An Intelligent 3D User Interface Adapting to User Control Behaviors

Transcription

1 An Intelligent 3D User Interface Aapting to User Control Behaviors Tsai-Yen Li Computer Science Department National Chengchi University Taipei, Taiwan 11623, R.O.C li@nccu.eu.tw Shu-Wei Hsu Computer Science Department National Chengchi University Taipei, Taiwan 11623, R.O.C g9204@cs.nccu.eu.tw ABSTRACT The WALK moe is one of the most common navigation interfaces for 3D virtual environments. However, ue to the limite view angle an low frame rate, users are often blocke by obstacles when they navigate in a cluttere virtual scene with such a moe. Intelligent 3D navigation interfaces with assisting mechanisms, such as motion planning methos or virtual force fiels, have been propose in the literature to improve navigation efficiency. Nevertheless, the applicability of these methos is subject to iniviual iscrepancy, an the control parameters of these methos are usually etermine by empirical means. In this paper, we propose an intelligent navigation interface with a personalizable assisting mechanism. We have esigne two methos, simulation experiment an ynamic ajustment, to fin the best control parameters for composing artificial forces for an iniviual in an off-line an on-line manner, respectively. The simulation experiment metho searches for the optimal control parameters for a user in a systematic manner while the ynamic ajustment metho makes the assisting mechanism aaptive to user control behaviors as well as environmental variations in real time. Our experiments show that both methos can further improve the navigation efficiency for a wier range of users. Categories an Subject Descriptors H.5.2 [Information Interfaces an Presentation]: User Interfaces Graphical User Interface, Input Devices an Strategies. I.3.6 [Computer Graphics]: Methoology an Techniques - Interaction Techniques. General Terms Algorithms, Design, Experimentation, Human Factors. Keywors Intelligent 3D interface, Artificial Force Fiel, Personalize User Interface Control, Aaptive Assisting Mechanism. Permission to make igital or har copies of all or part of this work for personal or classroom use is grante without fee provie that copies are not mae or istribute for profit or commercial avantage an that copies bear this notice an the full citation on the first page. To copy otherwise, or republish, to post on servers or to reistribute to lists, requires prior specific permission an/or a fee. IUI 04, January 13 16, 2004, Maeira, Funchal, Portugal. Copyright 2004 ACM /04/0001 $ INTRODUCTION Virtual Reality (VR) techniques aopt interactive 3D user interfaces to allow a user to have realistic experiences in virtual scenes. Such 3D user interfaces have a wie range of applications in computers, inustries, meicine, military, an eucation. For example, one can fin VR applications in Computer-Aie Design (CAD), flight simulation, virtual museums, an molecular structure analysis, an so on. Due to the constraints of 2D input an isplay evices, the esign of a goo user interface on a typical personal computer become a crucial an challenging problem. Toay 3D contents are becoming prevalent on the web via 3D browsers such as VRML browsers. The common navigation moes on such a browser inclue WALK, PAN, TURN, FLY, ROLL, an so on. The WALK moe is usually the efault moe that allows a user to walkthrough a 3D scene. However, ue to the limite fiel of view that can be isplaye on a screen, a viewpoint might easily get stuck at a corner of a scene since the boy of the avatar collies with the environmental obstacles. It usually takes several maneuvers for a user to escape from this type of ifficult region. In orer to overcome this problem, much previous work has propose assisting mechanisms, such as motion planning algorithms[8][11] an virtual forces[10], to improve navigation control. However, experiments show that user iscrepancy on the familiarity of controlling 3D interfaces greatly influences the effects of such assisting mechanisms. Novice users might nee more guiance from the system while expert users might want to have full control of the interface. Therefore, it is more esirable to have a system that can aapt to an iniviual user s control behavior an ajust the egrees of assistance accoringly. In this paper, we propose two mechanisms to make the intelligent 3D user interface with virtual forces aaptive to users control behaviors in orer to further improve the navigation performance. The rest of the paper is organize as follows. In the next section, we will review the work pertaining to intelligent 3D user interfaces. We will briefly review the assisting mechanism of virtual force fiel to ientify the control parameters that can be tune to fit users control behaviors. Next we will propose a mechanism to fin the optimal control parameters by simulation experiments. We will then present another mechanism to ajust these control parameters in an on-line manner. Experimental results will be presente an iscusse in the following section. Finally, we will conclue the paper at the en.

2 A C A D B Figure 1. Example of how to compute the virtual force for a viewpoint A accoring to the bounary cells of nearby obstacles. 2. RELATED WORK 3D user interface esign has attracte much attention in the virtual reality research community. Various kins of evices have been invente to facilitate the control of 3D scenes. For example, Hea Mounte Device (HUD), ata gloves, 3D tracking, an haptic evices are all examples of harware interfaces for 3D control. However, for most esktop computers, 2D mice remain the most accessible control evices [6][13]. However, it is a challenge to use a 2D mouse to control a 3D scene for a novice user since the controllable egrees of freeom (DOF) are limite by the DOF (2) of a mouse at a time. In aition, the mapping between them may not be very intuitive. Therefore, several researches have mae proposals to improve the 3D control interface, especially for 3D rotation [2][4][7][14]. Although many intelligent user interfaces have been propose in the literature, most of them are not for 3D manipulation [12]. Exceptions inclue using motion-planning techniques [8] to provie task-level controls. For example, Drucker an Zeltzer[3] argue that a task-level viewpoint control is crucial for exploring virtual scenes such as virtual museums since the users shoul be allowe to concentrate on scene viewing instea of being istracte by low-level navigation control. In aition, Artificial Intelligence (AI) techniques have been use to improve the efficiency of 3D navigation, especially for the WALK moe, at a control level. These techniques inclue motion planning algorithms [11] an virtual force fiel methos [10][5][16][17]. In the motion planning approach, efficient path planners are evoke to generate a collision-free path to take the viewpoint past the ifficult region where the user got stuck. In the virtual force approach, a repulsive force from obstacle bounary is compute to influence the viewpoint motion such that it is less likely to collie with the environmental obstacles. In other wors, the force fiel approach aims to prevent collisions while the motion planning approach can generate a feasible path for the viewpoint to the esire goal when collisions o happen. Since user iscrepancy is more significant for 3D interfaces, personalization of 3D interfaces is an important problem although not much work has been one in this irection. In [9], a reinforcement learning metho has been employe to realize user interface customization for manipulating 3D objects in a virtual environment. 3. VIRTUAL FORCE FIELD In this section, we will first review how virtual force is incorporate into the control loop of user navigation an then ientify the control parameters for composing the virtual force. 3.1 How Virtual Force Works In orer to compute the virtual force fiel, a iscrete representation of the virtual worl is commonly use. The objects in the worl are usually projecte into obstacle regions in a 2D gri. The bounary cells of obstacles generate repulsive forces (inversely proportional to the istance from the viewpoint) to influence the nearby cells. Each free cell in the gri will be given a virtual force vector compose from the forces generate by the bounary of all nearby obstacles aroun the viewpoint. For example, in Figure 1, the viewpoint A with an influence region of 7x7 will be affecte by cells v0, v1, an v2. If the virtual worl layout is static, the virtual force fiel can be compute in a preprocessing step an store in a 2D array for run-time lookups. In the regular control loop of a navigation interface, a user s mouse input (ragge vector) is translate into an offset vector for upating the next configuration of the viewpoint ( AB as in the example of Figure 2). With the influence of the virtual force fiel, a force vector is compute for the current configuration A an transforme into a ifference vector ( AC ). Then this vector is use to compose the final upating vector ( AD ) for the next viewpoint configuration. 3.2 Force Fiel Ajustment How the static force fiel is applie to affect a user s navigation is etermine by the interaction with the user. In [10], the virtual force is compute ynamically accoring to the following formula. Figure 2. Incorporating virtual force into a control loop. r r F p = µ user µ v µ F. (1) p F r r is the static force precompute for a give point p while p F p is the force ajuste accoring to user interaction an control history. The ajustment parameters, µ user, µ, an v µ, are efine as follows. µ = v kv, where µ is the velocity parameter, v is the v magnitue of the current velocity, an k is a constant. n n 1 n µ = η µ, where µ is the irectional parameter for step n, an η is etermine by the ot prouct of F r p an the forwar

3 Real User Source of Simulation Inputs Intention Analysis an Correction 3D Interface Simulation Result Virtual Force Assisting Moule Parameter Configuration Sample Input Commans 3D Display Parameter Configuration Generator Figure 3. Overall system architecture. facing vector. When the ot prouct is greater than zero, η = m, where m is a magnifying factor greater than 1. When the ot prouct is less than or equal to zero, η = n, where n is a shrinking factor less than 1. µ user = l is a user-preference factor specifie by the user. The greater value of µ means stronger force user influence. In [10], (l, k, m, n) are four fixe control parameters, calle parameter configuration, that are etermine by empirical means. However, it was reporte that the improvement of navigation efficiency with such assisting mechanisms is greatly affecte by scene variations an user characteristics. There is no single parameter configuration that is the best for everyone. Even if we want to fin a goo configuration for most people, teious experiments nee to be one. Therefore, it is highly esirable to have an automatic mechanism to fin the best parameter configuration for applying virtual forces to a given user. 4. DESIGNING ADJUSTABLE VIRTUAL FORCES Accoring to the experience of the above experiments, we know that the virtual force nees to be aaptive to a user in orer to achieve the best improvement. Hence, we have esigne two assisting mechanisms, simulation experiments an ynamic ajustment, which can be use to fin the best parameter configuration in an off-line an on-line manner, respectively. The overall system architecture propose in this paper is shown in Figure 3. The inputs to the system coul be from a real user or from simulate ata. The intention analysis an correction moule is use to moify the simulate inputs so that the user s high level intention is kept. The parameter configuration generator is use to come up with a goo configuration accoring to the simulation result. Then the virtual force assisting moule uses this parameter configuration to compute the virtual force for upating the next viewpoint location. Figure 4. Graphical interface of the experiments. The upper part is the 3D isplay moule while the mile portion is the 2D layout map. The lower portion shows the trening of the virtual force. The simulation experiment metho aims to search for the best parameter configuration for a user in a specific worl in an offline manner. We first ask the user to navigate through the worl by accomplish a given task an recor how the user controls the interface interactively uner the influence of virtual force. We then try to simulate the user s input in an off-line manner by feeing the sample input commans to the interface. In each simulation experiment, we use a fixe parameter configuration for the whole course an recor the number of steps taken to reach the goal location. The optimal parameter configuration can be foun after all possible configurations have been exhaustively teste. The ynamic ajustment metho is an on-line version of the simulation experiment. The main ifference is that the scale is smaller in orer to make it feasible for on-line applications. In this mechanism, the simulation experiment is one only for a segment of a few steps in the past control inputs. A greey algorithm is use, an only neighboring parameter configurations are evaluate to fin a better configuration for the next step. Although we only move the parameter configuration one step at a time, the configuration will converge to the best value graually after a few steps if there exists a unique optimal one. We will escribe the above two methos in more etails in the next two sections an compare the effectiveness of the two methos in Section SIMULATION EXPERIMENT METHOD 5.1 Recoring Navigation Process From [10] we know that the virtual force fiel mechanism can improve navigation efficient by reucing the number of steps taken to reach the goal from a given initial location. However, in orer to unerstan how the virtual force helps in various occasions, we have to recor the input commans uring navigation an analyze how the virtual forces are compute for these inputs. In orer to recor the user inputs, we have implemente an inter-

4 LanMark InputData A B C Sample Path Directional Curve B P R A D C Q Figure 5. Using milestones to moify the short-term goal for simulation experiments. When the viewpoint moves from A to D, the simulation might skip milestone B ue to a closer milestone C. face, as shown in Figure 4, to collect a user s inputs (ragge vectors) as well as the navigation trajectory uner the influence of virtual forces. With this raw ata, the user s navigation process can be simulate an reprouce. 5.2 Simulation Experiment The effectiveness of the virtual force mechanism is etermine by the four-parameter configuration (l, k, m, n), an it iffers greatly among users. We aim to fin a personalize configuration proviing an optimal navigation performance for a given user. However, we o not want the user to repeat the navigation experiment in person on the other han. In fact, a user s interface control behavior might change when a user gets familiar with a specific scene. Therefore, we propose to use simulation to repeat the experiments to search for the optimal configuration. By optimal configuration, we mean the parameter configuration that can prouce the virtual forces to assist the user to take the least number of steps to reach the goal location. 5.3 Intention Analysis an Input Correction Since we have recore the original input commans (ragge vectors) on the interactive interface, we can play back the navigation if all parameters remain the same. However, once we start to moify the parameter configuration, we are altering the virtual force applie to the interaction an the resulting path might not be the same as the original one. In a real experiment, a user can ajust the mouse input with visual feeback when the influence of virtual forces changes. However, in a simulation, if we keep the original inputs, the simulation might not generate prouctive or even useful movements. Instea, what we have to keep is the high-level intention behin the user s input. Therefore, we propose two methos, calle milestone an irectional curve, to ajust the input comman ynamically accoring to the current viewpoint location. The first metho creates milestones along the path by sampling points that the user passes through in the real navigation experiment. The milestones ivie the input commans into segments. We assume that these milestones are the mi-term goals in each segment liste in sequence that reflect the user s high-level intention. During the simulation, if the closest milestone is change to one in a future segment, then all input commans in this segment are skippe. For example, in Figure 5, when the input commans are still in the segment aroun milestone A an the viewpoint has Figure 6. Using irectional curve to moify orientation for simulation experiments. been move to location D, which is closer to milestone C than to milestone B. In this case, all input commans behin milestone C will be skippe an next comman will start in the segment containing milestone C. The secon metho, calle irectional curve, is esigne to fix the orientation problem of the input commans. When the viewpoint location eviates from the original path ue to ifferent virtual force influence, the input commans also nee to be ajuste to make the viewpoint move towar the right irection. When a user navigates, he/she controls the interface accoring to the scene in front of his/her viewpoint. One can imagine that there exists a short-term goal point in front of each viewpoint location. We call the list of virtual points along the path the irectional curve. We can use the points along this curve to ajust the moving irection of the viewpoint. For example, in Figure 6, we assume that the current viewpoint location is at P while the closest point along the sample path is Q. Assume that the corresponing point on the irectional curve for Q is point R. Then the vector PR becomes the ajuste irection for the viewpoint. In other wors, we are using an imaginary point in front of the viewpoint along the navigation path to guie the viewpoint movement in the simulation experiments. The low-level input commans are almost meaningless an cannot be use in simulation if it lacks the intention behin them. The two methos above are esigne to maintain the high-level intention of a human user in the simulation by analyzing the navigation path trace by the user. As we will show in later sections, the simulation experiments can run by itself with these two methos even if the virtual forces vary significantly. 6. DYNAMIC ADJUSTMENT METHOD Although the simulation experiments can fin the optimal parameter configuration automatically, it can only be one in an offline manner ue to the time-consuming simulation process. For example, if we ivie the range of each parameter into K intervals, the total number of simulation experiments nee to be one is K 4 before we can fin the optimal one in a brute-force search. In aition, the optimal configuration may only be goo for a specific scene even though it is likely to be a goo starting point for the user. Consequently, we investigate the possibility of bringing

5 Figure 7. Simple experimental scene for testing the effectiveness of intention analysis. the experiments on-line with a smaller scale. This metho is calle the ynamic ajustment metho. 6.1 Winow sampling for on-line search In the ynamic ajustment metho, we use a sliing winow of size N an maintain the input commans for the latest N steps. For example, assume that the user has taken K steps, an the system will recor the comman inputs for the N steps between MAX(K- N+1, 0) an K. The system will use the last N steps to run the simulation experiment an try to obtain a better configuration, if exists, among the neighbors for the next step. In other wors, we will simulate the last N steps with ifferent sets of parameters to see which one takes the least number of steps to reach a goal if it were use. For example, assume that the parameter configuration is enote by (A 1, A 2, A 3, A 4 ). Then the neighboring configurations will be (A 1-1, A 2, A 3, A 4 ), (A 1 + 1, A 2, A 3, A 4 ), (A 1, A 2-2, A 3, A 4 ),,(A 1, A 2, A 3, A ). The best neighbor will be selecte as the configuration for the next step. The configuration will converge to a local optimal one (if not global) graually if the user keeps his/her navigation habits. Since the parameters are not specifically esigne for a scene, the system shoul be able to ajust the configuration to a better one for any scenes in an online manner. In aition, when a user gets familiar with a scene or the navigation control, the egree of assistance that the user prefers might also change over time. 6.2 Aing the momentum mechanism Although the ynamic ajustment mechanism escribe above is able to ajust the parameters towar a better one, the user may feel inconsistent if the parameters are change too often simply because of scene variation. If we ajust the parameters accoring to the frame rate, which is typically above ten frames per secon, the fluctuation on the forces may be too frequent for a user to control the interface in a consistent way. Therefore, we have ae an inertia mechanism into the system such that unless enough momentum has been accumulate for a specific change, the on-line system will not ajust the configuration immeiately. We keep a counter for each neighbor of the current parameter configuration an increment the counter when we attempt to visit it every time. A change is actually mae only if the counter has reache a preefine threshol value, say 3. In other wors, unless the changes in that irection have accumulate enough momentum, we o not make the move on the parameter changes. Figure 8. The 3D scene use in the navigation experiments. The circle on 2D map is the goal location. With such a mechanism, we can maintain a more controllable interface by avoiing frequent fluctuation on the configuration as well as the resulting virtual forces. 7. EXPERIMENTS AND DISCUSSIONS We have implemente the aforementione virtual force ajustment mechanism in Java an integrate it into the 3D VRML [15] browser implemente by Blaxxun [1]. 7.1 Effects of Using Intention Analysis A key function in the simulation experiments is the intention analysis mechanism. In orer to verify the effect of the intention analysis methos, we have one some experiments on the scene shown in Figure 7. The scene is kept simple intentionally in orer to screen out other factors. In three simulation experiments with ifferent parameter configurations, the viewpoint all gets stuck at the corner when no intention analysis an ajustment are use. On the other han, if the intention analysis mechanism is use, the viewpoint can successfully reach the goal at the other en of the U-shape hallway in all experiments. This shows that the highlevel intention analysis is crucial for making the simulation feasible. 7.2 Comparisons of Simulation Experiments an Dynamic Ajustment Methos Ten subjects are invite to o the experiment of comparing navigation efficiencies uner ifferent assisting mechanisms. Two of them (users A an B) are familiar with the control interface of VRML browsers while another two subjects (C an D) are familiar with 3D game controls. Three of the subjects (E, F, an G) have a goo comman of computers but not 3D interfaces. Another two subjects (H an J) only use computers occasionally (once a month) an the last one (I) never uses computers. We use the scene shown in Figure 8 for each subject to o the experiments with the assisting mechanisms of simulation experiments an ynamic ajustment. In the simulation experiments, we first recor the user s input commans without any assisting mechanisms. Then we ask the

6 Table 1. The number of steps taken to reach the goal location with four ifferent mechanisms No Assisting Mechanisms Using Default Parameter Configuration Using Simulation Experiments Using Dynamic Ajustment User Step % step % step % step % A B C D E F G H I J Avg user to o two runs for each of the following two parameter configurations. One is the efault parameter configuration, an the other one is the optimal configuration foun for the user in the simulation experiments. We recor the numbers of steps taken to reach the goal in each run an compute their average for each configuration. The experimental results are summarize in first three columns in Table 1. The number of steps of the runs without any assistance is use as a base (100%) to compare with other mechanisms. The secon column shows the numbers of steps with the efault configuration chosen carefully while the thir column shows the performance with the optimal configuration. From these ata, we have verifie the effectiveness of using virtual force fiel as the assisting mechanism because the average number of steps is reuce to 45.4% with the efault set of parameters. We also foun that the performance can be further improve by 8% if we use the optimal configuration foun in simulation experiments. This effect seems to be more significant for novice users such as subject I, whose number of steps is reuce by aroun 15% if the optimal configuration is use. In the experiment of ynamic ajustment, we choose three initial configurations for each subject to run. For each of the three initial configurations, we ask the subject to repeat the navigation experiments three times an take the average of the night experiments at the en. The result is shown in the fourth column in Table 1. We o not expect the performance of the ynamic ajustment mechanism to be better than simulation experiments since its avantages are on the on-line characteristic an flexibility to aapt to the worl. We foun that the numbers of steps for four users (A, C, D, an H) are almost the same as the ones using the optimal configuration. One of them (Subject F) even has better performance with the ynamic ajustment mechanism. In orer to compare user iscrepancy on the optimal configuration, we have chosen two subjects (A an H) to o the comparisons. Each subject is aske to run the experiments with the optimal Figure 9. Comparing the effects of ifferent parameter configurations on ifferent users configuration for the other subject. For example, user A is aske to run the optimal configuration for user H an vice versa. The result is shown in Figure 9. We foun that the sensitivity of each user to the virtual force configuration might vary significantly. For novice users such as user H, using other people optimal configuration might be even worse than using the efault configuration. Therefore, we can conclue that user iscrepancy oes exist in aopting the assisting mechanisms an the aaptive interface can greatly benefit novice users in proviing just-in-time navigation assistance. 8. CONCLUSION AND FUTURE WORK In our previous work, we have shown that the assisting mechanism with virtual force can greatly improve navigation efficiency by reucing unnecessary collisions with environmental obstacles. However, the iscrepancy between users in aopting the intelligent interfaces has inspire the work reporte in this paper. We have esigne two aaptive mechanisms to fin a set of goo parameters automatically for a specific user. Our experiments show that the simulation experiments can fin the best virtual force parameters but the simulation can only be one in an offline manner ue to the time-consuming search process. The online ynamic ajustment metho, on the other han, oes not require any prior analysis an coul be as effective as using a fixe optimal configuration foun in simulation experiments. The aaptiveness of such a mechanism to scene variations an personal ifferences also make it attractable in more 3D applications. In orer to simplify the system experiments, we only stuy the effectiveness of virtual forces in the current system. However, we know that the path planning metho is also effective in guiing a user when a collision oes happen. Therefore, in the future we will try to incorporate the automatically generate escape path into the virtual force fiel to make the navigation even more efficient. In aition to improving the navigation efficiency passively by proviing virtual forces for a given scene, we are also working in the irection of reesigning a virtual scene with a better layout or more informative irection signs at appropriate places. We believe that a goo assisting mechanism on a well esigne scene

7 will provie an efficient an pleasant navigation for general way fining. 9. ACKNOWLEDGEMENT This work was partially supporte by grants from National Science Council uner contracts NSC E an NSC C E. 10. REFERENCES [1] Blaxxun3D. An embeing java applet VRML browser. Available at [2] Chen, Mountfor, an Sellen. A Stuy in Interactive 3D Rotation Using 2D Control Devices, Computer Graphics, 22(4): [3] Drucker, S. M. an Zeltzer, D., Intelligent Camera Control in a Virtual Environment, Graphics Interface 94, [4] Ewars, J. an Han, C. Maps: Movement an planning support for navigation in an immersive vrml browser. In Proceeings of VRML'97 Conference [5] Egbert, P. K., an Winkler, S. H. Collision-Free Object Movement Using Vector Fiels. IEEE Computer Graphics an Applications, 16(4):18-24, July, [6] Hong, L., Muraki, S., Kaufman, A., Bartz, D. an He, T. Virtual Voyage: Interactive Navigation in the Human Colon. In Proceeings of SIGGRAPH 97 Conference, [7] Jung. M. R., Paik, D., an Kim, D. A Camera Control Interface Base on the Visualization of Subspaces of the 6D Motion Space of the Camera. In Proceeings of IEEE Pacific Graphics [8] Latombe, J.-C., Robot Motion Planning, Kluwer Acaemic Publisher, Boston, MA, [9] Lenzmann, B., Wachsmuth, I. A User Aaptive Interface Agency for Interaction with a Virtual Environment. In Working Notes of the IJCAI-95 Workshop on Aaptation an Learning in Multiagent Systems, AAAI Press, [10] Li, T.Y., Chou, H.C. Improving Navigation Efficiency with Artificial Force Fiel. In Proceeings of th IPPR Conference on Computer Vision, Graphics, an Image Processing, Taiwan, [11] Li, T.Y., an Ting, H.K. An Intelligent User Interface with Motion Planning for 3D Navigation. In Proceeings of the IEEE Virtual Reality 2000 Conference, New Jersey, [12] Maybury, M. an Wahster, W. (es), Reaings in Intelligent User Interfaces, Morgan Kaufmann: Menlo Park, CA. [13] Neilson, an Olsen. Direct Manipulation Techniques for3d Objects Using 2D Locator Devices. In Proceeings of the 1986 Workshop on Interactive 3D Graphics, [14] Tan, Desney S., Robertson, George G., an Czerwinski, M. Exploring 3D Navigation: Combining Spee-couple Flying with Orbiting. In Proceeings of the SIGCHI Conference on Human Factors in Computing Systems, March [15] VRML97 International Stanar, URL: ions.htm [16] Wernert, Eric A. an Hanson, A. J. A Framework for Assiste Exploration with Collaboration. In Proceeings of IEEE Visualization'99. San Francisco, October [17] Xiao, D. an Hubbol, R. Navigation Guie by Artificial Force Fiels. In Proceeings of the ACM CHI 98 Conference, , 1998.