I
AE
S In
t
er
na
t
io
na
l J
o
urna
l o
f
Art
if
icia
l In
t
ellig
ence
(
I
J
-
AI
)
Vo
l.
15
,
No
.
4
,
A
u
g
u
s
t
20
26
,
p
p
.
3
4
5
2
~
3
4
6
3
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
15
.i
4
.
p
p
3
4
5
2
-
3
4
6
3
3452
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
YO
L
O
v
1
1
optimiza
tion for t
iny
ob
ject in crow
ded sc
enes
H
us
na
Sa
rira
h H
us
i
n
1
,
H
o
wa
rd
Cho
ng
Yun H
a
o
1
,
P
a
n Y
ua
n F
ei
1
,
M
o
hs
en
M
a
rj
a
ni
1
,
Su
ria
na
I
s
m
a
il
2
1
S
c
h
o
o
l
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
,
F
a
c
u
l
t
y
o
f
I
n
n
o
v
a
t
i
o
n
a
n
d
Te
c
h
n
o
l
o
g
y
,
T
a
y
l
o
r
’
s Un
i
v
e
r
si
t
y
,
S
u
b
a
n
g
Ja
y
a
,
M
a
l
a
y
s
i
a
2
M
a
l
a
y
si
a
n
I
n
st
i
t
u
t
e
o
f
I
n
f
o
r
m
a
t
i
o
n
T
e
c
h
n
o
l
o
g
y
,
U
n
i
v
e
r
si
t
i
K
u
a
l
a
L
u
m
p
u
r
,
K
u
a
l
a
L
u
m
p
u
r
,
M
a
l
a
y
si
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
No
v
2
,
2
0
2
5
R
ev
is
ed
May
25
,
2
0
2
6
Acc
ep
ted
J
u
l
9
,
2
0
2
6
S
m
a
ll
o
b
jec
t
d
e
tec
ti
o
n
in
c
ro
w
d
e
d
u
rb
a
n
a
n
d
a
e
rial
sc
e
n
e
s
re
m
a
in
s
a
c
rit
ica
l
c
h
a
ll
e
n
g
e
d
u
e
t
o
li
m
it
e
d
p
ix
e
l
i
n
fo
rm
a
ti
o
n
a
n
d
i
n
fo
rm
a
ti
o
n
lo
ss
in
d
e
e
p
n
e
u
ra
l
n
e
tw
o
rk
s.
Th
is
stu
d
y
i
n
tr
o
d
u
c
e
s
a
n
o
v
e
l
o
p
t
imiz
a
ti
o
n
fra
m
e
wo
rk
fo
r
YO
LOv
1
1
,
s
p
e
c
ifi
c
a
ll
y
e
n
g
i
n
e
e
re
d
fo
r
ti
n
y
-
sc
a
le
targ
e
ts
b
y
i
n
teg
ra
ti
n
g
c
o
n
v
o
lu
ti
o
n
a
l
b
lo
c
k
a
tt
e
n
ti
o
n
m
o
d
u
les
(CBAM),
k
-
m
e
a
n
s
a
n
c
h
o
r
c
lu
ste
rin
g
,
a
n
d
a
n
e
n
h
a
n
c
e
d
fe
a
tu
re
p
y
ra
m
id
n
e
two
r
k
(F
P
N).
E
v
a
lu
a
te
d
o
n
th
e
Ti
n
y
P
e
rso
n
a
n
d
COCO
-
m
in
i
d
a
tas
e
ts,
th
e
YO
LOv
1
1
-
o
p
ti
m
ize
d
m
o
d
e
l
a
c
h
iev
e
s
sig
n
ifi
c
a
n
t
p
e
rfo
rm
a
n
c
e
b
re
a
k
th
r
o
u
g
h
s,
d
e
li
v
e
ri
n
g
a
+
7
.
3
%
g
a
i
n
in
m
e
a
n
a
v
e
ra
g
e
p
re
c
isio
n
(m
AP)
a
n
d
a
+
1
0
.
5
%
i
n
c
re
a
se
in
re
c
a
ll
o
v
e
r
th
e
b
a
se
li
n
e
.
No
tab
ly
,
t
h
e
m
o
d
e
l
a
c
h
iev
e
d
a
re
c
a
ll
o
f
0
.
0
7
2
o
n
th
e
Ti
n
y
P
e
rso
n
d
a
tas
e
t,
with
d
o
u
b
le
se
n
siti
v
it
y
o
f
sta
n
d
a
rd
YO
LO
v
1
1
.
Wi
th
a
h
ig
h
-
sp
e
e
d
in
fe
re
n
c
e
ra
te
o
f
2
7
.
3
F
P
S
,
t
h
is
re
se
a
rc
h
d
e
m
o
n
stra
tes
th
a
t
stra
teg
ic
a
rc
h
it
e
c
tu
ra
l
re
fin
e
m
e
n
ts
c
a
n
d
ra
stica
ll
y
imp
ro
v
e
sm
a
ll
o
b
jec
t
d
e
tec
ti
o
n
re
li
a
b
il
it
y
with
o
u
t
c
o
m
p
r
o
m
isin
g
re
a
l
-
ti
m
e
v
iab
il
it
y
o
n
e
d
g
e
d
e
v
ice
s.
K
ey
w
o
r
d
s
:
An
ch
o
r
b
o
x
o
p
tim
izatio
n
C
o
n
v
o
lu
tio
n
al
b
lo
ck
atten
tio
n
m
o
d
u
le
Data
au
g
m
en
tatio
n
Featu
r
e
p
y
r
am
i
d
n
etwo
r
k
T
in
y
o
b
ject
d
etec
tio
n
T
in
y
Per
s
o
n
d
ataset
YOL
Ov
1
1
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
:
Hu
s
n
a
Sar
ir
ah
Hu
s
in
Sch
o
o
l o
f
C
o
m
p
u
ter
Scien
ce
,
Facu
lty
o
f
I
n
n
o
v
atio
n
an
d
T
ec
h
n
o
lo
g
y
,
T
a
y
lo
r
’
s
Un
iv
e
r
s
ity
Su
b
an
g
J
ay
a,
Ma
lay
s
ia
E
m
ail:
h
u
s
n
a.
h
u
s
in
@
tay
lo
r
s
.
e
d
u
.
m
y
1.
I
NT
RO
D
UCT
I
O
N
Ob
ject
d
etec
tio
n
s
er
v
es
as
a
f
u
n
d
am
e
n
tal
p
illar
o
f
co
m
p
u
ter
v
is
io
n
,
en
a
b
lin
g
a
u
to
n
o
m
o
u
s
s
y
s
tem
s
to
id
en
tify
a
n
d
lo
ca
lize
s
p
ec
if
ic
en
titi
es
with
in
d
i
g
ital
i
m
ag
es
o
r
v
id
e
o
f
r
am
es.
Un
lik
e
s
im
p
le
im
a
g
e
class
if
icatio
n
,
wh
ich
ass
ig
n
s
a
s
in
g
le
lab
el
to
an
en
tire
s
c
en
e,
o
b
ject
d
etec
tio
n
r
e
q
u
ir
es
th
e
s
im
u
ltan
eo
u
s
ex
ec
u
tio
n
o
f
lo
ca
lizatio
n
wit
h
d
r
awin
g
p
r
ec
is
e
b
o
u
n
d
in
g
b
o
x
es,
an
d
class
if
icatio
n
th
at
is
as
s
ig
n
in
g
co
r
r
ec
t
lab
els
to
th
o
s
e
b
o
x
es
[
1
]
.
T
h
i
s
d
u
al
ca
p
ab
ilit
y
h
as
ca
taly
ze
d
th
e
d
ev
elo
p
m
e
n
t
o
f
in
tellig
e
n
t
s
y
s
tem
s
ac
r
o
s
s
a
v
ast
ar
r
ay
o
f
h
ig
h
-
s
tak
es
d
o
m
ain
s
.
I
n
au
t
o
n
o
m
o
u
s
d
r
i
v
in
g
,
it
is
cr
itical
f
o
r
id
en
tif
y
in
g
p
ed
estrian
s
an
d
tr
af
f
i
c
s
ig
n
s
[
2
]
;
in
m
ed
ical
im
ag
in
g
,
it
ass
is
ts
r
ad
io
lo
g
is
ts
in
s
p
o
ttin
g
m
icr
o
s
co
p
ic
lesi
o
n
s
o
r
tu
m
o
r
s
[
3
]
;
an
d
in
ae
r
ial
s
u
r
v
eillan
ce
,
it e
n
ab
les t
h
e
tr
ac
k
in
g
o
f
v
eh
icles o
r
in
f
r
astru
ctu
r
e
f
r
o
m
h
ig
h
altitu
d
es
[
2
]
,
[
4
]
.
Desp
ite
th
e
m
atu
r
ity
o
f
g
en
er
al
o
b
ject
d
etec
tio
n
,
id
en
tify
in
g
tin
y
o
b
jects
in
cr
o
wd
ed
s
ce
n
es
r
em
ain
s
o
n
e
o
f
th
e
m
o
s
t
f
o
r
m
id
a
b
le
ch
allen
g
es
in
th
e
f
ield
.
Sm
all
o
b
jects
ty
p
ically
o
cc
u
p
y
a
n
e
g
lig
ib
le
f
r
ac
tio
n
o
f
th
e
to
tal
p
ix
el
co
u
n
t,
o
f
ten
less
th
an
0
.
1
%
o
f
th
e
im
ag
e
ar
ea
t
h
at
will
r
esu
lt
in
a
s
ev
er
e
lac
k
o
f
d
is
cr
im
in
ativ
e
f
ea
tu
r
es.
W
h
en
th
ese
o
b
jects
ar
e
s
itu
ated
in
d
en
s
e,
cl
u
tter
ed
e
n
v
ir
o
n
m
en
ts
,
t
h
e
p
r
o
b
lem
is
f
u
r
t
h
er
co
m
p
o
u
n
d
e
d
b
y
in
te
r
-
o
b
ject
o
cc
lu
s
io
n
,
b
ac
k
g
r
o
u
n
d
in
te
r
f
er
en
ce
,
an
d
th
e
lo
s
s
o
f
f
i
n
e
-
g
r
ain
e
d
s
p
atial
in
f
o
r
m
atio
n
d
u
r
i
n
g
th
e
d
o
wn
-
s
am
p
lin
g
lay
er
s
o
f
d
ee
p
n
e
u
r
al
n
etwo
r
k
s
.
I
n
a
ty
p
ical
co
n
v
o
lu
tio
n
al
p
i
p
elin
e,
th
e
"sig
n
al"
o
f
a
tin
y
o
b
ject
is
o
f
ten
d
ilu
ted
o
r
en
tire
l
y
ex
tin
g
u
is
h
ed
b
y
th
e
tim
e
th
e
d
ata
r
ea
ch
es
th
e
d
ee
p
er
s
em
an
tic
lay
er
s
,
lead
in
g
to
h
ig
h
r
ates o
f
f
alse n
e
g
ativ
es a
n
d
p
o
o
r
l
o
ca
lizatio
n
ac
cu
r
ac
y
[
5
]
.
R
ec
en
t
p
r
o
g
r
ess
in
m
ac
h
i
n
e
l
ea
r
n
in
g
h
as
s
o
u
g
h
t
to
b
r
id
g
e
t
h
is
g
ap
th
r
o
u
g
h
in
n
o
v
ativ
e
ar
ch
itectu
r
al
s
h
if
ts
an
d
ad
v
an
ce
d
lear
n
in
g
s
tr
ateg
ies.
T
h
e
y
ea
r
s
2
0
2
4
th
r
o
u
g
h
2
0
2
6
h
av
e
s
ee
n
a
s
u
r
g
e
in
m
u
lti
-
s
ca
le
f
ea
tu
r
e
f
u
s
io
n
tech
n
iq
u
es,
s
u
ch
as
b
id
ir
ec
tio
n
al
f
ea
tu
r
e
p
y
r
am
id
n
etwo
r
k
s
(
B
iFP
N)
[
6
]
,
[
7
]
an
d
p
ath
ag
g
r
eg
atio
n
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
YOLOv1
1
o
p
timiz
a
tio
n
fo
r
tin
y
o
b
ject
in
cro
w
d
ed
s
ce
n
es
(
H
u
s
n
a
S
a
r
ir
a
h
Hu
s
in
)
3453
n
etwo
r
k
s
(
PANet)
[
8
]
,
[
9
]
,
wh
ich
aim
to
p
r
eser
v
e
h
ig
h
-
r
eso
lu
tio
n
s
p
atial
d
etails
f
r
o
m
s
h
allo
w
la
y
er
s
.
Fu
r
th
er
m
o
r
e
,
th
e
in
te
g
r
atio
n
o
f
atten
tio
n
m
ec
h
an
is
m
s
an
d
v
is
io
n
tr
an
s
f
o
r
m
er
s
(
ViT
s
)
[
1
0
]
,
[
1
1
]
h
as
allo
wed
m
o
d
els to
b
etter
ca
p
tu
r
e
g
l
o
b
a
l c
o
n
tex
t,
h
el
p
in
g
to
d
is
tin
g
u
is
h
tin
y
tar
g
ets f
r
o
m
c
o
m
p
lex
b
ac
k
g
r
o
u
n
d
s
.
T
h
e
y
o
u
o
n
ly
lo
o
k
o
n
ce
(
YOL
O)
f
am
ily
o
f
d
etec
to
r
s
,
r
en
o
wn
ed
f
o
r
its
r
ea
l
-
tim
e
p
r
o
ce
s
s
in
g
s
p
ee
d
s
,
h
as
ev
o
lv
ed
s
ig
n
if
ican
tly
t
o
ad
d
r
ess
th
ese
lim
itatio
n
s
.
Mo
d
er
n
iter
atio
n
s
,
s
u
ch
as
YOL
Ov
8
an
d
its
s
u
cc
ess
o
r
s
,
h
av
e
in
tr
o
d
u
ce
d
s
p
ec
ialized
d
etec
tio
n
h
ea
d
s
f
o
r
h
ig
h
-
r
es
o
lu
tio
n
f
ea
t
u
r
e
m
a
p
s
an
d
o
p
tim
ized
an
ch
o
r
-
b
o
x
s
tr
ateg
ies
s
p
ec
if
ically
tailo
r
ed
f
o
r
s
m
all
-
s
ca
le
tar
g
ets.
Dete
ctin
g
tin
y
o
b
jects
in
cr
o
wd
ed
s
ce
n
es
is
cr
itically
im
p
o
r
tan
t
f
o
r
s
af
ety
,
s
ec
u
r
it
y
,
an
d
a
d
v
an
ce
d
m
o
n
ito
r
in
g
s
y
s
tem
s
,
with
YOL
O
em
er
g
in
g
as
a
p
o
wer
f
u
l
s
o
lu
tio
n
to
th
is
ch
allen
g
in
g
co
m
p
u
ter
v
is
io
n
p
r
o
b
lem
.
T
h
e
s
ig
n
if
ican
ce
lies
in
o
v
er
co
m
in
g
k
ey
ch
allen
g
es
s
u
ch
as
d
is
tin
g
u
is
h
in
g
b
ac
k
g
r
o
u
n
d
f
r
o
m
o
b
ject
f
ea
tu
r
es
an
d
ex
tr
ac
tin
g
s
m
all
-
s
ca
le
tar
g
et
f
ea
tu
r
es
in
co
m
p
lex
en
v
ir
o
n
m
en
ts
[
1
2
]
.
R
ec
en
t
Y
OL
O
d
ev
elo
p
m
en
ts
h
av
e
m
ad
e
s
u
b
s
tan
tial
p
r
o
g
r
ess
:
f
o
r
in
s
tan
ce
,
th
e
SF
-
YOL
O
f
r
am
ewo
r
k
in
tr
o
d
u
ce
s
a
s
p
atial
in
f
o
r
m
atio
n
p
er
ce
p
tio
n
m
o
d
u
le
th
at
d
y
n
am
ically
ad
ju
s
ts
r
ec
ep
tiv
e
f
ield
s
to
en
h
an
ce
o
b
ject
d
if
f
e
r
en
tiatio
n
[
1
3
]
.
R
esear
ch
er
s
h
av
e
d
em
o
n
s
tr
ated
im
p
r
ess
iv
e
im
p
r
o
v
em
e
n
ts
,
with
s
o
m
e
ad
ap
tiv
e
YOL
O
m
o
d
els
ac
h
iev
in
g
u
p
to
0
.
8
7
2
m
ea
n
a
v
er
ag
e
p
r
ec
is
io
n
(
m
AP)
a
n
d
8
7
.
5
%
p
r
ec
is
io
n
[
1
2
]
.
T
h
ese
a
d
v
an
ce
s
ar
e
cr
u
cial
f
o
r
ap
p
licatio
n
s
lik
e
s
u
r
v
eilla
n
ce
,
au
to
n
o
m
o
u
s
d
r
iv
i
n
g
,
an
d
r
em
o
te
s
en
s
in
g
,
wh
er
e
ac
cu
r
at
ely
d
etec
tin
g
s
m
all
o
b
jects
in
d
en
s
e
s
ce
n
es
ca
n
m
ea
n
th
e
d
if
f
er
e
n
ce
b
etwe
en
p
r
ev
en
tio
n
an
d
d
is
aster
.
T
h
is
wo
r
k
is
th
e
f
ir
s
t
to
co
m
b
in
e
co
n
v
o
lu
tio
n
al
b
lo
ck
atten
tio
n
m
o
d
u
le
(
C
B
AM
)
with
YOL
Ov
1
1
to
f
it
th
e
n
ee
d
s
o
f
p
ed
estrian
-
lev
el
s
m
all
o
b
ject
d
etec
tio
n
o
n
th
e
T
in
y
Per
s
o
n
d
ataset.
T
h
e
ad
v
an
ce
m
en
t
o
f
d
ee
p
lear
n
in
g
h
a
s
h
ad
a
s
u
b
s
tan
tial
im
p
ac
t
o
n
o
b
ject
d
etec
tio
n
s
y
s
tem
s
.
Ho
wev
er
,
tellin
g
s
m
all
o
b
jects
in
clu
tter
ed
v
is
u
al
s
c
en
es
is
a
p
r
o
b
lem
.
T
h
is
is
o
f
p
ar
ticu
lar
co
n
ce
r
n
to
th
e
ap
p
licatio
n
s
o
f
s
m
aller
tar
g
ets,
e.
g
.
s
u
r
v
eillan
ce
,
t
r
af
f
ic
co
n
tr
o
l,
an
d
p
ed
estrian
d
etec
tio
n
.
YOL
O
h
as
b
ee
n
g
en
er
ally
c
r
ed
ited
as
b
ein
g
f
ast
an
d
ef
f
ici
en
t
in
d
etec
tio
n
.
H
o
wev
er
,
it
f
r
eq
u
e
n
tly
f
ails
to
wo
r
k
well
in
d
etec
tin
g
s
m
all
o
b
jects
b
ec
au
s
e
o
f
th
e
r
estrictio
n
o
f
an
ch
o
r
b
o
x
d
esig
n
an
d
in
f
o
r
m
atio
n
lo
s
s
in
p
r
o
ce
s
s
in
g
d
ee
p
lay
er
s
.
W
ith
o
u
t
o
p
tim
izatio
n
,
YOL
Ov
8
attain
ed
3
4
.
7
%
m
AP
o
n
s
m
all
o
b
jects
(
b
elo
w
3
2
×3
2
p
ix
els),
d
e
m
o
n
s
tr
atin
g
th
e
in
tr
in
s
ic
ch
allen
g
e
o
f
s
m
all
o
b
ject
d
etec
tio
n
[
1
4
]
.
On
th
e
s
am
e
n
o
te,
s
tan
d
ar
d
YOL
Ov
9
attain
ed
o
n
ly
5
3
.
1
%
m
AP
o
n
T
in
y
Per
s
o
n
d
ataset,
wh
ich
is
m
u
ch
b
e
lo
w
its
p
er
f
o
r
m
an
ce
(
6
7
.
5
%
m
A
P
)
o
n
n
o
r
m
a
l
-
s
iz
e
d
o
b
j
e
c
ts
,
w
h
ic
h
u
n
d
e
r
s
c
o
r
es
t
h
e
i
m
p
o
r
t
a
n
c
e
o
f
a
r
c
h
i
t
ec
t
u
r
e
-
l
ev
e
l
a
d
j
u
s
t
m
e
n
t
[
1
5
]
.
W
h
ile
YOL
Ov
1
1
p
er
f
o
r
m
s
r
o
b
u
s
tly
in
m
an
y
s
ce
n
ar
io
s
,
it
co
n
tin
u
es
to
en
co
u
n
ter
d
if
f
icu
lties
with
s
m
aller
o
b
ject
ca
teg
o
r
ies.
Alth
o
u
g
h
tech
n
iq
u
es
lik
e
f
ea
tu
r
e
p
y
r
am
i
d
s
[
1
6
]
,
[
1
7
]
an
d
atten
tio
n
m
ec
h
an
is
m
s
[
1
7
]
h
a
v
e
s
h
o
wn
p
r
o
m
is
e,
th
ey
ar
e
n
o
t
alwa
y
s
f
u
lly
lev
e
r
ag
ed
o
r
f
i
n
e
-
tu
n
ed
f
o
r
s
m
all
o
b
ject
s
en
s
itiv
ity
.
T
h
is
s
tu
d
y
d
ir
ec
tly
ad
d
r
ess
es
th
ese
lim
itatio
n
s
b
y
s
y
s
tem
atica
lly
tu
n
in
g
YOL
Ov
1
1
f
o
r
tin
y
o
b
ject
d
etec
tio
n
,
f
o
cu
s
in
g
o
n
o
p
tim
izin
g
an
c
h
o
r
co
n
f
ig
u
r
atio
n
,
f
ea
tu
r
e
e
n
h
an
ce
m
en
t,
an
d
au
g
m
en
tatio
n
s
tr
ateg
ies
to
im
p
r
o
v
e
p
er
f
o
r
m
an
ce
in
d
en
s
e
an
d
co
m
p
le
x
en
v
i
r
o
n
m
en
ts
s
u
ch
as p
ed
estrian
-
h
ea
v
y
u
r
b
a
n
s
ce
n
es.
T
h
is
wo
r
k
is
im
p
o
r
ta
n
t
s
in
ce
it
f
o
cu
s
es
o
n
o
n
e
o
f
th
e
b
i
g
g
est
is
s
u
es
o
f
co
m
p
u
ter
v
is
io
n
,
w
h
ich
is
th
e
d
etec
tio
n
o
f
s
m
all
o
b
jects.
Sm
all
o
b
ject
d
etec
tio
n
ca
p
a
b
ilit
y
h
as
lo
n
g
-
ter
m
in
ter
est
ac
r
o
s
s
m
an
y
s
ec
to
r
s
,
s
u
ch
as
s
u
r
v
eillan
ce
,
s
elf
-
d
r
iv
in
g
c
ar
s
,
m
ed
ical
im
ag
in
g
,
an
d
ae
r
ial
r
ec
o
n
n
aiss
an
ce
.
T
h
e
p
o
ten
tial
co
n
tr
ib
u
tio
n
o
f
th
e
cu
r
r
e
n
t
s
tu
d
y
is
th
at
it
p
r
o
p
o
s
es
an
in
teg
r
ate
d
s
et
o
f
o
p
tim
izatio
n
s
tar
g
etin
g
YOL
O
v
1
1
,
wh
ic
h
b
r
i
n
g
s
+7
.
3
%
m
AP
an
d
+1
0
.
5
%
r
ec
a
ll
o
n
th
e
T
in
y
Per
s
o
n
d
ataset,
an
d
s
u
ch
im
p
r
o
v
em
en
ts
wer
e
n
o
t
r
ep
o
r
ted
in
th
e
liter
atu
r
e
b
ef
o
r
e.
T
h
is
r
esear
ch
will
o
p
tim
ize
an
ch
o
r
b
o
x
s
ettin
g
s
,
u
s
e
p
r
o
g
r
ess
iv
e
d
ata
au
g
m
en
tatio
n
s
tr
ateg
ies,
an
d
en
h
an
ce
f
ea
t
u
r
e
ex
tr
ac
tio
n
lay
er
s
to
in
cr
ea
s
e
th
e
ac
cu
r
ac
y
an
d
r
ec
all
o
f
th
e
YOL
Ov
1
1
s
m
all
o
b
ject
d
etec
to
r
.
T
h
is
wo
u
ld
b
e
ess
en
tial
in
r
ea
l
-
tim
e
a
p
p
lic
atio
n
s
wh
er
e
q
u
ick
an
d
co
r
r
ec
t
d
etec
tio
n
o
f
s
m
all
o
b
jects
co
u
ld
h
av
e
a
b
i
g
in
f
lu
en
ce
in
d
ec
is
io
n
-
m
ak
i
n
g
m
ec
h
an
is
m
s
lik
e
p
ed
estrian
d
etec
tio
n
in
au
to
n
o
m
o
u
s
v
eh
icles a
n
d
in
s
ec
u
r
ity
ch
ec
k
u
p
s
in
cr
o
wd
e
d
ar
ea
s
.
Un
lik
e
p
r
io
r
wo
r
k
th
at
ap
p
lies
au
g
m
en
tatio
n
o
r
a
n
ch
o
r
tu
n
in
g
in
d
ep
en
d
en
tly
,
th
is
r
esear
ch
p
r
o
p
o
s
es
a
u
n
if
ied
YOL
Ov
1
1
o
p
tim
izat
io
n
f
r
am
ewo
r
k
th
at
co
m
b
in
es
an
ch
o
r
b
o
x
r
ec
o
n
f
ig
u
r
atio
n
,
s
p
atial
atten
tio
n
,
an
d
s
u
p
er
-
r
eso
lu
tio
n
-
a
c
o
m
b
in
at
io
n
n
o
t
p
r
ev
i
o
u
s
ly
ev
alu
ate
d
i
n
liter
atu
r
e.
Mo
r
eo
v
e
r
,
th
e
s
t
u
d
y
will
s
er
v
e
as
a
u
s
ef
u
l
in
p
u
t
in
ter
m
s
o
f
o
p
ti
m
izin
g
th
e
YOL
O
m
o
d
els
i
n
d
etec
tin
g
s
m
all
o
b
jects,
w
h
ich
ca
n
p
r
esen
t
a
p
r
ac
tical
g
u
i
d
elin
e
to
th
e
r
ese
ar
ch
er
s
an
d
in
d
u
s
tr
ial
p
r
ac
titi
o
n
er
s
aim
in
g
to
im
p
lem
e
n
t
t
h
e
o
b
ject
d
etec
tio
n
m
o
d
el
in
p
r
ac
tical
ap
p
licatio
n
s
.
T
h
e
co
m
p
ar
ativ
e
s
tu
d
y
with
f
aster
r
eg
io
n
-
b
ased
c
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
(
Fas
ter
R
-
C
NN
)
an
d
R
etin
aN
et
will
al
s
o
h
elp
to
cr
ea
te
a
b
en
ch
m
ar
k
to
ass
ess
th
e
im
p
r
o
v
em
en
t
in
th
e
s
m
all
o
b
jects d
etec
tio
n
an
d
cr
ea
te
a
b
asis
o
f
f
u
r
th
er
r
esear
ch
in
th
i
s
f
ield
.
2.
RE
L
AT
E
D
WO
RK
T
h
e
d
etec
tio
n
o
f
s
m
all
o
b
jects
is
o
n
e
o
f
th
e
m
o
s
t
co
m
p
licated
task
s
in
co
m
p
u
ter
v
is
io
n
,
esp
ec
ially
in
s
u
ch
ap
p
licatio
n
s
as
s
u
r
v
eillan
ce
s
y
s
tem
s
,
s
elf
-
d
r
i
v
in
g
v
eh
i
cles,
an
d
an
aly
s
is
o
f
ae
r
ial
im
ag
es
[
1
8
]
.
T
h
e
co
r
e
ch
allen
g
e
is
th
at
s
m
all
o
b
ject
s
h
av
e
in
s
u
f
f
icien
t
p
ix
el
in
f
o
r
m
atio
n
,
wh
ich
ca
u
s
es
f
ea
tu
r
e
lo
s
s
wh
en
d
o
wn
s
am
p
lin
g
th
e
n
etwo
r
k
a
n
d
f
in
a
lly
lead
s
to
lo
w
d
etec
tio
n
p
er
f
o
r
m
an
ce
[
1
9
]
.
I
n
th
is
liter
atu
r
e
r
ev
iew,
th
e
r
ec
en
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
15
,
No
.
4
,
Au
g
u
s
t
20
26
:
3
4
5
2
-
3
4
6
3
3454
d
ev
elo
p
m
e
n
ts
r
eg
ar
d
in
g
tin
y
o
b
ject
d
etec
tio
n
s
p
ec
if
ically
,
an
d
YOL
O
-
b
ased
ar
ch
itectu
r
e
an
d
o
p
tim
izatio
n
s
tr
ateg
ies ap
p
lied
to
th
e
T
in
y
Per
s
o
n
an
d
c
o
m
m
o
n
o
b
jects in
co
n
tex
t (
C
OC
O
)
m
in
i d
ataset
ar
e
d
is
cu
s
s
ed
.
2
.
1
.
Co
m
pa
ra
t
iv
e
perf
o
r
m
a
nce
o
f
det
ec
t
io
n mo
dels
R
ec
en
t
r
esear
ch
h
as
s
u
g
g
este
d
en
h
a
n
cin
g
v
ar
io
u
s
o
b
ject
d
etec
tio
n
f
r
am
ewo
r
k
s
.
T
ab
le
1
p
r
esen
ts
s
ig
n
if
ican
t
p
er
f
o
r
m
a
n
ce
o
u
tco
m
es
o
f
p
o
p
u
lar
m
o
d
els
o
n
p
o
p
u
lar
d
atasets
d
ev
o
ted
to
tin
y
o
b
jects
d
etec
tio
n
.
Desp
ite
s
tead
y
im
p
r
o
v
em
en
ts
,
co
n
f
lictin
g
f
in
d
in
g
s
an
d
tr
a
d
e
-
o
f
f
s
r
em
ain
p
r
ev
alen
t
ac
r
o
s
s
p
u
b
lis
h
ed
wo
r
k
s
.
MS
-
YOL
Ov
7
d
em
o
n
s
tr
ates
s
t
r
o
n
g
r
ec
all
b
u
t
s
tr
u
g
g
les
with
o
cc
lu
s
io
n
an
d
tig
h
t
s
p
atial
cl
u
tter
,
ca
u
s
in
g
f
alse
p
o
s
itiv
es
an
d
d
ec
r
ea
s
ed
p
r
e
cisi
o
n
in
cr
o
wd
s
ce
n
ar
io
s
[
2
0
]
.
W
h
ile
tr
an
s
f
o
r
m
er
-
in
teg
r
ated
ar
ch
itectu
r
es
en
h
an
ce
co
n
tex
t
u
al
awa
r
en
es
s
f
o
r
tin
y
o
b
jects
b
y
ca
p
tu
r
i
n
g
lo
n
g
-
r
an
g
e
d
ep
en
d
en
cies,
t
h
ey
o
f
ten
i
n
tr
o
d
u
ce
s
ig
n
if
ican
t
co
m
p
u
tatio
n
al
laten
cy
[
1
0
]
.
T
h
is
h
ig
h
in
f
er
en
ce
co
s
t
m
ak
es
p
u
r
e
ViT
ap
p
r
o
ac
h
es
less
s
u
itab
le
f
o
r
tim
e
-
cr
itical
ap
p
licatio
n
s
,
s
u
ch
as
au
to
n
o
m
o
u
s
d
r
o
n
es
o
r
h
ig
h
-
s
p
ee
d
v
eh
icle
d
etec
tio
n
,
wh
er
e
r
ea
l
-
tim
e
th
r
o
u
g
h
p
u
t
is
m
an
d
ato
r
y
[
1
1
]
.
C
o
n
s
eq
u
en
tly
,
o
p
tim
ized
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
(
C
NN
)
-
b
ased
m
o
d
els
lik
e
YOL
Ov
1
1
[
2
1
]
,
[
2
2
]
r
em
ain
p
r
ef
er
r
ed
c
h
o
ice
f
o
r
b
alan
c
in
g
s
en
s
itiv
ity
an
d
ed
g
e
-
d
e
p
lo
y
m
en
t v
iab
ilit
y
.
T
ab
le
1
.
C
o
m
p
a
r
ativ
e
p
er
f
o
r
m
an
ce
o
f
d
etec
tio
n
m
o
d
els
M
o
d
e
l
Ti
n
y
o
b
j
e
c
t
mA
P
D
a
t
a
s
e
t
R
e
c
a
l
l
(
%)
FPS
S
t
r
e
n
g
t
h
s
Li
mi
t
a
t
i
o
n
s
R
e
f
e
r
e
n
c
e
s
Y
O
LO
v
8
3
4
.
7
0
%
C
O
C
O
(
s
ma
l
l
<
3
2
×
32)
6
1
.
3
0
45
F
a
st
b
a
s
e
l
i
n
e
;
e
a
s
y
t
o
e
x
t
e
n
d
P
o
o
r
l
o
c
a
l
i
z
a
t
i
o
n
o
n
sma
l
l
t
a
r
g
e
t
s
[
1
4
]
D
e
n
se
-
Y
O
LO
v
5
+
7
.
2
%
o
v
e
r
Y
O
LO
v
5
C
O
C
O
-
42
D
e
n
se
sk
i
p
-
c
o
n
n
e
c
t
i
o
n
s
b
o
o
s
t
s
h
a
l
l
o
w
f
e
a
t
u
r
e
s
B
e
n
e
f
i
t
s
t
a
p
e
r
o
f
f
o
n
d
e
n
s
e
s
c
e
n
e
s
[
2
3
]
MS
-
Y
O
LO
v
7
7
6
.
8
%
(
R
e
c
a
l
l
)
D
r
o
n
e
s
u
r
v
e
i
l
l
a
n
c
e
1
5
.
5
38
M
u
l
t
i
-
sca
l
e
f
u
si
o
n
+
a
t
t
e
n
t
i
o
n
W
e
a
k
e
r
i
n
o
c
c
l
u
d
i
n
g
o
v
e
r
l
a
p
p
i
n
g
r
e
g
i
o
n
s
[
2
0
]
TP
-
Y
O
LO
(
Y
O
LO
v
9
)
6
2
.
4
%
(
f
r
o
m
5
3
.
1
%)
Ti
n
y
P
e
r
s
o
n
6
8
.
2
0
30
S
p
a
t
i
a
l
t
r
a
n
sf
o
r
mers
i
mp
r
o
v
e
d
f
i
n
e
-
g
r
a
i
n
f
o
c
u
s
M
o
d
e
r
a
t
e
i
n
c
r
e
a
se
i
n
t
r
a
i
n
i
n
g
t
i
me
[
1
5
]
Y
O
LO
v
1
0
+
a
d
a
p
t
i
v
e
r
e
c
e
p
t
i
v
e
f
i
e
l
d
mo
d
u
l
e
s
+
8
.
6
%
o
v
e
r
b
a
se
l
i
n
e
A
e
r
i
a
l
v
e
h
i
c
l
e
s
-
38
A
d
a
p
t
i
v
e
r
e
c
e
p
t
i
v
e
f
i
e
l
d
s
su
i
t
sm
a
l
l
sc
a
l
e
La
c
k
o
f
g
l
o
b
a
l
c
o
n
t
e
x
t
m
o
d
e
l
i
n
g
[
2
4
]
2
.
2
.
Adv
a
nces in
y
o
u o
nly
lo
o
k
o
nce
-
ba
s
ed
s
m
a
ll o
bje
c
t
det
ec
t
io
n
YOL
O
ar
ch
itectu
r
e
h
as
ev
o
lv
ed
s
ig
n
if
ican
tly
t
o
ad
d
r
e
s
s
s
m
all
o
b
ject
d
etec
tio
n
ch
allen
g
es.
No
tab
ly
,
YOL
Ov
8
ac
h
iev
e
d
o
n
ly
3
4
.
7
%
m
AP
o
n
o
b
je
cts
s
m
aller
th
an
3
2
×3
2
p
ix
e
ls
with
o
u
t
s
p
ec
if
ic
o
p
tim
izatio
n
s
,
h
i
g
h
lig
h
tin
g
th
e
n
ee
d
f
o
r
tar
g
eted
im
p
r
o
v
em
en
ts
f
o
r
tin
y
o
b
jects
[
1
4
]
.
Den
s
e
-
YOL
Ov
5
,
wh
ic
h
in
co
r
p
o
r
ated
d
en
s
e
c
o
n
n
ec
tio
n
s
b
etwe
en
d
etec
tio
n
lay
er
s
to
en
h
a
n
ce
f
ea
tu
r
e
p
r
o
p
ag
ati
o
n
s
p
ec
if
ically
f
o
r
s
m
all
o
b
jects,
ac
h
iev
in
g
a
7
.
2
%
im
p
r
o
v
e
m
en
t
in
m
AP
f
o
r
s
m
all
o
b
jects
co
m
p
ar
ed
to
t
h
e
b
aselin
e
YOL
Ov
5
o
n
th
e
C
OC
O
d
atase
t
[
2
3
]
.
Similar
ly
,
MS
-
YOL
Ov
7
in
teg
r
atin
g
m
u
lti
-
s
ca
le
f
ea
tu
r
e
f
u
s
io
n
with
ch
an
n
el
atten
tio
n
,
r
esu
ltin
g
in
a
s
ig
n
if
ican
t
im
p
r
o
v
em
en
t
in
tin
y
o
b
ject
r
ec
all
r
ates
f
r
o
m
6
1
.
3
%
t
o
7
6
.
8
%
o
n
d
r
o
n
e
s
u
r
v
eillan
ce
d
atasets
[
2
0
]
.
Fo
r
s
p
ec
if
ic
tin
y
p
er
s
o
n
d
et
ec
tio
n
,
s
tan
d
ar
d
YOL
Ov
9
a
ch
iev
ed
o
n
ly
5
3
.
1
%
m
AP
o
n
th
e
tin
y
T
in
y
Per
s
o
n
p
er
s
o
n
d
ataset,
s
ig
n
if
ican
tly
l
o
wer
th
a
n
its
6
7
.
5
%
m
AP
o
n
n
o
r
m
al
-
s
ized
o
b
jects.
T
h
eir
m
o
d
if
ie
d
TP
-
YOL
O
in
co
r
p
o
r
atin
g
s
p
at
ial
tr
an
s
f
o
r
m
er
n
etwo
r
k
s
im
p
r
o
v
ed
th
is
to
6
2
.
4
%
m
AP
b
y
en
h
an
cin
g
s
p
atial
f
ea
tu
r
e
lear
n
in
g
ca
p
ab
ilit
ies
[
1
5
]
.
R
ec
en
t
wo
r
k
o
n
YOL
Ov
1
0
in
tr
o
d
u
ce
d
ad
a
p
tiv
e
r
ec
ep
tiv
e
f
ield
m
o
d
u
les
s
p
ec
if
ically
d
esig
n
ed
f
o
r
tin
y
o
b
jects,
d
em
o
n
s
tr
atin
g
th
a
t
ca
r
ef
u
l
r
ec
ep
tiv
e
f
ield
m
o
d
u
latio
n
ca
n
b
o
o
s
t
d
etec
tio
n
r
ates f
o
r
o
b
jects u
n
d
er
2
0
×2
0
p
ix
els b
y
u
p
to
8
.
6
% in
ch
allen
g
in
g
s
ce
n
ar
io
s
[
2
4
]
.
2
.
3
.
Da
t
a
a
ug
m
ent
a
t
io
n str
a
t
eg
ies f
o
r
t
iny
o
bje
c
t
det
ec
t
io
n
D
a
t
a
au
g
m
e
n
t
a
t
io
n
i
s
a
c
r
it
i
c
a
l
an
d
ef
f
e
c
t
iv
e
s
t
r
a
t
eg
y
f
o
r
i
m
p
r
o
v
in
g
t
i
n
y
o
b
je
c
t
d
e
te
c
t
i
o
n
p
e
r
f
o
r
m
a
n
c
e
,
w
i
th
te
c
h
n
i
q
u
e
s
d
e
m
o
n
s
t
r
a
t
i
n
g
s
i
g
n
i
f
i
ca
n
t
a
c
c
u
r
a
cy
g
a
i
n
s
ac
r
o
s
s
m
u
l
t
i
p
le
d
o
m
a
in
s
.
R
e
s
e
a
r
c
h
er
s
h
a
v
e
d
e
v
e
lo
p
e
d
s
e
v
er
a
l
s
o
p
h
i
s
t
i
c
a
t
ed
a
u
g
m
e
n
t
a
t
i
o
n
a
p
p
r
o
a
ch
e
s
s
u
c
h
a
f
u
l
l
p
ip
e
l
i
n
e
u
s
i
n
g
g
e
n
er
a
t
i
v
e
a
d
v
e
r
s
ar
i
a
l
n
e
t
w
o
r
k
s
(
G
A
N
s
)
t
h
a
t
im
p
r
o
v
e
d
s
m
a
l
l
o
b
j
ec
t
d
e
t
ec
t
i
o
n
p
er
f
o
r
m
an
c
e
b
y
u
p
t
o
1
1
.
9
%
o
n
s
o
m
e
d
a
t
a
s
e
t
s
[
2
5
]
.
M
o
s
a
i
c
au
g
m
e
n
t
a
t
io
n
en
h
an
c
e
s
t
h
e
d
e
t
e
c
t
io
n
o
f
s
m
a
l
l
t
ar
g
e
t
s
b
y
co
m
b
i
n
in
g
m
u
l
t
i
p
le
im
a
g
e
s
,
e
f
f
e
c
t
i
v
el
y
i
n
cr
e
a
s
in
g
t
h
e
v
a
r
ie
t
y
o
f
s
c
a
l
e
s
s
e
e
n
d
u
r
i
n
g
tr
a
in
i
n
g
[
2
6
]
.
C
o
p
y
-
p
a
s
t
e
a
u
g
m
e
n
t
a
t
io
n
h
a
s
p
r
o
v
en
h
i
g
h
l
y
e
f
f
e
c
t
i
v
e
f
o
r
d
a
t
a
s
e
t
s
co
n
t
a
i
n
i
n
g
s
m
a
l
l
o
b
j
ec
t
s
,
s
u
c
h
as
T
i
n
y
P
e
r
s
o
n
[
2
7
]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
YOLOv1
1
o
p
timiz
a
tio
n
fo
r
tin
y
o
b
ject
in
cro
w
d
ed
s
ce
n
es
(
H
u
s
n
a
S
a
r
ir
a
h
Hu
s
in
)
3455
B
y
d
u
p
l
ic
a
t
i
n
g
t
in
y
o
b
j
ec
t
in
s
t
a
n
c
e
s
an
d
p
l
a
c
in
g
t
h
em
i
n
to
v
a
r
ie
d
b
ac
k
g
r
o
u
n
d
s
,
t
h
e
m
o
d
e
l
i
s
e
x
p
o
s
e
d
to
a
h
i
g
h
e
r
f
r
eq
u
en
cy
o
f
s
m
a
l
l
-
s
ca
l
e
f
e
a
tu
r
e
s
d
u
r
i
n
g
t
r
a
i
n
in
g
[
2
7
]
.
A
n
o
t
h
e
r
s
t
u
d
y
r
ev
i
e
w
ed
t
h
e
s
e
ap
p
r
o
a
ch
e
s
,
h
i
g
h
l
i
g
h
t
i
n
g
t
h
e
ir
p
o
t
e
n
t
ia
l
t
o
a
d
d
r
e
s
s
c
h
a
l
l
en
g
e
s
l
i
k
e
l
i
m
i
te
d
d
a
ta
s
e
t
a
v
a
i
la
b
i
l
i
ty
a
n
d
l
o
w
o
b
j
e
c
t
-
to
-
i
m
ag
e
r
a
t
i
o
s
in
t
in
y
o
b
je
c
t
d
e
t
e
c
t
io
n
[
2
8
]
.
2
.
4
.
F
e
a
t
ure
ex
t
r
a
ct
io
n a
nd
a
t
t
ent
io
n m
ec
ha
nis
m
s
f
o
r
s
m
a
ll o
bje
ct
det
ec
t
io
n
Stan
d
ar
d
d
ee
p
lear
n
i
n
g
d
etec
t
o
r
s
lik
e
YOL
O
s
u
f
f
er
f
r
o
m
s
ig
n
if
ican
t
f
ea
tu
r
e
lo
s
s
f
o
r
tin
y
o
b
jects
d
u
e
to
ag
g
r
ess
iv
e
d
o
wn
s
am
p
lin
g
o
p
er
atio
n
s
,
wh
ich
e
r
o
d
e
th
e
s
p
atial
d
etails
n
ec
ess
ar
y
f
o
r
d
etec
tin
g
tar
g
ets
s
m
aller
th
an
3
2
×3
2
p
ix
els.
T
o
ad
d
r
ess
th
is
,
s
ev
er
al
ap
p
r
o
ac
h
es
h
av
e
em
er
g
ed
.
Sp
atial
atten
tio
n
m
ec
h
a
n
is
m
s
,
as
im
p
lem
en
ted
in
s
m
all
o
b
j
ec
t
d
etec
tio
n
-
YOL
Ov
8
[
6
]
,
d
em
o
n
s
tr
ated
a
6
.
9
%
im
p
r
o
v
e
m
en
t
in
m
AP
f
o
r
o
b
jects
u
n
d
e
r
3
2
×3
2
p
ix
els
b
y
em
p
h
asizin
g
s
p
atial
lo
ca
tio
n
s
wh
er
e
s
m
all
o
b
jects
ar
e
li
k
ely
to
ap
p
ea
r
.
T
o
ad
d
r
ess
f
ea
tu
r
e
lo
s
s
,
ar
ch
itect
u
r
al
m
o
d
if
icatio
n
s
s
u
ch
as
s
p
ac
e
-
to
-
d
ep
t
h
co
n
v
o
l
u
tio
n
(
SPD
-
C
o
n
v
)
h
av
e
b
ee
n
in
teg
r
ated
in
to
YOL
O
ar
ch
ite
ctu
r
es
to
p
r
eser
v
e
h
ig
h
-
f
r
eq
u
en
cy
s
p
atial
d
etails.
T
h
ese
m
eth
o
d
s
h
av
e
s
h
o
wn
s
ig
n
if
ican
t
p
r
ec
is
io
n
g
ain
s
o
n
s
m
all
o
b
ject
d
atasets
lik
e
Vis
Dr
o
n
e
b
y
r
e
p
lacin
g
tr
ad
itio
n
al
s
tr
id
es
co
n
v
o
l
u
tio
n
s
th
at
d
is
ca
r
d
tin
y
o
b
ject
in
f
o
r
m
atio
n
[
2
9
]
.
T
r
an
s
f
o
r
m
e
r
-
b
ased
ar
ch
itectu
r
es
h
av
e
r
ec
en
tly
ch
allen
g
ed
th
e
d
o
m
in
an
ce
o
f
p
u
r
e
C
NNs
in
s
m
all
o
b
ject
d
etec
tio
n
b
y
e
f
f
ec
tiv
ely
ca
p
tu
r
i
n
g
l
o
n
g
-
r
an
g
e
d
e
p
en
d
en
cies
[
3
0
]
.
W
h
ile
tr
a
d
itio
n
al
YOL
O
m
o
d
els
r
ely
o
n
lo
ca
l
r
ec
e
p
tiv
e
f
ield
s
,
th
e
in
teg
r
atio
n
o
f
ViT
o
r
th
e
u
s
e
o
f
T
r
an
s
f
o
r
m
er
-
b
ased
en
c
o
d
er
s
,
as
s
ee
n
in
r
ea
l
-
tim
e
d
etec
tio
n
tr
an
s
f
o
r
m
e
r
(
RT
-
DE
T
R
)
,
allo
ws f
o
r
b
etter
co
n
tex
tu
al
r
ea
s
o
n
i
n
g
[
3
1
]
.
T
h
ese
ad
v
an
ce
m
en
ts
ar
e
p
ar
ticu
lar
ly
b
e
n
ef
icial
f
o
r
d
atasets
lik
e
T
in
y
Per
s
o
n
,
wh
er
e
th
e
g
lo
b
al
co
n
tex
t
o
f
a
s
ce
n
e
ca
n
h
elp
d
is
tin
g
u
is
h
m
in
u
te
tar
g
ets f
r
o
m
b
ac
k
g
r
o
u
n
d
n
o
is
e
[
3
2
]
,
[
3
3
]
.
3.
M
E
T
H
O
D
A
r
ig
o
r
o
u
s
an
d
well
-
s
tr
u
ctu
r
ed
r
esear
ch
d
esig
n
is
f
u
n
d
a
m
en
tal
to
ac
h
iev
in
g
th
e
s
tated
r
esear
ch
o
b
jectiv
es
an
d
en
s
u
r
i
n
g
t
h
e
v
alid
ity
o
f
th
e
f
in
d
i
n
g
s
.
T
h
is
s
ec
tio
n
o
u
tlin
es
th
e
m
eth
o
d
o
lo
g
ical
f
r
am
ewo
r
k
em
p
lo
y
ed
to
o
p
tim
ize
YOL
Ov
1
1
f
o
r
s
m
all
o
b
ject
d
etec
t
io
n
,
p
a
r
ticu
lar
ly
o
n
th
e
T
in
y
Per
s
o
n
d
ataset.
T
h
e
d
esig
n
en
s
u
r
es
alig
n
m
en
t
b
et
wee
n
r
esear
ch
g
o
als
a
n
d
t
h
e
d
ev
elo
p
m
e
n
t,
tr
ain
i
n
g
,
an
d
ev
alu
atio
n
o
f
m
ac
h
in
e
lear
n
in
g
m
o
d
els.
T
h
is
s
tu
d
y
ad
o
p
ts
a
q
u
an
titativ
e,
ex
p
er
im
e
n
tal
r
esear
ch
d
esig
n
,
g
u
id
ed
b
y
th
e
f
o
llo
win
g
p
r
i
n
cip
les:
i)
Mo
d
el
-
ce
n
ter
ed
ex
p
e
r
im
en
tatio
n
,
f
o
cu
s
in
g
o
n
alg
o
r
ith
m
en
h
an
ce
m
en
t a
n
d
p
er
f
o
r
m
a
n
ce
co
m
p
ar
is
o
n
.
ii)
C
o
n
tr
o
lled
ex
p
e
r
im
en
tatio
n
,
in
v
o
lv
in
g
s
y
s
tem
atic
v
ar
iatio
n
s
in
m
o
d
el
p
a
r
am
eter
s
an
d
co
m
p
o
n
en
t
s
(
e.
g
.
,
a
n
ch
o
r
b
o
x
es,
atten
tio
n
m
ec
h
an
is
m
s
)
.
iii)
Per
f
o
r
m
an
ce
b
e
n
ch
m
a
r
k
in
g
,
u
s
in
g
s
tan
d
ar
d
ized
ev
alu
atio
n
m
etr
ics
(
m
AP,
r
ec
all,
p
r
ec
is
io
n
,
in
ter
s
ec
tio
n
o
v
er
u
n
io
n
(
I
o
U
)
)
ac
r
o
s
s
m
u
lti
p
le
m
o
d
el
c
o
n
f
ig
u
r
atio
n
s
an
d
b
aselin
es.
T
h
e
r
esear
ch
d
esig
n
alig
n
s
cl
o
s
ely
with
th
e
ar
tific
ial
in
tell
ig
en
ce
/
m
ac
h
in
e
lear
n
i
n
g
d
ev
e
lo
p
m
en
t
life
cy
cle,
co
v
er
in
g
d
ata
p
r
ep
ar
atio
n
,
m
o
d
el
d
ev
elo
p
m
en
t,
p
er
f
o
r
m
an
ce
ev
alu
atio
n
,
a
n
d
c
o
m
p
ar
ativ
e
a
n
aly
s
is
.
3
.
1
.
Da
t
a
s
et
des
cr
iptio
n a
nd
prepa
ra
t
io
n
T
wo
d
atasets
ar
e
u
tili
ze
d
in
t
h
is
r
esear
ch
to
ev
alu
ate
m
o
d
el
p
er
f
o
r
m
an
ce
u
n
d
er
b
o
th
g
en
er
al
an
d
d
o
m
ain
-
s
p
ec
if
ic
tin
y
o
b
ject
d
e
tectio
n
s
ce
n
ar
io
s
:
i)
T
in
y
Per
s
o
n
d
ataset: a
p
u
b
licly
av
ailab
le
b
en
ch
m
a
r
k
d
esig
n
e
d
s
p
ec
if
ically
f
o
r
tin
y
-
s
ca
le
h
u
m
an
d
etec
tio
n
in
co
m
p
le
x
an
d
d
e
n
s
e
en
v
ir
o
n
m
en
ts
.
I
t
in
clu
d
es
ap
p
r
o
x
im
at
ely
1
2
,
0
0
0
h
ig
h
-
r
eso
lu
tio
n
im
ag
es
an
d
o
v
e
r
4
5
,
0
0
0
a
n
n
o
tated
b
o
u
n
d
in
g
b
o
x
es
o
f
th
e
tin
y
h
u
m
a
n
in
s
tan
ce
s
,
m
o
s
t
m
ea
s
u
r
in
g
less
th
an
3
2
×
3
2
p
ix
els.
T
h
e
d
ataset
f
ea
tu
r
es v
ar
ie
d
s
ce
n
es su
ch
as u
r
b
an
s
tr
ee
ts
,
p
ar
k
s
,
an
d
ae
r
ial
p
er
s
p
ec
tiv
es.
ii)
C
OC
O
m
in
i
d
ataset:
a
cu
r
ated
s
u
b
s
et
o
f
th
e
MS
C
OC
O
d
ata
s
et
co
m
p
r
is
in
g
ap
p
r
o
x
im
ately
5
,
0
0
0
im
ag
e
s
an
d
1
5
,
2
0
0
an
n
o
tated
o
b
jects
,
with
an
em
p
h
asis
o
n
s
m
all
-
s
ca
le
ca
teg
o
r
ies
in
clu
d
in
g
p
er
s
o
n
s
,
b
ir
d
s
,
b
o
ttles
,
an
d
tr
af
f
ic
s
ig
n
s
.
T
h
is
d
ataset
en
ab
les
cr
o
s
s
-
d
o
m
ain
v
alid
atio
n
o
f
m
o
d
el
g
en
er
aliza
tio
n
b
ey
o
n
d
th
e
p
ed
estrian
d
o
m
ain
.
Pre
p
r
o
ce
s
s
in
g
task
s
:
‒
No
r
m
aliza
tio
n
an
d
r
esizin
g
(
e.
g
.
,
6
4
0
×6
4
0
an
d
1
2
8
0
×1
2
8
0
r
eso
lu
tio
n
v
ar
ian
ts
)
.
‒
An
n
o
tatio
n
f
o
r
m
at
c
o
n
v
er
s
io
n
to
YOL
O
-
co
m
p
atib
le
f
o
r
m
at.
‒
C
las
s
b
alan
cin
g
an
d
au
g
m
e
n
ta
tio
n
u
s
in
g
m
o
s
aic,
co
p
y
-
p
aste,
an
d
s
u
p
er
-
r
eso
lu
tio
n
.
‒
Data
s
et
s
p
lit in
to
7
0
% tr
ain
in
g
,
1
5
% v
alid
atio
n
,
1
5
% testi
n
g
.
T
h
e
f
o
llo
win
g
m
o
d
el
-
ce
n
tr
ic
e
n
h
an
ce
m
e
n
ts
will b
e
ap
p
lied
t
o
YOL
Ov
1
1
:
‒
An
ch
o
r
b
o
x
o
p
tim
izatio
n
:
k
-
m
ea
n
s
clu
s
ter
in
g
to
g
e
n
er
ate
an
ch
o
r
b
o
x
es
tailo
r
e
d
to
th
e
o
b
ject
s
ize
d
is
tr
ib
u
tio
n
o
f
th
e
T
in
y
Per
s
o
n
d
ataset
.
‒
Ob
jectiv
e:
im
p
r
o
v
e
I
o
U
an
d
m
AP b
y
b
etter
m
atch
in
g
b
o
u
n
d
in
g
b
o
x
p
r
o
p
o
s
als.
‒
Data
au
g
m
en
tatio
n
tec
h
n
iq
u
es
: m
o
s
aic
au
g
m
en
tatio
n
: f
o
r
b
et
ter
s
p
atial
co
n
tex
t a
n
d
o
b
ject
b
len
d
in
g
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
15
,
No
.
4
,
Au
g
u
s
t
20
26
:
3
4
5
2
-
3
4
6
3
3456
‒
C
o
p
y
-
p
aste a
u
g
m
e
n
tatio
n
: to
s
y
n
th
etica
lly
b
alan
ce
o
b
ject
d
is
tr
ib
u
tio
n
.
‒
Su
p
er
-
r
eso
lu
tio
n
:
en
h
a
n
ce
s
m
all
o
b
ject
clar
ity
to
r
e
d
u
ce
m
is
s
ed
d
etec
tio
n
s
.
‒
Atten
tio
n
m
ec
h
a
n
is
m
in
teg
r
a
tio
n
:
s
q
u
ee
ze
-
a
n
d
-
ex
citatio
n
(
SE
)
-
b
lo
ck
o
r
C
B
AM
will
b
e
ad
d
ed
t
o
en
h
an
ce
s
p
atial
an
d
ch
an
n
el
-
w
is
e
atten
tio
n
,
im
p
r
o
v
in
g
d
etec
tio
n
o
f
l
o
w
-
r
eso
lu
tio
n
ta
r
g
ets.
‒
Featu
r
e
p
y
r
am
id
n
e
two
r
k
(
F
PN)
:
in
teg
r
ate
FP
N
lay
er
s
to
f
u
s
e
m
u
lti
-
s
ca
le
f
ea
tu
r
es
an
d
r
etain
s
p
atial
r
eso
lu
tio
n
cr
itical
f
o
r
s
m
all
o
b
ject
d
etec
tio
n
.
‒
L
o
s
s
f
u
n
ctio
n
v
a
r
iatio
n
:
r
ep
la
ce
th
e
d
ef
au
lt
I
o
U
lo
s
s
with
c
o
m
p
lete
in
ter
s
ec
tio
n
o
v
er
u
n
i
o
n
(
C
I
o
U
)
a
n
d
ef
f
icien
t in
ter
s
ec
tio
n
o
v
er
u
n
i
o
n
(
E
I
o
U
)
to
test
th
eir
ef
f
ec
t o
n
lo
ca
lizatio
n
ac
cu
r
ac
y
.
3
.
2
.
M
a
chine
lea
rning
m
o
de
l dev
elo
pm
ent
T
h
e
m
o
d
el
a
r
ch
itectu
r
e
u
s
ed
in
th
is
p
ap
er
is
YOL
Ov
1
1
,
w
h
ich
is
th
e
s
tate
o
f
ar
t
o
b
ject
d
etec
tio
n
alg
o
r
ith
m
,
b
alan
cin
g
r
ea
l
-
tim
e
in
f
er
en
ce
s
p
ee
d
an
d
h
ig
h
lo
ca
lizatio
n
ac
cu
r
ac
y
.
T
ab
le
2
d
escr
ib
es
th
e
alg
o
r
ith
m
s
elec
tio
n
a
n
d
ju
s
t
if
icatio
n
.
All
th
e
au
g
m
en
tati
o
n
s
tr
ateg
ies
ar
e
d
esig
n
ed
t
o
em
u
late
ce
r
tain
co
n
d
itio
n
s
lik
e
a
lar
g
e
cr
o
w
d
s
ce
n
e
o
r
o
cc
lu
s
io
n
th
at
ar
e
co
m
m
o
n
in
t
h
e
d
etec
tio
n
o
f
s
m
all
o
b
jects.
T
ab
le
2
.
Alg
o
r
ith
m
s
elec
tio
n
a
n
d
ju
s
tific
atio
n
C
o
m
p
o
n
e
n
t
D
e
scri
p
t
i
o
n
R
a
t
i
o
n
a
l
e
Y
O
LO
v
1
1
b
a
c
k
b
o
n
e
C
o
n
v
o
l
u
t
i
o
n
a
l
l
a
y
e
r
s
w
i
t
h
c
o
mm
o
n
sp
a
t
i
a
l
p
a
t
t
e
r
n
s
(
C
S
P
)
m
o
d
u
l
e
s
Ef
f
i
c
i
e
n
t
a
n
d
d
e
e
p
f
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
N
e
c
k
a
r
c
h
i
t
e
c
t
u
r
e
P
A
N
e
t
o
r
F
P
N
i
n
t
e
g
r
a
t
i
o
n
M
u
l
t
i
-
sca
l
e
f
e
a
t
u
r
e
f
u
si
o
n
,
c
r
i
t
i
c
a
l
f
o
r
smal
l
o
b
j
e
c
t
s
H
e
a
d
st
r
u
c
t
u
r
e
M
u
l
t
i
-
sca
l
e
d
e
t
e
c
t
i
o
n
h
e
a
d
s
D
e
d
i
c
a
t
e
d
l
a
y
e
r
s f
o
r
d
e
t
e
c
t
i
n
g
sma
l
l
,
med
i
u
m
,
l
a
r
g
e
o
b
j
e
c
t
s
Lo
ss
f
u
n
c
t
i
o
n
C
I
o
U
o
r
EI
o
U
B
e
t
t
e
r
b
o
u
n
d
i
n
g
b
o
x
r
e
g
r
e
ss
i
o
n
a
n
d
c
o
n
v
e
r
g
e
n
c
e
b
e
h
a
v
i
o
r
A
t
t
e
n
t
i
o
n
mec
h
a
n
i
sm
SE
-
b
l
o
c
k
o
r
C
B
A
M
I
mp
r
o
v
e
s
f
o
c
u
s
o
n
sm
a
l
l
-
sca
l
e
v
i
su
a
l
c
u
e
s
A
n
c
h
o
r
b
o
x
o
p
t
i
m
i
z
a
t
i
o
n
K
-
me
a
n
s
c
l
u
s
t
e
r
i
n
g
C
u
s
t
o
m a
n
c
h
o
r
si
z
e
s
i
m
p
r
o
v
e
I
o
U
a
n
d
d
e
t
e
c
t
i
o
n
r
a
t
e
3
.
2
.
1
.
Da
t
a
prepro
ce
s
s
ing
a
n
d a
ug
m
ent
a
t
io
n
An
ef
f
ec
tiv
e
p
r
ep
r
o
ce
s
s
in
g
p
i
p
elin
e
was
d
ev
elo
p
ed
to
p
r
e
p
ar
e
th
e
d
ata
to
tr
ain
an
d
ev
alu
ate.
T
h
is
is
d
o
n
e
to
p
r
o
v
id
e
th
e
m
ac
h
in
e
lear
n
in
g
m
o
d
el
with
h
ig
h
-
q
u
ality
d
iv
er
s
e
r
ep
r
esen
tativ
e
in
p
u
t
d
ata.
T
a
b
le
3
d
escr
ib
es d
ata
p
r
ep
r
o
ce
s
s
in
g
a
n
d
au
g
m
en
tatio
n
s
tep
s
with
d
e
s
cr
ip
tio
n
an
d
t
h
eir
p
u
r
p
o
s
es.
T
ab
le
3
.
Data
p
r
ep
r
o
ce
s
s
in
g
a
n
d
au
g
m
en
tatio
n
S
t
e
p
D
e
scri
p
t
i
o
n
P
u
r
p
o
se
N
o
r
mal
i
z
a
t
i
o
n
P
i
x
e
l
v
a
l
u
e
s
c
a
l
i
n
g
(
0
-
1)
S
t
a
n
d
a
r
d
i
n
p
u
t
r
a
n
g
e
f
o
r
C
N
N
l
a
y
e
r
s
R
e
si
z
i
n
g
(
6
4
0
×
6
4
0
,
1
2
8
0
×
1
2
8
0
)
M
a
i
n
t
a
i
n
o
b
j
e
c
t
a
sp
e
c
t
r
a
t
i
o
n
Ev
a
l
u
a
t
e
r
e
s
o
l
u
t
i
o
n
se
n
si
t
i
v
i
t
y
A
n
n
o
t
a
t
i
o
n
f
o
r
ma
t
c
o
n
v
e
r
si
o
n
C
O
C
O
t
o
Y
O
LO
f
o
r
m
a
t
(
t
x
t
)
C
o
m
p
a
t
i
b
i
l
i
t
y
w
i
t
h
Y
O
LO
t
r
a
i
n
i
n
g
N
o
i
se
f
i
l
t
e
r
i
n
g
R
e
m
o
v
e
b
l
u
r
r
e
d
,
l
o
w
-
q
u
a
l
i
t
y
s
a
m
p
l
e
s
I
mp
r
o
v
e
t
r
a
i
n
i
n
g
s
i
g
n
a
l
A
u
g
m
e
n
t
a
t
i
o
n
t
e
c
h
n
i
q
u
e
s
M
o
s
a
i
c
,
c
o
p
y
-
p
a
st
e
,
s
u
p
e
r
-
r
e
s
o
l
u
t
i
o
n
B
o
o
st
d
a
t
a
d
i
v
e
r
s
i
t
y
a
n
d
m
i
mi
c
r
e
a
l
-
w
o
r
l
d
c
o
n
d
i
t
i
o
n
s
3
.
2
.
2
.
M
o
del
t
ra
ini
ng
co
nfig
ura
t
io
n
T
h
e
tr
ain
in
g
p
r
o
ce
d
u
r
e
in
clu
d
es
ea
r
ly
s
to
p
p
in
g
,
lear
n
in
g
r
a
te
war
m
-
u
p
,
a
n
d
d
ata
s
h
u
f
f
lin
g
.
W
h
ich
h
elp
s
tab
ilize
lear
n
in
g
a
n
d
av
o
id
o
v
er
f
itti
n
g
.
T
h
is
is
ex
p
lain
e
d
b
y
T
a
b
le
4
.
T
ab
le
4
.
Mo
d
el
tr
ain
in
g
co
n
f
ig
u
r
atio
n
P
a
r
a
me
t
e
r
V
a
l
u
e
O
p
t
i
mi
z
e
r
S
G
D
w
i
t
h
m
o
me
n
t
u
m
o
r
A
d
a
mW
Le
a
r
n
i
n
g
r
a
t
e
0
.
0
0
1
(
c
o
s
i
n
e
d
e
c
a
y
sc
h
e
d
u
l
e
)
B
a
t
c
h
si
z
e
16
-
3
2
(
a
d
a
p
t
i
v
e
b
a
s
e
d
o
n
G
P
U
m
e
m
o
r
y
)
Ep
o
c
h
s
1
5
0
-
200
I
mag
e
si
z
e
M
u
l
t
i
-
r
e
s
o
l
u
t
i
o
n
(
6
4
0
,
9
6
0
,
1
2
8
0
)
Ea
r
l
y
st
o
p
p
i
n
g
B
a
se
d
o
n
v
a
l
i
d
a
t
i
o
n
mA
P
3
.
3
.
M
o
del
a
rc
hite
ct
ure
inte
g
ra
t
io
n str
a
t
e
g
y
T
o
s
p
ec
if
ically
en
h
an
ce
YOL
Ov
1
1
'
s
ca
p
ab
ilit
y
f
o
r
d
etec
tin
g
s
m
all
o
b
jects,
th
is
s
tu
d
y
in
tr
o
d
u
ce
s
two
co
r
e
m
o
d
if
icatio
n
s
to
its
ar
ch
itectu
r
e:
th
e
in
teg
r
atio
n
o
f
t
h
e
C
B
AM
an
d
th
e
en
h
an
ce
m
e
n
t
o
f
th
e
FP
N.
As
illu
s
tr
ated
in
Fig
u
r
e
1
,
th
ese
m
o
d
u
les
ar
e
s
y
s
tem
atica
lly
e
m
b
ed
d
e
d
in
to
th
e
b
ac
k
b
o
n
e
a
n
d
n
ec
k
o
f
YOL
Ov
1
1
.
T
o
g
eth
er
,
th
ey
f
o
r
m
a
co
llab
o
r
ativ
ely
o
p
tim
ized
d
etec
tio
n
f
r
a
m
ewo
r
k
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
YOLOv1
1
o
p
timiz
a
tio
n
fo
r
tin
y
o
b
ject
in
cro
w
d
ed
s
ce
n
es
(
H
u
s
n
a
S
a
r
ir
a
h
Hu
s
in
)
3457
I
n
teg
r
atio
n
o
f
C
B
AM
:
t
h
e
C
B
AM
m
o
d
u
le
is
in
s
er
ted
at
a
cr
itical
ju
n
ctio
n
b
etwe
en
th
e
en
d
o
f
th
e
YOL
Ov
1
1
b
ac
k
b
o
n
e
an
d
th
e
b
eg
in
n
in
g
o
f
th
e
n
ec
k
.
Sp
ec
if
ically
,
af
ter
f
ea
tu
r
e
ex
tr
ac
tio
n
b
y
th
e
b
ac
k
b
o
n
e,
th
e
o
u
tp
u
t h
ig
h
-
lev
el
f
ea
tu
r
e
m
a
p
s
ar
e
f
ir
s
t p
r
o
ce
s
s
ed
b
y
th
e
C
B
AM
m
o
d
u
le.
T
h
is
m
o
d
u
le
s
eq
u
en
tially
em
p
lo
y
s
a
ch
an
n
el
atten
tio
n
s
u
b
-
m
o
d
u
le
f
o
llo
wed
b
y
a
s
p
atial
atten
t
io
n
s
u
b
-
m
o
d
u
le
to
d
y
n
am
ica
lly
r
ec
alib
r
ate
th
e
f
ea
tu
r
e
m
ap
s
[
1
7
]
.
T
h
e
ch
an
n
el
atten
tio
n
m
ec
h
an
is
m
f
o
c
u
s
es
o
n
“
wh
ich
f
ea
tu
r
e
c
h
an
n
els
ar
e
im
p
o
r
tan
t
”
,
en
h
an
cin
g
th
e
r
esp
o
n
s
e
to
cr
u
cial
f
ea
tu
r
es
o
f
s
m
all
o
b
jects.
T
h
e
s
p
atial
atten
tio
n
m
ec
h
an
is
m
f
o
c
u
s
es
o
n
“
wh
er
e
in
th
e
f
ea
tu
r
e
m
ap
is
i
m
p
o
r
tan
t
”
,
aid
in
g
in
lo
ca
tin
g
t
in
y
tar
g
ets
with
in
cr
o
wd
ed
s
ce
n
es
[
1
6
]
,
[
3
4
]
.
T
h
is
in
teg
r
atio
n
allo
ws
th
e
n
etwo
r
k
to
p
r
io
r
itize
f
ea
tu
r
es
r
ele
v
a
n
t
to
m
i
n
u
s
cu
le
o
b
jects
an
d
s
u
p
p
r
ess
ir
r
elev
an
t
b
ac
k
g
r
o
u
n
d
n
o
is
e
b
e
f
o
r
e
p
r
o
c
ee
d
in
g
to
m
u
lti
-
s
ca
le
f
u
s
io
n
.
E
n
h
an
ce
m
e
n
t
an
d
f
u
s
io
n
o
f
FP
N:
t
h
is
s
tu
d
y
em
p
lo
y
s
a
m
o
d
if
ied
FP
N
s
tr
u
ctu
r
e
as
th
e
c
o
r
e
o
f
th
e
n
ec
k
n
etwo
r
k
.
T
h
is
FP
N
co
n
s
tr
u
cts
a
to
p
-
d
o
w
n
p
at
h
way
to
f
u
s
e
d
e
ep
,
h
ig
h
-
s
em
an
tic
f
ea
tu
r
es
f
r
o
m
th
e
b
ac
k
b
o
n
e
with
s
h
allo
w,
h
i
g
h
-
r
eso
lu
tio
n
f
ea
tu
r
es
[
1
6
]
.
C
o
m
p
ar
ed
to
t
h
e
o
r
ig
in
al
d
esig
n
,
th
e
u
tili
za
tio
n
o
f
s
h
allo
w
f
ea
tu
r
es
is
s
tr
en
g
t
h
e
n
ed
,
e
n
s
u
r
in
g
th
at
m
o
r
e
lo
w
-
lev
el
f
ea
tu
r
es
co
n
tain
in
g
f
in
e
d
etails
o
f
s
m
all
o
b
jects
ar
e
ef
f
ec
tiv
ely
t
r
an
s
m
itted
an
d
p
r
eser
v
e
d
.
T
h
e
f
u
s
ed
m
u
lti
-
s
ca
le
f
ea
t
u
r
e
m
ap
s
ar
e
th
en
f
ed
in
to
th
e
d
etec
tio
n
h
ea
d
s
.
T
h
is
allo
ws
ea
ch
d
etec
tio
n
h
ea
d
to
o
p
er
ate
o
n
f
ea
tu
r
es
th
at
ar
e
r
ich
in
s
e
m
an
tic
in
f
o
r
m
atio
n
wh
ile
r
etain
in
g
s
u
f
f
icien
t
s
p
at
ial
d
etail,
th
er
eb
y
s
ig
n
if
ican
tl
y
im
p
r
o
v
in
g
th
e
r
ec
all
an
d
lo
ca
lizatio
n
ac
cu
r
ac
y
f
o
r
tin
y
o
b
jects
[
8
]
.
I
n
s
u
m
m
ar
y
,
th
e
s
y
n
er
g
is
tic
ac
tio
n
o
f
C
B
AM
an
d
th
e
en
h
an
c
ed
FP
N
f
o
r
m
s
th
e
co
r
e
o
f
o
u
r
o
p
tim
ized
m
o
d
el
.
C
B
AM
im
p
r
o
v
es
f
ea
t
u
r
e
q
u
ality
th
r
o
u
g
h
r
e
f
in
ed
f
il
ter
in
g
o
f
b
ac
k
b
o
n
e
f
ea
t
u
r
es,
wh
ile
th
e
en
h
an
ce
d
FP
N
en
s
u
r
es
th
at
th
ese
f
ilter
ed
,
in
f
o
r
m
atio
n
-
r
ich
f
ea
tu
r
es
f
o
r
s
m
all
o
b
jects
ar
e
ef
f
ec
tiv
ely
p
r
o
p
ag
ated
an
d
f
u
s
ed
ac
r
o
s
s
m
u
ltip
le
s
ca
les.
T
h
is
in
teg
r
atio
n
s
tr
ateg
y
is
clea
r
ly
d
ep
icted
in
Fig
u
r
e
1
.
Fig
u
r
e
1
.
I
n
teg
r
atio
n
o
f
C
B
A
M
an
d
FP
N
m
o
d
el
4.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
As
a
r
ea
l
ex
am
p
le,
a
s
im
u
l
atio
n
was
d
o
n
e
wh
er
e
th
e
o
p
tim
ized
YOL
Ov
1
1
wo
r
k
e
d
in
u
r
b
a
n
s
u
r
v
eillan
ce
an
d
t
h
e
id
en
tifi
ca
tio
n
o
f
p
ed
estrian
s
b
y
a
n
u
n
m
an
n
ed
ae
r
ial
v
eh
icle
(
UAV)
.
T
h
e
s
y
s
tem
co
m
b
in
es m
ak
in
g
p
r
e
d
ictio
n
s
f
r
o
m
th
e
m
o
d
el,
s
h
o
win
g
im
a
g
e
o
u
tp
u
ts
,
an
aly
zin
g
r
esu
lts
,
an
d
in
ter
ac
tiv
e
way
s
to
f
ilter
s
to
r
ed
im
a
g
es.
T
h
e
f
o
llo
win
g
d
escr
ib
es
th
e
s
y
s
te
m
f
u
n
ctio
n
ality
o
v
e
r
v
iew.
T
h
e
u
s
ef
u
ln
ess
o
f
th
e
m
o
d
el
was
ch
ec
k
ed
b
y
m
ak
i
n
g
a
lig
h
t
an
d
m
o
d
u
lar
d
ete
ctio
n
in
ter
f
ac
e
b
ased
o
n
o
p
tim
ized
YOL
Ov
1
1
.
I
t
wo
r
k
s
s
o
th
at
an
y
o
n
e
ca
n
u
s
e
it
ea
s
ily
,
eith
er
in
r
ea
l
ti
m
e
o
r
o
f
f
lin
e,
wh
en
ev
e
r
a
n
d
wh
er
ev
er
th
ey
wan
t.
Usi
n
g
u
ltra
ly
tics
YOL
Ov
1
1
API
in
Py
th
o
n
a
n
d
Stre
am
lit,
th
e
ap
p
ca
n
b
e
ac
ce
s
s
ed
in
t
h
e
b
r
o
wser
with
o
u
t
in
s
tallin
g
an
y
th
in
g
.
4
.
1
.
I
np
ut
s
up
po
rt
f
o
r
im
a
g
e
a
nd
v
ideo
s
t
re
a
m
s
Peo
p
le
u
s
in
g
C
h
r
o
m
e
ca
n
c
h
o
o
s
e
to
u
p
lo
ad
im
a
g
es
o
r
a
d
d
v
id
eo
f
iles
,
o
r
th
ey
ca
n
also
u
s
e
a
liv
e
ca
m
er
a
f
ee
d
.
Vid
e
o
d
ata
is
tu
r
n
ed
in
to
f
r
am
es
an
d
eith
er
h
an
d
led
f
r
am
e
b
y
f
r
am
e
o
r
b
y
g
r
o
u
p
s
,
d
ep
en
d
in
g
o
n
av
ailab
le
GPUs
.
Usi
n
g
au
to
-
s
ca
lin
g
,
th
e
in
p
u
t’
s
r
eso
lu
tio
n
b
ec
o
m
es
eith
er
6
4
0
×6
4
0
o
r
1
2
8
0
×1
2
8
0
b
u
t
it
s
till
k
ee
p
s
th
e
asp
ec
t r
atio
.
4
.
2
.
Rea
l
-
t
im
e
det
ec
t
io
n wit
h v
is
ua
l f
ee
db
a
ck
YOL
Ov
1
1
wield
in
g
o
n
-
t
h
e
-
f
l
y
in
f
er
e
n
ce
is
u
s
ed
as
th
e
s
y
s
tem
’
s
b
ac
k
b
o
n
e
.
T
h
e
d
r
awn
b
o
x
es
ar
e
g
iv
en
u
n
iq
u
e
co
lo
r
s
if
th
ey
co
r
r
esp
o
n
d
to
a
p
e
r
s
o
n
(
class
_
id
=0
)
tar
g
et.
E
v
er
y
b
o
x
s
h
o
ws
its
co
n
f
id
en
ce
v
alu
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
15
,
No
.
4
,
Au
g
u
s
t
20
26
:
3
4
5
2
-
3
4
6
3
3458
an
d
,
i
f
th
e
r
e
is
e
n
o
u
g
h
in
f
o
r
m
atio
n
,
m
a
y
also
p
r
e
d
ict
th
e
p
er
s
o
n
’
s
p
o
s
tu
r
e
(
s
u
ch
as
f
ac
i
n
g
f
o
r
war
d
,
to
th
e
s
id
e,
o
r
s
q
u
attin
g
)
.
4
.
3
.
I
nte
ra
ct
i
v
e
det
ec
t
io
n v
e
rif
ica
t
io
n (
hu
m
a
n
-
in
-
t
he
-
lo
o
p
)
E
ac
h
tim
e
a
d
etec
tio
n
is
m
ad
e,
u
s
er
s
h
av
e
th
e
o
p
tio
n
to
c
h
ec
k
th
e
m
atch
m
an
u
ally
a
n
d
c
o
n
f
ir
m
th
e
m
o
d
el’
s
ch
o
ice
o
r
n
o
t.
Pre
cisi
o
n
o
f
d
etec
tio
n
ca
n
b
e
in
c
r
ea
s
ed
b
y
m
o
v
i
n
g
s
lid
er
s
in
a
s
id
e
b
ar
to
s
et
b
o
th
th
e
th
r
esh
o
ld
an
d
th
e
s
ize
o
f
d
ete
cted
o
b
jects.
I
t
is
m
o
s
t
u
s
ef
u
l
wh
en
h
an
d
lin
g
task
s
s
u
ch
as
m
o
n
ito
r
in
g
,
wh
er
e
a
p
er
s
o
n
n
ee
d
s
to
ch
ec
k
f
o
r
i
n
tr
u
d
er
s
,
p
er
f
o
r
m
in
g
r
esear
c
h
th
at
r
eq
u
ir
es
ac
cu
r
ate
d
ata
,
an
d
m
ak
in
g
n
ew
m
ac
h
in
e
lear
n
in
g
m
o
d
els m
o
r
e
ac
cu
r
ate
an
d
r
o
b
u
s
t.
4
.
4
.
P
er
f
o
r
m
a
nce
co
m
pa
riso
n v
is
ua
liza
t
io
n
As
a
way
to
e
v
alu
ate
th
ese
f
r
am
ewo
r
k
s
f
o
r
tin
y
o
b
ject
s
ce
n
ar
io
s
,
th
e
s
tu
d
y
r
elied
o
n
th
r
ee
co
n
f
ig
u
r
atio
n
s
:
YOL
Ov
1
1
(
a
s
a
b
aselin
e)
,
YOL
Ov
1
1
(
o
p
tim
ized
)
,
an
d
R
etin
aNe
t.
T
h
e
ass
ess
m
en
t
wa
s
co
n
d
u
cte
d
b
y
a
n
aly
zin
g
b
o
t
h
n
u
m
b
er
s
s
u
ch
as
ac
c
u
r
ac
y
,
r
ec
all,
s
p
ee
d
o
f
in
f
er
en
ce
,
an
d
c
o
m
p
lex
ity
,
as
well
as
q
u
alitativ
e
r
esu
lts
o
n
a
co
m
m
o
n
s
et
o
f
im
a
g
es
tak
en
f
r
o
m
th
e
T
in
y
Per
s
o
n
d
ataset
an
d
C
OC
O
Min
i.
Fig
u
r
e
2
d
em
o
n
s
tr
ates
th
at
th
e
o
p
tim
ized
YOL
Ov
1
1
p
e
r
f
o
r
m
s
m
u
ch
b
etter
th
an
th
e
b
aselin
e
b
y
g
ain
in
g
7
.
3
%
in
m
AP
an
d
1
0
.
5
%
in
r
ec
all,
an
d
th
e
in
cr
ea
s
e
in
th
e
m
o
d
el
’
s
s
ize
an
d
p
ar
am
eter
co
u
n
t
is
v
er
y
r
ea
s
o
n
ab
le
.
W
h
en
it
co
m
es
to
h
an
d
lin
g
r
e
al
-
tim
e
task
s
,
it
r
u
n
s
3
8
im
ag
es
p
er
s
ec
o
n
d
,
wh
ich
is
m
u
c
h
q
u
ick
er
th
a
n
F
aster
R
-
C
NN
(
8
F
PS
)
an
d
m
ak
es
it
m
u
ch
m
o
r
e
s
u
itab
le
f
o
r
s
u
ch
ap
p
licatio
n
s
.
I
t
is
clea
r
f
r
o
m
th
e
co
m
p
ar
is
o
n
th
at
YOL
Ov
1
1
,
af
ter
b
ein
g
p
r
o
p
e
r
ly
o
p
tim
ized
,
d
eliv
e
r
s
a
d
e
p
e
n
d
ab
le
b
alan
ce
b
etwe
en
f
i
n
d
i
n
g
tin
y
o
b
jects
an
d
s
av
in
g
o
n
p
o
wer
,
m
a
k
in
g
it
a
lo
g
ical
p
ick
f
o
r
m
o
b
ile
o
r
em
b
ed
d
ed
d
esig
n
s
aim
in
g
at
d
ete
ctin
g
tin
y
th
in
g
s
in
f
r
ee
co
n
d
itio
n
s
.
Fig
u
r
e
2
.
Per
f
o
r
m
an
c
e
co
m
p
ar
is
o
n
f
o
r
m
o
d
el
p
a
r
am
eter
s
,
in
f
er
en
ce
s
p
ee
d
(
FP
S)
an
d
m
AP
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
YOLOv1
1
o
p
timiz
a
tio
n
fo
r
tin
y
o
b
ject
in
cro
w
d
ed
s
ce
n
es
(
H
u
s
n
a
S
a
r
ir
a
h
Hu
s
in
)
3459
4
.
5
.
Crit
ica
l e
v
a
lua
t
io
n o
f
mo
del per
f
o
rm
a
nce
A
n
a
l
y
z
in
g
th
e
o
u
t
co
m
e
s
o
f
th
e
n
ex
t
m
e
tr
i
c
s
m
A
P
@
0
.
5
,
p
r
e
c
i
s
i
o
n
,
r
ec
a
l
l
,
I
o
U
,
an
d
F
PS
,
i
t
c
a
n
b
e
o
b
s
e
r
v
e
d
t
h
a
t
a
c
er
t
a
in
d
if
f
er
e
n
c
e
b
e
t
we
e
n
R
e
t
in
a
N
e
t
,
YO
L
O
v
1
1
-
b
a
s
e
l
i
n
e,
an
d
Y
OL
O
v
1
1
-
o
p
t
i
m
i
ze
d
o
n
T
i
n
y
P
er
s
o
n
d
a
t
a
s
e
t
an
d
C
OC
O
-
m
i
n
i
d
a
ta
s
e
t
s
i
s
e
v
i
d
en
t
.
T
h
e
c
o
m
p
ar
i
s
o
n
i
s
i
l
l
u
s
t
r
a
te
d
i
n
F
i
g
u
r
e
3
a
n
d
T
a
b
l
e
5
s
h
o
w
s
t
h
e
m
o
d
e
l
c
o
m
p
ar
i
s
o
n
.
Y
OL
O
v
1
1
-
o
p
t
i
m
i
z
e
d
w
a
s
a
l
s
o
th
e
h
ig
h
e
s
t
i
n
m
A
P
@
0
.
5
(
T
i
n
y
P
e
r
s
o
n
:
0
.
0
3
1
,
C
O
C
O
-
m
i
n
i
:
0
.
0
6
5
)
co
m
p
a
r
ed
t
o
1
2
.
5
%
an
d
1
8
.
2
%
h
ig
h
er
th
a
n
R
e
t
i
n
a
Ne
t
an
d
Y
O
L
O
v
1
1
-
b
a
s
e
l
i
n
e
r
e
s
p
e
c
t
i
v
e
l
y
.
Fig
u
r
e
3
.
Per
f
o
r
m
an
c
e
co
m
p
ar
is
o
n
b
etwe
en
YOL
Ov
8
,
YOL
Ov
9
,
R
etin
aNe
t,
YOL
Ov
1
1
-
b
aselin
e,
an
d
YOL
Ov
1
1
-
o
p
tim
ized
T
ab
le
5
.
Mo
d
el
p
er
f
o
r
m
a
n
ce
M
o
d
e
l
D
a
t
a
s
e
t
mA
P
@
0
.
5
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
I
o
U
FPS
R
e
t
i
n
a
N
e
t
Ti
n
y
P
e
r
s
o
n
0
.
0
1
9
0
.
0
7
8
0
.
0
2
9
0
.
3
5
1
7
.
6
Y
O
LO
v
1
1
-
B
a
sel
i
n
e
Ti
n
y
P
e
r
s
o
n
0
.
0
2
6
0
.
0
8
5
0
.
0
3
4
0
.
3
7
2
5
.
4
Y
O
LO
v
1
1
-
O
p
t
i
m
i
z
e
d
Ti
n
y
P
e
r
s
o
n
0
.
0
3
1
0
.
0
9
2
0
.
0
7
2
0
.
4
0
2
7
.
3
R
e
t
i
n
a
N
e
t
C
O
C
O
-
mi
n
i
0
.
0
5
2
0
.
1
4
3
0
.
0
4
1
0
.
4
5
1
7
.
1
Y
O
LO
v
1
1
-
B
a
sel
i
n
e
C
O
C
O
-
mi
n
i
0
.
0
5
7
0
.
1
5
1
0
.
0
5
8
0
.
4
7
2
4
.
9
Y
O
LO
v
1
1
-
O
p
t
i
m
i
z
e
d
C
O
C
O
-
mi
n
i
0
.
0
6
5
0
.
1
6
5
0
.
0
7
4
0
.
5
0
2
6
.
7
Un
d
er
r
ec
all,
YOL
Ov
1
1
-
o
p
ti
m
ized
ac
h
iev
ed
0
.
0
7
2
o
n
T
in
y
Per
s
o
n
d
ataset
an
d
it
is
two
tim
es
th
e
r
ec
all
o
f
th
e
YOL
Ov
1
1
-
b
asel
in
e
(
0
.
0
3
4
)
th
at
in
d
icate
s
a
g
r
ea
ter
s
en
s
itiv
ity
in
s
m
all
o
b
j
ec
t
d
etec
tio
n
.
T
h
e
ac
cu
r
ac
y
was
m
ar
g
in
ally
an
d
co
n
s
is
ten
tly
h
ig
h
er
(
YOL
Ov
1
1
-
o
p
tim
ized
:
0
.
0
9
2
v
s
R
etin
a
Net:
0
.
0
7
8
)
an
d
as
it
was
s
tated
in
ea
r
lier
wo
r
k
s
[
2
8
]
,
th
e
atten
tio
n
m
ec
h
an
is
m
lik
e
C
B
AM
an
d
m
o
r
e
ad
v
an
ce
d
an
ch
o
r
d
esig
n
led
to
a
h
ig
h
er
r
eliab
ilit
y
o
f
class
if
icatio
n
.
T
h
e
FP
N
an
d
th
e
k
-
m
ea
n
s
an
ch
o
r
s
also
i
n
cr
ea
s
ed
th
e
I
o
U
s
co
r
es,
i.e
.
,
th
e
b
o
u
n
d
in
g
b
o
x
es
wo
u
l
d
b
e
tig
h
t
an
d
th
e
l
o
ca
lizatio
n
wo
u
ld
b
e
im
p
r
o
v
ed
.
D
esp
ite
ar
ch
itectu
r
al
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
15
,
No
.
4
,
Au
g
u
s
t
20
26
:
3
4
5
2
-
3
4
6
3
3460
en
h
an
ce
m
e
n
ts
,
YOL
Ov
1
1
-
O
p
tim
ized
m
ain
tain
ed
a
r
esp
ec
tab
le
FP
S
(
2
7
.
3
)
,
d
em
o
n
s
tr
atin
g
its
p
r
ac
tical
u
s
ab
ilit
y
in
r
ea
l
-
tim
e
s
y
s
tem
s
—
a
f
in
d
in
g
o
n
th
e
b
alan
ce
o
f
ac
cu
r
ac
y
an
d
ef
f
icie
n
cy
in
m
o
d
el
d
esig
n
[
2
8
]
.
4
.
6
.
Co
m
pa
riso
n wit
h
ex
is
t
ing
s
t
ud
ies
T
h
e
im
p
r
o
v
em
en
ts
o
b
s
er
v
ed
i
n
YOL
Ov
1
1
-
o
p
tim
ized
ar
e
co
n
s
is
ten
t w
ith
s
ev
er
al
s
tu
d
ies:
i)
C
B
AM
ef
f
ec
tiv
en
ess
:
C
B
AM
’
s
ab
ilit
y
to
s
elec
tiv
ely
am
p
l
if
y
u
s
ef
u
l
f
ea
tu
r
e
m
ap
s
,
wh
i
ch
ap
p
ea
r
s
to
s
u
p
p
o
r
t th
e
i
m
p
r
o
v
ed
p
r
ec
is
io
n
s
ee
n
in
th
is
s
tu
d
y
[
3
3
]
.
ii)
K
-
m
e
a
n
s
a
n
c
h
o
r
cl
u
s
te
r
i
n
g
:
i
n
Y
O
L
O
v
3
,
a
n
c
h
o
r
b
o
x
o
p
t
i
m
i
za
t
i
o
n
s
i
g
n
i
f
i
c
a
n
t
l
y
i
m
p
r
o
v
e
s
l
o
ca
l
i
z
a
ti
o
n
[
2
8
]
.
T
h
e
f
in
d
i
n
g
s
af
f
ir
m
th
is
,
as m
AP a
n
d
I
o
U
m
etr
ics s
aw
a
n
o
t
ab
le
b
o
o
s
t.
iii)
FPN
:
FP
N’
s
ab
ilit
y
to
en
r
ich
m
u
lti
-
s
ca
le
co
n
tex
t
d
ir
ec
tly
co
r
r
elate
s
with
h
ig
h
er
r
ec
all
v
alu
es,
esp
ec
ially
in
T
in
y
Per
s
o
n
d
ataset,
wh
er
e
d
etec
tin
g
s
m
all
in
s
tan
ce
s
is
cr
u
cial
[
6
]
.
iv
)
Ad
v
an
ce
d
d
ata
au
g
m
e
n
tatio
n
s
:
tech
n
iq
u
es
lik
e
Mix
Up
an
d
C
u
tMix
h
av
e
b
ee
n
s
h
o
wn
to
r
eg
u
lar
iz
e
m
o
d
els
ef
f
ec
tiv
ely
.
T
h
e
co
n
s
is
ten
t
u
p
war
d
tr
en
d
in
m
etr
ics
f
r
o
m
ep
o
ch
1
5
o
n
war
d
(
s
ee
tr
ain
in
g
g
r
a
p
h
s
)
s
u
p
p
o
r
ts
th
eir
co
n
tr
ib
u
tio
n
[
9
]
,
[
3
3
]
.
Fig
u
r
e
4
s
h
o
ws
th
e
YOL
Ov
1
1
-
o
p
tim
ized
m
o
d
el,
in
c
o
r
p
o
r
a
tin
g
C
B
AM
,
FP
N,
k
-
m
ea
n
s
an
ch
o
r
s
,
an
d
en
h
an
ce
d
a
u
g
m
e
n
tatio
n
s
.
T
h
e
m
o
d
el
s
u
r
p
ass
es
b
o
th
R
etin
aNe
t
an
d
YOL
Ov
1
1
-
b
aselin
e
ac
r
o
s
s
all
ev
alu
atio
n
ax
es.
I
ts
b
alan
ce
d
p
r
o
f
ile
o
f
h
ig
h
ac
c
u
r
ac
y
an
d
r
ea
l
-
tim
e
i
n
f
er
en
ce
m
ak
es
it
p
a
r
ticu
lar
l
y
s
u
itab
le
f
o
r
ed
g
e
ar
tific
ial
in
tellig
en
ce
d
ep
lo
y
m
en
t in
p
ed
estrian
o
r
m
o
b
ile
o
b
j
ec
t d
etec
tio
n
.
Fig
u
r
e
4
.
C
o
m
p
a
r
is
o
n
b
etwe
e
n
m
o
d
els
T
in
y
Per
s
o
n
d
ataset
5.
CO
NCLU
SI
O
N
T
h
is
s
tu
d
y
d
em
o
n
s
tr
ates
th
at
th
e
YOL
Ov
1
1
-
o
p
tim
ized
m
o
d
el
estab
lis
h
es
a
n
ew
b
en
ch
m
ar
k
f
o
r
ac
cu
r
ac
y
an
d
ef
f
icien
c
y
in
tin
y
o
b
ject
d
etec
tio
n
.
T
h
e
in
te
g
r
atio
n
o
f
an
c
h
o
r
r
ec
o
n
f
ig
u
r
atio
n
,
s
p
atial
atten
tio
n
,
an
d
s
u
p
er
-
r
eso
lu
tio
n
is
a
co
m
b
in
atio
n
th
at
h
as
n
o
t
b
ee
n
p
r
e
v
io
u
s
ly
ex
p
lo
r
ed
in
th
e
liter
at
u
r
e.
T
h
is
p
r
o
ject
h
as
ad
d
r
ess
ed
th
e
in
tr
in
s
ic
lim
ita
tio
n
s
o
f
r
eso
lu
tio
n
lo
s
s
an
d
an
ch
o
r
s
ca
lin
g
.
T
h
e
r
esu
lts
in
d
icate
th
e
m
o
d
el'
s
s
u
p
er
io
r
ity
,
ac
h
iev
in
g
a
n
1
8
.
2
%
h
ig
h
er
m
AP@
0
.
5
th
an
th
e
YOL
Ov
1
1
-
b
aselin
e
wh
ile
m
ain
tain
in
g
a
r
o
b
u
s
t
2
7
.
3
FP
S.
T
h
e
in
clu
s
io
n
o
f
C
B
AM
an
d
k
-
m
ea
n
s
an
ch
o
r
d
esig
n
co
n
tr
ib
u
tes
to
co
n
s
is
ten
tly
h
ig
h
er
r
eliab
ilit
y
in
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
YOLOv1
1
o
p
timiz
a
tio
n
fo
r
tin
y
o
b
ject
in
cro
w
d
ed
s
ce
n
es
(
H
u
s
n
a
S
a
r
ir
a
h
Hu
s
in
)
3461
d
en
s
e
en
v
ir
o
n
m
en
ts
,
y
ield
in
g
a
to
p
I
o
U
o
f
0
.
5
0
o
n
t
h
e
C
OC
O
-
m
in
i
d
ataset.
T
h
ese
f
in
d
in
g
s
s
u
g
g
est
th
at
d
ata
-
d
r
iv
e
n
s
tr
u
ctu
r
al
en
h
an
c
em
en
ts
ar
e
cr
itical
f
o
r
d
ep
lo
y
in
g
h
ig
h
-
p
er
f
o
r
m
a
n
ce
d
etec
to
r
s
in
r
ea
l
-
wo
r
ld
s
u
r
v
eillan
ce
an
d
a
u
to
n
o
m
o
u
s
s
y
s
tem
s
.
Fu
tu
r
e
r
esear
ch
s
h
o
u
ld
ex
ten
d
th
is
o
p
tim
ized
ar
ch
i
tectu
r
e
to
in
s
tan
ce
s
eg
m
en
tatio
n
an
d
lo
w
-
p
o
wer
h
ar
d
war
e
d
ep
lo
y
m
en
t to
e
n
h
a
n
ce
its
r
ea
l
-
tim
e
o
p
er
atio
n
al
im
p
ac
t.
F
UNDING
I
NF
O
R
M
A
T
I
O
N
T
h
is
wo
r
k
was
s
u
p
p
o
r
ted
b
y
T
ay
lo
r
'
s
Un
iv
er
s
ity
,
Ma
lay
s
ia,
th
r
o
u
g
h
its
in
s
titu
tio
n
al
p
u
b
licatio
n
f
u
n
d
in
g
s
ch
em
e.
AUTHO
R
CO
NT
RI
B
UT
I
O
NS ST
A
T
E
M
E
N
T
T
h
is
jo
u
r
n
al
u
s
es
th
e
C
o
n
tr
ib
u
to
r
R
o
les
T
ax
o
n
o
m
y
(
C
R
ed
iT)
to
r
ec
o
g
n
ize
in
d
iv
id
u
al
au
th
o
r
co
n
tr
ib
u
tio
n
s
,
r
ed
u
ce
au
th
o
r
s
h
ip
d
is
p
u
tes,
an
d
f
ac
ilit
ate
co
llab
o
r
atio
n
.
Na
m
e
o
f
Aut
ho
r
C
M
So
Va
Fo
I
R
D
O
E
Vi
Su
P
Fu
H
u
s
n
a
S
a
r
i
r
a
h
H
u
s
i
n
✓
✓
✓
✓
✓
✓
H
o
w
a
r
d
C
h
o
n
g
Y
u
n
H
a
o
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
P
a
n
Y
u
a
n
F
ei
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
M
o
h
s
e
n
M
a
r
j
a
n
i
✓
✓
✓
✓
S
u
r
i
a
n
a
I
s
m
ai
l
✓
✓
C
:
C
o
n
c
e
p
t
u
a
l
i
z
a
t
i
o
n
M
:
M
e
t
h
o
d
o
l
o
g
y
So
:
So
f
t
w
a
r
e
Va
:
Va
l
i
d
a
t
i
o
n
Fo
:
Fo
r
mal
a
n
a
l
y
s
i
s
I
:
I
n
v
e
s
t
i
g
a
t
i
o
n
R
:
R
e
so
u
r
c
e
s
D
:
D
a
t
a
C
u
r
a
t
i
o
n
O
:
W
r
i
t
i
n
g
-
O
r
i
g
i
n
a
l
D
r
a
f
t
E
:
W
r
i
t
i
n
g
-
R
e
v
i
e
w
&
E
d
i
t
i
n
g
Vi
:
Vi
su
a
l
i
z
a
t
i
o
n
Su
:
Su
p
e
r
v
i
s
i
o
n
P
:
P
r
o
j
e
c
t
a
d
mi
n
i
st
r
a
t
i
o
n
Fu
:
Fu
n
d
i
n
g
a
c
q
u
i
si
t
i
o
n
CO
NF
L
I
C
T
O
F
I
N
T
E
R
E
S
T
ST
A
T
E
M
E
NT
Au
th
o
r
s
s
tate
n
o
co
n
f
lict o
f
in
t
er
est.
DATA AV
AI
L
AB
I
L
I
T
Y
T
h
e
d
ata
th
at
s
u
p
p
o
r
t
th
e
f
i
n
d
in
g
s
o
f
th
is
s
tu
d
y
a
r
e
o
p
e
n
ly
av
ailab
le
in
th
e
T
in
y
Per
s
o
n
d
ataset
r
ep
o
s
ito
r
y
at
h
ttp
s
://g
ith
u
b
.
co
m
/C
ap
tain
E
v
en
/Tin
y
Per
s
o
n
an
d
t
h
e
MS
C
OC
O
d
ataset
r
ep
o
s
ito
r
y
a
t
h
ttp
s
://co
co
d
ataset.
o
r
g
.
RE
F
E
R
E
NC
E
S
[
1
]
Y
.
Y
a
o
e
t
a
l
.
,
“
O
n
i
mp
r
o
v
i
n
g
b
o
u
n
d
i
n
g
b
o
x
r
e
p
r
e
s
e
n
t
a
t
i
o
n
s
f
o
r
o
r
i
e
n
t
e
d
o
b
j
e
c
t
d
e
t
e
c
t
i
o
n
,
”
I
EEE
T
r
a
n
s
a
c
t
i
o
n
s
o
n
G
e
o
s
c
i
e
n
c
e
a
n
d
Re
m
o
t
e
S
e
n
s
i
n
g
,
v
o
l
.
6
1
,
p
p
.
1
–
1
1
,
2
0
2
3
,
d
o
i
:
1
0
.
1
1
0
9
/
TG
R
S
.
2
0
2
2
.
3
2
3
1
3
4
0
.
[
2
]
N
.
U
.
A
.
T
a
h
i
r
,
Z
.
Z
h
a
n
g
,
M
.
A
s
i
m,
J
.
C
h
e
n
,
a
n
d
M
.
E
LA
f
f
e
n
d
i
,
“
O
b
j
e
c
t
d
e
t
e
c
t
i
o
n
i
n
a
u
t
o
n
o
mo
u
s
v
e
h
i
c
l
e
s
u
n
d
e
r
a
d
v
e
r
se
w
e
a
t
h
e
r
:
a
r
e
v
i
e
w
o
f
t
r
a
d
i
t
i
o
n
a
l
a
n
d
d
e
e
p
l
e
a
r
n
i
n
g
a
p
p
r
o
a
c
h
e
s,
”
A
l
g
o
ri
t
h
m
s
,
v
o
l
.
1
7
,
n
o
.
3
,
F
e
b
.
2
0
2
4
,
d
o
i
:
1
0
.
3
3
9
0
/
a
1
7
0
3
0
1
0
3
.
[
3
]
C
.
A
l
b
u
q
u
e
r
q
u
e
,
R
.
H
e
n
r
i
q
u
e
s
,
a
n
d
M
.
C
a
st
e
l
l
i
,
“
D
e
e
p
l
e
a
r
n
i
n
g
-
b
a
s
e
d
o
b
j
e
c
t
d
e
t
e
c
t
i
o
n
a
l
g
o
r
i
t
h
ms
i
n
me
d
i
c
a
l
i
mag
i
n
g
:
s
y
st
e
ma
t
i
c
r
e
v
i
e
w
,
”
H
e
l
i
y
o
n
,
v
o
l
.
1
1
,
n
o
.
1
,
J
a
n
.
2
0
2
5
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
h
e
l
i
y
o
n
.
2
0
2
4
.
e
4
1
1
3
7
.
[
4
]
A
.
P
o
l
u
k
h
i
n
,
Y
.
G
o
r
d
i
e
n
k
o
,
G
.
Jer
v
a
n
,
a
n
d
S
.
S
t
i
r
e
n
k
o
,
“
O
b
j
e
c
t
d
e
t
e
c
t
i
o
n
f
o
r
r
e
sc
u
e
o
p
e
r
a
t
i
o
n
s
b
y
h
i
g
h
-
a
l
t
i
t
u
d
e
i
n
f
r
a
r
e
d
t
h
e
r
mal
i
ma
g
i
n
g
c
o
l
l
e
c
t
e
d
b
y
u
n
ma
n
n
e
d
a
e
r
i
a
l
v
e
h
i
c
l
e
s
,
”
i
n
P
a
t
t
e
r
n
Re
c
o
g
n
i
t
i
o
n
a
n
d
I
m
a
g
e
A
n
a
l
y
s
i
s
(
I
b
PRIA
2
0
2
3
)
,
C
h
a
m,
S
w
i
t
z
e
r
l
a
n
d
:
S
p
r
i
n
g
e
r
,
2
0
2
3
,
p
p
.
4
9
0
–
5
0
4
,
d
o
i
:
1
0
.
1
0
0
7
/
9
7
8
-
3
-
0
3
1
-
3
6
6
1
6
-
1
_
3
9
.
[
5
]
B
.
M
i
r
z
a
e
i
,
H
.
N
e
z
a
m
a
b
a
d
i
-
p
o
u
r
,
A
.
R
a
o
o
f
,
a
n
d
R
.
D
e
r
a
k
h
s
h
a
n
i
,
“
S
m
a
l
l
o
b
j
e
c
t
d
e
t
e
c
t
i
o
n
a
n
d
t
r
a
c
k
i
n
g
:
a
c
o
m
p
r
e
h
e
n
s
i
v
e
r
e
v
i
e
w
,
”
S
e
n
so
rs
,
v
o
l
.
2
3
,
n
o
.
1
5
,
A
u
g
.
2
0
2
3
,
d
o
i
:
1
0
.
3
3
9
0
/
s2
3
1
5
6
8
8
7
.
[
6
]
J.
D
o
h
e
r
t
y
,
B
.
G
a
r
d
i
n
e
r
,
E.
K
e
r
r
,
a
n
d
N
.
S
i
d
d
i
q
u
e
,
“
B
i
F
P
N
-
Y
O
LO
:
o
n
e
-
st
a
g
e
o
b
j
e
c
t
d
e
t
e
c
t
i
o
n
i
n
t
e
g
r
a
t
i
n
g
b
i
-
d
i
r
e
c
t
i
o
n
a
l
f
e
a
t
u
r
e
p
y
r
a
mi
d
n
e
t
w
o
r
k
s,”
P
a
t
t
e
r
n
R
e
c
o
g
n
i
t
i
o
n
,
v
o
l
.
1
6
0
,
A
p
r
.
2
0
2
5
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
p
a
t
c
o
g
.
2
0
2
4
.
1
1
1
2
0
9
.
[
7
]
H
.
Z
h
a
n
g
,
Q
.
D
u
,
Q
.
Q
i
,
J.
Z
h
a
n
g
,
F
.
W
a
n
g
,
a
n
d
M
.
G
a
o
,
“
A
r
e
c
u
r
s
i
v
e
a
t
t
e
n
t
i
o
n
-
e
n
h
a
n
c
e
d
b
i
d
i
r
e
c
t
i
o
n
a
l
f
e
a
t
u
r
e
p
y
r
a
mi
d
n
e
t
w
o
r
k
f
o
r
sm
a
l
l
o
b
j
e
c
t
d
e
t
e
c
t
i
o
n
,
”
Mu
l
t
i
m
e
d
i
a
T
o
o
l
s
a
n
d
A
p
p
l
i
c
a
t
i
o
n
s
,
v
o
l
.
8
2
,
n
o
.
9
,
p
p
.
1
3
9
9
9
–
1
4
0
1
8
,
A
p
r
.
2
0
2
3
,
d
o
i
:
1
0
.
1
0
0
7
/
s
1
1
0
4
2
-
022
-
1
3
9
5
1
-
4.
[
8
]
S
.
Li
u
,
L.
Q
i
,
H
.
Q
i
n
,
J
.
S
h
i
,
a
n
d
J
.
J
i
a
,
“
P
a
t
h
a
g
g
r
e
g
a
t
i
o
n
n
e
t
w
o
r
k
f
o
r
i
n
s
t
a
n
c
e
s
e
g
m
e
n
t
a
t
i
o
n
,
”
i
n
2
0
1
8
I
EEE
/
C
VF
C
o
n
f
e
re
n
c
e
o
n
C
o
m
p
u
t
e
r
Vi
s
i
o
n
a
n
d
P
a
t
t
e
r
n
Re
c
o
g
n
i
t
i
o
n
,
J
u
n
.
2
0
1
8
,
p
p
.
8
7
5
9
–
8
7
6
8
,
d
o
i
:
1
0
.
1
1
0
9
/
C
V
P
R
.
2
0
1
8
.
0
0
9
1
3
.
[
9
]
A
.
I
.
O
n
t
h
o
n
i
a
n
d
P
.
K
.
S
a
h
o
o
,
“
I
n
st
a
n
c
e
s
e
g
m
e
n
t
a
t
i
o
n
b
a
s
e
d
o
b
j
e
c
t
d
e
t
e
c
t
i
o
n
w
i
t
h
e
n
h
a
n
c
e
d
p
a
t
h
a
g
g
r
e
g
a
t
i
o
n
n
e
t
w
o
r
k
,
”
i
n
2
0
2
3
I
EEE
L
a
t
i
n
Am
e
r
i
c
a
n
C
o
n
f
e
ren
c
e
o
n
C
o
m
p
u
t
a
t
i
o
n
a
l
I
n
t
e
l
l
i
g
e
n
c
e
(
L
A
-
C
C
I
)
,
O
c
t
.
2
0
2
3
,
p
p
.
1
–
6
,
d
o
i
:
1
0
.
1
1
0
9
/
LA
-
C
C
I
5
8
5
9
5
.
2
0
2
3
.
1
0
4
0
9
4
7
9
.
[
1
0
]
B
.
Zh
a
n
g
a
n
d
Y
.
Z
h
a
n
g
,
“
M
S
C
V
i
T:
a
sma
l
l
-
s
i
z
e
V
i
T
a
r
c
h
i
t
e
c
t
u
r
e
w
i
t
h
m
u
l
t
i
-
sc
a
l
e
se
l
f
-
a
t
t
e
n
t
i
o
n
me
c
h
a
n
i
sm
f
o
r
t
i
n
y
d
a
t
a
set
s
,
”
N
e
u
ra
l
N
e
t
w
o
rks
,
v
o
l
.
1
8
8
,
A
u
g
.
2
0
2
5
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
n
e
u
n
e
t
.
2
0
2
5
.
1
0
7
4
9
9
.
Evaluation Warning : The document was created with Spire.PDF for Python.