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.
254
~
264
I
S
S
N
:
2252
-
8938
,
D
O
I
:
10.11591/
ij
a
i.
v
11
.i
1
.pp
254
-
264
254
Jou
r
n
al
h
om
e
page
:
ht
tp
:
//
ij
ai
.
ia
e
s
c
or
e
.c
om
Pr
i
vac
y
p
r
e
se
r
vi
n
g h
u
m
a
n
ac
t
i
vi
t
y r
e
c
ogn
i
t
i
on
f
r
am
e
w
or
k
u
si
n
g
an
op
t
i
m
i
z
e
d
p
r
e
d
i
c
t
i
on
al
gor
i
t
h
m
K
am
b
al
a
V
ij
aya Ku
m
ar
, Jon
n
ad
u
la
H
ar
ik
ir
an
S
c
hool
of
C
om
put
e
r
S
c
i
e
nc
e
a
nd E
ngi
ne
e
r
i
ng
, V
I
T
-
A
P
U
ni
ve
r
s
i
t
y, A
m
a
r
a
va
t
hi
,
I
ndi
a
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
J
ul
2
3
, 2021
R
e
vi
s
e
d
D
e
c
22
, 2021
A
c
c
e
pt
e
d
J
a
n 3
,
202
2
Human
activity
recognition,
in
computer
vision
research,
is
the
a
rea
of
growing
interest
as
it
has
plethora
of
real
-
world
applications.
In
ferring
actions
from
one
or
more
persons
captured
through
a
live
video
has
its
immense
utility
in
the
contempora
ry
era
.
Same
time,
protecting
priv
acy
of
humans
is
to
be
given
paramount
importance.
Many
researcher
s
contributed
towards
this
end
leading
to
privacy
preserv
ing
action
recognition
sy
stems.
However,
having
an
optimized
model
that
can
withstand
any
ad
versary
model
s
that
strives
to
disclose
privacy
information.
To
address
this
problem,
we
proposed
an
algorithm
known
optimized
prediction
algorithm
for
p
rivacy
preserving activity recogniti
on (OPA
-
PPAR) ba
sed on
deep
neur
al net
works.
It
anonymizes
video
content
to
have
adaptive
privacy
model
that
defeats
attacks fro
m adversari
es. The p
rivacy mo
del enhances
the pri
vacy of h
umans
while
permitting
highly
accurate
approach
towards
action
recognitio
n.
The
algorit
hm
is
implem
ented
to
realize
privacy
preserving
human
a
ctivity
recognition
framewor
k
(PPHARF).
The
visu
al
recognition
of
human
actions
is
made
using
an
underlying
adversarial
learning
process
whe
re
the
anonymi
zation
is
optimi
zed
to
have
an
adaptive
priv
acy
model.
A
dataset
named
human
metabolome
database
(HMDB51)
is
u
sed
for
empirical
study.
Our
experiments
with
using
Python
data
science
platform
reveal
that
the
OPA
-
PPAR outpe
rform
s existi
ng meth
ods.
K
e
y
w
o
r
d
s
:
A
a
da
pt
iv
e
pr
iv
a
c
y m
ode
l
A
dve
r
s
a
r
ia
l
le
a
r
ni
ng
D
e
e
p ne
ur
a
l
ne
twor
ks
H
um
a
n a
c
ti
on r
e
c
ogni
ti
on
V
is
ua
l
pr
iv
a
c
y
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
:
V
ij
a
ya
K
um
a
r
K
a
m
ba
la
S
c
hool
of
C
om
put
e
r
S
c
ie
nc
e
a
nd E
ngi
ne
e
r
in
g, V
e
ll
or
e
I
ns
ti
tu
te
of
T
e
c
hnol
ogy, VI
T
-
A
P
U
ni
ve
r
s
it
y
A
m
a
r
a
va
th
i,
V
ij
a
ya
w
a
da
, A
ndhr
a
P
r
a
de
s
h, I
ndi
a
E
m
a
il
:
kvkuma
r
@
pvps
id
dha
r
th
a
.a
c
.i
n
1.
I
N
T
R
O
D
U
C
T
I
O
N
V
id
e
o
ba
s
e
d
s
ur
ve
il
la
nc
e
ha
s
be
c
om
e
a
n
im
por
ta
nt
c
om
put
e
r
vi
s
io
n
a
ppl
ic
a
ti
on.
I
t
ha
s
pl
e
nt
y
of
a
ppl
ic
a
ti
ons
in
th
e
r
e
a
l
w
or
ld
.
W
hi
le
vi
de
o
ba
s
e
d
s
ur
ve
il
la
nc
e
in
di
f
f
e
r
e
nt
dom
a
in
s
is
us
e
f
ul
,
it
ha
s
pot
e
nt
ia
l
r
is
k
in
te
r
m
s
of
pr
iv
a
c
y
le
a
ka
ge
.
T
he
r
e
f
or
e
,
m
a
ny
r
e
s
e
a
r
c
he
r
s
c
ont
r
ib
ut
e
d
to
w
a
r
ds
pr
iv
a
c
y
pr
e
s
e
r
vi
ng
a
c
ti
on
r
e
c
ogni
ti
on.
H
um
a
n
a
c
ti
on
r
e
c
ogni
ti
on
is
a
n
im
por
ta
nt
r
e
s
e
a
r
c
h
a
r
e
a
w
it
h
r
ic
h
s
e
t
of
m
e
th
ods
w
it
h
m
a
c
hi
ne
le
a
r
ni
ng,
de
e
p
le
a
r
ni
ng
a
nd
ge
ne
r
a
ti
ve
a
dve
r
s
a
r
ia
l
ne
twor
k
(
G
A
N
)
ba
s
e
d
m
ode
ls
.
A
c
ti
on
r
e
c
og
ni
ti
on
us
in
g
de
e
p
le
a
r
ni
ng,
of
te
n s
uppor
te
d
by
pr
iv
a
c
y
pr
e
s
e
r
vi
ng
m
e
th
od,
a
r
e
e
xpl
or
e
d
in
[
1]
–
[
6]
.
L
yu
e
t
al
.
[
1]
p
r
opos
e
d
a
de
e
p
le
a
r
ni
ng
ba
s
e
d
m
e
th
od
f
or
pr
iv
a
c
y
pr
e
s
e
r
vi
ng
f
r
a
m
e
w
or
k
w
it
h
f
a
ir
a
nd
de
c
e
nt
r
a
li
z
e
d
a
ppr
oa
c
h.
R
a
s
im
e
t
al
.
[
2]
pr
opos
e
d
a
d
e
e
p
le
a
r
ni
ng
ba
s
e
d
m
ode
l
f
or
pr
i
va
c
y
pr
e
s
e
r
vi
ng
a
ppr
oa
c
h
to
pr
ot
e
c
t
pe
r
s
on
a
l
da
ta
.
W
e
ng
e
t
al
.
[
3]
pr
opos
e
d
a
de
e
p
l
e
a
r
ni
ng
m
ode
l
w
it
h
bl
o
c
kc
ha
in
f
or
pr
iv
a
c
y
pr
ot
e
c
ti
on.
L
yu
e
t
al
.
[
4]
s
tu
di
e
d
f
e
de
r
a
te
d
c
lo
ud
m
ode
ls
to
a
c
hi
e
ve
f
a
ir
a
nd
pr
iv
a
c
y
pr
e
s
e
r
vi
ng
a
ppr
oa
c
he
s
to
s
ol
ve
pr
obl
e
m
s
.
K
um
a
r
e
t
al
.
[
5]
e
xpl
or
e
d
de
e
p
le
a
r
ni
ng
a
lg
or
it
hm
s
a
nd
r
e
s
o
lu
ti
on
im
a
ge
s
be
s
id
e
s
s
pa
ti
a
l
r
e
la
ti
on
s
hi
ps
to
r
e
c
ogni
z
e
hum
a
n
a
c
ti
ons
.
R
a
jp
ur
e
t
al
.
[6
]
pr
opos
e
d
a
c
lo
ud
-
ba
s
e
d
s
e
r
vi
c
e
to
a
c
hi
e
ve
pr
iv
a
c
y
pr
e
s
e
r
vi
ng
a
c
ti
on r
e
c
ogni
ti
on us
in
g
de
e
p c
onvolut
io
n ne
ur
a
l
ne
twor
k
(
C
N
N
)
m
ode
l.
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
P
r
iv
ac
y
pr
e
s
e
r
v
in
g
hum
an ac
ti
v
it
y
r
e
c
ogni
ti
on f
r
am
e
w
or
k
u
s
in
g
…
(
K
am
bal
a V
ij
ay
a K
um
ar
)
255
T
he
r
e
a
r
e
m
a
ny
a
dve
r
s
a
r
ia
l
m
ode
ls
th
a
t
pa
ve
d
w
a
y
f
or
hum
a
n
a
c
ti
on
r
e
c
ogni
ti
on.
T
he
y
a
r
e
f
ound
in
[
7]
–
[
12]
to
m
e
nt
io
n
f
e
w
.
W
u
e
t
al
.
[
7]
pr
opos
e
d
a
pr
iv
a
c
y
-
pr
ot
e
c
ti
ve
-
ge
ne
r
a
ti
ve
a
dve
r
s
a
r
ia
l
ne
twor
k
(
PP
-
G
A
N
)
w
it
h
m
odul
e
s
s
uc
h
a
s
r
e
gul
a
to
r
a
nd
ve
r
if
ic
a
to
r
.
I
t
e
ns
ur
e
s
pr
ot
e
c
ti
on
of
pr
iv
a
c
y,
s
tr
uc
tu
r
e
s
im
il
a
r
it
y
a
nd
ut
il
it
y
of
th
e
a
ppr
oa
c
h.
D
e
bi
e
e
t
al
.
[
8]
pr
opos
e
d
a
pr
iv
a
c
y
pr
e
s
e
r
vi
ng
G
A
N
f
or
c
la
s
s
if
ic
a
ti
on
of
E
C
G
da
ta
. M
a
xi
m
ov
e
t
al
.
[
9]
p
r
opos
e
d a
G
A
N
ba
s
e
d s
y
s
t
e
m
known a
s
c
ondi
ti
ona
l
id
e
nt
it
y a
nonymi
z
a
ti
on
ge
ne
r
a
ti
ve
a
dve
r
s
a
r
ia
l
ne
twor
k
(
C
I
A
G
A
N
)
w
hi
c
h
s
uppor
ts
a
nonymi
z
a
ti
on
a
nd
r
e
c
ogni
ti
on
of
a
c
ti
ons
in
im
a
ge
a
nd
vi
de
o.
I
n
f
ut
ur
e
,
th
e
y
in
te
nd
to
e
nha
nc
e
it
w
it
h
f
ul
l
im
a
ge
a
nonymi
z
a
ti
on.
M
a
r
ti
ns
s
on
e
t
al
.
[
10]
pr
opos
e
d
a
n
a
dve
r
s
a
r
ia
l
r
e
pr
e
s
e
nt
a
ti
on
le
a
r
ni
ng
m
ode
l
w
i
th
e
f
f
ic
ie
nt
m
a
na
ge
m
e
nt
of
le
a
r
na
bl
e
pa
r
a
m
e
te
r
s
.
L
i
e
t
al
.
[
11]
us
e
d
a
pr
e
-
tr
a
in
e
d
G
A
N
ba
s
e
d
m
ode
l
f
or
pr
iv
a
c
y pr
ot
e
c
ti
on.
S
hi
r
a
i
a
nd
W
hi
te
hi
ll
[
12]
pr
opos
e
d
a
G
A
N
ba
s
e
d m
ode
l
f
or
r
e
c
ogni
ti
on of
f
a
c
e
s
.
F
r
om
th
e
li
te
r
a
tu
r
e
,
it
is
unde
r
s
to
od
th
a
t
th
e
r
e
a
r
e
pl
e
nt
y
of
de
e
p
le
a
r
ni
ng
ba
s
e
d
m
e
th
ods
f
or
a
c
ti
on
r
e
c
ogni
ti
on.
S
im
il
a
r
ly
,
th
e
r
e
a
r
e
m
a
ny
G
A
N
ba
s
e
d
a
ppr
oa
c
he
s
us
e
d
f
or
hum
a
n
a
c
ti
vi
ty
r
e
c
ogni
ti
on.
M
a
ny
of
th
e
de
e
p
le
a
r
ni
ng
a
nd
G
A
N
ba
s
e
d
m
e
th
od
s
a
r
e
e
qui
pp
e
d
w
it
h
pr
iv
a
c
y
pr
e
s
e
r
vi
ng
a
ppr
oa
c
he
s
to
pr
ot
e
c
t
da
t
a
.
H
ow
e
ve
r
,
th
e
r
e
is
ne
e
d
f
or
opt
im
iz
a
ti
on
of
a
c
ti
on
r
e
c
ogni
ti
on
m
e
th
od
w
it
h
pr
iv
a
c
y
budge
t
opt
im
iz
a
ti
on.
T
o
a
ddr
e
s
s
th
is
pr
obl
e
m
,
w
e
pr
opos
e
d
a
n
a
lg
or
it
hm
known
a
s
opt
im
iz
e
d
pr
e
di
c
ti
on
a
lg
or
it
hm
f
or
pr
iv
a
c
y
pr
e
s
e
r
vi
ng
a
c
ti
vi
ty
r
e
c
ogni
ti
on
(
O
P
A
-
P
P
A
R
)
ba
s
e
d
on
de
e
p
ne
ur
a
l
ne
twor
ks
.
I
t
a
nonymi
z
e
s
vi
d
e
o
c
ont
e
nt
to
ha
ve
a
da
pt
iv
e
pr
iv
a
c
y
m
ode
l
th
a
t
de
f
e
a
t
s
a
tt
a
c
k
s
f
r
om
a
dve
r
s
a
r
ie
s
.
T
he
pr
iv
a
c
y
m
ode
l
e
nha
nc
e
s
th
e
pr
iv
a
c
y
of
hum
a
ns
w
hi
le
pe
r
m
it
ti
ng
hi
ghl
y
a
c
c
ur
a
te
a
ppr
oa
c
h
to
w
a
r
ds
a
c
ti
on
r
e
c
ogni
ti
on.
T
he
a
lg
or
it
hm
is
im
pl
e
m
e
nt
e
d
to
r
e
a
li
z
e
pr
iv
a
c
y
pr
e
s
e
r
vi
ng
hum
a
n
a
c
ti
vi
ty
r
e
c
ogni
ti
on
f
r
a
m
e
w
or
k
(
P
P
H
A
R
F
)
.
T
he
vi
s
ua
l
r
e
c
ogni
ti
on
of
hum
a
n
a
c
ti
ons
is
m
a
de
us
in
g
a
n
unde
r
ly
i
ng
a
dve
r
s
a
r
ia
l
le
a
r
ni
ng
pr
oc
e
s
s
w
h
e
r
e
th
e
a
nonymi
z
a
ti
on
is
opt
im
iz
e
d
to
ha
ve
a
n
a
da
pt
iv
e
pr
iv
a
c
y
m
ode
l.
A
da
ta
s
e
t
na
m
e
d
H
M
D
B
51i
s
us
e
d
f
or
e
m
pi
r
ic
a
l
s
tu
dy.
O
ur
c
ont
r
ib
ut
io
ns
in
th
is
pa
p
e
r
a
r
e
:
i)
w
e
pr
opos
e
d
a
f
r
a
m
e
w
or
k
known
a
s
P
P
H
A
R
F
th
a
t
le
ve
r
a
ge
s
a
c
ti
on
r
e
c
ogni
ti
on
m
ode
l,
pr
iv
a
c
y
budge
t
m
ode
l
a
nd
a
nonymi
z
a
ti
on
m
ode
l
f
or
pr
iv
a
c
y
pr
e
s
e
r
vi
ng
w
it
h
a
dve
r
s
a
r
ia
l
s
e
tt
in
g
;
ii
)
w
e
pr
opos
e
d
a
n
a
lg
or
it
hm
known
a
s
O
P
A
-
P
P
A
R
ba
s
e
d
on
de
e
p
ne
ur
a
l
ne
twor
ks
;
a
nd
ii
i)
w
e
bui
lt
a
n
a
ppl
ic
a
ti
on
to
e
v
a
lu
a
te
th
e
P
P
H
A
R
F
a
nd
th
e
unde
r
ly
in
g
O
P
A
-
P
P
A
R
a
lg
or
it
hm
us
in
g
H
M
D
B
51 da
ta
s
e
t.
T
he
r
e
m
a
in
de
r
of
th
e
pa
pe
r
is
s
tr
uc
tu
r
e
d
in
:
s
e
c
ti
on
2
r
e
vi
e
w
di
f
f
e
r
e
nt
ki
nds
of
m
e
th
ods
us
e
d
f
or
a
c
ti
on
r
e
c
ogni
ti
on
a
nd
pr
iv
a
c
y
pr
e
s
e
r
va
ti
on.
S
e
c
ti
on
3
pr
e
s
e
nt
s
th
e
pr
opos
e
d
m
e
th
od
w
it
h
unde
r
ly
in
g
a
lg
or
it
hm
.
S
e
c
ti
on
4
pr
e
s
e
nt
s
e
xpe
r
im
e
nt
a
l
r
e
s
ul
ts
a
nd
e
va
lu
a
te
s
th
e
s
a
m
e
.
S
e
c
ti
on
5
c
onc
lu
de
s
th
e
pa
pe
r
a
nd
gi
ve
s
s
ugge
s
ti
ons
f
or
f
ut
ur
e
w
or
k.
2.
R
E
L
A
T
E
D
WORK
H
um
a
n
a
c
ti
on
r
e
c
ogni
ti
on
is
a
n
im
por
ta
nt
r
e
s
e
a
r
c
h
a
r
e
a
w
it
h
r
ic
h
s
e
t
of
m
e
th
ods
w
it
h
m
a
c
hi
ne
le
a
r
ni
ng,
de
e
p
le
a
r
ni
ng
a
nd
ge
ne
r
a
ti
ve
a
dve
r
s
a
r
ia
l
ne
twor
k
(
G
A
N
)
ba
s
e
d
m
ode
ls
.
M
a
ny
pr
iv
a
c
y
pr
e
s
e
r
vi
ng
de
e
p
le
a
r
ni
ng
te
c
hni
que
s
a
r
e
e
xpl
or
e
d
by
B
oul
e
m
ta
f
e
s
e
t
al
.
[
13]
.
M
a
le
kz
a
d
e
h
e
t
al
.
[
14]
pr
opos
e
d
pr
iv
a
c
y
pr
e
s
e
r
vi
ng ba
s
e
d a
ppr
oa
c
h t
ha
t
m
a
ke
s
us
e
of
de
e
p a
ut
oe
nc
od
e
r
. L
y
u
e
t
al
.
[
1]
pr
opos
e
d a
de
e
p l
e
a
r
ni
ng
-
ba
s
e
d
m
e
th
od
f
or
pr
iv
a
c
y
pr
e
s
e
r
vi
ng
f
r
a
m
e
w
or
k
w
it
h
f
a
ir
a
nd
de
c
e
nt
r
a
li
z
e
d
a
ppr
oa
c
h.
R
a
s
im
e
t
al
.
[
2]
pr
opos
e
d
a
de
e
p l
e
a
r
ni
ng
-
ba
s
e
d m
ode
l
f
or
pr
iv
a
c
y pr
e
s
e
r
vi
ng a
ppr
oa
c
h t
o pr
ot
e
c
t
pe
r
s
ona
l
da
ta
. W
e
ng
e
t
al
.
[
3]
pr
opos
e
d
a
de
e
p
l
e
a
r
ni
ng
m
ode
l
w
it
h
bl
oc
k
c
ha
in
f
or
pr
iv
a
c
y
pr
ot
e
c
ti
on.
Y
one
ta
ni
e
t
al
.
[
15]
in
ve
s
ti
ga
te
d
on
s
e
c
ur
it
y
u
s
in
g
doubly
pe
r
m
ut
e
d
hom
om
or
phi
c
e
nc
r
ypt
io
n
(
D
P
H
E
)
w
hi
c
h
is
m
e
a
nt
f
or
pr
ot
e
c
ti
ng
hi
gh
-
di
m
e
ns
io
na
l
da
ta
.
L
yu
e
t
al
.
[
4]
s
tu
di
e
d
f
e
de
r
a
te
d
c
lo
ud
m
ode
l
s
to
a
c
hi
e
ve
f
a
ir
a
nd
pr
iv
a
c
y
pr
e
s
e
r
vi
ng
a
ppr
oa
c
he
s
to
s
ol
ve
pr
obl
e
m
s
.
Y
a
ng
e
t
al
.
[
16]
e
m
pl
oye
d
m
a
c
hi
ne
le
a
r
ni
ng
(
M
L
)
m
ode
ls
f
or
hype
r
s
pe
c
tr
a
l
im
a
ge
c
la
s
s
if
ic
a
ti
on.
D
u
e
t
al
.
[
17]
p
r
opos
e
d
de
e
p
le
a
r
ni
ng
m
ode
ls
w
it
h
pr
iv
a
c
y
pr
e
s
e
r
vi
ng
a
nd
a
ls
o
a
ppr
oxi
m
a
te
a
ppr
oa
c
h
in
c
om
put
in
g.
J
hons
on
e
t
al
.
[
18]
f
oc
us
e
d
on
th
e
r
e
a
l
ti
m
e
s
ty
le
tr
a
ns
f
e
r
us
in
g
pe
r
c
e
pt
io
n
lo
s
s
a
nd
s
upe
r
-
r
e
s
ol
ut
io
n.
H
e
e
t
al
.
[
19]
pr
opos
e
d
a
m
e
th
od
f
or
im
a
ge
r
e
c
ogni
ti
on
ba
s
e
d
on
de
e
p
r
e
s
id
ua
l
le
a
r
ni
ng.
K
ue
hne
e
t
al
.
[
20]
w
or
ke
d on the
vi
de
o da
ta
ba
s
e
known a
s
H
M
D
B
t
ha
t
is
us
e
d f
or
huma
n a
c
ti
on r
e
c
ogni
ti
on.
Y
un
e
t
al
.
[
21]
f
oc
us
e
d
on
hum
a
n
a
c
ti
vi
ty
r
e
c
ogni
ti
on
us
in
g
m
ul
ti
pl
e
in
s
ta
nc
e
le
a
r
ni
ng
a
nd
body
pos
e
f
e
a
tu
r
e
s
.
H
e
e
t
al
.
[
22]
e
xpl
oi
te
d
de
e
p
r
e
s
id
ua
l
n
e
twor
ks
w
it
h
id
e
nt
it
y
m
a
ppi
ng.
S
z
e
ge
dy
e
t
al
.
[
23]
in
ve
s
ti
ga
te
d
on
de
e
p
c
onvolut
io
na
l
ne
twor
ks
w
it
h
a
c
ti
on
r
e
c
ogni
ti
on
us
in
g
pr
e
-
r
e
c
or
de
d
vi
de
os
.
L
e
e
ne
s
e
t
al
.
[
24]
s
tu
di
e
d
on
th
e
pr
iv
a
c
y
is
s
ue
s
a
s
s
oc
ia
te
d
w
i
th
da
ta
pr
ot
e
c
ti
on
D
a
i
e
t
al
.
[
25]
pr
opos
e
d
a
nove
l
m
e
th
od
to
w
a
r
ds
hum
a
n
a
c
ti
on
r
e
c
ogni
ti
on
w
it
h
pr
iv
a
c
y
pr
e
s
e
r
ve
d.
K
um
a
r
e
t
al
.
[
5]
e
xpl
or
e
d
de
e
p
le
a
r
ni
ng
a
lg
or
it
hm
s
a
nd
r
e
s
ol
ut
io
n
i
m
a
ge
s
be
s
id
e
s
s
p
a
ti
a
l
r
e
la
ti
ons
hi
ps
to
r
e
c
ogni
z
e
hum
a
n
a
c
ti
ons
.
O
r
e
kondy
et
al
.
[
26]
pr
opos
e
d
a
m
ode
l
f
or
vi
s
u
a
l
pr
iv
a
c
y
a
dvi
s
or
th
a
t
im
pr
ove
s
pr
iv
a
c
y
of
th
e
s
ys
te
m
.
P
it
ta
lu
ga
e
t
al
.
[
27]
f
oc
us
e
d
on
m
ot
io
n
r
e
c
ons
tr
uc
ti
on
o
f
v
id
e
os
by
us
in
g
di
f
f
e
r
e
nt
im
a
ge
de
s
c
r
ip
to
r
s
.
D
a
i
e
t
al
.
[
28]
us
e
d s
pa
ti
a
l
r
e
s
ol
ut
io
n c
a
m
e
r
a
s
a
nd e
xt
r
e
m
e
ly
l
o
w
t
e
m
por
a
l
r
e
s
ol
ut
io
ns
f
or
a
c
ti
vi
ty
r
e
c
ogni
ti
on
a
nd
pr
e
s
e
r
vi
ng
pr
iv
a
c
y.
D
o
s
ovi
ts
ki
y
a
nd
B
r
ox
[
29]
in
ve
s
ti
ga
t
e
d
on
c
onvolut
io
na
l
ne
twor
ks
f
or
in
ve
r
ti
ng
of
vi
s
ua
l
r
e
pr
e
s
e
nt
a
ti
on
s
.
L
yu
e
t
al
.
[
30]
pr
opos
e
d
c
ol
l
a
bor
a
ti
ve
de
e
p
l
e
a
r
ni
ng
m
ode
ls
f
or
hum
a
n
a
c
ti
vi
ty
r
e
c
ogni
ti
on.
W
e
in
z
a
e
pf
e
l
e
t
al
.
[
31]
e
xpl
oi
te
d l
oc
a
l
de
s
c
r
ip
to
r
s
i
n i
m
a
ge
s
t
o a
r
r
iv
e
a
t
r
e
c
ons
tr
uc
ti
on of
i
m
a
ge
s
f
or
vi
s
ua
l
qua
li
ty
.
R
yoo
e
t
al
.
[
32
]
us
e
d
s
upe
r
s
ti
ti
ous
vi
de
o
r
e
c
or
di
ngs
in
or
de
r
to
r
e
c
ogni
z
e
hum
a
n
a
c
ti
ons
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
2022: 254
-
264
256
f
r
om
e
xt
r
e
m
e
lo
w
-
r
e
s
ol
ut
io
n
vi
de
os
.
M
a
he
ndr
a
n
a
nd
V
e
da
ld
i
[
33]
e
xpl
or
e
d
on
th
e
vi
s
u
a
li
z
a
ti
on
of
C
N
N
s
by
us
in
g
na
tu
r
a
l
pr
e
-
im
a
ge
s
.
W
a
ng
e
t
al
.
[
34]
us
e
d
c
ode
d
a
pe
r
tu
r
e
vi
de
os
f
or
hum
a
n
a
c
ti
vi
ty
r
e
c
ogni
ti
on
w
it
h
pr
iv
a
c
y pr
e
s
e
r
ve
d.
M
a
c
hot
e
t
al
.
[
35]
in
ve
s
ti
ga
te
d
on
s
e
ns
or
da
ta
in
or
de
r
to
di
s
c
ove
r
uns
e
e
n
a
c
ti
vi
ti
e
s
a
s
s
oc
ia
te
d
w
it
h
hum
a
n
a
c
ti
on
r
e
c
ogni
ti
on.
P
it
ta
lu
ga
a
nd
K
oppa
l
[
36]
us
e
d
m
in
ia
tu
r
e
vi
s
io
n
s
e
ns
or
s
pr
opos
e
d
pr
iv
a
c
y
pr
e
s
e
r
vi
ng opti
c
s
t
o s
tr
ik
e
ba
la
nc
e
be
twe
e
n ut
il
it
y of
vi
de
os
a
nd pr
iv
a
c
y. I
t
ha
s
m
a
ny a
ppl
ic
a
ti
ons
l
ik
e
m
ot
io
n
tr
a
c
ki
ng,
de
pt
h
s
e
n
s
in
g
a
nd
bl
ob
de
te
c
ti
on.
P
it
ta
lu
ga
e
t
al
.
[
37]
di
d
s
im
il
a
r
ki
nd
of
w
or
k.
Z
ha
ng
e
t
al
.
[
38]
pr
opos
e
d
a
m
e
th
odol
ogy
to
id
e
nt
if
y
hum
a
n
a
c
ti
vi
ti
e
s
a
s
s
oc
ia
te
d
w
it
h
f
a
ll
de
te
c
ti
on
of
e
ld
e
r
ly
pe
opl
e
.
S
ur
e
t
al
.
[
39]
on
th
e
ot
he
r
ha
nd
pr
opos
e
d
a
te
c
hni
que
to
c
ha
r
a
c
te
r
iz
e
gi
ve
n
t
a
r
ge
t
us
in
g
M
I
M
O
r
a
da
r
.
R
a
jp
ur
e
t
al
.
[
40]
p
r
opos
e
d a
c
lo
ud
-
ba
s
e
d s
e
r
vi
c
e
t
o a
c
hi
e
ve
pr
i
va
c
y pr
e
s
e
r
vi
ng a
c
ti
on r
e
c
ogni
ti
on us
in
g de
e
p
C
N
N
m
ode
l.
C
he
ng
e
t
al
.
[
41]
us
e
d
a
de
e
p
le
a
r
ni
ng
a
ppr
oa
c
h
f
or
e
m
ot
io
n
r
e
c
ogni
ti
on.
R
ib
oni
a
nd
B
e
tt
in
i
[
42]
pr
ov
id
e
d
a
n
ont
ol
ogy
-
ba
s
e
d
a
ppr
oa
c
h
to
w
a
r
ds
c
ont
e
xt
a
w
a
r
e
a
c
ti
vi
ty
r
e
c
ogni
ti
on
s
uppor
te
d
by
hybr
i
d
r
e
a
s
oni
ng.
X
u
et
al
.
[
43]
de
f
in
e
d
a
n
a
r
c
hi
te
c
tu
r
e
f
or
hum
a
n
a
c
ti
vi
ty
r
e
c
ogni
ti
on
w
it
h
two
-
s
tr
e
a
m
s
pa
ti
ot
e
m
por
a
l
ne
twor
ks
f
ul
ly
c
oupl
e
d.
Z
ol
f
a
gha
r
i
e
t
a
l
.
[
44
]
pr
opos
e
d
s
m
a
r
t
a
c
ti
vi
ty
r
e
c
ogni
ti
on
f
r
a
m
e
w
or
k
(
S
A
R
F
)
th
a
t
he
lp
s
in
m
oni
to
r
in
g
hum
a
ns
th
a
t
pr
om
ot
e
a
m
bi
e
nt
a
s
s
is
te
d
li
vi
ng
(
A
A
L
)
.
Y
oun
e
t
al
.
[
45]
f
oc
us
e
d
on
pr
ognos
ti
c
s
a
nd
he
a
lt
h
m
a
na
ge
m
e
nt
th
a
t
in
vol
ve
s
s
e
ns
in
g
f
unc
ti
ons
,
r
e
a
s
o
ni
ng,
pr
ognos
ti
c
s
,
a
nd
he
a
lt
h
m
a
na
ge
m
e
nt
.
C
il
ib
e
r
to
e
t
al
.
[
46]
pr
opos
e
d
a
3D
m
ode
l
to
ha
v
e
a
c
ti
on
r
e
c
og
ni
ti
on
w
it
h
pr
iv
a
c
y
pr
e
s
e
r
ve
d.
C
ip
pi
te
ll
i
e
t
al
.
[
47]
us
e
d
s
ke
le
ta
l
da
ta
c
ol
le
c
te
d
f
r
om
s
e
n
s
or
s
to
d
e
te
c
t
hum
a
n
a
c
ti
ons
.
W
a
ng
e
t
al
.
[
48]
s
tu
di
e
d
on
ge
nd
e
r
bi
a
s
e
li
m
in
a
ti
on w
hi
le
m
a
ki
ng de
e
p i
m
a
ge
r
e
pr
e
s
e
nt
a
ti
ons
.
W
u
e
t
al
.
[
7]
pr
opos
e
d
a
pr
iv
a
c
y
-
pr
ot
e
c
ti
v
e
-
G
A
N
(
P
P
-
G
A
N
)
w
it
h
m
odul
e
s
s
uc
h
a
s
r
e
gul
a
to
r
a
nd
ve
r
if
ic
a
to
r
.
I
t
e
ns
ur
e
s
pr
ot
e
c
ti
on
of
pr
iv
a
c
y,
s
tr
uc
tu
r
e
s
im
il
a
r
it
y
a
nd
ut
il
it
y
o
f
th
e
a
ppr
oa
c
h.
I
t
ha
s
is
s
ue
s
w
it
h
di
f
f
e
r
e
nt
he
a
d
pos
e
s
of
hum
a
ns
in
te
r
m
s
of
f
a
c
e
r
e
c
ogni
ti
on
th
a
t
ne
e
ds
f
ur
th
e
r
im
pr
ove
m
e
nt
.
D
e
bi
e
e
t
al
.
[
8]
pr
opos
e
d
a
pr
iv
a
c
y
pr
e
s
e
r
vi
ng
G
A
N
f
o
r
c
la
s
s
if
ic
a
ti
on
of
E
C
G
da
ta
.
M
a
xi
m
ov
e
t
al
.
[
9]
p
r
opos
e
d
a
G
A
N
ba
s
e
d
s
y
s
te
m
known
a
s
C
I
A
G
A
N
w
hi
c
h
s
uppor
ts
a
nonymi
z
a
ti
on
a
nd
r
e
c
ogni
ti
on
of
a
c
ti
ons
in
i
m
a
ge
a
nd
vi
de
o.
I
n
f
ut
u
r
e
,
th
e
y
in
te
nd
to
e
nha
nc
e
it
w
it
h
f
ul
l
im
a
ge
a
nonymi
z
a
ti
on.
M
a
r
ti
ns
s
on
e
t
al
.
[
10]
pr
opos
e
d
a
n
a
dve
r
s
a
r
ia
l
r
e
pr
e
s
e
nt
a
ti
on
le
a
r
ni
ng
m
ode
l
w
it
h
e
f
f
ic
ie
nt
m
a
na
g
e
m
e
nt
of
le
a
r
na
bl
e
pa
r
a
m
e
te
r
s
.
T
s
e
ng
a
nd
W
u
[
49]
pr
opos
e
d
G
A
N
known
a
s
“
pr
iv
a
c
y
ge
ne
r
a
ti
ve
a
dve
r
s
a
r
ia
l
ne
twor
k
(
C
P
G
A
N
)
”
w
hi
c
h
is
a
le
a
r
ni
n
g
f
r
a
m
e
w
or
k
w
it
h
a
dve
r
s
a
r
ia
l
s
e
tt
in
gs
.
J
in
e
t
al
.
[
50]
pr
opos
e
d
A
s
ync
hr
onous
I
nt
e
r
a
c
ti
ve
G
A
N
w
hi
le
L
i
e
t
al
.
[
11]
u
s
e
d
a
pr
e
-
tr
a
in
e
d
G
A
N
ba
s
e
d
m
ode
l
f
or
pr
iv
a
c
y
pr
ot
e
c
t
io
n.
M
a
e
t
al
.
[
51]
de
f
in
e
d
ye
t
a
not
he
r
G
A
N
m
ode
l
known
a
s
f
us
io
n
G
A
N
w
hi
c
h
m
a
k
e
s
u
s
e
of
a
ga
m
e
be
t
w
e
e
n
ge
ne
r
a
to
r
a
nd
di
s
c
r
im
in
a
to
r
.
S
hi
r
a
i
a
nd
W
hi
te
hi
ll
[
12]
pr
opos
e
d
a
G
A
N
ba
s
e
d
m
ode
l
f
or
r
e
c
ogni
ti
on
of
f
a
c
e
s
.
L
iu
e
t
al
.
[
52]
e
xpl
or
e
d
a
dv
e
r
s
a
r
ia
l
ne
twor
ks
f
or
a
c
c
ur
a
c
y e
nha
nc
e
m
e
nt
a
nd pr
iv
a
c
y qua
nt
if
ic
a
ti
on.
W
u
e
t
al
.
[
53]
pr
opos
e
d
G
A
N
m
ode
l
f
or
vi
s
ua
l
r
e
c
ogni
ti
on
w
h
il
e
pr
e
s
e
r
vi
ng
pr
iv
a
c
y.
T
he
y
us
e
d
th
e
c
onc
e
pt
of
r
e
s
ta
r
ti
ng
a
nd
e
ns
e
m
bl
e
a
ppr
oa
c
he
s
to
le
ve
r
a
ge
pe
r
f
or
m
a
nc
e
.
R
oy
a
nd
B
odde
ti
[
54]
p
r
opos
e
d
a
non
-
z
e
r
o
s
um
ga
m
e
w
it
h
a
dve
r
s
a
r
ia
l
s
e
tt
in
gs
.
Z
ha
ng
e
t
al
.
[
55]
us
e
d
a
dve
r
s
a
r
ia
l
le
a
r
ni
ng
m
e
c
h
a
ni
s
m
to
r
e
duc
e
unw
a
nt
e
d
bi
a
s
e
s
in
M
L
a
ppl
ic
a
ti
ons
.
C
he
id
e
t
al
.
[
56]
pr
opos
e
d
a
p
r
ot
oc
ol
na
m
e
d
m
ul
t
i
-
pa
r
ty
c
la
s
s
if
ic
a
ti
on
th
a
t
he
lp
s
in
hum
a
n
a
c
ti
on
r
e
c
ogni
ti
on
w
it
h
pr
iv
a
c
y
pr
e
s
e
r
ve
d.
C
he
id
a
nd
C
ha
ll
a
l
[
57]
in
ve
s
ti
ga
te
d
on
hum
a
n
a
c
ti
vi
ty
ba
s
e
d
on
s
e
ns
or
ba
s
e
d
on
s
e
ns
o
r
s
a
nd
pr
iv
a
c
y
pr
e
s
e
r
vi
ng
pr
ot
oc
ol
s
.
O
h
e
t
al
.
[
58]
pr
opos
e
d
a
f
a
c
e
le
s
s
pe
r
s
on
r
e
c
ogni
ti
on
a
nd
in
ve
s
ti
ga
t
e
d
on
it
s
im
pl
ic
a
ti
ons
.
F
r
om
th
e
li
te
r
a
tu
r
e
,
it
i
s
unde
r
s
to
od
th
a
t
th
e
r
e
a
r
e
pl
e
nt
y
o
f
de
e
p
le
a
r
ni
ng
-
ba
s
e
d
m
e
th
o
ds
f
or
a
c
ti
on
r
e
c
ogni
ti
on.
S
im
il
a
r
ly
,
th
e
r
e
a
r
e
m
a
ny
G
A
N
ba
s
e
d
a
ppr
oa
c
he
s
us
e
d
f
or
hum
a
n
a
c
ti
vi
ty
r
e
c
og
ni
ti
on.
M
a
ny
of
th
e
de
e
p
le
a
r
ni
ng
a
nd
G
A
N
ba
s
e
d
m
e
th
od
s
a
r
e
e
qui
pp
e
d
w
it
h
pr
iv
a
c
y
pr
e
s
e
r
vi
ng
a
ppr
oa
c
h
e
s
to
pr
ot
e
c
t
da
t
a
.
H
ow
e
v
e
r
,
th
e
r
e
is
n
e
e
d
f
or
opt
im
iz
a
ti
on
of
a
c
ti
on
r
e
c
ogni
ti
on
m
e
th
od
w
i
th
pr
iv
a
c
y
budge
t
opt
im
iz
a
ti
on.
T
ow
a
r
ds
th
is
e
nd
a
f
r
a
m
e
w
or
k
is
pr
opos
e
d i
n t
hi
s
pa
pe
r
.
3.
M
A
T
E
R
I
A
L
S
A
N
D
M
E
T
H
O
D
3.1
.
P
r
ob
le
m
d
e
f
in
it
io
n
G
iv
e
n
a
vi
de
o
da
ta
s
e
t
(
r
a
w
vi
de
os
c
a
pt
ur
e
d)
,
de
not
e
d
a
s
X
,
w
h
ic
h
is
s
ubj
e
c
t
e
d
to
a
c
ti
on
r
e
c
ogni
ti
on
ta
s
k
T
w
it
h
a
pr
iv
a
c
y
budge
t.
T
he
d
a
ta
s
e
t
X
ha
s
s
e
t
of
c
la
s
s
la
b
e
ls
de
not
e
d
by
a
nd
th
e
pe
r
f
or
m
a
nc
e
of
ta
s
k
is
e
va
lu
a
te
d
u
s
in
g
a
c
os
t
f
unc
ti
on
de
not
e
d
a
s
.
A
n
e
xi
s
ti
ng
s
u
pe
r
vi
s
e
d
le
a
r
ni
ng
m
e
th
od
f
or
pr
e
di
c
ti
on
of
a
c
ti
ons
is
de
not
e
d
a
s
w
hi
c
h
is
e
nha
nc
e
d
to
s
uppor
t
w
hi
c
h
is
a
c
os
t
f
unc
ti
on
f
or
budge
t
a
s
s
oc
ia
te
d
w
it
h
pr
iv
a
c
y
le
a
ka
ge
a
nd
u
s
e
d
to
f
in
d
pr
iv
a
c
y
le
a
k
a
ge
.
S
m
a
ll
e
r
v
a
lu
e
of
in
di
c
a
te
s
th
a
t
th
e
in
put
d
a
ta
ha
s
le
s
s
pr
iv
a
te
i
nf
or
m
a
ti
on a
s
s
oc
ia
te
d w
it
h i
t.
T
a
bl
e
1
s
how
s
t
he
not
a
ti
ons
us
e
d i
n t
he
pa
p
e
r
.
P
r
ovi
de
d
X
,
de
f
in
e
a
n
a
nonymi
z
a
ti
on
f
unc
ti
on
∗
w
hi
c
h
tr
a
ns
f
or
m
s
X
in
to
a
nonymi
z
e
d
X
de
not
e
d a
s
∗
(
)
a
nd
a
ne
w
de
e
p
le
a
r
ni
ng
ba
s
e
d
a
c
ti
on
r
e
c
ogni
ti
on
m
ode
l,
de
no
te
d
a
s
∗
is
de
r
iv
e
d.
I
n
th
e
pr
oc
e
s
s
,
c
a
r
e
is
ta
ke
n
to
e
n
s
ur
e
th
a
t
th
e
f
unc
ti
on
of
is
a
f
f
e
c
te
d
m
in
im
a
ll
y.
T
hi
s
dua
l
go
a
l
is
to
be
a
c
hi
e
ve
d
is
c
ons
id
e
r
e
d
a
s
a
n
opt
im
iz
a
ti
on
pr
obl
e
m
e
xpr
e
s
s
e
d
in
(
1)
.
T
he
c
os
t
f
unc
ti
o
n
of
pr
iv
a
c
y
budge
t
is
dyna
m
ic
in
na
tu
r
e
a
s
it
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
P
r
iv
ac
y
pr
e
s
e
r
v
in
g
hum
an ac
ti
v
it
y
r
e
c
ogni
ti
on f
r
am
e
w
or
k
u
s
in
g
…
(
K
am
bal
a V
ij
ay
a K
um
ar
)
257
de
pe
nds
on
th
e
r
unt
im
e
ta
s
k.
T
he
r
e
f
or
e
,
(
1)
is
r
e
de
f
in
e
d
a
nd
e
xpr
e
s
s
e
d
a
s
in
(
2)
.
A
f
ix
e
d
s
tr
uc
tu
r
e
ne
ur
a
l
ne
twor
k,
de
not
e
d
a
s
,
is
de
f
in
e
d
in
or
de
r
to
ha
ve
f
in
it
e
s
e
a
r
c
h
s
pa
c
e
to
s
ol
ve
th
e
pr
obl
e
m
w
it
h
e
a
s
e
.
T
hi
s
m
odi
f
ic
a
ti
on
is
e
xpr
e
s
s
e
d
in
(
3)
.
I
n
or
de
r
to
e
nha
nc
e
pe
r
f
or
m
a
nc
e
of
th
e
de
e
p
le
a
r
ni
ng
m
ode
l,
w
e
pr
opos
e
d
a
n e
ns
e
m
bl
e
a
ppr
oa
c
h.
T
a
bl
e
1
.
N
ot
a
ti
ons
u
s
e
d i
n t
he
pa
p
e
r
N
ot
a
t
i
on
D
e
s
c
r
i
pt
i
on
∗
A
nonym
i
z
a
t
i
on f
unc
t
i
on opt
i
m
i
z
e
d
∗
T
he
ne
w
or
de
r
i
ve
d de
e
p l
e
a
r
ni
ng m
e
t
hod
A
n e
xi
s
t
i
ng de
e
p l
e
a
r
ni
ng m
e
t
hod f
or
pr
e
di
c
t
i
on
C
os
t
f
unc
t
i
on
A
nonym
i
z
a
t
i
on f
unc
t
i
on
X
R
a
w
vi
de
o da
t
a
s
e
t
S
e
t
of
l
a
be
l
s
of
X
C
os
t
f
unc
t
i
on f
or
pr
i
va
c
y budge
t
t
o f
i
nd pr
i
va
c
y l
e
a
ka
ge
∗
(
)
A
nonym
i
z
e
d i
nput
da
t
a
X
A
pr
i
va
c
y budge
t
m
ode
l
. I
t
i
s
a
f
i
xe
d s
t
r
uc
t
ur
e
ne
ur
a
l
ne
t
w
or
k
R
e
pr
e
s
e
nt
s
l
e
a
r
na
bl
e
pa
r
a
m
e
t
e
r
s
of
R
e
pr
e
s
e
nt
s
l
e
a
r
na
bl
e
pa
r
a
m
e
t
e
r
s
of
R
e
pr
e
s
e
nt
s
l
e
a
r
na
bl
e
pa
r
a
m
e
t
e
r
s
of
N
e
ga
t
i
ve
e
nt
r
opy
∗
,
∗
=
[
(
(
(
)
)
,
)
+
(
(
)
)
]
(
1)
∗
,
∗
=
(
,
)
[
(
(
(
)
)
,
)
+
(
(
(
)
)
,
)
]
(
2)
∗
,
∗
=
(
,
)
[
(
(
(
)
)
,
)
+
(
(
(
)
)
,
)
]
(
3)
3.2
.
P
r
op
os
e
d
f
r
a
m
e
w
or
k
We
pr
opos
e
d
a
f
r
a
m
e
w
or
k
na
m
e
d
P
P
H
A
R
F
w
hi
c
h
is
c
r
uc
ia
l
f
or
a
c
c
om
m
oda
ti
ng
th
e
unde
r
ly
in
g
m
e
c
ha
ni
s
m
s
a
nd
a
lg
or
it
hm
s
to
a
c
hi
e
ve
th
e
de
s
ir
e
d
dua
l
goa
l
of
th
e
s
ys
te
m
w
hi
c
h
e
nha
nc
e
s
th
e
a
c
ti
on
r
e
c
ogni
ti
on
w
it
h
pr
iv
a
c
y
le
a
ka
ge
pr
e
ve
nt
io
n
a
nd
ke
e
ps
th
e
c
a
pa
bi
li
ti
e
s
o
f
pr
e
di
c
ti
on
a
lg
o
r
it
hm
m
a
xi
m
a
l.
T
he
f
r
a
m
e
w
or
k
ha
s
di
f
f
e
r
e
nt
m
ode
ls
in
vol
ve
d.
T
he
y
a
r
e
known
a
s
t
he
a
c
ti
on
r
e
c
ogni
ti
on
m
ode
l
∗
,
a
n
opt
im
is
e
d
a
nonymi
z
a
ti
on
f
unc
ti
on
∗
a
nd
a
pr
iv
a
c
y
budge
t
m
ode
l
de
not
e
d
a
s
.
T
he
s
e
m
ode
ls
a
r
e
im
pl
e
m
e
nt
e
d
a
s
de
e
p
ne
ur
a
l
n
e
twor
ks
w
it
h
le
a
r
na
bl
e
pa
r
a
m
e
te
r
s
.
T
he
tr
a
in
in
g
of
th
e
e
nt
ir
e
m
ode
l
is
m
a
de
w
it
h
c
om
bi
na
ti
on
of
two
lo
s
s
f
unc
ti
ons
na
m
e
ly
a
nd
.
T
he
unde
r
ly
in
g
t
r
a
in
in
g
in
th
e
f
r
a
m
e
w
or
k
ha
s
a
dua
l
goa
l
c
ons
is
ti
n
g
of
a
c
hi
e
vi
ng
opt
im
iz
e
d
a
nonymi
z
a
ti
on
f
unc
ti
on
∗
w
hi
c
h
f
il
te
r
s
pr
iv
a
te
in
f
or
m
a
ti
on
pr
io
r
to
th
e
a
c
tu
a
l
ta
s
k
a
nd a
ls
o e
ns
ur
e
s
t
ha
t
∗
(
)
is
a
c
hi
e
ve
d w
it
hout
l
im
it
in
g f
unc
ti
ona
li
ty
of
a
c
ti
on r
e
c
ogni
ti
on mode
l.
A
s
pr
e
s
e
nt
e
d
in
F
ig
ur
e
1,
th
e
l
e
a
r
ne
d
a
nonymi
z
a
ti
on
m
odul
e
t
a
ke
s
X
a
s
in
put
a
nd
tr
a
ns
f
or
m
s
it
in
to
a
nonymi
z
e
d
vi
de
o
c
ont
e
nt
th
a
t
f
i
lt
e
r
s
out
p
r
iv
a
te
in
f
o
r
m
a
ti
on
a
nd
m
odi
f
ie
s
it
s
o
a
s
to
e
ns
ur
e
th
a
t
th
e
vi
de
o
c
ont
e
nt
is
u
s
e
f
ul
f
or
a
c
ti
on
r
e
c
ogni
ti
on,
but
uni
que
hum
a
n
i
de
nt
it
y
c
a
nnot
be
a
c
hi
e
ve
d.
T
he
a
nonymi
z
e
d
vi
de
o
is
s
ubj
e
c
te
d
to
a
c
ti
on
r
e
c
o
gni
ti
on
m
ode
l
w
hi
c
h
is
de
not
e
d
a
s
.
I
t
ha
s
it
s
c
os
t
f
unc
ti
on
d
e
not
e
d
a
s
.
I
n
th
e
s
a
m
e
f
a
s
hi
on,
th
e
a
nonymi
z
e
d
vi
de
o
c
ont
e
nt
is
s
ubj
e
c
te
d
to
pr
iv
a
c
y
budge
t
m
odul
e
w
he
r
e
a
not
he
r
c
os
t
f
unc
ti
on de
not
e
d a
s
. W
he
n both c
os
t
f
unc
ti
ons
a
r
e
c
om
bi
n
e
d
to
f
or
m
of
a
l
os
s
f
unc
ti
on w
hi
c
h c
ont
r
ol
s
th
e
it
e
r
a
ti
ve
pr
oc
e
s
s
of
th
e
f
r
a
m
e
w
or
k
a
nd
e
ns
ur
e
s
opt
im
iz
a
ti
on
of
a
c
ti
on
r
e
c
ogni
ti
on
w
hi
le
pr
e
s
e
r
vi
ng
pr
iv
a
c
y.
T
he
a
nonymi
z
a
ti
on
m
ode
l
is
im
pl
e
m
e
nt
e
d
a
s
a
f
r
a
m
e
le
ve
l
f
il
te
r
w
hi
c
h
is
ba
s
e
d
on
2D
c
onvolut
io
na
l
ne
twor
k.
T
he
a
c
ti
on
r
e
c
ogni
ti
on
m
odul
e
i
s
ta
ke
n
f
r
om
[
59]
a
nd
r
e
us
e
d
it
.
T
he
pr
iv
a
c
y
budge
t
m
ode
l
is
m
a
de
up
of
R
e
s
N
e
t.
F
or
th
e
s
a
m
e
of
a
c
ti
on
r
e
c
ogn
it
io
n,
th
e
vi
de
o
is
di
vi
de
d
in
to
num
be
r
of
f
r
a
m
e
s
(
vi
de
o c
li
ps
)
a
nd e
a
c
h f
r
a
m
e
i
s
uni
que
ly
i
de
nt
if
ie
d.
A
m
in
im
a
x
pr
obl
e
m
a
s
s
oc
ia
t
e
d
w
it
h
(
3
)
is
s
ol
ve
d
by
c
ons
id
e
r
in
g
di
f
f
e
r
e
nt
le
a
r
na
bl
e
pa
r
a
m
e
te
r
s
of
th
e
th
r
e
e
m
ode
ls
us
e
d
in
th
e
f
r
a
m
e
w
or
k.
T
he
le
a
r
na
bl
e
pa
r
a
m
e
te
r
s
,
,
a
nd
a
r
e
a
s
s
oc
ia
te
d
w
it
h
,
a
nd
r
e
s
pe
c
ti
ve
ly
.
I
n
or
de
r
to
s
ol
ve
t
he
m
in
im
a
x
pr
obl
e
m
,
w
e
c
ons
id
e
r
e
d
th
e
not
io
n
of
a
lt
e
r
na
ti
ve
m
in
im
iz
a
ti
on
f
ound
in
[6
0
]
.
I
t
is
e
xpr
e
s
s
e
d
a
s
in
(
4
)
-
(
6
)
.
T
he
n
th
e
two
lo
s
s
f
unc
ti
on
s
a
r
e
opt
im
iz
e
d
to
s
ol
ve
th
e
opt
im
iz
a
ti
on pr
obl
e
m
s
e
xpr
e
s
s
e
d i
n
(
7)
a
nd (
8)
. T
he
(
7
)
is
t
he
m
in
im
iz
a
ti
on pr
obl
e
m
w
hi
le
(
8
)
is
m
in
im
a
x
pr
obl
e
m
.
T
he
f
or
m
e
r
is
us
e
d
to
ha
ve
tr
a
in
in
g
of
a
nd
w
hi
le
th
e
la
tt
e
r
is
us
e
d
to
ke
e
p
tr
a
c
k
of
di
f
f
e
r
e
nt
pa
r
a
m
e
te
r
s
of
pr
iv
a
c
y
budge
t
m
ode
l.
I
n
or
de
r
to
s
ol
ve
two
l
os
s
f
unc
ti
ons
two
pa
r
a
m
e
te
r
upda
te
r
ul
e
s
a
r
e
e
xpr
e
s
s
e
d i
n
(
9)
-
(
12)
.
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
2022: 254
-
264
258
F
ig
ur
e
1
.
P
r
opos
e
d pr
iv
a
c
y pr
e
s
e
r
vi
ng huma
n a
c
ti
vi
ty
r
e
c
ogni
ti
on f
r
a
m
e
w
or
k w
it
h a
dve
r
s
a
r
ia
l
s
e
tt
in
g
←
−
∝
∇
(
(
,
)
−
(
,
)
)
(
4)
←
−
∝
∇
(
,
)
(
5)
←
−
∝
∇
(
,
)
(
6)
∗
,
∗
=
(
,
)
(
,
)
(
7)
∗
,
∗
=
(
,
)
(
8)
←
←
,
−
∝
∇
(
,
)
(
,
)
(
9)
←
∈
{
1
,
…
}
(
,
)
(
10)
,
+
∝
∇
(
,
)
(
11)
,
−
∝
∇
(
,
)
,
∀
∈
{
1
,
…
}
(
12)
W
e
f
ound
th
a
t
(
4
)
is
in
s
ta
bl
e
w
hi
c
h
c
a
n
be
s
ol
ve
d
by
c
ons
id
e
r
in
g
ne
ga
ti
ve
e
nt
r
opy
w
hi
c
h
is
in
c
or
por
a
te
d
to
ha
ve
a
ne
w
s
c
he
m
e
a
s
e
xpr
e
s
s
e
d
in
(
13
)
-
(
15
)
.
W
it
h
th
e
s
e
opt
im
iz
a
ti
ons
,
th
e
r
e
is
pos
s
ib
il
it
y
of
m
a
xi
m
iz
in
g
e
nt
r
opy
th
a
t
le
ve
r
a
ge
s
pe
r
f
or
m
a
nc
e
.
T
he
(
2
)
is
f
ur
th
e
r
opt
im
iz
e
d
w
it
h
e
ns
e
m
bl
e
a
ppr
oa
c
h
in
th
e
tr
a
in
in
g
pr
oc
e
s
s
to
im
pr
ove
m
ode
l
a
c
c
ur
a
c
y
a
s
e
xpr
e
s
s
e
d
i
n
(
16
)
.
T
he
e
ns
e
m
bl
e
m
ode
l
a
nd
opt
im
iz
e
d
pa
r
a
m
e
te
r
s
e
tt
in
gs
a
r
e
f
ur
th
e
r
i
m
pr
ove
d w
it
h a
s
c
he
m
e
e
xpr
e
s
s
e
d i
n
(
17
)
-
(
19
)
.
←
−
∝
∇
(
(
,
)
−
(
,
)
)
(
13)
,
←
,
−
∝
∇
,
(
,
)
(
14)
←
−
∝
∇
(
,
)
(
15)
∗
,
∗
=
(
,
)
[
(
(
(
)
)
,
)
+
∈
(
(
(
)
)
,
]
(
16)
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
P
r
iv
ac
y
pr
e
s
e
r
v
in
g
hum
an ac
ti
v
it
y
r
e
c
ogni
ti
on f
r
am
e
w
or
k
u
s
in
g
…
(
K
am
bal
a V
ij
ay
a K
um
ar
)
259
←
−
∝
∇
(
+
∈
−
(
,
)
)
(
17)
←
←
←
−
∝
∇
(
,
)
(
,
)
(
18)
,
−
∝
∇
(
,
)
,
∀
∈
{
1
,
…
}
(
19)
W
it
h
th
e
s
e
opt
im
iz
a
ti
ons
,
th
e
pr
opos
e
d
f
r
a
m
e
w
or
k
P
P
H
A
R
F
is
m
a
de
m
or
e
s
ophi
s
ti
c
a
te
d
in
te
r
m
s
of
hum
a
n
a
c
ti
on
r
e
c
ogni
ti
on
a
nd
pr
e
s
e
r
vi
ng
pr
iv
a
c
y
th
a
t
e
ns
ur
e
s
non
-
di
s
c
lo
s
ur
e
of
id
e
nt
it
y.
W
it
h
di
f
f
e
r
e
nt
m
odul
e
s
in
pl
a
c
e
,
th
e
f
r
a
m
e
w
or
k
ope
r
a
te
s
in
a
n
it
e
r
a
ti
ve
m
od
e
l
in
or
de
r
to
ha
v
e
be
tt
e
r
p
e
r
f
or
m
a
nc
e
.
W
it
h
c
om
bi
ne
d l
os
s
f
unc
ti
on i
t
c
a
n r
e
a
li
z
e
t
h
e
dua
l
goa
l
of
t
he
f
r
a
m
e
w
or
k a
f
or
e
m
e
nt
io
ne
d.
3.3. T
h
e
p
r
op
os
e
d
al
gor
it
h
m
We
pr
opos
e
d
a
n
a
lg
or
it
hm
known
O
P
A
-
P
P
A
R
ba
s
e
d
on
de
e
p
ne
ur
a
l
ne
twor
ks
.
I
t
a
nonymi
z
e
s
vi
de
o
c
ont
e
nt
to
ha
ve
a
da
pt
iv
e
pr
iv
a
c
y
m
ode
l
th
a
t
de
f
e
a
ts
a
tt
a
c
ks
f
r
om
a
dve
r
s
a
r
ie
s
.
T
h
e
pr
iv
a
c
y
m
ode
l
e
nha
nc
e
s
th
e
pr
iv
a
c
y
of
hum
a
n
s
w
hi
le
pe
r
m
it
ti
ng
hi
ghl
y
a
c
c
ur
a
te
a
ppr
o
a
c
h
to
w
a
r
ds
a
c
ti
on
r
e
c
ogni
ti
on.
T
he
a
lg
or
it
hm
is
im
pl
e
m
e
nt
e
d
to
r
e
a
li
z
e
P
P
H
A
R
F
.
T
he
vi
s
ua
l
r
e
c
ogni
ti
on
of
hum
a
n
a
c
ti
ons
is
m
a
de
us
in
g
a
n
unde
r
ly
in
g
a
dve
r
s
a
r
ia
l
le
a
r
ni
ng
pr
oc
e
s
s
w
he
r
e
th
e
a
nonymi
z
a
ti
on
i
s
opt
i
m
iz
e
d
to
ha
ve
a
n
a
da
pt
iv
e
pr
iv
a
c
y
m
ode
l.
A
da
ta
s
e
t
na
m
e
d H
M
D
B
51 i
s
us
e
d f
or
e
m
pi
r
ic
a
l
s
tu
dy.
A
lg
or
it
hm
1. O
pt
im
iz
e
d pr
e
di
c
ti
on a
lg
or
it
hm
f
or
pr
iv
a
c
y pr
e
s
e
r
vi
ng a
c
ti
vi
ty
r
e
c
ogni
ti
on
Inputs:
X
,
m
o
d
e
l
l
e
a
r
n
a
b
l
e
p
a
r
a
m
e
t
e
r
s
s
u
c
h
a
s
,
and
O
u
t
p
u
t
:
U
p
d
a
t
e
d
r
e
c
o
g
n
i
z
e
d
a
c
t
i
o
n
s
m
a
p
w
i
t
h
p
r
i
v
a
c
y
p
r
e
s
e
r
v
e
d
1.
I
n
i
t
i
a
l
i
z
e
f
r
a
m
e
s
v
e
c
t
o
r
F
2.
I
n
i
t
i
a
l
i
z
e
a
c
t
i
o
n
s
m
a
p
R
3.
F
S
p
l
i
t
V
i
d
e
o
(
X
)
4.
F
o
r
e
a
c
h
f
r
a
m
e
f
i
n
F
5.
R
e
p
e
a
t
6.
A
p
p
l
y
l
e
a
r
n
e
d
a
n
o
n
y
m
i
z
a
t
i
o
n
m
o
d
e
l
on f
7.
A
p
p
l
y
p
r
i
v
a
c
y
b
u
d
g
e
t
m
o
d
e
l
on f
8.
A
p
p
l
y
a
c
t
i
o
n
r
e
c
o
g
n
i
t
i
o
n
m
o
d
e
l
∗
optimized
by
a
n
d
9.
C
o
m
p
u
t
e
C
o
s
t
F
u
n
c
t
i
o
n
O
f
A
c
t
i
o
n
R
e
c
o
g
n
i
t
i
o
n
(
)
10.
C
o
m
p
u
t
e
C
o
s
t
F
u
n
c
t
i
o
n
O
f
P
r
i
v
a
c
y
L
e
a
k
a
g
e
(
)
11.
l
o
s
s
f
u
n
c
t
i
o
n
L
+
12.
U
s
e
l
e
a
r
n
a
b
l
e
p
a
r
a
m
e
t
e
r
s
,
and
13.
G
e
t
f
e
e
d
b
a
c
k
f
o
r
t
h
r
e
e
m
o
d
e
l
s
14.
U
n
t
i
l
C
o
n
v
e
r
g
e
n
c
e
15.
U
p
d
a
t
e
R
16.
End For
17.
R
e
t
u
r
n
R
A
s
pr
e
s
e
nt
e
d
in
A
lg
or
it
hm
1,
it
ta
ke
s
X
a
nd
m
ode
l
le
a
r
na
bl
e
pa
r
a
m
e
te
r
s
s
uc
h
a
s
,
,
a
nd
a
s
in
put
s
a
nd
pr
oduc
e
s
a
n
upd
a
te
d
r
e
c
ogni
z
e
d
a
c
ti
on
m
a
p
w
it
h
pr
iv
a
c
y
pr
e
s
e
r
ve
d.
I
n
s
te
p
1,
it
in
it
ia
li
z
e
d
f
r
a
m
e
s
ve
c
to
r
na
m
e
d
F
w
hi
c
h
hol
ds
f
r
a
m
e
s
(
not
hi
ng
but
s
pl
it
f
il
m
s
of
vi
de
o)
.
S
te
p
2
in
it
ia
li
z
e
s
a
c
ti
ons
m
a
p
th
a
t
w
il
l
be
upda
te
d
it
e
r
a
ti
ve
ly
a
nd
r
e
tu
ne
d
on
c
onve
r
ge
nc
e
.
S
te
p
3
s
pl
it
s
gi
ve
n
r
a
w
vi
de
o
in
to
s
om
e
f
r
a
m
e
s
.
A
n
it
e
r
a
ti
ve
pr
oc
e
s
s
is
e
xpr
e
s
s
e
d
in
s
te
ps
4
th
r
ough
s
te
p
16.
F
or
e
a
c
h
f
r
a
m
e
a
ga
in
,
th
e
r
e
is
a
n
it
e
r
a
ti
ve
pr
oc
e
s
s
th
a
t
a
ppl
ie
s
th
e
two
m
odul
e
s
a
s
gi
ve
n
in
s
te
p
6,
s
te
p
7
,
a
nd
s
te
p
8
r
e
s
pe
c
ti
ve
ly
.
T
w
o
ki
nds
of
c
os
t
f
unc
ti
ons
a
r
e
c
om
put
e
d i
n
s
te
p 9 a
nd s
te
p 10 r
e
s
pe
c
ti
ve
ly
. T
he
s
e
t
w
o c
os
t
f
unc
ti
ons
a
r
e
c
om
bi
ne
d i
n s
te
p 11 to a
r
r
iv
e
a
t
a
c
om
bi
ne
d
lo
s
s
f
unc
ti
on
th
a
t
is
us
e
d
in
th
e
tr
a
in
in
g
of
th
e
m
ode
ls
in
or
de
r
to
gi
ve
f
e
e
dba
c
k
a
nd
c
ont
in
ue
pr
oc
e
s
s
unt
il
c
onv
e
r
ge
nc
e
.
S
te
p
12
us
e
s
le
a
r
na
bl
e
pa
r
a
m
e
te
r
s
a
nd
s
te
p
13
gi
ve
s
f
e
e
db
a
c
k
ne
e
de
d
in
th
e
a
dve
r
s
a
r
ia
l
s
e
tt
in
g
of
th
e
pr
opos
e
d
f
r
a
m
e
w
or
k.
S
te
p
7
r
e
tu
r
n
s
f
in
a
l
r
e
s
ul
ts
th
a
t
a
r
e
obt
a
in
e
d
w
it
h
pr
iv
a
c
y
pr
e
s
e
r
ve
d.
4.
E
X
P
E
R
I
M
E
N
T
A
L
R
E
S
U
L
T
S
We
pr
opos
e
d
a
n
a
lg
or
it
hm
known
O
P
A
-
P
P
A
R
is
e
va
lu
a
te
d
us
in
g
H
M
D
B
51
d
a
ta
s
e
t.
T
he
r
e
s
ul
ts
of
O
P
A
-
P
P
A
R
is
c
om
pa
r
e
d
w
it
h
th
a
t
o
f
our
pr
io
r
w
or
k
na
m
e
d
m
ul
ti
-
ta
s
k
le
a
r
ni
ng
ba
s
e
d
hybr
id
pr
e
di
c
ti
o
n
a
lg
or
it
hm
(
M
T
L
-
H
P
A
)
a
nd
th
e
s
ta
te
of
th
e
a
r
t
m
e
th
od
na
m
e
d
gr
a
di
e
nt
r
e
ve
r
s
a
l
la
ye
r
(
G
R
L
)
[6
1
]
.
A
s
s
how
n
in
F
ig
ur
e
2,
th
e
r
e
a
r
e
51
a
c
ti
on s
a
m
pl
e
s
in
H
M
D
B
51
d
a
ta
s
e
t.
O
ut
of
th
e
m
100
s
a
m
pl
e
s
a
r
e
us
e
d
f
or
e
m
pi
r
ic
a
l
s
tu
dy i
n t
hi
s
pa
pe
r
. H
ow
e
ve
r
, r
e
s
ul
ts
a
r
e
pr
e
s
e
nt
e
d i
n t
hi
s
pa
p
e
r
f
or
10 a
c
ti
ons
. T
he
y i
nc
lu
de
c
li
m
b, e
a
t,
j
um
p,
ki
s
s
,
pus
h,
pu
s
hup,
r
un,
s
it
,
s
m
il
e
a
nd
w
a
lk
.
A
s
pr
e
s
e
nt
e
d
in
F
ig
ur
e
3,
th
e
in
put
im
a
ge
s
or
f
r
a
m
e
s
a
r
e
s
how
n
in
le
f
t
c
ol
um
n
a
nd
th
e
a
c
ti
on
r
e
c
ogni
z
e
d
a
nd
a
nonymi
z
e
d
f
r
a
m
e
a
r
e
s
how
n
in
s
e
c
ond
a
nd
th
ir
d
c
ol
um
ns
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
2022: 254
-
264
260
r
e
s
pe
c
ti
ve
ly
.
T
he
e
xp
e
r
im
e
nt
a
l
r
e
s
ul
ts
a
r
e
e
va
lu
a
te
d
in
te
r
m
s
of
pr
e
c
is
io
n,
r
e
c
a
ll
a
nd
F
1
-
S
c
or
e
.
T
he
pe
r
f
or
m
a
nc
e
va
lu
e
s
a
r
e
obt
a
in
e
d
w
it
h
hum
a
n
s
tu
dy
on
a
nonym
iz
e
d
s
a
m
pl
e
s
. T
he
gr
ound
tr
ut
h
a
nd
pr
e
di
c
ti
on
r
e
s
ul
ts
of
t
he
a
c
ti
on r
e
c
ogni
ti
on me
th
ods
a
r
e
s
ubj
e
c
te
d t
o e
va
lu
a
ti
on i
n t
e
r
m
s
of
t
he
m
e
a
s
ur
e
s
.
F
ig
ur
e
2. S
om
e
huma
n a
c
ti
on s
a
m
pl
e
s
pr
e
s
e
nt
i
n
H
M
D
B
51 d
a
ta
s
e
t
[
39]
F
r
a
m
e
A
c
t
i
on
A
nonym
i
z
e
d F
r
a
m
e
F
r
a
m
e
A
c
t
i
on
A
nonym
i
z
e
d F
r
a
m
e
C
l
i
m
b
E
a
t
J
um
p
K
i
s
s
P
us
h
P
us
hup
R
un
S
i
t
S
m
i
l
e
W
a
l
k
F
ig
ur
e
3
.
E
xpe
r
im
e
nt
a
l
r
e
s
ul
ts
f
or
t
he
s
e
le
c
ti
on a
c
ti
ons
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
P
r
iv
ac
y
pr
e
s
e
r
v
in
g
hum
an ac
ti
v
it
y
r
e
c
ogni
ti
on f
r
am
e
w
or
k
u
s
in
g
…
(
K
am
bal
a V
ij
ay
a K
um
ar
)
261
F
ig
ur
e
4 a
nd F
ig
ur
e
5
s
how
t
he
pe
r
f
or
m
a
nc
e
c
om
pa
r
is
on be
twe
e
n t
he
G
R
L
vs
. pr
opos
e
d m
e
th
od a
nd
M
T
L
-
H
P
A
vs
.
th
e
pr
opos
e
d
m
e
th
od.
I
n
bot
h
th
e
c
a
s
e
s
,
th
e
a
c
ti
on
r
e
c
ogni
ti
on
m
ode
ls
a
r
e
pr
e
s
e
nt
e
d
in
hor
iz
ont
a
l
a
xi
s
a
nd
ve
r
ti
c
a
l
a
xi
s
s
how
s
th
e
p
e
r
f
or
m
a
nc
e
(
%
)
.
O
bs
e
r
va
ti
ons
a
r
e
m
a
d
e
w
it
h
10
hum
a
n
a
c
ti
ons
.
F
or
e
a
c
h
hum
a
n
a
c
ti
on
100
e
xpe
r
im
e
nt
s
a
r
e
m
a
de
w
it
h
th
e
pr
ot
ot
ype
m
a
de
to
de
m
ons
tr
a
te
pr
oof
o
f
th
e
c
onc
e
pt
.
P
r
e
c
is
io
n,
r
e
c
a
ll
a
nd
F
1
-
s
c
or
e
a
r
e
c
om
put
e
d
ba
s
e
d
on
gr
ound
tr
ut
h
a
nd
th
e
r
e
s
ul
ts
of
th
e
a
c
ti
on
r
e
c
ogni
ti
on
m
ode
ls
.
T
he
f
in
a
l
e
va
lu
a
ti
on
r
e
s
ul
ts
a
r
e
obt
a
in
e
d
w
it
h
hum
a
n
s
tu
dy.
T
he
r
e
s
ul
ts
r
e
ve
a
le
d
th
a
t
th
e
pr
opos
e
d
m
e
th
od
O
P
A
-
P
P
A
R
out
pe
r
f
or
m
s
th
e
e
xi
s
ti
ng
m
e
th
o
ds
known
a
s
G
R
L
a
nd
M
T
L
-
H
P
A
.
T
he
M
T
L
-
H
P
A
s
how
e
d
s
ig
ni
f
ic
a
nt
ly
be
tt
e
r
pe
r
f
or
m
a
nc
e
ove
r
th
e
ba
s
e
li
ne
G
R
L
m
e
th
od.
T
he
e
xpe
r
im
e
nt
a
l
r
e
s
ul
t
s
r
e
ve
a
le
d
th
a
t
th
e
pr
opos
e
d
a
c
ti
on
r
e
c
ogni
ti
on
m
e
th
od
not
onl
y
pr
e
s
e
r
ve
s
pr
iv
a
c
y
a
nd
r
e
c
ogni
z
e
s
hum
a
n
a
c
ti
ons
but
a
ls
o
ha
s
opt
im
iz
a
ti
ons
in
te
r
m
s
of
pr
iv
a
c
y
budge
t
a
nd
a
c
om
bi
ne
d
lo
s
s
f
unc
ti
on
to
gu
id
e
th
e
r
e
c
ogni
ti
on
pr
oc
e
s
s
a
s
s
oc
ia
te
d
w
it
h
th
e
pr
opos
e
d
f
r
a
m
e
w
or
k
.
A
s
pr
e
s
e
nt
e
d
in
T
a
bl
e
2
a
nd
T
a
bl
e
3,
th
e
pe
r
f
or
m
a
nc
e
of
t
he
pr
opos
e
d m
e
th
od i
s
c
om
pa
r
e
d w
it
h t
ha
t
of
G
R
L
a
nd M
T
L
-
H
P
A
.
F
ig
ur
e
4. P
e
r
f
or
m
a
nc
e
c
om
pa
r
is
on of
a
c
ti
on r
e
c
ogni
ti
on mode
ls
G
R
L
a
nd O
P
A
-
P
P
A
R
F
ig
ur
e
5. P
e
r
f
or
m
a
nc
e
c
om
pa
r
is
on of
a
c
ti
on r
e
c
ogni
ti
on mode
ls
M
T
L
-
H
P
A
a
nd O
P
A
-
P
P
A
R
T
a
bl
e
2. R
e
s
ul
ts
of
t
he
pr
opos
e
d m
e
th
od
c
om
pa
r
e
d w
it
h t
ha
t
of
G
R
L
A
c
t
i
on
G
R
L
P
e
r
f
or
m
a
nc
e
O
P
A
-
P
P
A
R
P
e
r
f
or
m
a
nc
e
P
r
e
c
i
s
i
on
R
e
c
a
l
l
F1
-
S
c
or
e
P
r
e
c
i
s
i
on
R
e
c
a
l
l
F1
-
S
c
or
e
C
l
i
m
b
0.61
0.9
0.727152
0.90524
0.97308
0.937935
E
a
t
0.63
0.92
0.747871
0.93492
0.994704
0.963886
J
um
p
0.6
0.89
0.716779
0.8904
0.962268
0.92494
K
i
s
s
0.58
0.92
0.711467
0.86072
0.994704
0.922874
P
us
h
0.57
0.87
0.68875
0.86982
0.940644
0.903847
P
us
hup
0.62
0.93
0.744
0.92008
0.99603
0.95655
R
un
0.63
0.91
0.744545
0.93492
0.983892
0.958781
S
i
t
0.61
0.88
0.720537
0.90524
0.951456
0.927773
S
m
i
l
e
0.59
0.92
0.71894
0.87556
0.994704
0.931337
W
a
l
k
0.61
0.94
0.739871
0.90524
0.997152
0.948976
0
0.2
0.4
0.6
0.8
1
1.2
P
r
e
c
i
s
i
on
R
e
c
a
l
l
F
1-
S
c
or
e
P
r
e
c
i
s
i
on
R
e
c
a
l
l
F
1-
S
c
or
e
G
R
L
O
P
A
-
P
P
A
R
P
e
r
f
or
m
a
n
c
e
(
%
)
A
c
t
i
on
R
e
c
og
n
i
t
i
on
M
ode
l
s
P
e
r
f
or
m
a
n
c
e
C
om
pa
r
i
s
on
C
l
i
m
b
E
a
t
J
um
p
K
i
s
s
P
us
h
P
u
s
h
u
p
R
un
S
i
t
S
m
i
l
e
W
a
l
k
0
0.2
0.4
0.6
0.8
1
1.2
P
r
e
c
i
s
i
on
R
e
c
a
l
l
F
1-
S
c
or
e
P
r
e
c
i
s
i
on
R
e
c
a
l
l
F
1-
S
c
or
e
M
T
L
-
H
P
A
O
P
A
-
P
P
A
R
P
e
r
f
or
m
a
n
c
e
(
%
)
A
c
t
i
on
R
e
c
og
n
i
t
i
on
M
ode
l
s
P
e
r
f
or
m
a
n
c
e
C
om
pa
r
i
s
on
C
l
i
m
b
E
a
t
J
um
p
K
i
s
s
P
us
h
P
u
s
h
u
p
R
un
S
i
t
S
m
i
l
e
W
a
l
k
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
2022: 254
-
264
262
T
a
bl
e
3. R
e
s
ul
ts
of
t
he
pr
opos
e
d m
e
th
od
c
om
pa
r
e
d w
it
h t
ha
t
of
our
pr
io
r
m
e
th
od M
T
L
-
H
P
A
A
c
t
i
ons
M
T
L
-
H
P
A
O
P
A
-
P
P
A
R
P
r
e
c
i
s
i
on
R
e
c
a
l
l
F1
-
S
c
or
e
P
r
e
c
i
s
i
on
R
e
c
a
l
l
F1
-
S
c
or
e
C
l
i
m
b
0.854
0.918
0.884844
0.90524
0.97308
0.937935
E
a
t
0.882
0.9384
0.909326
0.93492
0.994704
0.963886
J
um
p
0.84
0.9078
0.872585
0.8904
0.962268
0.92494
K
i
s
s
0.812
0.9384
0.870636
0.86072
0.994704
0.922874
P
us
h
0.798
0.8874
0.840329
0.86982
0.940644
0.903847
P
us
hup
0.868
0.9486
0.906512
0.92008
0.99603
0.95655
R
un
0.882
0.9282
0.90451
0.93492
0.983892
0.958781
S
i
t
0.854
0.8976
0.875257
0.90524
0.951456
0.927773
S
m
i
l
e
0.826
0.9384
0.87862
0.87556
0.994704
0.931337
W
a
l
k
0.854
0.9588
0.903371
0.90524
0.997152
0.948976
5.
C
O
N
C
L
U
S
I
O
N
A
N
D
F
U
T
U
R
E
WORK
We
pr
opos
e
d
a
n
a
lg
or
it
hm
known
O
P
A
-
P
P
A
R
ba
s
e
d
on
de
e
p
ne
ur
a
l
ne
twor
ks
.
I
t
a
nonymi
z
e
s
vi
de
o
c
ont
e
nt
to
ha
ve
a
da
pt
iv
e
pr
iv
a
c
y
m
ode
l
th
a
t
de
f
e
a
ts
a
tt
a
c
ks
f
r
om
a
dve
r
s
a
r
ie
s
.
T
h
e
pr
iv
a
c
y
m
ode
l
e
nha
nc
e
s
th
e
pr
iv
a
c
y
of
hum
a
ns
w
hi
le
pe
r
m
it
ti
ng
hi
ghl
y
a
c
c
ur
a
te
a
ppr
o
a
c
h
to
w
a
r
ds
a
c
ti
on
r
e
c
ogni
ti
on.
T
he
a
lg
or
it
hm
is
im
pl
e
m
e
nt
e
d
to
r
e
a
li
z
e
P
P
H
A
R
F
.
T
he
vi
s
ua
l
r
e
c
ogni
ti
on
of
hum
a
n
a
c
ti
ons
is
m
a
de
us
in
g
a
n
unde
r
ly
in
g
a
dve
r
s
a
r
ia
l
le
a
r
ni
ng
pr
oc
e
s
s
w
he
r
e
th
e
a
nonymi
z
a
ti
on
i
s
opt
i
m
iz
e
d
to
ha
ve
a
n
a
da
pt
iv
e
pr
iv
a
c
y
m
ode
l.
A
da
ta
s
e
t
na
m
e
d
H
M
D
B
51
is
us
e
d
f
or
e
m
pi
r
ic
a
l
s
tu
dy.
O
ur
e
xpe
r
im
e
nt
s
w
it
h
us
in
g
P
yt
hon
da
ta
s
c
ie
nc
e
pl
a
tf
or
m
r
e
ve
a
l
th
a
t
th
e
O
P
A
-
P
P
A
R
out
pe
r
f
o
r
m
s
e
xi
s
ti
ng
m
e
th
ods
.
I
t
c
a
n
be
us
e
d
in
r
e
a
l
w
or
ld
a
ppl
ic
a
ti
ons
w
he
r
e
P
P
H
A
R
F
c
a
n
f
it
s
e
a
m
le
s
s
ly
.
T
he
e
xpe
r
im
e
nt
a
l
r
e
s
ul
ts
r
e
ve
a
le
d
th
a
t
th
e
pr
opos
e
d
a
c
ti
on
r
e
c
ogni
ti
o
n
m
e
th
od
not
onl
y
pr
e
s
e
r
ve
s
pr
iv
a
c
y
a
nd
r
e
c
ogni
z
e
s
hum
a
n
a
c
ti
ons
but
a
ls
o
ha
s
opt
im
iz
a
ti
ons
in
te
r
m
s
of
pr
iv
a
c
y
budge
t
a
nd
a
c
om
bi
ne
d
lo
s
s
f
unc
ti
on
to
gui
de
th
e
r
e
c
ogni
ti
on
pr
oc
e
s
s
a
s
s
o
c
ia
te
d
w
it
h
th
e
pr
opos
e
d
f
r
a
m
e
w
or
k.
T
he
pr
opos
e
d
m
e
th
od
pa
ve
s
w
a
y
f
or
f
ur
th
e
r
in
v
e
s
ti
ga
ti
ons
in
te
r
m
s
of
opt
im
iz
in
g
th
e
th
r
e
e
m
ode
ls
i
nvol
ve
d i
n t
he
s
ys
te
m
.
R
E
F
E
R
E
N
C
E
S
[
1]
L
.
L
yu
e
t
al
.
,
“
T
ow
a
r
ds
f
a
i
r
a
nd
de
c
e
nt
r
a
l
i
z
e
d
pr
i
va
c
y
-
pr
e
s
e
r
vi
ng
de
e
p
l
e
a
r
ni
ng
w
i
t
h
bl
oc
kc
ha
i
n,”
C
oR
R
,
pp.
1
–
4,
J
un.
2019
,
[
O
nl
i
ne
]
. A
va
i
l
a
bl
e
:
ht
t
p:
/
/
a
r
xi
v.or
g/
a
bs
/
1906.01167.
[
2]
R
.
M
.
A
l
gul
i
ye
v,
R
.
M
.
A
l
i
gul
i
ye
v,
a
nd
F
.
J
.
A
bdul
l
a
ye
va
,
“
P
r
i
va
c
y
-
pr
e
s
e
r
v
i
ng
de
e
p
l
e
a
r
ni
ng
a
l
gor
i
t
hm
f
or
bi
g
pe
r
s
ona
l
da
t
a
a
na
l
ys
i
s
,”
J
our
nal
of
I
ndus
t
r
i
al
I
nf
or
m
at
i
on I
nt
e
gr
at
i
on
, vol
. 15, pp. 1
–
14, S
e
p.
2019, doi
:
10.1016/
j
.j
i
i
.2019.07.002.
[
3]
J
.
W
e
ng,
J
.
W
e
ng,
J
.
Z
ha
ng,
M
.
L
i
,
Y
.
Z
ha
ng,
a
nd
W
.
L
uo,
“
D
e
e
pC
ha
i
n:
a
udi
t
a
bl
e
a
nd
pr
i
va
c
y
-
pr
e
s
e
r
vi
ng
de
e
p
l
e
a
r
ni
ng
w
i
t
h
bl
oc
kc
ha
i
n
-
ba
s
e
d
i
nc
e
nt
i
ve
,”
I
E
E
E
T
r
ans
ac
t
i
ons
on
D
e
pe
ndabl
e
an
d
Se
c
ur
e
C
om
put
i
ng
,
pp.
1
–
1,
2019,
doi
:
10.1109/
T
D
S
C
.2019.2952332.
[
4]
L
.
L
yu
e
t
al
.
,
“
T
ow
a
r
ds
f
a
i
r
a
nd
pr
i
va
c
y
-
pr
e
s
e
r
vi
ng
f
e
de
r
a
t
e
d
de
e
p
m
ode
l
s
,
”
I
E
E
E
T
r
ans
ac
t
i
ons
on
P
ar
al
l
e
l
and
D
i
s
t
r
i
but
e
d
Sy
s
t
e
m
s
, vol
. 31, no. 11, pp. 2524
–
2541, N
ov. 2020, doi
:
10.1109/
T
P
D
S
.2020.2996273.
[
5]
K
.
V
.
K
um
a
r
,
D
.
J
.
H
a
r
i
ki
r
a
n,
M
.
A
.
R
.
P
r
a
s
a
d,
a
nd
U
.
S
i
r
i
s
ha
,
“
P
r
i
va
c
y
-
pr
e
s
e
r
vi
ng
hum
a
n
a
c
t
i
vi
t
y
r
e
c
ogni
t
i
on
a
nd
r
e
s
ol
ut
i
on
i
m
a
ge
us
i
ng
de
e
p
l
e
a
r
ni
ng
a
l
gor
i
t
hm
s
s
pa
t
i
a
l
r
e
l
a
t
i
ons
hi
p
a
nd
i
nc
r
e
a
s
i
ng
t
he
a
t
t
r
i
but
e
va
l
ue
i
n
O
pe
nC
V
,”
I
nt
e
r
nat
i
onal
J
our
nal
of
A
dv
anc
e
d Sc
i
e
nc
e
and T
e
c
hnol
ogy
, vol
. 29, no. 7, pp. 514
–
523, 2020.
[
6]
A
.
S
.
R
a
j
put
,
B
.
R
a
m
a
n,
a
nd
J
.
I
m
r
a
n,
“
P
r
i
va
c
y
-
pr
e
s
e
r
vi
ng
hum
a
n
a
c
t
i
on
r
e
c
ogni
t
i
on
a
s
a
r
e
m
ot
e
c
l
oud
s
e
r
vi
c
e
us
i
ng
R
G
B
-
D
s
e
ns
or
s
a
nd d
e
e
p C
N
N
,
”
E
x
pe
r
t
Sy
s
t
e
m
s
w
i
t
h A
ppl
i
c
at
i
ons
, vol
. 152, A
ug. 2020, doi
:
10.1016/
j
.e
s
w
a
.2020.113349.
[
7]
Y
.
W
u,
F
.
Y
a
ng,
Y
.
X
u,
a
nd
H
.
L
i
ng,
“
P
r
i
va
c
y
-
pr
ot
e
c
t
i
ve
-
G
A
N
f
o
r
pr
i
va
c
y
pr
e
s
e
r
vi
ng
f
a
c
e
de
-
i
de
nt
i
f
i
c
a
t
i
on,”
J
our
nal
of
C
om
put
e
r
Sc
i
e
n
c
e
and T
e
c
hnol
ogy
, vol
. 34, no. 1, pp. 47
–
60, J
a
n. 2019, doi
:
10.1007/
s
11390
-
019
-
1898
-
8.
[
8]
E
.
D
e
bi
e
,
N
.
M
ous
t
a
f
a
,
a
nd
M
.
T
.
W
hi
t
t
y,
“
A
pr
i
va
c
y
-
pr
e
s
e
r
vi
ng
ge
ne
r
a
t
i
ve
a
dve
r
s
a
r
i
a
l
ne
t
w
or
k
m
e
t
hod
f
or
s
e
c
ur
i
ng
E
E
G
b
r
a
i
n
s
i
gna
l
s
,”
J
ul
. 2020, doi
:
10.1109/
I
J
C
N
N
48605.2020.9206683.
[
9]
M
.
M
a
xi
m
ov,
I
.
E
l
e
z
i
,
a
nd
L
.
L
e
a
l
-
T
a
i
xe
,
“
C
I
A
G
A
N
:
c
ondi
t
i
ona
l
i
de
nt
i
t
y
a
nonym
i
z
a
t
i
on
ge
ne
r
a
t
i
ve
a
dve
r
s
a
r
i
a
l
n
e
t
w
or
ks
,”
i
n
2020
I
E
E
E
/
C
V
F
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.
2020,
pp.
5446
–
5455,
doi
:
10.1109/
C
V
P
R
42600.2020.00549.
[
10]
J
.
M
a
r
t
i
ns
s
on,
E
.
L
.
Z
e
c
,
D
.
G
i
l
l
bl
a
d,
a
nd
O
.
M
ogr
e
n,
“
A
dve
r
s
a
r
i
a
l
r
e
pr
e
s
e
nt
a
t
i
on
l
e
a
r
ni
ng
f
or
s
ynt
he
t
i
c
r
e
pl
a
c
e
m
e
nt
of
p
r
i
va
t
e
a
t
t
r
i
but
e
s
,”
J
un. 2020, [
O
nl
i
ne
]
. A
va
i
l
a
bl
e
:
ht
t
p:
/
/
a
r
xi
v.or
g/
a
bs
/
2006.08039.
[
11]
Q
.
L
i
,
Z
.
Z
he
ng,
F
.
W
u,
a
nd
G
.
C
he
n,
“
G
e
ne
r
a
t
i
ve
a
dve
r
s
a
r
i
a
l
ne
t
w
or
ks
-
ba
s
e
d
pr
i
va
c
y
-
pr
e
s
e
r
vi
ng
3D
r
e
c
ons
t
r
uc
t
i
on,”
i
n
2020
I
E
E
E
/
A
C
M
28t
h
I
nt
e
r
nat
i
onal
Sy
m
pos
i
um
on
Q
ual
i
t
y
of
S
e
r
v
i
c
e
(
I
W
Q
oS)
,
J
un.
2020,
pp.
1
–
10,
doi
:
10.1109/
I
W
Q
oS
49365.2020.9213037.
[
12]
S
.
S
hi
r
a
i
a
nd
J
.
W
hi
t
e
hi
l
l
,
“
P
r
i
va
c
y
-
p
r
e
s
e
r
vi
ng
a
nnot
a
t
i
on
of
f
a
c
e
i
m
a
ge
s
t
hr
ough
a
t
t
r
i
but
e
-
pr
e
s
e
r
vi
ng
f
a
c
e
s
ynt
he
s
i
s
,”
i
n
2019
I
E
E
E
/
C
V
F
C
onf
e
r
e
nc
e
on C
o
m
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. 2019, vol
. 2
019
-
J
une
, pp. 21
–
29,
doi
:
10.1109/
C
V
P
R
W
.2019.00009.
[
13]
A
. B
oul
e
m
t
a
f
e
s
,
A
. D
e
r
ha
b,
a
nd Y
.
C
ha
l
l
a
l
,
“
A
r
e
vi
e
w
of
pr
i
va
c
y
-
pr
e
s
e
r
vi
ng t
e
c
hni
que
s
f
or
de
e
p
l
e
a
r
ni
ng,”
N
e
ur
oc
om
put
i
ng
,
vol
.
384, pp. 21
–
45, A
pr
. 2020, doi
:
10.1016/
j
.ne
uc
om
.2019.11.041.
[
14]
M
.
M
a
l
e
kz
a
de
h,
R
.
G
.
C
l
e
gg,
a
nd
H
.
H
a
dda
di
,
“
R
e
pl
a
c
e
m
e
nt
a
ut
oe
nc
ode
r
:
a
pr
i
va
c
y
-
pr
e
s
e
r
vi
ng
a
l
gor
i
t
hm
f
or
s
e
ns
or
y
da
t
a
a
na
l
ys
i
s
,”
P
r
oc
e
e
di
ngs
-
A
C
M
/
I
E
E
E
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on
I
nt
e
r
ne
t
of
T
hi
ngs
D
e
s
i
gn
and
I
m
pl
e
m
e
nt
at
i
on,
I
oT
D
I
2018
,
pp.
165
–
176, O
c
t
.
2017, doi
:
10.1109/
I
oT
D
I
.2018.00025.
[
15]
R
.
Y
one
t
a
ni
,
V
.
N
.
B
odde
t
i
,
K
.
M
.
K
i
t
a
ni
,
a
nd
Y
.
S
a
t
o,
“
P
r
i
va
c
y
-
pr
e
s
e
r
vi
ng
vi
s
ua
l
l
e
a
r
ni
ng
us
i
ng
doubl
y
pe
r
m
ut
e
d
hom
om
or
phi
c
e
nc
r
ypt
i
on,”
i
n
P
r
oc
e
e
di
ngs
of
t
he
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
,
O
c
t
.
2017,
vol
.
2017
-
O
c
t
ob,
pp.
2059
–
2069, doi
:
10.1109/
I
C
C
V
.2017.225.
[
16]
M
.
D
e
r
Y
a
ng,
K
.
S
.
H
ua
ng,
Y
.
F
.
Y
a
ng,
L
.
Y
.
L
u,
Z
.
Y
.
F
e
ng,
a
nd
H
.
P
.
T
s
a
i
,
“
H
ype
r
s
pe
c
t
r
a
l
i
m
a
ge
c
l
a
s
s
i
f
i
c
a
t
i
on
us
i
ng
f
a
s
t
a
n
d
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
P
r
iv
ac
y
pr
e
s
e
r
v
in
g
hum
an ac
ti
v
it
y
r
e
c
ogni
ti
on f
r
am
e
w
or
k
u
s
in
g
…
(
K
am
bal
a V
ij
ay
a K
um
ar
)
263
a
da
pt
i
ve
bi
di
m
e
ns
i
ona
l
e
m
pi
r
i
c
a
l
m
ode
d
e
c
om
pos
i
t
i
on
w
i
t
h
m
i
ni
m
um
noi
s
e
f
r
a
c
t
i
on,”
I
E
E
E
G
e
os
c
i
e
nc
e
and
R
e
m
ot
e
Se
ns
i
n
g
L
e
t
t
e
r
s
, vol
. 13, no. 12, pp. 1950
–
1954, D
e
c
. 2016, doi
:
10.1109/
L
G
R
S
.2016.2
618930.
[
17]
W
.
D
u
e
t
al
.
,
“
A
ppr
oxi
m
a
t
e
t
o
be
gr
e
a
t
:
c
om
m
uni
c
a
t
i
on
e
f
f
i
c
i
e
nt
a
nd
pr
i
va
c
y
-
pr
e
s
e
r
vi
ng
l
a
r
ge
-
s
c
a
l
e
di
s
t
r
i
bu
t
e
d
de
e
p
l
e
a
r
ni
ng
i
n
i
nt
e
r
ne
t
of
t
hi
ngs
,”
I
E
E
E
I
n
t
e
r
ne
t
of
T
hi
ngs
J
ou
r
nal
,
vol
.
7,
n
o.
12,
pp.
11678
–
11692,
D
e
c
.
2020,
doi
:
10.1109/
J
I
O
T
.2020.2999594.
[
18]
J
.
J
ohns
on,
A
.
A
l
a
hi
,
a
nd
L
.
F
e
i
-
F
e
i
,
“
P
e
r
c
e
pt
ua
l
l
o
s
s
e
s
f
or
r
e
a
l
-
t
i
m
e
s
t
yl
e
t
r
a
ns
f
e
r
a
nd
s
upe
r
-
r
e
s
o
l
ut
i
on
,”
i
n
L
e
c
t
u
r
e
N
ot
e
s
i
n
C
om
put
e
r
Sc
i
e
nc
e
(
i
nc
l
udi
ng
s
ubs
e
r
i
e
s
L
e
c
t
u
r
e
N
ot
e
s
i
n
A
r
t
i
f
i
c
i
al
I
nt
e
l
l
i
ge
nc
e
and
L
e
c
t
ur
e
N
ot
e
s
i
n
B
i
oi
nf
or
m
at
i
c
s
)
,
vol
.
9906,
S
pr
i
nge
r
I
nt
e
r
na
t
i
ona
l
P
ubl
i
s
hi
ng, 2016, pp. 694
–
711.
[
19]
K
.
H
e
,
X
.
Z
ha
ng,
S
.
R
e
n,
a
nd
J
.
S
un,
“
D
e
e
p
r
e
s
i
dua
l
l
e
a
r
ni
ng
f
or
i
m
a
ge
r
e
c
ogni
t
i
on,”
i
n
2016
I
E
E
E
C
onf
e
r
e
nc
e
on
C
o
m
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. 2016, pp. 770
–
778, doi
:
10.1109/
C
V
P
R
.2016.90.
[
20]
H
.
K
ue
hne
,
H
.
J
hua
ng,
E
.
G
a
r
r
ot
e
,
T
.
P
oggi
o,
a
nd
T
.
S
e
r
r
e
,
“
H
M
D
B
:
a
l
a
r
ge
vi
de
o
da
t
a
ba
s
e
f
or
hum
a
n
m
ot
i
on
r
e
c
ogni
t
i
on,”
i
n
2011 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
, N
ov. 2011, pp. 2556
–
2563
, doi
:
10.1109/
I
C
C
V
.2011.6126543.
[
21]
K
.
Y
un,
J
.
H
ono
r
i
o,
D
.
C
ha
t
t
opa
dhya
y,
T
.
L
.
B
e
r
g,
a
nd
D
.
S
a
m
a
r
a
s
,
“
T
w
o
-
pe
r
s
on
i
nt
e
r
a
c
t
i
on
de
t
e
c
t
i
on
us
i
ng
body
-
pos
e
f
e
a
t
ur
e
s
a
nd
m
ul
t
i
pl
e
i
ns
t
a
nc
e
l
e
a
r
ni
ng,”
i
n
2012
I
E
E
E
C
om
put
e
r
Soc
i
e
t
y
C
onf
e
r
e
n
c
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. 2012, pp. 28
–
35, d
oi
:
10.1109/
C
V
P
R
W
.2012.6239234.
[
22]
K
.
H
e
,
X
.
Z
ha
ng,
S
.
R
e
n,
a
nd
J
.
S
un,
“
I
de
nt
i
t
y
m
a
ppi
ngs
i
n
de
e
p
r
e
s
i
dua
l
ne
t
w
or
ks
,”
L
e
c
t
ur
e
N
ot
e
s
i
n
C
om
put
e
r
Sc
i
e
nc
e
(
i
nc
l
udi
ng
s
ubs
e
r
i
e
s
L
e
c
t
u
r
e
N
ot
e
s
i
n
A
r
t
i
f
i
c
i
al
I
nt
e
l
l
i
ge
nc
e
and
L
e
c
t
u
r
e
N
ot
e
s
i
n
B
i
oi
nf
or
m
at
i
c
s
)
,
v
ol
.
9908,
pp.
630
–
645,
M
a
r
.
2016, doi
:
10.1007/
978
-
3
-
319
-
46493
-
0_38.
[
23]
C
.
S
z
e
ge
dy
e
t
al
.
,
“
G
oi
ng
de
e
pe
r
w
i
t
h
c
onvol
ut
i
ons
,”
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. 1
–
9, doi
:
10.1109
/
C
V
P
R
.2015.7298594.
[2
4]
R
.
L
e
e
ne
s
,
R
.
v
a
n
B
r
a
ke
l
,
S
.
G
ut
w
i
r
t
h,
a
nd
P
.
D
e
H
e
r
t
,
D
at
a
P
r
ot
e
c
t
i
on
and
P
r
i
v
ac
y
:
(
I
n)
v
i
s
i
bi
l
i
t
i
e
s
and
I
nf
r
a
s
t
r
uc
t
u
r
e
s
,
vol
.
36
.
C
ha
m
:
S
pr
i
nge
r
I
nt
e
r
na
t
i
ona
l
P
ubl
i
s
hi
ng, 2017.
[
25]
J
.
D
a
i
,
B
.
S
a
gha
f
i
,
J
.
W
u,
J
.
K
onr
a
d,
a
nd
P
.
I
s
hw
a
r
,
“
T
ow
a
r
ds
pr
i
va
c
y
-
pr
e
s
e
r
vi
ng
r
e
c
ogni
t
i
on
of
hum
a
n
a
c
t
i
vi
t
i
e
s
,”
i
n
2015
I
E
E
E
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on I
m
age
P
r
o
c
e
s
s
i
ng (
I
C
I
P
)
, S
e
p. 2015, pp. 4238
–
42
42, doi
:
10.1109/
I
C
I
P
.2015.7351605.
[
26]
T
.
O
r
e
kondy,
B
.
S
c
hi
e
l
e
,
a
nd
M
.
F
r
i
t
z
,
“
T
ow
a
r
ds
a
vi
s
ua
l
pr
i
va
c
y
a
dvi
s
or
:
und
e
r
s
t
a
ndi
ng
a
nd
pr
e
di
c
t
i
ng
pr
i
va
c
y
r
i
s
ks
i
n
i
m
a
ge
s
,”
i
n
P
r
oc
e
e
di
ngs
of
t
he
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
,
O
c
t
.
2017,
vol
.
2017
-
O
c
t
ob,
pp.
3706
–
3715,
doi
:
10.1109/
I
C
C
V
.2017.398.
[
27]
F
.
P
i
t
t
a
l
uga
,
S
.
J
.
K
oppa
l
,
S
.
B
.
K
a
ng,
a
nd
S
.
N
.
S
i
nha
,
“
R
e
v
e
a
l
i
ng
s
c
e
ne
s
by
i
nve
r
t
i
ng
s
t
r
uc
t
ur
e
f
r
om
m
ot
i
on
r
e
c
ons
t
r
uc
t
i
ons
,”
P
r
oc
e
e
di
ngs
of
t
he
I
E
E
E
C
om
put
e
r
Soc
i
e
t
y
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
,
pp.
145
–
154,
A
p
r
.
2019,
doi
:
10.1109/
C
V
P
R
.2019.00023.
[
28]
J
.
D
a
i
,
J
.
W
u,
B
.
S
a
gha
f
i
,
J
.
K
onr
a
d,
a
nd
P
.
I
s
hw
a
r
,
“
T
ow
a
r
ds
pr
i
va
c
y
-
pr
e
s
e
r
vi
ng
a
c
t
i
vi
t
y
r
e
c
ogni
t
i
on
us
i
ng
e
xt
r
e
m
e
l
y
l
o
w
t
e
m
por
a
l
a
nd
s
pa
t
i
a
l
r
e
s
ol
ut
i
on
c
a
m
e
r
a
s
,”
i
n
I
E
E
E
C
om
put
e
r
Soc
i
e
t
y
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. 2015, pp. 68
–
76, doi
:
10.1109/
C
V
P
R
W
.2015.7301356.
[
29]
A
.
D
os
ovi
t
s
ki
y
a
nd
T
.
B
r
ox,
“
I
nve
r
t
i
ng
v
i
s
ua
l
r
e
pr
e
s
e
nt
a
t
i
ons
w
i
t
h
c
onvol
ut
i
ona
l
ne
t
w
or
ks
,”
i
n
P
r
oc
e
e
di
ngs
of
t
he
I
E
E
E
C
om
put
e
r
Soc
i
e
t
y
C
onf
e
r
e
n
c
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
,
J
un.
2016,
pp.
4829
–
4837,
doi
:
10.1109/
C
V
P
R
.2016.522.
[
30]
L
.
L
yu,
X
.
H
e
,
Y
.
W
.
L
a
w
,
a
nd
M
.
P
a
l
a
ni
s
w
a
m
i
,
“
P
r
i
va
c
y
-
pr
e
s
e
r
vi
ng
c
ol
l
a
bor
a
t
i
ve
de
e
p
l
e
a
r
ni
ng
w
i
t
h
a
ppl
i
c
a
t
i
on
t
o
hum
a
n
a
c
t
i
vi
t
y
r
e
c
ogni
t
i
on,”
i
n
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on
I
nf
or
m
at
i
on
and
K
no
w
l
e
dge
M
anage
m
e
nt
,
P
r
oc
e
e
di
ng
s
,
N
ov.
2017,
pp
.
1219
–
1228, doi
:
10.1145/
3132847.3132990.
[
31]
P
.
W
e
i
nz
a
e
pf
e
l
,
H
.
J
é
gou,
a
nd
P
.
P
é
r
e
z
,
“
R
e
c
ons
t
r
uc
t
i
ng
a
n
i
m
a
ge
f
r
om
i
t
s
l
oc
a
l
de
s
c
r
i
pt
or
s
,”
i
n
P
r
oc
e
e
di
ngs
of
t
he
I
E
E
E
C
om
put
e
r
So
c
i
e
t
y
C
onf
e
r
e
n
c
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
,
J
un.
2011,
pp.
337
–
344,
doi
:
10.1109/
C
V
P
R
.2011.5995616.
[
32]
M
.
S
.
R
yoo,
B
.
R
ot
hr
oc
k,
C
.
F
l
e
m
i
ng,
a
nd
H
.
J
.
Y
a
ng,
“
P
r
i
va
c
y
-
pr
e
s
e
r
vi
ng
hum
a
n
a
c
t
i
vi
t
y
r
e
c
ogni
t
i
on
f
r
om
e
xt
r
e
m
e
l
ow
r
e
s
ol
ut
i
on,”
A
pr
. 2016, [
O
nl
i
ne
]
. A
va
i
l
a
bl
e
:
ht
t
p:
/
/
a
r
xi
v.or
g/
a
bs
/
1604.03196.
[
33]
A
.
M
a
he
ndr
a
n
a
nd
A
.
V
e
da
l
di
,
“
V
i
s
ua
l
i
z
i
ng
de
e
p
c
onvol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
ks
us
i
ng
na
t
ur
a
l
pr
e
-
i
m
a
ge
s
,”
I
nt
e
r
nat
i
onal
J
our
nal
of
C
om
put
e
r
V
i
s
i
on
, vol
. 120, no. 3, pp. 233
–
255, M
a
y 2016, doi
:
10.1007/
s
112
63
-
016
-
0911
-
8.
[
34]
Z
. W
. W
a
ng, V
. V
i
ne
e
t
, F
. P
i
t
t
a
l
uga
, S
. N
.
S
i
nha
, O
.
C
os
s
a
i
r
t
, a
nd S
. B
. K
a
ng,
“
P
r
i
va
c
y
-
pr
e
s
e
r
vi
ng a
c
t
i
on r
e
c
ogni
t
i
on us
i
ng c
ode
d
a
pe
r
t
ur
e
vi
de
os
,”
i
n
I
E
E
E
C
om
put
e
r
Soc
i
e
t
y
C
onf
e
r
e
n
c
e
on
C
om
put
e
r
V
i
s
i
on
a
nd
P
at
t
e
r
n
R
e
c
ogni
t
i
on
W
or
k
s
hops
,
J
un.
2019,
pp.
1
–
10, doi
:
10.1109/
C
V
P
R
W
.2019.00007.
[
35]
F
.
A
l
M
a
c
hot
,
M
.
R
.
E
l
koba
i
s
i
,
a
nd
K
.
K
ya
m
a
kya
,
“
Z
e
r
o
-
s
hot
hum
a
n
a
c
t
i
vi
t
y r
e
c
ogni
t
i
on
us
i
ng
non
-
vi
s
ua
l
s
e
ns
or
s
,”
Se
ns
or
s
,
vol
.
20, no. 3, p. 825, F
e
b. 2020, doi
:
10.3390/
s
20030825.
[
36]
F
.
P
i
t
t
a
l
uga
a
nd
S
.
J
.
K
oppa
l
,
“
P
r
i
va
c
y
pr
e
s
e
r
vi
ng
opt
i
c
s
f
or
m
i
ni
a
t
u
r
e
vi
s
i
on
s
e
ns
or
s
,”
i
n
P
r
oc
e
e
di
ngs
of
t
he
I
E
E
E
C
om
put
e
r
Soc
i
e
t
y
C
onf
e
r
e
n
c
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
, J
un. 2015, pp. 314
–
324, doi
:
10.1109/
C
V
P
R
.2015.7298628.
[
37]
F
. P
i
t
t
a
l
uga
a
nd S
. J
. K
oppa
l
, “
P
r
e
-
c
a
pt
ur
e
pr
i
va
c
y f
or
s
m
a
l
l
vi
s
i
on s
e
ns
or
s
,
”
IE
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
n
e
I
nt
e
l
l
i
ge
nc
e
, vol
. 39, no. 11, pp. 2215
–
2226, N
ov. 2017, doi
:
10.1109/
T
P
A
M
I
.2
016.2637354.
[
38]
C
.
Z
ha
ng,
Y
.
T
i
a
n,
a
nd
E
.
C
a
pe
z
ut
i
,
“
P
r
i
va
c
y
pr
e
s
e
r
vi
ng
a
ut
om
a
t
i
c
f
a
l
l
de
t
e
c
t
i
on
f
or
e
l
de
r
l
y
us
i
ng
R
G
B
D
c
a
m
e
r
a
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
(
i
nc
l
udi
ng
s
ubs
e
r
i
e
s
L
e
c
t
ur
e
N
ot
e
s
i
n
A
r
t
i
f
i
c
i
al
I
nt
e
l
l
i
ge
nc
e
and L
e
c
t
u
r
e
N
ot
e
s
i
n B
i
oi
nf
or
m
at
i
c
s
)
, vol
.
7382, no. 1, S
pr
i
nge
r
B
e
r
l
i
n H
e
i
de
l
be
r
g, 2012, pp. 625
–
633.
[
39]
S
.
N
.
S
ur
,
S
.
B
e
r
a
,
S
. S
hom
e
,
R
.
B
e
r
a
,
a
nd
B
.
M
a
j
i
,
“
T
a
r
ge
t
c
ha
r
a
c
t
e
r
i
z
a
t
i
on
us
i
ng
M
I
M
O
r
a
da
r
,”
I
nt
e
r
nat
i
onal
J
our
nal
on
Sm
ar
t
Se
ns
i
ng and I
nt
e
l
l
i
ge
nt
Sy
s
t
e
m
s
, vol
. 13, no. 1, pp. 1
–
8, 2020, doi
:
10.21307/
i
j
s
s
i
s
-
2019
-
013.
[
40]
M
i
r
a
n
K
i
m
,
X
i
a
oqi
a
n
J
i
a
ng,
“
H
E
A
R
:
H
um
a
n
A
c
t
i
on
R
e
c
ogni
t
i
on
vi
a
N
e
ur
a
l
N
e
t
w
r
oks
on
H
om
om
or
phi
c
a
l
l
y
E
nc
r
ypt
e
d
D
a
t
a
"
,
a
r
X
i
v:
2104.0916v1 [
c
s
.C
R
]
.
[
41]
B
. C
he
ng
e
t
al
.
,
“
R
obu
s
t
e
m
ot
i
on r
e
c
ogni
t
i
on
f
r
om
l
ow
qua
l
i
t
y a
nd
l
ow
bi
t
r
a
t
e
vi
de
o:
a
de
e
p l
e
a
r
ni
ng a
ppr
oa
c
h,”
i
n
2017 S
e
v
e
nt
h
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on
A
f
f
e
c
t
i
v
e
C
om
put
i
ng
and
I
nt
e
l
l
i
ge
nt
I
nt
e
r
ac
t
i
on
(
A
C
I
I
)
,
O
c
t
.
2017,
pp.
65
–
70,
doi
:
10.1109/
A
C
I
I
.2017.8273580.
[
42]
D
.
R
i
boni
a
nd
C
.
B
e
t
t
i
ni
,
“
C
O
S
A
R
:
hybr
i
d
r
e
a
s
oni
ng
f
or
c
ont
e
xt
-
a
w
a
r
e
a
c
t
i
vi
t
y
r
e
c
ogni
t
i
on,”
P
e
r
s
onal
and
U
bi
qui
t
ous
C
om
put
i
ng
, vol
. 15, no. 3, pp. 271
–
289, M
a
r
. 2011, doi
:
10.1007/
s
00779
-
010
-
0331
-
7.
[
43]
M
. X
u, A
.
S
ha
r
ghi
, X
. C
he
n,
a
nd D
.
J
. C
r
a
nda
l
l
, “
F
u
lly
-
c
oupl
e
d t
w
o
-
s
t
r
e
a
m
s
p
a
t
i
ot
e
m
por
a
l
ne
t
w
or
ks
f
or
e
xt
r
e
m
e
l
y l
ow
r
e
s
ol
ut
i
on
a
c
t
i
on
r
e
c
ogni
t
i
on,”
i
n
2018
I
E
E
E
W
i
nt
e
r
C
onf
e
r
e
nc
e
on
A
ppl
i
c
at
i
ons
of
C
om
put
e
r
V
i
s
i
on
(
W
A
C
V
)
,
M
a
r
.
2018,
pp.
1607
–
1615,
doi
:
10.1109/
W
A
C
V
.2018.00178.
[
44]
S
. Z
ol
f
a
gha
r
i
a
nd M
. R
.
K
e
yva
npour
,
“
S
A
R
F
:
s
m
a
r
t
a
c
t
i
vi
t
y r
e
c
ogni
t
i
on f
r
a
m
e
w
or
k i
n a
m
bi
e
nt
a
s
s
i
s
t
e
d l
i
vi
ng,”
i
n
P
r
o
c
e
e
di
ngs
of
t
he
2016
F
e
de
r
at
e
d
C
onf
e
r
e
nc
e
on
C
om
put
e
r
Sc
i
e
nc
e
and
I
nf
or
m
at
i
on
Sy
s
t
e
m
s
,
F
e
dC
SI
S
2016
,
O
c
t
.
2016,
pp.
1435
–
1443,
doi
:
10.15439/
2016F
132.
[
45]
B
.
D
.
Y
ou
n
e
t
al
.
,
“
S
t
a
t
i
s
t
i
c
a
l
he
a
l
t
h
r
e
a
s
oni
ng
of
w
a
t
e
r
-
c
ool
e
d
pow
e
r
ge
ne
r
a
t
or
s
t
a
t
or
ba
r
s
a
ga
i
ns
t
m
oi
s
t
ur
e
a
bs
or
pt
i
on,”
I
E
E
E
T
r
ans
ac
t
i
ons
on E
ne
r
g
y
C
onv
e
r
s
i
on
, vol
. 30, no. 4, pp. 1376
–
1385, D
e
c
. 2015,
doi
:
10.1109/
T
E
C
.2015.2444873.
Evaluation Warning : The document was created with Spire.PDF for Python.