TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.6, Jun
e
201
4, pp. 4345 ~ 4
3
5
2
DOI: 10.115
9
1
/telkomni
ka.
v
12i6.458
2
4345
Re
cei
v
ed Se
ptem
ber 30, 2013; Revi
se
d De
ce
m
ber
16, 2013; Accepted Janu
ary 14, 201
4
ARFMS: An AR-based WYSIWYG Filmmaking System
Chen Ling*,
Wenjun Zh
a
ng
Schoo
l of F
ilm and T
V
Arts &
T
e
chnolog
y, S
han
gh
ai Un
iver
sit
y
149 Ya
nch
ang,
Shang
ha
i 200
072, Ch
in
a, Ph: 56332
03
0
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: lcrex@shu.
e
du.cn
A
b
st
r
a
ct
This paper pr
oposes
a
nov
el “What You S
ee Is
What You Get
”
fil
mmaking system
, c
a
lled as
ARF
M
S. It provides
users w
i
th a l
o
w
cost an
d easy to
use s
ystem to
make
mov
i
es. Us
ers
cann
ot on
ly re
al-
time v
i
sua
l
i
z
e
the perfor
m
an
ce w
i
th comp
uter-ge
nerat
ed
objects, but
also vis
u
a
l
effects films w
ill
b
e
finish
ed w
i
tho
u
t
post-pro
ducti
on. After finis
h
ed th
e scre
enp
lay, the
perfor
m
a
n
ce w
i
l
l
b
e
shot o
n
-site
us
in
g
an
ordi
nary c
a
mer
a
. CG o
b
j
e
cts w
ill be
si
mply a
n
d
effect
iv
ely co
ntrol
l
ed
b
y
natur
al
intera
ction. The
n
us
ers
can
app
ly th
e
DR-
marker
jus
t
like
an
actu
al
scen
e
e
l
e
m
e
n
t, becaus
e it t
a
kes the
adv
ant
ages
of
mark
e
r
-
base
d
an
d
ma
rkerless-b
a
se
d
registratio
n
a
ppro
a
c
hes. T
h
i
s
paper
i
m
pl
e
m
e
n
ts ARF
M
S to achiev
e pr
e-
visual
i
z
a
t
i
on a
nd rea
l
-time fin
i
sh the fil
m
usi
ng au
g
m
ent
ed
reality techn
o
l
ogy. T
he user
can perfor
m
w
i
th
the CG charact
e
r and stag
e pr
operty just lik
e real o
nes.
Ke
y
w
ords
: fil
m
mak
i
ng, vis
u
al effects, pre-visual
i
z
a
t
i
on, a
u
g
m
e
n
ted re
ali
t
y, dimin
i
sh
ing
reality
Copy
right
©
2014 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
Being the hi
ghe
st pea
k
of entertain
m
ent,
the film indust
r
y ha
s a hu
ge m
a
rket and
potential. Te
chn
o
logie
s
which in
crea
se
efficienc
y in
filmmaking
are re
qui
red.
In the mid 9
0's,
cre
a
tion
of fil
m
an
d TV
wa
s a
dopte
d
th
e Virtual
Real
ity (VR) tech
nology. 3
D
a
n
imation i
s
ai
ded
to sho
o
t in the pre
-
p
r
od
uct
i
on sta
ge. Th
is is
pr
e
-
visu
alizatio
n (Pre
Viz). PreVi
z
i
s
a
wide
sp
re
ad
pre
-
produ
ctio
n techniq
ue u
s
ed in vario
u
s
film produ
ct
ions to plan camera wo
rk b
e
fore the a
c
tu
al
sho
o
t. It allows a prepa
ra
tion of often
very comp
li
cated shot
s that have to be clear not onl
y to
the dire
ctor b
u
t also to the
directo
r
of ci
nemat
og
ra
ph
y, the set design
e
r, the lig
hting crew, a
nd
other
memb
e
r
s
of the film
team in
cludi
ng the
sp
eci
a
l effect
s an
d visual
effe
cts
units. Sin
c
e
2000, PreViz has be
com
e
an i
n
crea
si
ngly impo
rt
a
n
t techn
o
log
y
in the Holl
ywood. M
o
st
the
aca
demi
c
co
mmunity focu
se
s the PreV
iz usin
g the VR or 3D ga
me techn
o
log
y
[1]. Howev
e
r,
filmmakin
g
ca
n’t be the onl
y indoor
shot.
Unlike gam
e
visualization
s
that most of
ten con
c
e
n
tra
t
e
on
the re
prese
n
tation
of
the
virtual 3D wo
rl
d al
one. A n
o
ve
l PreViz, M
R
-PreViz,
using
augme
n
ted reality (AR) [2
]
[3] has been
prop
osed
[2
]
. In c
ontras
t
to VR aiming
for c
o
mpletely
repla
c
in
g the
natu
r
al
re
ce
ption of
ou
r
environ
ment
by the
provi
s
i
on of
artifici
al
input
cue
s
,
AR
has
always a
llowe
d the u
s
er to o
p
e
r
ate
from a
co
m
m
on a
nd
well
kno
w
gro
u
n
d
, only sparsely
modifying the
environm
ent
by additional
artificial vi
rtu
a
l conte
n
t [5]. Adding artifi
cial
content o
r
even supe
rim
posi
ng the
re
al co
ntent co
mpletely
is
something
whi
c
h h
a
s
bee
n
done i
n
the a
r
ea
of visualpe
rception for d
e
cade
s.
Therefore, a
c
tors ca
n perfo
rm
wit
h
com
puter-g
enerated obj
e
c
ts
in the
re
al e
n
vironm
ent a
nd the
outd
o
o
r. Sony, Stu
d
io O
u
tput
a
nd M
a
rshm
al
low
La
ser Fe
ast
tried to
pro
d
u
c
e th
ree
short movies
“G
re
at f
ilms fill ro
oms” (http://www.g
r
eatfilm
sfillroom
s.com
)
.
These movie
s
are all in
re
al-time an
d n
o
post-produ
ction.
In this pa
per,
we
ta
ke th
e
advantag
es o
f
the
M
R
-P
re
Viz a
nd
“G
re
at films th
e fil
l
ro
oms
”
to present a
n
e
w type
film
a
nd
TV
shoot system. We
call
it
a
s
A
R
Filmmaki
ng
Sy
stem (ARFM
S
).
It is combin
e
d
with co
mp
uter vision, h
u
man
-
co
mput
er interactio
n
,
VR and AR technol
ogy to
achi
eve WYS
I
WYG (What
You
See Is What You Get)
visual effect
s filmmakin
g
.
An overvie
w
of the ARFM
S is int
r
odu
ced in
S
e
ctio
n
2. In Se
ctio
n 3,
we d
e
scribe
ou
r
ARFMS and
pre
s
ent several results. In t
he end, we di
scuss po
ssibl
e
enha
nceme
n
ts.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4345 – 4
352
4346
2. ARFMS
On-Site P
r
eV
iz u
s
e
s
the
di
gital data
of the
ca
mer
a
-
w
o
r
ks
s
y
nc
h
r
on
iz
ed
re
co
r
ded
in
th
e
actual
shooti
ng to achiev
e regi
stration
. And
the green screen i
s
repl
aced with the final
VFX
conte
n
t. The
n
the di
re
cto
r
can
se
e th
e final
sy
nth
e
tic ima
ge to
determine
whether to retake.
Ho
wever, tra
cki
ng sen
s
ors are complex
and expe
nsiv
e. Now, filmm
a
kin
g
also take
s pla
c
e in t
he
actual
enviro
n
ment an
d is comm
only fo
r inde
pen
den
t filmmaking.
Since the i
n
trodu
ction of
DV
techn
o
logy, the me
an
s o
f
prod
uctio
n
ha
s be
co
m
e
mo
re d
e
m
ocratized.
Filmmakers
can
con
c
eiva
bly shoot and edit
a film, create
and edit
the sou
nd and m
u
si
c, and mix the final cut
on
a home com
puter. With In
ternet movie
distrib
u
tion, indep
ende
nt filmmakers ca
n excha
nge
with
each
othe
r. More and m
o
re amateu
rs
pro
d
u
c
e
thei
r o
w
n movie
s
. Ho
weve
r, i
t
is ha
rd fo
r
the
amateu
r to
m
a
ke
a VFX
m
o
vie with
out
high-end
dev
i
c
e
s
a
nd
profession
al
skill
s. Th
us, th
e
goal
of ARFMS i
s
to offer a m
e
thod to
ma
ke a VFX
film, whi
c
h i
s
ea
sy to u
s
e
for users
witho
u
t
profe
ssi
onal
skill
s. Wo
rkflow of filmma
king u
s
in
g
ARFMS is
as f
o
llows (Fi
g
u
r
e 1): (1
) p
r
ep
aring
the contin
uity, (2) pre
pari
n
g the virtual obje
c
ts, (3
) actual
shooti
ng and (4)
modificatio
n
and
compositing.
We
pro
p
o
s
e
d
a n
e
w filmmakin
g
syste
m
ARFMS.
Ho
wever, it i
s
a
hug
e p
r
oject to
impleme
n
t. In the expe
rim
ent, we i
m
ple
m
ent a
si
mpl
e
ARFMS
prototype. We
employ a
lap
t
op
comp
uter wit
h
two
2.13
G
H
z Intel
Core
i3 p
r
o
c
e
s
sors a
nd
2.5GB
RAM, an
o
r
di
nary in
expen
sive
came
ra, a
Kinect a
nd a
speci
a
l ma
rke
r
. The frame
w
ork
of ARF
M
S is
sho
w
n
inFigu
re2. A
fter
finishe
d
the
scre
enpl
ay an
d continuity, the p
e
rf
o
r
man
c
e will be sh
ot
on-site.
In fact,
we
u
s
e a
n
ordin
a
ry
cam
e
ra i
n
ste
ad o
f
the profe
s
si
onale
qui
pm
e
n
t, which i
s
l
o
w
co
st (ju
s
t
50
RMB in t
h
e
simulatio
n
)
a
nd ea
sy for
amateu
r to g
e
t. One of ARFMS go
als
is to satisfy amateu
rs. Th
en
virtual obj
ect
s
can
be
add
e
d
by DR-marker.
CG
obj
e
c
ts
ca
n be
bu
ilt in the p
r
e-p
r
odu
ction
sta
ge.
And they
can
be
simply
an
d effectively
controlle
d
by
Kinect. Th
en
ceforwa
r
d
the
VFX movie
will
be fini
she
d
re
al time. It wo
rks thro
ugh
out
the typical fil
mmaki
ng th
re
e sta
g
e
s
p
r
o
c
ess, which a
r
e
pre
-
produ
ctio
n, prod
uctio
n
and po
st-pro
ductio
n
sta
g
e
s
. After the directo
r
finishe
d
previo
us
work,
the film can
be shot by ARFMS.
In the remaining section, we
will introduce two import
ant
comp
one
nts
of ARFMS be
low.
Figure 1. Gen
e
ral ARF
M
S Wo
rkflo
w
Figure 2.The
Frame
w
o
r
k of ARFMS
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
ARFMS: An AR-b
ased WYSIWYG Filmm
a
king System
(Chen Li
ng)
4347
2.1. CG Ch
ar
acte
r
Con
t
ro
l
A
range of tools ori
g
inall
y
designe
d for high
-en
d
grap
hics are
used for PreV
iz: from
Maya and Mo
tionbuilde
r
, to
3D Studio M
a
x, Softimage XSI, and Pose
r
.
Each of these p
r
og
ra
ms
has its own strength
s
and wea
k
n
e
sse
s
but all of
them allow for some kind of staging of events
on virtual sets as
well a
s
a
definition of virtual
cam
e
ra angle
s
to
ward the
re
sulti
ng scen
es.
T
hey
are
widely
u
s
ed i
n
tra
d
itional PreV
iz
nowaday
s. Howeve
r
,
tra
d
i
t
ional techno
logy to han
d
l
e
motion data
use
s
motion
captu
r
ed d
a
ta whi
c
h are
attached to th
e CG charact
e
r
.
Then the
use
r
can natu
r
ally
control the
virtual characte
r
.
Howe
ver
,
the motion captu
r
ed
equipment is
expen
sive, inconve
n
ient a
nd un
comfo
r
table.
The
state of
the art i
s
u
s
ed several cameras i
n
th
e studi
o [6] to get the
mo
del an
d
action. But it need
s a p
r
of
ession
al studi
o. It does not
suit for am
ateur u
s
e
r
s.
W
e
take Kine
ct
as
a chea
p and easy motion
captu
r
ed devi
c
e. Kinect is attache
d
to the MikuMi
ku
Dan
c
e (MM
D
) to
control the CG cha
r
a
c
ter a
s
sho
w
n inFi
g
u
re3.
The to
p
right corner fi
gure i
s
the user
’
s
real actio
n
got by a Kinect, and it controls
CG boy to do the sa
m
e
action. In this way
,
we
ca
n easily gain the
roug
h a
c
tion.
It will be ap
proximately acceptable fo
r a
n
animatio
n
.
Then the
use
r
ca
n control the
cha
r
a
c
ter m
a
nually to get the better a
c
ti
on.
Then
the
action
will be
saved a
s
a m
o
tion file “vmd”.
The file will b
e
con
n
e
c
ted
with the dif
f
erent
cha
r
a
c
ter model
in
the actual sho
o
ting
on-site.
After finished
the virtual c
hara
c
te
r and moti
on, the user can use
them in the
ARFMS.
Firstly
,
the CG cha
r
a
c
ter is adde
d on the actual
on
-si
t
e image, its
size positio
n and dire
ction
will
be manu
ally adju
s
ted.
The
n
, the actor
can
play with the CG objec
t
s
in realtime.
Figure 3. Kinec
t is
Attac
hed to the MMD to
Control
Figure 4. The
Disa
ppe
are
d
Marker i
s
the
Action of the CG Characte
r unde
r the Virtual
C
h
ar
ac
te
r
2.2. DR-mark
e
r
There are two types of
AR regi
stratio
n
nowa
days:
AR regist
rati
on based on
marke
r
method [7] a
nd
AR regi
stration ba
sed o
n
marke
r
le
ss
method[8]. Due to highly visual feeli
n
g
s
, a
spe
c
ial ma
rker will be un
accepta
b
le. Hen
c
e t
he m
a
rkerl
e
ss wa
y will be popular
. Ho
weve
r
,
it
alway
s
o
c
cu
rs to
CG
o
b
ject
s’
flutter usin
g the
markerl
e
ss a
ppro
a
ch. Flu
tter mean
s t
h
e
comp
uter-ge
nerate
d
mod
e
ls will
shift their po
si
tion,
due to cal
c
ulation erro
r
.
Then
ce we use
marker meth
od.
Then we
use diminishing realit
y (DR) techn
o
lo
gy for removing the speci
a
l
marker fro
m
a live video stream of the use
r
’
s
re
al environm
ent. W
e
call this
approa
ch as
DR-
marker[9].O
u
r
appro
a
ch is based on a simple set
up a
nd neither re
quire
s any pre-p
r
o
c
e
ssi
ng nor
any informati
on on the st
ru
cture a
nd lo
cation of the object
s
to be removed.
While
A
R
ha
s alway
s
bee
n
rest
ricte
d
to addin
g
artifici
al co
ntent to the re
al enviro
n
ment,
DR allo
ws for removing re
a
l
world conte
n
t. Existing a
ppro
a
che
s
re
quire
compl
e
x setups an
d are
not applicabl
e in real-tim
e to seamle
ssly delete
v
i
sual content
from the observe
r
’
s vie
w
in
uncon
strain
e
d
enviro
n
me
nts. Ou
r DR-marke
r
a
pproach do
es
n
o
t have any
distan
ce
or 3D
environment struct
ure information and
uses a si
ngl
e ca
mera only
.
The illustration is shown
inFigu
re4 tha
t
a CG cha
r
a
c
ter i
s
su
peri
m
posed into
a real
sce
ne.
The marke
r
is und
er the
CG
model. Ho
we
ver
,
it is not been seen.
2.2.1. Marker
Dete
ction a
nd Projectin
g Model
DR-ma
r
ker b
egin
s
with
th
e ma
rker
det
ection. It
can
be fa
st an
d
stably d
e
tect
ed in
re
al
time usin
g th
e app
roa
c
h
ARTool
kit (ht
t
p://www.hitl
.wa
s
hin
g
ton.e
du/artool
kit/).
This i
s
the
most
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4345 – 4
352
4348
famous meth
ods in A
R
. After the
qua
drate ma
rker
d
e
tected,
we
can o
b
tain the
tran
sform
a
tion
matrix. Mean
while, the all
-
roun
d ma
rker coo
r
din
a
te
can be b
u
ilt. The coordinate
s
of the a
r
tificial
objects will transl
a
te into
the marker coordinate by
a transformation matrix.
In fact, the
m
a
rker ha
s
a l
o
t of p
r
ior
kn
owle
dge.
On
e of the
imp
o
rtant info
rmat
ion i
s
the
size an
d the l
o
catio
n
of th
e sq
ua
re ma
rker. A
c
co
rdin
g
to
this kn
o
w
led
ge,
the region of
interest
(ROI)
can
be
picke
d
u
p
. F
i
rstly, we
can
get th
e a
c
tu
al sce
ne i
m
a
ge
without
m
a
rker. S
e
con
d
ly,
we
put the
m
a
rker into th
e
image.
The
ma
rker will
b
e
dete
c
ted
an
d the ve
rtexe
s
V
i
(
i
=1, 2, 3
,
4)
are g
a
ined,
clea
rly sh
own in Figu
re5.
Here,
we d
e
cid
e
the ROI by the points a
r
ou
nd
the
vertexes. The
points
P
i
can
be defined u
s
ing the follo
wing formula:
00
/(
)
ii
i
i
PS
i
z
e
P
S
i
z
e
V
V
V
V
(1)
4
0
1
1/
4
i
i
VV
(2)
Whe
r
e Si
ze
(
V
i
) den
otes th
at the si
de l
e
ngth of the
m
a
rker. Si
ze
(
P
i
) de
note
s
tha
t
the sid
e
le
n
g
th
of the
ROI, which
is the
ou
ter recta
ngle
i
n
the
Figu
re
5.
V
0
pre
s
e
n
ts the
centroid
of the ve
rtex
es.
The
ROI i
s
bi
gger than
the
marke
r
. Th
e
rea
s
o
n
i
s
th
at the a
r
ea
a
r
oun
d the
ma
rke
r
ha
s a
pri
o
ri
kno
w
le
dge which
will be use
d
in the followin
g
se
ction. Becau
s
e the marker is squ
a
re, the
proje
c
ting
mo
del is th
e rectangula
r
mo
d
e
l. Once
the
ROI is
dete
r
minate, the i
m
age of th
e
ROI
will be
save
d
as the
texture
of the p
r
oje
c
ting mo
del
. As illustrated in
Figure 6, a
q
uadrate ma
rker
is on a pi
cture in Figure 6(c). The
n
ceforwardthe
ROI is co
mpute
d
and the texture in Figu
re
6(b
)
will su
pe
rimp
ose th
e marker. In this
wa
y, the
speci
a
l
marker ha
s been remove
d
from
the scene,
as sho
w
n inF
i
gure 6
(
a
)
.When the AR
marker i
s
det
ected first time, the ROI is picke
d
up.
The
affine tran
sfo
r
mation
matri
x
can
be
det
ermin
ed. Th
e
n
the
quad
ra
te ROI
will
b
e
tran
sfo
r
me
d,
whi
c
h ta
ke
s
place di
storti
on. The
imag
e of ROI will
be a
qua
drat
e mod
e
l. Thi
s
is the
proje
c
ting
model. Wh
en
the live vide
o strea
m
com
e
s, the tran
sformatio
n
matrix will be cal
c
ulate
d
fast a
fter
marker dete
c
tion.
And
the proje
c
ting mo
del
will
cha
n
g
e
it, just like
a com
pute
r-g
enerated m
o
del
in the AR app
lication.
Figure 5. The
spe
c
ial ma
rker of Figu
re
6
Figure 6. Proj
ecting mo
del
of DR-marke
r
(a)
marker
DR-m
arker h
a
s b
e
e
n
remove
d from
the scene; (b
) original im
ag
e; (c) the texture of
proje
c
ting m
o
del
2.2.2. Projecting Model Adjustmen
t
Due to lightin
g cha
nge, the
texture of th
e proje
c
ting
model will be
not con
s
iste
nt. So we
need to adj
ust the proje
c
tin
g
model in ti
me. Here we
adju
s
t the HS
V (Hue Satu
ration Intensit
y)
of the texture
to match th
e
environ
ment.
The
RO
I
is arou
nd
the m
a
rker and
it
i
s
big
ger. Wh
en
the came
ra
capture
s
the i
m
age
s, part
of ROI c
an b
e
see
n
. That mean
s the area is overl
a
p
ped
betwe
en th
e
real
scen
e
an
d the
proje
c
ting m
odel. So
we
adj
ust
th
e texture
HS
V via compa
r
ing
with the sam
e
points’
HSV in the texture and
the real i
m
age. It is de
termine
d
as f
o
llows:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
ARFMS: An AR-b
ased WYSIWYG Filmm
a
king System
(Chen Li
ng)
4349
4
1
()
/
(
)
HS
V
i
i
i
HS
V
P
HS
V
V
(3)
Whe
r
e
HSV
(
V
i
) is th
e
HSV of the vert
exes
of the
marker.
HSV
(
P
i
) is
HSV of
the ROI vert
exes.
λ
HSV
is the
co
efficient that
the p
r
oje
c
ting
model
will
ch
ange.
λ
HSV
ra
n
ges from
0
to
λ
max
.
λ
ma
x
i
s
t
he
s
u
perior limit
. In our experiment, it t
a
k
e
s
two. Onc
e
λ
HSV
is d
e
termin
ed, the ne
w texture
HSV
new
(
tex
j
) will be obtained by:
()
[
1
(
1
)
]
()
new
j
H
S
V
H
S
V
j
HS
V
t
ex
HS
V
t
ex
(4)
Her
e
HSV
(
tex
j
)
den
otes t
hat HSV of t
he pixel i
n
th
e proje
c
ting
model
and
HSV
ne
w
(
tex
j
) is the
new one.
α
HSV
is th
e weight
for different
function
s of
h
ue, saturatio
n
and
inten
s
it
y. In the actu
al
situation, the
intensity ch
a
nge
s the mo
st, so the
α
V
wi
ll be bigg
er t
hen
α
H
an
d
α
S
In this pap
er,
α
H
is 0.15,
α
S
is 0.15, and
α
V
is 1.In fact, light will not cha
nge at a
ll times, so the
projectin
g
mod
e
l is
not adj
uste
d
every fra
m
e.
In this pa
per,
we
ma
ke
th
e texture
adj
usted
auto
m
atically eve
r
y
30
frames. Note that the automatic
exposure of the camera shoul
d be turned off. Otherwi
se the
image
s will chang
e at any time. The
brig
htness will b
e
out of control
.
A DR-ma
r
ke
r simulatio
n
re
sult is exhi
bited in
Figu
re
7, which is
a real
-time cont
inuou
s
seq
uen
ce. According to th
e top left co
rner of
eve
r
y image, the fra
m
es p
e
r
se
cond a
r
e ave
r
age
30fps. T
h
e
r
e
is a
sp
eci
a
l
marker on
th
e groun
d in
F
i
gure
7(a). It i
s
a
ra
w
ca
ptured
fram
e.
And
the other thre
e figure
s
are other three p
o
st-p
ro
ce
ssin
g frame
s
, usi
ng DR-m
arke
r. In Figure 7
(
b)-
(d), an a
rro
w indicate
s the area
whi
c
h
the speci
a
l
marker exi
s
ts in the origin
al frame. If you
watch not
carefully, you will not
find the difference. It's worth no
ting that the projecting m
o
del
adjustment can fit the illumination envi
r
onm
ent whi
c
h is
relatively
stable. However, it is hard to
be suita
b
le in
the actual co
mplicate
d
out
door e
n
viro
n
m
ent.
Figure 7. Proj
ecting Mo
del
Adjustme
nt Simulati
on Results. (a) o
r
igin
al frame, (b
) the 31
st
frame
(the arro
w is
pointing to th
e r
egio
n
of DR-m
arke
r), (c) the 241
st
fra
m
e, and (d
) the 421
st
fram
e
2.2.3. Tracki
ng
Whe
n
the live video stre
a
m
come
s, the AR
marker will be detected. It processe
s fast.
However,
for
the influence
of t
he li
ght, the transform
a
tion m
a
trix
will happen t
o
change
a li
ttle.
The vis
ual feeling is
flutter. So we try to
wi
pe o
u
t it by
Euclide
an di
stance fo
rmula
as follows:
2
n
o
t
c
h
an
ge
T
ch
an
ge
T
ot
h
(1
,
2
,
3
,
4
)
ers
l
a
st
current
ii
VV
r
i
(5)
Her
e
T
is
the trans
formation matrix.
V
i
last
presents th
e vertexes of
the
marker i
n
the last fra
m
e
.
V
i
current
denotes th
e ve
rtexes
of the
ma
rker in
the
cu
rrent frame.
r
i
s
the
threshol
d of the
allo
wed
moving rang
e
,
which is kno
w
n in
Figu
re5.
Equation
(5
)
tells that if all
four ve
rtexe
s
move
a little
,
the transformation matri
x
will hold on. The expe
rimental result
show
s that the computer-
gene
rated m
odel
s do not flutter after this method p
r
o
c
e
s
ses.
In the film and TV sho
o
t, image
s will b
e
shot
steady
. That mean
s the pictu
r
e
s
will not
rapidly
ch
ang
e. And 80%
shots a
r
e
stati
c
. The
r
efo
r
e,
we utili
ze
a h
o
ld-o
n mo
de
to achi
eve thi
s
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4345 – 4
352
4350
goal. Figu
re
8 sho
w
s the
result of the hold-on
mo
d
e
. Whe
n
the
came
ra i
s
re
ady and
will not
move. The transfo
rmatio
n
matrix is
sav
ed. Then t
he
marker
co
ord
i
nates
will n
o
t cha
nge. In t
h
is
way, the m
a
rker can
be
re
moved
from t
he
scene
and
the
CG
obje
c
ts
ca
n al
so
be regi
stered
in
the scen
e. Of cou
r
se, the
marker
ca
n b
e
occlu
ded
cl
early sho
w
n i
n
Figu
re8. Althoug
h the ma
rke
r
is occlu
ded b
y
the hand, the CG obj
ect i
s
still able to registe
r
.
2.2.4. Virtual-to
-real O
ccl
usion
DR-ma
r
ker h
a
s
bee
n impl
emented
a
s
stated
above
.
In an
ordin
a
ry AR regi
stration,
grap
hical obj
ects a
r
e a
d
d
ed on the m
a
rker. On
th
e
other ha
nd, it means
real
object
s
will
be
occlud
ed by comp
uter-ge
nerate
d
one
s. Real obj
e
c
ts always app
ear o
ccl
ude
d
by virtual ones,
rega
rdl
e
ss of their actu
al spatial relatio
n
s
hip.
The
s
e
simple app
roa
c
he
s cann
ot handl
e occlu
s
ion
betwe
en diffe
rent type
s of
obje
c
ts in
the
scen
e.
Ho
we
ver, DR-m
arker i
s
al
ways the ba
ckg
r
ou
n
d
of the scene i
n
the pra
c
tica
l situation. If the projecti
n
g
model i
s
in front of the rea
l
object, it do
es
not meet the actual expe
rien
ce. It seems not
re
al
. So DR-ma
rker
con
s
id
ers the occlu
s
i
on
betwe
en proj
ecting mo
del
and the re
al scen
e.
Here, we
con
s
ide
r
that
det
ection
only in
front
of
previously
define
d
textured
pl
ane
s in
the re
al sce
n
e
. The m
e
th
od is in
spire
d
by F. Jan
et.al [10]. First, a refere
n
c
e b
a
ckg
r
o
u
nd is
selected
and saved. Th
en, once
a frame comes,
the reference
background will
compare with
the curre
n
t ca
mera ima
ge. We ad
opt “ad
aptive
HSV” criterio
n to ach
i
eve pixel co
mpari
s
o
n
:
(
,
)
(
,)
(
,
)
(
1
(
,)
)
(
,)
o
x
y
x
y
H
xy
xy
V
x
y
(
6
)
Whe
r
e
∆
H
(
x,
y
)is the
hue
di
fference
of pi
xel (x, y) i
n
t
he
referen
c
e
backg
rou
nd
and th
e
cu
rre
nt
came
ra im
ag
e.
∆
V
(
x,
y
) is t
he inten
s
ity difference of two ima
g
e
s
,
α
(
x,
y
) i
s
a wei
ght. Formul
a
(6)
is an
ada
ptive HSV me
thod. Ordina
ry ba
ckgr
o
u
nd subtracti
on only thin
ks th
e inten
s
ity
difference. T
hat means they simp
ly utilize the influence factor of
brightness. If the color of
the
foreg
r
ou
nd i
s
clo
s
e to
ba
ckgroun
d, it can’t be
re
cog
n
ize
d
. The
r
ef
ore, a
daptive
HSV criteri
o
n is
employed. It con
s
id
ers the
hue effect. The wei
ght fun
c
tion
s is dete
r
mine
d as foll
ows:
(
,
)
(
,)
/
(
(
,
)
(
,)
)
m
i
n
(
(
,
)
,
(
,)
)
re
f
c
urr
ent
x
y
H
xy
H
x
y
V
xy
V
x
y
V
xy
(7)
Her
e
,
V
ref
(
x,
y
) is the inten
s
ity of pixel (
x
, y) in
the referen
c
e ima
ge.
V
current
(
x,
y
) is the intensity
ofthe co
rre
sp
ondin
g
pixel in the cu
rre
nt
image. min(•
)
is minimi
ze
d value.
β
is a weight. In this
function, we kno
w
that
if min(
V
ref
(
x,
y
),
V
current
(
x,
y
)) i
s
bigge
r, the
α
(
x,
y
)
will be
bigger. It means
whe
n
the ima
ge is too bri
g
ht, the image will loo
k
pale.
The colo
r is
hard to be
re
cog
n
ized. So the
intensity is th
e prin
cip
a
l fa
ctor.
When th
e
∆
H
(
x,
y
) i
s
b
i
gger, the
α
(
x,y
) de
pen
ds
o
n
the color. In
a
s
i
milar
re
as
on
, w
h
en
∆
V
(
x,
y
) i
s
la
rge,
α
(
x,
y
) will
be
small. So we can utilize light
ness to
com
p
are.
For hole
s
in t
he
com
pari
s
o
n
ima
ge, it
sh
ould
be
sm
oo
thed. Th
en
we con
s
ide
r
whether the
pix
e
l
sho
u
ld be o
c
clud
ed or n
o
t:
1
(
,)
(
,
)
(
1
(
,)
)
(,
)
oth
r
s
0e
HV
ox
y
x
y
t
x
y
t
occlu
s
io
n
x
y
(8)
Fi
g
ure
9
.
I
llustration of Occlusio
n
Fi
g
ure 8. Hol
d
-on M
ode
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
ARFMS: An AR-b
ased WYSIWYG Filmm
a
king System
(Chen Li
ng)
4351
Whe
r
e
t
H
an
d
t
V
are the
respe
c
tive threshold
of hu
e and inte
nsi
t
y. After the formula
(8),
we
deci
de which
pixel need
s
not to be dra
w
n. Ultimatel
y
the proje
c
ting mod
e
l is
modified a
n
d
the
image will b
e
rend
ered. Th
e result is sh
own in
Fig
u
re
9. The DR-m
a
rker i
s
on the boo
k and h
a
rd
to be
re
cog
n
ize
d
by th
e
na
ked
eye.
The
finge
rs occlu
de th
e
DR-m
ar
ke
r.
A
f
t
e
r o
ccl
us
ion
algorith
m
, it looks like valid due to the a
c
tual spatial relation
ship.
3. Results a
nd Discu
ssi
on
In this p
ape
r, we e
m
ploy
Micro
s
oft Vis
ual Stu
d
io
2008 to
impl
ement the
software.
ARTool
kit, O
pen
CV a
nd
Open
GL
are
combi
ned
to
impleme
n
t th
e p
r
op
osed
method. A
sq
uare
marker,
like
i
n
Figu
re
6(b),
is
dete
c
ted fi
rst. Th
en
a p
r
ojectin
g
mo
d
e
l is built,
whi
c
h i
s
th
e texture
in Figure
6(c). Finally the output image will
not
see the A
R
marker in
Figure
6(a). Next the
artificial CG
obje
c
ts can b
e
sup
e
rim
p
o
s
ed on the p
r
ojectin
g
mod
e
l. The re
sult
can be
se
en
in
Figure 4 a
nd
Figure 7. Th
e
com
put
er-ge
nerate
d
cha
r
acter is j
u
st
t
he real
scene
without
seei
ng
spe
c
ial
markers.
The
virtu
a
l-to-real
o
ccl
usio
n i
s
sho
w
n in Fi
gure9.
Finge
rs are i
n
front
of the
AR
marker. After our
occlu
s
io
n algo
rithm, t
he real
o
b
je
cts will
not ap
pear o
ccl
ude
d by the virtu
a
l
obje
c
ts
as ot
her A
R
appli
c
ation
s
. In t
h
is
ca
se, th
e
proj
ectin
g
model
ha
s t
he a
c
tual
sp
atia
l
relation
shi
p
. Here, due to
a priori
kno
w
led
ge of
the marker, we
can ju
st con
s
ide
r
the are
a
of
marker, omitti
ng the extra p
i
xels
to s
i
mplify c
a
lc
ulation.
In the experi
m
ent, we ma
ke
some
sh
ort VFX f
ilms, and ea
ch o
ne i
s
abo
ut 2 min
u
tes. In
Table
1, m
a
rker d
e
tectio
n
and
tra
c
king
proces
s i
s
i
n
real tim
e
,
whi
c
h
co
sts
about
32.7m
s p
e
r
frame. After the proj
ectin
g
model first time, t
he total
pro
c
e
ss time
will be 39.7m
s. That mean
s
the fram
e p
e
r se
co
nd
of DR-m
arke
r i
s
25fps.
Ho
wev
e
r,
if we co
nsider
th
e
o
c
clu
s
ion
alg
o
rith
m,
due to the da
ta read time, it will have 211ms computi
ng time. So the frame pe
r se
con
d
will drop
to 4fps. In the future work,
we will o
perate GPU to reduce the
computing time.
DR-ma
r
ker ta
ke
s the adva
n
tage
s of two AR r
egist
rati
on method
s. Becau
s
e it u
s
e
s
the
marker m
e
th
od, DR-ma
r
ker is
stable
and low-co
m
putation. It will rarely occur to flut
ter.
Furthe
rmo
r
e,
DR-ma
r
ker
remove
s the
marker fr
om
the image.
It achieves t
he re
sult of
th
e
markerl
e
ss m
e
thod. Somet
i
mes it is hard to diminish
reality. Theref
ore, we
can p
u
t the marke
r
in
a sim
p
le
envi
r
onm
ent, such a
s
the
sim
p
le texture
d
e
sktop. Fo
r i
n
stan
ce, th
e
sofa i
s
only
one
colo
r, bla
ck,
in Figure 1
0
. In Figure
10(a
)
, there
is an AR
marker
abov
e the sofa.
It is
insufferable f
o
r users. The
n
we re
move the mark
er u
s
ing the pro
p
o
s
ed
DR-ma
r
ker. The re
sult
is
s
h
ow
n in
F
i
gu
r
e
1
0
(
b
)
.
N
e
xt, w
e
p
l
ac
e
a
CG
c
h
ar
ac
te
r
to s
i
t in
the
so
fa
in F
i
gu
r
e
1
0
(
c
)
.
So th
is
method can b
e
use
d
in filmmakin
g
syste
m
as we pro
p
o
se
d ARFMS
in the other p
aper.
Table 1. DR-marker P
r
oce
ssi
ng Power
Computing time
Frames pe
r second
Marker detection
and tracking
32.7ms
30.6fps
DR-ma
rker
w
i
tho
u
t occlusion total process time
39.7ms
25.2fps
DR-ma
rker
w
i
th
occlusion total pr
ocess time
211ms
4.7fps
Figure 10. Co
nce
p
tual Illust
ration
of DR-marker
(a) An
AR marker a
bove the sofa
; (b) The resu
lt
of taking out the marke
r
; (c) A CG ch
ara
c
ter ad
ded o
n
the image.
In fact, the
DR-m
arke
r me
thod is sim
p
li
city
of the o
p
e
rato
r for the
amateu
r u
s
e
r
s. T
hey
only need to
put a marker on the ground. The
n
he ca
n sho
o
t
their VFX film. First, as an
amateu
r, a u
s
er
want
s to m
a
ke
a
sho
r
t VFX film.
The
story i
s
that a
virt
ual CG b
o
y
dances in t
he
study. The
n
a Kine
ct is
a
ttached to
th
e MMD,
and
he d
a
n
c
e
s
by himself to cont
rol the
CG
cha
r
a
c
ter a
s
sho
w
n in Fig
u
re 3. A rou
gh actio
n
of CG boy is o
b
tained q
u
ickly. Afterwa
r
d the
use
r
co
ntrol
s
the characte
r manu
ally to get the
better action. The
motion file is finishe
d
. Next,
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4345 – 4
352
4352
the amate
u
r
use
r
p
u
ts a
speci
a
l ma
rker on the
boo
kcase, a
s
sh
own in the to
p l
e
ft corner im
age
in Figu
re 1
1
.
He u
s
e
s
a
la
ptop to remo
ve the marke
r
an
d ad
d th
e CG
boy o
n
the re
al on
-site
image,
who
s
e si
ze
po
sitio
n
an
d di
re
cti
on i
s
m
anu
all
y
adju
s
ted. T
hen th
e d
a
n
c
ing m
o
tion fil
e
is
con
n
e
c
ted to
the boy. Fi
n
a
lly, the
user ca
n shoot th
e film re
al ti
me that the
CG
boy da
nces i
n
the study in F
i
gure 11. If the VFX film is
wonder
ful, post-production modi
fication phase will ski
p
and the VFX
film has b
een
finished. If some det
ail
s
want to be
chang
ed, he
can utilize th
e
CG
boy and ra
w
seq
uen
ce to
synthe
size a better film. In
sum, it is very easy for an a
m
ateur to u
s
e
.
Figure 11. DR-m
arke
r Simulation Result. The top
left corne
r
imag
e is the ra
w frame, and the
other five ima
ges a
r
e real-ti
m
e pro
c
e
s
sin
g
output fram
es
4. Conclusio
n
s and Fu
tur
e
Work
This pa
per p
r
esents a n
e
w
type film
and TV
sho
o
t
system, ARFMS. Its goal is to
achi
eve WYS
I
WYG VFX filmmaki
ng. As the beginni
n
g
of the ARF
M
S proje
c
t, we imple
m
en
t a
n
ARFMS proto
t
ype. ARFMS demon
strate
s it is lo
w
co
st and e
a
sy t
o
use. T
he a
m
ateur u
s
e
r
will
mak
e
a VFX film with jus
t
an ordinary
came
r
a
, a
sp
ec
ia
l ma
rk
e
r
an
d s
o
ftw
ar
e
.
However, A
R
FMS still has
some
problems. It is
just
a prototype.
The first problem is the
pro
c
e
ssi
ng p
o
we
r of occl
usio
n algo
rithm. In
the prototype, due
to the data read time, the
occlu
s
ion
alg
o
rithm of th
e
DR-ma
r
ker i
s
not re
al ti
me, which is
4fps
. In the future work
, we will
operate GP
U to redu
ce
the comput
ing time.A
nother issue o
f
DR-m
a
rke
r
is illumination
con
s
i
s
ten
c
y. Due to lightin
g cha
nge, th
e texture
of the proje
c
ting
model will b
e
not co
nsi
s
t
ent.
Although p
r
oj
ecting m
odel
adju
s
tment wants to
so
lve
this proble
m
, the method
can be
only u
s
ed
in the illumi
nation environment is rel
a
tively st
able. It is hard
to be suitabl
e in the act
ual
compli
cate
d outdoo
r environment. The
next phase,
we will
con
s
id
er the outdo
o
r
appli
c
ation.
This
pap
er is just th
e b
egi
nning
of the
proje
c
t. We
wish mo
re
re
sea
r
che
r
s wo
uld p
a
y
clo
s
e attentio
n to the WYSIWYG filmma
king
system.
Referen
ces
[1] Nitsche
M.
Experi
m
e
n
ts in the use of game te
chno
lo
gy for
pre-visu
al
i
z
a
t
io
n.
Proceed
ings
of the 20
0
8
Confer
ence
on
F
u
ture Pla
y
: R
e
searc
h
, Pla
y
,
Share. 20
08: 1
60-1
65.
[2]
W
eng ENG, K
han
RU, Adr
u
c
e
SAZ
, et al.
Mobil
e
C
a
mer
a
as
a Hum
an
Visio
n
in
Augm
ented
Re
alit
y.
T
E
LKOMNIKA Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
n
g
.
2012; 1
1
(7): 3
568-
357
5.
[3]
Ismail AW
, Su
nar MS. Intu
itiv
eness
3D
o
b
je
cts Inte
raction
i
n
Au
gme
n
ted
Real
it
y Us
ing
S-PI Algor
ithm
.
T
E
LKOMNIKA Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
n
g
.
2013; 1
1
(7): 3
561-
356
7.
[4]
Ichikari R,
T
enmoku R, Shiba
t
a F
,
et al. Mixed r
eal
it
y
pr
e-v
i
sual
izati
on for f
ilm making: on-set camera-
w
o
rk auth
o
rin
g
and actio
n
reh
earsa
l.
Internat
ion
a
l Jour
na
l of Virtual Re
ality
.
2008; 7(4): 2
5
-
32.
[5] Sziebig
G.
Achievi
ng tota
l i
m
mers
ion: T
e
chno
logy tre
n
d
s be
hin
d
au
gmente
d
rea
lit
y - A survey
.
WSEAS Internatio
nal Co
nf
erenc
e. Proceedi
ngs.
Mathe
m
atics and C
o
mputers in
Scienc
e and
Engi
neer
in
g. 2009: 45
8-4
63.
[6]
Starck J, Maki
A, Nobuh
ara S
,
Hilton
A, Matsu
y
ama
T
.
T
he Multiple-
C
am
era 3-D Pro
d
u
c
tion Studi
o.
IEEE Transactions on C
i
rcuits
and Syste
m
s for Vide
o Techn
o
lo
gy
. 200
9; 19 (6): 856-8
69.
[7]
Kato H, Billi
ng
hurst M.
Marker T
r
acking
an
d HMD Ca
libr
a
tion for a Vid
e
o
-bas
ed Au
g
m
ented R
e
a
lity
Confer
enci
ng System.
2nd I
EEE and
ACM Internation
a
l
W
o
rkshop o
n
Augmente
d
Realit
y
.
Sa
n
F
r
ancisco. 19
9
9
: 85-94.
[8]
Klein G, Murr
ay
D.
Par
a
ll
el
tracking an
d ma
pp
ing for s
m
a
ll AR w
o
rkspaces.
6th IEEE and
A
C
M
Internatio
na
l Symp
osi
u
m on
Mixed a
nd
Aug
m
ented R
eal
it
y
.
2007: 22
5-23
4.
[9]
Lin
g
C, Z
hao J, Z
hang W
.
DR-Marker: A Novel Di
mi
ni
shin
g-Re
ality-B
a
sed AR Re
gi
stration
. 201
3
Seventh Inter
n
ation
a
l Co
nfere
n
ce onI
ma
ge a
nd Graph
ics. Qing
dao. 2
013:
716-
720.
[10]
F
Jan, R Hol
g
e
r
, B Gregor
y
.
D
e
tecting
dyn
a
m
ic occlus
io
n in f
r
ont
of static b
a
ckgro
unds f
o
r AR scen
e
s
.
Procee
din
g
s of
the
w
o
rksho
p
on V
i
rtu
a
l env
ir
onme
n
ts. Z
u
rich. 2003: 1
53-1
61.
Evaluation Warning : The document was created with Spire.PDF for Python.