The Influence Of Layout In Crossing-Based Interfaces

The Influence Of Layout In Crossing-Based Interfaces Authors list removed for anonymous review ABSTRACT In this paper we present the first empirical study of the parameters influencing the selection of two options of a simple crossing-based dialog box. Our results show that in the case of users interacting directly on the screen the parameters influencing the selection time include the index of difficulty of the connection as well as the angle between the target centerline and the horizontal. We use the results of our experiment to provide the first practical guidelines for laying out crossing-based dialog boxes. CATEGORIES AND SUBJECT DESCRIPTORS H.5. Graphical User Interfaces, Input Devices and Strategies; D.. User Interfaces; I.3.6 Interaction Techniques ADDITIONAL KEYWORDS AND PHRASES Crossing-based interfaces, command composition, fluid interaction, pen-computing INTRODUCTION As large format pen-based devices such as tablet computers are becoming more common, the need for interaction techniques specially adapted to this configuration is becoming more important. While the standard WIMP-interface (Windows, Icons, Menus, and Pointers) has proven to be very powerful in an indirect setting, it is more difficult to use in the direct setting offered by a typical tablet computer configuration. For example, while traditional interfaces are designed with a stable pointing device in mind, tablet computers use a pen which makes actions such a double-clicking difficult. More importantly, by focusing on clicking, the classical interfaces are ill-fitted for pen computing: while fluid creation of strokes is the natural mode of operation for a pen, current interfaces insist on segmenting user s interaction in a series of points and clicks. Encouraged by Accot s work on crossing-based interaction, we recently investigated the feasibility of crossing as an interaction paradigm for pen-based interfaces. Our prototype application, CrossY [4] is a Figure 1 A screenshot of CrossY, the first completely crossing-based interface. simple sketching interface for which all interactions are performed by crossing small goals on the screen with a pen. CrossY demonstrated that it is possible to duplicate the most common point-and-click widgets (such as menus, checkboxes, scrollbars, and dialog boxes) in a crossing environment. In fact, our experience with CrossY suggests that crossing is a powerful alternative to clicking. We were able to provide a rich set of interactions for each widget and demonstrated that crossing can also encourage the fluid composition of commands [4]. As we moved our focus from design exploration to a more systematic examination of how key design parameters might influence the efficiency of crossing-based interfaces (both in terms of speed and in terms of less tangible aspects such as comfort), we encountered a shortage of data about key aspects of the interface. In a typical configuration, tablet users are faced with somewhat different psychomotor requirements than desktop users: On the one hand, hand-eye coordination is more easily achieved in the direct setting of a tablet computer than in the indirect setting of most desktops. On the other hand, without the cursor acceleration, tablet users experience the full size of the screen first hand and have to cover comparatively large distances. Yet most studies of goal crossings have been in the indirect setting with high gain between movement of the pointing device and the movement of the cursor. Another problem we faced was that most study on crossing focused on axis aligned configurations. Yet as shown in 1

Figure 3 Here, the key variables for designing crossing-based widgets are shown. Figure One of CrossY's dialog boxes illustrating the need for more information about the typical landing points and crossing paths. The designer needs to know how big the different targets need to be and which spacing is necessary between widgets. Figure, even in the case of the simplest dialog box, fluid composition of command selection implies that users will have to perform selections between targets at an angle of each other. Preliminary observations suggested that users speed may be influenced not only by the distance between the two targets (as predicted by Fitts; law) but also by the angle of the centerline between targets The work presented here represents a first attempt towards a better understanding of the parameters that influence users performance in crossing-based interfaces. We compared the influence of the connection angle both in direct and indirect settings. Our results show that when moving from an indirect to a direct configuration, the angle between the centerline of the two targets and the horizontal influences crossing time. Our results also show that the influence of angle on crossing time is exacerbated in the case of vertical targets (a common case in CrossY). Based on our results, we provide the first practical guidelines for the design of efficient crossing-based dialog boxes such as the one presented in Figure. PREVIOUS WORK Accot s Steering law represents the first step in generalizing Fitts law to target crossing [1]. His work was later extended to a more systematic study of the crossing paradigm [3]. These initial findings were crucial in establishing the viability of the crossing paradigm. They also provided valuable data about the relative speed of different crossing configurations. For example, they explored the influence of the orientation of the target with respect to the main trajectory (either collinear or orthogonal), and the influence of a distractor between two targets. Our work extends these finding by considering the influence of the angle between the centerline of the targets and the horizontal. Further, we provide the first measurements in a direct configuration. A large body of work has studied Fitts law [7] and its application to computer interface design [11]. In particular, several authors have studied the possible influence of the angle between the target centerline and the horizontal. Card et al. [6] report a small influence of the angle when using a joystick, but none when using a mouse. MacKenzie [1] and Jagaconski [9] confirmed this finding for a mouse. It is important to note, that these results were gathered in an indirect setting with a gain between 1.5 and. Our results extend these findings to a direct setting, typical to tablet computers. Goal crossing has been used as the basis of interaction in several systems. Lotus Notes, [8] allows users to select several emails in a list by crossing them with a single stroke. Baudisch s toggle maps [5] let users toggle several toggle switches by simply painting over them. Recently, we implemented CrossY [4], the first drawing application solely based on crossing. The present study extends this design exploration by providing empirical data to better understand different design parameters in the crossing interface. EXPERIMENT There are many parameters that can influence users efficiency while using crossing-based interfaces. Based on our own observations while developing CrossY, we decided to focus on the parameters relevant for the design of dialog boxes such as the one shown in Figure. From a design perspective, their efficiency is influenced by the three parameters shown in Figure 3: 1) the height of the crossing bar (h); ) the center-to-center distance between two columns (dc); 3) the center to center distance between successive rows (dr). While intuitive from a design perspective, these parameters make the comparison with the traditional index of difficulty as stated in Fitts law problematic. Fitts law - D a + blog + 1 H - connection time is a function of the distance between the targets (D) and their height (H). The parameters a and b are determined by the specific experimental conditions. When

ID distance height 10 40 180 60 3 4 5 6 80 40 40 60 300 0 600 40 310 10 60 0 630 10 788 13 Figure 5 The target heights, distance between targets, and corresponding ID s for the experiment. designing CrossY, empirical evidence seemed to imply that the angle between the target centerline and the horizontal might influence the values of a and/or b. Based on this observation, we structured our experiment around the following set of parameters: 1) the height of the target (H), ) the distance between two target centers (D; subsequently, D and H are combined into an index of difficulty, ID), and 3) the angle (α) between the centerline of the targets and the horizontal. This choice allowed us to compare our results with previous studies (e.g. [9, 1]) and to explore the influence of α in the direct setting. In addition, we considered target orientation, another variable that is highly relevant for crossing interfaces. As shown in Figure 1, due to layout constraints in crossing interfaces, it is advantageous if the targets stay vertical. In contrast, previous experimental studies examined targets that were either orthogonal or collinear to the main direction of travel [3]. In summary, we identified 5 variables that may influence user performance in crossing interfaces: Distance: the distance, center to center, between two crossing targets. Height: the height of the targets Angle: the angle between the target centerline and the horizontal line Orientation: the angle between the target centerline and the targets themselves Setting: whether the task is performed in a direct setting (interaction directly on the display) or in an indirect setting (interaction on a tablet in front of the display). Design Fully crossing all these variables would, of course, be unpractical. Instead focused on the interaction of ID(Distance, Height) x Angle within three typical configurations of orientation and setting. For ID, we selected ten possible Distance-Height combinations corresponding to 5 indices of difficulty (, 3, 4, 5, 6) as shown in Figure 5. Our choice was constrained both by practical concerns (e.g. in dialogue boxes, targets need to be far enough apart to allow for a label), and the size of the screen of our tablets. For the Angle, we selected 0, 15, 30, 45, 60, 90 degrees. While CrossY was built around the idea of a right to left traversal (angles from 180 to 70 degrees), we felt that this might be too unusual for untrained users. Instead we focused on the upper right quadrant (0 to 90 degrees) as a compromise between limiting the potential for occlusion and familiarity. In selecting values for orientation and setting, our first goal was to ensure comparability with prior work. Because previous studies have focused on indirect settings with crossing targets orthogonal to the main direction of movement, we included this setting in our experiment. In this condition (IndOrtho), users were interacting on a tablet placed in front of the computer, and the targets were always orthogonal to the centerline between the two targets. To study the possible effects of interacting directly on the screen, we also considered the direct orthogonal condition (DirOrtho), which was similar to IndOrtho, except that the interaction took place directly on the display of the tablet computer. Finally, we were interested in a situation closely simulating a typical crossing-based interface on a tablet computer. In the direct vertical condition (DirVert) participants interacted directly on the screen and the targets were always vertical. The screen for the orthogonal orientation is shown in Figure 4. Task Users were asked to cross two targets with a single stroke and without lifting the pen. After crossing the first target, a Figure 4 Screen shot of our test application. Shown here is the setting with the targets orthogonal to the main direction of travel. In the vertical task setting, the crossing targets stayed vertical within the context of the screen, the rest of the screen looked identical. 3

1.5 Indirect Orthogonal Setting (IndOrtho) ID = ID = 3 ID = 4 ID = 5 ID = 6 Crossing Time (s) 1 0.5 Figure 6 The setup for our experiment. Left: indirect setting; Right: direct setting. feedback sound was played and the color of the target changed. Upon crossing the second target, the task was completed and the next task appeared. An error was counted when the user lifted the pen before crossing the second target, crossed the target in the wrong direction, or did strokes that involved no crossing at all. To avoid counting simple landing (i.e. touching the pen to the screen) as an error, we introduced a minimum stroke length of 10 pixels. Thus, a simple touch of the screen, as it may occur between blocks, did not count as an error. To provide the user with feedback about their performance we provided two scales in the upper left corner of the application window. One represented the cumulative error rate and the other the completion time for each connection. The first scale was a half-circle consisting of three evenly sized parts colored in blue, green and red with a representation of an error rate from 0% on the left side to 8% on the right side of the half circle. The second feedback widget showed a diagram where the y-axis reflected completion time and the x-axis represented the trial number. Parallel to the x-axis, a line indicated the average completion time for the task (based on pilot studies). After each trial, a dot appeared on the diagram and gave feedback about the current speed of the user. Method For this experiment we adopted a within-subject design as skill variation between subjects might be important and little or no asymmetrical skill transfer was expected. The presentation of different distance by height by angle combinations was randomized. To reduce measurement noise, each combination was repeated 10 times in a row. As a result, participants performed 600 connections in each device setting. To limit the influence of skill transfer, we used a fully counterbalanced design for the order of conditions (IndOrtho, DirOrtho and DirVert). Overall participants were asked to complete 1800 connections. Apparatus The main apparatus of the experiment was our test application running on two Toshiba Protégé Tablet PCs, both with 1.5 GB RAM and one with 1.7 GHz CPU frequency and the other with 1.5 GHz. To ensure that the 0 0 30 60 90 α (degrees) Figure 7 The effects of ID and Angle on Crossing Time in the indirect orthogonal setting. different CPU frequencies had no influence on our results we fully balanced the tasks that were done on the different computers. The diagonal of the screen was 307mm and the resolution was set to 1400 x 1050 pixels. The tablet PCs were used in the folded configuration when they appear as a slate to users (see Figure 6). For the indirect condition, we used a Wacom IntuosII tablet where the gain between pen movement and cursor movement was set to 1.8. This value corresponds to an average of the values reported in the literature [3, 1]. To prevent confusion and errors, the buttons on the pen for the Tablet PCs and the pen for the Wacom tablet were disabled. The test application was written in C#. Apart from presenting the different tasks, it also logged all actions performed by the users. Protocol The subjects were 1 students at the University of Maryland (7 male, 5 female; age range 18-30 years). One of the subjects was left-handed. The three configurations (IndOrtho, DirOrtho, DirVert) were presented in blocks. Before completing the actual task, users completed a training session of 3 connections per ID/angle combination. Subjects were asked to complete the crossing as fast as possible, but with a certain precision. They were asked to keep the error rate within the middle area of the error scale described above. This represents an error rate of 4%. Since we were measuring only the times from first crossing to second crossing, users had a chance to rest as soon as they completed a crossing. Users were wearing headphones connected to their tablet as some users were run in pairs in the same room. Users received $0 for their participation. RESULTS AND ANALYSIS For our analysis, we used the average times for the connections performed under each combination of angle, ID, and condition. Bonferroni adjustments for multiple comparisons were used. The overall error rate for all the settings, trials and users was 6.37%. The error rate for the indirect setting (8.54%) was higher than the one for the direct setting (5.8%). 4

1.5 Direct Orthogonal Setting (DirOtho) ID = ID = 3 ID = 4 ID = 5 ID = 6 1.5 Direct Vertical Setting (DirVert) ID = ID = 3 ID = 4 ID = 5 ID = 6 Crossing Time (s) 1 0.5 Crossing Time (s) 1 0.5 0 0 30 60 90 0 0 30 60 90 α (degrees) α (degrees) Figure 8 The effects of ID and Angle on Crossing Time in the direct orthogonal setting. Figure 9 The effects of ID and Angle on Crossing Time in the direct vertical setting. To evaluate the possibility of asymmetrical skill transfer [13], we performed a Condition (IndOrtho, DirOrtho, DirVert) x Order Analysis of Variance (ANOVA). There was no main effect of presentation order (F,15 = 1.61, p =.33) and no interaction between Condition and Presentation Order (F,33 = 0.587, p =.56). To examine the influence of different CPU frequencies we also performed an ANOVA with Condition as a within-subject factor and CPU frequency (1.7 vs. 1.5 GHz) as a between-subject factor with average crossing time as the dependent variable. We found no significant main effect of CPU frequency on total crossing time (F 1,0 = 1.44, p =.44) No interaction between CPU frequency and Condition was found (F 1,0 = 1.437, p =.45). Next we performed a Condition x ID x Angle withinsubject ANOVA on average crossing time. There was a significant main effect for Condition (F, = 15.7, p <.001). Post-hoc tests indicated that the indirect condition was significantly slower than either of the direct conditions (p <.005). There were no significant differences between the direct conditions (p =.34). These findings are not surprising, because hand-eye coordination provides a significant advantage in the case of the direct conditions. This is consistent with the comments of users who reported that they felt more control over the interaction in these settings. We also found a significant main effect of ID, (F 5,44 = 656, p <.001), reflecting Fitts law. In addition, there was a main effect of Angle (F 5,55 = 34.4, p <.001). This latter finding was unexpected and we will now provide a more detailed analysis for each condition to illustrate the nature of this effect. The indirect orthogonal condition We performed an ID x Angle within-subject ANOVA on average crossing time in the indirect orthogonal condition (Figure 7). It revealed a strong main effect of ID (F 4,44 = 639, p <.001) suggesting a direct linear dependency between total crossing time and ID which reflects Fitts law: 9 ID - 463, (r = 0.996) The main effect of Angle (F 5,55 = 1.6, p =.17) and the interaction between ID and angle (F 0,0 = 1.34, p =.3) failed to reach significance. In general, these findings are consistent with previous studies [9, 1], although in our case, the 45 degree angle does not seem to stand out. The direct orthogonal condition We performed an ID x Angle within-subject ANOVA on average crossing time in the indirect orthogonal condition (Figure 7). There was a main effect of Angle (F 5,55 = 7.89, p <.001) and an interaction between ID and angle (F 0,0 = 1.93, p =.01). Post-hoc comparisons suggested that the only significant differences were between angles lower than 15 degrees and the 90 degree condition (p <.03; mean difference around 115 ms.). Further, there was a main effect of ID (F 4,44 = 33, p <.001) reflecting that, as predicted by Fitts law, there was a strong linear dependency between the total crossing time and ID for each angle: 57 ID - 498, (r 45 ID - 444, (r = 0.956), Angle = 0 = 0.973), Angle = 15 74 ID - 539, (r = 0.96), Angle = 30 6 ID - 471, (r = 0.964), Angle = 60 98 ID - 544, (r = 0.975), Angle = 90 These results can be explained by the switch to a direct condition with no gain between hand movement and the cursor on the screen. With increasing angle, the movement shifts from a left-right direction to a bottom-top direction. 5

These different directions involve different muscle groups, particularly when ID (and hence the distance to travel) is high. Specifically, we observed that as the angle nears 90 degrees, the upper arm and the shoulder participate in the movement. These proximal joins have slower bit rates than more distal joins like the wrist [10] and this may account for the slower crossing times. The direct vertical condition We performed an ID x Angle within-subject ANOVA on average crossing time in the direct vertical condition (Figure 8). Again, there was a main effect for Angle (F 5,55 = 54., p <.001) and an interaction between ID and Angle (F 0,0 = 1.80, p =.03). The effect of angle was considerably stronger than in the direct orthogonal condition. For example the difference between Angle = 0 and Angle = 90 degree was 6 ms (p <.001). Setting aside the special case of angle=0 (for which we observed an unexplained bump at ID 6), we observed a strong correlation between angle, and the difference between the connection time at that angle vs. the connection time at 15 degrees (Figure 10). Again there was a strong main effect on ID (F 4,44 = 380, p <.001) reflecting Fitts law. For each angle we observed a strong linear dependency between the total crossing time and ID: 0 45 60 IndOrtho DirOrtho DirVert 66 ID - 546, (r = 0.94), Angle = 0 35 ID - 440, (r 43 ID - 404, (r = 0.968), Angle = 15 = 0.989), Angle = 30 54 ID - 39, (r = 0.989), Angle = 45 47 ID - 89, (r = 0.989), Angle = 60 18 ID - 91, (r = 0.991), Angle = 90 As for the DirOrtho condition, these results can be explained by the progressive involvement of more proximal joints in the movement. Yet, the effect is much stronger than in the DirOrtho condition and additional factors may play a role. One possible explanation might lie in the dynamic of the movement itself [14] (see Figure 11). As expected, the IndOrtho condition appears noisier than the DirOrtho - probably because of the high gain for this condition. However, the most interesting difference is between the DirOrtho and DirVert condition. Participant no longer take a straight path between the two crossing targets but start to use more S-like strokes to achieve shallower entry and exit angles into the crossing targets. We have yet to formally characterize this effect, but we are planning to explore this track in the future. User feedback The results of the user testing show that the crossing paradigm is easy to understand and to transfer into action even by novice users. Overall, the direct setting was in all cases preferred to the indirect setting. This suggests that direct pen-based interactions are easier to perform than indirect pen-based interactions. All subjects agreed that the indirect setting was more difficult to handle and more error prone. This is reflected in the increased crossing time (see Figure 7) and a higher error rate. DISCUSSION As discussed above, our results have important implications for the design of pen-based interfaces. In this 0.5 Additional Time compared to Fitts' law Additional direct crossing vertical setting time (DirVert) (as compared to α = 15) as a function of α 90 Crossing Time (s) R = 0.99 Figure 11 Users take different paths depending on the setting and path angle. In the first and second column (Orthogonal setting), paths are relatively straight for all angles. In the third column (Vertical setting) we see an increasing in curvature with increasing angles. The paths shown are the 9 th and 10 th strokes of each user for each setting for ID = 5 (Height = 0 pixels). 0 0 30 60 90 α (degrees) Figure 10 The influence of angle on additional crossing time averaging over all IDs. ( t ref = 15deg (angle) = 3.64(angle 15) + 15.9, r =0.99). 6

Figure 1 The design of a simple dialog box. section, we illustrate how our findings can be applied to develop guidelines for the layout of crossing-based applications. The Layout of a dialog box with two columns For illustration purposes, we consider the simple case of a two-column dialog box as shown Figure 1, and restrict ourselves to the case of interactions in the first quadrant (e.g., selecting the font "Georgia" and font size "1" in sequence). In designing such a dialog box, one first needs to consider how much space is required to comfortably perform this selection. This includes the space needed to land the pen before the connection, the minimum space required between the two columns to be sure that the S- shape connections observed at high angles will not accidentally trigger targets in the second column, and the space needed to take off the pen before crossing the dialog box boundary as this action might be used as a validation command [4]. Landing and takeoff space To better understand the user requirements for landing and takeoff we plotted the distribution of landing and takeoff points in the present study (Figure 1). As the target grows bigger, the users are able to comfortably land their pen further away. This can be explained by the fact that they are trying to maintain a comfortable angle through which they are seeing the target upon landing []. A similar behavior is observed for take-off. From a design perspective, this means that one has to reserve a 70 pixel (1.3 mm) margin to the left of the first column and a 100 pixel (17.5 mm) margin to the right of the second column if one assumes a typical crossing target height of 30 pixels, as in CrossY. We also measured the maximum bounding box for the S shape connection curve. For the 90 degree angle, this box is 100 pixels (17.5mm) by 70 pixels (1.3mm), which implies that the distance between columns should be at least 00 pixels (35mm). Figure 13 Distribution of the landing points (left column) and take-off points (right column) relative to the crossing target height and the connection angle. Black vertical lines represent the crossing target position. Points further than 15 pixels from target center are not shown (less than 10% of all points). Influence of the layout on performance. As discussed above, in the DirVert condition, the crossing time between two targets whose centerline is α degree from the horizontal is given by the following formula (for α 15 degree 1 ): 35 ID 440 + 3.64(α-15) We used this formula to draw the contour plot presented Figure 14b. Given two targets (of height 0), one with its center at the origin of the graph and one with its center at the (x, y) coordinate, the value read at the (x, y) will provide an estimation of the crossing time. For example if one places a target at (400, 386) one might expect a 800ms crossing time. As a reference, we also show in Figure 14 1 For α < 15 degrees our data were inconclusive suggesting that users performed similar as in the horizontal condition 7

Figure 14 Each contour plot shows the predicted crossing time (in ms) it will take a participant to connect two targets (with the first target at the origin and the second target inside the first quadrant). Left: the predicted times for the IndOrtho condition. Right: the predicted times for the DirVert condition. The gray area shows where in the case of dialog boxes, the movement might be limited by Accot s law (see text). the same contour plot for the indirect condition. This graph confirms out informal observations during the development of CrossY that as the dialog box becomes taller it also become more cumbersome to use. Correcting errors. Another interesting question is how users react to missing a target and how they correct for it. After missing the second target, instead of starting over again, some users simply made a loop back to the other side of the target and attempted to cross a second time. See Figure 11 (DirVert, 45 ) where one stroke extends longer than the others. This stroke was not counted as an error since, after missing the second target, the user traveled back and crossed it successfully. This suggests that a backtracking path (e.g. for checkboxes) has to be considered. Continuous versus segmented strokes In this experiment we decided to explore single stroke connections. This decision was both practical (because errors were easily excluded) and inspired by our work with CrossY, where a stroke is used as a scoping mechanism. However, not all interfaces have this requirement and relaxing this constraint will likely improve crossing speed (especially for more difficult connections) [3]. It is possible that for segmented strokes, the effect of the angle will be different. We have not yet investigated this configuration, but pilot studies conducted during the design of this Figure 15 The influence of the angle between the target and the horizontal and target height on entry angle. Figure 16 The influence of the angle between the target and the horizontal and target height on exit angle. 8

experiment, suggest that the effects will still be present in the multi-strokes configuration, though less pronounced. Beyond two columns Of course, applications may use more than columns in their dialog boxes. In the present study, we did not investigate situations in which more than two marks are crossed in succession. Yet, our data provide us with two important pieces of information: the target entry angles and the target exit angles. These values are graphed in Figure 15 and Error! Reference source not found.. The figures show that for low angles, the entry and exit angles roughly correspond to the angle of the target centerline. However, after the angle reaches 0 degrees, users limit the entry angles around 30 degrees. A simple additive formula for successive connections will at least require input and output angle to be similar (so that there will be very little influence from the first path to the second) Our data suggest that even within the restrictions of the first quadrant, interactions between successive connections are to be expected. Of course, the problem will be even more acute in the case of connections which require users to first go up and then go down. Additional experiments are needed to explore these cases. Relationship with the Steering law It is interesting to interpret our results in the context of the Steering law [1]. Our experiments did not constrain the movement of users (either visually or by creating a do-notcross barrier). In a more typical configuration (e.g., Figure 1) users will perceive a sort of tunnel between different columns of widgets. Conceivably, user may limit the speed at which they approach this tunnel. At first, this seems to limit the scope of our results. However, it is important to remember that according to Accot, there are upper bounds to the validity of the Steering law. He set the upper limit at 80 pixels and above (corresponding to 134 pixels in our screen resolution). To reflect this fact, we grayed out the left part of our contours plots (Figure 14).Furthermore, aside from the extreme case of 90 degrees, the vast majority of trajectories in our study were inside the bounding box created by the two marks on the screen. Therefore, we argue that the steering law did not have a significant influence on crossing time in the present setup. Traversing a cascading menu Importantly, our results may be used to estimate the time it takes to navigate through the first level of hierarchies within wide cascading menus (e.g., the Start menu in the Windows operating system). In such menus users are often allowed to exit the visual boundary of the menu without penalty. We believe that our results will complement the results presented by Accot in cases where no constraints restrict the natural movement envelope. Further experiments will be necessary to understand how the presence of a constraint not limiting the natural movement envelope may influence users behavior. E 1 E E 3 E 4 E 5 Figure 17 Interaction between standard event dispatch and crossing interfaces. As seen here, in cases where the cursor is moving rapidly, small crossing goals might be crossed without receiving any event to identify the crossing. Implementation issues When adapting our application to be usable across conditions, we discovered that crossing-based interfaces may require a different kind of dispatch mechanism. In current traditional windowing systems, mouse (or pen) events are dispatched one event at a time. As a result, if sub-windows on the screen are small, the dispatch mechanism does not guaranty that a given window will see the starting and ending point of a path crossing over that window. This is perfectly alright for a point-and-click interface, because users will eventually stop the mouse on the top of a button or any other widget. However, the same dispatch mechanism is problematic in the case of crossing interfaces. Especially when users are moving rapidly over a target, the sub-windows containing this target may not receive any events (Figure 17). To address this problem, our simple experimental application (with targets) dispatched each event to all sub-windows. Of course, this solution will not scale very well. For larger applications, we suggest to not dispatch one point at a time, but one segment at a time [previous_point, current_point], to all windows whose bounding boxes intersect with the segments bounding box. Like the current point-based dispatch, this approach can be applied recursively and can be computed efficiently. It will also take care of cases in which several widgets are crossed in one segment. Importantly, in the latter case, some kind of precedence mechanism will be needed, if the order in which goals are crossed is important. Finally, this approach will also generalize to cases in which the window border is an active goal (e.g., in CrossY dialog boxes [4]). FUTURE WORK This work is only the first step in our empirical exploration of crossing-based interfaces. In the future, we would like to explore the behavior of more realistic configurations. First, we would like to explore the behavior of the system when the users are allowed to lift the pen from the screen during a connection. We would also like to understand better how the presence of visual elements, 9

either in the path (such as labels), or visually bounding the path without limiting its natural bounding box (such as other crossing targets), may change user interaction. We would also like to investigate timing characteristic in the case of 3 or more goals. CONCLUSION In this paper we presented the first steps toward a better understanding of the parameters influencing the performance of crossing interfaces. We showed that for the direct configuration with vertical targets, connection time depended on the angle between the centerline of the two targets and the horizontal. We also showed that our empirical results can be applied to better understand the implications of a given layout on overall crossing performance. Finally, we show that our results might also apply in the case of classical interfaces such as large cascading menus. ACKNOWLEDGEMENTS Removed for anonymous review. REFERENCES 1. Accot, J. and S. Zhai. Beyond Fitts' Law: Models for Trajectory-Based HCI Tasks. Proceedings of CHI'97, pp. 95-30.. Accot, J., Les Tâches Trajectorielles en Interaction Homme-Machine Cas des tâches de navigation., PhD thesis, Université de Toulouse 1. 001 3. Accot, J. and S. Zhai. More than dotting the i's --- foundations for crossing-based interfaces. Proceedings of CHI'03, pp. 73-80. 4. Anonymous. CrossY: A crossing based drawing application. Proceedings of UIST'04, pp. (In press). 5. Baudisch, P. Don't click, paint! Using toggle maps to manipulate sets of toggle switches. Proceedings of UIST'98, pp. 65-66. 6. Card, S.K., W.K. English, and B.J. Burr, Evaluation of Mouse, Rate-Controlled Isometric Koystick, Step Keys and Text Keys for text selection on a CRT. Ergonomics, 1978. 1(8): p. 601-613. 7. Fitts, P.M., The infomation capacity of the human motor system in controlling amplitude of movement. Journal of Experimental Psychology, 1954. 47: p. 381-391. 8. IBM, Lotus Notes (http://www.lotus.com). 004. 9. Jagaconski, R.J. and D.L. Monk, Fitts' Law in Two Dimensions with Hand and Head Movements. Journal of Motor Behavior, 1985. 17(1): p. 77-95. 10. Langolf, G.D., D.B. Chaffin, and J.A. Foulke, An investigation of Fitts' law using a wide range of movement amplitudes. Journal of Motor Behavior, 1976. 8(): p. 113-18. 11. MacKenzie, I.S., Fitts' Law as a Research and Design Tool in Human-Computer Interaction. Human- Computer Interaction, 199. 7(1): p. 91-139. 1. MacKenzie, I.S. and W. Buxton. Extending Fitts' law to two-dimensional tasks. Proceedings of CHI'9, pp. 19-6. 13. Poulton, E.C., Unwanted asymmetrical transfer effects with balanced experimental designs. Psychological Bulletin, 1966. 66(1): p. 1-8. 14. Viviani, P. and C.A. Terzuolo, Trajectory determines movement dynamics. Neuroscience, 198. 7: p. 431-437. 10