
atio
n
s
f
o
r
ea
ch
c
h
an
n
el.
Hen
ce
,
th
is
o
p
er
atio
n
r
eq
u
i
r
es
m
o
r
e
m
em
o
r
y
th
a
n
th
e
r
eg
u
lar
co
n
v
o
lu
tio
n
o
p
er
atio
n
.
On
t
h
e
o
th
er
h
a
n
d
,
th
e
s
p
ee
d
ac
h
iev
ed
b
y
o
u
r
d
etec
to
r
o
f
1
0
.
6
4
FP
S
is
f
ea
s
ib
le
f
o
r
s
m
o
o
th
o
p
er
atio
n
o
n
ed
g
e
d
ev
ices
t
h
at
s
u
p
p
o
r
t
I
o
T
in
tellig
en
t
s
y
s
tem
s
.
T
h
e
p
r
io
r
ity
em
p
h
as
izes
f
all
d
etec
tio
n
p
er
f
o
r
m
an
ce
th
at
m
i
n
im
izes f
a
ll a
ctiv
ity
r
ec
o
g
n
itio
n
er
r
o
r
s
.
T
ab
le
2
.
C
o
m
p
a
r
is
o
n
o
f
th
e
p
r
o
p
o
s
ed
d
etec
to
r
with
o
th
e
r
lig
h
tweig
h
t Y
OL
O
d
etec
to
r
s
an
d
atten
tio
n
m
ec
h
an
is
m
m
eth
o
d
s
o
n
th
e
Fallen
d
ataset.
I
t a
ls
o
co
m
p
ar
es t
h
e
n
u
m
b
er
o
f
p
ar
a
m
eter
s
an
d
co
m
p
u
tatio
n
al
co
m
p
lex
ity
o
f
m
o
d
els
M
o
d
e
l
s
G
F
LO
P
S
P
a
r
a
me
t
e
r
s
mA
P
@
0
.
5
mA
P
@
0
.
5
:
0
.
9
5
Y
O
LO
v
3
t
i
n
y
1
9
.
0
1
2
,
1
3
3
,
6
7
0
0
.
6
8
0
.
3
6
5
Y
O
LO
v
5
n
7
.
2
2
,
5
0
9
,
0
4
9
0
.
7
1
1
0
.
3
8
7
Y
O
LO
v
6
n
1
1
.
9
4
,
2
3
8
,
4
4
1
0
.
7
3
5
0
.
4
1
0
Y
O
LO
v
7
t
i
n
y
1
3
.
2
6
.
0
2
0
.
4
0
0
0
,
6
9
0
0
,
3
2
0
Y
O
LO
v
8
n
8
.
2
3
,
0
1
1
,
4
3
3
0
,
7
2
7
0
,
4
0
5
Y
O
LO
v
1
0
n
6
.
2
2
,
6
9
5
,
1
9
6
0
.
7
1
1
0
.
3
9
6
Y
O
LO
V
8
h
5
.
3
9
0
7
,
4
3
3
0
.
7
0
8
0
.
3
8
7
Y
O
LO
v
8
h
-
LB
5
.
4
9
1
7
.
4
1
7
0
.
7
1
2
0
.
3
8
7
Y
OLOv
8
h
-
LB
-
C
S
E
5
.
6
1
.
1
9
4
.
6
3
5
0
.
7
3
2
0
.
3
9
8
T
ab
le
3
.
C
o
m
p
a
r
is
o
n
o
f
th
e
p
r
o
p
o
s
ed
d
etec
to
r
with
o
th
e
r
lig
h
tweig
h
t Y
OL
O
d
etec
to
r
s
an
d
atten
tio
n
m
ec
h
an
is
m
m
eth
o
d
s
o
n
th
e
L
e
2
i d
ataset.
I
t a
ls
o
co
m
p
a
r
es th
e
n
u
m
b
e
r
o
f
p
ar
am
eter
s
an
d
c
o
m
p
u
tatio
n
al
co
m
p
lex
ity
o
f
m
o
d
els
M
o
d
e
l
s
G
F
LO
P
S
P
a
r
a
me
t
e
r
s
mA
P
@
0
.
5
mA
P
@
0
.
5
:
0
.
9
5
Y
O
LO
v
3
t
i
n
y
1
9
.
0
1
2
,
1
3
3
,
1
5
6
0
.
8
5
0
.
5
7
6
Y
O
LO
v
5
n
7
.
2
2
,
5
0
8
,
8
5
4
0
.
8
8
9
0
.
5
8
4
Y
O
LO
v
6
n
1
1
.
9
4
,
2
3
8
,
3
4
2
0
.
8
5
9
0
.
5
4
9
Y
O
LO
v
7
t
i
n
y
1
3
.
2
6
,
0
1
7
,
6
9
4
0
.
8
9
8
0
.
5
3
2
Y
O
LO
v
8
n
8
.
2
3
,
0
1
1
,
2
3
8
0
,
9
1
0
0
.
5
7
1
Y
O
LO
v
8
h
5
.
3
9
0
7
,
2
3
8
0
,
9
0
7
0
.
5
7
4
Y
O
LO
v
1
0
n
6
.
2
2
,
6
9
5
,
1
9
6
0
.
8
9
3
0
.
5
2
5
Y
O
LO
v
8
h
-
LB
5
.
4
9
1
7
,
2
2
2
0
,
9
0
7
0
.
5
7
7
Y
O
LO
v
8
h
-
LB
-
SE
5
.
8
1
,
1
8
2
,
6
0
0
0
.
9
1
8
0
.
5
6
9
Y
O
LO
v
8
h
-
LB
-
C
B
A
M
5
.
8
1
,
1
8
4
,
3
9
2
0
.
9
1
3
0
.
5
8
3
Y
O
LO
v
8
h
-
LB
-
C
A
N
5
.
8
1
,
1
8
5
,
5
9
8
0
.
8
9
0
.
5
7
1
Y
O
LO
v
8
h
-
LB
-
E
LA
5
.
8
1
,
1
8
4
,
5
5
0
0
.
9
0
6
0
.
5
9
4
Y
O
LO
v
8
h
-
LB
-
DAN
5
.
8
1
,
2
1
9
,
4
0
0
0
.
9
0
9
0
.
5
7
7
Y
O
LO
v
8
-
h
-
LB
-
G
C
N
ET
5
.
8
1
,
1
8
4
,
5
7
5
0
.
8
9
8
0
.
5
7
1
Y
O
LO
v
8
h
-
LB
-
E
C
A
5
.
8
1
,
1
8
2
,
2
4
9
0
.
8
7
6
0
.
5
7
Y
OLOv
8
h
-
LB
-
C
S
E
5
.
6
1
.
1
9
4
.
4
4
0
0
,
9
1
6
0
,
6
0
3
Fig
u
r
e
7
.
Sp
ee
d
co
m
p
ar
is
o
n
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
with
o
th
er
lig
h
tweig
h
t Y
OL
O
m
o
d
els.
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
ac
h
iev
es 1
0
.
6
4
FP
S f
aster
th
an
YOL
Ov
3
-
tin
y
Evaluation Warning : The document was created with Spire.PDF for Python.