﻿Eurographics Symposium on Rendering 2003
Per Christensen and Daniel Cohen-Or (Editors)
Interactive Rendering of Translucent Deformable Objects
Tom Mertens †
Jan Kautz ‡
Philippe Bekaert †
Expertise Centre for Digital Media
Limburgs Universitair Centrum
Universitaire Campus, 3950 Diepenbeek, Belgium †
Hans-Peter Seidel ‡
Frank Van Reeth †
Max-Planck-Institut für Informatik ‡
Saarbrücken, Germany
Abstract
Realistic rendering of materials such as milk, fruits, wax, marble, and so on, requires the simulation of subsurface
scattering of light. This paper presents an algorithm for plausible reproduction of subsurface scattering effects.
Unlike previously proposed work, our algorithm allows to interactively change lighting, viewpoint, subsurface
scattering properties, as well as object geometry.
The key idea of our approach is to use a hierarchical boundary element method to solve the integral describing
subsurface scattering when using a recently proposed analytical BSSRDF model. Our approach is inspired by
hierarchical radiosity with clustering. The success of our approach is in part due to a semi-analytical integration
method that allows to compute needed point-to-patch form-factor like transport coefficients efficiently and
accurately where other methods fail.
Our experiments show that high-quality renderings of translucent objects consisting of tens of thousands of polygons
can be obtained from scratch in fractions of a second. An incremental update algorithm further speeds up
rendering after material or geometry changes.
Categories and Subject Descriptors (according to ACM
CCS): I.3.3 [Computer Graphics]: Picture/Image GenerationDisplay
algorithms I.3.7 [Computer Graphics]: Three-
Dimensional Graphics and Realism Color Color, Shading,
Shadowing and Texture;
1. Introduction
In our daily life, we are surrounded by many translucent
objects, such as milk, marble, wax, skin, paper, and so on.
The translucency is caused by light entering the material and
scattering inside it. This subsurface scattering diffuses the
incident light, making small surface detail look smoother.
Furthermore, light may scatter through an object, which
lights up thin geometric detail if illuminated from behind.
These effects create a distinct look that cannot be achieved
with simple surface reflection models. Even if a material
does not seem to be very translucent at first sight, at the appropriate
scale it might exhibit subsurface scattering effects
12 (see also color plate figure 4).
† {Tom.Mertens,Philippe.Bekaert,Frank.VanReeth}@luc.ac.be
‡ {jnkautz,hpseidel}@mpi-sb.mpg.de
c○ The Eurographics Association 2003.
Subsurface scattering effects can be rendered offline using
a wide range of methods proposed for participating media,
including finite element methods 25, 1, 26 , (bidirectional) path
tracing 8, 19 , photon mapping 14, 3 , and diffusion approximations
29 .
A major breakthrough was recently achieved with an analytical
model for subsurface scattering 15 , eliminating the
need for numerical simulation of subsurface light transport
in homogeneous highly scattering optically thick materials
such as milk and marble. In order to use this model in global
illumination algorithms, a hierarchical integration technique
was proposed by Jensen et al. 13 . Unfortunately, this technique
does not appear to allow interactive rendering. The
method proposed by 20 is interactive, but only for rigid objects
with fixed, possibly inhomogeneous, subsurface scattering
properties. Hao et al. 10 produce similar results with
their technique, also for fixed geometry. Our goal is to render
deformable, translucent objects at interactive rates under
varying lighting and viewing conditions (see color page for
examples).
The problem at hand is related to the widely studied problem
of real-time rendering with complex surface BRDFs
e.g. 11, 16, 21 . These techniques cannot handle subsurface scat
tering because they assume that light is directly reflected and
not scattered inside the object. Recent work on precomputed
radiance transfer 27 can be extended to handle subsurface
scattering, as it has already been demonstrated to work for
participating media, but it assumes static models.
We show that our goal can be reached with a hierarchical
boundary element method to solve the subsurface scattering
integral (4) of §2, in the spirit of hierarchical radiosity
with clustering 9, 28, 26, 31 . An outline of the method is given
in §3. The success of the method is in part due to an efficient
and accurate semi-analytic integration method for the
needed point-to-patch form-factor like transport coefficients
(§4). Design choices of our implementation are described in
§5 and results are presented and discussed in §6.
2. Background
2.1. The Subsurface Reflection Equation
We first introduce the necessary background on subsurface
scattering and the dipole source approximation recently
brought to the attention for rendering translucent materials
by Jensen et.al 15, 13 .
The shade of a surface point xo on a translucent object, observed
from a direction ωo, is computed with the following
integral:
L → � �
(xo,ωo) =
L ← (xi,ωi)S(xi,ωi;xo,ωo)(ωi · Ni)dωidxi (1)
S
Ω + (xi)
L → (xo,ωo) and L ← (xi,ωi) denote exitant/incident radiance
respectively. S is the object’s surface, Ω+(xi) is the hemisphere
above xi in normal direction Ni, and S(xi,ωi;xo,ωo)
is the bidirectional surface scattering distribution function
(BSSRDF). In general, the BSSRDF is an eight-dimensional
function, expressing what fraction of (differential) light energy
entering the object at a location xi from a direction ωi
leaves the object at a second location xo into direction ωo.
Because of its high dimensionality, it is infeasible to precompute
and store this term, especially since it depends on
the object’s geometry.
Previous techniques 8, 15 show that subsurface scattering
can be modeled adequately using two components: single
scattering and multiple scattering. Single scattering accounts
for only a small fraction of the outgoing radiance of a highly
scattering optically thick translucent material such as marble,
milk, . . . . The dominant term is the multiple scattering,
which we will focus on.
Multiple scattering diffuses the incident illumination,
such that there is almost no dependence on the incident and
outgoing direction anymore. Therefore it can be represented
at high accuracy as a four-dimensional function Rd(xi,xo),
which only depends on the incident and outgoing positions.
Additionally accounting for the Fresnel transmittance when
light enters and leaves the material, we get the following sub-
Mertens et al. / Interactive Rendering of Translucent Deformable Objects
surface scattering reflectance function:
S(xi,ωi;xo,ωo) = 1
π Ft(η,ωo)Rd(xi;xo)Ft(η,ωi). (2)
Substituting this into Equation 1, we get:
L → (xo,ωo) = 1
Ft(η,ωo)B(xo)
π
�
(3)
B(xo) =
E(xi) =
E(xi)Rd(xi,xo)dxi
S
�
Ω
(4)
+ L
(xi)
← (xi,ωi)Ft(η,ωi)(Ni · ωi)dωi (5)
In order to render translucent objects efficiently, one needs
an efficient way to solve this equation at every surface point.
2.2. Dipole Source Approximation
First, we need to decide on a model for the BSSRDF, and
thus a model for Rd. This model has to fulfill two main criteria:
it should be adaptable to different materials, and should
provide an efficient solution of equation (4).
An accurate method to determine Rd(xi;xo) is to use a full
simulation. Many different techniques have been proposed
(coming from the area of participating media), e.g. 8, 26, 19, 14 .
While these techniques are effective, they fulfill none of the
above criteria.
We chose to use a recently introduced model for the BSS-
RDF in homogeneous media as the basis for our work 15 .
This model is based on a dipole source approximation for
the solution of a diffusion equation 12, 29 that models light
transport in densely scattering media. We shall see that it
fulfills the above requirements.
The key idea behind the dipole source approximation for
Rd(xi,xo) 15 , is elegant and simple. An incoming ray at position
xi on a homogeneous, planar, and infinitely large and
thick medium is converted into a dipole source (i.e. two
sources) at the same position. One source of the dipole is
placed at a distance zv above the surface at xi. The second
source is at a distance zr below the surface at xi. Rd(xi,xo)
is then obtained by summing a sort of illumination contribution
at xo, due to the two sources near xi. The result is a
function of the distance r only, between xi and xo (see Table
1).
Although the dipole source approximation is only valid
for planar surfaces 12, 18 , mis-using it for curved surfaces
yields highly plausible results too, as has been shown in several
complex examples by Jensen et al. 15, 13 . We will use it
in the same spirit. The model is also inherently limited to
homogeneous materials. The appearance of heterogeneous
materials was simulated by means of texture mapping techniques
15, 20 in previous work and this could be done in the
work presented here as well.
c○ The Eurographics Association 2003.
Rd(xi,xo) = α′
�
zr(1 + σsr)
4π
e−σsr
s3 + zv(1 + σsv)
r
e−σsv
s3 �
v
zr = 1/σ ′ t, zv = zr + 4AD
sr = �xr − xo�, with xr = xi − zr · Ni
sv = �xv − xo�, with xv = xi + zv · Ni
1 + Fdr
A =
1 − Fdr
Fdr = − 1.440 0.710
+ + 0.668 + 0.0636η
η2 η
D = 1/3σ ′ �
t , σ = 3σaσ ′ t
σ ′ t = σa + σ ′ s, α ′ = σ ′ s/σ ′ t
σ ′ s = reduced scattering coefficient (given)
σa = absorption coefficient (given)
η = relative refraction index (given)
Ft(η,ω) = Fresnel transmittance factor
Table 1: The dipole source BSSRDF model. This paper
presents an efficient strategy for computing integrals of this
model, similar to point-to-patch form factors in radiosity, as
well as a hierarchical evaluation algorithm allowing interactive
rendering speeds.
3. Outline of the New Method
In this section, we outline a hierarchical boundary element
method to solve equation (4) efficiently, in the style of hierarchical
radiosity with clustering. Before elaborating the
details of the method in next sections, we compare with related
work at the end of this section.
3.1. Discretisation of Equation (4)
The additional surface integral (4) required for subsurface
scattering, makes the calculation of subsurface scattering
considerably more expensive than local light reflection calculations.
If we break the surfaces of our object to be rendered into
regions Ak, and if we assume small lighting variation over
each region, making E(xi) = Ek constant for all xi ∈ Ak,
equation (4) reduces to the following sum:
Mertens et al. / Interactive Rendering of Translucent Deformable Objects
B(xo) = ∑EkF(Ak,xo) k
�
(6)
F(Ak,xo) = Rd(xi,xo)dxi
(7)
We compute the factors F(A,xo) for the midpoint of each
element w.r.t. all other elements. The irradiance Ek is also
evaluated at the midpoints. The sum (6) can then be used
to evaluate exitant radiance at each elements midpoint and
taken for the constant radiance value for the whole element.
c○ The Eurographics Association 2003.
Ak
The factors F(A,xo) are similar to point-to-patch form
factors in the radiosity method 2 . There, the form factor usually
has a purely geometric meaning, whereas our form factor
also encodes the material properties. It is generally much
smoother, allowing us to treat E(xi) as constant for large object
patches, and even to ignore visibility.
It turns out that for all but the nearest elements, a singlesample
estimate of the form factor is sufficiently accurate.
The same criterion for selecting a number of samples can be
used as in 13 . We present in §4 a semi-analytic integration
method for the form factor. It is indispensable for efficiently
and accurately handling nearby elements, which would otherwise
sometimes require a large number of samples.
3.2. Clustering
Computing subsurface scattered radiance in this simple way
is unfortunately a quadratic procedure in terms of the number
of elements. This is too costly for interactive image synthesis,
except for the simplest models. A log-linear procedure
is obtained by grouping distant elements hierarchically
in so called clusters, in a similar way as in hierarchical radiosity
with clustering 9, 28, 26, 31 .
Each cluster represents a collection of faces, along with
a compact representation for these faces. This collection of
faces is made up of all faces deeper in the hierarchy. A cluster
also stores pointers to one or more child-clusters, which
themselves store a representation for all their child-faces.
The leaf nodes of this hierarchy correspond with the original
faces of which the object is composed. We discuss the
requirements and options in constructing the cluster hierarchies
further in Section 5.1.
Rather than computing the form factor between each pair
of face elements, we first create a candidate link between the
top level cluster containing all the faces of the object and
itself. Next, this link is subjected to a refinement oracle. If
the oracle decides that the link would not allow sufficiently
accurate integration, the candidate link is refined by opening
the cluster at the receiver or emitter side. As such, new candidate
links result, which are tested in turn. The refinement
oracle and strategy we used is described in Section 5.2.
After linking, we compute the irradiance at all leaf nodes
(see Section 5.3). The leaf node irradiance is then pulled up
to the top of the cluster hierarchy by averaging the irradiance
on child clusters. We then recursively traverse the cluster
hierarchy and gather the irradiance at each cluster element
from over its links. This involves multiplying the irradiance
at the emitter element with the form factor (7) corresponding
to the pair of linked elements and adding the product at
the receiver cluster. Clusters always receive contributions at
their midpoint. Gathering at vertices is possible too, but care
needs to be taken with vertices shared by sibling clusters.
The contributions gathered at all clusters are then pushed
Figure 1: Part of the cluster hierarchy relevant for calculating
subsurface scattered light at the dot. Each cluster
element level is depicted with a different shade. The surrounding
triangles are directly linked to the leaf containing
the dot, whereas the other, larger triangles are connected
through one of the leaf’s parents. We see that interactions
with distant geometry are handled at higher levels.
The cracks between the different levels do not cause visible
artifacts as long as the refinement criteria are met.
down the hierarchy, i.e. energy received at higher levels in
the hierarchy is distributed to the leaf clusters. The accumulated
results at the leaf nodes (the faces of the object) are
averaged at the vertices to allow Gouraud shading. More details
on the rendering are given in Section 5.4.
As before, a single sample estimate suffices for most form
factors. Form factors with nearby face or cluster elements,
which would require more than one sample according to
Jensen et al. 13 , are integrated using the semi-analytical technique
of Section 4.
Figure 1 shows the hierarchy of cluster elements, relevant
for calculating subsurface scattered light at the indicated
spot. Part of the illumination is gathered at a higher
level than at the leaf face node.
3.3. Comparison with Previous Work
Jensen’s Algorithm Jensen et al. 13 also presented a hierarchical
integration technique for (4) that allows to integrate
subsurface scattering in global illumination algorithms like
photon mapping. First, they cover the surface of the object
with a uniform cloud of irradiance sample point locations.
These sample points are then sorted into an octree data structure,
which also allows to treat distant sample points as a
group. Their integration technique is not mesh based and
therefore allows a wider class of object descriptions than our
method, such as procedurally generated geometry.
The method proposed in this paper aims at hardware accelerated
interactive rendering, which is very often done for
mesh-based objects of the kind shown throughout this paper.
Our method offers a speed advantage in three areas:
• Integration of the distant response: our cluster hierarchy
Mertens et al. / Interactive Rendering of Translucent Deformable Objects
allows to gather the response from distant elements at
higher levels, amortizing its cost over many leaf elements.
• Integration of the local response: the semi-analytic form
factor integration method of Section 4 is considerably
more accurate and cheaper than their sampling approach
for uniformly lit nearby areas (see Section 4.2). Only
if nearby elements are not approximately uniformly lit,
do they need to be broken into smaller pieces. Jensen’s
sampling approach would correspond to always breaking
up nearby elements to a predefined maximum resolution.
In our proof-of-concept implementation, only cluster elements
are broken up. Adaptive subdivision of face elements
still is to be incorporated.
• Hierarchy rebuilding after a material or geometry update:
as long as the topology of the object doesn’t change, the
only effect of geometry or material changes is that links
need to be demoted or promoted in the hierarchy 5 . It will
be shown in the results section, that our refinement oracle
and strategy are cheap enough to allow sufficiently
fast relinking. They are even fast enough to re-render from
scratch at interactive rates. Re-building the sample octree
seems to be a major bottleneck inhibiting interactive updates
in Jensen’s approach.
The quality offered by both approaches is in principle similar.
In Jensen’s method, it is controlled by the sample point
density. In our approach, it is controlled by the mesh density.
Our method may suffer from artifacts due to the use
of Gouraud shading, or due to a poor quality mesh. These
problems have been well studied in the context of radiosity
however and many solutions are available 2 .
Both Jensen’s method and ours are based on the dipole
source approximation BSSRDF model. We mentioned before
that this model is only correct for planar boundaries between
homogeneous materials, but yields plausible results if
mis-used for curved surfaces. A heterogeneous material appearance
can be simulated with texture mapping techniques
15, 20 .
Hierarchical Radiosity (HR) Radiosity and the subsurface
scattering problem share some similarities, which the authors
of 20 also noticed. A full surface to surface transfer
needs to be computed. Hierarchical methods can relieve the
computational load considerably, as HR demonstrates. Our
method however differs from HR in the following way:
• Interreflections do not need to be computed. I.e. contributions
collected in the hierarchy do not need to be redistributed.
• Whereas HR computes area-to-area transfers, we can afford
to just sample area at a single sample point for which
our form factor was designed.
• We always assume full visibility, allowing for faster link
refinement. Moreover it reduces possible artifacts caused
by large contributions collected at higher levels.
• Cluster roots can be linked to themselves and their children
(and vice versa). As a matter of fact, these links
c○ The Eurographics Association 2003.
Mertens et al. / Interactive Rendering of Translucent Deformable Objects
represent a very important contribution, i.e. the global response.
Lensch et al. 20 Comparing to this work, our technique is
superior in computing the global response since they rely on
a matrix multiplication to scatter from one patch to another.
In other words, this is a quadratic procedure for which the
interactive rendering rate can easily break down in practice.
4. Subsurface Scattering Form Factor
Our rendering algorithm is based on a discretization of the
object surfaces into planar polygons Ai that both fulfill
certain constraints on size (discussed below) and that can
be regarded as homogeneously lit (irradiance Ei). In that
case, the outgoing radiosity (4) at any point xo can be obtained
by summing contributions EiF(Ai,xo). The integral
F(Ai,xo) := �
Ai Rd(xi,xo)dxi is similar to a point-to-patch
form factor in radiosity: it expresses the fraction of light entering
an object through the polygon Ai, that gets scattered
to the point xo. We propose here an efficient, custom integration
technique for it.
For planar surfaces, Rd is symmetric in xo and xi. Hence,
the fraction of light that scatters through an object from a
patch Ai to a point xo, is the same as the fraction of light
entering the object at xo that gets scattered into Ai. This is
no longer true when using the dipole approximation formula
for curved surfaces, due to the fact that the dipole approximation
itself is in principle not valid in that case. However,
using the dipole approximation (with inverse interpretation)
will be shown to result in plausible and attractive images, as
demonstrated in previous work. 15
As mentioned before, Rd(xi,xo) is the sum of two terms:
the contribution from two sources in a dipole configuration.
The terms are similar so that we will focus on the integral of
the contribution of a single source point d of the dipole (see
Table 1):
�
I = z(1 + σs)
A
e−σs
dx. (8)
s3 s denotes the distance between the point x on the polygon
A to the considered dipole source d. We first assume that A
is a triangle. We shall see that our result straightforwardly
extends to the case of arbitrary planar polygons. Figure 2
and Table 2 illustrate and explain the symbols used in our
derivation.
We solve the integral in polar coordinates in A’s supporting
plane, using the orthogonal projection of d onto this
plane, called d ′ , as the pole:
� θmax � rmax(θ)
I =
z(1 + σs)
rmin(θ)
e−σs
r dr dθ (9)
s3 θmin
Consider, to start with, the case that one of A’s vertices
coincides with d ′ so that rmin(θ) = 0 for all polar
c○ The Eurographics Association 2003.
θ 2
τ supporting plane of triangle
d single source of dipole
d ′
d orthogonally projected on τ
d ′′
d ′ orthogonally projected on edge
R �d ′ − d ′′ �
S �d − d ′′ �
h �d − d ′ �, height of d w.r.t. τ
S 2
R 2 + h 2
t parameter on edge, t=0 indicates d ′′
t0,t1, t2 midpoint, start and end of edge, respectively
Δt t = t0 + Δt
r,s s 2 = r 2 + h 2 , r 2 = R 2 +t 2
L edge length, L = t2 −t1
A
Table 2: Symbols used in the form factor derivation.
t 2
d’’
.
S
s
R
t
r
θ1
t 1
d (single source)
h
d’
¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥
£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£
 ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ 
 ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ 
 ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ 
¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢
¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢
¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢
¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢
£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£
¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥
 ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ 
¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥
£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£
 ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ 
 ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ 
 ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ 
¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢
¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢
¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢
¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢
£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£
¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥
 ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ 
¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥
£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£
 ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ 
 ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ 
 ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ 
¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢
¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢
¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢
¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢
£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£
¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥
 ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ 
¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥
£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£
 ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ 
 ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ 
 ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ 
¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢
¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢
¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢
¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢
£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£
¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥
 ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ 
¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥
£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£
 ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ 
 ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ 
 ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ 
¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢
¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢
¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢
¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢
£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£¤£
¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥¤¥
 ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ 
Figure 2: Left: Depiction of used variables (for one edge).
Right: Integration over the triangle is performed by integrating
along the edges. The contribution of front facing edges
is subtracted from the contribution of back facing edges.
angles θ. Change of integration variable r to the distance
s = √ r 2 + h 2 , and substitution of u = σs, then yields:
� � θ2 s(θ)
I =
z(1 + σs)
θ1 0
e−σs
sdsdθ
s3 � θ2 �
=
−
θ1
zσ
u e−u� σs(θ)
σh dθ
= z
h e−σh � θ2 e
(θ2 − θ1) − z
−σs(θ)
dθ. (10)
s(θ)
s(θ) denotes the distance from d to points on the triangle
edge opposite to d ′ . Our area integral thus has been reduced
to an integral over one triangle edge.
For a general triangle, the sum of three such expressions
is obtained: one for each edge of the triangle. The contribution
of back facing edges (furthest away from d ′ ) is counted
positive, while the contribution of front facing edges is subtracted
(see Figure 2).
For any closed polygon, the first terms in (10), called I1
θ1
d’
from now on, cancel if d ′ is outside the polygon. If d ′ is
inside, their sum equals 2π z h e−σh .
The second terms, which we call I2, are more complicated.
Change of integration variable from θ to t, the distance
along the edge segment measured w.r.t. the orthogonal
projection d ′′ of the dipole source point d onto the edges
supporting line, yields:
Mertens et al. / Interactive Rendering of Translucent Deformable Objects
t(θ) = Rtanθ and dθ = Rdt
R2 +t 2
(11)
� θ2 e
I2 = z
θ1
−σs
Rdt
s R2 +t 2
(12)
� t2 1
1
= zR √
t1 S2 +t 2 R2 +t 2 e−σ√ S2 +t2 dt. (13)
t1 = Rtan(θ1) and t2 = Rtan(θ2) denote the distances from
the edge endpoints to d ′′ . We did not find an analytic solution
to this integral as such, but Taylor series expansion w.r.t. the
midpoint of the edge (parameter t0) of the three main factors
in the integrand allows to obtain a good approximation:
1
1
�
√ = 1 −
S2 +t 2 s0
1
�
2t0Δt
2 s2 +
0
� Δt � � �
2
+ ... (14)
s0
1
R2 1
=
+t 2 r2 � �
2t0Δt
1 −
0
r2 +
0
� Δt � � �
2
+ ...
(15)
r0
e −σs = e −σs0
�
� 2t0Δt
1 − σs0
s2 + (
0
Δt
)
s0
2� �
+ ... (16)
Here, s0 := s(t0) and r0 := r(t0).
The product of the Taylor approximations results in a simple
polynomial and its integration is straightforward:
I2 = zRe−σs0
r2 0s0 �
�
L 1 + ... . (17)
The higher order terms can be safely ignored by imposing
constraints on the edge length L = t2 −t1, as discussed in the
next section.
In short, the contribution of a single edge e due to one of
the dipole source points is Ie = I1 + I2. The form factor of a
triangle A with three edges (a,b,c) and the dipole is:
�
Rd(xi,xo)dxi =
A
α′ � r
±(Ia +I
4π
v a) ± (I r b +I v b) ± (I r c +I v c ) � ,
(18)
where I r and I v denote the contribution from each dipole
source point. The sign for a term in the sum is positive if it
is a back facing edge for xo, otherwise it is negative.
4.1. Error Analysis
The error due to ignoring second and higher order terms
in the Taylor series expansions (14) to (16) is straightforward
to analyze if one makes sure that the first order term is
sufficiently smaller than 1. The series expansions converge
rapidly in that case.
F( A,x)
0.6
0.5
0.4
0.3
0.2
0.1
Form Factor
MC 10000 samples
MC 100 samples
MC 12 samples
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 mm
Figure 3: This graph shows F(A,x) of a standard triangle A
with the tip at x = (0..1cm,0cm). The base edge is L = 1cm
and the height is 0.5cm. The x-axis of the graph shows the
distance of the triangle’s tip to O. The sample point xo lies
also in the origin (planar setting). Skim milk was used as
the material. The black curve is evaluated using our method,
the colored curves show Monte Carlo evaluations. Our form
factor and 10000 Monte Carlo samples result in virtually the
same curve.
Suppose, we take 2t0 Δt
s0 s0 + � �
Δt 2 1
< s0
2 for the first factor.
Inspecting the higher order terms, we then find that the relative
error is about 15% maximally. This condition will be
satisfied if we ensure that |Δt|
s0 < 1 5 .
A sufficient condition for fast convergence of the second
factor is |Δt|
r0 < 1 5 . This condition is stronger than the previ-
�
ous one since r0 ≤ s0 = r2 0 + h2 .
For the last factor we need to take |Δt|σ < 1 5 . With L =
2|Δt|, we have the following conditions:
L
s0
< L
<
r0
2
5
and L < 2
. (19)
5σ
If these conditions are satisfied, the relative error on the
integrand of I2 is maximally roughly 50%. This apparently
high error can only occur near the end-points of the edges.
On most part of the edges, the error is much lower and so is
the error on the integrals.
Edges not satisfying the above constraints are recursively
subdivided until the conditions hold. The error decreases exponentially
with subdivision.
Even when ignoring second and higher order terms, first
order terms still appear in the result (17) for I2. Taking into
account the same conditions above, it turns out that also the
contribution from the first order terms are bounded and can
be safely ignored.
The condition L < 2
5σ is a global one. For more translucent
materials this condition becomes less important. We noticed
no significant impact on quality when ignoring this condition.
c○ The Eurographics Association 2003.
# samples
1000
100
10
Mertens et al. / Interactive Rendering of Translucent Deformable Objects
Form Factor
MC Predicted
1
1 2 3 4 5 6 7 8 9 10 mm
Figure 4: This graph compares the number of samples we
have to take vs. Jensen and Buhler’s method 13 to evaluate
the BSSRDF over an incrementally scaled standard triangle.
We chose xo = O. The tip of the triangles lies at the
origin as well. The x-axis of the graph shows the distance
from the triangle’s base edge to origin. In this example we
chose the material coefficients for marble. Although Jensen
and Buhler’s method needs less samples for small triangles,
their predicted number of samples is too small to achieve
acceptable accuracy, see Figure 3.
Figure 3 compares the results of our method with Monte
Carlo integration. It shows that our method indeed produces
accurate results with the above conditions. Using only 12
samples as done in 13 , produces rather low accuracy for this
setting.
4.2. Discussion
Our integration strategy has several distinct properties. First,
it turns out that many of the edge contributions in a polygon
mesh will cancel out. Indeed, when integrating over neighboring
planar triangles, shared edges are iterated over twice,
once per sharing triangle, and in opposite direction. If the
triangles receive the same irradiance, the terms Γ := I r e + I v e
for each such edge cancel out. Only the contribution of outer
edges remains. For the same reason, the form factor for an
arbitrary planar polygon can be obtained by simply summing
the appropriately signed contributions Γ for all polygon
edges.
Our integration strategy is partly based on analytic integration,
and partly on numerical approximation through Taylor
series expansion. The level of accuracy can be controlled
by loosening or tightening the refinement conditions (Equation
19).
In figure 4, we show the number of samples that we need
to take for different triangles. In this case, a sample corresponds
to computing I2 for a line segment (there can be multiple
line segments for an edge, due to subdivision in order
to meet the conditions). We compare this to the number of
c○ The Eurographics Association 2003.
Figure 5: Comparison of visual quality with the form factor
procedure and point sampling for marble at scale 20cm. Low
frequency noise is clearly visible with point sampling.
samples needed by Jensen and Buhler 13 . The efficiency of
our integration strategy is evident.
For distant clusters, the form factor will be sufficiently
small to be computed with a single point sample at the midpoint.
As figure 4 indicates, there is nothing to gain by using
the form factor in these cases. Due to the steeply descending
nature of Rd as distance increases, we see that only for
nearby clusters the form factor is appropriate.
By calculating form factors directly with an object mesh,
we avoid the need to generate and keep track of a dense set
of uniformly distributed surface point samples. This set requires
costly operations to be performed each time the object
geometry is changed. Furthermore, our integration method
does not suffer from high-frequency noise like Monte Carlo
methods 15 . We compared an obvious alternative to our form
factor algorithm, i.e. a form factor using simple uniform
sampling. The amount of samples is chosen according to the
point sample distribution criteria used by Jensen et al. 13 . It
turns out that our form factor procedure is superior for both
quality and performance. For moderately translucent materials,
Rd is quite steep and more samples are needed in the
integration. While making a material less translucent we noticed
a low frequency noise and a serious frame rate drop.
See table 5 and figure 5 for details.
Our form factor algorithm assumes polygons with sufficiently
constant irradiance. For sampling from environment
maps and point/spot lights, this assumption is acceptable
since lighting does not tend to change dramatically. Care
must be taken in situations where sharp shadow borders occur,
although for highly scattering media this becomes less
of a problem due to the diffusing nature of multiple scattering.
5. Implementation
We will now discuss several implementation-related issues
of our approach.
5.1. Mesh Hierarchy Construction
Our algorithm is given a hierarchical structure such as a
multi-resolution mesh, a face cluster hierarchy or a subdivision
surface, but needs to ensure that finer clusters are
fully embedded in coarser levels. For instance, approximating
subdivision schemes and progressive meshes cannot be
applied here.
5.1.1. Face Cluster Hierarchy
Face cluster hierarchies have been used in different applications,
but they are best known from face cluster radiosity
31 . Such a hierarchy is built by merging neighboring clusters
until a single cluster is reached containing all original faces.
We build our hierarchy bottom-up with an algorithm similar
to 6 . Unlike their work, our measure only tries to cluster
faces as compactly as possible, ignoring surface curvature.
Instead we make sure during linking, i.e. at run-time where
geometry and curvature may change, that only near-planar
clusters are chosen, since our form factor is correct for planar
clusters only. To this end we take a simple curvature measure
into account during linking in order to reduce the error for
the assumption of planar clusters: A′
A , where A′ and A are
the total projected area and total area of the cluster triangles,
respectively. The form factor from a cluster is then approximated
by projecting all its triangles onto the plane defined
by the cluster’s midpoint and averaged normal, and working
with the resulting polygon.
5.1.2. Subdivision
Another method to obtain a hierarchy with the desired properties
is surface subdivision. Since the finer clusters have to
be fully embedded in the next coarser level, we chose to use
subdivision based on 4-to-1 splits.
Butterfly. Any subdivision scheme to be used must be interpolatory
to fulfill our hierarchy criteria. For example, the
butterfly algorithm 4, 32 can be used to generate hierarchies
fulfilling this criteria. An example can be seen in Figure 1.
Shrink-Wrapped. Using a similar algorithm to 17 , we
create hierarchies for arbitrary closed meshes. We first
project the mesh from its center onto a sphere, and relax
it using the umbrella operator. We then take a coarse base
mesh, whose vertices are shared with the original mesh, and
recursively split. Whenever a new vertex is introduced, we
project it onto the sphere, find in which triangle from the
original mesh it lies, and use the interpolated coordinates
from the original mesh as the new coordinates for the new
vertex.
5.1.3. Discussion
The hierarchy generated by subdivision is more efficient
than the face clustering method, since every cluster is represented
by exactly one triangle. But it imposes also more
Mertens et al. / Interactive Rendering of Translucent Deformable Objects
restrictions on the mesh. The butterfly method can only produce
simple shapes from a low polygon count base mesh. Although
the shrink-wrapping can work on arbitrary meshes,
the approach often leads to unevenly tessellated meshes.
Overall, it is a compromise between efficiency and quality.
5.2. Refinement Oracle and Strategy
We use Jensen’s maximum solid angle criteria 13 , as it is
cheap to evaluate and practical:
• When the maximum deviation of the solid angle ω from
R’s children’s midpoints to E’s collection of faces exceeds
a certain threshold ε1, split and move the link down the
hierarchy at R. Repeat, until the deviation is below ε1.
• Before connecting a link to an emitter cluster E, check ω
from R’s midpoint to E’s faces. If it is above a threshold
ε2, split and move the link down at E. Repeat, until it is
below ε2.
More advanced linking criteria which reflect the magnitude
of the actual form factor better, may be devised in order
to reduce link count.
5.3. Irradiance Sampling
Irradiance is computed at each leaf cluster (triangle) at runtime
for environment maps and point light sources. In the
latter case, shadows can be computed by employing a variant
of shadow mapping 30 . In the former case, we sample irradiance
from an environment based on its spherical harmonics
coefficients. Nine coefficient suffice since the incoming radiance
undergoes a cosine-weighted integration over the hemisphere
and is thus bandlimited 24 . Shadows however cannot
be taken into account, unless we assume rigid objects. 27 The
resulting irradiance value is obtained with a simple dot product.
Prefiltered environment 7 maps can also be used, but we
did not implement this alternative.
Note that irradiance is only sampled at leaf clusters;
higher clusters pull the averaged irradiance from their children.
5.4. Rendering
Our rendering algorithm provides diffuse exitant radiance at
each vertex, which is passed to a set of hardware shaders.
The final rendering includes reflection mapping or phong
shading. We added a Fresnel factor, as seen in equation 3,
which is also used for rendering reflections from environment
maps.
Everything is computed in high dynamic range at floating
point precision. A simple gamma curve per pixel is applied
to the final shading for the purpose of tone mapping. This
enhances the global response which is typically much lower
that the local response, and adds a lot to the realism.
c○ The Eurographics Association 2003.
5.5. Interactivity
Our algorithm operates at different levels of interactivity:
1. Render Mode. Shading is re-evaluated under varying
lighting conditions. Irradiance is recomputed at the cluster
leaf nodes and distributed along the links and down
the hierarchy to the receiver leaf nodes, where they are
used for rendering.
2. Incremental Mode. Lighting, material and geometry
can be altered. The algorithm incrementally updates the
links’ form factors which are connected to altered clusters.
In case the material changes, every link is tagged
to be updated. We use a simple demotion/promotion procedure
5 to maintain consistency in the hierarchy: links
which do not meet the conditions are moved one level up
or down. Interactivity can be controlled by bounding the
number of form factor computations per frame.
3. On-the-fly Mode. Each frame we traverse the hierarchy
to perform a full evaluation from scratch. This mode
avoids the overhead of keeping track of links, resulting in
quicker updates. Memory requirements are also reduced
significantly. However, the frame rate is of course lower
and may compromise interactivity on slower systems.
6. Results
We implemented our algorithm in C++ on a dual Intel Xeon
2.4Ghz 2Gb RAM configuration with an ATI Radeon 9700
graphics board.
Table 3 shows that our algorithm behaves roughly loglinear
in the number of triangles.
T L
Mertens et al. / Interactive Rendering of Translucent Deformable Objects
L
T on-the-fly incr. render
2016 89K 44.1 20 8 5
8160 388K 47.5 47 23 13
32736 1744K 53.3 188 68 43
131736 6905K 52.7 722 351 199
Table 3: Model complexity versus link count and running
time (ms). Looking at the links-per-triangle ratio, we notice
that there is roughly a log-linear correlation between the
number of triangles T and the number of links L. The experiment
was done for a marble butterfly subdivision model
scaled at 2.5cm.
In table 4 we illustrate the performance of our approach
with different models. We see that the number of links is
mesh-dependent. In this experiment we choose very conservative
thresholds for the refinement oracle such that the quality
of the solution is guaranteed for every situation.
Table 5 illustrates the efficiency of the form factor compared
to point sampling. We see that it behaves more robustly
in rendering time when the local response gains importance.
c○ The Eurographics Association 2003.
model type T L on-the-fly
horse FC 8764 992K 97
bust FC 10518 1258K 102
elk FC 11384 1743K 165
candle S 4474 435K 41
cube S 16380 1572K 154
Table 4: Overview of performance with different models.
Timings for on-the-fly mode are in ms. Material was marble
scaled at 10cm. The hierarchy types are indicated by:
(F)ace (C)lustering or (S)ubdivision.
scale (m) form factor (ms) point sampling (ms)
.05 160 162
.1 160 168
.2 161 222
.4 169 501
.8 340 1778
Table 5: Comparison of the form factor procedure and point
sampling on marble. The object’s bounding box is scaled at
different sizes. The duration of a full on-the-fly evaluation is
measured.
We will now discuss some results depicted in figure 6. All
screenshots were taken interactively in on-the-fly mode, at
a frame rate ranging from 3 to 15fps, roughly. Except for
figure 6.5, which is rendered in incremental mode at 7.5fps.
In figure 6.1-2-3 a horse is rendered with three different
materials: skim milk, ketchup and candle wax, respectively.
Notice the color shifts across the model in figure 6.1. The
scattering of light is very obvious for the thin geometric features
in figure 6.3 when the model is lit from behind.
In figure 6.4, the scale of whole milk changes from 1.0m,
10cm, 5cm, 2cm, 1cm, to 5mm, respectively. The hierarchy
for the tweety model here was generated using shrink wrapping.
In the leftmost image, the shading is practically the
same as with simple diffuse reflection. The awkward appearance
in regions where normals are orthogonal to the direction
of incoming light, is due to (gamma curve) tone mapping.
We interactively added a bump to a subdivided cube in
figure figure 6.5. The resulting model is lit with a spot light,
and enhanced with shadow mapping. Notice how the shadow
cast by the bump is ‘leaking’ further on the adjacent plane.
For that same cube we applied a Perlin noise 23 distortion
(figure 6.5): 3D noise sampled at each vertex and perturbs
its position along the normal. Each frame we increment the
offset to the noise’s sampling position, resulting in a full deformation
on 16K triangles. This situation is rendered inter
actively at roughly 4-5 frames per second, including irradiance
sampling (spot light and shadows).
Figure 6.7a depicts how a simple subdivision shape is
used to simulate the appearance of a candle. It is lit from
inside by a moving and flickering point light source. When
we apply a simple ‘twist’ deformation, the model gets thinner,
resulting in more light passing through from the inside
(figure 6.7b). As this is a very simple model of nearly 5K
triangles, this experiment runs in real-time (15fps).
We show a translucent toy elk to show that models are
not restricted to be of genus 0. The spherical wheel has been
dented interactively, resulting in more scattered light passing
through.
Figure 6.9a shows a visualization of the clusters used for
shading a triangle on the ear on the left. Interactions at the
different levels in the hierarchy are colored uniquely. Right
next to it (6.9b) we show the actual rendering.
The appearance of a marble bust in figure 6.10 has been
enriched with Fresnel reflection from an environment map.
7. Conclusion
7.1. Summary
We have presented an efficient technique for rendering
translucent deformable objects. It allows to interactively
change the object’s geometry as well as the subsurface scattering
parameters.
To this end we introduce a novel boundary element
method similar to hierarchical radiosity. Our approach differs
from existing work by using a mesh hierarchy. It has
been proven to be highly suitable for deforming geometry
and changing material properties. The integration over the
whole surface can be handled at different levels in the hierarchy,
thus reducing the complexity for the integration from
quadratic to log-linear. The derivation of a semi-analytical
form factor for the BSSRDF introduced by Jensen et al. 15
is given, for the purpose of handling the local response very
efficiently. It computes the total contribution of a triangle
scattering light onto a point. Our experiments show that it
is an improvement over existing point sampling schemes for
both performance and accuracy.
We achieve interactive rendering rates for complex objects
consisting of tens of thousands of polygons. A full evaluation
of the shading can be obtained from scratch in fractions
of a second. An incremental update routine assists in
increasing the degree of interactivity after material or geometry
changes.
7.2. Future Work
The current implementation works on a fixed mesh with a
fixed hierarchy. It is worthwhile to explore adaptive meshing
techniques to handle situations where irradiance is not
Mertens et al. / Interactive Rendering of Translucent Deformable Objects
constant over a triangle, such as sharp shadow borders. This
is a common problem in radiosity as has been studied thoroughly
2 . We would like to extend our technique to handle
the very local response for fine surface detail such as displacement
or bump maps, which it currently can only handle
with finely tesselated meshes. Currently our implementation
does not handle topology changes. An efficient algorithm
could be devised to update the hierarchy at runtime in this
case, which seems feasible in combination with on-the-fly
evaluation.
Better oracle functions may lead to more efficient linking,
reducing overall rendering cost. Alternatives which take the
magnitude of the form factor into account are suitable candidates.
Interpolation methods employing higher-order basis functions
will remove artifacts caused by Gouraud shading, and
may reduce the need of many links. Also, the idea of introducing
higher order basis functions during hierarchical integration,
as has been done for HR 2 , seems promising.
Our technique is currently restricted to homogeneous materials,
a limitation imposed by using the dipole source BSS-
RDF model. We would like to e.g. allow for impurities in
the material casting volumetric shadows inside the material
(e.g. marble). The single scattering term is currently ignored.
We would like to add a single scattering term to the renderings,
which can be done with a GPU-based implementation
of the Hanrahan-Krueger model 8 , as demonstrated by
NVIDIA 22 . It would be interesting to investigate what the
accuracy is of the underlying BSSRDF model (i.e. using the
dipole diffusion approximation), for curved surfaces and arbitrary
geometry as it is only valid for semi-infinite planar
media.
Acknowledgements
The authors would like to thank Jens Vorsatz for his valuable
help with the hierarchical meshes.
This work is partly financed by the European Regional Development
Fund. The first author was also supported by a
Marie Curie Doctoral Fellowship.
References
1. Philippe Blasi, Bertrand Le Saëc, and Christophe Schlick. A
Rendering Algorithm for Discrete Volume Density Objects.
Computer Graphics Forum, 12(3):201–210, 1993.
2. Michael F. Cohen and John R. Wallace. Radiosity and Realistic
Image Synthesis. Academic Press Professional, Cambridge,
MA, 1993.
3. Julie Dorsey, Alan Edelman, Justin Legakis, Henrik Wann
Jensen, and Hans Køhling Pedersen. Modeling and Rendering
of Weathered Stone. In Proceedings of SIGGRAPH 99,
pages 225–234, 1999.
c○ The Eurographics Association 2003.
Mertens et al. / Interactive Rendering of Translucent Deformable Objects
4. Nira Dyn, David Levin, and John A. Gregory. A Butterfly
Subdivision Scheme for Surface Interpolation with Tension
Control. ACM Transactions on Graphics, 9(2):160–169, April
1990.
5. David A. Forsyth, Chien Yang, and Kim Teo. Efficient Radiosity
in Dynamic Environments. In Fifth Eurographics Workshop
on Rendering, pages 313–323, June 1994.
6. Michael Garland, Andrew Willmott, and Paul S. Heckbert. Hierarchical
Face Clustering on Polygonal Surfaces. In 2001
ACM Symposium on Interactive 3D Graphics, pages 49–58,
March 2001.
7. Ned Greene. Applications of world projections. In Graphics
Interface 86, pages 108–Ű114, May 1986.
8. Pat Hanrahan and Wolfgang Krueger. Reflection from layered
surfaces due to subsurface scattering. In Proceedings of SIG-
GRAPH 93, pages 165–174, 1993.
9. Pat Hanrahan, David Salzman, and Larry Aupperle. A Rapid
Hierarchical Radiosity Algorithm. In Computer Graphics
(Proceedings of SIGGRAPH 91), volume 25, pages 197–206,
July 1991.
10. Xuejun Hao, Thomas Baby, and Amitabh Varshney. Interactive
Subsurface Scattering for Translucent Meshes. In Proceedings
2003 ACM Symposium on Interactive 3D Graphics,
page to appear, april 2003.
11. Wolfgang Heidrich and Hans-Peter Seidel. Realistic,
Hardware-accelerated Shading and Lighting. In Proceedings
of SIGGRAPH 99, pages 171–178, August 1999.
12. Akira Ishimaru. "Wave Propagation and Scattering in Random
Media", volume 1. Academic Press, 1978.
13. Henrik Wann Jensen and Juan Buhler. A Rapid Hierarchical
Rendering Technique for Translucent Materials. ACM Transactions
on Graphics, 21(3):576–581, July 2002.
14. Henrik Wann Jensen and Per H. Christensen. Efficient Simulation
of Light Transport in Scenes With Participating Media
Using Photon Maps. In Proceedings of SIGGRAPH 98, pages
311–320, 1998.
15. Henrik Wann Jensen, Stephen R. Marschner, Marc Levoy, and
Pat Hanrahan. A Practical Model for Subsurface Light Transport.
In Proceedings of SIGGRAPH 2001, pages 511–518,
August 2001.
16. Jan Kautz and Michael D. McCool. Interactive Rendering with
Arbitrary BRDFs using Separable Approximations. In Dani
Lischinski and Greg Ward Larson, editors, Tenth Eurographics
Rendering Workshop 1999, pages 281–292, June 1999.
17. Leif P. Kobbelt, Jens Vorsatz, Ulf Labsik, and Hans-Peter
Seidel. A Shrink Wrapping Approach to Remeshing Polygonal
Surfaces. Computer Graphics Forum, 18(3):119–130,
September 1999.
18. Jan J. Koenderink and Andrea J. van Doorn. Shading in the
Case of Translucent Objects. In Human Vision and Electronic
Imaging VI, pages 312–320. SPIE, 2001.
19. Eric P. Lafortune and Yves D. Willems. Rendering Participating
Media with Bidirectional Path Tracing. In Eurographics
Rendering Workshop 1996, pages 91–100, 1996.
c○ The Eurographics Association 2003.
20. Hendrik P. A. Lensch, Michael Goesele, Philippe Bekaert, Jan
Kautz, Marcus A. Magnor, Jochen Lang, and Hans-Peter Seidel.
Interactive Rendering of Translucent Objects. In Proceedings
of Pacific Graphics 2002, pages 214–224, October 2002.
21. David McAllister, Anselmo Lastra, and Wolfgang Heidrich.
Efficient Rendering of Spatial Bi-directional Reflectance Distribution
Functions. In Proceedings of Graphics Hardware
2002, pages 79–88, September 2002.
22. Nvidia home page. http://www.nvidia.com.
23. Ken Perlin. An image synthesizer. In Proceedings of the 12th
annual conference on Computer graphics and interactive techniques,
pages 287–296. ACM Press, 1985.
24. Ravi Ramamoorthi and Pat Hanrahan. An Efficient Representation
for Irradiance Environment Maps. In Proceedings of
SIGGRAPH 2001, Computer Graphics Proceedings, Annual
Conference Series, pages 497–500, August 2001.
25. Holly E. Rushmeier and Kenneth E. Torrance. Extending
the Radiosity Method to Include Specularly Reflecting
and Translucent Materials. ACM Transactions on Graphics,
9(1):1–27, 1990.
26. François X. Sillion. A unified hierarchical algorithm for
global illumination with scattering volumes and object clusters.
IEEE Transactions on Visualization and Computer
Graphics, 1(3):240–254, 1995.
27. Peter-Pike Sloan, Jan Kautz, and John Snyder. Precomputed
Radiance Transfer for Real-Time Rendering in Dynamic,
Low-Frequency Lighting Environments. In Proceedings
of SIGGRAPH 2002, pages 527–536, July 2002.
28. Brian Smits, James Arvo, and Donald Greenberg. A Clustering
Algorithm for Radiosity in Complex Environments. Computer
Graphics, 28(Annual Conference Series):435–442, July
1994.
29. Jos Stam. Multiple scattering as a diffusion process. In Eurographics
Rendering Workshop 1995, pages 41–50, 1995.
30. Lance Williams. Casting curved shadows on curved surfaces.
In Computer Graphics (SIGGRAPH ’78 Proceedings), volume
12, pages 270–274, August 1978.
31. Andrew Willmott, Paul S. Heckbert, and Michael Garland.
Face Cluster Radiosity. In Eurographics Workshop on Rendering
1999, pages 293–304, June 1999.
32. Denis Zorin, Peter Schröder, and Wim Sweldens. Interpolating
Subdivision for Meshes with Arbitrary Topology. In Proceedings
of SIGGRAPH 96, pages 189–192, August 1996.
Mertens et al. / Interactive Rendering of Translucent Deformable Objects
1 2 3
4
5 6 7b
8 9a 9b 10
Figure 6: Interactive rendering results. 1-2-3: Horse model with different materials. 4: Changing the scale of subsurface
scattering for whole milk. 5: Tesselated cube with bump casting shadow. 6: Deformation on skim milk cube using Perlin noise.
7: Twist-deformation on candle. 8: Example of a genus 1 model. 9: Visualization of the traversed clusters in the hierarchy for
shading the bunny’s ear. 10: Marble bust rendered in a high dynamic range environment.
7a
c○ The Eurographics Association 2003.
