I
nte
rna
t
io
na
l J
o
urna
l o
f
E
lect
rica
l a
nd
Co
m
pu
t
er
E
ng
ineering
(
I
J
E
CE
)
Vo
l.
1
6
,
No
.
1
,
Feb
r
u
ar
y
20
2
6
,
p
p
.
216
~
229
I
SS
N:
2088
-
8
7
0
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijece.
v
1
6
i
1
.
pp
2
1
6
-
2
2
9
216
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ec
e.
ia
esco
r
e.
co
m
AI SWLM
:
artifi
cia
l int
ellig
enc
e
-
b
a
sed sy
stem f
o
r
w
ildlife
mo
nitoring
Arun G
o
v
ind
a
n
K
rish
na
n
1
,
J
a
y
a
r
a
ma
n B
hu
v
a
na
2
,
M
irn
a
lin
ee
T
ha
ng
a
Na
da
r
T
ha
ng
a
T
h
ai
3
,
B
ha
ra
t
hk
um
a
r
Azha
g
iy
a
M
a
na
v
a
la
Ra
m
a
nu
j
a
m
4
1
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
,
S
I
V
ET
C
o
l
l
e
g
e
,
C
h
e
n
n
a
i
,
I
n
d
i
a
2
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
r
i
n
g
,
S
r
i
S
i
v
a
s
u
b
r
a
ma
n
i
y
a
N
a
d
a
r
C
o
l
l
e
g
e
o
f
E
n
g
i
n
e
e
r
i
n
g
,
C
h
e
n
n
a
i
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Ma
y
1
4
,
2
0
2
5
R
ev
is
ed
Sep
2
2
,
2
0
2
5
Acc
ep
ted
No
v
2
3
,
2
0
2
5
De
tec
ti
o
n
a
n
d
re
c
o
g
n
it
io
n
o
f
wild
a
n
ima
ls
a
re
e
ss
e
n
ti
a
l
fo
r
a
n
ima
l
su
rv
e
il
lan
c
e
,
b
e
h
a
v
i
o
r
m
o
n
it
o
r
in
g
a
n
d
sp
e
c
ies
c
o
u
n
ti
n
g
.
In
tr
u
sio
n
o
f
a
n
ima
ls
a
n
d
t
h
e
d
isa
ste
r
to
b
e
c
a
u
se
d
c
a
n
b
e
a
v
e
rted
b
y
t
h
e
ti
m
e
ly
re
c
o
g
n
it
i
o
n
o
f
in
tru
d
in
g
a
n
ima
ls.
An
a
rti
f
icia
l
in
tell
ig
e
n
c
e
-
b
a
se
d
s
y
ste
m
fo
r
wild
li
fe
m
o
n
it
o
r
in
g
(AI
S
WL
M
)
is
d
e
sig
n
e
d
a
n
d
imp
lem
e
n
ted
o
n
t
h
e
c
a
m
e
ra
trap
ima
g
e
s.
Th
e
c
h
a
ll
e
n
g
e
s
su
c
h
a
s
d
e
tec
ti
n
g
a
n
d
re
c
o
g
n
izi
n
g
a
n
ima
ls
o
f
d
iffere
n
t
siz
e
s,
sh
a
p
e
,
a
n
g
les
a
n
d
sc
a
le,
re
c
o
g
n
izin
g
t
h
e
a
n
ima
ls
o
f
sa
m
e
a
n
d
d
iffere
n
t
sp
e
c
ies
,
d
e
tec
ti
n
g
t
h
e
m
u
n
d
e
r
v
a
rio
u
s
il
l
u
m
in
a
ti
o
n
c
o
n
d
it
io
n
s,
wi
th
p
o
se
v
a
rian
ts
a
n
d
o
c
c
lu
si
o
n
a
re
a
d
d
re
ss
e
d
b
y
id
e
n
ti
fy
i
n
g
th
e
o
p
ti
m
a
l
we
ig
h
ts
o
f
th
e
d
e
e
p
lea
rn
i
n
g
a
rc
h
it
e
c
tu
r
e
,
AI
S
WL
M
.
M
o
d
e
ls
we
re
trai
n
e
d
u
si
n
g
G
o
ld
S
tan
d
a
rd
S
n
a
p
sh
o
t
S
e
re
n
g
e
t
i
d
a
tas
e
t
with
ra
n
d
o
m
we
ig
h
ts
a
n
d
t
h
e
b
e
st
we
ig
h
ts
o
f
m
o
d
e
l
we
re
u
se
d
a
s
in
it
ial
we
i
g
h
ts
f
o
r
trai
n
in
g
t
h
e
a
u
g
m
e
n
ted
d
a
ta.
Th
is
h
a
s
d
o
u
b
led
th
e
p
e
rfo
r
m
a
n
c
e
in
term
s
o
f
m
e
a
n
a
v
e
ra
g
e
p
re
c
isio
n
,
wh
ich
c
a
n
b
e
in
ter
p
re
ted
.
K
ey
w
o
r
d
s
:
An
im
al
i
n
tr
u
s
io
n
C
am
er
a
tr
ap
im
ag
es
C
SP
Den
s
eNe
t
Dee
p
l
ea
r
n
in
g
PANet
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
:
J
ay
ar
am
an
B
h
u
v
a
n
a
Dep
ar
tm
en
t o
f
C
o
m
p
u
ter
Scie
n
ce
an
d
E
n
g
in
ee
r
in
g
,
Sri
Siv
a
s
u
b
r
am
an
iy
a
Na
d
ar
C
o
lleg
e
o
f
E
n
g
in
ee
r
i
n
g
Kala
v
ak
k
am
,
C
h
en
n
ai,
I
n
d
ia
E
m
ail: b
h
u
v
a
n
aj@
s
s
n
.
ed
u
.
in
1.
I
NT
RO
D
UCT
I
O
N
R
ec
o
g
n
izin
g
an
im
als
ir
r
esp
ec
tiv
e
o
f
wild
o
r
d
o
m
esti
c
is
es
s
e
n
tial
in
a
v
ar
iety
o
f
ap
p
licatio
n
s
n
am
ely
s
p
ec
ies
co
u
n
tin
g
,
s
u
r
v
eillan
ce
o
f
tr
ess
-
p
ass
in
g
o
f
t
h
e
an
im
a
ls
,
an
d
m
o
n
ito
r
in
g
t
h
eir
b
e
h
a
v
io
r
s
f
o
r
m
an
ag
i
n
g
th
em
ef
f
ec
tiv
el
y
.
B
y
d
etec
tin
g
th
e
p
r
esen
ce
o
f
a
n
im
als,
t
h
e
d
is
aster
ca
u
s
ed
b
y
t
h
eir
i
n
tr
u
s
io
n
c
o
u
ld
b
e
r
ed
u
ce
d
.
Als
o
,
th
e
welf
ar
e
o
f
t
h
e
an
im
als
is
m
o
s
t
ess
en
tial
in
b
alan
cin
g
th
e
ec
o
s
y
s
tem
.
C
o
u
n
tin
g
th
e
an
im
als,
with
th
eir
s
p
ec
ies
m
an
u
ally
i
n
ap
p
licatio
n
s
lik
e
ce
n
s
u
s
wi
ll
b
e
tim
e
co
n
s
u
m
in
g
an
d
ex
p
en
s
iv
e
o
p
er
atio
n
.
I
n
v
o
lv
i
n
g
h
u
m
an
s
to
m
o
n
ito
r
th
e
in
tr
u
s
io
n
o
f
an
im
als will b
e
ted
io
u
s
an
d
r
is
k
y
.
Hu
m
an
s
p
er
ce
iv
e
,
v
is
u
alize
wh
at
th
ey
s
ee
a
n
d
ac
t
u
p
o
n
ac
co
r
d
in
g
l
y
.
Hu
m
an
v
is
u
al
r
ec
o
g
n
itio
n
s
y
s
tem
p
o
s
s
ess
e
s
o
b
ject
co
n
s
tan
cy
,
ab
ilit
y
to
r
ec
o
g
n
ize
o
b
j
ec
t
ac
r
o
s
s
d
if
f
er
en
t
v
iewp
o
in
t
co
n
d
itio
n
s
s
u
ch
as
o
r
ien
tatio
n
,
lig
h
tin
g
,
an
d
o
b
je
ct
s
ize
v
ar
iab
ilit
y
.
W
e
ca
n
in
t
er
p
r
et
th
e
e
n
titi
es
in
ea
ch
s
ce
n
e,
ir
r
esp
ec
tiv
e
o
f
th
eir
s
ize,
s
ca
le,
an
g
les,
r
o
tate
d
o
r
tr
an
s
lated
.
Sem
an
tic
m
ea
n
in
g
o
f
im
a
g
es
an
d
v
id
e
o
s
ar
e
u
s
ef
u
l
in
f
o
r
m
ati
o
n
f
o
r
an
y
s
ce
n
e
i
n
ter
p
r
etatio
n
with
s
ev
er
al
ap
p
licatio
n
s
in
v
o
lv
in
g
s
elf
-
d
r
iv
in
g
ca
r
s
,
n
a
v
ig
atio
n
in
m
o
b
ile
r
o
b
o
tics
,
s
tr
ee
t
tr
af
f
ic
o
b
s
er
v
a
tio
n
s
,
s
o
cc
er
g
am
e
a
n
aly
s
is
,
s
m
ar
t
r
o
o
m
ca
m
er
as,
m
o
n
ito
r
i
n
g
o
f
eld
er
ly
.
T
h
e
d
etec
tio
n
o
f
a
n
im
al
in
tr
u
s
io
n
c
an
b
e
m
o
d
elled
as o
b
ject
r
ec
o
g
n
itio
n
p
r
o
b
lem
.
An
im
al
class
if
icatio
n
an
d
r
ec
o
g
n
itio
n
p
la
y
a
m
ajo
r
r
o
le
in
s
u
r
v
eillan
ce
,
au
to
m
atic
ca
r
d
r
iv
in
g
to
p
r
ev
en
t
ac
cid
en
ts
,
an
im
al
p
o
p
u
latio
n
s
u
r
v
e
y
f
o
r
en
d
an
g
e
r
ed
s
p
ec
ies,
an
im
al
s
u
r
v
eillan
ce
.
T
o
b
alan
ce
th
e
wild
life
ec
o
lo
g
y
,
m
o
n
ito
r
in
g
an
d
s
u
r
v
eillan
ce
will
b
e
th
e
in
h
er
e
n
t
p
a
r
t
o
f
th
e
s
y
s
tem
.
T
h
e
r
e
ar
e
f
ew
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
A
I
S
WL
M:
a
r
tifi
cia
l in
tellig
e
n
ce
-
b
a
s
ed
s
ystem
fo
r
w
ild
life
mo
n
ito
r
in
g
(
A
r
u
n
G.
K
.
)
217
s
u
cc
ess
f
u
l
wo
r
k
s
to
m
ain
tain
th
e
b
io
d
iv
e
r
s
ity
o
f
th
e
b
ir
d
s
,
wh
er
e
f
o
r
wild
an
im
al
m
o
n
ito
r
in
g
s
u
c
h
s
o
p
h
is
ticated
s
y
s
tem
tech
n
iq
u
es
th
at
h
av
e
b
ee
n
d
ep
lo
y
ed
h
av
e
f
ew
in
h
er
e
n
t
d
r
awb
ac
k
s
s
u
ch
as
lack
in
r
o
b
u
s
tn
ess
,
co
v
er
ag
e
ar
e
a,
r
e
liab
ilit
y
o
f
th
e
eq
u
ip
m
en
t
an
d
th
e
d
elay
in
in
f
o
r
m
in
g
th
e
d
ec
is
io
n
s
to
th
e
au
th
o
r
ities
.
E
ar
ly
d
ec
is
io
n
-
m
a
k
in
g
s
y
s
tem
s
h
o
u
ld
b
e
in
p
la
ce
wh
er
ev
er
we
h
av
e
h
u
m
an
wild
co
n
f
lict.
T
h
e
s
o
lu
tio
n
to
h
a
n
d
le
th
is
is
s
u
e
is
to
im
itate
th
e
co
g
n
itiv
e
f
u
n
ctio
n
ality
o
f
b
r
ain
i
n
r
ec
o
g
n
izin
g
o
b
jects.
T
h
is
m
o
tiv
ated
u
s
to
in
v
esti
g
ate
d
if
f
er
en
t
th
eo
r
ies
t
o
d
esig
n
n
o
v
el
co
m
p
u
tatio
n
al
f
r
am
ewo
r
k
s
to
s
o
lv
e
s
ig
n
if
ican
t
v
is
u
al
p
er
ce
p
tio
n
task
s
.
T
h
er
e
is
a
g
r
o
win
g
n
ee
d
f
o
r
AI
-
b
ased
s
y
s
tem
s
th
at
ca
n
au
to
m
atica
lly
d
etec
t
an
d
class
if
y
wild
life
s
p
ec
ies
in
r
ea
l
-
wo
r
ld
e
n
v
ir
o
n
m
en
ts
f
o
r
p
r
o
a
ctiv
e
co
n
s
er
v
atio
n
a
n
d
to
ad
d
r
ess
h
u
m
an
-
a
n
im
al
co
n
f
lict.
W
e
aim
to
d
esig
n
an
au
to
m
atic
co
m
p
u
tatio
n
al
f
r
am
ewo
r
k
t
o
p
r
o
v
id
e
e
f
f
icien
t
s
o
lu
tio
n
s
f
o
r
an
im
al
r
ec
o
g
n
itio
n
,
p
er
f
o
r
m
ed
e
f
f
o
r
tl
ess
ly
b
y
a
h
u
m
an
b
ein
g
.
Sev
er
al
ch
allen
g
es
in
an
im
al
d
etec
tio
n
an
d
r
ec
o
g
n
itio
n
ar
e,
An
im
als
o
f
d
if
f
e
r
en
t
s
izes
(
s
m
all
an
d
lar
g
e
)
,
Occ
lu
s
io
n
,
m
u
ltip
le
s
p
ec
ies
in
s
am
e
f
r
a
m
e,
an
im
al
lo
o
k
in
g
s
im
ilar
to
b
ac
k
g
r
o
u
n
d
,
p
a
r
tially
v
is
ib
le
an
im
als
with
o
u
t
o
c
c
lu
s
io
n
,
co
u
n
tin
g
th
e
n
u
m
b
er
o
f
an
im
als
in
g
iv
e
n
f
r
am
e,
v
a
r
io
u
s
illu
m
in
atio
n
co
n
d
itio
n
s
,
with
p
o
s
e
v
ar
ian
t,
d
etec
tin
g
m
u
ltip
le
in
s
tan
c
es
o
f
s
am
e
s
p
ec
ies
an
im
als
in
a
s
in
g
le
f
r
am
e,
lo
ca
tin
g
th
e
d
etec
ted
a
n
im
al
in
a
c
lu
tter
ed
b
ac
k
g
r
o
u
n
d
.
T
h
is
p
ap
er
p
r
esen
ts
an
en
d
-
to
-
en
d
d
ee
p
lear
n
in
g
b
ased
ar
tific
ial
in
tellig
en
ce
-
b
ased
s
y
s
tem
f
o
r
wild
life
m
o
n
it
o
r
in
g
(
AI
SW
L
M)
f
o
r
an
im
al
d
etec
tio
n
f
r
o
m
ca
m
er
a
tr
ap
im
ag
es.
T
h
e
n
o
v
elty
o
f
t
h
e
wo
r
k
is
in
th
e
u
s
e
o
f
tr
an
s
f
er
lear
n
in
g
o
n
y
o
u
o
n
ly
lo
o
k
o
n
ce
v
e
r
s
io
n
5
(
YOL
Ov
5
)
v
a
r
ian
ts
with
a
class
-
b
alan
ce
d
au
g
m
en
tatio
n
s
tr
ateg
y
th
at
s
ig
n
if
ican
tl
y
im
p
r
o
v
es
p
er
f
o
r
m
an
ce
in
ter
m
s
o
f
m
ea
n
av
e
r
ag
e
p
r
ec
is
io
n
(
m
AP)
,
p
r
ec
is
io
n
,
an
d
r
ec
all
w
h
en
co
m
p
ar
e
d
with
th
e
ex
is
tin
g
ap
p
r
o
ac
h
es o
n
th
e
Ser
en
g
eti
d
ataset.
An
ex
ten
s
iv
e
s
et
o
f
ex
p
e
r
im
en
ts
wer
e
co
n
d
u
cted
to
i
d
en
tify
b
est s
u
itab
le
m
o
d
el
f
o
r
a
n
im
al
d
etec
tio
n
;
C
las
s
im
b
alan
ce
is
s
u
e
is
h
an
d
led
b
y
a
p
p
ly
in
g
au
g
m
e
n
tatio
n
an
d
th
e
b
est
weig
h
ts
ar
e
u
s
ed
to
in
itialize
th
e
tr
ain
in
g
o
f
th
e
en
h
an
ce
d
d
ataset;
d
etailed
q
u
alitativ
e
an
d
q
u
an
titativ
e
an
aly
s
is
wer
e
d
o
n
e
o
n
th
e
p
e
r
f
o
r
m
an
ce
o
f
th
e
p
r
o
p
o
s
ed
s
y
s
tem
.
T
h
e
ar
ticle
is
o
r
g
an
ized
as
f
o
l
lo
ws:
s
ec
tio
n
2
d
is
cu
s
s
es
th
e
ex
is
tin
g
s
y
s
tem
s
f
o
r
o
b
ject
d
e
tectio
n
an
d
class
if
icatio
n
f
o
r
an
im
als.
Sectio
n
3
p
r
o
p
o
s
es
AI
-
SW
L
M
ar
ch
itectu
r
e
an
d
th
e
d
esig
n
o
f
its
f
u
n
ctio
n
al
co
m
p
o
n
en
ts
.
Sectio
n
4
d
is
cu
s
s
es
th
e
im
p
lem
en
tatio
n
r
elate
d
co
n
ce
p
ts
o
f
AI
-
SW
L
M.
Sec
tio
n
5
p
r
o
v
id
es
th
e
p
lan
o
f
d
if
f
er
en
t
ex
p
er
im
en
ts
;
s
ec
tio
n
6
p
r
esen
ts
th
e
d
etai
led
q
u
alitativ
e
an
d
q
u
a
n
titativ
e
an
aly
s
is
o
f
th
e
r
esu
lts
an
d
co
m
p
a
r
is
o
n
f
o
llo
w
ed
b
y
t
h
e
co
n
cl
u
s
io
n
in
s
ec
tio
n
7
.
2.
SURVE
Y
O
F
E
XI
ST
I
NG
W
O
RK
S
Hu
m
an
s
p
er
ce
iv
e
v
is
u
al
in
f
o
r
m
atio
n
th
r
o
u
g
h
th
e
r
etin
a
,
wh
i
ch
is
tr
an
s
m
itted
v
ia
th
e
o
p
tical
n
er
v
e
to
th
e
b
r
ai
n
,
w
h
er
e
it
is
in
ter
p
r
et
ed
in
to
o
b
jects
an
d
s
ce
n
es.
R
e
s
ea
r
ch
er
s
h
av
e
f
o
u
n
d
th
at
n
e
u
r
o
n
al
f
ir
in
g
p
atter
n
s
in
th
e
in
f
e
r
io
r
tem
p
o
r
al
co
r
te
x
s
tr
o
n
g
ly
co
r
r
elate
with
s
u
cc
ess
f
u
l
o
b
ject
r
ec
o
g
n
itio
n
task
s
.
T
h
e
h
u
m
an
v
is
u
al
r
ec
o
g
n
itio
n
s
y
s
tem
in
clu
d
es
n
eu
r
o
n
al
r
ep
r
esen
tatio
n
s
ca
p
ab
l
e
o
f
p
atter
n
d
is
cr
im
in
atio
n
.
Ar
tific
ial
i
n
tellig
en
ce
(
AI
)
,
a
d
o
m
ai
n
o
f
co
m
p
u
ter
s
cien
ce
,
h
as
d
ev
elo
p
ed
m
ec
h
an
is
m
s
to
in
co
r
p
o
r
ate
s
u
c
h
i
n
tellig
en
ce
th
r
o
u
g
h
alg
o
r
ith
m
s
th
at
au
to
m
ate
h
u
m
an
-
lik
e
p
er
ce
p
tio
n
a
n
d
o
b
ject
r
ec
o
g
n
itio
n
.
I
n
co
r
p
o
r
atin
g
th
e
n
eu
r
o
n
r
ep
r
esen
tatio
n
p
atter
n
s
o
f
th
e
h
u
m
an
b
r
ain
i
n
to
co
m
p
u
tatio
n
al
alg
o
r
ith
m
s
ca
n
lead
to
ef
f
icien
t
o
b
ject
r
ec
o
g
n
itio
n
.
Ob
ject
d
etec
tio
n
r
em
ain
s
o
n
e
o
f
th
e
m
o
s
t
ch
allen
g
in
g
task
s
in
co
m
p
u
ter
v
is
io
n
,
r
e
q
u
ir
in
g
i
d
en
tific
atio
n
o
f
o
b
ject
in
s
tan
ce
s
v
ar
y
in
g
in
c
o
lo
r
,
s
h
a
p
e,
lo
ca
tio
n
,
p
o
s
e,
illu
m
in
atio
n
,
an
d
b
ac
k
g
r
o
u
n
d
.
I
t
s
er
v
es
as
th
e
f
o
u
n
d
atio
n
f
o
r
ap
p
licatio
n
s
s
u
ch
as
s
eg
m
en
tatio
n
,
ca
p
tio
n
i
n
g
,
o
b
ject
tr
ac
k
in
g
,
an
d
s
ce
n
e
u
n
d
er
s
tan
d
i
n
g
.
R
ea
l
-
wo
r
ld
a
p
p
licatio
n
s
in
clu
d
e
au
to
n
o
m
o
u
s
v
eh
icles
an
d
s
u
r
v
eillan
ce
s
y
s
tem
s
[
1
]
.
E
ar
lier
,
m
ac
h
in
e
lear
n
in
g
alg
o
r
ith
m
s
wer
e
wid
ely
u
s
ed
f
o
r
o
b
ject
d
etec
tio
n
.
T
h
e
w
o
r
k
in
[
2
]
f
o
c
u
s
es
o
n
ef
f
icien
t
m
u
ltis
ca
le
f
ea
tu
r
es
f
o
r
im
ag
e
r
etr
iev
al.
Ho
wev
er
,
s
h
ap
e
f
ea
tu
r
es
o
f
ten
s
tr
u
g
g
le
u
n
d
e
r
v
ar
y
in
g
s
h
ad
o
ws
an
d
illu
m
in
atio
n
.
E
x
tr
ac
tin
g
ed
g
es
in
wild
life
im
a
g
er
y
r
em
ai
n
s
d
if
f
icu
lt.
Mu
lti
-
r
eso
lu
tio
n
f
ea
tu
r
es
[
3
]
ar
e
well
s
u
ited
f
o
r
d
etec
tin
g
o
b
jects
o
f
v
ar
y
in
g
s
h
ap
es.
D
o
m
ain
g
e
n
er
aliza
tio
n
ch
allen
g
es
ar
e
ad
d
r
ess
ed
in
o
b
ject
d
etec
tio
n
,
esp
ec
ially
i
n
wild
life
d
atasets
wh
er
e
en
v
ir
o
n
m
en
tal
v
ar
iatio
n
af
f
ec
ts
th
e
p
er
f
o
r
m
a
n
ce
[
4
]
.
T
h
e
ev
o
lu
tio
n
o
f
d
ee
p
lear
n
in
g
alg
o
r
ith
m
s
an
d
s
u
p
p
o
r
ti
n
g
h
ig
h
-
en
d
s
y
s
tem
s
h
as
s
i
g
n
if
ican
tly
ad
v
an
ce
d
c
o
m
p
u
ter
v
is
io
n
.
V
ar
io
u
s
d
ee
p
lear
n
i
n
g
tech
n
iq
u
es
[
5
]
–
[
1
1
]
n
o
w
allo
w
au
to
m
atic
ex
tr
ac
tio
n
o
f
f
ea
tu
r
es
f
r
o
m
im
ag
es
an
d
v
i
d
eo
s
.
Prio
r
wo
r
k
o
n
ca
m
e
r
a
tr
ap
im
ag
es
ca
n
b
e
b
r
o
ad
l
y
class
if
ied
in
to
two
ca
teg
o
r
ies:
ap
p
licatio
n
o
f
p
r
e
-
t
r
ain
ed
m
o
d
els an
d
u
s
e
o
f
o
b
je
ct
d
etec
tio
n
an
d
r
ec
o
g
n
itio
n
m
o
d
els.
A
n
o
tab
le
ex
am
p
le
is
m
u
lti
-
task
g
en
er
ativ
e
ad
v
e
r
s
ar
ial
n
etwo
r
k
(
MT
GAN)
[
1
2
]
,
a
n
en
d
-
to
-
e
n
d
f
r
am
ewo
r
k
d
ev
elo
p
ed
to
d
etec
t
s
m
all
-
s
ca
le
o
b
jects,
in
wh
i
ch
a
g
en
er
at
o
r
u
p
s
ca
les
im
ag
e
r
eso
lu
tio
n
an
d
a
d
is
cr
im
in
ato
r
s
im
u
ltan
eo
u
s
ly
ev
alu
ate
a
u
th
en
ticity
an
d
th
e
p
r
esen
ce
o
f
th
e
o
b
ject.
T
h
is
is
ev
alu
ated
o
n
co
m
m
o
n
o
b
jects
in
c
o
n
tex
t
(
C
OC
O)
an
d
W
I
DE
R
FAC
E
d
atasets
,
th
eir
m
o
d
el
u
s
ed
R
esNet5
0
as
its
b
ac
k
b
o
n
e
an
d
in
co
r
p
o
r
ated
a
r
eg
r
ess
io
n
m
o
d
u
le
to
r
ef
in
e
d
etails.
T
h
is
m
u
lti
-
task
s
tr
u
ctu
r
e
h
elp
s
m
ain
tain
o
b
ject
-
lev
el
clar
ity
in
lo
w
-
r
eso
lu
tio
n
r
eg
io
n
s
,
m
ak
in
g
it we
ll
-
s
u
itab
le
f
o
r
wild
life
m
o
n
ito
r
in
g
ap
p
licatio
n
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
6
,
No
.
1
,
Feb
r
u
ar
y
20
2
6
:
2
1
6
-
229
218
Ma
s
k
R
eg
io
n
-
b
ased
co
n
v
o
l
u
tio
n
al
n
eu
r
al
n
etwo
r
k
(
R
-
C
NN)
[
1
3
]
,
d
e
r
iv
ed
f
r
o
m
f
aster
R
-
C
NN,
h
as
b
ee
n
u
s
ed
f
o
r
ca
ttle
d
etec
tio
n
an
d
c
o
u
n
tin
g
,
s
u
cc
ess
f
u
lly
h
a
n
d
lin
g
o
cc
lu
s
io
n
an
d
o
v
er
lap
b
y
lev
e
r
ag
in
g
b
in
ar
y
m
ask
class
if
icatio
n
.
Simp
ler
C
NN
-
b
ased
m
o
d
els
h
av
e
b
e
en
u
s
ed
to
class
if
y
im
ag
es
i
n
to
m
am
m
als
an
d
r
ep
tiles
[
1
4
]
,
o
r
m
o
r
e
g
r
an
u
la
r
ly
in
to
Sn
ak
es,
L
izar
d
s
,
an
d
T
o
ad
s
/Fro
g
s
[
1
5
]
.
C
am
er
a
tr
a
p
s
ar
e
wid
ely
u
s
ed
to
ca
p
tu
r
e
wild
life
im
ag
es
f
o
r
p
o
p
u
latio
n
s
u
r
v
ey
s
.
Ho
wev
er
,
th
ese
tr
ap
s
also
r
ec
o
r
d
h
u
m
an
s
an
d
f
alse
tr
ig
g
er
s
d
u
e
to
win
d
o
r
v
eg
etatio
n
[
1
6
]
.
T
o
class
if
y
s
u
ch
im
ag
es
i
n
to
wild
life
,
h
u
m
an
,
o
r
em
p
t
y
,
a
d
ee
p
lear
n
in
g
ap
p
r
o
ac
h
u
s
ed
Alex
Net
-
9
6
to
s
eg
m
en
t
f
o
r
e
g
r
o
u
n
d
o
b
jects
an
d
ad
d
r
ess
class
i
m
b
alan
ce
b
y
c
o
lo
r
au
g
m
en
tatio
n
,
ac
h
iev
in
g
7
3
.
1
3
% r
ec
all.
T
wo
-
lev
el
class
if
icatio
n
o
n
t
h
e
Sn
ap
s
h
o
t
Ser
en
g
eti
d
ataset
was
p
er
f
o
r
m
ed
in
[
1
7
]
.
T
h
e
f
ir
s
t
s
tag
e
was
a
b
in
ar
y
class
if
ier
f
o
r
an
im
al
p
r
esen
ce
,
f
o
llo
wed
b
y
m
u
lti
-
class
cla
s
s
if
icatio
n
in
to
2
6
s
p
ec
ies
u
s
in
g
p
r
e
-
tr
ain
ed
m
o
d
els
s
u
ch
as
Alex
Net,
v
is
u
al
g
eo
m
etr
y
g
r
o
u
p
(
VGG)
,
Go
o
g
L
eNe
t,
an
d
v
ar
io
u
s
R
esNet
v
er
s
io
n
s
,
ac
h
iev
in
g
9
3
.
6
%
with
en
s
e
m
b
le
m
eth
o
d
s
.
T
h
is
wo
r
k
u
s
ed
th
e
s
am
e
d
ataset
as
o
u
r
s
b
u
t
f
o
c
u
s
ed
o
n
class
if
icatio
n
,
n
o
t
o
b
ject
d
ete
ctio
n
.
Oth
er
ef
f
o
r
ts
u
s
ed
p
r
e
-
t
r
ain
ed
m
o
d
els
[
1
8
]
lik
e
Den
s
eNe
t2
0
1
,
I
n
ce
p
tio
n
-
R
esNet
-
V3
,
an
d
NASNetM
o
b
ile
to
class
if
y
3
5
an
im
al
s
p
ec
ies
in
th
e
Par
k
s
C
an
ad
a
d
ataset.
Au
g
m
en
tatio
n
tech
n
iq
u
es
h
elp
e
d
m
itig
ate
c
lass
im
b
alan
ce
,
im
p
r
o
v
in
g
p
er
f
o
r
m
a
n
ce
to
7
1
.
2
%
af
ter
e
n
s
em
b
le.
Similar
ly
,
R
esNet
-
1
8
was
em
p
lo
y
ed
in
[
1
9
]
to
class
if
y
an
im
als
ac
r
o
s
s
5
8
class
es
f
r
o
m
ca
m
er
a
tr
ap
im
ag
es
tak
en
in
ten
U.
S.
s
tate
s
.
Pre
-
tr
ain
ed
m
o
d
e
ls
lik
e
I
n
ce
p
tio
n
V3
,
Mo
b
ile
Net,
an
d
VGG
-
1
6
wer
e
u
s
ed
f
o
r
class
if
y
in
g
s
ix
an
im
al
ca
teg
o
r
ies
[
2
0
]
.
A
r
o
b
u
s
t,
lo
ca
tio
n
-
in
v
ar
ia
n
t
class
if
i
er
tr
ain
ed
o
n
d
atasets
lik
e
Fl
i
ck
R
an
d
iNatu
r
alis
t
was
p
r
o
p
o
s
ed
in
[
2
1
]
.
Usi
n
g
Ker
as
-
R
etin
aNe
t,
th
eir
m
o
d
els
ac
h
iev
ed
a
m
AP
o
f
8
2
.
3
3
%
–
8
8
.
5
9
%
wh
en
test
ed
o
n
Sn
ap
s
h
o
t
Ser
en
g
eti.
Facial
d
etec
tio
n
u
s
in
g
Fas
ter
-
R
C
N
N
was
ex
p
lo
r
ed
in
[
2
2
]
u
s
in
g
th
e
a
n
im
al
f
ac
e
d
atab
ase
(
AFD)
,
ac
h
iev
in
g
8
7
.
0
3
% a
cc
u
r
ac
y
.
YOL
Ov
2
was
u
s
ed
in
[
2
3
]
f
o
r
s
p
ec
ies r
ec
o
g
n
itio
n
.
R
ec
en
t
s
u
r
v
ey
s
an
d
m
o
d
el
in
n
o
v
atio
n
s
em
p
h
asize
th
e
g
r
o
win
g
r
o
le
o
f
d
ee
p
lear
n
in
g
in
ec
o
lo
g
ical
m
o
n
ito
r
in
g
.
Fo
r
in
s
tan
ce
,
Z
h
a
o
et
a
l.
[
2
4
]
p
r
o
v
i
d
es
a
d
etail
ed
r
ev
iew
o
f
C
NN
-
b
ased
wil
d
life
class
if
icatio
n
f
r
o
m
ca
m
er
a
tr
a
p
im
ag
es,
h
ig
h
lig
h
tin
g
ch
allen
g
es
s
u
ch
as
clas
s
im
b
alan
ce
an
d
f
e
atu
r
e
ex
tr
ac
tio
n
in
u
n
co
n
tr
o
lled
e
n
v
ir
o
n
m
en
ts
.
B
h
attac
h
ar
jee
et
a
l.
[
2
5
]
p
r
o
p
o
s
es
YOL
O
-
b
ased
ar
ch
itect
u
r
es
cu
s
to
m
ized
f
o
r
an
im
al
d
etec
tio
n
u
n
d
er
v
ar
y
in
g
en
v
ir
o
n
m
en
tal
co
n
d
itio
n
s
,
s
h
o
win
g
im
p
r
o
v
ed
d
etec
t
io
n
p
r
ec
is
io
n
an
d
r
o
b
u
s
tn
ess
ac
r
o
s
s
r
ea
l
-
wo
r
ld
d
atasets
.
3.
M
E
T
H
O
D
T
h
e
p
r
o
p
o
s
ed
ar
tific
ial
in
tellig
en
ce
-
b
ased
s
y
s
tem
f
o
r
wild
life
m
o
n
ito
r
in
g
(
A
I
SW
L
M)
will
r
ec
o
g
n
ize
th
e
ca
teg
o
r
y
o
f
th
e
s
p
ec
ies
i
n
th
e
g
i
v
en
ca
m
e
r
a
tr
ap
im
a
g
e.
T
h
e
o
b
ject
o
f
in
ter
est
is
th
e
an
im
al,
wh
ic
h
is
d
etec
ted
b
y
t
h
e
p
o
p
u
lar
a
n
d
e
f
f
icien
t
o
b
ject
d
etec
tio
n
alg
o
r
ith
m
YOL
Ov
5
.
T
h
e
im
ag
es
o
f
d
if
f
er
en
t
an
im
als
ca
p
tu
r
ed
in
tr
ap
ca
m
e
r
as
u
n
d
er
d
if
f
er
en
t
lig
h
tin
g
co
n
d
itio
n
s
ar
e
f
ed
to
tr
ain
th
e
p
r
o
p
o
s
ed
o
b
ject
d
etec
tio
n
m
o
d
el
f
o
r
l
o
ca
lizatio
n
o
f
a
n
i
m
al
s
p
ec
ies,
r
ec
o
g
n
itio
n
o
f
s
p
ec
ies
an
d
co
u
n
tin
g
o
f
s
p
ec
ies.
T
h
is
will
en
ab
le
u
s
to
m
o
n
ito
r
an
im
al
m
o
v
em
en
ts
,
lo
ca
tio
n
s
an
d
f
u
r
th
er
n
o
tify
th
e
r
esp
ec
tiv
e
f
o
r
est
d
e
p
ar
tm
e
n
ts
r
eg
ar
d
i
n
g
th
ei
r
m
o
v
em
en
t
n
ea
r
ag
r
icu
ltu
r
al
f
ield
s
an
d
r
esid
en
tial
ar
ea
s
.
Statis
t
ics
o
f
an
im
als
ca
n
b
e
u
s
ed
b
y
th
e
f
o
r
est
d
ep
ar
tm
en
t to
m
ain
tain
th
e
ec
o
s
y
s
tem
.
T
h
e
p
r
o
p
o
s
ed
AI
SW
L
M
ac
ce
p
ts
th
e
in
p
u
ts
in
th
e
f
o
r
m
o
f
im
ag
es
ca
p
tu
r
ed
an
d
ap
p
lies
th
e
YOL
OV5
ar
ch
itectu
r
e
th
at
h
as
a
b
ac
k
b
o
n
e
s
y
s
tem
,
n
ec
k
an
d
d
etec
tio
n
h
ea
d
to
lo
ca
lize
an
d
class
if
y
th
e
an
im
al.
T
h
e
in
p
u
t
im
ag
es
in
b
atch
es
will
b
e
p
r
o
ce
s
s
ed
th
r
o
u
g
h
th
e
b
ac
k
b
o
n
e,
n
ec
k
an
d
t
h
e
h
ea
d
o
u
tp
u
ts
th
e
lo
ca
lized
as
th
e
wild
an
im
als
alo
n
g
with
th
eir
n
am
es
an
d
c
o
u
n
t.
T
h
e
p
r
o
p
o
s
ed
AI
SW
L
M
s
y
s
tem
co
m
b
in
es
s
tan
d
ar
d
d
ee
p
lear
n
in
g
c
o
m
p
o
n
en
ts
s
u
ch
as
th
e
YOL
Ov
5
d
etec
tio
n
a
r
ch
itectu
r
e
with
n
o
v
el
en
h
an
ce
m
en
ts
in
clu
d
in
g
a
two
-
s
tag
e
tr
ain
in
g
p
r
o
ce
d
u
r
e
u
s
in
g
p
r
etr
ain
ed
weig
h
ts
,
class
-
b
alan
ce
d
d
ata
au
g
m
en
tatio
n
,
a
n
d
ev
alu
atio
n
ac
r
o
s
s
d
if
f
e
r
en
t
m
o
d
el
co
n
f
ig
u
r
atio
n
s
.
T
h
e
n
o
v
e
lty
is
in
th
e
s
tr
u
ctu
r
ed
au
g
m
e
n
tatio
n
p
i
p
elin
e
an
d
r
eu
s
in
g
b
est
-
tr
ain
ed
weig
h
ts
to
im
p
r
o
v
e
g
en
e
r
aliza
tio
n
o
f
t
h
e
wild
life
d
etec
tio
n
m
o
d
el
c
h
allen
g
es,
in
clu
d
in
g
p
o
o
r
illu
m
in
atio
n
,
clu
tter
ed
b
a
ck
g
r
o
u
n
d
s
,
an
d
d
i
f
f
er
en
t sp
ec
i
es.
T
h
e
f
u
n
ctio
n
al
co
m
p
o
n
en
t
s
o
f
AI
S
W
L
M
an
d
th
e
id
en
tify
in
g
th
e
b
est
s
u
itab
le
m
o
d
el
f
o
r
d
etec
tio
n
is
s
h
o
wn
in
Alg
o
r
ith
m
1
an
d
Fig
u
r
e
1
.
T
h
e
wo
r
k
in
g
o
f
f
u
n
ctio
n
al
c
o
m
p
o
n
en
ts
is
elab
o
r
ated
in
th
e
f
o
llo
win
g
3
s
u
b
s
ec
tio
n
s
.
Alg
o
r
ith
m
1
.
Ar
tific
ial
in
tellig
en
ce
-
b
ased
s
y
s
tem
f
o
r
wild
life
m
o
n
ito
r
in
g
(
AI
SW
L
M)
Input:
-
Training: Wildlife images, labels, bounding box coordinates
-
Testing: Images
Output:
-
Recognized objects, labels, bounding box coordinates, counted species
Step 1: Let X ← Original imbalanced training dataset with labels and bounding boxes
Function Main ()
1. WLM_O ← WLM(X)
2. WLM_A_RW ← WLM (X_Enhanced with random weights)
3. WLM_A_BW ← WLM (X_Enhanced with best weights of WLM_O)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
A
I
S
WL
M:
a
r
tifi
cia
l in
tellig
e
n
ce
-
b
a
s
ed
s
ystem
fo
r
w
ild
life
mo
n
ito
r
in
g
(
A
r
u
n
G.
K
.
)
219
4. WLM_O_Count ← Counting_Species (WLM_O)
5. WLM_A_RW_Count ← Counting_Species (WLM_A_RW)
6. WLM_A_BW_Count ← Counting_Species (WLM_A_BW)
7. AI_SWLM ← Performance_comparison (WLM_O, WLM_A_RW, WLM_A_BW)
Return: AI_WLM model
Function Augment(X)
1. X_Enhanced ← manual augment
2. X_imglevel ← Image_level_augment(X_Enhanced)
3. X_pixellevel ← Pixel_level_augment(X_Enhanced)
4. X_augmented ← X_imglevel + X_pixellevel
Return: X_augmented
Function WLM(X)
1. X.remove_duplicates ()
2. X.remove_corrupted ()
3. X_preprocess ← X.reshape (640, 640)
4. X_augmented ← Augment(X_preprocess)
5. X_featuremap ← CSP_Network (X_augmented)
6. X_featuremap ← Spatial_Pyramid_Pooling (X_featuremap)
7. X_feature_Pyramid ← PANet(X_featuremap)
8. X_PANet ← X_feature_Pyramid
9. (Class_prob, Obj_scores, b_boxes) ← Detection_Head(X_PANet)
Note:
-
Class_prob: class probabilities
-
Obj_scores: objectness scores
-
b_boxes: bounding boxes
Return: WLM model
Fig
u
r
e
1
.
Ov
e
r
v
iew
o
f
p
r
o
p
o
s
ed
AI
SW
L
M
3
.
1
.
F
e
a
t
ure
ex
t
r
a
ct
io
n net
w
o
rk
T
h
e
f
ea
tu
r
e
e
x
tr
ac
tio
n
p
ar
t
o
f
th
e
ar
ch
itectu
r
e
will
s
er
v
e
as
t
h
e
b
ac
k
b
o
n
e
a
n
d
h
elp
to
e
x
tr
a
ct
f
ea
tu
r
es
f
r
o
m
th
e
in
p
u
t
im
ag
es.
T
h
e
B
ac
k
b
o
n
e
o
f
th
e
AI
SW
L
M
h
as
th
e
cr
o
s
s
s
tag
e
p
ar
tial
n
et
wo
r
k
(
C
SP
Net)
an
d
s
p
atial
p
y
r
am
id
s
p
o
o
lin
g
as
th
e
m
ajo
r
f
u
n
ctio
n
al
u
n
its
th
at
ex
tr
ac
t
th
e
f
ea
tu
r
es
f
r
o
m
th
e
in
p
u
t
im
a
g
es.
T
h
e
r
ich
f
ea
tu
r
es
o
f
th
e
wild
life
s
p
ec
ies
in
ea
ch
f
r
a
m
e
will
b
e
ex
tr
ac
ted
u
s
in
g
a
lig
h
tweig
h
t
n
etwo
r
k
ca
lled
C
SP
Ne
t,
wh
er
e
th
e
f
ea
tu
r
e
m
a
p
is
d
iv
id
e
d
in
to
h
alv
es
an
d
ar
e
co
m
b
in
e
d
af
ter
p
ass
in
g
th
e
m
th
r
o
u
g
h
d
if
f
e
r
en
t
lay
er
s
.
Similar
ly
,
th
e
g
r
ad
ien
t
in
f
o
r
m
atio
n
is
also
m
ad
e
to
f
lo
w
th
r
o
u
g
h
d
if
f
e
r
en
t
p
ath
s
a
n
d
ar
e
co
n
ca
ten
ated
an
d
tr
an
s
itio
n
ed
w
h
ile
p
ass
in
g
d
u
r
in
g
th
e
b
ac
k
p
r
o
p
ag
at
io
n
.
T
h
e
b
asic
b
u
ild
in
g
b
lo
c
k
o
f
th
is
b
ac
k
b
o
n
e
s
tr
u
ctu
r
e
is
d
en
s
e
b
lo
ck
,
wh
ic
h
will
h
av
e
s
ev
er
al
d
en
s
e
lay
er
s
in
it.
I
n
a
d
e
n
s
e
b
lo
ck
th
e
in
p
u
t
o
f
o
n
e
d
en
s
e
lay
er
will
b
e
th
e
co
n
ca
ten
atio
n
o
f
p
r
e
v
io
u
s
d
en
s
e
lay
er
’
s
o
u
tp
u
t
an
d
its
in
p
u
t.
T
h
is
ar
r
an
g
em
en
t
will
h
elp
in
ac
cu
m
u
latin
g
k
n
o
wled
g
e
f
r
o
m
all
o
f
th
e
p
r
ev
i
o
u
s
lay
er
s
.
Mu
ltip
le
d
en
s
e
b
lo
ck
s
wi
ll
b
e
s
ep
ar
ated
b
y
tr
an
s
itio
n
al
lay
er
s
.
T
h
e
tr
an
s
itio
n
al
lay
er
h
as
s
et
o
f
co
n
v
o
lu
t
io
n
al
lay
er
s
an
d
an
av
er
a
g
e
p
o
o
lin
g
lay
er
o
f
1
×
1
an
d
2
×
2
r
esp
ec
tiv
ely
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
6
,
No
.
1
,
Feb
r
u
ar
y
20
2
6
:
2
1
6
-
229
220
C
SP
Ne
t
is
m
ad
e
u
p
o
f
p
ar
tial
Den
s
eNe
t
b
lo
ck
a
n
d
Par
tial
tr
an
s
itio
n
al
lay
er
s
.
Par
tial
Den
s
eNe
t
b
lo
ck
will
d
iv
id
e
th
e
f
ea
tu
r
e
m
ap
i
n
to
two
s
ay
f
i
an
d
f
j
f
r
o
m
t
h
e
b
ase
lay
er
,
wh
er
e
o
n
e
h
alf
f
i
wi
ll
p
ass
th
r
o
u
g
h
t
h
e
d
en
s
e
b
lo
ck
an
d
th
e
o
t
h
er
h
alf
will
b
e
co
n
ca
ten
ated
with
th
e
in
p
u
t
o
f
tr
an
s
itio
n
al
la
y
er
.
I
n
th
e
p
ar
tial
tr
an
s
itio
n
al
lay
er
s
,
th
e
f
ir
s
t
lay
er
ac
ce
p
ts
th
e
o
u
tp
u
t
o
f
its
p
r
e
v
io
u
s
p
ar
tial
d
en
s
e
b
lo
c
k
as
in
p
u
t.
T
h
e
o
u
t
p
u
t
o
f
th
e
tr
an
s
itio
n
al
la
y
er
is
n
o
w
co
n
ca
ten
ated
with
th
e
o
th
e
r
h
alf
o
f
th
e
f
ea
tu
r
e
m
a
p
f
j
an
d
s
er
v
ed
to
th
e
n
ex
t
tr
an
s
itio
n
al
lay
er
.
T
h
e
C
SP
N
et
with
it
s
h
ier
ar
ch
ical
f
ea
tu
r
e
f
u
s
io
n
ap
p
r
o
ac
h
will
s
tr
en
g
th
en
th
e
lear
n
in
g
ab
ilit
y
b
y
g
iv
in
g
th
e
in
n
er
m
o
s
t
lay
er
s
with
th
e
f
ea
tu
r
es
ex
tr
ac
ted
f
r
o
m
th
e
ea
r
ly
d
en
s
e
lay
er
s
.
Du
e
to
its
p
ar
tial
co
n
n
ec
tio
n
s
,
C
SP
Net
e
x
tr
ac
ts
v
er
y
d
iv
e
r
s
if
ied
s
et
o
f
f
ea
tu
r
es
th
at
will
h
elp
to
d
is
cr
im
in
ate
ag
ain
s
t
th
e
wild
life
o
f
s
am
e
an
d
d
if
f
er
e
n
t
s
p
ec
ies.
T
h
e
s
p
atial
p
y
r
am
i
d
p
o
o
lin
g
(
SP
P)
in
th
e
b
ac
k
b
o
n
e
s
tag
e
o
f
YOL
OV5
,
is
a
v
ar
ian
t
o
f
B
ag
o
f
W
o
r
d
s
(
B
o
W
)
m
o
d
el
r
em
o
v
es
th
e
lim
itatio
n
o
f
C
o
n
v
o
lu
tio
n
al
lay
er
s
wo
r
k
in
g
with
f
ix
ed
s
ized
in
p
u
ts
.
T
h
is
ch
ar
ac
ter
is
tic
o
f
th
e
SP
P
m
ak
es
th
e
m
o
d
el
s
ca
le
in
v
ar
ian
t
a
n
d
av
o
i
d
s
o
v
er
f
itti
n
g
.
T
h
e
o
u
t
p
u
t
f
r
o
m
C
SP
Net
is
p
ass
ed
to
SP
P
b
ef
o
r
e
th
e
f
ea
t
u
r
es
ar
e
s
en
t
to
th
e
Nec
k
p
h
a
s
e
o
f
th
e
n
etwo
r
k
.
SP
P
m
ak
es
m
u
ltip
le
co
p
ies
o
f
th
e
f
ea
tu
r
es
an
d
ap
p
lies
m
ax
p
o
o
lin
g
o
f
d
if
f
er
en
t
s
ized
k
e
r
n
e
ls
an
d
c
o
n
ca
ten
ated
th
em
an
d
ca
n
g
e
n
er
ate
o
u
tp
u
t
o
f
f
ix
e
d
len
g
th
ir
r
esp
ec
tiv
e
o
f
th
e
in
p
u
t size
u
s
in
g
t
h
e
m
u
lti
-
l
ev
el
s
p
atial
b
in
s
.
3
.
2
.
F
e
a
t
ure
py
ra
m
id pa
t
h a
g
g
re
g
a
t
io
n net
wo
rk
(
P
ANe
t
)
T
h
e
n
e
x
t
s
tep
in
o
b
je
ct
d
e
t
ec
t
io
n
o
f
an
im
a
l
s
is
th
e
co
n
s
tr
u
c
tio
n
o
f
f
ea
tu
r
e
p
y
r
am
i
d
s
b
y
p
a
th
ag
g
r
eg
at
io
n
n
e
two
r
k
(
P
A
Ne
t)
in
th
e
n
e
ck
s
ta
g
e
o
f
Y
OL
O
V5
.
P
A
Ne
t
p
er
f
o
r
m
s
th
e
in
s
t
an
ce
s
eg
m
en
ta
tio
n
th
a
t
s
er
v
e
s
a
s
th
e
n
e
ck
p
ar
t
o
f
th
e
s
in
g
le
s
tag
e
o
b
je
ct
d
e
tec
t
i
o
n
m
o
d
e
l.
T
h
e
p
u
r
p
o
s
e
o
f
th
e
f
ea
tu
r
e
p
y
r
a
m
id
i
s
to
g
en
er
al
iz
e
th
e
m
o
d
e
l
o
n
o
b
jec
t
s
ca
lin
g
an
d
to
s
eg
m
en
t
an
im
a
l
in
s
t
an
c
es
i
n
th
e
ca
m
er
a
tr
ap
im
ag
e
s
b
y
m
ain
ta
in
in
g
th
ei
r
s
p
a
ti
al
in
f
o
r
m
at
io
n
.
T
h
e
m
o
d
e
l
n
e
ed
s
to
d
ete
ct
th
e
s
am
e
w
il
d
l
if
e
s
p
ec
ie
s
in
d
i
f
f
er
en
t
s
iz
e
s
an
d
s
c
al
es
.
T
h
i
s
f
ea
tu
r
e
p
y
r
am
id
i
s
d
e
s
ig
n
ed
to
ex
t
r
ac
t
m
u
lt
i
-
s
c
al
e
f
ea
tu
r
e
m
ap
s
an
d
p
er
f
o
r
m
s
w
el
l
o
n
u
n
s
e
en
o
r
h
id
d
en
d
a
ta.
T
h
e
r
e
aso
n
wh
y
PA
Ne
t’
s
ch
o
s
en
i
s
b
ec
au
s
e
it
h
e
lp
s
in
p
r
o
p
er
lo
c
al
iza
t
io
n
o
f
p
ix
el
s
f
o
r
m
a
s
k
f
o
r
m
at
io
n
.
P
AN
et
h
elp
s
in
b
o
t
to
m
-
u
p
p
ath
a
u
g
m
e
n
ta
tio
n
,
ad
ap
t
iv
e
f
ea
tu
r
e
p
o
o
li
n
g
,
f
u
l
ly
co
n
n
ec
te
d
f
u
s
io
n
.
Featu
r
es
will
f
lo
w
v
ia
b
o
th
b
o
tto
m
-
u
p
an
d
to
p
-
d
o
wn
p
ath
way
s
th
at
wo
r
k
ar
o
u
n
d
t
h
e
s
p
atial
r
eso
lu
tio
n
b
ef
o
r
e
s
en
d
in
g
th
e
m
f
o
r
p
r
ed
ictio
n
s
tag
e
o
f
th
e
n
etwo
r
k
.
T
h
e
B
o
tto
m
-
u
p
n
etwo
r
k
u
s
es
R
es
Net
ar
ch
itectu
r
e,
th
r
o
u
g
h
wh
ic
h
th
e
f
ea
tu
r
es
f
lo
w,
th
at
h
elp
s
in
s
em
an
tic
d
etec
tio
n
an
d
r
ed
u
ce
s
th
e
s
p
atial
d
im
en
s
io
n
in
to
h
alf
.
T
h
e
to
p
-
d
o
wn
f
lo
w,
u
p
s
am
p
les
an
d
au
g
m
en
ts
th
e
p
r
ev
i
o
u
s
lay
er
’
s
o
u
tp
u
t
an
d
p
r
o
p
a
g
ates th
e
f
ea
tu
r
es th
at
ar
e
s
em
an
tically
s
ig
n
if
ican
t.
3
.
3
.
O
bje
c
t
lo
ca
liza
t
io
n a
nd
predict
io
n us
ing
det
ec
t
io
n h
ea
d
T
h
e
th
ir
d
s
tag
e
o
f
Pro
p
o
s
ed
AI
SW
L
M
is
th
e
h
ea
d
o
f
YOL
OV5
,
wh
ich
p
r
ed
icts
th
e
b
o
u
n
d
in
g
b
o
x
co
o
r
d
in
ates,
o
b
jectless
s
co
r
e
alo
n
g
with
th
e
lab
el
o
f
th
e
p
r
ed
icted
an
im
al.
I
t
ap
p
lies
an
ch
o
r
b
o
x
es
o
n
f
ea
tu
r
es
m
ap
s
f
r
o
m
PANet
an
d
g
en
e
r
at
es
f
in
al
o
u
t
p
u
t
v
ec
to
r
s
with
cl
ass
p
r
o
b
ab
ilit
ies,
o
b
jectless
s
co
r
es,
an
d
b
o
u
n
d
in
g
b
o
x
es.
Fr
o
m
t
h
e
d
etec
ted
a
n
im
als,
th
e
c
o
u
n
t
o
f
th
e
s
p
ec
ies
b
elo
n
g
in
g
to
th
e
s
am
e
o
r
d
if
f
e
r
en
t
wild
life
s
p
ec
ies
in
th
e
s
ce
n
e
is
p
r
o
ce
s
s
ed
wh
ich
ca
n
b
e
co
m
m
u
n
icate
d
to
t
h
e
au
th
o
r
ities
co
n
ce
r
n
ed
.
T
h
e
d
etec
tio
n
h
ea
d
will
h
av
e
3
lay
er
s
th
at
ac
ce
p
t
th
e
f
ea
tu
r
e
m
a
p
s
o
f
s
izes
n
am
ely
,
8
0
×
8
0
,
4
0
×
4
0
an
d
2
0
×
2
0
r
esp
ec
tiv
ely
to
d
etec
t
th
e
an
im
als
o
f
d
if
f
er
en
t
s
izes.
T
h
ese
d
etec
tio
n
lay
er
s
g
en
er
ate
an
o
u
tp
u
t
v
ec
to
r
with
p
r
e
d
icted
b
o
u
n
d
i
n
g
b
o
x
co
o
r
d
in
ates,
class
p
r
o
b
ab
ilit
y
an
d
ca
teg
o
r
y
o
f
th
e
a
n
im
al
p
r
e
d
icted
.
4.
I
M
P
L
E
M
E
NT
A
T
I
O
N
4
.
1
.
Da
t
a
s
et
d
escript
io
n
Sn
ap
s
h
o
t
Ser
en
g
eti
is
o
n
e
o
f
t
h
e
wo
r
ld
’
s
lar
g
est
ca
m
er
a
tr
ap
p
r
o
jects
with
7
.
1
m
illi
o
n
im
a
g
es
ac
r
o
s
s
1
2
s
ea
s
o
n
s
.
I
n
th
o
s
e
7
.
1
m
il
lio
n
im
ag
es,
o
v
er
7
6
%
o
f
i
m
ag
es
wer
e
em
p
ty
.
Ser
en
g
et
i
Natio
n
al
Par
k
in
T
an
za
n
ia
is
b
est
k
n
o
wn
f
o
r
t
h
e
m
ass
iv
e
an
im
al
m
ig
r
atio
n
s
o
f
W
ild
eb
ee
s
t,
Z
eb
r
a
th
at
d
r
iv
e
th
e
cy
cle
o
f
its
d
y
n
am
ic
ec
o
s
y
s
tem
.
T
h
e
m
o
s
t
co
m
m
o
n
wild
life
s
p
ec
ies
in
th
e
d
ataset
ar
e
W
ild
eb
ee
s
t,
Z
eb
r
a
an
d
Gaz
elle
T
h
o
m
p
s
o
n
s
.
T
o
tally
2
2
5
ca
m
er
as
wer
e
d
e
p
lo
y
ed
to
ca
p
tu
r
e
th
e
wild
life
im
ag
es
in
Ser
e
n
g
eti
Natio
n
al
Par
k
,
T
an
za
n
ia,
E
ast
Af
r
ica.
C
itizen
v
o
lu
n
teer
s
h
av
e
b
ee
n
in
v
o
lv
e
d
in
th
is
p
r
o
ject
to
an
n
o
tate
th
e
im
ag
es
th
at
h
av
e
4
8
class
es
o
f
wild
life
s
p
ec
ies
in
it.
T
h
e
lab
elled
d
ataset
n
am
ed
Go
ld
Stan
d
ar
d
Sn
a
p
s
h
o
t
Ser
en
g
eti
with
4
6
class
es
is
u
s
ed
in
th
is
wo
r
k
f
o
r
s
p
ec
ies
d
etec
tio
n
an
d
r
e
co
g
n
itio
n
.
B
o
u
n
d
in
g
b
o
x
c
o
o
r
d
in
ates
ar
o
u
n
d
ea
c
h
an
im
al
in
a
ca
m
er
a
tr
ap
im
ag
e
ar
e
p
r
o
v
id
ed
with
th
e
d
ata
s
et
.
4
.
2
.
Da
t
a
pre
-
pro
ce
s
s
ing
a
nd
a
ug
m
ent
a
t
io
n
T
h
e
Go
ld
Stan
d
ar
d
Sn
a
p
s
h
o
t
Ser
en
g
eti
h
as
th
e
wid
th
,
h
ei
g
h
t,
,
,
an
d
o
f
th
e
b
o
u
n
d
in
g
b
o
x
es
f
o
r
ea
ch
o
f
t
h
e
in
s
tan
ce
o
f
a
n
im
als
p
r
esen
t
in
an
im
ag
e.
As
th
e
f
ir
s
t
s
tep
o
f
p
r
e
-
p
r
o
ce
s
s
in
g
,
th
ese
m
ea
s
u
r
es
o
f
th
e
b
o
u
n
d
in
g
b
o
x
co
o
r
d
in
ates
ar
e
co
n
v
er
t
ed
in
to
,
,
wid
th
an
d
h
eig
h
t.
7
0
%
o
f
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
A
I
S
WL
M:
a
r
tifi
cia
l in
tellig
e
n
ce
-
b
a
s
ed
s
ystem
fo
r
w
ild
life
mo
n
ito
r
in
g
(
A
r
u
n
G.
K
.
)
221
im
ag
es
in
th
e
d
ataset
ar
e
u
s
e
d
f
o
r
tr
ain
in
g
th
e
s
p
ec
ies
d
etec
tio
n
m
o
d
el,
wh
er
e
2
0
%
o
f
im
ag
es
ar
e
u
s
ed
f
o
r
v
alid
atio
n
an
d
1
0
% o
f
im
ag
es
ar
e
u
s
ed
f
o
r
test
in
g
.
Data
s
et
im
ag
es
wer
e
in
d
if
f
er
en
t
s
izes,
s
o
th
ey
wer
e
co
n
v
e
r
ted
in
to
a
s
tan
d
ar
d
im
ag
e
s
i
ze
o
f
6
4
0
,
6
4
0
.
Data
clea
n
in
g
o
p
er
atio
n
s
wer
e
p
er
f
o
r
m
e
d
to
r
em
o
v
e
t
h
e
d
u
p
licate,
co
r
r
u
p
t
a
n
d
u
n
u
s
ed
im
ag
es.
Alo
n
g
with
th
at,
th
e
im
ag
es
f
o
r
wh
ich
th
e
lab
el
an
d
an
n
o
tatio
n
s
ar
e
n
o
t
g
iv
en
wer
e
r
em
o
v
ed
f
r
o
m
th
e
d
ataset.
Few
im
ag
es
with
m
is
m
atch
ed
f
ile
n
am
es
h
av
e
also
b
ee
n
r
em
o
v
ed
f
r
o
m
t
h
e
d
ataset
as
a
p
ar
t
o
f
d
ata
clea
n
s
in
g
.
Ad
d
itio
n
al
im
ag
es
with
lab
els
n
am
ely
s
co
u
t
g
u
ar
d
,
jeep
,
tr
u
ck
s
,
r
o
ck
s
,
s
k
y
an
d
im
ag
es
with
o
u
t
an
im
als
ar
e
r
em
o
v
ed
f
r
o
m
th
e
d
ataset.
Au
g
m
en
tatio
n
is
th
e
p
r
o
ce
s
s
o
f
ad
d
in
g
n
ew
lear
n
in
g
s
am
p
les
eith
er
f
r
o
m
th
e
ex
is
tin
g
d
ata
o
r
b
y
g
en
er
atin
g
n
ew
s
y
n
th
etic
d
ata
th
at
in
cr
ea
s
es
th
e
s
ize
o
f
th
e
d
ataset
to
en
h
an
ce
th
e
lear
n
in
g
o
f
th
e
m
o
d
el
an
d
h
en
ce
en
h
an
ce
th
e
p
e
r
f
o
r
m
an
ce
with
b
etter
ac
cu
r
ac
y
.
As
o
n
e
o
f
th
e
r
eg
u
lar
izatio
n
tech
n
iq
u
es,
d
ata
au
g
m
en
tatio
n
av
o
id
s
m
o
d
el
o
v
er
f
itti
n
g
b
y
in
tr
o
d
u
cin
g
a
d
iv
er
s
e
s
et
o
f
lear
n
in
g
i
n
s
tan
ce
s
.
Sev
er
al
d
ata
au
g
m
e
n
tatio
n
s
tech
n
iq
u
es
h
av
e
b
ee
n
d
e
v
el
o
p
ed
to
ac
h
iev
e
b
etter
p
er
f
o
r
m
an
ce
in
co
m
p
u
ter
v
is
io
n
p
r
o
b
lem
s
,
t
h
at
ca
n
b
e
ca
teg
o
r
ized
in
to
i
m
ag
e
lev
el
an
d
at
p
ix
el
lev
e
l
au
g
m
en
tatio
n
,
to
im
p
r
o
v
e
t
h
e
lear
n
in
g
ab
ilit
y
a
n
d
to
in
c
r
ea
s
e
th
e
v
ar
ian
ce
i
n
th
e
tr
ain
in
g
d
ata.
At
th
e
p
ix
el
lev
el
th
e
p
r
o
p
o
s
ed
AI
SW
L
M
s
y
s
tem
p
er
f
o
r
m
s
b
asic
lin
ea
r
an
d
a
f
f
in
e
tr
a
n
s
f
o
r
m
atio
n
s
n
am
el
y
th
e
f
li
p
p
in
g
,
r
o
tatio
n
,
clip
p
in
g
,
ad
ap
tiv
e
s
ca
lin
g
an
d
m
o
d
if
y
i
n
g
th
e
b
r
ig
h
t
n
ess
an
d
co
n
tr
a
s
t
d
ata
au
g
m
en
tatio
n
s
.
At
th
e
im
ag
e
lev
el,
t
h
e
p
r
o
p
o
s
ed
s
y
s
tem
em
p
lo
y
s
co
p
y
-
p
aste,
letter
b
o
x
to
m
ain
tain
asp
ec
t
r
atio
an
d
m
ix
-
u
p
,
m
o
s
aic
m
eth
o
d
s
f
o
r
d
ata
au
g
m
en
tatio
n
.
C
u
t
m
ix
a
n
d
m
o
s
aic
m
eth
o
d
s
co
m
b
in
e
2
a
n
d
4
im
ag
es
r
esp
ec
tiv
ely
to
g
eth
e
r
to
g
en
er
ate
a
n
ew
lear
n
in
g
s
am
p
le.
I
n
th
e
Go
ld
Stan
d
ar
d
Sn
ap
s
h
o
t
Ser
en
g
eti
d
ataset,
it h
as b
ee
n
o
b
s
er
v
ed
th
at
t
h
e
n
u
m
b
er
o
f
i
m
ag
es in
a
f
ew
o
f
t
h
e
class
es
is
n
o
t
s
u
f
f
icien
t
to
tr
ain
a
o
b
ject
d
etec
t
io
n
an
d
r
ec
o
g
n
itio
n
m
o
d
el,
s
in
ce
th
e
d
ataset
is
im
b
alan
ce
d
.
T
h
e
g
r
ea
ter
n
u
m
b
er
o
f
s
am
p
les
ar
e
f
o
u
n
d
in
th
e
class
W
ild
b
ea
s
t,
an
d
v
er
y
a
s
m
aller
n
u
m
b
e
r
o
f
s
am
p
les
ar
e
f
o
u
n
d
in
class
es
n
am
ely
Har
e,
W
ater
b
u
ck
,
Ver
v
et
Mo
n
k
ey
,
a
n
d
L
e
o
p
a
r
d
R
h
in
o
ce
r
o
s
.
T
o
en
h
an
ce
t
h
e
v
ar
ia
n
ce
o
f
th
e
tr
ain
in
g
d
ata,
im
ag
e
a
u
g
m
e
n
tatio
n
is
th
e
b
est
ap
p
r
o
ac
h
b
ef
o
r
e
ap
p
ly
in
g
an
y
d
ee
p
lear
n
in
g
f
r
am
ewo
r
k
.
Ap
ar
t
f
r
o
m
th
e
ab
o
v
e
tech
n
iq
u
es,
th
e
p
r
o
p
o
s
ed
AI
SW
L
M
also
au
g
m
en
ts
th
e
d
ataset
with
n
ew
im
ag
es
tak
en
f
r
o
m
th
e
web
to
f
o
r
m
en
h
an
ce
d
d
ata
s
et.
T
h
is
i
s
d
o
n
e
to
o
v
er
co
m
e
th
e
d
ata
im
b
alan
ce
p
r
o
b
lem
ac
r
o
s
s
all
th
e
cla
s
s
e
s
.
T
h
ese
m
an
u
ally
au
g
m
en
ted
i
m
ag
es
o
f
th
e
en
h
an
ce
d
d
atase
t
ar
e
an
n
o
tated
an
d
th
e
b
o
u
n
d
in
g
b
o
x
co
o
r
d
in
ates
ar
e
th
e
n
u
s
ed
b
y
th
e
p
r
o
p
o
s
ed
s
y
s
tem
d
u
r
in
g
tr
ain
in
g
.
T
h
e
en
h
a
n
ce
d
d
ataset
was
s
p
lit
in
to
3
p
ar
ts
in
th
e
s
am
e
p
r
o
p
o
r
tio
n
as
th
at
o
f
th
e
o
r
ig
in
al
d
ataset
an
d
u
s
ed
f
o
r
t
r
ain
in
g
,
v
alid
atio
n
an
d
test
in
g
th
e
an
im
al
o
b
ject
d
etec
tio
n
an
d
r
ec
o
g
n
itio
n
m
o
d
e
l,
AI
SW
L
M.
AI
SW
L
M
m
o
d
el
is
d
ev
el
o
p
e
d
a
n
d
t
r
ai
n
ed
in
a
NV
I
D
I
A
G
eF
o
r
ce
R
T
X
2
0
8
0
GPU
1
1
GB
s
y
s
tem
w
it
h
C
UDA
v
e
r
s
i
o
n
9
.
1
u
s
in
g
P
y
T
o
r
c
h
v
e
r
s
i
o
n
1
.
7
.
1
an
d
Py
th
o
n
v
e
r
s
i
o
n
3
.
6
.
1
0
.
T
h
e
m
o
d
e
ls
we
r
e
t
r
ai
n
ed
u
s
i
n
g
s
to
c
h
asti
c
g
r
ad
ie
n
t
d
es
ce
n
t
(
S
GD)
wit
h
a
n
in
iti
al
l
ea
r
n
i
n
g
r
at
e
o
f
0
.
0
1
a
n
d
m
o
m
e
n
t
u
m
o
f
0
.
9
3
7
.
L
o
s
s
was
co
m
p
u
te
d
u
s
i
n
g
GI
o
U
f
o
r
b
o
u
n
d
i
n
g
b
o
x
r
e
g
r
ess
i
o
n
a
n
d
b
in
a
r
y
c
r
o
s
s
-
e
n
t
r
o
p
y
f
o
r
cl
ass
i
f
ic
at
io
n
a
n
d
o
b
je
ct
n
ess
.
W
e
h
av
e
ch
o
s
e
n
YOL
Ov
5
f
o
r
its
s
p
ee
d
a
n
d
a
cc
u
r
ac
y
in
r
ea
l
-
tim
e
o
b
jec
t
d
et
ec
ti
o
n
tas
k
s
a
n
d
h
e
n
c
e
s
u
it
a
b
le
f
o
r
d
e
p
l
o
y
m
en
t
i
n
wil
d
l
if
e
s
u
r
v
eil
lan
ce
.
C
SP
N
et
e
n
h
a
n
c
es
t
h
e
l
ea
r
n
in
g
ca
p
a
b
i
lit
y
b
y
e
n
a
b
li
n
g
f
e
at
u
r
e
r
e
u
s
e
a
n
d
g
r
a
d
ie
n
t
f
lo
w.
PAN
et
h
e
lp
s
t
o
r
etai
n
s
p
ati
al
f
ea
t
u
r
es
a
n
d
i
m
p
r
o
v
e
l
o
c
ali
za
t
io
n
in
c
lu
tte
r
e
d
wil
d
li
f
e
s
c
e
n
es
.
T
o
ad
d
r
ess
s
e
v
e
r
e
c
lass
im
b
ala
n
c
e
i
n
t
h
e
S
er
e
n
g
eti
d
at
aset
,
a
tw
o
-
ti
er
a
u
g
m
e
n
ta
ti
o
n
s
t
r
at
e
g
y
w
a
s
u
s
ed
.
5.
E
XP
E
R
I
M
E
N
T
S
T
h
is
s
ec
tio
n
d
is
cu
s
s
es
th
e
v
ar
io
u
s
ex
p
er
im
en
ts
co
n
d
u
cted
t
o
id
en
tify
th
e
m
o
s
t
s
u
itab
le
AI
SW
L
M
m
o
d
el
f
o
r
d
etec
tin
g
th
e
an
im
al
s
.
a.
T
o
d
etec
t
an
d
r
ec
o
g
n
ize
th
e
a
n
im
als
in
th
e
Go
ld
Stan
d
ar
d
Sn
ap
s
h
o
t
Ser
en
g
eti
d
ataset
with
th
e
o
r
ig
i
n
al
s
et
o
f
im
ag
es,
th
at
ar
e
im
b
alan
ce
d
ac
r
o
s
s
th
e
class
es.
−
E
x
p
er
im
en
tatio
n
with
a
s
m
aller
m
o
d
el,
YOL
O
V5
m
r
ef
e
r
r
ed
to
as
wild
life
m
o
n
ito
r
in
g
with
o
r
ig
in
al
d
ataset
(
W
L
M
-
O1
)
.
−
E
x
p
er
im
en
tatio
n
with
lar
g
e
r
m
o
d
els with
o
r
ig
in
al
d
ataset,
YOL
O
V5
l r
ef
er
r
ed
as WLM
-
O2
.
b.
Dete
ctin
g
an
d
r
ec
o
g
n
izin
g
an
i
m
als with
au
g
m
en
ted
d
ataset
an
d
d
if
f
er
en
t w
eig
h
t i
n
itializatio
n
m
eth
o
d
s
.
−
E
x
p
er
im
en
tatio
n
with
s
m
alle
r
m
o
d
el,
YOL
O
V5
m
with
r
an
d
o
m
l
y
in
itialized
weig
h
ts
r
ef
er
r
ed
as
w
ild
life
m
o
n
ito
r
in
g
with
au
g
m
en
ted
d
ataset
an
d
r
an
d
o
m
w
eig
h
ts
(
W
L
M
-
A
-
R
W
1
)
.
−
E
x
p
er
im
en
tatio
n
with
lar
g
er
m
o
d
el
o
n
a
u
g
m
e
n
ted
d
atas
et,
YOL
O
V5
l
with
r
an
d
o
m
ly
in
itialized
weig
h
ts
r
ef
er
r
ed
(
W
L
M
-
A
-
R
W
2
)
.
−
E
x
p
er
im
en
tatio
n
with
s
m
aller
m
o
d
el
o
n
a
u
g
m
e
n
ted
d
ataset,
YOL
O
V5
m
with
u
s
in
g
b
est
weig
h
t
f
r
o
m
ex
p
er
im
en
t
W
L
M
-
O1
r
ef
e
r
r
ed
as
wild
life
m
o
n
ito
r
in
g
with
a
u
g
m
en
ted
d
ataset
an
d
b
est
tr
ain
ed
weig
h
ts
(
W
L
M
-
A
-
B
W
1
)
.
−
E
x
p
er
im
en
tatio
n
with
lar
g
er
m
o
d
el
o
n
a
u
g
m
en
ted
d
ata
s
et,
YOL
O
V5
l
with
b
est
weig
h
t
f
r
o
m
ex
p
er
im
en
t WLM
-
O2
r
e
f
er
r
e
d
as WLM
-
A
-
B
W
2
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
6
,
No
.
1
,
Feb
r
u
ar
y
20
2
6
:
2
1
6
-
229
222
5
.
1
.
B
uil
din
g
WL
M
-
O
1
m
o
d
el
wit
h o
rig
ina
l da
t
a
s
et
At
f
ir
s
t
th
e
o
r
ig
in
al
d
ataset
was
u
s
ed
to
tr
ain
th
e
YOL
O
V5
m
m
o
d
el
wh
ich
h
as
3
6
9
lay
er
s
.
T
h
is
m
o
d
el
u
s
ed
th
e
p
r
e
tr
ain
e
d
weig
h
ts
an
d
h
y
p
e
r
p
ar
am
eter
s
v
a
lu
es
o
f
YOL
O
V5
m
o
d
el.
T
h
e
n
th
e
YOL
O
V5
m
m
o
d
el
was
m
ad
e
to
r
u
n
f
o
r
5
0
0
ep
o
ch
s
with
1
6
as
b
atch
s
ize
.
T
h
e
m
o
d
el
tr
ain
in
g
s
to
p
p
ed
at
3
1
1
ep
o
ch
s
s
in
ce
it
h
as
s
h
o
wn
n
o
im
p
r
o
v
em
e
n
t
in
lear
n
in
g
af
ter
th
at
ep
o
ch
.
T
h
eir
b
est,
last
weig
h
ts
ar
e
s
av
ed
f
o
r
d
etec
tio
n
a
n
d
f
u
tu
r
e
u
s
e.
T
h
e
t
r
ain
ed
W
L
M
-
O1
m
o
d
el
is
u
s
ed
f
o
r
test
in
g
an
d
f
o
u
n
d
th
at
th
e
m
o
d
el
d
et
ec
ted
f
o
r
m
o
s
t
test
im
ag
es
an
d
f
ailed
to
d
etec
t
wild
life
s
p
ec
ies
in
th
e
f
ew
o
f
th
e
test
im
ag
es.
T
h
e
o
u
tco
m
e
o
f
th
e
W
L
M
-
O1
m
o
d
el
d
u
r
in
g
test
in
g
will
h
a
v
e
th
e
b
o
u
n
d
in
g
b
o
x
ar
o
u
n
d
th
e
d
etec
ted
a
n
im
als
an
d
co
u
n
ts
th
e
n
u
m
b
er
o
f
s
p
ec
ies
p
r
esen
t
in
th
e
test
im
a
g
e.
T
h
e
lay
er
s
o
f
YOL
O
V5
m
wer
e
n
o
t
s
u
f
f
icien
t
to
d
etec
t
t
h
e
an
im
al
in
s
tan
ce
s
with
p
o
o
r
lig
h
t c
o
n
d
itio
n
s
an
d
d
u
e
to
class
im
b
alan
ce
in
th
e
o
r
ig
in
al
d
ataset.
5
.
2
.
B
uil
din
g
WL
M
-
O2
m
o
d
el
wit
h o
rig
ina
l da
t
a
s
et
T
h
e
o
r
i
g
in
al
d
ataset
was
th
e
n
u
s
ed
to
tr
ain
th
e
YOL
O
V5
l
m
o
d
el
wh
ich
h
as
m
o
r
e
la
y
er
s
wh
en
co
m
p
ar
ed
to
th
e
YOL
O
V5
m
m
o
d
el.
YOL
O
V5
l h
as
4
6
8
la
y
er
s
.
T
h
is
m
o
d
el
also
u
s
ed
th
e
p
r
e
tr
ain
ed
weig
h
ts
an
d
h
y
p
er
p
ar
a
m
eter
s
v
alu
es
p
r
o
v
id
e
d
with
t
h
e
YOL
O
V5
m
o
d
el.
T
h
en
t
h
e
YOL
O
V5
l
m
o
d
el
was
m
ad
e
to
r
u
n
f
o
r
5
0
0
ep
o
ch
s
with
1
6
b
atch
s
ize.
T
h
e
m
o
d
el
tr
ai
n
in
g
s
to
p
p
ed
at
2
9
2
ep
o
ch
s
an
d
n
o
im
p
r
o
v
e
m
en
t
was
o
b
s
er
v
ed
in
lear
n
in
g
a
f
ter
th
a
t
ep
o
ch
.
T
h
eir
b
est,
last
weig
h
ts
ar
e
n
o
ted
f
o
r
d
etec
tio
n
a
n
d
f
u
tu
r
e
u
s
e.
W
h
en
W
L
M
-
O2
m
o
d
el
is
u
s
ed
f
o
r
test
in
g
,
an
aly
s
is
o
n
d
etec
tio
n
s
,
f
o
u
n
d
t
h
at
th
er
e
wer
e
wr
o
n
g
d
e
tectio
n
s
o
f
wild
life
s
p
ec
ies an
d
s
o
m
e
r
em
ain
ed
u
n
d
etec
ted
b
ec
au
s
e
th
e
tr
ain
in
g
was n
o
t su
f
f
icien
t b
ec
au
s
e
o
f
t
h
e
u
n
av
ailab
ilit
y
o
f
en
o
u
g
h
d
ata
ac
r
o
s
s
class
e
s
.
T
h
e
W
L
M
-
O2
m
o
d
el
was
ab
le
to
d
etec
t
s
p
ec
ies
u
n
d
er
p
o
o
r
lig
h
tin
g
co
n
d
itio
n
s
b
u
t sti
ll c
lass
im
b
alan
ce
h
as p
lay
ed
in
d
r
ag
g
i
n
g
th
e
p
er
f
o
r
m
a
n
ce
d
o
wn
.
5
.
3
.
B
uil
din
g
WL
M
-
A
-
RW1
m
o
del f
ro
m
s
cr
a
t
ch
wit
h a
u
g
m
ent
ed
d
a
t
a
s
et
T
h
e
au
g
m
en
te
d
d
ataset
was
u
s
ed
to
tr
ain
th
e
YOL
O
V5
m
m
o
d
el
lab
eled
as
W
L
M
-
A
-
R
W
1
f
r
o
m
s
cr
atch
with
o
u
t
u
s
in
g
an
y
s
p
e
cial
weig
h
t
in
itializatio
n
.
T
h
is
ex
p
er
im
e
n
t
also
u
s
ed
h
y
p
er
p
ar
am
eter
s
p
r
o
v
id
e
d
with
YOL
O
V5
.
T
h
en
th
e
m
o
d
el
was
m
ad
e
to
r
u
n
f
o
r
5
0
0
ep
o
ch
s
with
1
6
as
b
atch
s
ize.
Sin
ce
it
is
tr
ain
in
g
f
r
o
m
s
cr
atch
th
e
ea
r
l
y
s
to
p
p
i
n
g
was
n
o
t
u
s
ed
an
d
th
e
m
o
d
el
was
r
u
n
f
o
r
co
m
p
lete
5
0
0
ep
o
ch
s
.
T
h
e
tr
ain
in
g
r
esu
lts
o
f
W
L
M
-
A
-
R
W
1
h
av
e
s
h
o
wn
g
o
o
d
p
er
f
o
r
m
an
ce
in
ter
m
s
o
f
lea
r
n
in
g
an
d
v
ar
i
an
ce
b
etwe
en
th
e
an
im
als.
T
h
e
test
r
esu
lts
o
f
th
e
ex
p
er
im
e
n
t
s
h
o
w
th
at
it
h
as
d
etec
ted
th
e
s
p
ec
ies
an
d
lab
e
led
th
em
c
o
r
r
ec
tl
y
with
o
u
t
an
y
p
r
o
b
lem
.
T
h
o
u
g
h
th
e
tr
ain
in
g
a
n
d
d
etec
tio
n
wer
e
g
o
o
d
,
th
e
d
r
awb
ac
k
s
w
er
e
th
at
it
was
n
o
t
d
etec
tin
g
a
f
ew
m
u
ltip
le
s
p
e
cies
in
th
e
s
am
e
im
ag
e,
it
j
u
s
t
d
etec
ted
o
n
e
o
r
two
s
p
ec
i
es
an
d
i
g
n
o
r
ed
t
h
e
r
em
ain
in
g
.
An
d
o
b
s
er
v
e
d
th
at
YOL
O
V5
m
lay
er
s
wer
e
n
o
t
en
o
u
g
h
to
d
etec
t
s
p
ec
ies
in
f
ew
im
ag
es
with
p
o
o
r
q
u
ality
an
d
lig
h
tin
g
c
o
n
d
itio
n
s
.
5
.
4
.
B
uil
din
g
WL
M
-
A
-
RW2
m
o
del f
ro
m
s
cr
a
t
ch
wit
h a
u
g
m
ent
ed
da
t
a
s
et
T
h
e
s
am
e
p
r
o
c
ed
u
r
e
as
in
ex
p
er
im
en
t
in
W
L
M
-
A
-
R
W
1
wer
e
u
s
ed
in
YOL
O
V5
l
f
r
o
m
lab
elled
as
W
L
M
-
A
-
R
W
2
.
T
h
e
o
n
ly
p
o
s
i
tiv
e
in
th
e
W
L
M
-
A
-
R
W
2
m
o
d
el
is
th
at
it
d
etec
ted
s
p
ec
ies
ev
en
in
im
ag
es
with
p
o
o
r
q
u
ality
an
d
lig
h
tin
g
co
n
d
itio
n
s
.
L
ik
e
W
L
M
-
A
-
R
W
1
,
th
is
m
o
d
el
lack
s
p
er
f
o
r
m
a
n
ce
b
y
n
o
t
d
etec
tin
g
a
f
ew
m
u
ltip
le
s
p
ec
ies
in
th
e
s
am
e
im
ag
e.
T
h
e
ad
d
itio
n
al
o
b
s
er
v
atio
n
m
ad
e
is
th
at
th
e
W
L
M
-
A
-
R
W
2
m
o
d
el
d
etec
ts
a
f
ew
s
p
ec
ies
wr
o
n
g
l
y
.
Oth
er
th
a
n
th
e
f
ew
d
r
awb
a
c
k
s
th
e
W
L
M
-
A
-
R
W
2
s
h
o
wed
g
o
o
d
p
e
r
f
o
r
m
an
ce
wh
en
co
m
p
ar
ed
to
th
e
p
r
ev
io
u
s
ly
b
u
ilt o
n
es.
5
.
5
.
B
uil
din
g
WL
M
-
A
-
B
W1
m
o
del us
ing
bes
t
weig
ht
f
ro
m
WL
M
-
O
1
t
ha
t
us
ed
o
rig
ina
l da
t
a
s
et
T
h
e
au
g
m
en
ted
d
ataset
was
o
n
ce
ag
ain
u
s
ed
to
tr
ain
t
h
e
Y
OL
O
V5
m
m
o
d
el
r
ef
er
r
e
d
as
W
L
M
-
A
-
B
W
1
.
T
h
is
tim
e
th
e
m
o
d
el
wa
s
g
iv
en
t
h
e
b
est
weig
h
ts
o
f
W
L
M
-
O1
wh
ich
was
tr
ain
ed
o
n
th
e
o
r
ig
i
n
al
d
ataset
with
s
am
e
h
y
p
er
p
ar
am
eter
s
.
Sin
ce
it
u
s
es
weig
h
ts
f
r
o
m
th
e
p
r
ev
io
u
s
m
o
d
el,
we
u
s
ed
ea
r
ly
s
to
p
p
in
g
to
s
to
p
th
e
m
o
d
el
wh
e
n
th
e
r
e
is
n
o
im
p
r
o
v
em
e
n
t
in
lear
n
i
n
g
.
T
h
e
m
o
d
el
s
to
p
p
ed
tr
ain
in
g
at
3
8
8
e
p
o
ch
s
.
T
h
e
t
r
ain
in
g
r
esu
lts
wer
e
en
co
u
r
a
g
in
g
in
te
r
m
s
o
f
lear
n
in
g
.
T
h
e
test
in
g
r
e
s
u
lt
s
h
o
ws
th
e
b
est
p
er
f
o
r
m
an
ce
,
wh
en
c
o
m
p
ar
e
d
with
th
e
p
r
ev
io
u
s
au
g
m
e
n
ted
m
o
d
els.
Mu
ltip
le
s
p
ec
ies
d
etec
tio
n
was
also
f
o
u
n
d
to
b
e
im
p
r
o
v
ed
b
u
t
s
till
p
er
f
o
r
m
ed
p
o
o
r
l
y
o
n
i
m
ag
es
with
p
o
o
r
q
u
ality
,
lig
h
tin
g
co
n
d
itio
n
s
an
d
an
o
m
alies.
5
.
6
.
B
uil
din
g
WL
M
-
A
-
B
W2
m
o
del us
ing
bes
t
weig
ht
f
ro
m
WL
M
-
O
2
t
ha
t
us
ed
o
rig
ina
l da
t
a
s
et
T
h
e
f
in
al
e
x
p
er
im
e
n
t
was
b
u
ild
in
g
W
L
M
-
A
-
B
W
2
m
o
d
el
u
s
in
g
au
g
m
en
ted
d
ataset
u
s
in
g
th
e
b
est
weig
h
ts
f
r
o
m
W
L
M
-
O2
th
at
was
tr
ain
ed
o
n
th
e
o
r
ig
i
n
al
d
ataset.
T
h
e
m
o
d
el
was
th
en
m
ad
e
to
r
u
n
f
o
r
5
0
0
ep
o
c
h
s
with
1
6
as
b
atch
s
ize.
T
h
e
m
o
d
el
s
to
p
p
ed
its
t
r
ain
in
g
at
4
1
6
e
p
o
ch
s
with
n
o
im
p
r
o
v
em
e
n
ts
in
lear
n
in
g
af
ter
t
h
at.
T
h
e
test
r
e
s
u
lts
s
h
o
wed
th
at
W
L
M
-
A
-
B
W
2
h
as
g
iv
en
b
etter
r
esu
lts
f
o
r
im
ag
es
with
p
o
o
r
q
u
ality
,
lig
h
tin
g
c
o
n
d
itio
n
s
an
d
an
o
m
alies.
Mu
ltip
le
s
p
ec
ies
d
etec
tio
n
s
wer
e
also
im
p
r
o
v
e
d
,
an
d
t
h
e
m
is
class
if
icatio
n
was d
r
asti
ca
l
ly
r
ed
u
ce
d
in
W
L
M
-
A
-
B
W
2
m
o
d
el
wh
en
co
m
p
a
r
ed
to
all
t
h
e
p
r
ev
i
o
u
s
m
o
d
els.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
A
I
S
WL
M:
a
r
tifi
cia
l in
tellig
e
n
ce
-
b
a
s
ed
s
ystem
fo
r
w
ild
life
mo
n
ito
r
in
g
(
A
r
u
n
G.
K
.
)
223
6.
RE
SU
L
T
S
6
.
1
.
Q
ua
ntit
a
t
iv
e
a
na
l
y
s
is
W
ild
life
s
p
ec
ies
d
etec
tio
n
an
d
id
en
tific
atio
n
m
o
d
el
d
etec
t
s
th
e
an
im
als
an
d
r
ec
o
g
n
izes
th
em
b
y
b
o
u
n
d
in
g
b
o
x
es
an
d
g
en
e
r
ates
o
b
jectiv
en
ess
s
co
r
e
alo
n
g
with
class
n
am
es.
Qu
an
titativ
e
p
e
r
f
o
r
m
a
n
ce
an
aly
s
is
is
p
er
f
o
r
m
ed
to
e
v
alu
ate
t
h
e
m
ea
s
u
r
ab
le
f
ac
t
o
r
s
o
f
th
e
r
es
u
lts
g
en
er
ated
b
y
th
e
d
etec
tio
n
an
d
id
e
n
tific
atio
n
m
o
d
el
o
n
th
e
test
s
et.
Pre
cisi
o
n
,
R
ec
all,
m
AP q
u
an
titativ
e
m
ea
s
u
r
es a
r
e
u
s
ed
f
o
r
th
e
ev
al
u
a
tio
n
.
T
h
e
th
r
ee
lo
s
s
es
ca
lcu
lated
ar
e
b
o
u
n
d
in
g
b
o
x
lo
s
s
,
o
b
jectn
ess
lo
s
s
an
d
class
if
icatio
n
lo
s
s
f
o
r
b
o
th
tr
ain
in
g
an
d
v
alid
atio
n
.
B
o
u
n
d
in
g
b
o
x
lo
s
s
is
th
e
lo
s
s
co
m
p
u
ted
f
o
r
th
e
lo
ca
lizatio
n
p
h
ase
o
f
an
im
al
d
etec
tio
n
wh
er
e
it
ca
lcu
lates
th
e
m
ea
n
s
q
u
ar
ed
er
r
o
r
b
etwe
en
t
h
e
g
r
o
u
n
d
tr
u
th
a
n
d
th
e
p
r
ed
icted
b
o
x
.
T
h
e
p
r
o
b
ab
ilit
y
o
f
th
e
b
o
u
n
d
i
n
g
b
o
x
h
av
in
g
an
a
n
im
al
is
ca
lcu
lated
b
y
th
e
o
b
jectn
ess
s
co
r
e.
B
in
ar
y
C
r
o
s
s
-
E
n
tr
o
p
y
was
u
s
ed
t
o
co
m
p
u
te
th
e
class
if
icatio
n
lo
s
s
d
u
r
in
g
a
n
im
al
s
p
ec
ies p
r
ed
ictio
n
.
T
h
e
tr
ain
i
n
g
lo
s
s
am
o
n
g
t
h
ese
m
o
d
els s
h
o
ws
a
d
ec
lin
e
f
r
o
m
0
.
0
3
to
less
th
an
0
.
0
2
5
,
s
im
ilar
ly
th
e
o
b
jectiv
en
ess
lo
s
s
h
as a
l
s
o
g
o
t r
ed
u
ce
d
f
r
o
m
0
.
0
1
5
to
less
th
an
0
.
0
1
.
T
h
e
class
if
icatio
n
lo
s
s
h
as
also
d
ec
r
ea
s
ed
to
0
.
0
0
5
f
o
r
W
L
M
-
A
-
B
W
2
.
Fro
m
th
e
r
esu
lts
,
we
o
b
s
er
v
ed
th
at
th
e
p
r
ec
is
io
n
v
alu
es
wer
e
d
eter
io
r
atin
g
with
th
e
s
am
p
le
o
f
th
e
o
r
ig
in
al
d
ataset.
W
h
en
th
e
au
g
m
en
ted
d
ataset
is
u
s
ed
th
e
p
r
ec
is
io
n
v
alu
es
ar
e
c
o
n
s
is
ten
t
an
d
in
cr
ea
s
in
g
f
o
r
m
o
s
t
o
f
th
e
tim
e
an
d
r
ea
ch
in
g
ab
o
v
e
0
.
8
f
o
r
W
L
M
-
A
-
B
W
2
.
Similar
b
eh
av
io
r
w
as
n
o
ticed
f
o
r
r
ec
all
v
alu
es
am
o
n
g
th
e
m
o
d
els.
Fro
m
th
e
a
n
aly
s
is
it
is
u
n
d
e
r
s
to
o
d
th
at
in
cr
ea
s
ed
p
r
ec
is
io
n
an
d
r
ec
all
v
alu
es
lead
to
b
etter
o
b
ject
d
etec
tio
n
r
esu
lts
o
f
W
L
M
-
A
-
B
W
2
.
C
o
n
s
id
er
in
g
th
e
m
AP
v
al
u
es
o
b
tain
ed
f
o
r
th
e
s
am
e
th
r
ee
YOL
O
V5
l
m
o
d
els
f
o
r
two
d
if
f
er
en
t
th
r
esh
o
ld
v
alu
es,0
.
5
an
d
0
.
5
:
0
.
9
5
,
m
AP
v
alu
es
o
b
tain
ed
f
r
o
m
th
e
o
r
i
g
in
al
d
ataset
wer
e
n
o
t
c
o
n
tin
u
o
u
s
ly
in
cr
ea
s
in
g
.
Fro
m
th
is
it
ca
n
b
e
u
n
d
er
s
to
o
d
th
at
th
e
d
etec
tio
n
o
b
tain
ed
f
o
r
th
e
ex
p
er
im
e
n
t
with
th
e
o
r
ig
in
al
d
ataset
was
n
o
t
b
etter
.
B
u
t
wh
en
th
e
m
AP
v
alu
es
o
f
th
e
o
th
er
two
m
o
d
els
n
am
ely
W
L
M
-
A
-
R
W
1
an
d
W
L
M
-
A
-
B
W
2
ar
e
o
b
s
er
v
ed
,
th
ey
ar
e
co
n
tin
u
o
u
s
ly
in
cr
ea
s
in
g
an
d
b
ec
o
m
e
co
n
s
tan
t
af
ter
s
o
m
e
tim
e.
T
h
o
u
g
h
b
o
th
m
o
d
el’
s
m
AP
v
alu
es
i
n
cr
ea
s
e
an
d
b
ec
o
m
e
co
n
s
tan
t,
th
e
m
AP
v
alu
es
o
f
th
e
ex
p
er
im
e
n
t
with
au
g
m
en
te
d
d
ataset
u
s
in
g
b
est
weig
h
ts
,
W
L
M
-
A
-
B
W
2
wer
e
s
lig
h
tly
b
etter
wh
en
co
m
p
a
r
ed
to
th
e
o
th
er
.
W
ith
th
is
it
is
f
o
u
n
d
th
at
ex
p
e
r
im
en
ts
with
au
g
m
en
ted
d
atasets
u
s
in
g
alr
ea
d
y
tr
ain
ed
weig
h
ts
g
iv
e
b
et
ter
d
etec
tio
n
wh
en
co
m
p
ar
ed
to
all
o
th
er
e
x
p
er
im
en
tal
m
o
d
els.
As
ca
n
b
e
s
ee
n
f
r
o
m
th
e
v
alu
es
r
e
p
o
r
te
d
in
T
ab
le
1
,
m
AP
v
alu
es
ar
e
v
er
y
lo
w
f
o
r
W
L
M
-
O1
an
d
W
L
M
-
O2
.
T
h
is
was
al
s
o
o
b
s
er
v
ed
f
r
o
m
th
e
d
etec
tio
n
o
f
t
h
ese
m
o
d
els,
wh
er
e
m
an
y
a
n
im
als
wer
e
lef
t u
n
id
e
n
tifie
d
an
d
m
an
y
wer
e
f
alsely
d
etec
ted
,
an
d
th
ese
m
o
d
els
co
u
ld
n
o
t
d
etec
t
m
a
n
y
ch
allen
g
in
g
im
a
g
es a
s
well.
T
ab
le
1
.
Per
f
o
r
m
an
ce
m
etr
ics o
f
all
AI
SW
L
M
m
o
d
els
M
o
d
e
l
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
mA
P
[
0
.
5
:
0
.
9
5
]
W
LM
-
O1
6
6
.
4
1
5
0
.
5
4
3
2
.
8
5
W
LM
-
O2
7
6
.
1
7
6
1
.
9
1
3
5
.
9
4
W
LM
-
A
-
R
W
1
8
0
.
5
1
7
4
.
5
0
6
2
.
3
6
W
LM
-
A
-
R
W
2
8
0
.
4
4
7
6
.
5
5
6
2
.
6
9
W
LM
-
A
-
B
W
1
7
7
.
4
3
7
7
.
4
7
6
3
.
9
7
W
LM
-
A
-
B
W
2
8
1
.
2
8
7
7
.
8
8
6
4
.
2
7
W
LM
-
O1
6
6
.
4
1
5
0
.
5
4
3
2
.
8
5
Fro
m
th
e
d
etec
tio
n
s
m
ad
e
b
y
W
L
M
-
O1
an
d
W
L
M
-
O2
m
o
d
els,
it wa
s
o
b
s
er
v
ed
th
at
th
e
d
e
tectio
n
h
as
s
ev
er
al
f
alse
p
o
s
itiv
es
wh
er
e
b
u
f
f
alo
es
we
r
e
d
etec
te
d
as
wil
d
b
ea
s
ts
with
p
o
o
r
o
b
jectiv
en
ess
s
co
r
e
an
d
m
an
y
an
im
als
wer
e
n
o
t
d
etec
ted
d
u
e
to
lo
w
m
AP
v
alu
es.
Fro
m
t
h
e
T
ab
le
1
,
t
h
e
m
AP
v
alu
es
f
o
r
th
e
a
u
g
m
en
te
d
m
o
d
els
s
u
ch
as
W
L
M
-
A
-
R
W
1
an
d
W
L
M
-
A
-
R
W
2
ar
e
twice
h
ig
h
er
th
a
n
n
o
n
-
au
g
m
e
n
ted
W
L
M
-
O1
an
d
W
L
M
-
O2
m
o
d
els
wh
er
e
th
eir
Fals
e
Po
s
i
tiv
es
wer
e
co
m
p
ar
ativ
ely
r
ed
u
ce
d
with
W
L
M
-
O1
an
d
W
L
M
-
O2
.
An
d
we
f
o
u
n
d
s
o
m
e
a
n
im
als
ar
e
n
o
t
d
etec
ted
in
im
ag
es
wi
th
m
u
ltip
le
s
p
ec
ies.
T
h
o
u
g
h
t
h
e
m
AP
v
alu
es
ar
e
r
elativ
ely
h
ig
h
b
u
t
n
o
t
s
u
f
f
icie
n
t
to
im
p
r
o
v
e
th
e
d
etec
tio
n
f
o
r
m
u
ltip
le
s
p
ec
ies
in
a
s
in
g
le
ca
m
er
a
tr
ap
im
ag
e.
Fig
u
r
e
2
(
a
)
to
(
c)
s
h
o
ws
th
e
p
er
f
o
r
m
a
n
ce
o
f
W
L
M
-
A
-
B
W
2
u
n
d
er
v
ar
ied
b
ac
k
g
r
o
u
n
d
co
n
d
itio
n
s
:
Fig
u
r
e
2
(
a
)
clea
r
-
s
k
y
illu
m
in
atio
n
,
Fig
u
r
e
2
(
b
)
d
en
s
e
f
o
r
est
,
an
d
Fig
u
r
e
2
(
c
)
s
h
ad
o
w
d
o
m
in
ated
s
ce
n
es
.
Fro
m
th
e
d
etec
tio
n
s
o
f
W
L
M
-
A
-
R
W
1
an
d
W
L
M
-
A
-
R
W
2
m
o
d
els,
it
was
clea
r
ly
s
ee
n
th
at
th
e
d
etec
tio
n
o
f
f
alse
p
o
s
itiv
es
was
r
ed
u
ce
d
with
in
cr
ea
s
e
in
m
AP
v
alu
es
b
u
t
s
till
s
ev
er
al
s
p
ec
i
es
ar
e
n
o
t
d
etec
ted
wh
en
th
er
e
ar
e
m
u
ltip
le
s
p
ec
ies
in
a
s
in
g
le
im
ag
e.
Fo
r
th
e
m
o
d
els,
W
L
M
-
A
-
B
W
1
an
d
W
L
M
-
A
-
B
W
2
th
at
u
s
ed
au
g
m
en
ted
d
ataset
an
d
th
e
f
etch
ed
b
est
weig
h
ts
f
r
o
m
W
L
M
-
O1
an
d
W
L
M
-
O2
m
o
d
els,
it
is
o
b
s
er
v
ed
th
at
th
e
m
AP
v
alu
es
ar
e
r
elativ
ely
h
ig
h
er
th
an
th
e
o
th
er
m
o
d
els.
T
h
e
d
etec
tio
n
f
r
o
m
th
ese
m
o
d
els
wer
e
b
etter
th
an
th
e
p
r
e
v
io
u
s
m
o
d
els.
An
d
we
f
o
u
n
d
th
at
th
ese
m
o
d
els
o
v
er
c
o
m
e
ch
allen
g
es
s
u
c
h
as
m
is
class
if
icatio
n
,
less
o
b
jectn
ess
s
co
r
e
an
d
m
u
lt
ip
le
s
p
ec
ies d
etec
tio
n
th
at
o
cc
u
r
r
ed
i
n
o
th
er
m
o
d
els.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
6
,
No
.
1
,
Feb
r
u
ar
y
20
2
6
:
2
1
6
-
229
224
Fro
m
th
ese
im
ag
es
o
n
e
ca
n
cl
ea
r
ly
s
ee
th
at
th
e
d
etec
tio
n
o
f
m
u
ltip
le
s
p
ec
ies
in
a
s
in
g
le
i
m
ag
e
was
im
p
r
o
v
e
d
,
wh
ich
in
tu
r
n
c
o
n
tr
ib
u
ted
to
th
e
i
n
cr
ea
s
e
in
m
AP
v
alu
es.
T
h
e
co
n
f
u
s
io
n
m
at
r
ices
in
Fig
u
r
e
3
s
h
o
w
th
at
W
L
M
-
A
-
B
W
2
as
s
h
o
wn
in
Fig
u
r
e
3(
a
)
p
er
f
o
r
m
s
s
ig
n
if
ican
tly
b
etter
s
p
ec
ies
r
ec
o
g
n
it
io
n
th
an
W
L
M
-
O1
as
s
h
o
wn
in
Fig
u
r
e
3(
b
).
T
h
e
au
g
m
en
tatio
n
a
n
d
th
e
u
s
e
o
f
th
e
b
est
weig
h
ts
f
o
r
tr
ain
in
g
W
L
M
-
A
-
B
W
2
lead
to
b
etter
p
er
f
o
r
m
an
ce
as sh
o
w
n
in
th
e
d
iag
o
n
al
o
f
th
e
co
n
f
u
s
io
n
m
atr
ix
.
T
o
o
m
a
n
y
s
p
ec
ies ar
e
lef
t u
n
d
etec
ted
b
y
th
e
W
L
M
-
O1
m
o
d
el
d
u
e
to
th
e
u
n
av
ailab
ilit
y
o
f
th
e
s
u
f
f
icien
t le
ar
n
in
g
s
am
p
les ac
r
o
s
s
th
e
class
es.
(
a)
(
b
)
(
c)
Fig
u
r
e
2
.
W
L
M
-
A
-
B
W
2
in
d
if
f
er
en
t b
ac
k
g
r
o
u
n
d
s
(
a
)
a
c
lear
s
k
y
(
b
)
f
o
r
est
,
an
d
(
c)
s
h
a
d
o
w
(
a)
(
b
)
Fig
u
r
e
3
.
C
o
n
f
u
s
io
n
m
atr
i
x
o
f
(
a)
W
L
M
-
O1
m
o
d
el
a
n
d
(
b
)
W
L
M
-
A
-
B
W
2
m
o
d
el
6
.
2
.
Q
ua
lit
a
t
iv
e
a
na
ly
s
is
T
h
e
d
etec
tio
n
r
esu
lts
o
b
tain
ed
f
r
o
m
all
th
e
6
m
o
d
els
wer
e
an
aly
ze
d
in
th
is
s
ec
tio
n
wi
th
r
esp
ec
t
to
ch
allen
g
in
g
s
itu
atio
n
s
n
am
el
y
d
if
f
e
r
en
t
illu
m
in
atio
n
co
n
d
itio
n
s
,
b
ac
k
g
r
o
u
n
d
,
clu
tter
,
s
am
e
s
p
ec
ies
s
in
g
le
in
s
tan
ce
,
d
if
f
er
e
n
t sp
ec
ies d
if
f
er
en
t in
s
tan
ce
s
.
6
.
2
.
1
.
Dif
f
er
ent
illu
m
ina
t
io
n
co
nd
it
io
ns
a
nd
s
im
ila
r
ba
ck
g
ro
un
d
W
ild
life
s
p
ec
ies
wer
e
s
h
o
t
u
n
d
er
d
if
f
er
en
t
lig
h
tin
g
co
n
d
itio
n
s
(
i.e
.
,
d
if
f
e
r
en
t
illu
m
in
atio
n
s
)
in
th
at
m
an
y
s
p
ec
ies
wer
e
p
ict
u
r
ed
u
n
d
er
p
o
o
r
illu
m
in
atio
n
co
n
d
itio
n
s
.
T
o
d
etec
t
im
ag
es
u
n
d
er
th
e
p
o
o
r
illu
m
in
atio
n
co
n
d
itio
n
s
was
o
n
e
o
f
th
e
m
ajo
r
ch
allen
g
es
f
ac
e
d
b
y
th
e
o
b
ject
d
etec
tio
n
m
o
d
els.
W
ith
th
e
p
r
esen
ce
o
f
C
SP
Dar
k
n
et
with
4
6
8
lay
er
s
in
its
b
ac
k
b
o
n
e
m
ad
e
th
e
SW
L
M
m
o
d
els
p
o
s
s
ib
le
to
o
v
er
co
m
e
t
h
e
p
o
o
r
illu
m
in
atio
n
ch
allen
g
e
.
T
h
e
an
im
als d
etec
ted
in
Fig
u
r
e
4
ar
e
th
e
r
esu
lts
o
f
th
e
b
est p
er
f
o
r
m
in
g
W
L
M
-
A
-
B
W
2
m
o
d
el
th
at
h
as
b
ee
n
tr
ain
e
d
o
n
th
e
a
u
g
m
e
n
ted
d
ata
an
d
u
s
ed
th
e
b
est
weig
h
ts
f
r
o
m
W
L
M
-
OM
1
m
o
d
el.
I
t
ca
n
b
e
o
b
s
er
v
e
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
A
I
S
WL
M:
a
r
tifi
cia
l in
tellig
e
n
ce
-
b
a
s
ed
s
ystem
fo
r
w
ild
life
mo
n
ito
r
in
g
(
A
r
u
n
G.
K
.
)
225
th
at
th
e
test
im
ag
es
ar
e
tak
en
at
n
ig
h
t
with
d
if
f
er
en
t
illu
m
in
atio
n
s
an
d
im
ag
es
th
at
ar
e
d
if
f
icu
lt
to
d
if
f
er
e
n
tiate
f
r
o
m
th
e
b
ac
k
g
r
o
u
n
d
.
W
L
M
-
A
-
B
W
2
ca
n
d
etec
t a
n
im
als wi
th
ch
allen
g
in
g
b
ac
k
g
r
o
u
n
d
s
s
u
ch
as c
lo
u
d
y
,
s
u
n
s
et
an
d
s
u
n
r
is
e.
T
h
ese
d
etec
tio
n
s
ar
e
attr
ib
u
ted
to
th
e
wo
r
k
in
g
o
f
C
SP
Dar
k
n
et
th
at
clea
r
ly
d
if
f
er
en
tiates
b
ac
k
g
r
o
u
n
d
an
d
f
o
r
eg
r
o
u
n
d
in
f
o
r
m
atio
n
d
u
r
in
g
d
etec
tio
n
.
T
h
is
ca
p
ab
ilit
y
o
f
th
e
SW
L
M
m
o
d
el
will
allo
w
th
e
AI
SW
L
M
to
d
etec
t th
e
an
im
a
l e
v
en
o
u
ts
id
e
th
e
f
o
r
est o
r
co
u
n
tr
y
s
id
e
ir
r
esp
ec
tiv
e
o
f
th
e
b
ac
k
g
r
o
u
n
d
.
Fig
u
r
e
4
.
W
L
M
-
A
-
B
W
2
p
er
f
o
r
m
an
ce
in
p
o
o
r
illu
m
in
atio
n
c
o
n
d
itio
n
s
6
.
2
.
2
.
Clutt
er
C
lu
tter
ed
im
ag
es
h
a
v
e
th
e
f
o
cu
s
o
n
th
e
d
if
f
er
e
n
t
o
b
jects
th
an
o
n
t
h
e
d
esire
d
o
b
jects.
So
,
th
e
wild
s
p
ec
ies
ca
p
tu
r
ed
in
clu
s
ter
ed
im
ag
es
ar
e
eith
er
b
lu
r
r
e
d
o
r
n
o
t
s
ee
n
b
r
ig
h
tly
.
Dete
ctin
g
s
p
ec
ies
in
th
e
clu
tter
ed
im
ag
es
is
th
e
n
ex
t
ch
allen
g
e
o
f
AI
-
SW
L
M.
T
h
e
p
r
esen
ce
o
f
PANet
as
its
n
ec
k
in
th
e
ar
ch
itectu
r
e
o
f
A
I
SW
L
M
p
lay
s
a
m
ajo
r
r
o
le
in
d
etec
tin
g
s
p
ec
ies
in
clu
tter
e
d
i
m
ag
es
an
d
m
ak
es
it
p
o
s
s
ib
le
t
o
r
ec
o
g
n
ize
ea
ch
o
f
th
e
an
im
als
in
th
e
clu
tter
.
T
h
e
b
i
-
d
ir
ec
tio
n
al
f
ea
tu
r
e
f
u
s
io
n
tech
n
iq
u
e
h
elp
s
th
e
n
etwo
r
k
tr
ain
o
n
d
if
f
er
en
t
in
p
u
t
f
ea
t
u
r
es.
Dete
ctio
n
o
f
m
u
ltip
le
wild
an
im
al
s
p
ec
ies
in
clu
tter
ed
im
ag
es
b
y
W
L
M
-
A
-
B
W
2
m
o
d
el
is
s
h
o
wn
in
Fig
u
r
e
5
.
Fig
u
r
e
5
.
Dete
ctio
n
o
f
clu
tter
e
d
im
ag
es b
y
W
L
M
-
A
-
BW2
6
.
2
.
3
.
S
i
n
g
l
e
s
p
e
c
i
e
s
s
i
n
g
l
e
in
s
t
a
n
c
e
,
s
i
n
g
l
e
s
p
e
c
i
e
s
m
u
lt
i
p
l
e
i
n
s
t
a
n
c
e
s
a
n
d
d
i
f
f
e
r
e
nt
s
p
e
c
i
e
s
m
u
l
t
i
p
l
e
i
n
s
t
a
n
c
e
s
T
h
e
o
th
er
c
h
allen
g
e
o
f
AI
-
SW
L
M
s
y
s
tem
is
to
d
etec
t th
e
s
in
g
le
s
p
ec
ies in
a
s
in
g
le
in
s
tan
ce
.
C
SP
an
d
YOL
O
d
etec
tio
n
h
ea
d
p
lay
th
e
m
ain
r
o
le
in
th
e
s
am
e
s
p
ec
ies
in
a
s
in
g
le
in
s
tan
ce
an
d
ca
n
b
e
s
ee
n
in
Fig
u
r
e
6
.
B
asic
r
ec
o
g
n
itio
n
o
f
an
im
als
is
p
er
f
o
r
m
ed
well
b
y
W
L
M
-
A
-
B
W
2
m
o
d
el
with
g
o
o
d
o
b
jectiv
en
ess
s
co
r
e
g
r
ea
ter
th
an
9
5
%
an
d
b
etter
m
AP
wh
en
co
m
p
ar
ed
with
o
th
er
m
o
d
els
u
n
d
er
co
n
s
id
er
atio
n
th
r
o
u
g
h
th
eir
q
u
an
titativ
e
m
ea
s
u
r
es
s
u
ch
as
p
r
ec
is
io
n
an
d
r
ec
all.
T
h
e
lear
n
in
g
ab
ilit
y
o
f
th
e
b
est
p
er
f
o
r
m
in
g
m
o
d
el
is
ac
h
iev
ed
d
u
e
t
o
d
ata
au
g
m
e
n
tatio
n
an
d
h
a
v
in
g
m
u
ltip
les
lay
er
s
o
f
C
SP
Dar
k
n
et
as
its
b
ac
k
b
o
n
e.
Par
tial
tr
an
s
itio
n
lay
er
in
C
SP
Dar
k
n
et,
with
its
f
ea
tu
r
e
f
u
s
io
n
s
tr
ateg
y
in
a
h
ier
ar
ch
ical
f
ash
io
n
co
n
tr
ib
u
ted
to
class
if
icatio
n
o
f
m
u
ltip
le
an
i
m
al
s
p
ec
ies
in
t
h
e
ca
m
er
a
tr
ap
im
ag
es.
T
h
e
p
er
f
o
r
m
a
n
ce
o
f
W
L
M
-
A
-
B
W
2
m
o
d
el
in
d
etec
tin
g
th
e
s
in
g
l
e
an
im
al
s
p
ec
ies
wi
th
m
u
ltip
le
in
s
tan
ce
s
i
s
s
h
o
wn
in
Fig
u
r
e
7
an
d
d
if
f
er
en
t
an
im
al
s
p
ec
ies with
m
u
ltip
le
in
s
tan
ce
s
in
Fig
u
r
e
8
.
Evaluation Warning : The document was created with Spire.PDF for Python.