Results

Results relating to gaze direction identification and subjective ratings were analysed independently.

Gaze Direction Identification

A repeated measures two-way ANOVA was performed on the gaze direction identification data, with the four oculesic conditions and the twelve gaze angles (numbers on clock face) as factors. This overall measure combined data from all three cameras, and indicated a main effect of both oculesic behaviour (F(3,11)=4.19;P<0.01), and gaze angle (F(3,11)=11.26; P<.001). Post-hoc Tukey tests determined that the eyelid animation and vergence condition performed significantly poorer than the classes of: eyelid animation, no vergence(F(1,11)=11.73;P<.001) and no eyelid animation, no vergence (F(1,11)=4.03; P<.005). An interaction effect was found between oculesic condition and gaze angle during the latter comparison (F(1,11)=2.02;P<.05). The column entitled ’All Cameras’ in Table7.1shows this overall decrease in accuracy of identification of gaze direction as the behavioural fidelity of the agent increases.

7.1. Oculesic Behaviour Experiment 183 Table 7.1: Mean (and standard deviation) gaze direction identification accuracy for conditions and cameras.

Condition All Cameras Centre Camera Side Cameras

No Eyelid Animation, No Vergence 80.7% (0.72) 89.2% (0.53) 76.5% (0.79) No Eyelid Animation, Vergence 79.4% (0.73) 92.2% (0.43) 73.0% (0.83) Eyelid Animation, No Vergence 78.9% (0.58) 93.8% (0.24) 71.4% (0.67) Eyelid Animation, Vergence 73.0% (0.64) 89.8% (0.54) 64.7% (0.67)

The most basic oculesic behaviour is seen to perform best, but the difference is only significant compared to the eyelid animation and vergence condition. This corresponds to the above ANOVA calculations, suggesting that, when analysed on this macro scale considering ratings from all three cameras, the in- clusion, and in particular the combination, of the two oculesic cues of eyelid animation and vergence are detrimental to observers’ ability to accurately judge gaze direction, thereby opposing part of the stated hypothesis.

The influence of camera angle on the accuracy of gaze direction identification was then explored. As shown in Table7.1, the centre camera formed one group, and the two symmetrical side cameras were evaluated together. Starting with the centre camera, a repeated measures two-way ANOVA was per- formed with the four oculesic conditions and the twelve gaze angles as factors. A main effect was found between the oculesic conditions (F(3,11)=3.26; P<.05), and an interaction effect was found between factors (F(3,11)=1.73; P<.01). Post-hoc Tukey tests exposed that the differences lay between the eyelid animation, no vergencecondition and two other classes: eyelid animation and vergence (F(1,11)=7.64; P<.01) and no eyelid animation, no vergence (F(1,11)=6.99; P<.01). The same tests were then per- formed on the data for the two side cameras, again finding a significant difference between oculesic conditions (F(3,11)=4.02; P<.01). Similar to the overall camera analysis, differences were found be- tween the class of eyelid animation and vergence and the two other classes of eyelid animation, no vergence(F(1,11)=6.92; P<.01) and no eyelid animation, no vergence (F(1,11)=11.83; P<.001). An interaction effect was found during the latter comparison (F(1,11)=1.98; P<.05). Table7.1highlights consistently high accuracy from the centre camera over the four conditions, and lower, more variable accuracy, when judging from the side cameras.

The above analysis shows that accuracy of gaze direction identification is influenced by the agent’s actual gaze direction, in combination with the observing camera angle. Therefore, analysis between symmetrical gaze angle ‘pairs’ was performed in order to isolate scenarios and expose specific combi- nations of gaze direction and camera angle with high impact. The experimental design of a clock face with one central and two symmetrical side cameras granted identical states for opposite gaze angles ac- cording to the side camera considered. For instance, a one o’clock gaze viewed from the right camera is equivalent to an 11 o’clock gaze from the left camera. Accordingly, 2 left = 10 right, 3 left = 9 right, 4 left = 8 right, 5 left = 7 right. Gaze angles 6 and 12 were treated independently, as the pair is not equivalent. These logical pairings of symmetrical gaze angles allowed direction of observed gaze to be classified more generally as toward and away: terms more meaningful to interaction in AMC as demon- strated during gaze analysis in the telecommunication experiments documented in Chapters4–6. In this

7.1. Oculesic Behaviour Experiment 184

Figure 7.6: Overall combined condition gaze identification accuracy from centre (C), toward (T), and away (A) cameras for symmetrical gaze angle pairs. Note that the 0◦/180◦ vertical pair (6 and 12) remain separate, as the agent eye representation differs greatly.

context, toward gaze indicated that the agent was looking in the hemisphere in which the observer lies, and away gaze indicated that direction of gaze was in the opposite direction of the perspective of the ob- server. Prior to this analysis, the validity of the pairings was confirmed by performing repeated measures two-way ANOVA on the data for the two gaze angles within each pair, and no significant differences were found. Overall accuracy levels for gaze angle pairs are shown in Figure7.6for centre, toward, and away perspectives. Certain combinations of gaze angle and camera angle show a significant and negative impact. In particular, 30◦offset from vertical (pairs 1/11, 5/7) when observers are positioned in the toward-hemisphere, and the 60◦offset from vertical (pairs 2/10, 4/8) when observers are positioned in the away-hemisphere.

Subjective Rating

The questionnaire sought to elicit judgements regarding the agent’s realism and participants’ ratings of self-performance during the gaze direction identification task for each condition. Question 1 (“The be- haviour of the avatar’s eyes appeared natural”) and Question 4 (“The general appearance of the avatar was realistic”) directly addressed perception of realism. A critical requirement of the experimental de- sign was to isolate animation to the eyes and surrounding areas. Thus any variation in reported levels of perceived realism could be related solely to changes in oculesic representation. A one-way ANOVA evaluation of combined data from Questions 1 and 4 exposed a highly significant difference between conditions (P<.001). Post-hoc Tukey tests exposed differences to lie between the two classes featuring eyelid animation, and the two that did not, (P<.001). Table7.2presents the Likert scale responses, showing how the inclusion of eyelid animation is able to enhance perceived realism, while vergence was

7.1. Oculesic Behaviour Experiment 185 Table 7.2: Mean response and (standard deviation) for Questions 1, 4 and 5 on the 1..7 (nega- tive..positive) Likert scale questionnaire.

Condition Question 1 Question 4 Question 5

No eyelid animation, no vergence 3.75 (1.48) 4.29 (1.32) 4.87 (1.24) No eyelid animation, vergence 3.52 (1.53) 4.35 (1.38) 4.53 (1.36) Eyelid animation, no vergence 5.69 (1.06) 5.98 (1.13) 5.42 (0.82) Eyelid animation, vergence 5.33 (1.17) 5.85 (1.14) 5.16 (1.10)

Figure 7.7: Detailed screen captures of the agent’s eyes when looking at number 5 (150◦offset from vertical gaze angle 12) from the toward (right) camera. Top: no eyelid animation, no vergence. Bottom: eyelid animation and vergence.

not a determining factor. Responses to Question 4 also demonstrate the influence of eyelid animation on the agent’s holistic realism. There was no significant difference found between the two classes featuring eyelid animation (P=.14), adding to the evidence that vergence was not a factor affecting the agent’s realism. Hence, the subjective component of the original hypothesis is supported by findings related to the impact of eyelid animation, but not supported by findings related to vergence.

One-way ANOVA evaluations of responses to Questions 2 and 3 (“I could easily tell where the avatar was looking from the centre view / side views”) did not expose significant differences between conditions despite significant differences in accuracy found during analysis of gaze direction identifica- tion data. However, a significant difference was found between conditions for Question 5 (“How well do you think you completed the identification task?”) (P<.05), with eyelid animation conditions earning superior ratings. This suggests that confidence in identifying the agent’s direction of gaze was raised by the enhanced realism afforded by the addition of eyelid animation, despite actual gaze identification accuracy not always corresponding to this perception. An example rendering of the agent’s eyes in the eyelid animation, vergencecondition and the no eyelid animation, no vergence condition is presented in Figure7.7.

7.1.5

Discussion

The experiment presented in this section explored how varying conditions of an autonomous agent’s oculesic behaviour is seen to influence both accuracy of gaze direction identification and perceived real-

7.2. Eyelid Kinematics Models and Experiments 186