Introduction

The focus of this project is to grasp a better understanding of how the human visual system works. More specifically, the question is, "Why does change blindness occur in certain situations?" In order to tackle that question, a distinction must be made between visual search for change and change blindness. Visual search for change is the process of looking for a change in scenery when one knows there is a change. Change blindness is the inability to detect a change when one is not searching for a change. Even with this distinction in mind, many different attributes can be responsible for change blindness.

The hypothesis to be proven in this research project is that the presence or absence of an object is detected diffrently than the movement of an object in a visual search for change experiment.

Table Of Contents


Background

There were several experiments done on change blindness in the past 5 years. Prior to these experiments, little has been done to research this phenomenon. Each experiment discovered a different approach to observing change blindness. Of the experiments done, only a handful have utilized eye tracking as a means of experimenting.

Flicker Paradigm

One of the most important experiments done on change blindness was about interruptions between changed images. The experiment setup was simple. An original picture would be displayed on a screen and alternated with a similar image with one obvious change from the original. In between the succession of images, a gray blank screen would be placed to disrupt the clean transition from one image to another. Subjects were asked to then report when they detected a change between the two images. The images were alternated as many times as were needed for the subject to find the change. By varying the times of the gray disruption image, a trend was found showing that most subjects of the experiment were poor at detecting changes when the interstimulus interval was greater than 70 to 100 msec. [4] The interstimulus interval is the amount of time the disruption image is shown.

Diagram 1: Flicker Paradigm Sequence

Diagram 1 above shows the transition form one image to another. This experiment, involving a disruption image, is often known as the flicker paradigm. In addition to the flicker paradigm, other factors need to be taken into consideration.

 

 

Types of Change

The type of change in a pair of images is also important when discussing change blindness. In general, there three major types of changes that can be done to an image. [2] All three types of change are illustrated in diagram 1.

Diagram 2: Appearance/Disappearance

Diagram 2 shows the first type of change, which is an appearance/disappearance change. Oddly enough, the size of the change does not always matter.

Diagram 3: Color Change

The tv's color change in diagram 3 is another example of a type of change.

Diagram 4: Position Change

Diagram 4 shows the last kind of change, which is position change. The dish in the book shelf has been slightly shifted over. The magnitude of any kind of change is not the main factor in change detection, however.[4]

Central and Marginal Interest

The main factor in change detection is the amount of attention one pays to an object. Different scenes have different objects with different amounts of attention being drawn to them. A quantitative method for measuring an objects interest has not yet been created. Experimentally. however, one can determine in specific scenes, what objects are of peak interest and which are not. Several subjects can be given the task of describing a picture. Objects mentioned by most of the subjects can be considered Central Interest (CI) objects and objects not mentioned by most of the subjects can be considered Marginal Interest (MI) objects.[4]

Mudsplashes

Simliar to the flicker paradigm, 'mudsplashes' are a form of transient, which also deter one's ability to detect change. This experiment was done in the exact same fashion as the flicker paradigm experiment with the exception of only blocking small parts of an image instead of the entire image.[1]

Diagram 5: Sample Mudsplash

Diagram 5 shows the disruption image that replaces the gray image in the flicker paradigm. The result of this change is amazingly the same result. The squares, however, were only placed on top of locations where a change did not occur. Apparently, the boxes are enough of a distraction to prevent a subject from detecting change in the remainder of the image. [1]

Research related to Specific Aim

The interesting part of change blindness is that one can look at an object, but not see what that object is. [5] Attenion seems to be intimately related with change blindness.

The introduction of an eye tracker to change blindness is not a new idea. Eye trackers are usually not necessary for most change blindness experiments. Subjects in most experiments are told to simply speak when they have detected a change. The eye tracker can be utilized, however, to determine what a subject is looking at and what they are percieving simultaneously. Verbal confirmation of change detection in conjunction with eye tracking a subject can achieve this task.

Table Of Contents


Experimental Design and Methods

Design

The experiment was designed as a visual search for change experiment. Although the subject is aware that a change is present, the subject exhibits change blindness until the object being changed is detected. The hypothesis of this experiment is that there is some difference between the way we look at different kinds of changes. More specifically, appearance/disappearance and movement changes will be targeted.

The experiment itself consists of 18 pairs of images. A pre-experiment was held to determine approximately equivalent difficulties in images. The trial types were dispersed evenly throughout the experiment. The stand-alone eye tracker was used in conjunction with a computer to collect a data stream of pixel locations. In addition, a Hi-8 video recorded the experiments for further analysis.

Subjects were asked to volunteer for this experiment, and were offered candy as a reward. All subjects ranged from ages 13 to 23. 11 of the 12 subjects were college students, and 9 of those 11 college students were imaging science students.

Subjects were told that an image pair was going to be presented to them with a short flash of a gray screen in between. The two similar images within the pair would cycle back and forth, one at a time. Their task was to verbally describe a single change when they detected it. The subjects were also told that the orientation of the images would not be a factor to detect. When the subject detected the correct change, the experimenter would stop the trial and then a new pair would be presented until all 18 trials were completed.

The experiment was written in Flash v5.0 for reasons of mac and pc compatibility. Variables were set to count the number of alternations back and forth of the image pairs individually. At the end of the experiment, the list of "flips" were displayed on the screen next to their corresponding trials. The experiment can be found online by clicking on the TRY ME link on the sidebar to the left.

Chart 1: Types of Trials

Movement
Short Descriptions
Trial 1
Apartment with right side windows moving left and right
Trial 2
A male student with their arms up in the air and a BBQ moving in the bottom right corner
Trial 4
Apartment with pair of top windows moving up and down
Trial 7
A female student sitting at a computer with a speaker moving from one monitor to another
Trial 8
A store Window with the % sign moving from one number to another
Trial 10
Yellow cone with red and black lettering switching places
Trial 13
Fire extinguisher moving up and down
Trial 14
CIS coffee room with a basket moving left and right
Trial 16
A kids train ride with a moving plaque on the left
   
Absence/Presence
Short Descriptions
Trial 3
A parking lot where the cement parking blocks disappear
Trial 5
The CIS building with disappearing windows on the right
Trial 6
An empty fountain at the mall with disappearing plants at the top of the image
Trial 9
A guitar with disappearing fret markers (white dots on neck of guitar)
Trial 11
Two street cones with disappearing arrows
Trial 12
A mall directional sign with disappearing lights on the left
Trial 15
A kids ride area with a white sign disappearing on the post in the middle
Trial 17
Friendly's with a large disappearing window on the right
Trial 18
Apartment (identical to trial 1) with a disappearing walkway

Chart 1 is a full listing of all the image pairs used in the experiment. The dispersion of the different types of trials was relatively uniform.

 

Methods of Analysis

The data collected consisted primarily of two types. The first type was the non-tracking data. Non-tracking data was the data collected by writing down the number of iterations the viewing program cycled through. Tracking data consisted of coding the Hi-8 video and analyzing the data stream. The data stream analysis was used souly for cosmetic purposes such as displaying a track for a poster. The Hi-8 video coding is the major source of information to be used for this experiment. Observations of the Hi-8 video were also invaluable in understanding the search pattern.

Coding the Hi-8 video was the most time consuming of all the analysis. The coding procedure consisted of creating a ruler to measure the distance in degrees of vision, and going through each trial and measuring the last 5 eye movements with the ruler. To create the ruler a simple triangular equivalence method was used comparing the large display tv and the small coding tv. The idea behind the measurement of the last 5 eye movements of a trial was that a low level mental process may be persent in determining the location of the change and that a pattern of eye movement might be present there, if anywhere at all.

Table Of Contents


Results

Chart 2: Alternations per Trial with Standard Error Bars

Chart 2 shows the number of alternations for each image pair for all 12 subjects. The blue bars on the left are the averages, while the red bars on the right are the medians. The marks on the tips of each bar represent the standard errors for each trial. The trial numbers are sequential and correspond directly with the Flash version of the experiment. The trials were reordered in the graph as to put similar types of trials together as indicated in Chart 1, respectively. Most of the averages are higher than their corresponding medians, indicating that a few outliers may have skewed the averaged data. The 'move' and 'absence' trials appear at first glance to be relatively equivalent. To further analyze this hypothesis, an unpaired two-tailed Student t-test was done on the data. A two-tailed test was chosen, because no predictions of the greater distribution could be made. The data points were also not correlated on a one-to-one basis, so the unpaired test seemed appropriate. The level of significance was 0.05. The data sets were approximately normally distributed.

The result that the t-test gave was a t-value of 0.02820 and a two-tail t-critical value of 2.1314. The null hypothesis of the t-test was that the 'move' and 'absence' groups were identically distributed. Since the t-value was less than the critical value, the null hypothesis was accepted, meaning that there are no statistically significant differences (at the 95% confidence level) between the move and absence groups.

Figure 1: Sample Point Plot of a Trial

Figure 1 shows the data stream collected from a subject on a sample trial. In this trial, the red and black letterings were switched. A concentration of eye movements are fixated on the cone in the image. Note that the subject looked at the object, but did not notice that it was changing on the first pass. Information like Figure 1 is useful for determining where the marginal and central interest objects are located in the image. Otherwise, Figure 1 has very little use other than for visualizing an entire trial at once. Time cannot truly be measured by the system, nor can a beginning or end be determined.

Graph 1: Distance from Last Eye Fixation versus Time Prior to Last Eye Fixation for Move Trials

 

Graph 2: Distance from Last Eye Fixation versus Time Prior to Last Eye Fixation for Absent Trials

Graphs 1 and 2 show the results of the subjects last 5 eye fixations compared to the time since the last fixation. A time of zero, correlates to a distance of zero. Each color cloud represents a set of data points generated by a combination of all the subject's fixations with relation to time. Each cloud of data points was linearly regressed to determine the slopes and y-intercept (i.e. distance at time = 0 ) of each trial. A linear regression was used instead of other regressional methods because the data seemed too scattered to have any other pattern.

Chart 3: Slopes and Intercepts of Linearly Regressed Data Points

Trial
Slope
Intercept
 
Trial
Slope
Intercept
move 1
0.0012
2.38
 
absent 1
0.0002
3.2566
move 2
0.0001
6.4613
 
absent 2
0.00003
3.1009
move 3
0.001
7.3805
 
absent 3
0.001
3.2262
move 4
0.0015
4.4134
 
absent 4
0.0015
0.7265
move 5
0.0006
2.67
 
absent 5
0.001
1.7977
move 6
0.0007
0.7365
 
absent 6
-0.0001
7.3663
move 7
0.0007
2.8945
 
absent 7
-0.0004
6.6692
move 8
0.00005
3.5637
 
absent 8
0.00001
2.5047
move 9
0.0014
3.6056
 
absent 9
0.0003
4.8461

 

The slopes and y-intercepts of the move and absent trial groups were then compared to see if there were any differences between them. The hope was to see a distinct difference between the groups that could possibly distinguish them. An unpaired Student's t-test was used to try to make this distinction. The hypothesis to test was that the 'move' slopes were equivalent to the 'absent' slopes. The result of this t-test was a t stat value of 1.51638 and a 2-tailed t-critical value of 2.13145. Although this result still proves that the data sets are similar, the t-value is not very far from the t-critical value. The result of the unpaired t-test on the y-intercepts was a t-value of 0.0678 and a 2-tailed t-critical value of 2.1199. This result is fairly clear as the t-value is nowhere near the t-critical value. Both t-tests were done with a significance level of 5%.

Table Of Contents


Discussion

The results of the experiment showed no significance between absence changes and location changes. The two-tailed t-statistic was successful in determining that the two groups were not statistically different from each other.

The only questionable result was the two-tailed t-test result of the slopes of the two groups. The statistic showed that they were still roughly equivalent, but did so on a borderline area of the t-test. The closeness of the t-value to the t-critical value means that within some kind of error, the t-test could have gone either way. This result probably means that more data should be collected to determine a more concrete finding.

On the other hand, a statistical difference in slope may not be a very significant find. A steep slope means that a large distance from the change was covered in a short amount of time. In many of the images however, the change was not limited to a single point, but to a single set of features. This may have been a flaw in the design of the experiment, because the distance to the last fixation point was not always accurate in determining whether or not a subject was taking into account the full change or only part of the change. As a result, distances that appear farther may actually be the same object being first detected. Further experiments may be needed to test this hypothesis.

The amount of concentration needed to detect change seems very high. Several subjects had trouble verbalizing the object being changed. Almost every subject stuttered when they discovered a change that had been stumping them. The fire extinguisher, for instance, was referred to as a fire hydrant on two accounts. This seems to verify the link between attention span and change blindness. No data other than observations were collected on this phenomenon.

Table Of Contents


Conclusion

Absence and movement changes have no significant statistical differences within a 5% confidence interval. The experimental design was not tailored enough to meet the specific needs of the experiment, but did prove to have some useful information. The hypothesis of this experiment was rejected. A possible differentiation may have been found in the move and absence groups through slope categorization of the linear regression. Further experiments may be needed to make this differentiation, if it even exists at all.

Improvement on this experiment could be made in several areas. One change in the experiment would be the creation of images with only single objects being moved, not features. By creating feature changes such as disappearing pavement blocks, a measurement of distance is very difficult to establish. The problem is that there is no one real focus of the change, and so finding a point to measure to from the eye position is difficult. Another useful change to this experiment would be coding not only the distance, but also the direction to the change. By coding the direction to the change, one could visualize the search pattern and see if there were any tendencies towards looking in a particular direction between 'flips'. Despite these changes, the experiment was still successful in attempting to find a difference between the two experimental groups.

In future experiments, several things could be investigated. Memory, for instance, seems to be very good for retaining the changes in the image pairs. A more focused experiment in memory retention and change blindness may have some interesting results. Another experiment that may be interesting is a threshold of change experiment. A movement experiment, for example, could be designed to make very subtle changes between the image pair until an object is in a completely different location, like a movie with interruptions.

 

Table Of Contents