OBJECTIVE QUALITY SELECTION FOR HYBRID LOD MODELS

Tom Jehaes, Wim Lamotte

Hasselt University, Expertise Centre for Digital Media

Interdisciplinary institute for BroadBand Technology (IBBT)

Wetenschapspark 2, 3590 Diepenbeek, Belgium

Nicolaas Tack

IMEC, Kapeldreef 75, 3001 Leuven, Belgium

Keywords:

Progressive Geometry, Image-Based Rendering, Level-of-Detail, Objective Quality Selection.

Abstract:

The problem of rendering large virtual 3D environments at interactive framerates has traditionally been solved

by using polygonal Level-of-Detail (LoD) techniques, for which either a series of discrete models or one pro-

gressive model were determined during preprocessing. At runtime, several metrics such as distance, projection

size and scene importance are used to scale the objects to such a resolution that a target framerate is maintained

in order to provide a satisfactory user experience. In recent years however, image-based techniques have re-

ceived a lot of interest from the research community because of their ability to represent complex models in a

compact way, thereby also decreasing the time needed for rendering. One of the questions however that should

be given some more consideration is when to switch from polygonal rendering to image-based rendering. In

this paper we explore this topic further and provide a solution using an objective image quality metric which

tries which we use to optimize render quality. We test the presented solution on both desktop and mobile

systems.

1 INTRODUCTION AND

RELATED WORK

In order to efﬁciently render large virtual 3D environ-

ments at interactive framerates, application designers

have traditionally resorted to using polygonal Level-

of-Detail (LoD) techniques. Usually, either a series

of discrete models or one progressive model are de-

termined during preprocessing. At runtime, a target

framerate is maintained in order to provide a satisfac-

tory user experience. This is achieved by employing

several metrics such as distance, projection size and

scene importance to scale the objects a suitable res-

olution. In recent years however, image-based tech-

niques have received a lot of interest from the research

community because of their ability to represent com-

plex models in a compact way, thereby also decreas-

ing the time needed for rendering.

An important question however, is when to switch

from polygonal rendering to image-based rendering.

Previously, a subjective quality metric was often used

for determining the switching distance at runtime. In

this paper we explore this topic further and provide

a solution using an objective image quality metric

which tries to optimize render quality.

Our following related work discussion is divided

into three categories: ﬁrstly, polygonal LoD tech-

niques which have been used traditionally, secondly,

image-based rendering techniques which have been

researched in previous years as an alternative render-

ing method for distant objects and ﬁnally, mixed sys-

tems, beneﬁting from the strengths of each system:

polygonal rendering for objects close by in high qual-

ity, faster image-based rendering for objects further

away providing sufﬁcient detail.

1.1 Polygonal LoD

The ﬁrst techniques for choosing the Level of Detail

(LoD) of polygonal objects have used static heuris-

tics based on the screen size of an average face of

the object (Funkhouser et al., 1992) and/or the dis-

tance of the object to the user (Blake, 1987; Rossignac

and Borrel, 1993). These simple heuristics improve

the frame rate in many cases, but cannot guarantee a

regulated or bounded execution time. To guarantee a

241

Jehaes T., Lamotte W. and Tack N. (2008).

OBJECTIVE QUALITY SELECTION FOR HYBRID LOD MODELS.

In Proceedings of the Third International Conference on Computer Graphics Theory and Applications, pages 241-248

DOI: 10.5220/0001099102410248

 SciTePress

bounded frame rate, (Funkhouser and S

equin, 1993)

have presented a predictive technique that uses an es-

timate of the execution time for the correct choice of

the LoD to use. The core of this work is a multiple

choice knapsack problem that maximizes the visual

quality for a given maximum execution time. The

greedy solution of Funkhouser and S

equin however

only guarantees to be half as good as the optimal so-

lution. Therefore, (Gobbetti and Bouvier, 2000) have

proposed to use convex optimisation (interior point al-

gorithm) with a guaranteed speciﬁed accuracy.

1.2 Image-based Rendering

As a solution for faster rendering of large geomet-

rical models, Image-Based Rendering (IBR) has re-

ceived many followers over the last years. The render

times of image-based models are fairly constant and

rely on the resolution of the reference images instead

of polygon count. This means that for high polygon

count models, image-based rendering can be a much

faster render solution without introducing too much

visual degradation. For closeup viewing of detailed

models however, conventional polygon rendering is

often still preferred. McMillan ﬁrst proposed the fun-

damental 3D warping equation together with an oc-

clusion compatible warping order to efﬁciently render

new views based on a series of reference depth im-

ages (McMillan and Bishop, 1995). Extensions to this

have been presented in (Shade et al., 1998), (Oliveira

and Bishop, 1999) and (Chang et al., 1999) whereby

separate models are represented by a cluster of refer-

ence images.

More recently, (Oliveira et al., 2000) shows that

a factorisation of the 3D warping equation, called

relief-texture mapping (RTM), enables the use of fast

graphics hardware for part of the calculation to speed

up the warping. Warping is done in two steps: a pre-

warp followed by a simple texture mapping. This

delivers a signiﬁcant speedup because the pre-warp

is implemented using a fast two-pass reconstruction

algorithm and the texture mapping can be done on

fast graphics hardware. Layered Relief Textures pre-

sented in (Parilov and Stuerzlinger, 2002) combine

the idea of storing multiple samples per pixel with fast

hardware assisted warping of RTM. Finally, (Fujita

and Kanai, 2002) incorporate dynamic shading into

the RTM approach by using per-pixel shading hard-

ware.

1.3 Qos for Mixed Systems

Conventional texture mapping can be seen as a very

simple form of image-based rendering that does not

take depth information into account. Based on this,

(Maciel and Shirley, 1995) focus on maintaining a

high framerate by replacing clusters of objects with

simple texture mapped primitives. Similarly, (Shade

et al., 1996) use a BSP tree scene representation for

which they cache images of nodes that were rendered

in previous frames. Taking frame-to-frame coherence

into account they reused these cached images for ren-

dering subsequent frames, thereby gaining a signiﬁ-

cant rendering speedup.

Later on, making use of depth image representa-

tions, (Rafferty et al., 1998) extend a portal culling

renderer in which they determine the view through a

portal by warping a precalculated Layered Depth Im-

age (LDI) that captures the view through that portal.

This approach is generalized for massive model ren-

dering in (Aliaga and Lastra, 1999). A grid of view-

points is constructed for which the far geometry is de-

termined after which LDIs are created that represent

this far geometry. The renderer can then ﬁrst render

far geometry from LDIs followed by polygon based

rendering of the near geometry. The MMR system

(Aliaga et al., 1999) replaces the LDIs in the previous

approach by Textured Depth Meshes (TDMs) to make

optimal use of current graphics hardware.

Two fundamental drawbacks of the former hy-

brid techniques are the often very long preprocess-

ing times and huge storage requirements. Therefore,

a system that creates image-based representations on

demand without the need for additional storage was

presented in (Hidalgo and Hubbold, 2002). They em-

ploy a dual renderer setup in which the hybrid ren-

derer (HR) can request reference depth images rep-

resenting the current far geometry from a reference

image generator (RIG) that runs in parallel. While

the RIG is working on the requested data, the HR can

use warping on the previous reference depth image to

render its frames. Prediction is used to request opti-

mal reference images.

A speciﬁc optimization algorithm for a terrain ﬂy-

over application was proposed in (Zach et al., 2002).

Interesting in this approach is that they not only use

discrete and continuous Level Of Detail for polygonal

rendering but also point-based rendering for the trees

on the terrain.

A major drawback of these approaches is that they

are not well suited for handling dynamic scenes for

which the contents is not known beforehand. This

is the case for instance for the increasingly popu-

lar MMORPGs. We have therefore selected to use

the hybrid rendering technique presented in (Jehaes

et al., 2004), which was extended to mobile devices

in (Jehaes et al., 2005). This technique uses a com-

bination of progressive geometry and relief texture

GRAPP 2008 - International Conference on Computer Graphics Theory and Applications

242

mapped objects for representing separate objects in

the scene. The authors have however only used a sub-

jective quality metric for selecting at run-time which

render method should be used for each object in-

stance. In this paper we therefore present a objective

metric based on image quality.

In the next section we give some more details

about hybrid object representation. Following that,

we present our solution for an objective LoD selec-

tion metric which is meant to maintain a high level of

image quality. Finally, we present our test setup and

results on both desktop and mobile systems and end

with conclusions and some pointers for future work.

2 HYBRID MODELS

The hybrid geometric/image-based rendering scheme,

which was presented in (Jehaes et al., 2004), makes

use of geometry simpliﬁcation based on (Hoppe,

1996), and the relief-texture mapping technique

which was introduced by Oliveira et al. (Oliveira

et al., 2000). This last technique is used because

it makes efﬁcient use of the texturing capabilities of

current graphics cards to speed up rendering of the

image-based models. It also integrates easily into the

standard geometry rendering pipeline. An example

of relief texture mapped model representations can be

seen in ﬁgure 7(b). The relief textures capture the

appearance of the model as seen from each side of

the bounding box. During rendering, each visible re-

lief texture is pre-warped into a texture which is sub-

sequently mapped onto the corresponding bounding

box quad, resulting in a correct view of the model.

Because the pre-warp equation can be efﬁciently cal-

culated, the total render time will be small and mostly

depends on the resolution of the relief textures.

The render scheme also supports animated objects

such as avatars. Often, the animations for objects fur-

ther away do not add much to the visual quality, so

they can be suppressed. Furthermore, these distant

animated objects can then be replaced by RTM ob-

jects, using the same LoD rendering scheme.

3 OBJECTIVE HYBRID LOD

SELECTION

Triangular rendering is nowadays one of the most ef-

ﬁcient methods for real-time rendering of 3D con-

tent. However, in some cases, triangles can become

so small that the processing of their three vertices and

the rasterization are done for shading only one pixel.

In this case, other rendering algorithms, such as IBR,

become more efﬁcient. Therefore, Hybrid LoD selec-

tion also selects the most efﬁcient rendering method

next to the correct LoD for the rendered object. Sub-

sections 3.1 and 3.2 respectively discuss LoD selec-

tion for the polygonal and IBR objects. Subsection

3.3 then discusses the selection of the most efﬁcient

rendering algorithm.

3.1 Polygonal LoD Selection

The LoD selection mechanism for polygonal objects

uses a Pareto optimisation (Tack et al., 2006) for the

selection of the optimal LoDs in terms of quality and

render time. Central in this approach is the use of the

Pareto plot, which is a collection of Pareto optimal

points for which it is impossible to improve the vi-

sual quality without increasing the cost. The Pareto

plots are measured - for each separate polygonal ob-

ject in the scene - in off-line preprocessing steps and

encoded with the objects. The online steps of the opti-

mization use a gradient descent algorithm to combine

the Pareto plots of the visible objects in the scene with

a minimum of online parameters (e.g. object distance)

to ﬁnd the optimal trade-off between quality and cost.

3.2 LoD Selection for Relief Texture

Mapping

Advanced LoD selection mechanisms use a model for

estimating the performance and the quality (Gobbetti

and Bouvier, 2000). However, these systems control

the render time of polygonal rendering, which is a

very regular algorithm and hence easy to model. Re-

lief Texture Mapping has an irregular ﬂow and its per-

formance model is much more complicated, i.e. it is

necessary to track a lot of online parameters to obtain

a good estimate of the render time. Consequently, us-

ing this performance model at run-time has a nega-

tive effect on frame render times. We have therefore

chosen a different approach for the RTM control algo-

rithm which requires much less run-time processing:

the selection of the switching distance and resolution

of the relief texture mapped objects by using a target

quality of the rendered image as the control parame-

ter.

In order to specify a target quality, we have used

two objective quality metrics for measuring the vi-

sual quality of an RTM model: the Peak Signal-

to-Noise Ratio (PSNR) and Structural SIMilarity

(SSIM) (Wang et al., 2004) metrics. These metrics

compare a reference image with the image that con-

tains the rendered RTM object. The reference im-

age is obtained by rendering the polygonal object at

OBJECTIVE QUALITY SELECTION FOR HYBRID LOD MODELS

243

Figure 1: The PSNR as function of the viewing angles ϕ

and θ (30

◦

steps) for three different distances to the object

for a single resolution RTM object.

full quality. Next, the quality is sampled for different

viewing angles around the object. The viewpoint is

ﬁrst located at a ﬁxed distance and both the viewing

angles ϕ and θ are set to 0, the quality is derived and

the viewing angle ϕ is increased by a preset amount.

This is repeated until the viewing angle ϕ is 360

◦

then ϕ is reset to 0 and θ is increased. The procedure

is stopped at θ equal to 180

◦

and ϕ equal to 360

◦

. This

procedure is then repeated for different distances.

Figure 1 illustrates the result for an RTM object

(ﬂamingo) with a resolution of 512x512. The PSNR is

plotted as a function of the viewing angles ϕ and θ and

distance between the observer and the object. From

ﬁgure 1, one can derive the parameters onto which

the quality depends:

• The quality increases with a larger distance be-

tween the viewpoint (observer) and the object.

• The quality is the highest for viewpoints near to

the viewpoints for which the Relief Texture Maps

were derived: ϕ equal to 0

◦

, 90

◦

, 180

◦

and

270

◦

and θ equal to 0

◦

, 90

◦

and 180

◦

. For

in-between viewpoints, interpolations between

RTMs are needed and the quality decreases be-

cause of occlusions in the used relief texture map.

An additional parameter, which is not shown in

ﬁgure 1 but that also inﬂuences the quality is the res-

olution of the RTMs. This is illustrated in Figure 2,

which shows the PSNR as a function of the distance to

the viewpoint for an RTM object at a single viewing

angle and different resolutions.

Figure 1 and Figure 2 can be summarised in a ta-

ble, which stores the resolution of the RTM object as

a function of quality, distance and viewing angle (see

Table 1 for a single viewing angle). This table is de-

rived in preprocessing and is used at run-time for the

selection of the correct RTM resolution.

Figure 2: The quality (PSNR) as a function of distance to

the viewpoint and resolution. For small distances, a high

resolution is needed for a acceptable target quality of 30

dB; for large distances, this needed resolution decreases.

Table 1: The selection of the RTM resolution as a function

of requested quality and distance.

dist/Q 20 dB 25 dB 30 dB 35 dB

5 m 16x16 128x128 512x512 -

10 m 16x16 16x16 64x64 256x256

20 m 16x16 16x16 16x16 64x64

30 m ... ... ... ...

3.3 Hybrid LoD Selection

To be able to select the most efﬁcient rendering algo-

rithm, a comparison between polygonal rendering and

RTM is needed. Figure 3 therefore shows the render

time as a function of the quality (PSNR), object dis-

tance and the used rendering technique. The quality

is varied by changing the resolution of the relief tex-

tures for RTM and the number of triangles for polyg-

onal rendering. The render time was measured on a

Pentium M processor with a frequency of 1.86GHz

and an Intel 915GM graphics accelerator. The time

reported in Figure 3 for RTM is the total time needed

for a complete warping operation, of which the pre-

warping takes 90% (Oliveira et al., 2000). However,

the pre-warping is only done when the relative posi-

tion of the camera to the object changes and the real

impact of RTM on the total render time will therefore

depend on the user navigation and object animations.

Figure 3 shows an exponential relationship be-

tween the execution time and quality for RTM, while

the relationship is linear for polygonal rendering.

From Figure 3, it is clear that high resolution relief

textures (points with high execution time) must be

used for content which is close to the observer.E.g.,

for the distance of 2.8m (RTM:2.8), the technique

gives a quality of 23 dB for an execution time of 23

ms. Even if the pre-warping is not executed for ev-

GRAPP 2008 - International Conference on Computer Graphics Theory and Applications

244

Figure 3: Polygonal rendering versus RTM: execution time

as a function of quality and distance (2.8m, 10m and

22.4m).

ery frame, its impact is still too high to use RTM

for close content. However, for distant content (e.g.

RTM:22.4), low resolution relief textures give an ac-

ceptable quality (> 30dB) for a low execution time.

Even if the pre-warping is executed every frame, the

render time is still comparable with the polygonal ren-

dering time. If the pre-warping is not needed, then

the rendering is limited to rendering of the texture

mapped bounding box and the rendering time is neg-

ligible.

Our hybrid LoD selection ﬁrst uses table 1 to se-

lect the resolution of the RTM object to meet the re-

quested quality. If this resolution is higher than a hard

limit (e.g. 128x128), polygonal rendering is used for

that object. The RTM objects are then rendered and

the render time is measured. This render time is then

subtracted from the render time budget and the Pareto

optimisation for polygonal rendering is executed for

maximising the quality for the polygonal objects for

the computed time budget.

In order to optimize user experience and interac-

tivity we set a target framerate and make adjustments

to the target quality based on the current framerate.

When the current framerate is either higher or lower

than the target framerate, we increase or decrease the

target quality. Since our LoD selection is directly

linked to the target image quality, switching distances

and resolutions are automatically adjusted resulting in

a higher or lower framerate.

4 RESULTS

In order to determine the usefulness of our image-

based quality metric in an actual application, we set

up a test for both desktop and mobile systems. Con-

centrating ﬁrst on the desktop scenario, we set out

to render a scene consisting of 400 object instances

which were randomly chosen from 8 different base

models. Each model representation consisted of a

progressive mesh representation, with a correspond-

ing Pareto plot, and a RTM representation for which

the resolutions and switching distances were deter-

mined by using the image-based metric presented

above. For reasons of comparison, we deﬁned a cam-

era path through the scene which would be used for

each of the 4 render scenarios: full resolution geom-

etry, image-based only (IBR), progressive mesh only

(PM) and mixed representations. As can be seen in

ﬁgure 5(a), using full resolution geometry rendering

would result in a maximal triangle count of more than

1600000 for which we got a framerate of about 4 fps.

Note that the overly large frametimes for the full ge-

ometry walkthrough were left out of the graphs for

reasons of readability. Fortunately, we can increase

the framerate considerably by using one of the other

representation types as shown in ﬁgure 4(a). The tar-

get frametime for both the PM and mixed scenarios

was set to 20ms, while the IBR scenario would render

at maximum speed. We could have also regulated the

IBR framerate by increasing the RTM resolution, but

we wanted to show the speedup beneﬁts of using the

IBR approach.

During our experiments we tested with both the

PSNR and SSIM image quality metrics and found that

the SSIM metric gave results which were more con-

sistent compared to our own subjective evaluation of

the image quality. After performing some tests com-

paring the results of both metric, we clearly noticed

that the some PSNR results were inconsistent with

what was to be expected. The SSIM metric exhib-

ited a much more consistent result, so we therefore

switched over to this metric. By combining the frame-

time graphs with the results of image quality measure-

ments shown in table 2, which were taken at the corre-

sponding points indicated on ﬁgure 4(a), we can con-

clude that our mixed representation scenario, using

the image-based metric results in higher image quality

at comparable frametimes which was what we set out

to achieve. The difference in image quality between

the four scenarios is indicated in ﬁgure 6. The use

of the mixed scenario clearly allows for much higher

resolution meshes to be used for objects close to the

viewer. As our camera moves through the scene, less

objects are in the view and the PM and mixed sce-

narios converge to the same image quality, which can

also be concluded from ﬁgures 5(a) and 5(c).

For our second test we set up a scene consisting

of 60 object instances, using the same base models as

in the desktop test. The application was deployed on

Dell X51v PDAs, which incorporate the Intel 2700G

GPU. From ﬁgure 5(b) it can be seen that the maximal

triangle count at application start is about 230000 for

OBJECTIVE QUALITY SELECTION FOR HYBRID LOD MODELS

245

(a) Desktop (b) PDA

Figure 4: Frametimes (ms) measured during a walkthrough (Target frametime was set to 20ms for desktop and 66ms for

PDA). Image quality was measured at each of the speciﬁed points (A-E).

(a) PM only (Desktop) (b) PM only (PDA)

Figure 5: Number of PM triangles rendered for desktop and PDA during the walkthrough. Maximum and minimum represent

lowest and highest resolutions for all objects combined. Current denotes the selected object quality at runtime.

(a) (b) (c) (d)

Figure 6: Desktop screenshots: (a) Full geometry (b) IBR (c) PM only (d) Mixed PM and IBR.

GRAPP 2008 - International Conference on Computer Graphics Theory and Applications

246

Table 2: SSIM quality measurements for desktop.

SSIM IBR PM MIXED

A 0.8756 0.9747 0.9804

B 0.9318 0.9865 0.9912

C 0.9278 0.9882 0.9921

D 0.9408 0.9935 0.9992

E 0.9266 1.0000 1.0000

Table 3: SSIM quality measurements for PDA.

SSIM IBR PM MIXED

A 0.9686 0.9775 0.9812

B 0.9696 0.9852 0.9855

C 0.9694 0.9884 0.9913

D 0.9791 0.9975 0.9977

E 0.9683 0.9996 0.9996

which we got a framerate of 0.41 fps (not shown) us-

ing full geometry rendering. Again, by applying one

of the other representation types, we can substantially

increase the framerate. Note that the target frametime

for this mobile application was set to 66ms. When

we compare the resulting frametimes graph of ﬁgure

4(b) to ﬁgure 4(a), we can see that on the PDA, it is

much more difﬁcult to stay close to the target frame-

time. This is a result of the inferior processing capa-

bilities of the mobile device. Furthermore, for both

the IBR and mixed scenarios, we see a sudden in-

crease in frametime at two periods during the walk-

through. These two periods consist of camera move-

ment, thereby resulting in the need for updates to the

RTM objects. When looking at ﬁgures 4(b) and 5(d)

we can see however that the mixed scenario is much

less affected during the second movement period be-

cause at this moment much less objects are visible

and rendered using the RTM technique than during

the ﬁrst period.

Another difference can be seen when comparing

the graphs in ﬁgures 5(c) and 5(d). On the PDA we

are never able to render the maximal object quality,

while on the desktop, objects near to the viewer are

always rendered at maximum resolution. Even so,

from table 3 and the screenshots presented in ﬁgure

7 it is obvious that by using the mixed representation

we can greatly increase image quality while rendering

at comparable framerates. Furthermore, notice that

when using the PM only approach, the application is

unable to bring the frametime down to the target fram-

etime, even by using the lowest resolution versions

for all object instances while the mixed approach can

render at the target frametime using higher resolution

versions for the near objects.

5 CONCLUSIONS

We presented our solution for determining when to

switch from standard geometry representations to

image-based representations during rendering of a

complicated scene. During preprocessing, an image

quality metric is used for determining at which dis-

tance the image-based representation provides sufﬁ-

cient image quality compared to the geometrical ver-

sion. These measurements are performed for a sub-

set of viewpoints on the viewing sphere and stored

for later use. We successfully tested the usefulness of

this metric during a test that was performed on both a

desktop and a mobile system.

With regard to future work, we will be looking

into other image metrics for determining the quality

difference between the images rendering using the

different representation types because we have no-

ticed that, even by using the SSIM metric, the im-

age metric can sometimes differ from the quality per-

ceived by users themselves.

ACKNOWLEDGEMENTS

This work has been done under IBBT project

A4MC3.

REFERENCES

Aliaga, D. G., Cohen, J., Wilson, A., Baker, E., Zhang, H.,

Erikson, C., III, K. E. H., Hudson, T., Sturzlinger, W.,

Bastos, R., Whitton, M. C., Jr., F. P. B., and Manocha,

D. (1999). MMR: an interactive massive model ren-

dering system using geometric and image-based ac-

celeration. In Symposium on Interactive 3D Graphics,

pages 199–206. ACM.

Aliaga, D. G. and Lastra, A. (1999). Automatic im-

age placement to provide a guaranteed frame rate.

In Rockwood, A., editor, Siggraph 1999, Computer

Graphics Proceedings, pages 307–316, Los Angeles.

Addison Wesley Longman.

Blake, E. (1987). A metric for computing adaptive detail in

animated scenes using object-oriented programming.

In Eurographics Conference Proceedings, pages 295–

307.

Chang, C.-F., Bishop, G., and Lastra, A. (1999). Ldi tree: A

hierarchical representation for image-based rendering.

In Rockwood, A., editor, Siggraph 1999, Computer

Graphics Proceedings, pages 291–298, Los Angeles.

Addison Wesley Longman.

Fujita, M. and Kanai, T. (2002). Hardware-assisted relief

texture mapping. Eurographics 2002 short paper pre-

sentation, pages 257–262.

OBJECTIVE QUALITY SELECTION FOR HYBRID LOD MODELS

247

(a) (b) (c) (d)

Figure 7: PDA screenshots: (a) Full geometry (b) IBR (c) PM only (d) Mixed PM and IBR.

Funkhouser, T. A. and S

equin, C. H. (1993). Adaptive dis-

play algorithm for interactive frame rates during visu-

alization of complex virtual environments. Computer

Graphics, 27(Annual Conference Series):247–254.

Funkhouser, T. A., Squin, C. H., and Teller, S. J. (1992).

Management of large amounts of data in interactive

building walkthroughs. In 1992 Symposium on Inter-

active 3D Graphics, pages 11–20.

Gobbetti, E. and Bouvier, E. (2000). Time-critical multires-

olution rendering of large complex models. Journal of

Computer-Aided Design, 32(13):785–803.

Hidalgo, E. and Hubbold, R. (2002). Hybrid geometric -

image based rendering. Computer Graphics Forum,

21(3).

Hoppe, H. (1996). Progressive meshes. Computer Graphics

(ACM SIGGRAPH ’96 Proceedings), 30(Annual Con-

ference Series):99–108.

Jehaes, T., Quax, P., and Lamotte, W. (2005). Adapt-

ing a large scale networked virtual environment for

display on a pda. In ACE ’05: Proceedings of the

2005 ACM SIGCHI International Conference on Ad-

vances in computer entertainment technology, pages

217–220, New York, NY, USA. ACM Press.

Jehaes, T., Quax, P., Monsieurs, P., and Lamotte, W. (2004).

Hybrid representations to improve both streaming

and rendering of dynamic networked virtual environ-

ments. In Proceedings of the 2004 International Con-

ference on Virtual-Reality Continuum and its Applica-

tions in Industry (VRCAI2004).

Maciel, P. W. C. and Shirley, P. (1995). Visual navigation of

large environments using textured clusters. In Sympo-

sium on Interactive 3D Graphics, pages 95–102, 211.

McMillan, L. and Bishop, G. (1995). Plenoptic modeling:

An image-based rendering system. SIGGRAPH 95

Conference Proceedings, pages 39–46.

Oliveira, M. M. and Bishop, G. (1999). Image-based ob-

jects. In Symposium on Interactive 3D Graphics,

pages 191–198.

Oliveira, M. M., Bishop, G., and McAllister, D. (2000). Re-

lief texture mapping. In Akeley, K., editor, Siggraph

2000, Computer Graphics Proceedings, pages 359–

368. ACM Press / ACM SIGGRAPH / Addison Wes-

ley Longman.

Parilov, S. and Stuerzlinger, W. (2002). Layered relief tex-

tures. Journal of WSCG, 10(2):357–364. ISSN 1213-

6972.

Rafferty, M. M., Aliaga, D. G., Popescu, V., and Lastra,

A. A. (1998). Images for accelerating architectural

walkthroughs. IEEE Computer Graphics and Appli-

cations, 18(6):38–45.

Rossignac, J. and Borrel, P. (1993). Multi-resolution 3d ap-

proximations for rendering complex scenes. In Con-

ference on Geometric Modeling in Computer Graph-

ics, pages 455–465.

Shade, J., Lischinski, D., Salesin, D. H., DeRose, T.,

and Snyder, J. (1996). Hierarchical image caching

for accelerated walkthroughs of complex environ-

ments. Computer Graphics, 30(Annual Conference

Series):75–82.

Shade, J. W., Gortler, S. J., He, L.-W., and Szeliski, R.

(1998). Layered depth images. Computer Graphics,

32(Annual Conference Series):231–242.

Tack, N., Lafruit, G., Catthoor, F., and Lauwereins, R.

(2006). Platform independent optimization of multi-

resolution 3d content for enabling universal media ac-

cess. The Visual Computer, 22(8):577–590.

Wang, Z., Bovik, A. C., Sheikh, H. R., and Simoncelli, E. P.

(2004). Image quality assessment: From error visi-

bility to structural similarity. IEEE Transactions on

Image Processing, 13(4):600–612.

Zach, C., Mantler, S., and Karner, K. (2002). Time-critical

rendering of discrete and continuous levels of detail.

In VRST ’02: Proceedings of the ACM symposium

on Virtual reality software and technology, pages 1–

8, New York, NY, USA. ACM Press.

GRAPP 2008 - International Conference on Computer Graphics Theory and Applications

248