GENERATING STYLIZED DANCE MOTION FROM

LABANOTATION BY USING AN AUTONOMOUS DANCE AVATAR

Worawat Choensawat and Kozaburo Hachimura

School of Science and Engineering, Ritsumeikan University, Kyoto, Japan

Keywords:

Dance Notation, LabanEditor, 3D CG Animation, Associate Memory, Knowledge-based Systems.

Abstract:

When producing the animation of a body motion from the dance notation, the dance knowledge is a key for

accomplishing high-quality movement. This knowledge enables the dancer to know how to perform the cor-

rect movement from a movement notation score. This paper presents an approach for automatically simulating

a CG animation from Labanotation scores. We achieve this goal by the integration of a CG animation with a

dance-style interpretation module and it is called an autonomous dance avatar. In our experiment, we imple-

mented an autonomous dance avatar to perform a Japanese stylized traditional dance such as Noh-Plays. The

result shows that the autonomous dance avatar can reproduce Noh-Play correctly from Labanotation after it

has been trained with the Noh-Play knowledge.

1 INTRODUCTION

Dance community, mainly in Western countries, has

widely accepted Labanotation as a graphical notation

scheme for describing human body movement. Sim-

ilar to music score, Labanotation uses staff and sym-

bols for the purpose of recording human movements

in the ﬁelds of choreography and dance education

(Hutchinson Guest, 1977).

Labanotation does not represent the nuances of a

performance and exact movements of any particular

dancer. However, it does capture the choreographer’s

creative idea, so that any person might interpret and

perform those ideas again. Based on the aforemen-

tioned, with the same notation score, different dancers

may perform a movement differently depend on their

experience.

Labanotation is rich in symbols, and by using the

full set of symbols; almost all of our body movements

can be described. However, the resulting notation

would become extremely complicated and difﬁcult to

comprehend. For that reason, the fundamental sym-

bols have usually been used. The question is: how

can we realize a method of describing peculiar fea-

tures and nuances of artistic, traditional dance move-

ments while suppressing the complexity in the nota-

tion score?

Hachimura and his research team have developed

a system, named LabanEditor (Kojima et al., 2002),

for editing Labanotation score and displaying the CG

character animation of its score. LabanEditor uses a

motion template for generating a CG animation from

the fundamental elements of Labanotation (as illus-

trated in Section 2). The motion template describes

the relationship between Labanotation symbols and

the rotation and translation of the corresponding joint.

However, using a single motion template, the system

cannot reproduceslightly distinct poses that are some-

times deﬁned with the same symbol.

For a current version of LabanEditor (Choensawat

et al., 2010), the system uses the method of dynamic

templates in order to represent the nuances of dance

movements. With the dynamic template method, the

system allows users to describe a single Labanotation

score with multiple templates. However,this will load

a user task for describing every single dance motion.

Because the creation of motion templates is a dif-

ﬁcult task for users, this burden on the users can

be lessened by having a knowledge based of dance

styles embedded in a character model. In this pa-

per, we present a dance-style interpretation module

embedded in the character model, called Autonomous

Dance Avatar. The embedded module enabled an

autonomous dance avatar to encode the pattern of

Labanotation score and select an appropriate dance

movement to the pattern from the learned knowledge.

The proposed module is built in LabanEditor.

535

Choensawat W. and Hachimura K. (2012).

GENERATING STYLIZED DANCE MOTION FROM LABANOTATION BY USING AN AUTONOMOUS DANCE AVATAR.

In Proceedings of the International Conference on Computer Graphics Theory and Applications, pages 535-542

DOI: 10.5220/0003946305350542

 SciTePress

2 LABANOTATION

A Labanotation score is drawn in the form of vertical

staff where each column corresponds to a body part.

Figure 1(a) is an example of Labanotation scores cor-

responding to dance motion. Figure 1(b) shows the

basic arrangement of columns in the staff. The hor-

izontal dimension of the staff represents the parts of

the body, and the vertical dimension represents time.

The center line of the staff represents the center of the

body: Columns on the right represent the right side

of the body, and columns on the left, the left side of

the body. Symbols are placed in the columns of the

staff. The vertical length of a symbol shows the du-

ration of the movement, from its beginning to its end

(Hutchinson Guest, 1977).

^ƵƉƉŽƌƚ

ZŝŐŚƚ>ĞŐ'ĞƐƚƵƌĞ

>ĞĨƚ>ĞŐ'ĞƐƚƵƌĞ

ŽĚǇ

>ĞĨƚƌŵ

ZŝŐŚƚƌŵ

,ĞĂĚ

>ĞĨƚ ZŝŐŚƚ

ϵĐŽůƵŵŶƐ

ϰϯ

;ĂͿ ;ďͿ

Figure 1: Labanotation scores: (a) example of Labanota-

tion scores, (b) columns of Labanotation representing body

parts.

Figure 2(a) shows direction symbols, used for de-

scribing the direction of movement of body parts. The

shape of a symbol represents the horizontal direction

of motion. Shading within a direction symbol shows

the level of a movement, i.e. vertical direction of

movement (low, middle, and high), as shown in Fig-

ure 2(a). Figure 2(b) shows the rotation signs and re-

lationship pins respectively. The motion of each body

part is expressed by a sequence of symbols placed in

Figure 2: Symbols and signs used in Labanotation: (a) di-

rection symbols, and (b) rotation signs.

the corresponding column.

Labanotation is rich in symbols, all type of move-

ment ranging from the simplest to the complex can be

accurately described. Its usefulness are not limited to

dancers and choreographer; the system has also been

successfully applied to every ﬁeld in which there is

the need for recording human body motions e.g. ath-

letics, anthropology, and physiotherapy.

3 RELATED WORK

3.1 Utilizing Labanotation for Dance

Communities

To date, several graphics applications have been de-

veloped for preparing Labanotation scores and gener-

ating the body movement.

LabanWriter (Fox, 2000) is currently the most

widely used Labanotation editor. The system is only

for preparing Labanotation scores and recording them

in digital form. It does not provide a function for dis-

playing character animations corresponding to the no-

tation. The latest version of LabanWriter can handle

about 700 Labanotation symbols.

There have been several attempts to generate

CG animation from Labanotation. The CG anima-

tion generator transforms Labanotation scores, which

were prepared with LabanWriter, to the animation via

the commercial software LifeForms (Coyle et al.,

2002). However, LifeForms can only support the fun-

damental symbols of Labanotation.

LabanDancer (Wilke et al., 2005) is a Laban-

Writer scores to 3D animation translation tool. Like

LifeForms, LabanDancer does not have any functions

for preparing Labanotation scores and supports only

a limited number of symbols.

LabanChoreographer (Zhang et al., 2006) is intro-

duced for choreographing by retrieving the most sim-

ilar motions from a motion capture database. Labano-

tation is used as an index tool for retrieval. Character

animation is produced from motion capture data but

not from the notation.

Practically, the above application software has

some restrictions as follows. First, they lack of the

integration of both creating Labanotation scores and

producing 3D CG character animation. This is be-

cause they were separately designed and developed.

Secondly, their software mainly focused on Western

dances, and it takes no particular account of stylized

dance motions of other cultures.

We decided to implement a dance-style interpreta-

tion embedded in the character model in LabanEditor

GRAPP 2012 - International Conference on Computer Graphics Theory and Applications

536

because LabanEditor has the capability of preparing a

Labanotation score and displaying the 3D CG charac-

ter animation associated with the score.

3.2 Associative Memory

Simulating human-like learning or cognitive learning,

we focus on an associative memory as referred to

a content-addressable memory (Haykin, 1998). The

content-addressable memory is a memory organiza-

tion in which the memory is accessed by its content.

If a pattern is presented to an associative memory, it

returns whether this pattern coincides with a stored

pattern. The coincidence need not be an exact match.

An associative memory may also return a stored pat-

tern that is similar to the presented one, so that noisy

input can also be recognized.

An associative memory is used in information re-

trieval; for example, it is used for creating a memory

of related keywords to produce thesauri (or knowl-

edge bases) as introduced in (Lu et al., 2008; Chen

et al., 2003; Chen et al., 1993). These automatic

thesauri were then integrated with some existing

manually-created thesauri for assisting concept explo-

ration and query reﬁnement. This associated mem-

ory was implemented with Hopﬁeld networks (Ad´an-

Coello et al., 2007). However, other models such as

Boltzmann machine (Mairal et al., 2010), hypernet-

works (Zhang and Kim, 2006; Zhang, 2008) have

been applied to a search system of the word dictio-

nary and a sentence completion when missing some

words.

For the application of associative memories , we

are interested in the string matching problem. If there

is a given pattern, the problem consists in ﬁnding one

or more usually all the occurrences of a pattern in a

text. This problem has commonly occurred in many

applications involving information retrieval such as

bibliographic search and molecular biology. The con-

cept of associative memories can be used for solving

this problem. Such hypernetworks are used for learn-

ing the higher-order associations of the words from

a text corpus. As described by Zhang et al. (Zhang,

2008), the hypernetwork memory is used for gener-

ating a text dialogue for a given movie scene image.

The hypernetwork memory has the recall and recog-

nition capability. For example, a training sentence

“You need to wear it” and its source from the movie

“24”. In the recall task, the hypernetwork is given,

say, “? need to wear it” and should complete the miss-

ing word to produce ”You need to wear it”. In the

recognition task, the hypernetwork is to output “24”

as the source of the sentence.

As suggested above, we can adopt an associative

memory for generating a dance motion from Labano-

tation. In terms of Labanotation, a posture (or pose)

is a position of the body as represented by a combi-

nation of Labanotation symbols. Then, a pose can be

deﬁned as a smallest unit of a Labanotation score. In

an analogous manner, a Labanotation score would be

comparable to a sentence of which words are equiva-

lent of Labanotation units and characters are Laban-

otation symbols. Lastly, the source of sentences are

comparable to a dance-style movement. The method-

ology of the dance-movement generation by using the

associative memory will be described in Section 5.

4 LABANEDITOR

LabanEditor is an interactive graphical editor for edit-

ing Labanotation scores and displaying the 3D CG

character animation associated with scores. The in-

teractive interface for preparing Labanotation score

allows users to input and edit the score by drag-and-

drop techniques. When replaying the Labanotation

score, users can observe the animation with a red hor-

izontal line cursor moving upward corresponding to

the animation progresses, as shown in Figure 3.

In the LabanEditor system, Labanotation scores

can be represented as a simple format called Labano-

tation Data (LND), which uses alphanumeric charac-

ters to represent basic symbols. The example of LND

representation is shown in Figure 4(b). The lines that

begin with “#” indicate the fundamentalparameters of

Labanotation. The movement of a body part is spec-

iﬁed in the line followed by a command “direction”,

which corresponds to the Labanotation direction sym-

bols. Figure 4 illustrates how a Labanotation score is

convert to LND structure.

Figure 3: LabanEditor

LND describes a pose of the body at each timing

just like key-frame body postures for animation, so

GENERATING STYLIZED DANCE MOTION FROM LABANOTATION BY USING AN AUTONOMOUS DANCE

AVATAR

537

, Ğ Ă Ě Ğ ƌ

^ Ǉ ŵ ď Ž ůƐ

 Ž Ě Ǉ  Ɖ Ă ƌ ƚ Ɛ

 ŝƌ Ğ Đ ƚ ŝŽ Ŷ > Ğ ǀ Ğ ů ^ ƚĂ ƌ ƚ  Ŷ Ě

η ď Ğ Ă ƚ ϰ ͬ ϰ

η ƚ Ğ ŵ Ɖ Ž ϭ Ϯ Ϭ

Ě ŝ ƌ Ğ Đ ƚ ŝŽ Ŷ ůͺ Ɛ Ƶ Ɖ Ɖ Ž ƌ ƚ Ɖ ů Ă Đ Ğ ŵ ŝĚ Ϭ ͘ Ϭ Ϭ ͘ Ϭ

Ě ŝ ƌ Ğ Đ ƚ ŝŽ Ŷ ƌ ͺ Ɛ Ƶ Ɖ Ɖ Ž ƌ ƚ Ɖ ů Ă Đ Ğ ŵ ŝĚ Ϭ ͘ Ϭ Ϭ ͘ Ϭ

Ś Ž ůĚ Đ Ğ Ŷ ƚĞ ƌ Ϭ ͘ Ϭ

Ě ŝ ƌ Ğ Đ ƚ ŝŽ Ŷ ƌ ͺ Ă ƌ ŵ ĨŽ ƌ ǁ Ă ƌ Ě ŵ ŝĚ Ϭ ͘ Ϭ Ϯ ͘ Ϭ

Ě ŝ ƌ Ğ Đ ƚ ŝŽ Ŷ ůͺ Ă ƌ ŵ ůĞ Ĩ ƚ ůŽ ǁ Ϯ ͘ Ϭ ϰ ͘ Ϭ

Ě ŝ ƌ Ğ Đ ƚ ŝŽ Ŷ ƌ ͺ Ă ƌ ŵ ĨŽ ƌ ǁ Ă ƌ Ě Ś ŝ Ő Ś Ϯ ͘ Ϭ ϰ ͘ Ϭ

;ď Ϳ

ϭ Ϯ

;Ă Ϳ

Figure 4: Relationship between Labanotation score and

LND.

that we can produce motion of a body part by sim-

ply applying interpolation between start and end key-

frame poses. A key-frame pose of a body part at a

time corresponding to an end of a symbol is deﬁned

by a Labanotation symbol. The system converts di-

rection symbols into animation key-frames by using

a motion template for a mapping between the symbol

and its corresponding pose of the body part.

For generating a animation, the system converts

LND into animation key-frames by using a motion

template ﬁle for a mapping between the symbol and

its corresponding pose of the body part. The motion

template ﬁle describes the relationship between a di-

rection symbol at the particular column and the rota-

tion and translation of the corresponding joint.

Figure 5 shows a notation and description in a mo-

tion template ﬁle, and the resulting pose. The symbol

marked “A” in Figure 5 (a) is mapped to the descrip-

tion of the part marked “A” in the motion template

ﬁle shown in Figure 5 (b), which indicates a target

pose of the right arm achieved by rotating the right

shoulder joint 90 degree counterclockwise around the

y-axis from the standard pose as shown in Figure 5

(c).

5 AUTONOMOUS DANCE

AVATAR

An autonomousdance avatar is a character model em-

bedded with an capability of dance-style interpreta-

tion. The interpretation of dance styles is the re-

call and classiﬁcation process of stored Labanota-

tion scores and their associated motion templates. To

achieve that, we started with teaching a dance avatar

to have a dance-style memory, which is an associative

memory between Labanotation scores and the corre-

Figure 5: Relationship between user input symbols and a

template ﬁle; (a) user input symbol, (b) part of a template

ﬁle, and (c) target pose corresponding to the template in (b).

sponding movements. The dance-style memory is an

associative function between scores and their move-

ments as shown in Figure 6. Figure 6 is an example

of stylized traditional dances, Noh-Plays where the

leftmost column shows Labanotation scores related to

four unit movements (called Kata) and the snapshots

of the corresponding CG animation.

Figure 6: CG character animation of four Kata: (a) Hiraki,

(b) Tachi, (c) Shitai-tome, and (d) Ougi-kazashi.

GRAPP 2012 - International Conference on Computer Graphics Theory and Applications

538

The dance-style memory can be designed and

implemented with a two-layered, associative mem-

ory. The ﬁrst layer involves a recall process

of known poses. For example, as shown in

Table 1, a combination of Labanotation sym-

bols associated with pose#20 comprises three sym-

bols of (l

support,place,mid), (r support,place,mid),

(r arm,forward,mid), which these symbols are related

to Labanotation symbols #1, #2, and #4 as shown in

Figure 4 (a). After that, the second layer classiﬁes a

sequence of poses to a trained dance style; for exam-

ple, a sequence of pose#20 and #10 is classiﬁed as

motion#3.

Table 1: Dance-style interpretation with a two-layered, as-

sociative memory.

Layer Query Recall/

Recognition

Labanotation unit Pose No.

(l support,place,mid)

pose#20(r support,place,mid)

(r arm,forward,mid)

(l arm,left,low)

pose#10

(r arm,forward,high)

Sequence of poses Movement

(#20, #10) motion#3

(#5, #7, #30) motion#8

To summarize, we divide a task of autonomous

dance avatar into two sequential subtasks as follows:

1. Decompose a Labanotation score into a number

of units, and

2. Store and retrieve dance styles in/from the two-

layered, associate memory.

The interpretation consists of storing and retriev-

ing processes as described in Algorithms 2 and 2, re-

spectively. First of all, both algorithms must start with

the decomposition of Labanotation scores into units,

and then follow by training or testing stages as shown

in Algorithms 2 and 2), respectively. Algorithm 2 de-

scribes the implementation of an associative memory

for storing Labanotation units and ,then, dance-style

patterns which a pattern can be formed as a concate-

nation of units while Algorithm 2 shows the retrieving

method of dance-style patterns.

Given an unknown Labanotation score, we can as-

sign a set of motion templates to it by applying Algo-

rithm 2. After decomposing the Labanotation score,

Algorithm 2 starts with retrieving a stored unit from

the 1

-layer, associative memory for all units. Lastly,

we adopt a concept of string matching for searching a

input : A set of Labanotation scores and its

corresponding motion templates

output: A storage of dance-style patterns

1 Decompose Labanotation scores into units;

2 Create a two-layered, associative memory for

storing dance-style patterns;

: storing Labanotation units

: storing a sequence of Labanotation units

and its associated motion tempates

Algorithm 1: Developing the dance-style pat-

tern storage.

set of motion templates. A sequence of units is analo-

gous to a text. That is to ﬁnd an occurrence of a set of

patterns (deﬁned in Algorithm 2) in the text. This can

be implemented by the 2

-layer, associative memory.

input : An unknown Labanotation score

output: A set of motion templates associated

with its score

1 Decompose the Labanotation score into units;

2 for each Labanotation unit do

3 Retrieve a most matching Labanotation unit

from the 1

-layer, associative memory;

4 end

5 Retrieve a set of motion templates from the

-layer, associative memory;

Algorithm 2: Retrieving a dance-style pattern

from the storage.

5.1 Decomposition of a Labanotation

Score

A Labanotation score is a set of symbols aligned

along a time line as explained in Section 2. Given

a set of symbols, we can ﬁnd a minimum number

of subsets where a subset must compose of coinci-

dent symbols. Each subset represents as a unit. Al-

gorithm 3 shows how to break down a score into a

number of units as similar to the minimum clique par-

tition problem where the problem and its solution is

described in the graph algorithms and applications as

found in (Cenek and Stewart, 2003).

Given a score, vertices can be represented by sym-

bols appearing in the score. Let l

, c

, u

be lowest,

center, and highest points of symbol i, respectively.

Edges e

will be one if and only if

≤ c

≤ u

OR l

≤ c

≤ u

GENERATING STYLIZED DANCE MOTION FROM LABANOTATION BY USING AN AUTONOMOUS DANCE

AVATAR

539

For example, we will show a decomposition by

a simple example. Given a Labanotation score as

shown in Figure 7, we can ﬁnd four cliques by ap-

plying Algorithm 3 as shown in Table 2.

1 2 3

5 7

1 2

3 4

8 9

;ĂͿ ;ďͿ ;ĐͿ

Figure 7: Transform a Labanotation score to an undirected

graph G(V, E): (a) a Labanotation score, (b) Labeling sym-

bols with 1 to N where N is a number of symbols in the

score, and (c) the corresponding graph V = {1, . . . , N};N =

5.2 Associative Memory

The purpose to use of an associative memory is for

building a dance knowledge. The knowledge can be

constructed by using a training set of input and target.

After training, the memory is equivalent to a mapping

function between inputs and the associative outputs.

This can be accomplished by using the Bayesian sta-

tistical theory (Agrawal and Srikant, 1994; Liu et al.,

1998). We are implementing an associative mem-

ory for interpreting Labanotation scores to motions.

The interpretation from scores to motions must pass

throughout poses as an in-between data as described

in Table 1. Subsequently, our model is a two-layered,

associative memory where the ﬁrst layer is for match-

ing between Labanotation symbols and a Labanota-

input : A Labanotation score

output: A set of Labanotation units

1 Draw an undirected graph G(V, E), where V is

a set of all symbols appearing in the score;

2 Determine E = {e

} by using the equation

below;



1 if c

∈ [l

, u

]

0 otherwise

3 Partition V into a minimum number of cliques;

Algorithm 3: Decomposing a Labanotation

score to a set of minimum units.

Table 2: Example of a list of Labanotation units.

Labanotation score: Figure 7(a)

No. of symbols: 9

No. of units: 9 (minimum of cliques)

Clique No. Set of symbols

{1, 2, 3, 4}

{5, 6}

{6, 7}

{8, 9}

Motion template No. Pattern

tion unit, and the second is for assigning a sequence

of Labanotation units with a motion template ﬁle.

For the ﬁrst-layer, the memory is used for convert-

ing a Labanotation score to a sequence of poses. We

ﬁrst store the Labanotation units referred as poses, af-

ter that the memory can recall the stored poses. For

unknown pose in the recall process, the memory will

try to search the most similar pose even if some miss-

ing symbols are occurred. The implementation is

based on the statistical theory of the joint probability

of a Labanotation symbol S and a Labanotation unit

U as shown below:

P(S, U|W

) = P(x

, x

) = P(x|W

) (1)

where W

is the training parameters and x = (x

, x

)

is the training pattern consisting of a Labanotation

symbol x

and a Labanotation unit x

. The imple-

mentation can be achieved by using a Bayesian clas-

siﬁer (Witten et al., 2011; Hall et al., 2009).

The second-layer, associate memory is used for

matching between a sequence of Labanotation units

and its corresponding motion template. Similar to the

ﬁrst layer, the implementation of the second layer is

as shown below and let a sequence of units be Q and

a motion template be T.

P(Q, T|W

) = P(y

, y

) = P(y|W

) (2)

where W

is the training parameters and y = (y

, y

)

is the training pattern consisting of a sequence of units

and templates x

6 EXPERIMENT AND

PRELIMINARY RESULT

In experiment, we test the dance-style interpreta-

tion of Noh-Plays. For a brief description, Noh-

plays are one of the most famous and characteristic

Japanese traditional performing arts. Noh movements

GRAPP 2012 - International Conference on Computer Graphics Theory and Applications

540







Figure 8: Snapshot of the CG animation for a Labanotation score comparing between the dance avatar with Noh knowledge

and the normal avatar.

are highly stylized and unique. ‘Shimai’ is a short but

principal performance extracted from the whole Noh

play. In principle, each Shimai is composed of a num-

ber of prescribed movement units known as ‘Kata’, or

form.

For the preparation of a dance-style database, we

recorded the Noh-Plays (performed by Mr. Toyohiko

Sugiura, who is the master of Kanze Noh School) by

using three video cameras in the following angles:

front, side, and perspective views, respectively. We

have the videos of 6 Shimai with 32 unique Kata in

total. By precisely observing the videos, we described

each Kata with Labanotation and their associated mo-

tion templates. We use these Labanotation score and

their associated motion templates to built the dance

knowledge.

The aforementioned process involves with the

preparation of training data that consists of Labanota-

tion scores and the associated motion templates. We

implemented the autonomous dance avatar embed-

ded in LabanEditor and used Java classes of Bayesian

classiﬁer (Witten et al., 2011) for developing the two-

layered, associate memory. After training a dance

avatar, it can perform a Noh play correctly where an

example is shown in Figure 8. Figure 8 shows the

snapshots of the autonomous dance avatar embedded

with a dance-style interpretation module comparing

with a normal avatar. The autonomous dance avatar

can move its body according to their stored motion

patterns while the normal avatar just moved its body

following a standard movement. Even though two

avatars put their arm besides their body, they have

different postures of their left arm. The autonomous

dance avatar has its slightly bent left arm while the

other has its left arm straight out. While their move-

ment according to Labanotation unit D (Figure 8), the

autonomous dance avatar rotating its right hand differ

from that of the normal avatar.

7 CONCLUSIONS AND FUTURE

WORK

Since using the fundamental description of Labano-

tation cannot describe a detail of human body move-

ment, we have to create the motion template that de-

scribes the relationship between Labanotation sym-

bols and the rotation and translation of the corre-

sponding joint. In this paper, we present an au-

tonomous dance avatar in which a dance-style inter-

pretation module embedded. The embedded mod-

ule enabled an dance avatar to encode the pattern of

Labanotation score and select an appropriate dance

movement to the pattern.

The contribution of this paper is a proposed frame-

work for developing the autonomous dance avatar

which includes the following mechanisms:

1. a mechanism based on a minimum clique partition

for ﬁnding minimum independent units,

2. a mechanism incorporated an associative memory

for storing and retrieving Labanotation units, and

searching a motion template from a sequence of

Labanotation units.

GENERATING STYLIZED DANCE MOTION FROM LABANOTATION BY USING AN AUTONOMOUS DANCE

AVATAR

541

In our experiment, we test our approach against

the stylized Japanese traditional dance, Noh-Plays.

We create the database of Noh-Plays acquired from

a recorded video of a Noh expert. The experimen-

tal results shows that the autonomous dance avatar

can remember the Noh movement pattern. Compar-

ing with a normal avatar, the autonomousdance avatar

can pose its body according to Noh style. This is an

preliminary result.

As beneﬁcial for Noh players/learners, the au-

tonomous avatar in LabanEditor can be used for the

following goals:

• Self Studying: Noh beginners have the possibil-

ity of studying body motions on their own via the

notation and CG animation.

• Expressing Idea: They can use the system as a

presentation tool for their idea about the choreog-

raphy of the performance and display in 3D CG

animation.

• Choreographing a Noh Play: Ability to chore-

ograph a Noh play without having to have the

knowledge of Labanotation.

In future work, the dance-style knowledge of the

autonomous dance avatar will expand to cover other

dance styles. The system of a variety of dance-style

knowledge will be implemented and evaluated in both

user and expert domains. The scope of our evaluation

will be related with the usefulness of the system, the

accuracy and quality of 3D character animation.

REFERENCES

Ad´an-Coello, J., Tobar, C., de Freitas, R., and Marin, A.

(2007). Hopﬁlter: an agent for ﬁltering web pages

based on the hopﬁeld artiﬁcial neural network model.

Data Management. Data, Data Everywhere, pages

164–167.

Agrawal, R. and Srikant, R. (1994). Fast algorithms for

mining association rules in large databases. In 20th

International Conference on Very Large Data Bases,

pages 478–499. Morgan Kaufmann, Los Altos, CA.

Cenek, E. and Stewart, L. (2003). Maximum indepen-

dent set and maximum clique algorithms for overlap

graphs. Discrete Applied Mathematics, 131(1):77–91.

Chen, H., Lally, A., Zhu, B., and Chau, M. (2003). Help-

fulmed: intelligent searching for medical information

over the internet. Journal of the American Society for

Information Science and Technology, 54(7):683–694.

Chen, H., Lynch, K., Basu, K., and Ng, T. (1993). Gener-

ating, integrating, and activating thesauri for concept-

based document retrieval. IEEE Expert, 8(2):25–34.

Choensawat, W., Takahashi, S., Nakamura, M., Choi, W.,

and Hachimura, K. (2010). Description and reproduc-

tion of stylized traditional dance body motion by us-

ing labanotation. Transactions of the Virtual Reality

Society of Japan, 15(3):379 – 388.

Coyle, M., Maranan, D., and Calvert, T. (2002). A tool for

translating dance notation to animation. In Proceed-

ings of Western Computer Graphics Symposium.

Fox, I. (2000). Documentation technology for the 21st cen-

tury. Proceedings of World Dance, pages 136–142.

Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reute-

mann, P., and Witten, I. (2009). The weka data min-

ing software: an update. ACM SIGKDD Explorations

Newsletter, 11(1):10–18.

Haykin, S. (1998). Neural Networks: A Comprehensive

Foundation. Pearson Education.

Hutchinson Guest, A. (1977). Labanotation. New York:

Routledge, Chapman y Hall.

Kojima, K., Hachimura, K., and Nakamura, M. (2002). La-

baneditor: Graphical editor for dance notation. In

Robot and Human Interactive Communication, 2002.

Proceedings. 11th IEEE International Workshop on,

pages 59–64. IEEE.

Liu, B., Hsu, W., and Ma, Y. (1998). Integrating classiﬁca-

tion and association rule mining. In Fourth Interna-

tional Conference on Knowledge Discovery and Data

Mining, pages 80–86. AAAI Press.

Lu, W., Lin, R., Chan, Y., and Chen, K. (2008). Using web

resources to construct multilingual medical thesaurus

for cross-language medical information retrieval. De-

cision Support Systems, 45(3):585–595.

Mairal, J., Bach, F., and Ponce, J. (2010). Task-driven dic-

tionary learning. Pattern Analysis and Machine Intel-

ligence, IEEE Transactions on, pages 1–1.

Wilke, L., Calvert, T., Ryman, R., and Fox, I. (2005). From

dance notation to human animation: The labandancer

project. Computer Animation and Virtual Worlds,

16(3-4):201–211.

Witten, I., Frank, E., and Hall, M. (2011). Data Min-

ing: Practical machine learning tools and techniques.

Morgan Kaufmann.

Zhang, B. (2008). Hypernetworks: A molecular evolu-

tionary architecture for cognitive learning and mem-

ory. Computational Intelligence Magazine, IEEE,

3(3):49–63.

Zhang, B. and Kim, J. (2006). Dna hypernetworks for infor-

mation storage and retrieval. DNA computing, pages

298–307.

Zhang, S., Li, Q., Yu, T., Shen, X., Geng, W., and Wang,

P. (2006). Implementation of a notation-based motion

choreography system. Interactive Technologies and

Sociotechnical Systems, pages 495–503.

GRAPP 2012 - International Conference on Computer Graphics Theory and Applications

542