Semantic Depth of Field
`
`Silvia Miksch
`Robert Kosara
`Vienna University of Technology
`http://www.asgaard.tuwien.ac.at/
`(cid:0)rkosara,silvia(cid:2)@asgaard.tuwien.ac.at
`
`Helwig Hauser
`VRVis Research Center, Austria
`http://www.VRVis.at/vis/
`hauser@VRVis.at
`
`Abstract
`
`We present a new technique called Semantic Depth of
`Field (SDOF) as an alternative approach to focus-and-
`context displays of information. We utilize a well-known
`method from photography and cinematography (depth-of-
`field effect) for information visualization, which is to blur
`different parts of the depicted scene in dependence of their
`relevance. Independent of their spatial locations, objects of
`interest are depicted sharply in SDOF, whereas the context
`of the visualization is blurred. In this paper, we present a
`flexible model of SDOF which can be easily adopted to var-
`ious types of applications. We discuss pros and cons of the
`new technique, give examples of application, and describe
`a fast prototype implementation of SDOF.
`
`Keywords: Depth of Field, Focus and Context, Information
`Visualization
`
`1. Introduction
`
`Whenever large amounts of data are to be investigated, vi-
`sualization potentially becomes a useful solution to provide
`insight into user data. Especially for exploration and anal-
`ysis of very large data-sets, visualization not only needs to
`provide an easy-to-read visual metaphor, but also should en-
`able the user to efficiently navigate the display, allowing for
`flexible investigation of arbitrary details.
`Focus and Context (F+C) techniques enable the user to
`investigate specific details of the data while at the same
`time also providing an overview over the embedding of
`the data under investigation within the entire dataset. But
`F+C encompasses a number of very different techniques
`that achieve similar goals in very different ways.
`
`1.1. Different Kinds of Focus and Context
`
`The most prominent group of F+C methods are distortion-
`oriented [12] or spatial methods. The geometry of the dis-
`play is distorted to allow a magnification of interesting in-
`
`It
`formation without losing the (less magnified) context.
`is thus possible to navigate information spaces that are far
`too large to be displayed on a screen. Examples are fish-
`eye views [5, 20], hyperbolic trees [9, 10, 18], stretchable
`rubber sheets [21], etc. Distortion-oriented techniques are
`usually used in an explicit way, by actively bringing the in-
`teresting objects into focus, e.g. by clicking on objects or
`dragging them around.
`For smaller numbers of objects that have a lot of data
`associated with them, a visualization method is useful that
`shows just a limited number of data dimensions, and allows
`the user to select which of the objects are to be shown in
`more detail – we call these dimensional methods. The con-
`text in this case are not only the other objects, but also the
`remaining data dimensions. This type of method also shows
`more detail, but in terms of data dimensions, not screen size.
`Examples are magic lenses [22] and tool glasses [2], where
`the user moves a window over the display, the objects inside
`which are displayed in more detail.
`The third type of focus and context allows the user to
`select objects in terms of their features, not their spatial re-
`lations; usually by assigning a certain visual cue to them –
`we therefore call these methods cue methods. They make
`it possible to query the data for information which is not
`immediately visible in the initial visualization, while keep-
`ing the original layout, and thus not destroying the user’s
`mental map [17]. An example for such a system is a Ge-
`ographic Information System (GIS) that makes it possible
`to display crime data, certain cities, or hospitals [14]. This
`data is displayed in the same context as before, but the rel-
`evant parts of the display have a higher color saturation and
`opacity than the rest. This leads the viewer’s attention to
`the relevant objects easily without removing context infor-
`mation.
`In contrast to distortion-oriented techniques and
`magic lenses, with this type of method, the user first selects
`the criteria, and then is shown all the objects fulfilling them.
`The technique presented in this paper is of the third type,
`but we use a different visual cue for discriminating focus
`and context.
`
`Authorized licensed use limited to: Fish & Richardson PC. Downloaded on June 13,2024 at 01:52:17 UTC from IEEE Xplore. Restrictions apply.
`Proceedings of the IEEE Symposium on Information Visualization 2001 (INFOVIS’01)
`1522-4048/01 $17.00 © 2001 IEEE
`
`1
`
`APPLE 1008
`
`

`

`Depth of Field (SDOF) for information visualization, which
`renders objects sharp of blurred, depending on their current
`relevance. It thus makes use of the phenomena described
`above to guide the viewer’s attention.
`
`2 Related Work
`
`There have been surprisingly few attempts to use DOF or
`blur in visualization at all; the ones relevant to this work are
`shortly summarized here.
`In a system for the display of time-dependent cardio-
`vascular data [25], a stereoscopic 3D display is included
`that is controlled by the viewer’s eyes. Like a microscope,
`only one thin slice through the data appears sharp, all oth-
`ers are blurred and therefore almost invisible. Eye track-
`ing equipment determines what the user is looking at, and
`that point is brought into focus. This makes it possible to
`concentrate on one detail without the surrounding structures
`confusing the viewer. Later work [26] describes “non-linear
`depth cues”, which means displaying structures that cur-
`rently are of interest (like single organs) in focus, and other
`objects out of focus, not based on their distance from the
`camera, but on their importance.
`The Macroscope [13] is a system for displaying several
`zoom levels of information in the same display space. For
`this purpose, the images on all levels are drawn over each
`other, with the more detailed ones drawn “in front”, i.e.,
`drawn over the less magnified layers. The layers’ trans-
`parency can be changed so that the background (context)
`can be more or less visible. The less detailed layers are
`blurred so as to not distract the viewer, but serve as context.
`The most interesting existing approach for this work is
`a display of geographical information [3]. In this system,
`up to 26 layers of information can be displayed at the same
`time. Each layer has an interest level associated with it that
`the user can change. The interest level is a combination of
`blur and transparency, making less interesting layers more
`blurred and more transparent at the same time. This work
`does not seem to have been followed up on recently.
`In this paper, we describe a general model of SDOF,
`i.e., of selectively using sharpness vs. blur to empha-
`size/deemphasize certain parts of the data. We clearly em-
`bed SDOF within the scope of information visualization and
`computer graphics. In addition to the above examples, we
`provide a flexible solution which easily is adopted to vari-
`ous kinds of applications, as demonstrated later on.
`
`3. Semantic Depth of Field (SDOF)
`
`SDOF allows the user to select relevant parts of a visual-
`ization that are then pointed out by deemphasizing all the
`rest through blur. The discrimination between relevant and
`
`Figure 1. A lantern with a bridge as context.
`
`1.2. The Uses of Blur and Depth of Field
`
`Blur is usually considered to be an imperfection: it makes
`features harder to recognize and can hide information com-
`pletely. But the difference between sharp and blurred parts
`of an image is a very effective means of guiding the viewer’s
`attention. In photography, the depth-of-field (DOF) effect
`leads to some parts of the image being depicted sharply,
`while others are blurred [1]. The viewer automatically
`looks at the sharp parts, while the blurred parts provide non-
`disturbing context for the objects of interest (see Fig. 1 for
`an example). The same effect is also used in cinematog-
`raphy [8], where focus changes can guide the audience’s
`attention from one character to another, from a character to
`an object he or she just noticed, etc.
`Because the human eye (like every lens system) also has
`limited DOF [6], an important characteristic of human vi-
`sion is that whenever we get interested in a specific part of
`our environment, we 1) bring the the object of interest into
`the center of our eye (where the area of most acute vision,
`the fovea centralis, is located), and 2) focus on that object.
`From the above applications of DOF (photography and cin-
`ematography), we know that this process is easily inverted:
`If we display sharp objects in a blurred context, the viewer’s
`attention is automatically guided to the sharp objects. This
`also gives us reason to believe that DOF is perceived preat-
`tentively, i.e. within 50ms of exposure to the stimulus, and
`without serial search [23]. This means, it very efficiently
`makes use of the bandwidth of the human visual system to
`convey a lot of information in very little time.
`We have developed an F+C technique we call Semantic
`
`Authorized licensed use limited to: Fish & Richardson PC. Downloaded on June 13,2024 at 01:52:17 UTC from IEEE Xplore. Restrictions apply.
`Proceedings of the IEEE Symposium on Information Visualization 2001 (INFOVIS’01)
`1522-4048/01 $17.00 © 2001 IEEE
`
`2
`
`

`

`Data
`
`Spatial
`Arrangement
`2D
`3D
`
`Relevance
`and Blurring
`Selection
`Distance
`...
`
`Viewing and
`Camera Model
`Photorealistic
`Adaptive
`...
`
`b
`
`1
`
`1
`
`g
`
`t
`
`h
`
`maxb
`
`r
`
`0
`
`Figure 2. SDOF Building Blocks.
`
`Figure 3. The Blur function.
`
`irrelevant objects can be binary (an object is either relevant
`or irrelevant) or continuous (an object can have a relevance
`value between the two extremes).
`Different relevance metrics for objects have to be of-
`fered by the application, that have to deal with the specific
`information and tasks the application is made for. Exam-
`ples for binary relevance measures are the set of chessmen
`that threaten a specific piece in a chess tutoring system (see
`Fig. 5c and the accompanying video), the layer containing
`roads in a GIS application (Fig. 5d), or all incidents related
`to high blood glucose in a graphical patient record. Contin-
`uous functions could express the age of files in a file system
`viewer (Fig. 5a), the recent performance of stocks in a stock
`market browser, or the distance of cities from a specified
`city in terms of flight hours.
`The building blocks of SDOF are discussed in the fol-
`lowing subsections, and are summarized in Fig. 2 and
`Tab. 1.
`
`3.1. Spatial Arrangement
`
`In information visualization, usually some kind of layout al-
`gorithm is used to arrange objects in the visualization space
`(typically 2D or 3D). The special challenge of information
`visualization is the fact that data often does not have any
`inherent structure that naturally translates to a layout. Map-
`ping functions are a very important part of visualization be-
`cause they determine how well the user can build a mental
`map that he or she can use to understand and navigate the
`visualization. Changing the layout often means having to
`learn a new layout, and thus losing one’s ability to navigate
`easily.
`In our model, the spatial mapping function is called
`(cid:3)(cid:4)(cid:5); it translates from the original data space to an in-
`termediate visualization space (2D or 3D).
`
`3.2. Relevance and Blurring
`
`Independently of the spatial arrangement, the blur level of
`each object is determined. This is done in two steps: First,
`each object is assigned a relevance value  by the relevance
`function (cid:2). The value of  is in the interval (cid:0)(cid:2)(cid:3) (cid:4)(cid:5), where 1
`means the object is maximally relevant, and 0 means the ob-
`ject is completely irrelevant. This relevance value is trans-
`lated into a blur value (cid:2) through the blur function (cid:4) .
`The relevance function is application-specific and thus
`can be very different between applications (see Sect. 5.2 for
`examples). The (cid:4)  function can theoretically also take on
`any shape, but we have found the function depicted in Fig. 3
`to be sufficient for our current uses. The user can specify
`the threshold value , the step height (cid:4), and the maximum
`blur diameter (cid:2) (cid:2)(cid:3). The gradient (cid:5) is then calculated by the
`application.
`
`3.3. Viewing and Camera Models
`
`In order to provide a consistent model, and to embed the
`idea of SDOF in existing work in computer graphics, we
`discuss camera models for generating images with SDOF.
`Depending on whether the visualization space is two- or
`three-dimensional, different camera models can be used to
`finally achieve the SDOF effect. The camera provides two
`functions: (cid:0)(cid:2) (cid:4)(cid:2) projects data values from an intermedi-
`ate space (where the information was laid out by the (cid:2)(cid:0)(cid:4)
`function) to screen space; and (cid:6) (cid:8), which calculates the blur
`level of each data item depending on its (cid:6) coordinate and the
`(cid:6)(cid:4) (cid:6)  value the camera is currently focused at.
`In the following, we describe two camera models: a
`regular photo-realistic camera ((cid:0)(cid:2) (cid:4)(cid:2) ) can be used in
`the 2D case;
`for 3D, we present the adaptive camera
`((cid:0)(cid:2) (cid:4)(cid:2)(cid:10)).
`
`3.3.1 2D SDOF and the Photo-realistic Camera
`In the 2D case, objects get a third coordinate in addition to
`their (cid:7) and (cid:8) values. This additional (cid:6) value depends on
`
`Authorized licensed use limited to: Fish & Richardson PC. Downloaded on June 13,2024 at 01:52:17 UTC from IEEE Xplore. Restrictions apply.
`Proceedings of the IEEE Symposium on Information Visualization 2001 (INFOVIS’01)
`1522-4048/01 $17.00 © 2001 IEEE
`
`3
`
`

`

`(4)
`
`(cid:3) (cid:4)
`
`(cid:5)
`(cid:0)(cid:3)
`(cid:2)
`
`
`
`(cid:0)(cid:0) (cid:3)  (cid:3) (cid:0) (cid:3) (cid:6)
`(cid:8) (cid:3)(cid:4) (cid:0) (cid:3)
`
`An example for an adaptive camera is splatting [7, 24],
`which is a volume rendering technique, but which also can
`be used for information visualization. By changing the size
`of the splat kernel depending on the (cid:3) value of a data point,
`SDOF can be implemented easily.
`Another possibility is to use pre-blurred billboards
`(Sect. 6 and [16]). Objects are rendered into memory, the
`images are then blurred and mapped onto polygons in the
`scene.
`
`4. Properties and Applicability
`
`This section discusses some high-level properties of SDOF,
`how it can be principally applied, and what challenges it
`brings with it.
`
`4.1. Properties
`
`SDOF, being yet another F+C highlighting technique, has
`the following properties that make it an addition to the cur-
`rent toolbox:
`
`(cid:0) SDOF does not distort geometry. It is therefore usable
`when sizes (of objects or parts of objects (glyphs)) and
`positions are used as visual parameters. We also be-
`lieve that it is easier to recognize blurred icons than
`distorted ones.
`(cid:0) SDOF does not alter colors. If color is used to convey
`meaning, or the visualization is to be used by color-
`blind people, SDOF can be used instead of color high-
`lighting. This also means that SDOF is independent of
`color, and can therefore be used when only gray-scale
`is available (e.g., printed images).
`(cid:0) SDOF changes the irrelevant objects, rather than the
`relevant ones. It is therefore useful whenever the rel-
`evant objects contain a lot of information whose per-
`ception might be impeded by changes.
`
`4.2. Applicability
`
`SDOF requires concrete queries to the data (which can be
`simple, but have to be formulated nonetheless), and is there-
`fore useful for analyzing and presenting data.
`SDOF can serve as an additional aid to guide the
`viewer’s attention, together with brighter colors, etc., or as
`a completely separate dimension of data display. Because
`blur is very naturally associated with importance (even
`more than color), we do not believe that it is suitable for
`true multi-variate data visualization. It can, however, add
`
`irrelevant
`
`focus plane
`
`relevant
`2D SDOF scene
`
`3D intermediate
`scene
`
`rendered image
`
`Figure 4. The photo-realistic camera and 2D SDOF.
`
`the intended blur diameter (cid:0) of the object: If the camera is
`focused at (cid:2)(cid:0) (cid:3) , an object with intended blur (cid:0) has to be
`moved to a distance of (cid:2) from the lens of the camera (see
`Fig. 4):
`
`(1)
`
`(2)
`
`(cid:0)(cid:0)(cid:0)(cid:0)
`
`(cid:0)(cid:0)(cid:0)(cid:0)
`
`(cid:4)
`
`(cid:5) (cid:2)(cid:0) (cid:3)  (cid:0)
`(cid:0)(cid:0) (cid:3) (cid:0) (cid:2) 
`(cid:5)  (cid:3)
` (cid:3)
`(cid:0)
`(cid:2)(cid:0) (cid:3) 
`(cid:4)
`
`(cid:0) (cid:0) (cid:0) (cid:3) (cid:2)(cid:3) (cid:2)(cid:0) (cid:3)  (cid:0)
`
`(cid:0) (cid:3) (cid:0) (cid:3) (cid:6)
` (cid:3)(cid:4) (cid:0)(cid:0) (cid:3)  (cid:3)
`
`where (cid:5) is the effective lens diameter as defined in the thin
`lens model [11], and (cid:2) is the focal length of the lens in use.
`The above equations apply to camera models such as dis-
`tribution ray tracing [4], linear post-filtering [19], etc.
`If the camera uses perspective projection, objects also
`have to be scaled (and possibly moved) to compensate for
`depth effects that are not desired in this case.
`
`3.3.2 3D SDOF and the Adaptive Camera
`In the 3D case, of course, it is not possible to directly map
`blur factors to depth values, because the spatial arrangement
`of data items already contains a third dimension. However,
`using a simple extension of the photo-realistic camera, it is
`possible to also handle the 3D case.
`The adaptive camera is a modification of a photo-
`realistic camera that can change its focus for every object
`point to be rendered. This is easily possible with object-
`order rendering, but can also be achieved when rendering
`in image order. In contrast to the photo-realistic camera,
`the adaptive camera can render SDOF in 2D and 3D scenes.
`The photo-realistic camera is, in fact, a special case of the
`adaptive camera (which simply stays focused at the same
`distance for the whole image).
`Function (cid:0) (cid:3) (cid:8) is defined like (cid:0) (cid:3)  in Eq. 1. Different to
`the 2D case, now the inversion of (cid:0) (cid:3) (cid:8) must be resolved for
`(cid:0)(cid:0) (cid:3) -values:
`
`(cid:3) (cid:3) (cid:0) (cid:3) (cid:8)(cid:0)(cid:4) (cid:0)(cid:0) (cid:3)  (cid:3) (cid:0) (cid:3) (cid:0)(cid:4) (cid:0)(cid:0) (cid:3) 
`
`(3)
`
`Authorized licensed use limited to: Fish & Richardson PC. Downloaded on June 13,2024 at 01:52:17 UTC from IEEE Xplore. Restrictions apply.
`Proceedings of the IEEE Symposium on Information Visualization 2001 (INFOVIS’01)
`1522-4048/01 $17.00 © 2001 IEEE
`
`4
`
`

`

`(cid:4)(cid:3) (cid:5)(cid:3)
`
`
`
` (cid:0)(cid:7) (cid:7)(cid:3)(cid:7)(cid:4) (cid:8)
`
`(cid:4)(cid:5)(cid:6)
`
`(cid:3) (cid:4)
`
`(cid:6) (cid:4) (cid:5) (cid:7) (cid:0)
` (cid:5)
`
`(cid:0)(cid:2)(cid:3)
`
`(cid:6)(cid:7)(cid:5)(cid:8)(cid:0)(cid:2)
` (cid:0)
`
`(cid:6) (cid:8)(cid:3)
`
` (cid:0)
`
`(cid:4)(cid:5)(cid:6)
`
`(cid:3)(cid:3)
`(cid:3)(cid:4)
`
`(cid:5) (cid:4) (cid:0) 
`
`(cid:4)(cid:3) (cid:5)(cid:3)(cid:6)
`
`(cid:7)(cid:4) (cid:8)
` (cid:0)(cid:7) (cid:7)(cid:3)
`
`(cid:4)(cid:5)(cid:5)(cid:6)
`
`(cid:3) (cid:4) (cid:6)
`
`(cid:6)(cid:3) (cid:5)  (cid:4) (cid:5) (cid:7) (cid:0)
`(cid:8) (cid:5)
`
`(cid:0)(cid:2)(cid:2)(cid:3)
`
`(cid:6)(cid:7)(cid:5)(cid:8)(cid:5)(cid:2)
` (cid:0)
`
`(cid:6) (cid:8)(cid:3)
`(cid:6)
` (cid:0)
`
`(cid:4)(cid:5)(cid:5)(cid:6)
`
`(cid:3)(cid:3)
`(cid:3)(cid:4)
`(cid:3)(cid:6)
`(cid:5) (cid:4) (cid:0) 
`
`(cid:0)(cid:2)(cid:3)
`
`(cid:3)(cid:4)(cid:5)(cid:0)(cid:2)
` (cid:0)
`
`(cid:2)
` (cid:0)
`
`(cid:4) 
` (cid:0)
`
`(cid:0)(cid:2)(cid:2)(cid:0)(cid:0)(cid:2)
`
`(cid:0)(cid:2)(cid:2)(cid:3)
`
`(cid:3)(cid:4)(cid:5)(cid:5)(cid:2)
` (cid:0)
`
`(cid:0)(cid:2)(cid:2)(cid:0)(cid:0)(cid:2)
`
`(cid:2)
` (cid:0)
`
`(cid:4) 
` (cid:0)
`
`Table 1. All steps necessary for visualizing data values (cid:0)(cid:2)(cid:2)(cid:0)(cid:0)(cid:2) with 2D (top) and 3D SDOF (bottom).
`
`another dimension for a short time, when the displayed data
`is to be queried.
`Blurring needs space, so when a lot of very small objects
`are depicted, it is only of little use. The application can deal
`with this problem by drawing the objects in such an order
`that sharp objects are drawn on top of blurred ones. But this
`can introduce artifacts, where parts of the display appear
`sharp only because of the contrast between sharp objects
`and the background.
`
`from a sharp depiction – this value translates to the step
`height (cid:7) in Fig. 3; b) the minimal difference in blur that can
`be distinguished – this value can be used to calculate (cid:8), if
`the smallest difference between any two  values is given.
`Because this is generally not the case, the blur function is
`adapted for every image after examining the  values of all
`objects. These values can vary with the use of the generated
`image (printing out, projecting onto a wall, etc.), the use of
`different screens, etc.
`
`4.3. Challenges
`
`5.2. User Interaction
`
`SDOF images depend on the output device (similar to tone
`mapping [15], for example). The reason for this is that blur
`is not an absolute measure, but depends on the viewing an-
`gle that the image covers – this is also the reason why small
`images look sharper than larger ones: the circles of confu-
`sion are not visible in the smaller version, or at least to a
`smaller extent. We use a calibration step at program startup
`to account for this problem (see Sect. 5.1).
`Images that contain SDOF effects are also problematic
`when lossy compression is used (like MPEG, JPEG, etc.).
`In this case, artifacts can be introduced that create a high
`contrast in a blurred area, and thus distracting the user.
`But SDOF is most useful in interactive applications, so this
`problem should play no big role in practice.
`
`5. Parameterization
`
`Parameterization of SDOF consists of two parts: Adapta-
`tion to current viewing parameters and user interaction to
`change the relevance mapping.
`
`5.1. Output Adaptation
`
`We ask the user to select two blur levels on program startup:
`a) the minimal blur level that can be easily distinguished
`
`Interaction is a key part of SDOF. Blurred objects are un-
`natural, and it is therefore important for the user to be able
`to change the relevance mapping and blur function quickly,
`and to return to a depiction that shows all objects in focus.
`Depending on the application, there are different usage
`patterns. In many applications, it is useful to be able to point
`at an object and say “Show me all objects that are older
`than this”, “Show me all chessmen that cover this piece”
`(Fig. 5e), or “Show me the cities weighed by their railway
`distance from this city”.
`Another way is to select values independently of objects:
`“Show me all threatened chess pieces of my color”, “Show
`me all files that were changed today” (Fig. 5b), or “Show
`me all current patients weighed by their need for drug xyz”.
`An additional feature we believe is useful is the auto fo-
`cus. After a pre-specified time, it makes all objects appear
`sharp again, thus making examination of all objects easier
`(this function can be turned off).
`Transitions between different displays are always ani-
`mated to enable the user to follow the change and immedi-
`ately see which objects are relevant in the new display. This
`is another reason for separating  and (cid:2) (see section 3.2):
`The animation is done between the old and the new (cid:2) values,
`rather than the  values. This is because the (cid:0)  function
`can contain discontinuities that can lead to jumps between
`
`Authorized licensed use limited to: Fish & Richardson PC. Downloaded on June 13,2024 at 01:52:17 UTC from IEEE Xplore. Restrictions apply.
`Proceedings of the IEEE Symposium on Information Visualization 2001 (INFOVIS’01)
`1522-4048/01 $17.00 © 2001 IEEE
`
`5
`
`

`

`Tree Depth
`
`Tree Depth
`
`Filenames
`
`Filenames
`
`a) A file browser showing the age of files
`through blur. Continuous relevance function.
`
`b) A file browser showing today’s files sharply,
`older ones blurred. Binary relevance function.
`
`c) A chess tutoring system showing the
`chessmen that threaten the knight on e3.
`
`d) A Geographic Information System (GIS) showing
`the roads layer in focus.
`
`e) A chess tutoring system showing the chessmen that cover the knight on e3.
`
`Figure 5. SDOF in action. See Sect. 5.2 and 3 for details.
`
`Authorized licensed use limited to: Fish & Richardson PC. Downloaded on June 13,2024 at 01:52:17 UTC from IEEE Xplore. Restrictions apply.
`Proceedings of the IEEE Symposium on Information Visualization 2001 (INFOVIS’01)
`1522-4048/01 $17.00 © 2001 IEEE
`
`6
`
`

`

`1
`1
`1
`
`1
`1
`1
`
`1
`1
`1
`
`0.16
`0.16
`0.4 0.4 0.4
`0.4
`1
`1
`1
`0.4
`0.4
`1
`1
`1
`0.4
`1
`1
`1
`0.4
`0.4
`0.16 0.4 0.4 0.4 0.16
`
`n = 3
`
`n = 3.8
`
`Figure 6. The Box Filter (left), and the generalized
`box filter for arbitrary sizes (right).
`
`blur levels of objects, and are therefore undesirable.
`
`6. Implementation
`
`A method in information visualization should not only be
`visually effective, but also fast, so that it can be used in-
`teractively. Blurring used to be a very slow operation be-
`cause it involves a sum of three color components of many
`neighboring pixels for every single pixel in the image, and is
`still not supported by hardware except in high-end graphics
`workstations. We have implemented SDOF using texture
`mapping hardware, which makes it fast on currently avail-
`able consumer PCs. The described method is an implemen-
`tation of the adaptive camera model (see Sect. 3.3.2).
`Blur can be understood as a convolution operation of the
`image with a blur kernel. In photography, this blur kernel
`ideally is round, but usually is a regular polygon with six to
`eight vertices, due to the shape of the aperture.
`The more common type of blur kernel in computer sci-
`ence is the box filter (Fig. 6, left). It has the big advantage of
`being separable [16], which reduces its computational cost
`from (cid:0) to (cid:3), where  is the filter size. It can also
`be generalized quite easily to arbitrary sizes (Fig. 6, right)
`other than just odd integers. This implementation directly
`uses (cid:3) as its filter size .
`Using graphics hardware is different from a software im-
`plementation of a filter in that it does not sum up the color
`values of surrounding pixels for every single pixel. Rather,
`it adds the whole image to the frame buffer in one step by
`drawing it onto a textured polygon (this is done by blending
`with a special blend function). When the image is drawn
`in different positions (with one pixel distance between the
`images), several image pixels are added to the same frame
`buffer pixel. Because of the limited accuracy of the frame
`buffer (typically eight bits per color component), this can
`only be done for small values of  (we have found  (cid:0) (cid:4) to
`yield acceptable images).
`
`For larger blur diameters, we use a two-step approach.
`First, we sum up four images into the frame buffer, with
`their color values scaled so that the sum uses the entire eight
`bits. We then transfer this image to texture memory (this is
`a fast operation) and use this auxiliary sum as the operand
`for further calculations. The auxiliary sum already contains
`the information from four addition steps, so when summing
`them up further, only one quarter of the addition steps is
`needed. Because all the values in the box filter (except for
`the border, which is treated separately) are equal, all auxil-
`iary sums are equal – they are only displaced. This means,
`that the auxiliary sum only needs to be computed once (as
`well as another auxiliary for the borders). Summing up aux-
`iliary sums is therefore not only more accurate, it is also
`faster.
`For blur diameters larger than 20 pixels, we first scale
`the image to one quarter of its size, then blur with half the
`diameter, and then scale it back (“quarter method”).
`Using the described method, it is possible to run applica-
`tions – like the ones shown in the images and the accompa-
`nying video – at interactive frame rates (at leat 5 frames per
`second) on cheap consumer graphics hardware. This num-
`ber is likely to increase with some further optimizations as
`well as the use of multi-texturing (which is supported by
`more and more consumer graphics cards).
`
`7. Evaluation
`
`To show that SDOF is actually perceived preattentively, and
`to demonstrate its usefulness in applications, we are cur-
`rently performing a user study with 16 participants. We
`want to find out a) if SDOF is, indeed, perceived preat-
`tentively, which includes the detection and localisation of
`targets, as well as the estimation of the number of targets on
`screen (as a number relative to all objects in the image) in
`the presence of distractors; b) how many blur levels people
`can distinguish, and how blur is perceived (e.g., linear, ex-
`ponential, etc.); c) how blur compares to other visual cues
`which are known to be perceived preattentively (such as
`color and orientation); and d) how well SDOF can be used
`to solve simple problems with simple applications (where
`the emphasis is on the use of SDOF). This study is still in
`progress at the time of this writing, but we will publish the
`results as soon as they are available.
`
`8. Conclusions and Future Work
`
`We have presented an extension to the well-known depth-
`of-field effect that allows objects to be blurred depending on
`their relevance rather than on their distance from the cam-
`era. This technique makes it possible to point the user to
`relevant objects, without distorting the geometry and other
`features of the visualization.
`
`Authorized licensed use limited to: Fish & Richardson PC. Downloaded on June 13,2024 at 01:52:17 UTC from IEEE Xplore. Restrictions apply.
`Proceedings of the IEEE Symposium on Information Visualization 2001 (INFOVIS’01)
`1522-4048/01 $17.00 © 2001 IEEE
`
`7
`
`

`

`Because of the similarity to the familiar depth-of-field ef-
`fect, and the fact that DOF is an intrinsic part of the human
`eye, we believe that it is a quite natural metaphor for visu-
`alization and can be used quite effortlessly by most users.
`SDOF can be used when analyzing and presenting data,
`and also seems to be effective as a tool for pointing infor-
`mation out in tutoring systems.
`We expect to learn a lot about SDOF’s properties dur-
`ing our user study, and will use this information to define
`criteria when and how SDOF can be best used.
`As one of the next steps, we want to investigate the ap-
`plicability of SDOF to other areas of scientific visualization,
`like volume and flow visualization.
`We also want to find out how SDOF can be applied to
`human computer interaction, to enable the user to grasp im-
`portant information faster, and to be alerted to important
`changes without being distracted too much.
`
`9. Acknowledgments
`
`We would like to thank Markus Hadwiger for his help in
`coming up with a fast method for rendering SDOF im-
`ages. This work is part of the Asgaard Project, which
`is supported by Fonds zur F ¨orderung der wissenschaft-
`lichen Forschung (Austrian Science Fund), grant P12797-
`INF. Parts of this work have been carried out in the scope
`of the basic research on visualization at the VRVis Re-
`search Center (http://www.VRVis.at/vis/) in Vi-
`enna, Austria, which is funded by an Austrian governmental
`research program called Kplus.
`
`References
`
`[1] A. Adams. The Camera. Little Brown & Company, 1991.
`[2] E. A. Bier, M. C. Stone, K. Pier, W. Buxton, and T. D.
`DeRose. Toolglass and magic lenses: The see-through in-
`terface. Computer Graphics (Proceedings SIGGRAPH’93),
`27(Annual Conference Series):73–80, 1993.
`[3] G. Colby and L. Scholl. Transparency and blur as selective
`cues for complex visual information. In SPIE Vol. 1460, Im-
`age Handling and Reproduction Systems Integration, pages
`114–125, 1991.
`[4] R. L. Cook, T. Porter, and L. Carpenter. Distributed ray
`tracing. Computer Graphics (Proceedings SIGGRAPH’84),
`18(3):137–145, July 1984.
`[5] G. W. Furnas. Generalized fisheye views. In M. M. Mantei
`and P. Orbeton, editors, Proceedings of the ACM Confer-
`ence on Human Factors in Computer Systems, SIGCHI Bul-
`letin, pages 16–23, New York, U.S.A., 1986. Association for
`Computer Machinery.
`[6] E. B. Goldstein. Sensation and Perception. Brooks/Cole
`Publishing Company, 5(cid:2) edition, June 1998.
`[7] J. Huang, K. Mueller, N. Shareef, and R. Crawfis. Fastsplats:
`Optimized splatting on rectilinear grids. In Proc. Visualiza-
`tion 2000, Salt Lake City, USA, Oct. 2000. IEEE.
`
`[8] S. D. Katz. Film directing shot by shot: Visualizing from
`concept to screen. Focal Press, 1991.
`[9] M. Kreuseler, N. L´opez, and H. Schumann. A scalable
`framework for information visualization. In Proc. Informa-
`tion Vizualization, Salt Lake City, USA, Oct. 2000. IEEE.
`[10] J. Lamping, R. Rao, and P. Pirolli. A focus+context tech-
`nique based on hyperbolic geometry for visualizing large
`hierarchies. In Proceedings CHI’95. ACM, 1995.
`[11] H.-C. Lee. Review of image