|Home | About | Journals | Submit | Contact Us | Français|
Is it easier to track objects that you have seen repeatedly? We compared repeated blocks, where identities were the same from trial to trial, to unrepeated blocks, where identities varied. People were better in tracking objects that they saw repeatedly.
We tested four hypotheses to explain this repetition benefit. First, perhaps the repeated condition benefits from consistent mapping of identities to target and distractor roles. However, the repetition benefit persisted even when both the repeated and the unrepeated conditions used consistent mapping. Second, repetition might improve the ability to recover targets that have been lost, or swapped with distractors. However, we observed a larger repetition benefit for color-color conjunctions, which do not benefit from such error recovery processes, than for unique features, which do. Furthermore, a repetition benefit was observed even in the absence of distractors. Third, perhaps repetition frees up resources by reducing memory load. However, increasing memory load by masking identities during the motion phase reduced the repetition benefit. The fourth hypothesis is that repetition facilitates identity tracking, which in turn improves location tracking. This hypothesis is consistent with all our results. Thus, our data suggest that identity and location tracking share a common resource.
There are many real world tasks that require us to keep track of moving objects. These range from specialized tasks such as air traffic control and radar operation (Allen, McGeorge, Pearson, & Milne, 2004) to everyday tasks such as driving in traffic and monitoring your children at the beach.
Our research question concerns the observer’s experience with the identity of the tracked objects. Specifically, is it easier to track objects that you have seen repeatedly? Imagine you place your pet fish in an aquarium with your neigbor’s pet fish. Is it easier to monitor your own pet fish than your neighbor’s? The answer to this question holds straightforward practical implications (for starters, perhaps you should not let others monitor your pet fish when it really matters), but it also elucidates the mechanisms that drive monitoring moving objects.
The laboratory method for this sort of tracking is the multiple object tracking paradigm (MOT, Pylyshyn & Storm, 1988). In a typical MOT trial, the observer is presented with an array of identical objects (e.g., gray disks). The targets are briefly highlighted, after which all objects move independently for several seconds. The observer must then indicate the location of the targets (see Figure 1). During the moving phase, all the objects are identical, so the observer can do the task only by continuously attending to the targets. In most cases, observers can track 3–5 targets (Cowan, 2001; Intriligator & Cavanagh, 2001; Yantis, 1992), though tracking performance can be modulated by the speed of the objects and how close they come to one another (Alvarez & Franconeri, 2007; Franconeri, Lin, Pylyshyn, Fisher, & Enns, 2008).
While the standard MOT paradigm has taught us a great deal about object tracking (for reviews, see Cavanagh & Alvarez, 2005; Scholl, 2009), it is not naturalistic, in the sense that in the real world tracked objects are often distinct rather than identical. The first investigation of MOT where the targets had distinct identities was performed by Pylyshyn (2004). The paradigm was similar to that described above except that, at the start of each trial, each target was labeled with a digit. The digits then disappeared and were not visible during the motion phase. At the end of the trial, the observer was asked to indicate first which items were the targets and then which digit went with which target. Pylyshyn found that the capacity for target identities was substantially lower than the capacity for target locations. These data suggest that there is a binding problem (Wolfe & Cave, 1999) for target identities in MOT (see also Imaruoka, Saiki, & Miyauchi, 2005; Saiki, 2002; Saiki, 2003).
One could argue that Pylyshyn’s (2004) paradigm is still unrealistic because in the real world target identities are usually continuously visible. Oksama and Hyönä (2004) addressed this issue in their multiple identity tracking (MIT) experiments in which the stimuli were distinct objects (or pseudo-objects) that were continuously visible. At the end of the trial, all the objects were masked, and the observer was asked to report the identity of a randomly selected target. They found that observers could track approximately four objects without confusing their identities, much more than the capacity reported by Pylyshyn (2004), suggesting that the identity-location binding problem is at least partially ameliorated by making the targets continuously visible during the moving phase.
If continuous visibility completely eliminated the binding problem, then we would expect tracking capacity for MIT to be the same as that for MOT. However, Horowitz et al. (2007) reported that this was not the case. In their experiments the stimuli were cartoon animals. In the specific condition, observers were asked to identify the location of particular objects, similar to the Oksama and Hyönä (2004) setup. In the standard condition, observers were asked to identify the locations of all the targets, without specifying which target was where. Capacity in the specific condition ranged from 1.4 to 2.6 objects, whereas in the standard condition it ranged from 2.3 to 3.4 objects. Thus, even when the object identities were continuously visible during the movement phase, the tracking capacity for MIT was still less than that for MOT.
As MOT studies move away from identical stimuli towards more ecologically valid studies that utilize distinct objects, we need to consider not only the influence of stimulus factors on tracking accuracy (e.g., velocity, Alvarez & Franconeri, 2007; spacing, Franconeri, et al., 2008; color and shape, Makovski & Jiang, 2009) but also the interplay between tracking and memory (Allen, Mcgeorge, Pearson, & Milne, 2006; Fougnie & Marois, 2006; Oksama & Hyona, 2004; Saiki, 2003).
Oksama and Hyönä’s (2008) study is a clear example of the importance of non-stimulus factors driving tracking performance. Their data suggests that it is easier to track objects that you are more familiar with. They found tracking performance to be significantly better for real, common objects (drawings from Snodgrass & Vanderwart, 1980) than for pseudo-objects (drawings of object-like stimuli, from Kroll & Potter, 1984). This finding directly relates to our research question since it suggests that it is easier to track objects that you have had more exposure to. However, their results were not conclusive. First, in their design, familiarity is confounded by nameability. Nameable, real objects might have been easier to remember, thereby reducing memory load and consequently improving tracking performance. Second, there might be low-level visual differences between the real and the pseudo-objects Oksama and Hyönä used. These differences could have contributed to the observed improvement in tracking performance.
We avoided these issues by using the same set of stimuli in all conditions, thus equating for nameability and perceptual effects. We then compared performance under repeated (same target and distractors throughout a block) and unrepeated (targets and distractors randomly sampled throughout a block) conditions. We found that repeating item identities significantly improved tracking performance. We considered four possible explanations. First, identity repetition might reduce target-distractor interference, since distractor identities never become target identities and vice versa. Second, identity repetition might make it easier to recover lost targets. Third, identity repetition might reduce working memory load, freeing up resources for tracking. Finally, perhaps resources can flexibly be distributed between identity and location tracking. It could be that identity repetition makes identity tracking more efficient, thus freeing more resources for location tracking.
Identity tracking here is really comprised of two processes: recognizing and maintaining the representation of an identity; and binding this representation to a position or spatial index. Thus, the visual system must maintain a visual representation of the zebra as it moves, and bind the zebra image to a particular spatial locus. It is important to think about this visual representation of the zebra as something different from simply knowing that one is tracking a zebra, a distinction that will become important in Experiment 5.
Our results are most consistent with the hypothesis that repetition improves identity tracking. Whether it does so through facilitating the visual representations of targets or their binding to positions, or both, cannot be determined from our experiments. Furthermore, while we show that each of the other hypotheses is insufficient to explain the repetition benefit, neither can we exclude the possiblity that they may contribute to the effect. In the General Discussion, we discuss the implication of these findings for models of MIT.
As noted above, Oksama and Hyönä’s data suggest that familiarity with the stimuli enhances tracking performance, but their result is potentially confounded with nameability and the perceptual qualities of the stimuli. The goal of our first experiment was to demonstrate that merely repeating target identities improves tracking performance.
We compared performance when all stimuli changed from trial to trial within a block (the unrepeated condition), to a situation where the targets and distractors remained the same on each trial throughout the block (the repeated condition). Crucially, each object had an equal chance of being assigned to the repeated set for all observers. Thus, there should be no visual or semantic differences between the repeated and unrepeated objects.
Eight naïve, paid volunteers, ranging in age from 18 to 50 years (average 27.5 years) participated in this experiment. All participants had normal or corrected-to-normal vision.
Stimuli were presented on a 21-inch monitor set to a resolution of 1024 by 768 at a refresh rate of 75 Hz, controlled by a Macintosh G5 computer running Mac OS 10.4. The experiment was programmed in Matlab 7.5 (The MathWorks) using the Psychophysics Toolbox routines (Brainard, 1997; Pelli, 1997). Participants were seated approximately 57.4 cm from the monitor; at this distance, 1 cm on the screen subtends one degree of visual angle (°). The stimulus field consisted of eight cartoon animals (2.37° x 2.37°). Animals moved at a speed of 17.8°/s in straight lines, except when they bounced off each other or the sides of an imaginary window (32.7° x 22.5°). At the end of a trial, each animal was masked by a red (luminance: 10.5 cd/m2, as measured with a Tektronix photometer, and CIE (x, y) coordinates of (.617, .349)) 2.37° x 2.37° square. The background was white (.289, .317, luminance: 63.2 cd/m2).
The experiment consisted of 8 blocks of 30 trials. On all trials, participants tracked four targets among four distractors. There was a set of 22 cartoon animals (see Appendix 1). In four of the blocks, targets and distractors randomly varied from trial to trial (the ‘unrepeated’ condition); in the other four blocks, both targets and distractors remained the same on every trial (the ‘repeated’ condition). Half of the participants started with the repeated condition, the other half started with the unrepeated condition. The basic trial procedure is illustrated in Figure 2. Each trial started with highlighting the targets in red boxes (3.57° x 3.57°, luminance: 10.5 cd/m2, (.617, .349)) for 3 seconds. All items then moved for a randomly determined time between 2 and 12 seconds. The length of the trial was unpredictable in order to ensure that participants were continually tracking rather than adopting a strategy of locating the targets just before the end of the trial. All items were then masked with red squares. Participants were asked to click on each specific target in turn (e.g., ‘where is the zebra?,’ followed by ‘where is the turtle,’ etc.). Participants were instructed to take their time and respond as accurately as possible. The experiment was performed without breaks and took approximately 90 minutes.
We analyzed two dependent measures: location accuracy and identity accuracy. Location accuracy refers to participants’ ability to distinguish target from distractor locations, whereas identity accuracy refers to their ability to know which target is where. Imagine that a participant is asked to track the alligator, the camel, the fox, and the zebra, among distractors that include the turtle, the elephant, the horse, and the tiger. If the participant was asked to click on, say, the zebra, and she clicked on the camel, this would be considered a hit with respect to location accuracy, because it indicates that she knew that the item was a target, but as a miss with respect to identity accuracy, because she did not know which target it is. Note that with this definition, identity accuracy can never exceed location accuracy.
Location and identity accuracies are plotted in Figure 3 as a function of repetition. Performing a 2-way within subjects ANOVA, we found that location accuracy was significantly greater than identity accuracy (F(1,7) = 22.46, MSE = 75.48, p < 0.005, η2 = .762), and accuracy for repeated objects was greater than accuracy for unrepeated objects (F(1,7) = 17.52, MSE = 56.72, p < 0.005, η2 = .715). Furthermore, these two effects interacted, such that the effect of repetition was greater for identity accuracy than for location accuracy (F(1,7) = 16.53, MSE = 12.10, p = 0.005, η2 = .702). Finally, post-hoc tests showed that both identity accuracy and location accuracy were significantly higher in the repeated case (F(1,7) = 22.31, p < 0.005, η2 = .761 and F(1,7) = 6.85, p < 0.05, η2 = .494 respectively).
The main result of Experiment 1 is that identity accuracy increased from roughly 64% when tracking unrepeated targets to roughly 81% when tracking repeated targets. This indicates that it is indeed easier to track unique targets if you have seen them more often (an effect that we will refer to as the repetition benefit). Another noteworthy result is that location accuracy also increased for repeated items. Thus, participants were not only better at tracking identities, but also at tracking locations.
What is the source of the repetition benefit demonstrated in Experiment 1? One possibility is that this benefit has nothing to do with repetition per se, but more with consistency of target and distractor roles. In the unrepeated condition in our experiments, any given item could serve as both a distractor and a target, on different trials, while in the repeated condition an item’s role was always fixed. This aspect of the design could potentially create a repetition benefit by itself in a number of ways. For example, visual search studies have shown that inconsistent mapping (where targets and distractors can swap roles) leads to impaired performance relative to consistent mapping (Schneider & Shiffrin, 1977). Other visual search studies have demonstrated that a distractor is more distracting if it served as a target on the previous trial (target-distractor priming, Pinto, et al., 2005). Conversely, negative priming studies (Tipper, 1985; Tipper, et al., 1998) have also demonstrated that ignored distractors are harder to respond to when they show up as targets on the next trial. Thus, it is possible that the repetition benefit has nothing to do with identity repetition per se, but is simply a consequence of a design flaw in our experiments.
In Experiment 2 we addressed this concern by introducing two new conditions. In both conditions target and distractor identities were chosen from separate pools, ensuring that targets never served as distractors and vice versa. If the repetition benefit is really a benefit for consistent mapping, then we would expect that tracking performance in these new conditions would be as good as tracking in the repeated condition. However, if the repetition benefit is indeed due to the repetition of the target identity than we still expect tracking performance to be the best in the repeated condition.
Eight participants took part as paid volunteers. Their ages ranged from 18 to 50 (average 31.6) years, and all had normal or corrected-to-normal vision.
The stimuli employed in this experiment were different from the previous experiment. The stimuli were 168 pictures from the MIT objects database containing 2400 pictures of objects (http://cvcl.mit.edu/MM/download.html).
There were four conditions, repeated, standard unrepeated, unrepeated Consistent Mapping 1 (CM1), and unrepeated Consistent Mapping 2 (CM2). The repeated condition was similar to that of Experiment 1, where throughout the block targets and distractors had fixed identities (the 8 identities of targets and distractors were randomly selected from the pool of 168 pictures). Similarly, the standard unrepeated condition was designed to mimic the unrepeated condition from the previous experiment. We first randomly selected 40 of the 160 remaining pictures. Then, on each trial the four target and four distractor identities were randomly chosen from this pool of 40 pictures. In the CM1 condition we randomly selected 80 of the remaining 120 pictures. 40 pictures were then randomly assigned to the target pool and the remaining 40 pictures to the distractor pool. On each trial, the four target identities were randomly drawn from the target pool, whereas the four-distractor identities were randomly drawn from the distractor pool. This ensured that the targets could never serve as distractor and vice versa. Furthermore the number of target repetitions in this condition was the same as the number of target repetitions in the standard unrepeated condition. Finally, in the CM2 condition the remaining group of 40 pictures was randomly divided in a group of 20 target pictures and 20 distractor pictures. Other than the number of pictures in each pool being smaller than in the CM1 condition, the CM2 and CM1 condition were the same. Thus, the CM1 condition matched the standard unrepeated condition in terms of number of target repetitions, but used twice as many stimuli, while the CM2 condtion matched the standard unrepeated condition in terms of number of stimuli, but with twice the number of target repetitions. The distribution of pictures over the different pools was done independently for each participant.
The experiment consisted of 4 blocks of 40 trials in a Latin Square design. Two participants were run on each of the four possible orders. The experiment was performed without breaks and took approximately 60 minutes.
The data are presented in Figure 4. We analyzed identity and location accuracy separately. In both cases, we compared the unrepeated and CM conditions to the repeated condition. With regards to location accuracy, the trend was towards accuracy being the highest in the repeated condition, and the lowest in the unrepeated condition, but this did not reach significance for any of the comparisons. With regards to identity accuracy, a one-way ANOVA with the factor condition (repeated, standard unrepeated, CM1 and CM2) revealed a main effect (F(3,21) = 12.83, MSE = 28.774, p < .0001, η2 = .647). A within-subjects contrast, contrasting all conditions to the repeated condition, revealed that accuracy in the repeated condition was higher than in all other conditions (for all conditions F > 7.5, all p < 0.05, all η2 > .5).
Thus, the effects of consistent mapping do not account for the entire repetition benefit. Identity accuracy was significantly higher in the repeated condition than in both of the consistent mapping conditions. Even when target and distractor identities were strictly separated, so that inconsistent mapping, target-distractor priming, and negative priming were excluded, tracking performance was still better in the repeated than in the unrepeated condition. This shows that the repetition benefit is, at least partly, due to repeating target identities.
Another possible explanation is that the repetition benefit has the same cause as the uniqueness benefit. Previous studies (Horowitz, et al., 2007; Makovski & Jiang, 2009) have shown that it is easier to track unique objects, relative to identical objects. Both studies suggested that this is because if the observer lost track of a target, he would know that he had lost, for example, the zebra, and could then search for a zebra-like item in the field to recover that target. Perhaps repeated targets are easier to track because repetition improves observers' memory for the target set, making it easier both to recognize which target is missing and to search for the missing target.
Alternatively, the repetition benefit might reflect a shift in resources from identity tracking to position tracking. If we posit that the location and identity tracking aspects of the task draw on a common resource, then anything that makes identity tracking easier would free up resources for improved position tracking. For instance, if the same identity is repeatedly bound to the same position, or object file (Kahneman, et al., 1992; Pylyshyn, 2001), this operation may become easier over time, allowing the observer to concentrate more resources on position tracking.
Contrasting these two hypotheses required a change from the cartoon animal and real object stimuli to a more controlled stimulus set, following Makovski and Jiang (2009).
The purpose of Experiment 3 was twofold. First, we wanted to determine whether the repetition benefit would replicate with simple stimuli that presumably result in a lower binding load. Second, we wanted to test whether the repetition benefit could be entirely explained by improved search for lost targets. We therefore designed a condition in which each stimulus was defined by a unique feature, and a condition in which each stimulus was defined by a conjunction of two features.
In the feature condition, each stimulus was a square of a unique color: red, blue, yellow, green, black, white, orange, or purple. In the conjunction condition, stimuli were squares divided vertically such that the left and right halves were of different colors. Each stimulus was a unique combination of two colors. In the feature condition the binding load is lower than in the conjunction condition (Treisman, 2006), so if the repetition benefit is based on improved binding, the repetition benefit should be substantially lower in the feature condition than in the conjunction condition.
The target recovery hypothesis predicts an opposite result. Makovski and Jiang (2009) found that tracking performance increased when targets had feature identities but not when they had conjunction identities (compared to when all targets were identical), suggesting that conjunction identities cannot be used for recovering lost targets. At any rate, searching for a lost feature target should be substantially easier than searching for a lost conjunction target, especially a color x color conjunction (Wolfe, et al., 1990). Thus, if the repetition benefit is based on searching for lost targets, this benefit should be eliminated, or at least reduced, in the conjunction condition, compared to the feature condition.
Therefore we have two contrasting predictions. According to the binding hypothesis, the repetition benefit should be larger in the conjunction condition. The target recovery hypothesis predicts the opposite: the repetition benefit should be observed in the feature condition, but not (or only minimally) in the conjunction condition.
Sixteen participants took part as paid volunteers. They ranged in the age from 18 to 50 years old (average 30.1 years), and all had normal or corrected to normal vision.
In both conditions the stimuli had the same dimensions as in Experiment 1 (i.e. 2.37° x 2.37°). In the feature condition there were eight different basic colors: red (0.641, 0.341, luminance: 11.5 cd/m2), blue (0.148, 0.072, luminance: 4.82 cd/m2), yellow (0.405, 0.521, luminance: 54.4 cd/m2), green (0.293, 0.606, luminance: 26.7 cd/m2), black (0, 0, luminance: 0 cd/m2), white (0.289, 0.317, luminance: 60.6 cd/m2), orange (0.482, 0.444, luminance: 24.0 cd/m2), purple (0.277, 0.142, luminance: 4.36 cd/m2). The background (in both conditions) was gray (.282, .308, 10.8 cd/m2). In the conjunction condition there were 24 possible color x color conjunctions. All conjunctions were two half-rectangles (1.19° x 2.37°) put together to form a 2.37° x 2.37° square. Each half-rectangle was made up of one of the basic colors of the feature condition. In order to ensure that no item would have a unique color with a given display, we restricted ourselves to a subset of the potential color combinations. The allowed conjunctions were red-blue, red-green, red-yellow, yellow-blue, yellow-green, green-blue, black-white, black-purple, black-orange, white-orange, white-purple and orange-purple. Each conjunction could appear in two mirror-reversed versions (e,g. red-blue and blue-red were different stimuli). See Figure 5 for an example of the feature and conjunction stimuli we used.
The experiment consisted of two sessions of 8 blocks of 30 trials each, for a total of 480 trials per participant. Each session took approximately 75 minutes, and the sessions could be performed on separate days. Within a session all items were either defined by unique colors (feature session), or by color conjunctions (conjunction session). Each session comprised four unrepeated blocks and four repeated blocks, as defined in Experiment 1. Half of the participants started with the feature session, while the other half started with the conjunction session. Within each session, half of the participants started with the repeated condition, and the other half with the unrepeated condition. If a participant started with the repeated condition on the first session, she would also start with the repeated condition on the second session.
Figure 6 plots identity and location accuracy as a function of repetition and session. A two-way ANOVA on identity accuracy with the factors repetition (unrepeated versus repeated) and stimulus type (feature or conjunction) as independent variables, and identity accuracy as dependent variable revealed that there was a main effect of trial type (F(1,15) = 35.61, MSE = 71.96, p < 0.001, η2 = .704) as accuracy was higher in the feature condition, a main effect of repetition (F(1,15) = 58.06, MSE = 53.40, p < 0.001, η2 = .795) as accuracy was higher in the repeated blocks, and a significant interaction (F(1,15) = 5.27, MSE = 84.86, p < 0.05, η2 = .260) showing that the repetition benefit was larger for conjunctions than for features. The same pattern of results held for location accuracy, except now there was only a trend to a larger repetition benefit for conjunctions (F(1,15) = 2.91, MSE = 8.21, p = 0.11, η2 = .163). Furthermore, for both conjunctions and features, there was a significant repetition benefit, for both identity and location accuracy (Conjunctions: identity: F(1,15) = 36.50, MSE = 80.86, p < 0.001, η2 = .709; location: F(1,15) = 12.59, MSE = 15.54, p < 0.005, η2 = .456; Features: identity: F(1,15) = 10.39, MSE = 57.40, p < 0.01, η2 = .409; location: F(1,15) = 5.58, MSE = 8.96, p < 0.05, η2 = .271).
We found that location accuracy was significantly greater than identity accuracy in both conditions (Conjunctions: F(1,15) = 47.61, MSE = 131.86, p < 0.001, η2 = .760; Features: F(1,15) = 44.28, MSE = 58.29, p < 0.001, η2 = .747), demonstrating a binding problem even for simple color stimuli. We also replicated Makovski and Jiang's (2009) finding that tracking is better with feature stimuli than with conjunction stimuli.
Importantly, repetition improved tracking for conjunctions more than it improved tracking for features. This result directly contradicts the target recovery hypothesis, and is in accordance with the binding hypothesis. Since the binding load in the feature condition is lower than in the conjunction condition, repetition has a larger effect in the latter case.
However, we could account for these results without the binding hypothesis. Instead of assuming that repetition directly improves target recovery, perhaps repetition improves memory for the target set. This in turn improves tracking performance because the better an observer knows the target set, the less likely she is to accidentally start tracking a distractor instead of a target. Thus, prevention of target-distractor swaps is more robust.
This hypothesis can explain why performance improves more for conjunction stimuli than for feature stimuli (since the latter are already very easy to remember, and there is little room for improvement). We designed Experiment 4 to test this hypothesis.
If the repetition benefit is due to improved prevention of target-distractor swaps, we should be able to eliminate it by eliminating the distractors. One advantage of using unique stimuli for tracking is that we can do that without making the task completely trivial (see Experiment 7 of Horowitz, et al., 2007).
Eight participants took part as paid volunteers. Their ages ranged from 21 – 50 years (average 32.25 years), and all had normal or corrected-to-normal vision.
The stimuli and procedure were similar to those of Experiment 1, except that there were five targets and no distractors. The experiment consisted of 4 blocks of 30 trials and was performed without breaks. The entire experiment took approximately 45 minutes.
The results are plotted in Figure 7. There was a significant repetition benefit (F(1,7) = 9.754, MSE = 5.41, p < .05, η2 = .582). Thus, the target-distractor swap hypothesis cannot completely explain the repetition effect. However, note that the repetition benefit seems to be reduced in comparison to the other experiments (in this experiment the repetition benefit is approximately 5%, whereas in the other experiments it hovers around 15%). This suggests that repetition does serve to reduce the probability of target-distractor swaps. Nevertheless, there remains a significant repetition benefit even when there are no distractors. How can we explain this?
Perhaps repetition improves tracking performance by reducing the memory load. There is certainly evidence that MOT and visual working memory are related, from both neural and behavioral sources (Allen, et al., 2006; Drew & Vogel, 2008; Fougnie & Marois, 2006; Howe, Horowitz, Morocz, Wolfe, & Livingstone, 2009; Oksama & Hyona, 2004). Repetition might make it easier to encode stimuli or to transfer stimuli from working memory to long-term memory, thereby freeing up resources that could then be devoted to the tracking task. We designed Experiment 5 to test this hypothesis.
In Experiment 5, we contrasted the method used in the previous experiments (and Horowitz, et al., 2007; Makovski & Jiang, 2009; Oksama & Hyona, 2004; Oksama & Hyönä, 2008), in which identities are constantly visible, with the original method of Pylyshyn (2004), in which identities are presented at the start of the trial and then masked during the motion phase. If repetition improves MIT via processes unrelated to the tracking of moving identities (such as reducing the memory load) then the repetition benefit should also show up in the masked condition. In contrast, if the repetition benefit derives primarily from processes involved in identity tracking, such as more efficient binding of identities and positions during tracking, then it should only be observed when the identities are visible during tracking.
Note that it is likely that identity accuracy will be greater in the visible identities condition, regardless of which account is correct. However, the memory load hypothesis predicts a larger repetition benefit in the masked condition, whereas the improved identity tracking hypothesis expects a larger repetition benefit when all items are visible during the motion phase.
Sixteen participants took part as paid volunteers. Their ages ranged from 21 to 50 (average 27.8) years, and all had normal or corrected-to-normal vision.
There were two conditions, visible and masked. The visible condition was identical to Experiment 1. The masked condition was similar except that during the motion phase, the items were replaced with the same masks used during the response phase (see Figure 8).
The experiment consisted of two sessions. Each session started with a practice block of 10 trials, followed by 4 blocks of 30 trials. Half of the participants started with the visible session in which the identities of all items remained visible throughout the trial, the other half of the participants started with the masked session, in which the identities were visible for eight seconds before the start of the trial, but once the items started moving identities disappeared. Within each session half of the participants started with the repeated condition, and the other half with the unrepeated condition. If a participant started with the repeated condition on the first session, she would also start with the repeated condition on the second session (and similarly if she would start with the unrepeated condition). The sessions were performed on different days. One session would be performed without breaks and took approximately 50 minutes.
The data are presented in Figure 9. We analyzed the results for identity and location accuracy separately. In both cases, a two-way ANOVA with factors condition (masked vs. visible) and repetition (repeated vs. unrepeated) revealed a main effect for condition (identity: F(1,15) = 54.43, MSE = 128.58, p < .0001, η2 = .784; location: F(1,15) = 51.96, MSE = 54.94, p < .0001, η2 = .776) as performance was higher when all items were visible, a main effect for repetition (identity: F(1,15) = 29.68, MSE = 42.00, p < 0.001, η2 = .664; location: F(1,15) = 18.72, MSE = 19.62, p = 0.001, η2 = .555) as performance was higher when the targets were repeated, and a significant interaction (identity: F(1,15) = 10.71, MSE = 50.74, p = 0.005, η2 = .417; location: F(1,15) = 6.70, MSE = 13.76, p < 0.05, η2 = .309) as repetition improved performance more in the visible than in the masked condition. Follow-up tests revealed that, for both identity and location accuracy, there was a repetition benefit in the visible condition (Fs > 26.3, ps < 0.001, η2 > .64), but no significant repetition benefit in the masked condition (Fs < 2.55, ps > .13, η2 < .15). We then performed a two-way ANOVA on the results of the visible condition, with the factors condition (repeated or unrepeated) and accuracy (location or identity). This interaction contrast was significant (F(1,15) = 14.41, MSE = 15.43, p < .005, η2 = .490), indicating that the increase in identity accuracy was larger than the increase in location accuracy, thereby replicating the results of Experiment 1. A similar analysis of the masked condition revealed no significant effects. These data disconfirm the memory load hypothesis, and are in line with the predictions of the binding hypothesis.
One problem with interpreting these results is that overall accuracy was lower in the masked condition. Since we don’t know the shape of the accuracy x difficulty curve for this task, it is possible, if not necessarily likely, that the masked condition lies in a domain where accuracy is less sensitive to difficulty than in domain of the visible condition data. Therefore we ran a control experiment with eight paid volunteers, where we adjusted the speed in order to produced equivalent accuracy in both conditions. Observers ran in a block of 60 trials of the repeated condition in which speed was adjusted using a QUEST routine (King-Smith, Grigsby, Vingrys, Benes, & Supowit, 1994; Watson & Pelli, 1983) to obtain 85% accuracy. We then ran all four conditions of the main experiment at this speed (visible condition: mean = 13.7°/s (standard error = 2.4°/s), masked condition: 5.9°/s (1.1°/s)).
The control experiment replicated the results of the main experiment (see Figure 10). With regards to identity accuracy, we again found a significant interaction, indicating that the repetition benefit was larger in the visible than in the masked condition (F(1,7) = 6.99, MSE = 173.66, p < 0.05, η2 = .500), a repetition benefit in the visible condition, but not in the masked case (F(1,7) = 6.51, MSE = 426.59, p < 0.05, η2 = .482; and F < .3, p > 0.6, η2 < .05 respectively). With regards to location accuracy the data trended in the same direction, but none of the effects reached significance (interaction, indicating a larger repetition benefit in the visible condition: F(1,7) = 4.20, MSE = 3.49, p = 0.08, η2 = .375; visible condition, repetition benefit: F(1,7) = 2.136, MSE = 2.14, p = 0.19, η2 = .234; masked condition, repetition deficit: F(1,7) = 1.379, MSE = 0.87, p = 0.28, η2 = .165).
In five experiments we consistently found that repeating target identities improved both identity and position tracking. In Experiment 1, using cartoon animals, we found a solid repetition benefit. We found that identity accuracy was more than 15% higher with repeated than with unrepeated targets. Experiment 2 showed that even when distractor and target identities were never swapped, performance was still better when targets were repeated. In Experiment 3 we found that the repetition benefit was larger when the identities were defined by color-color conjunctions than if the identities were defined by unique colors. Experiment 4 showed that the repetition benefit was still observed even when all stimuli were targets. Finally, Experiment 5 showed that the repetition benefit is strongly reduced when target identities are not visible during the motion phase.
We proposed four hypotheses to explain the repetition benefit. First, in the repeated condition target and distractor roles are consistent across trials, whereas in the unrepeated condition, targets on one trial could become distractors on a subsequent trial and vice versa. This consistent mapping might account for the repetition benefit. Experiment 2 compared the repeated condition to unrepeated conditions with consistent mapping. In these unrepeated conditions targets and distractors were selected from separate stimulus pools, assuring that targets could never become distractors (and vice versa). Experiment 2 still showed better tracking performance in the repeated condition. Thus, consistent mapping cannot explain the entire repetition benefit.
Second, repetition might improve tracking by improving the ability to recover lost targets. This explanation is difficult to reconcile with the results of Experiment 3, in which easy to search for feature stimuli show a smaller repetition benefit than difficult to search for color-color conjunctions. A modification of this hypothesis suggests that repetition allows observers to better prevent the swapping of a target for a distractor, because memory for the target set is improved. Experiment 4 showed that this hypothesis cannot fully explain the repetition benefit, since a benefit was observed even in a target-only tracking task, in which the probability of swapping targets and distractors is nil. Since the benefit was reduced relative to the previous experiments, we can conclude that repetition does help reduce target-distractor swaps, but this cannot be the whole story.
Third, repetition might reduce the memory load, freeing up shared resources which could then be diverted to the tracking task. This hypothesis was tested in Experiment 5 by masking the objects during the moving phase. If repetition improves tracking by improving memory, then this manipulation should increase the repetition benefit, or at least leave it unchanged. However, the manipulation reduced the repetition benefit, disconfirming this account.
The fourth hypothesis, consistent with all of the results, is that repetition improves performance, because it improves processes of identity tracking, either through improving recognition and maintenance of target identities, or through facilitating identity-position binding during tracking. For example, if we assume that MIT proceeds via a cyclical updating of identity + position bindings (Oksama & Hyönä, 2008), the repetition of target identities might facilitate recognition of identities by changing the long-term representation of the object (Ellis, Shepherd, & Davies, 1979; Leveroni, et al., 2000; Wolfe & Horowitz, 2004), reducing the time or neural resources needed to complete the identity-location binding operation on each cycle.
Summarizing, our data decisively show that repeating target identities increases performance on a MIT task. Experiment 5 shows that this repetition benefit is mainly due to processes involved when the different identities are in motion. This strongly suggests that repeating target identities improves identity tracking. Experiment 2, 3 and 4 are all consistent with the notion that this improved identity tracking is due to improved identity-location binding. However, our data do not exclude that other processes could also play a role.
In the current study we have found that you are better at tracking a stimulus when you have had more exposure to this stimulus. In general, when you see a stimulus more often, you become more familiar with this stimulus, but the two terms are not interchangeable. For example, in Experiment 1 of our study, participants may have become more familiar with the repeated cartoon animals than the cartoon stimuli in the unrepeated condition. However, it would be hard to defend the proposition that our participants were truly "unfamiliar" with a blue square simply because it was presented in the unrepeated condition of Experiment 3. On the other hand, while not all objects you have seen more often are more familiar, the reverse does hold. You have to have seen an object (or something similar to it) repeatedly before it could be considered familiar to you. This assertion is congruent with the findings of Oksama and Hyönä (2008). They compared tracking familiar to tracking unfamiliar objects (where familiar objects were real objects and unfamiliar objects were pseudo-objects), and observed an advantage for tracking familiar objects. They also replicated the finding using face stimuli where the "familiar" faces were famous faces (e.g. Albert Einstein and Saddam Hussein) likely to be well known to their observers. Recall that in the introduction we outlined that several mechanisms may have caused the observations of Oksama and Hyönä (2008), such as their familiar objects being more nameable. We speculate that the findings in both studies are caused by the same underlying mechanism, namely improved identity tracking (i.e. improved maintenance and recognition of identities or improved identity-location binding) of familiar (or repeated) objects.
One way to summarize our results is that tracking identities in addition to locations is effortful (because repeating identities makes it easier), and pulls some resources away from tracking locations alone (because we get significant increases in both location and identity accuracy).
These results rule out a simple class of models of multiple identity tracking in which identities come along for free as long as locations are tracked. This model, for example, was assumed by Pylyshyn (2004) when he designed his studies. It also underlies the "impletion" idea behind object file theory (Kahneman, et al., 1992). The object file theory suggests that directing attention to an object creates a file that contains all relevant data regarding this object, including its position. So the object file theory is equivalent to models where identities come along for free, but now, after identities have been acquired, locations get a free ride. In neither case would it be possible to trade off performance between identity and location tracking.
The fact that identity tracking is effortful (for instance because it requires updating position and identity bindings) is consistent with either a unitary account of MIT, in which the same neural system is responsible for tracking positions and identities or a dual account such as that proposed by Horowitz et al. (2007). However, if position tracking and binding are carried out by separate systems, then making binding easier should not affect position tracking. In all of our experiments, improved identity accuracy was accompanied by improved location accuracy. Thus, our data support a unitary account of MIT.
Our data are furthermore consistent with a limited resource model. If we assume that tracking positions and binding identities to positions are processes which share some neural resource, then if repeating targets makes binding more efficient, some resources are freed up for tracking, thus improving both identity and location performance. A limited resource model of MOT has been proposed by Alvarez and Franconeri (2007). According to this model, tracking is accomplished via an array of flexibly allocated indexes (or FLEXes), such that when few targets are tracked, they receive many indexes, and when many targets are tracked, each target receives fewer indexes (the FLEX model can also be implemented in a serial fashion). Thus tracking performance degrades smoothly as tracking load increases. However, under this account, the limited resource is really spatial resolution: when few targets are tracked, each target receives more resources (FLEXes), and the resolution increases so that it can be tracked at a higher speed. In order to explain our data, we have to add the assumption that the visual system can trade off spatial resolution for identity information (speculatively, in the form or featural resolution), so that when recognizing objects becomes easier, spatial resolution improves. Thus, these data suggest a more general limited resource account of MOT and MIT.
One way to understand this tradeoff might be to assume that we can trade off spatial resolution for resolution in a non-spatial dimension. For example, if observers track colored squares, making the tracking task more difficult might reduce the resolution of color information (see Zhang & Luck, 2008 for a method to measure color resolution).
Another approach is to assume that MIT consists of a series of update cycles. On each cycle, the system updates both the positions and the bindings. If binding is more efficient, the update cycles can proceed at an increased rate, leading to reduced position error, such as in the MOMIT model (Oksama & Hyönä, 2008).
Performance in a multiple identity tracking task improves when targets are repeated often, providing the targets are visible when they are in motion. This effect is most likely due to identity tracking improving because of target repetition. Our results indicate that tracking positions and identities are not independent tasks. Our data thus argue against simple models in which identity comes for free with position tracking, and also against dual-system models in which position tracking and identity tracking are accomplished by different neural hardware. Instead, we suggest that a limited resource model, in which position tracking and identity binding compete for the same resource pool, may explain tracking.