• Ei tuloksia

3 Methods

4.3 Study 3: Do small changes in instructions change the way people seek

4.3.3 Analyses

4.3.3.1 Defining semantic regions of interest (ROIs) and areas fixated

The semantic ROIs were defined from the eye movements during the memory task. The fixated areas were estimated from a fixation distribution map, as described in Chapter 3.6. The cut-off point for the z-axis of the resulted fixation density map (FDM) was 0.25, which our qualitative examination of the distributions showed to be the value that best suited the most images. The areas fixated were calculated for each image across the observers in a similar manner from both quality and difference-estimation tasks, with a cut-off point of 0.02.

64 4.3.3.2 Defining salient areas

The low-level salient areas were defined using Walter and Koch’s (2006) Saliency Toolbox 2.2 (http://www.saliencytoolbox.net/ downloaded in July 2011). It is based on the modelling work of Itti and Koch (2000) and calculates the saliency map for the image using information on contrast, colour and orientation weighted with the winner-take-all maps. We used the toolbox’s default settings and because our images contained humans we added a skin-colour feature with the weight of one. The areas with positive values were considered salient.

4.3.4 Results

There were no differences between the groups in the estimations of magnitude the participants gave for the images (Wald χ²(1)=1.0, p>0.05). However, there were differences in viewing strategies (Table 5 presents the main results). The participants engaged in the difference-estimation task looked at the images for a longer time (Wald χ²(1)=9.2, p<0.01) and needed more fixations (Wald χ²(1)=7.2, p<0.05) than those doing the quality-estimation task. The average fixation durations per image per participant did not differ according to the task (Wald χ²(1)=1.1, p>0.05), but there was an interaction between the task and the content (Wald χ²(5)=19.8, p<0.01) as well as between the task and the manipulation (Wald χ²(6)=13.6, p<0.01). This indicates that the influence of a task becomes visible in fixation duration only if the type of test material is taken into account.

We further examined the duration of the first fixations, which are used to plan subsequent eye movements throughout the scene (Castelhano & Henderson, 2007) and in this case when the content is known they yield information about the processing of the first actively chosen fixation point (Holmqvist et al., 2011).

The first fixations were longer in the quality task than in the difference task (Wald χ²(1)=7.5, p<0.01), meaning that planning where to look next took longer in the former than in the latter. The average saccades amplitude per image was also longer in the difference task (Wald χ²(1)=8.7, p<0.01), and the task-content interaction was significant (Wald χ²(5)=15.4, p<0.01). Therefore, the fixations were further apart in the difference task than in the quality task, and there was less detailed examination of one area with repeated fixations.

65

Table 5. The medians of the variables describing strategy in the quality and difference tasks, as well as the significance of the between-task comparisons

Quality

ns = non-significant, *=p<0.5, **= p<0.01, ***=p<0.001

We also examined the spatial distribution of the fixations. First, a larger area was fixated on in the difference task than in the quality task (Wald χ²(1)=232.2, p<0.001). The medians of the areas fixated on were 26.1 and 15.2 per cent in the difference and quality tasks, respectively. The fixations in the quality task covered 16.7 per cent of the image area on average, and in comparisons between the two tasks, only 4.1 per cent of the area fixated on in the quality-estimation task was not fixated on also in the difference task. It seems that those engaged in the difference-estimation task fixated on the same important areas as those engaged in the quality task, but these areas were not enough: a further large area was needed to cover the search in the difference task. Next we analysed what kinds of areas were fixated on in these different tasks.

To define the types of areas in the different image contents we calculated the salient areas from the low-level image features as well as the semantic regions of interest (ROIs) from the eye movements recorded in the memory task. In these analyses we estimated the group-level differences in the areas fixated on. The salient areas were widely distributed across the images, and the areas considered salient differed depending on the contents. In the content “woman”, which depicts a woman sitting by a table with many different objects around her, only 6.4 per cent of the image area was considered salient. The corresponding figure for the content “scenery”, which depicts a nature scene with water, forest, rocks and sky, was 10.4 per cent, the largest proportion of all the contents. Similarly,

66

the size of the semantic ROIs depended on the content: the semantic ROIs covered 6.1 per cent of the image area in the portrait of a boy, compared with 41 per cent in a busy image of an outside cafeteria. Therefore, the areas considered salient did not vary in size according to the image contents as much as the areas considered semantically important.

The time spent looking at these areas differed between the tasks. A higher proportion of time was spent looking at semantic ROIs in the quality task than in the difference task (Wald χ²(1)=251.0, p<0.001), and vice versa in the salient areas (Wald χ²(1)=4.3, p<0.05) (Figure 11). The proportion of time spent fixated on a certain area also depended on the interaction between the task and the contents (semantic ROIs: Wald χ²(5)=118.2, p<0.001; salient areas: Wald χ²(5)=52.6, p<0.001). The biggest between-task differences in attention allocation concerned contents with strong attention attracters, such as faces or large areas considered semantically important (the content “cafeteria”). The information from the strong attention attracters seemed to be enough in the quality-estimation task, whereas attention was also actively allocated outside this area in the difference task. It therefore seems that semantically important image areas are more important in the quality-estimation than in the difference task.

Such areas are fixated on in the latter task, as is a large area in addition.

Figure 11. The average proportions of time spent on semantic ROIs (a) and salient areas (b) of all the time spent looking at images per image content.

67

To shed further light on the relationship between salient areas and semantic ROIs we examined their common relationship and the areas fixated on. The salient areas that were fixated on were often also within the semantic ROIs (Figure 12), the proportions falling outside being only 1.7 per cent in the quality task and 3.5 per cent in the difference task. The area that was both salient and in the semantic ROIs comprised less than five per cent of the whole image area on average, but it nevertheless accounted for 13.8 per cent of the fixations in the quality task and 12.0 per cent in the difference task (Figure 12a). It thus seems that the salient areas of an image are important only if they are also semantically important. This supports the notion that saliency models work, because most objects are salient (Einhäuser, Spain, et al., 2008).

Figure 12. The proportions (a) and the numbers (b) of fixations on different image areas for the two tasks

The results show that a small change in the instructions influences viewing behaviour, even when comparing magnitude-estimation tasks. The viewing times were shorter, and the viewing concentrated more on the semantically important areas in the quality-estimation than in the difference-estimation task. The semantically important areas were attended to in the difference task as well, but this information was not sufficient and other large areas were also fixated on. The salient areas were important if they were also semantically important.

68

4.4 Study 4: Are individual differences in viewing behavior related to different estimation rules in a quality-estimation task?

The findings from Studies 1 and 2 revealed that, depending on the material, people use different decision-making rules when estimating quality. It was shown in Study 3 that the instructions influence viewing behaviour even in two quite similar magnitude-estimation tasks. In the case of high-quality images quality estimation tends to be a preference task and therefore subjective. Subjectivity means that there are individual differences, which have been linked to the use of different deduction rules (Kruglanski & Gigerenzer, 2011). We posited that quality estimation in the case of high-quality images could be a task in which subjective differences are related to different deduction rules. We further assumed that estimation rules on which quality estimations are based could be used to access deduction rules. Here, we used here the term estimation rules to refer to the set of attributes on which people base their estimations.

Individual differences in eye movements have also been reported (Andrews &

Coppola, 1999; Boot et al., 2009; Castelhano & Henderson, 2008; Rayner et al., 2007). We wanted further to find out whether such individual differences are related to differences in estimation rules. We conducted two experiments to examine this relationship. In the first one we used the IBQ method, to elicit estimation rules: the method allows people to describe freely on what they base their preference estimations. The aim of second experiment was to confirm the results of the first using a larger image set and a different method of measuring the estimation rules. The participants estimated how important the different attributes collected in Experiment 1 were for their preference estimation. Both experiments involved the recording of eye movements and analysis of the relationship between individual viewing tendencies and estimation rules.