Presentation from ACM AVI 2012 in Capri, Italy on gravity navigation. Gravity navigation (GravNav) is a family of multi-scale navigation techniques that use a gravity-inspired model for assisting navigation in large visual 2D spaces based on the interest and
salience of visual objects in the space. GravNav is an instance of topology-aware navigation, which makes use of the structure of the visual space to aid navigation. We have performed a controlled study comparing GravNav to standard zoom and pan navigation, with and without variable-rate zoom control. Our results show a significant improvement for GravNav over standard navigation, particularly when coupled with variable-rate zoom. We also report findings on user behavior in multi-scale navigation.
1. GravNav: Using a Gravity Model
for Multi-Scale Navigation
Waqas Javed, Sohaib Ghani, Niklas Elmqvist
Purdue University
West Lafayette, IN
USA
Presented By: Waqas Javed
AVI 2012
May 21-24, 2012 ▪ Capri Island, Italy
1
7. Related Work
• General and Multi-scale Navigation
– Pan and zoom (Furnas and Bederson 2005)
– Speed-dependent automatic zooming (Igarashi and Hinckley 2000)
– OrthoZoom (Appert and Fekete 2006)
• Assisted Navigation
– Topology aware navigation (Moscovich et al. 2009, Ghani et al. 2011)
– Content aware scrolling (Ishak and Feiner 2006)
• Pointing
– Semantic pointing (Blanch et al. 2004)
– Sticky targets (Mandryk and Gutwin 2008)
– Force-enhanced targets (Ahlström et al. 2006)
7
13. Task
• Zoomed-out overview to a zoomed-in detail view
of a particular target
• The target was surrounded by distractor objects
• 2xN distractor objects at a random position on
the periphery of every Nth imaginary circle
• The target was a square the size of 10% of the
viewport size at 1:1 zoom factor
• The square was red whenever viewed at less than
full scale factor, and blue otherwise
• A collection of successively larger concentric rings
centered around the target
13
19. Experimental Design
• Participants: 12
• Navigation Technique T: 2
– SN (Standard Navigation)
– GN (Gravity Navigation)
• Zoom Control Technique Z: 2
– SZ (Standard Zoom Control)
– OZ (OrthoZoom)
• Index of Difficulty ID: 5
– 10, 15, 20, 25, 30
• Repetitions: 5
19
20. Experimental Design
• Trials were organized in blocks based on zoom
control technique Z
• Within each block T and ID factors were
randomized
• Measures during each trial
– completion time
– cinematic interaction data
• Minimum five training tasks
20
21. Experimental Hypothesis
• H1: GN will be faster than SN
• H2: OZ will be faster than SZ
• H3: GN will benefit more from OZ than SN
21
26. Subjective Feedback
• Participants were generally favorable in regards
to OrthoZoom
• Few participants felt that they often overshot the
target with OrthoZoom
– “the [OrthoZoom] task sometimes got out of control.”
• A couple of participants stated that some trials
were easier than others
– “sometimes it felt as if my cursor snapped onto the
target, whereas other times not so much.”
– Another thought that targets seemed to “pull in the
cursor.”
26
27. Summary of Results
• Gravity navigation exhibited significantly faster
completion times than standard navigation (confirming
H1)
• Variable-rate zoom control using OrthoZoom resulted
in significantly faster completion times than standard
constant rate zoom control (confirming H2)
• Gravity navigation with OrthoZoom was significantly
faster than gravity navigation with constant-rate zoom
control whereas no significant such difference was
found for standard navigation (confirming H3)
27
28. Conclusion
• A novel family of multi-scale navigation
techniques that we call GravNav
• GravNav utilizes the topology of the
underlying visual space to assist navigation
• Quantitative evaluation of the GravNav
technique
• Study results confirm the usefulness of the
new technique
28
29. Thank You!
Questions???
Contact Information:
Waqas Javed
School of Electrical & Computer Engineering
Purdue University
West Lafayette, IN, USA
E-mail: wjaved@purdue.edu
29 http://engineering.purdue.edu/pivot/
Notas del editor
In this presentation I will talk about GravNav, a multiscalenavigation technique that uses a gravity inspired model for assisting navigation in 2D visual spaces.This work is done in collaboration with SG, and NE at Purdue University.
Imagine zooming into a large multiscale map to a small town like lafayette, IN, from where I came.A small disparity in mouse position may have massive impact when quickly zooming into the space to see details. [click]In other words, you may end up in large corn fields surrounding the city instead of the downtown area.
To avoid such situations, we often have to adjust the viewport to reach our target[click] In this case downtown lafayette. However, the process of adjusting the viewport by the users results in more time consuming navigation than necessary. We believe that the Knowledge of the underlying visual space can be used to avoid this extra navigational cost. For example in this scenario, we know that it is often the case that the user is interested in the city than in the surrounding corn fields.
In fact Human attention has many similarities to the theory of gravity: regions of high visual interest attract attention just like moths are drawn to an open flame at night, and interest wanes with distance.[click] For example in this figure we tend to concentrate more on the irregular portions than plain portions.
The concept of visual salience is key in directing our visual attention. A recent work shows that map attention is centered on major cities and coastlines.Despite these facts, it is curious that so few navigation techniques for computer applications take advantage of this phenomenon.[click] Making use of the visual saliency or the topology of the underlying visual space could increase the efficiency of the navigation process.For example while panning and zooming into this image; ease with which viewport can be adjusted on the regions of interest is important.
In this work we present the concept of gravity navigation (GravNav) where an interest-based gravity model is used to ease navigation in large visual spaces. The basic idea is that objects in the visual space create “gravity wells” that attract the user’s viewport during navigation.
[click]This work draws upon existing work on general and multiscale navigation, assisted navigation, and pointing.[click] For general and multiscale navigation; starting with simple pan and zoom techniques more recent work includes techniques like Speed-dependent automatic zooming and orthoZoomscroller where a variable-rate zooming is used to scroll across large distances.[click] For assisted navigation recent work includes topology aware navigation techniques, where the structure of the space is used to assist navigationAnd content-aware scrolling, where scrolling in a document is modulated by the document content itself.Our gravity navigation technique also comes under this umbrella of assisted navigation techniques.[click] The pointing equivalent of gravity navigation is semantic pointing, where knowledge about the underlying user interface elements on the screen can be used to improve pointing performanceOther related work in this domain includes sticky targets that ease acquisition,and force-enhanced targets that attract the pointer
The intuition behind gravity navigation (GravNav) is to use a degree-of-interest (DOI) function for the local neighborhood of visual items around the current position of the viewport
[click] DOI takes the shape of a function [on slide] that returns the interest (a real value) for any position p on the visual space. [click] Most commonly, the DOI function is defined on a set S of discrete visual shapes s in the space as DOI(s) = 1.[click] For a structured visual space such a map, we can utilize our higher-level knowledge of the space and its salient features to define the DOI function. Consider the following example function for a map consisting of sets of cities C and roads R.[click] While our emphasis in this work is on assisting navigation in structured visual spaces, we may also want to apply gravity navigation to entirely unstructured visual spaces. In such situations, one can calculate the visual saliency of the graphical features in the space to find the regions that are perceptually important to the human visual system. [click] In this figure we use gradient magnitude, an edge detection algorithm to calculate saliency of a photograph.
Given a DOI function next step is to calculate a resulting attention gravity vector around a point pc. We used a formula similar to the Newton’s gravity field model and hence the name of the technique. [click] Figure shows a simple visual space consisting of five square shapes with connecting paths[click] Here we visualize the gravity field sampled at regular intervals in the visual space.Unlike the gravity wells caused by heavenly bodies, our attention gravity also has gravity troughs arising from paths in the visual space
[click] The last step of the process is to use the attention gravity vector computed from the DOI function to guide multi-scale navigation.our approach is to manipulate the control display (CD) mapping to achieve this.[click] For panning, we take the user’s input vector in motor space and add the attention gravity vector, suitably scaled, to create an actual navigation vector n’ that is used to transform the viewport’s position.[click] For zooming, we view the navigation vector n from motor space as a 3D vector pointing into (for zoom-in) or out of (for zoomout) the screen along the z axis.Again, we perform a vector addition with the attention gravity vector, resulting in an actual navigation vector n’. It is often advantageous, however, to not enable gravity effects for zoom-out operations.
We conducted a user study to measure the effectiveness of gravity navigation in comparison to traditional multi-scale navigation techniques.
[click] we designed a task where the objective was to navigate the viewport from an initial zoomed-out overview to a zoomed-in detail view of a particular target and then click on it.[click] The target was surrounded by distractor objects, the number of which was linearly dependent on the distance from the target.[click] In particular, we placed 2 N objects at a random position on the periphery of every N-th imaginary circle, centered at the target.[click] The target itself was a square the size of 10% of the viewport size at zoom factor 1;[click] the square was red in color whenever viewed at less than full scale factor, and blue when viewed at zoom factor 1 or closer. The target could only be clicked while in the blue stage.[click] Because the target was too small to see in the zoomed-out views, we added a collection of successively larger concentric rings centered around the target; thesewere placed and sized so that at least one was always visible at any zoom level.
This video show an example task for the experiment.Starting with a zoom out view user zoom-in to click on the target. Cocentric circles help in locating the target. While Distractor objects, similar to the target are placed on random locations around the target.
We included three factors in our experiment[click] Navigation Technique (T)[click] Zoom Control Technique (Z)[click] Index of Difficulty (ID)
We used two navigation technique[click] A standard pan and zoom implementation where a left mouse drag panned the view in world space proportional to the amount of pixelsdragged in screen space, and a right mouse drag zoomed the view in and out based on right and left movement respectively.[click] For gravity navigation we used the gravity zooming and panning operations. The pan and zoom interactions were similar to the standard navigation case, but our gravity model modulated both the speed and direction of all operations.
We suspected that zoom control may be a significant factor for navigation performance because of the tradeoff between high speed and error correction. Thus we included two levels of this factor[click] For the standard zoom control the zoom interactions caused a constant amount of zooming (zoom-constant = 0.01) for the same movement along the horizontal axis.[click] For variable-rate zoom control, we adopted the OrthoZoom technique from an earlier work.In our implementation, drag movement along the horizontal axis is scaled by the pointer’s vertical distance from the initial button press position while dragging horizontally and transformed into zoom operations.
The Index of Difficulty (ID) captures the distance between the initial position of the center point of the viewport and the target to navigate to.[click] Given an ID value, Using Fitts’ law, one can simply find the distance D to travel using this formula, where we treat W as constant target width.[click] A recent trend has been to study very large indices of difficulty that correspond to navigation in multi-scale spaces. Therefore we use five different values for this factor.
With 12 participants We used a full-factorial within-participants design
[click] We organized the trials in blocks based on zoom control technique so that participants would only have to utilize one type of zoom technique at a time.[click] Block order and the order of navigation technique and ID levels within each block were randomized across participants to counteract learning effects[click] During the experiment we collected the time it took participants to complete a task. We also collected full cinematic interaction data during each trial.[click] Participants first received general instructions about the experiment and the trials. They were then given a demonstration on how to solve a trial using the different zoom techniques. As is customary in similar experiments, we did not inform participants that the experiment incorporated a navigation assistance technique.
For the experiment we wanted to test three hypothesis[click] H1: GN will be faster than SN because gravity navigation corrects for the small pointing errors that grow to large navigation errors over long distances[click] H2: Based on previous results we predict OZ will be faster than SZ[click] H3: We also speculate that the combination of gravity navigation and OrthoZoom will be particularly effective since it will allow the user to make large-scale zoom movements without worrying about overshooting or losing the target.
Our results come in several forms that is quantitative completion times, dynamic navigation behavior, and subjective feedback.
We averaged the completion times for each condition and participant across repetitions and performed a repeated-measures analysis of variance.[click] Figure shows boxplots of completion times for different conditions[click] From our analysis we find a significant effect of navigation technique, zoom control technique, and index of difficulty on the completion time.
[click] Our analysis shows that Gravity navigation was an average of 25.6% faster than standard navigation across all indices of difficulty[click] Using a posthoc test (Tukey HSD) we found that the condition GN + OZ was significantly (p < :05) faster than all other combinations, followed by GN + SZ (p < :05), and then SN + OZ and SN + SZ having no significant time differences
[click] From the user study data we alsoobserved that the dynamic navigation behavior for gravity navigation was different from standard navigation. Figure shows a boxplot of the number of clutch operations (unique sequences of mouse press-drag-release) for the different conditions.The number was significantly lower for GravNav technique.[click] This figure shows aggregated performance for the different techniques in a space-scale diagram. While this figure does not incorporate timing, we found that gravity navigation seems to converge faster towards the target.
We did not inform participants about the gravity navigation technique and we randomly chose between SN and GN navigations in all blocks.Here I summarize the subjective feedback collected during the experiment.[click] Participants were generally favorable in regards to OrthoZoom (which we explicitly had to train participants in using and also used for blocking), saying that itgave them much more fine-grained control over zooming than for constant-rate zoom.[click] On the other hand, a few participants felt that they often overshot the target with OrthoZoom“the [OrthoZoom] task sometimes got out of control.”[click] A couple of participants stated that some trials were easier than others “sometimes it felt as if my cursor snapped onto the target, whereas other times not so much.”Another thought that targets seemed to “pull in the cursor.”We believe these trials were the one with GN technique.
To summarize the finding of our user study[click] Gravity navigation exhibited significantly lower completion times than standard navigation (confirming H1)[click] Variable-rate zoom control using OrthoZoom [2] resulted in significantly faster completion times than standard constant rate zoom control (confirming H2)[click] Gravity navigation with OrthoZoom was significantly faster than gravity navigation with constant-rate zoom control whereas no significant such difference was found for standard navigation; in addition, gravity navigation with constant-rate zoom control was faster than standard navigation with OrthoZoom (confirming H3)
To conclude my talk[click] In this work we presented A novel family of multi-scale navigation techniques that we call GravNav.[click] The technique utilizes the topology of the underlying visual space to assist navigation[click] We also presents findings of a controlled quantitative user study[click] Results from the user study confirm the usefulness of the new technique