I
A
E
S
I
n
t
e
r
n
at
io
n
al
Jou
r
n
al
of
A
r
t
if
ic
ia
l
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
V
ol
.
11
, N
o.
1
,
M
a
r
c
h
2022
, pp.
1
02
~
1
09
I
S
S
N
:
2252
-
8938
,
D
O
I
:
10.11591/
ij
a
i.
v
11
.i
1
.pp
1
02
-
1
09
102
Jou
r
n
al
h
om
e
page
:
ht
tp
:
//
ij
ai
.
ia
e
s
c
or
e
.c
om
Im
age
an
d
vi
d
e
o
f
ac
e
r
e
t
r
i
e
val
w
i
t
h
q
u
e
r
y i
m
age
u
si
n
g
c
on
vol
u
t
i
on
al
n
e
u
r
al
n
e
t
w
or
k
f
e
at
u
r
e
s
I
m
an
e
H
ac
h
c
h
an
e
1
,
A
b
d
e
lm
aj
id
B
ad
r
i
1
, A
c
h
a S
ah
e
l
1
, I
lh
a
m
E
l
m
ou
r
ab
it
1
, Y
as
s
in
e
R
u
ic
h
e
k
2
1
L
a
bor
a
t
oi
r
e
d’
E
l
e
c
t
r
oni
que
, E
ne
r
gi
e
, A
ut
om
a
t
i
que
a
nd T
r
a
i
t
e
m
e
nt
de
l
’
I
nf
o
r
m
a
t
i
on (
E
E
A
a
nd T
I
)
, F
a
c
ul
t
de
s
S
c
i
e
nc
e
s
e
t
T
e
c
hni
que
s
M
oha
m
m
e
di
a
, U
ni
ve
r
s
i
t
H
a
s
s
a
n I
I
C
a
s
a
bl
a
nc
a
,
M
oha
m
m
e
di
a
,
M
or
oc
c
o
2
I
R
T
E
S
-
L
a
bor
a
t
oi
r
e
S
E
T
, U
ni
ve
r
s
i
t
de
T
e
c
hnol
ogi
e
de
B
e
l
f
or
t
M
ont
b
l
i
a
r
d, B
e
l
f
or
, F
r
a
nc
e
A
r
t
ic
le
I
n
f
o
A
B
S
T
R
A
C
T
A
r
ti
c
le
h
is
to
r
y
:
R
e
c
e
iv
e
d
M
a
y
7
, 2021
R
e
vi
s
e
d
D
e
c
2
2
, 2021
A
c
c
e
pt
e
d
D
e
c
29
,
2021
This
paper
addresses
the
issue
of
image
and
video
face
retrieval.
The
aim
of
this
work
is
to
be
able
to
retrieve
images
and/or
videos
of
specific
person
from a dataset
of images
and videos
if we have
a query i
mage of that
person.
The
methods
proposed
so
far
either
focus
on
images
or
videos
and
us
e
hand
crafted
features.
In
this
work
we
built
an
end
-
to
-
end
pipeline
for
both
image
and
video
face
retrieval
where
we
use
convolutional
neural
network
(CNN)
features
from
an
off
-
line
feature
extractor.
And
we
exploit
the
object
proposals
learned
by
a
region
proposal
network
(RPN)
in
the
online
filter
ing
and
re
-
ranking
steps.
Moreover,
we
study
the
impact
of
finetuni
ng
the
networks,
the
impact
of
sum
-
pooling
and
max
-
pooling,
and
the
im
pact
of
different similarity metrics. The
results that we
were able
to achieve ar
e very
promising.
K
e
y
w
o
r
d
s
:
C
la
s
s
if
ic
a
ti
on
C
onvolut
io
na
l
ne
ur
a
l
ne
twor
k
F
a
s
te
r
R
-
C
N
N
I
m
a
ge
a
nd vide
o r
e
tr
ie
va
l
I
m
a
ge
pr
oc
e
s
s
in
g
I
m
a
ge
t
o vi
de
o i
ns
ta
nc
e
r
e
tr
ie
va
l
O
bj
e
c
t
r
e
c
ogni
ti
on
This is an
open
acce
ss artic
le unde
r the
CC BY
-
SA
license.
C
or
r
e
s
pon
di
n
g A
u
th
or
:
I
m
a
ne
H
a
c
hc
ha
ne
L
a
bor
a
to
ir
e
d’
E
le
c
tr
oni
que
,
E
ne
r
gi
e
,
A
ut
om
a
ti
que
a
nd
T
r
a
it
e
m
e
nt
de
l’
I
nf
or
m
a
ti
on
(
E
E
A
a
nd
T
I
)
,
F
a
c
ul
t
de
s
S
c
ie
nc
e
s
e
t
T
e
c
hni
que
s
M
oha
m
m
e
di
a
, U
ni
v
e
r
s
it
H
a
s
s
a
n I
I
C
a
s
a
bl
a
n
c
a
M
oha
m
m
e
di
a
, M
or
oc
c
o
E
m
a
il
:
ha
c
hc
ha
ne
im
a
ne
@
gm
a
il
.c
om
1.
I
N
T
R
O
D
U
C
T
I
O
N
T
he
m
a
s
s
iv
e
a
dva
nc
e
s
in
in
te
r
ne
t
te
c
hnol
ogi
e
s
a
nd
th
e
pr
ol
if
e
r
a
ti
on
of
s
m
a
r
tp
hone
s
,
di
gi
ta
l
c
a
m
e
r
a
s
a
nd
s
to
r
a
ge
d
e
vi
c
e
s
le
d
to
a
n
in
c
r
e
a
s
e
in
th
e
popula
r
it
y
of
vi
s
u
a
l
s
e
a
r
c
h
a
ppl
ic
a
ti
ons
s
uc
h
a
s
im
a
g
e
r
e
tr
ie
va
l,
vi
de
o
r
e
tr
ie
va
l
or
pr
e
c
is
e
ly
in
s
ta
nc
e
s
e
a
r
c
h.
B
y
c
om
pa
r
in
g
a
que
r
y
a
ga
in
s
t
a
da
ta
ba
s
e
,
in
s
t
a
nc
e
s
e
a
r
c
h
is
u
s
e
d
to
e
xt
r
a
c
t
im
a
ge
s
or
vi
de
os
of
a
pa
r
ti
c
ul
a
r
obj
e
c
t
f
r
om
la
r
ge
da
t
a
ba
s
e
s
.
I
t
ha
s
be
e
n
c
om
m
onl
y
us
e
d
in
pr
oduc
t
r
e
c
ogni
ti
on, pr
ope
r
ty
i
de
nt
if
ic
a
ti
on, a
nd othe
r
a
ppl
ic
a
ti
ons
[
1]
–
[
3]
.
W
e
s
houl
d
not
e
th
a
t
in
one
ha
nd,
im
a
g
e
-
to
-
im
a
ge
r
e
tr
ie
va
l
is
a
w
e
ll
-
known
f
ie
ld
w
he
r
e
la
r
ge
-
s
c
a
l
e
f
a
c
e
im
a
ge
r
e
tr
ie
va
l
ha
s
r
e
c
e
nt
ly
a
tt
r
a
c
te
d
a
tt
e
nt
io
n,
a
nd
a
w
id
e
va
r
ie
ty
of
m
e
th
ods
ha
ve
be
e
n
pr
opos
e
d
f
or
f
a
c
e
r
e
c
ogni
ti
on
a
nd
r
e
tr
ie
va
l
[
4]
–
[7
]
.
F
ol
lo
w
in
g
pr
ope
r
a
da
pt
a
ti
on,
w
e
ll
-
known
te
c
hni
que
s
f
or
im
a
ge
r
e
tr
ie
va
l
w
e
r
e
us
e
d
f
or
f
a
c
e
r
e
c
ogni
ti
on/
r
e
tr
ie
va
l,
s
uc
h
a
s
b
a
g
-
of
-
vi
s
ua
l
w
or
ds
(
B
oV
W
)
.
O
th
e
r
r
e
c
e
nt
s
tu
di
e
s
us
e
d c
onvolut
io
na
l
ne
ur
a
l
ne
twor
k
(
C
N
N
)
f
or
t
he
f
e
a
tu
r
e
e
xt
r
a
c
ti
on t
a
s
k
[
6]
.
O
n
th
e
ot
he
r
h
a
nd,
im
a
ge
-
to
-
vi
de
o
r
e
tr
ie
va
l
[
8]
–
[
10]
is
a
n
a
s
ym
m
e
tr
ic
pr
obl
e
m
w
he
r
e
th
e
la
c
k
of
te
m
por
a
l
in
f
or
m
a
ti
on
in
im
a
ge
s
s
to
ps
us
f
r
om
us
in
g
s
ta
nda
r
d
te
c
hni
que
s
f
or
e
xt
r
a
c
ti
ng
vi
de
o
de
s
c
r
ip
to
r
s
[
11]
–
[
14]
.
T
r
a
di
ti
ona
ll
y,
im
a
ge
-
to
-
vi
de
o
r
e
tr
ie
va
l
te
c
hni
que
s
a
r
e
ba
s
e
d
on
a
c
la
s
s
ic
e
xt
r
a
c
ti
on
m
e
th
od
e
s
of
ha
nd
-
c
r
a
f
te
d f
e
a
tu
r
e
s
s
c
a
le
i
nva
r
ia
nt
f
e
a
tu
r
e
t
r
a
ns
f
or
m
(
S
I
F
T
)
[
15]
,
a
nd bina
r
y r
obus
t
in
de
pe
nde
nt
e
le
m
e
nt
a
r
y
f
e
a
tu
r
e
s
(
B
R
I
E
F
)
[
16]
.
S
m
a
ll
e
r
e
f
f
or
t
ha
s
be
e
n
m
a
de
to
a
d
a
pt
de
e
p
le
a
r
ni
ng
te
c
hni
que
s
.
W
e
c
a
n
a
ppl
y
s
ta
nda
r
d
f
e
a
tu
r
e
s
f
or
im
a
ge
r
e
tr
ie
va
l
[
17]
–
[
20
]
by
pr
oc
e
s
s
in
g
e
a
c
h
f
r
a
m
e
a
s
a
n
in
de
pe
nde
nt
im
a
ge
.
M
or
e
r
e
c
e
nt
w
or
ks
s
how
e
d t
ha
t
i
s
pos
s
ib
le
t
o us
e
C
N
N
f
or
f
e
a
tu
r
e
e
xt
r
a
c
ti
on w
he
n w
or
ki
ng on videos
[
21]
, [
22]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
I
m
age
and v
id
e
o f
ac
e
r
e
t
r
ie
v
al
w
it
h que
r
y
i
m
age
u
s
in
g
…
(
I
m
a
ne
H
ac
hc
hane
)
103
B
ut
not
m
uc
h
w
or
k
ha
s
be
e
n
done
in
c
om
bi
ni
ng
bot
h,
m
e
a
ni
n
g
ha
vi
ng
one
pi
pe
li
ne
f
or
bot
h
i
m
a
ge
r
e
tr
ie
va
l
a
nd
vi
de
o
r
e
tr
ie
va
l
us
in
g
one
que
r
y
im
a
ge
.
H
e
nc
e
,
in
th
is
pa
pe
r
,
w
e
in
ve
s
ti
ga
te
th
is
is
s
ue
.
W
e
a
r
e
tr
yi
ng
to
r
e
tr
ie
ve
th
e
t
op
N
m
os
t
r
e
le
va
nt
im
a
ge
s
a
nd/
or
vi
de
os
of
a
n
in
s
ta
nc
e
f
r
om
a
s
in
gl
e
im
a
ge
que
r
y
in
s
ta
nc
e
. M
or
e
s
pe
c
if
ic
a
ll
y, w
e
a
r
e
w
or
ki
ng on f
a
c
e
r
e
t
r
ie
va
l.
I
n ot
he
r
w
or
ds
, gi
vi
ng a
n
i
ns
ta
nc
e
of
a
f
a
c
e
i
n a
que
r
y
im
a
ge
,
w
e
a
r
e
tr
yi
ng
to
r
e
tr
ie
ve
th
e
to
p
N
m
os
t
r
e
le
v
a
nt
im
a
ge
in
s
ta
nc
e
s
a
nd/
or
vi
de
o
in
s
t
a
nc
e
s
f
r
om
our
da
ta
ba
s
e
of
vi
de
os
a
nd i
m
a
g
e
s
of
t
ha
t
s
p
e
c
if
ic
f
a
c
e
.
T
he
m
a
in
c
ont
r
ib
ut
io
n of
t
hi
s
pa
pe
r
i
s
t
o
bui
ld
a
n e
nd
-
to
-
e
nd p
i
pe
li
ne
, f
or
bot
h
i
m
a
ge
a
nd vide
o
f
a
c
e
r
e
tr
ie
va
l
us
in
g
one
que
r
y
im
a
ge
.
T
he
pi
pe
li
ne
ta
ke
s
a
dva
nt
a
ge
of
of
f
-
th
e
-
s
he
lf
a
nd
f
in
e
-
tu
ne
d
f
e
a
tu
r
e
s
f
r
o
m
a
n obje
c
t
de
te
c
ti
on
C
N
N
. W
e
t
e
s
te
d t
he
i
m
pa
c
t
of
m
ul
ti
pl
e
s
im
il
a
r
it
y m
e
tr
ic
s
, di
f
f
e
r
e
nt
ne
twor
k a
r
c
hi
te
c
tu
r
e
s
,
m
a
x
-
pool
in
g a
nd s
um
-
pool
in
g a
s
w
e
ll
a
s
t
he
i
m
pa
c
t
of
m
os
t
c
o
m
m
on r
e
r
a
nki
ng s
tr
a
te
gi
e
s
.
2.
R
E
L
A
T
E
D
WORK
V
is
ua
l
s
e
a
r
c
h
a
nd
r
e
tr
ie
va
l
a
r
e
in
ge
n
e
r
a
l
a
n
in
de
xi
ng
a
nd
que
r
yi
ng
pr
obl
e
m
f
or
vi
s
ua
l
da
ta
,
w
hi
c
h
c
a
n
be
f
ur
th
e
r
di
vi
de
d
in
to
c
a
te
gor
ie
s
de
pe
ndi
ng
on
th
e
que
r
y
ty
pe
a
nd
da
ta
ba
s
e
us
e
d. T
he
m
os
t
s
tu
di
e
d
f
ie
ld
in
vi
s
ua
l
r
e
tr
ie
va
l
is
im
a
ge
-
to
-
im
a
g
e
r
e
tr
ie
va
l,
w
he
r
e
w
e
us
e
a
que
r
y
im
a
ge
to
f
in
d
th
e
m
os
t
r
e
le
va
nt
im
a
ge
s
f
r
om
a
n i
m
a
ge
da
ta
s
e
t
[
23]
, [
24
]
.
G
e
ne
r
a
ll
y s
pe
a
ki
ng, vis
ua
l
s
e
a
r
c
h a
nd r
e
tr
ie
va
l
r
e
m
a
in
s
a
n i
s
s
ue
of
i
nde
xi
ng
a
nd
que
r
yi
ng
vi
s
ua
l
da
ta
.
T
hi
s
is
s
ue
c
a
n
b
e
c
a
te
gor
iz
e
d
de
pe
nd
in
g
on
th
e
ty
pe
of
que
r
ie
s
a
nd
da
ta
ba
s
e
s
u
s
e
d.
T
he
m
os
t
s
tu
di
e
d
a
r
e
a
in
vi
s
u
a
l
r
e
tr
ie
va
l
is
im
a
ge
-
to
-
im
a
ge
r
e
tr
ie
va
l,
w
e
r
e
a
w
e
u
s
e
a
qu
e
r
y
im
a
ge
to
r
e
tr
ie
ve
th
e
m
os
t
r
e
le
va
nt
im
a
ge
s
f
r
om
a
n
im
a
ge
da
t
a
s
e
t
[
23]
,
[
24]
.
A
not
he
r
a
r
e
a
of
vi
s
ua
l
r
e
tr
ie
va
l
is
vi
d
e
o
-
to
-
vi
de
o
r
e
tr
ie
va
l
w
he
r
e
a
que
r
y vi
de
o i
s
us
e
d t
o r
e
tr
ie
ve
r
e
le
va
nt
v
id
e
os
f
r
om
a
vi
de
o da
ta
s
e
t
[
25]
.
A
f
ur
th
e
r
va
r
ia
nt
i
s
vi
de
o
-
to
-
im
a
ge
r
e
tr
ie
va
l
in
w
hi
c
h
w
e
us
e
a
que
r
y
vi
de
o
to
s
e
a
r
c
h
a
da
ta
s
e
t
of
im
a
ge
s
[
26]
,
it
is
us
ua
ll
y
us
e
d
in
a
ugm
e
nt
e
d
r
e
a
li
ty
.
A
nd
of
c
our
s
e
w
e
h
a
ve
th
e
im
a
ge
-
to
-
vi
de
o
r
e
tr
ie
va
l
w
he
r
e
w
e
s
e
a
r
c
h
a
da
ta
ba
s
e
of
vi
de
os
us
in
g
a
que
r
y
im
a
ge
[
21
]
.
I
n
th
is
pa
pe
r
,
w
e
m
e
r
ge
two
of
th
os
e
a
r
e
a
s
:
I
m
a
ge
-
to
-
im
a
ge
r
e
tr
ie
va
l
a
nd
im
a
ge
-
to
-
vi
de
o r
e
tr
ie
va
l.
W
e
f
oc
us
on
bot
h
i
m
a
ge
a
nd vide
o r
e
t
r
ie
va
l
us
in
g one
que
r
y i
m
a
ge
. M
or
e
pr
e
c
is
e
ly
,
w
e
a
r
e
t
a
r
ge
ti
ng f
a
c
e
r
e
tr
ie
va
l.
M
e
a
ni
ng, givi
ng a
que
r
y f
a
c
e
i
m
a
ge
w
e
a
r
e
t
r
yi
ng t
o r
e
tr
ie
ve
t
he
m
os
t
r
e
le
va
nt
im
a
ge
s
a
nd/
or
vi
de
os
of
t
ha
t
s
pe
c
if
ic
f
a
c
e
.
F
a
c
e
r
e
tr
ie
va
l
is
a
di
f
f
ic
ul
t
ta
s
k
be
c
a
us
e
it
is
ha
r
d
to
a
da
pt
tr
a
di
ti
ona
l
im
a
ge
r
e
tr
ie
va
l
m
e
th
ode
s
(
li
ke
ba
g
of
w
or
ds
)
a
r
e
di
f
f
ic
ul
t
to
a
ppl
y
to
th
e
f
ie
ld
of
f
a
c
e
r
e
s
e
a
r
c
h
[
27]
.
B
e
c
a
us
e
th
e
tr
a
di
ti
ona
l
de
s
c
r
ip
to
r
ba
s
e
d
on
th
e
de
te
c
ti
on
of
ke
y
poi
nt
s
(
li
ke
S
I
F
T
)
of
te
n
f
a
il
s
due
to
th
e
s
m
oot
h
s
ur
f
a
c
e
of
th
e
f
a
c
e
.
P
r
e
vi
ous
w
or
k,
us
in
g
a
pr
e
vi
ou
s
ly
tr
a
in
e
d
im
a
ge
c
la
s
s
if
ic
a
ti
on
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
k
a
s
a
f
e
a
tu
r
e
e
xt
r
a
c
to
r
,
s
ho
w
e
d
th
a
t
it
is
m
or
e
a
ppr
opr
ia
te
to
us
e
a
f
ul
ly
c
onne
c
te
d
la
ye
r
f
or
im
a
ge
r
e
tr
ie
va
l
[
17]
.
R
a
z
a
vi
a
n
e
t
al
.
[
28]
I
m
pr
ove
d
r
e
s
ul
ts
by
c
om
bi
ni
ng
f
ul
ly
c
onne
c
te
d
la
y
e
r
s
e
xt
r
a
c
te
d
f
r
om
di
f
f
e
r
e
nt
im
a
ge
s
ubm
a
tc
he
s
.
L
a
te
r
,
th
e
ne
w
w
or
k
f
ound
th
a
t
th
e
c
onvolut
io
na
l
la
ye
r
is
s
ig
ni
f
ic
a
nt
ly
be
tt
e
r
th
a
n
th
e
f
ul
ly
c
onne
c
te
d
la
ye
r
in
im
a
ge
r
e
tr
ie
va
l
ta
s
ks
[
3]
, [
28]
.
W
he
n
w
or
ki
ng
on
im
a
ge
-
to
-
im
a
ge
r
e
tr
ie
va
l,
a
va
r
ie
ty
of
C
N
N
-
ba
s
e
d
obj
e
c
t
de
te
c
ti
on
pi
pe
li
ne
s
ha
ve
be
e
n
pr
opos
e
d.
I
n
th
is
pa
pe
r
,
w
e
a
r
e
in
te
r
e
s
te
d
in
F
a
s
te
r
R
-
C
N
N
[
29]
,
a
C
N
N
ne
twor
k
c
r
e
a
te
d
by
R
e
n
e
t
al
.
T
he
y
us
e
d
a
r
e
gi
on
pr
opos
a
l
ne
twor
k
(
R
P
N
)
[
30]
in
F
a
s
te
r
R
-
C
N
N
to
r
e
m
ove
th
e
d
e
pe
nde
nc
e
of
obj
e
c
t
pr
opos
it
io
ns
th
a
t
e
xi
s
ts
in
ol
de
r
C
N
N
obj
e
c
t
de
te
c
ti
on
s
ys
t
e
m
s
.
A
nd,
e
ve
n
th
ough
F
a
s
te
r
R
-
C
N
N
is
de
s
ig
n
e
d
to
de
te
c
t
ge
nr
a
l
obj
e
c
ts
,
J
i
a
ng
a
nd
L
e
a
r
ne
d
-
M
il
le
r
[
31]
w
e
r
e
a
bl
e
to
hi
ghl
ig
ht
it
s
im
pr
e
s
s
iv
e
f
a
c
e
de
te
c
ti
on
pe
r
f
or
m
a
nc
e
,
e
s
pe
c
ia
ll
y w
he
n
r
e
tr
a
in
e
d
on a
s
ui
ta
bl
e
f
a
c
e
de
t
e
c
ti
on
tr
a
in
in
g
s
e
t
[
6]
.
T
he
c
ur
r
e
nt
pi
pe
li
n
e
,
th
a
t
w
e
a
r
e
w
or
ki
ng
on,
us
e
s
of
f
-
th
e
-
s
he
lf
a
nd
f
in
e
ly
tu
ne
d
f
e
a
tu
r
e
s
of
F
a
s
te
r
R
-
C
N
N
'
s
e
nd
-
to
-
e
nd
obj
e
c
t
de
te
c
ti
on
a
r
c
hi
te
c
tu
r
e
to
e
xt
r
a
c
t
gl
oba
l
a
nd
lo
c
a
l
c
onvolut
io
na
l
f
e
a
tu
r
e
s
in
one
pa
s
s
a
nd
te
s
t
th
e
ir
ut
il
it
y
f
o
r
im
a
ge
a
nd
vi
de
o
f
a
c
e
r
e
tr
ie
va
l
us
in
g
one
que
r
y
f
a
c
e
im
a
ge
.
W
e
a
ls
o
te
s
t
th
e
im
pa
c
t
of
di
f
f
e
r
e
nt
s
im
il
a
r
it
y
m
e
tr
ic
s
, ne
twor
k a
r
c
hi
te
c
tu
r
e
s
, m
a
x
-
pool
in
g a
nd s
um
-
pool
in
g, a
s
w
e
ll
a
s
r
e
r
a
nki
ng
s
tr
a
te
gi
e
s
.
3.
M
E
T
H
O
D
O
L
O
G
Y
3.1. CN
N
-
b
as
e
d
r
e
p
r
e
s
e
n
t
at
io
n
s
I
n
our
ne
w
pi
pe
li
ne
,
F
ig
ur
e
1,
w
e
e
xa
m
in
e
th
e
im
por
ta
nc
e
of
us
in
g
lo
c
a
l
a
nd
gl
oba
l
C
N
N
f
e
a
tu
r
e
s
e
xt
r
a
c
te
d
f
r
om
pr
e
-
tr
a
in
e
d
F
a
s
te
r
R
-
C
N
N
m
ode
ls
[
29]
f
or
im
a
ge
a
nd
vi
de
o
f
a
c
e
r
e
tr
ie
va
l.
W
e
u
s
e
bounding
boxs
a
bove
our
que
r
y
im
a
ge
s
to
de
f
in
e
th
e
in
s
ta
n
c
e
s
th
a
t
w
e
a
r
e
lo
oki
ng
f
or
.
F
a
s
te
r
R
-
C
N
N
ha
d
two
m
a
jo
r
pa
r
ts
th
a
t
s
ha
r
e
a
c
onvolut
io
na
l
la
ye
r
.
T
he
f
i
r
s
t
one
is
R
P
N
;
it
is
a
s
m
a
ll
ne
ur
a
l
ne
twor
k
th
a
t
g
li
de
s
ove
r
th
e
la
s
t
f
e
a
tu
r
e
m
a
p
of
th
e
c
onvolut
io
n
la
ye
r
s
to
pr
e
di
c
t
w
he
th
e
r
a
n
obj
e
c
t
is
pr
e
s
e
nt
or
not
,
as
w
e
ll
as
th
e
bounding
box
of
th
os
e
obj
e
c
t
s
c
a
ll
e
d
w
in
dow
s
.
T
he
s
e
c
ond
o
ne
is
th
e
c
la
s
s
if
ie
r
th
a
t
le
a
r
ns
to
la
be
l
e
a
c
h
of
th
os
e
obj
e
c
ts
a
s
one
of
t
he
c
la
s
s
e
s
i
n t
he
l
e
a
r
ni
ng da
ta
s
e
t
[
3]
.
A
s
w
it
h
e
a
r
li
e
r
w
or
ks
[
3]
,
[
32]
,
a
nd
[
33]
our
obj
e
c
ti
ve
is
to
de
r
iv
e
a
c
om
pa
c
t
im
a
ge
r
e
pr
e
s
e
nt
a
ti
on
f
r
om
F
a
s
te
r
R
-
C
N
N
a
c
ti
va
ti
ons
.
W
e
c
on
s
tr
uc
t
th
e
gl
oba
l
de
s
c
r
ip
to
r
by
ig
nor
in
g
a
ll
of
F
a
s
te
r
R
-
C
N
N
'
s
l
a
ye
r
s
th
a
t
w
or
k
w
it
h
obj
e
c
t
pr
opos
it
io
ns
,
a
nd
w
e
d
e
r
iv
e
f
e
a
tu
r
e
s
f
r
om
th
e
la
s
t
c
onvolut
io
na
l
la
ye
r
.
T
a
ki
ng
th
e
e
xt
r
a
c
te
d
a
c
ti
va
ti
on
s
of
th
e
c
onvolut
io
n
la
ye
r
f
or
a
n
im
a
ge
or
a
f
r
a
m
e
in
to
c
on
s
id
e
r
a
ti
on,
w
e
gr
oup
th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
11
, N
o.
1
,
M
a
r
c
h 20
22
:
1
02
-
1
09
104
a
c
ti
va
ti
ons
of
e
a
c
h
f
il
te
r
to
f
or
m
a
n
im
a
ge
de
s
c
r
ip
to
r
w
it
h
th
e
s
a
m
e
di
m
e
ns
io
n
a
s
th
e
num
be
r
of
f
il
te
r
s
in
th
e
c
onvolut
io
n l
a
ye
r
.
W
he
n
w
or
ki
ng
on
c
ons
tr
a
c
ti
ng
th
e
lo
c
a
l
de
s
c
r
ip
to
r
,
th
e
r
e
gi
on
pool
in
g
la
ye
r
a
tt
a
c
he
d
to
th
e
la
s
t
c
onvolut
io
na
l
la
ye
r
is
us
e
d
to
e
xt
r
a
c
t
th
e
c
onvolut
io
na
l
a
c
t
iv
a
ti
ons
f
or
e
a
c
h
of
th
e
obj
e
c
t
pr
opos
it
io
n
s
ga
th
e
r
e
d
by
th
e
R
P
N
f
or
th
e
lo
c
a
l
de
s
c
r
ip
to
r
.
T
hi
s
pr
ovi
de
s
th
e
c
a
pa
bi
li
ty
of
c
r
e
a
ti
ng
a
lo
c
a
l
de
s
c
r
ip
to
r
f
or
e
ve
r
y
w
in
dow
pr
opos
a
l
by
a
ggr
e
ga
ti
ng
th
e
a
c
ti
va
ti
ons
of
th
a
t
w
in
dow
in
th
e
R
oI
pool
in
g
la
ye
r
.
S
um
-
pool
e
d
f
e
a
tu
r
e
s
a
r
e
l
2
-
nor
m
a
li
z
e
d i
n a
m
a
nne
r
s
im
il
a
r
t
o t
hos
e
de
s
c
r
ib
e
d by s
e
ve
r
a
l
ot
he
r
a
ut
hor
s
[
18]
, [
32]
, f
ol
lo
w
e
d
by w
hi
te
ni
ng a
nd a
s
e
c
ond r
ound of
l
2
-
nor
m
a
li
z
a
ti
on, while
m
a
x
-
pool
e
d f
e
a
tu
r
e
s
a
r
e
onl
y l
2
-
nor
m
a
li
z
e
d onc
e
w
it
hout
a
ny w
hi
te
ni
ng.
F
ig
ur
e
1. P
r
opos
e
d pi
pe
li
ne
’
s
a
r
c
hi
te
c
tu
r
e
3.2. Vid
e
o an
d
i
m
age
r
e
t
r
ie
val
T
he
f
e
a
tu
r
e
e
xt
r
a
c
ti
ng
is
done
of
f
li
ne
w
he
r
e
w
e
c
r
e
a
te
th
e
d
e
s
c
r
ip
to
r
s
f
or
th
e
im
a
g
e
s
,
th
e
vi
de
o
f
r
a
m
e
s
a
nd
th
e
que
r
y
im
a
ge
s
.
A
t
te
s
ti
ng
ti
m
e
(
th
e
onl
in
e
por
ti
on
of
th
e
pi
pe
li
ne
)
w
e
f
ol
lo
w
th
e
r
a
ki
ng
s
tr
a
te
gi
e
s
de
s
c
r
ib
e
d i
n t
hi
s
s
e
c
ti
on. W
e
s
t
a
r
t
w
it
h a
f
il
te
r
in
g s
te
p, w
he
r
e
t
he
que
r
y f
e
a
tu
r
e
s
a
r
e
c
om
pa
r
e
d t
o a
ll
th
e
da
ta
s
e
t
it
e
m
s
a
n
d
th
e
n
r
a
nke
d
us
in
g
a
s
im
il
a
r
it
y
m
e
a
s
ur
e
.
A
t
th
is
s
te
p,
w
e
a
r
e
s
ti
ll
c
ons
id
e
r
in
g
th
e
e
nt
ir
e
f
r
a
m
e
a
s
a
que
r
y.
A
f
te
r
th
e
f
il
te
r
in
g
s
te
p,
w
e
lo
c
a
ll
y
a
na
ly
z
e
a
nd
r
e
-
r
a
nk
th
e
N
uppe
r
e
le
m
e
nt
s
.
I
t
is
th
e
s
pa
ti
a
l
r
e
-
r
a
nki
ng.
L
a
s
t,
w
e
u
s
e
que
r
y
e
xpa
ns
io
n
(
Q
E
)
,
in
w
hi
c
h
w
e
c
om
bi
ne
th
e
d
e
s
c
r
ip
to
r
s
of
th
e
M
hi
ghe
r
e
le
m
e
nt
s
of
t
he
f
ir
s
t
r
a
nki
ng w
it
h t
he
que
r
y de
s
c
r
ip
to
r
t
o c
onduc
t
a
ne
w
s
e
a
r
c
h
(
M
=
5)
.
4.
E
X
P
E
R
I
M
E
N
T
S
4.1. Ut
il
iz
e
d
d
at
as
e
t
s
T
o
te
s
t
our
m
e
th
od
s
,
w
e
ne
e
d
to
u
s
e
a
da
ta
s
e
t
of
im
a
ge
s
a
nd
v
id
e
os
.
W
e
c
oul
d
not
f
in
d
one
,
s
o
w
e
de
c
id
e
d t
o m
e
r
ge
t
w
o e
xi
s
ti
ng one
s
. T
he
s
e
a
r
e
t
he
d
a
ta
s
e
t
s
w
e
u
s
e
d:
−
Y
ouT
ube
f
a
c
e
s
da
ta
ba
s
e
[
34]
:
T
h
e
da
t
a
s
e
t
c
ont
a
in
s
3,425
vi
d
e
os
of
1,595
pe
opl
e
,
a
ll
of
w
hi
c
h
w
e
r
e
dow
nl
oa
de
d
f
r
om
Y
ouT
ube
.
T
he
d
a
ta
ba
s
e
c
ont
a
in
s
a
n
a
ve
r
a
ge
of
2.15
vi
de
os
f
or
e
a
c
h
s
ubj
e
c
t,
w
it
h
48
f
r
a
m
e
s
be
in
g t
he
s
hor
te
s
t
c
li
p a
nd 6,070 f
r
a
m
e
s
b
e
in
g t
he
l
onge
s
t.
−
F
a
c
e
S
c
r
ub
[
35]
:
22,507
unc
ons
tr
a
in
e
d
f
a
c
e
im
a
ge
s
a
m
a
s
s
e
d
f
r
om
th
e
I
nt
e
r
ne
t.
W
e
a
dde
d
a
f
r
a
m
in
g
box
to
t
he
que
r
y i
m
a
ge
s
t
o s
ur
r
ounde
t
he
t
a
r
ge
t
f
a
c
e
s
.
T
he
da
ta
s
e
ts
w
e
us
e
d t
o f
in
e
-
tu
ne
t
he
ne
twor
k:
−
F
E
R
E
T
[
36]
:
T
hi
s
da
ta
s
e
t
ha
s
3,528
im
a
g
e
s
.
W
e
pr
ovi
de
a
f
r
a
m
in
g
box
to
th
e
que
r
y
im
a
ge
s
in
or
de
r
to
s
ur
r
ounding t
he
t
a
r
ge
t
f
a
c
e
s
.
−
F
A
C
E
S
94
[
37]
:
T
hi
s
da
ta
s
e
t
ha
s
2,809 im
a
g
e
s
.
W
e
a
ls
o
us
e
d
th
e
55,127
unc
ons
tr
a
in
e
d
f
a
c
e
im
a
ge
s
of
th
e
or
ig
in
a
l
F
a
c
e
S
c
r
ub
da
ta
s
e
t
to
f
in
e
-
tu
ne
th
e
ne
twor
k. W
he
n t
e
s
ti
ng, we
u
s
e
d 111 que
r
y i
m
a
ge
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
I
m
age
and v
id
e
o f
ac
e
r
e
t
r
ie
v
al
w
it
h que
r
y
i
m
age
u
s
in
g
…
(
I
m
a
ne
H
ac
hc
hane
)
105
4.2. E
xp
e
r
im
e
n
t
al
s
e
t
u
p
A
c
c
or
di
ng
to
pr
e
vi
ous
w
or
ks
[
3]
,
[
6]
,
[
21
]
de
e
pe
r
ne
twor
ks
a
c
hi
e
ve
d
be
tt
e
r
pe
r
f
or
m
a
nc
e
in
e
xt
r
a
c
ti
ng
gl
oba
l
a
nd
lo
c
a
l
f
e
a
tu
r
e
s
.
T
he
r
e
f
or
e
,
w
e
de
c
id
e
d
to
u
s
e
th
e
V
G
G
16
a
r
c
hi
te
c
tu
r
e
s
of
F
a
s
te
r
R
-
C
N
N
a
nd
c
om
p
a
r
e
it
w
it
h
th
e
Z
F
a
r
c
hi
te
c
tu
r
e
to
te
s
t
th
e
va
li
di
ty
of
th
e
th
e
or
y
w
he
n
w
or
ki
ng
on
bot
h
i
m
a
ge
a
nd
vi
de
o
r
e
tr
ie
va
l.
W
he
n
w
or
ki
ng
w
it
h
th
e
V
G
G
16
a
r
c
hi
te
c
tu
r
e
,
th
e
gl
oba
l
de
s
c
r
ip
to
r
s
a
r
e
e
xt
r
a
c
te
d
f
r
om
th
e
la
s
t
c
onvolut
io
n
la
ye
r
“
c
onv5_3”
a
nd
a
r
e
of
di
m
e
ns
io
n
512.
A
nd
w
he
n
w
or
ki
ng
w
it
h
th
e
Z
F
a
r
c
hi
te
c
tu
r
e
,
th
e
gl
oba
l
de
s
c
r
ip
to
r
s
a
r
e
e
xt
r
a
c
te
d
f
r
om
th
e
la
s
t
c
onvolut
io
n
la
y
e
r
“
c
onv5”
a
nd
a
r
e
of
di
m
e
ns
io
n
256.
F
or
th
e
lo
c
a
l
f
e
a
tu
r
e
s
,
w
e
gr
oup
th
e
m
f
r
om
th
e
F
a
s
te
r
R
-
C
N
N
R
oI
c
l
us
te
r
in
g
la
ye
r
.
T
he
gl
oba
l
de
s
c
r
ip
to
r
s
f
or
th
e
V
G
G
16 a
r
c
hi
te
c
tu
r
e
a
r
e
e
xt
r
a
c
te
d f
r
om
t
he
l
a
s
t
c
onvolut
io
n
l
a
y
e
r
“
c
onv5_3
”
a
nd a
r
e
of
di
m
e
ns
io
n 512, while
th
e
gl
oba
l
de
s
c
r
ip
to
r
s
f
or
th
e
Z
F
a
r
c
hi
te
c
tu
r
e
a
r
e
e
xt
r
a
c
te
d
f
r
om
th
e
la
s
t
c
onvolut
io
n
la
ye
r
“
c
onv5
”
a
nd
a
r
e
of
di
m
e
ns
io
n 256. W
e
gr
oup loc
a
l
f
e
a
tu
r
e
s
u
s
in
g t
he
F
a
s
te
r
R
-
C
N
N
r
e
gi
on of
i
nt
e
r
e
s
t
(
R
oI
)
c
lu
s
te
r
in
g l
a
ye
r
.
W
e
a
ls
o
e
xpe
r
im
e
nt
e
d
w
it
h
w
id
e
ly
us
e
d
s
im
il
a
r
it
y
m
e
tr
ic
s
to
s
e
e
w
hi
c
h
one
is
m
or
e
s
ui
ta
bl
e
f
or
ou
r
pi
pe
li
ne
.
W
e
te
s
te
d
th
e
f
ol
lo
w
in
g
s
im
il
a
r
it
y
m
e
tr
ic
s
:
C
o
s
in
e
s
im
il
a
r
it
y
m
e
tr
ic
,
E
uc
li
di
e
n
s
im
il
a
r
it
y
m
e
tr
ic
,
M
a
nha
ta
n
s
im
il
a
r
it
y
m
e
tr
ic
,
C
he
byc
he
v
s
im
il
a
r
it
y
m
e
tr
ic
,
M
in
kow
s
ki
s
im
il
a
r
it
y
m
e
tr
ic
,
C
a
nbe
r
r
a
s
im
il
a
r
it
y
m
e
tr
ic
,
a
nd
C
or
r
ol
a
ti
on
s
im
il
a
r
it
y
m
e
tr
ic
.
T
he
f
ol
lo
w
in
g
s
p
e
c
if
ic
a
ti
ons
w
e
r
e
us
e
d
f
or
th
e
e
xp
e
r
im
e
nt
s
:
P
r
oc
e
s
s
or
:
I
nt
e
l(
R
)
C
or
e
(
T
M
)
i7
-
7700K
C
P
U
4.20
G
H
z
,
R
A
M
:
16
G
B
,
O
S
:
U
bunt
u
16.04,
G
r
a
phi
c
s
c
a
r
d:
N
V
I
D
I
A
G
e
F
or
c
e
G
T
X
1070.
W
e
s
houl
d
not
e
th
e
e
xt
r
a
c
ti
on
ti
m
e
f
or
th
e
V
G
G
16
r
e
qui
r
e
d
a
n
a
ve
r
a
ge
of
16h
11m
in
30
s
e
c
ond
s
c
om
pa
r
e
d
to
a
n
a
ve
r
a
ge
of
7h
34m
in
a
nd
22
s
e
c
onds
w
he
n
us
i
ng
Z
F
.
T
hi
s
ti
m
e
di
f
f
e
r
e
nc
e
c
a
n
be
e
xpl
a
in
e
d
by
th
e
s
iz
e
s
of
th
e
ne
twor
ks
. T
he
r
a
nki
ng
to
ok
on
a
ve
r
a
ge
2
s
e
c
onds
pe
r
que
r
y
im
a
ge
;
th
e
r
e
-
r
a
nki
ng
to
ok
a
n
a
ve
r
a
ge
of
16 s
e
c
onds
p
e
r
que
r
y i
m
a
ge
, a
nd w
he
n us
in
g t
he
Q
E
, t
he
r
e
-
r
a
nki
ng t
ook a
n a
ve
r
a
ge
of
17 s
e
c
o
nds
pe
r
que
r
y i
m
a
ge
.
4.3. Of
f
-
t
h
e
-
s
h
e
lf
C
N
N
f
e
at
u
r
e
s
I
n
th
is
s
e
c
ti
on
w
e
e
va
lu
a
te
u
s
in
g
F
a
s
te
r
R
-
C
N
N
f
e
a
tu
r
e
s
f
or
f
a
c
e
im
a
ge
a
nd/
or
vi
de
o
r
e
tr
ie
va
l.
W
e
ha
ve
te
s
te
d
di
f
f
e
r
e
nt
s
im
il
a
r
it
y
m
e
tr
ic
s
,
as
de
t
a
il
e
d
a
bove
. T
he
r
e
s
ul
ts
,
di
s
pl
a
ye
d
in
T
a
bl
e
1,
w
e
r
e
s
im
il
a
r
a
nd
c
lo
s
e
,
but
th
e
be
s
t
r
e
s
ul
ts
w
e
r
e
obt
a
in
e
d
u
s
in
g
th
e
c
o
s
in
e
a
nd
t
he
e
uc
li
di
e
n
s
im
il
a
r
it
y
m
e
tr
ic
s
c
om
bi
ne
d
w
it
h
our
r
e
-
r
a
nki
ng s
tr
a
te
gi
e
s
w
it
h a
pr
e
c
is
io
n of
55.4%
. B
ut
w
it
h t
h
e
ot
he
r
s
im
il
a
r
it
y m
e
tr
ic
s
, t
he
que
r
y e
xpa
ns
io
n
a
nd t
he
s
pa
ti
a
l
r
e
r
a
nki
ng did not
i
m
pr
ove
t
he
r
e
s
ul
ts
.
M
or
e
ove
r
,
a
c
om
pa
r
a
ti
ve
s
tu
dy
of
th
e
s
um
a
nd
m
a
x
pool
in
g
s
tr
a
te
gi
e
s
of
im
a
g
e
-
w
is
e
a
nd
r
e
gi
on
-
w
is
e
de
s
c
r
ip
to
r
s
w
a
s
a
ls
o
c
onduc
te
d,
w
it
h
th
e
r
e
s
ul
t
s
s
um
m
a
r
iz
e
d
in
T
a
bl
e
1.
S
um
-
pool
in
g
is
be
tt
e
r
th
a
n
m
a
x
-
pool
in
g,
a
c
c
or
di
n
g
to
our
te
s
ts
.
I
t
a
ls
o
c
onf
ir
m
s
th
a
t
F
a
s
t
e
r
R
-
C
N
N
w
it
h
a
V
G
G
16
a
r
c
hi
te
c
tu
r
e
tr
a
in
e
d
on
pa
s
c
a
l
V
O
C
d
a
ta
s
e
ts
pe
r
f
or
m
e
d
be
s
t,
w
hi
c
h
i
s
c
on
s
is
te
nt
w
it
h
pr
e
vi
ous
r
e
s
e
a
r
c
h
th
a
t
ha
d
de
m
ons
tr
a
te
d
th
a
t
de
e
p ne
twor
ks
c
oul
d de
li
ve
r
be
tt
e
r
r
e
s
ul
ts
w
he
n e
xt
r
a
c
ti
ng
gl
o
ba
l
a
nd l
oc
a
l
f
e
a
tu
r
e
s
.
4.4. F
in
e
-
t
u
n
in
g t
h
e
C
N
N
M
or
e
im
por
ta
nt
ly
,
w
e
in
ve
s
ti
ga
te
d
th
e
e
f
f
e
c
ts
of
f
in
e
-
tu
ni
ng
a
pr
e
-
tr
a
in
e
d
ne
twor
k
on
r
e
c
ove
r
y
pe
r
f
or
m
a
nc
e
w
it
h t
he
que
r
y obje
c
ts
t
o r
e
tr
ie
ve
. W
e
us
e
d t
he
m
o
de
l
V
G
G
16 of
F
a
s
te
r
R
-
C
N
N
pr
e
-
tr
a
in
e
d w
it
h
th
e
pa
s
c
a
l
V
O
C
obj
e
c
ts
.
W
e
r
e
f
in
e
d i
t
us
in
g t
w
o da
ta
s
e
ts
:
−
W
e
r
e
f
in
e
d
th
e
f
ir
s
t
ne
twor
k
w
it
h
F
E
R
E
T
a
nd
F
a
c
e
s
9
4
da
ta
s
e
t
s
a
nd
w
e
c
a
ll
e
d
it
V
G
G
16
(
F
e
r
e
t
a
nd F
a
c
e
s
94)
. B
e
c
a
us
e
of
t
he
ir
s
m
a
ll
s
iz
e
, t
he
F
e
r
e
t
a
nd t
he
F
a
c
e
s
94 da
ta
s
e
t
s
w
e
r
e
c
om
bi
ne
d, a
nd
th
e
ne
twor
k’
s
out
put
la
ye
r
w
a
s
m
odi
f
ie
d
to
r
e
tu
r
n
422
c
la
s
s
pr
oba
bi
li
ti
e
s
a
nd
th
e
ir
c
or
r
e
s
ponding
bounding
box
c
oor
di
na
te
s
[
6]
(
th
e
422
c
ount
s
f
or
th
e
269
c
la
s
s
e
s
in
th
e
F
E
R
E
T
da
ta
s
e
t
a
nd
th
e
152
c
la
s
s
e
s
i
n t
he
F
a
c
e
s
94 da
ta
s
e
t,
pl
us
one
a
ddi
ti
ona
l
c
la
s
s
f
or
t
he
ba
c
kgr
ound)
.
−
W
e
r
e
f
in
e
d
th
e
s
e
c
ond
ne
twor
k
w
it
h
u
s
in
g
th
e
F
a
c
e
S
c
r
ub
da
ta
s
e
t.
W
e
c
a
ll
e
d
it
V
G
G
16
(
F
a
c
e
s
c
r
ub)
.
F
or
th
is
ne
twor
k
th
e
out
put
la
ye
r
w
a
s
m
odi
f
ie
d
to
r
e
tu
r
n
530
c
la
s
s
pr
oba
bi
li
ti
e
s
a
nd
th
e
ir
c
or
r
e
s
ponding
bounding box coor
di
na
te
s
(
530 c
la
s
s
e
s
, pl
us
on
e
a
ddi
ti
ona
l
c
la
s
s
f
or
t
he
ba
c
kgr
ound)
.
T
he
in
it
ia
l
pa
r
a
m
e
t
e
r
s
of
F
a
s
te
r
R
-
C
N
N
a
s
d
e
s
c
r
ib
e
d
in
[
19]
di
d
not
c
ha
nge
,
but
due
to
a
r
e
duc
e
d
num
be
r
of
tr
a
in
in
g
s
a
m
pl
e
s
,
th
e
num
be
r
of
it
e
r
a
ti
ons
w
a
s
r
e
duc
e
d
f
r
om
80,000
to
20,000.
W
e
us
e
th
e
r
e
f
in
e
d
ne
twor
ks
of
th
e
tu
ni
ng
s
tr
a
te
gy
(
V
G
G
16
(
F
e
r
e
t
a
nd
F
a
c
e
s
94)
a
nd
V
G
G
16
(
F
a
c
e
s
c
r
ub)
)
on
our
im
a
ge
a
nd
vi
de
o
da
ta
s
e
t
to
e
xt
r
a
c
t
th
e
de
s
c
r
ip
to
r
s
a
nd
pe
r
f
or
m
im
a
ge
a
nd
vi
de
o
f
a
c
e
r
e
tr
ie
va
l.
T
hos
e
r
e
s
ul
ts
a
r
e
pr
e
s
e
nt
e
d
in
T
a
bl
e
2.
T
hi
s
ti
m
e
th
e
M
a
nha
tt
a
n
s
im
il
a
r
it
y
m
e
tr
ic
,
a
ls
o
c
a
ll
e
d
c
it
y
bl
oc
k,
pr
oduc
e
d
th
e
be
s
t
r
e
s
ul
ts
.
W
e
s
houl
d
a
ls
o
not
e
th
a
t
th
e
que
r
y
e
xpa
n
s
io
n
a
nd
s
p
a
ti
a
l
r
e
r
a
nki
ng
s
li
ght
ly
im
pr
ove
d
th
e
r
e
s
ul
ts
.
W
he
n
c
om
pa
r
in
g
th
e
s
um
-
pool
in
g
s
tr
a
te
gi
e
to
th
e
m
a
x
-
pool
in
g
s
tr
a
te
gi
e
of
th
e
im
a
ge
-
w
is
e
a
nd
r
e
gi
on
-
w
is
e
de
s
c
r
ip
to
r
s
,
s
um
-
pool
in
g
ga
ve
be
tt
e
r
r
e
s
ul
ts
th
a
n
m
a
x
-
pool
in
g
w
it
h
m
os
t
s
im
il
a
r
it
y
m
e
tr
ic
s
.
B
ut
m
a
x
-
poo
li
ng
ga
ve
t
he
be
s
t
r
e
s
ul
ts
w
he
n u
s
e
d w
it
h t
he
M
a
nh
a
tt
a
n s
im
il
a
r
it
y m
e
tr
ic
w
it
h a
n a
c
c
ur
a
c
y of
76.2%
.
W
e
a
ls
o
c
om
pa
r
e
d
di
f
f
e
r
e
nt
F
a
s
te
r
R
-
C
N
N
a
r
c
hi
te
c
tu
r
e
s
tr
a
in
e
d
on
di
f
f
e
r
e
nt
da
ta
s
e
ts
.
W
e
de
te
r
m
in
e
d
th
a
t
de
e
pe
r
ne
twor
ks
ga
ve
be
tt
e
r
r
e
s
ul
ts
,
w
hi
c
h
is
c
ons
i
s
te
nt
w
it
h
th
e
li
te
r
a
tu
r
e
.
W
e
a
ls
o
not
ic
e
d
th
e
da
ta
s
e
ts
, on whic
h t
he
ne
twor
k w
a
s
pr
e
vi
ous
ly
t
r
a
in
e
d, ha
d
th
e
m
os
t
im
pa
c
t
on t
he
r
e
s
ul
ts
. A
s
w
e
c
a
n s
e
e
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
11
, N
o.
1
,
M
a
r
c
h 20
22
:
1
02
-
1
09
106
w
he
n
w
or
ki
ng
w
it
h
of
f
-
th
e
-
s
he
lf
ne
twor
ks
,
th
e
n
e
twor
ks
tr
a
in
e
d
on
pa
s
c
a
l
V
O
C
g
a
ve
a
ve
r
a
ge
r
e
s
ul
ts
.
B
ut
th
e
be
s
t
r
e
s
ul
ts
w
e
r
e
obt
a
in
e
d w
he
n w
or
ki
ng w
it
h t
he
ne
twor
ks
t
r
a
in
e
d f
or
f
a
c
e
c
la
s
s
if
ic
a
ti
on, me
a
ni
ng t
r
a
in
e
d on
F
a
s
e
c
r
ub
a
nd
F
e
r
e
t
a
nd
F
a
c
e
s
94
in
our
c
a
s
e
.
O
n
th
a
t
a
c
c
ount
,
t
he
V
G
G
16
tr
a
in
e
d
on
F
a
c
e
s
c
r
ub
ga
ve
th
e
be
s
t
r
e
s
ul
ts
be
c
a
u
s
e
th
e
na
tu
r
e
of
th
e
ph
ot
o
s
in
th
is
da
ta
s
e
t
is
m
or
e
s
im
il
a
r
to
th
e
da
ta
s
e
t
th
a
t
w
e
a
r
e
w
or
ki
ng
on.
F
e
r
e
t
a
nd
F
a
c
e
s
94
im
a
ge
s
w
e
r
e
ta
ke
n
in
a
c
ont
r
ol
le
d
e
nvi
r
on
m
e
nt
,
but
F
a
s
e
c
r
ub
im
a
ge
s
w
e
r
e
a
m
a
s
s
e
d
f
r
om
th
e
w
e
b
a
nd
s
how
c
a
s
e
th
e
s
ubj
e
c
t
in
di
f
f
e
r
e
nt
pos
it
io
ns
w
it
h
d
if
f
e
r
e
nt
li
ght
i
ng
s
e
tu
ps
a
nd
f
a
c
ia
l
e
xpr
e
s
s
io
n
s
w
hi
c
h
is
c
lo
s
e
s
t
to
w
ha
t
vi
de
o
s
c
a
n
be
.
T
ha
t
is
w
hy
th
e
V
G
G
16
tr
a
in
e
d
on
F
a
c
e
s
c
r
ub
ga
ve
th
e
b
e
s
t
r
e
s
ul
ts
w
he
n
us
e
d
f
or
r
e
tr
ie
vi
ng
f
a
c
e
im
a
ge
s
a
nd
vi
de
os
f
r
om
a
da
ta
s
e
t
of
im
a
ge
s
a
nd
vi
de
os
us
in
g
one
que
r
y
im
a
ge
w
it
h a
pr
e
c
i
s
io
n of
76.2%
. S
o, w
e
w
e
r
e
a
bl
e
t
o i
m
pr
ove
t
he
r
e
s
ul
ts
w
it
h 13.7%
.
T
a
bl
e
1. M
e
a
n a
ve
r
a
g
e
pr
e
c
is
io
n (
m
A
P
)
of
pr
e
-
tr
a
in
e
d F
a
s
te
r
R
-
C
N
N
m
ode
ls
t
r
a
in
e
d w
it
h m
ic
r
os
of
t
C
O
C
O
or
pa
s
c
a
l
V
O
C
M
e
t
r
i
c
s
M
ode
l
s
P
ool
i
ng
R
a
nki
ng
Re
-
r
a
nki
ng
QE
C
os
i
ne
s
i
m
i
l
a
r
i
t
y m
e
t
r
i
c
V
G
G
16 (
P
a
s
c
a
l
V
O
C
)
s
um
0.551
0.551
0.554
m
a
x
0.538
0.545
0.544
V
G
G
16 (
M
i
c
r
os
of
t
C
O
C
O
)
s
um
0.545
0.521
0.516
m
a
x
0.524
0.525
0.522
Z
F
(
P
a
s
c
a
l
V
O
C
)
s
um
0.550
0.539
0.538
m
a
x
0.534
0.544
0.540
E
uc
l
i
di
e
n
s
i
m
i
l
a
r
i
t
y m
e
t
r
i
c
V
G
G
16 (
P
a
s
c
a
l
V
O
C
)
s
um
0.551
0.551
0.554
m
a
x
0.538
0.545
0.544
V
G
G
16 (
M
i
c
r
os
of
t
C
O
C
O
)
s
um
0.545
0.521
0.516
m
a
x
0.524
0.525
0.522
Z
F
(
P
a
s
c
a
l
V
O
C
)
s
um
0.550
0.539
0.538
m
a
x
0.534
0.544
0.540
M
a
nha
t
a
n
s
i
m
i
l
a
r
i
t
y m
e
t
r
i
c
V
G
G
16 (
P
a
s
c
a
l
V
O
C
)
s
um
0.550
0.550
0.545
m
a
x
0.540
0.543
0.538
V
G
G
16 (
M
i
c
r
os
of
t
C
O
C
O
)
s
um
0.543
0.513
0.507
m
a
x
0.527
0.529
0.526
Z
F
(
P
a
s
c
a
l
V
O
C
)
s
um
0.547
0.535
0.530
m
a
x
0.538
0.549
0.546
C
he
byc
he
v
s
i
m
i
l
a
r
i
t
y
m
e
t
r
i
c
V
G
G
16 (
P
a
s
c
a
l
V
O
C
)
s
um
0.497
0.482
0.493
m
a
x
0.470
0.451
0.469
V
G
G
16 (
M
i
c
r
os
of
t
C
O
C
O
)
s
um
0.513
0.465
0.487
m
a
x
0.488
0.437
0.453
Z
F
(
P
a
s
c
a
l
V
O
C
)
s
um
0.518
0.515
0.517
m
a
x
0.499
0.459
0.490
M
i
nkow
s
ki
s
i
m
i
l
a
r
i
t
y
m
e
t
r
i
c
V
G
G
16 (
P
a
s
c
a
l
V
O
C
)
s
um
0.551
0.551
0.544
m
a
x
0.538
0.545
0.544
V
G
G
16 (
M
i
c
r
os
of
t
C
O
C
O
)
s
um
0.545
0.521
0.516
m
a
x
0.524
0.525
0.522
Z
F
(
P
a
s
c
a
l
V
O
C
)
s
um
0.550
0.544
0.536
m
a
x
0.534
0.544
0.540
C
a
nbe
r
r
a
s
i
m
i
l
a
r
i
t
y m
e
t
r
i
c
V
G
G
16 (
P
a
s
c
a
l
V
O
C
)
s
um
0.547
0.544
0.539
m
a
x
0.528
0.516
0.518
V
G
G
16 (
M
i
c
r
os
of
t
C
O
C
O
)
s
um
0.538
0.516
0.512
m
a
x
0.526
0.524
0.524
Z
F
(
P
a
s
c
a
l
V
O
C
)
s
um
0.540
0.538
0.537
m
a
x
0.524
0.530
0.521
C
or
r
ol
a
t
i
on s
i
m
i
l
a
r
i
t
y
m
e
t
r
i
c
V
G
G
16 (
P
a
s
c
a
l
V
O
C
)
s
um
0.551
0.551
0.544
m
a
x
0.539
0.549
0.548
V
G
G
16 (
M
i
c
r
os
of
t
C
O
C
O
)
s
um
0.545
0.520
0.524
m
a
x
0.524
0.522
0.517
Z
F
(
P
a
s
c
a
l
V
O
C
)
s
um
0.549
0.544
0.545
m
a
x
0.537
0.542
0.537
4.5. Com
p
ar
is
on
I
n
th
is
s
e
c
ti
on
w
e
pr
e
s
e
nt
a
c
om
pa
r
a
ti
ve
s
tu
dy
be
twe
e
n
our
r
e
s
ul
ts
a
nd
ot
he
r
r
e
s
ul
ts
obt
a
in
e
d
us
in
g
f
is
he
r
ve
c
to
r
(
F
V
)
a
nd
ba
g
of
vi
s
ua
l
w
or
d
(
B
O
V
W
)
.
W
he
n
w
or
ki
ng
on
vi
de
o
r
e
tr
ie
va
l
a
nd
im
a
ge
a
nd
v
id
e
o
r
e
tr
ie
va
l,
our
pi
pe
li
ne
,
w
hi
c
h
ut
il
iz
e
s
r
a
w
f
a
s
te
r
R
-
C
N
N
f
e
a
t
ur
e
s
,
out
pe
r
f
or
m
e
d
a
ll
ot
he
r
te
c
hni
qu
e
s
.
T
he
r
e
s
ul
ts
a
r
e
di
s
pl
a
ye
d i
n
T
a
bl
e
3.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
I
m
age
and v
id
e
o f
ac
e
r
e
t
r
ie
v
al
w
it
h que
r
y
i
m
age
u
s
in
g
…
(
I
m
a
ne
H
ac
hc
hane
)
107
T
a
bl
e
2. M
e
a
n a
ve
r
a
g
e
pr
e
c
is
io
n (
m
A
P
)
of
t
he
f
in
e
-
tu
ne
d F
a
s
te
r
R
-
C
N
N
m
ode
ls
w
it
h V
G
G
16 a
r
c
hi
te
c
tu
r
e
s
f
in
e
-
tu
ne
d w
it
h F
a
c
e
s
c
r
ub or
F
e
r
e
t
a
nd F
a
c
e
s
9 r
e
s
p
e
c
ti
ve
ly
M
e
t
r
i
c
s
M
ode
l
s
P
ool
i
ng
R
a
nki
ng
Re
-
r
a
nki
ng
QE
C
os
i
ne
s
i
m
i
l
a
r
i
t
y m
e
t
r
i
c
V
G
G
16(
F
a
c
e
s
c
r
ub)
.
s
um
0.757
0.737
0.706
m
a
x
0.738
0.731
0.756
V
G
G
16(
F
e
r
e
t
a
nd F
a
c
e
s
94)
s
um
0.577
0.570
0.563
m
a
x
0.554
0.564
0.572
E
uc
l
i
di
e
n s
i
m
i
l
a
r
i
t
y m
e
t
r
i
c
V
G
G
16(
F
a
c
e
s
c
r
ub)
.
s
um
0.757
0.737
0.706
m
a
x
0.738
0.731
0.756
V
G
G
16(
F
e
r
e
t
a
nd F
a
c
e
s
94)
s
um
0.577
0.570
0.563
m
a
x
0.554
0.564
0.572
M
a
nha
t
a
n s
i
m
i
l
a
r
i
t
y m
e
t
r
i
c
V
G
G
16(
F
a
c
e
s
c
r
ub)
.
s
um
0.738
0.695
0.734
m
a
x
0.750
0.746
0.762
V
G
G
16(
F
e
r
e
t
a
nd F
a
c
e
s
94)
s
um
0.565
0.561
0.553
m
a
x
0.562
0.573
0.580
C
he
byc
he
v s
i
m
i
l
a
r
i
t
y m
e
t
r
i
c
V
G
G
16(
F
a
c
e
s
c
r
ub)
.
s
um
0.545
0.555
0.562
m
a
x
0.564
0.579
0.605
V
G
G
16(
F
e
r
e
t
a
nd F
a
c
e
s
94)
s
um
0.504
0.513
0.514
m
a
x
0.495
0.501
0.500
M
i
nkow
s
ki
s
i
m
i
l
a
r
i
t
y m
e
t
r
i
c
V
G
G
16(
F
a
c
e
s
c
r
ub)
.
s
um
0.757
0.727
0.747
m
a
x
0.738
0.731
0.756
V
G
G
16(
F
e
r
e
t
a
nd F
a
c
e
s
94)
s
um
0.577
0.570
0.560
m
a
x
0.554
0.564
0.572
C
a
nbe
r
r
a
s
i
m
i
l
a
r
i
t
y m
e
t
r
i
c
V
G
G
16(
F
a
c
e
s
c
r
ub)
.
s
um
0.742
0.742
0.760
m
a
x
0.723
0.731
0.737
V
G
G
16(
F
e
r
e
t
a
nd F
a
c
e
s
94)
s
um
0.567
0.569
0.568
m
a
x
0.556
0.558
0.552
C
or
r
ol
a
t
i
on s
i
m
i
l
a
r
i
t
y m
e
t
r
i
c
V
G
G
16(
F
a
c
e
s
c
r
ub)
.
s
um
0.757
0.728
0.749
m
a
x
0.741
0.731
0.748
V
G
G
16(
F
e
r
e
t
a
nd F
a
c
e
s
94)
s
um
0.577
0.570
0.563
m
a
x
0.557
0.568
0.573
T
a
bl
e
3. C
om
pa
r
a
ti
ve
s
tu
dy w
it
h
ot
he
r
t
e
c
hni
que
s
.
R
e
s
ul
t
s
pr
ovi
de
d a
s
m
A
P
M
e
t
hod
Y
ouT
ube
F
a
c
e
s
D
a
t
a
ba
s
e
+F
a
c
e
s
c
r
ub
(
a
n i
m
a
ge
a
nd vi
de
o da
t
a
s
e
t
)
Y
ouT
ube
F
a
c
e
s
D
a
t
a
ba
s
e
(
a
vi
de
o da
t
a
s
e
t
)
F
E
R
E
T
(
a
n i
m
a
ge
da
t
a
s
e
t
)
O
ur
pi
pe
l
i
ne
0.762
0.903
0.8913
F
a
s
t
e
r
R
-
C
N
N
f
e
a
t
ur
e
s
+F
V
[
21]
0.006
0.006
-
F
a
s
t
e
r
R
-
C
N
N
f
e
a
t
ur
e
s
+B
O
V
W
[
21]
-
0.001
-
L
og I
C
A
I
I
+K
N
N
[
38]
-
-
0.3553
L
og I
C
A
I
+K
N
N
[
38]
-
-
0.3608
L
G
H
P
de
s
c
r
i
pt
or
[
7]
-
-
0.5460
5.
C
O
N
C
L
U
S
I
O
N
I
n
th
is
pa
pe
r
,
w
e
de
m
ons
tr
a
te
how
to
us
e
C
N
N
f
e
a
tu
r
e
s
f
r
om
a
n
obj
e
c
t
de
te
c
ti
on
ne
twor
k
f
o
r
im
a
ge
a
nd
vi
de
o
f
a
c
e
r
e
tr
ie
va
l
u
s
in
g
one
qu
e
r
y
im
a
ge
.
W
e
us
e
d
F
a
s
te
r
R
-
C
N
N
f
e
a
tu
r
e
s
a
s
our
gl
oba
l
a
nd
lo
c
a
l
de
s
c
r
ip
to
r
s
in
our
e
nd
-
to
-
e
nd
pi
pe
li
ne
.
W
e
d
e
m
ons
tr
a
te
d
th
a
t
th
e
be
s
t
s
im
il
a
r
it
y
m
e
tr
ic
to
us
e
w
it
h
th
e
of
f
-
th
e
-
s
he
lf
f
e
a
tu
r
e
is
th
e
c
os
in
e
s
im
il
a
r
it
y
m
e
tr
ic
,
a
nd
th
a
t
th
e
be
s
t
one
to
u
s
e
w
it
h
r
e
f
in
e
d
n
e
twor
ks
is
th
e
M
a
nha
tt
a
n
s
im
il
a
r
it
y
m
e
tr
ic
.
W
e
a
l
s
o
f
ound
th
a
t
s
um
-
pool
in
g
ge
ne
r
a
ll
y
pe
r
f
or
m
s
be
tt
e
r
,
but
w
he
n
u
s
in
g
th
e
f
in
e
-
tu
ne
d ne
twor
ks
w
it
h t
he
M
a
nha
tt
a
n s
im
il
a
r
it
y m
e
tr
ic
s
, m
a
x
-
pool
in
g ga
ve
t
he
be
s
t
r
e
s
ul
ts
. W
e
e
s
ta
bl
is
he
d
th
a
t
r
e
r
a
nki
ng
s
tr
a
te
gi
e
s
c
a
n
im
pr
ove
th
e
r
e
s
ul
ts
.
M
os
t
im
por
ta
nt
ly
,
w
e
pr
ove
d
th
a
t
f
in
e
tu
ne
d
ne
twor
ks
gi
ve
th
e
be
s
t
r
e
s
ul
ts
.
S
o,
w
he
n
w
or
ki
ng
on
im
a
ge
a
nd
vi
de
o
f
a
c
e
r
e
tr
ie
va
l
us
in
g
one
que
r
y
im
a
ge
,
w
e
f
ound
th
e
be
s
t
r
e
s
ul
ts
w
e
r
e
obt
a
in
e
d
u
s
in
g
a
f
in
e
-
tu
ne
d
ne
twor
k
c
o
m
bi
ne
d
w
it
h
m
a
x
-
pool
in
g,
a
ll
our
r
e
r
a
nki
ng
s
tr
a
te
gi
e
s
a
nd
us
in
g
th
e
M
a
nha
tt
a
n
s
im
il
a
r
it
y
m
e
tr
ic
.
W
e
de
te
r
m
in
e
d
th
a
t
F
in
e
tu
ne
d
C
N
N
f
e
a
tu
r
e
c
a
n
gi
ve
gr
e
a
t
r
e
s
ul
ts
(
76,2%
)
in
r
e
a
l
ti
m
e
(
17
s
e
c
onds
pe
r
que
r
y
im
a
ge
)
w
he
n
w
or
ki
ng
on
im
a
ge
a
nd
vi
de
o
f
a
c
e
r
e
tr
ie
va
l
us
in
g a
que
r
y i
m
a
ge
.
A
C
K
N
O
WL
E
D
G
E
M
E
N
T
S
T
hi
s
w
or
k
f
a
ll
s
w
it
hi
n
th
e
s
c
op
e
of
B
ig
D
a
t
a
a
nd
C
onne
c
te
d
O
bj
e
c
t
(
B
D
C
O
)
.
W
e
w
oul
d
li
ke
to
th
a
nk t
he
H
a
s
s
a
n I
I
U
ni
ve
r
s
it
y of
C
a
s
a
bl
a
n
c
a
f
or
f
in
a
nc
in
g t
hi
s
pr
oj
e
c
t.
R
E
F
E
R
E
N
C
E
S
[
1]
D
.
F
e
ng,
M
.
-
G
.
L
i
a
ng,
F
.
G
a
o,
Y
.
-
C
.
H
ua
ng,
X
.
-
F
.
Z
ha
ng,
a
nd
L
.
-
Y
.
D
ua
n,
“
T
ow
a
r
ds
l
a
r
ge
-
s
c
a
l
e
obj
e
c
t
i
ns
t
a
nc
e
s
e
a
r
c
h:
A
m
ul
t
i
-
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
11
, N
o.
1
,
M
a
r
c
h 20
22
:
1
02
-
1
09
108
bl
oc
k
N
-
a
r
y
T
r
i
e
,”
I
E
E
E
T
r
ans
ac
t
i
ons
on
C
i
r
c
ui
t
s
and
Sy
s
t
e
m
s
f
o
r
V
i
de
o
T
e
c
hnol
ogy
,
vol
.
31,
no.
1,
pp.
372
–
386,
J
a
n.
2021,
doi
:
10.1109/
T
C
S
V
T
.2020.2966541.
[
2]
S
.
S
.
T
s
a
i
e
t
a
l
.
,
“
M
obi
l
e
pr
oduc
t
r
e
c
ogni
t
i
on,”
i
n
P
r
oc
e
e
di
ngs
of
t
he
i
nt
e
r
nat
i
o
nal
c
onf
e
r
e
nc
e
on
M
ul
t
i
m
e
di
a
-
M
M
’
10
,
2010,
A
r
t
.
no. 1587, doi
:
10.1145/
1873951.1874293.
[
3]
A
.
S
a
l
va
dor
,
X
.
G
i
r
o
-
i
-
N
i
e
t
o,
F
.
M
a
r
que
s
,
a
nd
S
.
S
a
t
oh,
“
F
a
s
t
e
r
R
-
C
N
N
f
e
a
t
ur
e
s
f
or
i
ns
t
a
nc
e
s
e
a
r
c
h,”
i
n
2016
I
E
E
E
C
onf
e
r
e
nc
e
on C
om
put
e
r
V
i
s
i
on and P
at
t
e
r
n R
e
c
ogni
t
i
on W
or
k
s
hop
s
(
C
V
P
R
W
)
,
J
un. 2016,
pp. 394
–
401, doi
:
10.1109/
C
V
P
R
W
.2016.56.
[
4]
C.
-
W
.
L
i
n
a
nd
S
.
H
ong,
“
H
i
gh
-
or
de
r
hi
s
t
ogr
a
m
-
ba
s
e
d
l
oc
a
l
c
l
us
t
e
r
i
ng
pa
t
t
e
r
ns
i
n
pol
a
r
c
oor
di
na
t
e
f
or
f
a
c
i
a
l
r
e
c
ogni
t
i
on
a
n
d
r
e
t
r
i
e
va
l
,”
T
he
V
i
s
ual
C
om
put
e
r
, M
a
r
. 2021, doi
:
10.1007/
s
00371
-
021
-
02102
-
9.
[
5]
F.
-
C
.
L
i
n,
H
.
-
H
.
N
go,
a
nd
C
.
-
R
.
D
ow
,
“
A
c
l
oud
-
ba
s
e
d
f
a
c
e
vi
de
o
r
e
t
r
i
e
va
l
s
ys
t
e
m
w
i
t
h
de
e
p
l
e
a
r
ni
ng,”
T
he
J
our
nal
o
f
Supe
r
c
om
put
i
ng
, vol
. 76, no. 11, pp. 8473
–
8493, N
ov. 2020, doi
:
10.1007/
s
112
27
-
019
-
03123
-
x.
[
6]
I
.
H
a
c
hc
ha
ne
,
A
.
B
a
dr
i
,
A
.
S
a
he
l
,
a
nd
Y
.
R
ui
c
he
k,
“
N
e
w
f
a
s
t
e
r
R
-
C
N
N
ne
ur
ona
l
a
ppr
oa
c
h
f
or
f
a
c
e
r
e
t
r
i
e
va
l
,”
i
n
L
e
c
t
ur
e
N
ot
e
s
i
n
N
e
t
w
or
k
s
and Sy
s
t
e
m
s
, S
pr
i
nge
r
I
nt
e
r
na
t
i
ona
l
P
ubl
i
s
hi
ng, 2019, pp. 113
–
120.
[
7]
S
.
R
.
D
ube
y,
“
L
oc
a
l
di
r
e
c
t
i
ona
l
r
e
l
a
t
i
on
pa
t
t
e
r
n
f
or
unc
on
s
t
r
a
i
ne
d
a
nd
r
obus
t
f
a
c
e
r
e
t
r
i
e
va
l
,”
M
ul
t
i
m
e
di
a
T
ool
s
and
A
ppl
i
c
at
i
ons
,
vol
. 78, no. 19, pp. 28063
–
28088, O
c
t
. 2019, doi
:
10.1007/
s
11042
-
019
-
07908
-
3.
[
8]
L
.
L
i
u,
J
.
L
i
,
L
.
N
i
u,
R
.
X
u,
a
nd
L
.
Z
ha
ng,
“
A
c
t
i
vi
t
y
i
m
a
ge
-
to
-
vi
de
o
r
e
t
r
i
e
va
l
b
y
di
s
e
nt
a
ngl
i
ng
a
ppe
a
r
a
nc
e
a
nd
m
ot
i
on,”
i
n
P
r
oc
e
e
di
ngs
of
t
he
A
A
A
I
C
onf
e
r
e
nc
e
on A
r
t
i
f
i
c
i
al
I
nt
e
l
l
i
ge
nc
e
, 2021, pp. 2145
–
2153.
[
9]
J
.
P
ont
-
T
us
e
t
,
P
.
A
r
be
l
a
e
z
,
J
.
T
.
B
a
r
r
on,
F
.
M
a
r
que
s
,
a
nd
J
.
M
a
l
i
k,
“
M
ul
t
i
s
c
a
l
e
c
om
bi
na
t
or
i
a
l
gr
oupi
ng
f
or
i
m
a
ge
s
e
gm
e
nt
a
t
i
on
a
nd
obj
e
c
t
pr
opos
a
l
ge
ne
r
a
t
i
on,”
I
E
E
E
T
r
ans
ac
t
i
ons
on
P
at
t
e
r
n
A
nal
y
s
i
s
and
M
ac
hi
ne
I
nt
e
l
l
i
ge
nc
e
,
vol
.
39,
no.
1,
pp.
128
–
140
,
J
a
n. 2017, doi
:
10.1109/
T
P
A
M
I
.2016.2537320.
[
10]
J
.
L
ong,
E
.
S
he
l
ha
m
e
r
,
a
nd
T
.
D
a
r
r
e
l
l
,
“
F
ul
l
y
c
onvol
ut
i
ona
l
ne
t
w
or
ks
f
or
s
e
m
a
nt
i
c
s
e
gm
e
nt
a
t
i
on,”
i
n
2015
I
E
E
E
C
onf
e
r
e
nc
e
o
n
C
om
put
e
r
V
i
s
i
on and P
at
t
e
r
n R
e
c
ogni
t
i
on (
C
V
P
R
)
,
J
un. 2015, pp. 3431
–
3440,
doi
:
10.1109/
C
V
P
R
.2015.7298965.
[
11]
A
.
F
.
D
e
A
r
a
uj
o,
“
L
a
r
ge
-
s
c
a
l
e
vi
de
o
r
e
t
r
i
e
va
l
us
i
ng
i
m
a
ge
que
r
i
e
s
a
di
s
s
e
r
t
a
t
i
on
s
ubm
i
t
t
e
d
t
o
t
he
de
p
a
r
t
m
e
nt
of
e
l
e
c
t
r
i
c
a
l
e
ngi
ne
e
r
i
ng
a
nd
t
he
c
om
m
i
t
t
e
e
on
gr
a
dua
t
e
s
t
udi
e
s
of
s
t
a
nf
or
d
uni
ve
r
s
i
t
y
i
n
pa
r
t
i
a
l
f
ul
f
i
l
l
m
e
nt
o
f
t
he
r
e
qui
r
e
m
e
nt
s
f
o
r
t
he
d
e
gr
e
e
of
doc
t
or
of
phi
l
os
,”
2016.
[
12]
J
oe
Y
ue
-
H
e
i
N
g,
M
.
H
a
us
kne
c
ht
,
S
.
V
i
j
a
ya
na
r
a
s
i
m
ha
n,
O
.
V
i
nya
l
s
,
R
.
M
o
ng
a
,
a
nd
G
.
T
ode
r
i
c
i
,
“
B
e
yond
s
hor
t
s
ni
ppe
t
s
:
D
e
e
p
ne
t
w
or
ks
f
or
vi
de
o
c
l
a
s
s
i
f
i
c
a
t
i
on,”
i
n
2015
I
E
E
E
C
onf
e
r
e
nc
e
on
C
om
put
e
r
V
i
s
i
on
and
P
at
t
e
r
n
R
e
c
ogni
t
i
on
(
C
V
P
R
)
,
J
un.
2015,
pp. 4694
–
4702, doi
:
10.1109/
C
V
P
R
.2015.7299101.
[
13]
K
.
S
i
m
onya
n
a
nd
A
.
Z
i
s
s
e
r
m
a
n,
“
T
w
o
-
s
t
r
e
a
m
c
onvol
ut
i
ona
l
ne
t
w
or
ks
f
or
a
c
t
i
on
r
e
c
ogni
t
i
on
i
n
vi
de
os
,”
J
un.
2014,
[
O
nl
i
ne
]
.
A
va
i
l
a
bl
e
:
ht
t
p:
/
/
a
r
xi
v.or
g/
a
bs
/
1406.2199.
[
14]
D
.
T
r
a
n,
L
.
B
our
de
v,
R
.
F
e
r
gus
,
L
. T
or
r
e
s
a
ni
,
a
nd
M
.
P
a
l
u
r
i
,
“
L
e
a
r
ni
ng
s
pa
t
i
ot
e
m
por
a
l
f
e
a
t
ur
e
s
w
i
t
h
3D
c
on
vol
ut
i
ona
l
ne
t
w
or
ks
,”
i
n
2015 I
E
E
E
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on C
om
put
e
r
V
i
s
i
on (
I
C
C
V
)
, D
e
c
. 2015, pp. 4489
–
4497, doi
:
10.1109/
I
C
C
V
.2015.510.
[
15]
D
.
G
.
L
ow
e
,
“
D
i
s
t
i
nc
t
i
ve
i
m
a
ge
f
e
a
t
ur
e
s
f
r
om
s
c
a
l
e
-
i
nva
r
i
a
nt
ke
ypoi
nt
s
,”
I
nt
e
r
nat
i
onal
J
our
nal
of
C
om
put
e
r
V
i
s
i
on
,
vol
.
60,
no.
2, pp. 91
–
110, N
ov. 2004, doi
:
10.1023/
B
:
V
I
S
I
.0000029664.99615.94.
[
16]
M
.
C
a
l
onde
r
,
V
.
L
e
p
e
t
i
t
,
C
.
S
t
r
e
c
ha
,
a
nd
P
.
F
ua
,
“
B
R
I
E
F
:
B
i
na
r
y
r
obus
t
i
nde
pe
nde
nt
e
l
e
m
e
nt
a
r
y
f
e
a
t
ur
e
s
,”
i
n
L
e
c
t
ur
e
N
ot
e
s
i
n
C
om
put
e
r
Sc
i
e
n
c
e
, S
pr
i
nge
r
B
e
r
l
i
n H
e
i
de
l
be
r
g, 2010, pp. 778
–
792.
[
17]
A
.
B
a
be
nko,
A
.
S
l
e
s
a
r
e
v,
A
.
C
hi
gor
i
n,
a
nd
V
.
L
e
m
pi
t
s
ky,
“
N
e
ur
a
l
c
ode
s
f
or
i
m
a
ge
r
e
t
r
i
e
va
l
,”
S
pr
i
nge
r
I
nt
e
r
na
t
i
ona
l
P
ub
l
i
s
hi
ng,
2014, pp. 584
–
599.
[
18]
Y
. K
a
l
a
nt
i
di
s
,
C
. M
e
l
l
i
na
, a
nd S
.
O
s
i
nde
r
o, “
C
r
os
s
-
di
m
e
ns
i
on
a
l
w
e
i
ght
i
ng f
or
a
ggr
e
ga
t
e
d de
e
p c
onvol
ut
i
ona
l
f
e
a
t
ur
e
s
,”
i
n
L
e
c
t
ur
e
N
ot
e
s
i
n C
om
put
e
r
Sc
i
e
nc
e
, S
pr
i
nge
r
I
nt
e
r
na
t
i
ona
l
P
ubl
i
s
hi
ng, 2016, pp. 685
–
701.
[
19]
A
.
S
.
R
a
z
a
vi
a
n,
J
.
S
ul
l
i
va
n,
S
.
C
a
r
l
s
s
on,
a
nd
A
.
M
a
ki
,
“
V
i
s
ua
l
i
ns
t
a
nc
e
r
e
t
r
i
e
va
l
w
i
t
h
de
e
p
c
onvol
ut
i
ona
l
ne
t
w
or
ks
,”
I
T
E
T
r
ans
ac
t
i
ons
on M
e
di
a T
e
c
hnol
ogy
and A
ppl
i
c
at
i
ons
, vol
. 4, no. 3, pp. 251
–
25
8, 2016, d
oi
:
10.3169/
m
t
a
.4.251.
[
20]
L
.
W
u,
Y
.
W
a
ng,
Z
.
G
e
,
Q
.
H
u,
a
nd
X
.
L
i
,
“
S
t
r
uc
t
ur
e
d
de
e
p
ha
s
hi
ng
w
i
t
h
c
onvol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
ks
f
or
f
a
s
t
pe
r
s
on
r
e
-
i
de
nt
i
f
i
c
a
t
i
on,”
C
om
put
e
r
V
i
s
i
on and I
m
age
U
nde
r
s
t
andi
ng
, vol
. 167, pp. 63
–
73, F
e
b. 2018, doi
:
10.1016/
j
.c
vi
u.2017.11.009.
[
21]
I
.
H
a
c
hc
ha
ne
,
A
.
B
a
dr
i
,
A
.
S
a
he
l
,
a
nd
Y
.
R
ui
c
he
k,
“
L
a
r
ge
-
s
c
a
l
e
i
m
a
ge
-
to
-
vi
de
o
f
a
c
e
r
e
t
r
i
e
va
l
w
i
t
h
c
onvol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
k
f
e
a
t
ur
e
s
,”
I
A
E
S
I
nt
e
r
nat
i
onal
J
our
nal
of
A
r
t
i
f
i
c
i
al
I
nt
e
l
l
i
ge
nc
e
(
I
J
-
A
I
)
,
vol
.
9,
no.
1,
pp.
40
–
45,
M
a
r
.
2020
,
doi
:
10.11591/
i
j
a
i
.v9.i
1.pp40
-
45.
[
22]
C
.
Z
ha
ng,
B
.
H
u,
Y
.
S
uo,
Z
.
Z
ou,
a
nd
Y
.
J
i
,
“
L
a
r
ge
-
s
c
a
l
e
vi
de
o
r
e
t
r
i
e
va
l
vi
a
de
e
p
l
oc
a
l
c
onvol
ut
i
ona
l
f
e
a
t
ur
e
s
,”
A
dv
anc
e
s
i
n
M
ul
t
i
m
e
di
a
, vol
. 2020, pp. 1
–
8, J
un. 2020, doi
:
10.1155/
2020
/
7862894.
[
23]
L
.
Z
he
ng,
Y
.
Y
a
ng,
a
nd
Q
.
T
i
a
n,
“
S
I
F
T
m
e
e
t
s
C
N
N
:
A
d
e
c
a
de
s
ur
ve
y
of
i
ns
t
a
nc
e
r
e
t
r
i
e
va
l
,”
I
E
E
E
T
r
ans
ac
t
i
ons
on
P
at
t
e
r
n
A
nal
y
s
i
s
and M
ac
hi
ne
I
nt
e
l
l
i
ge
nc
e
, vol
. 40, no. 5, pp. 1224
–
1244, M
a
y 2018, d
oi
:
10.1109/
T
P
A
M
I
.2017.2709749.
[
24]
S
.
Z
ha
ng,
M
.
Y
a
ng,
T
.
C
our
,
K
.
Y
u,
a
nd
D
.
N
.
M
e
t
a
xa
s
,
“
Q
ue
r
y
s
pe
c
i
f
i
c
r
a
nk
f
us
i
on
f
or
i
m
a
ge
r
e
t
r
i
e
va
l
,”
I
E
E
E
T
r
ans
ac
t
i
ons
o
n
P
at
t
e
r
n A
nal
y
s
i
s
and M
ac
hi
ne
I
nt
e
l
l
i
ge
nc
e
, vol
. 37, no. 4, pp. 803
–
815, A
pr
. 2015, doi
:
10.1109/
T
P
A
M
I
.2014.2346201.
[
25]
S
.
P
oul
l
ot
,
S
.
T
s
uka
t
a
ni
,
A
.
P
huong
N
guye
n,
H
.
J
é
gou,
a
nd
S
.
S
a
t
oh,
“
T
e
m
po
r
a
l
m
a
t
c
hi
ng
ke
r
ne
l
w
i
t
h
e
xpl
i
c
i
t
f
e
a
t
ur
e
m
a
ps
,”
i
n
P
r
oc
e
e
di
ngs
of
t
he
23r
d A
C
M
i
nt
e
r
nat
i
onal
c
onf
e
r
e
nc
e
on M
ul
t
i
m
e
di
a
, O
c
t
. 20
15, pp. 381
–
390, doi
:
10.1145/
2733373.2806228.
[
26]
D
.
M
.
C
h
e
n
a
nd
B
.
G
i
r
od,
“
A
hybr
i
d
m
obi
l
e
vi
s
ua
l
s
e
a
r
c
h
s
ys
t
e
m
w
i
t
h
c
o
m
pa
c
t
gl
oba
l
s
i
gna
t
ur
e
s
,
”
I
E
E
E
T
r
ans
ac
t
i
ons
on
M
ul
t
i
m
e
di
a
, vol
. 17, no. 7, pp. 1019
–
1030, J
ul
. 2015, doi
:
10.1109
/
T
M
M
.2015.2427744.
[
27]
C
.
H
e
r
r
m
a
nn
a
nd
J
.
B
e
ye
r
e
r
,
“
F
a
s
t
f
a
c
e
r
e
c
ogni
t
i
on
by
us
i
ng
a
n
i
nve
r
t
e
d
i
nd
e
x,”
i
n
P
r
oc
e
di
ngs
SP
I
E
94
05,
I
m
age
P
r
oc
e
s
s
i
ng:
M
ac
hi
ne
V
i
s
i
on A
ppl
i
c
at
i
ond V
I
I
I
, 940507
,
F
e
b. 2015, A
r
t
. no. 940507, doi
:
10.
1117/
12.2078988.
[
28]
A
.
S
.
R
a
z
a
vi
a
n,
H
.
A
z
i
z
pour
,
J
.
S
ul
l
i
va
n,
a
nd
S
.
C
a
r
l
s
s
on,
“
C
N
N
f
e
a
t
ur
e
s
of
f
-
t
he
-
s
he
l
f
:
A
n
a
s
t
oundi
ng
ba
s
e
l
i
ne
f
or
r
e
c
ogni
t
i
on,”
i
n
20
14
I
E
E
E
C
onf
e
r
e
nc
e
on
C
om
put
e
r
V
i
s
i
on
and
P
at
t
e
r
n
R
e
c
ogni
t
i
on
W
or
k
s
hops
,
J
un.
2014,
pp.
512
–
519,
doi
:
10.1109/
C
V
P
R
W
.2014.131.
[
29]
S
. R
e
n, K
. H
e
, R
.
G
i
r
s
hi
c
k, a
nd J
. S
un, “
F
a
s
t
e
r
R
-
C
N
N
:
T
ow
a
r
ds
r
e
a
l
-
t
i
m
e
obj
e
c
t
de
t
e
c
t
i
on w
i
t
h r
e
gi
on pr
opos
a
l
ne
t
w
or
ks
,”
I
E
E
E
T
r
ans
ac
t
i
ons
on
P
at
t
e
r
n
A
nal
y
s
i
s
and
M
ac
hi
ne
I
nt
e
l
l
i
ge
nc
e
,
vol
.
3
9,
no.
6,
pp.
1137
–
1149,
J
un.
2017,
doi
:
10.1109/
T
P
A
M
I
.2016.2577031.
[
30]
R
. G
i
r
s
hi
c
k,
“
F
a
s
t
R
-
C
N
N
,”
i
n
2015 I
E
E
E
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on C
om
pu
t
e
r
V
i
s
i
on (
I
C
C
V
)
, D
e
c
.
2015, pp.
1440
–
144
8, doi
:
10.1109/
I
C
C
V
.2015.169.
[
31]
H
.
J
i
a
ng
a
nd
E
.
L
e
a
r
ne
d
-
M
i
l
l
e
r
,
“
F
a
c
e
de
t
e
c
t
i
on
w
i
t
h
t
he
f
a
s
t
e
r
R
-
C
N
N
,”
i
n
2017
12t
h
I
E
E
E
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on
A
ut
om
at
i
c
F
ac
e
and G
e
s
t
ur
e
R
e
c
ogni
t
i
on (
F
G
2017)
, M
a
y 2017, pp. 650
–
657, doi
:
10.1109/
F
G
.2017.82.
[
32]
A
.
B
.
Y
a
nde
x
a
nd
V
.
L
e
m
pi
t
s
ky,
“
A
ggr
e
ga
t
i
ng
l
oc
a
l
de
e
p
f
e
a
t
ur
e
s
f
or
i
m
a
ge
r
e
t
r
i
e
va
l
,”
i
n
2015
I
E
E
E
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on C
om
put
e
r
V
i
s
i
on (
I
C
C
V
)
, D
e
c
. 2015, pp. 1269
–
1277, doi
:
10.1109/
I
C
C
V
.20
15.150.
[
33]
G
.
T
ol
i
a
s
,
R
.
S
i
c
r
e
,
a
nd
H
.
J
gou,
“
P
a
r
t
i
c
ul
a
r
obj
e
c
t
r
e
t
r
i
e
va
l
w
i
t
h
i
nt
e
gr
a
l
m
a
x
-
pool
i
ng
of
C
N
N
a
c
t
i
va
t
i
ons
,”
C
om
put
e
r
V
i
s
i
on
and P
at
t
e
r
n R
e
c
ogni
t
i
on
, N
ov. 2015, [
O
nl
i
ne
]
. A
va
i
l
a
bl
e
:
ht
t
p:
/
/
a
r
xi
v.or
g/
a
bs
/
1511.05879.
[
34]
L
.
W
ol
f
,
T
.
H
a
s
s
ne
r
,
a
nd
I
.
M
a
o
z
,
“
F
a
c
e
r
e
c
ogni
t
i
on
i
n
unc
ons
t
r
a
i
ne
d
vi
de
o
s
w
i
t
h
m
a
t
c
he
d
ba
c
kgr
ound
s
i
m
i
l
a
r
i
t
y,”
i
n
C
V
P
R
2011
, J
un. 2011, pp. 529
–
534, doi
:
10.1109/
C
V
P
R
.2011.5995566.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
I
m
age
and v
id
e
o f
ac
e
r
e
t
r
ie
v
al
w
it
h que
r
y
i
m
age
u
s
in
g
…
(
I
m
a
ne
H
ac
hc
hane
)
109
[
35]
H.
-
W
.
N
g
a
nd
S
.
W
i
nkl
e
r
,
“
A
da
t
a
-
dr
i
ve
n
a
ppr
oa
c
h
t
o
c
l
e
a
ni
ng
l
a
r
ge
f
a
c
e
da
t
a
s
e
t
s
,”
i
n
2014
I
E
E
E
I
nt
e
r
n
at
i
onal
C
onf
e
r
e
nc
e
on
I
m
age
P
r
oc
e
s
s
i
ng (
I
C
I
P
)
, O
c
t
. 2014, pp. 343
–
347, doi
:
10.1109/
I
C
I
P
.2014.702
5068.
[
36]
P
.
J
.
P
hi
l
l
i
ps
,
H
.
W
e
c
hs
l
e
r
,
J
.
H
ua
ng,
a
nd
P
.
J
.
R
a
us
s
,
“
T
he
F
E
R
E
T
da
t
a
ba
s
e
a
nd
e
va
l
ua
t
i
on
pr
oc
e
dur
e
f
or
f
a
c
e
-
r
e
c
ogni
t
i
on
a
l
gor
i
t
hm
s
,”
I
m
age
and V
i
s
i
on C
om
put
i
ng
, vol
. 16, no. 5, pp. 295
–
306, A
pr
. 1998, doi
:
10.1016/
S
0262
-
8856(
97
)
00070
-
X.
[
37]
D
. L
. S
pa
c
e
k, “
F
a
c
e
s
94 a
f
a
c
e
r
e
c
ogni
t
i
on da
t
a
s
e
t
,”
2007.
[
38]
M
.
K
.
B
how
m
i
k,
P
.
S
a
ha
,
A
.
S
i
ngha
,
D
.
B
ha
t
t
a
c
ha
r
j
e
e
,
a
nd
P
.
D
ut
t
a
,
“
E
nha
nc
e
m
e
nt
of
r
obus
t
ne
s
s
of
f
a
c
e
r
e
c
ogni
t
i
on
s
y
s
t
e
m
t
hr
ough
r
e
duc
e
d
ga
us
s
i
a
ni
t
y
i
n
L
og
-
I
C
A
,”
E
x
pe
r
t
Sy
s
t
e
m
s
w
i
t
h
A
ppl
i
c
at
i
ons
,
vol
.
116,
pp.
96
–
107,
F
e
b.
2019,
doi
:
10.1016/
j
.e
s
w
a
.2018.08.047.
B
I
O
G
R
A
P
H
I
E
S
O
F
A
U
T
H
O
R
S
Imane
Hachchane
is
a
Ph
.
D
.
student
in
Image
processing
at
th
e
EEA
and
TI
Laboratory,
Hassan
II
University
Casablanca,
Faculty
of
Scienc
es
and
Technology
of
Mohammedia
(FSTM)
in
Morocco.
She
received
her
Software
Engin
eering
Degree
from
th
e
National School
of Applied Sciences
of
K
en
itra, Moroc
co
in 2016. Sh
e’s cur
rently wor
king on
Facial
Large
Scale
Image
Retrieval
under
the
supervision
of
Pr.
A.
B
adri.
Her
main
research
interest
is
to
enhance
the
accura
cy
and
speed
of
largesca
le
image
and
video
face
retrieva
l
using
neural
networks
a
nd
deep
learning.
She
can
be
contacted
at
email:
hachchaneimane@gmail.com
.
Abdelmajid
Badri
is
a
holder
of
a
doctorate
in
Electronics
and
Im
age
Processing
in
1992
at
the
University
of
Poitiers
–
Franc
e.
In
1996,
he
obtain
ed
the
diploma
of
the
authorizat
ion
to
Manage
Researches
(Habilitation
à
Diriger
des
Re
cherches:
HDR)
to
the
University
of
Poitiers
–
Franc
e,
on
the
image
processing.
He
is
a
direct
or
at
the
Higher
School
of Technology (ES
T) at Casablanca and he
is
a
University
Professor (
PES
-
C) at the
Univer
sity
Hassan
II
-
Casablan
ca
-
Morocco
(FSTM).
He
is
a
member
of
the
laboratory
EEA
and
TI
(Electronics,
Energy,
Automatic
and
informatio
n
Processing)
which
he
managed
since
1996.
He
managed
several
doctoral
theses.
He
is
a
co
-
author
of
several
n
ational
and
international
publications.
He
is
responsible
for
several
research
projects
financed
by
the
ministry
or
by
the
industrialists.
He
was
member
of
several
committees
of
programs
of
i
nternationa
l
confere
nces
and
president
of
three
international
congresses
in
the
same
domain.
He
is
a
member
and
co
-
responsibl
e
in
several
scientific
association
s
in
touch
with
his
domai
n
of
research.
H
e
can
be
contacted
at
email:
abdelmaji
d_badri@
yahoo.fr
.
Aïcha
Sahel
is
a
holder
of
a
doctorate
in
Electronics
and
Image
Processing
in
1996
at
the
University
of
Poitiers
-
Franc
e.
She
is
a
university
Prof
essor
at
the
University
Hassan
II
-
Casablan
ca
-
Morocco
(FSTM)
She
is
a
member
of
the
labo
ratory
EEA
and
TI.
The
research
works
of
A.
Sahel
concern
the
Communicat
ion
and
I
nformation
Technology
(Ele
ctronics
System
s,
Signal
/Image
Processi
ng
and
Telecommunication).
She
co
-
supervises
doctoral
theses
and
she
is
a
co
-
author
of
several
national
and
internati
onal
publications.
She
is
a
member
in
financed
research
projects.
She
was
a
member
of
steering
commi
ttees
of
three
internationa
l
congresse
s
in
the
same
domain
of
research
.
Sh
e
can
be
contacted
at
email:
sahel_ai@yahoo.fr.
Ilham
Elmourabit
is
a
hold
er
of
a
doctorate
in
Telecomm
unication
and
information
engineer
ing
in
2011
at
the
University
Hassan
II
-
Casabl
anca
-
Morocco
(FSTM)
.
She
is
a
university
Professor
at
the
Hassan
II
University
Casablanca,
Faculty
of
Sciences
and
Technology
of
Mohammedia
(FSTM)
in
Morocco
.
She
is
a
member
of
the
laboratory
EEA
and
TI.
The
research
works
of
I
.
Elmourabit
concern
the
Commun
ication
and
Information
Technology
.
She
co
-
supervises
doctoral
theses
and
she
is
a
co
-
author
of
several
national
and
internationa
l publications
. Sh
e
can be
contacted
at
email
:
elmourabi
t.ilham
@
gmail.co
m
.
Yassine
Ruichek
(Senior
Member,
IEEE)
received
the
Ph.D.
de
gree
in
control
and
computer
engineering
and
the
Habilitation
à
Diriger
des
Reche
rches
(HDR)
degree
in
physic
science
from
the
University
of
Lille,
France,
in
1997
and
2
005,
respectively.
Since
2007,
he
has
been
a
Full
Professor
with
the
University
of
Technology
of
Belfort
-
Montbéliard
(UTBM).
His
research
interests
include
computer
vision,
image
p
rocessing
and
analysis,
pattern recognition, data fusion,
and localization,
with applications
in i
ntelligent transportat
ion
systems and video surve
illance.
He
can be cont
acted at
email
:
yassine.ruichek@ut
bm.fr
.
Evaluation Warning : The document was created with Spire.PDF for Python.