I
AE
S In
t
er
na
t
io
na
l J
o
urna
l o
f
Art
if
icia
l In
t
ellig
ence
(
I
J
-
AI)
Vo
l.
9
,
No
.
1
,
Ma
r
ch
2
0
2
0
,
p
p
.
40
~
45
I
SS
N:
2
2
5
2
-
8938
,
DOI
: 1
0
.
1
1
5
9
1
/i
j
ai.
v
9
.i
1
.
p
p
40
-
45
40
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
i
a
esco
r
e.
co
m
La
rg
e
-
sca
le i
m
a
g
e
-
to
-
v
ideo
face re
t
riev
a
l
w
ith
co
nv
o
lutiona
l
neura
l net
w
o
rk
f
e
a
tures
I
m
a
ne
H
a
chcha
ne
1
,
Ab
del
ma
j
id B
a
dri
2
,
Acha
Sa
hel
3
,
Ya
s
s
ine R
u
iche
k
4
1,
2,
3
L
a
b
o
ra
to
ry
o
f
El
e
c
tro
n
ics
,
E
n
e
rg
y
,
A
u
to
m
a
ti
o
n
&
In
f
o
rm
a
ti
o
n
P
r
o
c
e
ss
in
g
,
F
a
c
u
lt
y
o
f
S
c
ien
c
e
s a
n
d
T
e
c
h
n
iq
u
e
s
M
o
h
a
m
m
e
d
ia,
Un
iv
e
rsit
y
Ha
ss
a
n
II
Ca
sa
b
lan
c
a
,
M
o
h
a
m
m
e
d
ia,
M
o
ro
c
c
o
4
IRT
ES
-
L
a
b
o
ra
to
r
y
S
ET
,
Un
iv
e
rs
it
y
o
f
T
e
c
h
n
o
l
o
g
y
Be
l
f
o
rt
M
o
n
t
b
é
li
a
rd
,
F
ra
n
c
e
Art
icle
I
nfo
AB
ST
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Au
g
2
3
,
2
0
1
9
R
ev
i
s
ed
Oct
1
0
,
2
0
1
9
A
cc
ep
ted
No
v
3
,
201
9
Co
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
fe
a
tu
re
s
a
re
b
e
c
o
m
in
g
th
e
n
o
r
m
i
n
in
sta
n
c
e
re
tri
e
v
a
l.
T
h
is
w
o
rk
in
v
e
sti
g
a
tes
t
h
e
re
lev
a
n
c
e
o
f
u
sin
g
a
n
o
f
th
e
sh
e
lf
o
b
jec
t
d
e
tec
ti
o
n
n
e
tw
o
rk
,
li
k
e
F
a
ste
r
R
-
CNN
,
a
s
a
f
e
a
tu
re
e
x
tr
a
c
to
r
f
o
r
a
n
im
a
g
e
-
to
-
v
id
e
o
f
a
c
e
re
tri
e
v
a
l
p
ip
e
li
n
e
in
st
e
a
d
o
f
u
sin
g
h
a
n
d
-
c
ra
f
ted
f
e
a
tu
re
s.
We
u
se
th
e
o
b
jec
ts
p
ro
p
o
sa
ls
lea
rn
e
d
b
y
a
Re
g
io
n
P
ro
p
o
sa
l
Ne
tw
o
rk
(RP
N)
a
n
d
t
h
e
ir
a
ss
o
c
iate
d
re
p
re
se
n
tatio
n
s
tak
e
n
f
ro
m
a
CNN
f
o
r
th
e
f
il
terin
g
a
n
d
th
e
re
-
ra
n
k
in
g
ste
p
s.
M
o
re
o
v
e
r,
w
e
stu
d
y
th
e
re
le
v
a
n
c
e
o
f
f
e
a
tu
re
s
f
ro
m
a
f
in
e
tu
n
e
d
n
e
tw
o
rk
.
In
a
d
d
it
i
o
n
to
th
a
t
w
e
e
x
p
lo
re
t
h
e
u
se
o
f
f
a
c
e
d
e
te
c
ti
o
n
,
f
ish
e
r
v
e
c
to
r
a
n
d
b
a
g
o
f
v
isu
a
l
w
o
rd
s
w
it
h
th
o
se
sa
m
e
CNN
f
e
a
tu
re
s.
We
a
ls
o
tes
t
th
e
im
p
a
c
t
o
f
d
i
ff
e
re
n
t
si
m
il
a
rit
y
m
e
tri
c
s.
T
h
e
re
su
lt
s
o
b
tain
e
d
a
re
v
e
r
y
p
ro
m
isin
g
.
K
ey
w
o
r
d
s
:
B
OVW
C
las
s
i
f
icatio
n
C
NN,
Fas
ter
R
-
C
NN
Face
r
etr
iev
al
FV
I
m
ag
e
p
r
o
ce
s
s
i
n
g
I
m
ag
e
-
to
-
v
id
eo
in
s
ta
n
ce
r
etr
iev
al
Ob
j
ec
t r
ec
o
g
n
itio
n
Vid
eo
r
etr
iev
al
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
:
I
m
an
e
Hac
h
c
h
an
e,
L
ab
o
r
ato
r
y
o
f
E
lec
tr
o
n
ics,
E
n
er
g
y
,
Au
to
m
at
io
n
&
I
n
f
o
r
m
ati
o
n
P
r
o
ce
s
s
in
g
,
Facu
lt
y
o
f
Sc
ien
ce
s
an
d
T
ec
h
n
iq
u
es M
o
h
a
m
m
ed
ia,
Un
i
v
er
s
it
y
Has
s
an
I
I
C
a
s
ab
lan
ca
,
Mo
h
a
m
m
ed
ia,
Mo
r
o
cc
o
.
E
m
ail:
h
ac
h
c
h
a
n
ei
m
an
e
@
g
m
a
il.c
o
m
1.
I
NT
RO
D
UCT
I
O
N
I
n
th
i
s
w
o
r
k
w
e
ad
d
r
ess
th
e
ta
s
k
o
f
i
m
ag
e
to
v
id
eo
f
ac
e
r
etr
i
ev
al.
W
ith
b
il
lio
n
s
o
f
i
m
a
g
es
an
d
v
id
eo
s
cr
ea
ted
ea
ch
d
a
y
,
it
i
s
e
s
s
e
n
ti
al
to
b
u
ild
to
o
ls
f
o
r
ac
ce
s
s
i
n
g
an
d
r
etr
ie
v
in
g
m
u
lti
m
ed
ia
co
n
ten
t
e
f
f
icien
tl
y
.
I
n
th
e
co
n
te
x
t
o
f
r
etr
ie
v
al,
i
m
ag
e
-
to
-
v
id
eo
f
ac
e
r
etr
iev
a
l
is
t
h
e
task
o
f
id
en
ti
f
y
in
g
a
s
p
ec
i
f
ic
f
r
a
m
e
o
r
s
ce
n
e
i
n
a
v
id
eo
o
r
a
co
llectio
n
o
f
v
id
eo
s
f
r
o
m
a
s
p
ec
if
ic
f
ac
e
i
n
s
ta
n
ce
i
n
a
s
tatic
i
m
a
g
e.
On
o
n
e
h
an
d
,
i
m
ag
e
-
to
-
v
id
e
o
r
etr
iev
al
is
an
as
y
m
m
etr
ic
p
r
o
b
lem
.
I
m
a
g
es
o
n
l
y
co
n
t
ain
s
tat
ic
in
f
o
r
m
atio
n
b
u
t
v
id
eo
s
h
a
v
e
m
u
c
h
r
ich
er
v
is
u
al
in
f
o
r
m
a
tio
n
,
lik
e
o
p
tical
f
lo
w
.
Du
e
to
t
h
e
lack
o
f
te
m
p
o
r
al
in
f
o
r
m
atio
n
,
s
ta
n
d
ar
d
tech
n
iq
u
es
u
s
ed
f
o
r
e
x
tr
ac
ti
n
g
v
id
eo
d
escr
ip
to
r
s
[1
-
4]
ca
n
n
o
t
b
e
d
ir
ec
tl
y
u
s
ed
o
n
s
tatic
i
m
a
g
es.
B
u
t,
s
tan
d
ar
d
f
ea
t
u
r
es
f
o
r
i
m
a
g
e
r
etr
iev
a
l
[5
-
8]
ca
n
b
e
ap
p
lied
to
v
id
eo
d
ata
b
y
p
r
o
ce
s
s
in
g
ea
c
h
f
r
a
m
e
as
an
in
d
ep
en
d
en
t
i
m
a
g
e.
T
em
p
o
r
al
in
f
o
r
m
atio
n
is
u
s
u
all
y
c
o
m
p
r
es
s
ed
eith
er
b
y
r
ed
u
ci
n
g
th
e
n
u
m
b
er
o
f
lo
ca
l
f
ea
t
u
r
es
o
r
b
y
en
co
d
in
g
m
u
lt
i
p
le
f
r
a
m
es
i
n
to
a
s
in
g
le
g
lo
b
a
l
r
ep
r
esen
tatio
n
.
On
t
h
e
o
th
er
h
an
d
,
f
ac
e
r
etr
iev
al
r
e
m
ain
s
a
c
h
alle
n
g
i
n
g
tas
k
b
ec
au
s
e
co
n
v
e
n
tio
n
al
i
m
a
g
e
r
e
tr
iev
al
ap
p
r
o
ac
h
es,
s
u
c
h
a
s
b
ag
o
f
v
is
u
al
w
o
r
d
s
(
B
OVW
)
,
ar
e
d
if
f
ic
u
lt to
ad
ap
t to
th
e
f
ac
e
d
o
m
ai
n
[
9
]
.
T
r
a
d
itio
n
all
y
,
i
m
ag
e
-
to
-
v
id
eo
r
etr
iev
al
o
r
f
ac
e
r
etr
iev
al
m
et
h
o
d
s
[
1
0
-
12]
ar
e
b
ased
o
n
h
an
d
-
cr
a
f
ted
f
ea
t
u
r
es
(
SIFT
[
1
3
]
,
B
R
I
E
F
[
1
4
]
,
etc.
)
an
d
n
o
t
m
u
c
h
e
f
f
o
r
t
h
as
b
ee
n
p
u
t
s
o
f
ar
in
to
t
h
e
ad
ap
tatio
n
o
f
d
ee
p
lear
n
in
g
tec
h
n
iq
u
es,
s
u
c
h
as
c
o
n
v
o
l
u
tio
n
al
n
eu
r
al
n
et
w
o
r
k
s
(
C
NN)
.
C
NN
s
tr
ain
ed
w
it
h
lar
g
e
a
m
o
u
n
t
s
o
f
d
ata
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
A
r
ti
f
I
n
tell
I
SS
N:
2252
-
8938
La
r
g
e
-
s
ca
le
ima
g
e
-
to
-
vid
e
o
fa
ce
r
etri
ev
a
l wi
th
co
n
vo
lu
tio
n
a
l n
eu
r
a
l n
etw
o
r
k
fea
tu
r
es
(
I
ma
n
e
Ha
ch
ch
a
n
e
)
41
ca
n
lear
n
f
ea
tu
r
es
g
e
n
er
ic
e
n
o
u
g
h
to
b
e
u
s
ed
to
s
o
lv
e
tas
k
s
f
o
r
w
h
ic
h
th
e
n
et
w
o
r
k
h
a
s
n
o
t
b
ee
n
tr
ai
n
ed
[
1
5
]
.
Fo
r
im
a
g
e
r
etr
iev
al,
i
n
p
ar
ticu
lar
,
m
a
n
y
w
o
r
k
s
in
th
e
li
ter
atu
r
e
[7
,
1
6
]
h
av
e
ad
o
p
ted
s
o
lu
tio
n
s
b
ased
o
n
s
tan
d
ar
d
f
ea
tu
r
e
s
e
x
tr
ac
ted
f
r
o
m
a
p
r
etr
ai
n
ed
C
N
N
f
o
r
i
m
ag
e
cla
s
s
i
f
icatio
n
[
1
7
]
,
ac
h
i
ev
i
n
g
en
co
u
r
ag
i
n
g
p
er
f
o
r
m
a
n
ce
s
.
Ma
n
y
C
NN
-
b
a
s
ed
o
b
j
ec
t
d
etec
tio
n
p
ip
elin
es
h
av
e
b
ee
n
p
r
o
p
o
s
ed
,
b
u
t
w
e
ar
e
m
o
r
e
in
ter
es
ted
in
t
h
e
late
s
t
o
n
e
s
.
Fa
s
ter
R
-
C
NN
[
1
8
]
u
s
es
a
R
e
g
io
n
P
r
o
p
o
s
al
Net
w
o
r
k
(
R
P
N)
th
at
r
e
m
o
v
es
t
h
e
d
ep
en
d
e
n
ce
o
f
o
b
j
ec
t
p
r
o
p
o
s
als
f
r
o
m
o
ld
er
C
NN
o
b
j
ec
t
d
etec
tio
n
s
y
s
te
m
s
.
I
n
Fa
s
ter
R
-
C
N
N,
R
P
N
s
h
ar
es
f
ea
t
u
r
es
w
i
th
th
e
o
b
j
ec
t
-
d
etec
tio
n
n
et
w
o
r
k
in
[
1
9
]
to
s
i
m
u
lta
n
eo
u
s
l
y
lear
n
p
r
o
m
i
n
en
t
o
b
j
ec
t
p
r
o
p
o
s
itio
n
s
an
d
t
h
ei
r
ass
o
ciate
d
class
p
r
o
b
ab
ilit
ies.
A
lt
h
o
u
g
h
th
e
Fa
s
ter
R
-
C
N
N
is
d
esig
n
ed
f
o
r
g
en
er
ic
o
b
j
ec
t
d
etec
tio
n
[
2
0
]
.
De
m
o
n
s
tr
ated
t
h
at
it
ca
n
ac
h
iev
e
i
m
p
r
ess
i
v
e
f
ac
e
d
etec
tio
n
p
er
f
o
r
m
a
n
ce
e
s
p
ec
iall
y
w
h
en
r
etr
ain
ed
o
n
a
s
u
itab
le
f
ac
e
d
etec
tio
n
tr
ai
n
in
g
s
et
[
2
1
]
.
I
n
th
i
s
p
ap
er
w
e
tr
y
to
f
ill
t
h
is
g
ap
b
y
ex
p
lo
r
in
g
t
h
e
r
elev
a
n
c
e
o
f
o
n
-
t
h
e
-
s
h
el
f
an
d
f
i
n
e
-
t
u
n
e
d
f
ea
tu
r
es
o
f
an
o
b
j
ec
t d
etec
tio
n
C
NN
f
o
r
i
m
ag
e
-
to
-
v
id
eo
f
ac
e
r
etr
iev
al
.
W
e
ex
p
lo
it th
e
f
ea
t
u
r
es o
f
a
s
tate
-
of
-
t
h
e
-
ar
t p
r
e
-
tr
ain
ed
o
b
j
ec
t
d
etec
tio
n
C
NN
ca
lled
Fas
ter
R
-
C
NN.
W
e
u
s
e
h
i
s
e
n
d
-
to
-
en
d
o
b
j
ec
t
d
etec
tio
n
ar
ch
i
tectu
r
e
to
ex
tr
ac
t
g
lo
b
al
an
d
lo
ca
l
co
n
v
o
lu
tio
n
a
l
f
ea
tu
r
es
in
a
s
i
n
g
le
f
o
r
w
ar
d
p
ass
a
n
d
test
t
h
eir
r
ele
v
an
ce
f
o
r
i
m
a
g
e
-
to
-
v
id
eo
f
ac
e
r
etr
iev
al.
W
e
also
ex
p
lo
r
e
th
e
u
s
e
o
f
f
ac
e
d
etec
tio
n
,
Fis
h
er
Vec
to
r
(
FV)
[
4
]
an
d
B
OVW
w
o
r
d
s
w
it
h
th
o
s
e
s
a
m
e
C
NN
f
ea
t
u
r
e
s
.
T
h
e
r
est
o
f
t
h
is
p
ap
er
is
o
r
g
an
ized
as
f
o
llo
w
s
:
Sect
io
n
2
p
r
esen
ts
o
u
r
r
esear
c
h
m
et
h
o
d
,
in
clu
d
in
g
o
u
r
f
ea
t
u
r
e
s
ex
tr
ac
tio
n
m
e
th
o
d
an
d
th
e
r
ak
in
g
an
d
r
er
an
k
in
g
s
tr
ate
g
ies
.
Sectio
n
3
p
r
esen
ts
o
u
r
r
esu
lt
s
an
d
d
is
cu
s
s
io
n
s
.
Fi
n
all
y
,
w
e
p
r
ese
n
t o
u
r
co
n
cl
u
s
i
o
n
s
in
Sect
io
n
4
.
2.
M
E
T
H
O
DO
L
O
G
Y
2
.
1
.
Da
t
a
s
et
s
ex
plo
it
ed
W
e
ev
alu
ate
o
u
r
m
eth
o
d
o
lo
g
i
es u
s
in
g
t
h
e
f
o
llo
w
i
n
g
d
atase
ts
:
Yo
u
T
u
b
e
C
eleb
r
ities
Face
T
r
ac
k
in
g
a
n
d
R
ec
o
g
n
it
io
n
Data
(
Y
-
C
eleb
)
[
2
2
]
:
T
h
e
d
ataset
co
n
tain
s
1
9
1
0
s
eq
u
en
ce
s
o
f
4
7
s
u
b
j
ec
ts
.
All
v
id
eo
s
ar
e
en
co
d
e
d
in
MP
E
G4
at
2
5
f
p
s
r
ate.
Yo
u
T
u
b
e
Face
s
Data
b
ase
[
2
3
]
:
T
h
e
d
ata
s
et
co
n
tain
s
3
,
4
2
5
v
id
eo
s
o
f
1
,
5
9
5
d
if
f
er
en
t
p
eo
p
le.
A
ll
th
e
v
id
eo
s
w
er
e
d
o
w
n
lo
ad
ed
f
r
o
m
Yo
u
T
u
b
e.
A
n
a
v
er
ag
e
o
f
2
.
1
5
v
id
eo
s
ar
e
a
v
ailab
le
f
o
r
ea
c
h
s
u
b
j
ec
t.
T
h
e
s
h
o
r
test
clip
d
u
r
atio
n
is
4
8
f
r
a
m
es,
th
e
lo
n
g
e
s
t
clip
is
6
,
0
7
0
f
r
a
m
es,
an
d
th
e
av
er
a
g
e
len
g
th
o
f
a
v
id
eo
clip
is
1
8
1
.
3
f
r
am
es.
T
h
e
d
atasets
u
s
ed
to
f
i
n
et
u
n
e
t
h
e
n
et
w
o
r
k
:
FERET
[
2
4
]
:
3
5
2
8
im
a
g
es,
in
clu
d
in
g
5
5
Qu
er
y
i
m
a
g
es.
A
f
r
a
m
i
n
g
b
o
x
s
u
r
r
o
u
n
d
in
g
th
e
tar
g
et
f
ac
e
is
p
r
o
v
id
ed
f
o
r
q
u
er
y
i
m
ag
e
s
.
F
A
C
E
S9
4
[
2
5
]
:
2
8
0
9
im
a
g
e
s
2
8
0
9
im
ag
e
s
,
in
c
lu
d
i
n
g
5
5
Q
u
er
y
i
m
a
g
e
s
.
A
f
r
a
m
i
n
g
b
o
x
s
u
r
r
o
u
n
d
in
g
t
h
e
tar
g
et
f
ac
e
i
s
p
r
o
v
id
ed
f
o
r
q
u
er
y
i
m
a
g
es.
Face
Scr
u
b
[
2
6
]
: 5
5
1
2
7
im
ag
e
s
2
.
2
.
Video
re
t
rie
v
a
l st
ra
t
e
g
y
:
T
h
is
s
ec
tio
n
d
escr
ib
es t
h
e
th
r
e
e
m
aj
o
r
s
tep
s
in
o
u
r
p
ip
elin
e,
w
e
u
s
ed
:
1.
Fil
ter
in
g
s
tep
.
W
e
cr
ea
te
im
a
g
e
d
escr
ip
to
r
s
f
o
r
q
u
er
y
a
n
d
d
atab
ase
f
r
am
e
s
u
s
in
g
C
N
N
f
ea
tu
r
e
s
.
A
t
test
i
n
g
ti
m
e,
th
e
d
escr
ip
to
r
o
f
th
e
q
u
er
y
is
co
m
p
ar
ed
to
all
it
e
m
s
in
t
h
e
d
atab
ase,
w
h
ic
h
ar
e
th
e
n
r
an
k
ed
ac
co
r
d
in
g
to
a
s
i
m
i
lar
it
y
m
ea
s
u
r
e.
A
t t
h
i
s
s
ta
g
e,
th
e
e
n
tire
f
r
a
m
e
is
co
n
s
id
er
ed
a
s
a
q
u
er
y
.
2.
Sp
atial
r
e
-
r
an
k
i
n
g
.
Af
ter
t
h
e
f
i
lter
in
g
s
tep
,
th
e
N
u
p
p
er
ele
m
en
ts
ar
e
an
al
y
ze
d
lo
ca
ll
y
a
n
d
r
e
-
r
an
k
ed
.
3.
Qu
er
y
e
x
p
an
s
io
n
(
QE
)
.
W
e
av
er
ag
e
t
h
e
f
r
a
m
e
d
escr
ip
to
r
s
o
f
th
e
M
h
i
g
h
er
ele
m
e
n
ts
o
f
t
h
e
f
ir
s
t
r
an
k
i
n
g
w
it
h
q
u
er
y
d
escr
ip
to
r
to
ca
r
r
y
o
u
t
a
n
e
w
s
ea
r
c
h
.
2
.
3
.
CNN
-
ba
s
ed
r
epre
s
ent
a
t
io
ns
W
e
ex
p
lo
r
e
th
e
r
elev
a
n
ce
o
f
u
s
i
n
g
C
NN
f
ea
t
u
r
es
f
o
r
f
ac
e
i
m
ag
e
to
v
id
eo
f
ac
e
r
etr
ie
v
al.
T
h
e
q
u
er
y
in
s
ta
n
ce
is
d
ef
i
n
ed
b
y
a
b
o
u
n
d
in
g
b
o
x
ab
o
v
e
th
e
q
u
er
y
i
m
a
g
e.
W
e
u
s
e
th
e
f
ea
tu
r
es
e
x
tr
a
cted
f
r
o
m
Fa
s
ter
R
-
C
NN
p
r
e
-
tr
ai
n
ed
m
o
d
els
[
1
8
]
as
o
u
r
g
lo
b
al
a
n
d
lo
ca
l
f
ea
t
u
r
es.
Fas
ter
R
-
C
NN
h
a
s
a
r
eg
io
n
p
r
o
p
o
s
al
n
et
w
o
r
k
th
at
g
i
v
es
t
h
e
lo
ca
tio
n
s
i
n
t
h
e
i
m
a
g
e
w
h
ich
h
a
v
e
h
ig
h
er
p
r
o
b
ab
ilit
ies
o
f
h
a
v
i
n
g
a
n
o
b
j
ec
t,
an
d
a
class
if
ier
th
at
lab
els
ea
ch
o
f
th
o
s
e
o
b
j
ec
t
p
r
o
p
o
s
als
as
o
n
e
o
f
th
e
clas
s
es
in
th
e
lear
n
in
g
d
ataset
[
2
7
]
.
W
e
ex
tr
ac
t
co
m
p
ac
t
f
ea
t
u
r
es
f
r
o
m
t
h
e
ac
tiv
a
tio
n
s
o
f
a
co
n
v
o
lu
tio
n
al
la
y
er
i
n
a
C
NN
[
2
7
-
28]
.
Fas
ter
R
-
C
N
N
is
f
as
ter
o
n
a
g
lo
b
al
an
d
lo
ca
l
s
ca
le.
W
e
b
u
ild
a
g
l
o
b
al
f
r
a
m
e
d
escr
ip
to
r
b
y
ig
n
o
r
in
g
all
t
h
e
la
y
er
s
t
h
at
w
o
r
k
w
i
th
o
b
j
ec
t
p
r
o
p
o
s
als
an
d
ex
tr
ac
t
f
ea
t
u
r
es
f
r
o
m
th
e
l
ast
co
n
v
o
l
u
tio
n
a
l
la
y
er
.
C
o
n
s
i
d
er
in
g
th
e
e
x
tr
ac
ted
ac
tiv
atio
n
s
o
f
a
co
n
v
o
lu
tio
n
la
y
er
f
o
r
a
f
r
a
m
e,
w
e
g
r
o
u
p
th
e
ac
tiv
atio
n
s
o
f
ea
ch
f
ilter
to
cr
ea
te
a
f
r
am
e
d
escr
ip
to
r
w
it
h
t
h
e
s
a
m
e
d
i
m
en
s
io
n
as
th
e
n
u
m
b
er
o
f
f
i
lter
s
in
th
e
co
n
v
o
l
u
tio
n
la
y
er
,
to
d
o
s
o
b
o
th
m
ax
a
n
d
s
u
m
p
o
o
l
-
i
n
g
s
tr
ate
g
ie
s
ar
e
co
n
s
id
er
ed
an
d
co
m
p
ar
ed
in
s
ec
tio
n
3
.
W
e
ag
g
r
eg
a
te
th
e
ac
tiv
atio
n
s
o
f
ea
c
h
w
i
n
d
o
w
s
u
g
g
esti
o
n
i
n
th
e
R
o
I
P
o
o
lin
g
la
y
er
to
cr
ea
te
r
eg
io
n
a
l d
escr
ip
tio
n
s
[
2
1
]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8938
I
n
t J
A
r
ti
f
I
n
tell
,
Vo
l.
9
,
No
.
1
,
Ma
r
ch
20
20
:
40
–
45
42
W
e
u
s
e
t
h
e
V
GG1
6
ar
ch
ite
ctu
r
e
o
f
Fas
ter
R
-
C
N
N
to
ex
tr
ac
t
t
h
e
g
lo
b
al
an
d
lo
ca
l
f
ea
t
u
r
es.
W
e
ch
o
o
s
e
t
h
at
ar
c
h
itect
u
r
e
b
ec
au
s
e
it
p
er
f
o
r
m
s
b
etter
.
I
t
h
as
b
ee
n
s
h
o
w
n
in
p
r
ev
io
u
s
w
o
r
k
s
i
n
th
e
liter
at
u
r
e
[
2
1
,
2
7
]
th
at
t
h
e
ca
p
ab
ilit
ies
o
f
d
ee
p
er
n
et
w
o
r
k
s
ac
h
iev
e
b
etter
p
er
f
o
r
m
a
n
ce
.
T
h
e
g
lo
b
al
d
escr
ip
to
r
s
ar
e
ex
tr
ac
ted
f
r
o
m
th
e
la
s
t c
o
n
v
o
l
u
tio
n
la
y
er
“
co
n
v
5
_
3
”
an
d
ar
e
o
f
d
i
m
e
n
s
io
n
5
1
2
.
T
h
e
lo
ca
l f
ea
tu
r
es a
r
e
g
r
o
u
p
ed
f
r
o
m
t
h
e
Fas
ter
R
-
C
NN
R
o
I
clu
s
ter
i
n
g
l
a
y
er
.
A
ll e
x
p
er
i
m
e
n
t
s
w
er
e
p
er
f
o
r
m
ed
o
n
a
N
v
id
ia
GT
X
GP
U.
2
.
4
.
F
ine
-
t
un
ing
F
a
s
t
er
R
-
C
NN
Fin
e
t
u
n
in
g
t
h
e
Fas
ter
R
-
C
N
N
n
et
w
o
r
k
allo
w
s
as
to
o
b
tain
f
ea
t
u
r
es
s
p
ec
if
ic
to
f
ac
e
r
etr
iev
al
an
d
s
h
o
u
ld
h
elp
i
m
p
r
o
v
e
t
h
e
p
er
f
o
r
m
a
n
ce
o
f
s
p
atial
a
n
al
y
s
i
s
a
n
d
r
e
-
r
an
k
i
n
g
.
T
o
ac
h
iev
e
th
i
s
,
w
e
ch
o
o
s
e
to
f
i
n
e
-
tu
n
e
Fa
s
ter
R
-
C
NN
to
d
etec
t
t
h
e
q
u
er
y
f
ac
es.
T
h
e
r
esu
l
tin
g
n
et
w
o
r
k
s
w
il
l
b
e
u
s
ed
to
e
x
tr
ac
t
b
etter
lo
ca
l
a
n
d
g
lo
b
al
r
ep
r
esen
tatio
n
s
,
an
d
w
il
l b
e
u
s
ed
to
p
er
f
o
r
m
s
p
atia
l r
er
an
k
i
n
g
.
W
e
ch
o
s
e
to
r
ef
in
e
th
e
m
o
d
el
VGG1
6
Fa
s
ter
R
-
C
N
N,
p
r
e
-
t
r
ain
ed
w
it
h
t
h
e
o
b
j
ec
ts
o
f
P
as
ca
l
VOC,
w
it
h
t
w
o
d
ef
er
en
t
d
atasets
.
T
h
e
f
ir
s
t
n
et
w
o
r
k
w
a
s
r
ef
i
n
ed
u
s
i
n
g
FERET
an
d
Face
s
9
4
d
a
tasets
,
w
e
co
m
b
i
n
e
th
e
m
to
cr
ea
te
o
n
e
b
ig
g
er
d
ata
s
et
.
W
e
m
o
d
i
f
y
t
h
e
o
u
tp
u
t
la
y
er
in
t
h
e
n
et
w
o
r
k
to
r
etu
r
n
4
2
2
class
p
r
o
b
ab
ilit
ies
(
2
6
9
p
eo
p
le
in
th
e
FERET
d
ataset
p
lu
s
1
5
2
p
eo
p
le
in
th
e
F
ac
es9
4
d
ataset,
p
lu
s
o
n
e
ad
d
it
io
n
al
clas
s
f
o
r
th
e
b
ac
k
g
r
o
u
n
d
)
an
d
th
e
ir
co
r
r
esp
o
n
d
in
g
b
o
u
n
d
ed
b
o
u
n
d
b
o
x
co
o
r
d
in
ates
[
2
1
]
.
T
h
is
n
e
w
r
e
f
i
n
ed
n
et
w
o
r
k
w
ill
b
e
ca
lled
VGG(
F
-
F),
t
h
e
tr
ain
i
n
g
p
r
o
ce
s
s
to
o
k
2
h
o
u
r
s
4
7
m
i
n
u
te
s
.
T
h
e
s
ec
o
n
d
n
et
w
o
r
k
w
as
r
ef
i
n
ed
u
s
in
g
Face
Scr
u
b
d
ataset.
We
m
o
d
i
f
y
t
h
e
o
u
tp
u
t
la
y
er
in
t
h
e
n
et
wo
r
k
to
r
etu
r
n
5
3
0
class
p
r
o
b
ab
ilit
ies
(
5
3
0
p
eo
p
le,
p
lu
s
o
n
e
ad
d
itio
n
al
c
lass
f
o
r
th
e
b
ac
k
g
r
o
u
n
d
)
an
d
t
h
eir
co
r
r
esp
o
n
d
in
g
b
o
u
n
d
ed
b
o
u
n
d
b
o
x
co
o
r
d
in
ates.
O
u
r
s
ec
o
n
d
r
ef
i
n
ed
n
et
w
o
r
k
w
il
l b
e
ca
lled
VGG(
F
-
S)
[
2
1
]
,
th
e
tr
ain
i
n
g
to
o
k
2
h
o
u
r
s
3
0
m
i
n
u
te
s
.
W
e
k
ep
t th
e
Fas
ter
R
-
C
NN
’
s
o
r
ig
in
al
p
ar
a
m
eter
s
d
escr
ib
ed
in
[
1
9
]
,
b
u
t d
u
e
to
o
u
r
s
m
aller
n
u
m
b
er
o
f
tr
ain
i
n
g
s
a
m
p
les
we
d
ec
r
ea
s
e
d
th
e
n
u
m
b
er
o
f
iter
atio
n
s
f
r
o
m
8
0
,
0
0
0
t
o
2
0
,
0
0
0
.
W
e
u
s
e
th
e
r
ef
i
n
ed
n
et
w
o
r
k
s
o
f
th
e
tu
n
i
n
g
s
tr
ateg
y
(
VGG(
F
-
S)
&
VGG(
F
-
F))
o
n
all
d
atasets
to
ex
tr
ac
t
i
m
ag
e
a
n
d
r
eg
io
n
d
escr
ip
to
r
s
to
p
er
f
o
r
m
a
f
ac
e
r
etr
iev
al.
2
.
5
.
F
a
s
t
er
R
-
CNN
f
ea
t
ures
& F
a
ce
det
ec
t
io
n
W
e
ev
a
lu
a
te
t
h
e
i
m
p
ac
t
o
f
u
s
i
n
g
a
f
ac
e
d
etec
tio
n
alg
o
r
it
h
m
o
n
o
u
r
d
ataset
s
a
n
d
q
u
er
ies
b
ef
o
r
e
u
s
in
g
Fas
ter
R
-
C
NN
f
o
r
f
ea
tu
r
e
ex
tr
ac
tio
n
an
d
th
e
r
a
n
k
in
g
an
d
r
er
an
k
i
n
g
s
tr
ate
g
ies a
s
d
escr
ib
ed
p
r
ev
io
u
s
l
y
.
2
.
6
.
F
a
s
t
er
R
-
CNN
f
ea
t
ures
& F
Vs
T
o
ex
p
lo
r
e
th
e
r
elev
an
ce
o
f
u
s
in
g
FV
s
o
n
C
NN
f
ea
tu
r
e,
f
o
r
th
e
i
m
a
g
e
-
to
-
v
id
eo
f
ac
e
r
etr
iev
al
tas
k
,
w
e
f
ir
s
t
e
x
tr
ac
t
t
h
e
C
NN
f
e
atu
r
es
o
f
ea
c
h
f
r
a
m
e.
W
e
t
h
en
ap
p
l
y
P
r
in
cip
al
C
o
m
p
o
n
e
n
t
An
al
y
s
i
s
(
P
C
A
)
,
Gau
s
s
ia
n
m
ix
t
u
r
e
m
o
d
el
(
GM
M)
,
L
2
n
o
r
m
al
izatio
n
o
n
th
o
s
e
f
ea
t
u
r
es
b
ef
o
r
e
u
s
i
n
g
o
u
r
F
V
f
u
n
ct
io
n
.
Fi
n
all
y
,
as d
escr
ib
ed
b
ef
o
r
e,
w
e
co
m
p
u
te
th
e
s
i
m
ilar
it
y
m
ea
s
u
r
e
a
n
d
u
s
e
t
h
e
r
an
k
in
g
an
d
r
er
an
k
i
n
g
s
tr
ateg
ie
s
.
2
.
7
.
F
a
s
t
er
R
-
CNN
f
ea
t
ures
& B
O
VW
T
o
ex
p
lo
r
e
th
e
r
elev
an
ce
o
f
u
s
in
g
B
OVW
w
it
h
C
NN
f
ea
t
u
r
e,
f
o
r
th
e
i
m
ag
e
-
to
-
v
id
eo
f
ac
e
r
etr
iev
al
task
,
w
e
f
ir
s
t
e
x
tr
ac
t
t
h
e
C
NN
f
ea
t
u
r
es
o
f
ea
c
h
f
r
a
m
e.
T
h
en
w
e
ap
p
l
y
t
h
e
cl
u
s
ter
i
n
g
,
v
ec
to
r
q
u
an
tizatio
n
an
d
in
v
er
ted
i
n
d
ex
i
n
g
s
tep
s
.
Fi
n
all
y
,
as
d
escr
ib
ed
b
ef
o
r
e,
w
e
co
m
p
u
te
t
h
e
s
i
m
ilar
it
y
m
e
asu
r
e
an
d
u
s
e
t
h
e
r
er
an
k
i
n
g
s
tr
ateg
ie
s
.
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
W
e
ev
alu
ate
t
h
e
u
s
e
o
f
Fa
s
ter
R
-
C
NN
f
ea
t
u
r
es
f
o
r
f
a
ce
i
m
ag
e
to
v
id
eo
f
ac
e
r
etr
iev
al.
W
e
ex
p
er
i
m
e
n
ted
w
it
h
s
ix
d
if
f
er
en
t
s
i
m
ilar
it
y
m
etr
ics.
T
h
e
r
esu
lt
s
w
er
e
s
i
m
ilar
a
n
d
clo
s
e
b
u
t
o
v
er
all
co
s
i
n
e
p
er
f
o
r
m
ed
b
etter
.
T
ab
le
1
s
h
o
w
s
a
n
e
x
a
m
p
le
o
f
o
u
r
r
es
u
lts
w
h
e
n
u
s
i
n
g
f
ea
tu
r
es
f
r
o
m
a
n
o
n
th
e
s
h
el
f
n
et
w
o
r
k
w
it
h
VG
G1
6
ar
ch
itectu
r
e
tr
ai
n
ed
o
n
p
ascal
d
ataset.
W
e
ca
r
r
ied
o
u
t
a
co
m
p
ar
ati
v
e
s
tu
d
y
o
f
th
e
s
u
m
an
d
m
ax
-
p
o
o
lin
g
s
tr
ate
g
ie
s
o
f
th
e
i
m
a
g
e
-
w
i
s
e
an
d
r
eg
io
n
-
w
is
e
d
escr
ip
to
r
s
.
T
ab
le
2
s
u
m
m
ar
ize
s
m
o
s
t
o
f
o
u
r
r
esu
lt
s
.
A
cc
o
r
d
in
g
to
o
u
r
e
x
p
er
i
m
en
ts
,
t
h
e
s
u
m
-
p
o
o
lin
g
g
i
v
e
s
b
etter
p
er
f
o
r
m
a
n
ce
th
a
n
t
h
e
m
ax
-
p
o
o
lin
g
.
I
t
also
s
h
o
w
s
th
e
p
er
f
o
r
m
an
ce
o
f
Fas
ter
R
-
C
NN
w
it
h
a
VGG1
6
ar
ch
itect
u
r
es
t
r
ain
ed
o
n
t
w
o
d
if
f
er
e
n
t
d
atase
ts
(
P
ascal
VOC
an
d
C
O
C
O)
,
VGG1
6
tr
a
in
ed
o
n
C
OC
O
p
er
f
o
r
m
ed
b
etter
b
ec
au
s
e
th
e
d
ataset
is
b
ig
g
er
an
d
m
o
r
e
d
iv
er
s
e.
Mo
r
eo
v
er
,
it
p
r
esen
t
s
th
e
i
m
p
ac
t
o
f
s
p
atial
r
er
an
k
i
n
g
an
d
q
u
er
y
e
x
p
an
s
io
n
.
Usi
n
g
th
e
g
lo
b
al
f
e
atu
r
es
o
f
Fa
s
ter
R
-
C
NN
o
n
th
eir
o
w
n
w
it
h
o
u
t
an
y
r
er
an
k
i
n
g
s
tr
ate
g
y
g
i
v
es
th
e
b
est
r
e
s
u
lt
s
.
Sp
atial
r
er
an
k
i
n
g
&
QE
h
ad
n
o
p
o
s
itiv
e
i
m
p
ac
t
o
n
th
e
r
esu
lts
.
W
e
s
h
o
u
ld
n
o
te
t
h
at
in
av
er
a
g
e
t
h
e
o
f
f
li
n
e
f
ea
t
u
r
e
e
x
tr
ac
tio
n
t
o
o
k
2
9
.
7
m
i
n
u
tes
w
h
ile
t
h
e
o
n
li
n
e
r
an
k
i
n
g
s
tep
s
to
o
k
3
.
7
s
ec
o
n
d
s
a
n
d
t
h
e
r
e
r
an
k
i
n
g
s
tr
ate
g
y
to
o
k
7
m
i
n
u
tes
f
o
r
Y
-
C
eleb
d
atas
et
.
Fo
r
Yo
u
T
u
b
e
Face
s
Data
b
ase,
th
e
o
f
f
li
n
e
f
ea
t
u
r
e
ex
tr
ac
tio
n
to
o
k
2
0
h
o
u
r
s
w
h
il
e
th
e
o
n
li
n
e
r
a
n
k
in
g
s
tep
s
to
o
k
o
n
l
y
8
5
s
ec
o
n
d
s
an
d
th
e
r
er
an
k
in
g
s
tr
ateg
y
to
o
k
2
1
m
in
u
te
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
A
r
ti
f
I
n
tell
I
SS
N:
2252
-
8938
La
r
g
e
-
s
ca
le
ima
g
e
-
to
-
vid
e
o
fa
ce
r
etri
ev
a
l wi
th
co
n
vo
lu
tio
n
a
l n
eu
r
a
l n
etw
o
r
k
fea
tu
r
es
(
I
ma
n
e
Ha
ch
ch
a
n
e
)
43
T
ab
le
1
.
Me
an
A
v
er
a
g
e
P
r
ec
is
io
n
(
m
A
P
)
Of
P
r
etr
ain
ed
Fas
te
r
R
-
C
NN
Mo
d
els
w
it
h
Vg
g
1
6
A
r
c
h
itect
u
r
es o
n
P
ascal
Data
s
et
Usi
n
g
Di
f
f
er
e
n
t Si
m
i
lar
it
y
Me
a
s
u
r
e
s
Usi
n
g
Y
-
C
eleb
Data
s
et.
S
i
mi
l
a
r
i
t
y
me
t
r
i
c
P
o
o
l
i
n
g
R
a
n
k
i
n
g
R
e
r
a
n
k
i
n
g
QE
C
o
si
n
e
max
0
.
8
8
8
0
.
8
6
0
0
.
5
5
0
su
m
0
.
9
1
5
0
.
8
4
6
0
.
6
0
0
M
a
n
h
a
t
t
a
n
max
0
.
9
0
0
0
.
8
6
9
0
.
5
7
0
su
m
0
.
9
0
5
0
.
8
4
1
0
.
4
2
8
Eu
c
l
i
d
i
a
n
max
0
.
8
8
8
0
.
8
6
0
0
.
5
5
0
su
m
0
.
9
1
5
0
.
8
4
6
0
.
6
0
0
C
i
t
y
B
l
o
c
k
max
0
.
9
0
0
0
.
8
6
9
0
.
5
7
0
su
m
0
.
9
0
5
0
.
8
4
1
0
5
7
8
L1
max
0
.
9
0
0
0
.
8
6
9
0
.
5
7
0
su
m
0
.
9
0
5
0
.
8
4
1
0
.
5
7
8
L2
max
0
.
8
8
8
0
.
8
6
0
0
.
5
5
0
su
m
0
.
9
1
5
0
.
8
4
6
0
.
6
0
3
Y
-
C
eleb
-
Face
s
co
l
u
m
n
p
r
ese
n
t
th
e
r
e
s
u
lt
s
o
f
u
s
i
n
g
f
ac
e
d
ete
ctio
n
o
n
t
h
e
Y
-
C
eleb
d
ataset.
As
w
e
ca
n
s
ee
in
T
ab
le
2
f
ac
e
d
etec
tio
n
d
id
n
o
t
im
p
r
o
v
e
th
e
r
esu
lt
s
.
W
e
s
h
o
u
ld
n
o
te
th
at
w
e
w
er
e
ab
le
to
r
ed
u
ce
th
e
r
an
k
i
n
g
ti
m
e
to
2
.
4
s
ec
o
n
d
s
o
n
av
er
ag
e.
T
ab
le
2
s
h
o
w
t
h
at
th
e
r
ef
i
n
ed
f
ea
tu
r
es
s
li
g
h
tl
y
ex
ce
ed
ed
th
e
r
a
w
f
ea
t
u
r
es
i
n
th
e
s
p
atial
r
er
an
k
in
g
a
n
d
t
h
e
QE
s
ta
g
es.
B
u
t
s
till
,
t
h
e
g
lo
b
al
f
ea
t
u
r
es
o
f
F
aster
R
-
C
NN
f
r
o
m
VGG1
6
tr
ain
ed
o
n
C
OC
O
u
s
e
d
w
it
h
o
u
t a
n
y
r
er
an
k
i
n
g
s
tr
ate
g
y
g
i
v
e
th
e
b
est r
es
u
lt
s
.
T
ab
le
2
.
Me
an
A
v
er
a
g
e
P
r
ec
is
io
n
(
m
A
P
)
o
f
p
r
e
-
tr
ain
ed
Fas
t
er
R
-
C
NN
m
o
d
els
w
i
th
V
GG1
6
ar
ch
itectu
r
es.
(
P)
,
(
C
)
,
(
F
-
S)
A
N
D
(
F
-
F)
d
en
o
te
w
h
eth
er
t
h
e
n
et
w
o
r
k
w
as t
r
ain
ed
w
i
th
P
ascal
VO
C
,
Mic
r
o
s
o
f
t CO
C
O,
Face
Scr
u
b
o
r
Fer
et
&
Face
s
9
4
i
m
ag
e
s
,
r
esp
ec
tiv
e
l
y
.
W
ith
a
co
m
p
ar
is
o
n
b
et
w
ee
n
s
u
m
a
n
d
m
ax
p
o
o
lin
g
s
tr
ateg
ie
s
.
W
h
en
i
n
d
icate
d
,
Q
E
is
ap
p
lied
w
ith
M
=
5
N
e
t
w
o
r
k
P
o
o
l
i
n
g
Y
-
C
e
l
e
b
Y
o
u
T
u
b
e
F
a
c
e
s Dat
a
b
a
se
Y
-
C
e
l
e
b
-
F
a
c
e
s
R
a
n
k
i
n
g
R
e
r
a
n
k
i
n
g
QE
R
a
n
k
i
n
g
R
e
r
a
n
k
i
n
g
QE
R
a
n
k
i
n
g
R
e
r
a
n
k
i
n
g
QE
V
G
G
1
6
(P)
max
0
.
8
8
8
0
.
8
6
0
0
.
5
5
0
0
.
8
9
2
0
.
8
7
7
0
.
8
8
2
0
.
5
7
4
0
.
5
1
6
0
.
5
4
2
su
m
0
.
9
1
5
0
.
8
4
6
0
.
6
0
0
0
.
8
9
7
0
.
8
8
6
0
.
8
9
1
0
.
6
1
8
0
.
4
8
6
0
.
5
1
1
V
G
G
1
6
(
C
)
max
0
.
9
1
1
0
.
8
8
8
0
.
5
2
2
0
.
8
9
2
0
.
8
7
8
0
.
8
8
9
0
.
6
2
2
0
.
5
7
4
0
.
6
1
7
su
m
0
.
9
2
6
0
.
8
0
7
0
.
5
1
2
0
.
9
0
3
0
.
8
8
2
0
.
8
9
6
0
.
7
0
5
0
.
5
3
8
0
.
5
5
1
V
G
G
1
6
(F
-
S)
max
0
.
8
0
9
0
.
7
7
7
0
.
4
5
7
0
.
8
4
8
0
.
8
3
4
0
.
8
3
8
0
.
4
7
7
0
.
4
2
3
0
.
4
5
0
su
m
0
.
9
1
7
0
.
8
4
3
0
.
5
7
8
0
.
8
8
2
0
.
8
7
3
0
.
8
7
4
0
.
6
3
5
0
.
5
0
9
0
.
5
1
9
V
G
G
1
6
(F
-
F)
max
0
.
9
1
5
0
.
8
7
4
0
.
5
5
4
0
.
8
9
4
0
.
8
8
4
0
.
8
8
7
0
.
6
6
6
0
.
6
5
6
0
.
6
8
2
su
m
0
.
9
2
4
0
.
8
9
9
0
.
6
2
1
0
.
8
9
6
0
.
8
9
2
0
.
8
9
3
0
.
7
1
5
0
.
6
1
2
0
.
6
4
6
W
h
en
u
s
i
n
g
FVs
w
it
h
Fas
ter
R
-
C
NN
f
ea
t
u
r
es
w
e
ca
n
s
a
y
th
at
m
ax
p
o
o
lin
g
p
er
f
o
r
m
ed
b
etter
,
as
s
h
o
w
n
i
n
T
ab
le
3
,
b
u
t
it
is
c
l
ea
r
th
at
u
s
i
n
g
FV
s
i
s
n
o
t
a
g
o
o
d
id
ea
.
T
h
e
m
A
P
is
v
er
y
lo
w
(
b
elo
w
1
0
%).
W
e
co
u
ld
n
’
t te
s
t o
n
t
h
e
Yo
u
T
u
b
e
Face
s
Data
b
ase
d
u
e
to
a
Me
m
o
r
y
E
r
r
o
r
ca
u
s
ed
b
y
th
e
s
ize
o
f
t
h
e
d
ataset
a
n
d
t
h
e
li
m
ita
tio
n
o
f
t
h
e
h
ar
d
w
ar
e.
T
ab
le
3
.
Me
an
A
v
er
a
g
e
P
r
ec
is
io
n
(
m
A
P
)
o
f
p
r
e
-
tr
ain
ed
Fas
t
er
R
-
C
NN
m
o
d
els
w
i
th
V
GG1
6
ar
ch
itectu
r
es.
(
P
)
an
d
,
(
C
)
d
en
o
te
w
h
e
th
er
t
h
e
n
et
w
o
r
k
w
as tr
ai
n
ed
w
it
h
P
ascal
VOC o
r
Mic
r
o
s
o
f
t
C
OC
O
.
W
ith
a
co
m
p
ar
is
o
n
b
et
w
ee
n
s
u
m
a
n
d
m
ax
p
o
o
lin
g
s
tr
ate
g
ie
s
.
W
h
en
in
d
icate
d
,
QE
is
ap
p
lied
w
it
h
M
=
5
N
e
t
w
o
r
k
P
o
o
l
i
n
g
Y
-
C
e
l
e
b
R
a
n
k
i
n
g
R
e
r
a
n
k
i
n
g
QE
V
G
G
1
6
(
P
)
max
0
.
0
9
7
0
.
1
0
2
0
.
0
9
7
su
m
0
.
0
9
7
0
.
1
0
0
0
.
1
0
2
V
G
G
1
6
(
C
)
max
0
.
0
9
7
0
.
1
0
2
0
.
0
9
8
su
m
0
.
0
9
7
0
.
0
9
7
0
.
0
9
7
W
h
en
u
s
i
n
g
o
n
B
OVW
w
ith
Fas
ter
R
-
C
NN
f
ea
tu
r
e
s
w
e
co
u
ld
n
’
t
a
n
al
y
ze
th
e
f
u
ll
r
es
u
lts
b
ec
au
s
e
we
k
ep
t
r
u
n
n
i
n
g
in
to
a
Me
m
o
r
y
E
r
r
o
r
ca
u
s
ed
by
th
e
s
izes
o
f
t
h
e
d
atasets
a
n
d
th
e
li
m
itat
io
n
o
f
th
e
h
ar
d
w
ar
e
i
n
ad
d
itio
n
to
th
at
th
e
r
esu
l
t
o
b
tain
ed
w
er
e
n
o
t
th
at
en
co
u
r
ag
in
g
.
T
ab
le
4
p
r
esen
t
th
e
r
esu
lts
t
h
at
w
e
w
er
e
ab
le
to
g
et.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8938
I
n
t J
A
r
ti
f
I
n
tell
,
Vo
l.
9
,
No
.
1
,
Ma
r
ch
20
20
:
40
–
45
44
T
ab
le
4
.
Me
an
A
v
er
a
g
e
P
r
ec
is
io
n
(
m
A
P
)
o
f
p
r
e
-
tr
ain
ed
Fas
t
er
R
-
C
NN
m
o
d
els
w
i
th
V
GG1
6
ar
ch
itectu
r
es.
(
C
)
d
en
o
te
th
at
t
h
e
n
et
w
o
r
k
was tr
ain
ed
w
it
h
Mic
r
o
s
o
f
t
C
O
C
O.
W
ith
a
co
m
p
ar
is
o
n
b
et
w
e
en
s
u
m
a
n
d
m
a
x
p
o
o
lin
g
s
tr
ate
g
ies.
W
h
e
n
i
n
d
icate
d
,
QE
is
ap
p
lied
w
it
h
M
=
5
N
e
t
w
o
r
k
P
o
o
l
i
n
g
Y
-
C
e
l
e
b
R
a
n
k
i
n
g
R
e
r
a
n
k
i
n
g
QE
V
G
G
1
6
(
C
)
max
0
.
0
3
2
0
0
.
0
9
7
su
m
0
.
0
3
2
-
-
Fin
all
y
,
w
e
ca
n
clea
r
l
y
s
ee
i
n
th
at
t
h
e
r
a
w
f
a
s
ter
R
-
C
N
N
f
ea
tu
r
es
lar
g
el
y
o
u
tp
er
f
o
r
m
e
d
th
e
o
th
er
s
tr
ateg
ie
s
w
it
h
a
m
A
P
o
f
9
2
.
6
%.
T
ab
le
5
s
h
o
w
c
o
m
p
ar
is
o
n
w
it
h
S
tate
-
of
-
th
e
-
ar
t
.
T
ab
le
5
.
C
o
m
p
ar
is
o
n
w
i
th
S
tat
e
-
of
-
t
h
e
-
ar
t.
R
es
u
lts
p
r
o
v
id
ed
as
m
A
P
.
M
e
t
h
o
d
Y
-
C
e
l
e
b
Y
o
u
T
u
b
e
F
a
c
e
s Dat
a
b
a
se
NN
[
2
3
]
-
0
.
1
4
5
O
-
S
B
o
F
[
2
9
]
-
0
.
4
7
1
RN
-
B
O
F
[
3
0
]
-
0
.
4
6
5
F
a
st
e
r
R
-
C
N
N
f
e
a
t
u
r
e
s
0
.
9
2
6
0
.
9
0
3
F
a
st
e
r
R
-
C
N
N
f
e
a
t
u
r
e
s +
F
V
0
.
0
9
7
0
.
0
0
6
F
a
st
e
r
R
-
C
N
N
f
e
a
t
u
r
e
s +
B
O
V
W
0
.
0
3
2
0
.
0
0
1
4.
CO
NCLU
SI
O
N
T
h
is
ar
ticle
ex
p
lo
r
es
t
h
e
u
s
e
o
f
f
ea
tu
r
e
s
f
r
o
m
a
n
o
b
j
ec
t
d
etec
tio
n
C
N
N
f
o
r
i
m
a
g
e
-
to
-
v
id
eo
f
ac
e
r
etr
iev
al.
I
t
u
s
e
s
Fas
ter
R
-
C
N
N
f
ea
t
u
r
es
as
g
lo
b
al
a
n
d
lo
ca
l
d
escr
ip
to
r
s
.
W
e
h
a
v
e
s
h
o
w
n
th
at
t
h
e
co
m
m
o
n
s
i
m
ilar
it
y
m
etr
ic
g
iv
e
s
s
i
m
ilar
r
esu
lt
s
.
W
e
al
s
o
f
o
u
n
d
t
h
at
s
u
m
-
p
o
o
lin
g
p
er
f
o
r
m
s
b
etter
th
an
m
ax
-
p
o
o
lin
g
i
n
m
o
s
t
ca
s
e
s
,
an
d
co
n
tr
ar
y
to
o
u
r
p
r
ev
io
u
s
w
o
r
k
[
2
1
]
f
in
e
tu
n
i
n
g
d
o
es
n
o
t
i
m
p
r
o
v
e
th
e
r
esu
lt
s
.
Mo
r
e
i
m
p
o
r
tan
tl
y
,
w
e
f
o
u
n
d
t
h
at
ap
p
ly
in
g
t
h
e
s
i
m
ilar
it
y
m
ea
s
u
r
e
d
ir
ec
tly
o
n
th
e
C
NN
f
ea
t
u
r
e
o
f
an
o
f
f
-
t
h
e
-
s
h
e
lf
C
NN
tr
ain
ed
o
n
a
lar
g
e
a
n
d
d
iv
er
s
e
d
ata
s
et
g
a
v
e
t
h
e
b
es
t
r
esu
lt
s
,
a
n
d
t
h
at
u
s
in
g
FVs
o
r
B
OVW
is
m
e
m
o
r
y
co
n
s
u
m
i
n
g
an
d
is
n
o
t
s
u
itab
le
f
o
r
C
NN
f
ea
tu
r
e
s
in
t
h
is
ca
s
e.
ACK
NO
WL
E
D
G
E
M
E
NT
S
T
h
i
s
w
o
r
k
f
all
s
w
it
h
i
n
t
h
e
s
c
o
p
e
o
f
B
ig
Data
an
d
C
o
n
n
ec
ted
Ob
j
ec
t
(
B
DC
O)
.
W
e
w
o
u
ld
li
k
e
to
th
an
k
t
h
e
Has
s
an
I
I
Un
iv
er
s
it
y
o
f
C
asab
la
n
ca
f
o
r
f
i
n
a
n
cin
g
t
h
is
p
r
o
j
ec
t.
RE
F
E
R
E
NC
E
S
[1
]
A
.
F
il
g
u
e
iras
De
A
r
a
u
jo
,
“
L
a
rg
e
-
S
c
a
le
V
id
e
o
Re
tri
e
v
a
l
Us
in
g
I
m
a
g
e
Qu
e
ries
A
Diss
e
rtatio
n
S
u
b
m
it
ted
T
o
T
h
e
De
p
a
rtme
n
t
Of
El
e
c
tri
c
a
l
En
g
in
e
e
rin
g
A
n
d
T
h
e
Co
m
m
it
te
e
On
G
r
a
d
u
a
te
S
tu
d
ies
Of
S
tan
f
o
rd
Un
iv
e
rsit
y
In
P
a
rt
ial
F
u
lf
il
lm
e
n
t
Of
T
h
e
Re
q
u
irem
e
n
ts F
o
r
T
h
e
De
g
re
e
Of
Do
c
to
r
Of
P
h
il
o
s,”
2
0
1
6
.
[2
]
J.
Y.
-
H.
Ng
,
M
.
Ha
u
sk
n
e
c
h
t,
S
.
Vijay
a
n
a
ra
si
m
h
a
n
,
O.
V
i
n
y
a
ls,
R.
M
o
n
g
a
,
a
n
d
G
.
T
o
d
e
rici,
“
Be
y
o
n
d
S
h
o
r
t
S
n
ip
p
e
ts:
De
e
p
Ne
tw
o
rk
s f
o
r
V
id
e
o
Clas
sif
ica
ti
o
n
,
”
M
a
r.
2
0
1
5
.
[3
]
K.
S
im
o
n
y
a
n
a
n
d
A
.
Zi
ss
e
r
m
a
n
,
“
Tw
o
-
S
trea
m
Co
n
v
o
lu
ti
o
n
a
l
Ne
tw
o
rk
s
f
o
r
A
c
ti
o
n
Re
c
o
g
n
it
io
n
i
n
Vid
e
o
s,”
J
u
n
.
2
0
1
4
.
[4
]
D.
T
ra
n
,
L
.
Bo
u
rd
e
v
,
R.
F
e
rg
u
s,
L
.
T
o
rre
sa
n
i,
a
n
d
M
.
P
a
lu
r
i,
“
L
e
a
rn
in
g
S
p
a
ti
o
tem
p
o
ra
l
F
e
a
tu
re
s
w
it
h
3
D
Co
n
v
o
l
u
ti
o
n
a
l
Ne
tw
o
rk
s,”
De
c
.
2
0
1
4
.
[5
]
A
.
Ba
b
e
n
k
o
,
A
.
S
les
a
re
v
,
A
.
Ch
ig
o
rin
,
a
n
d
V.
L
e
m
p
it
sk
y
,
“
Ne
u
ra
l
Co
d
e
s f
o
r
Im
a
g
e
Re
tri
e
v
a
l,
”
A
p
r.
2
0
1
4
.
[6
]
Y.
Ka
lan
ti
d
is,
C.
M
e
ll
in
a
,
a
n
d
S
.
Os
in
d
e
ro
,
“
Cro
ss
-
d
im
e
n
sio
n
a
l
W
e
ig
h
ti
n
g
f
o
r
A
g
g
re
g
a
ted
De
e
p
Co
n
v
o
lu
t
io
n
a
l
F
e
a
tu
re
s,”
De
c
.
2
0
1
5
.
[7
]
A
.
S
.
R
a
z
a
v
ian
,
J.
S
u
ll
iv
a
n
,
S
.
Ca
rlsso
n
,
a
n
d
A
.
M
a
k
i,
“
V
isu
a
l
In
sta
n
c
e
Re
tri
e
v
a
l
w
it
h
D
e
e
p
Co
n
v
o
lu
ti
o
n
a
l
Ne
tw
o
rk
s,”
De
c
.
2
0
1
4
.
[8
]
L
.
W
u
,
Y.
W
a
n
g
,
Z.
G
e
,
Q.
Hu
,
a
n
d
X.
L
i,
“
S
tru
c
tu
re
d
d
e
e
p
h
a
sh
in
g
w
it
h
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
t
w
o
rk
s
f
o
r
f
a
s
t
p
e
rso
n
re
-
i
d
e
n
ti
f
ica
ti
o
n
,
”
C
o
mp
u
t
.
Vi
s.
Im
a
g
e
Un
d
e
rs
t.
,
v
o
l.
1
6
7
,
p
p
.
6
3
–
7
3
,
F
e
b
.
2
0
1
8
.
[9
]
C.
He
rrm
a
n
n
a
n
d
J.
Be
y
e
re
r,
“
F
a
s
t
f
a
c
e
re
c
o
g
n
it
io
n
b
y
u
sin
g
a
n
i
n
v
e
rted
in
d
e
x
,
”
2
0
1
5
,
v
o
l.
9
4
0
5
,
p
.
9
4
0
5
0
7
.
[1
0
]
R.
A
ra
n
d
jelo
v
ic
a
n
d
A
.
Zi
ss
e
rm
a
n
,
“
T
h
re
e
th
i
n
g
s
e
v
e
r
y
o
n
e
sh
o
u
l
d
k
n
o
w
to
im
p
ro
v
e
o
b
jec
t
re
tri
e
v
a
l,
”
in
2
0
1
2
I
EE
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
,
2
0
1
2
,
p
p
.
2
9
1
1
–
2
9
1
8
.
[1
1
]
J.
P
o
n
t
-
T
u
se
t,
P
.
A
rb
e
lae
z
,
J.
T
.
Ba
rro
n
,
F
.
M
a
rq
u
e
s,
a
n
d
J.
M
a
li
k
,
“
M
u
lt
isc
a
le
Co
m
b
in
a
to
rial
G
ro
u
p
in
g
f
o
r
I
m
a
g
e
S
e
g
m
e
n
tatio
n
a
n
d
Ob
jec
t
P
r
o
p
o
sa
l
G
e
n
e
ra
ti
o
n
,
”
M
a
r.
2
0
1
5
.
[1
2
]
J.
L
o
n
g
,
E.
S
h
e
lh
a
m
e
r,
a
n
d
T
.
Da
rre
ll
,
“
F
u
ll
y
Co
n
v
o
l
u
ti
o
n
a
l
Ne
tw
o
rk
s f
o
r
S
e
m
a
n
ti
c
S
e
g
m
e
n
tatio
n
,
”
No
v
.
2
0
1
4
.
[1
3
]
D.
G
.
L
o
w
e
,
“
Distin
c
ti
v
e
I
m
a
g
e
F
e
a
tu
re
s
f
ro
m
S
c
a
le
-
In
v
a
rian
t
Ke
y
p
o
in
ts,”
In
t.
J
.
Co
m
p
u
t
.
Vi
s.
,
v
o
l.
6
0
,
n
o
.
2
,
p
p
.
91
–
1
1
0
,
N
o
v
.
2
0
0
4
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
A
r
ti
f
I
n
tell
I
SS
N:
2252
-
8938
La
r
g
e
-
s
ca
le
ima
g
e
-
to
-
vid
e
o
fa
ce
r
etri
ev
a
l wi
th
co
n
vo
lu
tio
n
a
l n
eu
r
a
l n
etw
o
r
k
fea
tu
r
es
(
I
ma
n
e
Ha
ch
ch
a
n
e
)
45
[1
4
]
M
.
Ca
lo
n
d
e
r,
V
.
L
e
p
e
ti
t,
C.
S
tr
e
c
h
a
,
a
n
d
P
.
F
u
a
,
“
BRIEF
:
Bin
a
r
y
Ro
b
u
st
In
d
e
p
e
n
d
e
n
t
El
e
m
e
n
tar
y
F
e
a
tu
re
s,”
S
p
rin
g
e
r,
Be
rli
n
,
He
id
e
l
b
e
rg
,
2
0
1
0
,
p
p
.
7
7
8
–
7
9
2
.
[1
5
]
J.
Y.
H.
Ng
,
M
.
Ha
u
sk
n
e
c
h
t,
S
.
Vijay
a
n
a
ra
si
m
h
a
n
,
O.
V
i
n
y
a
ls,
R.
M
o
n
g
a
,
a
n
d
G
.
T
o
d
e
rici,
“
Be
y
o
n
d
sh
o
rt
sn
ip
p
e
ts:
De
e
p
n
e
tw
o
rk
s
f
o
r
v
id
e
o
c
las
si
f
i
c
a
ti
o
n
,
”
Pro
c
.
IEE
E
Co
mp
u
t.
S
o
c
.
Co
n
f
.
Co
mp
u
t.
V
is.
Pa
t
ter
n
Rec
o
g
n
it
.
,
v
o
l
.
0
7
-
12
-
Ju
n
e
,
p
p
.
4
6
9
4
–
4
7
0
2
,
2
0
1
5
.
[1
6
]
A
.
A
ra
u
jo
a
n
d
B.
G
iro
d
,
“
L
a
rg
e
-
S
c
a
le
V
id
e
o
Re
tri
e
v
a
l
Us
in
g
Im
a
g
e
Qu
e
ries
,
”
IEE
E
T
ra
n
s.
Circ
u
it
s
S
y
st.
Vi
d
e
o
T
e
c
h
n
o
l
.
,
v
o
l
.
X
X,
n
o
.
c
,
p
p
.
1
–
1
,
2
0
1
7
.
[1
7
]
G
.
De
Oliv
e
ira Barra
,
M
.
L
u
x
,
a
n
d
X
.
G
iro
-
I
-
Nie
to
,
“
L
a
rg
e
sc
a
le c
o
n
ten
t
-
b
a
se
d
v
id
e
o
re
tri
e
v
a
l
w
it
h
L
Iv
RE,
”
Pro
c
.
-
In
t.
W
o
rk
.
C
o
n
te
n
t
-
Ba
se
d
M
u
lt
im
e
d
.
In
d
e
x
.
,
v
o
l.
2
0
1
6
-
Ju
n
e
,
2
0
1
6
.
[1
8
]
S
.
Re
n
,
K.
He
,
R.
G
irsh
ick
,
a
n
d
J.
S
u
n
,
“
F
a
ste
r
R
-
CNN
:
T
o
wa
rd
s
Re
al
-
T
i
m
e
Ob
jec
t
De
t
e
c
ti
o
n
w
it
h
Re
g
io
n
P
r
o
p
o
sa
l
Ne
tw
o
rk
s,”
IEE
E
T
ra
n
s.
Pa
tt
e
rn
An
a
l.
M
a
c
h
.
I
n
tell.
,
v
o
l.
3
9
,
n
o
.
6
,
p
p
.
1
1
3
7
–
1
1
4
9
,
2
0
1
7
.
[1
9
]
R.
G
irsh
ick
,
“
F
a
st
R
-
CNN
,
”
in
2
0
1
5
IEE
E
In
ter
n
a
ti
o
n
a
l
Co
n
fer
e
n
c
e
o
n
Co
m
p
u
ter
Vi
si
o
n
(
ICCV)
,
2
0
1
5
,
p
p
.
1
4
4
0
–
1
4
4
8
.
[2
0
]
H.
Jia
n
g
a
n
d
E.
L
e
a
rn
e
d
-
M
il
ler,
“
F
a
c
e
De
t
e
c
ti
o
n
w
it
h
th
e
F
a
ste
r
R
-
CNN
,
”
Pro
c
.
-
1
2
t
h
IEE
E
In
t.
Co
n
f.
A
u
t
o
m.
Fa
c
e
Ge
stu
re
Rec
o
g
n
it
io
n
,
FG
2
0
1
7
-
1
st
I
n
t.
W
o
rk
.
Ad
a
p
t.
S
h
o
t
L
e
a
rn
.
Ge
stu
re
Un
d
e
rs
t.
Pro
d
.
A
S
L
4
GU
P
2
0
1
7
,
Bi
o
me
trics
W
il
d
,
Bwi
ld
2
0
1
7
,
He
t
e
ro
g
e
,
p
p
.
6
5
0
–
65
7
,
2
0
1
7
.
[2
1
]
I.
Ha
c
h
c
h
a
n
e
,
A
.
Ba
d
ri,
A
.
S
a
h
e
l,
a
n
d
Y.
Ru
ich
e
k
,
“
Ne
w
F
a
ste
r
R
-
CNN
Ne
u
ro
n
a
l
A
p
p
ro
a
c
h
f
o
r
F
a
c
e
Re
tri
e
v
a
l,
”
in
L
e
c
tu
re
No
tes
in
Ne
tw
o
rk
s a
n
d
S
y
ste
ms
,
v
o
l.
6
6
,
2
0
1
9
,
p
p
.
1
1
3
–
1
2
0
.
[2
2
]
M
in
y
o
u
n
g
Kim
,
S
.
Ku
m
a
r,
V
.
P
a
v
lo
v
ic,
a
n
d
H.
Ro
w
le
y
,
“
F
a
c
e
tra
c
k
in
g
a
n
d
re
c
o
g
n
it
i
o
n
w
it
h
v
isu
a
l
c
o
n
stra
i
n
ts
in
re
a
l
-
w
o
rld
v
id
e
o
s,”
i
n
2
0
0
8
I
EE
E
Co
n
fer
e
n
c
e
o
n
Co
m
p
u
ter
V
isio
n
a
n
d
Pa
tt
e
rn
Rec
o
g
n
it
io
n
,
2
0
0
8
,
p
p
.
1
–
8.
[2
3
]
L
.
W
o
lf
,
T
.
Ha
s
sn
e
r,
a
n
d
I.
M
a
o
z
,
“
F
a
c
e
re
c
o
g
n
it
io
n
i
n
u
n
c
o
n
stra
i
n
e
d
v
id
e
o
s
w
it
h
m
a
tch
e
d
b
a
c
k
g
ro
u
n
d
sim
il
a
rit
y
,
”
in
CVP
R
2
0
1
1
,
2
0
1
1
,
p
p
.
5
2
9
–
5
3
4
.
[2
4
]
P
.
J.
P
h
il
li
p
s
,
H.
W
e
c
h
sle
r,
J.
H
u
a
n
g
,
a
n
d
P
.
J.
Ra
u
ss
,
“
T
h
e
F
E
RET
d
a
tab
a
se
a
n
d
e
v
a
lu
a
ti
o
n
p
r
o
c
e
d
u
re
f
o
r
f
a
c
e
-
re
c
o
g
n
it
io
n
a
lg
o
rit
h
m
s,”
Ima
g
e
V
is.
Co
mp
u
t.
,
v
o
l.
1
6
,
n
o
.
5
,
p
p
.
2
9
5
–
3
0
6
,
A
p
r.
1
9
9
8
.
[2
5
]
D.
L
.
S
p
a
c
e
k
,
“
F
a
c
e
s9
4
a
f
a
c
e
re
c
o
g
n
it
io
n
d
a
tas
e
t,
”
2
0
0
7
.
[2
6
]
H.
-
W
.
Ng
a
n
d
S
.
W
in
k
ler,
“
A
d
a
ta
-
d
riv
e
n
a
p
p
r
o
a
c
h
t
o
c
lea
n
in
g
l
a
rg
e
f
a
c
e
d
a
tas
e
ts,”
in
2
0
1
4
I
EE
E
In
ter
n
a
ti
o
n
a
l
Co
n
fer
e
n
c
e
o
n
Ima
g
e
Pro
c
e
ss
in
g
(
ICIP)
,
2
0
1
4
,
p
p
.
3
4
3
–
3
4
7
.
[2
7
]
A
.
S
a
l
v
a
d
o
r,
X
.
G
iro
-
I
-
Nie
to
,
F
.
M
a
rq
u
e
s,
a
n
d
S
.
S
a
to
h
,
“
F
a
ste
r
R
-
CNN
F
e
a
tu
re
s
f
o
r
In
sta
n
c
e
S
e
a
rc
h
,
”
in
IEE
E
Co
mp
u
ter
S
o
c
iety
Co
n
fer
e
n
c
e
o
n
Co
mp
u
ter
V
isio
n
a
n
d
Pa
t
ter
n
Rec
o
g
n
it
io
n
W
o
rk
sh
o
p
s
,
2
0
1
6
,
p
p
.
3
9
4
–
4
0
1
.
[2
8
]
G
.
T
o
li
a
s,
R.
S
icre
,
a
n
d
H.
Jé
g
o
u
,
“
P
a
rti
c
u
lar
o
b
jec
t
re
tri
e
v
a
l
w
it
h
in
teg
ra
l
m
a
x
-
p
o
o
li
n
g
o
f
CNN
a
c
ti
v
a
ti
o
n
s,”
No
v
.
2
0
1
5
.
[2
9
]
N.
P
a
ss
a
li
s a
n
d
A
.
T
e
f
a
s,
“
S
p
a
ti
a
l
Ba
g
o
f
F
e
a
tu
re
s L
e
a
rn
in
g
f
o
r
L
a
r
g
e
S
c
a
le F
a
c
e
I
m
a
g
e
R
e
tri
e
v
a
l,
”
2
0
1
7
,
p
p
.
8
–
1
7
.
[3
0
]
N.
P
a
ss
a
li
s
a
n
d
A
.
T
e
fa
s,
“
L
e
a
rn
in
g
Ne
u
ra
l
Ba
g
-
of
-
F
e
a
tu
re
s
f
o
r
Larg
e
-
S
c
a
le
I
m
a
g
e
R
e
tri
e
v
a
l,
”
IEE
E
T
ra
n
s.
S
y
st.
M
a
n
,
Cy
b
e
rn
.
S
y
st.
,
v
o
l.
4
7
,
n
o
.
1
0
,
p
p
.
2
6
4
1
–
2
6
5
2
,
2
0
1
7
.
Evaluation Warning : The document was created with Spire.PDF for Python.