Object-based attentional selection in scene viewing

Abstract

Two contrasting views of visual attention in scenes are the visual salience and the cognitive relevance hypotheses. They fundamentally differ in their conceptualization of the visuospatial representation over which attention is directed. According to the saliency model, this representation is image-based, while the cognitive relevance framework advocates an object-based representation. Previous research has shown that (1) viewers prefer to look at objects over background and that (2) the saliency model predicts human fixation locations significantly better than chance. However, it could be that saliency mainly acts through objects. To test this hypothesis, we investigated where people fixate within real objects and saliency proto-objects. To this end, we recorded eye movements of human observers while they inspected photographs of natural scenes under different task instructions. We found a preferred viewing location (PVL) close to the center of objects within naturalistic scenes. Compared to the PVL for real objects, there was less evidence for a PVL for human fixations within saliency proto-objects. There was no evidence for a PVL when only saliency proto-objects that did not spatially overlap with annotated real objects were analyzed. The results suggest that saccade targeting and, by inference, attentional selection in scenes is object-based.

Introduction

What we see and understand about the visual world around us is tightly related to where we place our eyes. Eye movements thus serve as a window into the operation of the attentional system (Henderson, 2003; Rayner, 1998, 2009). Recent research has emphasized two general classes of factors that may control eye guidance in scenes: bottom-up image properties and cognitive knowledge structures used in a top-down manner. The present research is concerned with the following question: What is the role of higher-level stimulus structure, such as objects, compared to low-level image features, such as contrast, color, and orientation, in attentional guidance? Specifically, we ask whether viewers select and prioritize objects in scenes as opposed to saliency proto-objects. 

Understanding the processes that determine where we attend and look in scenes is a fundamental goal in studying scene perception (Henderson & Hollingworth, 1998; Rayner, 1998, 2009). Theoretical models of visual attention allocation in scenes can be contrasted as primarily bottom-up versus top-down, with these views forming the extremes of a continuum (e.g., Baddeley & Tatler, 2006; Itti & Koch, 2000; Parkhurst & Niebur, 2003; Reinagel & Zador, 1999; Tatler, Baddeley, & Vincent, 2006; Torralba, Oliva, Castelhano, & Henderson, 2006; Underwood & Foulsham, 2006). The dominant view in the computational vision literature has been the visual saliency hypothesis according to which bottom-up stimulus-based information generated from the image drives the allocation of visual attention and thus the placement of fixations in a scene. The most prominent model is the computational model of saliency developed by Itti and Koch (2000, 2001; see also Itti, Koch, & Niebur, 1998) and based on ideas proposed by Koch and Ullman (1985). The model analyzes natural images in terms of intensity, color, and orientation. For each of these visual features, a separate conspicuity map is computed by searching for change relative to adjacent regions. These separate maps are then combined into a unified single saliency map. For a given scene region, a change in any of the three general features results in an increase in the saliency value assigned to that region of the image. The values from the saliency map, in combination with a winner-take-all network and inhibition of return, produce a sequence of predicted fixations that scan the scene in order of decreasing saliency. 

A growing body of research suggests that fixation placement in scene viewing is also strongly affected by cognitive factors. A compelling demonstration of top-down influences on gaze control during scene viewing originates from task effects. Early qualitative demonstrations (Buswell, 1935; Yarbus, 1967) and recent controlled experiments (Castelhano, Mack, & Henderson, 2009) indicate that viewing task biases the choice of fixation locations in scenes. Interestingly, while there is a correlation of saliency with fixations during free viewing (e.g., Parkhurst, Law, & Niebur, 2002; Peters, Iyer, Itti, & Koch, 2005), recent studies suggest that during visual search, saliency has only a minor or no impact on fixation patterns (e.g., Foulsham & Underwood, 2007; Henderson, Brockmole, Castelhano, & Mack, 2007; Henderson, Malcolm, & Schandl, 2009). 

Recent updates of the bottom-up saliency model include some top-down information (Navalpakkam & Itti, 2005). Another model combining both bottom-up and top-down information is the Contextual Guidance model for object search in real-world scenes (Torralba et al., 2006). In this model, fixation selection is driven from a bottom-up salience computation modulated by contextual priors or task constraints: A salience-plus-context searcher is assumed to fixate the peak of the bottom-up salience map, weighted by the context-based prior (Torralba et al., 2006). On the other side of the continuum is the cognitive relevance hypothesis, according to which fixation locations are selected based on the needs of the cognitive system in relation to the current task (Henderson et al., 2007, 2009). According to this hypothesis, eye movements are primarily controlled by task goals interacting with a semantic analysis of the scene and memory for comparable viewing episodes (Hayhoe & Ballard, 2005). Although a representation of potential attention and fixation targets is also generated from the visual input, locations in this representation are “flat” in the sense that they are not ranked on the basis of visual salience. Once a figure–ground representation has been generated, saccade targets are ranked based on their cognitive relevance and attention is then directed to regions according to the ranking (see Henderson et al., 2007, 2009, for discussion). Consistent with the cognitive relevance hypothesis, a recent study reported that nonsalient search targets were fixated more frequently and earlier than highly salient but search-irrelevant regions (Henderson et al., 2009). 

The visual salience and cognitive relevance hypotheses fundamentally differ in their conceptualization of the visuospatial representation over which attention is directed. According to the saliency model, this representation is image-based, while the cognitive relevance framework advocates an object-based representation. According to the cognitive relevance hypothesis, the scene input is used to generate a flat (unranked) visuospatial landscape of potential object regions. If the two hypotheses differ in such a fundamental aspect, why is there empirical evidence supporting both visual salience and cognitive relevance? This could be due to a correlation between salient regions and objects (e.g., Einhäuser, Spain, & Perona, 2008; Henderson et al., 2007). First of all, viewers prefer to look at objects over background (Buswell, 1935; Henderson, 2003; Yarbus, 1967). Second, objects differ from scene background in their image properties. As a consequence, the predictive power of the saliency model could be mainly due to salience being a (weak) proxy for objects. With the present work, we test this alternative hypothesis: Rather than selecting salient regions, viewers select and prioritize objects in scenes. As a novel approach, we investigate landing positions ¹ within fixated objects or regions to make inferences about attentional selections in scenes. 

First, we investigate landing positions within fixated objects in photographs of real-world scenes. If attentional selection in real-world scenes is object-based, viewers should prefer to saccade to the center of objects. This working hypothesis is motivated by the finding of a Preferred Viewing Location (PVL, Rayner, 1979; see also Rayner, Liversedge, Nuthmann, Kliegl, & Underwood, 2009) in reading, which was shown to generalize to an object-viewing task (Henderson, 1993). Compared to picture viewing, reading takes place in a highly structured visual environment. The “scene” consists of well-defined word objects, and most of the time the eyes simply move from left to right through a line of text. Current theories of gaze control in reading assume that eye guidance in reading is word-based (cf., Radach & Kennedy, 2004). The central piece of evidence in support of word-based eye guidance is the observation of a PVL (see Yang & Vitu, 2007, for discussion). The PVL is derived from calculating the distributions of letter-based landing positions within words. These distributions resemble truncated Gaussian distributions with a mean that is typically at or somewhat left of word center. The mean of the distribution is the PVL. The phenomenon has been replicated many times (e.g., McConkie, Kerr, Reddix, & Zola, 1988; McDonald & Shillcock, 2004; Nuthmann, Engbert, & Kliegl, 2005; Vitu, O'Regan, & Mittau, 1990). Accordingly, most reading theories (e.g., E-Z Reader: Reichle, Rayner, & Pollatsek, 2003; SWIFT: Engbert, Nuthmann, Richter, & Kliegl, 2005) assume that the eyes target the center of the word as the optimal viewing position (OVP). The variance in landing positions around the OVP/PVL is thought to be due to visuomotor constraints (McConkie et al., 1988). Empirical research has shown that the length of the fixated word has a moderate effect on the PVL: For longer words, the PVL moves somewhat to the left, toward a location between the beginning and the middle of a word (Rayner, 1979). By far the largest influence on mean and standard deviation of the Gaussian PVL curve is exhibited by launch site distance (McConkie et al., 1988; Nuthmann et al., 2005; Rayner, Sereno, & Raney, 1996). Finally, the parameters of the PVL slightly differ for different types of fixations like first-pass single fixations, refixation cases, or fixations following inter-word regressions (e.g., McDonald & Shillcock, 2004; Nuthmann & Kliegl, 2009). 

Notably, the PVL generalizes to a complex visual-cognitive task involving the recognition of objects instead of words (Henderson, 1993). In this study, participants viewed arrays of line drawings of different objects while their eye movements were recorded. Both horizontal and vertical components of within-object fixation locations were normally distributed and clustered around the center of the object. These basic findings were confirmed by a recent study where participants searched for a target among an array of eight symmetrically arranged photorealistic objects (Foulsham & Underwood, 2009). Further, the within-object PVL was not influenced by the saliency of an adjacent distractor object. In the real world, however, objects rarely occur in isolation. With the present work, we explore whether the PVL extends to naturalistic objects within photorealistic scenes. Importantly, finding a two-dimensional PVL close to the center of objects would be evidence for object-based saccade targeting and, by inference, object-based attentional selection in scenes. 

Second, we compare the influence of cognitive relevance and visual salience in a more direct manner. Specifically, we test the hypothesis that saliency does not drive attention directly but through its correlation with object properties (cf., Einhäuser et al., 2008; Henderson et al., 2007). To this end, we perform distributional analyses of human fixation locations within saliency proto-objects. The saliency model only computes bottom-up information and has no notion of what an object is. It models selective attention to salient locations in a given image. However, Walther and Koch (2006) extended the saliency model by a process of inferring the extent of a proto-object at the salient location. The notion of proto-objects goes back to Rensink (2000a, 2000b). In his coherence theory of visual cognition, proto-objects are defined as volatile units of visual information that can be bound into a coherent and stable object when accessed by focused attention. In the extension of the saliency model as proposed by Walther and Koch (2006), the extent of attended proto-objects is determined based on the maps that are used to compute the saliency map, using feedback connections in the saliency computation hierarchy. It is important to note that such proto-objects do not necessarily have a one-to-one correspondence to objects. If viewers select objects rather than salient regions for attention and fixation, there should be less evidence for a PVL for human fixations within saliency proto-objects, compared to a pronounced PVL for real objects. There should be no evidence for a PVL when only saliency proto-objects that do not spatially overlap with annotated real objects are analyzed. Conversely, if viewers prioritize salient regions we should find a strong PVL for human fixations within saliency proto-objects and the weakest PVL for real objects that do not overlap with the saliency proto-objects. To test these hypotheses, we used an approach combining corpus-analytical and experimental techniques. To ensure high statistical power, we collected gaze data from 36 participants. They each viewed 135 semantically rich color photographs of natural scenes under three instruction sets: memorization, preference judgment, and visual search. 

Methods

Participants, apparatus, and materials

Thirty-six participants (12 males; mean age = 22.2 years) were recruited at the University of Edinburgh. They each viewed 135 unique full-color photographs (800 × 600 pixels) of real-world scenes from a variety of scene categories. Scenes were presented on a 21-inch CRT monitor at a distance of 90 cm for 8 s each. During scene presentation, participants' eye movements were recorded using an SR Research EyeLink 2K eye tracker operating at 1000 Hz. Viewing was binocular, but only the right eye was tracked. Stimulus presentation and response recording were controlled via Experiment Builder (SR Research, Canada). The computer kept a complete record of the duration, sequence, and location of each eye fixation. 

Gaze data analysis

Saccades were defined with a 50°/s velocity threshold using a 9-sample saccade detection model. Raw data were converted into a fixation sequence matrix using SR Research Data Viewer. Fixations were excluded from analysis if they were preceded by or co-occurred with blinks, were the first or last fixation in a trial, or had durations less than 50 ms or longer than 1200 ms. 

Procedure

The 135 scenes were divided into three blocks of 45 scenes. During each block, participants were instructed to perform one of three tasks: (1) search for an object, (2) memorize the scene for a subsequent memory test, or (3) make an aesthetic preference judgment. All scenes were presented for 8 s. During the Search task, an object name was presented (Ariel font, vertical height of 1.62°) for 800 ms followed by a central fixation cross for 200 ms and the scene for 8 s. If the participant located the search target, they responded by pressing a trigger on a controller (Microsoft Sidewinder). The scene remained on the screen until the 8 s were over. Eye movements were only analyzed up until the first fixation on the search target. During the Aesthetic Preference block, participants were instructed to examine each scene and then rate how much they liked it. After each scene, a response screen appeared asking the participant to respond on a 1–4 scale how much they liked the scene (1 = Dislike, 4 = Like). Responses were made via four buttons on the controller. Both the Search block and the Aesthetic Preference block were preceded by three practice trials. In the Memorization task, the participants had to encode the scene to prepare for an old/new recognition test administered at the end of the experiment. The order of tasks and the scenes used in each task were rotated across 9 subject groups using a dual Latin-square design. This ensured that every order of task and combination of scene with task was represented at least once across the 9 subject groups. Four instances of each subject group were used across the 36 participants. After the first phase constituting the three tasks (135 unique scenes in total) was complete, participants were offered a voluntary break and then a Memory Test block began. In this block, 134 of the original scenes were presented with half of them mirrored horizontally. A further 22 new scenes were presented while half of these had also been flipped horizontally. Each scene was presented for 3 s. Participants were informed that their memory would be tested for all the scenes they had previously encountered. Their task was to identify whether a scene was (1) old (old scenes from any task block in identical form), (2) altered (old scenes from any task block that have been mirrored horizontally), or (3) new. 

Search objects

Each scene contained one pre-defined search object (e.g., a basket in Figure 1c). Search targets were chosen such that they occurred only once in the scene, did not appear at the center of the scene, were not occluded, and were large enough to be easily recognized. Object size was determined by drawing a rectangular box around the search target. Attempts were made to identify search targets that were similar in size. The resulting mean object width was 2.82° (SD = 0.84°), and the average object height was 2.77° (SD = 0.88°). The height of some objects exceeded their width, or vice versa; the mean ratio of object width to object height was 1.08 (SD = 0.35; Table 1). Figure 1a shows the distribution and size of search target boxes across the scenes (135 scenes, 1 search target per scene). Across scenes, search objects cover the whole screen area, excluding the central area where participants began each trial. 

Figure 1

 

(a) Distribution and size of search target boxes across the scenes (135 scenes, 1 search target per scene). The red cross indicates the center of the screen, and the red box marks the object box displayed in (c). (b) Pixel-based object map representing the distribution of all annotated objects across all scenes; see text for details. (c) Example scene with the red object box surrounding the search target (a basket). (d) Zooming into the upper left part of the example scene, containing the search target and the imaginary box around it. The intersection of the two broken lines marks the center of the object. See text for details on the landing position analysis.

Table 1

 

Mean size of real objects (search objects and annotated objects) as well as saliency proto-objects in the study.

	Real objects		Saliency proto-objects
	Search objects	Annotated objects
N	135	730	654 (5 * 135 − 21)
Width	M = 2.82° (SD = 0.84°)	M = 2.31° (SD = 1.24°)	M = 3.27° (SD = 1.41°)
Height	M = 2.77° (SD = 0.88°)	M = 2.64° (SD = 1.35°)	M = 3.05° (SD = 1.14°)
Ratio of object width to object height	M = 1.08 (SD = 0.35)	M = 0.97 (SD = 0.50)	M = 1.14 (SD = 0.55)

Saliency proto-objects

For each scene, a corresponding saliency map was generated using the SaliencyToolbox 2.2 for MATLAB (http://www.saliencytoolbox.net/). The toolbox implements a bottom-up model of visual attention: It computes the saliency map for an image, simulates serial scanning of the image with the focus of attention (FOA), and determines the extent of proto-objects around the FOA (Walther & Koch, 2006). For the purpose of this study, the standard settings were used. In particular, the extent of the respective proto-object was inferred from the feature maps (rather than the saliency map). The actual salience map has a resolution of 1/16 of the original image, and an enlarged example is included in Figure 2b where brighter areas indicate higher salience. We analyzed the first five peaks predicted by the model. As the present study was not concerned with the order of fixation, the rank of the five selected regions was irrelevant. For an example scene, in Figures 2a and 2b red crosses mark the foci of attention, and the yellow contours in Figure 2b display the outline of the surrounding proto-objects. For fixation location analyses, the width and height of a given proto-object were defined by the closest fitting rectangle (see Figure 2, red rectangles). Twenty-one proto-objects were centrally located such that the start fixation of a given trial would be located within the object. These objects were excluded from analysis, leaving 654 saliency-based proto-objects. The saliency proto-objects were on average somewhat larger than the real objects, and there was more variation in their size (Table 1). 

Figure 2

 

(a) Example scene with annotated objects (green boxes) and five saliency proto-objects (red boxes). The saliency proto-objects are further characterized by the simulated fixation positions (red crosses) and rank numbers. (b) Corresponding saliency map including the first five model predictions. Red crosses mark the foci of attention, and the yellow contours display the outline of the surrounding proto-objects. For analyses, the width and height of a given proto-object were defined by the closest fitting rectangle (in red).

Results

Preferred viewing location for objects in scenes

The goal of the first analysis was to determine whether the phenomenon of a preferred viewing location (PVL) generalizes to naturalistic objects within photorealistic scenes. 

Selection criteria

As outlined in the Methods section, every scene contained one pre-defined search object. The initial analysis of landing positions was restricted to these search object interest areas to minimize possible effects of object size on the PVL (recall that word length affects the PVL in reading; Rayner, 1979). The search objects were relatively consistent in size (see Search objects section, Table 1, Figure 1). In addition, the initial analysis was restricted to the Memorization and Preference Tasks, where the locations of fixations that naturally fell on these objects during viewing were analyzed. In the Search Task, participants pressed a button as soon as they located the search object in the scene. Therefore, the fixation location data from this task were not suitable for fixation analysis. In the study, scenes were rotated through the three viewing tasks across groups of participants. As a consequence, all participants viewed the same scenes; however, a third of them viewed a given scene in the search task, a third in the memorization task, and a third in the preference task. In summary, for the initial analysis we examined fixations on objects defined as the search targets in the Search Task but restricted analysis to the Memorization and Preference Tasks. Preliminary analyses not reported here indicated that global eye-movement parameters like average fixation duration (memory encoding: 267 ms, preference judgment: 265 ms) and saccade amplitude (memory encoding: 4.6°, preference judgment: 4.8°) were very similar in the Memorization and Preference Tasks. Therefore, the data were collapsed across these two tasks. 

Landing position analyses were conducted for different object-based fixation types. To maximize data power, the main analyses considered all valid fixations including immediate refixations and later revisits of the object. Further analyses considered first fixations and single fixations. First fixations were the first fixations that landed on an object regardless of whether or not it was the only fixation or the first of multiple fixations on that object in the first pass. Single fixations were fixations on objects that were fixated exactly once during the first pass. The first pass comprises all fixations on an object between first entering the object and first leaving it (Henderson & Hollingworth, 1998). 

Normalized landing positions within objects

A given fixation location FL has horizontal (x) and vertical (y) components. Fixation locations are originally given in screen coordinates (x coordinate measured from the left screen border, y coordinate measured from the upper screen border). For the PVL analysis, the original fixation sequence matrix was reduced to valid fixations that fell on objects. The locations of these fixations were recalculated as within-object landing positions: The horizontal component was calculated relative to the left border of the object box, and the vertical component was calculated relative to the upper border of the object (see Figure 1d for an example). 

These two-dimensional landing positions were then normalized to take differences in object size into account. In reading, eye movements are effectively one-dimensional along the horizontal axis (except for return sweeps to the next line of text). When printed in a certain font size, all words of the text share the same height while being of variable length. In scene viewing, saccades are most frequently initiated in the horizontal direction, but vertical and less frequently oblique saccades also occur (e.g., Tatler & Vincent, 2008). Further, objects typically not only vary in width but also in height. The landing position area for a big object necessarily covers a broader range of fixation positions than the area for a small object. Therefore, the horizontal landing position axis was standardized by dividing the horizontal landing position component by the width of the object box, leading to landing positions ranging between 0 and 1. ² Likewise, vertical landing positions were normalized by dividing them by the height of the object, again leading to landing positions ranging between 0 and 1. As an example, the relative within-object fixation location of the blue fixation in Figure 1d is given by FL _xy = [0.29, 0.45]. Accordingly, a two-dimensional within-object fixation location FL _xy = [0.5, 0.5] refers to the center of the object (see Figure 1d). 

PVL results

The following analyses consider all valid fixations (see Gaze data analysis section). The horizontal and vertical components of within-object fixation locations were examined separately (Figure 3a) and jointly (Figure 3b). In a first step, horizontal and vertical components of landing positions were analyzed separately. Distributions were calculated based on 10 bins of equal size. Within-object landing position distributions for both horizontal and vertical components were Gaussian in shape and peaked around the center of the object (Figure 3a). To obtain estimates for the mean (μ) and standard deviation (σ) of the horizontal and vertical landing position distributions, we applied a grid search method (in steps of 0.01) with a minimum χ ² criterion (cf., Nuthmann et al., 2005; Nuthmann, Engbert, & Kliegl, 2007). Best-fitting lines are shown in Figure 3a; fit parameters are listed in Table 2. Mean and standard deviation of the fitted normal curve characterize the PVL curve. The actual PVL is indexed by the mean of the fitted normal curve. How much the eyes deviate from the PVL is indexed by the standard deviation of the Gaussian distribution. 

Figure 3

 

Preferred viewing location for objects in scenes. (a) Distribution of the horizontal component of landing positions for objects (red circles) and the corresponding distribution of the vertical component of within-object landing positions (blue squares). Experimentally observed distributions were fitted using truncated Gaussians. The vertical broken line marks a landing position at the center of the object. (b) Corresponding smoothed two-dimensional viewing location histogram. The intersection of the two broken lines marks the center of the object.

Table 2

 

Horizontal and vertical landing position distributions for objects in scenes: Estimates of the means and standard deviations (SDs) of the fitted normal curve.

	Horizontal PVL			Vertical PVL			n
	Mean	SDs	χ 2	Mean	SDs	χ 2
All fixations	0.50	0.29	0.001059	0.50	0.34	0.000433	1874
First fixations	0.50	0.40	0.001503	0.45	0.43	0.000771	979
Single fixations	0.49	0.46	0.001932	0.42	0.47	0.001251	707

 

Note: χ ² denotes sum of squared residuals. Each value in the n column denotes the number of observations for a given distribution.

For the base analysis, data were collapsed across initial fixations, immediate refixations, and later revisits. We cannot exclude the possibility that the global PVL distribution with a peak at the center of the object results from the summation of distributions dependent on the type of fixation. For example, it could be that the PVL is largely driven by refixations within an object. We therefore examined whether the distinct PVL pattern qualitatively replicates when restricting analyses to first fixations and/or single fixations in first-pass viewing. The results can be summarized as follows (descriptive statistics: Table 2). First, the horizontal PVL is again right at or slightly left of object center. Second, the vertical PVL is moved toward a location slightly above the center of the object. We cannot qualify these deviations from object center in terms of a tendency to overshoot or undershoot the center of the object, as the present analyses do not take the direction of the incoming saccade into account (but see Decomposition of the PVL: The effect of saccade direction section). Third, these distributions generally have a larger spread than the ones obtained in the base analysis using all valid fixations. Fourth, they are somewhat less stable due to the fewer number of observations (see also the increased χ ² values, Table 2). In sum, the PVL qualitatively remains when only fixations resulting from saccades with launch locations outside of the object are examined. 

A full account of fixation locations within objects needs to accommodate the two-dimensional nature of the data. Therefore, in a next step we performed distributional analyses in two-dimensional space. For the data that went into the base analysis presented in Figure 3a, Figure 3b shows the corresponding two-dimensional fixation location histogram. The histogram was created using narrow bins and additional smoothing. We used 50 histogram bins in either direction (horizontal and vertical). The smoothing of the two-dimensional histogram was based on an algorithm by Eilers and Goeman (2004), using penalized likelihood estimation. The smoothing parameter λ was set to 10 (values close to zero lead to a plot that is essentially just the raw data, higher values lead to more smoothing, see Eilers & Goeman, 2004). The frequency information is displayed as variations in color (Figure 3b), with colors ranging from blue (few fixations) to red (many fixations) and passing through cyan, yellow, and orange. The data revealed a “hot spot” close to the center of the object (FL _xy = [0.5 0.5]). The results thus confirm the separate analyses of horizontal and vertical landing position components. To summarize, the analyses reported in this section established a PVL for real objects in scenes: Viewers preferred to send their eyes to the center of objects within naturalistic scenes. The results support the conclusion that the eye-movement control system directs the eyes in terms of object units during scene viewing. 

Landing positions within saliency model proto-objects

A second set of analyses considered the distribution of human observers' viewing locations within saliency proto-objects (Walther & Koch, 2006). If viewers indeed select objects rather than salient regions, then there should be less or no evidence for a PVL for saliency proto-objects compared to the object analysis. However, it could still be that saliency has an indirect effect on attention, acting through its correlation with objects (Einhäuser et al., 2008; Henderson et al., 2007). To estimate the overlap between real objects and saliency proto-objects, additional objects were annotated in the scenes (see below). This allowed us to examine the fixation location distributions for saliency proto-objects that do not spatially overlap with real objects. If attentional selection is object-based, there should be no evidence for a PVL when only saliency proto-objects that do not spatially overlap with annotated real objects are analyzed. 

Object annotation

Annotation was conducted by an independent annotator who was naive to the purpose of the research. In the real world, objects are difficult to delineate. For example, does woodland in a landscape scene constitute one object or does it consist of multiple objects (i.e., individual trees)? Further, objects in real-world scenes are often partially occluded. Based on these considerations, the annotator was instructed to select (1) objects that most people would agree about, (2) objects that were not occluded by other scene elements, and (3) objects that were of moderate size. Note that we did not aim at an exhaustive object annotation of the scenes. Some scenes contained more items fitting these criteria than others; therefore, the number of annotated objects varied across scenes (see below). Annotation was implemented using Experiment Builder (SR Research, Canada). The position and spatial extent for each object was defined via a rectangular box whose coordinates were saved on disk for offline analysis (see Figure 2a for a example scene and tagged objects). This work resulted in 742 annotated objects. Objects located at the center of the scene (n = 12) were excluded from analysis, leaving 730 objects (M = 5.4 objects per scene, SD = 1.9). The average size of these objects was comparable to that of the search targets, but there was more variability in sizes (Table 1). 

A final analysis assessed the distribution of objects across the scenes. Figure 1b presents a summed object map. For a given image, all image pixel locations that belonged to a tagged object were coded with 1, while all other locations were coded with 0. These pixel-based maps were then summed across all images to obtain a single object map. Figure 1b displays the map as a pseudocolor plot with colors ranging from blue (no objects) to red (many objects). Annotated objects were located throughout the scenes, with some tendency to cluster on a ring around the center of the screen. 

General analysis method

For the saliency proto-object analysis, we used the first five fixated proto-objects generated by the saliency model (Walther & Koch, 2006) for each scene (see Saliency proto-objects section). To investigate fixation location distributions within saliency proto-objects, we analyzed the instances where human observers actually fixated on these proto-objects. In accordance with the main analysis in the previous section, all valid fixations were included in the analysis. When considering horizontal and vertical landing position components separately, analyses were again based on 10 relative landing zones. Landing position proportions sum to one. Thus, if landing positions were to follow a uniform distribution, all landing-position-contingent fixation probabilities should have an equal probability of 0.10. Visually, in Figure 4 (left panels) the empirical data points would then follow the broken horizontal line at y = 0.10. 

Figure 4

 

Landing position distributions for human fixations within saliency proto-objects. Observers' fixation locations during a memorization and/or preference judgment scene-viewing task (A and B) are compared with the same subjects' data for an object search task (C and D). Further, distributions for all saliency proto-objects (A and C) are compared with distributions for proto-objects that do not overlap with real objects (B and D). (Left) Distributions for the horizontal (red dots, solid line) and vertical (blue dots, broken line) components of landing positions within saliency proto-objects. Vertical bars denote error bars, representing 1 ± SE. To facilitate comparison, the same scaling of the y-axis was used as in Figure 3a. (Right) Corresponding smoothed two-dimensional landing position histograms.

For statistical evaluation, we tested whether fixation probabilities varied as a function of relative landing zone. In a repeated-measures analysis of variance (ANOVA), relative landing zone was specified as a within-subject factor. Polynomial trends of fixation probabilities over fixation position inform us about the shape of the landing position curve. Negative quadratic trends (i.e., convex curvature) indicate a preference for the center of objects. Linear trends indicate a significant preference for targeting either end of the object. 

The spatial overlap between saliency proto-objects and real objects was determined using the rectangular boxes surrounding the proto-objects and tagged objects (see Figure 2). It was determined algorithmically whether a given proto-object box overlapped with any of the tagged objects. Recall that there were on average 5.4 tagged objects per scene, while for a given scene the first five proto-objects fixated by the saliency model went into the analysis. For example, in the scene displayed in Figure 2a, the fifth proto-object overlaps with three tagged objects. The proto-object comprises the cardboard box almost entirely, and it partially overlaps with two other real objects. We note that this procedure is an approximation because (1) scenes were not tagged exhaustively, and (2) the algorithm did not evaluate the extent of the overlap between proto-objects and tagged objects. This procedure is therefore liberal as a test of the influence of saliency proto-objects on attention allocation and eye-movement targeting. The overlap analysis revealed that 41.0% of the 654 saliency proto-objects overlapped with 730 annotated objects. Conversely, 37.9% of the annotated objects overlapped with saliency proto-objects. The exact numbers need to be interpreted with caution, as we cannot exclude the possibility that some of the proto-objects overlapped with real objects that were not tagged. 

Memorization and preference

An analysis including all saliency proto-objects was contrasted with an analysis restricted to proto-objects that did not overlap with real objects. Figure 4A displays the results for the pooled memorization and preference data in the analysis that included all saliency proto-objects. The data suggest that there is a PVL for fixation locations within proto-objects, but the preference for object centers over object ends was not as clearly pronounced as it was for real objects (Figure 3). Further, the vertical component of the PVL was situated somewhat toward the top of object center. The results of the ANOVA are presented in Table 3 and can be summarized as follows: For both horizontal and vertical landing positions, the probability of fixations differed between landing zones. For horizontal landing positions, only the quadratic trend was reliable, indicating a PVL at the center of objects. For vertical landing positions, both the quadratic and linear trends were reliable; together they generate the slightly asymmetric curve shown in Figure 4A. 

Table 3

 

Effect of horizontal and vertical landing positions on fixation probability within saliency proto-objects.

	All saliency proto-objects				Saliency proto-objects that do not overlap with real objects
	Horizontal		Vertical		Horizontal		Vertical
	F	η 2	F	η 2	F	η 2	F	η 2
Memorization/preference
Overall effect	28.2**	0.447	12.5**	0.263	6.8**	0.163	2.6*	0.070
Quadratic trend	230.6**	0.868	28.9**	0.453	51.2**	0.594	1.9	0.051
Linear trend	0.03	0.001	21.6**	0.382	6.9**	0.165	8.7*	0.200
Search
Overall effect	2.9*	0.078	1.0	0.029	1.8	0.050	0.923	0.026
Quadratic trend	11.2*	0.242	0.7	0.021	5.3*	0.132	0.0	0.000
Linear trend	1.0	0.027	0.1	0.002	0.1	0.002	0.03	0.001

 

Note: *p < 0.05; **p < 0.001.

Importantly, when saliency proto-objects that overlapped with real objects were excluded from analysis, the PVL almost disappeared (Figure 4B). The results of the ANOVA are shown in Table 3. For vertical landing positions, there was only a significant linear trend but no quadratic trend. For horizontal landing positions, there were significant quadratic and linear trends. However, a comparison of effect size revealed that the quadratic trend was weaker than in the analysis that included all proto-objects (Table 3). Therefore, it appears that when the influence of actual objects is removed from the saliency proto-object regions, little evidence for a PVL remains. This result suggests that saliency proto-objects are not selected for attention and fixation. 

Search

For the reasons described above, we were not able to examine the PVL for the search targets in the Search Task. We therefore constrained the above analyses for both objects and saliency proto-objects to the Memorization and Preference Tasks. However, for completeness we also examined the PVL for the saliency proto-objects in the Search Task. Participants indicated that they had localized the target after an average of 3618 ms. The search target was correctly identified in, on average, 88% of the trials. When participants searched for a particular object in a scene, the PVL results for the proto-objects were even more dramatic. In an analysis that included all saliency proto-objects (Figure 4C), vertical landing position did not significantly modulate fixation probability (Table 3). For horizontal landing positions, there was a significant quadratic trend, yet the joint consideration of horizontal and vertical landing position components (Figure 4C, right panel) did not indicate a clear PVL. 

In addition, when saliency proto-objects that overlapped with real objects were excluded from analysis, the PVL completely disappeared (Figure 4D). The smoothed two-dimensional landing position distribution (right panel) showed little overall and no systematic variation in color and thus indicates an almost uniform distribution. The results of the ANOVA showed that there was no significant effect of landing position on either horizontal or vertical fixation probabilities (Table 3). 

Further analyses

The central analyses in the two previous sections examined human fixations within saliency proto-objects that did not overlap with real objects. A diametrically opposed analysis considered human fixations within real objects that did not overlap with saliency proto-objects. Different than in the initial object analysis, which was based on the search target objects (Preferred viewing location for objects in scenes—Evidence for object-based attentional selection section), the present analysis was based on the annotated objects. Objects that overlapped with saliency proto-objects were excluded. The tagged objects were more variable in size than the search objects, yet the results qualitatively replicated the object-centered PVL (Figure 5A, pooled data from the memorization and preference judgment tasks). Using the tagged objects, we also examined within-object fixation locations observers generated when searching for an object in the scene (excluding the actual search objects). Notably, the present data provide a baseline for the proto-object analyses presented above: While there was no distinct PVL when observers' search fixations that fell onto saliency proto-objects were analyzed (Figures 4C and 4D), there was a PVL for search fixations on real objects (Figure 5B). Compared to the combined data from the other two viewing tasks (Figure 5A), the PVL for the object search task showed a wider spread in landing positions (Figure 5B). While fixation locations were only analyzed up until the button press, the scene remained on the screen until the 8-s presentation time was over. Interestingly, once analyses were restricted to fixations that were made after the button press the spread of within-object landing positions was back to a level that is comparable to the memorization and preference tasks (n = 2077, horizontal: PVL _M,SD = [0.52, 0.31], vertical: PVL _M,SD = [0.47, 0.33]). These results tentatively suggest that when viewers adopt a “search mode”, their fixation placement within objects is somewhat less accurate than in other tasks. 

Figure 5

 

Distribution of human fixations within real objects that do not overlap with saliency proto-objects. (A) Pooled data from the memorization and preference judgment tasks. (B) Data from the object search task. (Left) Distributions for the horizontal (red circles) and vertical (blue squares) components of landing positions. The best-fitting normal curve for each distribution is also presented; in addition, means (Ms) and standard deviations (SDs) of the fitted normal curve are listed. (Right) Corresponding smoothed two-dimensional landing position histograms.

All analyses reported above concerned fixations made by human observers. For completeness, we also examined the distribution of fixations generated by the saliency model. The model simulates a sequence of fixation positions across the scene and then determines the extent of proto-objects around these fixation positions (Walther & Koch, 2006; see Figure 2b for visualization). In this analysis, we sought to determine how saliency model fixations are distributed within these proto-objects. The results indicated that model fixations cluster closely around the center of proto-objects (Figure 6A). More precisely, model fixations are horizontally centered on the object and vertically placed somewhat above the object center. There is much less variance in within-object viewing locations generated by the model than for human observers. In addition, we examined the distribution of saliency model fixations within real objects. Naturally, the model can only generate fixations on real objects if these spatially overlap with saliency proto-objects. Again, we observed a clear and sharp PVL at around object center (Figure 6B); note that the distributions are based on 122 observations only. 

Figure 6

 

(A) Distribution of saliency model fixations within proto-objects. (B) Distribution of saliency model fixations within real objects that overlap with saliency proto-objects. (Left) Distributions for the horizontal (red dots, solid line) and vertical (blue dots, broken line) components of landing positions. (Right) Corresponding smoothed two-dimensional landing position histograms.

Decomposition of the PVL: The effect of saccade direction

In the previous sections, we considered the effect of fixation type and, to some extent, the effect of viewing task instruction on the PVL for objects in scenes. Strikingly, the peak of the two-dimensional PVL histograms was consistently found to be at a location very close to the center of the object. This clustering around centers of objects is seemingly inconsistent with the undershooting propensity of the saccadic system (e.g., Henson, 1979; McConkie et al., 1988). Therefore, further analyses tested the possibility that the two-dimensional PVL histograms reflect composite distributions that can be decomposed by splitting the data contingent on modulating variables. Factors that are likely to modulate the PVL include the orientation of the incoming saccade, the size of the object, and the launch site distance. Here, we focus on the orientation of the incoming saccade. To maximize data power, the pooled data from the memorization and preference judgment tasks were analyzed, using the annotated objects. Analyses considered initial fixations in first-pass viewing, that is saccades that were launched from outside a given object and led to a within-object fixation, irrespective of whether it was followed by an immediate refixation or not. 

We first examined from which direction the eyes preferably entered the objects. How are launch sites from outside the objects distributed relative to the center of objects? To answer this question, incoming saccades were categorized according to the angle between the horizontal plane and the vector connecting the center of the object and the launch site (Figure A1a). The resulting directional distribution showed a horizontal preference (Figure A1b). Most saccades entered the object from the left or from the right. Naturally, the direction of the incoming saccade is partly determined by the location of the object in the scene; for example, objects close to the left border of the scene cannot be entered from the left (see Figure 1 for object maps). 

Based on these results, for further PVL analyses object boxes were divided into four directional segments, each subtending 45° (toward either side) around each of the four major axes connecting to the center of the object (left, right, up, and down; see Figure A1). Accordingly, we distinguished saccades entering objects from the left (n = 1871), right (n = 2021), top (n = 986), or bottom (n = 973). Every saccade launched from outside a given object and leading to a within-object fixation was assigned to one of these quadrants, according to the angle between the launch site and the center of the object. The distributional analyses of landing positions showed that saccades coming from a given direction tend to undershoot the center of the object on the axis of the movement (Figure 7). For example, if the eyes enter the object from the bottom, they tend to fall short of the vertical center of the object. To summarize, these analyses demonstrate that the global PVL distribution with a peak at the center of the object results from the summation of distributions dependent on the orientation of the incoming saccade. 

Figure 7

 

PVL for objects in scenes, contingent on the direction of the incoming saccade. The analysis distinguished saccades entering the object from the left, right, top, or bottom. Every saccade launched from outside a given object and leading to a within-object fixation was assigned to one of these directional segments, according to the angle between the launch site and the center of the object (see Figure A1 for illustration). The outer panels display the resulting two-dimensional landing position distributions, while the central panel shows the composite distribution.

General discussion

The present study contributes to reconciling two seemingly conflicting views of visual attention in scenes: On the one hand, according to the saliency hypothesis the visuospatial representation over which attention is directed in a scene is image-based (e.g., Itti & Koch, 2000; Parkhurst et al., 2002; Walther & Koch, 2006). On the other hand, according to the cognitive relevance hypothesis this representation is object-based (Henderson et al., 2007, 2009). With the present work, we sought to test between these two alternatives by analyzing the distribution of fixation locations within real objects and saliency proto-objects. 

On the role of objects in real-world scenes

The visual world around us presents a scene, which typically contains a number of well-defined and well-located objects. The majority of fixations during scene perception are centered on or very close to objects rather than on empty space or background (Buswell, 1935; Henderson, 2003; Yarbus, 1967). Studies of scene perception and memory demonstrate that these fixations are functional: Close or direct fixation is typically needed to perceive, identify, and encode objects into memory (Henderson, 2003). For example, in trans-saccadic change blindness experiments where an object in a scene changes during a saccade, object changes are far more likely to be detected when the changing object is fixated before and after the change (Henderson & Hollingworth, 1999; Hollingworth & Henderson, 2002; Hollingworth, Williams, & Henderson, 2001). Further, attentional guidance in scenes is also influenced by object and scene semantics, with more fixations on scene-inconsistent objects than their consistent counterparts (see Võ & Henderson, 2009, for a review). 

Objects and saliency

The main sources of empirical support for the saliency model are (1) correlations between human fixation positions and model-determined visual saliency and (2) differences in scene statistics at fixated and nonfixated locations. In a recent study, Elazary and Itti (2008) argued that the saliency model could predict “interesting” objects in photographic scenes. They used a large existing database of labeled images (Russell, Torralba, Murphy, & Freeman, 2008). The notion of interesting objects was based on the assumption that observers who were asked to label “objects” in a natural scene would predominantly do so for objects that they found “interesting.” Results showed that low-level saliency, as computed from Itti et al.'s model (Itti & Koch, 2000; Itti et al., 1998), was a highly significant predictor of which objects humans chose to label. The results were interpreted in terms of a selection bias to submit and to label objects that are inherently salient compared with the domain of all possible objects. However, the saliency model has no notion of object. Thus, it could be that saliency maps predict fixations only indirectly, because objects tend to be more salient than background, rather than because fixations depend directly on saliency. Accordingly, with the present work we directly tested the hypothesis that observers select objects rather than salient regions in the image. 

Preferred viewing location for objects in scenes—Evidence for object-based attentional selection

To investigate the object-based nature of eye guidance in scene perception, we analyzed the distribution of fixation locations within objects in photorealistic scenes. Distributional analyses were performed on (1) real objects and (2) proto-objects generated by the saliency model of visual attention (Walther & Koch, 2006). We found a two-dimensional preferred viewing location for objects in scenes. The horizontal and vertical components of normalized landing position distributions were approximately normal in shape. The means of the fitted normal curves, indexing the PVL, were consistently found to be close to the center of objects. Accordingly, two-dimensional distributional analyses revealed a “hot spot” close to the centers of objects. Thus, viewers preferred to send their eyes to the center of objects within naturalistic scenes. Based on this tendency, we argue that the eye-guidance system directs the eyes in terms of object units during scene viewing. 

The findings were further corroborated by an analysis of fixated proto-objects generated by the saliency model of visual attention (Walther & Koch, 2006). Compared to the distinct PVL for real objects, there was less evidence for a PVL for human fixations within saliency proto-objects. There was no evidence for a PVL when only saliency proto-objects that did not spatially overlap with annotated real objects were analyzed. Taken together, these results suggest that saccade targeting and, by inference, attentional selection in scenes is object-based. Saliency only has an indirect effect on attention, acting through its correlation with objects (cf., Einhäuser et al., 2008). 

Accepting that eye guidance in scene perception is object-based, the decision about where to send the eyes next can be analyzed in a hierarchical way: First, which object is selected as the target for the next saccade, and second, where do the eyes actually land given the selection of a target object for a particular saccade? With regard to the first question, it has been suggested that objects are selected based on cognitive relevance: They are prioritized for attention and fixation primarily on the basis of cognitive knowledge structures interacting with task goals (Henderson et al., 2009). Concerning the second question, based on the present data we suggest the following: When observers move their eyes to a new object, they send them to a functional target location at or near the center of the object as an optimal viewing location. The eyes tend to deviate from that location due to two sources of error: (1) a systematic error by which the system tends to undershoot the target location, and (2) a random error that produces the spread in Gaussian landing position distributions (see McConkie et al., 1988, for words in reading). 

We conclude this section with a few remarks qualifying the scope of the present findings. We do not wish to claim that viewers only ever look at objects, as empirical evidence suggests otherwise. Previous research has shown that viewers prefer to look at objects over background (Buswell, 1935; Henderson, 2003; Yarbus, 1967). Available evidence is of qualitative nature, probably because quantitative estimations are hard to achieve. The concept of objects is a hierarchical one, and what constitutes an object may depend on the task and mindset of the observer (Henderson & Hollingworth, 1998; Henderson, Weeks, & Hollingworth, 1999). Further, objects in real-world scenes are often partly occluded. Clearly, when viewing a scene a number of fixations fall on background elements, with some of them possibly being intermediary fixations from one semantically rich scene region to another. Other fixations might be strategically placed at a location between surrounding objects. Finally, we found that landing position distributions for objects in scenes can be approximated by normal distributions. However, these distributions were truncated at object boundaries, suggesting that some fixations are mislocated in the sense that they fall short of the intended target object (cf., Engbert & Nuthmann, 2008, for words in reading). In sum, we arrive at the conclusion that attentional selection in scenes has a strong object-based component: Observers parse the scene into constituent objects, and then send the eyes to the center of a selected target object. 

Replication of the PVL over fixation types and viewing tasks

The present data are the first to establish that the preferred viewing location phenomenon generalizes from words in sentences (Rayner, 1979) and isolated objects (Henderson, 1993) to naturalistic objects within photorealistic scenes. In addition, exploratory analyses examined the effect of three variables on the object-based PVL in scene viewing: the type of fixation on the object, the direction from where the eyes enter the object, and the instruction that was given to observers when inspecting the scene. 

First, while the main analyses considered all valid fixations, including immediate refixations as well as later revisits of the object, the finding of a PVL was replicated when restricting analyses to first fixations and/or single fixations in first-pass viewing. Second, we found that the global two-dimensional PVL histograms with peaks at the centers of objects reflect composite distributions that can be decomposed by splitting the data contingent on the direction of the incoming saccade. Saccades entering the object from a given direction tend to undershoot the center of the object on the axis of the movement. The data thus corroborate a general tendency for saccades to undershoot their target location (e.g., Henson, 1979; McConkie et al., 1988). 

Third, the object-based PVL in scenes was replicated across different viewing tasks. In the experiment, participants viewed a set of scenes under three different instructions: memorization, preference judgment, and search. Due to a great similarity of global eye-movement parameters in the Memorization and Preference Tasks, the data were collapsed across these two tasks. An object-based PVL was established for these data and also for the search data. The rather subtle task differences can be summarized as follows: First, there was a broader spread in landing position distributions when observers searched for a target object in the scene (Figure 5). Second, the predicted lack of PVL for saliency proto-objects was more dramatic in the Search Task than in the other tasks. Compared to real objects, the PVL for saliency proto-objects more or less disappeared in search, whereas it was attenuated for the two other tasks (Figure 4). Based on the present data, we can only speculate about these task differences. When searching for an object in a scene, all objects that are not the target object can be seen as distracter objects. Detail encoding is only required to a degree that allows the viewer to reject these objects as nontargets. Accurate fixation placement on these objects might be of less importance than in a task that calls for a greater amount of detail encoding, like memorization. A related argument concerns a possible relationship between saccade-targeting accuracy and fixation durations. Precision in saccade programming is thought to be greater following longer saccade latencies (e.g., Coëffé & O'Regan, 1987). Specifically, the duration of the preceding fixation can affect the “random error” that produces the Gaussian shape of the landing position distribution and is indexed by the standard deviation of that distribution. For reading data, it has been reported that the variance in landing position distributions on words is reduced following longer fixations (Nuthmann, 2006; but see McConkie et al., 1988). A fixation duration analysis of the present data (Nuthmann, Smith, Engbert, & Henderson, 2010) showed that average fixation durations were shorter in search (232 ms) than in memorization (267 ms), replicating prior research (e.g., Henderson et al., 1999; but see Castelhano et al., 2009). Numerical simulations of these data were suggestive of shorter saccade latencies in search than in memorization (Nuthmann et al., 2010). Thus, the increased variance of landing position distributions in the Search Task could be due to shorter fixation durations and saccade latencies at the prior fixation location. Clearly, further research is required to confirm such reasoning. 

Further research will be needed to investigate what other factors modulate the PVL for objects in scenes, such as object size and launch site distance. Furthermore, it would be informative to see whether other landing-position related phenomena observed in reading generalize from words in sentences to objects in scenes, notably the refixation Optimal Viewing Position effect (e.g., McConkie, Kerr, Reddix, Zola, & Jacobs, 1989; Nuthmann et al., 2005) and the fixation-duration Inverted-Optimal Viewing Position effect (e.g., Nuthmann et al., 2005; Vitu, McConkie, Kerr, & O'Regan, 2001). 

Conclusion

The study presented here allows us to draw important conclusions regarding the nature of attentional selection in scenes. First, we find that the preferred viewing location phenomenon generalizes to naturalistic objects within photorealistic scenes. Viewers prefer to send their eyes to the center of objects in scenes. Second, we provide evidence that saliency does not drive attention directly but through its correlation with objects. Taken together, the results support the conclusion that the eye-guidance system directs the eyes in terms of object units during scene viewing. 

Appendix A

Directional distribution of saccades toward objects in scenes

Figure A1. 

Figure A1

 

The direction of initial saccades directed toward objects in scenes was examined as the angle between the horizontal plane and the vector connecting the center of the object and the launch site. (a) The calculation of the angle is directional, with the arrow pointing toward the launch site. (b) The resulting directional distribution shows a horizontal preference. Probability densities were computed from 36 bins. The figure also illustrates the procedure for analyzing the object-based PVL contingent on the direction of the incoming saccade (Figure 7). The respective rectangular object box was divided into four quadrants to consider saccades entering the object from the left, right, top, or bottom (see text for further details).

Acknowledgments

We thank Tim J. Smith, Charles Schandl, and Maria T. Kozłowska for research assistance. Data collection was supported by a grant from the Economic and Social Research Council (ESRC) of the UK to JMH (RES-062-23-1092). 

Commercial relationships: none. 

Corresponding author: Antje Nuthmann. 

Email: [e-mail]. 

Address: Psychology Department, University of Edinburgh, 7 George Square, Edinburgh, EH8 9JZ, UK. 

Footnotes

Footnotes

1   ¹The terms landing position and viewing location are synonyms for fixation location.

Footnotes

2   ²A similar procedure has been used in reading to compare landing positions within words of different lengths (Nuthmann et al., 2005, 2007).