Dance Motion Segmentation Method based on Choreographic

Primitives

Narumi Okada

, Naoya Iwamoto

, Tsukasa Fukusato

and Shigeo Morishima

Waseda University, JST, Tokyo, Japan

Waseda Research Institute for Science and Engineering, JST, Tokyo, Japan

Keywords: Choreographic Primitives, Motion Segmentation, Motion Capture, Dance Motion.

Abstract: Data-driven animation using a large human motion database enables the programing of various natural

human motions. While the development of a motion capture system allows the acquisition of realistic

human motion, segmenting the captured motion into a series of primitive motions for the construction of a

motion database is necessary. Although most segmentation methods have focused on periodic motion, e.g.,

walking and jogging, segmenting non-periodic and asymmetrical motions such as dance performance,

remains a challenging problem. In this paper, we present a specialized segmentation approach for human

dance motion. Our approach consists of three steps based on the assumption that human dance motion is

composed of consecutive choreographic primitives. First, we perform an investigation based on dancer

perception to determine segmentation components. After professional dancers have selected segmentation

sequences, we use their selected sequences to define rules for the segmentation of choreographic primitives.

Finally, the accuracy of our approach is verified by a user-study, and we thereby show that our approach is

superior to existing segmentation methods. Through three steps, we demonstrate automatic dance motion

synthesis based on the choreographic primitives obtained.

1 INTRODUCTION

The recent digitization of multimedia content has

evolved streaming technologies, and the number of

character animations has been increasing. Among

them, dance animation has been attracting world-

wide attention, because of their sophistication and

artistry. Character dance animation is created by two

main approaches as follows.

The first approach seeks to interpolate key-

frames, i.e., a character’s fundamental postures. This

method enables the creation of high-quality

animation by editing character postures. However,

creating character dance motion by this approach is

difficult, because it requires a match of the rhythm

between music and motion.

The second approach seeks to capture a dancer’s

dance motion using a motion capture system. This

method enables a smooth and high-fidelity

representation of character dance motion. However,

the method requires the added cost of a physical

image capturing space. Moreover, dancer motion

must be recaptured to create a new dance animation.

Thus, this approach is costly and time-consuming.

From these considerations, an approach that creates

realistic dance motion animation automatically and

efficiently by reusing the available dance motion

data is highly desirable. Such an approach could

proceed by connecting dance motion segments that

have been segmented from a series of previously

captured dance motions; thus, a high-quality new

dance motion can be easily generated.

Traditional human motion segmentation methods

have focused on the periodic motion or angular

velocity of each joint. However, these methods are

not applicable to non-periodic and asymmetricaly

motions such as a dance motion. In particular, a

dance motion is composed of various elements of

choreography, which we call “choreographic

primitives” in this paper. Generally, professional

human choreographers create an entire dance

performance by combining choreographic primitives.

Dividing a dance motion into choreographic

primitives is expected to improve the quality of the

dance motion synthesized by a data-driven approach.

In this study, our goal is to segment dance

motion into the framework of choreographic

primitives. First, we perform a perception-based

332

Okada N., Iwamoto N., Fukusato T. and Morishima S..

Dance Motion Segmentation Method based on Choreographic Primitives.

DOI: 10.5220/0005304303320339

In Proceedings of the 10th International Conference on Computer Graphics Theory and Applications (GRAPP-2015), pages 332-339

ISBN: 978-989-758-087-1

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

investigation using professional dancers to

accurately detect the segmentation points of a

particular dance motion. Then, we define rules for

detecting the boundaries of choreographic primitives

based on the results of the initial investigation.

Finally, we verify the accuracy of our approach by a

user-study. Our approach seeks to enable automatic,

unsupervised dance motion segmentation according

to the evaluated motion and specific features of the

accompanying music. We further seeks to improve

the accuracy of the detection of choreographic

primitives.

The remainder of this paper is organized as

follows. In Section 2, we review related work. In

Section 3, we discuss the factors inherent in the

segmentation process based on choreographic

primitives. In Section 4, we describe the main ideas

underlying the algorithms used in the proposed

method. Section 5 presents the results. In Section 6,

we conclude the paper and discuss limitations and

future work.

2 RELATED WORK

Accompanying the continuous increase in the

abundance of computer graphic content, numerous

studies have considered the reuse of the existing

motion data. In particular, automatic motion

generation systems have been researched by Kovar

et al. (2002), Arikan et al. (2002), and Beaudoin et al.

(2008). These researches have sought to calculate

the similarity of postures in motion databases and to

connect the postures whose values are high, thereby

creating new natural motions.

Recently research on automatic motion

generation, especially for a dance motion have been

proposed by Kim et al. (2003), Alankus et al. (2005)

and Shiratori et al. (2006). These researchers have

attempted to generate a new series of dance motions

by connecting the dance motion segments derived

from a database. Considering the temporal

information of motion for synchronizing to the input

music is necessary. To detect beats from the dance

motion data, Kim et al. (2003) and Alankus et al.

(2005) used the moments of directional changes of

the motion. Shiratori et al. (2006) detected the local

minimum points of Weight Effort which is the sum

of the absolute angular velocities of the joints. They

inspired by a theory developed by Laban et al.

(1971). This theory is based on an impression of a

human motion from the standpoint of effort and

shape; effort represents the power of the movements,

and shape represents the silhouette of the

movements. Furthermore, Nakata et al. (2002)

determined that relation between the strength of a

motion and human feelings was more strongly

correlated with effort than shape. Then, using the

minimum points of Weight Effort, considering the

stopping points of the bodily motion easily is

possible.

From this viewpoint, the fundamental step is to

collect motion segments in the form of a database to

enable the reuse of the motion data and to create a

new series of motions.

To construct a database, numerous studies have

been performed for the detection and segmentation

of motion images into a series of some type of a

motion primitive. Osaki et al. (2000) proposed a

method based on joint velocities. Barbic et al.

(2004) proposed PCA (Principal Component

Analysis) based method. Zhou et al. (2008) and

Zhou et al. (2013) proposed a method based on an

extension of kernel k-means clustering and Vögele

et al. (2014) proposed the methods based on

the

positions of the head, wrists and ankles.

These methods are valid for cyclic motions such

as walking; however, in dance motions, because of

the complex posture and noncyclic motions, these

methods are not applicable.

Dance motion segmentation frameworks have

been mainly based on musical beats. Kim et al.

(2007) and Fan et al. (2012) focused on the feature

that dance choreography is made up of motions

synchronized with the music and proposed a method

that used the musical tempo. Lee et al. (2013)

divided a dance motion based on points where the

musical features change. However, to generate

variety of a new series of dance motions, segmenting

the existing dance motion into choreographic

primitives is necessary. The length of each

choreographic primitive would generally be different

and the choreography often differs even if the same

melodies exist in the music. Therefore, an accurate

segmentation is difficult task without including

motion information.

To detect choreographic primitives from a series

of dance motion, Rennhak et al. (2010) used the

acceleration, velocity and power of each joint. To

divide a dance motion with a consideration for the

time series, Masurelle et al. (2013) suggested a

method segmenting the point when footstep impacts

are detected, for the dance which has features in

footsteps like salsa. These methods consider only

motion information. Therefore, these are valid for

the novice dancers’ motions, because dancers who

are not an experimented dancer will not be

accurately synchronized in time with the music.

DanceMotionSegmentationMethodbasedonChoreographicPrimitives

333

However, without incorporating musical

information, when attempting to create a new

motion, the output dance motion is not always

matched to the musical beats.

To segment a dance motion into basic temporal

primitives, Shiratori et al. (2004) used the velocities

of the body parts in coordination with musical

information. This method, succeeded in gaining a

high-accuracy recognition rate of segmentation

points. However, the segmentation rules were made

up for a Japanese traditional dance, and the time

series was not considered. Generally, a dance cannot

always be segmented according to the timing when

the body’s velocities change, but the timing is

decided by the time-series of the choreography.

Therefore, a segmentation method that can

consider the musical tempo and the primitives of a

dance motion and time series of motions has not yet

been developed. Here we assume that a dance

motion is made of segments of choreographic

primitives that are synchronized with music.

Therefore, to accurately segment a dance motion,

considering the primitives of the choreography,

while simultaneously accommodating the musical

tempo would be natural. Owing to this segmentation,

a natural stream of a dance motion is expected to be

created by reusing the existing dance motion data.

3 INVESTIGATION

In this section, we discuss the perception-based

investigation we performed using dancers to

determine choreographic primitives. On the basis of

the results, we defined rules for dance motion

segmentation with a consideration for choreographic

primitives and musical information and improved

the quality of the dance motion composed of the

existing motion data.

3.1 Subjective Experiment

For dance motion animation, we employ 28

character joints, as shown in Figure 1. We use 1

motion data whose length is about 1 minute captured

from a professional female dancer with music

synchronized the motion from the web page

http://perfume-global.com/project.html.

Generally, professional human dancers create the

entire dance motion synchronized to 8 counts (2 bars

of music). Here, we assumed that the boundaries of

the choreographic primitives occur at the musical

beats. We recruited 5 professional dancers(2 men

and 3 women) as participants, and asked them to

Figure 1: Skeletal model based on 28 joints. For the

perception-based investigation, we focused on elbows and

toes (Section4).

Figure 2: The results of the investigation. 5 professional

dancers selected segmentation points from 112 possible

boundary points (i.e., musical beats). The participants A,

B are male, and C, D, E are female.

assign the counts wherein the choreographic

primitive boundaries occurred from the 112 musical

beats. The dance motion is analysed using skinned

mesh animation, because evaluating choreographic

primitives based only on skeletal animation is

difficult.

3.2 Results of the Investigation

Figure 2 shows the results of the subjective

experiment, in which the number of each participant

selecting a particular boundary point (i.e., musical

0 50 100 150

The number of possible boundary

points

(i.e., musical beats)

The answer of each participant

All participants selected

4 participants including hem/her selected

3 participants including hem/her selected

2 participants including hem/her selected

Only this participant selected

This participant did not selected

All participants did not selected

GRAPP2015-InternationalConferenceonComputerGraphicsTheoryandApplications

334

beat) is represented. The results suggest the

existence of an empirical rule for the selection of

choreographic primitives because all participants

selected 25 points as choreographic primitive

boundaries and none of the participants selected 47

points.

In this paper, we focus on some of the most

significant factors for dance motion segmentation,

rather than on the individual variations of the

answers provided. Therefore, we assumed that those

beats selected by all dancers were the accurate

segmentation points representing the choreographic

primitive boundaries.

3.3 Extracting the Important Factors

We examined the tendencies of the segmented points

that all dancers had selected as choreographic

primitive boundaries. Based on this examination, we

determined that the segmentation timing was caused

by the three factors as follows; musical beats,

motion symmetry, and the intervals of footsteps. We

focused on not only music-based detection, but also

the motion features. In doing so, we determined that

two types of choreographic primitives were in

evidence. The first type reflected symmetry

movement, whereas the second type represented

asymmetry movement, as shown in Figure 3. If we

take note of footsteps, each choreographic primitive

consisted of footsteps whose intervals were of the

same length, as shown in Figure 4.

4 DANCE MOTION

SEGMENTATION

Based on the factors indicated by the investigation

(Section 3), we defined the rules for the

segmentation of choreographic primitives. In this

study, the rules consist of the three factors: musical

beats, motion symmetry and the intervals of

footsteps. Then, we suggested a method for

obtaining segmentation points automatically. When

we performed investigation (Section 3), we use

skinned character, though we use stick figures in this

section.

4.1 Musical Beats

A dance motion is generally synchronized with

music; therefore, musical information is essential for

acquisition of choreographic primitives. As observed

in the previous investigation (Section 3), 8 counts of

the musical beats are particularly important because

Figure 3: Examples of symmetrical dance postures (left)

and asymmetrical postures (right) during a dance

performance.

Figure 4: Examples of a single segment. Each footstep has

equal time intervals.

when choreographers create a dance motion, they

are usually aware of the 8 counts. We therefore

calculated the frames of 8 counts, and all the

calculated 8 counts are segmentation points. In this

study, the input music tempo is known.

4.2 Motion Symmetry

To detect segmentation points based on changes

from symmetrical (asymmetrical) to asymmetrical

(symmetrical) motions in a dance motion series,

Figure 5: Detecting symmetrical motions. The red circles

are regarded as false detection because the length of

sequences are less than 7 frames.

DanceMotionSegmentationMethodbasedonChoreographicPrimitives

335

Figure 6: Refinement of the point candidates using

musical beats. Candidate point A and C is shifted to the

nearest musical beat, because these are located within a

specified region near the beats.

knowing whether the motion is symmetrical is, of

course, necessary. However, detecting symmetrical

motions based on posture is difficult because some

motions appear to be symmetrical, but are not

symmetrical in reality. For example, some right and

left legs movements are similar in a time series,

although they are stepping alternately. Then, we

focus on the elbows, because we assume that the

shape of the elbows are important when we judge

the symmetry of figures. Unlike the legs, arms have

many symmetrical movements by frames, and the

degree of freedom is one, except for the twist.

Therefore, detecting symmetrical motions easily is

possible. We calculated the difference between the

degrees of the right and left elbows, 



, which

is given as follows:





 



















(1)

where, 









is the degree of the right elbow and





 is that of the left elbow in the i-th frame. If





 is under a particular threshold, the posture in

the i-th frame is regarded as symmetrical.

To avoid the false detection of symmetry, we

defined a minimum choreographic primitive length

of 7 frames because some frames happened to be

judged as symmetrical on the way of the

choreographic primitive. Setting a minimum length

allows us to better consider the temporal continuity.

In accordance with these considerations, the points

at which a motion changes from symmetrical

(asymmetrical) to asymmetrical (symmetrical) are

regarded as segmentation candidate points. Figure 5

shows an overview of this algorithm.

4.3 Intervals of Footsteps

We detected footsteps on the assumption that if the

length of frames between one footstep and the next

is stable, then these footsteps are closely associated

with a single choreographic primitive.

To detect footsteps, we use the velocities of foot

joints (joints C and D in Figure 1) based on Shiratori

et al.’s (2004) study because foot velocity is nearly

zero when the foot is on the ground. However, using

velocity information alone results in many false

detections, because the motion capture data suffers

from numerous fluctuations

To avoid these errors, we incorporate some

constraints for foot coordinate detection.

 To condition of touching the ground is assumed

when the foot velocity is less than the threshold

value.

 The condition of leaving the grounds is

assumed when the foot acceleration is greater

than the threshold value.

 If the Y coordinate (as shown in Figure 1) is

under the threshold value, the foot is judged as

touching the ground, even if it was judged as

leaving the ground by the other constraint.

In a dance motion, footsteps are usually

synchronized with the musical beats, and we use the

moment when the foot is about to touch the ground

as a candidate frame.

4.4 Refinement of Candidates using

Musical Beats

The candidate segmentation points based on motion

symmetry and footsteps (Sections 4.2 and 4.3) are

not always synchronized with the musical beats.

Therefore, we refine candidate points based on

choreographic primitive boundaries that are

synchronized to musical beats. We employ the

method developed by Shiratori et al. (2004) to shift

the candidate points based on musical beats. If a

musical beat is located within a specified region near

a candidate point based on motion symmetry or

footsteps, we shift the candidate point to the beat.

However, if no musical beats are observed within

the specified region, we regard the boundary point as

a false detection. If no musical beats are observed in

the specified region around the candidate point

based on footsteps, we use the frame as it is.

Thereafter, we calculate the interval of the next

footstep at all the detected frames, if the interval has

the same length as the next interval, the area around

these footsteps is regarded as belonging to a single

segment. Although there are segmentation candidate

points in the area, we revise the points as false

detections. In our method, the length of the specified

region is set to one half of the musical tempo. Figure

6 shows an overview of this algorithm.

GRAPP2015-InternationalConferenceonComputerGraphicsTheoryandApplications

336

5 RESULTS

5.1 Results and Verification of Our

Method

The segmentation candidate points detected by the

musical beats and symmetry movement are shown in

Table 1. The metrics of the evaluation were the

“Recall Rate” and “Precision Rate.” In accordance

with the result of the investigation report in Section

3, the frames on 8 counts are always segmented;

thus, the result of using the musical beats shows

100% precision rate. However, the recall rate is still

low, that is, the choreographic primitives are not

detected.

For the detection of the intervals of footsteps

(Section 4.3), all detected points are successfully

placed in the same segment. Using only footsteps,

some points are not in the same segment. Then, by

prioritizing the musical beats, we succeeded in

improving the segmentation accuracy.

To verify the accuracy of our method, we

compared the results of our method with those of the

two existing methods. One is the most popular

segmentation method for dance motions that uses the

musical information of 4 beats (Kim et al. (2007)

and Fan et al. (2012) used). The other is proposed by

Shiratori et al. (2004) that uses motion velocities and

musical information.

As the dataset, we use the same dance motion

which we used in Section 3 consisting of 25

choreographic primitives (decided by the 5

professional dancers reported in Section 3.2). To

compare the accuracy of the results, we revised the

detected segmentation points by existing methods

using the musical beats. Table 2 shows the

respective segmentation results.

The results indicate that our method attended a

higher accuracy than the existing methods. The

recall rate of our method was 80.0%, which is a

sufficient accuracy for segmentation. Although the

precision rate is still 62.5%, the results of our

method represent an improvement, owing to the

Table 1: Accuracy of the candidate points.

Methodology Recall[%] Precision[%]

Musical 8 beats 60.0 100

Motion symmetry 40.0 40.0

Table 2: Verifying the accuracy of our method.

Methodology Recall[%] Precision[%]

Musical 4 beats 60.0 50.0

Shiratori et al. (2004) 4.0 7.7

Our method 80.0 62.5

consideration of the motion symmetry and footstep

intervals. This will lead to an improved generation

of the new series of a dance motion. The results of

the method proposed by Shiratori et al. (2004)

indicate low recall and precision rates, because the

using motion data was not Japanese dance and the

boundaries of choreographic primitives are not

always placed at the changing points of body

velocities.

When we focus on the false detection of

segmentation points, 12 normal beats were detected,

3 points are one count later than the accurate points,

and 1 point was the timing for which 4 of the 5

dancers had segmented. Few segments whose

lengths are the same as one count are expected to

exist. If two such points reside next to each other,

only one of the two should be selected. Other

detected points that were selected as accurate points

by 5 dancers appear to represent a dancers’

individual judgement.

5.2 Synthesizing Dance Motion using

the Proposed Segmentation Method

We also performed an experiment to verify that our

method is valid for reusing motions through a user-

study, because we assumed that the appropriate

segmentation of the dance motion enables the

creation of a new series of dance motions with high-

quality. We recruited 12 participants consisting of

10 men 20-30 years of age and 2 women 20-25 years

Figure 7: Example of the generation of a new series of dance motion.

You can watch the examples of a new series of dance motions in the demo as the paper attachment.

DanceMotionSegmentationMethodbasedonChoreographicPrimitives

337

Table 3: Results of the evaluation experiment for the

automatic creation of a new dance motions.

Segmentation Method Percentage[%]

Musical 4 beats 33.3

Our method 66.7

of age. We showed them a series of dance motions

formed by connecting the short motion segments

detected by proposed method and musical beats

method wherein segmentation is regularly performed

every four counts. The participants were asked to

select what they considered the better of the two

resulting motions. The segments were selected at

random to do justice to both segmentation methods.

Figure 7 shows as example of a generated new series

of dance motions.

Table 3 shows the results of the experiment. The

results show that the 66.7% of the participants felt

that the dance motion based on the proposed

approach provided more natural dance transitions,

because of the choreographic primitives used.

Some participants indicated dissatisfaction with

the dance motions based on the musical beats

method, for the reason; the character changes to

another motion in the middle of a choreographic

primitive. Therefore, when we generate a new series

of dance motions, the segmentation phase is highly

significant, which reflects the general appreciation

shown in the experiment for the results of our

method.

6 CONCLUSIONS

In this paper, we proposed a segmentation method

for a dance motion that established choreographic

primitives based on an investigation into the

perceptions of actual dancers. We defined

segmentation rules based on the musical beats, the

symmetry of motion, and the timing of footsteps.

Higher accuracy percentages were obtained in our

method than those in existing methods.

In the proposed method, if the music at times

deviates from the standard 8 counts, the existing 8

counts cannot be detected accurately; e.g., in case of

songs that have four counts before their chorus.

However if the timing of the beginning of the chorus

is input, irregular counts can be detected. Therefore,

we intend to apply the proposed segmentation

method to any genre of music. Though if the music

has except for 4 counts in its bar, the way of making

the choreography may be different. To apply our

method for these music is our future work.

The segments were sorted randomly when

comparing the new series of dance motions

generated by the dance segmentation methods

considered. If we incorporate sorting rules,

generating more natural dance transition is possible.

On a future work, we plan to construct sorting rules

based on similarities of posture and to synchronize a

dance motion with the atmosphere of the input

music. Finally, we intend to construct an automatic

dance motion generation system based on the

accurate segmentation of the dance motion data

collected from the Internet.

REFERENCES

Alankus, G., Bayazit, A. A., and Bayazit, O. B., 2005.

Automated Motion Synthesis for Dancing Characters.

Computer Animation and Virtual Worlds, 16(3-4),

259-271.

Arikan, O., and Forsyth, D., 2002. Interactive Motion

Generation from Examples. ACM Transaction on

Graphics, 21(3): 483-490.

Barbic, J., Pan, J. Y., Faloutsos, C., Hodginsj, K., and

Pollard, N., 2004. Segmenting motion capture data

into distinct behaviors. In Proc. of Graphics Interface,

2(6): 185-194.

Beaudoin, P., Coros, S., Panne, M., and Poilin, P., 2008.

Motion-motif graphs. In Proceedings of the ACM

SIGGRAPH/Eurographics Symposium on Computer

Animation, 117-126.

Fan, R., Xu, S., and Geng, W., 2012. Example-based

Automatic Music-Driven Conventional Dance Motion

Synthesis. IEEE Transactions on Visualization and

Computer Graphics, 18(3): 501-515.

Kim, G., Wang, Y., and Seo, H., 2007. Motion Control of

a Dancing Character with Music. In Proceedings of

the 6th IEEE/ACIS International Conference on

Computer and Information Science (ICIS), 930–936.

Kim, T.H., Park, S.I., and Shin, S.Y., 2003. Rhythmic

Motion Synthesis based on Motion-Beat Analysis.

ACM Transaction on Graphics, 22(3): 392-401.

Kovar, L., Gleicher, M., and Pighin, F., 2002. Motion

Graphs. ACM Transaction on Graphics, 21(3): 473-

482.

Laban, R., and Ullmann, L., 1971. The Mastery of

Movement. Pre Textos.

Lee, M., Lee, K., and Park, J., 2013. Music Similarity-

based Approach to Generating Dance Motion

Sequence. Multimedia Tools and Applications, 62(3):

895-912.

Masurelle, A., Essid, S., and Richard, G., 2013.

Multimodal Classification of Dance Movements using

Body Joint Trajectories and Step Sounds. In Proc. 14

International Workshop on Image and Audio Analysis

for Multimedia Interactive Services (WIAMIS), 1-4.

GRAPP2015-InternationalConferenceonComputerGraphicsTheoryandApplications

338

Nakata, T., Mori, T., and Sato, T., 2002. Analysis of

Impression of Robot Bodily Expression. The Journal

of Robotics and Mechatronics, 14(1): 27-36.

Osaki, R., Shimada, M., and Uehara, K., 2000. A Motion

Recognition Method by Using Primitive Motions.

VDB 5 Proceedings of the Fifth Working Conference

on Visual Database Systems: Advances in Visual

Information Management, 117-128.

Rennhak, B., Shiratori, T., Kudoh, S., Vinayavekhin, P.

and Ikeuchi, K., 2010. Detecting Dance Motion

Structure Using Body Components and Turning

Motions. In Proc. IEEE/RSJ International Conference

on Intelligent Robots and Systems (IROS 2010), 2264-

2269.

Shiratori, T., Nakazawa, A., and Ikeuchi, K., 2004.

Detecting Dance Motion Structure through Music

Analysis. In Proc. IEEE International Conference on

Automatic Face and Gesture Recognition (FGR 2004),

857-862.

Shiratori, T., Nakazawa, A., and Ikeuchi, K., 2006.

Dancing-to-Music Character Animation. Computer

Graphics Forum, 25(3): 449-458.

Vögele, A., Krüger, B., and Klein, R., 2014. Efficient

unsupervised Temporal Segmentation of Human

Motion. In Proceedings of the ACM SIGGRAPH /

Eurographics Symposium on Computer Animation.

Zhou, F., Torre, F., and Hodgins J. K., 2008. Aligned

cluster analysis for temporal segmentation of human

motion. In IEEE Conference on Automatic Face and

Gestures Recognition. 1(2).

Zhou, F., Torre, F., and Hodgins J. K., 2013. Hierarchical

aligned cluster analysis for temporal clustering of

human motion. IEEE Trans. on Pattern Analysis and

Machine Intelligence (PAMI). 35(3): 582-596.

DanceMotionSegmentationMethodbasedonChoreographicPrimitives

339