Internati
o
nal
Journal of Ele
c
trical
and Computer
Engineering
(IJE
CE)
Vol
.
5
,
No
. 3,
J
une
2
0
1
5
,
pp
. 56
9~
57
8
I
S
SN
: 208
8-8
7
0
8
5
69
Jo
urn
a
l
h
o
me
pa
ge
: h
ttp
://iaesjo
u
r
na
l.com/
o
n
lin
e/ind
e
x.ph
p
/
IJECE
Toward Cont
ent-
Aware Video
Partitioning Methods for
Distributed HEVC Video Encoding
My
un
gh
oo
n
J
e
on,
B
y
oun
g
-
D
ai
L
e
e
Department o
f
C
o
mputer Scie
n
c
e, K
y
onggi University
, Korea
Article Info
A
B
STRAC
T
Article histo
r
y:
Received
Ja
n 16, 2015
Rev
i
sed
Ap
r
20
, 20
15
Accepte
d
May 4, 2015
Recen
tl
y, c
l
oud
com
puting has
em
erged as
a potentia
l platform
for
distributed v
i
deo
encoding due to its adva
ntag
es
in term
s
of cos
t
s
as
well as
performance. For distributed
video en
cod
i
ng, the input vid
e
o must be
partit
ioned into s
e
vera
l s
e
gm
ents
,
each of which i
s
proces
s
e
d over dis
t
ributed
resources. Th
is paper descr
i
bes the e
ffect of
differen
t
video
partitionin
g
schemes on overall
encoding
p
e
rforman
ce
in the distribu
ted
encoding of
High-Effic
i
enc
y
Video Coding (HEVC). In addition
,
we explor
e
d
performances of
video partitioning sche
mes on the basis of the ty
pes of the
conten
t to
be en
coded.
Keyword:
C
l
ou
d c
o
m
put
i
n
g
Co
n
t
en
t-
a
w
ar
en
e
s
s
Di
st
ri
b
u
t
e
d e
n
c
odi
ng
HEVC
Vid
e
o
p
a
rtitio
nin
g
Copyright ©
201
5 Institut
e
o
f
Ad
vanced
Engin
eer
ing and S
c
i
e
nce.
All rights re
se
rve
d
.
Co
rresp
ond
i
ng
Autho
r
:
Byoung-Dai L
ee
Depa
rtem
ent of Com
puter Sci
e
nce,
K
yon
gg
i U
n
i
v
er
sity,
I
u
i-do
ng
, Y
e
ong
ton
g
-g
u, Suwo
n-
si,
G
y
eongg
i-d
o
,
K
o
r
ea.
Em
a
il: b
l
ee@kyo
ng
g
i
.ac.kr
1.
INTRODUCTION
High-E
fficienc
y
Video Codi
ng (HEVC
)
[1] is the
m
o
st recent vide
o coding standa
rd, and it has been
desi
g
n
e
d
t
o
ad
dress i
n
crease
d
vi
de
o re
sol
u
t
i
on, s
u
ch as
u
l
t
r
a-hi
g
h
defi
ni
t
i
on (
U
H
D
) c
o
nt
ent
.
H
o
wev
e
r, t
h
e
sup
e
ri
o
r
i
t
y
of HEVC
e
n
co
di
ng
per
f
o
r
m
a
nce over t
h
at
o
f
exi
s
t
i
ng st
an
da
rd co
di
n
g
t
ech
nol
ogi
es com
e
s wi
t
h
hi
g
h
com
put
at
i
onal
com
p
l
e
xi
t
y
t
h
at
pro
duce
s
hi
g
h
enc
odi
n
g
bi
t
rat
e
s an
d
l
ong e
n
c
odi
ng
t
i
m
e
s. An ess
e
nt
i
a
l
t
a
sk f
o
r
HE
V
C
enco
der
s
, t
h
eref
ore
,
i
s
t
o
r
e
sol
v
e t
h
e i
ssu
e of
hi
g
h
c
o
m
put
i
n
g
po
we
r
du
ri
n
g
t
h
e e
n
c
odi
ng
process. Many studies
ha
ve bee
n
de
vot
ed to solv
ing these
proble
m
s
. Recen
t initiatives include
the
im
pro
v
em
ent
of
per
f
o
rm
ance by
ap
pl
y
i
ng
cl
ou
d c
o
m
put
i
ng t
o
vi
de
o
codi
ng
t
ech
no
l
ogy
.
In
pa
rt
i
c
ul
ar
,
con
s
i
d
era
b
l
e
a
t
t
e
nt
i
on ha
s b
een pai
d
t
o
di
st
ri
but
e
d
enc
o
di
n
g
, i
n
w
h
i
c
h enc
o
di
n
g
j
o
bs t
h
at
re
q
u
i
r
e hi
g
h
com
put
at
i
onal
com
p
l
e
xi
t
y
are di
st
ri
b
u
t
e
d
d
u
r
i
n
g vi
deo e
n
c
odi
ng
by
usi
n
g
t
h
e u
n
l
i
m
i
t
e
d com
put
i
ng
res
o
u
r
ces
avai
l
a
bl
e i
n
a cl
ou
d com
put
i
ng e
n
vi
ro
nm
ent
.
Thi
s
a
p
proach m
o
re effe
ctively reduce
s
encoding time tha
n
doe
s t
h
e
use
o
f
exi
s
t
i
n
g st
an
d
-
al
one e
n
c
o
d
e
rs
.
1
Thi
s
pa
per
p
r
ovi
des a
per
f
o
r
m
a
nce eval
ua
t
i
on
of
di
st
ri
b
u
t
e
d
HE
VC
e
n
co
di
n
g
.
It
f
o
cuses
o
n
t
h
e
effects
of the
video
pa
rtitioning m
e
thod on ove
rall
distri
bute
d
encoding perform
a
nce.
T
h
e vide
o partit
ioni
ng
m
e
t
hods a
d
opt
ed i
n
t
h
i
s
pa
p
e
r can
be cat
e
g
o
r
i
zed i
n
t
o
t
w
o t
y
p
e
s. T
h
e
fi
rst
i
s
a
gen
e
ral
m
e
t
h
o
d
, i
n
w
h
i
c
h
v
i
d
e
o
is u
n
i
fo
rmly
p
a
rtitio
n
e
d
on
th
e b
a
sis o
f
th
e nu
m
b
er o
f
nod
es th
at perfo
r
m
en
co
d
i
ng
task
s. Th
e seco
nd
is
v
i
d
e
o
p
a
rtitio
nin
g
o
n
th
e
b
a
si
s o
f
GOP t
h
at co
nsid
ers a v
i
deo
en
cod
i
ng
op
tio
n.
Alon
g
with
em
p
i
rical a
n
alysis
o
f
two
well-kn
own
v
i
d
e
o
partitio
n
i
ng
sche
m
e
s, we ex
ten
d
e
d th
e
GOP-b
a
sed v
i
d
e
o partitio
n
i
ng
sche
m
e
in
su
ch
a
way th
at d
y
n
a
m
i
c p
a
rts of th
e inp
u
t
vid
e
o
are
p
a
r
titio
n
e
d
i
n
to
m
o
re seg
m
en
ts th
an static p
a
rts.
In
d
o
i
ng
so, we ca
n re
duce the e
n
coding tim
e
with m
oderate
in
creases in
th
e o
v
e
rall b
itrat
e. In summ
ary, we
1
Th
is
p
a
p
e
r is an
ex
tend
ed v
e
rsio
n of th
e IC
ONI
2
014
.[6
]
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E V
o
l
.
5, No
. 3,
J
u
ne 2
0
1
5
:
56
9 – 5
7
8
57
0
perform
e
d distributed enc
o
ding
by ap
plying two pa
rtitioni
ng m
e
thods a
n
d
com
p
ared t
h
e perform
a
nce results.
The com
p
aris
on e
n
a
b
led us
to iden
tify a vide
o pa
rtitioning m
e
thod that
efficiently proce
sses distributed
en
cod
i
ng
,
and
we
can
th
en
reco
mmen
d
con
s
id
eration
s
for vid
e
o
p
a
rtitio
n
i
ng
.
2.
RELATED WORK
As vi
deo c
o
di
ng t
e
c
h
n
o
l
o
gy
adva
nces
, t
h
e i
ssue
of
hi
g
h
c
o
m
put
at
i
onal
com
p
l
e
xi
t
y
i
n
enco
di
n
g
a
n
d
deco
di
n
g
p
r
oc
esses ha
s ari
s
e
n
. T
h
i
s
has
p
r
om
pt
ed n
u
m
e
rous
resea
r
c
h
er
s t
o
resol
v
e c
o
nse
que
nt
pr
obl
em
s i
n
diffe
re
nt way
s
.
One
o
f
t
h
e t
y
pi
cal
ap
pr
oache
s
p
r
o
p
o
ses i
m
pr
ovi
ng
o
v
e
r
al
l
p
e
rf
orm
a
nce by
enha
nci
n
g
exi
s
t
i
ng
vi
de
o
technologies. Research on distribute
d
pr
ocessi
n
g
t
h
at
uses cl
o
u
d
co
m
put
i
ng t
echn
o
l
o
gy
has al
s
o
bee
n
p
r
op
o
s
ed
. In
[2
],
Had
oop
[3
] propo
sed sp
lit an
d m
e
rg
e arch
itectu
r
e to
p
r
o
cess
H.2
6
4
en
co
d
i
n
g
in
a
di
st
ri
b
u
t
e
d m
a
nne
r.
In
[4]
,
a
paral
l
e
l
i
z
i
ng
vi
deo t
r
a
n
sc
o
d
i
n
g sy
st
em
was put
f
o
rwa
r
d
wi
t
h
t
h
e ap
pl
i
cat
i
on
of
th
e Map
Red
u
ce fram
e
wo
rk [3
]. A h
e
u
r
istic alg
o
r
ith
m
for th
e efficien
t sch
e
du
ling
o
f
p
a
rtitio
n
e
d
v
i
deo
segm
ent
s
was pr
op
ose
d
, s
o
t
h
at
t
r
ansc
od
i
ng ca
n be e
x
ecut
e
d
whi
l
e
con
s
i
d
eri
ng t
a
sk o
v
e
r
hea
d
a
nd t
h
e
com
put
i
ng ca
paci
t
y
of
a s
y
st
em
. The di
st
ri
but
e
d
vi
de
o
pr
ocessi
ng i
n
t
h
e
s
e st
u
d
i
e
s f
o
cuse
d m
a
i
n
l
y
o
n
enco
di
n
g
t
i
m
e
, but
beca
use p
r
oces
si
n
g
spee
d an
d effi
ci
ent
bi
t
rat
e
cont
ro
l
m
e
t
hod
s hav
e
been hi
ghl
i
g
ht
ed a
s
com
pone
nts
of HE
VC c
o
ding
technol
ogy for processi
ng
UHD quality videos, studyi
ng video
enc
o
ding from
vari
ous
vi
e
w
p
o
i
nt
s i
s
nece
ssar
y
.
3.
SYSTE
M
ARCHITECT
URE
To m
easure
di
st
ri
but
e
d
e
n
co
di
n
g
p
e
r
f
o
r
m
a
nce,
we
devel
o
ped
a di
st
ri
but
ed H
E
VC
e
n
c
odi
ng
sy
st
em
base
d o
n
t
h
e
H
a
do
o
p
f
r
am
ewor
k. Fi
gu
re
1 s
h
o
w
s t
h
e
de
vel
ope
d sy
st
em
ar
chi
t
ect
ure a
nd
schem
a
t
i
c
workfl
ow
.
Th
e system workflow larg
ely co
nsists o
f
p
a
rt
itio
n
i
n
g
, en
codin
g
,
an
d
co
m
b
i
n
ing
step
s.
In
t
h
e p
a
rtitio
n
i
ng
step
,
a v
i
d
e
o
p
a
rtitio
n
e
r p
a
rtitio
n
s
an
in
pu
t v
i
d
e
o
fo
r d
i
stri
b
u
t
ed
en
co
d
i
n
g
. In th
e en
cod
i
ng
step
, p
a
rtitio
n
e
d
v
i
d
e
o
segm
ents are t
r
ans
f
erre
d to
mappers
whe
r
e
enc
odi
ng is
perform
e
d. An encode
d vi
de
o from
each m
a
p tas
k
is
t
h
en
m
e
rged
t
h
ro
u
g
h
t
h
e
re
du
cer
t
o
cre
a
t
e
a fi
nal
HE
VC
e
n
code
d vi
deo fi
l
e
. Gi
ve
n t
h
at
the de
vel
o
ped s
y
st
em
is base
d on
Hadoop e
x
ecuti
on, eac
h m
a
p task recei
ves a
split vide
o
on standby from the vi
de
o pa
rt
itioner
wh
en
t
h
e cu
rren
t en
cod
i
ng
task
is co
m
p
lete.
Fi
gu
re
1.
Sy
st
em
archi
t
ect
ure
of
di
st
ri
but
e
d
e
n
co
di
n
g
f
o
r
H
E
VC
.
4.
VIDE
O P
A
R
T
ITIONI
NG
SCHE
ME
4.
1.
No P
a
rti
t
ionin
g
Sc
heme
Item
B
in
Figure 2 d
e
no
tes the exp
ected
resu
lt for th
e en
cod
i
ng
task witho
u
t
p
a
rtition
i
ng
A (e.g.,
raw
vi
de
o)
. As sh
o
w
n i
n
B
,
t
h
e I
-
f
r
am
e and G
O
P are ge
ne
ra
t
e
d fol
l
o
wi
n
g
t
h
e enc
odi
ng
o
p
t
i
on s
p
eci
fi
ed
at
t
h
e
l
o
we
r ri
g
h
t
si
de of t
h
e fi
g
u
re
. The i
n
t
e
r-c
o
d
i
n
g pr
ocesse
s
t
h
ro
ug
h t
h
e p
i
ct
ure refe
renc
e perf
orm
e
d i
n
t
h
i
s
process are indicated by an
arrow.
Th
at is, B is a
b
a
sic criterio
n
th
at
can com
p
are distributed enc
odi
ng
per
f
o
r
m
a
nce.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
To
ward Con
t
en
t-Aw
a
r
e Vi
d
e
o
Pa
rtitio
n
i
n
g
Meth
o
d
s
f
o
r
Di
strib
u
t
ed HEVC Vid
e
o …
(Byoung-Dai
Lee)
57
1
4.
2.
No P
a
rti
t
ionin
g
Sc
heme
Nod
e
cou
n
t
–
b
ased
v
i
d
e
o
partitio
n
i
ng
is
a b
a
sic
p
a
rtiti
o
n
i
n
g
sch
e
m
e
u
s
ed
i
n
m
a
n
y
ex
isti
ng
di
st
ri
b
u
t
e
d en
c
odi
ng st
udi
es.
It
em
C
i
n
Fi
gu
re 2 sh
o
w
s an
exam
pl
e of en
codi
ng
resul
t
s
deri
ved
by
uni
fo
rm
l
y
p
a
rtitio
n
i
n
g
a
v
i
d
e
o
in
t
o
n
nod
es.
As a
v
i
d
e
o
is p
a
rtitio
n
e
d, a p
a
rtitio
n
i
ng area (PA) is
created
.
Th
is PA is an
i
n
p
u
t
u
n
i
t
of
di
st
ri
but
e
d
enc
o
d
i
ng.
I
n
C
,
t
h
e
r
e
f
o
r
e,
(x/
n
) P
A
s
are creat
e
d
, w
h
i
c
h i
s
cal
cul
a
t
e
d by
di
vi
di
ng
no
de
co
un
ts
(
e
.g
., n) b
y
th
e nu
m
b
er
o
f
fr
am
es in
A (e.g
., x). Here, th
e en
cod
i
ng
resu
lt of C
d
i
ffers fro
m
th
at o
f
B in
two
ways. Fi
rst, th
e resu
lt o
f
C is l
i
m
ited
in
term
s
o
f
th
e us
e of the refe
re
nce pictur
es. For exam
ple, GOP 2 in
B
can
re
fere
nc
e
any pictures in GOP 2, so
that
a
n
op
t
i
m
um
estim
at
ed val
u
e f
o
r i
n
t
e
r-c
odi
ng
ca
n
be
d
e
ri
ve
d
.
In
C,
h
o
wev
e
r, GOP 2
is
p
a
rtitio
n
e
d
in
to
GOP
2
'
; th
u
s
, the referen
ce
p
i
ctu
r
ed
canno
t be u
s
ed
i
n
th
e sa
m
e
m
a
nner t
h
at
i
s
use
d
f
o
r B
.
I
n
C
,
t
h
e
r
ef
o
r
e,
l
i
m
i
t
a
t
i
ons on
a usa
b
l
e
refe
r
e
nce pi
ct
ure c
a
n ari
s
e, l
e
a
d
i
ng t
o
inefficie
n
t inte
r-c
odi
ng.
A la
yout c
h
ange
in the
I-
fram
e
can
occ
u
r in accorda
n
ce
with the
PA
form
ed,
constituting anothe
r diffe
re
nc
e. Because
the
first fram
e
of the form
ed PA
should be the
I-fram
e
when intra-
codi
ng i
s
bei
n
g
perf
orm
e
d, t
h
e overal
l
bi
t
rat
e
i
n
creases as
th
e nu
m
b
er o
f
I-fram
es with
h
i
g
h
b
it rates rises. In
su
mm
ary, p
a
rtitio
n
e
rs are co
m
p
elled
to
u
s
e referen
ce
p
i
ctu
r
es in
a
li
mited
m
a
n
n
e
r b
ecau
s
e o
f
v
i
d
e
o
p
a
rtitio
n
i
n
g
, and
an
in
crease i
n
I-fram
e
s can
cau
se a rise i
n
th
e b
it
rates
o
f
en
cod
e
d
v
i
d
e
o
.
4.
3.
GOP
-b
ase
d
P
a
rti
t
i
o
ni
n
g
Scheme
W
i
t
h
reg
a
rd
to GOP-b
ased
v
i
d
e
o
p
a
rtition
i
ng
sch
e
m
e
, D in
Figu
re
2
sh
ows th
e ex
p
ected
resu
lt of
GOP-b
a
sed
v
i
d
e
o partitio
n
i
ng
an
d en
cod
i
ng
.
Du
e t
o
th
e
p
a
rtitio
n
i
n
g
, this sch
e
m
e
also in
vo
l
v
es the l
i
m
i
ted
use
of re
fere
nc
e pictures
bet
w
een GOPs
. For exam
ple,
G
O
P 1
i
n
D
ca
n
not refe
re
nce
t
h
e I-
fram
e
in GO
P 2
(which c
oul
d be used
pre
v
iously) b
ecause
of partitioni
ng. Howe
ver, this
schem
e
is different from
the node
-
co
un
t
b
a
sed
sch
e
m
e
in
th
at the GOP itself can
n
o
t
b
e
p
a
rtitio
n
e
d
,
as sh
own
in GOP
2
'
, GOP
3
'
, and
GOP 4
'
in
C. Co
m
p
ared
with
th
e resu
lt o
f
B,
th
erefo
r
e, th
e resu
lt o
f
D sho
w
s
no
ad
d
ition
a
l in
crease in
I-fram
e
s, and
thus pre
v
ents an
increas
e i
n
bitrates.Ne
v
e
r
theless,
GOP
-based
p
a
rtitio
n
i
n
g
m
a
y p
a
rtiti
o
n
seg
m
en
ts t
h
at are
hi
g
h
er
o
r
l
o
we
r t
h
a
n
t
h
e
n
u
m
b
er
of
n
o
d
es t
h
at
pe
rf
orm
en
codi
ng
,
depe
n
d
i
n
g o
n
c
o
nt
en
t
t
y
pe. Thi
s
si
t
u
at
i
on
di
m
i
ni
shes t
h
e
effi
ci
ency
of
t
h
e enc
odi
ng s
p
eed
, w
h
i
c
h i
s
l
o
wer t
h
a
n
t
h
at
achi
e
ved
wi
t
h
no
de c
o
u
n
t
-
base
d
p
a
rtitio
n
i
n
g
.
Fig
u
re 2
.
Exp
e
cted
resu
lts o
f
d
i
stribu
ted
en
co
d
i
n
g
with
v
i
deo
p
a
rtitio
n
i
n
g
.
5.
EVAL
UATI
O
N
We de
vel
o
pe
d
a di
st
ri
b
u
t
e
d
enco
di
n
g
sy
s
t
em
by
usi
n
g
t
h
e Ha
do
o
p
fram
e
wor
k
an
d pe
rf
orm
e
d
enco
di
n
g
vi
a t
h
e HE
VC
t
e
st
m
odel
(HM
1
4
.
0
)
[5]
.
T
h
e
m
a
xim
u
m
num
ber of exec
u
t
i
on n
o
d
es u
s
e
d
i
n
t
h
e
expe
ri
m
e
nt
i
s
38
. As t
h
e e
n
c
odi
ng
opt
i
o
n o
f
HM
i
s
14.
0,
GO
P was set
t
o
3
2
. Pre
d
i
c
t
i
o
n bet
w
ee
n f
r
a
m
es used
t
h
e "R
an
dom
-Access" m
ode,
and "
D
ec
odi
n
g
R
e
f
r
esh
T
y
p
e
"
used
t
h
e C
R
A m
ode. I
n
a
d
di
t
i
on, a
n
i
n
p
u
t
raw
v
i
d
e
o
used
a 4K v
i
d
e
o co
m
p
risin
g
1
,
4
0
0
frames. To
ev
al
uate th
e effect
of th
e
v
i
d
e
o
p
a
rtitio
n
i
n
g
sch
e
me o
n
enco
di
n
g
pe
rf
o
r
m
a
nce, we co
nd
uct
e
d e
x
peri
m
e
nt
s i
n
whi
c
h cha
nge
s i
n
b
i
t
rat
e
and PS
NR
were m
easure
d
.
Figure
3 s
h
ows
the c
h
anges i
n
enc
odi
ng tim
e
accordi
n
g to
t
h
e
num
ber
of node
s.
When t
h
e num
b
er of
nodes i
s
1
,
t
h
is nu
m
b
er refers to th
e
en
cod
i
ng
resu
lt with
v
i
d
e
o
partitio
n
i
ng
.
Wh
en th
e
nu
m
b
er of
n
o
d
e
s is two or
m
o
re, th
is n
u
m
b
er refers t
o
th
e d
i
stribu
ted en
cod
i
ng
result. Th
e no
p
a
rt
itio
n
i
n
g
sch
e
m
e
ex
h
i
b
its th
e lo
ng
est
en
cod
i
ng
tim
e
(see Figu
re
3
)
.
As th
e nu
m
b
er o
f
nod
es
in
creases, th
e no
d
e
-b
ased
p
a
rtitionin
g
sch
e
m
e
u
s
es less
en
cod
i
ng ti
m
e
th
an
d
o
e
s th
e GOP-b
a
sed partitio
n
i
ng
sche
m
e
. Th
is
findin
g
is attribu
t
ed
t
o
th
e fact
th
at the
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E V
o
l
.
5, No
. 3,
J
u
ne 2
0
1
5
:
56
9 – 5
7
8
57
2
n
u
m
b
e
r
of PAs fo
rm
ed
v
i
a
n
o
d
e
-b
ased p
a
rtitio
n
i
ng
is
t
h
e sam
e
as th
e nu
m
b
er of
no
d
e
s
u
s
ed. Thu
s
, all
encoding tas
k
s
are
perform
ed in pa
rallel.
Fi
gu
re
3.
Enc
o
di
n
g
Ti
m
e
Th
e GOP-b
a
sed
p
a
rtitio
n
i
ng
sch
e
m
e
fo
rm
s
4
4
PA
s,
wh
ich
ex
ceeds th
e nu
m
b
er o
f
m
a
x
i
m
u
m
allo
wab
l
e
n
odes. As a
resu
lt, th
is p
a
rtitio
n
i
n
g
cann
o
t
b
e
perfo
r
m
e
d
with
i
n
on
e
p
a
rallel ex
ecu
tion
an
d
n
e
eds
stan
db
y PAs.
Th
is
resu
lts in an en
cod
i
ng
t
i
m
e
th
at is
long
er th
an
t
h
at
ob
serv
ed
i
n
n
ode-b
ased p
a
rtitio
n
i
n
g
sch
e
m
e
. N
o
te
th
at wh
en th
e
n
u
m
b
e
r
o
f
nodes f
a
lls to
b
e
t
w
een 10
and
20
, t
h
e two
sche
m
e
s ex
h
i
b
it si
m
i
lar
en
cod
i
ng
tim
e. Th
is
resu
lt is
attrib
u
t
ed to
t
h
e sim
u
lta
n
e
o
u
s
pro
cessing
cap
a
b
ility o
f
stan
db
y PAs. Figu
re 4
sh
ows the ch
an
g
e
s in
b
it rat
e
s an
d
en
co
d
i
ng
ti
m
e
s. Th
e no
p
a
rtitio
n
i
ng
sch
e
m
e
ex
h
i
b
its th
e lo
west
b
it rate.
Th
is find
ing
is d
u
e
to
th
e fact th
at th
ere is
n
o
in
creas
e in
the num
ber of I-fram
e
s, and there is an abs
e
nce of
li
mitatio
n
s
o
n
t
h
e use of referen
ce p
i
ct
u
r
es.
Th
e
GOP-b
a
sed
p
a
rtitio
n
i
ng
sch
e
m
e
creates h
i
gh
er
b
it rates th
an
d
o
e
s th
e no
p
a
rtitio
n
i
n
g
sch
e
me, b
u
t
no
ch
an
g
e
in
b
it rates o
c
cu
rs ev
en
with
an
i
n
creased
n
u
m
b
e
r
o
f
nod
es.
Fi
gu
re 4.
B
i
t
r
at
e
Th
is resu
lt is attrib
u
t
ed
t
o
th
e fact th
at the
num
ber o
f
PAs r
e
m
a
i
n
s
con
s
t
a
nt
, re
ga
r
d
l
e
ss o
f
t
h
e
n
u
m
b
e
r o
f
n
odes.
In
con
t
rast,
in
th
e n
o
d
e
-based
p
a
rtitio
n
i
n
g
sch
e
m
e
,
th
e
n
u
m
b
e
r o
f
PAs
ch
ang
e
s with
th
e
num
ber of n
o
d
e
s,
so
t
h
at
t
h
e num
ber of I-
fr
am
es
and refe
r
e
nces pi
ct
ures
chan
ges
a
s
wel
l
.
Thi
s
feat
ure
resul
t
s
in
ch
ang
e
s in
bit rate. In
p
a
rticu
l
ar, th
e nod
e-b
a
sed
p
a
rtitio
n
i
ng
sch
e
m
e
e
x
h
i
b
its lo
wer
bit rates th
an
d
o
es th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
To
ward Con
t
en
t-Aw
a
r
e Vi
d
e
o
Pa
rtitio
n
i
n
g
Meth
o
d
s
f
o
r
Di
strib
u
t
ed HEVC Vid
e
o …
(Byoung-Dai
Lee)
57
3
GOP-b
a
sed
p
a
rtitio
n
i
ng
sch
e
me b
y
u
p
to
27
nod
es. Be
yon
d
th
is n
u
m
b
e
r, n
o
d
e
-b
ased
p
a
rtitio
n
i
n
g
exh
i
b
its
h
i
gh
er b
it rates. Th
is find
i
n
g is attrib
u
t
ed
to th
e in
cr
ease i
n
t
h
e
num
ber
of I-fram
es and
reference
lim
its a
s
the
num
ber o
f
n
o
d
es i
n
crea
ses.
Al
t
h
o
u
gh t
h
e
PSNR
ex
hi
bi
t
s
som
e
change
s
,
they are
m
i
n
i
m
a
l (see Figure 5).
Th
ese
d
i
fferences in
b
it rate an
d
en
co
d
i
n
g
ti
me b
a
sed
on
the v
i
d
e
o
p
a
rtitio
n
i
ng
sch
e
m
e
in
d
i
cate th
at th
e v
i
d
e
o
p
a
rtitio
n
i
n
g
sch
e
m
e
is a crit
i
cal facto
r
th
at
affects
o
v
e
ra
ll en
cod
i
ng
p
e
rform
a
n
ce. It is p
a
rticu
l
arly i
m
p
o
r
tant
to consi
d
er tha
t
the overall encodi
ng
performance ch
anges according to whet
her eac
h partitioni
ng sc
hem
e
is
ap
p
lied. Hen
c
e, it is
n
ecessary to
u
s
e a v
i
d
e
o
p
a
rtitio
n
i
ng sch
e
m
e
th
at c
a
n
gu
aran
tee efficien
cy in
ord
e
r to
i
m
p
r
ov
e th
e ov
er
all en
cod
i
ng
p
e
rf
or
m
a
n
ce. Howev
e
r, it
is h
a
rd
t
o
d
e
term
in
e wh
ich of th
e two
p
r
evio
u
s
l
y
men
tio
n
e
d
p
a
rtitio
n
i
n
g
sch
e
mes, (i.e., on
e b
a
sed
on
th
e n
u
m
b
e
r of no
d
e
s and
ano
t
h
e
r
b
a
sed
on
GOP)
gua
ra
nt
ees m
o
re effi
ci
e
n
t
per
f
o
r
m
a
nce. Thi
s
ki
n
d
o
f
p
r
o
b
l
e
m
occurs
due t
o
t
h
e di
ffe
renc
e bet
w
ee
n h
o
w
t
h
e
two
p
a
rtitio
n
i
ng
sch
e
m
e
s co
nsid
er t
h
e
v
i
d
e
o p
a
rtitio
n
i
ng
p
r
o
cess. Ev
en tho
ugh
th
e two
p
a
rtitio
n
i
n
g
sch
e
m
e
s
account
for the
enc
odi
ng vi
de
o a
n
d system
e
nvi
ronm
ent in
the
process of partitioni
ng
, they do not acc
ount
for
the vi
deo itself and thus
cannot gua
r
antee e
f
fici
ent
enc
o
di
n
g
per
f
o
r
m
a
nce
fo
r di
ve
rse vi
d
e
o.
Fi
gu
re 5.
PS
N
R
We m
easured
the efficiency
change
of the two
partitioni
ng schem
e
s according to c
ont
ent type in
o
r
d
e
r to
clari
f
y sp
ecific reaso
n
s
for th
is pro
b
l
em
. Mo
reov
er, we reco
gn
ize th
at a p
a
rtitio
n
i
n
g
sch
e
me th
a
t
doe
s not
acc
o
unt
f
o
r t
h
e
vi
deo c
ont
e
n
t
i
s
pro
b
l
e
m
a
t
i
c
,
and
we pr
o
v
i
d
e an anal
y
s
i
s
of t
h
e re
sul
t
s
of t
h
e
measurem
ent (see bel
o
w for s
p
ecific
descri
ptions
).
6.
CONTENT T
Y
PE
Th
is ch
ap
ter
describ
e
s th
e chan
g
e
of
p
a
rtitio
n
i
n
g
sch
e
m
e
p
e
rform
a
n
ce base
d
on
th
e characteristics
of
t
h
e
co
nt
ent
o
f
t
h
e
vi
de
o.
Fi
rst
,
we
de
fi
ned
vi
de
o c
o
n
t
ent
as st
at
i
c
and
dy
nam
i
c. Next
,
we
per
f
o
rm
ed
en
cod
i
ng
b
y
ap
p
l
ying
the
p
a
rtitio
n
i
ng
sch
e
me b
a
sed on
no
d
e
nu
m
b
er and
th
e partitio
n
i
n
g
sch
e
m
e
b
a
sed
o
n
GOP size
on
each
o
f
th
e d
e
fi
n
e
d co
n
t
en
t typ
e
s. Fin
a
lly
,
we co
m
p
ared the efficien
cy of th
e t
w
o p
a
rtitio
n
i
n
g
schem
e
s according to the
content type
by
com
p
aring the
measurem
ent res
u
lts.
In
cases
wh
ere th
e v
i
d
e
o
con
t
en
t was sim
p
le an
d
h
a
d
little
v
a
riation
,
we defin
e
d
its co
n
t
en
t typ
e
as
static, i.e., stat
ic-type cha
r
act
eristics m
ean the
vide
o is
c
o
m
posed of rel
a
t
i
v
el
y
sim
p
le scenes
when vi
sually
obs
erved. M
o
reove
r, static-type c
onte
n
t has
the c
h
aracte
r
istic o
f
a relativ
ely lo
w b
it
rate an
d en
cod
i
ng
tim
e.
As the s
u
bject
of t
h
e content does
not vary that
m
u
ch
, t
h
e sim
i
larity b
e
tween
the frames is h
i
g
h
,
wh
ich
increases i
n
ter-codi
ng
process
efficien
cy. Fi
gure 6
presents
a sim
p
le exa
m
ple of t
h
is cha
r
acteristic. Fra
m
es 1
to
60
in
th
e
p
i
ctu
r
e tend
to
fo
rm
a relativ
el
y lo
w b
it rate
an
d
en
co
d
i
n
g
t
i
m
e
d
i
strib
u
tion
co
m
p
ared
to o
t
h
e
r
i
n
t
e
rval
s. M
o
r
e
ove
r,
w
h
en
vi
sual
l
y
det
e
rm
ini
n
g t
h
e c
o
m
p
lexi
t
y
of t
h
e
vi
deo
by
act
ual
l
y
pl
ay
i
ng t
h
e
r
e
l
e
vant
i
n
t
e
rval
(
w
here
a
m
onkey
m
oves sl
o
w
l
y
at
t
h
e cent
e
r o
f
t
h
e screen
), i
t
can be c
o
nfi
r
m
e
d t
h
at
i
t
i
s
a rel
a
t
i
v
el
y
sim
p
le type video. He
nce, by aggre
g
ating these char
acteristics, we d
e
term
in
ed
th
e co
n
t
en
t th
at p
o
ssessed
v
i
su
al ch
aracteristics si
m
ilar t
o
fram
e
1
to
60 of th
e v
i
d
e
o was th
e static typ
e
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E V
o
l
.
5, No
. 3,
J
u
ne 2
0
1
5
:
56
9 – 5
7
8
57
4
Fi
gu
re
6.
R
e
sul
t
s of
enc
o
di
n
g
.
6.
1. Vi
de
o Co
nten
t
T
y
pe De
ci
si
on
In o
r
der t
o
de
f
i
ne t
h
e cont
e
n
t
t
y
pe of t
h
e vi
deo
,
we use
d
t
h
e fol
l
owi
ng
m
e
t
hod. Fi
rst
,
we pe
rf
orm
e
d
enco
di
n
g
o
n
a
num
ber o
f
vi
d
e
os t
h
at
co
nt
ai
n di
f
f
ere
n
t
t
y
pes of co
nt
e
n
t
and c
o
l
l
ect
ed bi
t
rat
e
and enc
o
di
n
g
tim
e
results. We the
n
visua
lly conf
irm
e
d the cha
r
acterist
i
cs of the
cont
ent by playing each vi
de
o.
We the
n
col
l
ect
ed rel
e
v
a
nt
i
n
fo
rm
ati
on.
Fi
nal
l
y
, t
h
e
vi
de
o c
ont
e
n
t
t
y
pe was
de
fi
n
e
d
usi
n
g t
h
e
c
o
l
l
ect
ed i
n
fo
r
m
at
i
on.
The s
p
ecific c
h
aracteristics of
each
of the
de
fined c
o
ntent ty
pes a
r
e
descri
bed
below.
To explain eac
h type of c
o
ntent, we
prese
n
t the enc
o
di
n
g
re
sul
t
s
of a p
a
rt
i
c
ul
ar vi
deo as
an exam
pl
e.
The
vide
o use
d
for e
n
c
odi
ng was
a
4K
vide
o c
o
m
pos
ed of 200 fram
es and th
e 32
GOP size.
Random
access
m
ode was
use
d
t
o
e
n
co
de t
h
e
vi
de
o
6.
2. St
ati
c
-T
y
p
e
C
o
n
t
en
t
We im
p
l
e
m
en
ted
th
e en
cod
i
n
g
o
f
static-typ
e con
t
en
t
u
sing
a p
a
rtitio
n
i
ng
sch
e
m
e
b
a
sed
on
the
n
u
m
b
e
r
o
f
nod
es an
d a p
a
rtitio
n
i
n
g
sch
e
me b
a
sed
on
GOP.
We
d
e
scrib
e
d
th
e characteristics of each
partitioni
ng sc
hem
e
according to the res
u
lts.
W
e
pe
rfor
m
e
d enc
odi
ng s
u
ccessively in
a single node
without
d
i
stribu
ted
processin
g
of t
h
e PA to
m
easu
r
e th
e
p
e
rf
orm
a
n
ce o
f
th
e v
i
deo
'
s
p
a
rtitio
n
i
ng
asp
ects. Figu
re 7
shows the e
n
c
odi
ng
result of video that is partitioned
by
applying each of the partitioni
ng sc
hem
e
s wi
th a 70
fram
e
vi
deo co
m
posed of st
at
i
c
-t
y
p
e cont
e
n
t
.
Fi
gu
re 7-
A p
r
esent
s
bi
t
rat
e
and enc
odi
n
g
t
i
m
e
di
st
ri
but
i
on
o
f
th
e en
cod
i
ng
im
p
l
e
m
en
tatio
n
resu
lts wh
ere static- typ
e
co
n
t
en
t was no
t p
a
rtitio
n
e
d
.
Fra
m
es 1
,
3
2
,
and
64
in
the figure s
h
ow the l
o
cation
of
t
h
e I-fram
e. Since the
I-fra
me is basi
cally
in
tra-cod
e
d, it h
a
s sho
r
ter
enco
d
i
ng
ti
m
e
co
m
p
ared
to
o
t
h
e
r fram
e
s
th
at are in
ter-cod
ed
. Meanwh
ile, in
term
s o
f
b
it rate d
i
stribu
tio
n, th
e I-fram
e
fo
rm
s a consi
d
erabl
y
hi
g
h
bi
t
rat
e
com
p
ared
t
o
ot
her
fram
e
s. He
nce, i
n
t
h
e di
st
ri
but
e
d
enco
di
n
g
pr
oc
ess of
static-type content, I-fram
e
s
can be id
e
n
tifi
e
d as a critical
factor t
h
at affe
ct
s t
h
e ove
ral
l
bi
t
rat
e
and e
n
codi
ng
ti
m
e
. Fig
u
r
e 7-C sh
ows th
e
en
cod
i
ng
im
p
l
e
m
en
tatio
n
resu
lt wh
ere th
e
p
a
rtitio
n
i
n
g
sch
e
m
e
b
a
sed
o
n
th
e
num
ber o
f
n
o
d
e
s i
s
appl
i
e
d us
i
ng t
w
o n
ode
s.
C
form
s t
w
o PAs t
h
at
ha
ve a l
e
ngt
h
of
35 f
r
a
m
e
s. Si
nce t
h
e 36t
h
fram
e
is encoded as a
n
I-fra
m
e, an
in
crease in
b
it rate o
c
cu
rs. In
C,
th
e b
it rate is 36
5Kb
and
th
e enco
d
i
ng
t
i
m
e
i
s
8,87
8 seco
nds
, w
h
i
c
h sh
ows a 1
5
.
9
% i
n
crease i
n
bit rate and a 2% decr
ea
se in encoding tim
e
co
m
p
ared
to
A. Figure 7
-
B
sh
ows th
e enco
d
i
n
g
resu
lts
o
f
ap
p
l
ying
a p
a
rtitio
n
i
n
g
sch
e
m
e
b
a
sed
on
GOP.
Un
lik
e
Figu
re
7
-
C, B sh
ows
n
o
ch
an
g
e
in
I-fram
e fo
llo
win
g
t
h
e p
a
rtitio
n
i
ng
, an
d
th
e i
n
ter-cod
ing
pro
cess
also
sho
w
s little ch
an
g
e
.
When
co
m
p
aring
th
e
m
easu
r
ed
b
it rate an
d
en
cod
i
ng
ti
m
e
,
we see th
e b
it rate
i
n
crease
d
0.
6%
an
d t
h
e
enc
o
di
ng
t
i
m
e
decrea
sed
2.3% c
o
m
p
ared to
A.
Mo
reo
v
e
r, we i
m
p
l
e
m
en
ted
en
cod
i
ng
u
s
ing
a
p
a
rtitio
n
i
n
g
sch
e
m
e
b
a
sed
o
n
no
d
e
s
and
ch
ang
e
d
th
e
n
u
m
b
e
r
o
f
nod
es to fiv
e
. Th
is was to
acco
u
n
t
for th
e ch
aracteristics
o
f
t
h
e
p
a
rtition
i
ng
sch
e
m
e
b
a
sed
on
n
o
d
e
s, wh
ich
variab
ly form
s
a v
i
d
e
o
'
s PA
based
on
th
e num
b
e
r o
f
n
o
d
e
s
rath
er than
th
e p
a
rtitio
n
i
ng
sch
e
m
e
b
a
sed
o
n
GOP. Figu
re
7-D sho
w
s th
e en
cod
i
n
g
resu
lts of t
h
e p
a
rtitio
n
i
ng
sch
e
m
e
b
a
sed
on
fi
v
e
no
d
e
s.
In
D, it
is clear that five I-fram
e
s are gene
rated a
s
fi
ve PAs
are
formed. He
nce, c
h
anges i
n
bit rate and e
n
c
odi
ng tim
e
fo
llow.
Th
e enco
d
i
n
g
resu
lts
in
d
i
cate a
b
it rate o
f
40
9Kb
an
d en
cod
i
ng
t
i
m
e
o
f
8,094
seco
nd
s. Co
m
p
ar
ing
with
A, th
e b
i
t
rate showed an
in
c
r
ease of
30%, while
the
encoding tim
e
showe
d
a
d
ecrease of 10.6%
.
Eve
n
th
ou
gh
it is possib
l
e to
say t
h
at effi
ciency increase
d
in
D co
m
p
ared t
o
C in term
s of
encoding tim
e,
the bit
rate in
creased
to
up
to
3
0
% ov
erall.
As a resu
lt, it can
no
t be said
to
b
e
ef
ficien
t. Fin
a
lly, in
p
a
rtition
i
ng
static-
typ
e
co
n
t
en
t,
alth
o
ugh
th
e
partitio
n
i
ng
sche
m
e
was
b
a
sed
on
the nu
mb
er
o
f
no
des
an
d
can
d
i
m
i
n
i
sh
th
e
en
cod
i
ng
tim
e
co
m
p
ared
t
o
the p
a
rtition
i
ng
sch
e
m
e
b
a
sed
on
GOP, it h
a
d
a d
i
sad
v
a
n
t
ag
e in
th
at it m
a
d
e
the
b
it rate m
o
re inefficien
t.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
To
ward Con
t
en
t-Aw
a
r
e Vi
d
e
o
Pa
rtitio
n
i
n
g
Meth
o
d
s
f
o
r
Di
strib
u
t
ed HEVC Vid
e
o …
(Byoung-Dai
Lee)
57
5
(A)
No Partition
i
ng
Sch
e
m
e
(B) GOP-size b
a
sed
Partitio
nin
g
Sch
e
m
e
(C)
N
o
d
e
-count b
a
sed
Partitio
n
i
ng
Sch
e
m
e
:
2N
o
d
es
(D)
N
o
d
e
-count b
a
sed
Partitio
n
i
ng
Sch
e
m
e
:
5N
o
d
es
Fig
u
re
7
.
Resu
l
t
s o
f
static-type con
t
en
t en
cod
i
ng
with
v
i
d
e
o
p
a
rtitio
n
i
n
g
.
6.
3. D
y
n
a
mi
c-
T
y
pe Co
nte
n
t
Dy
nam
i
c-t
y
pe cont
e
n
t
re
fer
s
t
o
c
ont
e
n
t
w
h
ose c
h
a
r
acteris
tics are
opposi
te those
of t
h
e
static-type.
That
i
s
, dy
nam
i
c-t
y
pe cont
e
n
t
i
s
co
m
pose
d
o
f
o
b
ject
s t
h
at
h
a
ve com
p
l
e
x s
h
ape
s
an
d m
a
ny
al
t
e
rat
i
ons. Hence
,
dynam
i
c-type conte
n
t is com
pos
ed
of c
o
m
p
licated scenes
whe
n
visually
assessed, a
nd i
t
has the
cha
r
a
c
teristic
of
ha
vi
n
g
a hi
gh
bi
t
rat
e
an
d
l
o
n
g
enc
o
di
n
g
t
i
m
e
i
n
t
h
e encodi
ng
p
r
oces
s
.
Fo
r e
x
am
pl
e, t
h
e i
n
t
e
r
v
al
be
t
w
een
fram
e
80 and
20
0 i
n
Fi
g
u
r
e
6 fo
rm
s a rel
a
t
i
v
el
y
hi
gh bi
t
rat
e
and enc
o
di
n
g
t
i
m
e
co
m
p
are
d
t
o
st
at
i
c
t
y
pe
cont
e
n
t
.
When
obse
r
vi
ng t
h
e
vi
deo'
s com
p
lexi
t
y
by
pl
ay
ing a
n
i
n
t
e
rval
of t
h
e vi
deo
w
h
ere a t
i
g
er ca
uses a
h
u
g
e
water sp
ray as it rap
i
d
l
y sways its tai
l
in
th
e water, a d
r
am
atic ch
an
g
e
i
n
th
e
o
b
j
ects was con
f
irm
e
d
.
Hence
,
we de
f
i
ned t
h
e t
y
pe
of c
ont
e
n
t
t
h
at
has vi
sual
ch
aract
eri
s
t
i
c
s si
m
i
l
a
r t
o
fram
e
s 80 t
o
20
0 a
s
t
h
e
d
y
n
a
m
i
c typ
e
. We
p
e
rform
e
d
d
y
n
a
m
i
c-typ
e
co
n
t
en
t en
cod
i
n
g
b
y
app
l
yin
g
a p
a
rtitio
n
i
ng
sch
e
m
e
b
a
sed
o
n
the
n
u
m
b
e
r
o
f
nodes and
a
p
a
rtitio
n
i
n
g
sch
e
m
e
b
a
sed
on
GO
P. Figu
re 8 show
s th
e en
cod
i
n
g
resu
lt of app
l
yin
g
each
of t
h
e pa
rtitioning sc
he
mes on a
70-fram
e dyna
m
i
c-
type vide
o c
o
ntent. Figure
8
shows t
h
e e
n
c
odi
ng
result of a
pply
i
ng each
of the partitioning
schem
e
s on a
70-fram
e
dynamic-type vi
deo content. Fi
gure 8-A
sh
ows th
e
resu
lt of stand-alo
n
e
en
cod
i
ng
i
m
p
l
e
m
en
tatio
n
o
n
d
y
n
a
m
i
c-
typ
e
con
t
en
t with
ou
t p
a
rtitio
n
i
n
g
.
It
can
be c
o
n
f
i
r
m
e
d t
h
at
t
h
e
dy
n
a
m
i
c-t
y
pe cont
ent
i
s
had
hi
gh
bi
t
rat
e
di
st
ri
b
u
t
i
on
ove
ral
l
.
Howev
e
r, un
lik
e th
e
prev
i
o
usly
m
e
n
tio
n
e
d
static typ
e
co
nten
t, th
e
b
it rat
e
o
f
th
e
I-frame do
es
n
o
t
show a
distinc
tive diffe
re
nce
com
p
ared
to
othe
r b-fram
es
and
p-fram
es.
For
exam
ple, the interval be
twee
n
fram
e
s 1
5
to
25
in
A is co
m
p
o
s
ed
o
f
fram
es
th
at h
a
v
e
a b
it rate si
milar to
o
r
h
i
gh
er th
an th
at o
f
th
e I-fra
m
e.
In
part
i
c
ul
ar,
f
r
am
e 17 an
d f
r
a
m
e
25
have
b
i
t
rat
e
hi
g
h
er
t
h
an t
h
e
I-frame, which is ca
used
by a large
change
in
th
e d
y
n
a
m
i
c
-
typ
e
obj
ect com
p
o
s
itio
n
.
Th
i
s
in
creases t
h
e co
m
p
u
t
atio
n
co
m
p
lex
ity o
f
in
ter-cod
ing
pro
cess,
resu
lting
in
a hig
h
b
it rate. Hen
ce, wh
en
p
a
rt
itio
n
i
n
g
d
y
n
a
mic-typ
e
co
n
t
en
t, th
e o
v
e
rall b
it rate an
d
en
cod
i
ng
tim
e
are not
only affe
cted
by the I-
fram
e but a
r
e als
o
i
n
flue
nced by t
h
e b-fram
e and p-fram
e. Figure
8-B
sh
ows t
h
e resu
lt o
f
en
cod
i
ng
d
y
n
a
m
i
c-typ
e
co
n
t
en
t
b
y
ap
p
l
ying
th
e
partitio
n
i
ng
sche
m
e
b
a
sed
on GOP.
Co
m
p
arin
g th
e en
cod
i
ng
results with
A, th
e b
it rate
i
n
c
r
ea
sed by
0.
4%,
a
n
d
t
h
e
enc
o
ding tim
e decreas
ed
by
2.
7%. T
h
i
s
res
u
l
t
do
es n
o
t
di
ffe
r m
u
ch fr
o
m
t
h
at
of t
h
e
p
r
evi
ousl
y
m
easure
d
st
at
i
c
-t
y
p
e
cont
e
n
t
.
Hen
ce, w
e
can
say t
h
at the p
a
rtitio
n
i
ng
sch
e
m
e
b
a
sed
on
GOP
h
a
s little
d
i
fferen
ce in
efficien
cy i
n
term
s o
f
co
n
t
en
t
typ
e
.
Fig
u
re
8
-
C shows th
e resu
lt of a
n
o
d
e
-b
ased p
a
rtition
i
ng
sch
e
m
e
u
s
in
g
t
w
o
n
o
d
e
s. Co
mp
ari
n
g
t
h
e en
co
d
i
n
g
p
e
rform
a
n
ce with
A, th
e
b
itrat
e increase
d
by
3.7%, and t
h
e
en
coding tim
e decrease
d
by
2.1%
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E V
o
l
.
5, No
. 3,
J
u
ne 2
0
1
5
:
56
9 – 5
7
8
57
6
(A
)
No Partition
i
ng
Sch
e
m
e
(B) GOP-size b
a
sed
Partitio
nin
g
Sch
e
m
e
(C)
N
o
d
e
-count b
a
sed
Partiti
o
n
i
ng
Sch
e
m
e
: 2
N
od
es
(D)
N
o
d
e
-count b
a
sed
Partiti
o
n
i
ng
Sch
e
m
e
: 5
N
od
es
Fig
u
re
8
.
Resu
l
t
s o
f
d
y
n
a
m
i
c-t
y
p
e
con
t
en
t
enco
d
i
n
g
with v
i
d
e
o p
a
rtitio
n
i
ng
This
result i
n
dicates a big di
ffe
rence
in e
f
ficiency in
term
s o
f
b
it rate wh
en co
m
p
ared with
t
h
e previous
static-typ
e co
nten
t resu
lts. Th
is is b
ecau
s
e
th
e I-fram
e h
a
s a lo
w im
p
act o
n
th
e
o
v
e
rall b
it rate in
d
y
n
a
m
i
c-
typ
e
co
n
t
en
t, un
lik
e static-typ
e co
n
t
en
t. It also
showed
sim
i
lar efficiency in ter
m
s of enc
o
di
n
g
t
i
m
e
. Fi
gure
8-
D sho
w
s th
e
resu
lt
o
f
encod
i
ng
b
y
ap
p
l
y
i
n
g
t
h
e
p
a
rtiti
o
n
i
n
g
sch
e
m
e
b
a
sed
on
a no
d
e
wit
h
fi
v
e
no
d
e
s.
Co
m
p
arin
g th
e en
cod
i
ng
results with
A, th
e b
it rate
i
n
c
r
ea
sed by
3.
7%,
a
n
d
t
h
e
enc
o
ding tim
e decreas
ed
by
1
1
.1%. Co
m
p
aring
th
is resu
lt with
th
e resu
lt o
f
th
e sta
tic-typ
e con
t
en
t i
m
p
l
e
m
en
tati
on, an efficiency increase
is co
nfirm
e
d
in
ter
m
s o
f
bo
th
o
f
b
it rate an
d
en
cod
i
ng
. M
o
reo
v
e
r, co
m
p
ared
with
th
e
p
a
rtitio
n
i
n
g
sch
e
me th
at
use
d
t
w
o
no
de
s, t
h
e o
n
e wi
t
h
fi
ve n
o
d
es sh
owe
d
a co
ns
i
d
erable dec
r
eas
e in encodi
ng
t
i
m
e
d
e
sp
ite a si
m
i
lar
b
it rate. Th
is is b
ecau
s
e in
ter-cod
i
ng
h
a
s a relativ
ely efficien
t allo
catio
n
o
f
I-fram
e
s in
cases of co
m
p
licated
fram
e
s becaus
e
five PAs are
form
ed. Fo
r exa
m
ple, five PAs are form
ed in
D and fram
e 1
5
is intra-code
d int
o
I-
fram
e
. Ho
we
ver
,
as was m
e
nt
i
one
d i
n
A, a
s
fram
e
15
expresse
d relatively com
p
lica
t
ed scenes, it was
one
of
th
e fram
es th
at o
r
ig
i
n
ally form
ed
a h
i
gh
bit rate
that implem
ented inefficien
t
i
n
t
e
r-c
odi
ng
. H
e
nce
,
eve
n
though D is e
n
code
d at a high bit rate beca
use fram
e 15
is in
tra-co
d
e
d
,
t
h
ere is no
large in
crease in
bit rat
e
si
nce t
h
e f
r
am
e ori
g
i
n
al
l
y
had a hi
gh
bi
t
rat
e
. H
o
we
ve
r, the enc
odi
ng tim
e
showed a large diffe
r
ence
.
C
o
m
p
ari
ng t
h
e
enc
odi
ng
res
u
l
t
s
of f
r
am
e 15
i
n
A a
n
d
D, t
h
e enc
odi
ng t
i
m
e was
27
0 sec
o
nds
an
d
45
sec
o
n
d
s,
resp
ectiv
ely.To
su
m
u
p
th
ese ch
aracterist
i
cs, a p
a
r
titio
n
i
ng
sch
e
m
e
b
a
sed
o
n
th
e
n
u
m
b
e
r of nod
es can
produce c
h
anges in efficienc
y
accord
i
ng to content type, and it can im
ple
m
ent an efficient pa
rtition for
dy
nam
i
c-t
y
pe cont
e
n
t
.
6.
4. E
ffi
ci
enc
y
Ch
an
ge
Th
e efficien
cy ch
ang
e
of th
e two
p
a
rtitio
n
i
ng
sch
e
m
e
s acc
o
r
d
i
ng
to
th
e co
n
t
en
t typ
e
d
e
scrib
e
d
in
th
e
p
r
ev
iou
s
secti
o
n
can
b
e
summarized
as fo
llows.
Fi
rst, th
e p
a
rtitio
n
i
n
g
sch
e
m
e
b
a
sed
on
GOP
sh
owed
con
s
i
s
t
e
nt
e
n
c
odi
ng
pe
rf
o
r
m
a
nce
wi
t
h
out
a
l
a
rge
cha
n
ge
of
ef
fi
ci
ency
i
n
t
e
rm
s of
c
ont
ent
t
y
pe.
He
nc
e, i
t
can
g
u
a
ran
t
ee relativ
ely stab
le en
co
d
i
n
g
p
e
rfo
rman
ce wh
en
partitio
n
i
ng
v
i
d
e
o
th
at is co
m
p
o
s
ed
of d
i
v
e
rse co
n
t
en
t
typ
e
s. However, a
p
a
rtitio
n
i
ng
sch
e
m
e
b
a
sed
o
n
nu
m
b
er
of nod
es sho
w
ed
d
i
fferen
t
efficien
cy with
d
i
fferent
co
n
t
en
t typ
e
. It h
a
d
in
efficien
t en
cod
i
ng
p
e
rfo
r
m
a
n
ce d
u
ri
n
g
th
e
p
r
o
cess o
f
p
a
rtitio
n
i
ng
static-typ
e co
n
t
en
t
because
on inc
r
ease in hi
gh bit rate occ
u
rre
d
.
Howe
ve
r, when
pa
rtitioning dynam
i
c-type content, it showe
d
efficient enc
oding pe
rform
a
nce ove
rall, as the dec
r
ease
of encoding-tim
e
exceede
d
the increase of bit
rate.
This c
h
aracteri
s
tic are confirmed in Fi
gu
re 9, w
h
i
c
h pre
s
e
n
t
s
t
h
e cha
n
ge of bi
t
rat
e
a
n
d enco
di
n
g
t
i
m
e
of
t
h
e
two
partitioni
ng schem
e
s according to t
h
e c
onte
n
t type ex
pressed as a
percenta
ge. The
vide
o length
use
d
in
t
h
e pi
ct
ure
was
7
0
fram
e
s
an
d
t
h
e res
o
l
u
t
i
o
n was 4
K
. We
di
d not
i
m
pl
em
ent
di
st
ri
b
u
t
e
d
e
n
co
di
n
g
t
o
e
x
a
m
i
n
e
th
e ch
an
g
e
i
n
en
cod
i
ng
tim
e
an
d
b
it rate cau
sed
b
y
th
e
p
a
rtitio
n
i
ng
of the v
i
d
e
o
itself.
Hen
c
e, th
e en
co
d
i
n
g
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
To
ward Con
t
en
t-Aw
a
r
e Vi
d
e
o
Pa
rtitio
n
i
n
g
Meth
o
d
s
f
o
r
Di
strib
u
t
ed HEVC Vid
e
o …
(Byoung-Dai
Lee)
57
7
tim
e
in Figure
9 is the s
u
m
of the en
c
o
ding tim
e
of each
of
the PAs. Fi
gu
re 9-A s
h
ows t
h
e change
of
bit
rate
an
d
en
co
d
i
n
g
t
i
m
e
o
f
th
e two p
a
rtitio
n
i
n
g
sch
e
m
e
s o
n
static-typ
e con
t
en
t. Th
e
p
a
rtitio
n
i
n
g
sch
e
m
e
b
a
sed
on
GO
P di
d n
o
t
p
r
o
d
u
ce any
c
h
a
nge i
n
bi
t
rat
e
and e
n
c
o
di
n
g
t
i
m
e
regar
d
l
e
ss
of t
h
e n
u
m
b
er
of
n
odes
.
Ho
w
e
ver
,
in
th
e p
a
rtitio
n
i
ng
schem
e
b
a
sed
on
th
e
nu
m
b
er o
f
no
des, th
e en
cod
i
n
g
resu
lts ch
an
g
e
d
acco
r
d
i
ng
to
the
chan
gi
n
g
num
ber of n
ode
s. Al
t
h
o
u
gh
t
h
e enco
di
n
g
t
i
m
e
decrease
s
as
the num
b
er of nodes i
n
creas
es, the
ove
rall efficiency of t
h
e e
n
c
odi
ng wa
s low due to the
i
n
crease in th
e
h
i
gh
b
it rate.
Fig
u
re
9
-
B
sho
w
s th
e
p
e
rform
a
n
ce ch
ang
e
o
f
t
h
e two
p
a
rtitio
n
i
ng
sch
e
m
e
s o
n
d
y
n
a
m
i
c-typ
e
co
n
t
en
t. Th
e en
cod
i
ng
resu
lt
s o
f
t
h
e
p
a
rtitio
n
i
n
g
sch
e
m
e
b
a
sed
on GOP sh
owed
a si
m
ilar ch
ang
e
to
t
h
o
s
e
of
th
e static-typ
e co
n
t
en
t. Hen
c
e, th
e
p
a
rtitio
n
i
n
g
sch
e
m
e
b
a
sed
on GOP can
b
e
said
to
h
a
v
e
a
relativ
ely s
m
al
l ch
ang
e
i
n
en
cod
i
ng
efficien
cy
b
a
sed
o
n
t
h
e con
t
en
t
typ
e
. Ho
wev
e
r, th
e
p
a
rtitio
nin
g
sch
e
m
e
b
a
sed
on
th
e
num
b
e
r o
f
nod
es sh
owed
a d
i
fferen
t
resul
t
f
r
o
m
t
h
at
of st
at
i
c
-t
y
p
e
cont
e
n
t
.
I
n
part
i
c
ul
ar, i
n
case
of
dy
nam
i
c-t
ype c
ont
e
n
t
wi
t
h
t
w
o
no
des a
n
d f
o
u
r
n
o
d
e
s, th
e
b
it rate d
ecreased
m
o
re th
an th
e
static-typ
e conten
t, an
d th
e en
cod
i
ng
tim
e
decreased e
v
en
m
o
re.
As a
res
u
lt, the efficie
n
cy greatly
increased. Howeve
r, wit
h
a
steady i
n
crease in t
h
e
num
ber of nodes, eve
n
though t
h
e encodi
ng tim
e
decreased,
the
efficiency e
v
e
n
tually decrea
se
d
as t
h
e
b
it rate in
creased
.
Th
e
efficien
cy ch
an
g
e
of th
e two p
a
rtitio
n
i
n
g
sch
e
m
e
s b
a
sed
on
con
t
en
t typ
e
can
b
e
su
mm
a
r
ized
as
fo
llows. Th
e
p
a
rtitio
n
i
n
g
sch
e
m
e
b
a
sed
on GOP is ab
le to
prov
i
d
e re
lativ
ely stab
le enco
d
i
n
g
efficiency reg
a
rd
less
o
f
t
h
e
co
n
t
en
t typ
e
.
Howev
e
r, th
e
p
a
rtitio
n
i
n
g
sch
e
m
e
b
a
sed
on
th
e
nu
m
b
er o
f
nod
es
h
a
d d
i
fferen
t lev
e
ls of
efficiency accordi
ng to c
o
ntent type.
In pa
rticular, it had
high efficiency
following the i
n
cr
easi
ng number of
nodes in
dyna
mic-type cont
ent. Ho
we
ve
r, if the num
ber of
nodes e
x
ceed
ed a ce
rtain level, the bit rate
in
creased
,
wh
i
c
h
ev
en
tu
ally cau
sed
less efficien
t en
cod
i
ng
.
Hen
c
e, ev
en th
ou
gh
a
p
a
rtitio
n
i
ng
sch
e
m
e
b
a
sed
o
n
th
e
nu
m
b
er
o
f
nod
es
p
r
ovid
e
s un
stab
le en
cod
i
ng
ef
f
i
cie
n
cy overall, it has the
ch
a
r
act
eristic of
bei
n
g able
t
o
p
r
ovi
de
opt
i
m
al
encodi
n
g
e
ffi
ciency for dynamic-type
conte
n
t.
To
su
m
u
p
,
app
l
yin
g
a
p
a
rtiti
o
n
i
n
g
sch
e
m
e
for d
i
st
ribu
ted
en
cod
i
ng
t
h
at d
o
e
s no
t acco
u
n
t
fo
r
v
i
deo
cont
e
n
t
ca
nn
ot
g
u
ara
n
t
ee
opt
im
al
encodi
ng
pe
rf
orm
a
nce,
an
d i
t
ca
n c
r
eat
e a p
r
o
b
l
e
m
of l
o
we
r e
n
codi
ng
efficiency.
Fig
u
re 9
.
En
cod
i
ng
resu
lts o
f
co
n
t
en
t
typ
e
with
v
i
d
e
o
p
a
rtitio
n
i
ng
7.
CO
NCL
USI
O
N
AN
D F
U
T
U
RE W
O
R
K
Thi
s
pa
per
des
c
ri
be
d t
h
e effe
ct
of vi
de
o pa
rt
i
t
i
oni
ng sc
hem
e
s whe
n
pe
rf
or
m
i
ng di
st
ri
b
u
t
e
d enc
odi
ng
,
an
d it listed
sev
e
ral
po
in
ts
th
at v
i
d
e
o
p
a
rtitio
n
i
n
g
sch
e
mes sh
ou
ld co
n
s
i
d
er fo
r efficien
t en
cod
i
n
g
.
We
classified
th
e
vid
e
o
p
a
rtitio
n
i
n
g
sch
e
m
e
s in
to
(a) a
v
i
d
e
o
partitio
n
i
ng
sche
m
e
b
a
sed on
th
e nu
m
b
er
o
f
n
o
d
e
s
an
d
(b) a p
a
rtitio
n
i
ng
sch
e
m
e
b
a
sed
o
n
GOP.
W
e
also
po
in
ted
o
u
t
th
at th
ese two
typ
e
s o
f
v
i
d
e
o
p
a
rtit
io
n
i
ng
sch
e
m
e
s are prob
lem
a
t
i
c in
th
at th
ey
d
o
no
t accoun
t for th
e im
ag
e itself du
ri
n
g
th
e
p
a
rtitio
n
i
n
g
pro
cess.
There
f
ore,
we measured the c
h
ange
in enc
o
ding efficiency
of the pa
r
titioning schem
e
s
according to c
onte
n
t
type. The m
e
a
s
urem
ent resul
t
s indicated change
s in e
ffici
ency according to conten
t type, and we
de
scribe
d
th
e po
ten
tial prob
lem
s
th
at c
a
n
o
c
cur wh
en th
e con
t
en
t
typ
e
s are no
t con
s
id
ered. As
fo
r
th
e fu
tur
e
resear
ch,
we are c
u
rrent
l
y working on the partitioni
ng sc
hem
e
s th
at take into account the
im
a
g
e's content type and
devel
opi
ng a c
ont
e
n
t
t
y
pe di
f
f
ere
n
t
i
a
t
i
ng al
go
ri
t
h
m
t
h
at
u
s
es col
o
r di
f
f
er
ence com
pone
nt
s of
vi
de
o an
d t
h
e
n
u
s
es
v
a
riab
le
partitio
n
i
ng
sche
m
e
s b
a
sed on
th
is.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E V
o
l
.
5, No
. 3,
J
u
ne 2
0
1
5
:
56
9 – 5
7
8
57
8
ACKNOWLE
DGE
M
ENTS
T
h
i
s
w
o
r
k
w
a
s
s
u
p
p
o
r
t
e
d
b
y
K
y
o
n
g
g
i
U
n
i
v
e
r
s
i
t
y
‘
s
G
r
a
d
u
a
t
e
R
e
s
e
a
r
c
h
A
s
s
i
s
t
a
n
t
s
h
i
p
2014.
REFERE
NC
ES
[1]
Sullivan,
Gar
y
J,
et a
l
.
, “Overview of the High Ef
ficiency
Vid
e
o
Coding (HEVC) Standard
”,
IEEE Transactions
o
n
Circuits and
Sys
t
ems for Video
Technolog
y
, vol.
22, no
. 12
, pp
. 1
649-1668, Dec 2
012.
[2]
Pereira, R
a
fael,
et al
l.
, “An Architecture for Dis
t
ributed
High Performance Vid
e
o Processing in
the Cloud
”,
Pro
c
.
of 2010 I
E
EE 3r
d International
Conference on
C
l
oud Computing
, 2010, pp. 482-4
89
[3]
http://hadoop
.ap
ache.org/
[4]
Lao, Feng
,
et al
.
, “Parallelizing
Video Transcod
ing Us
ing Map-
Reduce-B
a
sed C
l
oud Computing”,
Proc. of 2012
IEEE Internatio
nal Symposiu
m
onCircuits and
S
y
stems,
pp
. 2905
-2908, May
201
2.
[5]
http://hevc.hhi.fr
a
unhofer.d
e/
[6]
M
y
unghoon Jeo
n
,
et
al
., “Tow
ard video
par
titioning schemes
in di
stribu
ted
video processin
g
for HEVC”,
International Co
nference on
Internet (
I
CONI)
,
Dec 2014
.
BIOGRAP
HI
ES OF
AUTH
ORS
M
y
unghoon Jeon receiv
e
d his B.S and M.S. degr
ees in Co
m
puter Science
from
Ky
onggi
University
, Korea in 2004 and 20
11 respectively
.
He
is a Ph.D. stu
d
ent working at
department of
K
y
onggi Univer
sity
B
y
oung-Dai
Lee is an assistant
professor at the
department of K
y
onggi Un
iv
ersity
, Korea. He
received his B.S. and M.S. degr
ees in Computer
Science from Yonsei University, Korea in 1996
and 1998 respectively
.
H
e
receiv
ed his Ph.D. de
gree in Computer Science and En
gineer
ing from
University
of Minnesota,
twin c
ities, U.S.A.
in 20
03. Before
jo
ining the K
y
onggi
University
, h
e
worked at Samsung Electron
i
cs, Co
., Ltd
as
a s
e
nior
engineer f
r
om 2003 to 20
10. During
the
period, h
e
has
participated in
ma
ny
c
o
mme
rc
ia
l
i
z
a
t
i
on projects related to mobile bro
a
dcast
s
y
ste
m
s.
His rese
a
r
c
h
inte
re
sts inc
l
ude
se
nsor
networks, open mobile
platfo
rm, and mobile
m
u
ltim
edia bro
a
dcast.
Evaluation Warning : The document was created with Spire.PDF for Python.