Five Factors that Guide Attention in Visual Search

Abstract

How do we find what we are looking for? Fundamental limits on visual processing mean that even when the desired target is in our field of view, we often need to search, because it is impossible to recognize everything at once. Searching involves directing attention to objects that might be the target. This deployment of attention is not random. It is guided to the most promising items and locations by five factors discussed here: Bottom-up salience, top-down feature guidance, scene structure and meaning, the previous history of search over time scales from msec to years, and the relative value of the targets and distractors. Modern theories of search need to specify how all five factors combine to shape search behavior. An understanding of the rules of guidance can be used to improve the accuracy and efficiency of socially-important search tasks, from security screening to medical image perception.

How can a texting pedestrian walk right into a pole, even though it is clearly visible ¹? At any given moment, our attention and eyes are focused on some aspects of the scene in front of us, while other portions of the visible world go relatively unattended. We deploy this selective visual attention because we are unable to fully process everything in the scene at the same time. We have the impression of seeing everything in front of our eyes, but over most of the visual field we are probably seeing something like visual textures, rather than objects ^2,3 Identifying specific objects and apprehending their relationships to each other typically requires attention, as our unfortunate texting pedestrian can attest.

Figure 1 illustrates this point. It is obvious that this image is filled with Ms and Ws in various combinations of red, blue, and yellow, but it takes attentional scrutiny to determine whether or not there is a red and yellow M.

Figure 1:

On first glimpse, you know something about the distribution of colors and shapes but not how those colors and shapes are bound to each other. Find ‘M’s that are red and yellow.

The need to attend to objects in order to recognize them raises a problem. At any given moment, the visual field contains a very large, possibly uncountable number of objects. We can count the Ms and Ws of Figure 1, but imagine looking at your reflection in the mirror. Are you an object? What about your eyes or nose or that small spot on your chin? If object recognition requires attention, and if the number of objects is uncountable, how do we manage to get our attention to a target object in a reasonable amount of time? Attention can process items at a rate of, perhaps, 20-50 items per second. If you were looking for a street sign in an urban setting containing a mere 1000 possible objects (every window, tire, door handle, piece of trash, etc.), it would take 20-50 seconds just to find that sign. It is introspectively obvious that you routinely find what you are looking for in the real world in a fraction of that time. To be sure, there are searches of the needle-in-a-haystack, Where’s Waldo? variety that take significant time, but routine searches for the saltshaker, the light switch, your pen, and so forth, obviously proceed much more quickly. Search is not overwhelmed by the welter of objects in the world because search is guided to a (often very small) subset of all possible objects by several sources of information. The purpose of this article is to briefly review the growing body of knowledge about the nature of that guidance.

We will discuss five forms of guidance:

1. Bottom-up, stimulus-driven guidance in which the visual properties of some aspects of the scene attract more attention than others.

2. Top-down, user-driven guidance in which attention is directed to objects with known features of desired targets.

3. Scene guidance in which attributes of the scene guide attention to areas likely to contain targets.

4. Guidance based on the perceived value of some items or features.

5. Guidance based on the history of prior search.

Measuring Guidance

We can operationalize the degree of guidance in a search for a target by asking what fraction of all items can be eliminated from consideration. One of the more straight-forward methods to do this is to present observers with visual search displays like those in Figure 2 and measure the reaction time (RT) required for them to report whether or not there is a target (here a “T), as a function of the number of items (set size). The slope of the RT x set size function is a measure of the efficiency of search. For a search for a T among Ls (Fig 2A), the slope would be in the vicinity of 20-50 msec/item ⁴. We believe that this reflects serial deployment of attention from item to item ⁵ though this need not be the case ⁶.

Figure 2:

The basic visual search paradigm. A target (here a ‘T‘) is presented amidst a variable number of distractors. Search ‘efficiency’ can be indexed by the slope of the function relating reaction time (RT) to the visual set size. If the target in 2B is a red T, the slope for 2B will be half of that for 2A because attention can be limited to just half of the items in 2B.

In Fig. 2B, the target is a red T. This search would be faster and more efficient ⁷ because attention can be guided to the red items. If half the items are red (and if guidance is perfect), the slope will be reduced by about half, suggesting that, at least in this straightforward case, slopes index the amount of guidance.

The relationship of slopes to guidance is not entirely simple, even for arrays of items like those in Fig 2 ⁸ but see ⁹. Matters become far more complex with real world scenes where the visual set size is not easily defined ^10,11. However, if the slope is cut in half when half the items acquire some property, like the color red in 2B, it is reasonable to assert that search has been guided by that property ⁹.

The problem of distractor rejection

As shown in Figure 2, a stimulus attribute can make search slopes shallower by limiting the number of items in a display that need to be examined. However, guidance of attention is not the only factor that can modulate search slopes. If observers are attending to each item in the display (in series or in parallel), the slope of the RT x set size function can also be altered by changing how long it takes to reject each distractor. Thus, if we markedly reduced the contrast of Figure 2A, the RT x set size function would become steeper, not because of a change in guidance but because it would now take longer to decide if any given item was a T or an L.

Bottom-up guidance by stimulus salience

Attention is attracted to items that differ from their surroundings, if those differences are large enough and if those differences occur in one of a limited set of attributes that guide attention. The basic principles are illustrated in Figure 3.

Figure 3:

Which items ‘pop-out’ of this display, and why?

Three items ‘pop-out’ of this display. The purple item on the left differs from its neighbors in color. It is identical to the purple item just inside the upper right corner of the image. That second purple item is not particularly salient even though it is the only other item in that shade of purple; its neighbors are close enough in color that the differences in color do not attract attention. The bluish item to its left is salient by virtue of an orientation difference. The square item a bit further to the left is salient because of the presence of a ‘closure’ feature ¹² or the absence of a collection of line terminations ¹³. We call properties like color, orientation, or closure basic (or guiding) features, because they can guide the deployment of attention. Other properties may be striking when one is directly attending to an item, and may be important for object recognition, but they do not guide attention. For example, the one ‘plus’ in the display is not salient, even though it possesses the only X-intersection in the display, because intersection type is not a basic feature ¹⁴. The ‘pop-out’ we see in Figure 3 is not just subjective phenomenonology. Pop-out refers to extremely effective guidance, and is diagnosed by a near-zero slope of the RT x set size function; though there may be systematic variability even in these ‘flat’ slopes ¹⁵.

There are two fundamental rules of bottom-up salience ¹⁶. Salience of a target increases with difference from the distractors (target-distractor – TD- heterogeneity) and with the homogeneity of the distractors (distractor-distractor –DD- homogeneity) along basic feature dimensions. Bottom-up salience is the most extensively modeled aspect of visual guidance nicely reviewed in ¹⁷. The seminal modern work on bottom-up salience is Koch and Ullman’s ¹⁸ description of a winner-take-all network for deploying attention. Subsequent decades have seen the development of several influential bottom-up models e.g. ^19,20–22. However, bottom-up salience is just one of the factors guiding attention. By itself, it does only modestly well in predicting the deployment of attention (usually indexed by eye fixations). Models do quite well predicting search for salience, but not as well predicting search for other sorts of targets ¹⁷. This is quite reasonable. If you are looking for your cat in the bedroom, it would be counterproductive to have your attention visit all the shiny, colorful objects first. Thus, a bottom-up saliency model will not do well if the observer has a clear top-down goal ²³. One might think that bottom-up salience would dominate if observers free-viewed a scene in the absence of such a goal, but bottom-up models can be poor at predicting fixations even when observers “free view” scenes without specific instructions ²⁴. It seems that observers generate their own, idiosyncratic tasks, allowing other guiding forces to come into play. It is worth noting that salience models work better if they are not based purely on local features but acknowledge the structure of objects in the field of view ²⁵. For instance, while the most salient spot in an image might be the edge between the cat’s tail and the white sheet on the bed, fixations are more likely to be directed to middle of the cat ^26,27.

Top-down Feature Guidance

Returning to Figure 1, if you search for Ws with yellow elements, you can guide your attention to yellow items and subsequently determine if they are Ws or Ms ⁷. This is feature guidance, sometimes referred to as feature-based attention ²⁸. Importantly, it is possible to guide attention to more than one feature at a time. Thus, search for a big, red, vertical item can benefit from our knowledge of its color, size, and orientation ²⁹. Following the TD heterogeneity rule, search efficiency is dependent on the number of features shared by targets and distractors ²⁹, and observers appear to be able to guide to multiple target features simultaneously ³⁰. This finding raises the attractive possibility that search for an arbitrary object among other arbitrary objects would be quite efficient because objects would be represented sparsely in a high-dimensional space. Such sparse coding has been invoked to explain object recognition ^31,32. However, search for arbitrary objects turns out not to be particularly efficient ^11,33. By itself, guidance to multiple features does not appear to be an adequate account of how we search for objects in the real world (see the section on scene guidance, below).

What are the guiding attributes?

Feature guidance bears some metaphorical similarity to your favorite computer search engine. You enter some terms into the search box and an ordered list of places to attend is returned. A major difference between internet search engines and the human visual search engine is that human search uses only a very small vocabulary of search terms (i.e., features). The idea that there might be a limited set of features that could be appreciated “preattentively” ³⁴ was at the heart of Treisman’s “Feature Integration Theory” ³⁵. She predicted that targets defined by unique features would pop-out of displays. Subsequent theorists modified this proposal to suggest that features could guide the deployment of attention ^{7 36}.

There are probably only a couple dozen attributes that can guide attention. The visual system can detect and identify a vast number of stimuli, but it cannot use arbitrary properties to guide attention the way that Google or Bing can use arbitrary search terms. A list of guiding attributes is found in Table 1. This article does not list all of the citations that support each entry. Many of these can be found in older versions of the list ^37,38. Recent changes to the list are marked in color in Table 1 and citations are given for those.

Table 1:

The guiding attributes for feature search

Undoubted Guiding Attributes
Color	Motion	Orientation	Size (incl. length, spatial freq., and apparent size 39)
Probable Guiding Attributes
Luminance onset (flicker) but see 40	Luminance polarity	Vernier offset	Stereoscopic depth & tilt
Pictorial depth cues But see 41.	Shape	Line termination	Closure
Curvature	Topological status
Possible Guiding Attributes
Lighting direction (shading)	Expansion / Looming	Number	Glossiness (luster)
Aspect ratio	eye of origin / binocular rivalry
Doubtful cases
Novelty	Letter Identity Alphanumeric Category	Familiarity – over-learned sets, in general 42
Probably Not Guiding Attributes
Intersection	Optic flow	Color change	3-D volumes (eg. geons)
Luminosity	Material type	Scene Category	Duration
Stare-in-crowd 43,44	Biological motion	Your name	Threat
Semantic Category (Animal, artifact, etc)	Blur 45	Visual rhythm 46	Animacy/Chasing 47
Threat 48
Faces are a complicated issue
Faces among other objects	Familiar Faces	Emotional faces	Schematic Faces
Factors that modulate search
Cast shadows	Amodal Completion	Apparent Depth

Attributes like color are deemed to be “undoubted” because multiple experiments from multiple labs attest to their ability to guide attention. “Probable” feature dimensions may be merely probable because we are not sure how to define the feature. Shape is the most notable entry here. It seems quite clear that something about shape guides attention ⁴⁹. It is less clear exactly what that might be, though the success of deep learning algorithms in enabling computers to classify objects may open up new vistas for understanding human search for shape ⁵⁰.

The attributes described as “possible” await more research. Often these attributes only have a single paper supporting their entry on the list, as in the case of numerosity: Can you direct attention to the pile with “more” elements in it, once you eliminate size, density, and other confounding visual factors? Perhaps ⁵¹, but it would be good to have converging evidence. Search for the magnitude of a digit (e.g. “find the highest number”) is not guided by the semantic meaning of the digits but by their visual properties ⁵²

The list of attributes that do not guide attention is, of course, potentially infinite. Table 1 lists a few plausible candidates that have been tested and found wanting. For example, there has been considerable interest recently in what could be called “evolutionarily motivated” candidates for guidance. What would enhance our survival if we could find it efficiently? Looking at a set of moving dots on a computer screen, we can perceive that one is “chasing” another ⁵³. However, this aspect of animacy does not appear to be a guiding attribute ⁴⁷. Nor does “threat” (defined by association with electric shock) seem to guide search ⁴⁸.

Some caution is needed here because a failure to guide is a negative finding and it is always possible that, were the experiment done correctly, the attribute might guide after all. Thus, early research ⁵⁴ found that binocular rivalry and eye-of-origin information did not guide attention, but more recent work ^55,56 suggests that it may be possible to guide attention to interocular conflict, and our own newer data ⁵⁷ indicates that rivalry may guide attention if care is taken to suppress other signals that interfere with that guidance. Thus, binocular rivalry was listed under “doubtful cases & probable non-features” in ³⁷, but is now listed under “possible guiding attributes” in Table 1.

Faces remain a problematic candidate for feature status, with a substantial literature yielding conflicting results and conclusions. Faces are quite easy to find among other objects ^58,59 but there is dispute about whether the guiding feature is “face-ness” or some simpler stimulus attribute ^60,61. A useful review by Frischen et al. ⁶² argues that “preattentive search processes are sensitive to and influenced by facial expressions of emotion”, but this is one of the cases where it is hard to reject the hypothesis that the proposed feature is modulating the processing of attended items, rather than guiding the selection of which items to attend. Suppose that, once attended, it takes 10 msec longer to disengage attention from an angry face than from a neutral face. The result would be that search would go faster (10 msec/item faster) when the distractors were neutral than when they were angry. Consequently, an angry target among neutral distractors would be found more efficiently than a neutral face among angry. Evidence for guidance by emotion would be stronger if the more efficient emotion searches were closer to pop-out than to classic inefficient, unguided searches, e.g., T among Ls,⁶³. Typically, this is not the case. For example, Gerritsen et al ⁶⁴ report that “Visual search is not blind to emotion” but, in a representative finding, search for hostile faces produced a slope of 64 msec/item, which is quite inefficient, even if somewhat more efficient than the 82 msec/item for peaceful target faces (p1054).

There are stimulus properties that, while they may not be guiding attributes in their own right, do modulate the effectiveness of other attributes. For example, apparent depth modulates apparent size, and search is guided by that apparent size ⁶⁵. Finally, there are properties of the display that influence the deployment of attention. These could be considered aspects of “scene guidance” (see the next major section, below). For example, attention tends to be attracted to the center of gravity in a display ⁶⁶. Elements like arrows direct attention even if they, themselves do not pop-out ⁶⁷. As discussed by Rensink ⁶⁸, these and related factors can inform graphic design and other situations where the creator of an images wants to control how the observer consumes that image.

There have been some general challenges to the enterprise of defining specific features, notably the hypothesis that many of the effects attributed to the presence or absence of basic features are actually produced by crowding in the periphery ³. For example, is efficient search for cubes lit from one side among cubes lit from another side evidence for preattentive processing of 3D shape and lighting ⁶⁹, or merely a by-product of the way these stimuli are represented in peripheral vision ⁴¹? Resolution of this issue requires a set of visual search experiments with stimuli that are “uncrowded”. This probably means using low set sizes; for example, see the evidence that material type is not a guiding attribute ⁷⁰.

A different challenge to the preattentive feature enterprise is the possibility that too many discrete features are proposed. Perhaps many specific features form a continuum of guidance by a single, more broadly defined attribute. For instance, the cues to the 3D layout of the scene include stereopsis, shading, linear perspective and more. These might be part of a single attribute describing the 3D disposition of an object. Motion, onsets, and flicker might be part of a general dynamic change property ⁷¹. Most significantly, we might combine the spatial features of line termination, closure, topological status, orientation, and so forth into a single shape attribute with properties defined by the appropriate layer of the right convolutional neural net (CNN). Such nets have shown themselves capable of categorizing objects, so one could imagine a preattentive CNN guiding attention to objects as well ⁷². At this writing, such an idea remains a promissory note. Regardless of how powerful CNNs may become, humans cannot guide attention to entirely arbitrary/specific properties in order to find particular types of object ⁷³ and it is unknown if some intermediate representation in a CNN could capture the properties of the human search engine. If it did, we might well find that such a layer represented a space with dimensions corresponding to attributes like size, orientation, line termination, vernier offset, and so forth, but this remains to be seen.

Guidance by scene properties

While the field of visual search has largely been built on search for targets in arbitrary 2D arrays of items, most real world search takes place in structured scenes, and this structure provides a source of guidance. To illustrate, try search for any humans in Figure 4. Depending on the resolution of the image as you are viewing it, you may or may not be able to see legs poking out from behind the roses by the gate. Regardless, what should be clear is that the places you looked were strongly constrained. Biederman, Mezzanotte, and Rabinowitz ⁷⁴ suggested a distinction between semantic and syntactic guidance.

Figure 4:

Scene Guidance: Where is attention guided if you are looking for humans? What if the target was a bird?

Syntactic guidance has to do with physical constraints. You don’t look for people on the front surface of the wall or in the sky because people typically need to be supported against gravity. Semantic guidance refers to the meaning of the scene. You don’t look for people on the top of the wall, not because they could not be there but because they are unlikely to be there given your understanding of the scene, whereas you might scrutinize the bench. Scene guidance would be quite different (and less constrained) if the target were a bird. The use of the terms “semantic” and “syntactic” should not be seen as tying scene processing too closely to linguistic processing, nor should the two categories be seen as neatly non-overlapping ^{75 76}. Nevertheless, the distinction between syntactic and semantic factors, as roughly defined here, can be observed in electrophysiological recordings: scenes showing semantic violations (e.g., a bar of soap sitting next to the computer on the desk) produce different neural signatures than scenes showing syntactic violations (e.g., a computer mouse on top of the laptop screen) ⁷⁷. While salience may have some influence in this task ⁷⁸, it does not appear to be the major force guiding attention here ^24,79. But note that feature guidance and scene guidance work together. People certainly could be on the lawn, but you do not scrutinize the empty lawn in Figure 4 because it lacks the correct target features.

Extending the study of guidance from controlled arrays of distinct items to structured scenes poses some methodological challenges. For example, how do we define the set size of a scene? Is “rose bush” an item in Figure 4, or does each bloom count as an item? In bridging between the world of artificial arrays of items and scenes, perhaps the best we can do is to talk about the “effective set size” ^{80 10}, the number of items/locations that are treated as candidate targets in a scene give a specific task. If you are looking for the biggest flower, each rose bloom is part of the effective set. If you are looking for a human, those blooms are not part of the set. While any estimate of effective set size is imperfect, it is a very useful idea and it is clear that, for most tasks, the effective set size will be much smaller than the set of all possible items ¹¹.

Preview methods have been very useful in examining the mechanisms of scene search ⁸¹. A scene is flashed for a fraction of a second and then the observer searches for a target. The primary data are often eye tracking records. Often, these experiments involve searching while the observer’s view of the scene is restricted to a small region around the point of fixation (“gaze-contingent” displays). Very brief exposures (50-75 msec) can guide deployment of the eyes once search begins ⁸². A preview of the specific scene is much more useful than a preview of another scene of the same category, though the preview scene does not need to be the same size as the search stimulus ⁸¹. Importantly, the preview need not contain the target in order to be effective ⁸³. Search appears to be more strongly guided by a relatively specific scene ‘gist’ ^{80 84}, an initial understanding of the scene that does not rely on recognizing specific objects ⁸⁵. The gist includes both syntactic (e.g., spatial layout) and semantic information, and this combination can provide powerful search guidance. Knowledge about the target provides an independent source of guidance ^86,87. These sources of information provide useful ‘priors’ on where targets might be (“If there is a vase present, it’s more likely to be on a table than in the sink”), that are more powerful than memory for where a target might have been seen ^{88 89,90} in terms of guiding search.

Preview effects may be fairly limited in search of real scenes. If the observer searches a fully visible scene rather than being limited to a gaze-contingent window, guidance by the preview is limited to the first couple of fixations ⁹¹. Once search begins, guidance is presumably updated based on the real scene, rendering the preview obsolete. In gaze-contingent search, the effects last longer because this updating cannot occur. This updating can be seen in the work of Hwang et al. ⁷⁶, where, in the course of normal search, the semantic content of the current fixation in a scene influences the target of the next fixation.

Modulation of search by prior history

In this section, we summarize evidence showing that the prior history of the observer, especially the prior history of search, modulates the guidance of attention. We can organize these effects by their time scale, from within a trial (on the order of 100s of ms) to lifetime learning (on the order of years).

A number of studies have demonstrated the preview benefit: when half of the search array is presented a few hundred msec before the rest of the array, the effective set size is reduced, either because attention is guided away from the old “marked” items (visual marking ⁹²) and/or toward the new items (onset prioritization ⁹³).

On a slightly longer timescale, priming phenomena are observed from trial to trial within an experiment, and can be observed over seconds to weeks. The basic example is “priming of pop-out” ⁹⁴, in which an observer might be asked to report the shape of the one item of unique color in a display. If that item is the one red shape among green on one trial, responses will be faster if the next trial repeats red among green as compared to a switch to green among red; though the search in both cases will be a highly efficient, color pop-out search. More priming of pop-out is found if the task is harder ⁹⁵. Note that it is neither the response nor the reporting feature which is repeated in priming of pop-out, but the target-defining or selection feature.

More generally, seeing the features of the target makes search faster than reading a word cue describing the target, even for overlearned targets. This priming by target features takes about 200 msec to develop ⁹⁶. Priming by the features of a prior stimulus can be entirely incidental; simply repeating the target from trial to trial is sufficient ⁹⁷. More than one feature can be primed at the same time ^97,98 and both target and distractor features can be primed ^97,99. Moreover, it is not just that observers are more ready to report targets with the primed feature; priming actually boosts sensitivity (i.e., d’) ¹⁰⁰. Such priming can last for at least a week ¹⁰¹.

Observers can also incidentally learn information over the course of an experiment that can guide search. In contextual cueing ¹⁰², a subset of the displays are repeated across several blocks of trials. While observers do not notice this repetition, RTs are faster for repeated displays than for novel, unrepeated displays ¹⁰³. The contextual cueing effect is typically interpreted as an abstract form of scene guidance: just as you learn that, in your friend’s kitchen, the toaster is on the counter next to the coffeemaker, you learn that, in this configuration of rotated Ls, the T is in the bottom left corner. However, evidence for this interpretation is mixed. RT x set size slopes are reduced for repeated displays ¹⁰² in some experiments, but not in others ¹⁰⁴. Contextual cueing effects can also be observed in cases such as pop-out search and ¹⁰⁵, attentionally-cued search, ¹⁰⁶, where guidance is already nearly perfect. Kunar et al. ¹⁰⁴ suggested that contextual cueing reflects response facilitation, rather than guidance. Again, the evidence is mixed. There is a shift towards a more liberal response criterion for repeated displays ¹⁰⁷, but this is not correlated with the size of the contextual cueing RT effect. In pop-out search, sensitivity to the target improves for repeated displays without an effect on decision criterion ¹⁰⁵. It seems likely that observed contextual cueing effects reflect a combination of guidance effects and response facilitation, the mix depending on the specifics of the task. Oculomotor studies show that the context is often not retrieved and available to guide attention until a search has been underway for several fixations ^108,109. Thus, the more efficient the search, the greater the likelihood that the target will be found before the context can be retrieved. Indeed, in simple letter displays, search does not become more efficient even when the same display is repeated several hundred times ¹¹⁰, presumably because searching de novo is always faster than waiting for context to become available. Once the task becomes more complex (e.g., searching for that toaster) ¹¹¹, it becomes worthwhile to let memory guide search ^{112 113}.

Over years and decades, we become intimately familiar with, for example, the characters of our own written language. There is a long-running debate about whether familiarity (or, conversely, novelty) might be a basic guiding attribute. Much of this work has been conducted with overlearned categories like letters. While the topic is not settled, semantic categories like “letter” probably do not guide attention ^114,115, though mirror-reversed letters may stand out against standard letters ^{116 117}. Instead, items made familiar in LTM can modulate search ^42,118 though there are limits on the effects of familiarity in search ^{119 120}.

Modulation of search by the value of items

In the past few years, there has been increasing interest in the effects of reward or value on search. Value proves to be a strong modulator of guidance. For instance, if observers are rewarded more highly for red items than for green, they will subsequently guide attention toward red, even if this is irrelevant to the task ¹²¹. Note, color is the guiding feature; value modulates its effectiveness. The learned associations of value do not need to be task relevant or salient in order to have their effects ¹²² and learning can be very persistent with value-driven effects being seen half a year after acquisition ¹²³. Indeed, the effects of value may be driving some of the long-term familiarity effects described in the previous paragraph ⁴².

Visual search is mostly effortless. Unless we are scrutinizing aerial photographs for hints to North Korea’s missile program, or hunting for signs of cancer in a chest radiograph, we typically find what we are looking for in seconds or less. This remarkable ability is the result of attentional guidance mechanisms. While thirty-five years or so of research has given us a good grasp of the mechanisms of bottom-up salience, top-down feature-driven guidance and how those factors combine to guide attention ^124,125, we are just beginning to understand how attention is guided by the structure of scenes and the sum of our past experiences. Future challenges for the field will include understanding how discrete features might fit together in a continuum of guidance and extending our theoretical frameworks from two-dimensional scenes to immersive, dynamic, three-dimensional environments.