Obscurance-based Volume Rendering Framework.

Full text

Page 1

IEEE/ EG Symposium on Volume and Point-Based Graphics (2008)
H.- C. Hege, D. Laidlaw, R. Pajarola, O. Staadt (Editors)
Obscurance-based Volume Rendering Framework
M. Ruiz1, I. Boada1, I. Viola2, S. Bruckner3, M. Feixas1, and M. Sbert1
1Graphics and Imaging Laboratory, University of Girona, Spain
2Department of Informatics, University of Bergen, Norway
3Institute of Computer Graphics and Algorithms, Vienna University of Technology, Austria
Abstract
Obscurances, from which ambient occlusion is a particular case, is a technology that produces natural-looking
lighting effects in a faster way than global illumination. Its application in volume visualization is of special interest
since it permits us to generate a high quality rendering at a low cost. In this paper, we propose an obscurance-
based framework that allows us to obtain realistic and illustrative volume visualizations in an interactive manner.
Obscurances can include color bleeding effects without additional cost. Moreover, we obtain a saliency map from
the gradient of obscurances and we show its application to enhance volume visualization and to select the most
salient views.
Categories and Subject Descriptors (according to ACM CCS): I.3.7 [Computer Graphics]: Three-Dimensional
Graphics and Realism
1. Introduction
Global illumination is a well-known technique for produc-
ing realistic scenes. However, although it might play a de-
cisive role in 3D volume visualization since it provides vi-
sual cues that enhance data interpretation, its application is
still challenging in direct volume rendering. The main lim-
iting factor is the computational cost of simulating global
effects of light in a volume, making interactive exploration
difficult [Max95]. On the other hand, illustrative methods
aim at creating visualizations which convey information to
the viewer as opposed to physically correct light interaction.
Volume illustration enhances the expressiveness of volume
rendering by highlighting important features within a vol-
umewhilesubjugatinginsignificantdetailsandrenderingthe
result in a way that resembles an illustration [ER00] . Ide-
ally, a volume rendering system should be able to support
both realistic and illustrative renderings.
Obscurances have been introduced by Zhukov et
al. [ZIK98] and Iones et al. [IKSZ03] as an efficient tech-
nique that gives perceptually similar results to global illu-
mination with a small fraction of the computational cost.
Moreover, without adding computational cost, obscurances
also allow us to compute color bleeding, which consists in
the effect that the objects around another object with intense
coloration are dyed with this color [MSC03]. The obscu-
rance technique was first used in video-game environments.
Its application to volume rendering, called vicinity shading,
was introduced by Stewart [Ste03].
Inthispaper,wepresentanobscurance-basedvolumeren-
dering system that allows to obtain realistic and illustrative
volume visualizations in an interactive manner. One impor-
tant aspect of our work shows that obsurances are not only
useful for realistic depiction but also for illustrative render-
ing. As obscurances can be interpreted as general informa-
tion about the neighborhood of a voxel, they can be used as
a bias for the generation of more expressive illustrative de-
pictions of a data set (see Figure 1).
Saliency typically arises from contrasts between items
and their neighborhood [IK01,TIR05,vdWGB06] and it is
considered that the most salient voxels in a 3D data set will
attract the attention of the viewer. In our approach, voxel
saliency is determined by the obscurance gradient, which
measures the maximum variation of the obscurance field.
Once the saliency of the volume is obtained, we implicitly
have the saliency map of any structure contained in the vol-
ume. This saliency map can be applied to viewpoint selec-
tion and to enhance visualization. This can help to discover
relevant characteristics of the model otherwise unnoticed by
the observer.
c ? The Eurographics Association 2008.

Page 2

M. Ruiz, I. Boada, I. Viola, S. Bruckner, M. Feixas, and M. Sbert / Obscurance-based Volume Rendering Framework
(a) (b)(c)
Figure 1: CT-human body data set rendered with the pro-
posed obscurance-based volume rendering framework. The
images have been obtained by modifying interactively the
transfer function and the way in which obscurances are ap-
plied to the model.
2. Background
In this section, obscurances, ambient occlusion, and related
illumination models are described.
2.1. Obscurances and ambient occlusion
Zhukovetal.introducedambientocclusionwiththetermob-
scurances [ZIK98,IKSZ03]. Roughly speaking, obscurance
measures the part of the hemisphere obscured by the neigh-
boring surfaces. For instance, a corner of a room is more
obscured than the center. From the physics of light trans-
port, obscurance expresses the lack of secondary (reflected)
light rays coming to the specific parts of the scene, thus mak-
ing them darker. Computation was done as a preprocess and
the obscurance values were used as an ambient term during
rendering. Since the obscurance computation was a property
of the geometry and not of the lighting conditions, results
could be combined with an arbitrary direct illumination. The
method was also useful for interactive applications because
the results were independent from the viewpoint. Landis de-
tailed how ambient occlusion could be used to add realism
to models [Lan02]. For a survey see [MFS08].
The obscurance O of a point p is defined as the integral
O(p) =1
π
where ρ is a function of the distance d(p,ω) of the first in-
tersection of a ray shot from point p with direction ω, p is a
surface point, θ is the angle between the normal vector at p
and direction ω, and the integration is over the hemisphere
oriented according to the surface normal. We only consider
a neighborhood of p, i.e. function ρ is set to 1 for distances
greater than a maximum distance dmax. Therefore, the inte-
gral function O(p) captures occlusion (or openness) infor-
mation of the environment of point p. Considering extreme
cases, an obscurance value 1 means that the point is com-
pletely open, i.e, not occluded and a value 0 means that it is
completely occluded.
Z
Ωρ(d(p,ω))cosθdω,
(1)
Ambient occlusion [Lan02] is a simplified version of the
obscurances illumination model. Ambient occlusion
A(p) =1
π
substitutes the ρ function in the obscurances equation (1)
by the visibility function V(p,ω) that has value 0 when no
geometry is visible in direction ω and 1 otherwise.
Z
ΩV(p,ω)cosθdω,
(2)
Color bleeding consists in the effect that the objects
around another object with intense coloration are dyed with
this color. To obtain color bleeding, Méndez et al. [MSC03]
included in Equation (1) the diffuse reflectivity R(q):
W(p) =1
π
whereqisthefirstpointindirectionωthatoccludes p.When
no occlusion is found within dmax, the average reflectivity is
used. Observe that adding color bleeding to obscurances is
almost free.
Z
ΩR(q)ρ(d(p,ω))cosθdω,
(3)
2.2. Volumetric shadowing
A volumetric version of the obscurances technique, called
vicinity shading, was proposed by Stewart [Ste03]. Vicinity
shading simulates illumination of isosurfaces by taking into
account neighboring voxels. An occlusion volume is com-
puted and stored in a shading texture that is accessed dur-
ing rendering. This volume has to be re-computed each time
that the rendering parameters are modified and the method
does not support color bleeding. Since this first work, sev-
eral models to illuminate the isosurfaces have been pro-
posed. Wyman et al. [WPSH06] presented a method that
supports the simulation of direct lighting, shadows and in-
terreflections by storing pre-computed global illumination
in an additional volume to allow viewpoint, lighting and
isovalue changes. Despite the improvements achieved with
these methods they still have a main limitation, they only
allow to represent one of the surfaces of the volume. This
limitation is overcome by Ropinski et al. [RMSD∗08] and
Hernell et al. [HLY07] using a local volumetric shadowing
effect. Ropinski et al. compute a local histogram for each
voxel from the voxel’s neighbourhood, by accumulating in-
tensities weighted by inverse squared distances. These local
histograms can be combined interactively with the user de-
fined transfer function to give an effect similar to local ambi-
ent lighting. Hernell et al. [HLY07] obtain the incident light
intensity, arriving at a voxel, by integrating for each voxel
and within a sphere surrounding it the attenuated transfer
function density. This comes to compute, in the usual way,
the visibility arriving at a voxel, using the opacities, aver-
aged for all directions.
It is important to note at this point the twofold difference
between these local volumetric shadowing effects and the
classic obscurances (or ambient occlusion) used in our ap-
proach. Firstly, obscurances technique uses a ρ function (see
discussion in Section 3.2) to modulate the effect of the oc-
clusion with the distance. Secondly, obscurances compute
c ? The Eurographics Association 2008.

Page 3

M. Ruiz, I. Boada, I. Viola, S. Bruckner, M. Feixas, and M. Sbert / Obscurance-based Volume Rendering Framework
Figure2:Localvolumetricshadowing(option1)andobscu-
rance computation (option 2) for a volume model consisting
of three concentric spheres with densities d1< d2< d3.
explicitly visibility tests. This means that, although the test
can be made up to a predefined maximum distance, if an
occlusion is found, the rest of the environment in this di-
rection is ignored, contrarily to local volumetric shadowing
which integrates for the whole distance. In Figure 2, the es-
sential difference between the local volumetric shadowing
and the obscurances approach is illustrated. Observe that ob-
scurances only take into account the distance from a voxel
to the next occluding one, not what is in between. This is
indeed different from Hernell’s algorithm, which considers
the accumulated visibility of the whole environment, being
nearer to the physical realism (or at least more coherent with
the transfer functions).
Obscurances (and later ambient occlusion) never claimed
to be physically realistic, it was introduced as a fast photo-
realistic approximation to indirect illumination. Local volu-
metric methods have a much higher cost, with complexity
proportional to the square of the number of voxels, against
the complexity of obscurances computation, proportional to
the number of voxels times the number of directions. Thus,
on a scale of physical realism (and cost) the different strate-
gies can be sorted into increasing order as follows: ambient
occlusion (the lowest), obscurances, Hernell’s [HLY07] and
Ropinski’s [RMSD∗08] approaches, and global illumination
(the highest).
3. Obscurances for volume rendering
In this section we go further into the obscurance-based vol-
ume rendering by testing different distance functions for
computing obscurances and providing discussion on quality
assumptions.
3.1. Algorithm
We take as a basic implementation of the obscurance-based
volume rendering approach the one proposed by Stewart
in [Ste03]. First, the volume data set is centered in a sphere
built from the recursive subdivision of an icosahedron and
the lines from each vertex to the center of the volume are
taken as the directions to consider (12, 42, and 162 direc-
tions have been taken in our experiments). Then, for each
direction, the volume is swept using Bresenham’s algorithm.
This is equivalent to casting parallel rays covering all vox-
els. Obscurance computation for a given voxel is based on
the presence (and the distance) of occluders within a cer-
tain radius along the processed direction. A visibility test
compares the densities of two voxels which can be inten-
sity values, which are independent of the transfer function,
or opacity values assigned by the transfer function. In each
case, we say that voxel vioccludes vi−1if the density of vi
is greater than that of vi−1. To process the voxels we use a
stack which stores the previously visited and yet unoccluded
voxels in a density-based decreasing order. All the voxels
in a ray are traversed and for each one we check if it is the
nearest occluder to one of the previous unoccluded voxels
(i.e. the ones stored in the stack), and in the occlusion test
we check the distance. In each step, we start to check if the
current voxel vioccludes the one on the top of the stack vs.
If vsis occluded then we can remove it from the stack so that
it will not be processed anymore and continue applying the
same procedure to vs−1. If vs is unoccluded, the rest of the
stack voxels do not need to be processed since vsis the voxel
of the stack with lower density. Then, the next voxel of the
ray is processed. This pre-computed obscurance is stored as
vicinity shading values in a separate texture volume which
is used during rendering.
In order to integrate color bleeding effects, we multiply
the obscurance value by the color of the occluding voxel,
and add an ambient color constant (in our case, white) to
the unoccluded voxels. In this way, spectral obscurances are
accumulated.
3.2. Analysis of ρ function
In this section, the meaning and shape of the ρ function in
the obscurances definition (1) is discussed. First, this func-
tion should be a monotonically increasing function of d.
Second, this function is bounded from above. This reflects
the fact that normally ambient lighting of a given point is
primarily affected by its neighborhood. This is especially
true for scenes without bright light sources that may affect
the illumination at large distances. From 0 to a determined
value dmax, the function increases from 0 to 1, and for val-
ues greater than dmax the returned value is 1. This means
that only a limited environment around the point p is con-
sidered and beyond this the occlusions will not be taken into
account.
The shape of the ρ function is deduced from the fact that
c ? The Eurographics Association 2008.

Page 4

M. Ruiz, I. Boada, I. Viola, S. Bruckner, M. Feixas, and M. Sbert / Obscurance-based Volume Rendering Framework
Figure 3: Different ρ(d) functions.
we are interested in what happens in the vicinity of a point
(or voxel). The nearer the occlusion, the higher the influence
it should have. This influence diminishes with increasing
distance. This is reinforced by interpreting the obscurance
model with the ρ(d) = (1−e−τd) function as the illumina-
tion at the non reflecting boundaries of a nonscattering gas
with opacity τ and constant volume emittance. If we con-
sider the occlusions of the environment as having a similar
damping opacity effect over the ambient light, we should
use a function as similar as possible to ρ(d) = (1−e−τd)
[IKSZ03]. Some candidate functions (see Figure 3) are: (a)
All-or-nothing, ρ(d) = 0, used in the ambient occlusion ap-
proach (it does not allow color bleeding as the contribution
of a hit is 0); (b) Linear, ρ(d) =
ity shading approach; (c) Exponential, ρ(d) = 1−e−
and (d) Square root, ρ(d) =
Note that we have considered the exponential function with-
out normalization since the normalized exponential would
become very similar to the linear function.
d
dmax, used in the vicin-
d
dmax;
?
d
dmax, introduced in [MSC03].
Different data sets have been used to analyze the effect
of each function in the final visualization. The obtained im-
ages are shown in Figure 4 where each column represents
a different ρ: (a) all-or-nothing, (b) linear, (c) exponential,
and (d) square root. In the first row, we show the behavior
of the above functions (including color bleeding) on a cu-
bic phantom model with 1283voxels, where except for three
adjacent faces with opacity = 1 (left wall: green, right wall:
blue, floor: white), the rest of the voxels are transparent.
The effects at the corner illustrate the behaviour of the dis-
tance function. The linear function ρ (column (b)) corrects in
somewaythenon-smoothedeffectoftheall-or-nothingfunc-
tion (column (a)), but it considers a wider than necessary en-
vironment. The exponential ρ (column(c)) produces darker
values, similar to the all-or-nothing case, due to the discon-
tinuity leap at d = dmax. The square root function (column
(d)) is a good compromise, as it considers a nearer environ-
ment and the darkness decays smoothly. This is appropriate
to enhance the details of the model. We also study the results
of applying the different ρ functions on a CT-human body of
256×256×415 voxels. The obscurance volumes and the
visualizations obtained using an obscurance-based illumina-
Data
Size12
0:36
1:14
42
1:56
4:09
162
6:56
15:02
Aneurism
CT-body
256×256×256
256×256×415
Table 1: Time cost (minutes:seconds) for computing obscu-
rances using 12, 42, and 162 viewing directions.
tion model (see Section 5.1) are shown, respectively, in the
second and third rows of Figure 4. They have been obtained
considering 162 viewing directions and a maximum distance
equivalent to 64 voxels. Ambient occlusion (column (a)),
linear (column (b)), and exponential (column (c)) become
darker, and square root (column (d)) appears less dark and
more pleasant to the eye.
In the obscurances computation, dmax is also a key pa-
rameter to be considered since it controls the number of
voxels in a determined direction that have to be taken into
account to compute occlusions (see Equation 1). Therefore,
if dmax has a high value, the probability of finding an oc-
clusion increases, and hence the obscurance value, leading
to darker images. Conversely, if dmax has a low value, the
probability to be occluded decreases, leading to low obscu-
rance values and hence lighter images. Figure 5 illustrates
this effect in two data sets considering different dmax val-
ues (8, 64, and 256, respectively). As expected, we can ob-
serve how the darkness of the image increases when dmax
increases. The other images in this paper have been com-
puted using dmax= 64. In Figure 5, the effect of the number
of directions in the obscurances values is also shown. Re-
sults for three different number of directions (12, 42, and
162) are given. Observe that although 12 directions could
be considered for fast editing, we need at least 42 directions
for a good quality final image. We have used the high qual-
ity obscurances given by 162 directions for the rest of the
images shown in this paper. While the obscurances volume
of the CT-body has been computed from its opacity (given
by the transfer function), the obscurances of the aneurism
have been computed from its intensity values. The time cost
for computing obscurances for the CT-whole body and the
aneurism is shown in Table 1. Times are given for a CPU
Intel(R) Core(TM)2 Quad CPU Q6600 at 2.40GHz with 2
GB of memory. Note that, in accordance with the algorithm
of Section 3.1, the time cost is proportional to the number
of voxels times the number of directions. In the worst case,
where densities are found in decreasing order, all the vox-
els in a ray would be pushed to the stack, and then each one
would be popped, giving a cost proportional to the number
of voxels in a ray. Thus, the cost of the algorithm is indepen-
dent of dmax.
4. Volume saliency
The human visual system is able to reduce the amount of
incoming visual data to a small but relevant amount of infor-
mation for higher-level cognitive processing. Different com-
c ? The Eurographics Association 2008.

Page 5

M. Ruiz, I. Boada, I. Viola, S. Bruckner, M. Feixas, and M. Sbert / Obscurance-based Volume Rendering Framework
(a) (b)(c)(d)
Figure 4: From top to bottom, obscurances with color bleeding for a synthetic model, obscurances for the CT-human body data
set, and a rendering of this data set using an obscurance-based illumination model, all of them obtained considering different
ρ functions. From left to right: (a) ρ(d) = 0, (b) ρ(d) =
d
dmax, (c) ρ(d) = 1−e−
d
dmax and (d) ρ(d) =
?
d
dmax.
putational models have been proposed to interpret the se-
lective visual attention. The biologically-inspired model of
bottom-up attention of Itti et al. [IK01] permits us to under-
stand our ability to interpret complex scenes in real time.
The selection of a subset of available sensory information
before further processing appears to be implemented in the
form of a spatially circumscribed region of the visual field,
called focus of attention, while some information outside the
focus of attention is suppressed. This selection process is
controlled by a saliency map which is a topographic repre-
sentation of the instantaneous saliency of the visual scene
and shows what humans find interesting in visual scenes.
Inspired by Itti’s work, Lee et al. [LVJ05] introduced the
concept of mesh saliency, a measure of regional importance
for 3D meshes, computed using a center-surround mecha-
nism that is able to identify regions that are different from
their surrounding context. Mesh saliency is captured from
surface curvatures and is used in graphics applications such
as mesh simplification and viewpoint selection. Feixas et
al. [FSG] defined a view-based saliency of a polygon as
the average information-theoretic dissimilarity between this
polygon and its neighbors. In the volume rendering field,
Kim et al. [KV06] presented a visual-saliency-based oper-
ator to enhance human perception of the volume data by
guiding the viewer’s attention to selected regions. A defini-
tion of voxel saliency is not provided and it is assumed that a
saliency value is assigned to each voxel by using a user spec-
ification, eye-tracking data, or feature computation. In dif-
ferent works on saliency, it has been shown that attention is
attracted by changes in luminance, color, curvature, texture,
shape,etc.[TIR05].Thatis,salientfeaturesaregenerallyde-
termined from the local differential structure of images and
different operators such as color or luminance gradient have
c ? The Eurographics Association 2008.

Page 6

M. Ruiz, I. Boada, I. Viola, S. Bruckner, M. Feixas, and M. Sbert / Obscurance-based Volume Rendering Framework
nd=162, dmax=8
nd=162, dmax=64 nd=162,dmax=256nd=12,dmax=64 nd=42,dmax=64
nd=162, dmax=8
nd=162, dmax=64nd=162, dmax=256 nd=12, dmax=64 nd=42, dmax=64
Figure 5: The obscurance volumes of the aneurism (first row) and CT-human body (second row) are visualized considering
different dmaxvalues and number of directions nd. The square root function has been used in all the cases. Computation times
are given in Table 1.
been used [vdWGB06]. In Gonzalez et al. [GSF08], from an
information theory perspective, ambient occlusion has been
defined as the occlusion information associated with each
polygon of the model.
In this paper, a definition of voxel saliency based on the
gradient of obscurance field is proposed. Considering that
obscurance represents occlusion information associated with
a voxel, its variation with respect to its surround can in-
deed be perceptually salient, i.e. it can be considered as a
salient feature of the volume. This saliency would be most
noticeableatedges,occlusionvariations,andcorners.Onthe
other hand, a smooth or uniform region would produce low
saliency values, as is intuitively expected.
The voxel saliency is defined as the magnitude of the gra-
dient of obscurancesestimated by using the4D linear regres-
sion method proposed in [NCKG00]:
2?
where voxel p is located at the origin of the coordinate sys-
tem, and A, B, and C are the components of the obscu-
rance gradient [A,B,C]. These components are computed
as A = ∑kw(k)O(k)x(k), B = ∑kw(k)O(k)y(k), and C =
∑kw(k)O(k)z(k), where k stands for the voxels in the neigh-
borhood centered at voxel p, w(k) is the distance between
voxels p and k, O(p) is the obscurance of voxel p, and
x(k), y(k), and z(k) are, respectively, the x, y, and z com-
ponents of the vector from voxel p to voxel k. In our exper-
iments, the neighborhood of p is given by a cube of 53vox-
els, since smoother results are obtained than by using a cube
of 33voxels as in [NCKG00]. For each data volume, the
saliency has been scaled ranging from 0 to 1. Analogously to
S(p) =
A2+B2+C2,
(4)
mesh saliency [LVJ05], the gradient of obscurances is scale-
dependent (i.e., the saliency value depends on the size of
the neighborhood considered). We have to emphasize that
our definition of saliency can be generalized to the local
volumetric shadowing methods [HLY07, RMSD∗08]. Fig-
ures 6(b), (d) and (f) show the color-coded saliency maps ob-
tained for the CT-human body corresponding to Figure 6(a).
5. Applications
In this section, we describe how obscurances can be applied
to volume rendering to interactively produce realistic and il-
lustrative images. Two applications of saliency maps are also
presented.
5.1. Realistic and illustrative rendering
To apply the obscurances to the visualization, we use the
Blinn-Phong shading model where the color resulting from
the local lighting of each voxel x is multiplied by its obscu-
rance value:
I(x) = (kdN(x)·L+ks(N(x)·H)n)O(x)
where kdand ksare the diffuse and specular lighting coeffi-
cients, N(x) the normal of the voxel, L the light vector, H the
half-angle vector between L and the direction to the viewer,
n is the Phong exponent, and O(x) the obscurance of voxel
x which has been adjusted to the range [0,1]. Figure 4 (third
row) illustrates the result of applying the obscurance-based
Blinn-Phong model on the CT-human body data set.
(5)
We also introduce two parameters, low and high, such that
c ? The Eurographics Association 2008.

Page 7

M. Ruiz, I. Boada, I. Viola, S. Bruckner, M. Feixas, and M. Sbert / Obscurance-based Volume Rendering Framework
Figure 6: (a) Original CT-body data set. (b, d, f) Color-coded (from blue to red) saliency maps corresponding to the most
salient views. (c, e, g) Illustrative visualizations obtained with a saliency-based opacity modulation.
from 0 to low obscurances are set to 0 (making the voxel
completely black), from low to high they follow a linear dis-
tribution (preserving their original value), and from high to 1
their value is set to 1 (thus the voxel becomes completely un-
obscured). Increasing the low threshold will turn more vox-
els black, and this can be used to increase the contrast of the
resulting image. Decreasing the high threshold means that
more voxels are not darkened by their obscurance, and so
become brighter, thus increasing slightly the contrast too. In
the limit we could set low and high to the same value to
have only some voxels with obscurance 0 and others with
obscurance 1. Thus, there would be voxels with their own
color (modified only by local lighting if applied) and the
others would be black. The user can modify low and high
parameters to obtain the desired effect interactively. Figure
1 shows different renderings of the CT-body data set. Figure
1(a) shows the visualization of the model resulting from the
application of obscurances with low = 0 and high = 1. Fig-
ure 1(b) has been obtained with low = 0.4 and high = 0.6,
and making the skeleton transparent. Finally, in Figure 1(c)
all the structures have been set to white and the obscurances
assignmenthasbeenadjustedwithlow=0.6andhigh=0.7.
5.2. Saliency
As a measure of importance, the volume saliency is applied
to obtain the most salient views and to enhance volume vi-
sualization by modifying the transfer function according to
the computed saliency.
Similar to [LVJ05], where mesh saliency was used to se-
lect the best views, a method to calculate the saliency of a
viewpoint is proposed. Given the saliency of all the vox-
els, we can find the viewpoint which maximizes the visible
saliency. The viewpoint saliency is defined by
S(v) =∑
p∈P
S(p)V(p),
(6)
where v is a given viewpoint, P is the set of voxels of the
volume data, S(p) is the saliency of voxel p, andV(p) is the
visibility of voxel p from v.
We also present an automated technique to enhance vol-
umevisualizationbyemphasizing(increasingtheopacityof)
the most salient voxels and de-emphasizing (reducing the
opacity of) the least salient ones. So, the viewer’s attention
is guided towards the most salient parts of the model.
In Figure 6, (a) the original CT-body data set, (b-c) the
most salient view, (d-e) the least salient view, and (f-g) the
most salient view per unit area are shown. Images (c), (e)
and (g) have been obtained by multiplying the opacity by
the saliency. Figure 7 shows (a) the original CT-body data
set and (b-c) two different renderings obtained by scaling the
opacity according to the saliency values. In Figure 7(b), vox-
els with saliency lower than 0.2 have been made transparent
and the opacity of the most salient ones has been preeserved.
In Figure 7(c), the voxels with saliency lower than 0.2 have
been made transparent while the opacity of the most salient
ones has been doubled.
6. Conclusions
In this paper, we have analyzed obscurance-based volume
rendering by evaluating the main parameters involved in its
computation, such as the obscurance function and the num-
ber of viewing directions. From this study, we conclude that
the square root function gives better results than other ana-
lyzed functions and that 42 directions are enough to obtain
obscurances of a certain quality, although for high quality
results we have used 162 directions. In addition, we have in-
troduced two new applications of obscurances. The first is a
technique to obtain illustrative renderings and the second is
a method to compute the saliency map as the gradient of ob-
scurances. Saliency has been used to enhance visualization
c ? The Eurographics Association 2008.

Page 8

M. Ruiz, I. Boada, I. Viola, S. Bruckner, M. Feixas, and M. Sbert / Obscurance-based Volume Rendering Framework
(a) (b)(c)
Figure 7: (a) Original CT-human body data set. (b, c)
Images obtained by scaling the opacity according to the
saliency values.
and to select the most salient views. All our proposals have
been integrated in a common framework and tested on sev-
eral volume data sets. As future work, we plan to programme
a GPU version of the obscurances algorithm to obtain real-
time or interactive obscurance computation.
Acknowledgements
This work has been supported by TIN2007-68066-C04-01
and TIN2007-67982-C02 of the Ministry of Education and
Science (Spanish Government), by the MedViz Initiative in
Bergen (medviz.uib.no), and by the Austrian Science Fund
(FWF) grant no. P18322.
References
[ER00]
non-photorealistic rendering of volume models. In VIS
’00: Proceedings of the conference on Visualization ’00
(Los Alamitos, CA, USA, 2000), IEEE Computer Society
Press, pp. 195–202.
EBERT D., RHEINGANS P.: Volume illustration:
[FSG]
information-theoretic framework for viewpoint selection
and mesh saliency. ACM Transactions on Applied Per-
ception. In press.
FEIXAS M., SBERT M., GONZALEZ F.: A unified
[GSF08]
Viewpoint-based ambient occlusion.
Graphics and Applications 28, 2 (2008), 44–51.
GONZALEZ
F.,SBERT
M.,FEIXAS
IEEE Computer
M.:
[HLY07]
cient ambient and emissive tissue illumination using lo-
cal occlusion in multiresolution volume rendering.
Eurographics/IEEE-VGTCSymposiumonVolumeGraph-
ics 2007 (2007), pp. 1–8.
HERNELL F., LJUNG P., YNNERMAN A.: Effi-
In
[IK01]
sual attention. Nature Reviews Neuroscience 2, 3 (2001),
194–203.
ITTI L., KOCH C.: Computational modeling of vi-
[IKSZ03]
S.: Fast, realistic lighting for video games. IEEE Comput.
Graph. Appl. 23, 3 (2003), 54–64.
[KV06]KIM Y., VARSHNEY A.:
hancement for volume visualization.
Comput. Graph. 12, 5 (2006), 925–932.
[Lan02]LANDIS H.: Production-ready global illumina-
tion. In Course 16 notes, SIGGRAPH 2002 (2002).
[LVJ05]LEE C. H., VARSHNEY A., JACOBS D. W.:
Mesh saliency. ACM Trans. Graph. 24, 3 (2005), 659–
666.
[Max95]MAX N.: Optical models for direct volume ren-
dering. IEEE Transactions on Visualization and Com-
puter Graphics 1, 2 (1995), 99–108.
[MFS08]MENDEZ-FELIU A., SBERT M.: From obscu-
rances to ambient occlusion: A survey. The Visual Com-
puter (2008).In press, published online 15 February
2008.
[MSC03]MÉNDEZ A., SBERT M., CATÁ J.: Real-time
obscurances with color bleeding. In SCCG ’03: Proceed-
ings of the 19th spring conference on Computer graphics
(2003), ACM, pp. 171–176.
[NCKG00]NEUMANN L., CSÉBFALVI B., KÖNIG A.,
GRÖLLER E.: Gradient estimation in volume data using
4d linear regression. Comput. Graph. Forum 19, 3 (2000),
351–358.
[RMSD∗08]ROPINSKI
T.,
DIEPENBROCK S., MENSMANN J., HINRICHS K. H.:
Interactive volume rendering with dynamic ambient
occlusion and color bleeding. Computer Graphics Forum
(Eurographics 2008) 27, 2 (2008), 567–576.
[Ste03]STEWART A. J.: Vicinity shading for enhanced
perception of volumetric data. In VIS ’03: Proceedings of
the 14th IEEE Visualization 2003 (VIS’03) (2003), IEEE
Computer Society, pp. 355–362.
[TIR05]TSOTSOS J. K., ITTI L., REES G.: A brief and
selective history of attention. In Neurobiology of Atten-
tion, Itti L., Rees G., Tsotsos J. K., (Eds.). Elsevier, San
Diego, CA, 2005.
[vdWGB06]
DANOV A. D.: Boosting color saliency in image feature
detection. IEEE Trans. Pattern Anal. Mach. Intell. 28, 1
(2006), 150–156.
[WPSH06]WYMAN C., PARKER S. G., SHIRLEY P.,
HANSEN C. D.: Interactive display of isosurfaces with
global illumination. IEEE Trans. Vis. Comput. Graph. 12,
2 (2006), 186–196.
[ZIK98]ZHUKOV S., IONES A., KRONIN G.: An am-
bient light illumination model. In Proceedings of Euro-
graphics Rendering Workshop ’98 (1998), Springer Wien,
pp. 45–56.
IONES A., KRUPKIN A., SBERT M., ZHUKOV
Saliency-guided en-
IEEE Trans. Vis.
MEYER-SPRADOW
J.,
VAN DE WEIJER J., GEVERS T., BAG-
c ? The Eurographics Association 2008.