Managing Visual Clutter: A Generalized Technique for Label
Transcription
Managing Visual Clutter: A Generalized Technique for Label
Managing Visual Clutter: A Generalized Technique for Label Segregation using Stereoscopic Disparity Stephen Peterson∗ Magnus Axholt† Stephen R. Ellis‡ Department of Science and Technology Department of Science and Technology Human Systems Integration Division Linköping University Linköping University NASA Ames Research Center A BSTRACT We present a new technique for managing visual clutter caused by overlapping labels in complex information displays. This technique, “label layering”, utilizes stereoscopic disparity as a means to segregate labels in depth for increased legibility and clarity. By distributing overlapping labels in depth, we have found that selection time during a visual search task in situations with high levels of overlap is reduced by four seconds or 24%. Our data show that the depth order of the labels must be correlated with the distance order of their corresponding objects. Since a random distribution of stereoscopic disparity in contrast impairs performance, the benefit is not solely due to the disparity-based image segregation. An algorithm using our label layering technique accordingly could be an alternative to traditional label placement algorithms that avoid label overlap at the cost of distracting motion, symbology dimming or label size reduction. Keywords: Label placement, user interfaces, stereoscopic displays, augmented reality, air traffic control. Index Terms: H.5.2 [Information Systems]: User Interfaces; I.3 [Computing Methodologies]: Computer Graphics 1 I NTRODUCTION As information systems convey more and more data in confined spaces such as computer screens, care must be taken in the user interface to manage the resulting visual clutter. In cluttered displays, information may be obscured, fragmented or ambiguous, negatively affecting system usability. Labels, textual annotations containing object data, are one important source of visual clutter, as they overlay background layers containing their associated objects. Since legible labels need to occupy a certain minimum screen space, they may occlude or obscure other information, including other labels. Because labels are generally associated with objects or features in the background, their placement is linked to the spatial projection of their corresponding objects on the display plane. In certain cases, such as some information visualization applications, the underlying data can be spatially or temporally rearranged to simplify labeling and data interpretation. However, in applications like see-through Augmented Reality (AR), the background normally consists of real objects directly observed by the system user; accordingly all underlying display elements cannot be adjusted freely to simplify the labeling task. The application domain explored below is an AR display for Air Traffic Control (ATC) towers, in which tower and apron controllers operate to maintain safe aircraft separation at the airport. In our ∗ e-mail: [email protected] [email protected] ‡ e-mail: [email protected] † e-mail: IEEE Virtual Reality 2008 8-12 March, Reno, Nevada, USA 978-1-4244-1971-5/08/$25.00 ©2008 IEEE environment a Head-Up Display (HUD) system could use AR techniques to process position data and overlay controlled aircraft with labels, “data tags”, presenting vital flight information such as callsigns. This type of display could minimize controllers’ head-down time and attention shifts required to scan traditional radar displays. Despite the elevation of the control tower cab, typically about 50 meters above ground level, the lines of sight to controlled aircraft towards the local horizon are greatly compressed due to their relatively large distance from the tower, which could surpass 3 km. Therefore, the associated overlaid aircraft labels will frequently be subject to visual clutter in a HUD as they would likely overlap other aircraft and labels, especially at busy airports with distant taxiways and runways. Traditional label placement algorithms evaluate available 2D screen space to find optimal label locations without overlap, e.g. in cartography [7, 24], scientific illustration [10] and ATC radar interfaces [6, 9, 16]. This approach to label placement is not limited to a 2D presentation medium, since it includes AR and virtual environment interfaces [2, 3, 25, 21]. While these techniques generally avoid visual overlap, they introduce another interface design issue: which label belongs to which object? Despite the fact that a label may be connected to its background object with a line, there may be confusion as labels move according to the motion of their corresponding objects. Such confusion occurs especially if label lines intersect or are forced to overlap due to imperfect performance of label placement algorithms. Moreover, motion from automatic rearrangement of label positions can disturb or distract the user [1]. Other approaches aim at reducing visual clutter without spatial rearrangement, e.g. information filtering [15] or symbology dimming [12, 13] of data unimportant to the current task. However, automated importance classification and subsequent display suppression can entail a safety risk. Furthermore, declutter algorithms generally do not totally avoid the confusing overlap; they merely reduce it. We propose an alternative approach to reduce the visual clutter associated with label overlap: label layering. This approach does not rearrange labels in 2D screen space, nor does it filter or dim any information. Instead it extends the design space and utilizes the depth dimension, available in e.g. stereoscopic AR displays. More specifically, our technique entails placing labels in a certain number of predetermined depth layers located between the observer and the observed objects, with droplines connecting each label to its corresponding object in depth. While the general technique of reducing visual clutter using stereoscopic disparity is not novel in itself, as discussed later on, it is to our knowledge the first application and rigorous evaluation of the technique concerning the specific problem of label placement. In this work the label layering technique is instantiated in a HUD for control towers; however, it could potentially be applied to any user interface equipped with a stereoscopic display device. The human vision system interprets depth through a series of depth cues, which combine to give the observer both relative and absolute object depth information [5]. One of these cues, retinal (or binocular) disparity, interprets the difference (disparity) in binocular parallax on the retinas to trigger depth perception. The sensation 169 confusing display situation with multiple layers of overlapping labels. We manage the clutter by adjusting the labels’ stereoscopic disparity. This paper reports on an experiment testing whether human performance improves in a system with label layering based on stereoscopic disparity. It also assesses whether the disparity differences themselves are sufficient to provide practical benefit or whether consistency in depth order of the objects and their labels is also important. Section 2 presents related work to the experiment described in section 3. The results of the experiment are provided in section 4. Section 5 discusses these results while future work on an automated label layering algorithm is outlined in section 6. 2 Figure 1: Photographs taken from the subjects’ right eye viewpoint in three different viewing conditions (a-c), showing the traffic display overlaid with labels and droplines rendered on the HUD. No polarized filter was attached to the camera, making both the left and right eye image visible on the HUD. In (a) the scenario was rendered with fixed far disparity, in (b) with random disparity, and in (c) with ordered disparity. In each viewing condition the lines marked α , β and γ show the uncrossed stereoscopic disparities in the HUD imagery; a long line indicates a large disparity and subsequently a large apparent distance behind the screen. The lines marked α indicate the disparity of the label associated with the farthest object, β of an intermediate object, and γ of the closest object. The lines marked δ are constant in all viewing conditions, showing the disparity where the label droplines connect to their corresponding objects. of depth in a stereoscopic display is an effect of retinal disparity produced by the disparity in the left and right eye stereo images. The stereoscopic disparity of two objects at different depth is defined as the difference between convergence (parallax) angles from the left and right eye points. Though stereoscopic disparity is mostly known as a depth cue, it is possibly more important for its role in image segregation through its support for camouflage breaking. Three-dimensional objects that share identical visual textures with their background can be revealed by telltale disparity gradients visible in stereographic viewing systems [11]. Aerial stereography takes advantage of this fact in uncovering camouflaged military targets, invisible in a single camera view. The random dot stereogram, invented by Julesz [14], could be considered the “perfect camouflage” since hidden objects appear only if viewed by a stereographic system such as human vision; disparity is the only means to separate objects from the background noise. Our label layering approach uses the same concept in a cluttered 170 R ELATED W ORK There have been previous approaches of information segregation using depth layering to reduce visual clutter. The emphasis has generally been, however, on segregating information of different types in depth, i.e. information layering. This approach is different from our label layering technique, since we aim to distribute information of the same type, labels, in different depth layers. These two approaches, however, share many principles. Information layering, using the dual physical display planes in a multilayered display, has been investigated and compared to traditional, single layer, display devices. Although no performance benefit was found in the simple task conditions, significant performance improvements were detected using the multilayered display under demanding task conditions [23]. The information layering approach was also found to effectively offset the effects of added visual clutter in a visual tracking task on a flight display [19]. The clutter was added to the display in form of noise (making the display visually crowded), and in the form of absence of color coding for the tracking and target symbols (making the stimuli and task ambiguous). The noise, tracking and target symbols were only present in three segregated depth planes, reducing the noise impact and providing a cue to clearly identify the tracking and target symbols. Work using random dot stereograms has shown that it is possible to perceive three transparent layers concurrently, extending up to five in optimal conditions with low layer complexity and high interlayer disparity (> 5.7 arcmin) [22]. Although the display format of random dot stereograms greatly differs from stereoscopic AR displays, the findings could be supportive of the proposed depth layer design and distribution, suggesting practical limits to the number of layers that may be used. Object motion has been found to interact with stereo in various ways. Motion detection thresholds have been found to be higher with stereoscopic vision (e.g. [4, 18]), a phenomenon known as stereomotion suppression. Conversely, depth segregation is facilitated by motion (e.g. [17]). As relative motion may aid segregation of labels, the benefits of disparity-based label segregation could be reduced in situations with moving objects. Consequently, the overall ability to segregate labels based on moving stereoscopic display elements is perceptually complex and difficult to predict in general. Practical implementations require empirical investigation. 3 M ETHOD We constructed an experimental setup which allowed us to fulfill the following experimental goals: (i) Simulate realistic traffic at a major airport from the viewing position of an air traffic control tower. Render this simulation on a screen placed approximately at optical infinity relative to the observer, a distance where optical properties of visual stimuli are similar to those at the relatively large distances in a real airport environment. (ii) Overlay the airport traffic with labels identifying each object, and evaluate whether depth segregation of overlapping labels helps declutter the display and reduce users’ visual search and selection time. (iii) Render the overlay on a stereoscopic HUD, located at a realistic distance from the user considering a normal tower environment while minimizing the accommodation-vergence mismatch characteristic of the stereoscopic display format. 3.1 Hardware Setup The user was seated on a stool 2.0 m from the HUD (marker ”2” in fig. 2), a semi-transparent polarization-preserving projection screen, where the overlay graphics were rendered using passive stereo techniques. The distance was chosen so that it would be consistent with a realistic airport tower environment while the difference in accommodative demand between any two targets would always be less than 0.5 diopters. The active screen area resolution was 900×450 pixels with a 24.8◦ ×14.2◦ Field-of-View (FoV), giving each pixel a size of 1.7×1.7 arcmin. Two floor-mounted 6500 ANSI lumen projectors with linear polarization filters were used for projection. Previous measurements with full projector intensity have yielded center-screen contrast values (Michelson) over 0.99 and luminance values reaching 1000 cd/m2 [20]; approximately four times brighter than a regular computer monitor. This was considerably brighter than the background traffic display, so the projector intensities were lowered to increase visibility of the traffic display and to remove visual crosstalk in the HUD. The HUD was driven by an NVIDIA Quadro FX 4500 stereo graphics card on a dual Intel Xeon workstation running Ubuntu Linux. Although seated, the users were free to move their upper body and head. Head tracking data was fed through an IS-900 tracking system to the HUD, which meant that no height adjustment of the stool was needed for each user. The tracker sensor was attached to a lightweight and comfortable spectacle frame which was mounted on the user’s head. Traffic data, providing the HUD information about the current location of objects in the traffic simulation, were communicated over a dedicated LAN with a 1 Hz refresh rate (normal ground radar performance). Positions for intermediate frames, approximately 60 per second, were determined using linear extrapolation from previously known positions. The room was darkened and the windows were covered with thick black plastic film. The experimental code was written in C++ using OpenSceneGraph for scenario construction and VR Juggler for tracking and display configuration. 3.2 Task The task of each experimental trial was to identify and select by mouse click one aircraft in an airport traffic scenario on the traffic display, based on a given target label in the HUD. Figure 3: A portion of a rendered scenario showing an overlap situation in the top left corner. Labels and droplines (red) were rendered on the HUD, while the aircraft objects and ground plane were rendered on the traffic display. Figure 2: The experimental setup showing (1) a subject wearing a head tracker and polarized glasses, (2) the HUD rendering the overlay and (3) the traffic display rendering the traffic simulation. The bottom left half of the photo has been digitally enhanced to reveal detail; in reality the room was as dark as the upper right half. The traffic display, an opaque projection screen, was mounted 6.4 m from the user in the user’s line of sight through the center of the HUD (marker ”3” in fig. 2). This distance renders an accommodative demand of 0.33 diopters between the HUD and the traffic display. The active screen area displayed 1400×770 pixels with a 20.3◦ ×11.3◦ FoV, giving a pixel size of 0.9×0.9 arcmin. One 3500 ANSI lumen projector was used for image projection. The traffic was not rendered in stereo since it simulated airport traffic located at least 500 m from the observer, a distance at which relative disparities of physical objects are negligible. The traffic display was driven by an Intel Centrino laptop running Ubuntu Linux. The labels consisted of a 6-character airline callsign, 3 letters and 3 digits in sequence. The letter sequence was randomly selected within 5 possible combinations, corresponding to real airline identifiers starting with the letter A. The number sequence was randomly generated. The last digit, however, was only randomly generated between trials. By keeping the start letter and end digit constant within a trial, the subject was required to read the 4 centermost characters for object identification, thus limiting the possibility of using methods of exclusion. By randomly generating the included callsigns in a traffic scenario as described here, the subject could not recall the callsign and its location from a previous trial, thereby minimizing training effects due to scenario familiarity. The labels were rendered using the Century Gothic font in full red color. All labels had the same size and intensity on the screen plane. The labels were approximately 3.2◦ ×0.7◦ in size, where the total width varied slightly depending on the character glyphs. Each label was located approximately 1.2◦ above its corresponding object. Thus, the height in the visual field of the objects and labels was a cue to object distance. The objects in the traffic display, representing airport traffic, were rendered as blue cones with the base perpendicular to the ground plane and apex pointing in the direction of motion. Only the upper half of the cones was visible as the lower portions were occluded by the ground plane. Due to perspective, the width of the closest objects was ∼1.3◦ in the screen plane, while the farthest were about a third of that size. The mouse pointer used for selection in the traffic display was 0.6◦ in height. 171 Each traffic scenario was an extract from a simulated 24-hour airport traffic dataset at Paris CDG1 . The total number of visible objects was between 9 and 14 when each scenario was initialized, but as the scenarios evolved the number would ultimately range between 7 and 15. A screenshot of a rendered traffic scenario with the superimposed HUD graphics is shown in figure 3. Photographs of the scenarios as they were presented to the experiment participants are shown in figure 1, although they were not taken with polarized filters (glasses) and therefore include both left and right stereo HUD imagery (labels and droplines). Moreover the labels shown in the photos differ from the real stimuli in that the glyphs are considerably thicker due to lens flare; in reality they were perceived as shown in figure 3. 3.3 Participants We recruited 17 subjects, with ages ranging from 25 to 60. Because visual accommodation has substantially degraded after the age of about 40, we used approximately balanced age subgroups: eight subjects were 25-39 years of age, while nine were 40-60. Three subjects were female, of which two were over age 40. All participants were staff, contractors or students at the EUROCONTROL Experimental Centre. Participation was voluntary and no compensation was given. All subjects passed a stereo vision test presented on a computer monitor with a red-cyan anaglyph technique. Eight sets of random dot anaglyph images, each set with 4 images, were displayed on a computer screen. One image per set was distinguished from the rest by the retinal disparity cue only visible to subjects with stereo vision. The subjects’ task was to identify the distinguished image and tell whether the contained square shape was perceived in front or behind the screen. The retinal disparity required for passing the test was 3.2 arcmin. 3.4 Procedure Each subject was provided written experiment instructions before the trials. The instructions stressed that accuracy and response time were both important, but accuracy should be prioritized. We measured each subject’s inter-pupillary distance using a mirror and ruler. This value was used to calibrate the stereo disparity of the HUD. The subjects then moved to the stool, and mounted the position tracker and lightweight polarized glasses on their head. Before the experimental trials the subjects were given four test trials. During these trials the subjects were accustomed to the task, and screened for correct stereoscopic depth perception by which they confirmed that the closest and farthest labels were perceived at approximately the intended distances. The HUD registration with the background was calibrated by aligning four white spheres visible in the HUD with the corners of the traffic display. The calibration was performed by manually adjusting the position of the head tracker. This calibration assured that each dropline extended to the center of the corresponding object with an acceptable registration error of up to 0.4◦ . Before each trial the target callsign was presented on the HUD. The subjects pressed the spacebar on a keyboard on a table in front of the stool. This action removed the target callsign and initialized the scenario which appeared after a few seconds. The subjects scanned the labels, identified the target callsign, and visually followed a dropline extending from the label to the corresponding object on the traffic display. The label end of the dropline had the same disparity as the label, while the object end had approximately the same disparity as the traffic display. The subjects then selected 1 The data was simulated using the TAAM software from Preston Aviation Solutions Pty Ltd. 172 the object on the traffic display using a mouse. They finally confirmed their selection by pressing the spacebar, which cleared the scenario and presented the next callsign. The trials were divided into two blocks of 36 trials. During the break between the blocks, the subjects could rest, and even walk around. In case subjects removed the head tracker during the break, the HUD calibration was performed before recommencing. The subjects were also asked to report any discomfort during the break. After the trials each subject completed a questionnaire. The total time including preparations, trials and questionnaire was approximately 60 minutes per subject. 3.5 Independent Variables Viewing Condition Each label included in a trial was rendered on the HUD with a certain stereoscopic disparity, making the label appear to the subject at a distinct depth. The spatial distribution of stereoscopic disparities for all labels in a trial was the independent variable determining the principal viewing conditions of the experiment. Four viewing conditions were used for label presentation; i) ordered disparity, ii) random disparity, iii) fixed near disparity and iv) fixed far disparity. Three of these viewing conditions are illustrated in figure 1. (i) In the ordered disparity condition the labels were separated in depth into N discrete layers according to a logarithmic function: dn = dmax − log (N − n + 1) × dmax − dmin log N (1) where n is the label layer index (1 ≤ n ≤ N) and N is the total number of label layers, which in this experiment was fixed at 15. The closest label (n = 1) was located approximately at the HUD (dn = dmin = 2.2 m) while the farthest (n = N) was located near the traffic display (dn = dmax = 6.0 m). The distance between the intermediate labels increases according to the logarithmic function, in order to approximately equalize inter-layer stereoscopic disparity (see fig. 4). Even though the maximum number of labels was not always present in the view, 15 layers were prepared for each trial to take care of labels that enter the view at a later time, avoiding potentially disturbingly visible depth rearrangement. Given an IPD of 64 mm and 15 label layers, the inter-layer stereoscopic disparity was 6.0 ± 1.2 arcmin. The stereoscopic disparity difference between overlapping labels could be larger if located in nonadjacent label layers. The depth order of the labels matched the distance order to the corresponding objects. I.e., the label of the closest object in the traffic scenario was assigned to the first label layer; subsequent object labels were assigned to label layers based on their distance order. (ii) In the random disparity condition the labels were segregated as in the ordered condition except that their depth order was randomized with respect to that of the corresponding objects. We included this condition as a control case to determine if segregation by image disparity itself had a benefit, regardless of depth order. If so, the results would show that both the ordered and random disparity conditions aided search and target designation through visual clutter reduction. (iii) In the fixed near disparity condition, all labels were placed in a single apparent depth layer approximately at the HUD, (dn = 2.2 m). (iv) In the fixed far disparity condition, all labels were placed in a single apparent depth layer approximately at the traffic display (dn = 6.0 m). Response Error The target designation error was recorded. The response was erroneous if the target callsign differed from the required response. The experiment’s difficulty level was adjusted to ensure a level of erroneous identifications below 10%. 3.7 Experimental Design Figure 4: Distance (dn ) for each label layer (n) in the ordered and random disparity viewing conditions, given by equation 1. In the ordered disparity condition the label layer order correlates with the object distance order; in the random disparity condition the label layer order is randomized with respect to the object distance order. Object Motion The aircraft objects in each trial were either static or dynamic, i.e. in motion. When in motion, they moved according to the traffic simulation. Although objects exhibited varying motion, depending on their situation, depth and corresponding aircraft size, horizontal screen motion was generally below 0.3◦ /s. Objects could have higher speed, e.g. when landing or taking off, but such objects were not selected as targets. Overlap Level The overlap level could be high, medium or low as seen by the subjects. High overlap level indicates that the target object had two other objects in its immediate proximity, meaning its label would likely be in an overlap situation with two other labels. An object was defined to be in the immediate proximity of the target when within 1.4◦ from the target location as seen by the subject. In the medium overlap level the target object had one other object in its immediate proximity. Low overlap level indicates that no other objects were in the target object’s immediate proximity. This distinction is illustrated in the scenario shown in figure 3, where object AAG183 would be considered to be in the high overlap level, AIB603 in the medium overlap level, and AAL783 in the low overlap level. If objects were in motion the overlap level could change over time; however, the designated level corresponded to the situation in the first few seconds of each dynamic scenario. The overlap situations were not identical throughout the experimental conditions but randomly sampled from matched sets having common overlap properties. There were six specific scenarios per overlap level. By sampling the overlap situations from a larger set as described here, we reduced the number of times an overlap situation was re-used, thereby minimizing training effects due to scenario familiarity. Repetition Each combination of independent variables was repeated three times per subject with a blocking designed to let sequence effects, due to shifts of the viewing conditions with varied disparity distributions, abate. In this way the problem of possible asymmetric transfer in repeated measures experiments was avoided, as demonstrated in the post-experimental analysis. 3.6 Dependent Variables Response Time We recorded the duration from the instant the scenario was displayed until subjects generated the required response. The users responded by identifying and selecting the target aircraft with a mouse click, and subsequently confirming the selection by pressing the spacebar. Each subject saw all combinations of the independent variables in a repeated measures within-subject design. We made a total of 1224 data recordings, 72 per subject (4 viewing conditions × 2 object motions × 3 overlap levels × 3 repetitions). The viewing condition was blocked in groups of three, allowing any transient effects to abate when viewing conditions changed. The three blocked repetitions should not be confused with the independent variable repetition where all independent variables were fixed. Only the viewing condition was blocked, all other conditions were randomized for each subject. Accordingly, the three trials in a specific block were not exact replications of the same conditions. The blocks were constructed using a partial Latin squares design and were re-used for all subjects. The presentation order of the blocks between subjects was counterbalanced through random permutations. 4 R ESULTS Initial analysis of the data indicated that there were transition effects on response time within the blocks. Post-hoc Scheffé comparisons showed that there was a significant difference (F(3, 26) = 19.4, p < 0.01) in response time between the first and second block element, for each of the main viewing conditions, while there was no significant difference between the second and third repetition which were approximately equal. The transition effects only showed up in the ordered and random disparity viewing conditions. We believe that the transition effects are due to changes in the pattern of ocular vergence required for each of the main viewing conditions. For example, when viewing the ordered disparity condition the subjects would be required to change vergence from the far screen to various positions between the near and far screens. In contrast, when the labels were presented in the fixed disparity conditions, no vergence changes would be required while the subject looked from one label to another. In order to avoid these transition effects we report results based only on analyses of the third element within each block. All reported statistically significant effects were, however, present even when all three elements of the block were included. By limiting ourselves to only the third in-block repetition we are able to filter out effects and interactions that might be attributed to the transition effects. Using regression analysis on the median response time of each trial we found no significant effect for trial number on response time (r = −0.32, d f = 22, ns), meaning that subjects showed no general improvement of performance throughout the trials (see fig. 5). This confirms not only that our measures to minimize training effects due to scenario familiarity were successful, but also that there was no significant training effect due to skill development. Two subjects, despite having passed the stereo test and the initial depth perception screening, reported significant difficulties with stereo accommodation, both orally during the break and in the questionnaire. These subjects were therefore excluded from the analyses. Their exclusion from analysis did not have any effect on the pattern of statistically significant results, but we believe it improves the accuracy of the data since our experiment presupposes good stereoscopic vision. The questionnaire data is not included in this paper but will be reported at a later date. 173 (F(2, 26) = 63.9, p < 0.001), which confirmed that our initial scenario design worked as intended (see fig. 6). The post-hoc Scheffé analysis showed significant increases in response time between low and medium overlap levels (F(2, 177) = 12.5, p < 0.01) and between the medium and high overlap levels (F(2, 177) = 56.8, p < 0.01). No significant main effect was found for viewing condition on response time (F(2, 26) = 1.30, ns). Initially we had anticipated this main effect to be significant, however, as discussed later, subsequent analysis showed significant interaction effects with overlap level. Furthermore there were no significant main effects of age (F(1, 13) = 0.61, ns) or object motion (F(1, 13) = 0.06, ns) on response time. Figure 5: Trial number had no significant effect on response time, showing that no overall training effect was present. The diamond shapes show each subject’s recording per trial, while the larger squares show the median of these recordings per trial. The line represents the linear regression through the median response time of each trial, with its equation given above. 4.1 Response Time The results were analyzed using analysis of variance (ANOVA), using a repeated measures design with subject as a random variable and a fixed model for all other independent variables. When analyzing for effects, we applied a logarithmic transformation to the response time in order to remove the strong positive skew of the response time data, which otherwise would have produced a violation of the analytic assumption of homogeneity of variance. Figure 7: Mean response time for each viewing condition, grouped by overlap level. As the main result of this experiment, an interaction effect was found in the high overlap level, where ordered disparity showed significantly lower response times than the other conditions. As the main result of this experiment, a significant effect was found for the interaction of viewing condition and overlap level on response time (F(4, 52) = 4.63, p < 0.005), where a significantly lower response time was found for the ordered disparity condition when overlap levels were high (fig. 7). This means that in situations when the target object was in close proximity to two or more objects, ordered disparity was significantly faster (24.0%, 4.1 s) than fixed disparity (F(2, 57) = 9.72, p < 0.05). Similarly it was significantly faster (37.2%, 7.6 s) than random disparity (F(2, 57) = 21.4, p < 0.01). There was no significant interaction effect of object motion and viewing condition on response time. Figure 6: Overlap level had a significant main effect on response time. Initial analysis of the data showed an interaction effect for viewing condition and overlap level on response time (F(6, 78) = 2.49, p < 0.05). Post-hoc Scheffé comparisons showed no significant differences between the two fixed disparity conditions in either levels of high overlap (F(2, 57) = 0.29, ns), medium overlap (F(2, 57) = 0.03, ns) or low overlap (F(2, 57) = 0.21, ns). Since we are interested in the effects of the varying disparity scheme against the traditional “2D” layout with fixed stereoscopic disparity, and all these pairs of means did not vary by more than 2.7%, we collapsed these two conditions into one condition called fixed disparity. Subsequent analyses and diagrams are based on this approach. We found a main effect for overlap level on response time 174 4.2 Response Error The mean error rate was 6.1%. Analyzing the data using χ 2 contingency tables (see table 1) we found no significant effect of viewing condition on response error (χ 2 (2) = 0.36, ns). We suspected that there could be an increase in accuracy through training on the first two block repetitions, so we made an analysis on the full block data (see table 2). However, no effect was found for viewing condition, looking both at contingency tables of proportional and equal error distributions (χ 2 (2) = 4.26, ns). When analyzing the effects of object motion we found that the mean error rate was 8.3% for static scenes and 3.9% for dynamic scenes, but the effect was found not to be significant (χ 2 (1) = 3.1, ns). Generally, the error rate was too low for the number of subjects tested to make any further analyses of effects on error rate. How- Table 1: Contingency table showing response error per viewing condition, third block repetition. Response Correct Incorrect Disparity Fixed Ordered 189 96 15 6 Random 96 6 an input to an automatic label layering algorithm. An important issue to take into account, when designing an interface with layered information, is that the font glyphs must be thin enough for characterizing features (lines, strokes, etc) to protrude from underlying layers. In some pilot studies we used thicker font glyphs, yielding fewer available features from each depth layer, which hampered stereoscopic fusion. 6 Table 2: Contingency table showing response error per viewing condition, all block repetitions. Response Correct Incorrect Disparity Fixed Ordered 572 294 40 12 Random 282 24 ever, this low rate confirms that the subjects understood the instruction with respect to prioritizing accuracy over response time. 5 D ISCUSSION We have shown that ordered label layering significantly improves decision time, over 24% on average, in complex, realistically moving scenes with high degrees of label overlap, compared to cases without label layering. The effects were not significant in the less complex cases with little or no overlap, which was expected since in these cases the target labels did not need segregation from other labels. These results are similar to the findings of Wong et al. [23] mentioned previously, where the effects of information layering were only significant under demanding task conditions. We had initially assumed that the label layering technique, regardless of depth order, would have a positive impact on performance. However, results showed that only the ordered label layering, where label layer order corresponded to object distance order, improved performance. Conversely, the random label layer order significantly decreased user performance, even when compared to cases without label layering. This is likely due to the mismatch in depth cues in the random case; the height in visual field of the labels and background objects, as well as the objects’ relative size differences, are inconsistent with labels’ stereoscopic disparity order, negatively affecting label-object integration. If the label layering technique were applied to a flat 2D interface with an orthogonal view, like a traditional computer screen, there would be no depth conflict so we now hypothesize that random disparity would yield better performance, perhaps similar to ordered disparity. Motion had no main effect on response time, as shown in the results section. This may be due to the fact that motion can influence response time both ways. It could decrease response time since the relative motion could help clutter breaking and make identification faster. Alternatively it could increase response time as the motion could encourage the subject to wait for the traffic to evolve, making the target more visible. We therefore suspected that motion could have a main effect on error, however that effect was not found to be significant either, possibly due to our low baseline error rates. It would be interesting to perform a more statistically powerful, indepth study where motion is isolated as a perceptual cue, and cannot be used as a “wait-and-see” factor as in this experiment. It would also be interesting to analyze the interaction of motion and overlap level on response time, which was not possible with the current data. This would reveal if motion is more or less effective in the high overlap situations through its effect of user sensitivity to disparity. A better understanding of the role of motion could serve as F UTURE W ORK The current study justifies and provides empirical baseline data for development of a future, automated, label layering algorithm using stereoscopic disparity, which would subsequently be evaluated against traditional 2D label placement algorithms. As label layer order is an important factor for performance, at least in a perspective display format, such an algorithm must handle situations where the depth order of the individual objects changes over time in order to maintain the required label-object depth correlation. There was no such differential movement in depth in the present experiment. It would be interesting to evaluate the depth distribution function of label layers. The logarithmic distribution used in this experiment did segregate labels in an effective way; however, many alternative separation functions are available. We hypothesize that a constant inter-layer disparity would be preferential, instead of the 6.0 ± 1.2 arcmin disparity range used in this experiment. It would also be relevant to study the maximum number of perceivable overlapping label layers. If there is an upper limit, perhaps analogous to the five perceivable transparent layers in random dot stereogram stimuli [22], the label layering algorithm could be combined with techniques of traditional 2D algorithms for solving particularly complex situations, yielding a fully three-dimensional label placement algorithm. With a limit on the number of label layers, larger inter-layer disparities (> 10 arcmin) are possible. The optimal inter-layer disparity in such a case would require further empirical investigation. Even though our new approach effectively expands the design space and could alleviate the need for compromises required by traditional declutter methods, such as filtering, dimming, aggregation or label size reduction, it may be the case that the introduced depth movement of labels could be distracting and capture unnecessary attention. Indeed, research has shown that looming objects are much more likely to capture attention than receding ones [8]. The looming stimuli used in that study were however provided through relative size adjustments, with stereoscopic disparity fixed. Conversely, our system would vary the stereoscopic disparity of labels while keeping relative label size fixed. The practical effect in our system requires empirical investigation. We speculate however that even if similar patterns of attention capture were present for disparity-based object looming, the threshold of capture would be affected by the speed and smoothness of depth motion. Given that label layering is only effective in overlap situations, as shown in this paper, the default non-overlap label positions would be in a single depth layer. Since receding objects capture less attention than looming ones, imminent label overlap situations should be resolved through a receding motion. To satisfy the disparity ordering constraint, discovered in this experiment, the label corresponding to the farthest object should recede. Restoring the label to its default position could be implemented with a much slower loom, below the threshold of attention capture. This attention avoidance is e.g. analogous to the washout phase in motion-based flight simulators, where the desired sense of acceleration is achieved by tilting the motion platform backwards, while the restoration to the initial neutral position is performed slow enough to be below the human motion sensor thresholds. In addition, prediction algorithms could be used to resolve future label overlap situations in advance, reducing the risk of rapid depth movement. 175 ACKNOWLEDGEMENTS Stephen Peterson and Magnus Axholt were supported by PhD scholarships from the Innovative Research Programme at the EUROCONTROL Experimental Centre, Brétigny-sur-Orge, France. R EFERENCES [1] K. R. Allendoerfer, J. Galushka, and R. H. Mogford. Display system replacement baseline research report. Technical Report DOT/FAA/CT-TN00/31, William J. Hughes Technical Center, Atlantic City International Airport (NJ), 2000. [2] R. Azuma and C. Furmanski. Evaluating label placement for augmented reality view management. In Proceedings of IEEE/ACM International Symposium on Mixed and Augmented Reality (ISMAR 2003), pages 55–75, Tokyo, Japan, October 2003. [3] B. Bell, S. Feiner, and T. Höllerer. View management for virtual and augmented reality. In UIST ’01: Proceedings of the 14th annual ACM symposium on User interface software and technology, pages 101– 110, Orlando, Florida, 2001. [4] K. R. Brooks and L. S. Stone. Stereomotion suppression and the perception of speed: Accuracy and precision as a function of 3d trajectory. Journal of Vision, 6:1214–1223, 2006. [5] J. E. Cutting. How the eye measures reality and virtual reality. In Behavior Research Methods, Instrumentation, and Computers, volume 29, pages 29–36, 1997. [6] A. Dorbes. Requirements for the implementation of automatic and manual label anti-overlap. Technical Report 21/00, EUROCONTROL Experimental Centre (EEC), 2000. [7] S. Edmondson, J. Christensen, J. Marks, and S. Shieber. A general cartographic labeling algorithm. Cartographica, 33(4):13–23, 1996. [8] S. L. Franconeri and D. J. Simons. Moving and looming stimuli capture attention. Perception & Psychophysics, 65(7):999–1010, 2003. [9] K. Hartmann, K. Ali, and T. Strothotte. Floating labels: Applying dynamic potential fields for label layout. In Proceedings of 4th International Symposium on Smart Graphics, pages 101–113, Berlin, 2004. Springer Verlag. [10] K. Hartmann, T. Götzelmann, K. Ali, and T. Strothotte. Metrics for functional and aesthetic label layouts. In Proceedings of 5th International Symposium on Smart Graphics, pages 115–126, Berlin, 2005. Springer Verlag. [11] M. B. Holbrook. Breaking camouflage: stereography as the cure for confusion, clutter, crowding, and complexity - three-dimensional photography. Photographic Society of America Journal, 8, 1998. [12] M. S. John, B. A. Feher, and J. G. Morrison. Evaluating alternative symbologies for decluttering geographical displays. Technical Report 1890, Space and Naval Warfare System Center, San Diego, CA, 2002. [13] M. S. John, H. Smallman, D. I. Manes, B. A. Feher, and J. G. Morrison. Heuristic automation for decluttering tactical displays. The Journal of the Human Factors and Ergonomics Society, 47(3):509–525, 2005. [14] B. Julesz. Foundations of Cyclopean Perception. The University of Chicago Press, Chicago, 1971. ISBN: 0-226-41527-9. [15] S. Julier, M. Lanzagorta, L. Rosenblum, S. Feiner, and T. Höllerer. Information filtering for mobile augmented reality. In Proceedings of ISAR 2000, pages 3–11, Munich, Germany, October 2000. [16] S. Kakos and K. J. Kyriakopoulos. The navigation functions approach for the label anti-overlapping problem. In Proceedings of the 4th EUROCONTROL Innovative Research Workshop, Paris, France, 2005. [17] M. J. M. Lankheet and M. Palmen. Stereoscopic segregation of transparent surfaces and the effect of motion contrast. Vision Research, 38(5):659–668, 1998. [18] S. P. McKee, S. N. J. Watamaniuk, J. M. Harris, H. S. Smallman, and D. G. Taylor. Is stereopsis effective in breaking camouflage? Vision Research, 37:2047–2055, 1997. [19] R. V. Parrish, S. P. Williams, and D. E. Nold. Effective declutter of complex flight displays using stereoptic 3-d cueing. Technical Report 3426, NASA, 1994. 176 [20] S. Peterson, M. Axholt, and S. R. Ellis. Very large format stereoscopic head-up display for the airport tower. In Proceedings of the 16th Virtual Images Seminar, Paris, January 2007. [21] E. Rosten, G. Reitmayr, and T. Drummond. Real-time video annotations for augmented reality. In International Symposium on Visual Computing, 2005. [22] I. Tsirlin, R. S. Allison, and L. M. Wilcox. On seeing transparent surfaces in stereoscopic displays. Master’s thesis, York University, Canada, 2006. [23] B. L. W. Wong, R. Joyekurun, H. Mansour, P. Amaldi, A. Nees, and R. Villanueva. Depth, layering and transparency: Developing design techniques. In Proceedings of the Australasian Computer-Human Interaction Conference (OZCHI), Canberra, Australia, 2005. [24] M. Yamamoto, G. Camara, and L. A. N. Lorena. Tabu search heuristics for point-feature cartographic label placement. GeoInformatica, 6(1):77–90, 2002. [25] F. Zhang and H. Sun. Dynamic labeling management in virtual and augmented environments. In Proceedings of the 9th International Conference on Computer Aided Design and Computer Graphics (CAD/CG), 2005.