Keywords

1 Introduction

A revolution in manufacturing is approaching, and it could change the style of fabrication. Extracting affects in personal fabrication will become indispensable in enhancing the recent advances in the field, because we can provide that information to the fabricators and let the fabricators enjoy the experience. In this study, we proposed an extraction method of affect in personal fabrication based on Laban Movement Analysis (LMA) [1] by using the motion data of fabricators.

2 Related Studies

Many studies on extracting an affect in various situations have been performed, but most of them limit the target situations. This means that such studies are not sufficiently suitable for personal fabrication, where we use various tools and various movements are dependent on the tools.

LMA is known as a method for interpreting human movement. It is a tool used by dancers and actors. Moreover, it can be applied to movement generation of CG characters and robots. For example, Chi et al. have developed EMOTE, which is a 3D character animation system based on LMA [2]. Nakata et al. have adopted the method for generating the motions of a robot and analyzed its impressions [3]. Here, we should note that both studies are applied to generating motions.

However, though LMA was proposed as a method to extract affects from body movement, it has been difficult. LMA was composed of several characteristic elements, such as Space (Direct/Indirect), Weight (Light/Strong), and Time (Sudden/Sustained), although they were not defined numerically. Recently, as the technology on motion sensors has shown progress, research has been conducted on analyzing motions and computing the elements [4]. In addition, some methods of extracting affects are reported [5, 6]. However, they could not be applied to personal fabrication.

3 Experiment

We conducted an experiment on personal fabrication before proposing an extraction method because we do not have data sets of motions and affects. Here, we performed an experiment by using the electronic building blocks termed littleBits. We asked participants to arrange their original synthesizer in pairs by using littleBits. We shot the motions of participants by using a motion capture system (Bonita 10, Vicon) and video camera (HVR-A1J, SONY), as shown in Fig. 1. We also measured the heart rate of the participants (WHS-2, UNION TOOL). The participants were six male and six female Japanese students.

Fig. 1.
figure 1

Example scene of making synthesizer

Full size image

We classified affects by using the affect grid by Russel [7], and asked participants to recall and select a suitable affect and its intensity (5 grades) after fabrication. Figure 2 shows an example of the result. We classified the results into strong and weak based on the average intensity. The blue bar shows a strong affect by one of the participants.

Fig. 2.
figure 2

An example of affect (Color figure online)

Full size image

4 Affect Extraction Method

4.1 Amount of Characteristics of Laban Movement Analysis

We then defined the amount of characteristics of Laban Movement Analysis, Space, Weight and Time as shown in Fig. 3. Space was computed by the area between the head and both wrists. Weight was the vertical position of head. Time was the sequence of the speed of the wrists for 60 s.

Fig. 3.
figure 3

The amount of characteristics for Laban Movement Analysis; Space, Weight and Time

Full size image

4.2 Estimation of Interval of Strong Affect

First, we estimated the interval of a strong affect. Here, we used the parameters of 4.1 and heart rate. We introduced a decision tree by using the value of Space, Weight, Time, and heart rate for each person. For each value, we normalized all the parameters in advance. We used Weka and J48 decision tree. As a result, the estimation of the strong interval of affects was 71.3 % of recall (Fig. 4).

Fig. 4.
figure 4

Example of estimation of strong affect

Full size image

4.3 Classification of Affect

We then classified the affects by using a decision tree based on Space, Weight and Time for each participant. Figure 5 shows an example of the tree. As shown in the figure, we could classify different motions at the same affect. As a result, the average of the correctly classified instances was 58.1 %. In addition, 46.2 % of incorrect results were observed when the affect of the participants changed. This means that the correctly extracted affect would be approximately 80 %.

Fig. 5.
figure 5

An example of the decision tree

Full size image

Table 1 shows the result of the classified affect. As shown in the table, concentration and boredom could be classified at approximately 70 %. However, classifying surprise was difficult.

Table 1. Classified affect by proposed method
Full size table

5 Estimation by Human

As a reference, we performed an experiment for affect estimation by human participants. Here, we asked observers to classify the affect of participants, as shown in Fig. 6. Ten Japanese students participated in the experiment.

Fig. 6.
figure 6

Experimental scenery

Full size image

As a result, the correct rate was 21.1 %, which was lower than that of the proposed method. Table 2 shows the result of the classified affects. As shown in the table, humans could classify pleasant or unpleasant affects. However, it was difficult to distinguish between activated and deactivated in Russel’s affect grid.

Table 2. Classified affect by human
Full size table

6 Summary

We proposed a novel affect extraction method in personal fabrication based on Laban Movement Analysis by using the motion data of fabricators. We proposed a method for extracting affects by defining the amount of characteristics of LMA, and by introducing a decision tree. The method was relatively simple, and the results of the experiments demonstrated the effectiveness of the method.