PLEASE SCROLL DOWN FOR ARTICLE This article was downloaded by: [Huff, Markus] On: 28 September 2010 Access details: Access Details: [subscription number 927315895] Publisher Psychology Press Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House, 37- 41 Mortimer Street, London W1T 3JH, UK Visual Cognition Publication details, including instructions for authors and subscription information: http://www.informaworld.com/smpp/title~content=t713683696 Eye movements across viewpoint changes in multiple object tracking Markus Huff a ; Frank Papenmeier a ; Georg Jahn b ; Friedrich W. Hesse a a Knowledge Media Research Center, Tübingen, Germany b Department of Psychology, University of Greifswald, Greifswald, Germany First published on: 01 September 2010 To cite this Article Huff, Markus , Papenmeier, Frank , Jahn, Georg and Hesse, Friedrich W.(2010) 'Eye movements across viewpoint changes in multiple object tracking', Visual Cognition, 18: 9, 1368 — 1391, First published on: 01 September 2010 (iFirst) To link to this Article: DOI: 10.1080/13506285.2010.495878 URL: http://dx.doi.org/10.1080/13506285.2010.495878 Full terms and conditions of use: http://www.informaworld.com/terms-and-conditions-of-access.pdf This article may be used for research, teaching and private study purposes. Any substantial or systematic reproduction, re-distribution, re-selling, loan or sub-licensing, systematic supply or distribution in any form to anyone is expressly forbidden. The publisher does not give any warranty express or implied or make any representation that the contents will be complete or accurate or up to date. The accuracy of any instructions, formulae and drug doses should be independently verified with primary sources. The publisher shall not be liable for any loss, actions, claims, proceedings, demand or costs or damages whatsoever or howsoever caused arising directly or indirectly in connection with or arising out of the use of this material.
25
Embed
Eye movements across viewpoint changes in multiple object tracking
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
PLEASE SCROLL DOWN FOR ARTICLE
This article was downloaded by: [Huff, Markus]On: 28 September 2010Access details: Access Details: [subscription number 927315895]Publisher Psychology PressInforma Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK
Visual CognitionPublication details, including instructions for authors and subscription information:http://www.informaworld.com/smpp/title~content=t713683696
Eye movements across viewpoint changes in multiple object trackingMarkus Huffa; Frank Papenmeiera; Georg Jahnb; Friedrich W. Hessea
a Knowledge Media Research Center, Tübingen, Germany b Department of Psychology, University ofGreifswald, Greifswald, Germany
First published on: 01 September 2010
To cite this Article Huff, Markus , Papenmeier, Frank , Jahn, Georg and Hesse, Friedrich W.(2010) 'Eye movements acrossviewpoint changes in multiple object tracking', Visual Cognition, 18: 9, 1368 — 1391, First published on: 01 September2010 (iFirst)To link to this Article: DOI: 10.1080/13506285.2010.495878URL: http://dx.doi.org/10.1080/13506285.2010.495878
Full terms and conditions of use: http://www.informaworld.com/terms-and-conditions-of-access.pdf
This article may be used for research, teaching and private study purposes. Any substantial orsystematic reproduction, re-distribution, re-selling, loan or sub-licensing, systematic supply ordistribution in any form to anyone is expressly forbidden.
The publisher does not give any warranty express or implied or make any representation that the contentswill be complete or accurate or up to date. The accuracy of any instructions, formulae and drug dosesshould be independently verified with primary sources. The publisher shall not be liable for any loss,actions, claims, proceedings, demand or costs or damages whatsoever or howsoever caused arising directlyor indirectly in connection with or arising out of the use of this material.
Eye movements across viewpoint changes in multiple
object tracking
Markus Huff and Frank Papenmeier
Knowledge Media Research Center, Tubingen, Germany
Georg Jahn
Department of Psychology, University of Greifswald, Greifswald, Germany
Friedrich W. Hesse
Knowledge Media Research Center, Tubingen, Germany
Observers can visually track multiple objects that move independently even if thescene containing the moving objects is rotated in a smooth way. Abrupt scene rota-tions yield tracking more difficult but not impossible. For nonrotated, stable dy-namic displays, the strategy of looking at the targets’ centroid has been shown to beof importance for visual tracking. But which factors determine successful visualtracking in a nonstable dynamic display? We report two eye tracking experimentsthat present evidence for centroid looking. Across abrupt viewpoint changes, gazeon the centroid is more stable than gaze on targets indicating a process of realigningtargets as a group. Further, we show that the relative importance of centroid look-ing increases with object speed.
Downloaded By: [Huff, Markus] At: 05:14 28 September 2010
different camera angles. Hence, viewers have to keep track of the attended
players and the ball across abrupt viewpoint changes. In this study, we show
using a multiple object tracking (MOT) paradigm that gaze behaviour across
filmic cuts relies on mechanisms ensuring visual stability. When trackingthree objects across abrupt viewpoint changes, observers focus on a virtual
object (such as the invisible centre of mass of three objects) that includes as
vertices the tracked targets in the scene (Yantis, 1992). Most important, the
time spent looking at targets decreased after a viewpoint change, whereas
gaze time on the centroid remained more constant in total duration thr-
oughout the trial. Centroid looking that was proven as more important at
higher object speeds might enable humans to successfully track multiple
objects across abrupt viewpoint changes.Research on visual attention showed that observers can visually track
multiple objects that move independently (e.g., Alvarez & Franconeri, 2007;
Cavanagh & Alvarez, 2005; Pylyshyn & Storm, 1988). In standard multiple
object tracking displays, the frame or background around the moving ob-
jects remains stable throughout the entire trial. More recently, visual track-
ing was examined in nonstable scenes (Huff, Jahn, & Schwan, 2009; Huff,
Meyerhoff, Papenmeier, & Jahn, 2010; Liu et al., 2005; Seiffert, 2005). The
scene containing the objects remained no longer stable on the display butunderwent changes like rotations, zooms, or translations. It was shown, that
continuity of both object and scene motion seems to support visual tracking.
Abrupt changes such as translations of the whole scene or abrupt rotations
impair tracking performance considerably (Huff et al., 2009; Seiffert, 2005).
However, tracking performance was still above chance level after discontin-
uous translations and even if the reference frame was rotated abruptly by
308. Because in these conditions retinocentric coordinates suddenly change,
targets have to be recollected by means of scene-based coordinates. A group-ing strategy such as centroid looking codes a target’s location in relation to
other targets and may support tracking even if the scene as a whole moves
and retinocentric coordinates are lost.
How does the visual system manage to keep track of multiple objects?
First, human observers can only gaze at one object at a time. When tracking
more than one object there are at least two reasonable strategies: Target
jumping (Landry, Sheridan, & Yufik, 2001) or looking at a location within
the target group that minimizes the distance to each target. Recently, severalstudies examined gaze behaviour in MOT (Fehd & Seiffert, 2008; Zelinsky
& Neider, 2008; see also Doran, Hoffman, & Scholl, 2009). Those studies
manipulated the number of targets in either a standard MOT task (Fehd
& Seiffert, 2008) or in tracking sharks in a more realistic underwater scene
(Zelinsky & Neider, 2008). In both studies, MOT stimuli were presented in
a frame that remained stable throughout the trial and eye tracking results
were compared between conditions with different numbers of targets. In the
EYE MOVEMENTS ACROSS VIEWPOINT CHANGES 1369
Downloaded By: [Huff, Markus] At: 05:14 28 September 2010
Gaze behaviour. Eye tracking data in previous studies (Fehd & Seiffert,
2008; Zelinsky & Neider, 2008) were analysed with the so-called ‘‘shortest
distance rule’’, which assigns each gaze to its nearest object. This method is
based on the assumption that every gaze during a trial is directed towards anobject. This is fine for stable scenarios but becomes problematic if the scene
moves abruptly. In that case, not every gaze may be directed towards an
object. In fact, it seems very likely that in trials with abrupt scene motion,
some eye movements are not directed towards objects, at least in a short
interval after abrupt scene motion. Thus, we applied an alternative method
of coding and analysing eye movements. We defined every moving object
and the moving centroid of the target objects as separate areas of interest
(AOI). Only looking on the objects or the centroid was defined as a match.Eye movements not related with a defined AOI were coded as a miss.
As we were interested in how human observers manage keeping track of
multiple targets across abrupt viewpoint changes, we selected the trials, in
which all targets were identified correctly. Based on these criteria, 31.9% of
all trials were discarded. The 20 repetitions per condition ensured that at
least five trials per condition entered the final analysis for each participant.
Each moving object and the targets’ three-dimensional centroid were defined
as separate dynamic AOIs subtending 1.38 to 2.28 of visual angle dependenton their location on the floor plane. Overall, 43.5% of gaze was explained by
the a priori defined AOIs (28/s: 46.6%, 48/s: 41.9%, 68/s: 42.0%; see Figure 2).
Every 20 ms (50 Hz) we registered whether the participant’s gaze matched an
AOI. Those matches were summed up for each AOI (three targets, seven
distractors, and the targets’ centroid) in 500 ms intervals. The measure ‘‘pro-
portion matched’’ was calculated by dividing the sum of valid matches
per AOI by the total amount of possible matches. Finally, averages were
calculated for each AOI type, if a type consisted of more than one AOI. Forinstance, the proportions of gaze positions matching single distractors
were averaged to yield the mean proportion of gaze positions matching
a distractor.
Figure 2 shows the mean proportions of gaze positions matching the
AOI types separately for the 500 ms intervals in each condition. To test for
differences in gaze behaviour across the intervals statistically, we calculated
separate one-way ANOVAs including the within-subjects factor ‘‘interval’’
for each AOI-type (targets, distractors, and centroid) in each condition aswell as a planned contrast (t-test) between the 500 ms intervals on either side
of the abrupt viewpoint change in the viewpoint change conditions. The
results of these analyses are displayed in Table 2. In all conditions without
a viewpoint change, gaze behaviour did not change significantly across trial
intervals, FsB1.35, ps�.217, hp2sB.09. With viewpoint changes, gaze be-
haviour showed a drop and recovery of gaze on targets but not on the
targets’ centroid.
1374 HUFF ET AL.
Downloaded By: [Huff, Markus] At: 05:14 28 September 2010
Independent of object speed, viewpoint changes caused a significant dec-
line of gaze on target AOIs, F(9, 126)�3.00, p�.003, hp2�.18, F(9, 126)�
2.44, p�.013, hp2�.17, F(9, 126)�3.32, p�.001, hp
2�.19 for 2, 4, and 68/s
Figure 2. Time course analysis based on 500 ms intervals in Experiment 1. Results show stable gaze
allocation in the 08 viewpoint change condition. In the 208 viewpoint change conditions, significantly
less gaze on the targets was observed in the interval right after the abrupt viewpoint change. Most
importantly, gaze on the centroid dropped less than gaze on targets.
EYE MOVEMENTS ACROSS VIEWPOINT CHANGES 1375
Downloaded By: [Huff, Markus] At: 05:14 28 September 2010
object speed, respectively. This decline was followed by a significant increase
in the second interval after the viewpoint change up to the level before the
viewpoint change, pairwise t-tests pB.05, The time spent gazing on the cen-
troid differed less between the intervals before and after the viewpoint
change with 2 and 48/s object speed, F(9, 126)B1, p�.889, hp2�.06, and
F(9, 126)�1.29, p�.250, hp2�.08, respectively. At 68/s object speed, we
observed higher variability of gaze on the centroid resulting in a significant
main effect, F(9, 126)�3.19, p�.002, hp2�.19. Pairwise t-tests, however,
evinced that gaze on the centroid was not influenced by the abrupt viewpoint
change. Instead, significant differences were observed between intervals
4 and 5 as well as between intervals 9 and 10 (pairwise t-tests pB.05).Further analysis examined the total amount of gaze on centroid and tar-
gets in Experiment 1. First, we observed a significant interaction between
AOI type (targets and centroid) and object speed, F(2, 28)�10.41, pB.001,
hp2�.42. Planned post hoc analysis showed that centroid looking clearly
grew in importance with increasing object speed, F(2, 28)�9.00, pB.001,
hp2�.39, whereas higher object speed led to a decline of gaze on targets,
TABLE 2ANOVA results of the time course analyses of the data plotted in Figure 3; t-tests
reflect planned contrasts between 500 ms intervals before and after the abruptviewpoint change in Experiment 1
Downloaded By: [Huff, Markus] At: 05:14 28 September 2010
F(2, 28)�7.40, p�.002, hp2�.35 (see Figure 3). Linear trend analyses
confirmed this finding, F(1, 14)�18.49, p�.001, hp2�.57, and F(1, 14)�
10.79, p�.005, hp2�.44, respectively.
Taken together, results from the analysis of gaze behaviour across trial
intervals showed that gaze behaviour is stable across successful track-
ing trials in conditions without abrupt viewpoint changes. Viewpoint chan-
ges caused a decline of gaze on targets but not on the centroid following
the viewpoint change. Centroid looking was observed at all object
speeds and became more important with increasing object speeds as
hypothesized.
A large proportion of gaze was directed towards targets and the centroid.
In contrast, looking towards distractors was rare. In the conditions with
Figure 3. Gaze on target and centroid AOIs in Experiment 1. With increasing object speed, gaze on
the centroid increased while gaze on targets decreased. Error bars indicate the standard error of the
mean.
EYE MOVEMENTS ACROSS VIEWPOINT CHANGES 1377
Downloaded By: [Huff, Markus] At: 05:14 28 September 2010
viewpoint change, gaze on the targets decreased significantly in the interval
right after the viewpoint change, whereas gaze on the centroid did not. After
these 500 ms, gaze behaviour was the same as before the viewpoint change.
While reorienting and realigning the targets, participants look at the cen-
troid of the target group.
Centroid looking as observed with stable scenes (Fehd & Seiffert, 2008;
Zelinsky & Neider, 2008) was replicated in the present experiment with scenes
undergoing abrupt viewpoint changes. Data from Experiment 1 suggest gaze
to be more stable on the centroid than on target objects. However, we
observed high variability of gaze on the centroid in the higher object speed
conditions. One reason for this increased variability could be that the
exclusion criterion (only trials with perfect tracking performance were inclu-
ded in the analysis) was not strict enough. Maybe participants sometimes
tracked only two targets and guessed the third correctly by chance. This
might have happened more frequently at higher object speeds and could
have increased gaze variability. In the second experiment we added a con-
fidence rating asking participants if they were able to track all objects. Only
trials in which participants tracked all objects correctly and in which
they stated that they were able to track all objects were included in the
analysis. Higher spatial and temporal resolutions as provided by the
SMI iView X Hi-Speed eye tracking system deployed in Experiment 2
should reduce gaze variability further. Furthermore, the improved temporal
resolution allowed us to look at saccade measures to further clarify our
results.
Another improvement in Experiment 2 concerns the predictability of the
viewpoint change. Centroid looking right after the viewpoint change could
be the result of an anticipation strategy. As the centroid is relatively stable (it
moves slower and its degree of movement is restricted) and the viewpoint
change occurred after 3 s in each trial, it might be possible that participants
directed their gaze strategically towards the centre of the targets right before
the viewpoint change. Such a strategy would not reflect spontaneous gaze
behaviour. Hence, we varied the viewpoint change time randomly to prevent
anticipation.
EXPERIMENT 2
Our objective in Experiment 2 was to replicate the findings of Experiment 1
for a wider range of object speeds. We manipulated object speed in two levels
(4 and 108/s). Additionally, we varied the time point of the abrupt viewpoint
change (2, 3, 4 s after the beginning of the scene) in order to prevent
anticipation. Gaze behaviour was recorded with a SMI iView X Hi-Speed
eye tracking system.
1378 HUFF ET AL.
Downloaded By: [Huff, Markus] At: 05:14 28 September 2010
Method
Participants. Eighteen students of the University of Tubingen (13 female,
five male; mean age�24.2 years) participated in this experiment and rep-
orted normal vision. They received compensation for their participation.
Apparatus. Stimuli were created the same way as in Experiment 1 and
were presented via a 19-inch monitor in a 31.68 (wide)�25.58 (high) viewable
area at a resolution of 1280�1024 pixels. Eye movements were recorded
with a SMI iView X Hi-Speed eye tracking system providing high spatial and
temporal resolutions (500 Hz). Mean calibration accuracy was 0.228 for the
x-coordinate and 0.298 for the y-coordinate. Gaze data were prepared with
DynAOI, a tool designed for matching gaze data to a priori defined AOIs in
3-D scenes (Papenmeier & Huff, 2010). A chin- and headrest made sure that
the participants kept a distance of 66 cm from the display.
Stimuli and procedure. Stimulus material was similar to Experiment 1
with four exceptions. First, object speed was either 4 or 108/s. Second, in
order to prevent anticipation of the viewpoint change, we added some
uncertainty with respect to the time point of the abrupt viewpoint change.
Viewpoint changes occurred 2, 3, or 4 s following the beginning of object
motion. Third, the training phase was two-tailed. In the first part consisting
of 24 trials, participants were given feedback about their tracking per-
formance. In the second part, consisting of four trials, participants had to
answer the following question as confidence rating: ‘‘Did you manage to
track all targets correctly?’’ Participants used the mouse to click on the
corresponding answer: ‘‘Yes’’ or ‘‘no’’. Fourth, in the experimental phase
participants were not provided with any feedback. Instead, they were asked
to answer the confidence rating after each trial. Taken together, each
participant performed 120 experimental trials in a 2 (4 vs. 108/s object
speed)�2 (with vs. without viewpoint change)�30 repetitions per condition
within-subjects design.
Results and discussion
Tracking performance. Results were similar to Experiment 1. Tracking
performance as shown in Table 3 declined with increasing object speed,
F(1, 17)�207.61, pB.001, hp2�.92, and was lower with viewpoint changes,
F(1, 17)�54.76, pB.001, hp2�.76. The interaction of object speed and
viewpoint change was not significant, F(1, 17)�1.87, p�.189, hp2�.10.
Gaze behaviour. Compared to Experiment 1, we tightened the selection
criterion for the gaze behaviour analysis. We selected the trials, in which all
EYE MOVEMENTS ACROSS VIEWPOINT CHANGES 1379
Downloaded By: [Huff, Markus] At: 05:14 28 September 2010
targets were identified correctly and participants stated that they were able to
keep track of all targets in the confidence rating. The 30 repetitions per
condition ensured that at least three trials per condition entered the final
analysis for each participant*except for the viewpoint change condition
with 108/s object speed. In this specific condition, we had to exclude four
participants from analysis because they had less than three trials with correct
tracking performance and perfect confidence. In sum, we discarded 58.1% of
all trials (25.4% in the no viewpoint change condition with 48/s object speed,
55.9% in the viewpoint change condition with 48/s object speed, 67.0% in
the no viewpoint change conditions with 108/s object speed, and 83.9% in
the viewpoint change condition with 108/s object speed). Overall, 40.8% of
gaze was explained by the a priori defined AOIs (48/s: 42.9%, 108/s: 35.8%;
see Figure 4).
In Experiment 1, viewpoint changes occurred after 3 s in each trial. In
Experiment 2, the time points of the viewpoint change varied between 2 and
4 s after the beginning of the trial. Hence, we cannot analyse gaze behaviour
across the entire tracking interval. Therefore, we analysed gaze behaviour
across the 500 ms intervals on either side of the viewpoint change plus the
second 500 ms interval after the viewpoint change (see Figure 4 and Table 4).
We calculated separate within subjects ANOVAs with the independent
variable ‘‘time interval’’ and the dependent measure proportion matched for
each AOI type in each condition. In all conditions without a viewpoint
change, gaze behaviour did not change significantly across trial intervals (all
hp2sB.16; see Table 4). For conditions with viewpoint changes, we were able
to replicate the finding in the condition with 48/s object speed. Here, gaze
behaviour showed a similar drop and recovery of gaze on targets as in
Experiment 1, F(2, 34)�9.60, pB.001, hp2�.36. In contrast, gaze on the
centroid was not affected by the viewpoint change, F(2, 34)B1, p�.440,
hp2�.05.
We observed different gaze behaviour in conditions with 108/s object
speed. Here, gaze on both centroid and target AOIs was not influenced
across the three time intervals (all hp2sB.02). As tracking three objects
moving in straight lines at 108/s is very error prone (we had to discard
TABLE 3Tracking performance in Experiment 2 (with SD in parentheses)
Speed Viewpoint change Proportion correct
48/s 08 .93 (.04)
208 .83 (.07)
108/s 08 .76 (.09)
208 .68 (.08)
1380 HUFF ET AL.
Downloaded By: [Huff, Markus] At: 05:14 28 September 2010
83.89% of the trials for this analysis) this analysis is based on very few trials
resulting in noisy data (see Figure 4 and Table 4).
As in Experiment 1, we examined the relationship between object speed
and gaze spent on centroid and mean target AOIs for the three intervals
around the viewpoint change. First, we observed a significant interaction
between AOI type (targets and centroid) and object speed, F(1, 13)�4.97,
p�.044, hp2�.28, replicating the findings from Experiment 1. Centroid
looking grew in importance with increasing object speed. Although there
was no difference between centroid and mean target AOIs in conditions with
48/s object speed, with 108/s we recorded significantly more gaze on the
centroid than on the mean target AOIs (pairwise Holm-corrected t-tests, pB
.05; see Figure 5). In addition, compared to mean target AOIs, more gaze
was recorded on the centroid AOI, F(1, 13)�12.17, p�.004, hp2�.48.
Finally, there was no main effect of object speed on the total amount of gaze
spent on centroid and mean target AOIs, FB1.
Figure 4. Gaze behaviour across 500 ms intervals in Experiment 2.
EYE MOVEMENTS ACROSS VIEWPOINT CHANGES 1381
Downloaded By: [Huff, Markus] At: 05:14 28 September 2010
Because the degrees of movement of the centroid are restricted*it is
located close to the centre of the floor plane most of the time*it is necessary
to exclude two possible alternative explanations of the findings. First, if
participants reorient by looking at the centre of the floor plane after the
abrupt viewpoint change, the observed centroid looking could be an artifact.
In order to exclude this possibility, we summed up the gaze positions with-
in a region of the size of a single object (subtending 1.75 degrees of visual
angle) on the centre of the floor plane and calculated the corresponding
ANOVA with the independent variable ‘‘interval’’. There was no influence
of the ‘‘interval’’ on gaze on the centre, F(2, 34)�2.35, p�.112, hp2�.12.
In particular, there was no increase of gaze after the abrupt viewpoint
change.
Second, results may be influenced by variability of eye movement right
after the abrupt viewpoint change. For example, if participants focus on the
centroid right before the abrupt viewpoint change and if participants do not
saccade at all but keep their eyes still after the abrupt viewpoint change,
one might expect that gaze stays within the centroid AOI for a considerable
amount of time. Accordingly, the observed result pattern would not indicate
centroid looking. In order to control for gaze variability across trials, we cal-
culated the mean number of saccades within the relevant three 500 ms
intervals (one interval before and two intervals after the viewpoint change).
Saccades were detected using an eye-velocity criterion of 308/s, Data as
reported in Table 5 were submitted to within-subject ANOVAs including the
independent variable ‘‘time interval’’ and the dependent variable ‘‘mean
TABLE 4ANOVA results of the gaze behaviour across trial intervals in Experiment 2; plannedcontrasts between time intervals before and after the abrupt viewpoint change were
Downloaded By: [Huff, Markus] At: 05:14 28 September 2010
number of saccades’’ that were calculated separately for each condition. We
did not find any influence of the viewpoint change on the number of
saccades, all hp2sB.08. Hence, we can exclude the possibility of participants
just staring at the screen after the abrupt viewpoint change. Taken together,
centroid looking in the interval after the abrupt viewpoint change is real and
not a consequence of mere staring at the centre of the screen.
Transitions across abrupt viewpoint changes
Abrupt viewpoint changes suddenly displace the retinocentric representa-
tion of the 3-D scene. Do participants rematch their gaze position to the
displaced position of the previously fixated visible or invisible object across
Figure 5. Gaze on target and centroid AOIs in Experiment 2. With increasing object speed, gaze on
the centroid increased while gaze on targets decreased. Error bars indicate the standard error of the
mean.
EYE MOVEMENTS ACROSS VIEWPOINT CHANGES 1383
Downloaded By: [Huff, Markus] At: 05:14 28 September 2010
the viewpoint change or are there transitions between objects, for example
from a target to the centroid, or vice versa? If gaze positions were equally
likely rematched to the displaced position of a specific target or the centroid,
the drop of gaze on the targets might be caused by the larger target dis-
placement than centroid displacement possibly requiring more time to
re-match the gaze position to the displaced target locations than centroid
locations. For this analysis, we examined gaze behaviour in successful trials
of the 48/s condition with viewpoint change across all participants (perfect
tracking and perfect confidence rating). We defined transitions across
viewpoint changes as pairs of two object AOIs: The last object AOI that
gaze was matched with in the interval 500 ms before the viewpoint change
and the first object AOI that gaze was matched with in the interval 500 ms
following the viewpoint change. In order to identify valid transitions, we
removed gaze matches with the floor plane and misses as well as gaze
matches with object AOIs lasting less than 10 ms, thus removing spurious
matches caused by noise or bypassing saccades. Additionally, we removed
all transitions taking less than 10 ms because these did not result from
programmed saccades. As we were particularly interested in the difference
between gaze on target AOIs and the centroid AOI, we analysed transitions
starting and ending on target AOIs or the centroid AOI only. In total, there
were more transitions to the same object AOI (n�44) than to an object AOI
of the opposite category (target-centroid and centroid-target transitions, n�19), x2(1, N�63)�9.92, p�.002. Separating transitions as to whether they
started on a target AOI or centroid AOI, they did not differ in their
preference for ending on the very same object AOI or the object AOI of the
TABLE 5Mean number of saccades using a 308/s eye velocity criterion in Experiment 2 (withSD in parentheses); in conditions with 208 viewpoint change, the viewpoint change
occurred between intervals 1 and 2
Viewpoint change Speed (8/s) Interval Saccades
08 4 1 1.01 (0.56)
2 1.01 (0.55)
3 0.93 (0.47)
10 1 0.97 (0.52)
2 0.93 (0.72)
3 1.08 (0.91)
208 4 1 1.00 (0.66)
2 0.98 (0.47)
3 1.08 (0.59)
10 1 1.12 (0.74)
2 1.08 (0.84)
3 1.10 (0.70)
1384 HUFF ET AL.
Downloaded By: [Huff, Markus] At: 05:14 28 September 2010
opposite category, x2(1, N�63)�1.15, p�.283 (see Table 6). We thus
compared the time following the viewpoint change needed to re-match gaze
from a target AOI to the same target AOI (M�244 ms) and centroid AOI to
the same centroid AOI (M�217 ms) using a Wilcoxon rank sum test and
found no significant difference, W�227.5, p�.504. Hence despite larger
target displacements and the equal likelihood of making a transition to the
same object AOI rather than the object AOI of the opposite category for
targets and the centroid, the drop on target AOIs as described in the gaze
behaviour analysis cannot be explained by re-matching processes consuming
more time for targets than for the centroid.
Saccade analyses
So far, we showed a drop of gaze on target AOIs opposed to the centroid
AOI in the 500 ms following the viewpoint change. Furthermore, transitions
from a target AOI to the same target AOI take as long as transitions from
the centroid AOI to the centroid AOI. In order to gain further insights on
the difference of gaze to centroid and target AOIs following a viewpoint
change, we analysed the first saccades elicited by and therefore starting
within the 500 ms following the viewpoint change in the 48/s object speed
conditions irrespective of the previously fixated object AOIs across all
participants (perfect tracking and perfect confidence rating). We used a 308/seye-velocity threshold criterion for detection of the saccades. Consider-
ing the first saccade following the viewpoint change, there were 18 saccades
to the centroid AOI and 44 saccades to any target AOI. A Wilcoxon rank
sum test reveals that the time until the first saccade to the centroid AOI
(M�183 ms) is significantly shorter than the time until the first saccade to a
target AOI (M�266 ms), W�251, p�.024, thus showing that the centroid
of the target set can be computed very quickly in successful trials. This
difference cannot be explained in terms of saccade amplitude as the amp-
litude of saccades directed to the centroid AOI (M�2.688) did not differ
TABLE 6Total number of transitions across the abrupt viewpoint change for targets and the
centroid to either the same object AOI or an object AOI of the opposite category (targetto centroid or centroid to target transitions) in Experiment 2
AOI before viewpoint change
AOI after viewpoint change Target Centroid
Same object 31 13
Opposite category 10 9
EYE MOVEMENTS ACROSS VIEWPOINT CHANGES 1385
Downloaded By: [Huff, Markus] At: 05:14 28 September 2010
significantly from the amplitude of saccades directed to a target AOI (M�2.158) as revealed by a Wilcoxon rank sum test, W�495, p�.125.
Further, we analysed all saccades to target and centroid AOIs in the
500 ms intervals surrounding the viewpoint change. Although there wasan equal number of saccades to the centroid AOI (n�19) and to the mean
target AOI (n�16.67) in the 500 ms before the viewpoint change, x2(1, N�35.67)�0.15, p�.696, and in the second 500 ms interval following view-
point change (n�20 for centroid AOI, n�17 for mean target AOI), x2(1,
N�37)�0.24, p�.621, things look different in the interval right after
the viewpoint change. For the first 500 ms following the viewpoint change
there were more saccades to the centroid AOI (n�30) than the mean tar-
get AOI (n�17), x2(1, N�47)�3.60, p�.058. Summarizing, results fromthe saccade analyses suggest that the centroid can be computed very
quickly and that the centroid as a stable reference for the target set is
used when it comes to relocate the target objects after abrupt viewpoint
changes.
The goal of Experiment 2 was to replicate the findings of Experiment 1
showing that gaze on the centroid is more stable than gaze on targets. In
Experiment 2 we were able to replicate the central findings using an eye
tracking system with high temporal and spatial resolution. This enabledus to calculate additional analysis, like gaze transitions across the ab-
rupt viewpoint change and saccade goals after the viewpoint change.
Further, we tightened the selection criterion in Experiment 2 by analys-
ing trials with perfect tracking performance and perfect confidence rating
only. That is, participants stated that they guessed no target. Moreover,
we ruled out the alternative hypothesis according to which centroid looking
as described in Experiment 1 might have been the result of an anticipa-
tion strategy because the viewpoint change occurred always after 3 s. Al-though this is unlikely as trials with and without viewpoint change were
presented intermixed, we added some uncertainty with respect to the time
point of the viewpoint change in Experiment 2. Finally, we wanted to test
if the observed positive relation between object speed and gaze time spent
on the centroid also holds for a wider range of object speeds (4 and
108/s).
Despite all these changes, we were able to replicate the central findings of
Experiment 1. Viewpoint changes caused a decline of gaze on targets but notthe centroid. Centroid looking was observed at all object speeds and became
more important with increasing object speeds as hypothesized.
Although we were able to replicate the central findings of Experiment 1 in
the 48/s condition (drop and recovery of gaze on targets in the first 500 ms
interval after the abrupt viewpoint change), there was no influence of the
viewpoint change on gaze behaviour at all in the 108/s condition. Pre-
sumably, the tracking task was too difficult.
1386 HUFF ET AL.
Downloaded By: [Huff, Markus] At: 05:14 28 September 2010
GENERAL DISCUSSION
The reported experiments studied effects of scene motion and object speed
on eye movements. Observers’ gaze behaviour was recorded while they trac-
ked multiple moving objects across abrupt viewpoint changes. Abrupt
changes of the display holding the objects shift retinocentric coordina-
tes, which usually results in lower tracking performance (Huff et al., 2009;
Seiffert, 2005). How do human observers recollect targets after abrupt scene
motion? A reasonable strategy is perceptual grouping (Yantis, 1992). Look-
ing at the centroid that is defined as the targets’ centre of mass could indicate
such a strategy. Consistent with this presumption, gaze on the centroid
declined less after abrupt viewpoint changes than gaze on targets, which was
reduced for about 500 ms after the abrupt viewpoint change. Furthermore,
we demonstrated for the first time within the same experiment that centroid
looking gains in importance at higher object speeds.
Gaze behaviour during MOT was examined in two recent studies (Fehd
& Seiffert, 2008; Zelinsky & Neider, 2008). In both studies, a considerable
amount of gaze was directed towards the centre of mass of the target group
when three targets were tracked. This so called ‘‘centroid looking strategy’’
was stronger in the Fehd and Seiffert (2008) study, which employed higher
object speed than in the Zelinsky and Neider (2008) study. These results
suggest a positive correlation between object speed and amount of cen-
troid looking: When tracking gets more strenuous more gaze related with
the centroid was observed. We confirmed this assumption in both of our
experiments.
We hypothesized that after an abrupt viewpoint change the targets are
recollected by aligning the target group to the new viewpoint indicated by
gaze related with the centroid. We confirmed this hypothesis in Experiment 1
and Experiment 2 by showing that the centroid was looked at even right
after a viewpoint change when observers successfully recollected shifted
targets.
There are several benefits that centroid looking might provide in tracking
and recollecting targets. First, pursuing the centroid that represents the
target group by a single spatial coordinate prevents that any target moves far
from focal vision where peripheral vision quickly blurs. This may be an
advantage over a target jumping strategy, in which a single focused target
benefits from highest visual resolution. However, the distance from this
target located in the fovea to the nonfocused targets is larger than the mean
distance between the targets and the corresponding centroid. As a result,
with target jumping the nonfocused targets are located further in the
periphery increasing the probability of being lost. Second, the centroid’s
velocity and degrees of movement are restricted compared to single targets.
EYE MOVEMENTS ACROSS VIEWPOINT CHANGES 1387
Downloaded By: [Huff, Markus] At: 05:14 28 September 2010
And finally, the average displacement caused by an abrupt viewpoint change
is less for the centroid than for targets.
In the current study, with increasing task difficulty*either due to the
occurrence of an abrupt viewpoint change (no matter if its time point couldbe anticipated as in Experiment 1 or not as in Experiment 2) or due to
increasing object speed (Experiment 1 and 2)*the relative amount of
centroid looking increased. Presumably, our participants made use of some
or all of the benefits of centroid looking.
The present analysis of eye tracking data differs substantially from pre-
our study gaze data were analysed with the dynamic AOI approach (every
gaze on an a priori defined AOI was coded as a match and all other gazeswere coded as a miss), previous studies employed the ‘‘shortest distance
rule’’ that assigns every gaze to its nearest object. The dynamic AOI ap-
proach as implemented in this study explained about 35�47% of gaze. This is
a reasonable proportion because the total size of the AOIs (targets, dis-
tractors, and centroid) is far less than 35% of the display size (see Figure 1).
One way of increasing the amount of gaze explained by the dynamic AOI
approach could be to increase the size of the dynamic AOIs. In our study the
AOI size equalled the object size. However, increasing AOIs becomesproblematic as this would result in a lot of overlaps of the different AOIs.
In this context, it should be noted that the proportion of gaze identified as
centroid looking differs substantially between the previous studies (Fehd &
Seiffert, 2008; Zelinsky & Neider, 2008) and our study. Whereas in previ-
ous studies the proportion was 66% and 39% (Fehd & Seiffert, 2008, and
Zelinsky & Neider, 2008, respectively), in our study about 10% of gaze was
identified as centroid looking. However, considering the lower overall
proportion of gaze explained by the dynamic AOI approach, this value isclose to the proportion of centroid looking found in previous studies. Taken
together, although the eye tracking methodology applied in this study
differed in various aspects from previous experiments, it seems to be quali-
tatively equivalent to the ‘‘shortest distance rule’’.
Both methods*‘‘dynamic AOI’’ and ‘‘shortest distance rule’’*need the
a priori definition of objects and do not consider any other strategies. Maybe
gaze on targets and centroid do not account for all gaze in multiple object
tracking. For example, it is conceivable that some gaze is not only relatedwith real and virtual objects like the centroid. Instead, gaze could be
distributed more in the picture plane between the targets. In this case, the
dynamic AOI approach as applied in this study would be a good estimator of
the real gaze behaviour as it does not falsely assign gaze to the nearest
object.
How does the visual system calculate the positions of the centroid after
the viewpoint change without knowing the exact locations of the targets?
1388 HUFF ET AL.
Downloaded By: [Huff, Markus] At: 05:14 28 September 2010
From the additional analyses of Experiment 2 we know that centroid look-
ing is not the result of just staring at the centre of the floor plane or reduced
saccade frequencies after the viewpoint change. Further, we know that the
gaze position across the viewpoint change is most likely to be rematched tothe displaced position of the previously looked at object and that this
rematching process takes as long for targets as the centroid. At first sight,
these results suggest that there was no preference for centroid looking when
it comes to relocating targets following a viewpoint change. Having a closer
look at the data by taking all saccades in the intervals surrounding the
viewpoint change into account, we showed that there is a preference for
saccades to the centroid AOI over the mean target AOI in the first 500 ms
following the viewpoint change only. That is, there is a preference in lookingat the centroid for relocating the targets. Even more interesting in respect to
the process underlying centroid calculation is the analysis of the first saccade
following the viewpoint change. We found that the first saccade following
the viewpoint change is more quickly directed towards the centroid than the
targets and that the saccade to the centroid starts in mean only 183 ms
following the viewpoint change. Considering research on the relationship of
attention and saccades showing that attention shifts to the target of a sac-
cade prior to the occurrence of the saccade (e.g., Hoffman & Subramaniam,1995; Posner, 1980; Shepherd, Findlay, & Hockey, 1986), our results suggest
that the centroid is calculated almost instantly following the viewpoint
change. Thus, the process underlying centroid calculation must be working
very efficiently and might be supported by the recognition of the target set as
a perceptual group (Yantis, 1992). As the centroid is a stable representation
of the targets (it moves slower and its degrees of movements are restricted), it
is the most favourite goal of gaze when visual tracking gets harder and
especially after abrupt viewpoint changes.From an evolutionary point of view, the visual system cannot be adapted
to abrupt and unpredictable viewpoint changes yet, since the time period
from the introduction of the first edited movies (around 1900; Kuleshov,
1920/1974) is too short. A recent study has shown that there is a difference
between first time film viewers and experienced viewers in understanding
viewpoint changes within a scene (Schwan & Ildirar, 2010). Whereas
experienced viewers do understand such viewpoint changes first time viewers
do not. Therefore, it would be worth comparing gaze data of experiencedand inexperienced viewers during an MOT task including abrupt viewpoint
changes.
The abrupt scene motion employed in the reported experiments was
a rotational viewpoint change. Hence, target translations were larger than
centroid translations. Further studies with equidistant translations of all
objects (including targets, distractors, and centroid) are needed to gain fur-
ther insight in the gaze behaviour while tracking across abrupt changes in
EYE MOVEMENTS ACROSS VIEWPOINT CHANGES 1389
Downloaded By: [Huff, Markus] At: 05:14 28 September 2010
general. This could be done by simple horizontal translations of the floor
plane. An important question will be if the relative amount of centroid
looking is comparable to the one shown in this study. A first hypothesis
about gaze behaviour while tracking across abrupt translations can bededuced from research on saccadic localization. If objects are presented in
the periphery and subjects were instructed to look at the targets as a whole,
saccades typically land at the centre of gravity of the shape even if this is
located outside the shape (McGowan et al., 1998; Vishwanath & Kowler,
2003). If visual tracking relies on grouping the single targets into an object
of higher order, it seems plausible that gaze behaviour with equidistant
translations of targets and centroid is comparable to gaze behaviour with
rotational viewpoint changes; observers should be able to efficiently use thecentroid in recollecting target objects.
REFERENCES
Alvarez, G. A., & Franconeri, S. L. (2007). How many objects can you track? Evidence for
a resource-limited attentive tracking mechanism. Journal of Vision, 7(13), 1�10. doi:10.1167/
7.13.14.
Alvarez, G. A., & Oliva, A. (2008). The representation of simple ensemble visual features outside
the focus of attention. Psychological Science, 19(4), 392�398. doi:10.1111/j.1467-9280.
2008.02098.x.
Cavanagh, P., & Alvarez, G. A. (2005). Tracking multiple targets with multifocal attention.
Trends in Cognitive Sciences, 9(7), 349�354. doi:10.1016/j.tics.2005.05.009.
Doran, M., Hoffman, J. E., & Scholl, B. (2009). The role of eye fixations in concentration and
amplification effects during multiple object tracking. Visual Cognition, 17(4), 574�597.
doi:10.1080/13506280802117010.
Fehd, H. M., & Seiffert, A. E. (2008). Eye movements during multiple object tracking: Where do