I
A
E
S
I
n
t
e
r
n
at
i
on
al
Jou
r
n
al
o
f
A
r
t
i
f
i
c
i
a
l
I
n
t
e
l
l
i
ge
n
c
e
(
I
J
-
AI
)
V
ol
. 15, N
o. 1, F
e
br
ua
r
y 2026
, pp.
350
~
360
I
S
S
N
:
2252
-
8938
,
D
O
I
:
10.11591/
i
j
a
i
.v
15
.i
1
.pp
350
-
360
350
Jou
r
n
al
h
om
e
page
:
h
t
t
p
:
/
/
i
j
a
i
.i
ae
s
c
or
e
.c
om
H
u
m
an
ac
t
i
vi
t
y r
e
c
ogn
i
t
i
on
u
s
i
n
g s
e
l
e
c
t
i
ve
k
e
r
n
e
l
n
e
t
w
or
k
-
2D
c
on
vol
u
t
i
on
al
n
e
u
r
al
n
e
t
w
or
k
w
i
t
h
A
r
c
F
ac
e
l
os
s
B
an
u
s
h
r
i
S
r
i
n
i
vas
ai
ah
, Jagad
e
e
s
h
a R
am
e
gow
d
a
D
e
pa
r
t
m
e
nt
of
C
om
pu
t
e
r
S
c
i
e
nc
e
a
nd
E
ngi
ne
e
r
i
ng, I
m
pa
c
t
C
ol
l
e
ge
of
E
ngi
ne
e
r
i
ng a
nd
A
ppl
i
e
d
S
c
i
e
nc
e
s
,
B
e
nga
l
ur
u
, I
nd
i
a
D
e
pa
r
t
m
e
nt
of
C
om
pu
t
e
r
S
c
i
e
nc
e
a
nd
E
ngi
ne
e
r
i
ng,
V
i
s
ve
s
v
a
r
a
ya
T
e
c
hno
l
ogi
c
a
l
U
ni
ve
r
s
i
t
y, B
e
l
a
ga
vi
, I
nd
i
a
A
r
t
i
c
l
e
I
n
f
o
A
B
S
T
R
A
C
T
A
r
t
i
c
l
e
h
i
s
t
o
r
y
:
R
e
c
e
i
ve
d
M
a
r
7, 2025
R
e
vi
s
e
d
N
ov 3, 2025
A
c
c
e
pt
e
d
J
a
n 10, 2026
Hum
a
n
ac
t
ivi
t
y
recog
n
i
t
io
n
(HAR)
is
a
widely
adop
t
ed
t
ec
hn
iq
u
e
i
n
applicat
io
n
s
req
u
iri
n
g
acc
u
ra
t
e
ide
nt
ifica
t
io
n
of
hum
a
n
ac
t
io
n
s
.
However
,
HAR
approac
h
es
of
t
e
n
fa
ce
c
h
alle
n
ges
i
n
ge
n
eralizi
n
g
a
cross
co
m
plex
dat
ase
t
s
wi
th
mu
l
t
i
-
view
varia
t
io
n
s
,
res
u
l
t
i
n
g
i
n
red
u
ced
classifica
t
io
n
accu
racy
.
Exis
t
i
n
g
classifiers
face
s
h
or
t
co
m
i
n
gs
i
n
predic
t
i
n
g
hum
a
n
act
ivi
t
ies
d
u
e
t
o
th
e
prese
n
ce
of
irreleva
nt
video
fra
m
es
,
leadi
n
g
t
o
freq
u
e
nt
m
isclassifica
t
io
n
s
.
T
h
is
researc
h
proposes
a
selec
t
ive
ker
n
el
n
e
t
work
-
2D
con
vol
ut
io
n
al
ne
u
ral
n
e
t
work
wi
th
add
i
t
ive
a
n
g
u
l
ar
m
argi
n
loss
for
deep
face
recogn
i
t
io
n
(
SKN
-
2D
-
CNN
wi
th
ArcFace
lo
ss
)
t
o
recog
n
iz
e
hum
a
n
ac
t
ivi
t
y
effect
ively
.
SKN
dy
na
m
ically
adap
t
s
the
recep
t
ive
field
for
l
ear
n
i
n
g
mu
l
t
i
-
scale
spa
t
ial
fea
tu
res,
e
nh
a
n
ci
n
g
th
e
recog
n
i
t
io
n
of
i
nt
rica
t
e
hum
a
n
ac
t
ivi
t
ies
with
va
ryin
g
m
o
t
io
n
scales
.
I
n
th
e
e
m
beddi
n
g
space,
ArcFace
los
s
i
nt
rod
u
ces
an
a
n
g
u
lar
m
argi
n
pe
n
al
t
y
th
a
t
i
m
proves
i
nt
er
-
class
separabili
t
y
a
n
d
i
nt
ra
-
class
com
pac
tn
ess
for
recog
n
i
t
io
n
.
Toge
th
er,
th
e
proposed
m
e
th
od
m
i
n
i
m
izes
m
isclassifica
t
io
n
i
n
hum
a
n
ac
t
ivi
t
y
by
i
m
provi
n
g
the
rob
u
s
tn
ess
of
fea
tu
re
represe
nt
a
t
io
n.
Fea
tu
re
ex
t
rac
t
io
n
u
si
n
g
vis
u
al
geo
m
e
t
ry
gro
u
p
19
(
VGG19
)
cap
tu
res
spa
t
ial
fea
tu
res
like
edges
,
t
ex
tu
res
a
n
d
s
h
apes
fro
m
video
fram
es
,
e
nh
a
n
ci
n
g
th
e
m
odel’s
abili
t
y
t
o
di
ffere
nt
ia
t
e
be
t
wee
n
com
plex
hum
a
n
ac
t
ivi
t
ies.
T
h
e
proposed
m
e
th
od
ac
h
ieves
h
ig
h
acc
u
racy
of
99.
16
a
n
d
98.
75%
o
n
th
e
UCF101
a
n
d
HMDB
-
51
da
t
ase
t
s
,
respec
t
iv
ely,
wh
e
n
co
m
pared
wi
th
exis
t
i
n
g
m
e
th
ods
s
u
c
h
as
CNN
a
n
d
bidirec
t
io
n
al
ga
t
ed
recu
rre
nt
un
i
t
(BiGRU)
.
K
e
y
w
or
d
s
:
2D
c
onvo
l
u
t
i
ona
l
n
e
ur
a
l
ne
t
w
or
k
A
r
c
F
a
c
e
l
os
s
B
i
di
r
e
c
t
i
on
a
l
ga
t
e
d
r
e
c
ur
r
e
n
t
u
ni
t
H
um
a
n a
c
t
i
v
i
t
y r
e
c
ogni
t
i
on
S
e
l
e
c
t
i
ve
ke
r
ne
l
ne
t
w
or
k
This
is
an
open
access
article
under
the
CC
BY
-
SA
license
.
C
or
r
e
s
pon
di
n
g A
u
t
h
or
:
B
a
nus
hr
i
S
r
i
n
i
va
s
a
i
a
h
D
e
pa
r
t
m
e
n
t
of
C
om
put
e
r
S
c
i
e
nc
e
a
nd E
ngi
ne
e
r
i
ng, I
m
pa
c
t
C
ol
l
e
ge
o
f
E
ngi
ne
e
r
i
ng a
nd A
ppl
i
e
d S
c
i
e
nc
e
s
S
a
ha
ka
r
na
ga
r
,
B
e
nga
l
ur
u
-
560092,
I
ndi
a
E
m
a
i
l
:
ba
nus
hr
i
914@
gm
a
i
l
.c
om
1.
I
N
T
R
O
D
U
C
T
I
O
N
H
um
a
n
a
c
t
i
vi
t
y
r
e
c
ogn
i
t
i
on
(
H
A
R
)
ha
s
e
m
e
r
ge
d
a
s
a
n
a
c
t
i
ve
r
e
s
e
a
r
c
h
a
r
e
a
due
t
o
i
t
s
vi
t
a
l
r
ol
e
i
n
vi
de
o
unde
r
s
t
a
nd
i
ng
[
1
]
,
[
2
]
.
T
he
H
A
R
pr
oc
e
s
s
a
l
s
o
de
t
e
c
t
s
a
c
t
i
vi
t
i
e
s
ba
s
e
d
on
a
no
m
a
l
i
e
s
i
n
da
i
l
y
r
out
i
ne
s
,
w
hi
c
h
a
r
e
of
t
e
n
phys
i
c
a
l
l
y
l
i
nke
d
t
o
a
pe
r
s
on’
s
ps
yc
ho
l
ogi
c
a
l
s
t
a
t
e
a
nd
pe
r
s
ona
l
i
t
y
[
3
]
.
L
e
ve
r
a
gi
ng
t
he
s
e
a
c
t
i
on
-
s
pe
c
i
f
i
c
c
ha
r
a
c
t
e
r
i
s
t
i
c
s
e
nha
nc
e
s
bot
h
t
he
e
f
f
i
c
i
e
nc
y
a
nd
a
c
c
ur
a
c
y
of
t
he
a
c
t
i
v
i
t
y
r
e
c
ogn
i
t
i
on
pr
oc
e
s
s
.
F
ur
t
he
r
m
or
e
, t
he
hum
a
n body’
s
j
oi
n
t
s
a
r
e
hi
e
r
a
r
c
h
i
c
a
l
l
y or
ga
n
i
z
e
d, pr
ovi
d
i
ng a
s
t
r
uc
t
ur
e
d c
onne
c
t
i
v
i
t
y t
ha
t
i
s
e
f
f
e
c
t
i
ve
l
y
e
xpl
oi
t
e
d
us
i
ng
de
e
p
l
e
a
r
n
i
ng
(
D
L
)
t
e
c
hni
qu
e
s
[
4]
–
[
6]
.
V
i
de
o
-
ba
s
e
d
a
nd
s
e
ns
or
-
ba
s
e
d
t
e
c
hnol
og
i
e
s
r
e
pr
e
s
e
n
t
t
he
t
w
o
pr
i
m
a
r
y
a
ppr
oa
c
he
s
t
o
a
c
t
i
vi
t
y
r
e
c
ogn
i
t
i
on
s
ys
t
e
m
s
.
V
i
d
e
o
-
ba
s
e
d
s
ys
t
e
m
s
pr
oc
e
s
s
vi
s
ua
l
da
t
a
f
r
o
m
c
a
m
e
r
a
f
oo
t
a
ge
t
o
r
e
c
ogni
z
e
hu
m
a
n
a
c
t
i
vi
t
i
e
s
,
w
hi
l
e
s
e
ns
o
r
-
ba
s
e
d
s
ys
t
e
m
s
us
e
e
xt
e
r
na
l
s
e
ns
or
s
t
o c
a
p
t
ur
e
m
obi
l
i
t
y da
t
a
a
nd
m
oni
t
or
a
c
t
i
v
i
t
y pa
t
t
e
r
n
s
[
7]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
t
i
f
I
nt
e
l
l
I
S
S
N
:
2252
-
8938
H
um
an ac
t
i
v
i
t
y
r
e
c
ogni
t
i
on us
i
ng s
e
l
e
c
t
i
v
e
k
e
r
ne
l
ne
t
w
or
k
…
(
B
anus
hr
i
Sr
i
n
i
v
as
a
i
ah
)
351
D
L
t
e
c
hni
que
s
ha
ve
ou
t
pe
r
f
or
m
e
d
t
r
a
di
t
i
ona
l
ha
ndc
r
a
f
t
e
d
f
e
a
t
u
r
e
-
ba
s
e
d
m
e
t
hods
,
de
m
ons
t
r
a
t
i
ng
s
i
gni
f
i
c
a
n
t
s
uc
c
e
s
s
a
c
r
os
s
va
r
i
ous
c
om
put
e
r
vi
s
i
on
t
a
s
ks
[
8]
.
T
he
s
e
l
f
-
l
e
a
r
ni
ng
c
a
pa
b
i
l
i
t
y
of
D
L
ne
t
w
or
ks
e
na
bl
e
s
t
he
m
t
o
pr
oc
e
s
s
c
om
p
l
e
x
r
e
pr
e
s
e
nt
a
t
i
ons
o
f
vi
s
ua
l
da
t
a
,
m
a
k
i
ng
t
he
m
pa
r
t
i
c
ul
a
r
l
y
s
u
i
t
a
b
l
e
f
or
vi
de
o
-
ba
s
e
d
H
A
R
[
9]
,
[
10]
.
T
r
a
d
i
t
i
ona
l
m
a
c
h
i
ne
l
e
a
r
ni
ng
a
ppr
oa
c
he
s
r
e
qu
i
r
e
s
t
e
ps
s
uc
h
a
s
f
e
a
t
ur
e
e
x
t
r
a
c
t
i
on
a
nd
s
e
l
e
c
t
i
on
f
or
t
r
a
i
ni
ng,
w
he
r
e
a
s
m
ode
r
n
D
L
m
ode
l
s
u
t
i
l
i
z
e
k
e
r
ne
l
-
ba
s
e
d
f
i
l
t
e
r
s
t
o
pr
o
c
e
s
s
d
a
t
a
t
hr
ough
c
onvol
ut
i
on
ope
r
a
t
i
on
s
,
a
u
t
om
a
t
i
c
a
l
l
y
e
xt
r
a
c
t
i
ng
r
e
l
e
va
n
t
f
e
a
t
ur
e
s
[
11]
,
[
12]
.
H
ow
e
ve
r
,
e
xt
e
r
na
l
c
ondi
t
i
ons
s
uc
h
a
s
l
i
ght
i
ng
va
r
i
a
t
i
on
s
,
c
a
m
e
r
a
pos
i
t
i
oni
ng
a
nd
s
ubj
e
c
t
di
s
t
a
nc
e
c
a
n
ne
ga
t
i
ve
l
y
i
m
pa
c
t
r
e
c
ogni
t
i
on
pe
r
f
or
m
a
nc
e
[
13
]
.
A
dva
nc
e
d
D
L
a
ppr
oa
c
he
s
c
a
p
t
ur
e
s
pa
t
i
a
l
de
pe
nd
e
nc
i
e
s
by
l
e
a
r
ni
ng
hi
e
r
a
r
c
hi
c
a
l
f
e
a
t
ur
e
r
e
pr
e
s
e
n
t
a
t
i
ons
t
h
r
ough c
onvol
ut
i
ona
l
pr
o
c
e
s
s
e
s
t
ha
t
m
ode
l
bot
h g
l
oba
l
a
nd l
oc
a
l
s
pa
t
i
a
l
c
or
r
e
l
a
t
i
ons
be
t
w
e
e
n
pi
xe
l
s
.
A
ddi
t
i
ona
l
l
y,
ne
ur
a
l
-
c
on
t
r
ol
l
e
d
di
f
f
e
r
e
n
t
i
a
l
e
qua
t
i
on
ne
t
w
or
ks
a
r
e
c
a
pa
bl
e
of
c
a
p
t
ur
i
ng
i
nt
e
g
r
a
l
e
xpr
e
s
s
i
on
s
of
hum
a
n
a
c
t
i
vi
t
y,
c
ont
r
i
but
i
ng
t
o
i
m
pr
ove
d
m
ode
l
i
ng
[
14
]
,
[
15]
.
R
e
c
ogni
z
i
ng
vi
s
ua
l
l
y
s
i
m
i
l
a
r
a
c
t
i
v
i
t
i
e
s
r
e
m
a
i
ns
a
s
i
gn
i
f
i
c
a
nt
c
ha
l
l
e
nge
due
t
o
s
ubt
l
e
va
r
i
a
t
i
ons
i
n
hum
a
n
a
c
t
i
on
s
t
ha
t
r
e
p
r
e
s
e
nt
di
f
f
e
r
e
nt
be
ha
vi
or
s
.
A
c
c
ur
a
t
e
di
f
f
e
r
e
nt
i
a
t
i
on
i
s
c
r
i
t
i
c
a
l
f
or
r
e
l
i
a
bl
e
de
c
i
s
i
on
-
m
a
ki
ng
i
n
a
pp
l
i
c
a
t
i
ons
s
uc
h
a
s
s
ur
ve
i
l
l
a
nc
e
, w
he
r
e
a
s
m
i
s
c
l
a
s
s
i
f
yi
ng
s
i
m
i
l
a
r
a
c
t
i
ons
m
a
y r
e
s
u
l
t
i
n ove
r
l
ooki
ng a
bno
r
m
a
l
or
r
i
s
ky be
ha
v
i
or
.
A
hm
a
d
e
t
al
.
[
16
]
i
nt
r
oduc
e
d
a
c
onvol
ut
i
ona
l
ne
u
r
a
l
ne
t
w
or
k
(
C
N
N
)
c
om
b
i
ne
d
w
i
t
h
a
b
i
di
r
e
c
t
i
on
a
l
ga
t
e
d
r
e
c
ur
r
e
n
t
uni
t
(
B
i
G
R
U
)
f
or
H
A
R
us
i
ng
vi
s
ua
l
da
t
a
.
T
he
C
N
N
w
a
s
e
m
pl
oye
d
t
o
e
xt
r
a
c
t
de
e
p
f
e
a
t
u
r
e
s
f
r
o
m
f
r
a
m
e
s
e
que
nc
e
s
of
hu
m
a
n
a
c
t
i
v
i
t
y
v
i
de
o
s
.
T
he
m
os
t
s
i
gni
f
i
c
a
nt
f
e
a
t
ur
e
s
w
e
r
e
s
e
l
e
c
t
e
d
t
o
i
m
p
r
ove
pe
r
f
o
r
m
a
nc
e
, a
n
d
B
i
G
R
U
w
a
s
us
e
d t
o
l
e
a
r
n t
e
m
po
r
a
l
m
o
t
i
on
s
a
c
r
os
s
f
r
a
m
e
s
. T
h
i
s
a
p
pr
oa
c
h pr
i
m
a
r
i
l
y a
i
m
e
d t
o
e
nha
nc
e
c
l
a
s
s
i
f
i
c
a
t
i
on a
c
c
ur
a
c
y a
nd e
f
f
e
c
t
i
v
e
l
y
l
e
a
r
ni
ng
l
ong
-
dur
a
t
i
on t
e
m
po
r
a
l
a
c
t
i
ons
. S
i
nha
a
nd K
u
m
a
r
[
17]
pr
opos
e
d
a
H
A
R
f
r
a
m
e
w
or
k
f
o
c
us
e
d
on
i
m
p
r
ovi
ng
c
l
a
s
s
i
f
i
c
a
t
i
on
pe
r
f
o
r
m
a
nc
e
.
T
he
m
e
t
hod
i
nvol
ve
d
s
e
gm
e
nt
i
ng
i
m
a
ge
s
i
n
t
o
s
m
a
l
l
e
r
r
e
gi
ons
f
or
f
e
a
t
ur
e
e
xt
r
a
c
t
i
on,
w
he
r
e
g
r
e
y
l
e
ve
l
c
o
-
oc
c
ur
r
e
nc
e
m
a
t
r
i
x
(
G
L
C
M
)
a
nd
l
oc
a
l
gr
a
d
i
e
nt
t
hr
e
s
hol
d
pa
t
t
e
r
n
(
L
G
T
P
)
w
e
r
e
a
pp
l
i
e
d
f
or
f
e
a
t
ur
e
e
x
t
r
a
c
t
i
on,
a
nd
c
l
a
s
s
i
f
i
e
r
s
l
i
ke
B
i
G
R
U
, C
N
N
, a
nd
l
ong s
hor
t
-
t
e
r
m
m
e
m
or
y
(
L
S
T
M
)
w
e
r
e
ut
i
l
i
z
e
d t
o a
c
h
i
e
ve
a
c
c
ur
a
t
e
c
l
a
s
s
i
f
i
c
a
t
i
on.
K
us
hw
a
ha
e
t
al
.
[
18]
de
s
i
gne
d
a
de
e
p
C
N
N
ba
s
e
d
on
m
ul
t
i
-
s
c
a
l
e
pr
oc
e
s
s
i
ng
f
or
H
A
R
.
A
s
m
a
l
l
m
i
c
r
o
-
ne
t
w
or
k
w
a
s
i
nt
r
oduc
e
d
t
o
e
xt
r
a
c
t
e
xc
l
us
i
v
e
di
s
c
r
i
m
i
na
t
i
ve
f
e
a
t
u
r
e
s
of
hum
a
n
obj
e
c
t
s
s
uc
h
a
s
pos
e
,
or
i
e
nt
a
t
i
on,
a
nd
ob
j
e
c
t
s
i
z
e
.
H
ow
e
ve
r
,
C
N
N
s
s
t
r
uggl
e
t
o
pr
oc
e
s
s
l
a
r
ge
,
da
t
a
-
s
pe
c
i
f
i
c
hu
m
a
n
-
c
e
n
t
r
i
c
f
e
a
t
ur
e
s
,
w
hi
c
h
c
a
n
l
e
a
d
t
o
ove
r
f
i
t
t
i
ng
w
he
n
t
he
da
t
a
s
e
t
l
a
c
ks
d
i
ve
r
s
i
t
y
i
n
po
s
e
s
,
o
r
i
e
n
t
a
t
i
ons
,
or
obj
e
c
t
s
i
z
e
s
.
V
a
r
s
hne
y
a
nd
B
a
ka
r
i
ya
[
19
]
de
ve
l
ope
d
a
de
e
p
C
N
N
f
or
H
A
R
i
n
vi
de
o
s
e
que
n
c
e
s
by
i
nt
e
g
r
a
t
i
ng
m
u
l
t
i
pl
e
C
N
N
s
t
r
e
a
m
s
, i
nc
l
ud
i
ng s
pa
t
i
a
l
a
nd t
e
m
por
a
l
c
o
m
pone
nt
s
. T
he
s
pa
t
i
a
l
s
t
r
e
a
m
e
xt
r
a
c
t
s
a
c
t
i
v
i
t
y
r
e
pr
e
s
e
nt
a
t
i
ons
f
r
om
R
G
B
f
r
a
m
e
s
,
w
hi
l
e
t
he
t
e
m
por
a
l
s
t
r
e
a
m
c
a
pt
u
r
e
d
m
o
t
i
on
-
r
e
l
a
t
e
d
i
nf
o
r
m
a
t
i
on.
H
ow
e
ve
r
,
H
A
R
m
ode
l
s
f
a
c
e
d
l
i
m
i
t
a
t
i
ons
i
n
h
a
ndl
i
ng
e
nv
i
r
on
m
e
nt
a
l
va
r
i
a
t
i
ons
a
nd
oc
c
l
us
i
ons
,
a
s
t
he
y
r
e
l
i
e
d
on
f
i
xe
d
s
pa
t
i
a
l
a
nd
t
e
m
por
a
l
f
e
a
t
ur
e
s
.
A
hm
a
d
a
nd
W
u
[
20]
i
nt
r
oduc
e
d
s
pa
t
i
a
l
de
e
p
f
e
a
t
ur
e
s
i
nc
or
po
r
a
t
i
on
us
i
ng
a
m
ul
t
i
l
a
ye
r
G
R
U
f
or
H
A
R
.
T
hi
s
m
e
t
hod
e
xt
r
a
c
t
e
d
s
pa
t
i
a
l
a
nd
de
e
p
f
e
a
t
ur
e
s
f
r
om
f
r
a
m
e
s
e
que
nc
e
s
of
hu
m
a
n
a
c
t
i
v
i
t
y
vi
de
os
,
l
e
ve
r
a
g
i
ng
l
i
gh
t
w
e
i
ght
M
obi
l
e
N
e
t
V
2
m
ode
l
.
T
he
e
xt
r
a
c
t
e
d
f
e
a
t
ur
e
s
w
e
r
e
s
ubs
e
que
n
t
l
y
pa
s
s
e
d
t
hr
ough a
m
u
l
t
i
l
a
ye
r
G
R
U
, w
hi
c
h pr
oc
e
s
s
e
d da
t
a
s
e
que
nc
e
s
a
nd c
a
pt
ur
e
d
t
e
m
po
r
a
l
de
p
e
nde
nc
i
e
s
a
c
r
os
s
vi
de
o
f
r
a
m
e
s
.
T
he
e
x
i
s
t
i
ng
c
l
a
s
s
i
f
i
e
r
s
e
nc
ount
e
r
e
d
d
i
f
f
i
c
ul
t
i
e
s
i
n
a
c
c
ur
a
t
e
l
y
pr
e
d
i
c
t
i
ng
hu
m
a
n
a
c
t
i
v
i
t
i
e
s
due
t
o
t
he
pr
e
s
e
nc
e
of
i
r
r
e
l
e
v
a
nt
vi
de
o f
r
a
m
e
s
, l
e
a
di
ng t
o m
i
s
c
l
a
s
s
i
f
i
c
a
t
i
on. I
n or
de
r
t
o a
ddr
e
s
s
t
hi
s
c
ha
l
l
e
nge
, t
h
i
s
s
t
udy
pr
opos
e
s
a
s
e
l
e
c
t
i
ve
ke
r
n
e
l
ne
t
w
or
k
-
2D
c
onvol
ut
i
ona
l
n
e
ur
a
l
n
e
t
w
or
k w
i
t
h A
r
c
F
a
c
e
lo
s
s
(
S
K
N
-
2D
-
C
N
N
w
i
t
h
A
r
c
F
a
c
e
lo
s
s
)
by
i
nc
o
r
por
a
t
i
ng
a
dyna
m
i
c
ke
r
ne
l
s
e
l
e
c
t
i
on
m
e
t
hod.
I
n
c
ont
r
a
s
t
t
o
t
r
a
di
t
i
ona
l
C
N
N
s
,
t
he
s
e
l
e
c
t
i
ve
ke
r
ne
l
e
na
bl
e
d t
he
m
ode
l
t
o
a
da
pt
i
ve
l
y
a
d
j
us
t
i
t
s
r
e
c
e
pt
i
v
e
f
i
e
l
d
ba
s
e
d
on
i
nput
f
r
a
m
e
s
,
a
l
l
ow
i
ng
i
t
t
o
f
oc
us
on
t
he
m
os
t
i
nf
o
r
m
a
t
i
ve
s
p
a
t
i
a
l
f
e
a
t
u
r
e
s
w
hi
l
e
s
uppr
e
s
s
i
ng
i
r
r
e
l
e
va
n
t
ba
c
kgr
ound
i
nf
o
r
m
a
t
i
on.
A
ddi
t
i
ona
l
l
y,
t
he
i
nt
e
gr
a
t
i
on
o
f
A
r
c
F
a
c
e
l
os
s
e
nha
n
c
e
s
i
n
t
r
a
-
c
l
a
s
s
c
om
pa
c
t
ne
s
s
a
nd
i
n
t
e
r
-
c
l
a
s
s
s
e
pa
r
a
b
i
l
i
t
y,
r
e
s
ul
t
i
ng
i
n
m
or
e
d
i
s
c
r
i
m
i
na
t
i
ve
f
e
a
t
ur
e
r
e
pr
e
s
e
nt
a
t
i
ons
.
T
h
i
s
c
o
m
bi
na
t
i
on
e
ns
u
r
e
s
t
h
a
t
onl
y
t
he
m
os
t
di
s
t
i
nc
t
i
ve
a
nd
r
e
l
e
va
nt
vi
de
o
f
r
a
m
e
s
c
ont
r
i
but
e
t
o
a
c
t
i
vi
t
y
r
e
c
ogni
t
i
on,
t
he
r
e
by
i
m
pr
ovi
ng
t
he
ove
r
a
l
l
pe
r
f
or
m
a
nc
e
a
nd r
obu
s
t
ne
s
s
of
t
h
e
m
ode
l
c
om
pa
r
e
d t
o c
onve
n
t
i
ona
l
C
N
N
a
ppr
oa
c
he
s
.
S
pa
t
i
a
l
de
e
p f
e
a
t
u
r
e
i
nc
o
r
por
a
t
i
on
ut
i
l
i
z
i
ng a
m
u
l
t
i
l
a
y
e
r
G
R
U
i
s
us
e
d f
or
H
A
R
. T
hi
s
m
e
t
hod e
xt
r
a
c
t
s
s
pa
t
i
a
l
a
nd
de
e
p
f
e
a
t
ur
e
s
f
r
o
m
f
r
a
m
e
s
e
que
n
c
e
s
of
hu
m
a
n
a
c
t
i
vi
t
y
vi
de
os
by
l
e
ve
r
a
g
i
ng
t
he
l
i
gh
t
w
e
i
ght
M
obi
l
e
N
e
t
V
2
m
ode
l
.
T
he
e
xt
r
a
c
t
e
d
f
e
a
t
ur
e
s
a
r
e
s
ubs
e
que
nt
l
y
pa
s
s
e
d
t
hr
ough
a
m
ul
t
i
l
a
ye
r
G
R
U
,
w
hi
c
h
pr
oc
e
s
s
e
s
t
he
da
t
a
s
e
que
nc
e
a
nd
c
a
pt
ur
e
s
t
e
m
por
a
l
de
pe
nde
nc
i
e
s
a
c
r
o
s
s
vi
de
o
f
r
a
m
e
s
.
E
xi
s
t
i
ng
c
l
a
s
s
i
f
i
e
r
s
f
a
c
e
di
f
f
i
c
ul
t
i
e
s
i
n
pr
e
d
i
c
t
i
ng
hum
a
n
a
c
t
i
vi
t
i
e
s
due
t
o
i
r
r
e
l
e
va
n
t
vi
d
e
o
f
r
a
m
e
s
,
l
e
a
d
i
ng
t
o
i
na
c
c
u
r
a
t
e
c
l
a
s
s
i
f
i
c
a
t
i
on.
T
o
a
ddr
e
s
s
t
hi
s
i
s
s
ue
,
a
S
K
N
-
2D
-
C
N
N
w
i
t
h
A
r
c
F
a
c
e
l
os
s
i
s
pr
opo
s
e
d
by
i
nc
o
r
por
a
t
i
ng
a
dyna
m
i
c
ke
r
ne
l
s
e
l
e
c
t
i
on
m
e
t
hod.
I
n
t
r
a
d
i
t
i
ona
l
C
N
N
s
,
t
he
s
e
l
e
c
t
i
ve
k
e
r
ne
l
e
na
b
l
e
s
t
he
m
ode
l
t
o
a
da
pt
i
ve
l
y
a
dj
us
t
i
t
s
r
e
c
e
pt
i
ve
f
i
e
l
d
de
p
e
ndi
ng
on
t
he
i
npu
t
f
r
a
m
e
s
,
f
oc
us
i
ng
on
t
he
m
os
t
i
n
f
or
m
a
t
i
ve
s
p
a
t
i
a
l
f
e
a
t
ur
e
s
w
hi
l
e
s
uppr
e
s
s
i
ng
i
r
r
e
l
e
v
a
nt
ba
c
kgr
ound.
M
o
r
e
ove
r
,
i
n
t
e
gr
a
t
i
on
of
A
r
c
F
a
c
e
l
os
s
e
nha
nc
e
s
i
nt
r
a
-
c
l
a
s
s
c
om
pa
c
t
ne
s
s
a
nd
i
nt
e
r
-
c
l
a
s
s
s
e
pa
r
a
bi
l
i
t
y,
m
a
k
i
ng
t
h
e
l
e
a
r
ne
d
f
e
a
t
u
r
e
r
e
pr
e
s
e
n
t
a
t
i
ons
m
or
e
di
s
c
r
i
m
i
na
t
i
ve
.
T
hi
s
c
om
b
i
na
t
i
on
e
ns
u
r
e
s
t
ha
t
onl
y
t
he
m
os
t
d
i
s
t
i
nc
t
i
ve
a
nd
r
e
l
e
va
nt
vi
de
o
f
r
a
m
e
s
c
on
t
r
i
but
e
t
o
a
c
t
i
v
i
t
y
r
e
c
ogni
t
i
on,
t
he
r
e
by
e
nha
nc
i
ng
t
he
ove
r
a
l
l
pe
r
f
or
m
a
nc
e
a
nd
r
obus
t
ne
s
s
of
t
he
m
ode
l
c
o
m
pa
r
e
d
t
o
s
t
a
nda
r
d C
N
N
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
t
i
f
I
nt
e
l
l
,
V
ol
. 15, N
o. 1, F
e
br
ua
r
y 2026
:
350
-
360
352
T
he
ke
y c
ont
r
i
bu
t
i
on o
f
t
hi
s
r
e
s
e
a
r
c
h s
t
udy
is
a
s
f
o
l
l
ow
s
:
i)
F
e
a
t
ur
e
e
xt
r
a
c
t
i
on
us
i
ng
vi
s
ua
l
ge
o
m
e
t
r
y
gr
oup
19
(
VGG
19
)
e
xt
r
a
c
t
s
hi
e
r
a
r
c
hi
c
a
l
f
e
a
t
ur
e
s
f
r
om
de
e
p
l
a
ye
r
s
,
c
a
p
t
ur
i
ng
bot
h
l
ow
a
nd
hi
gh
-
l
e
ve
l
da
t
a
a
nd
r
e
duc
i
ng
vi
de
o
f
r
a
m
e
s
i
z
e
s
t
o
m
a
t
c
h
t
he
i
npu
t
di
m
e
ns
i
ons
.
ii)
I
n
t
r
a
di
t
i
ona
l
2D
-
C
N
N
,
S
K
N
i
s
i
nc
or
por
a
t
e
d
t
o
dyna
m
i
c
a
l
l
y
a
d
j
us
t
t
h
e
r
e
c
e
pt
i
ve
f
i
e
l
d,
a
l
l
ow
i
ng
t
he
m
ode
l
t
o
f
oc
u
s
on
t
he
m
os
t
r
e
l
e
va
nt
s
pa
t
i
a
l
f
e
a
t
u
r
e
s
by
a
voi
d
i
ng
i
r
r
e
l
e
va
nt
ba
c
kgr
ound
f
r
a
m
e
s
.
T
h
i
s
a
da
pt
a
bi
l
i
t
y
i
m
p
r
ove
s
t
h
e
ne
t
w
or
k’
s
a
b
i
l
i
t
y
t
o
i
de
nt
i
f
y
i
nt
r
i
c
a
t
e
hum
a
n
a
c
t
i
v
i
t
i
e
s
w
i
t
h
c
o
m
pl
e
x
s
pa
t
i
a
l
pa
t
t
e
r
n
s
, l
e
a
d
i
ng t
o e
nh
a
nc
e
d r
e
c
ogn
i
t
i
on a
c
c
ur
a
c
y.
i
i
i
)
T
he
A
r
c
F
a
c
e
l
os
s
f
unc
t
i
on
e
nha
nc
e
s
di
s
c
r
i
m
i
na
t
i
ve
a
b
i
l
i
t
y
by
i
nt
r
oduc
i
ng
a
n
a
ngul
a
r
m
a
r
gi
n,
w
hi
c
h
e
na
bl
e
s
gr
e
a
t
e
r
i
nt
e
r
-
c
l
a
s
s
s
e
pa
r
a
bi
l
i
t
y a
nd i
n
t
r
a
-
c
l
a
s
s
c
om
pa
c
t
ne
s
s
i
n t
he
f
e
a
t
ur
e
e
m
be
ddi
ng s
pa
c
e
. T
hi
s
r
e
s
ul
t
s
i
n m
or
e
a
c
c
ur
a
t
e
r
e
c
ogn
i
t
i
on
a
nd
r
obus
t
pe
r
f
or
m
a
nc
e
f
or
hu
m
a
n
a
c
t
i
v
i
t
i
e
s
w
i
t
h s
ubt
l
e
va
r
i
a
t
i
ons
be
t
w
e
e
n c
l
a
s
s
e
s
.
T
he
pa
pe
r
i
s
or
g
a
ni
z
e
d
a
s
f
ol
l
ow
s
.
S
e
c
t
i
on
2
de
t
a
i
l
s
t
he
f
un
c
t
i
on
i
ng
o
f
t
he
pr
opos
e
d
m
e
t
hodol
ogy.
S
e
c
t
i
on 3 pr
e
s
e
nt
s
t
he
r
e
s
ul
t
s
a
nd di
s
c
us
s
i
on, w
hi
l
e
s
e
c
t
i
on 4 c
onc
l
ude
s
t
he
r
e
s
e
a
r
c
h.
2.
P
R
O
P
O
S
E
D
M
E
T
H
O
D
T
he
S
K
N
-
2D
-
C
N
N
w
i
t
h
A
r
c
F
a
c
e
l
os
s
e
f
f
i
c
i
e
n
t
l
y
c
a
p
t
ur
e
s
m
ul
t
i
-
s
c
a
l
e
f
e
a
t
ur
e
s
,
e
nha
nc
i
ng
t
he
m
ode
l
’
s
a
bi
l
i
t
y t
o ha
ndl
e
a
c
t
i
v
i
t
i
e
s
w
i
t
h di
ve
r
s
e
s
pa
t
i
a
l
a
nd t
e
m
por
a
l
c
om
pl
e
xi
t
i
e
s
.
T
hi
s
ne
t
w
or
k
dyna
m
i
c
a
l
l
y
a
da
pt
s
t
o t
he
i
npu
t
da
t
a
a
nd
e
f
f
e
c
t
i
ve
l
y
pr
oc
e
s
s
e
s
s
pa
t
i
a
l
i
n
f
or
m
a
t
i
on t
o
r
e
c
ogn
i
z
e
a
c
t
i
v
i
t
i
e
s
w
hi
l
e
p
r
oduc
i
ng
hi
ghl
y
d
i
s
c
r
i
m
i
na
t
i
ve
f
e
a
t
u
r
e
e
m
be
ddi
ngs
,
t
he
r
e
by
i
m
pr
ov
i
ng
c
l
a
s
s
i
f
i
c
a
t
i
on
a
c
c
u
r
a
c
y.
I
ni
t
i
a
l
l
y
,
d
a
t
a
i
s
c
ol
l
e
c
t
e
d f
r
om
t
h
e
U
F
C
101 a
nd
H
M
D
B
-
51
da
t
a
s
e
t
s
a
nd pr
e
-
pr
oc
e
s
s
e
d
t
hr
ough
nor
m
a
l
i
z
a
t
i
on a
nd
r
e
m
ova
l
of
unw
a
nt
e
d
da
t
a
.
F
e
a
t
ur
e
e
x
t
r
a
c
t
i
on
us
i
ng
V
G
G
19
c
a
pt
ur
e
s
s
pa
t
i
a
l
f
e
a
t
ur
e
s
l
i
ke
e
dge
s
,
t
e
xt
u
r
e
s
a
nd
s
ha
pe
s
f
r
om
v
i
de
o
f
r
a
m
e
s
,
f
u
r
t
he
r
e
nha
nc
i
ng
t
he
m
od
e
l
’
s
cap
a
b
i
l
i
t
y
t
o
di
f
f
e
r
e
nt
i
a
t
e
be
t
w
e
e
n
c
o
m
pl
e
x
hum
a
n
a
c
t
i
v
i
t
i
e
s
. F
i
gur
e
1
i
l
l
us
t
r
a
t
e
s
t
he
p
i
pe
l
i
ne
of
t
he
pr
opos
e
d m
e
t
hodol
ogy.
F
i
gur
e
1.
P
i
pe
l
i
ne
o
f
pr
opos
e
d
m
e
t
hod
2.1.
D
at
as
e
t
T
he
vi
de
o f
r
a
m
e
s
r
e
qui
r
e
d f
or
hu
m
a
n a
c
t
i
vi
t
y c
l
a
s
s
i
f
i
c
a
t
i
on a
r
e
ga
t
he
r
e
d
f
r
om
t
w
o publ
i
c
l
y a
va
i
l
a
bl
e
da
t
a
s
e
t
s
na
m
e
l
y,
H
M
D
B
-
51
a
nd
U
C
F
101.
T
he
U
C
F
101
da
t
a
s
e
t
i
nc
l
ude
s
a
c
t
i
on
r
e
c
ogn
i
t
i
on
da
t
a
f
r
om
r
e
a
l
i
s
t
i
c
be
ha
vi
or
vi
d
e
os
w
i
t
h 101
a
c
t
i
v
i
t
y
c
a
t
e
gor
i
e
s
, c
ol
l
e
c
t
e
d
f
r
om
Y
ouT
ube
. M
os
t
of
t
he
vi
de
o c
l
i
ps
i
n t
he
H
M
D
B
-
51
da
t
a
s
e
t
a
r
e
s
our
c
e
d
f
r
o
m
m
ovi
e
s
,
w
i
t
h
a
s
m
a
l
l
e
r
por
t
i
on
e
x
t
r
a
c
t
e
d
f
r
o
m
publ
i
c
pl
a
t
f
o
r
m
s
s
uc
h
a
s
G
oogl
e
V
i
de
o, Y
ouT
ube
, a
nd
t
he
P
r
e
l
i
nge
r
A
r
c
h
i
ve
.
A
de
t
a
i
l
e
d e
xpl
a
na
t
i
on of
t
h
e
s
e
da
t
a
s
e
t
s
i
s
p
r
ovi
de
d
.
2.1.1. U
C
F
-
101 d
at
as
e
t
T
he
U
C
F
-
101
da
t
a
s
e
t
c
ont
a
i
ns
a
c
t
i
on
r
e
c
ogni
t
i
on
da
t
a
f
r
o
m
r
e
a
l
i
s
t
i
c
be
h
a
vi
or
v
i
de
os
ga
t
he
r
e
d
f
r
om
Y
ouT
ube
,
e
nc
om
pa
s
s
i
ng
101
a
c
t
i
vi
t
y
c
a
t
e
go
r
i
e
s
[
21
]
.
I
t
i
s
a
n
e
xt
e
ns
i
on
of
t
he
U
C
F
50
da
t
a
s
e
t
,
c
om
pr
i
s
i
ng
13,320
vi
de
os
a
c
r
os
s
t
he
s
e
101
c
a
t
e
gor
i
e
s
.
U
C
F
-
101
of
f
e
r
s
a
w
i
d
e
va
r
i
e
t
y
of
hu
m
a
n
a
c
t
i
ons
w
i
t
h
no
t
a
bl
e
va
r
i
a
t
i
ons
i
n
c
a
m
e
r
a
m
o
t
i
on,
t
a
r
ge
t
a
ppe
a
r
a
nc
e
,
t
a
r
ge
t
s
c
a
l
e
,
a
nd
v
i
e
w
i
ng
a
ng
l
e
s
.
T
he
101
a
c
t
i
on
c
a
t
e
gor
i
e
s
a
r
e
or
ga
ni
z
e
d i
n
t
o 25 gr
oups
, e
a
c
h c
ont
a
i
ni
ng 4
t
o 7 a
c
t
i
on vi
de
os
.
2.1.2. H
M
D
B
-
51 d
at
as
e
t
M
os
t
of
t
he
vi
d
e
o
c
l
i
p
s
i
n
t
he
H
M
D
B
-
51
da
t
a
s
e
t
[
22
]
a
r
e
s
our
c
e
d
f
r
o
m
m
ot
i
on
pi
c
t
u
r
e
s
,
w
i
t
h
a
s
m
a
l
l
e
r
por
t
i
on
ob
t
a
i
ne
d
f
r
om
publ
i
c
pl
a
t
f
o
r
m
s
l
i
ke
G
oogl
e
V
i
de
o,
Y
ouT
ube
, a
nd
t
h
e
P
r
e
l
i
nge
r
A
r
c
h
i
ve
.
T
he
da
t
a
s
e
t
c
on
t
a
i
ns
a
t
o
t
a
l
of
6,849
c
l
i
ps
.
T
he
a
c
t
i
v
i
t
i
e
s
a
r
e
c
a
t
e
gor
i
z
e
d
i
nt
o
f
i
ve
c
a
t
e
go
r
i
e
s
:
ge
ne
r
a
l
body
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
t
i
f
I
nt
e
l
l
I
S
S
N
:
2252
-
8938
H
um
an ac
t
i
v
i
t
y
r
e
c
ogni
t
i
on us
i
ng s
e
l
e
c
t
i
v
e
k
e
r
ne
l
ne
t
w
or
k
…
(
B
anus
hr
i
Sr
i
n
i
v
as
a
i
ah
)
353
m
ove
m
e
n
t
s
,
body
m
ove
m
e
nt
s
i
nvol
v
i
ng
obj
e
c
t
i
n
t
e
r
a
c
t
i
on,
ob
j
e
c
t
-
i
nf
l
ue
nc
e
d
f
a
c
i
a
l
be
ha
vi
o
r
,
ge
ne
r
a
l
f
a
c
i
a
l
be
ha
vi
or
,
a
nd
body
m
ove
m
e
nt
s
i
nvol
v
i
ng
hum
a
n
i
nt
e
r
a
c
t
i
on.
T
a
b
l
e
1
pr
ov
i
de
s
a
de
s
c
r
i
pt
i
on
of
t
he
U
C
F
-
101
a
nd
H
M
D
B
-
51
da
t
a
s
e
t
s
us
e
d
f
or
v
i
de
o
-
ba
s
e
d
H
A
R
,
de
t
a
i
l
i
ng
t
he
num
be
r
of
a
c
t
i
v
i
t
i
e
s
,
vi
de
o
c
l
i
p
s
,
da
t
a
s
our
c
e
s
, a
nd
t
r
a
i
n
i
ng
-
t
e
s
t
i
ng s
p
l
i
t
s
.
T
a
bl
e
1. D
e
t
a
i
l
e
d
de
s
c
r
i
pt
i
on of
da
t
a
s
e
t
s
us
e
d
f
or
vi
de
o
b
a
s
e
d H
A
R
D
a
t
a
s
e
t
H
M
D
B
-
51
U
C
F
-
101
N
um
be
r
of
a
c
t
i
vi
t
i
e
s
51
101
V
i
d
e
o
c
l
i
p
s
6766
13,320
T
r
a
i
ni
ng
s
et
70 vi
d
e
o c
l
i
ps
14,900
S
our
c
e
s
Y
ouT
ube
Y
ouT
ube
, M
ovi
e
c
l
i
ps
,
a
nd
G
oogl
e
v
i
d
e
o
T
e
s
t
s
e
t
30
v
i
d
e
o c
l
i
ps
6
,
360
2.2.
P
r
e
-
p
r
oc
e
s
s
i
n
g
T
hi
s
s
e
c
t
i
on
de
s
c
r
i
be
s
t
he
da
t
a
pr
e
-
pr
oc
e
s
s
i
ng
ph
a
s
e
,
w
he
r
e
a
s
l
i
di
ng
w
i
ndow
a
ppr
oa
c
h
i
s
c
ons
i
de
r
e
d
a
nd
nor
m
a
l
i
z
a
t
i
on
i
s
a
pp
l
i
e
d
t
o
e
qua
l
i
z
e
t
he
f
e
a
t
u
r
e
ve
c
t
or
.
T
o
a
voi
d
l
a
r
ge
-
s
c
a
l
e
f
e
a
t
ur
e
s
f
r
om
dom
i
na
t
i
ng
t
he
da
t
a
s
e
t
,
nor
m
a
l
i
z
a
t
i
on
i
s
us
e
d
t
o
put
a
l
l
f
e
a
t
ur
e
s
i
nt
o
a
c
om
pa
r
a
b
l
e
r
a
nge
.
M
i
n
-
m
a
x
nor
m
a
l
i
z
a
t
i
on
s
i
gn
i
f
i
c
a
nt
l
y
e
nha
nc
e
s
t
he
a
c
c
u
r
a
c
y
of
a
m
a
c
hi
ne
l
e
a
r
ni
ng
m
ode
l
by
s
c
a
l
i
ng
d
a
t
a
s
e
t
va
l
ue
s
t
o
t
he
[
0, 1]
r
a
nge
.
I
n
(
1)
r
e
p
r
e
s
e
n
t
s
t
he
t
r
a
ns
f
or
m
a
t
i
on f
unc
t
i
on.
∗
=
(
−
−
)
(
1)
W
he
r
e
,
,
de
not
e
t
he
o
r
i
g
i
na
l
, m
a
x
i
m
u
m
a
n
d m
i
n
i
m
um
v
a
l
u
e
s
o
f
t
he
g
i
ve
n
f
e
a
t
u
r
e
s
, r
e
s
pe
c
t
i
ve
l
y
.
W
he
r
e
∗
d
e
no
t
e
s
t
he
no
r
m
a
l
i
z
e
d
va
l
ue
o
f
,
s
c
a
l
e
d
b
e
t
w
e
e
n
t
h
e
r
a
nge
[
0
,1]
.
T
h
e
m
i
n
i
m
u
m
a
nd
m
a
x
i
m
um
va
l
ue
s
a
r
e
e
xt
r
a
c
t
e
d f
r
o
m
t
he
t
r
a
i
ni
ng s
e
t
a
nd
us
e
d t
o no
r
m
a
l
i
z
e
bo
t
h
t
h
e
t
r
a
i
n
i
ng
a
nd
t
e
s
t
i
ng
da
t
a
s
e
t
s
.
i)
I
n
H
A
R
us
i
ng
s
e
ns
or
s
,
t
he
s
l
i
di
ng
w
i
ndow
t
e
c
hn
i
que
ge
ne
r
a
t
e
s
t
i
m
e
-
s
e
r
i
e
s
da
t
a
,
w
he
r
e
d
e
pe
nde
nc
i
e
s
e
xi
s
t
be
t
w
e
e
n
pr
e
vi
ous
a
nd
r
e
c
e
nt
va
l
ue
s
.
A
n
e
f
f
e
c
t
i
ve
f
e
a
t
u
r
e
ge
ne
r
a
t
i
on
m
e
c
ha
ni
s
m
i
s
e
s
s
e
nt
i
a
l
i
n
H
A
R
t
o pr
e
c
i
s
e
l
y c
a
p
t
ur
e
t
e
m
po
r
a
l
de
pe
nde
n
c
i
e
s
.
ii)
T
he
s
l
i
di
ng
w
i
ndow
m
e
t
hod
gr
oups
s
e
n
s
or
r
e
a
di
ngs
,
w
i
t
h
e
a
c
h
w
i
ndow
c
ont
a
i
n
i
ng
m
u
l
t
i
pl
e
f
e
a
t
ur
e
s
f
r
om
t
he
s
a
m
e
t
i
m
e
s
t
e
p.
A
s
t
he
w
i
ndow
m
ove
s
f
or
w
a
r
d
by
s
t
e
ps
,
t
he
ne
xt
s
a
m
pl
e
i
nc
l
ude
s
r
e
a
di
ng
s
f
r
om
s
+1
to
s
+
w
+1
,
w
he
r
e
r
e
f
e
r
s
t
o
s
t
e
p s
i
z
e
a
nd w
de
not
e
s
t
he
w
i
ndow
s
i
z
e
.
i
i
i
)
T
o e
ns
ur
e
e
f
f
e
c
t
i
ve
da
t
a
ha
nd
l
i
ng,
t
he
t
e
s
t
s
e
t
i
s
pr
e
-
p
r
oc
e
s
s
e
d be
f
or
e
b
e
i
ng f
e
d
f
or
f
e
a
t
u
r
e
e
xt
r
a
c
t
i
on.
2.3.
F
e
at
u
r
e
e
xt
r
ac
t
i
on
T
he
V
G
G
19
i
s
a
t
ype
of
C
N
N
c
ons
i
s
t
i
ng
of
19
l
a
ye
r
s
,
i
nc
l
ud
i
ng
16
c
onvol
u
t
i
on
(
c
onv
)
l
a
ye
r
s
a
nd
3
f
ul
l
y
c
onne
c
t
e
d
(
F
C
)
l
a
ye
r
s
,
de
s
i
gne
d
t
o
c
l
a
s
s
i
f
y
H
A
R
i
nt
o
1000
c
a
t
e
gor
i
e
s
.
I
t
i
s
t
r
a
i
ne
d
on
i
m
a
ge
da
t
a
,
ut
i
l
i
z
i
ng
m
u
l
t
i
pl
e
3×
3
f
i
l
t
e
r
s
i
n
e
a
c
h
c
onv
l
a
ye
r
.
T
he
16
-
c
onv
l
a
ye
r
pe
r
f
or
m
s
f
e
a
t
u
r
e
e
x
t
r
a
c
t
i
on,
w
hi
l
e
t
he
f
i
na
l
3 l
a
ye
r
s
ha
ndl
e
c
l
a
s
s
i
f
i
c
a
t
i
on. F
e
a
t
u
r
e
s
a
r
e
e
xt
r
a
c
t
e
d i
n 5 gr
oups
, e
a
c
h f
ol
l
ow
e
d by a
m
a
x
-
pool
i
ng l
a
ye
r
.
T
he
m
ode
l
r
e
c
e
i
v
e
s
a
n
i
m
a
ge
o
f
t
he
s
i
z
e
224×
224
,
w
i
t
h
out
c
o
m
e
s
c
or
r
e
s
pondi
ng
t
o
t
he
a
c
t
i
on
r
e
c
ogni
t
i
on
l
a
be
l
[
23]
,
[
24
]
.
T
he
V
G
G
19
f
e
a
t
ur
e
s
a
s
t
r
uc
t
ur
e
d
de
s
i
gn
c
o
m
pr
i
s
i
ng
m
ul
t
i
pl
e
c
onne
c
t
e
d
c
onv
l
a
ye
r
s
a
nd
f
ul
l
y
c
onne
c
t
e
d
l
a
ye
r
s
.
I
t
i
nvo
l
ve
s
a
n
a
l
t
e
r
na
t
i
ve
a
r
r
a
nge
m
e
nt
of
c
onv
l
a
y
e
r
s
a
nd
non
-
l
i
ne
a
r
a
c
t
i
va
t
i
on
l
a
ye
r
s
,
out
pe
r
f
o
r
m
i
ng
s
i
ngl
e
-
l
a
ye
r
s
t
r
uc
t
ur
e
s
by
e
f
f
e
c
t
i
ve
l
y
e
x
t
r
a
c
t
i
ng
i
m
a
ge
f
e
a
t
u
r
e
s
.
M
a
x
-
poo
l
i
ng
i
s
us
e
d f
or
dow
n
s
a
m
pl
i
ng, w
hi
l
e
a
m
odi
f
i
e
d
r
e
c
t
i
f
i
e
d l
i
ne
a
r
uni
t
(
R
e
L
U
)
a
c
t
i
va
t
i
on f
unc
t
i
on s
e
l
e
c
t
s
t
he
l
a
r
ge
s
t
va
l
ue
s
w
i
t
h
i
n a
r
e
gi
on
a
s
pool
e
d
va
l
ue
s
.
T
he
dow
ns
a
m
p
l
i
ng
l
a
y
e
r
e
nha
nc
e
s
f
e
a
t
u
r
e
e
xt
r
a
c
t
i
on
by
r
e
t
a
i
ni
ng
ke
y
de
t
a
i
l
s
w
hi
l
e
r
e
duc
i
ng
t
he
num
b
e
r
of
pa
r
a
m
e
t
e
r
s
.
T
he
i
ni
t
i
a
l
16
-
c
onv
l
a
ye
r
s
of
V
G
G
19
e
xt
r
a
c
t
f
e
a
t
ur
e
s
,
w
hi
l
e
t
he
s
ubs
e
que
nt
t
w
o l
a
y
e
r
s
c
l
a
s
s
i
f
y t
he
m
. E
a
c
h f
e
a
t
ur
e
e
xt
r
a
c
t
i
on gr
oup i
s
s
e
pa
r
a
t
e
d by a
m
a
x
-
pool
i
ng l
a
ye
r
. A
f
t
e
r
f
e
a
t
ur
e
e
xt
r
a
c
t
i
on, da
t
a
i
s
pa
s
s
e
d t
o c
l
a
s
s
i
f
i
c
a
t
i
on
l
a
ye
r
s
f
or
H
A
R
i
de
nt
i
f
i
c
a
t
i
on.
2.4.
C
l
as
s
i
f
i
c
at
i
on
T
hi
s
s
e
c
t
i
on
i
n
t
r
oduc
e
s
t
he
S
K
N
-
2D
-
C
N
N
w
i
t
h
A
r
c
F
a
c
e
l
o
s
s
f
unc
t
i
on
,
w
hi
c
h
e
f
f
i
c
i
e
n
t
l
y
c
a
p
t
ur
e
s
m
ul
t
i
-
s
c
a
l
e
f
e
a
t
ur
e
s
,
a
l
l
ow
i
ng
t
he
m
ode
l
t
o
ha
ndl
e
a
c
t
i
vi
t
i
e
s
w
i
t
h
va
r
yi
ng
s
pa
t
i
a
l
a
nd
t
e
m
por
a
l
c
om
pl
e
xi
t
i
e
s
.
I
n
t
r
a
di
t
i
ona
l
2D
-
C
N
N
,
S
K
N
i
s
i
nc
o
r
por
a
t
e
d
t
o
e
na
b
l
e
t
he
ne
t
w
or
k
t
o
dyna
m
i
c
a
l
l
y
a
dj
us
t
i
t
s
r
e
c
e
p
t
i
ve
f
i
e
l
d
ba
s
e
d
on
i
nput
f
e
a
t
ur
e
s
t
h
r
ough
a
s
pl
i
t
,
f
us
e
,
a
nd
s
e
l
e
c
t
p
r
oc
e
s
s
.
I
n
t
he
s
pl
i
t
s
t
a
g
e
,
i
nput
f
e
a
t
u
r
e
m
a
ps
a
r
e
c
onvol
ve
d
us
i
ng m
ul
t
i
p
l
e
ke
r
ne
l
s
of
d
i
f
f
e
r
e
nt
s
i
z
e
s
,
s
uc
h
a
s
3×
3
a
nd
5×
5,
t
o
c
a
pt
ur
e
d
i
ve
r
s
e
s
pa
t
i
a
l
pa
t
t
e
r
ns
.
D
ur
i
ng t
he
f
us
e
s
t
a
ge
, t
he
s
e
out
put
s
a
r
e
i
nt
e
g
r
a
t
e
d us
i
ng gl
oba
l
a
ve
r
a
ge
pool
i
ng a
nd f
ul
l
y c
onne
c
t
e
d l
a
y
e
r
s
t
o
pr
oduc
e
c
om
pa
c
t
,
c
ha
nn
e
l
-
w
i
s
e
a
t
t
e
n
t
i
on
de
s
c
r
i
p
t
or
s
.
I
n
t
he
s
e
l
e
c
t
s
t
a
ge
,
a
s
o
f
t
a
t
t
e
n
t
i
on
m
e
c
ha
n
i
s
m
i
s
a
ppl
i
e
d
t
o s
e
l
e
c
t
t
he
m
os
t
a
ppr
opr
i
a
t
e
f
e
a
t
ur
e
s
a
c
r
os
s
t
he
d
i
f
f
e
r
e
n
t
ke
r
n
e
l
s
c
a
l
e
s
,
he
l
pi
ng c
a
pt
ur
e
f
i
ne
-
g
r
a
i
ne
d
i
nf
or
m
a
t
i
on
t
ha
t
d
i
s
t
i
ngui
s
he
s
i
nt
r
i
c
a
t
e
hum
a
n
a
c
t
i
vi
t
i
e
s
.
T
he
A
r
c
F
a
c
e
l
os
s
f
unc
t
i
on
i
s
i
nt
e
g
r
a
t
e
d
i
n
t
o
t
he
c
l
a
s
s
i
f
i
c
a
t
i
on
pha
s
e
t
o
e
nha
nc
e
t
he
d
i
s
c
r
i
m
i
na
t
i
ve
a
bi
l
i
t
y
of
t
he
m
ode
l
[
25]
.
I
ns
t
e
a
d
of
us
i
ng
E
uc
l
i
de
a
n
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
t
i
f
I
nt
e
l
l
,
V
ol
. 15, N
o. 1, F
e
br
ua
r
y 2026
:
350
-
360
354
di
s
t
a
nc
e
, A
r
c
F
a
c
e
l
os
s
i
nt
r
odu
c
e
s
a
n a
ngul
a
r
m
a
r
g
i
n pe
na
l
t
y be
t
w
e
e
n f
e
a
t
ur
e
e
m
be
ddi
ng
s
of
di
f
f
e
r
e
n
t
c
l
a
s
s
e
s
.
T
hi
s
m
a
r
g
i
n
i
m
pr
ove
s
i
n
t
e
r
-
c
l
a
s
s
s
e
pa
r
a
b
i
l
i
t
y
a
nd
i
nt
r
a
-
c
l
a
s
s
c
om
pa
c
t
ne
s
s
,
r
e
s
ul
t
i
ng
i
n
m
o
r
e
c
ons
i
s
t
e
nt
a
nd
r
obus
t
f
e
a
t
ur
e
l
e
a
r
ni
ng. T
oge
t
he
r
, t
he
pr
opos
e
d m
e
t
hod e
ns
ur
e
s
hi
gh r
e
c
ogni
t
i
on pe
r
f
o
r
m
a
nc
e
by f
oc
us
i
ng on
t
he
m
os
t
r
e
l
e
va
nt
s
pa
t
i
a
l
f
e
a
t
u
r
e
s
a
nd
ge
ne
r
a
t
i
ng
hi
gh
l
y
s
e
pa
r
a
bl
e
r
e
pr
e
s
e
n
t
a
t
i
ons
i
n
t
he
e
m
b
e
ddi
ng
s
pa
c
e
.
I
ni
t
i
a
l
l
y,
t
he
s
e
t
up
us
e
s
2
ke
r
ne
l
s
of
va
r
i
ous
s
i
z
e
s
but
i
s
e
a
s
i
l
y
e
xt
e
nda
b
l
e
t
o
m
u
l
t
i
pl
e
br
a
nc
he
s
.
T
he
a
ut
om
a
t
i
c
a
l
l
y
i
n
i
t
i
a
l
i
z
e
d
s
pl
i
t
f
e
a
t
u
r
e
m
a
p
∈
ℝ
′
×
′
×
′
c
onduc
t
s
2
t
r
a
ns
f
or
m
a
t
i
on
s
̃
:
→
̃
∈
ℝ
×
×
us
i
ng
ke
r
ne
l
s
of
s
i
z
e
3×
3
a
nd
5
×
5
.
B
ot
h
̃
a
nd
̂
ha
ve
e
f
f
i
c
i
e
n
t
l
y
g
r
oupe
d
c
onvo
l
ut
i
on
a
nd
ba
t
c
h
nor
m
a
l
i
z
a
t
i
on i
n
s
e
que
nc
e
.
T
he
f
us
e
ope
r
a
t
i
on
a
l
l
ow
s
ne
u
r
ons
f
o
r
a
d
j
us
t
i
ng
R
F
s
i
z
e
s
ba
s
e
d
on
t
he
c
ont
e
x
t
of
t
he
i
nput
s
t
i
m
ul
i
.
T
he
c
or
e
i
de
a
i
s
t
o
ut
i
l
i
z
e
ga
t
i
ng
m
e
c
ha
n
i
s
m
s
t
o
c
ont
r
o
l
t
he
f
l
ow
of
da
t
a
f
r
om
s
e
ve
r
a
l
br
a
nc
he
s
,
e
a
c
h
c
a
r
r
yi
ng
i
nf
or
m
a
t
i
on
a
t
d
i
f
f
e
r
e
nt
s
c
a
l
e
s
,
i
nt
o
t
he
ne
u
r
ons
of
s
ubs
e
que
nt
l
a
ye
r
s
.
T
he
i
n
i
t
i
a
l
f
us
i
on
o
f
ou
t
put
s
f
r
om
t
he
b
r
a
nc
he
s
i
s
pe
r
f
or
m
e
d
e
l
e
m
e
n
t
-
w
i
s
e
,
a
s
de
f
i
ne
d i
n (
2)
.
H
e
r
e
,
gl
ob
a
l
i
nf
o
r
m
a
t
i
on
i
s
e
m
be
dd
ed
by
a
ppl
yi
ng
gl
ob
a
l
a
ve
r
a
g
e
pool
i
ng
f
o
r
ge
ne
r
a
t
i
ng
c
ha
nne
l
-
w
i
s
e
s
t
a
t
i
s
t
i
c
s
,
r
e
p
r
e
s
e
nt
e
d
a
s
∈
ℝ
.
T
he
−
ℎ
e
l
e
m
e
n
t
of
i
s
e
va
l
ua
t
e
d by r
e
du
c
i
ng
a
c
r
os
s
t
h
e
s
pa
t
i
a
l
d
i
m
e
n
s
i
on
×
, a
s
e
xpr
e
s
s
e
d by (
3)
.
=
̃
+
̂
(
2)
=
(
)
=
1
×
∑
∑
=
1
(
,
)
=
1
(
3)
W
he
r
e
de
t
e
r
m
i
ne
s
f
e
a
t
ur
e
m
a
p
s
,
̃
+
̂
r
e
f
e
r
s
t
o
f
e
a
t
ur
e
m
a
ps
f
r
o
m
t
w
o
br
a
nc
he
s
of
di
f
f
e
r
e
n
t
ke
r
ne
l
s
i
z
e
s
,
r
e
pr
e
s
e
nt
s
a
c
ha
nne
l
,
a
nd
de
not
e
t
he
he
i
ght
a
nd w
e
i
ght
of
s
pa
t
i
a
l
di
m
e
ns
i
on
s
, a
nd
(
,
)
de
not
e
s
t
he
a
c
t
i
va
t
i
on
va
l
ue
a
t
s
pa
t
i
a
l
l
oc
a
t
i
on
(
,
)
i
n
c
ha
nne
l
of
t
he
f
e
a
t
ur
e
m
a
p.
A
ddi
t
i
ona
l
l
y,
a
c
om
pa
c
t
f
e
a
t
ur
e
ℝ
×
1
i
s
c
ons
t
r
uc
t
e
d
f
o
r
e
na
b
l
i
ng
a
da
p
t
i
ve
a
nd
pr
e
c
i
s
e
s
e
l
e
c
t
i
on
s
us
i
ng
F
C
l
a
ye
r
w
i
t
h
di
m
e
ns
i
ona
l
i
t
y
r
e
duc
t
i
on,
i
m
p
r
ovi
ng
e
f
f
i
c
i
e
nc
y.
T
he
R
e
L
U
f
unc
t
i
on
i
s
r
e
pr
e
s
e
nt
e
d
by
a
nd
ℝ
×
1
.
T
o
e
va
l
ua
t
e
t
he
i
m
pa
c
t
of
on
m
ode
l
e
f
f
e
c
t
i
ve
ne
s
s
,
a
de
c
r
e
a
s
e
d
r
a
t
i
o
i
s
i
nt
r
oduc
e
d
t
o
r
e
gu
l
a
t
e
i
t
s
va
l
u
e
s
a
s
e
xpr
e
s
s
e
d
i
n
(
4)
t
o
(
5
)
.
W
he
r
e
i
ndi
c
a
t
e
s
t
he
m
i
ni
m
um
di
s
t
a
nc
e
t
ypi
c
a
l
l
y
s
e
t
t
o
32.
A
s
of
t
a
t
t
e
n
t
i
on
m
e
c
ha
ni
s
m
i
s
ut
i
l
i
z
e
d t
hr
ough c
ha
nne
l
s
f
o
r
t
he
s
e
l
e
c
t
i
on
of
s
e
ve
r
a
l
s
pa
t
i
a
l
da
t
a
s
c
a
l
e
s
us
i
ng
f
e
a
t
ur
e
.
T
he
S
of
t
M
a
x
i
s
us
e
d
f
or
c
ha
nne
l
-
w
i
s
e
p
r
oc
e
s
s
, a
s
r
e
pr
e
s
e
nt
e
d
i
n (
6)
a
nd (
7)
.
=
ℱ
(
)
=
(
(
)
)
(
4)
=
(
,
)
(
5)
=
+
,
=
+
(
6)
=
.
̃
+
.
̂
,
+
=
1
(
7)
W
he
r
e
,
de
not
e
s
t
h
e
a
t
t
e
n
t
i
on
w
e
i
ght
f
or
c
ha
nne
l
i
n
b
r
a
nc
he
s
A
a
nd
B
,
a
nd
r
e
pr
e
s
e
nt
s
E
ul
e
r
’
s
num
be
r
.
E
a
c
h
bl
oc
k
ha
s
2
c
onv
l
a
ye
r
s
,
w
he
r
e
t
h
e
i
ni
t
i
a
l
l
a
ye
r
e
m
pl
oy
s
1×
1
ke
r
ne
l
s
t
o
e
nha
nc
e
f
e
a
t
ur
e
r
e
pr
e
s
e
n
t
a
t
i
on,
w
hi
l
e
t
he
s
e
c
ond
l
a
ye
r
u
s
e
s
us
e
r
ke
r
n
e
l
s
o
f
s
i
z
e
3
t
o
a
ggr
e
ga
t
e
c
ons
e
c
u
t
i
ve
i
nf
or
m
a
t
i
on
a
nd
e
xt
r
a
c
t
i
nc
r
e
a
s
i
ng
l
y
c
o
m
pl
i
c
a
t
e
d
r
e
l
a
t
i
ons
.
E
a
c
h
c
onv
l
a
ye
r
i
s
f
ol
l
ow
e
d
us
i
ng
ba
t
c
h
nor
m
a
l
i
z
a
t
i
on
a
nd
a
R
e
L
U
a
c
t
i
va
t
i
on.
A
f
t
e
r
e
a
c
h
2
c
on
v
bl
o
c
ks
,
a
F
C
l
a
ye
r
w
i
t
h
dr
opou
t
i
s
a
pp
l
i
e
d t
o
ge
ne
r
a
t
e
s
e
que
nt
i
a
l
f
e
a
t
ur
e
ve
c
t
or
(
,
)
∈
ℝ
.
T
he
c
onc
a
t
e
na
t
e
d
s
e
que
n
t
i
a
l
ve
c
t
or
(
,
)
w
i
t
h
us
e
r
e
m
be
dd
i
ng
t
he
n
c
onne
c
t
s
t
he
out
c
om
e
s
l
a
ye
r
w
i
t
h
|
|
node
t
o t
he
l
o
s
s
f
unc
t
i
on.
T
he
2D
-
C
N
N
w
i
t
h
A
r
c
F
a
c
e
f
unc
t
i
on
e
ns
ur
e
s
t
ha
t
e
ve
r
y
c
l
a
s
s
e
l
i
m
i
na
t
e
s
i
r
r
e
l
e
va
nt
f
e
a
t
ur
e
s
,
e
nha
nc
i
ng
d
i
s
c
r
i
m
i
na
t
i
ve
a
b
i
l
i
t
y
of
l
e
a
r
n
e
d
e
m
be
ddi
ng
f
e
a
t
ur
e
dur
i
ng
c
l
a
s
s
i
f
i
c
a
t
i
on.
T
he
2D
-
C
N
N
m
ode
l
e
va
l
ua
t
e
s
s
i
m
i
l
a
r
i
t
y
be
t
w
e
e
n
t
h
e
i
nput
f
e
a
t
u
r
e
e
m
be
dd
i
ng
s
a
nd
a
l
l
c
l
a
s
s
es
i
s
e
va
l
u
a
t
e
d
ut
i
l
i
z
i
ng
c
o
s
i
ne
s
i
m
i
l
a
r
i
t
y
s
c
or
e
f
o
r
pr
e
d
i
c
t
i
on.
A
r
c
F
a
c
e
l
os
s
f
unc
t
i
on
i
n
t
r
oduc
e
s
a
n
a
ngul
a
r
m
a
r
gi
n
t
o
t
he
c
os
i
ne
s
i
m
i
l
a
r
i
t
y
s
c
or
e
of
t
he
gr
ound
t
r
u
t
h
c
l
a
s
s
,
w
hi
c
h
i
nc
r
e
a
s
e
s
a
ngul
a
r
s
e
p
a
r
a
t
i
on
a
m
ong
d
i
f
f
e
r
e
nt
c
l
a
s
s
e
m
be
ddi
ngs
.
T
hi
s
m
a
r
gi
n
i
s
i
nc
o
r
por
a
t
e
d
i
nt
o
t
he
l
og
-
S
o
f
t
M
a
x
f
or
m
ul
a
t
i
on
t
o
nor
m
a
l
i
z
e
s
i
m
i
l
a
r
i
t
y
s
c
o
r
e
s
,
p
r
oduc
i
ng
a
va
l
i
d
pr
oba
bi
l
i
t
y
di
s
t
r
i
but
i
on
a
c
r
os
s
a
l
l
c
l
a
s
s
e
s
.
B
y
f
oc
us
i
ng
on
a
ngul
a
r
r
e
l
a
t
i
ons
h
i
ps
r
a
t
he
r
t
ha
n
t
he
a
bs
o
l
ut
e
m
a
gni
t
ud
e
s
of
e
m
be
ddi
ngs
,
A
r
c
F
a
c
e
l
os
s
s
i
m
p
l
i
f
i
e
s
t
he
opt
i
m
i
z
a
t
i
on
pr
oc
e
s
s
by
l
e
ve
r
a
gi
ng
c
os
i
ne
s
i
m
i
l
a
r
i
t
y
s
c
or
e
s
,
w
hi
c
h
a
r
e
i
nhe
r
e
n
t
l
y
bounde
d
b
e
t
w
e
e
n
-
1
a
nd
1,
m
a
k
i
ng
t
r
a
i
n
i
ng
m
o
r
e
e
f
f
e
c
t
i
ve
a
nd
s
t
a
b
l
e
.
D
ur
i
ng
t
r
a
i
n
i
ng,
t
he
m
ode
l
l
e
a
r
ns
t
o
s
e
pa
r
a
t
e
e
m
be
ddi
ngs
of
s
e
ve
r
a
l
c
l
a
s
s
e
s
i
n
a
ngu
l
a
r
s
pa
c
e
,
r
e
s
u
l
t
i
ng i
n
pr
e
di
c
t
i
ons
t
ha
t
a
r
e
p
r
e
dom
i
na
nt
l
y
di
s
t
r
i
but
e
d
be
t
w
e
e
n
0
a
nd
1.
T
hi
s
a
pp
r
oa
c
h i
m
p
r
ove
s
t
he
r
e
l
i
a
bi
l
i
t
y
of
t
he
f
uz
z
y l
ogi
c
s
ys
t
e
m
a
s
e
xpr
e
s
s
e
d
i
n
(
8)
.
W
he
r
e
r
e
pr
e
s
e
nt
s
t
he
t
e
m
pe
r
a
t
ur
e
,
w
hi
c
h
r
e
gu
l
a
t
e
s
t
he
pr
oba
b
i
l
i
t
y
di
s
t
r
i
but
i
on
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
t
i
f
I
nt
e
l
l
I
S
S
N
:
2252
-
8938
H
um
an ac
t
i
v
i
t
y
r
e
c
ogni
t
i
on us
i
ng s
e
l
e
c
t
i
v
e
k
e
r
ne
l
ne
t
w
or
k
…
(
B
anus
hr
i
Sr
i
n
i
v
as
a
i
ah
)
355
a
t
out
put
,
de
not
e
s
i
nput
f
or
ℎ
c
l
a
s
s
,
a
nd
de
t
e
r
m
i
ne
s
t
he
t
o
t
a
l
nu
m
be
r
of
c
l
a
s
s
e
s
.
T
he
hi
ghe
r
va
l
u
e
s
of
m
a
ke
t
he
ou
t
put
l
a
ye
r
d
i
s
t
r
i
bu
t
i
on c
l
os
e
r
t
o a
uni
f
or
m
d
i
s
t
r
i
bu
t
i
on.
(
)
=
/
∑
/
=
1
(
8)
3.
E
X
P
E
R
I
M
E
N
T
A
L
A
N
A
L
Y
S
I
S
T
hi
s
s
e
c
t
i
on
pr
e
s
e
n
t
s
a
n
e
va
l
ua
t
i
on
o
f
t
he
pr
opos
e
d
2D
-
C
N
N
w
i
t
h
A
r
c
F
a
c
e
l
os
s
f
unc
t
i
on
s
i
m
u
l
a
t
e
d
us
i
ng
P
yt
hon
ve
r
s
i
on
3.11,
W
i
ndow
s
10
ope
r
a
t
i
ng s
ys
t
e
m
(
O
S
)
,
6
G
B
G
P
U
,
16
G
B
R
A
M
,
i
7
pr
oc
e
s
s
o
r
,
a
nd
1
G
B
m
e
m
or
y.
T
he
pr
opos
e
d
m
e
t
hodol
ogy
i
s
a
s
s
e
s
s
e
d
ba
s
e
d
on
t
he
pe
r
f
or
m
a
nc
e
m
e
a
s
u
r
e
s
o
f
F1
-
s
c
or
e
,
a
c
c
ur
a
c
y,
r
e
c
a
l
l
,
a
nd
pr
e
c
i
s
i
on
,
a
s
s
how
n
i
n
(
9)
t
o
(
12)
.
W
he
r
e
de
not
e
s
t
r
ue
ne
ga
t
i
ve
,
de
not
e
s
t
r
ue
pos
i
t
i
ve
,
de
not
e
s
f
a
l
s
e
ne
ga
t
i
ve
, a
nd
de
not
e
s
f
a
l
s
e
po
s
i
t
i
ve
.
=
+
+
+
+
(
9)
=
+
(
10)
=
+
(
11)
1
−
=
2
×
×
+
(
12)
3.1.
P
e
r
f
or
m
an
c
e
an
al
ys
i
s
T
a
bl
e
2
pr
e
s
e
nt
s
a
n
e
va
l
ua
t
i
on
of
di
f
f
e
r
e
n
t
f
e
a
t
ur
e
e
xt
r
a
c
t
i
on
m
e
t
hods
a
c
r
os
s
bot
h
da
t
a
s
e
t
s
.
T
he
pe
r
f
or
m
a
nc
e
a
n
a
l
ys
i
s
o
f
V
G
G
19
i
s
p
r
ove
n
m
or
e
e
f
f
i
c
i
e
n
t
i
n
c
a
p
t
ur
i
ng
c
o
m
pl
e
x
pa
t
t
e
r
ns
a
nd
e
x
t
r
a
c
t
i
ng
hi
e
r
a
r
c
h
i
c
a
l
f
e
a
t
u
r
e
ba
s
e
on
t
he
da
t
a
s
e
t
.
T
he
f
e
a
t
ur
e
e
x
t
r
a
c
t
i
on
pe
r
f
or
m
a
nc
e
of
t
h
e
pr
opos
e
d
m
ode
l
i
s
c
om
pa
r
e
d
w
i
t
h e
xi
s
t
i
ng
m
e
t
hods
l
i
ke
E
f
f
i
c
i
e
n
t
N
e
t
,
M
ob
i
l
e
N
e
t
,
R
e
s
N
e
t
50
,
a
nd
V
G
G
16.
T
he
s
ugge
s
t
e
d m
ode
l
a
c
hi
e
ve
s
a
be
t
t
e
r
a
c
c
ur
a
c
y of
99.16 a
nd 98.75
%
on t
he
U
F
C
101 a
nd
H
M
D
B
-
51
da
t
a
s
e
t
s
,
r
e
s
pe
c
t
i
ve
l
y
.
T
a
bl
e
2. P
e
r
f
or
m
a
nc
e
e
v
a
l
ua
t
i
on of
d
i
f
f
e
r
e
nt
f
e
a
t
ur
e
e
x
t
r
a
c
t
i
on m
e
t
hods
on bo
t
h da
t
a
s
e
t
M
e
t
hod
s
D
a
t
a
s
e
t
A
c
c
ur
a
c
y (
%
)
P
r
e
c
i
s
i
on (
%
)
R
e
c
a
l
l
(
%
)
F1
-
s
c
or
e
(
%
)
E
f
f
i
c
i
e
nt
N
e
t
U
F
C
101
92.14
91.67
90.82
91.24
H
M
D
B
-
51
89.24
88.76
87.58
88.17
M
obi
l
e
N
e
t
U
F
C
101
90.67
90.21
89.15
89.68
H
M
D
B
-
51
86.53
85.89
85.23
85.56
R
e
s
N
e
t
50
U
F
C
101
94.25
93.87
92.96
93.41
H
M
D
B
-
51
90.56
89.98
89.12
89.54
V
G
G
16
U
F
C
101
93.74
93.12
92.25
92.68
H
M
D
B
-
51
88.94
88.32
87.76
88.04
P
r
opos
e
d
V
G
G
19
U
F
C
101
99.16
99.10
99.11
99.10
H
M
D
B
-
51
98.75
99.41
96.03
97.54
T
he
pe
r
f
or
m
a
nc
e
a
na
l
ys
i
s
of
t
h
e
c
l
a
s
s
i
f
i
c
a
t
i
on
us
i
ng
S
K
N
-
2D
-
C
N
N
i
s
c
om
pa
r
e
d
w
i
t
h
e
xi
s
t
i
ng
m
e
t
hods
s
uc
h
a
s
m
u
l
t
i
l
a
ye
r
pe
r
c
e
p
t
r
on
(
M
L
P
)
,
de
e
p
ne
ur
a
l
ne
t
w
or
k
(
D
N
N
)
,
R
e
s
N
e
t
,
a
nd
V
G
G
.
C
l
a
s
s
i
f
i
c
a
t
i
on
us
i
ng
S
K
N
-
2D
-
C
N
N
a
c
hi
e
ve
s
hi
ghe
r
a
c
c
ur
a
c
y
of
99.16
a
nd
98.75%
on
t
he
U
C
F
101
a
nd
H
M
D
B
-
51
da
t
a
s
e
t
s
,
r
e
s
pe
c
t
i
ve
l
y,
a
s
s
how
n
i
n
T
a
bl
e
3.
T
he
A
r
c
F
a
c
e
l
os
s
out
pe
r
f
or
m
s
t
r
a
di
t
i
ona
l
l
os
s
f
unc
t
i
on
s
by
e
n
s
ur
i
ng
t
ha
t
e
m
be
ddi
ng
s
a
r
e
bot
h
w
e
l
l
-
s
e
pa
r
a
t
e
d
a
c
r
os
s
c
l
a
s
s
e
s
a
nd c
om
pa
c
t
w
i
t
h
i
n
e
a
c
h
c
l
a
s
s
.
T
a
bl
e
4 pr
e
s
e
n
t
s
t
h
e
r
e
s
u
l
t
s
o
f
t
he
A
r
c
F
a
c
e
l
os
s
f
unc
t
i
on ba
s
e
d on e
a
c
h da
t
a
s
e
t
.
T
a
bl
e
3. P
e
r
f
or
m
a
nc
e
a
n
a
l
ys
i
s
of
d
i
f
f
e
r
e
nt
c
l
a
s
s
i
f
i
c
a
t
i
on m
e
t
hods
M
e
t
hod
s
D
a
t
a
s
e
t
A
c
c
ur
a
c
y (
%
)
P
r
e
c
i
s
i
on (
%
)
R
e
c
a
l
l
(
%
)
F1
-
s
c
or
e
(
%
)
M
L
P
U
F
C
101
84.37
82.56
81.78
82.16
H
M
D
B
-
51
80.24
79.45
78.36
78.9
DNN
U
F
C
101
91.52
90.78
89.34
90.05
H
M
D
B
-
51
87.96
87.42
86.34
86.87
R
e
s
N
e
t
U
F
C
101
93.42
92.86
91.78
92.32
H
M
D
B
-
51
89.74
89.15
88.42
88.78
VGG
U
F
C
101
92.81
92.12
91.06
91.58
H
M
D
B
-
51
88.45
87.89
87.23
87.55
S
K
N
-
2D
-
C
N
N
U
F
C
101
99.16
99.10
99.11
99.10
H
M
D
B
-
51
98.75
99.41
96.03
97.54
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
t
i
f
I
nt
e
l
l
,
V
ol
. 15, N
o. 1, F
e
br
ua
r
y 2026
:
350
-
360
356
T
a
bl
e
4. P
e
r
f
or
m
a
nc
e
e
v
a
l
ua
t
i
on of
d
i
f
f
e
r
e
nt
l
os
s
f
unc
t
i
on
M
e
t
hod
s
D
a
t
a
s
e
t
A
c
c
ur
a
c
y (
%
)
P
r
e
c
i
s
i
on (
%
)
R
e
c
a
l
l
(
%
)
F1
-
s
c
or
e
(
%
)
H
ube
r
l
os
s
U
F
C
101
91.82
91.23
90.54
90.88
H
M
D
B
-
51
87.34
86.85
85.73
86.28
P
i
nba
l
l
l
os
s
U
F
C
101
89.56
89.01
87.85
88.42
H
M
D
B
-
51
85.92
85.37
84.23
84.79
C
a
t
e
gor
i
c
a
l
c
r
os
s
e
nt
r
opy
U
F
C
101
93.54
92.98
91.85
92.41
H
M
D
B
-
51
88.67
88.09
87.25
87.66
F
oc
a
l
l
os
s
U
F
C
101
94.83
94.37
93.24
93.8
H
M
D
B
-
51
90.24
89.65
89.02
89.33
A
r
c
f
a
c
e
l
os
s
U
F
C
101
99.16
99.10
99.11
99.10
H
M
D
B
-
51
98.75
99.41
96.03
97.54
F
i
gur
e
2
r
e
pr
e
s
e
nt
s
t
he
a
c
c
ur
a
c
y
vs
e
po
c
h
gr
a
ph,
de
m
ons
t
r
a
t
i
ng
m
ode
l
’
s
a
c
c
ur
a
c
y
ove
r
t
r
a
i
ni
ng
e
poc
hs
f
or
t
r
a
i
n
i
ng
a
nd
va
l
i
d
a
t
i
on
on
bo
t
h
da
t
a
s
e
t
s
:
F
i
gur
e
2(
a
)
f
or
U
C
F
101
a
nd
F
i
gur
e
2(
b)
f
or
H
M
D
B
-
51.
T
he
t
r
a
i
ni
ng
a
nd
va
l
i
da
t
i
on
a
c
c
ur
a
c
y
i
n
c
r
e
a
s
e
s
t
e
a
di
l
y,
t
he
r
e
by
e
l
i
m
i
na
t
i
ng
t
he
r
i
s
k
of
unde
r
f
i
t
t
i
ng.
F
i
gur
e
3
s
how
s
t
he
l
os
s
vs
. e
poc
h gr
a
phs
f
or
pr
opos
e
d
m
e
t
hod, w
i
t
h F
i
gu
r
e
3(
a
)
r
e
p
r
e
s
e
nt
i
ng U
C
F
101 a
nd F
i
gur
e
3(
b)
r
e
pr
e
s
e
n
t
i
ng
H
M
D
B
-
51
,
i
l
l
us
t
r
a
t
i
ng
how
t
he
m
ode
l
'
s
l
os
s
c
ha
nge
s
ove
r
t
he
t
r
a
i
n
i
ng
e
poc
hs
a
nd
how
c
l
os
e
l
y
t
he
m
ode
l
'
s
p
r
e
di
c
t
i
ons
m
a
t
c
h
t
he
t
r
ue
l
a
be
l
s
.
T
he
R
O
C
c
ur
ve
gr
a
phi
c
a
l
l
y
i
l
l
us
t
r
a
t
e
s
a
c
l
a
s
s
i
f
i
e
r
’
s
pe
r
f
or
m
a
nc
e
by
p
l
ot
t
i
ng
t
he
t
r
ue
pos
i
t
i
ve
r
a
t
e
(
T
P
R
)
a
ga
i
ns
t
t
he
f
a
l
s
e
po
s
i
t
i
ve
r
a
t
e
(
F
P
R
)
a
c
r
o
s
s
di
f
f
e
r
e
nt
t
hr
e
s
ho
l
d s
e
t
t
i
ngs
.
(
a
)
(
b)
F
i
gur
e
2. P
e
r
f
o
r
m
a
nc
e
a
na
l
y
s
i
s
o
f
a
c
c
ur
a
c
y v
s
e
poc
hs
f
o
r
pr
opos
e
d
m
e
t
hod of
(
a
)
U
C
F
101
(
b)
H
M
D
B
-
51
(
a
)
(b
)
F
i
gur
e
3. P
e
r
f
o
r
m
a
nc
e
a
na
l
y
s
i
s
o
f
l
os
s
v
s
e
poc
h f
o
r
pr
opos
e
d
m
e
t
hod
of
(
a
)
U
C
F
101
(
b)
H
M
D
B
-
51
F
i
gur
e
4
p
r
e
s
e
n
t
s
t
h
e
pe
r
f
or
m
a
n
c
e
a
na
l
ys
i
s
ba
s
e
d
on
a
r
e
a
unde
r
t
he
c
u
r
ve
(
A
U
C
)
m
e
a
s
ur
e
s
f
or
U
C
F
101
a
s
s
how
n
i
n
F
i
gur
e
4(
a
)
a
nd
H
M
D
B
-
51
a
s
s
how
n
i
n
F
i
gu
r
e
4(
b
)
.
H
i
ghe
r
A
U
C
va
l
ue
s
,
c
l
os
e
r
t
o
1,
i
ndi
c
a
t
e
be
t
t
e
r
c
l
a
s
s
di
s
c
r
i
m
i
na
t
i
on.
T
he
pe
r
f
or
m
a
nc
e
a
na
l
y
s
i
s
us
i
ng
t
he
c
onf
us
i
on
m
a
t
r
i
x
i
s
s
how
n
i
n
F
i
gur
e
5,
w
i
t
h
F
i
gur
e
5(
a
)
f
or
U
C
F
101
a
nd
F
i
gur
e
5(
b)
f
or
H
M
D
B
-
51,
r
e
s
pe
c
t
i
ve
l
y.
I
n
e
a
c
h
m
a
t
r
i
x,
r
ow
s
r
e
pr
e
s
e
n
t
t
he
a
c
t
ua
l
c
l
a
s
s
e
s
,
w
hi
l
e
c
o
l
um
n
s
r
e
p
r
e
s
e
nt
t
he
p
r
e
di
c
t
e
d
c
l
a
s
s
e
s
,
he
l
p
i
ng
t
o
i
de
nt
i
f
y
m
i
s
c
l
a
s
s
i
f
i
c
a
t
i
ons
a
c
r
o
s
s
di
f
f
e
r
e
nt
a
c
t
i
v
i
t
y c
a
t
e
gor
i
e
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
t
i
f
I
nt
e
l
l
I
S
S
N
:
2252
-
8938
H
um
an ac
t
i
v
i
t
y
r
e
c
ogni
t
i
on us
i
ng s
e
l
e
c
t
i
v
e
k
e
r
ne
l
ne
t
w
or
k
…
(
B
anus
hr
i
Sr
i
n
i
v
as
a
i
ah
)
357
(
a
)
(
b)
F
i
gur
e
4. P
e
r
f
o
r
m
a
nc
e
a
na
l
y
s
i
s
o
f
T
P
R
vs
F
P
R
f
or
pr
opos
e
d m
e
t
hod
of
(
a
)
U
C
F
101
(
b)
H
M
D
B
-
51
(
a
)
(
b)
F
i
gur
e
5. P
e
r
f
o
r
m
a
nc
e
a
na
l
y
s
i
s
o
f
c
onf
us
i
on m
a
t
r
i
x
f
or
pr
opo
s
e
d m
e
t
hod
of
(
a
)
U
C
F
101 (
b)
H
M
D
B
-
51
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
t
i
f
I
nt
e
l
l
,
V
ol
. 15, N
o. 1, F
e
br
ua
r
y 2026
:
350
-
360
358
3.2. C
om
p
ar
at
i
ve
an
al
ys
i
s
T
a
bl
e
5
de
m
ons
t
r
a
t
e
s
t
he
c
om
pa
r
a
t
i
ve
a
na
l
ys
i
s
of
t
he
pr
op
os
e
d
m
e
t
hod
w
i
t
h
e
xi
s
t
i
ng
m
e
t
hods
.
T
he
pr
op
os
e
d
S
K
N
-
2D
-
C
N
N
w
i
t
h
A
r
c
f
a
c
e
l
os
s
m
e
t
ho
d
i
s
c
om
pa
r
e
d
w
i
t
h
e
xi
s
t
i
ng
m
e
t
hods
:
C
N
N
-
B
i
G
R
U
[
16]
, B
i
G
R
U
,
C
N
N
a
nd
L
S
T
M
[
17]
,
C
N
N
[
18]
,
a
nd
M
obi
l
e
N
e
t
V
2
[
20]
.
T
he
pr
opos
e
d
S
K
N
-
2D
-
C
N
N
w
i
t
h
A
r
c
f
a
c
e
l
os
s
de
m
ons
t
r
a
t
e
s
a
s
upe
r
i
or
a
c
c
ur
a
c
y
of
99.16
a
nd
98
.75%
on
U
F
C
101
a
nd
H
M
D
B
-
51
da
t
a
s
e
t
s
.
T
he
S
K
N
a
l
l
o
w
s
t
he
m
od
e
l
t
o
a
da
pt
t
o
di
ve
r
s
e
s
pa
t
i
a
l
pa
t
t
e
r
ns
,
i
m
pr
ovi
ng
r
e
c
ogni
t
i
on
f
or
c
om
pl
e
x
a
c
t
i
vi
t
i
e
s
a
nd
a
r
c
f
a
c
e
l
os
s
,
e
ns
ur
i
ng
t
ha
t
f
e
a
t
ur
e
e
m
be
ddi
ng
r
e
duc
e
s
m
i
s
c
l
a
s
s
i
f
i
c
a
t
i
on
f
or
a
c
t
i
vi
t
i
e
s
w
i
t
h
hi
gh
i
nt
e
r
-
c
l
a
s
s
s
i
m
i
l
a
r
i
t
y.
T
a
bl
e
5. C
om
pa
r
a
t
i
ve
a
na
l
ys
i
s
of
pr
opo
s
e
d m
e
t
hod w
i
t
h e
x
i
s
t
i
ng m
e
t
hods
M
e
t
hod
s
D
a
t
a
s
e
t
A
c
c
ur
a
c
y (
%
)
C
N
N
-
B
i
G
R
U
[
16]
U
F
C
101
91.79
H
M
D
B
-
51
71.89
B
i
G
R
U
, C
N
N
a
n
d
L
S
T
M
[
17]
U
F
C
101
98.80
H
M
D
B
-
51
NA
C
N
N
[
18]
U
F
C
101
98.01
H
M
D
B
-
51
97.45
M
obi
l
e
N
e
t
V
2
[
20]
U
F
C
101
92.93
H
M
D
B
-
51
80.61
P
r
opos
e
d
S
K
N
-
2D
-
C
N
N
w
i
t
h
A
r
c
F
a
c
e
l
os
s
m
e
t
hod
U
F
C
101
99.16
H
M
D
B
-
51
98.75
3.3. D
i
s
c
u
s
s
i
on
T
he
m
e
r
i
t
s
of
t
he
pr
opos
e
d
S
K
N
-
2D
-
C
N
N
w
i
t
h
A
r
c
F
a
c
e
l
os
s
a
nd
t
he
l
i
m
i
t
a
t
i
ons
of
e
x
i
s
t
i
ng
t
e
c
hni
que
s
l
i
ke
H
A
R
,
w
hi
c
h
m
a
i
nl
y
f
oc
us
on
i
m
pr
ovi
ng
t
he
a
c
c
ur
a
c
y
a
nd
l
e
a
r
ni
ng
l
ong
-
s
e
que
nc
e
t
e
m
por
a
l
a
c
t
i
ons
a
r
e
di
s
c
us
s
e
d
i
n
t
hi
s
s
e
c
t
i
on
.
T
he
H
A
R
pr
oc
e
s
s
t
ypi
c
a
l
l
y
w
or
ks
on
G
P
U
s
but
s
t
r
uggl
e
s
t
o
pr
e
di
c
t
a
c
t
i
vi
t
i
e
s
on
i
nt
e
r
ne
t
of
t
h
i
ngs
(
I
oT
)
de
vi
c
e
s
.
G
L
C
M
a
nd
L
G
T
P
a
r
e
us
e
d
f
or
f
e
a
t
ur
e
e
xt
r
a
c
t
i
on,
w
hi
l
e
c
l
a
s
s
i
f
i
e
r
s
l
i
ke
C
N
N
,
B
i
G
R
U
, a
nd
L
S
T
M
of
f
e
r
a
c
c
ur
a
t
e
c
l
a
s
s
i
f
i
c
a
t
i
on. T
he
S
of
t
M
a
x c
l
a
s
s
i
f
i
e
r
i
s
ut
i
l
i
z
e
d f
or
a
c
t
i
vi
t
y
c
l
a
s
s
i
f
i
c
a
t
i
on.
T
he
r
e
l
i
a
nc
e
on
s
pe
c
i
f
i
c
hum
a
n
-
c
e
nt
r
i
c
f
e
a
t
ur
e
s
l
e
a
ds
t
o
ove
r
f
i
t
t
i
ng
i
f
t
he
da
t
a
s
e
t
l
a
c
ks
di
ve
r
s
i
t
y
i
n
pos
e
s
,
or
i
e
n
t
a
t
i
ons
,
o
r
obj
e
c
t
s
i
z
e
s
.
T
he
H
A
R
m
ode
l
be
c
om
e
s
c
om
pl
e
x
unde
r
e
nvi
r
onm
e
nt
a
l
va
r
i
a
t
i
ons
or
oc
c
l
us
i
ons
,
a
s
i
t
i
s
ba
s
e
d
on
f
i
xe
d
s
pa
t
i
a
l
a
nd
t
e
m
por
a
l
f
e
a
t
ur
e
s
t
ha
t
m
a
y
not
a
da
pt
w
e
l
l
t
o
dyna
m
i
c
or
c
l
ut
t
e
r
e
d
ba
c
kgr
ounds
.
T
he
c
om
bi
na
t
i
on
of
M
obi
l
e
N
e
t
V
2
f
or
s
pa
t
i
a
l
f
e
a
t
ur
e
e
xt
r
a
c
t
i
on
a
nd
a
m
ul
t
i
l
a
ye
r
G
R
U
f
or
t
e
m
por
a
l
m
ode
l
l
i
ng
e
na
bl
e
s
t
he
m
ode
l
t
o
c
a
p
t
ur
e
bot
h
s
pa
t
i
a
l
de
t
a
i
l
s
a
nd t
e
m
por
a
l
de
pe
nde
nc
i
e
s
.
4.
C
O
N
C
L
U
S
I
O
N
T
hi
s
r
e
s
e
a
r
c
h
pr
opos
e
s
a
n
S
K
N
-
2D
-
C
N
N
m
ode
l
w
i
t
h
A
r
c
F
a
c
e
l
os
s
f
or
t
he
e
f
f
e
c
t
i
ve
c
a
pt
u
r
e
of
m
ul
t
i
-
s
c
a
l
e
s
pa
t
i
a
l
f
e
a
t
u
r
e
s
, l
e
ve
r
a
g
i
ng
i
t
s
h
i
ghl
y
d
i
s
c
r
i
m
i
na
t
i
ve
a
b
i
l
i
t
y f
or
H
A
R
.
T
he
ke
y
f
i
nd
i
ngs
s
how
t
ha
t
i
nt
e
gr
a
t
i
ng
t
he
s
e
l
e
c
t
i
ve
ke
r
ne
l
m
e
c
ha
n
i
s
m
e
na
bl
e
s
t
he
ne
t
w
or
k
t
o
dyna
m
i
c
a
l
l
y
a
dj
us
t
t
he
r
e
c
e
pt
i
ve
f
i
e
l
d,
f
oc
us
i
ng
on
t
he
m
os
t
i
nf
o
r
m
a
t
i
ve
s
pa
t
i
a
l
p
a
t
t
e
r
ns
a
nd
m
i
ni
m
i
z
i
ng
t
he
i
m
pa
c
t
o
f
i
r
r
e
l
e
va
nt
f
r
a
m
e
s
.
M
o
r
e
ove
r
,
t
he
A
r
c
F
a
c
e
l
os
s
f
unc
t
i
on
e
m
pl
oys
a
n
a
ngul
a
r
m
a
r
gi
n
pe
na
l
t
y
t
ha
t
e
nha
nc
e
s
i
nt
r
a
-
c
l
a
s
s
c
om
pa
c
t
ne
s
s
a
nd
i
nt
e
r
-
c
l
a
s
s
s
e
pa
r
a
b
i
l
i
t
y, r
e
s
u
l
t
i
ng i
n s
upe
r
i
o
r
pe
r
f
o
r
m
a
nc
e
. T
he
pr
a
c
t
i
c
a
l
i
m
pl
i
c
a
t
i
on of
t
hi
s
r
e
s
e
a
r
c
h l
i
e
s
i
n i
t
s
a
bi
l
i
t
y
t
o
e
nha
nc
e
t
he
r
e
l
i
a
b
i
l
i
t
y
a
nd
r
obus
t
ne
s
s
of
vi
de
o
-
ba
s
e
d
H
A
R
,
w
hi
c
h
i
s
s
i
gni
f
i
c
a
nt
f
or
s
ur
v
e
i
l
l
a
nc
e
a
ppl
i
c
a
t
i
ons
.
C
om
pa
r
e
d
t
o
e
xi
s
t
i
ng
m
e
t
hods
l
i
k
e
C
N
N
-
B
i
G
R
U
,
t
he
pr
opos
e
d
S
K
N
-
2D
-
C
N
N
w
i
t
h
A
r
c
F
a
c
e
l
os
s
a
c
h
i
e
ve
s
s
upe
r
i
or
a
c
c
ur
a
c
i
e
s
of
99.16
a
nd
98.75
%
on t
he
U
F
C
101
a
nd
H
M
D
B
-
51
da
t
a
s
e
t
s
,
r
e
s
pe
c
t
i
ve
l
y.
I
n
t
he
f
ut
ur
e
,
i
m
pr
ove
d
op
t
i
m
i
z
a
t
i
on
-
ba
s
e
d
m
e
t
hod
s
w
i
l
l
be
us
e
d
t
o
s
e
l
e
c
t
t
he
m
os
t
a
ppr
op
r
i
a
t
e
f
e
a
t
ur
e
s
,
f
ur
t
he
r
e
nha
nc
i
ng
m
ode
l
p
e
r
f
o
r
m
a
nc
e
i
n H
A
R
.
F
U
N
D
I
N
G
I
N
F
O
R
M
A
T
I
O
N
A
ut
hor
s
s
t
a
t
e
no f
und
i
ng i
nvo
l
ve
d.
A
U
T
H
O
R
C
O
N
T
R
I
B
U
T
I
O
N
S
S
T
A
T
E
M
E
N
T
T
hi
s
j
our
n
a
l
u
s
e
s
t
he
C
ont
r
i
but
o
r
R
ol
e
s
T
a
xonom
y
(
C
R
e
di
T
)
t
o
r
e
c
ogn
i
z
e
i
nd
i
vi
dua
l
a
u
t
hor
c
ont
r
i
bu
t
i
ons
,
r
e
duc
e
a
u
t
hor
s
h
i
p di
s
pu
t
e
s
, a
nd
f
a
c
i
l
i
t
a
t
e
c
o
l
l
a
bo
r
a
t
i
on.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
t
i
f
I
nt
e
l
l
I
S
S
N
:
2252
-
8938
H
um
an ac
t
i
v
i
t
y
r
e
c
ogni
t
i
on us
i
ng s
e
l
e
c
t
i
v
e
k
e
r
ne
l
ne
t
w
or
k
…
(
B
anus
hr
i
Sr
i
n
i
v
as
a
i
ah
)
359
N
am
e
of
A
u
t
h
or
C
M
So
Va
Fo
I
R
D
O
E
Vi
Su
P
Fu
B
a
nus
hr
i
S
r
i
n
i
va
s
a
i
a
h
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
J
a
ga
d
e
e
s
ha
R
a
m
e
gow
da
✓
✓
✓
✓
✓
✓
✓
C
:
C
onc
e
pt
ua
l
i
z
a
t
i
on
M
:
M
e
t
ho
d
ol
ogy
So
:
So
f
t
w
a
r
e
Va
:
Va
l
i
d
a
t
i
on
Fo
:
Fo
r
m
a
l
a
na
l
ys
i
s
I
:
I
nve
s
t
i
ga
t
i
on
R
:
R
e
s
our
c
e
s
D
:
D
a
t
a
C
ur
a
t
i
on
O
:
W
r
i
t
i
ng
-
O
r
i
g
i
na
l
D
r
a
f
t
E
:
W
r
i
t
i
ng
-
R
e
vi
e
w
&
E
d
i
t
i
ng
Vi
:
Vi
s
ua
l
i
z
a
t
i
on
Su
:
Su
pe
r
vi
s
i
on
P
:
P
r
oj
e
c
t
a
d
m
i
ni
s
t
r
a
t
i
on
Fu
:
Fu
nd
i
ng a
c
qui
s
i
t
i
on
C
O
N
F
L
I
C
T
O
F
I
N
T
E
R
E
S
T
S
T
A
T
E
M
E
N
T
A
ut
hor
s
s
t
a
t
e
no c
onf
l
i
c
t
of
i
nt
e
r
e
s
t
.
D
A
T
A
A
V
A
I
L
A
B
I
L
I
T
Y
T
he
da
t
a
s
e
t
s
ge
ne
r
a
t
e
d
dur
i
ng
a
nd/
or
a
na
l
yz
e
d
dur
i
ng
t
he
c
ur
r
e
n
t
s
t
udy
a
r
e
a
va
i
l
a
b
l
e
i
n
t
he
H
M
D
B
-
51
at
ht
t
ps
:
/
/
w
w
w
.ka
ggl
e
.c
om
/
da
t
a
s
e
t
s
/
e
a
s
on
l
l
l
/
H
M
D
B
-
51
a
nd
U
F
C
101
da
t
a
s
e
t
s
a
t
ht
t
ps
:
/
/
w
w
w
.ka
ggl
e
.c
om
/
da
t
a
s
e
t
s
/
m
a
t
t
h
e
w
j
a
ns
e
n/
u
c
f
101
-
a
c
t
i
on
-
r
e
c
ogn
i
t
i
on.
R
E
F
E
R
E
N
C
E
S
[
1]
M
.
A
. K
ha
n
e
t
al
.
,
“
H
um
a
n
a
c
t
i
on
r
e
c
ogni
t
i
on
us
i
ng
f
us
i
on
of
m
ul
t
i
vi
e
w
a
nd d
e
e
p f
e
a
t
ur
e
s
:
a
n
a
ppl
i
c
a
t
i
on
t
o
vi
d
e
o
s
ur
ve
i
l
l
a
nc
e
,
”
M
ul
t
i
m
e
di
a T
ool
s
and A
pp
l
i
c
a
t
i
ons
, vo
l
. 83, no. 5, pp. 14885
–
14911, 2024, d
oi
:
10.1007/
s
11042
-
020
-
08806
-
9.
[
2]
A
.
H
us
s
a
i
n,
S
.
U
.
K
ha
n,
N
.
K
ha
n,
M
.
S
ha
ba
z
,
a
nd
S
.
W
.
B
a
i
k,
“
AI
-
d
r
i
ve
n
be
ha
vi
o
r
b
i
om
e
t
r
i
c
s
f
r
a
m
e
w
or
k
f
or
r
obus
t
hum
a
n
a
c
t
i
vi
t
y
r
e
c
ogni
t
i
on
i
n
s
ur
ve
i
l
l
a
nc
e
s
ys
t
e
m
s
,
”
E
ngi
ne
e
r
i
ng
A
ppl
i
c
a
t
i
ons
of
A
r
t
i
f
i
c
i
al
I
n
t
e
l
l
i
ge
nc
e
,
vo
l
.
127,
2024,
d
oi
:
10.1016/
j
.e
nga
ppa
i
.2023.107218.
[
3]
Y
.
K
a
ya
a
n
d
E
.
K
.
T
opuz
,
“
H
um
a
n
a
c
t
i
vi
t
y
r
e
c
ogni
t
i
on
f
r
om
m
ul
t
i
p
l
e
s
e
ns
or
s
d
a
t
a
us
i
ng
d
e
e
p
C
N
N
s
,
”
M
ul
t
i
m
e
di
a
T
ool
s
and
A
ppl
i
c
at
i
ons
, vol
. 83, no. 4, pp. 10815
–
10838, 2023, d
oi
:
10.1007
/
s
11042
-
023
-
15830
-
y.
[
4]
M
.
S
.
R
a
j
,
S
.
N
.
G
e
or
ge
,
a
nd
K
.
R
a
j
a
,
“
L
e
ve
r
a
gi
ng
s
pa
t
i
o
-
t
e
m
por
a
l
f
e
a
t
ur
e
s
us
i
ng
gr
a
ph
ne
ur
a
l
ne
t
w
or
ks
f
or
hu
m
a
n
a
c
t
i
vi
t
y
r
e
c
ogni
t
i
on,
”
P
at
t
e
r
n R
e
c
ogni
t
i
on
, vo
l
. 150, 2024, d
oi
:
10.1016/
j
.pa
t
c
og.2024.110301.
[
5]
Y
. Z
hou, J
. X
i
e
, X
. Z
ha
ng, W
. W
u,
a
nd
S
. K
w
ong,
“
E
ne
r
gy
-
e
f
f
i
c
i
e
nt
a
nd
i
nt
e
r
pr
e
t
a
bl
e
m
ul
t
i
s
e
ns
or
hu
m
a
n a
c
t
i
vi
t
y r
e
c
ogn
i
t
i
on vi
a
d
e
e
p f
us
e
d
l
a
s
s
o ne
t
,
”
I
E
E
E
T
r
ans
ac
t
i
ons
on E
m
e
r
gi
ng
T
opi
c
s
i
n
C
om
put
at
i
ona
l
I
n
t
e
l
l
i
ge
nc
e
, vo
l
. 8, no.
5, pp.
3576
–
3588, 2024,
d
oi
:
10.1109/
T
E
T
C
I
.2024.3430008.
[
6]
P
.
L
a
l
w
a
ni
a
nd
G
. R
a
m
a
s
a
m
y,
“
H
um
a
n
a
c
t
i
vi
t
y
r
e
c
ogni
t
i
on
us
i
ng
a
m
u
l
t
i
-
br
a
n
c
he
d
C
N
N
-
B
i
L
S
T
M
-
B
i
G
R
U
m
od
e
l
,
”
A
ppl
i
e
d
Sof
t
C
om
put
i
ng
, vol
. 154, 2024, d
oi
:
10.1016/
j
.a
s
oc
.2024.111344.
[
7]
G
.
P
a
r
e
e
k,
S
.
N
i
g
a
m
,
a
nd
R
.
S
i
ngh,
“
M
od
e
l
i
ng
t
r
a
ns
f
or
m
e
r
a
r
c
h
i
t
e
c
t
ur
e
w
i
t
h
a
t
t
e
nt
i
on
l
a
ye
r
f
or
hum
a
n
a
c
t
i
vi
t
y
r
e
c
ogni
t
i
on,
”
N
e
ur
al
C
om
put
i
ng and A
ppl
i
c
a
t
i
ons
, vo
l
. 36, no. 10, pp. 5515
–
5528, 2024, d
oi
:
10.1007/
s
00521
-
023
-
09362
-
7.
[
8]
M
.
E
z
z
e
l
d
i
n,
A
.
S
.
G
hone
i
m
,
L
.
A
bd
e
l
ha
m
i
d
,
a
nd
A
.
A
t
i
a
,
“
M
ul
t
i
-
m
od
a
l
hybr
i
d
hi
e
r
a
r
c
h
i
c
a
l
c
l
a
s
s
i
f
i
c
a
t
i
on
a
ppr
o
a
c
h
w
i
t
h
t
r
a
ns
f
or
m
e
r
s
t
o
e
nha
nc
e
c
om
pl
e
x
hum
a
n
a
c
t
i
vi
t
y
r
e
c
ogni
t
i
on,
”
Si
gnal
,
I
m
age
and
V
i
de
o
P
r
oc
e
s
s
i
ng
,
vol
.
18,
pp.
9375
–
9385,
2024, d
oi
:
10.1007/
s
11760
-
024
-
03552
-
z.
[
9]
H
.
P
a
r
k,
G
.
H
.
L
e
e
,
J
.
H
a
n,
a
n
d
J
.
K
.
C
hoi
,
“
M
ul
t
i
c
l
a
s
s
a
ut
oe
nc
od
e
r
-
ba
s
e
d
a
c
t
i
ve
l
e
a
r
ni
ng
f
or
s
e
n
s
or
-
ba
s
e
d
hum
a
n
a
c
t
i
vi
t
y
r
e
c
ogni
t
i
on,
”
F
ut
ur
e
G
e
ne
r
at
i
on C
om
put
e
r
Sy
s
t
e
m
s
, vol
. 151, pp. 71
–
84, 2024, d
oi
:
10.1016
/
j
.f
ut
ur
e
.2023.09.029.
[
10]
Z
.
Y
a
ng,
K
.
L
i
,
a
n
d
Z
.
H
ua
ng,
“
M
F
C
A
N
N
:
a
f
e
a
t
ur
e
d
i
ve
r
s
i
f
i
c
a
t
i
on
f
r
a
m
e
w
or
k
ba
s
e
d
on
l
oc
a
l
a
n
d
gl
oba
l
a
t
t
e
n
t
i
on
f
or
hum
a
n
a
c
t
i
vi
t
y
r
e
c
ogni
t
i
on,
”
E
ngi
ne
e
r
i
ng
A
ppl
i
c
at
i
ons
o
f
A
r
t
i
f
i
c
i
a
l
I
nt
e
l
l
i
g
e
nc
e
,
vo
l
.
133,
J
ul
.
2024,
d
oi
:
10.1016/
j
.e
nga
ppa
i
.2024.108110.
[
11]
Y
.
C
.
L
a
i
,
Y
.
C
.
K
a
n,
K
.
C
.
H
s
u,
a
n
d
H
.
C
.
L
i
n,
“
M
ul
t
i
pl
e
i
nput
s
m
od
e
l
i
ng
of
hybr
i
d
c
onvol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
ks
f
or
hum
a
n
a
c
t
i
vi
t
y r
e
c
ogni
t
i
on,
”
B
i
om
e
di
c
a
l
Si
gna
l
P
r
oc
e
s
s
i
ng and C
ont
r
o
l
, vol
. 92, 2024, d
oi
:
10.1016/
j
.bs
pc
.2024.106034.
[
12]
M
.
A
.
A
l
-
qa
ne
s
s
,
A
.
D
a
hou,
M
.
A
.
E
l
a
z
i
z
,
a
nd
A
.
M
.
H
e
l
m
i
,
“
H
um
a
n
a
c
t
i
vi
t
y
r
e
c
ogni
t
i
on
a
nd
f
a
l
l
d
e
t
e
c
t
i
on
us
i
ng
c
onvol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
k
a
nd
t
r
a
ns
f
or
m
e
r
-
ba
s
e
d
a
r
c
hi
t
e
c
t
ur
e
,
”
B
i
om
e
di
c
al
Si
gnal
P
r
oc
e
s
s
i
ng
and
C
ont
r
ol
,
vol
.
95,
2024,
d
oi
:
10.1016/
j
.bs
pc
.2024.106412.
[
13]
A
.
B
oud
j
e
m
a
,
F
.
T
i
t
ouna
,
a
nd
C
.
T
i
t
ouna
,
“
A
R
e
N
e
t
:
c
a
s
c
a
d
e
l
e
a
r
n
i
ng
of
m
u
l
t
i
br
a
nc
h
c
onvol
ut
i
ona
l
ne
u
r
a
l
ne
t
w
or
ks
f
or
hum
a
n
a
c
t
i
vi
t
y
r
e
c
ogni
t
i
on,
”
M
u
l
t
i
m
e
di
a
T
ool
s
and
A
ppl
i
c
at
i
ons
,
vol
.
83,
no.
17,
pp.
51099
–
51128,
N
ov.
2023,
d
oi
:
10.1007/
s
11042
-
023
-
17496
-
y.
[
14]
Q
.
L
i
n,
X
.
L
i
,
K
.
C
a
i
,
M
.
P
r
a
ka
s
h,
a
nd
D
.
P
a
ul
r
a
j
,
“
S
e
c
u
r
e
i
nt
e
r
ne
t
of
m
e
d
i
c
a
l
t
hi
n
gs
(
I
oM
T
)
ba
s
e
d
on
E
C
M
Q
V
-
M
A
C
a
ut
he
nt
i
c
a
t
i
on
pr
ot
oc
ol
a
nd
E
K
M
C
-
S
C
P
bl
oc
kc
ha
i
n
ne
t
w
or
ki
ng,
”
I
n
f
or
m
at
i
on
S
c
i
e
nc
e
s
,
vol
.
654,
2024,
d
oi
:
10.1016/
j
.i
ns
.2023.119783.
[
15]
T
.
T
e
ng,
J
.
W
a
n,
a
nd
X
.
Z
ha
ng,
“
G
A
N
C
D
E
:
ne
ur
a
l
ne
t
w
or
ks
ba
s
e
d
on
gr
a
phs
a
nd
a
t
t
e
n
t
i
on
ne
ur
a
l
c
ont
r
ol
d
i
f
f
e
r
e
nt
i
a
l
e
qua
t
i
ons
f
or
hum
a
n
a
c
t
i
vi
t
y
r
e
c
ogni
t
i
on,
”
K
now
l
e
dge
and
I
n
f
or
m
at
i
on
Sy
s
t
e
m
s
,
vol
.
66,
pp.
6213
–
6240,
J
un.
2024,
d
oi
:
10.1007
/
s
10115
-
024
-
02154
-
y.
[
16]
T
.
A
hm
a
d
,
J
.
W
u,
H
.
S
. A
l
w
a
ge
e
d
,
F
.
K
ha
n,
J
.
K
ha
n,
a
n
d
Y
.
L
e
e
,
“
H
um
a
n
a
c
t
i
vi
t
y
r
e
c
ogni
t
i
on
ba
s
e
d
on
d
e
e
p
-
t
e
m
por
a
l
l
e
a
r
ni
ng
us
i
ng
c
onvol
ut
i
on
ne
u
r
a
l
ne
t
w
or
ks
f
e
a
t
ur
e
s
a
nd
bi
d
i
r
e
c
t
i
ona
l
ga
t
e
d
r
e
c
ur
r
e
nt
uni
t
w
i
t
h
f
e
a
t
ur
e
s
s
e
l
e
c
t
i
on,
”
I
E
E
E
A
c
c
e
s
s
,
vol
.
11,
pp. 33148
–
33159, 2023, d
oi
:
10.1109/
A
C
C
E
S
S
.2023.3263155.
[
17]
K
.
P
.
S
i
nha
a
nd
P
.
K
um
a
r
,
“
H
um
a
n
a
c
t
i
vi
t
y
r
e
c
ogn
i
t
i
on
f
r
om
U
A
V
vi
d
e
os
us
i
ng
a
n
opt
i
m
i
z
e
d
hybr
i
d
d
e
e
p
l
e
a
r
n
i
ng
m
od
e
l
,
”
M
ul
t
i
m
e
di
a T
ool
s
and A
pp
l
i
c
a
t
i
ons
, vo
l
. 83, no. 17, pp. 51669
–
51698, 2024, d
oi
:
10.1007/
s
11042
-
023
-
17289
-
3.
[
18]
A
.
K
us
hw
a
ha
,
A
.
K
ha
r
e
,
a
nd
O
.
P
r
a
ka
s
h,
“
M
i
c
r
o
-
ne
t
w
or
k
-
ba
s
e
d
d
e
e
p
c
onvol
ut
i
ona
l
ne
u
r
a
l
ne
t
w
or
k
f
or
hum
a
n
a
c
t
i
vi
t
y
r
e
c
ogni
t
i
on
f
r
om
r
e
a
l
i
s
t
i
c
a
nd
m
ul
t
i
-
vi
e
w
vi
s
ua
l
d
a
t
a
,
”
N
e
ur
al
C
o
m
put
i
ng
and
A
ppl
i
c
at
i
on
s
,
vol
.
35,
no.
18,
pp.
13321
–
13341,
2023, d
oi
:
10.1007/
s
00521
-
023
-
08440
-
0.
Evaluation Warning : The document was created with Spire.PDF for Python.