Publication - The feasibility of gaze tracking for “mind reading” during search

1

The feasibility of gaze tracking for “mind reading”
during search

Andreas Lennartz & Marc Pomplun

Department of Computer Science, University of Massachusetts at Boston,

100 Morrissey Blvd., Boston, MA 02125, USA

We perform thousands of visual searches1,2 every day, for example, when selecting

items in a grocery store or when looking for a specific icon in a computer display3.

During search, our attention and gaze are guided toward visual features similar to

those in the search target4-6. This guidance makes it possible to infer information

about the target from a searcher’s eye movements. The availability of compelling

inferential algorithms could initiate a new generation of smart, gaze-controlled

interfaces that deduce from their users’ eye movements the visual information for

which they are looking. Here we address two fundamental questions: What are the

most powerful algorithmic principles for this task, and how does their

performance depend on the amount of available eye-movement data and the

complexity of the target objects? While we choose a random-dot search paradigm
for these analyses to eliminate contextual influences on search7, the proposed

techniques can be applied to the local feature vectors of any type of display. We

present an algorithm that correctly infers the target pattern up to 66 times as often

as a previously employed method and promises sufficient power and robustness for

interface control. Moreover, the current data suggest a principal limitation of

target inference that is crucial for interface design: If the target patterns exceed a

certain spatial complexity level, only a subpattern tends to guide the observers' eye

movements, which drastically impairs target inference.

2

Eye movements can reveal a wealth of information about the content of a person’s

visual consciousness, such as the current interpretation of an ambiguous figure8 or the

geometry of mental imagery9. During visual search, our eye movements are attracted by

visual information resembling the search target4-6, causing the image statistics near our

fixated positions to be systematically influenced by the shape10 and basic visual

features5 of the target. One study found that the type of object sought, of two possible

categories, can be inferred from such systematic effects 11. If such inference were

possible for a larger set of candidate objects, a new generation of smart, gaze-controlled

human-computer interfaces12 could become reality. Gaining information about an

interface user’s object of interest, even in its absence, would be invaluable for the

interface to provide the most relevant feedback to its users.

To explore the potential of algorithmically inferring the search target from a

searcher’s eye fixations, we conducted two experiments of visual search in random-dot

patterns (see Fig. 1). Subjects searched a large random-dot array for a specific 3×3

pattern of squares in two (Experiment 1) or three (Experiment 2) luminance levels while

their eye movements were measured. Our aim was to devise algorithms that received a

subject’s gaze-fixation positions and the underlying display data and inferred the actual

target pattern with the highest possible probability. Fixation and display data from the

actual target pattern in the search display was excluded, because the disproportionate

fixation density at the end of a search would have made target inference trivial. A

variety of inferential algorithms (see Methods) was devised and tuned based on ten

subjects’ gaze-position data and evaluated on another ten subjects’ data for each

experiment. The current paradigm was well-suited for a first quantitative exploration of

this field, because it minimized the influence of semantic factors7 on eye movements

and supplied fixed numbers of equally probable target patterns, 29 = 512 in Experiment

1 and 39 = 19683 in Experiment 2. At the same time, this paradigm challenged the

algorithms to the extreme, not only due to these huge numbers of target candidates, but

3

also because they were not shown as discrete objects but formed a contiguous pattern

whose elements barely exceeded the spatial resolution of the eye-tracking system.

Our development and evaluation of inferential algorithms resulted in the

discovery of two particularly powerful mechanisms, whose combination outperformed

all other methods for both Experiments 1 and 2 without modifying its parameters

between experiments. In the following, we will describe these two components and

compare their performance with an approach adapted from a previous study10. In that

study10, the statistical distribution of display luminance in a window centered on a

subject’s fixation positions was measured and in some cases found to roughly resemble

the search target. To apply this method to the current task, for every fixation, we

extracted the display data from a 3×3-square window whose center square was placed

over the square on which the fixation landed. We computed the frequency of each

feature (black, gray, and white) in each square and subtracted the average frequency of

that feature across the nine squares. The feature with the highest value in each square

entered the estimated target pattern. This algorithm, which we termed ‘gaze-centered

feature map,’ outperformed all other methods analyzing feature statistics in individual

squares relative to fixation.

Our first newly developed technique, ‘pattern voting,’ is based on the assumption,

derived from a previous study6, that the strongest attractors of observers’ eye

movements during search are local patterns that are very similar to the search target. We

operationally defined the similarity between two 3×3 patterns as the number of

matching features in corresponding squares, resulting in a range of similarity values

from zero to nine. The voting algorithm keeps score of the votes for every possible 3×3

pattern. For each fixated square, a 3×3 window is placed over it nine times so that each

of its squares lands on the fixated square once. Each time, the patterns whose similarity

to the pattern in the window is eight (high-similarity patterns) receive one vote.

4

Identical patterns (similarity nine) do not receive votes for the benefit of a ‘fair’

evaluation, since neither the actual target nor the fixations on it are visible to the

algorithm. The pattern receiving the most votes is the estimated target pattern.

Interestingly, placing only the window center over fixated squares or weighting

this center position more heavily leads to reduced performance of the voting algorithm.

While this effect may partially be due to noise in gaze-position measurement, it is also

possible that subjects do not always fixate on the center of a suspected target.

Depending on how they memorize the target, their gaze may be attracted by a specific

position within similar patterns – a ‘gaze anchor’ position from where they compare the

local pattern with the memorized one. If we could estimate the most likely gaze anchor

positions, we could improve the pattern voting algorithm by assigning greater weights

to the votes received at the corresponding window positions relative to fixation. These

window positions should be indicated by greater consistency of their high-similarity

patterns, that is, stronger preference of some patterns over others. Our analyses showed

that the most effective weights are obtained by computing separately for the nine

window positions the votes for individual patterns as above, divide them by the average

number of votes for that position, and apply an exponent. The final score for a pattern is

the sum of its weights across the nine positions, and the highest score determines the

estimated target pattern. The exponent, which rewards high frequencies of patterns in

specific positions, should increase when more gaze samples are provided in order to

exploit the greater signal-to-noise ratio. The final ‘weighted pattern voting’ algorithm

computes the final score sn for pattern n as follows:

 V 
ln  e + r 
R
 ⋅ v r ,n   c 
sn = ∑    for n = 1,..., , (1)
r =1  Vr  

where is the total number of patterns (512 or 19683 in this study), R is the

number of distinct window positions relative to fixation (here, R = 9), vr,n is the number

5

of votes given by the pattern voting algorithm to pattern n in window position r, Vr is

the sum of votes for all patterns in r, and c is a constant whose optimal value was found

near 600 for both current experiments.

To evaluate the performance of the algorithms as a function of the number of

available search fixations, we resampled those fixation datasets that were not used for

developing the algorithms, that is, we repeatedly selected random subsets of them. Fig.

2 illustrates that pattern voting clearly outperforms the gaze-centered feature map. In

Experiment 1, even after only 20 fixations (about 5 s of search), the voting algorithm’s

probability of picking the correct target is already 18.6 times above chance level, while

it is only 0.2 times above chance for the feature map. After approximately 180 fixations,

the weighted pattern voting starts surpassing the basic voting algorithm and maintains a

steeper increase until the final 1800 fixations, where its performance reaches 31.9%,

outperforming the voting algorithm (22.5%), p < 0.01, which in turn exceeded the

performance of the gaze-centered feature map (0.5%), p < 0.001. This pattern is similar

in Experiment 2 (0.139%, 0.095%, and 0.023%, respectively, for 1800 fixations, both ps

< 0.05) but the superiority of the pattern voting algorithms over the feature map

approach is less pronounced. Fig. 3 illustrates the top ranking choices made by the

weighted pattern voting algorithm.

Even if we compensate for the difference in pattern set size, weighted pattern

voting still performs clearly better in Experiment 1 than in Experiment 2, as indicated

by greater performance-to-chance level proportion (163.5 versus 27.4, respectively),

and sensitivity d’ (2.54 versus 0.92, respectively) according to signal detection theory13

(see Methods) , p < 0.01, for 1800 fixations. If the reason for this discrepancy were

poorer memorization of the more complex target patterns in Experiment 2 and, as a

result, greater noise in the guidance of eye movements, then subjects should detect the

target less often than they do in Experiment 1. However, the mean target detection rate

6

is 43% in Experiment 1 and 47.3% in Experiment 2. Another possible explanation is

that the higher target complexity leads to subjects’ eye movements being guided by only

a part of the target pattern, and whenever this part is detected, a complete verification of

the local pattern is conducted. To test this hypothesis, we used resampling (1800

fixations) to rank all 2×2 patterns according to their frequency of being fixated, and

calculated the probability that any of the four 2×2 subpatterns of the target (see Fig. 4a)

was the top-ranked one. While the absolute hit rate does not differ statistically between

Experiments 1 and 2 (68.1% versus 51.6%, respectively), p > 0.3, both the hit rate-to-

chance level proportion (2.72 versus 10.44, respectively) and sensitivity d’ (0.65 versus

1.19, respectively), are greater in Experiment 2, p < 0.01, supporting our hypothesis

(Fig. 4b).

The present data suggest that the mechanisms underlying the weighted pattern

voting algorithm are robust enough for a useful target estimation in a variety of human-

computer interfaces. The proposed mechanisms can be adapted to various display types,

since image filters commonly used in computer vision14 and behavioral studies5,15 can
transform any display into a matrix of feature vectors. Moreover, the current data

advocate that the future designers of smart, gaze-controlled human-computer interfaces

should keep the spatial complexity of display objects low in order to induce more

distinctive patterns of eye movements for individual search targets.

Methods

Experiments. The same twenty subjects, aged 19 to 36 and having normal or corrected-

to-normal vision, participated in each experiment after giving informed consent. Their

eye movements were measured using an EyeLink-II head mounted eye tracker (SR

Research, Mississauga, Canada) with an average accuracy of 0.5° and a sampling rate of

500 Hz. At the start of each trial in Experiment 1, subjects were presented with their

7

search target - a 3×3 array of squares (width 0.6° of visual angle), each of which was

randomly chosen to be either black (1.2 cd/m2) or white (71.2 cd/m2) . In Experiment 2,

a third luminance level (gray, 36.2 cd/m2) was added. Subjects had six seconds to

memorize this pattern before it was replaced with the search display consisting of 40×40

squares of the same size and luminance levels as those in the target. Each search display

contained the target pattern exactly once (see Fig. 1). Subjects were instructed to find

the target as quickly as possible, then fixate on it and press a designated button to

terminate the trial. If the distance between gaze position and target object during the

button press was less than 1°, successful target detection was counted. If no response

occurred within 60 s after the onset of the search display, the trial also terminated. In

each experiment, every subject performed 30 trials during which the search target

remained the same.

Algorithms. Several other algorithms were implemented, their parameters and

components fitted – including modifications for gaze-anchor estimation leading to

relative performance gains of up to 72% - using ten subjects’ data and evaluated on the

other ten subjects’ data. These algorithms and their performance based on 1800

fixations in Experiments 1 and 2, respectively, were: Pattern voting with votes weighted
by similarity (30.9% and 0.122%), pattern voting weighted by fixation duration (24.1%

and 0.115%), pattern voting with lower (7) similarity threshold (21.5% and 0.102%),

Bayesian inference based on similarity metric (18.5% and 0.098%), 2×2 subpattern

voting (9.2% and 0.115%), 3×1 and 1×3 subpattern voting (7.1% and 0.103%), feature

map based on most frequently fixated patterns (6.5% and 0.08%), voting based on

feature correlation between neighboring squares (6% and 0.093%), and average

luminance in gaze-centered window (0.26% and 0.0069%).

Sensitivity computation. To compare the inferential performance of algorithms

between decision spaces of different sizes, we employed the sensitivity measure d’ for

8

the situation in which a technical device or human observer has to make a choice among

a known number of alternatives13,16. Although this measure assumes independence of

signals, which is not warranted in the present scenario, it provides a useful

approximation that has been applied to similar problems before13. In the subpattern

analysis (Fig. 4), we further make the simplifying assumption that all subpatterns of a

target are fixated with the same probability.

1. Wolfe, J.M. in Attention, H. Pashler, Ed. (Psychology Press, Hove, UK, 1998), pp.

13-71.

2. Najemnik, J. & Geisler, W.S. ature 434, 387-391 (2005).

3. Hornof, A. J. Human-Computer Interaction 19, 183-223 (2004).

4. Wolfe, J.M. Psychon. Bull. Rev. 1, 202-238 (1994).

5. Pomplun, M. Vision Res. 46, 1886-1900 (2006).

6. Shen, J., & Reingold, E. M. In Proc. of the Twenty-First Annual Conf. of the Cog.

Sci. Soc., M. Hahn & S. C. Stoness, Eds.(Erlbaum, Mahwah, NJ. 1999), pp. 649-652.

7. Henderson, J.M. & Hollingworth, A. Ann. Rev. Psych. 50, 243-271 (1999).

8. Pomplun, M., Velichkovsky, B.M. & Ritter, H. Perception 25, 931-948 (1996)

9. Mast. F.W. & Kosslyn, S.M. Trends in Cog. Sci.6, 271-272.

10. Rajashekar J., Bovik, L.C. & Cormack, A.K., J. of Vision 6, 379–386 (2006).

11. Zelinsky, G., Zhang, W., & Samaras, D. [Abstract]. J. of Vision 8, 380.

12. Sears, S. & Jacko, J.A. The Human-Computer Interaction Handbook:

Fundamentals, Evolving Technologies, and Emerging Applications (CRC Press,

Lincoln, USA, 2003).

9

13. Macmillan, N.A. & Creelman, C.D. Detection Theory: A User’s Guide (Cambridge

University Press, New York, 1991).

14. Gonzalez, R.E. & Woods, R.C. Digital Image Processing (Prentice Hall, Upper

Saddle River, 2002).

15. Zelinsky, G. J. Psych. Rev. 115, 787-835 (2008).

16. Hacker, MJ, & Ratcliff, R. Perception & Psychophysics 26, 168-170 (1979).

Acknowledgement. The project was supported by Grant Number R15EY017988 from the National Eye

Institute to M.P.

Correspondence and requests for materials should be addressed to M.P. (marc@cs.umb.edu).

10

Fig. 1. Search targets (left) and cut-outs from corresponding visual search

displays with a human subject’s scanpath superimposed on it. Actual displays

consisted of 40×40 squares. Red discs indicate fixation positions, consecutive

fixations are connected by straight lines, and the initial fixation is marked with a

blue dot. A green square indicates the position of the target in the search

display. a, Experiment 1; b, Experiment 2.

Fig. 2. Comparison of inferential performance of the gaze-centered feature map

algorithm adapted from a related study10 and the two pattern voting algorithms

proposed in the current study. Performance is measured as the probability of

correctly inferred target objects as a function of the number of gaze fixations

provided to the algorithms. This probability was approximated by repeated

resampling (20,000 and 100,000 times for Experiments 1 and 2, respectively) of

subjects’ fixation data. Notice that the number of potential target patterns is 512

in Experiment 1 and 19683 in Experiment 2. a, Experiment 1; b, Experiment 2.

Fig 3. Actual targets (green frame) and the three patterns ranked highest by the

weighted pattern voting algorithm in Experiment 1 (left) and Experiment 2 (right)

based on all recorded fixations. Actual target objects appearing in the first three

ranks are marked by red frames. While all target patterns in Experiment 1

occupy either rank one or two (out of 512 candidates), the average rank of the

target patterns in Experiment 2 is 1514 (out of 19683 candidates).

Fig. 4. Analysis of target subpattern frequencies near fixation. a, Each target is

decomposed into four 2×2 subpatterns. b, Probability of any of the four target

subpatterns to receive the most fixations among all 2×2 subpatterns (16

patterns in Experiment 1 and 81 patterns in Experiment 2). Error bars indicate

standard error of the mean across ten subjects.

Publication - The feasibility of gaze tracking for “mind reading” during search

More Related Content

What's hot (18)

Similar to Publication - The feasibility of gaze tracking for “mind reading” during search (20)

More from A. LE (9)

Recently uploaded (20)

Publication - The feasibility of gaze tracking for “mind reading” during search