Interactive Multi-Perspective Views of Virtual 3D

Landscape and City Models

Haik Lorenz, Matthias Trapp, Jürgen Döllner

Hasso-Plattner-Institute, University of Potsdam, Prof.-Dr.-Helmert-Strasse

2-3, 14482 Potsdam, Germany,
[[Link], [Link], doellner]@[Link]

Markus Jobst

Vienna University of Technology, Erzherzog-Johannplatz 1, A-1040 Vi-

enna, Austria, markus@[Link]

Fig. 1a. A panorama map Fig. 1b. Multi-perspective Fig. 1c. Multi-perspective
painted by H.C. Berann view of a virtual 3D city focus & context visualiza-
(used with permission) model inspired by (a) tion for walk-throughs

Fig. 1. A historic panorama map and examples of interactive multi-perspective

views of 3D city models
Abstract

Based on principles of panorama maps we present an interactive visualiza-

tion technique that generates multi-perspective views of complex spatial
environments such as virtual 3D landscape and city models. Panorama
maps seamlessly combine easily readable maps in the foreground with 3D
views in the background – both within a single image. Such nonlinear,
non-standard 3D projections enable novel focus & context views of com-
plex virtual spatial environments. The presented technique relies on global
space deformation to model multi-perspective views while using a stan-
dard linear projection for rendering which enables single-pass processing
by graphics hardware. It automatically configures the deformation in a
view-dependent way to maintain the multi-perspective view in an interac-
tive environment. The technique supports different distortion schemata be-
yond classical panorama maps and can seamlessly combine different visu-
alization styles of focus and context areas. We exemplify our approach in
an interactive 3D tourist information system.

Keywords: multi-perspective views, focus & context visualization, global

space deformation, virtual 3D city models, virtual 3D landscape models,
geovisualization

1 Introduction and Motivation

Virtual spatial environments based on 3D landscape and city models are

common tools for an increasing number of commercial and scientific ap-
plications and are applied as interactive space and context for planning,
simulation, and visualization tasks. One key requirement represents the ef-
ficient rendering of large amounts of data based on level-of-detail tech-
niques and multiresolution models. Another key requirement is the effec-
tive presentation of the environment and its contents, e.g., by providing
detail views for important areas while giving a coarse overview of their
spatial context.
While a single-perspective view depicts a scene from a single view-
point, “a multi-perspective rendering combines what is seen from several
viewpoints into a single image.” (Yu and McMillan, 2004) Mathemati-
cally, multi-perspective views rely on non-linear 3D projections or,
equivalently, non-planar reference shapes, used to map 3D world space on
2D image space. In this way occlusions become resolvable, scales at which
objects are depicted are adjustable, and spatial context information can be
included in a single image.
With these techniques, multi-perspective views can visually emphasize or
clarify an area of interest while retaining or extending its surrounding area,
achieving an effective information transfer (Keahey, 1998). Furthermore,
they utilize the available screen real estate to a high degree. Their charac-
teristics make multi-perspective views a tool for focus & context visualiza-
tion. Well-known examples include fisheye maps, which emphasize im-
portant information by magnification, or spherical maps, which add
context information by non-uniformly integrating a full 360° view.

1.1 Multi-Perspective Views for Maps

Multi-perspective views have been developed particularly in landscape de-

piction and Cartography. Chinese landscape painters used multi-
perspective views in the 11th century already (Vallance and Calder, 2001).

Fig. 2. Painting of Venice, Italy (about 1550) (Whitfield, 2005). It exhibits a pano-
ramic effect and includes labels

Another example, a 360° panorama view of the London skyline consisting

of six separate paintings, was created in the late 18th century. The incorpo-
ration of cartographic information yields panorama maps. Fig. 2 shows an
early example of Venice, Italy (about 1550). H.C. Berann, an Austrian art-
ist and panorama maker, pioneered one particular kind of panorama map.
Beginning in the early 1930’s he created a deformation and painting style
(Fig. 1(a)), known as “Berann panorama”, which became the de-facto
standard for tourist maps in recreational areas. This style seamlessly com-
bines a highly detailed image of the area of interest with a depiction of the
horizon including major landmarks. The area of interest is shown in the
foreground from a high viewpoint, whereas the horizon is shown in the
background from a low perspective. The environment is depicted with
“natural realism” (Patterson, 2000) and key information such as trails or
slopes is superimposed in an abstracted, illustrated fashion. As a result of
the high viewpoint the foreground shows key information top-down, i.e.,
free from obstructions and clearly visible. At the same time, the map user
can easily orient the map using the horizon, which is visible due to the
changed perspective, as reference without the need for a compass. For
these reasons, panorama maps are useful specifically to unskilled map
readers.
In general, the creation of panorama maps is time consuming and re-
quires a skilled artist. It includes proper viewpoint selection, partial land-
scape generalization, identification of landmarks, their integration into the
map with recognizable shapes, and a smooth transition between the fore-
ground and background perspective (Patterson, 2000). Even with the sup-
port of digital tools and digital 3D geodata, panorama creation still remains
a tedious manual process (Premoze, 2002). Despite their effectiveness,
panorama maps are rarely created, and the creation techniques can hardly
be transferred to interactive systems where the user manipulates the view-
point.

1.2 Multi-Perspective Views for Spatial 3D Environments

Multi-perspective views can be used to visualize 3D landscape models,

e.g., mountainous regions with the mountain peaks providing a distinctly
recognizable background for orientation purposes. Similarly, they can
visualize 3D city models, using the skyline of the city as background. In
today’s applications, interactive visualization is required to support the
user in exploring and analyzing the virtual 3D environment. With respect
to the usability of such applications, the navigation and orientation aids
represent key issues because users frequently “get lost in space” without
guidance (Buchholz et al., 2005). Here, the inclusion of a fixed horizon or
skyline similar to a Berann panorama offers an additional orientation cue
in the sense of a focus & context visualization.
To obtain an automatic, real-time enabled solution, we need to focus on
the projection as major tool for orientation and neglect artistic aspects such
as landmark depiction and selective generalization.
Computer graphics knows three approaches to achieve a panorama ef-
fect: multi-perspective images, deformations, and reflections on non-planar
surfaces (Vallance and Calder, 2001). Multi-perspective images either use
non-linear, non-uniform projections or combine multiple images from dif-
ferent viewpoints to create the final rendering. Deformations distort the
landscape before rendering the final image using a standard projection,
which implies recomputation of all geometric data for every image. Fi-
nally, reflections on non-planar surfaces use standard projections showing
an intermediate object that in turn reflects the landscape.

1.3 Interactive Multi-Perspective Views

Techniques implementing multi-perspective views can be classified as

multi-pass or single-pass. Multi-pass techniques create several intermedi-
ate images that are blended in a final compositing step. Each intermediate
image requires separate data processing, which is rather expensive when it
comes to complex spatial 3D environments. Specifically, out-of-core algo-
rithms can incur additional penalties because rendering of intermediate im-
ages often significantly reduces caching efficiency. Additionally, image
quality suffers due to resampling in the compositing step. Single-pass
techniques do not exhibit these disadvantages, yet they require customiza-
tion of the rendering process available only in software rendering (e.g., ray
tracing) until recently.
With the advent of a programmable rendering pipeline on GPUs the im-
plementation of interactive single-pass multi-perspective view techniques
becomes feasible. We demonstrate a technique that implements a dynamic
global deformation and shifts this task to the GPU. This approach exploits
best the optimization of current graphics hardware for standard projections
both in terms of image quality and speed. We apply our technique to an in-
teractive application that visualizes complex virtual 3D city models in the
context of a tourist information system. Our global deformation is not only
used to mimic Berann panoramas but also for a novel viewing technique
that enables looking ahead the current route in a pedestrian’s view.
An important aspect of this contribution is the analysis of view parame-
ters. In contrast to an artist choosing viewpoints for map creation, users of
interactive applications are inherently free to move. We analyze how to de-
fine the multi-perspective view and how to dynamically adjust our defor-
mation accordingly during the user’s navigation. In addition, we discuss
the implications for common 3D navigation techniques.
The paper is structured as follows. Section 2 discusses related work.
Section 3 explains techniques for interactive multi-perspective views and
their use for focus & context visualization. Section 4 describes the imple-
mentation. Section 5 discusses the test application and its performance.
Section 6 concludes the paper and outlines future work.
2 Related Work

The work of H.C. Berann includes maps, panoramas and fine art (Berann,
2007). His way of creating panorama maps and techniques are described in
(Patterson, 2000). (Premoze, 2002) introduces a first approach for imple-
menting these techniques except for multi-perspective views by means of
3D computer graphics. Additionally, instructions for manual creation of
panorama maps using various tools are available online.
Besides the artistic and visual quality of a Berann panorama, panoramic
depictions use a concept known as focus & context in the field of visuali-
zation. In general, such visualization not only contains the actual subject
but also its embedding context with the goal of supporting the user’s inter-
pretation process. Traditionally, focus & context has been regarded as dis-
tortion-based view of 2D or 3D information where emphasis is achieved
through varying magnification and screen real estate allocation. See
(Leung and Apperley, 1994) for a survey of different approaches and
(Carpendale and Montagnese, 2001) for a general definition. (Vallance and
Calder, 2001) presents ideas on the use of multi-perspective views for fo-
cus & context and their different creation techniques. Recently, this con-
cept has been extended to include other methods for emphasis, such as
generalization, rendering style, blur, or transparency (Hauser, 2003). (Kea-
hey, 1998) generalizes focus & context to providing separate information
dimensions.
Multi-perspective views have been analyzed mainly in the context of ray
tracing, which allows for easy manipulation of the camera model. (Yu and
McMillan, 2004) defines general linear cameras as affine combinations of
3 sample rays. (Löffelmann and Gröller, 1996) proposes a camera model
based on arbitrary surfaces to define viewing rays and a projection surface.
For real-time environments, (Yang et al., 2005) describe 3D view deforma-
tions as postprocessing step to achieve multi-perspective views and
nonlinear perspective projections. (Spindler et al., 2006) improves on this
method by integrating the view deformation directly into the image forma-
tion process through a camera texture. (Glassner, 2004, parts 1 and 2) de-
scribe an interesting non-interactive semiautomatic method to transfer the
artistic Cubism style to computer generated images. Applications of multi-
perspective views include, among others, story-telling, image processing
with the goal of creating panoramas from multiple images or video footage
(Roman et al., 2004), recovery of 3D information (Li et al., 2004), and im-
age-based rendering (Levoy and Hanrahan, 1996).
Deformation is a well-established field in geometric modeling. (Barr,
1984) is one of the first describing deformation operators. Such operators
are used frequently in current modeling tools. Research topics include vol-
ume preservation, avoidance of self-intersection, or deformation control.
Implementations can be classified as shape deformation or space deforma-
tion. Recent examples for the former approach are (Angelidis et al., 2004;
von Funck et al., 2006), which result in interactive deformations for mod-
erately sized models. The latter approach is useful for ray casting or ray
tracing. (Kurzion and Yagel, 1997) presents space deformations for hard-
ware-assisted volume rendering.
A prerequisite for our implementation is rendering of spatial 3D envi-
ronments, which includes terrain rendering (e.g., (Asirvatham and Hoppe,
2005; Hwa et al., 2004; Lindstrom and Pascucci, 2002)) and rendering of
large scenes. Approaches for the latter include out-of-core algorithms (e.g.,
(Buchholz and Döllner, 2005; Gobbetti and Marton, 2005)), specialized
visibility detection algorithms (e.g., (Chhugani, 2005; Wonka et al.,
2001)), and level-of-detail algorithms (e.g., (Sander and Mitchell, 2006)).
Additionally, interaction and navigation within a spatial 3D environment is
necessary. (Buchholz et al., 2005) contains both, a survey of navigation
techniques and improvements to common navigations.

3 Effective Presentation of Spatial 3D Environments

Multi-perspective views facilitate the implementation of effective presen-

tation of spatial 3D environments. They can add valuable cues by seam-
lessly integrating multiple perspectives in the resulting images and, there-
fore, make efficient use of the image space.
In the following, we present two related deformation techniques that
implement multi-perspective views:

1. The bird’s eye view deformation, which mimics Berann’s panorama

maps used to visualize mountain areas.
2. The pedestrian’s view deformation, which swaps the role of fore-
ground and background by presenting a low altitude perspective view
in front of a top view of distant city parts.

In general, both deformations need to ensure the user’s location awareness

during navigation and interaction. Even experienced users get disoriented
if the current perspective does not contain sufficient points of reference or
if the image sequence does not provide spatio-temporal coherence. For
these reasons, both techniques provide a seamless combination of different
views in a single image and achieve interactive frame rates.
We describe both deformation techniques using a reference plane T, a
usually horizontal plane. This plane can be elevated, e.g., to define the roof
of the average building as the horizon or to reduce distortion artifacts. A
point P of the virtual 3D city model not lying in that reference plane is as-
signed a reference point PT in T. Deformation is then calculated using PT
and applied to P. PT can be either a simple vertical projection of P onto T
or – if an object’s shape is to be kept free from distortion – a single refer-
ence point for the whole object.

3.1 Bird’s Eye View Deformation

Fig. 3. The bird’s eye view deformation shows a top view and the horizon simul-
taneously

Similar to Berann’s panorama maps, this deformation is based on

• a depiction of the area of interest using a bird’s eye view, which would
not permit a visible sky,
• a view of the horizon and sky, and
• a smooth transition between both perspectives.

As a result, the landscape appears to be separated into two planar sections

connected by a curved transition zone with the focus lying on the bird’s
eye view part in the foreground or lower image part (Fig. 3). Nevertheless,
the area of interest is not strictly separated from the transition zone but of-
ten reaches into the curved section.
For a painted panorama map, the map designer decides on relevant pa-
rameters such as the two view points and the transition in between. In an
interactive application the user can move the camera. To keep the three
key properties of this multi-perspective view regardless of the camera’s
orientation or position, we define fixed image areas separated by horizon-
tal lines for the bird’s eye view, transition zone, and horizon (Fig. 4). This
fixation results in a transition zone curvature that depends on the viewing
angle, yet the fixed horizon provides strong temporal coherence and eases
orientation tracking during navigation, whereas the ever-changing shape of
the landscape does not lead to distraction. In addition, this implicit defini-
tion of the horizon’s perspective permits the user to navigate relative to
and interact with the focus area using standard metaphors for virtual envi-
ronments while the visual context is adjusted automatically.

Fig. 4. Fixed image separation for the bird’s eye view deformation

In general, painted panoramas exhibit a horizontal horizon. In contrast, an

interactive application can permit rolling of the camera. In this case, the
horizon should provide feedback about the roll angle. In the following de-
scriptions, we assume no rolling.

Fig. 5. Schematic side view of the bird’s eye view deformation

The image subdivision results in the following set of viewing parameters:

• C – camera position
• ν – viewing angle of the reference plane
• bi – line separating focus area and transition zone in the image
• ri – line of the horizon in the image

Fig. 5 sketches a typical setting assuming a perspective projection. In our

implementation the transition zone is guided by a quadratic Bézier spline
due to its continuity properties at the borders. The exact computation is de-
scribed in Section 4. The line b – the projection of bi onto T – marks the
beginning of the transition zone. The half-plane following the transition
zone, which leads to the horizon depiction, is referred to by T’. It is com-
puted as a rotation of T by an angle β about the line r, the projection of ri
onto T. Thus, an object’s shape is maintained outside the transition zone.
Within this zone an object’s shape is preserved only if it uses a single ref-
erence point.
Typically, both bi and ri are rarely changed while C and ν reflect the
user’s navigation. To make efficient use of the screen space, the amount of
visible sky should be minimized while retaining a recognizable skyline.
Placing ri in the upper quarter of the screen generally gives good results.
The location of bi determines the curvature of the transition zone. Placing
bi in the lower half of the screen gives a good compromise between smooth
transition and visibility of the focus area.

3.2 Pedestrian’s View Deformation

Fig. 6. The pedestrian’s view deformation combines a realistic view of the user’s
vicinity with a top view of distant areas
The bird’s eye view deformation supports answering questions such as
“Which direction am I looking to?” without the need for a compass. For
pedestrian’s views, which occur in walk-through scenarios, the question
changes to “Where am I going to?”, e.g., if users want to look ahead the
path along they are currently walking. Due to the low viewing angle, how-
ever, users can generally not obtain an effective overview without chang-
ing the perspective or navigation mode because large parts of the spatial
3D environment are occluded.
To counter this effect, the pedestrian’s view deformation bends upwards
distant parts of the reference plane (Fig. 6). Compared to the technique
proposed in (Vallance and Calder, 2001), which deforms the reference
plane to fit the inside of a cylinder, the pedestrian’s view deformation has
the advantage of using a planar and, hence, clear and undistorted view of
distant regions in the background. In terms of focus & context, the promi-
nent sky in a pedestrian’s view, which provides only little information, is
replaced by a top view of the region ahead, resulting in a more efficient
use of screen space.
Similar to the bird’s eye view deformation, the landscape is separated
into two planar sections connected by a curved transition zone, yet the im-
age-based deformation definition is not appropriate as it does not lead to
comprehensible context behavior. We observe a more effective visualiza-
tion with the Pedestrian’s view deformation when using a fixed orientation
of T’ relative to the reference plane T in world space. Particularly, this en-
ables an intuitive “looking-up” operation to reveal more of the context in-
formation in the background. As a consequence, the curvature of the tran-
sition zone does not depend on the viewing angle but can be defined
independently. Nevertheless, the deformation follows the camera such that
the rotation axis r has a fixed distance and orientation relative to the cam-
era.
With this definition, again the user is relieved from explicitly control-
ling the multi-perspective view. Standard navigation metaphors within the
foreground remain applicable. Only interaction with the background, e.g.,
for the click-and-fly navigation (Mackinlay et al., 1990), needs to be aware
of our deformation for correct object identification.
Fig. 7. Schematic side view of the pedestrian’s view deformation

We use the following set of parameters to specify this multi-perspective

view (cp. Fig. 7):

• C – camera position
• β – angle between T and T’
• db – distance between CT (C projected onto T) and b
• ds – width of the transition zone’s source area

Analog to the bird’s eye view deformation T’ is a rotation of T about r and

the transition zone follows a quadratic Bézier spline. The line b marking
the beginning of the transition zone is always parallel to the image plane
and keeps a distance db from the camera’s vertical projection CT onto T.
We define the line r to be the center line of the transition zone’s source
area. Thus, it is parallel to b at a distance of ds / 2. This definition simpli-
fies the implementation shown in Section 4.
The parameters except for C again change rarely. They control two main
characteristics of the pedestrian’s view deformation: the amount of avail-
able orientation reference in the focus area through db and the amount of
visible context information through β. Setting β to values less than 90°
trades magnification in the context area for visible space and allows for
looking ahead a route farther. The parameter ds directly controls the transi-
tion zone’s curvature where a small transition zone and thus rather high
curvature shows good results.

3.3 Graphical Representation of Focus and Context

Both deformations presented in Section 3 smoothly and seamlessly com-

bine focus and context. Due to the view dependent nature of both deforma-
tions, the user might loose distinction between geometrically correct in-
formation in the focus area and deformed information in the context during
navigation. This might lead to misinterpretations, lost orientation, or erro-
neous navigation (Zanella et al., 2002). Specifically, the pedestrian’s view
deformation, permitting views without visible focus area and, hence, with-
out navigation reference, is prone to such effects. Solutions require visual
cues, e.g., iconic navigation aids or distinct rendering styles for focus and
context such as context color desaturation.
Besides the more effective use of screen space, in focus & context visu-
alization the two constituents can serve different purposes and thus are to
display different information dimensions beyond change of rendering style
(Keahey, 1998; Stone et al., 1994). Whereas the focus gives core informa-
tion, the context shows supporting information.
Panorama maps as inspiration for our bird’s eye view deformation use
this principle by adding thematic information such as trails to the focus
area while the landscape depiction style is constant for the whole image.
We demonstrate an extension showing a map with 3D landmarks in the fo-
cus area. The context remains a complete and photorealistic depiction
since the skyline is required to be recognizable. Nevertheless, generaliza-
tion techniques such as (Döllner et al., 2005) might prove useful. Addi-
tionally, context information can be enriched by labeling landmarks as
seen in some of Berann’s panorama maps.
The pedestrian’s view deformation permits displaying more important
information in the context. In fact, the focus area is limited to serve as
navigation reference and location marker within the spatial 3D environ-
ment whereas the context generally receives the larger screen space and
exhibits less occlusion. According to this observation, our sample visuali-
zation (cp. Fig. 1(c)) shows a photorealistic view in the focus and a map as
context for visual distinction. On top, the current travel route is highlighted
spanning both parts and thus allowing for a route preview.
Rendering such composite depictions does not require multi-pass tech-
niques. Instead, the deformation implementation presented in Section 4
provides a vertex-based interpolation value q ∈ [0; 1] with q = 0 within the
focus area, q = 1 within the context area, and a smooth transition in be-
tween. The rendering styles are then interpolated per pixel based on this
value q.

4 Real-Time Deformation Implementation

The implementation shifts the deformation task to the GPU. Changing ge-
ometry on the GPU, however, has major consequences for standard appli-
cation-based rendering optimizations such as occlusion culling or view
frustum culling.
Our deformation scheme does not introduce new vertices. Rendering arti-
facts due to insufficient tessellation can only appear in the curved transi-
tion zone. This confined nature allows for a straightforward solution: a
Level-of-Detail algorithm selects a more detailed object representation
within the transition zone. Alternatively, on-demand tessellation using
techniques such as generic mesh refinement (Boubekeur and Schlick,
2005) or the newly introduced geometry shaders can be used.
The GPU is a highly parallel streaming processor, thus each vertex
needs to be processed independently. This is achieved by formulating the
deformation of an individual point P as a function f D : P a P′ which is
computed by a vertex program. For efficient computation we want this
function to perform an affine transformation MD on P, where the 4x4-
transformation matrix MD depends only on the reference point PT. Thus,
we reformulate fD as P’ = MD(PT) · P. Both deformations described in Sec-
tion 3 share the same underlying construction, allowing us to use a single
function MD(PT).

Fig. 8. Deformation parameters and definitions. Only objects located in TT become

distorted

For our unified deformation, we divide the reference plane T into three
sections:

1. The undeformed part TF , which becomes the foreground or focus of

the image,
2. The transition zone TT , which becomes curved, and
3. The remainder TB , which becomes the background or context of the
image by rotating T about r.

These three sections are separated by the line b between TF and TT and the
line e between TT and TB. The lines b, r, and e are parallel and equidistant.
Finally, β denotes the angle between T and T’. Fig. 8 sketches this setting.
With these three sections, MD(PT) can be formulated depending on the lo-
cation of PT. In addition, the rendering style interpolation value q can be
derived:

PT ∈ TF : MD(PT) is the identity matrix, since this section is not to be de-

formed; q = 0 .
PT ∈ TT : MD(PT) needs to specify a transformation that transforms PT
and its frame of reference to the corresponding point PT’ on a
quadratic Bézier spline in the tangential frame of reference. A
suitable transformation consisting of a scaling followed by a
rotation based on the de Casteljau algorithm (Gallier, 1999) is
described in the following paragraph; q equals the Bézier
spline parameter t.
PT ∈ TB : MD(PT) is a rotation matrix about r with an angle β; q = 1 .

Fig. 9. The de Casteljau algorithm constructs a Bézier spline point through linear
interpolations. It also provides the point’s tangent

Fig. 9 shows the de Casteljau algorithm for the profile Bézier spline. The
axes b, r, e, and e’ appear as points in this side view, where b, r, and e’ be-
come control points of the spline. Since b − r = e′ − r , the resulting
spline is symmetrical. For a quadratic Bézier spline C (t ) with t ∈ [0; 1] ,
the algorithm uses linear interpolations to construct two intermediate
points rb (t ) = (1 − t )b + tr and re (t ) = (1 − t )r + te and the resulting point
C (t ) = (1 − t )rb (t ) + tre (t ) .
For a given point PT ∈ TT , the corresponding point on the Bézier spline
is found as PT′ = C ( PT − b e − b ) . This mapping does not define an arc
length parameterization of C and thus introduces an unwanted flattening of
objects within the transition zone. The suitable reparameterization of C can
be achieved using a lookup table, which is left for future work.
For our purpose, the key property of the de Casteljau algorithm is the
implicit tangent construction formed by the line through rb and re. To com-
pensate for the variable length contraction along the curve, i.e., the missing
arc length parameterization, a scaling along e − b with a factor
PT′ − rb PT − rb centered at rb is necessary. Then, a rotation of the scaled
PT about rb onto the tangent creates the correct tangential frame of refer-
ence for PT′ . This completes the definition of MD(PT).
The parameters T, b, e, and β of this unified computation depend on the
camera location. Thus, within an interactive application, they need to be
derived from the original (camera-independent) deformation parameters
described in Section 3 on a frame-by-frame basis. Also, especially for
scene graph based systems, the current frame of reference needs to be
taken into account. The most efficient and robust solution is to perform the
deformation in the camera’s frame of reference since it is constant during
image generation.

5 Performance

Fig. 10a. Bird’s eye view deformation Fig. 10b. Pedestrian’s view deforma-
showing public transport lines tion with highlighted route

Fig. 10. Sample multi-perspective images. The insets show the corresponding
standard perspectives

Fig. 10 shows sample images of the bird’s eye view deformation and pe-
destrian’s view deformation, respectively. For comparison, the insets show
a standard perspective projection using identical camera settings to high-
light the effects of our focus & context visualization.
We extended an existing 3D tourist information system for Berlin, Ger-
many, with our technique. Despite the additional data handling overhead
for two rendering styles, we were able to achieve interactive frame rates.
Table 1 summarizes average frame rates for two sample camera paths per
view deformation. The measurements were made on a PC with an AMD
Athlon 64 X2 (2.3 GHz), 2 GB main memory, and a NVidia GeForce
7900GT with 256 MB video memory. The test application does not utilize
the second CPU core. The sample dataset comprises the inner city of Ber-
lin with about 16,000 generically textured buildings, about 100 landmarks,
a 3 GB color aerial photo, and a 250 MB grayscale map image on top of a
digital terrain model.

Table 1. Performance measurements for different screen resolutions and con-

figurations
Resolution Configuration Path frames/ frames/sec without
sec bend.
1600x 1200 Pedestrian’s view 1 11.72 12.95
2 19.33 15.69
Bird’s eye view 3 8.35 29.86
4 6.73 17.64
1024x 768 Pedestrian’s view 1 17.85 15.63
2 22.75 17.69
Bird’s eye view 3 8.87 27.24
4 5.42 16.07
800x 600 Pedestrian’s view 1 20.54 16.49
2 23.94 18.42
Bird’s eye view 3 8.74 27.26
4 8.48 19.52
The frame rate without deformation is largely resolution independent sug-
gesting texture access as main bottleneck in our test application. To deal
with the texture amount, an out-of-core algorithm is used to load texture on
demand in sufficient resolution from disk. Table 2 shows the average
number of bytes read from hard disk per frame for our test setting at reso-
lution 1600x1200.
Table 2. Average hard disk access per frame with / without deformation at resolu-
tion 1600x1200
Configuration Path bytes/ bytes/frame without
frame bend.
Pedestrian’s view 1 260,207 407,822
2 122,729 215,398
Bird’s eye view 3 5,720,803 190,824
4 2,555,602 243,067
The exceptionally high load rates for the bird’s eye view deformation are
caused by the visible horizon. Compared to the corresponding standard
perspective projection, more terrain is visible and thus more texture re-
quires loading – even though at low quality. At the same time, changing
the view direction invalidates more texture. Hence, caching efficiency is
reduced dramatically. With the pedestrian’s view deformation, only a frac-
tion of the terrain is visible compared to a standard view, but due to the de-
formation distant terrain requires a significantly higher texture resolution
leading to only a slight reduction in texture load overhead.

6 Conclusions

We have demonstrated the concept and implementation of interactive

multi-perspective views for spatial 3D environments. They are inspired by
the well-known panorama maps and aim to increase the effectiveness of
interactive applications by using the principle of focus & context visualiza-
tion. Our implementation is based on a global space deformation processed
by graphics hardware and permits the seamless combination of different
graphical representations for focus and context areas. To verify its applica-
bility we have successfully integrated our technique into an existing inter-
active 3D tourist information system.
The visual quality in the transition zone can be further improved by in-
corporation of on-demand geometry tessellation, e.g., through the use of
geometry shaders, or by adaptation of a more advanced bending scheme.
In contrast to the currently used simple static lighting, dynamic lighting
and shadowing within a deformed 3D landscape model remains an inter-
esting open question. While this contribution describes the underlying
technology, user studies about the effectiveness and/or expressiveness of
our visualization approach, different rendering style combinations, and
navigation in a deformed 3D landscape model remain future work.
Acknowledgements

We would like to thank 3D Geo GmbH ([Link]) for providing the

implementation platform LandXplorer, an authoring and presentation sys-
tem for virtual 3D city models and landscape models. We also thank Mat-
thias Troyer for his permission to use the panorama map shown in Fig.
1(a).

References

The world of H.C. Berann (accessed 2007), url: [Link]

Angelidis, A., Cani, M.-P., Wyvill, G. & King S. (2004), Swirling sweepers: Con-
stant-volume modeling, in Proceedings of the 12th Pacific Conference on
Computer Graphics and Applications, IEEE Computer Society, Washington,
DC, USA, pp. 10-15.
Asirvatham, A. & Hoppe, H. (2005), Terrain Rendering Using GPU-Based Ge-
ometry Clipmaps, in M. Pharr (ed.), GPU Gems 2, Addison-Wesley, pp. 27-
45.
Barr, A. H. (1984), Global and Local Deformations of Solid Primitives, in
SIGGRAPH '84: Proceedings of the 11th annual conference on Computer
graphics and interactive techniques, ACM, New York, NY, USA, pp. 21-30.
Boubekeur, T. & Schlick, C. (2005), Generic Mesh Refinement on GPU, in Pro-
ceedings of ACM SIGGRAPH/Eurographics Graphics Hardware 2005, ACM,
pp. 99-104.
Buchholz, H.; Bohnet, J. & Döllner, J. (2005), Smart and Physically-Based Navi-
gation in 3D Geovirtual Environments, in IV '05: Proceedings of the Ninth In-
ternational Conference on Information Visualisation, IEEE Computer Society,
Washington, DC, USA, pp. 629-635.
Buchholz, H. & Döllner, J. (2005), View-Dependent Rendering of Multiresolution
Texture-Atlases, in Proceedings Information Visualization 2005, pp. 215-222.
Carpendale, M. S. T. & Montagnese, C. (2001), A Framework For Unifying Pres-
entation Space, in UIST '01: Proceedings of the 14th annual ACM symposium
on User interface software and technology, ACM, New York, NY, USA, pp.
61-70.
Chhugani, J.; Purnomo, B.; Krishnan, S.; Cohen, J.; Venkatasubramanian, S. &
Johnson, D. S. (2005), vLOD: High-Fidelity Walkthrough of Large Virtual
Environments, IEEE Transactions on Visualization and Computer Graphics
11(1), pp. 35-47.
Döllner, J.; Buchholz, H.; Nienhaus, M. & Kirsch, F. (2005), Illustrative Visuali-
zation of 3D City Models, in Robert F. Erbacher; Jonathan C. Roberts; Matti .
T. Gröhn & Katy Börner, ed., Visualization and Data Analysis 2005, pp. 42-
51.
von Funck, W., Theisel, H. & Seidel, H.-P. (2006), Vector field based shape de-
formations, in Proceedings ACM SIGGRAPH 2006, ACM, New York, NY,
USA, pp. 1118-1125.
Gallier, J. (1999), Curves and Surfaces in Geometric Modeling: Theory and Algo-
rithms, Morgan Kaufmann Publishers Inc., San Francisco, CA, USA.
Glassner, A. (2004), Digital Cubism, IEEE Computer Graphics and Applications
24(3), pp. 82-90.
Glassner, A. (2004), Digital Cubism, Part 2, IEEE Computer Graphics and Appli-
cations 24(4), pp. 84-95.
Gobbetti, E. & Marton, F. (2005), Far Voxels: A Multiresolution Framework for
Interactive Rendering of Huge Complex 3D Models on Commodity Graphics
Platforms, ACM Trans. Graph. 24(3), pp. 878-885.
Hauser, H. (2003), Generalizing Focus+Context Visualization, in Scientific Visu-
alization: The Visual Extraction of Knowledge from Data (Proc. of the
Dagstuhl 2003 Seminar on Scientific Visualization), Springer, pp. 305-327.
Hwa, L. M.; Duchaineau, M. A. & Joy, K. I. (2004),Adaptive 4-8 Texture Hierar-
chies, in VIS '04: Proceedings of the Conference on Visualization '04, IEEE
Computer Society, Washington, DC, USA, pp. 219-226.
Keahey, A. (1998),The Generalized Detail-In-Context Problem, in INFOVIS '98:
Proceedings of the 1998 IEEE Symposium on Information Visualization,
IEEE Computer Society, Washington, DC, USA, pp. 44-51.
Kurzion, Y. & Yagel, R. (1997), Interactive Space Deformation with Hardware-
Assisted Rendering, IEEE Computer Graphics and Applications 17(5), pp.
66-77.
Leung, Y. K. & Apperley, M. D. (1994), A Review and Taxonomy of Distortion-
Oriented Presentation Techniques, ACM Transaction on Computer-Human In-
teraction 1(2), pp. 126-160.
Levoy, M. & Hanrahan, P. (1996), Light Field Rendering, in Proceedings ACM
SIGGRAPH 1996, ACM, New York, NY, USA, pp. 31-42.
Li, Y.; Shum, H.; Tang, C. & Szeliski, R. (2004), Stereo Reconstruction from
Multiperspective Panoramas, IEEE Trans. Pattern Anal. Mach. Intell. 26(1),
pp. 45-62.
Lindstrom, P. & Pascucci, V. (2002), Terrain Simplification Simplified: A Gen-
eral Framework for View-Dependent Out-of-Core Visualization, IEEE Trans-
actions on Visualization and Computer Graphics 8(3), pp. 239-254.
Löffelmann, H. & Gröller, E. (1996), Ray Tracing with Extended Cameras, Jour-
nal of Visualization and Computer Animation 7(4), pp. 211-227.
Mackinlay, J. D.; Card, S. K. & Robertson, G. G. (1990), Rapid Controlled
Movement Through a Virtual 3D Workspace, in SIGGRAPH '90: Proceedings
of the 17th Annual Conference on Computer graphics and Interactive Tech-
niques, ACM, New York, USA, pp. 171-176.
Patterson, T. (2000), A View From on High: Heinrich Berann's Panoramas and
Landscape Visualization Techniques For the US National Park Service, Car-
tographic Perspectives 36, pp. 38-65.
Premoze, S. (2002), Computer Generated Panorama Maps, in Proceedings 3rd
ICA Mountain Cartography Workshop. Mt. Hood, Oregon.
Roman, A.; Garg, G. & Levoy, M. (2004), Interactive Design of Multi-Perspective
Images for Visualizing Urban Landscapes, in VIS '04: Proceedings of the con-
ference on Visualization '04, IEEE Computer Society, Washington, DC, USA,
pp. 537-544.
Sander, P. V. & Mitchell, J. L. (2006), Progressive Buffers: View-Dependent Ge-
ometry and Texture LOD Rendering, in SIGGRAPH '06: ACM SIGGRAPH
2006 Courses, ACM, New York, USA, pp. 1-18.
Spindler, M., Bubke, M., Germer, T. & Strothotte, T. (2006), Camera textures, in
Proceedings of the 4th international conference on Computer graphics and in-
teractive techniques in Australasia and Southeast Asia, ACM, New York,
USA, pp. 295-302.
Stone, M. C., Fishkin, K. & Bier, E. A. (1994), The movable filter as a user inter-
face tool, in Proceedings of the Conference on Human Factors in Computing
Systems, ACM, New York, USA, pp. 306-312.
Vallance S. & Calder, P. (2001), Multi-perspective images for visualization, in
ACM International Conference Proceeding Series, Vol. 147, ACM, New
York, USA, pp. 69-76.
Whitfield, P. (2005), Cities of the World. A History in Maps, The British Library,
London.
Wonka, Peter; Wimmer, Michael & Francois Sillion (2001), Instant Visibility, in
A. Chalmers & T.-M. Rhyne, ed.,Proceedings of Eurographics 2001, The Eu-
rographics Association and Blackwell Publishers, pp. 411-421.
Yang, Y.; Chen, J. X. & Beheshti, M. (2005), Nonlinear Perspective Projections
and Magic Lenses: 3D View Deformation, IEEE Computer Graphics and
Applications 25(1), pp. 76-84.
Yu, J. & McMillan, L. (2004), A Framework for Multiperspective Rendering, in
Alexander Keller & Henrik Wann Jensen, ed., Rendering Techniques 2004,
Proceedings of Eurographics Symposium on Rendering 2004,
EUROGRAPHICS Association, pp. 61-68.
Zanella, A., Carpendale, M. S. T. & Rounding, M. (2002), On the effects of view-
ing cues in comprehending distortions, in Proceedings of the second Nordic
conference on Human-computer interaction, ACM, New York, USA, pp. 119-
128.