Hierarchical Route Maps for Efficient Navigation

Home , Route (command)

Hierarchical Route Maps for Efficient Navigation Fangzhou Wang 1 Yang Li 2 Daisuke Sakamoto 1 Takeo Igarashi 1,3 [email protected] [email protected] [email protected] [email protected] 1The University of Tokyo 2Google Research 3JST ERATO Igarashi Design Interface Project a b c Route&Tree Genera,on

Figure 1: An example of Route Tree. The proposed algorithm automatically extracts meaningful multi scale views of a route and organizes those views into a hierarchical structure to support efficient navigation. (a) shows the original route and a complete overview of the Route Tree generated form the route, (b) and (c) show close-up view of the Route Tree. Note that these are visualization of the internal data structure and the end-users do not see them directly. ABSTRACT One of the difficulties with standard route maps is Author Keywords Route Map Visualization; View Extraction; Multi-scale accessing to multi-scale routing information. The user Navigation. needs to display maps in both a large scale to see details and a small scale to see an overview, but this requires ACM Classification Keywords tedious interaction such as zooming in and out. We propose H5.2 [Information interfaces and presentation]: User to use a hierarchical structure for a route map, called a Interfaces. - Graphical user interfaces; H.3.3 [Information “Route Tree”, to address this problem, and describe an storage and retrieval]: Information Search and Retrieval. - algorithm to automatically construct such a structure. A Retrieval models. Route Tree is a hierarchical grouping of all small route segments to allow quick access to meaningful large and INTRODUCTION small-scale views. We propose two Route Tree applications, A route map presents route information to guide users from “RouteZoom” for interactive map browsing and “TreePrint” one place to another. Various systems have been developed for route information printing, to show the applicability and for automatically generating route maps. These systems usability of the structure. We conducted a preliminary user have been conventionally implemented in car navigation study on RouteZoom, and the results showed that systems with a local map database. The use of web-based RouteZoom significantly lowers the interaction cost for route map generation systems, such as Google Maps [9], obtainingFigure 2: informationAn example fromof Route a map Tree. compared The proposed to a traditional algorithm automaticallyhas become extracts commonplace meaningful multi for scale various views types of a route of and users interactiveorganizes map those. views into a hierarchical structure to support efficientrecently navigation as smart. (a) phoneshows sthe and original tablet route PCs and become a complete popular. overview of the Route Tree generated form the route, (b) and (c)Ho showwever, close several-up view usability of the Route problems Tree. Notestill remainthat these with are these visualization of the internal data structure andmap the endsystems.-users do not see them directly.

Permission to make digital or hard copies of all or part of this work for One problem with standard map systems is the lack of personal or classroom use is granted without fee provided that copies are support for presenting multi-scale information. For example, not made or distributed for profit or commercial advantage and that copies when users want to see an overview of the entire route, a bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. smaller scale view is preferred. On the other hand, when Abstracting with credit is permitted. To copy otherwise, or republish, to they want to get detailed information on points of interest post on servers or to redistribute to lists, requires prior specific permission (POIs) such as turning points or junctions, they would and/or a fee. Request permissions from [email protected]. prefer a larger scale view to see more details. However, IUI'14, February 24–27, 2014, Haifa, Israel. Copyright © 2014 ACM 978-1-4503-2184-6/14/02...$15.00. standard paper-printed route maps or interactive route map http://dx.doi.org/10.1145/2557500.2557514 systems in a modern map system [9, 19] require the user to manually change both scale and position in the map to get a desirable view, which can be very tedious. Therefore, it is information such as landmarks around them or the shape of desirable to provide an easy way to access various the road. Furthermore, the distorted representation makes it meaningful views with different scales and positions in the difficult to use the map in combination with advanced route. services such as aerial photo views. We propose to pre-compute a hierarchical structure for a Several existing web-based map systems [9, 19] are able to route map (called a “Route Tree”) to provide quick access print a step-by-step guide showing all POIs and to such meaningful views (Figure 1). The important corresponding verbal directions along the route. Karnick et features of our hierarchical structure are as follows: al. [16] proposed a method of putting detailed views of POIs in the route around the complete overview of the route. Feature 1. Each node represents a part of its parent view in This can be problematic because the user might want to see a larger scale, while the root node represents a complete the map in some intermediate scales. Furthermore, some of view of the whole route. the detailed views are almost the same and therefore Feature 2. For every part of the route, there is a node that redundant because the method shows a detailed view for visualizes the part in the necessary scale. each POI even in a case where it is possible to show multiple POIs in a single view. Feature 3. The number of nodes in the tree is made as small as possible, while keeping an appropriate view size Several previous studies [22, 25] attempted to exploit the (size of the rendered result on the screen or print) for each user’s prior knowledge about the route in order to produce a node. more simplified, useful route for users. Other studies [5, 11] propose methods that combine videos or images with We present an algorithm for automatically constructing a conventional route maps to enrich the user’s knowledge Route Tree. It starts with an over-segmented tree and about the route. There are also systems that generate Goal- iteratively simplifies it by taking balance between oriented route maps such as tourist maps [10] or compactness and readability. A Route Tree itself is raw tree Destination-oriented maps [18] rather than Start-to-Goal data, and it can be used in various applications. In this route maps. paper, we implemented an interactive map browsing system called RouteZoom to demonstrate the effectiveness of a Hierarchical map structure RouteTree. There are several studies on hierarchical structures for map Our method is designed to address the situation where a visualization. However, most of them discuss how to user needs to actively learn or explore an area or route. It is automatically make a series of multi-scale map images from different from but complements automatic uses of maps geographic data while including information in an systems such as Turn-by-Turn navigation used in car appropriate level of detail (e.g., [20, 27]). They do not use navigation in which a user passively receives step-by-step them for route map visualization. Studies have been made instructions from the system. The contributions of this work on hierarchical structures and route maps, but most of them are summarized as follows. focus not on visualization but on human cognition on routes and its application to generating verbal route directions [17, 1. A novel hierarchical structure for route maps called a 26]. Our work differs from these studies in that our method Route Tree. generates a series of maps with appropriate detail in 2. A fast algorithm to construct the Route Tree with a few different scales for a route map, and uses them for parameters given by the user to control the results. interactive browsing or compact printing.

3. Two useful route map applications that exploit Route Map browsing interface Trees, respectively improved a conventional route map Various methods have been proposed for interactively browsing system and a route map printing system. browsing maps in different scales. Some of them [2, 4] integrated maps of multiple scales into one screen using RELATED WORK Focus+Context techniques such as Furnas’ fisheye [7]. Route map visualization Others adopt Overview+Detail techniques to arrange Agrawala et al. [1] have proposed a system to render a route pictures of different levels of focus and size [12, 15]. map that mimics handwriting maps made by humans. This Another approach is to extend conventional panning and approach focuses on conveying precise information about zooming interfaces. Igarashi et al. [13] proposed a method POIs and an abstract overview of the complete route. To to automatically change zoom level depending on the achieve the goal, it distorts the geographic shape and length scrolling speed for map navigation. Zhao et al. [28] of the route and takes away detailed context around the proposed to implicitly bookmark user’s view on the map to route. However, the lack of contextual information outside accelerate browsing speed. Both of the latter two techniques the route makes it difficult for users to turn back to the rely only on general user input and do not leverage precise route when they mistakenly choose a wrong one, or semantic information associated with routes. Pindat et al. to determine their precise position in the route map from [23] proposed content-aware adaptive fisheye lens that changes its way of zooming depending on its focus on map. POI NoPOI Turn Branch Continue However, the study focuses on interface design and does Road not discuss in detail how to pre-compute the necessary Motorway 7 11 17 10 semantic information it requires from a map. Primary 9 13 17 12 Robbins et al. [24] proposed a system called “ZoneZoom”, Secondary 10 14 17 13 which provides an effective way of navigation on a mobile Local 14 18 17 17 display. It divides the screen into several areas and allows the user to recursively zoom into them. Though they … … … … … suggested in their study that their method could be well Table 1: Example look-up table for Type(R) applied to map navigation, they did not discuss a method to automatically compute zoomable areas. We apply their in the map. For example, motorway segments are visible in method as the basis of our user interface for RouteZoom a small scale, but small street segments need to be in a application, but the zoomable areas are automatically larger scale to be visible, because in smaller scale it is computed by our method. usually not rendered on the map. A scale value is an integer ranging from 0 to around 20 (depending on the map) and ROUTE TREE CONSTRUCTION determines the approximate translation ratio from The construction process consists of three steps: 1) Scale latitude/longitude coordinates to pixel coordinates in the Evaluation, 2) Initial Tree Construction, and 3) map as Optimization (Figure 2). The input route data is taken from (x, y) = ( f (lat,lon), g(lat,lon)) a route-search engine (e.g., Google Directions API [8], 2Scale+8 NavEngine [21]), which includes information such as the f (lat,lon) = ⋅(lon − A) road’s geographic shape, the type of road, and POIs. First, 360 2Scale+8 we extract a set of “route segments” from the input route g(lat,lon) = ⋅(B − lat) data, which represent short, straight lines in the route, and 360 ⋅ cos(lat) use this information to assign scale values to the segments where x and y respectively indicate horizontal and vertical (Scale Evaluation). Second, we construct an initial tree by coordinates in the map image, and lat and lon indicate hierarchically organizing the route segments depending on latitude and longitude coordinates. A and B respectively their scales (Initial Tree Construction). Since this tree is indicate longitude and latitude at the top left of the map over-segmented, we optimize the nodes to an appropriate image [3]. number and size of view to get the final tree (Optimization). We use the following function to evaluate the scale: Scale Evaluation

The route shape is initially given as a polyline, and we Scale(R) = max{Type(R), Dist(R)} divide the polyline into short, straight segments with nearly where R indicates a route segment. equal lengths. Next, we evaluate a scale value for every route segment. A scale value indicates the minimum scale Type(R) returns a scale value of the segment defined by the necessary to present enough information about the segment road type and the type of POI it contains. It depends on the input map data. Table 1 shows the definition for Google Maps. A larger scale value is defined for segments that Input Data Evaluation Result contain POIs, because those segments usually need a higher scale to be visualized. It returns different values depending 1) Scale Evaluation on the type of POI (e.g., turning point, branch of roads, continue to a different street), because each of them has a different level of importance for the user and a different scale value for visualization. Dist(R) is added in order to 2) Initial Tree prevent neighboring POIs from being too close in the Construction rendered view. It measures the distance between R and Initial Tree Final Result neighboring POIs, and returns a minimum scale value to make the distance larger than a predefined value (we use 30 pixels in our implementation) on the screen. 3) Optimization Initial Tree Construction Figure 3 shows an overview of the method in this step. We first prepare a copy of segmented routes for each scale value within the defined range (Figure 3a). We remove all segments whose scale is smaller than the scale level Figure 2: Overview of Route Tree construction process. associated with each copy (Figure 3b). We then group a. Prepare Copied Routes b. Remove Segments nodesize(n) 1 Copies of routes a b Scale Value 0

small large return value

-1 Mmin view size p・Mmax Mmax small large

c. Create Nodes d. Resultant Tree where a, b, and p are values ranging from 0.0 to 1.0, Mmin and Mmax are respectively the minimum and maximum limit of the node size, and s(n) represents the size of node n. All these parameters are application-specific i.e. different values are set for different applications. The function sets the ideal node size to p • Mmax and gives a negative value as a penalty when the size exceeds its defined limit Mmax too small large small large much.

Figure 3: Initial tree construction process. The definition of childsize(n) are as follows: adjacent route segments in each scale level, making nodes #(M / s'(n)− m)2 & childsize(n) exp max (Figure 3c). Each node is associated with a scale level. = % 2 ( $ 2 ⋅σ ' Finally, an initial Route Tree is created by connecting each node on the basis of the containment relationship among the where s´(n) is the size of node n in its parent view. The segments (Figure 3d). function sets the ideal child node size to be 1/m of the Mmax in its parent view, and thus avoids the problem of the area a Optimization child node indicates being too large or too small in its The initial tree is over-segmented and each node is too parent view. small at this point. We therefore optimize the tree by minimizing a cost function to reduce the number and GA Optimization increase the size of nodes. We use a genetic algorithm (GA) for the optimization. The basic idea is to repeatedly apply the following four Cost Functions operations to the tree in order to reduce the number and The definition of the cost function is as follows: increase the size of nodes.

" 15 $ 10 (score(n) ≤ 0) 1. Integrate a node to its child nodes (Figure 4a). cost(t) = ∑# n∈t %$ 1/ score(n) (score(n) > 0) 2. Integrate a node to its adjacent node (Figure 4b). score(n) = α ⋅ nodesize(n)+ β ⋅ childsize(n) 3. Increment the scale value of a node (Figure 4c). where t represents a Route Tree and n represents its node. 4. Add 10 segments (5 for each side) to a node (Figure 4d). We set α=0.5 and β=0.5 for weighting parameters α, β in our implementation. The score function nodesize(n) is a a.Integrate to child nodes b. Integrate to adjacent node restriction against the view size of node n and childsize(n) parent node parent node is one against that of the child nodes of n. Increasing parameter α or decreasing β will better preserve every node size to be similar but randomize the size of area every child nodes indicates in its parent view, and vice vasa. The greater the returned score(n) value is and the smaller the child node number of nodes n in tree t is, the less is the cost(t) that is child nodes returned. The definition and graph plot of nodesize(n) are as c. Scale value +1 d. Add segments follows: parent node parent node $ & exp(A⋅(s(n)− p⋅ M max )) (s(n) < M min ) & & 2 nodesize(n) = % B⋅(s(n)− p⋅ M max ) +1.0 (M min ≤ s(n) ≤ M max ) & child node child node & 2 C ⋅(s(n)− p⋅ M max ) +1.0 (M max < s(n)) '& log(a) a −1.0 b −1.0 Figure 4: Operations in optimization process. A = , B = 2 ,C = 2 M min − p⋅ M max (M min − p⋅ M max ) (M min − p⋅ M max ) The algorithm first prepares 500 candidate trees by (a) User Interface randomly applying the four operations to the initial tree. It then iteratively reduces the cost of the trees by applying Main window crossover and mutation. The probabilities for crossover and mutation are respectively 0.5 and 0.3. Crossover randomly mixes two selected trees and mutation randomly applies one of the four operations to each node in the selected tree with Suggestions for a probability of 0.1. We use a simple roulette-wheel Smart zoom selection process to make successive generations, and the Commands for fitness of each seed is defined as the inverse of the cost Smart slide Command for evaluated by the Route Tree cost function, After repeating Smart zoom-out the process 1000 times, the algorithm returns the tree with the lowest score as the final solution. (b) Smart zoom & Smart slide

ROUTE TREE APPLICATIONS We implemented two Route Tree applications: Zoom-in “RouteZoom” for interactively browsing route maps on a display, and “TreePrint” for printing routing information. Zoom-out We will describe the details of each application respectively in the following two subsections. Next view RouteZoom Real-time navigation systems automatically move the map view according to the current position. However, the users Previous may want to browse the map under certain circumstances, view e.g., to allow them to 1) memorize directions for parts of the route they are unfamiliar with, and 2) see information Figure 5: (a) User interface of RouteZoom (b) Smart zoom for the next few turns to prepare their mind during driving. and smart slide functions in RouteZoom. Therefore, it is still important to improve techniques for browsing route maps for developing better navigation next or previous to the current viewport. From the gathered systems. nodes it then chooses one target node that has the nearest Figure 5 shows the user interface and functions of scale value to the current view scale. If the difference RouteZoom. The system has two main functions, the first of between the scale value of a target node and that of the which is “smart zoom”. With this function, the system current view is less than 3, it simply slides the view to the automatically suggests areas in the main map that the user target node. In case the difference is greater than or equal to may want to zoom in on to see the details. The user has 3 we maintain the current scale and slide the view to the only to select a suggested area to get a detailed view of it. next or previous POI’s position. The user can also get an appropriate zoomed-out view by right-clicking on the map or selecting the (-) command at TreePrint Although real-time navigation systems are very common the bottom right corner. Note that this function not only these days, there are still strong needs to print out route scales the view but also slides it to efficiently show the maps in some situations. We conducted an informal survey route, while in a standard map application the user has to to ask 20 daily map users whether they use a printed route adjust the view position manually after changing the scale. map. Surprisingly, 16 of them answered, “Sometimes I may The second function is “smart slide”. With this function, the need to print out a route map” and two answered, “I often user can move forward or backward along the route by print out a route map”, while only two answered, “Not selecting a command button on the top or bottom of the necessary at all”. They listed several cases when they parent view. The position and scale are automatically needed to print out maps such as traveling abroad in a adjusted after the slide. rented car without a car-navigation system, making a The construction process is as follows. First, we construct a backup for a cell phone low on battery power or a low- Route Tree from the route. We set construction parameters accuracy GPS, or just wanting to save money for their cell- Mmax =800x600px, Mmin=50x50px, a=0.1, b=0.0, p=0.9, phone Internet connection. m=5, σ=1.0. Second, we trace the parent-child relationship The TreePrint system produces a collection of map images in the tree to construct the smart-zoom function. Third, we for showing route information in three levels of views. It scan the Route Tree from the root to the leaves and connect produces a complete overview of a route, detailed views for each POI in the route to nodes that contain it. The smart- each POI, and mid-scale views between them. Figure 6a slide function gathers nodes connected to the POIs that are (a) TreePrint layout views clarify the spatial relationships between multiple POIs, such as the road shape or the distance between them. This is a clear improvement over conventional systems that have difficulty in visualizing this information when multiple POIs are concentrated in a short part within a long route.

15. Turn left onto Bridge St go 0.9 mi About 3 mins total 19.0 mi The TreePrint application first produces a Route Tree for the route it prints. The root node of the tree, nodes with

16. Turn right onto St Peter St go 0.1 mi total 19.1 mi depth 1, and nodes with depth 2 respectively represent a complete overview, mid-scale view, and POI-level views.

No. Page No. Page 17. Turn left onto Brown St We set construction parameters Mmax and Mmin depending 1 2-3 5 7 2 4 6 8 on the depth of the node, which is defined by the map size 3 5 7 9 4 6 8 10-11 on each view level. We set the same a, b, p, σ values used to construct RouteZoom, while setting larger m value m=16 Overview mid-scale & POI view to reduce the depth of the tree. After Route Tree construction, we set the depth of all leaf nodes to be 2 by (b) Google Maps layout iteratively removing parent nodes of leaf nodes whose depths are larger than 2, and adding a child node for leaf nodes whose depths are less than 2. We set the scale value as 15 for newly added children with depth 1 and one greater 17. Turn left onto Brown St go 0.1 mi total 19.2 mi than its parent node’ s scale for depth 2.

18. Slight left at Kimball Ct go 115 ft total 19.2 mi RESULTS Figure 7a shows an example of view transitions in

Unknown road RouteZoom and Figure 7b shows an example of printouts generated by TreePrint. These results are generated from identical route from The Community College of Baltimore to Sinai Hospital of Baltimore. In Figure 7a, views for which a higher scale is needed to see the details are automatically suggested in each of the Overview mid-scale & POI view views. These views are appropriately connected and the user can use the smart-zoom and smart-slide functions to Figure 6: Comparison of (a) TreePrint layout (b) Google Maps achieve easy and quick browsing of the route map. The red printing layout. For comparison, POI markers in Google Maps layout overview are manually added. The texts are enlarged arrows in the figure suggest zooming transitions between from original size for better visual presentation. the parent and child views while the blue arrows suggest sliding transitions. shows an example of a layout printed using map images produced by TreePrint. A complete overview of the route is In Figure 7b, the mid-scale view on pages 2-7 effectively printed on the first page. Mid-scale views are indicated on visualizes the spatial relationship between each of the POIs, the view with their index numbers. The user can refer to the and enables each POI view to have a larger scale to give table shown at the bottom to jump to pages with more detailed information without losing context information of detailed information about the indicated view. Currently our the route. The clear hierarchical indexing between complete route tree implementation only generates individual map overview, mid-scale view, and detailed POI view extract images and the final layout with text description is views necessary to the user. manually generated. However, it will not be difficult to automatically generate this simple layout given the EVALUATION sufficient text instruction and icon image resources in We conducted a user study for RouteZoom by comparing modern map systems. them with conventional methods. RouteZoom is basically an extension of a conventional route map browsing The advantage of enabling a Route Tree to print is that it technique, and they both use similar types of map images optimizes the map views depending on the information it for presenting information. Therefore, there is essentially contains, while the conventional printing method in a no difference in the amount of information they can provide. modern map system produces map images with fixed scale The difference is in the speed at which correct information regardless of the information they contain. The optimization can be gotten and the interaction cost needed to get it. Thus, process avoids the producing of multiple similar views or our goal in this user study is to confirm the following two views with excessively small scale. Furthermore, mid-scale hypotheses: a zooming transition sliding transition

8. Turn left onto Aliceanna St/Van Lill St go 33 ft total 4.5 mi 15. Turn left go 0.1 mi About 45 secs total 12.5 mi 9. Take the 1st right onto Aliceanna St go 0.8 mi About 3 mins total 5.4 mi

5. Continue onto Poncabird Pass go 0.2 mi 10. At the traffic circle, take11. the Continue 1st exit ontoonto I-83S go 0.6 mi go 4.6 mi Unknown road About 7 mins total 5.9 mi total 10.5 mi total 2.5 mi About 3 mins Unknown road Take the 1st exit onto S 1. Head west About 3 mins go 269 ft total 269 ft

2. Turn left toward Delvale Ave go 167 ft total 436 ft

3. Turn right onto Delvale A 6. Continue onto S Ponca St go 0.3 mi go 0.2 mi ve 4 total 0.4 mi total 2.7 mi About 1 min

12. Take exit 9B for Cold Spring Lane W go 0.3 mi total 10.9 mi

13. Merge onto W Cold Spring Ln go 0.7 mi About 2 mins total 11.6 mi

No. Page No. Page 4. Turn left onto Holabird Ave go 1.8 mi 7. Turn left onto Boston St go 1.8 mi About 6 mins total 2.3 mi About 5 mins total 4.5 mi 1 2 4 5 7 2 3 5 6-7

14. Turn right onto Greenspring Ave go 0.8 mi 3 4 About 2 mins total 12.4 mi

1 2 3 6

1 2 3 4 - 7 Figure 7: (a) Resultant view transitions on RouteZoom. Note that the user can also zoom-out to parent view by right click or clicking (-) command button on bottom-right of the map. (b) Example printing layout using map images generated by TreePrint.

H The user can gain accurate route information in a shorter 1 Interfaces time using RouteZoom than using Baseline. Two interfaces were compared in the study: H2 The user can gain accurate route information with a RouteZoom: The RouteZoom interface allows the smaller number of interactions and in a shorter interaction participants to fully use the functions of the application that time using RouteZoom than using Baseline interface. we described in the previous section. To confirm these two hypotheses, we asked participants to Baseline: The Baseline interface allows the participants to work on a route information-searching task to measure the pan the map by mouse dragging and to zoom by double task completion time to operationalize H , and the number 1 clicking or scrolling the mouse wheel. of interactions and the interaction time taken to complete the task to operationalize H2. Note that H1 concerns the Although more advanced browsing interfaces such as overall time for completing a task while H2 only concerns multiview or fisheye exist, we chose a simple one-view the portion of the task in which the user interacts with the panning and zooming interface as the Baseline interface. route map. We recruited sixteen participants (twelve males, This is because 1) the Baseline interface is the one having four females) from an IT corporate environment to conduct the most basic and popular interaction style among this study. conventional map systems [9, 19] and 2) our method of adding semantic view extraction to navigation interface is orthogonal to advanced methods such as multiview or fisheye. They are not mutually exclusive and it is possible to combined with our method. Our study thus complements existing work [6] showing the advantages and disadvantages of such advanced methods.

Tasks and Apparatus The task was to browse a given route map using one of the two interfaces and answer a questionnaire consisting of five item questions about the route information such as “For this turn X, which is the right direction to take”, or “What is the approximate distance from turn X to turn Y?” Figure 8 Figure 8: Experimental system for user study. The texts are shows the screenshot of the system we used in the study. enlarged from original size for better visual presentation. The left panel shows the map and the right panel shows the questionnaire. Each questionnaire item has four possible session, the interaction time does not include time for answers. We prepared the questionnaires before inspecting reading maps and answering questions. We also asked the resulting Route Tree generated by the system to avoid participants to answer a questionnaire with 7-point Likert potential bias. scale and conducted an interview for subjective usability measurements. Each participant was asked to complete two sessions sequentially, one session using the Baseline interface and Procedures the other using RouteZoom. Each session consisted of four Before starting each session, the participants were trials, for each of which the task was to browse a route map instructed to complete two practice trials to familiarize and answer a questionnaire about the route. All participants themselves with the system and the browsing interface for worked on the same eight questionnaires, but the order and the session. condition assignments were balanced among the In each trial in the session, the participants began by participants. reading the in-session questionnaire for 30 seconds (the The study was conducted on a Lenovo S20 PC running map was hidden during this preparation time) and then they Ubuntu Linux, with an Intel Xeon 3690 CPU and a 24” interactively browsed the route map while answering the monitor with 1920x1200px resolution. The participants questionnaire within the next 90 seconds. used a standard mouse (Logitech G400) with a mouse wheel to interact with the system. After completing each session, the participants were asked to answer a post-session questionnaire about the browsing Measurements interface used in the session. Finally, we conducted a post- We measured the performance of each participant by 1) session interview to collect feedbacks. Session completion time, 2) Answer accuracy of the questionnaire, 3) Number of interactions such as clicking or Evaluation Results mouse wheel rolling, and 4) Interaction time for operating Figure 9a shows the quantitative results for mean session the map. Note that while the session completion time completion time, answer accuracy, number of interactions, addresses the total time taken for completing all tasks in the and map interaction time taken. The post-session Result$of$ques7onnaire$with$7>likert$scale

Baseline$ RouteZoom$ *"p"<".05""""**"p"<".01 Results$of$Ques6onnaire$with$7=point$Likert$Scale 7$ 7$ ** ** ** * p$=$.007863 p$=$.0000597 p$=$.006897Interac1on&Time&per&Session&[s]&p$=$.1706 p$=$.1287 No.%of%Interac4ons%per%Session%p$=$.0101 6$ 80"" ** 200"" ** 70"" 5$ 6$ 60"" 150"" 50"" 4$ 40"" 100"" 30"" 3$ 20"" 50"" 10"" 2$ 5$ 55.021&& 32.724&& 136.313%% 70.125%% 0"" 0"" 4.88$ 5.94$ 4.19$ 5.94$ 4.69$ 3.06$ 2.75$ 3.50$ 4.38$ 3.31$ 3.88$ 5.38$ 1$ Q1$ Q2$ Q3$ Q4$ Q5$ Q6$ Session'Comple3on'Time'[s]' Answer%Accuracy%[%]% 4$ 360"" 100"" No. Questions p

300"" 80"" Q1 I can quickly browse route maps with this method. .008 240"" Q2 It is easy to browse route maps with this method. .000 60"" 180"" Q3 This method requires much operation to browse a route map. .007 3$ 40"" 120"" Q4 A lot of training is necessary to use this method. .171

60"" 20"" Q5 I get tired when operating maps with this method. .129 321.055'' 321.242'' 84.688%% 80.000%% 0"" 0"" Q6 I want to use this method in the future. .010 2$ (a) Objective Measurements (b) Subjective Evaluations Figure 9: Results for RouteZoom user study. 4.88$ 5.94$ 4.19$ 5.94$ 4.69$ 3.06$ 2.75$ 3.50$ 4.38$ 3.31$ 3.88$ 5.38$ 1$ Q1$ Q2$ Q3$ Q4$ Q5$ Q6$ questionnaire results (7-point Likert scale) are shown in interface (Figure 9b). In particular, they felt RouteZoom Figure 9b along with the questions asked. We respectively was a “quicker” way to browse the route maps than the used a paired t-test and a Wilcoxon signed-rank test to Baseline method. This is probably because the users had to analyze the quantitative and qualitative results. constantly move their hands and fingers to change the view with the Baseline method and thus felt the time taken for Objective Measurements the interaction was the bottleneck. RouteZoom removed There were no significant differences in the session this bottleneck and was positively received by the users. completion time for a session (p=.979), meaning that H 1 The results also show that the users felt it was easier to was not supported. On the other hand, there were browse route maps with fewer operations when using significant differences in both map interaction time taken (p=.000) and number of interactions (p=.000), while there RouteZoom. This corresponds to the reduced interaction- were no significant differences in the answer accuracy cost shown in the quantitative results. With respect to the (p=.141). This demonstrates hypothesis H was supported. learning effort needed for using these two methods, the 2 results for Q4 show that although the participants were Table 2 shows detail of interaction time taken and number already familiar with the conventional method, for them of interactions per session for each interface. The results there was no significant difference between the two shows that more zoom interactions are replaced by methods. This indicates that RouteZoom is easy to learn. automatic zooming ((za-zb)/za=0.70 in time and 0.68 in The participants also indicated they would like to use number) than pan interactions by automatic panning ((pa- RouteZoom in the future. p )/p =0.50 and 0.57). On the other hand, the replaced b a In addition to answering the structured post-session interaction cost is reduced more efficiently by automatic questionnaire, participants were asked to give comments panning (pb’/(pa-pb)=0.20 in time and 0.12 in number) than automatic zooming (z ’/(z -z )=0.42 and 0.31). and suggestions about RouteZoom in the post-study b a b interview. We received many positive comments, such as One possible reason H1 was not supported is that the users' that RouteZoom was “quick and easy” and that it provided behavior was mostly affected by the given time constraint “good positioning to appropriate views.” They also pointed (90 sec per question set) and not by the technique. The out usability issues that we can improve on, such as users apparently used the extra time due to the reduced “Sometimes suggestions in the map were too small to read interaction time to work on problems longer. This did not easily.” It is worth mentioning that while several contribute to improved answer accuracy because most of participants highly preferred the smart slide function, the mistakes were the result of fundamental several others never used it during the study and misunderstandings about the route information. commented that they did not want to shift their focus from the map to the command button at the bottom left corner. Although H was not supported in our study, the reduced 1 We can easily address this issue by providing a keyboard number of interactions and interaction time by RouteZoom shortcut for triggering the function. would have considerable impact on navigation safety, because less interaction cost will greatly help to prevent the DESIGN IMPLICATIONS user from being distracted from their driving or walking. It A Route Tree is a general representation with various may also have a positive impact on the session completion possible applications, which is not limited to two time in terms of practical conditions, because many car applications we proposed in this paper. For example, navigation systems adopt more adverse input conditions advanced browsing techniques such as multiview control than those in our user study (high performance wheel and fisheye zooming can be applied to a Route Tree. We mouse and large display). can enable users to select suggested areas on the parent view to instantly change not only the position but also the Subjective Evaluations zoom level of the detailed view. By enabling the Route The post-session questionnaire results showed that on the Tree to support the fisheye technique, we can allow the whole the participants preferred RouteZoom to the Baseline system users to automatically change the zoom level Baseline (a) RouteZoom (b) depending on the position of the fisheye lens on the map. Method Time (s) No. Time (s) No. A Route Tree also enables the user to control the route map by giving voice commands such as “Zoom in to area 1”, Zoom (z) 28.33 90.44 8.42 29.38 “Zoom out”, or “(Slide to) next”. This would be much Pan (p) 26.69 45.88 13.22 19.75 faster and less distracting for the users than existing voice- Auto zoom (z’) - - 8.33 18.88 enhanced browsing methods (e.g., [14]), because most of Auto pan (p’) - - 2.76 3.06 them require the user to constantly focus on the display. A Route Tree could also improve the automatic view Table 2: Comparison of interaction costs between two interfaces. Note that smart slide function in RouteZoom not transition function provided by real-time step-by-step only automatically pans the view but may also zoom it. navigation system. It would reduce the number of view transitions by optimizing the number of views to a interfaces with and without an overview. ACM TOCHI 9, minimum while preserving necessary information for the 4 (2002), 362–389. users. This would provide a less stressful navigation 13. Igarashi, T. and Hinckley, K. Speed-dependent experience to the user. automatic zooming for browsing large documents. In Proc. UIST 2000, ACM Press (2000), 139-148. CONCLUSION We proposed a hierarchical structure for route maps called 14. Igarashi.T. and Hughes, J.F. "Voice as Sound: Using a “Route Tree.” It helps users to move back and forth across Non-verbal Voice Input for Interactive Control" In multiple views with different scale levels and positions in a Proc. UIST 2001, ACM Press (2001), 155-156. map. We also proposed an algorithm to automatically 15. Javed, W., Ghani, S., and Elmqvist, N. PolyZoom: construct a Route Tree and two Route Tree applications. Multiscale and Multifocus Exploration in 2D Visual These are “RouteZoom”, an interactive map application Spaces. In Proc. CHI 2012, ACM Press (2012), 287-296. where the user interactively browses a hierarchical map, 16. Karnick, P., Cline, D., Jeschke, S., Razdan, A., and and “TreePrint”, which uses the hierarchical structure to Wonka, P. Route visualization using detail lenses. IEEE generate 3 kinds of map images suitable for printing out. TVCG 16, 2 (2009), 235-247. We conducted a user study with RouteZoom. The results showed that the application effectively reduces users’ map 17. Klippel, A., Hansen, S., Richter, K., and Winter, S. interaction costs and is well perceived by the user in terms Urban granularities—a data structure for cognitively of its quickness and easiness to use. ergonomic route directions. GeoInformatica 13, 2 (2008), 223-247. REFERENCES 18. Kopf, J., Agrawala, M., Bargeron, D. Salesin, D., and 1. Agrawala, M. and Stolte, C. Rendering effective route Cohen, M. Automatic generation of destination maps. maps. In Proc. SIGGRAPH 2001, ACM Press (2001), ACM TOG 29, 6 (2010), 158:1–158:12. 241-249. 19. MapQuest. http://www.mapquest.com/. 2. Appert, C., Chapuis, O., and Pietriga, E. High-precision 20. Martinez, A. Automated insetting: An expert component magnification lenses. In Proc. CHI 2010, ACM Press embedded in the census bureau’s map production (2010), 273–282. system. In Proc. AutoCarto 9, CaGIS (1989), 181-190. 3. Bing Maps Tile System. 21. NavEngine. http://developers.cloudmade.com/projects/ http://msdn.microsoft.com/en-us/library/bb259689.aspx. show/navengine. 4. Carpendale, M. S. T., Ligh, J. and Pattison, E. 22. Patel, K., Chen, M.Y., Smith, I., and Landay, J.A. Achieving higher magnification in context. In Proc. Personalizing routes. In Proc. UIST 2006, ACM Press UIST 2004, ACM Press (2004), 71–80. (2006), 187-190. 5. Chen, B., Neubert, B., Ofek, E., Deussen, O., and Cohen, 23. Pindat, C., Pietriga, E., Chapuis, O., and Puech, C. M.F. Integrated videos and maps for driving directions. JellyLens: content-aware adaptive lenses. In Proc. UIST In Proc. UIST 2009, ACM Press (2009), 223-232. 2012, ACM Press (2012), 261-270. 6. Cockburn, A., Karison, A., and Bederson, B.B. A 24. Robbins, D.C., Cutrell, E., Sarin, R., and Horvitz, E. review of overview+detail, zooming, and focus+context ZoneZoom. In Proc. AVI 2004, ACM Press (2004), 231- interfaces. ACM Computing Surveys 41, 1 (2008), 2:1- 234. 2:31. 25. Schmid, F. Knowledge-based wayfinding maps for 7. Furnas, G.W. Generalized fisheye views. ACM SIGCHI small display cartography. JLBS 2, 1 (2008), 57-83. Bulletin 17, 4 (1986), 16-23. 26. Tenbrink, T. and Winter, S. Variable Granularity in 8. Google Directions API. https://developers.google.com/ Route Directions. Spatial Cognition & Computation 9, 1 maps/documentation/directions/. (2009), 64-93. 9. Google Maps. http://maps.google.com/. 27. Timpf, S. Hierarchical Structures in Map Series, Ph.D 10. Grabler, F., Agrawala, M., Sumner, R.W., and Pauly, M. Thesis. Technical University of Vienna, Institute for Automatic generation of tourist maps. ACM TOG 27, 3 Geoinformation, 1998. (2008), 100:1–100:11. 28. Zhao, J., Wigdor, D., and Balakrishnan, R. TrailMap: 11. Hirtle, S.C. and Srinivas, S. Enriching spatial facilitating information seeking in a multi-scale digital knowledge through a multiattribute locational system. In map via implicit bookmarking. In Proc. CHI 2013, Proc. ICSC 2010, Springer-Verlag (2010), 279-288. ACM Press (2013), 3009-3018. 12. Hornbæk, K. Bederson, B. B. and Plaisant, C. Navigation patterns and usability of zoomable user