User-guided Modulation of Rendering Techniques for Detail Inspection

Ankit Sharma and Subodh Kumar

Department of Computer Sc. and Engg., Indian Institute of Technology Delhi, New Delhi, India

Keywords:

Feature based shading, Detail enhancement.

Abstract:

Understanding intricate details of carved models, for example ones prevalent in cultural heritage applications,

is often difﬁcult from renderings using traditional illumination models. A number of illustrative rendering

techniques are known, but each works well only for some models. We present a rendering system that com-

bines these techniques in an attempt to make the visualization more comprehensible given any context. In

particular, our system learns user’s visual preferences using exemplars from a domain and applies an appropri-

ate combination of the basis techniques to new meshes from that domain. Given a polygonal mesh, the system

applies different rendering techniques to different parts based on local features in order to enhance the overall

appearance.

1 INTRODUCTION

Perception of shape and surface details from com-

puter generated renderings of 3D objects is of sig-

niﬁcant interest in such applications as the study of

ancient artifacts and archaeology. Traditional illumi-

nation models, e.g., Lambertian, Phong, sub-surface

scattering, etc., can wash out ﬁne details or make

them hard to recognize in many cases. In the con-

text of archaeological relic illustration, it is crucial

that people be able to study and decipher the engrav-

ings. Techniques like stippling (Deussen et al., 2000)

are quite useful for an overall aesthetic view, but can

also mask some ﬁne details. Inspired by traditional

hand drawing, many rendering techniques have fo-

cused on determining an appropriate set of lines to

depict shape. In contrast, other techniques mainly

use shading, i.e., intensity gradients across the sur-

face. The most popular of these simulate lighting and

occlusion shading. Still others combine both (Wang

et al., 2010). The intent is to highlight features more

than to simulate photo-realism.

The importance of illustrative rendering is well

recognized (Bartz et al., 2005). As a result, a va-

riety of techniques have been proposed in the liter-

ature, each with its own strengths and weaknesses.

Different methods are suitable for different scenar-

ios, which may be hard to characterize. A versatile

‘master’ shading technique that caters to a wide va-

riety of 3D models remains elusive. We instead in-

vestigate ways to choose, or combine, techniques that

may suit a given situation and exaggerate features of

choice. The main problem then is to determine the

most suitable combination of known detail enhance-

ment techniques. Complications arise because usabil-

ity depends not only on the object structure but also

on the purpose and users’ preferences, which cannot

be expressed objectively. In most cases, an intuitive

deﬁnition may be available but a rigorous set of re-

quirements may not be established. There has been

research on automatic tweaking of rendering param-

eters for creating better views using entropy based

methods for increasing visual information recovery

(Gumhold, 2002; Takahashi et al., 2005; V

azquez and

Sbert, 2003; Wang and Shen, 2011)

We propose to learn the user’s notion of ‘good ren-

dering’ and relate it to the geometry of a surface. We

have attempted two methodologies. Both start with

a user provided library of basis shading techniques:

{T }. The ﬁrst creates a set of possible renderings

by generating a parametrized set of canonical shapes

and then shading each using a combination of tech-

niques from T . This massive collection of images

is then pruned by discarding images that are likely

to convey less visual information in terms of image

entropy (Wang and Shen, 2011). The users then se-

lect from this reduced set, the images that conform

to their notion of being ‘visually good’. The sec-

ond method is model-centric. It renders user-provided

sample polygonal models using different combina-

tions in different parts of the model. The user selects

regions that are satisfactory. The ﬁrst method has the

247

Sharma A. and Kumar S..

User-guided Modulation of Rendering Techniques for Detail Inspection.

DOI: 10.5220/0004652802470254

In Proceedings of the 9th International Conference on Computer Graphics Theory and Applications (GRAPP-2014), pages 247-254

ISBN: 978-989-758-002-4

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

potential to be more general, but in our experience

the second produces more useful renderings (see Sec-

tion 3.4). In each method, the collected data is ﬁnally

used to predict the appropriate rendering technique

for each vertex of any new mesh.

2 RELATED WORK

Non-photorealistic rendering techniques like stylized

shading and feature line drawings have been used to

depict surface detail (Wang et al., 2010). Interactive

systems that permit the user to change the lighting

and view direction for better exploration have been

designed (Halle and Meng, 2003) as well. Line draw-

ings do enhance feature detail but by themselves do

not provide complete visualization. We instead focus

on shading techniques and describe the most relevant

work next.

2.1 Light Source Placement

Proper lighting is crucial to comprehension of shape,

depth, and orientation. Improper light source place-

ment, for instance, can mislead us into thinking that a

convex object is concave, or vice versa. Thus effective

light source placement is well recognized for percep-

tually enhanced rendering. Two classes of approaches

are common: inverse lighting and information maxi-

mization.

Inverse lighting methods assume that the user has

prior knowledge of the shape and material properties

of the objects and speciﬁes how the object should ap-

pear. The algorithm then automatically computes the

light positions and intensities using a conﬁguration

optimization framework (Costa et al., 1999), or a di-

rect speciﬁcation of highlights and shadows (Poulin

and Fournier, 1992; Poulin et al., 1997). Monte-Carlo

method has also been used (Jolivet et al., 2002) for se-

lection of light positions according to a user-deﬁned

declarative model. These are powerful methods but

require extensive user involvement in the visualiza-

tion design.

Information maximization methods try to position

light sources in order to maximize the ‘information’

revealed to the user. Gumhold (Gumhold, 2002)

presents a method using this approach that uses an

entropy-based function, the lighting entropy. The in-

formation content of n random variables taking up

values from {v

,...,v

} is given by Shannon’s

source coding theorem as:

H = −

∑

i=1

logp

(1)

For a given illumination of a scene viewed with n cov-

ered pixels, the probabilities p

are computed from

the fraction of color values falling into the i

bin.

The lighting entropy is then calculated using equa-

tion 1. Information maximization approaches rely on

maximum entropy measures or perception-based op-

timization to position light sources. Their technique

applies to static models but does not address well in-

teractive object inspection. Lights need to be reposi-

tioned as camera moves to maximize the quality met-

ric. This repositioning can be distracting (Halle and

Meng, 2003).

We employ this technique. The light source is

placed at different positions on a bounding sphere of

the object and its entropy is measured. The position

with the highest entropy is selected (refer Figure 1).

This method is general and can be applied to calcu-

late entropies for any shading algorithm, as the calcu-

lations are performed on the resulting image.

Figure 1: Entropy based Light Source Placement: Compar-

ison of an image with low entropy (left) and high entropy.

2.2 Shadows and Occlusion Shading

In the context of rendering of statues and such other

artifacts, a local lighting model is not satisfactory.

Shadows and darkening of inaccessible areas is useful

(Anderson and Levoy, 2002). A related technique is

ambient occlusion.

Shadow is determined by several factors simulta-

neously: the direction of the light, the shape of the

object, the surface relief on which it falls as well as

the relative position of the light source, the object and

the receiving surface. The human visual system can

recover the object’s shape given all the other parame-

ters (Cavanagh and Leclerc, 1989).

Ambient Occlusion is widely used for shape de-

piction through shading (Vergne et al., 2011). It mea-

sures at each point, the fraction of hemisphere direc-

tions that are occluded. This value is used to mod-

ulate the diffuse shading term. The result is that the

crevices of the model are darkened, and the exposed

GRAPP2014-InternationalConferenceonComputerGraphicsTheoryandApplications

248

parts of the model appear brighter. The ambient oc-

clusion shading model offers a better perception of the

3D shape of the displayed objects. Perceptual exper-

iments show (Langer and B

uLthoff, 2000) that depth

discrimination under diffuse uniform sky lighting is

superior to that predicted by a direct lighting model.

In the context of archaeological models with high fre-

quency and detailed carved surface features, shadows

are hardly useful in enhancing perception. Ambient

occlusion on the other hand works well in most cases.

Figure 2: Ambient Occlusion (left and Mean Curvature

Shading (right)).

2.3 Stylized Shading

Stylized shading has been used to depict shape by ex-

aggerating surface details (Rusinkiewicz et al., 2006).

Simple modiﬁcations to the surface normals can fur-

ther enhance (Cignoni et al., 2005) geometric fea-

tures of an object. A multi-scale Lambertian shad-

ing of the models has also been applied (Rusinkiewicz

et al., 2006). They use successively smoother geome-

try with each shading pass and employ multiple local

lights for illumination. The multi-scale renderings are

weighed using user-tunable parameters and combined

together to produce the ﬁnal rendering. Kindlmann et

al. use a simpler approach (Kindlmann et al., 2003).

The 3D mesh colors are scaled by the value of the

mean curvature at each vertex (see Figure 2). The

technique highlights such surface features as ridges,

valleys, saddles etc. on the surface.

3 AUTOMATIC PARAMETER

SELECTION FOR OPTIMAL

RENDERING

Although we have discussed a few shading tech-

niques, many more are possible, including those yet

to be discovered. There isn’t a straightforward way

for an application designer to choose the technique

most effective in enhancing detail in that application.

Worse yet, different parts of a model may require dif-

ferent techniques. We instead are interested in ways

to explore multiple techniques in a common frame-

work and then seek a suitable combination on a per

vertex basis. This is our main contribution and this

section explains our approach.

3.1 Motivation

Comprehensibility depends on the object, the pur-

pose, and the users’ preferences, and cannot be easily

expressed mathematically. Therefore, we often ﬁnd

the best combination by trial and error for each object

or purpose. Little prior research has been reported

that addresses this problem: automatically yet mean-

ingfully choose the combination of a set of rendering

techniques to apply. The techniques discussed in Sec-

tion 2 are meant for shape depiction. Our tests also

focus on shape depiction, but our general framework

is useful for a variety of applications and a variety of

techniques suitable for that context.

3.2 Approach

Our method employs a supervised learning approach

to learn users’ preferences. Important issues include:

• What values should be learned.

• What surface features and other properties do the

values depend on.

• How should training be effected.

We explore this space. We report two related tech-

niques, one that directly learns the intensity values at

the vertices of a mesh and another that learns the vi-

sualization techniques to be used.

We begin by choosing a library of basis rendering

techniques for shape depiction. Our selection is based

only on intuition but in general, a well researched set

can be included; our framework is not speciﬁc to the

set. While rendering a 3D mesh, the vertex color pro-

duces by each rendering technique mainly depends on

the geometry. We combine these geometric factors

into a vertex feature vector. Consequently, the best

technique at each vertex is deemed to be a function

of its feature vector. We acquire a training data set

consisting of feature vectors and their corresponding

techniques obtained via user input. Then we apply an

appropriate machine learning algorithm to predict the

technique that should be used for a new feature vector.

As a simpliﬁcation, we also try to directly learn a

per-vertex diffuse color instead of the technique and

apply the Lambertian model. We describe our method

in more detail next.

User-guidedModulationofRenderingTechniquesforDetailInspection

249

3.2.1 Shading Techniques

For our purpose, we have selected the following tech-

niques, which appear better suited to ‘archaeological’

models, our domain of interested.

Diffuse and Specular Lighting: We apply the

commonly used Lambertian diffuse shading model

and Blinn-Phong specular shading model.

Curvature Based Shading: Local surface geometry

can be adequately described using the principal cur-

vatures (κ

and κ

) at the mesh vertices.

The diffuse color intensity at each vertex is scaled

by the mean curvature value and normalized.

Ambient Occlusion: The vertex color intensity is

scaled by the ambient occlusion.

Entropy: We use entropy as a measure of infor-

mation content in our visualization. High entropies

usually imply perceptually better renderings (Section

2.1).

3.2.2 Vertex Feature Vector

The techniques enumerated in section 3.2.1 compute

the intensity at each vertex of the 3D model, which

depends on the occlusion factor, the surface curvature

and the signed normal at the vertex as well as the view

direction and the light direction. We represent this

per-vertex data as a feature vector. This vertex feature

vector is deﬁned as:

v = {θ,φ, κ

,κ

,o} (2)

where,

θ = angle between light direction and surface nor-

mal at the vertex

φ = angle between view direction and surface nor-

mal at the vertex

,κ

= principal curvatures at the vertex, maxi-

mum and minimum

o = precomputed ambient occlusion factor

To capture the local context, we augment the vertex

feature vector with neighborhood mean curvature val-

ues mc

, computed as follows. Let ~v

and ~v

be the

principal directions at a given vertex that deﬁne a 2D

coordinate system X spanning plane P. All vertices

within a λ-ring neighborhood of the given vertex are

projected on P. (We choose λ = 2.) Then mc

is the

average mean-curvature value of the vertices that have

projections lying in the i

quadrant of X.

The augmented vertex feature vector is deﬁned as:

v = {φ,κ

,κ

,o, mc

,mc

}. (3)

Note that we have removed the angle between the

light direction and the surface normal at the vertex

as a feature in the augmented vertex feature vector.

Instead, the per-vertex light direction itself is learned

via entropy maximization.

The intensity at a vertex is a function of the ver-

tex feature vector, i.e., I = f (v). Formally, one can

directly learn f or one can learn technique h, such

that I = g ◦ h(v), where g is the universal algorithm to

compute the intensity, given h.

3.3 Obtaining the Data Set for Training

Creating the training data-set is a non-trivial under-

taking and the number of possibilities are potentially

immense. In principle, a cross product of all possi-

ble surface properties and all possible combinations

of rendering techniques have to be observed to select

the best among them. We describe our interface and

the process to simplify the task.

First the set T = {t

},1 ≤ i ≤ n, of n rendering

techniques are selected by a domain expert. Each

combination is then represented by a vector (s

..s

)

chosen such that the ﬁnal intensity of a point on the

surface is learned as H =

∑

i=n

i=1

, where

∑

i=n

i=1

= 1

and intensity C

is obtained if technique t

is used for

that point. The system then learns scalars s

. For our

purpose we have chosen seven techniques represented

by their indices as shown in Table 1.

Table 1: Class Identiﬁers for Different Techniques used for

Support Vector Classiﬁcation

Class (T) Technique

1 Ambient Occlusion

2 Diffuse Lighting

3 Mean Curvature Shading

4 Ambient Occlusion + Diffuse Lighting

5 Diffuse Lighting + Mean Curvature Shad-

ing

6 Ambient Occlusion + Mean Curvature

Shading

7 Ambient Occlusion + Diffuse Lighting +

Mean Curvature Shading

We found that it is not beneﬁcial to directly learn

the color. In particular, often the inherent reﬂectance

properties of the surface is known. One can instead

learn a modulating parameter. For example, in our

experiments we learn an appropriate scale for the re-

ﬂectance. In other words the diffuse reﬂectance of the

surface is scaled by the learned value H.

To explore the parameter space, we have experi-

mented with two techniques. In the ﬁrst, the user is

required to select a representative model. The trainer

is then presented with a sequence of candidate ren-

GRAPP2014-InternationalConferenceonComputerGraphicsTheoryandApplications

250

derings of the model to choose from. In the second

technique, the system generates a parameterized set

of canonical shapes rendered using various candidate

techniques. The canonical shape is generated around

a central vertex with the desired features, In this case

the choice is binary: if a visualization is selected, the

corresponding central vertex is added to the training

set. In the ﬁrst case, the trainer is allowed to use

a brush interface to select the regions of the model

where the visualization is satisfactory. All selected

vertices – their features that is – are added to the train-

ing set in one shot.

In either case, the number of candidates is inﬁnite.

Hence the system must sample the parameter space in

a structured fashion and ﬁlter out cases not likely to

be “perceptually” attractive. For the canonical model

technique, we sample the parameter space uniformly.

In this sense the ﬁrst technique is likely to be more ex-

haustive, although the second one attempts to restrict

attention to the more relevant ones and also presents

a more “in-context” non-local perception. (Figure 4

shows a screen-shot of our interface). To further ﬁlter

the set of parameters presented to the user, we employ

the entropy method. Renderings that do not meet an

entropy threshold are not presented to the trainer.

Figure 3: User Interface: Allows user to set light to

achieve maximum entropy and linearly combine different

techniques by assigning them weights and then clicking on

“Recompute Colors”

The light and view direction vectors are uniformly

distributed over the surface of a unit sphere centered

at the origin. To cover the range of possible com-

binations of light and view directions, we render the

models while varying these directions over a uniform

sampling of the unit sphere. The light is positioned in

order to maximize entropy.

3.3.1 Smoothing

We do not expect the trainer to be a visualization ex-

pert, but rather a domain expert or even a novice user.

Furthermore, since the training data is manually gen-

erated and the process can take an hour or even more,

it is possible to generate conﬂicting training data. In

particular, for the same or similar features multiple

Figure 4: Feature Selection User Interface with Test Object:

Four different renderings can be viewed, bottom right shows

blue selected vertices.

techniques may be chosen over the course of the train-

ing. To ensure consistency, we employ a smoothing

pass to the data. This is done by assigning weights

to the feature-technique correspondence. For exam-

ple, if the same feature occurs twice in the training-set

with different learned scalars, we reduce the weight of

each rule symmetrically so that they sum to 1.0. We

further use a similarity threshold ε to ensure that sim-

ilar features with differing rules have proportionally

reduced conﬁdence. A function that achieves this is

as follows. For a feature f if there exists a set of other

features { f

},1 ≤ i ≤ k such that d

= | f − f

| < ε,

where | | denotes the L

norm, we assign f a weight

w =

∑

3.3.2 Learning and Prediction

We have used Support Vector Machines (SVM) pro-

vided in libSVM (Chang and Lin, 2011) for regres-

sion and classiﬁcation.

In our regression approach, the intensity (I) at

a vertex is assumed to be a function (f ) of the

vertex feature vector (v), i.e., I = f (v) where I ∈

[0,1] , v = {θ, φ,c

,o} ∈ R

and f : R

→ [0, 1]

in the basic feature vector case, for example.

A plot of the feature space is shown in Figure 5

which suggests a linear model may be sufﬁcient. We

use ε-SVR (Support Vector Regression (Chang and

Lin, 2011)) to estimate the linear function f. For each

visible vertex of the 3D model to be visualized, an in-

tensity value is predicted by the SVM. At the render-

ing time, the base diffuse component of each vertex

is scaled by this intensity value to produce the visual-

ization.

We also employ a simpliﬁed classiﬁcation only

approach, where only the best suited technique for a

User-guidedModulationofRenderingTechniquesforDetailInspection

251

Figure 5: Feature Space of Learning Data Set for Armadillo

model (see Figure 4): The three axes are the principal cur-

vatures c

, c

and the θ, the angle between the light direc-

tion and the surface normal. The color of a feature point

corresponds to a particular intensity.

given model is predicted. In other words, only one

of the scalars is allowed to be 1.0 and the others set

to 0. For this classiﬁcation, we employ C-SVC (Sup-

port Vector Classiﬁers (Chang and Lin, 2011)) rather

than the regression. The technique number (T, refer

Table 1) is the class label. The classiﬁer provides a

function (f ) that gives the class number for a vertex

feature vector, i.e., T = f (v) where T ∈ {1..7} , v =

{θ,φ, c

,o} ∈ R

and f : R

→ {1..7}.

At the rendering time, the technique number is

computed for each visible vertex of the 3D model and

the ﬁnal color is computed accordingly.

3.4 Results

For the ﬁrst experiment, we use the canonical test

objects (see Figure 6) to obtain the dataset with the

standard vertex feature vector. Only mean curvature

shading and diffuse lighting are used. Figure 6 shows

3 views that produce high entropies with the light

placed directly pointing into the plane of the paper.

Figure 6: Simple Test Object: Mean Curvature Shading and

Diffuse Lighting.

The regression approach is then applied to obtain

a visualization of the Armadillo model (shown in Fig-

ure 7).

For the second experiment, the trainer selects parts

of the test model that appear better with a partic-

ular technique. The test model is the Armadillo

shown in Figure 4. All three techniques viz. diffuse

lighting, mean curvature shading, ambient occlusion

and their combinations are explored (refer Table 1).

Figure 7: Regression Based Approach Results for Ar-

madillo 3D Model. Four renderings are shown, separating

the left and the right halves for better comparison: from the

left – Ambient Occlusion, Our Approach, Diffuse Lighting

and Mean Curvature Shading. Training set is derived from

canonical models as shown in Figure 6.

Classiﬁcation based learning approach is used. The

learned scalar values are then applied to visualize the

Cuneiform Tablet as shown in Figure 8.

Figure 8: Classiﬁcation Approach Results for Cuneiform

Tablet 3D Model: Our approach (top left), Ambient Occlu-

sion (top right), Diffuse Lighting (bottom left), Mean Cur-

vature Shading (bottom right). Training data obtained from

Armadillo model in Figure 4.

Table 2 lists the time taken to predict vertex color

values using the two approaches.

Table 2: Results: Time taken vs number of vertices in 3D

model and number of feature vectors in training data set.

The number of Support Vectors (SV) is also listed.

Training

Dataset

Size

Num

Number

of Ver-

tices

Time(s)

Regression

(Fig. 7)

310005 297775 172974 874

Classiﬁcation

(Fig. 8)

1015 694 1861168 39

We next repeat the second experiment using the

augmented vertex feature vector. The light position is

GRAPP2014-InternationalConferenceonComputerGraphicsTheoryandApplications

252

Figure 9: Classiﬁcation using Augmented Feature Vector

applied to Cuneiform2. We have applied the scalars learned

from the Armadillo model (Figure 4) to visualize three dif-

ferent models. Four images are shown for each model (two

in the next two ﬁgures). Our approach is used for each im-

age on the top left, Ambient Occlusion is used on the top

right, Diffuse Lighting+AO+Mean Curvature on the bottom

left, and Mean Curvature Shading on bottom right.

learned via entropy maximization. The visualizations

are shown in Figures 9 to 11.

We can see that even though the individual steps

of our technique can be improved, the visualizations

it produces are meaningful. Augmented feature vec-

tor generally outperforms the basic vector. An infor-

mal study of ten graduate students is indicative of the

method’s perceptual effectiveness. Seven renderings

Figure 10: Classiﬁcation using Augmented Feature Vector

applied to 3D Mural. Our approach is on top left.

Figure 11: Classiﬁcation using Augmented Feature Vector

applied to Cuneiform3. Our approach is on top left.

of four models (three Cuneiform tablets and one Mu-

ral) were shown to each user. Seven were chosen from

the library of techniques and one was the results of the

learned technique. They were asked to assign a score

between 1 and 5 to each. Although the learned tech-

nique did not score the highest mark in each of the 40

cases, it was ranked the highest in 31 cases.

4 CONCLUSION AND FUTURE

WORK

We propose a novel way to inspect cultural artifacts

with the aid of machine-learning techniques. An auto-

mated approach to combine multiple rendering tech-

niques has been presented. The approach is promis-

ing: as the survey shows, a per-vertex combination

of rendering techniques can outrank each individual

component applied globally in a model. Further, the

weights of basis techniques vary substantially from

vertex to vertex. It is hard to choose this manually

on a per vertex basis. Our approach uses supervised

machine learning to learn users’ preferences and pre-

dict shading values for new models. The advantages

of this approach are:

• Using only a few test models, the approach gives

reasonably good results for new models.

• The technique can capture non-local context as

the users’ notion of a good rendering is based on

the overall perception of the complete model.

• Due to per-vertex computations being carried out,

each part of the model gets optimally shaded for

each view conﬁguration.

User-guidedModulationofRenderingTechniquesforDetailInspection

253

Our results show that learning based visualization

is a promising approach, even if the technique learned

by our current algorithm is not always the best in

our experiments. We have presented only early re-

sults and there is much scope for further study in

this direction. There is a need for to devise a theo-

retical framework for determining useful parameters

to learn. Algorithmic work is also required to al-

low faster computation of view-dependent training re-

sults for interactive manipulation. Also, other render-

ing techniques like specular lighting and cast shadows

could be added to the learning set. A more efﬁcient

training set generation would also be useful in making

the technique user friendly.

ACKNOWLEDGEMENTS

We thank the Department of Science and Technology

for funding this research and Lissy Verma for imple-

menting several basis rendering techniques. We also

thank the reviewers for helping improve the presenta-

tion of the paper.

REFERENCES

Anderson, S. and Levoy, M. (2002). Unwrapping and vi-

sualizing cuneiform tablets. Computer Graphics and

Applications, IEEE, 22(6):82 – 88.

Bartz, D., Hagen, H., Interrante, V., Ma, K.-L., and Preim,

B. (2005). Illustrative rendering techniques for visu-

alization: Future of visualization or just another tech-

nique? In Visualization, 2005. VIS 05. IEEE, pages

715–718.

Cavanagh, P. and Leclerc, Y. (1989). Shape from shadows.

Journal of Experimental Psychology: Human Percep-

tion and Performance, 15:13–27.

Chang, C.-C. and Lin, C.-J. (2011). Libsvm: A library

for support vector machines. ACM Trans. Intell. Syst.

Technol., 2(3):27:1–27:27.

Cignoni, P., Scopigno, R., and Tarini, M. (2005). A

simple normal enhancement technique for interactive

non-photorealistic renderings. Computer & Graphics,

29(1):125–133.

Costa, A. C., De Sousa, A. A., and Ferreira, F. N. (1999).

Lighting design: A goal based approach using opti-

mization. In Lischinski, D. and Larson, G. W., editors,

Rendering Techniques, pages 317–328. Springer.

Deussen, O., Hiller, S., van Overveld, C., and Strothotte,

T. (2000). Floating points: A method for computing

stipple drawings. Computer Graphics Forum, 19:40–

51.

Gumhold, S. (2002). Maximum entropy light source place-

ment. In Visualization, 2002. VIS 2002. IEEE, pages

275 –282.

Halle, M. and Meng, J. (2003). Lightkit: a lighting system

for effective visualization. In Visualization, 2003. VIS

2003. IEEE, pages 363 –370.

Jolivet, V., Plemenos, D., and Poulingeas, P. (2002). Inverse

direct lighting with a monte carlo method and declar-

ative modeling. In Proceedings of the International

Conference on Computational Science-Part II, ICCS

’02, pages 3–12, London, UK, UK. Springer-Verlag.

Kindlmann, G., Whitaker, R., Tasdizen, T., and M

oller,

T. (2003). Curvature-based transfer functions for di-

rect volume rendering: Methods and applications.

In Proceedings of the 14th IEEE Visualization 2003

(VIS’03), VIS ’03, pages 67–, Washington, DC, USA.

IEEE Computer Society.

Langer, M. S. and B

uLthoff, H. H. (2000). Depth discrim-

ination from shading under diffuse lighting. Percep-

tion, 29(6):649–660.

Poulin, P. and Fournier, A. (1992). Lights from highlights

and shadows. In Proceedings of the 1992 symposium

on Interactive 3D graphics, I3D ’92, pages 31–38,

New York, NY, USA. ACM.

Poulin, P., Ratib, K., and Jacques, M. (1997). Sketching

shadows and highlights to position lights. In Proceed-

ings of the 1997 Conference on Computer Graphics

International, CGI ’97, pages 56–, Washington, DC,

USA. IEEE Computer Society.

Rusinkiewicz, S., Burns, M., and DeCarlo, D. (2006).

Exaggerated shading for depicting shape and detail.

ACM Trans. Graph., 25(3):1199–1205.

Takahashi, S., Fujishiro, I., Takeshima, Y., and Nishita, T.

(2005). A feature-driven approach to locating optimal

viewpoints for volume visualization. In Visualization,

2005. VIS 05. IEEE, pages 495 – 502.

azquez, P.-P. and Sbert, M. (2003). Perception-based il-

lumination information measurement and light source

placement. In Proceedings of the 2003 international

conference on Computational science and its appli-

cations: PartIII, ICCSA’03, pages 306–316, Berlin,

Heidelberg. Springer-Verlag.

Vergne, R., Pacanowski, R., Barla, P., Granier, X., and

Shlick, C. (2011). Improving shape depiction under

arbitrary rendering. IEEE Transactions on Visualiza-

tion and Computer Graphics, 17(8):1071–1081.

Wang, C. and Shen, H.-W. (2011). Information theory in

scientiﬁc visualization. Entropy, 13(1):254–273.

Wang, S., Cai, K., Lu, J., Liu, X., and Wu, E. (2010). Real-

time coherent stylization for augmented reality. The

Visual Computer, 26(6-8):445–455.

GRAPP2014-InternationalConferenceonComputerGraphicsTheoryandApplications

254