I
AE
S
I
nte
rna
t
io
na
l J
o
urna
l o
f
Ro
bo
t
ics a
nd
Aut
o
m
a
t
io
n (
I
J
RA)
Vo
l.
1
0
,
No
.
2
,
J
u
n
e
2
0
2
1
,
p
p
.
1
3
3
~1
4
3
I
SS
N:
2722
-
2586
,
DOI
: 1
0
.
1
1
5
9
1
/
i
j
r
a
.
v
1
0
i
2
.
p
p
1
3
3
-
143
133
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
r
a
.
ia
esco
r
e.
co
m
Dete
c
ting A
frican
hoo
fed
a
ni
m
a
ls
i
n aeria
l i
m
a
g
ery
using
co
nv
o
lutiona
l neural netw
o
rk
Yunfe
i F
a
ng
1
,
Sh
eng
zhi D
u
2
,
L
a
rbi B
o
ub
chir
3
,
K
a
ri
m
Dj
o
ua
ni
4
1,
2,
4
De
p
a
rtm
e
n
t
o
f
El
e
c
tri
c
a
l
En
g
in
e
e
rin
g
,
T
sh
w
a
n
e
Un
iv
e
rsit
y
o
f
Tec
h
n
o
l
o
g
y
,
S
o
u
th
A
f
rica
3
L
I
A
S
D,
Un
iv
e
rsit
y
o
f
P
a
ris
8
,
F
r
a
n
c
e
Art
icle
I
nfo
AB
ST
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Au
g
31
,
2
0
2
0
R
ev
i
s
ed
Dec
1
,
2
0
2
0
A
cc
ep
ted
Feb
1
2
,
2
0
2
1
S
m
a
ll
u
n
m
a
n
n
e
d
a
e
rial
v
e
h
icle
s
a
p
p
li
c
a
ti
o
n
s
h
a
d
e
ru
p
ted
i
n
m
a
n
y
f
ield
s
in
c
lu
d
in
g
c
o
n
se
rv
a
ti
o
n
m
a
n
a
g
e
m
e
n
t.
A
u
to
m
a
ti
c
o
b
jec
t
d
e
t
e
c
ti
o
n
m
e
th
o
d
s
f
o
r
su
c
h
a
e
rial
ima
g
e
r
y
w
e
r
e
in
h
ig
h
d
e
m
a
n
d
to
f
a
c
il
it
a
te
m
o
re
e
fficie
n
t
a
n
d
e
c
o
n
o
m
ica
l
w
il
d
li
f
e
m
a
n
a
g
e
m
e
n
t
a
n
d
re
se
a
rc
h
.
T
h
is
p
a
p
e
r
a
i
m
e
d
to
d
e
tec
t
h
o
o
f
e
d
a
n
im
a
ls
in
a
e
rial
i
m
a
g
e
s
t
a
k
e
n
f
ro
m
a
q
u
a
d
-
ro
to
r
i
n
S
o
u
th
e
rn
Af
rica
.
Ob
jec
ts
c
a
p
tu
re
d
in
th
is
w
a
y
w
e
r
e
s
m
a
ll
b
o
th
i
n
a
b
so
l
u
te
p
ix
e
ls
a
n
d
f
ro
m
a
n
o
b
jec
t
-
to
-
im
a
g
e
ra
ti
o
p
o
in
t
o
f
v
iew
,
w
h
ich
we
re
n
o
t
p
e
rf
e
c
tl
y
su
it
fo
r
g
e
n
e
ra
l
p
u
r
p
o
se
d
o
b
jec
t
d
e
tec
to
rs.
W
e
p
ro
p
o
se
d
a
m
e
th
o
d
b
a
se
d
o
n
t
h
e
ico
n
ic
F
a
ste
r
re
g
io
n
-
b
a
se
d
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
s
(
R
-
CNN
)
f
r
a
m
e
w
o
rk
w
it
h
a
tro
u
s
c
o
n
v
o
lu
ti
o
n
lay
e
rs
in
o
rd
e
r
to
re
t
a
in
th
e
sp
a
ti
a
l
re
so
lu
ti
o
n
o
f
th
e
fe
a
tu
re
m
a
p
to
d
e
tec
t
sm
a
ll
o
b
jec
ts.
A
g
o
o
d
c
h
o
ice
o
f
a
n
c
h
o
rs
w
a
s
o
f
p
rim
e
i
m
p
o
rtan
c
e
i
n
d
e
tec
ti
n
g
sm
a
ll
o
b
jec
ts.
T
h
e
p
e
rf
o
rm
a
n
c
e
o
f
th
e
p
ro
p
o
se
d
F
a
ste
r
R
-
CN
N w
it
h
a
tro
u
s
c
o
n
v
o
l
u
ti
o
n
a
l
f
il
ters
in
th
e
b
a
c
k
b
o
n
e
n
e
tw
o
rk
wa
s
p
ro
v
e
n
to
b
e
o
u
tstan
d
in
g
in
o
u
r
sc
e
n
a
rio
b
y
c
o
m
p
a
rin
g
to
o
th
e
r
o
b
jec
t
d
e
tec
ti
o
n
a
rc
h
it
e
c
tu
re
s.
K
ey
w
o
r
d
s
:
An
c
h
o
r
d
esig
n
An
i
m
al
d
etec
tio
n
A
tr
o
u
s
co
n
v
o
lu
t
io
n
Fas
ter
R
-
C
NN
S
m
all
o
b
j
ec
t d
etec
tio
n
T
h
is
is
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
:
Yu
n
f
ei
Fan
g
Dep
ar
t
m
en
t o
f
E
lectr
ical
E
n
g
i
n
ee
r
in
g
T
s
h
w
a
n
e
U
n
iv
er
s
it
y
o
f
T
ec
h
n
o
lo
g
y
Staats
ar
till
er
ie
R
o
a
d
,
P
r
eto
r
ia
w
e
s
t,
P
r
eto
r
ia,
0
0
0
1
,
So
u
th
Af
r
ica
E
m
ail:
f
an
g
y
u
n
f
ei0
8
@
1
6
3
.
co
m
1.
I
NT
RO
D
UCT
I
O
N
Un
m
an
n
ed
ae
r
ial
v
eh
ic
les
(
U
A
V
s
)
as
a
co
n
v
e
n
ie
n
t
a
n
d
ea
s
y
-
to
-
g
et
d
ata
ac
q
u
is
i
tio
n
to
o
l
h
as
b
ee
n
ap
p
lied
in
m
an
y
w
ild
li
f
e
co
n
s
er
v
atio
n
a
n
d
r
esear
ch
tas
k
s
.
Au
to
m
a
tic
a
n
al
y
s
i
s
o
f
s
u
c
h
a
er
ial
i
m
a
g
er
y
is
o
f
s
ig
n
i
f
ica
n
t i
m
p
o
r
ta
n
ce
as
s
u
c
h
d
ata
a
m
o
u
n
ts
d
r
a
m
atica
ll
y
.
O
b
j
ec
t d
etec
tio
n
f
o
r
m
s
t
h
e
b
asi
s
o
f
m
a
n
y
co
m
p
u
ter
v
is
io
n
ap
p
licatio
n
s
.
W
o
r
k
h
as
b
ee
n
d
o
n
e
in
d
etec
ti
n
g
ter
r
es
tr
ial
[
1
]
,
m
ar
i
n
e
[
2
]
,
an
d
ce
lest
ial
[
3
,
4
]
s
p
ec
ies
f
r
o
m
ae
r
ial
i
m
ag
er
y
in
d
i
f
f
er
e
n
t
en
v
ir
o
n
m
e
n
ts
.
Af
r
ica
h
o
ld
s
a
v
ar
iet
y
o
f
u
n
iq
u
e
h
o
o
f
ed
w
il
d
lif
e
s
p
ec
ies,
an
d
a
n
u
m
b
er
o
f
th
e
m
ar
e
u
n
d
e
r
th
r
ea
t,
s
o
m
e
at
cr
itical
r
is
k
o
f
ex
ti
n
ctio
n
.
U
A
V
s
co
m
b
in
ed
w
i
th
co
m
p
u
ter
v
is
io
n
tech
n
iq
u
es c
a
n
as
s
is
t c
o
n
s
er
v
a
tio
n
w
o
r
k
er
s
an
d
r
esear
c
h
er
s
t
o
a
g
r
ea
t e
x
ten
t.
C
o
m
p
u
ter
v
is
io
n
ap
p
licatio
n
s
d
if
f
er
f
r
o
m
ca
s
e
to
ca
s
e
d
ep
en
d
in
g
o
n
th
e
s
ce
n
ar
io
an
d
u
n
iq
u
e
ch
ar
ac
ter
is
tic
s
o
f
t
h
e
p
r
o
v
id
ed
d
ataset
.
C
h
al
len
g
es
o
f
o
b
j
ec
t
d
etec
tio
n
i
n
ae
r
ial
i
m
ag
er
y
w
er
e
s
u
m
m
ar
is
ed
i
n
[
5
]
,
s
u
ch
as
s
m
all
o
b
j
ec
t
s
ize,
lar
g
e
s
ca
le
v
ar
iatio
n
s
,
cr
o
w
d
ed
in
s
ta
n
ce
s
a
n
d
v
ar
io
u
s
o
r
ie
n
tatio
n
s
.
I
n
a
m
o
n
g
w
h
ic
h
th
e
m
o
s
t
ch
alle
n
g
i
n
g
i
s
th
e
s
ize
an
d
s
ca
le
p
r
o
b
le
m
.
T
h
e
UA
V
h
as
to
o
p
er
ate
at
a
ce
r
ta
in
altitu
d
e
in
o
r
d
er
to
p
r
o
v
id
e
a
b
ig
f
ield
o
f
v
ie
w
,
av
o
id
d
is
t
u
r
b
in
g
t
h
e
tar
g
eted
an
d
o
th
er
lo
ca
l
s
p
ec
ies.
T
h
e
d
is
tan
ce
m
a
k
e
s
th
e
a
n
i
m
als
ca
p
t
u
r
ed
in
s
m
all
s
ca
le.
T
h
e
f
u
ll
i
m
a
g
e
r
eso
l
u
t
io
n
s
h
o
u
ld
b
e
b
ig
e
n
o
u
g
h
to
r
etain
t
h
e
ab
s
o
lu
te
o
b
j
ec
t
s
ize
i
n
p
ix
e
ls
s
o
t
h
at
it
ca
r
r
ies
s
u
f
f
icie
n
t
in
f
o
r
m
at
io
n
.
T
h
is
m
a
k
es
th
e
r
atio
o
f
th
e
o
b
j
ec
t
s
ize
to
th
e
f
u
ll
i
m
a
g
e
s
ize
q
u
i
te
s
m
all
co
m
p
ar
ed
to
o
b
j
ec
ts
in
g
r
o
u
n
d
-
p
er
ce
iv
ed
i
m
a
g
er
y
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
2586
I
A
E
S
I
n
t
J
R
o
b
&
A
u
to
m
,
Vo
l
.
1
0
,
No
.
2
,
J
u
n
e
2
0
2
1
:
1
3
3
–
143
134
T
o
illu
s
tr
ate
th
is
,
s
u
m
m
ar
ie
s
o
f
th
e
o
b
j
ec
t
s
ize
m
ea
s
u
r
ed
b
y
t
h
e
b
o
u
n
d
i
n
g
-
b
o
x
ar
ea
s
ar
e
g
iv
e
n
i
n
Fig
u
r
e
1
.
Ou
r
d
ataset
w
a
s
co
m
p
ar
ed
to
th
e
p
o
p
u
lar
g
en
er
ic
o
b
j
ec
t
d
etec
tio
n
d
atasets
P
A
SC
AL
-
VO
C
[
6
]
an
d
MS
-
C
O
C
O
[
7
]
,
w
h
ic
h
co
n
tai
n
s
e
v
er
y
d
a
y
o
b
j
ec
ts
s
u
c
h
a
s
c
ar
s
,
p
ed
estrian
s
a
n
d
p
ets
.
Fig
u
r
e
1
(
a)
s
h
o
w
s
t
h
e
h
is
to
g
r
a
m
o
f
t
h
e
ab
s
o
l
u
te
o
b
j
ec
t
s
ize
in
p
ix
el
s
.
Ob
j
ec
t
s
ize
w
a
s
r
ep
r
esen
ted
b
y
th
e
s
q
u
ar
e
r
o
o
t
o
f
t
h
e
b
o
u
n
d
in
g
-
b
o
x
ar
ea
.
T
h
e
b
in
s
ar
e
n
o
r
m
a
lized
to
co
m
p
ar
e
th
e
d
atase
ts
o
n
t
h
e
s
a
m
e
s
ca
le.
Ma
j
o
r
ity
o
f
t
h
e
o
b
j
ec
ts
in
o
u
r
w
i
ld
lif
e
d
ata
s
e
t
lies
i
n
ed
g
e
le
n
g
t
h
n
o
b
i
g
g
er
th
an
1
2
8
p
ix
el
s
.
W
h
ile
t
h
e
o
b
j
ec
t
s
ize
o
f
o
u
r
d
ataset
is
s
m
aller
th
a
n
th
e
o
t
h
er
d
ataset
s
,
t
h
e
f
u
ll
i
m
a
g
e
s
ize
is
m
u
c
h
b
i
g
g
er
.
T
h
i
s
m
a
k
es
th
e
r
atio
o
f
t
h
e
o
b
j
ec
t
s
ize
a
g
ai
n
s
t
th
e
f
u
ll
i
m
ag
e
s
i
ze
e
v
en
s
m
aller
,
a
s
s
h
o
w
n
o
n
F
ig
u
r
e
1
(
b
)
.
Fi
g
u
r
e
2
s
h
o
w
s
s
o
m
e
ex
a
m
p
le
ze
b
r
as
cr
o
p
p
e
d
f
r
o
m
t
h
e
ae
r
i
al
i
m
a
g
er
y
w
e
ca
p
t
u
r
ed
f
r
o
m
th
e
w
ild
n
atu
r
e.
Fi
g
u
r
e
2
(
a)
s
h
o
w
s
t
h
e
v
ar
io
u
s
o
r
ien
tatio
n
s
o
f
t
h
e
ze
b
r
as
f
r
o
m
t
h
e
b
ir
d
’
s
e
y
e
v
ie
w
.
I
n
Fi
g
u
r
e
2
(
b
)
,
th
e
ze
b
r
as
w
er
e
f
o
u
n
d
d
i
f
f
icu
l
t
to
b
e
d
is
tin
g
u
is
h
ed
f
r
o
m
o
t
h
er
an
i
m
als d
u
e
to
th
e
s
m
all
s
ca
le
an
d
i
llu
m
i
n
atio
n
co
n
d
it
io
n
s
.
(
a)
(
b
)
Fig
u
r
e
1
.
C
h
ar
ac
ter
is
t
ics o
f
o
b
j
ec
t size
in
o
u
r
d
ataset,
P
A
S
C
AL
-
VO
C
,
an
d
MS
-
C
O
C
O
:
h
is
to
g
r
a
m
o
f
(
a)
ab
s
o
lu
te
o
b
j
e
ct
s
ize
in
ed
g
e
le
n
g
t
h
an
d
(
b
)
r
elativ
e
o
b
j
ec
t size
(
o
b
j
ec
t size/i
m
a
g
e
s
ize)
,
s
h
o
w
i
n
g
o
b
j
ec
ts
in
o
u
r
d
ataset
o
cc
u
p
ies o
n
l
y
l
ittl
e
p
o
r
tio
n
o
f
th
e
f
u
ll i
m
a
g
e
(
a)
(
b
)
Fig
u
r
e
2
.
E
x
a
m
p
le
o
f
ze
b
r
as c
r
o
p
p
e
d
f
r
o
m
o
u
r
d
ataset
:
(
a)
d
if
f
er
en
t o
r
ien
tat
io
n
s
o
f
th
e
ze
b
r
as f
r
o
m
ae
r
ial
p
er
s
p
ec
tiv
e
an
d
(
b
)
ze
b
r
as in
d
is
tan
ce
u
n
d
er
v
ar
io
u
s
ill
u
m
i
n
a
tio
n
Dee
p
lear
n
in
g
f
r
a
m
e
w
o
r
k
s
s
u
ch
as
r
ec
u
r
r
e
n
t
n
e
u
r
al
n
e
t
w
o
r
k
(
R
NN)
[
8
]
an
d
co
n
v
o
l
u
tio
n
al
n
eu
r
al
n
et
w
o
r
k
(
C
NN
)
[
9
]
h
av
e
b
o
o
s
ted
m
ac
h
i
n
e
lear
n
in
g
ap
p
licati
o
n
s
to
an
o
t
h
er
lev
el
i
n
r
ec
en
t
y
ea
r
s
.
R
e
s
ea
r
ch
i
n
f
ield
s
lik
e
n
a
tu
r
e
la
n
g
u
a
g
e
p
r
o
ce
s
s
in
g
(
N
L
P
)
[
1
0
]
,
m
ac
h
i
n
e
tr
an
s
la
tio
n
[
1
1
]
,
an
d
co
m
p
u
ter
v
is
io
n
[
1
2
,
1
3
]
h
av
e
b
ee
n
d
o
m
in
ated
b
y
d
ee
p
lear
n
in
g
.
T
h
e
w
o
r
d
“
d
ee
p
”
r
ef
er
s
to
th
e
d
ep
th
o
f
i
n
f
o
r
m
at
io
n
th
at
t
h
e
n
e
u
r
al
n
et
w
o
r
k
ex
tr
ac
ts
f
r
o
m
t
h
e
r
a
w
d
ata.
Mo
d
er
n
C
NN
f
ac
ilit
ated
o
b
j
ec
t
d
etec
to
r
s
d
er
iv
e
th
e
f
ea
t
u
r
e
r
ep
r
esen
tatio
n
s
b
y
s
tac
k
in
g
a
s
eq
u
en
ce
o
f
co
n
v
o
l
u
tio
n
a
l
la
y
er
s
o
v
er
th
e
in
p
u
t
i
m
a
g
e.
Du
r
i
n
g
t
h
e
p
r
o
ce
s
s
,
th
e
f
ea
t
u
r
e
m
ap
s
ize
w
as
co
n
tin
u
o
u
s
l
y
r
ed
u
ce
d
to
e
x
tr
ac
t
m
o
r
e
ab
s
tr
ac
t
in
f
o
r
m
a
tio
n
a
n
d
allo
w
tr
a
n
s
lat
io
n
-
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t
J
R
o
b
&
A
u
to
m
I
SS
N:
2722
-
2586
Dete
ctin
g
A
fr
ica
n
h
o
o
fe
d
a
n
i
ma
ls
in
a
eria
l ima
g
ery
u
s
in
g
c
o
n
vo
lu
tio
n
a
l n
e
u
r
a
l n
etw
o
r
k
(
Yu
n
fei
F
a
n
g
)
135
in
v
ar
ia
n
ce
.
T
h
is
w
a
s
n
o
r
m
all
y
ac
h
iev
ed
b
y
s
tr
id
ed
co
n
v
o
lu
ti
o
n
o
r
m
ax
-
p
o
o
lin
g
.
S
m
all
o
b
jects
w
er
e
f
o
u
n
d
to
b
e
d
if
f
ic
u
lt
to
h
a
n
d
le
as
t
h
eir
lo
ca
tio
n
in
f
o
r
m
atio
n
co
u
ld
b
e
lo
s
t
af
ter
d
o
w
n
-
s
a
m
p
lin
g
.
T
o
d
etec
t
th
e
s
m
all
o
b
j
ec
ts
,
th
e
f
ea
tu
r
e
m
ap
h
as
to
r
etain
a
r
ea
s
o
n
ab
le
r
eso
l
u
tio
n
,
a
n
d
at
th
e
s
a
m
e
ti
m
e
b
ein
g
s
e
m
a
n
tical
l
y
ex
p
r
ess
i
v
e.
Fo
r
t
h
is
p
u
r
p
o
s
e,
w
e
ap
p
lied
a
s
eq
u
en
ce
o
f
atr
o
u
s
co
n
v
o
l
u
tio
n
a
l
la
y
er
s
to
k
ee
p
th
e
f
ea
tu
r
e
m
ap
at
th
e
d
es
ir
ed
r
eso
lu
tio
n
.
T
h
is
f
its
t
h
e
s
i
n
g
le
-
la
y
er
r
e
p
r
ese
n
ta
tio
n
t
h
at
d
etec
ts
o
b
j
ec
ts
o
f
a
ll
s
ca
le
s
f
r
o
m
o
n
e
s
in
g
le
f
ea
t
u
r
e
la
y
er
,
s
u
c
h
as
t
h
e
t
w
o
-
s
tag
e
f
a
s
ter
r
eg
io
n
b
as
ed
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
e
t
wo
r
k
s
(
R
-
C
NN
)
.
T
h
e
m
atc
h
in
g
q
u
al
ities
o
f
t
h
e
a
n
ch
o
r
s
an
d
g
r
o
u
n
d
-
tr
u
t
h
w
as
f
o
u
n
d
to
b
e
h
ig
h
l
y
co
r
r
elate
d
to
th
e
d
etec
tio
n
p
er
f
o
r
m
a
n
ce
.
A
f
i
n
e
f
ea
t
u
r
e
s
tr
id
e
an
d
co
m
p
r
eh
e
n
s
i
v
e
s
et
o
f
an
c
h
o
r
s
h
elp
s
i
m
p
r
o
v
e
d
et
ec
tio
n
p
er
f
o
r
m
a
n
ce
,
esp
ec
iall
y
f
o
r
th
e
s
m
all
o
b
j
ec
ts
.
Ho
w
e
v
er
,
it
is
a
tr
ad
e
-
o
f
f
b
e
t
w
ee
n
d
etec
tio
n
ac
c
u
r
ac
y
a
n
d
co
m
p
u
tatio
n
co
s
t.
Fas
ter
R
-
C
NN
[
1
4
]
is
t
h
e
r
e
p
r
esen
tati
v
e
o
f
a
“
t
w
o
-
s
ta
g
e”
o
b
j
ec
t
d
etec
to
r
,
w
h
ic
h
tr
ai
n
e
d
a
r
eg
io
n
p
r
o
p
o
s
al
n
et
w
o
r
k
(
R
P
N
)
to
g
en
er
ate
o
b
j
ec
t
ca
n
d
id
ates.
T
h
e
ca
n
d
id
ates
w
er
e
th
en
p
ass
ed
o
n
to
an
o
th
er
n
et
w
o
r
k
f
o
r
m
u
lt
i
-
cla
s
s
c
lass
i
f
icatio
n
an
d
b
o
u
n
d
in
g
-
b
o
x
f
i
n
e
-
tu
n
i
n
g
.
I
n
t
h
e
s
ec
o
n
d
s
ta
g
e,
“ROI
ali
g
n
m
e
n
t”
[
1
5
]
(
“
R
OI
p
o
o
lin
g
”
i
n
ea
r
l
y
v
er
s
io
n
)
cr
o
p
p
ed
th
e
f
ea
t
u
r
es
f
o
r
th
e
o
b
j
ec
t
p
r
o
p
o
s
als
an
d
f
it
t
h
e
m
in
to
th
e
s
a
m
e
s
ize.
“An
ch
o
r
s
”
w
er
e
a
s
et
o
f
p
r
e
-
d
ef
i
n
ed
b
o
u
n
d
i
n
g
-
b
o
x
e
s
th
at
s
er
v
e
a
s
o
b
j
ec
t
p
r
o
p
o
s
als
f
o
r
t
h
e
R
P
N.
T
h
e
o
u
tp
u
t
b
o
u
n
d
i
n
g
-
b
o
x
w
er
e
d
er
iv
ed
b
y
p
r
ed
ictin
g
th
e
“
o
f
f
s
et
s
”
to
th
e
an
c
h
o
r
s
.
Du
r
in
g
tr
ain
i
n
g
,
an
ch
o
r
s
g
en
er
ate
p
o
s
iti
v
e
an
d
n
eg
at
iv
e
ex
a
m
p
le
s
ac
co
r
d
in
g
to
its
i
n
te
r
s
ec
tio
n
w
it
h
th
e
g
r
o
u
n
d
-
tr
u
t
h
.
Sk
ip
p
in
g
t
h
e
p
r
o
p
o
s
al
g
e
n
er
at
in
g
p
r
o
ce
s
s
,
s
i
n
g
le
s
h
o
t
o
b
j
ec
t
d
etec
to
r
(
SS
D
)
p
r
o
p
o
s
ed
to
m
ak
e
f
in
al
p
r
ed
ictio
n
s
o
n
clas
s
lab
el
an
d
b
o
u
n
d
in
g
-
b
o
x
co
o
r
d
in
ates
o
f
f
s
ets
d
ir
ec
tl
y
f
r
o
m
th
e
f
e
at
u
r
e
m
ap
s
,
u
n
li
k
e
f
aster
R
-
C
NN
t
h
at
h
a
n
d
les
o
b
j
ec
ts
o
f
all
s
ca
les
o
n
th
e
s
a
m
e
f
ea
t
u
r
e
m
ap
.
SS
D
w
o
r
k
ed
o
n
a
h
ier
ar
ch
ical
f
ea
tu
r
e
p
y
r
a
m
id
a
n
d
ea
ch
f
ea
t
u
r
e
la
y
er
w
a
s
d
esi
g
n
ated
to
o
b
j
ec
ts
o
f
o
n
e
s
ca
le.
YO
L
O
(
Yo
u
O
n
l
y
L
o
o
k
O
n
ce
)
[
1
6
]
d
iv
id
ed
th
e
i
m
a
g
e
in
to
a
g
r
id
.
E
ac
h
g
r
id
ce
ll
w
as
r
esp
o
n
s
ib
l
e
f
o
r
p
r
ed
ictin
g
t
h
e
o
b
j
ec
ts
w
h
o
s
e
b
o
u
n
d
in
g
-
b
o
x
ce
n
tr
e
lie
s
i
n
t
h
is
ce
ll.
T
h
e
cl
ass
lab
el,
co
n
f
id
e
n
ce
a
n
d
b
o
u
n
d
in
g
-
b
o
x
co
o
r
d
in
ates
w
er
e
i
n
teg
r
ated
a
s
a
s
in
g
le
r
eg
r
ess
io
n
p
r
o
b
le
m
,
w
h
ic
h
g
a
in
ed
p
r
o
ce
s
s
i
n
g
s
p
ee
d
.
B
u
t
o
n
e
o
b
v
io
u
s
s
h
o
r
tco
m
i
n
g
is
to
d
ea
l
w
ith
o
cc
lu
d
ed
o
b
j
ec
ts
w
h
o
s
e
ce
n
tr
es
lie
i
n
t
h
e
s
a
m
e
g
r
id
ce
ll.
Dete
cti
n
g
s
m
all
o
b
j
ec
ts
w
as
also
f
o
u
n
d
n
o
t
ea
s
y
,
a
s
t
h
e
g
r
id
d
iv
is
io
n
w
as
co
ar
s
e.
I
n
u
p
g
r
ad
ed
v
er
s
io
n
s
o
f
YO
L
O
[
1
7
,
1
8
]
,
an
ch
o
r
s
w
er
e
i
n
tr
o
d
u
c
ed
to
i
m
p
r
o
v
e
th
e
p
er
f
o
r
m
a
n
ce
o
n
lo
ca
tio
n
p
r
ed
i
ctio
n
.
So
m
e
r
ec
en
t
w
o
r
k
p
r
o
p
o
s
ed
to
r
ep
r
esen
t
th
e
o
b
j
ec
t
as
co
o
r
d
in
ate
p
o
in
ts
,
an
d
m
a
k
e
p
r
ed
ictio
n
s
b
y
g
r
o
u
p
in
g
t
h
e
p
o
in
ts
[
1
9
,
2
0
]
.
A
f
i
n
e
f
ea
t
u
r
e
m
ap
r
eso
lu
ti
o
n
is
n
ee
d
ed
to
i
m
p
r
o
v
e
d
etec
tio
n
p
er
f
o
r
m
a
n
ce
o
n
s
m
al
l
o
b
j
ec
ts
.
Me
th
o
d
s
f
o
r
r
ec
o
v
er
in
g
s
p
ati
al
r
eso
lu
tio
n
w
h
ile
k
ee
p
in
g
s
e
m
an
tic
i
n
f
o
r
m
atio
n
w
er
e
i
m
p
o
r
ted
f
r
o
m
i
m
a
g
e
s
eg
m
e
n
tatio
n
as
it
b
y
n
at
u
r
e
r
eq
u
ir
es
d
en
s
e
p
r
ed
ictio
n
o
n
p
ix
el
lev
e
l.
A
co
m
m
o
n
p
r
ac
tice
w
as
to
u
s
e
lin
ea
r
up
-
p
o
o
lin
g
o
r
tr
an
s
p
o
s
e
co
n
v
o
lu
tio
n
(
also
ca
lled
“
d
e
-
co
n
v
o
lu
tio
n
”)
[
2
1
]
af
ter
co
n
ti
n
u
o
u
s
d
o
w
n
-
s
a
m
p
li
n
g
.
f
ea
t
u
r
e
p
y
r
a
m
id
n
et
w
o
r
k
(
FP
N
)
[
2
2
]
later
ally
co
n
n
ec
ted
t
h
e
u
p
-
s
a
m
p
led
la
y
er
s
to
th
e
p
r
ev
io
u
s
la
y
er
s
to
r
ein
f
o
r
ce
th
e
i
n
f
o
r
m
atio
n
,
esp
ec
iall
y
f
o
r
th
e
s
h
allo
w
la
y
er
s
.
T
h
is
f
lex
ib
le
s
tr
u
ct
u
r
e
co
u
ld
s
er
v
e
as
b
ac
k
b
o
n
e
n
et
w
o
r
k
to
m
a
n
y
d
etec
t
io
n
s
c
h
e
m
e
s
.
Fo
r
ex
a
m
p
le,
R
eti
n
aN
et
[
2
3
]
is
ap
p
r
o
x
im
a
tel
y
a
co
m
b
in
at
io
n
o
f
SS
D
an
d
FP
N
,
w
it
h
a
m
o
d
i
f
ied
lo
s
s
to
m
i
tig
a
te
in
f
l
u
en
ce
o
f
o
v
er
w
h
el
m
i
n
g
n
u
m
b
er
s
o
f
ea
s
y
n
e
g
ati
v
e
ex
a
m
p
les.
I
n
co
n
tr
as
t
to
tr
a
n
s
p
o
s
e
co
n
v
o
l
u
tio
n
a
n
d
u
p
-
p
o
o
lin
g
,
a
tr
o
u
s
co
n
v
o
l
u
tio
n
(
also
ca
ll
ed
d
ilated
co
n
v
o
lu
tio
n
o
r
“h
o
le”
alg
o
r
ith
m
)
d
o
n
o
t
d
o
w
n
-
s
a
m
p
le
th
e
o
r
ig
i
n
al
i
m
a
g
e
b
u
t
ap
p
l
y
a
p
y
r
a
m
id
o
f
atr
o
u
s
f
ilter
s
w
it
h
d
if
f
er
en
t
d
ilatio
n
r
ates
to
ex
tr
ac
t
f
ea
tu
r
es
f
r
o
m
d
if
f
er
en
t
s
ca
le
s
.
A
d
ilat
i
o
n
f
ilter
is
a
n
o
r
m
al
co
n
v
o
lu
tio
n
al
f
ilter
i
n
s
er
ted
b
y
ze
r
o
s
.
Di
latio
n
r
ate
is
th
e
d
i
s
tan
ce
to
i
n
s
er
t
ze
r
o
s
,
w
h
ic
h
c
o
n
tr
o
ls
t
h
e
e
f
f
ec
tiv
e
r
ec
ep
tiv
e
f
ield
o
f
t
h
e
f
ilter
.
T
h
is
tec
h
n
iq
u
e
w
as
ad
o
p
te
d
in
o
b
j
ec
t
d
etec
tio
n
a
n
d
a
p
p
lied
in
n
u
m
er
o
u
s
o
cc
asio
n
s
,
s
u
ch
as
r
o
ad
lan
e
d
etec
tio
n
in
[
2
4
]
an
d
b
r
i
d
g
e
cr
a
ck
d
etec
tio
n
in
[
2
5
]
.
I
n
[
2
6
]
,
A
tr
o
u
s
co
n
v
o
l
u
tio
n
w
a
s
r
ep
o
r
ted
to
h
av
e
i
m
p
r
o
v
ed
d
etec
tio
n
p
er
f
o
r
m
a
n
ce
o
f
s
m
al
l
o
b
j
ec
ts
.
W
ith
a
d
i
f
f
er
e
n
t
s
ca
le
o
f
“
s
m
all”
o
b
j
ec
ts
,
th
e
au
th
o
r
u
s
ed
SS
D,
w
h
ic
h
w
e
f
o
u
n
d
d
if
f
ic
u
lt
to
m
atch
t
h
e
an
c
h
o
r
s
.
An
d
th
e
y
co
n
s
tr
u
ct
e
x
tr
a
la
y
er
s
at
th
e
en
d
o
f
t
h
e
C
N
N,
w
h
ile
w
e
ap
p
l
y
atr
o
u
s
co
n
v
o
lu
tio
n
i
n
in
ter
m
ed
iate
la
y
er
s
.
A
v
er
y
s
i
m
ilar
co
n
te
x
t
as
o
u
r
w
o
r
k
,
Af
r
ican
m
a
m
m
als
w
e
r
e
d
etec
ted
f
r
o
m
ae
r
ial
i
m
a
g
e
s
in
[
2
7
]
,
w
h
er
e
t
w
o
s
ib
li
n
g
n
e
t
w
o
r
k
s
w
er
e
co
n
s
tr
u
cted
.
O
n
e
p
r
ed
icts
clas
s
p
r
o
b
ab
ilit
y
f
o
r
ea
ch
f
ea
tu
r
e
m
ap
ce
ll,
th
e
o
th
er
o
u
tp
u
t
s
b
o
u
n
d
i
n
g
-
b
o
x
c
o
o
r
d
in
ates.
T
h
e
i
m
ag
e
s
w
er
e
cr
o
p
p
e
d
in
to
s
m
a
ll
p
iece
s
a
n
d
d
etec
tio
n
w
as
m
ad
e
in
e
ac
h
p
iece
.
T
h
is
is
a
co
m
m
o
n
p
r
ac
tice
to
d
ea
l
w
i
th
ex
tr
em
el
y
h
i
g
h
-
r
eso
l
u
tio
n
r
e
m
o
te
s
en
s
i
n
g
i
m
a
g
es
[
2
8
]
.
Ob
v
io
u
s
l
y
,
th
er
e
w
i
ll
b
e
a
b
u
n
c
h
o
f
o
b
j
ec
ts
cu
t
in
to
d
if
f
e
r
en
t
p
ar
ts
a
n
d
t
h
is
is
a
p
r
o
b
le
m
w
h
e
n
p
r
ep
ar
in
g
tr
ain
i
n
g
d
ata
an
d
s
t
itch
in
g
t
h
e
p
atch
es b
ac
k
to
g
e
th
er
to
f
o
r
m
a
u
n
i
f
ied
d
etec
tio
n
.
2.
RE
S
E
ARCH
M
E
T
H
O
D
2
.
1
.
B
a
ck
bo
ne
net
w
o
rk
R
esNet
[
2
9
]
ad
d
r
ess
ed
th
e
d
eg
r
ad
atio
n
p
r
o
b
lem
o
f
v
er
y
d
e
ep
n
eu
r
al
n
et
w
o
r
k
s
,
w
h
er
e
ac
cu
r
ac
y
g
et
s
s
atu
r
ated
w
h
e
n
t
h
e
n
et
w
o
r
k
g
o
es
to
o
d
ee
p
.
B
y
ad
d
in
g
t
h
e
in
p
u
t
to
t
h
e
o
u
tp
u
t,
v
er
y
d
ee
p
n
et
w
o
r
k
g
ai
n
s
d
etec
tio
n
ac
cu
r
ac
y
.
Fi
g
u
r
e
3
s
h
o
w
s
t
h
e
r
esid
u
a
l
u
n
i
t
t
h
at
p
er
f
o
r
m
s
a
co
n
v
o
l
u
tio
n
o
p
er
atio
n
a
n
d
“
s
h
o
r
tcu
t”
co
n
n
ec
tio
n
b
et
w
ee
n
t
h
e
i
n
p
u
t
an
d
o
u
tp
u
t.
T
h
e
id
ea
o
f
r
esid
u
al
lear
n
i
n
g
w
as
u
s
ed
in
m
a
n
y
o
th
er
ar
ch
itect
u
r
es
lik
e
Dar
k
n
et
-
5
3
o
f
YO
L
O
-
v3
an
d
I
n
ce
p
to
n
-
v
3
o
f
Go
o
g
le
Net
[
3
0
]
.
T
h
e
n
et
w
o
r
k
co
u
ld
g
r
o
w
v
er
y
d
ee
p
b
y
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
2586
I
A
E
S
I
n
t
J
R
o
b
&
A
u
to
m
,
Vo
l
.
1
0
,
No
.
2
,
J
u
n
e
2
0
2
1
:
1
3
3
–
143
136
s
tack
i
n
g
a
b
u
n
ch
o
f
co
n
v
o
lu
tio
n
a
l
la
y
er
s
.
Ho
w
e
v
er
,
c
o
m
p
u
tatio
n
al
co
s
t
h
a
m
p
er
s
th
e
u
s
e
o
f
h
ea
v
y
ar
ch
itect
u
r
es f
o
r
b
ig
i
m
a
g
es i
n
o
u
r
d
ataset.
Her
ein
,
w
e
u
s
ed
R
esNet
-
5
0
th
at
co
n
tain
s
5
0
co
n
v
o
l
u
tio
n
a
l
la
y
er
s
.
T
h
e
m
ai
n
b
o
d
y
o
f
R
e
s
Net
-
5
0
co
n
tain
s
4
b
lo
ck
s
th
at
w
er
e
co
m
p
o
s
ed
b
y
3
,
4
,
6
an
d
3
la
y
er
s
o
f
t
h
e
r
esid
u
al
u
n
its
as
s
h
o
w
n
in
Fi
g
u
r
e
3
,
r
e
s
p
ec
tiv
el
y
.
Or
i
g
i
n
all
y
,
t
h
e
s
p
atial
r
eso
l
u
tio
n
s
o
f
t
h
e
b
lo
ck
s
w
er
e
1
/4
,
1
/8
,
1
/1
6
an
d
1
/3
2
o
f
th
e
in
p
u
t
i
m
ag
e
s
ize.
I
n
s
tead
o
f
co
n
t
in
u
o
u
s
l
y
r
ed
u
ce
t
h
e
s
p
atial
r
eso
l
u
tio
n
af
ter
B
lo
ck
2
,
th
e
o
u
tp
u
t
s
tr
id
e
w
a
s
k
ep
t
at
1
/8
o
f
th
e
i
n
p
u
t
i
m
ag
e
s
ize.
T
o
e
x
tr
ac
t
ab
s
tr
ac
t
f
ea
t
u
r
es,
A
ll
t
h
e
3
x
3
co
n
v
o
lu
tio
n
al
f
il
ter
s
i
n
B
lo
ck
s
2
,
3
an
d
4
ar
e
r
ep
lace
d
b
y
atr
o
u
s
co
n
v
o
lu
t
io
n
al
f
ilter
s
.
E
s
s
en
ce
o
f
atr
o
u
s
co
n
v
o
lu
tio
n
w
a
s
to
ca
tch
i
n
f
o
r
m
atio
n
f
r
o
m
a
b
i
g
g
er
ar
ea
an
d
s
k
ip
s
o
m
e
in
b
et
w
ee
n
b
y
s
e
t
tin
g
“
h
o
le
s
”
o
n
t
h
e
f
ilt
er
.
T
h
e
f
o
r
m
u
latio
n
o
f
o
n
e
-
d
i
m
en
s
io
n
al
s
ig
n
al
s
is
:
[
]
∑
[
]
[
]
w
h
er
e
[
]
is
th
e
o
u
tp
u
t
o
f
i
n
p
u
t
[
]
co
n
v
o
lv
ed
w
it
h
f
ilter
[
]
w
it
h
le
n
g
th
.
is
th
e
s
tr
id
e
to
s
a
m
p
le
[
]
,
ca
lled
“
d
ilatio
n
r
ate”
.
A
f
i
n
e
f
ea
t
u
r
e
m
ap
s
p
atial
r
eso
l
u
tio
n
co
n
tr
ib
u
tes
to
a
v
o
id
in
g
o
m
itt
in
g
th
e
v
er
y
s
m
a
ll
o
b
j
ec
ts
an
d
m
a
k
es
a
n
ch
o
r
s
b
etter
m
atc
h
in
g
t
h
e
g
r
o
u
n
d
-
tr
u
th
.
B
u
t
it
i
s
a
tr
ad
e
-
o
f
f
c
h
o
ic
e
as
th
e
n
u
m
b
er
o
f
an
ch
o
r
s
i
n
cr
ea
s
e
ex
p
o
n
e
n
tiall
y
o
n
a
d
o
u
b
le
s
ized
f
ea
tu
r
e
m
ap
.
A
s
th
e
atr
o
u
s
co
n
v
o
l
u
tio
n
al
la
y
er
s
ar
e
s
tack
e
d
co
n
s
ec
u
tiv
e
l
y
an
d
th
e
ab
s
tr
ac
t
lev
el
ac
cu
m
u
late
s
in
a
v
er
y
d
ee
p
n
et
w
o
r
k
,
w
e
p
r
o
p
o
s
e
to
u
s
e
th
e
s
a
m
e
d
ilatio
n
r
ate
eq
u
als
to
2
o
th
er
s
th
a
n
p
r
o
g
r
ess
i
v
el
y
e
n
lar
g
e
th
e
d
ilatio
n
r
ate
.
A
b
latio
n
o
n
t
h
e
d
ilatio
n
r
ate
w
ill
b
e
d
o
n
e
in
Sectio
n
5
.
B
r
ac
k
et
p
ar
t
in
Fig
u
r
e
4
illu
s
tr
ate
s
th
e
m
o
d
i
f
i
ed
R
esNet
-
5
0
in
d
etail,
i
n
clu
d
in
g
la
y
er
s
izes
a
n
d
d
ep
th
s
.
T
h
e
r
est d
ep
icts
th
e
f
lo
w
o
f
t
h
e
f
ea
tu
r
e
s
d
er
iv
ed
f
r
o
m
th
e
b
ac
k
b
o
n
e
n
et
w
o
r
k
in
Fas
t
er
R
-
C
NN.
Fig
u
r
e
3
.
An
ex
a
m
p
le
o
f
a
r
esi
d
u
al
u
n
it o
f
R
esNet,
t
h
e
1
x
1
f
i
lter
s
w
er
e
u
s
ed
to
ad
j
u
s
t th
e
f
e
atu
r
e
d
ep
th
s
Fig
u
r
e
4
.
A
r
ch
itectu
r
e
o
f
t
h
e
p
r
o
p
o
s
ed
m
eth
o
d
,
th
e
r
ed
an
d
w
h
ite
b
o
x
es
r
ep
r
ese
n
t
th
e
atr
o
u
s
f
ilter
s
t
h
at
ar
e
ap
p
lied
in
th
e
b
ac
k
b
o
n
e
n
et
w
o
r
k
; t
h
e
n
u
m
b
er
s
o
n
t
h
e
to
p
s
h
o
w
t
h
e
d
ep
th
o
f
la
y
er
s
, t
h
e
n
u
m
b
er
s
at
th
e
b
o
tto
m
ar
e
th
e
s
p
atial
r
eso
lu
t
io
n
o
f
t
h
e
la
y
er
s
T
h
e
o
u
tp
u
t
s
tr
id
e
is
k
ep
t
at
8
f
o
r
s
ev
er
al
r
ea
s
o
n
s
.
O
n
e
is
th
at
th
e
s
m
alle
s
t
o
b
j
ec
t
s
ize
in
o
u
r
d
ataset
is
ab
o
u
t
1
1
x
1
1
,
n
o
m
in
al
l
y
it
w
o
u
ld
p
r
o
j
ec
t
to
at
leas
t
o
n
e
-
p
i
x
el
o
n
th
e
f
ea
tu
r
e
m
ap
at
t
h
e
o
u
tp
u
t
s
tr
id
e
o
f
8
.
An
o
th
er
is
t
h
at
s
m
all
o
b
j
ec
ts
n
ee
d
a
s
m
all
s
tr
id
i
n
g
s
tep
to
b
etter
m
atc
h
t
h
e
a
n
c
h
o
r
s
.
T
h
e
m
atc
h
in
g
q
u
alit
y
f
o
r
ea
ch
o
b
j
ec
t
s
ca
le
h
as
g
r
ea
t
i
m
p
ac
t
o
n
t
h
e
d
etec
tio
n
r
es
u
lt
s
.
As
t
h
e
n
u
m
b
er
o
f
s
m
all
o
b
j
ec
ts
co
m
p
o
s
e
o
n
l
y
s
m
al
l
p
o
r
tio
n
o
f
th
e
d
ataset.
Misali
g
n
m
e
n
t
b
et
w
ee
n
t
h
e
a
n
ch
o
r
s
an
d
g
r
o
u
n
d
-
tr
u
t
h
f
o
r
s
e
v
er
al
ex
a
m
p
le
s
is
n
o
t
af
f
o
r
d
ab
le
f
o
r
th
e
s
m
all
o
b
jects.
T
h
e
to
tal
n
u
m
b
er
o
f
a
n
ch
o
r
s
s
h
o
u
ld
also
b
e
co
n
s
i
d
er
ed
.
I
t
is
f
u
r
th
e
r
d
is
cu
s
s
ed
later
to
g
eth
er
w
i
th
t
h
e
d
etec
tio
n
ar
ch
itec
tu
r
e
a
n
d
an
ch
o
r
s
etti
n
g
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t
J
R
o
b
&
A
u
to
m
I
SS
N:
2722
-
2586
Dete
ctin
g
A
fr
ica
n
h
o
o
fe
d
a
n
i
ma
ls
in
a
eria
l ima
g
ery
u
s
in
g
c
o
n
vo
lu
tio
n
a
l n
e
u
r
a
l n
etw
o
r
k
(
Yu
n
fei
F
a
n
g
)
137
2.
2
.
Det
ec
t
io
n
a
rc
hite
ct
ure
T
h
e
d
etec
tio
n
ar
ch
itect
u
r
e
f
o
l
lo
w
s
Fas
ter
R
-
C
NN.
A
3
x
3
c
o
n
v
o
l
u
tio
n
al
f
il
ter
w
ill
s
lid
e
o
n
th
e
la
s
t
la
y
er
o
f
th
e
b
ac
k
b
o
n
e
n
et
w
o
r
k
an
d
d
er
iv
es
a
f
ea
t
u
r
e
m
ap
.
E
ac
h
ce
ll
o
f
t
h
e
f
ea
t
u
r
e
m
ap
o
u
t
p
u
ts
p
r
ed
ictio
n
s
f
o
r
th
e
p
r
e
-
d
e
f
in
ed
s
et
o
f
a
n
ch
o
r
s
.
Su
p
p
o
s
e
t
h
er
e
ar
e
k
an
c
h
o
r
s
f
o
r
ea
ch
ce
ll,
t
h
e
p
r
ed
ictio
n
s
will b
e
2
k
o
b
j
ec
t a
n
d
n
o
n
-
o
b
j
ec
t sco
r
es a
n
d
4
k
co
o
r
d
in
ate
o
f
f
s
et
s
f
o
r
R
P
N.
Featu
r
es
o
f
th
e
o
b
j
ec
t
p
r
o
p
o
s
als
g
e
n
er
ated
f
r
o
m
t
h
e
R
P
N
ar
e
d
er
iv
ed
b
y
p
r
o
j
ec
tin
g
th
e
b
o
u
n
d
in
g
-
b
o
x
es
to
th
e
f
ea
t
u
r
e
m
ap
.
T
h
en
th
r
o
u
g
h
R
OI
alig
n
m
en
t,
th
e
f
ea
t
u
r
es
ar
e
cr
o
p
p
ed
in
to
s
ize
o
f
1
4
x
1
4
u
s
i
n
g
b
i
-
lin
ea
r
i
n
ter
p
o
latio
n
a
n
d
f
u
r
t
h
e
r
d
o
w
n
-
s
a
m
p
led
to
7
x
7
,
as
s
h
o
w
n
i
n
F
ig
u
r
e
5
.
Feat
u
r
es
o
f
d
if
f
er
e
n
t
o
b
j
ec
ts
ar
e
alig
n
ed
to
th
e
s
a
m
e
s
ize
th
at
ca
n
g
o
th
r
o
u
g
h
F
C
la
y
er
s
a
n
d
o
u
tp
u
t
th
e
f
i
n
al
p
r
ed
ictio
n
s
.
Fro
m
Fig
u
r
e
1
,
th
e
m
o
s
t
p
o
p
u
lar
o
b
j
ec
t
s
ize
in
o
u
r
d
ataset
lies
in
b
et
w
ee
n
4
8
x
4
8
an
d
6
4
x
6
4
,
w
h
ich
o
n
l
y
p
r
o
j
ec
ts
to
6
x
6
to
8
x
8
o
n
th
e
f
ea
t
u
r
e
m
ap
.
I
t
is
f
ar
s
m
aller
t
h
an
1
4
x
1
4
.
W
e
ex
p
er
im
en
ted
w
it
h
s
m
aller
R
OI
ali
g
n
m
e
n
t
s
ize,
s
u
c
h
as
7
x
7
an
d
2
x
2
an
d
d
id
n
o
t
ac
h
i
ev
e
b
etter
r
es
u
lt
s
.
Di
v
id
i
n
g
f
e
atu
r
es
o
f
a
s
m
all
o
b
j
e
cts
in
to
f
i
n
e
g
r
ain
ed
p
iece
s
h
elp
s
d
escr
ib
e
th
e
o
b
j
ec
t
in
d
etail.
I
n
ter
s
ec
tio
n
s
o
f
u
n
io
n
(
I
o
Us)
b
et
w
ee
n
t
h
e
an
c
h
o
r
s
a
n
d
th
e
g
r
o
u
n
d
-
tr
u
t
h
w
er
e
u
s
ed
as
m
e
tr
ics
to
s
elec
t
p
o
s
itiv
e
a
n
d
n
e
g
ati
v
e
e
x
a
m
p
l
es.
Fo
r
o
t
h
er
i
m
p
le
m
e
n
tatio
n
d
etails
o
f
Fa
s
ter
R
-
C
NN,
w
e
r
ec
o
m
m
en
d
r
ef
er
r
i
n
g
to
th
e
o
r
ig
i
n
al
w
o
r
k
.
Fig
u
r
e
5
.
R
OI
alig
n
m
en
t o
f
th
e
s
ec
o
n
d
s
tag
e
;
s
m
all
o
b
j
ec
ts
also
g
ai
n
b
en
ef
its
b
y
d
i
v
id
in
g
f
ea
tu
r
es to
f
i
n
e
g
r
ain
ed
p
iece
s
2
.
2
.
Ancho
r
s
et
t
ing
s
An
c
h
o
r
s
w
er
e
e
m
p
ir
icall
y
c
h
o
s
en
i
n
m
o
s
t
o
f
th
e
d
etec
tio
n
ar
ch
itect
u
r
es.
I
n
Fa
s
ter
R
-
C
NN
w
h
er
e
p
r
ed
ictio
n
s
w
er
e
m
ad
e
o
n
a
s
in
g
le
f
ea
t
u
r
e
m
ap
,
t
h
e
a
n
c
h
o
r
s
m
u
s
t
co
v
er
all
t
h
e
o
b
j
ec
t
s
c
ales.
I
n
SS
D
w
h
er
e
p
r
ed
ictio
n
s
w
er
e
m
ad
e
o
n
m
u
ltip
le
f
ea
t
u
r
e
m
ap
s
,
ea
c
h
f
ea
tu
r
e
la
y
er
r
eq
u
ir
es
a
s
p
ec
i
f
ic
d
e
s
ig
n
ed
o
b
j
ec
t
s
ca
le
an
d
o
v
er
all,
th
e
an
c
h
o
r
s
s
h
o
u
l
d
m
atc
h
all
t
h
e
o
b
j
ec
t
s
ca
les,
in
th
e
o
r
ig
i
n
al
w
o
r
k
o
f
SS
D,
t
h
e
au
t
h
o
r
s
u
s
ed
th
e
f
o
llo
w
in
g
f
o
r
m
u
latio
n
to
as
s
i
g
n
th
e
a
n
ch
o
r
s
f
o
r
ea
ch
f
ea
tu
r
e
la
y
er
:
(
)
[
]
m
is
th
e
n
u
m
b
er
o
f
la
y
er
s
,
an
d
is
th
e
m
in
i
m
u
m
a
n
d
m
ax
i
m
u
m
a
n
ch
o
r
s
ca
le
th
a
t
w
a
s
ass
i
g
n
ed
to
th
e
lo
w
e
s
t
an
d
h
ig
h
est
f
ea
t
u
r
e
la
y
er
,
r
esp
ec
tiv
el
y
.
B
ec
au
s
e
t
h
e
o
b
j
ec
t
s
ca
le
o
f
o
u
r
d
ataset
is
v
er
y
s
m
al
l a
n
d
s
p
r
ea
d
i
n
a
n
ar
r
o
w
r
an
g
e,
w
h
ich
is
r
ef
lecte
d
i
n
F
ig
u
r
e
1
(
b
)
,
m
a
n
u
all
y
c
h
o
o
s
in
g
t
h
e
a
n
ch
o
r
s
is
m
o
r
e
f
le
x
ib
le.
Fo
r
ex
a
m
p
le,
t
h
e
r
ati
o
o
f
a
1
6
x
1
6
an
ch
o
r
b
o
x
to
a
1
0
8
0
x
1
9
2
0
i
m
a
g
e
is
0
.
0
0
0
1
2
,
m
ak
i
n
g
s
a
v
er
y
s
m
al
l
n
u
m
b
er
.
T
h
er
e
w
er
e
s
e
m
i
-
au
to
m
atic
a
n
c
h
o
r
ass
i
g
n
i
n
g
m
et
h
o
d
s
,
s
u
c
h
as
u
s
i
n
g
c
en
tr
es
o
f
K
-
m
ea
n
s
clu
s
ter
i
n
g
o
n
th
e
g
r
o
u
n
d
-
tr
u
t
h
b
o
x
es
as
th
e
an
ch
o
r
s
.
Ob
v
io
u
s
l
y
,
K
-
m
ea
n
s
cl
u
s
ter
i
n
g
o
n
l
y
r
ev
ea
ls
t
h
e
s
tatis
tic
p
atter
n
s
o
f
t
h
e
o
b
j
ec
t
s
izes,
b
u
t
h
as
n
o
clu
e
o
n
t
h
e
f
ea
tu
r
e
s
tr
id
e
s
tep
,
as
w
ell
as
th
e
m
atc
h
i
n
g
q
u
alit
y
b
et
w
ee
n
th
e
an
c
h
o
r
s
an
d
g
r
o
u
n
d
-
tr
u
th
.
Featu
r
e
s
tr
id
e
s
tep
is
o
f
g
r
ea
t
i
m
p
o
r
tan
ce
o
n
th
e
s
m
all
o
b
jects.
An
d
ca
r
ef
u
ll
y
b
alan
cin
g
t
h
e
f
ea
t
u
r
e
m
ap
s
ize
an
d
m
atc
h
in
g
q
u
a
l
it
y
i
s
n
ee
d
e
d
f
o
r
v
ar
io
u
s
d
etec
tio
n
ar
ch
ite
ctu
r
es.
E
x
h
a
u
s
t
iv
e
s
ea
r
c
h
i
n
g
o
f
t
h
e
h
y
p
er
-
p
ar
a
m
eter
s
s
u
ch
as
i
m
a
g
e
s
ize,
o
u
tp
u
t
s
tr
id
e
an
d
an
c
h
o
r
s
ca
les
w
a
s
i
m
p
le
m
e
n
ted
f
o
r
b
o
th
th
e
s
in
g
le
-
an
d
m
u
lti
-
la
y
er
o
b
j
e
ct
d
etec
to
r
.
T
h
e
n
u
m
b
er
o
f
an
ch
o
r
s
a
n
d
th
e
I
o
U
m
at
c
h
in
g
q
u
al
ities
f
o
r
ea
ch
o
b
j
ec
t
s
ca
le
ar
e
in
ter
ested
f
ac
to
r
s
to
b
e
lo
o
k
ed
u
p
o
n
.
T
a
b
le
1
s
h
o
w
s
s
o
m
e
r
esu
lt
s
f
o
r
th
e
s
in
g
le
-
la
y
er
d
etec
to
r
,
w
e
s
u
m
m
ar
ize
t
h
e
p
o
r
tio
n
o
f
th
e
b
ig
g
est
I
o
U
th
at
ex
ce
ed
s
th
r
es
h
o
ld
s
0
.
5
,
0
.
6
an
d
0
.
7
u
n
d
er
th
e
co
m
b
in
at
io
n
s
o
f
i
m
a
g
e
s
ize,
o
u
tp
u
t
s
tr
id
e
an
d
an
ch
o
r
s
ca
les.
T
h
e
asp
ec
t
r
atio
is
s
et
to
[
0
.
5
,
1
,
2
]
.
A
n
c
h
o
r
s
ca
les
ar
e
r
ep
r
esen
ted
b
y
th
e
ed
g
e
le
n
g
t
h
,
f
o
r
ex
a
m
p
le,
an
c
h
o
r
s
ca
le
8
r
ep
r
es
en
ts
a
n
ch
o
r
b
o
x
o
f
s
ize
8
x
8
.
Fo
r
t
h
e
s
a
m
e
i
m
a
g
e
s
iz
e,
t
h
e
an
c
h
o
r
s
ca
le
s
s
ee
m
n
o
t
h
a
v
e
m
u
ch
in
f
l
u
e
n
ce
o
n
t
h
e
I
o
U
m
a
tch
i
n
g
q
u
alities
,
p
r
o
b
ab
ly
d
u
e
to
s
m
a
ll o
u
tp
u
t s
tr
id
e
s
tep
an
d
f
i
n
e
d
iv
is
io
n
o
f
t
h
e
an
c
h
o
r
s
ca
les.
T
ab
le
2
s
h
o
w
s
s
o
m
e
o
f
th
e
r
esu
lt
s
f
o
r
I
o
U
m
atc
h
in
g
q
u
a
liti
es
f
o
r
m
u
l
ti
-
la
y
er
d
etec
to
r
s
.
T
h
e
o
u
tp
u
t
s
tr
id
e
lev
el
i
s
t
h
e
o
r
d
er
o
f
2
b
y
w
h
ic
h
t
h
e
f
ea
tu
r
e
m
ap
s
ize
was r
ed
u
ce
d
.
Featu
r
e
la
y
er
s
o
f
l
ev
el
2
-
6
ar
e
[
2
2
, 2
3
,
2
4
,
2
5
.
2
6
]
ti
m
es
s
m
aller
th
a
n
th
e
in
p
u
t
i
m
a
g
e
s
ize.
A
n
e
x
tr
a
an
ch
o
r
s
ca
le
in
b
et
w
ee
n
t
wo
an
c
h
o
r
s
ca
les
w
a
s
ad
d
ed
b
y
t
h
e
ch
o
ice
i
n
d
icate
d
b
y
“
i
n
ter
m
ed
iate
s
ca
le”.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
2586
I
A
E
S
I
n
t
J
R
o
b
&
A
u
to
m
,
Vo
l
.
1
0
,
No
.
2
,
J
u
n
e
2
0
2
1
:
1
3
3
–
143
138
T
ab
le
1
.
A
n
ch
o
r
s
ett
in
g
s
a
n
d
m
atc
h
in
g
q
u
al
ities
o
f
s
i
n
g
le
-
la
y
er
o
b
j
ec
t
d
etec
to
r
I
mag
e
si
z
e
O
u
t
p
u
t
st
r
i
d
e
F
e
a
t
u
r
e
ma
p
si
z
e
A
n
c
h
o
r
sca
l
e
s
N
u
mb
e
r
o
f
a
n
c
h
o
r
s
I
o
U
>
0
.
7
I
o
U
>
0
.
6
I
o
U
>
0
.
5
1
1
0
8
0
x
1
9
2
0
8
(
1
3
5
,
2
4
0
)
[
8
,
1
6
,
3
2
,
6
4
,
1
2
8
]
4
8
6
0
0
0
4
3
.
9
8
%
7
2
.
7
8
%
9
9
.
5
3
%
2
1
0
8
0
x
1
9
2
0
8
(
1
3
5
,
2
4
0
)
[
8
,
1
6
,
3
2
,
6
4
,
1
2
8
，
2
5
6
]
5
8
3
2
0
0
4
3
.
9
9
%
7
2
.
8
1
%
9
9
.
6
4
%
3
1
0
8
0
x
1
9
2
0
8
(
1
3
5
,
2
4
0
)
[
1
6
,
3
2
,
6
4
,
1
2
8
]
3
8
8
8
0
0
4
3
.
9
8
%
7
2
.
7
8
%
9
9
.
5
3
%
4
1
0
8
0
x
1
9
2
0
8
(
1
3
5
,
2
4
0
)
[
1
6
,
3
2
,
6
4
,
1
2
8
,
2
5
6
]
4
8
6
0
0
0
4
3
.
9
9
%
7
2
.
8
1
%
9
9
.
6
4
%
5
1
0
8
0
x
1
9
2
0
8
(
1
3
5
,
2
4
0
)
[
3
2
,
6
4
,
1
2
8
]
2
9
1
6
0
0
4
3
.
9
1
%
7
2
.
4
8
%
9
8
.
6
3
%
6
1
0
8
0
x
1
9
2
0
8
(
1
3
5
,
2
4
0
)
[
3
2
,
6
4
,
1
2
8
,
2
5
6
]
3
8
8
8
0
0
4
3
.
9
2
%
7
2
.
5
0
%
9
8
.
7
3
%
7
5
4
0
x
9
6
0
8
(
6
8
,
1
2
0
)
[
8
,
1
6
,
3
2
,
6
4
,
1
2
8
]
1
2
2
4
0
0
2
7
.
6
3
%
6
0
.
0
7
%
9
3
.
1
2
%
8
5
4
0
x
9
6
0
8
(
6
8
,
1
2
0
)
[
8
,
1
6
,
3
2
,
6
4
,
1
2
8
，
2
5
6
]
1
4
6
8
8
0
2
7
.
6
3
%
6
0
.
0
7
%
9
3
.
1
2
%
9
5
4
0
x
9
6
0
8
(
6
8
,
1
2
0
)
[
1
6
,
3
2
,
6
4
,
1
2
8
]
9
7
9
2
0
2
7
.
6
1
%
6
0
.
0
1
%
9
2
.
8
3
%
10
5
4
0
x
9
6
0
8
(
6
8
,
1
2
0
)
[
1
6
,
3
2
,
6
4
,
1
2
8
,
2
5
6
]
1
2
2
4
0
0
2
7
.
6
1
%
6
0
.
0
1
%
9
2
.
8
3
%
11
5
4
0
x
9
6
0
8
(
6
8
,
1
2
0
)
[
3
2
,
6
4
,
1
2
8
]
7
3
4
4
0
2
4
.
8
0
%
4
9
.
7
3
%
6
8
.
9
1
%
12
5
4
0
x
9
6
0
8
(
6
8
,
1
2
0
)
[
3
2
,
6
4
,
1
2
8
,
2
5
6
]
9
7
9
2
0
2
4
.
8
0
%
4
9
.
7
3
%
6
8
.
9
1
%
T
ab
le
2
.
A
n
ch
o
r
s
ett
in
g
s
a
n
d
m
atc
h
in
g
q
u
al
ities
o
f
m
u
lti
-
la
y
er
o
b
j
ec
t
d
etec
to
r
I
mag
e
si
z
e
O
u
t
p
u
t
s
t
r
i
d
e
l
e
v
e
l
A
n
c
h
o
r
sca
l
e
s
I
n
t
e
r
me
d
i
a
t
e
sca
l
e
N
u
mb
e
r
o
f
a
n
c
h
o
r
s
I
o
U
>
0
.
7
I
o
U
>
0
.
6
I
o
U
>
0
.
5
1
1
0
8
0
x
1
9
2
0
2
-
6
[
8
,
1
6
,
3
2
,
6
4
,
1
2
8
]
N
5
0
6
4
3
0
9
.
5
5
%
3
4
.
3
0
%
7
8
.
0
8
%
2
1
0
8
0
x
1
9
2
0
2
-
6
Y
1
0
1
2
8
6
0
2
8
.
6
3
%
7
4
.
3
5
%
9
7
.
8
5
%
3
1
0
8
0
x
1
9
2
0
2
-
6
[
1
6
,
3
2
,
6
4
,
1
2
8
,
2
5
6
]
N
5
0
6
4
3
0
3
3
.
3
9
%
7
0
.
2
9
%
9
9
.
4
5
%
4
1
0
8
0
x
1
9
2
0
2
-
6
Y
1
0
1
2
8
6
0
7
7
.
0
1
%
9
9
.
9
0
%
1
0
0
.
0
0
%
5
1
0
8
0
x
1
9
2
0
2
-
6
[
3
2
,
6
4
,
1
2
8
,
2
5
6
,
5
1
2
]
N
5
0
6
4
3
0
4
7
.
8
7
%
7
3
.
2
8
%
9
8
.
7
4
%
6
1
0
8
0
x
1
9
2
0
2
-
6
Y
1
0
1
2
8
6
0
9
2
.
2
1
%
9
7
.
4
9
%
9
8
.
7
5
%
7
1
0
8
0
x
1
9
2
0
3
-
7
[
8
,
1
6
,
3
2
,
6
4
,
1
2
8
]
N
1
2
6
6
7
5
2
.
3
1
%
8
.
8
0
%
2
4
.
1
1
%
8
1
0
8
0
x
1
9
2
0
3
-
7
Y
2
5
3
3
5
0
7
.
4
3
%
2
4
.
1
0
%
5
4
.
9
0
%
9
1
0
8
0
x
1
9
2
0
3
-
7
[
1
6
,
3
2
,
6
4
,
1
2
8
,
2
5
6
]
N
1
2
6
6
7
5
9
.
5
5
%
3
4
.
3
1
%
7
8
.
1
5
%
10
1
0
8
0
x
1
9
2
0
3
-
7
Y
2
5
3
3
5
0
2
8
.
6
3
%
7
4
.
3
4
%
9
7
.
8
3
%
11
1
0
8
0
x
1
9
2
0
3
-
7
[
3
2
,
6
4
,
1
2
8
,
2
5
6
,
5
1
2
]
N
1
2
6
6
7
5
3
3
.
9
1
%
6
9
.
7
1
%
9
8
.
2
1
%
12
1
0
8
0
x
1
9
2
0
3
-
7
Y
2
5
3
3
5
0
7
3
.
4
8
%
9
7
.
0
9
%
9
8
.
7
2
%
3.
RE
SU
L
T
S AN
D
AN
AL
Y
SI
S
3
.
1
.
Da
t
a
s
et
T
h
e
d
ataset
w
as
co
llected
in
s
e
m
i
-
d
eser
t
ar
ea
s
in
s
o
u
t
h
e
r
n
Na
m
ib
ia
u
s
i
n
g
DJ
I
p
h
an
t
o
m
3
an
d
p
h
an
to
m
4
in
Dec
e
m
b
er
(
s
u
m
m
er
in
Na
m
ib
ia)
.
W
e
to
o
k
n
u
m
er
o
u
s
f
li
g
h
ts
i
n
d
i
f
f
er
e
n
t
ti
m
es
d
u
r
in
g
t
h
e
d
a
y
f
o
r
a
w
ee
k
,
a
n
d
t
h
e
d
u
r
atio
n
o
f
th
e
v
id
eo
s
a
m
o
u
n
ted
to
s
e
v
er
al
h
o
u
r
s
.
T
h
r
ee
s
p
ec
ies
wer
e
co
v
er
ed
in
t
h
is
d
ataset:
b
lu
e
w
ild
eb
ee
s
t
(
g
n
u
)
,
g
e
m
s
b
o
k
(
o
r
y
x
)
,
a
n
d
ze
b
r
a.
T
h
ese
s
p
ec
ies
w
er
e
ch
o
s
e
n
b
ec
au
s
e
th
e
y
s
p
r
ea
d
w
id
el
y
i
n
s
o
u
th
er
n
Af
r
ica,
h
a
v
in
g
w
e
ll
r
ep
r
esen
tat
io
n
o
f
th
e
h
o
o
f
ed
an
i
m
al
s
o
n
th
i
s
la
n
d
.
A
d
d
itio
n
al
l
y
,
t
h
e
y
ar
e
g
r
eg
ar
io
u
s
,
m
a
k
i
n
g
it
ea
s
y
to
r
ec
o
r
d
a
n
u
m
b
er
o
f
in
s
ta
n
ce
s
i
n
o
n
e
s
h
o
t.
A
ls
o
ad
d
i
n
g
ch
alle
n
g
es
to
th
e
d
etec
to
r
b
y
in
tr
o
d
u
ci
n
g
cr
o
w
d
ed
an
d
o
cc
lu
d
ed
o
b
j
ec
ts
.
T
h
e
r
eso
lu
tio
n
o
f
t
h
e
i
m
a
g
e
i
s
1
0
8
0
x
1
9
2
0
.
Fra
m
es
w
er
e
ta
k
en
f
r
o
m
th
e
v
id
eo
s
e
v
er
y
ce
r
tain
i
n
ter
v
al
o
f
s
ec
o
n
d
s
an
d
w
er
e
d
i
v
id
e
d
r
an
d
o
m
l
y
in
to
tr
ain
i
n
g
a
n
d
test
in
g
s
ets.
1
6
9
3
f
r
a
m
es
w
it
h
ar
o
u
n
d
2
0
0
0
0
in
s
ta
n
ce
s
w
er
e
u
s
ed
f
o
r
tr
ain
i
n
g
,
a
n
d
3
8
9
f
r
a
m
es
w
it
h
4
0
1
7
in
s
ta
n
ce
s
w
er
e
u
s
ed
as
test
in
g
ex
a
m
p
le
s
.
T
h
e
f
li
g
h
t
h
eig
h
t
w
as
1
0
-
2
0
m
et
er
s
,
w
it
h
s
ig
h
ts
f
r
o
m
v
ar
io
u
s
an
g
le
s
to
th
e
o
b
j
ec
t.
T
h
e
en
v
ir
o
n
m
en
t
a
n
d
illu
m
i
n
atio
n
co
n
d
itio
n
s
w
er
e
d
iv
er
s
i
f
ied
o
n
p
u
r
p
o
s
e
to
allo
w
g
en
er
aliza
tio
n
.
3
.
2
.
B
a
ck
bo
ne
net
w
o
rk
R
esNet
-
5
0
w
as
ta
k
e
n
as
th
e
b
asic
f
ea
t
u
r
e
e
x
tr
ac
ti
n
g
b
ac
k
b
o
n
e
n
e
t
w
o
r
k
f
o
r
its
r
o
b
u
s
tn
es
s
an
d
m
o
d
er
ate
s
ize.
T
o
r
etain
th
e
o
u
tp
u
t
s
tr
id
e
at
8
,
atr
o
u
s
co
n
v
o
lu
tio
n
a
l
la
y
er
s
w
it
h
s
tr
id
e
2
r
ep
lace
d
th
e
n
o
r
m
al
co
n
v
o
lu
tio
n
al
la
y
er
s
a
n
d
t
h
e
s
tr
id
e
2
d
o
w
n
-
s
a
m
p
li
n
g
o
p
er
atio
n
s
.
T
h
e
la
s
t
la
y
er
o
f
t
h
e
B
l
o
ck
4
w
as
ta
k
en
a
s
th
e
f
ea
t
u
r
e
m
ap
i
n
t
h
e
Fas
ter
R
-
C
NN
d
etec
t
io
n
s
ch
e
m
e.
F
ig
u
r
e
6
(
a)
is
a
s
ch
e
m
atic
o
f
t
h
e
b
ac
k
b
o
n
e
n
et
w
o
r
k
.
T
h
e
h
ig
h
lig
h
ted
la
y
er
s
ar
e
th
e
ac
tiv
atio
n
f
ea
t
u
r
e
m
ap
s
.
T
h
e
FP
N
s
tr
u
ctu
r
e
r
ec
o
v
er
s
th
e
s
p
atial
r
eso
lu
tio
n
o
f
th
e
f
ea
tu
r
e
m
ap
b
y
u
p
-
s
a
m
p
li
n
g
a
n
d
liter
al
co
n
n
ec
tio
n
as
s
h
o
w
n
i
n
Fi
g
u
r
e
6
(
b
)
.
T
h
e
t
w
o
n
et
w
o
r
k
s
tr
u
ctu
r
e
s
co
u
ld
b
o
th
b
e
u
s
ed
in
Fa
s
ter
R
-
C
NN
th
a
t
m
a
k
e
p
r
ed
ictio
n
s
o
n
o
n
e
s
i
n
g
le
f
ea
t
u
r
e
m
ap
.
Fo
r
m
u
lti
-
la
y
er
d
etec
to
r
s
,
d
ir
ec
tl
y
m
a
k
in
g
p
r
ed
ictio
n
s
f
r
o
m
t
h
e
s
h
allo
w
la
y
er
s
is
n
o
t
f
ea
s
ib
le.
I
t
is
b
etter
to
en
h
an
ce
t
h
e
s
e
m
atic
in
f
o
r
m
atio
n
b
y
in
tr
o
d
u
cin
g
a
FP
N
s
tr
u
ct
u
r
e.
SS
D
ta
k
es
t
h
e
o
u
tp
u
t
la
y
er
s
o
f
t
h
e
FP
N
u
p
o
n
R
esNet
-
5
0
as
s
h
o
w
n
in
Fig
u
r
e
6
(
c)
.
YOL
O
-
v
3
u
s
ed
an
c
h
o
r
s
to
o
v
er
co
m
e
d
ef
icie
n
c
y
o
f
lo
ca
tio
n
ac
cu
r
ac
y
i
n
ea
r
l
ier
v
er
s
io
n
s
o
f
YOL
O.
P
r
ed
ictio
n
s
w
er
e
m
a
d
e
o
n
a
p
y
r
a
m
id
o
f
3
f
ea
t
u
r
e
la
y
er
s
s
tar
ti
n
g
f
r
o
m
o
u
tp
u
t
s
tr
id
e
8
.
B
ec
au
s
e
o
f
th
e
s
q
u
ee
ze
-
a
n
d
-
e
x
cite
m
e
n
t
s
tr
u
c
tu
r
e
o
f
th
e
n
et
w
o
r
k
d
o
es
s
i
m
il
ar
w
o
r
k
as
in
FP
N,
no
m
o
d
if
ica
tio
n
w
as
m
ad
e
o
n
th
e
n
et
w
o
r
k
s
tr
u
ctu
r
e
i
n
o
u
r
e
x
p
er
i
m
e
n
t.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t
J
R
o
b
&
A
u
to
m
I
SS
N:
2722
-
2586
Dete
ctin
g
A
fr
ica
n
h
o
o
fe
d
a
n
i
ma
ls
in
a
eria
l ima
g
ery
u
s
in
g
c
o
n
vo
lu
tio
n
a
l n
e
u
r
a
l n
etw
o
r
k
(
Yu
n
fei
F
a
n
g
)
139
(
a)
(
b
)
(
c)
Fig
u
r
e
6
.
A
cti
v
atio
n
f
ea
t
u
r
e
m
ap
o
f
v
ar
io
u
s
n
et
w
o
r
k
s
tr
u
ct
u
r
es: (
a)
s
tr
aig
h
t f
o
r
w
ar
d
R
esNe
t
w
it
h
atr
o
u
s
f
ilter
s
,
(
b
)
FP
N
s
tr
u
ct
u
r
e,
an
d
(
c)
SS
D
s
t
y
le
p
r
ed
ictio
n
s
o
n
F
P
N
lay
er
s
3
.
4
.
Ancho
rs
Fo
r
s
m
all
o
b
j
ec
ts
in
a
b
ig
i
m
a
g
es,
b
etter
m
atc
h
i
n
g
o
f
t
h
e
an
c
h
o
r
s
w
it
h
th
e
g
r
o
u
n
d
-
tr
u
th
co
u
ld
o
n
l
y
b
e
ac
h
iev
ed
b
y
elab
o
r
ate
d
iv
is
io
n
o
f
a
n
c
h
o
r
s
ca
les
a
n
d
a
f
in
e
r
eso
lu
tio
n
f
ea
t
u
r
e
m
ap
.
T
o
f
u
r
th
er
r
ev
ea
l
t
h
e
I
o
U
m
atc
h
in
g
f
o
r
t
h
e
o
b
j
ec
t
s
ca
les,
Fi
g
u
r
e
7
p
lo
t
th
e
b
ig
g
es
t
I
o
U
o
f
ea
ch
g
r
o
u
n
d
-
tr
u
t
h
ag
ai
n
s
t
t
h
e
o
b
j
ec
t
s
ize.
T
h
e
h
is
to
g
r
a
m
s
h
o
w
s
th
e
n
u
m
b
er
o
f
o
b
j
ec
ts
lie
i
n
ea
c
h
o
b
j
ec
t
s
ca
le.
E
x
a
m
p
les
ar
e
ta
k
en
f
r
o
m
T
ab
le
2
,
th
e
h
y
p
er
-
p
ar
am
eter
s
ar
e
o
n
th
e
to
p
o
f
ea
ch
f
i
g
u
r
e
(
Fi
g
u
r
e
7
(
a)
an
d
(
b
)
)
.
I
n
ter
m
ed
iate
an
c
h
o
r
s
ar
e
i
n
s
er
ted
in
th
e
r
i
g
h
t
co
lu
m
n
co
m
p
ar
ed
to
th
e
lef
t
(
Fig
u
r
e
7
(
b
)
t
o
(
a)
an
d
Fig
u
r
e
7
(
d
)
to
(
c
)
)
.
I
o
U
m
atc
h
i
n
g
i
s
im
p
r
o
v
ed
b
y
ad
d
in
g
in
ter
m
ed
iate
an
c
h
o
r
s
ca
les,
wh
ich
ca
n
b
e
s
ee
n
b
y
co
m
p
ar
i
n
g
Fi
g
u
r
e
7
(
b
)
to
(
a)
an
d
(
d
)
t
o
(
c)
.
I
o
U
p
ea
k
s
ar
e
f
o
r
m
ed
ar
o
u
n
d
th
e
n
e
w
i
n
s
er
t
ed
an
ch
o
r
s
ca
le.
Fro
m
th
e
to
p
r
o
w
o
f
Fig
u
r
e
7
to
th
e
b
o
tto
m
r
o
w
(
(
a)
to
(
c)
an
d
Fig
u
r
e
7
(
b
)
to
(
d
)
)
,
I
o
U
m
a
t
ch
in
g
i
s
i
m
p
r
o
v
ed
b
y
m
o
v
in
g
th
e
f
ea
tu
r
e
m
ap
o
n
e
la
y
er
u
p
,
w
h
ich
i
n
d
icate
a
f
i
n
er
f
ea
t
u
r
e
m
ap
r
eso
lu
t
io
n
.
T
h
e
p
o
r
tio
n
o
f
I
o
U>
0
.
5
ac
h
ie
v
es
1
0
0
%
at
f
ea
t
u
r
e
le
v
el
2
-
6
w
ith
i
n
ter
m
e
d
iate
an
c
h
o
r
s
as
s
h
o
w
n
i
n
Fig
u
r
e
7
(
d
)
.
Ho
w
ev
er
,
t
h
e
to
tal
n
u
m
b
er
o
f
an
c
h
o
r
s
i
s
ab
o
u
t
1
M,
w
h
ile
t
h
e
n
u
m
b
er
o
f
a
n
ch
o
r
s
o
n
t
h
e
s
a
m
e
f
ea
t
u
r
e
lev
els
w
it
h
o
u
t
in
ter
m
ed
iate
an
ch
o
r
s
ca
les
is
ab
o
u
t
5
0
0
K.
A
t
f
ea
t
u
r
e
lev
el
3
-
7
w
it
h
in
ter
m
ed
iate
an
ch
o
r
s
,
t
h
e
to
tal
n
u
m
b
er
o
f
an
ch
o
r
s
i
s
o
n
l
y
1
2
0
K.
Fo
r
s
in
g
le
-
la
y
er
o
b
j
ec
t
d
etec
to
r
s
,
th
e
o
p
tim
u
m
ch
o
ice
is
R
o
w
3
o
f
T
ab
le
1
,
th
e
a
n
ch
o
r
s
ca
les
ar
e
[
1
6
,
3
2
,
6
4
,
1
2
8
]
w
it
h
th
e
o
r
ig
i
n
al
i
m
a
g
e
s
ize.
Fo
r
th
e
m
u
lti
-
la
y
er
d
etec
to
r
s
,
R
o
w
4,
an
d
R
o
w
6
o
f
T
ab
le
2
h
as
co
m
p
ar
ab
le
I
o
U
m
atc
h
i
n
g
r
ate
a
n
d
R
o
w
6
g
en
er
ates
le
s
s
a
n
ch
o
r
s
b
y
o
m
itti
n
g
an
c
h
o
r
s
ca
le
o
f
1
6
.
Ho
w
e
v
er
,
th
e
p
o
o
r
ly
m
atc
h
ed
o
b
j
ec
ts
a
r
e
ce
n
tr
alize
d
o
n
t
h
e
s
m
a
ll
o
b
j
ec
ts
.
A
s
s
h
o
w
n
i
n
Fi
g
u
r
e
8
,
th
e
o
b
j
ec
t
s
w
it
h
b
i
g
g
e
s
t
I
o
Us<0
.
5
ar
e
a
ll
b
elo
w
3
2
x
3
2
.
T
h
is
in
f
l
u
en
c
es
t
h
e
d
etec
tio
n
f
o
r
s
m
al
l o
b
j
ec
ts
.
So
h
y
p
er
-
p
ar
a
m
eter
s
in
R
o
w
4
is
c
h
o
s
en
f
o
r
S
SD.
(
a)
(
b
)
(
c)
(
d
)
Fig
u
r
e
7
.
Vis
u
al
izatio
n
o
f
t
h
e
o
b
j
ec
t size
an
d
its
b
est
m
atch
e
d
I
o
U
(
b
lu
e
d
o
ts
)
:
(
a)
r
esu
lts
f
o
r
f
ea
tu
r
e
le
v
els 3
-
7
an
d
(
b
)
its
co
m
p
ar
atio
n
af
ter
i
n
ter
m
ed
iate
a
n
ch
o
r
s
ar
e
ad
d
ed
,
an
d
(
c)
f
ea
tu
r
e
lev
el
s
2
-
6
an
d
(
d
)
its
co
m
p
ar
atio
n
af
ter
in
ter
m
ed
iate
an
ch
o
r
s
ar
e
ad
d
ed
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
2586
I
A
E
S
I
n
t
J
R
o
b
&
A
u
to
m
,
Vo
l
.
1
0
,
No
.
2
,
J
u
n
e
2
0
2
1
:
1
3
3
–
143
140
Fig
u
r
e
8
.
An
ch
o
r
s
h
av
i
n
g
h
i
g
h
I
o
U
m
atc
h
i
n
g
r
ate,
b
u
t p
o
o
r
m
atc
h
es a
r
e
ce
n
tr
alize
d
o
n
s
m
all
o
b
j
ec
ts
3
.
5
.
M
o
del e
v
a
lua
t
i
o
n
T
h
e
m
o
d
els
w
er
e
tr
ain
ed
f
r
o
m
s
cr
atch
o
n
t
h
e
tr
ain
i
n
g
s
et
o
f
o
u
r
d
ataset.
MS
-
C
O
C
O
m
etr
ics
w
er
e
u
s
e
d
to
e
v
alu
a
te
t
h
e
d
etec
tio
n
p
er
f
o
r
m
a
n
ce
.
O
n
l
y
t
h
e
o
v
er
all
m
A
P
(
m
ea
n
a
v
er
ag
e
p
r
ec
is
io
n
)
an
d
m
AP
f
o
r
t
h
e
lar
g
e,
m
ed
iu
m
,
an
d
s
m
a
ll
o
b
j
e
cts
w
er
e
co
n
s
id
er
ed
,
n
eg
lect
in
g
th
e
co
r
r
esp
o
n
d
in
g
m
AR
(
m
e
an
av
er
ag
e
r
ec
all)
.
T
ab
le
3
lis
ted
th
e
ev
al
u
atio
n
r
esu
lt
s
f
o
r
th
e
v
ar
io
u
s
d
etec
tio
n
m
o
d
el
s
,
in
c
lu
d
i
n
g
t
w
o
-
s
ta
g
e
d
etec
to
r
Fas
ter
R
-
C
NN
a
n
d
o
n
e
-
s
ta
g
e
S
SD
a
n
d
YOL
O
-
v
3
.
T
h
e
asp
ec
t
r
atio
s
w
er
e
all
s
et
to
[
0
.
5
,
1
,
2
]
.
Fo
r
Fas
ter
R
-
C
NN,
t
h
e
b
ac
k
b
o
n
e
n
et
w
o
r
k
s
w
er
e
b
ase
d
o
n
R
esNet
-
5
0
.
H
y
p
er
-
p
ar
a
m
eter
s
HP
1
an
d
HP
2
d
ef
in
ed
th
e
i
m
a
g
e
s
ize
an
d
an
ch
o
r
s
ca
le
s
.
HP
1
w
as
t
h
e
s
e
ttin
g
s
w
i
th
t
h
e
o
p
ti
m
u
m
a
n
ch
o
r
ch
o
ice
b
alan
cin
g
t
h
e
I
o
U
m
a
tch
i
n
g
r
ate
a
n
d
th
e
to
tal
n
u
m
b
er
o
f
an
ch
o
r
s
.
Un
d
er
th
e
s
a
m
e
b
ac
k
b
o
n
e
n
et
w
o
r
k
w
ith
atr
o
u
s
f
il
ter
s
.
HP
1
o
u
tp
er
f
o
r
m
s
HP
2
as
th
e
I
o
U
m
atc
h
i
n
g
r
ate
b
et
w
ee
n
th
e
an
c
h
o
r
s
an
d
g
r
o
u
n
d
-
tr
u
t
h
is
h
i
g
h
er
.
R
a
te1
an
d
r
ate
2
d
ef
i
n
ed
th
e
d
if
f
er
en
t
d
ilatio
n
r
ates.
Usi
n
g
t
h
e
s
a
m
e
s
m
alles
t
d
ilatio
n
r
ate
[
2
,
2
,
2
]
led
to
b
etter
r
es
u
lts
t
h
an
en
l
ar
g
in
g
t
h
e
d
ilatio
n
r
ate
f
o
r
ea
ch
b
lo
ck
o
f
R
e
n
et
-
5
0
.
T
h
is
co
u
ld
b
e
ca
u
s
ed
b
y
th
e
n
ar
r
o
w
r
an
g
e
o
f
t
h
e
o
b
j
ec
t
s
ca
le
co
m
b
in
ed
w
it
h
elab
o
r
ate
d
iv
is
io
n
o
f
t
h
e
a
n
c
h
o
r
s
th
at
m
a
k
es
en
lar
g
i
n
g
th
e
“
r
ec
ep
tiv
e
f
ield
”
n
o
t
n
ec
e
s
s
ar
y
.
HP
2
u
s
e
s
i
m
a
g
e
s
ize
th
at
i
s
h
al
f
o
f
t
h
e
o
r
ig
i
n
a
l
i
m
a
g
e,
w
h
ic
h
lead
s
to
m
u
ch
lig
h
ter
n
e
t
w
o
r
k
.
B
u
t
t
h
e
p
er
f
o
r
m
a
n
ce
is
i
n
f
er
io
r
d
u
e
to
th
e
w
o
r
s
e
an
c
h
o
r
m
atc
h
in
g
r
ate
a
n
d
lo
s
s
o
f
i
n
f
o
r
m
atio
n
.
T
ab
le
3
.
E
v
alu
atio
n
o
f
v
ar
io
u
s
m
o
d
els
D
e
t
e
c
t
i
o
n
m
o
d
e
l
B
a
c
k
b
o
n
e
Hy
p
e
r
-
p
a
r
a
m
e
t
e
r
s
m
A
P
m
A
P
(
l
a
r
g
e
)
m
A
P
(
m
e
d
i
u
m
)
m
A
P
(
s
m
a
l
l
)
F
a
st
e
r
R
-
C
N
N
R
e
sN
e
t
-
5
0
+
a
t
r
o
u
s
H
P
1
+
r
a
t
e
1
0
.
5
9
0
.
6
9
0
.
6
0
.
3
6
R
e
sN
e
t
-
5
0
+
a
t
r
o
u
s
H
P
1
+
r
a
t
e
2
0
.
4
2
0
.
6
2
0
.
5
8
0
.
3
3
R
e
sN
e
t
-
5
0
+
a
t
r
o
u
s
H
P
2
+
r
a
t
e
1
0
.
5
1
0
.
6
7
0
.
5
2
0
.
2
3
R
e
sN
e
t
-
5
0
+
F
P
N
H
P
1
0
.
4
8
0
.
6
4
0
.
4
9
0
.
1
7
S
S
D
R
e
sN
e
t
-
5
0
+
F
P
N
S
H
P
0
.
3
2
0
.
3
4
0
.
3
3
0
.
2
8
Y
O
L
O
-
v3
D
a
r
k
n
e
t
-
52
YHP
0
.
3
4
0
.
5
2
.
0
.
3
6
0
.
3
1
H
P
1
:
i
mag
e
si
z
e
1
0
8
0
x
1
9
2
0
,
a
n
c
h
o
r
s
c
a
l
e
s:
[
1
6
,
3
2
,
6
4
,
1
2
8
]
H
P
2
:
i
mag
e
si
z
e
5
4
0
x
9
6
0
,
a
n
c
h
o
r
sca
l
e
s:
[
8
,
1
6
,
3
2
,
6
4
]
R
a
t
e
1
:
d
i
l
a
t
i
o
n
r
a
t
e
[
2
,
2
,
2
]
R
a
t
e
2
:
d
i
l
a
t
i
o
n
r
a
t
e
[
2
,
4
,
8
]
S
H
P
1
:
i
mag
e
si
z
e
:
1
0
8
0
x
1
9
2
0
,
f
e
a
t
u
r
e
l
e
v
e
l
:
3
-
7
,
a
n
c
h
o
r
sc
a
l
e
s:
[
1
6
,
3
2
,
6
4
,
1
2
8
,
2
5
6
]
,
i
n
t
e
r
me
d
i
a
t
e
sc
a
l
e
:
Y
e
s
Y
O
L
O
-
v
3
:
i
mag
e
si
z
e
:
1
0
8
0
x
1
9
2
0
,
a
n
c
h
o
r
sca
l
e
s [1
6
,
6
4
,
2
5
6
]
SS
D
co
u
ld
o
n
l
y
tak
e
t
h
e
f
ea
t
u
r
e
la
y
er
s
f
r
o
m
a
FP
N
s
tr
u
ct
u
r
e.
SHP
,
th
e
h
y
p
er
-
p
ar
a
m
ete
r
s
f
o
r
SS
D
w
a
s
t
h
e
o
p
ti
m
u
m
o
p
tio
n
t
h
at
w
a
s
an
a
l
y
s
ed
p
r
ev
io
u
s
l
y
.
T
h
e
d
etec
tio
n
p
er
f
o
r
m
a
n
ce
w
a
s
i
n
f
er
io
r
f
o
r
all
o
b
j
ec
t
s
ca
les t
h
a
n
t
h
e
t
w
o
-
s
ta
g
e
Fa
s
t
er
R
-
C
NN.
Fin
e
d
i
v
i
s
io
n
o
f
t
h
e
an
ch
o
r
s
ca
les
w
a
s
co
m
p
u
ls
o
r
y
f
o
r
s
u
c
h
d
ata
t
h
at
w
er
e
b
o
th
o
b
j
ec
tiv
el
y
an
d
r
ela
tiv
el
y
s
m
all.
Ho
w
ev
er
,
d
esi
g
n
in
g
a
p
r
o
p
er
an
ch
o
r
s
et
f
o
r
a
s
eq
u
en
ce
o
f
f
ea
t
u
r
e
la
y
er
s
i
s
n
o
t
ea
s
y
.
Go
o
d
I
o
U
m
atc
h
in
g
f
o
r
th
e
o
b
j
ec
t
s
ca
les
o
n
b
o
th
th
e
s
m
a
ll
an
d
t
h
e
lar
g
e
en
d
s
is
d
i
f
f
icu
lt
to
b
e
m
et
at
t
h
e
s
a
m
e
ti
m
e.
YO
L
O
-
v
3
o
u
tp
u
t
s
t
h
r
ee
f
ea
tu
r
e
l
a
y
er
s
,
t
h
e
an
c
h
o
r
s
ca
le
s
ar
e
s
et
to
[
1
6
,
6
4
,
2
5
6
]
.
P
er
f
o
r
m
a
n
ce
i
m
p
r
o
v
e
m
e
n
t
w
a
s
g
ai
n
ed
o
n
all
o
b
j
ec
t scale
s
o
v
er
SS
D
w
it
h
FP
N
s
tr
u
ct
u
r
e.
Fig
u
r
e
9
g
iv
e
s
v
is
u
aliza
tio
n
s
o
f
s
o
m
e
g
o
o
d
d
etec
tio
n
r
es
u
lt
s
.
Fi
g
u
r
e
9
(
a)
-
(
c)
,
(
d
)
-
(
f
)
,
an
d
(
g
)
-
(
j
)
ar
e
r
esu
lt
s
f
o
r
t
h
e
s
p
ec
ie
s
:
ze
b
r
a,
o
r
y
x
a
n
d
b
lu
e
w
ild
eb
ee
s
t
,
r
es
p
ec
tiv
el
y
.
T
h
e
s
ce
n
ar
io
is
w
h
er
e
th
e
an
i
m
a
ls
ar
e
clea
r
an
d
p
ar
ted
,
clu
tter
ed
,
an
d
o
cc
lu
d
ed
b
y
o
n
e
an
o
t
h
er
,
in
d
is
tan
ce
a
n
d
s
m
all
in
s
ize
a
n
d
th
e
co
r
r
esp
o
n
d
in
g
r
esu
lt
s
ar
e
Fi
g
u
r
e
9
(
a)
,
(
d
)
,
a
n
d
(
g
)
;
Fi
g
u
r
e
9
(
b
)
,
(
e)
,
an
d
(
h
)
;
an
d
F
ig
u
r
e
9
(
c)
,
(
f
)
,
an
d
(
i)
;
r
esp
ec
tiv
el
y
.
T
h
e
p
ictu
r
es
ar
e
cr
o
p
p
ed
an
d
r
esiz
ed
f
o
r
d
is
p
lay
co
n
v
e
n
ie
n
ce
an
d
d
o
n
o
t
r
ev
ea
l
th
ei
r
r
ea
l
s
ize
o
f
th
e
f
u
ll
i
m
ag
e
.
T
h
e
o
b
j
ec
ts
th
at
ar
e
d
en
s
el
y
c
lu
tter
ed
an
d
s
m
all
i
n
s
ize
ar
e
n
o
to
r
io
u
s
l
y
d
if
f
ic
u
lt
to
d
etec
t,
o
u
r
m
et
h
o
d
s
h
o
w
s
g
o
o
d
r
esu
lts
i
n
s
o
m
e
p
lace
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t
J
R
o
b
&
A
u
to
m
I
SS
N:
2722
-
2586
Dete
ctin
g
A
fr
ica
n
h
o
o
fe
d
a
n
i
ma
ls
in
a
eria
l ima
g
ery
u
s
in
g
c
o
n
vo
lu
tio
n
a
l n
e
u
r
a
l n
etw
o
r
k
(
Yu
n
fei
F
a
n
g
)
141
E
x
a
m
p
le
s
o
f
f
al
s
e
d
etec
tio
n
s
a
r
e
s
h
o
w
n
i
n
Fig
u
r
e
10
.
Fig
u
r
e
1
0
(
a)
-
(
c)
ar
e
f
alse
p
o
s
itiv
es
f
o
r
ze
b
r
as,
o
r
y
x
an
d
b
lu
e
w
ild
eb
ee
s
t,
Fi
g
u
r
e
1
0
(
d
)
-
(
f
)
ar
e
th
e
an
i
m
a
ls
n
o
t
b
ein
g
d
etec
ted
(
f
alse
n
eg
ati
v
es).
T
h
e
f
als
e
n
eg
at
iv
e
s
ar
e
m
ai
n
l
y
th
e
s
m
all
o
b
j
ec
ts
.
A
d
d
in
g
m
o
r
e
tr
ain
i
n
g
ex
a
m
p
le
s
o
f
th
e
s
m
all
o
b
j
ec
ts
in
th
e
f
u
t
u
r
e
m
a
y
h
elp
th
e
s
itu
a
tio
n
.
(
a)
(
b
)
(
c)
(
d
)
(
e)
(
f
)
(
g
)
(
h
)
(
i)
Fig
u
r
e
9
.
Dete
ctio
n
r
esu
lts
of
(
a)
o
r
y
x
w
h
e
n
th
e
y
ar
e
ap
ar
t,
(
b
)
cr
o
w
d
ed
an
d
o
cc
lu
d
ed
,
an
d
(
c)
f
o
r
th
e
s
m
al
l
o
b
j
ec
ts
; (
d
)
ze
b
r
a
w
h
e
n
th
e
y
a
r
e
ap
ar
t,
(
e
)
cr
o
w
d
ed
an
d
o
cc
lu
d
ed
,
an
d
(
f
)
f
o
r
th
e
s
m
all
o
b
jects; an
d
(
g
)
b
lu
e
w
ild
eb
ee
s
t
w
h
e
n
th
e
y
ar
e
ap
ar
t,
(
h
)
cr
o
w
d
ed
an
d
o
cc
lu
d
ed
,
an
d
(
i)
f
o
r
th
e
s
m
all
o
b
j
ec
ts
(
a)
(
b
)
(
c)
(
d
)
(
e)
(
f
)
Fig
u
r
e
10
.
Fals
e
p
o
s
itiv
e
d
etec
tio
n
s
f
o
r
(
a)
ze
b
r
a
,
(
b
)
o
r
y
x
,
a
n
d
(
c)
b
lu
e
w
ild
eb
ee
s
t; a
n
d
f
al
s
e
n
e
g
ati
v
e
d
etec
tio
n
s
f
o
r
(
d
)
ze
b
r
a,
(
e
)
,
o
r
y
x
,
a
n
d
(
c)
b
lu
e
w
ild
eb
ee
s
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
2586
I
A
E
S
I
n
t
J
R
o
b
&
A
u
to
m
,
Vo
l
.
1
0
,
No
.
2
,
J
u
n
e
2
0
2
1
:
1
3
3
–
143
142
4.
CO
NCLU
SI
O
N
W
e
b
r
o
u
g
h
t
u
p
th
e
p
r
o
b
lem
o
f
d
etec
tin
g
A
f
r
ica
n
h
o
o
f
ed
m
a
m
m
als
u
s
i
n
g
ae
r
ial
i
m
ag
er
y
ta
k
en
f
r
o
m
UAVs.
I
n
a
R
e
s
Net
-
5
0
b
ac
k
b
o
n
e
n
et
w
o
r
k
,
a
s
eq
u
e
n
ce
o
f
a
tr
o
u
s
co
n
v
o
lu
tio
n
al
f
ilter
s
w
e
r
e
u
s
ed
to
k
ee
p
th
e
f
ea
t
u
r
e
m
a
p
r
eso
lu
tio
n
at
ce
r
tain
le
v
el
an
d
at
th
e
s
a
m
e
ti
m
e
co
n
ti
n
u
e
ex
tr
ac
ti
n
g
d
ee
p
er
ab
s
tr
ac
t
s
e
m
an
t
ic
f
ea
t
u
r
es.
T
h
e
d
etec
tio
n
p
er
f
o
r
m
an
ce
f
o
r
s
u
ch
d
ata
is
s
en
s
iti
v
e
to
th
e
m
atc
h
in
g
q
u
a
lit
y
b
et
w
ee
n
t
h
e
an
c
h
o
r
an
d
th
e
g
r
o
u
n
d
-
tr
u
t
h
.
T
h
e
i
m
ag
e
s
ize,
o
u
tp
u
t
s
tr
id
e
an
d
an
ch
o
r
ch
o
ices
co
m
b
in
ed
to
g
et
h
er
d
eter
m
in
e
t
h
e
I
o
U
m
atc
h
in
g
q
u
a
liti
e
s
.
T
h
is
f
ea
t
u
r
e
ex
tr
ac
tio
n
tech
n
iq
u
e
w
a
s
p
r
o
v
en
r
o
b
u
s
t
i
n
d
etec
ti
n
g
s
m
al
l
o
b
j
ec
ts
b
y
co
m
p
ar
i
n
g
to
FP
N.
T
h
e
tw
o
-
s
tag
e
Fa
s
ter
R
-
C
N
N
s
u
r
p
as
s
s
in
g
le
-
s
ta
g
e
d
etec
to
r
s
in
d
etec
tin
g
s
m
all
o
b
j
ec
t
s
,
esp
ec
iall
y
f
o
r
o
u
r
d
ataset
w
h
er
e
th
e
o
b
j
ec
t size
is
o
f
s
m
al
l r
atio
to
th
e
f
u
ll i
m
ag
e.
ACK
NO
WL
E
D
G
E
M
E
NT
S
T
h
e
v
id
eo
s
u
s
ed
in
t
h
is
p
ap
er
w
er
e
co
llected
u
n
d
er
th
e
ass
is
t
an
ce
o
f
Go
n
d
w
a
n
a
C
o
llectio
n
,
Na
m
ib
ia.
RE
F
E
R
E
NC
E
S
[1
]
B.
Ke
ll
e
n
b
e
rg
e
r,
M
.
V
o
lp
i,
a
n
d
D.
T
u
ia,
"
F
a
st
a
n
im
a
l
d
e
tec
ti
o
n
i
n
UA
V
im
a
g
e
s
u
sin
g
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
s,"
2
0
1
7
IE
EE
I
n
ter
n
a
t
io
n
a
l
Ge
o
sc
ien
c
e
a
n
d
Rem
o
te S
e
n
si
n
g
S
y
mp
o
si
u
m
(
IGARS
S
)
,
F
o
r
t
W
o
rth
,
T
X
,
USA
,
2
0
1
7
,
p
p
.
8
6
6
-
8
6
9
.
[2
]
C.
Ch
e
n
a
n
d
K.
L
iu
,
"
S
ti
n
g
ra
y
d
e
tec
ti
o
n
o
f
a
e
rial
i
m
a
g
e
s
w
it
h
r
e
g
io
n
-
b
a
se
d
c
o
n
v
o
l
u
ti
o
n
n
e
u
ra
l
n
e
tw
o
rk
,
"
2
0
1
7
IEE
E
In
ter
n
a
ti
o
n
a
l
C
o
n
fer
e
n
c
e
o
n
Co
n
su
me
r E
lec
tro
n
ics
-
T
a
iw
a
n
(
ICCE
-
T
W
)
,
T
a
ip
e
i,
T
a
i
wa
n
,
2
0
1
7
,
p
p
.
1
7
5
-
1
7
6
.
[3
]
L
.
B.
Bo
u
d
a
o
u
d
,
F
.
M
a
u
ss
a
n
g
,
R.
G
a
re
ll
o
,
a
n
d
A
.
Ch
e
v
a
ll
ier,
"
M
a
rin
e
Bir
d
De
te
c
ti
o
n
Ba
se
d
o
n
De
e
p
L
e
a
rn
in
g
u
sin
g
Hig
h
-
Re
so
lu
ti
o
n
A
e
rial
I
m
a
g
e
s,"
OCEANS
2
0
1
9
-
M
a
rs
e
il
le
,
M
a
rse
il
le,
F
ra
n
c
e
,
2
0
1
9
,
p
p
.
1
-
7
.
[4
]
S.
-
J.
Ho
n
g
,
Y.
Ha
n
,
S
.
-
Y.
Kim
,
A
.
-
Y.
Lee
,
a
n
d
G
.
Kim
,
"
A
p
p
li
c
a
ti
o
n
o
f
d
e
e
p
-
lea
rn
i
n
g
m
e
th
o
d
s
t
o
b
ird
d
e
tec
ti
o
n
u
sin
g
u
n
m
a
n
n
e
d
a
e
ri
a
l
v
e
h
icle
ima
g
e
r
y
,
"
S
e
n
so
rs
,
v
o
l.
1
9
,
n
o
.
7
,
p
p
.
1
-
16
,
2
0
1
9
.
[5
]
G
.
-
S
.
X
ia,
e
t
a
l
.
,
"
DO
TA
:
A
lar
g
e
-
sc
a
le
d
a
tas
e
t
f
o
r
o
b
jec
t
d
e
tec
ti
o
n
in
a
e
rial
im
a
g
e
s,
"
Pro
c
e
e
d
in
g
s
o
f
th
e
IEE
E
Co
n
fer
e
n
c
e
o
n
Co
m
p
u
ter
Vi
si
o
n
a
n
d
Pa
tt
e
rn
Rec
o
g
n
it
io
n
(
CVP
R)
,
2
0
1
8
,
p
p
.
3
9
7
4
-
3
9
8
3
.
[6
]
M
.
Ev
e
rin
g
h
a
m
,
L
.
V
a
n
G
o
o
l,
C.
K.
W
il
li
a
m
s,
J.
W
in
n
,
a
n
d
A
.
Zi
ss
e
r
m
a
n
,
"
T
h
e
p
a
sc
a
l
v
isu
a
l
o
b
jec
t
c
las
se
s
(
v
o
c
)
c
h
a
ll
e
n
g
e
,
"
In
ter
n
a
ti
o
n
a
l
j
o
u
rn
a
l
o
f
c
o
mp
u
ter
v
isio
n
,
v
o
l
.
8
8
,
n
o
.
2
,
p
p
.
3
0
3
-
3
3
8
,
2
0
1
0
.
[7
]
J.
De
n
g
,
W
.
Do
n
g
,
R.
S
o
c
h
e
r,
L
.
L
i,
K
.
L
i,
a
n
d
F
.
F
.
L
i,
"
Im
a
g
e
Ne
t:
A
lar
g
e
-
sc
a
le
h
iera
rc
h
ica
l
ima
g
e
d
a
tab
a
se
,
"
2
0
0
9
IE
EE
C
o
n
fer
e
n
c
e
o
n
Co
m
p
u
ter
Vi
sio
n
a
n
d
Pa
tt
e
rn
Rec
o
g
n
it
i
o
n
,
M
iam
i,
F
L
,
USA
,
2
0
0
9
,
p
p
.
2
4
8
-
2
5
5
.
[8
]
Y
.
Su
a
n
d
C.
-
C.
J
.
Ku
o
,
"
On
e
x
ten
d
e
d
lo
n
g
sh
o
rt
-
term
m
e
m
o
r
y
a
n
d
d
e
p
e
n
d
e
n
t
b
id
irec
ti
o
n
a
l
r
e
c
u
rre
n
t
n
e
u
ra
l
n
e
tw
o
rk
,
"
Ne
u
ro
c
o
mp
u
ti
n
g
,
v
o
l.
3
5
6
,
p
p
.
1
5
1
-
1
6
1
,
2
0
1
9
.
[9
]
A
.
Kriz
h
e
v
sk
y
,
I.
S
u
tsk
e
v
e
r,
a
n
d
G
.
E.
Hin
to
n
,
"
Im
a
g
e
n
e
t
c
l
a
ss
i
f
ic
a
ti
o
n
w
it
h
d
e
e
p
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
s,"
Co
mm
u
n
ica
ti
o
n
s o
f
th
e
ACM
,
v
o
l
.
6
0
,
n
o
.
6
,
p
p
.
8
4
-
9
0
,
2
0
1
7
.
[1
0
]
T
.
Yo
u
n
g
,
D.
Ha
z
a
ri
k
a
,
S
.
P
o
ri
a
,
a
n
d
E.
Ca
m
b
ria,
"
Re
c
e
n
t
T
re
n
d
s
i
n
De
e
p
L
e
a
rn
in
g
Ba
se
d
Na
tu
ra
l
L
a
n
g
u
a
g
e
P
r
o
c
e
ss
in
g
,
"
in
IEE
E
Co
mp
u
t
a
ti
o
n
a
l
I
n
telli
g
e
n
c
e
M
a
g
a
zin
e
,
v
o
l
.
1
3
,
n
o
.
3
,
p
p
.
5
5
-
7
5
,
A
u
g
.
2
0
1
8
.
[1
1
]
Y
.
S
u
,
K
.
F
a
n
,
N
.
Ba
c
h
,
C.
-
C.
J
.
Ku
o
,
a
n
d
F
.
H
u
a
n
g
,
"
Un
su
p
e
rv
i
se
d
m
u
lt
i
-
mo
d
a
l
n
e
u
ra
l
m
a
c
h
in
e
tran
sla
ti
o
n
,
"
in
Pro
c
e
e
d
in
g
s
o
f
t
h
e
IEE
E
Co
n
fer
e
n
c
e
o
n
C
o
mp
u
ter
Vi
sio
n
a
n
d
P
a
tt
e
rn
Rec
o
g
n
it
io
n
,
2
0
1
9
,
p
p
.
1
0
4
8
2
-
1
0
4
9
1
.
[1
2
]
T
.
N.
S
a
in
a
th
,
A
.
M
o
h
a
m
e
d
,
B.
Kin
g
sb
u
ry
,
a
n
d
B.
Ra
m
a
b
h
a
d
ra
n
,
"
De
e
p
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
s
f
o
r
L
V
CS
R,
"
2
0
1
3
IEE
E
In
ter
n
a
ti
o
n
a
l
Co
n
fer
e
n
c
e
o
n
Aco
u
stics
,
S
p
e
e
c
h
a
n
d
S
ig
n
a
l
Pr
o
c
e
ss
in
g
,
V
a
n
c
o
u
v
e
r,
BC,
Ca
n
a
d
a
,
2
0
1
3
,
p
p
.
8
6
1
4
-
8
6
1
8
.
[1
3
]
H
.
-
C
.
S
h
i
n
,
e
t
a
l
.
,
"
De
e
p
Co
n
v
o
lu
ti
o
n
a
l
Ne
u
ra
l
Ne
tw
o
rk
s
f
o
r
Co
m
p
u
ter
-
A
id
e
d
De
tec
ti
o
n
:
CNN
A
r
c
h
it
e
c
tu
re
s,
Da
tas
e
t
Ch
a
ra
c
teristics
a
n
d
T
ra
n
sf
e
r
L
e
a
rn
in
g
,
"
in
IEE
E
T
ra
n
s
a
c
ti
o
n
s o
n
M
e
d
ica
l
Im
a
g
i
n
g
,
v
o
l.
3
5
,
n
o
.
5
,
p
p
.
1
2
8
5
-
1
2
9
8
,
M
a
y
2
0
1
6
.
[1
4
]
S
.
Re
n
,
K.
He
,
R.
G
irsh
ick
,
a
n
d
J.
S
u
n
,
"
F
a
ste
r
r
-
c
n
n
:
T
o
w
a
rd
s
re
a
l
-
ti
m
e
o
b
jec
t
d
e
tec
ti
o
n
w
it
h
r
e
g
io
n
p
ro
p
o
sa
l
n
e
tw
o
rk
s,"
IEE
E
tra
n
sa
c
ti
o
n
s
o
n
p
a
tt
e
rn
a
n
a
lys
is
a
n
d
ma
c
h
i
n
e
i
n
telli
g
e
n
c
e
,
v
o
l
.
3
9
,
n
o
.
6
,
p
p
.
1
1
3
7
-
1
1
4
9
,
2
0
1
6
.
[1
5
]
K
.
He
,
G
.
Gk
io
x
a
ri,
P
.
Do
l
lar,
a
n
d
R
.
G
irsh
ick
,
"
M
a
sk
r
-
c
n
n
,
"
in
P
ro
c
e
e
d
in
g
s
o
f
t
h
e
IEE
E
i
n
ter
n
a
t
io
n
a
l
c
o
n
fer
e
n
c
e
o
n
c
o
m
p
u
ter
v
isi
o
n
,
2
0
1
7
,
p
p
.
2
9
6
1
-
2
9
6
9
.
[1
6
]
J
.
Re
d
m
o
n
,
S
.
Div
v
a
la,
R
.
G
irsh
ick
,
a
nd
A
.
F
a
rh
a
d
i
,
"
Yo
u
o
n
ly
lo
o
k
o
n
c
e
:
Un
if
ied
,
re
a
l
-
ti
m
e
o
b
jec
t
d
e
tec
ti
o
n
,
"
in
Pro
c
e
e
d
in
g
s
o
f
t
h
e
IEE
E
c
o
n
fer
e
n
c
e
o
n
c
o
mp
u
ter
v
isio
n
a
n
d
p
a
t
ter
n
re
c
o
g
n
i
ti
o
n
,
2
0
1
6
,
p
p
.
7
7
9
-
7
8
8
.
[1
7
]
M
.
J
.
S
h
a
f
iee
,
B.
Ch
y
w
l,
F
.
L
i,
a
n
d
A
.
W
o
n
g
,
"
F
a
st
YO
L
O:
A
f
a
st
y
o
u
o
n
ly
lo
o
k
o
n
c
e
sy
ste
m
f
o
r
re
a
l
-
ti
m
e
e
m
b
e
d
d
e
d
o
b
jec
t
d
e
tec
ti
o
n
in
v
id
e
o
,
"
a
rXiv p
re
p
ri
n
t
a
rXiv:1
7
0
9
.
0
5
9
4
3
,
2
0
1
7
.
[1
8
]
J.
Re
d
m
o
n
a
n
d
A
.
F
a
rh
a
d
i,
"
Yo
lo
v
3
:
A
n
in
c
re
m
e
n
tal
im
p
ro
v
e
m
e
n
t,
"
a
rXiv p
re
p
rin
t
a
rXiv:
1
8
0
4
.
0
2
7
6
7
,
2
0
1
8
.
[1
9
]
H.
L
a
w
a
n
d
J.
De
n
g
,
"
Co
rn
e
rn
e
t:
De
tec
ti
n
g
o
b
jec
ts
a
s
p
a
ired
k
e
y
p
o
in
ts,
"
in
Pr
o
c
e
e
d
in
g
s
o
f
th
e
Eu
ro
p
e
a
n
Co
n
fer
e
n
c
e
o
n
Co
m
p
u
ter
Vi
si
o
n
(
ECCV
)
,
2
0
1
8
,
p
p
.
7
3
4
-
7
5
0
.
[2
0
]
K.
Du
a
n
,
S
.
Ba
i,
L
.
X
ie,
H.
Qi,
Q.
Hu
a
n
g
,
a
n
d
Q.
T
ian
,
"
Ce
n
tern
e
t:
Ke
y
p
o
in
t
tri
p
lets
f
o
r
o
b
jec
t
d
e
tec
ti
o
n
,
"
in
Pro
c
e
e
d
in
g
s
o
f
t
h
e
IEE
E
In
ter
n
a
t
io
n
a
l
C
o
n
fer
e
n
c
e
o
n
Co
m
p
u
ter
Vi
sio
n
,
2
0
1
9
,
p
p
.
6
5
6
9
-
6
5
7
8
.
[2
1
]
L
.
Ch
e
n
,
G
.
P
a
p
a
n
d
re
o
u
,
I.
K
o
k
k
in
o
s,
K.
M
u
r
p
h
y
,
a
n
d
A
.
L
.
Yu
il
le,
"
De
e
p
Lab
:
S
e
m
a
n
ti
c
Im
a
g
e
S
e
g
m
e
n
tatio
n
w
it
h
De
e
p
Co
n
v
o
lu
ti
o
n
a
l
Ne
ts,
A
tro
u
s
Co
n
v
o
lu
ti
o
n
,
a
n
d
F
u
ll
y
Co
n
n
e
c
ted
CRF
s,"
in
IEE
E
T
ra
n
sa
c
ti
o
n
s
o
n
Pa
tt
e
r
n
An
a
lys
is
a
n
d
M
a
c
h
i
n
e
In
tell
ig
e
n
c
e
,
v
o
l.
4
0
,
n
o
.
4
,
p
p
.
8
3
4
-
8
4
8
,
1
Ap
ril
2
0
1
8
.
[2
2
]
T
.
-
Y
.
L
i
n
,
P
.
D
o
l
lar
,
R
.
G
ir
s
h
ic
k
,
K.
He
,
B
.
Ha
r
i
h
a
ra
n
,
a
n
d
S
.
Be
l
o
n
g
i
e
,
"
F
e
a
t
u
re
p
y
ra
m
id
n
e
t
w
o
rk
s
f
o
r
o
b
je
c
t
d
e
tec
t
i
o
n
,
"
in
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
I
EE
E
c
o
n
fer
e
n
c
e
o
n
c
o
m
p
u
ter
v
is
i
o
n
a
n
d
p
a
t
ter
n
re
c
o
g
n
it
i
o
n
,
2
0
1
7
,
p
p
.
2
1
1
7
-
2
1
2
5
.
Evaluation Warning : The document was created with Spire.PDF for Python.