A QUALITATIVE EXPERT KNOWLEDGE APPROACH TO

RENDERING OPTIMIZATION

D. Vallejo-Fernandez, C. Gonzalez-Morcillo and L. Jimenez-Linares

Escuela Superior de Informatica, Universidad de Castilla-La Mancha, Paseo de la Universidad 4, Ciudad Real, Spain

Keywords:

Rendering, Qualitative Expert Knowledge, Fuzzy Systems, Fuzzy Sets, Multi-Agent Systems.

Abstract:

The rendering process allows the developer to obtain a raster 2D image from the deﬁnition of a 3D scene. This

process is computationally intensive if the source scene has a certain complexity or high-quality images are

required. Therefore, a lot of time is spent and many computational resources are needed.

In this paper, a novel approach called QUEKARO (standing for a QUalitative Expert Knowledge Approach to

Rendering Optimization) is presented for adjusting some relevant parameters involved in the rendering process

by using expert systems. This way, the developer can obtain optimized results which reduce the time spent in

the rendering process and, in most cases, do not affect the ﬁnal quality of the raster 2D image. These results

will be exposed on the result section, in which different optimizations will be studied.

As we discuss on the ﬁnal section of this paper, the use of expert systems in the rendering process involves

a novel approach which reduces drastically the resources used and provides us with a high-scalable system.

Using these arguments, we will justify the inclusion of expert systems in this area and will study future works.

1 INTRODUCTION

Fotorrealistic image synthesis can be interpreted as

the process which pursues the creation of synthetic

images which cannot be distinguished from the im-

ages obtained from the real world. This process is

divided into several phases such as modelling, set-

ting materials and textures, placing the virtual light

sources, and, ﬁnally, rendering.

The latter lies basically in generating a 2D image

from an abstract description, involving the geometry

of the scene, the deﬁnition of lights, the camera po-

sitions, and the use of materials. Rendering is usu-

ally the most computationally intensive phase of the

whole process and, therefore, it takes a long time to

be done. In addition, if the scene to be rendered is

complex or high-quality realistic images are required,

such process involves a lot of time.

To overcome this problem, the rendering phase

has experimented an important evolution. First, re-

searchs tried very hard to solve basic problems, such

as the detection of visible surfaces or the basic shad-

ing. Time passed and different rendering algorithms

were developed, ranging from simple and fast to more

complex and accurate, which simulate the light be-

haviour faithfully. Such methods are usually classi-

ﬁed into two main categories: local and global illu-

mination algorithms. Kajiya pointed out all rendering

algorithms aim to model the light behaviour over var-

ious types of surfaces and try to solve the rendering

equation (Kajiya, 1986), which forms the mathemati-

cal basis of all rendering algorithms.

As a result of the huge amount of time spent in

the rendering phase, it is often considered a bottle-

neck in photorealistic projects. In fact, the generation

of a single high-quality image may take several hours

up to several days, even on fast computers. In addi-

tion, the user normally sets the rendering parameters

so that the ﬁnal rendering time increases without ob-

taining substantial quality improvements, that is, the

user tends to optimize this process excessively.

This paper presents a novel approach called

QUEKARO based on the use of fuzzy systems to op-

timize the rendering phase. This way, ﬁnal results can

be improved by entities which use qualitative expert

knowledge over a distributed platform. With this de-

336

Vallejo-Fernandez D., Gonzalez-Morcillo C. and Jimenez-Linares L. (2007).

A QUALITATIVE EXPERT KNOWLEDGE APPROACH TO RENDERING OPTIMIZATION.

In Proceedings of the Ninth International Conference on Enterprise Information Systems - AIDSS, pages 336-341

DOI: 10.5220/0002361503360341

 SciTePress

sign we have obtained a robust, scalable, and ﬂexible

system which allows the ﬁnal user to optimize the ren-

dering phase without lossing perceptual quality of the

resulting image.

This paper is structured as follows. First, Section

2 overviews the state of the art related to rendering

optimizations. Section 3 describes QUEKARO, dis-

cussing the underlying architecture and studying in

depth the fuzzy system used in QUEKARO. Section

4 shows empirical results obtained with this approach

by using different conﬁgurations. Finally, section 5

resumes main conclusions and suggests future works.

2 RELATED WORK

The huge amount of time required by the rendering

process implies all related optimizations pursue the

reduction of the ﬁnal time spent. These optimizations

can be classiﬁed into three general types (or into a mix

of them) (Sundstedt et al., 2005):

1. Optimizations related to the use of powerful pro-

cessing elements (hardware optimizations).

2. Optimizations related to the use of distributed

systems (pararellel/distributed optimizations

(Chalmers et al., 2002)).

3. Optimizations related to the use of multi-agent

systems.

The ﬁrst type of optimization is related to the use

of powerful processing elements. This way, it is worth

pointing out that some researchers use programmable

GPUs (Graphics Processing Units) so that these pow-

erful processors run several portions of code of the

rendering algorithm. The main problem of this ap-

proach is the low-abstraction level used, that is, the al-

gorithms must be speciﬁcally designed for each hard-

ware architecture.

The second type of optimization is based on the

’divide and conquer’ principle. If we adapt this basic

principle for rendering, we can employ several com-

putational units to reduce the ﬁnal calculation time.

There are many related approaches in this research

line, such as the work made by Fernandez-Sorribes

et al. (Fernandez-Sorribes et al., 2006), which uses a

grid architecture for distributed rendering. However,

most of them are not able to achieve effective load

balancing and, therefore, the ﬁnal time required de-

pends on the most complex task.

The third type of optimization is based on the use

of multi-agent systems. In this ﬁeld, different propos-

als have been presented, such as the work exposed by

Rangel-Kuoppa et al. (Kuoppa et al., 2003). How-

ever, this approach does not make use of the multi-

agent technology because there are not interaction

mechanisms among agents. Another related work is

proposed by Schlechtweg et al. (Schlechtweg et al.,

2005), which makes use of a multi-agent system for

rendering artistic styles such as stippling and hatch-

ing.

2.1 Comparison to the Quekaro

Approach

The approach described in this paper is based on a

fuzzy system. This way, we can represent the qual-

itative expert knowledge by using a set of linguistic

rules, which allows each computational unit to adjust

a number of settings related to the rendering process.

Throught these adjustments, ﬁnal render times are re-

duced by solving the problem related to the user’s ten-

dency to use rendering parameters that are not opti-

mum, because these adjustments increase the render-

ing time without obtaining important quality beneﬁts.

QUEKARO has been satisfactorily used in MA-

gArRO, standing for a Multi-Agent Architecture for

Rendering Optimization. This multi-agent system

(designed under the principles of the set of FIPA

standards (Foundation for Intelligent Physical Agents

FIPA, 2002)) is composed of an indeterminate num-

ber of specialized agents, which compete for the pur-

chase of work units which make up the complete

work. A work unit can be interpreted as a part of a

scene or as a frame in an animation.

The approach used in QUEKARO implies sev-

eral important improvements regarding other research

lines:

• It supplies us with a high-level of abstraction due

to the use of a fuzzy system.

• The descriptive power of the fuzzy system used

allows the expert to model the knowledge relating

to rendering easily (Tanaka, 1998).

• The use of a fuzzy system combined with a multi-

agent system permits each agent to employ differ-

ent models of expert knowledge. This way, we

could employ different specialized agents by us-

ing different sets of rules.

• It complements the use of expert knowledge in

hardware or paralell rendering based alternatives.

3 PRESENTATION OF THE

QUEKARO APPROACH

As we mentioned in previous sections, we have used

a fuzzy system to optimize relevant parameters in-

A QUALITATIVE EXPERT KNOWLEDGE APPROACH TO RENDERING OPTIMIZATION

337

volved in the rendering process. The choice of sets

of rules as the method for modelling the qualitative

expert knowledge allows us to describe and extend

this knowledge easily. This way, the system allows

the user to incorporate different rendering methods,

such as Raytracing or bidirectional Pathtracing, each

one modelated by using a different set of rules. In

this paper, we will study the set of rules made for

the Pathtracing rendering method. Therefore, we will

have several entities specialized in this method. The

choice of this method is because it is one of the most

used, owing to its good rendering time-quality ratio.

Next, we will study the different elements which

compose QUEKARO. First, we must specify how

the rendering parameters are adjusted by using

QUEKARO. This is done by obtaining the values re-

lated to the output parameters of the ﬁred rules (the

consequents) and by using a method to adjust the ren-

dering parameters correctly. Second, we must deﬁne

how these rules are ﬁred. This is done by matching

the input parameters with the antecedents of the rules.

Last, we must deﬁne the set of rules which governs

the QUEKARO behaviour. All these elements will be

studied carefully.

The output parameters, which repay with the con-

sequents of the rules, are conﬁgured to decrease the ﬁ-

nal time required to ﬁnish the rendering process with-

out obtaining important losses of the perceptible qual-

ity. The linguistic labels used are VS (very small),

S (small), N (normal), B (big), and VB (very big)

(Zadeh, 1975). Next, we will explain the meaning

of these parameters:

• Recursion Level (Rl): this parameter is deﬁned

over the linguistic labels {VS, S, N, B, VB}. It

refers to the global recursion level in Raytracing,

that is, the number of light bounces.

• Light Samples (Ls): this parameter is deﬁned over

the linguistic labels {VS, S, N, B, VB}. It rep-

resents the number of samples per light on the

scene.

• Interpolation Band Size (Ibs): this parameter is

deﬁned over the linguistic labels {VS, S, N, B,

VB}. It deﬁnes the interpolation band size in pix-

els among adjacent work units, as shown in Figure

1. Moreover, this parameter has a relevant impor-

tance because it is used in the ﬁnal composition

process of the image.

We have used classic trapezoidal functions to de-

ﬁne the membership of fuzzy sets related to the output

variables, as we can see in Figure 2.

The output parameters deﬁned (recursion level,

light samples, and interpolation band size) have a

strong dependency with the Pathtracing rendering

Figure 1: Deﬁnition of the interpolation band size between

tasks.

method. However, the parameters related to the an-

tecedents of the rules can be used with other render-

ing methods, depending on their characteristics. We

will only need to change the descriptions of the rules.

Most of these parameters are directly related to the

characteristics of the scene:

• Complexity (C): this parameter is deﬁned over the

linguistic labels {VS, S, N, B, VB}. It represents

the complexity-size ratio of the work unit. This

complexity is calculated depending on parameters

such as the materials used to deﬁne the elements

which compose the model.

• Size (S): this parameter is deﬁned over the linguis-

tic labels {S, N, B}. It refers to the size of the

work unit in pixels.

• Neighbour Difference (Nd): this parameter is de-

ﬁned over the linguistic labels {VS, S, N, B, VB}.

It represents the complexity difference among the

current work unit and its neighbour work units.

To take this difference into account is important

not to make quality differences among neighbour

work units out.

• Optimization Level (Op): this parameter is de-

ﬁned over the linguistic labels {VS, S, N, B, VB}.

Its value is selected by the user, and it allows

him/her to apply different levels of optimization

(more or less agressive than the user predeﬁned

optimization).

The deﬁnition of the fuzzy sets related to the in-

put variables, with the exception of the optimization

level, is made dynamically, that is, in run-time. The

intervals of these sets are deﬁned depending on the

characteristics of the input scene and are made by lin-

ear distribution. For example, different intervals of

the complexity variable will be created depending on

the complexity of the scene. This way, we can fo-

cus on each problem independently. Another impor-

tant question is related to the inﬂuence of the output

variables in relation to the ﬁnal rendering parameters.

When we defuzziﬁcate and obtain crisp values for the

recursion level and the light samples, we do not use

ICEIS 2007 - International Conference on Enterprise Information Systems

338

Figure 2: Deﬁnition of the output variables.

this values directly in the rendering. On the contrary,

we weight these obtained values with the values pre-

deﬁned by the user. This way, we can apply local

optimizations so that the ﬁnal result does not show

important losses of the perceptible quality. In fact,

this is essence of QUEKARO, that is, to optimize the

rendering process depending on the values deﬁned by

the user.

Once we have studied the different variables, the

next step is to present the set of rules. We have used

if-then rules due to their resemblance to human rea-

soning. This set, which shows an expert’s knowledge

in Pathtracing, has 28 rules, but we will study some of

them, which represent the main types of rules used:

1. If C is {B, VB} and S is {B, N} and Op is VB

then Ls is VS and Rl is VS.

2. If C is {S, VS} and S is {B, N} and Op is VB then

Ls is S and Rl is S.

3. ...

4. If C is {B, VB} and Nd is B then Ls is VB.

5. If C is {B, VB} and Nd is N then Ls is B.

6. If Nd is VB then Ibs is VB.

7. If Nd is B then Ibs is B.

8. ...

The ﬁrst rule takes into account parameters related

to the complexity of the scene together with the opti-

mization level deﬁned by the user. As this optimiza-

tion level is very aggressive (VB), the output parame-

ters suffer a big reduction (VS). The fourth rule uses

the neighbour difference parameter to determine the

value of the light samples parameter. The rest of this

type of rules are focused on solving the problem re-

lated to the interpolation among work units. The sev-

enth rule is related to the second rule, and it adjusts

the interpolation band size depending on the neigh-

bour difference.

All the variables and all the rules have been de-

scribed by using the XML metalanguage. This way,

we can model QUEKARO easily by using the de-

scriptive power of XML. In our next research devel-

opment, we will only need to add some linguistic vari-

ables and some rules to use more rendering meth-

ods, as well as implementing the code which man-

ages these parameters. Next, we have a piece of our

description:

<?xml version=’1.0’ encoding=’UTF-8’?>

...

</linguisticvar>

...

</rule>

...

</system>

4 EXPERIMENTAL RESULTS

The results exposed here have been obtained by us-

ing QUEKARO with the MAgArRO system. All the

implementations related to these systems are avail-

able for download under the GPL Free Software Li-

cense (http://code.google.com/p/masyro06/). To test

the system presented in this paper we have used 8

computers with the same hardware characteristics.

Each of them has a Pentium Intel Centrino 1,6 GHz,

1 GB RAM, and the Debian GNU/Linux operating

system. The rendering method used in all cases was

the Pathtracing method, using the Yafray 0.0.9 render

engine, an 8 oversampling level, a 5 recursion level

in global conﬁguration for Raytracing, an image res-

olution of 1000x600 pixels, and 1024 light samples.

Making the rendering in one machine and using this

conﬁguration, a rendering time of 1 hour, 27 minutes,

and 27 seconds (mm:ss format from now on) was re-

quired to obtain the 2D image shown in Figure 4.a.

The most important parameter in the optimization

phase of this system is the optimization level param-

eter. Its value is deﬁned by the user depending on

A QUALITATIVE EXPERT KNOWLEDGE APPROACH TO RENDERING OPTIMIZATION

339

Figure 4: The result of the rendering without using optimizations (a) and using a small optimization level (b).

Figure 3: Rendering times by using different optimization

levels.

the level chosen, ranging from VS (minimum value)

to VB (maximum value). As well as managing this

parameter in our tests, we will play by using a vari-

able number of agents, trying to establish an optimal

conﬁguration (resources used and time needed). Ta-

ble 1 shows the time required to render the scene by

using different optimization levels (Figure 3). As we

can see, the rendering time spent with very small and

small optimizations in the case of using only an agent

are lightly bigger than the time obtained without op-

timizations. These negative results are as a conse-

quence of using interpolation bands and repeating the

rendering of some part of the work units,as well as

analyzing the input scene and composing the ﬁnal re-

sult.

Table 1: Results obtained with different number of agents

and different optimization levels.

Agents VS S N B VB

1 93:46 90:54 86:54 70:29 67:02

2 48:15 45:29 41:15 35:53 29:20

4 22:08 21:47 19:31 16:24 14:21

8 15:58 15:31 12:17 10:50 09:47

However, this situation can be got around because

our intention is to use this system with more than one

agent. In fact, excellent results are obtained when four

or more agents are used. For example, in the case of

applying a small optimization level (Figure 4.a and

Figure 4.b), the time required to render the scene is

just about 22 minutes by using 4 agents, whereas the

rendering time without using QUEKARO was about

87 minutes, that is, we have reduced the time to the

fourth. In addition, the time decreases if we use

more agents. However, the ﬁnal rendering time of the

scene always depends on the most complex work unit.

Moreover, it is important to compare this overall ap-

proach to the paralell computing approach in relation

to the rendering times required, because the ﬁrst ap-

proach uses a partition scheme which allows us to ob-

tain work units with a similar complexity. Whereas,

the last approach is extremely subject to the more

complex work unit.

As a ﬁnal remark, using aggressive optimizations

may result in a lose of quality. This problem can ap-

pear in several parts of the image (depending on the

scene) and can be avoided by using lighter optimiza-

tions.

4.1 Comparing Empirical Results

As well as comparing the time obtained using differ-

ent optimizations, it is also very important to compare

the quality of the different results. However, this qual-

ity is a subjective topic because there are many factors

which determine a ﬁnal evaluation. For example, a

non-expert person may think a 2D image resulting of

a rendering has an excellent quality although this re-

sult has several differences in relation to a rendering

without optimizations. Therefore, we could establish

a metric in which the evaluation of an expert person

is the reference when we are comparing empirical re-

sults. However, and to automate this evaluation, we

have deﬁned a metric based on the color difference

between pixels, that is, a result using optimizations

will be worse than another if this last has a lower dif-

ference than the ﬁrst regarding the result without op-

timizations.

Implementing a program which uses this metric

ICEIS 2007 - International Conference on Enterprise Information Systems

340

is quite easy by employing an API. Our implementa-

tion has the two 2D images to compare as input and

a gray-scale image as output. This gray-scale image

represents the difference between pixels of the input

images. A pixel of black colour represents there is

no difference between the pixels corresponding to the

other two images, that is, the minimum difference,

and a pixel of white colour represents there is the

maximum difference. In the Figure 4.c we can see

the difference of quality (depending on the proposed

metric) between the results exposed in the Figure 4.a

and the Figure 4.b. In this paper we have inverted

the colours so that the lector can appreciate the dif-

ferences between the two images clearly. In this case,

they only differ 1.21 %.

5 CONCLUSIONS AND FUTURE

WORKS

The obtaining of a system that allows the user to op-

timize the rendering process is an important step for-

ward because rendering times are usually very high.

In addition, the inclusion of QUEKARO into a multi-

agent system makes it specially important to tackle

the problems related to the rendering process. This

overall system can work along the Internet to create a

grid system which would be an efﬁcient tool to deal

with complex works.

The experimental results shown in the previous

section prove the inclusion of fuzzy systems opens

very promising research lines in the rendering prob-

lem. Moreover, the use of a fuzzy system has pro-

vided us with several advantages:

• It allows us to model expert knowledge easily.

• It provides us with a hightly scalable system.

• It provides us with an understandable human rep-

resentation.

An important problem to be solved, and one of

our main research line, is to use different rendering

techniques in different work units. Under this ap-

proach, we could use complex realistic techniques

only if needed, reducing the rendering time drasti-

cally. On the one hand, we may use the Pathtracing

method in a complex work unit which needs a high-

quality result because it contains an important part

of the model, such as a glass of wine. On the other

hand, we may use the Scanline method, which is a

much faster method, in those parts that do not require

a high-detail level, such as a wall. However, this idea

implies we must adjust the interpolation band size ef-

ﬁciently to obtain a high quality in the ﬁnal 2D image.

Another important research line is to use more

rendering parameters, such as the spacial resolution of

the image, the oversampling level, radiosity parame-

ters (patch size, number of elements, etc), the number

of photons, and so on. To include all these parame-

ters demands that we model QUEKARO carefully to

obtain high-quality results.

ACKNOWLEDGEMENTS

The authors would like to thank the anonymous re-

viewers for their insightful comments that have sig-

niﬁcantly improved the quality of this paper. This

work has been funded by the Junta de Comunidades

de Castilla-La Mancha under the Research Projects

PBC06-0064 and PAC-060141.

REFERENCES

Chalmers, A., Davis, T., and Reinhard, E. (2002). Practical

Parallel Rendering. AK Peters Ltd.

Fernandez-Sorribes, J. A., Gonzalez-Morcillo, C., and

Jimenez-Linares, L. (2006). Grid architecture for dis-

tributed rendering. In Proceedings of Ibero-American

Symposium in Computer Graphics 2006, pages 141–

148.

Foundation for Intelligent Physical Agents FIPA (2002).

FIPA Abstract Architecture Speciﬁcation. Founda-

tion for Intelligent Physical Agents FIPA, Geneva,

Switzerland.

Kajiya, J. T. (1986). The rendering equation. Computer

Graphics, 20(4):143–150. Proc. of SIGGRAPH’86.

Kuoppa, R. R., Cruz, C. A., and Mould, D. (2003). Dis-

tributed 3d rendering system in a multi-agent plat-

form. In Proceedings of the Fourth Mexican Inter-

national Conference on Computer Science (ENC’03),

pages 168–176.

Schlechtweg, S., Germer, T., and Strothotte, T. (2005). Ren-

derbots – multiagent systems for direct image genera-

tion. In Computer Graphics Forum, pages 137–148.

Sundstedt, V., Debattista, K., Longhurst, P., Chalmers, A.,

and Troscianko, T. (2005). Visual attention for efﬁ-

cient high-ﬁdelity graphics. In Spring Conference on

Computer Graphics (SCCG 2005), pages 162–168.

Tanaka, K. (1998). An introduction to fuzzy logic for prac-

tical applications. Springer.

Zadeh, L. A. (1975). The concept of a linguistic variable

and its applications to approximate reasoning. Part i,

ii, iii. Information Science.

A QUALITATIVE EXPERT KNOWLEDGE APPROACH TO RENDERING OPTIMIZATION

341