Eurographics Workshop on Natural Phenomena (2007)

D. Ebert, S. Mrillou (Editors)

Eulerian Motion Blur

Doyub Kim and Hyeong-Seok Ko

Seoul National University

Abstract
This paper describes a motion blur technique which can be applied to rendering fluid simulations that are carried
out in the Eulerian framework. Existing motion blur techniques can be applied to rigid bodies, deformable solids,
clothes, and several other kinds of objects, and produce satisfactory results. As there is no specific reason to
discriminate fluids from the above objects, one may consider applying an existing motion blur technique to render
fluids. However, here we show that existing motion blur techniques are intended for simulations carried out in the
Lagrangian framework, and are not suited to Eulerian simulations. Then, we propose a new motion blur technique
that is suitable for rendering Eulerian simulations.
Categories and Subject Descriptors (according to ACM CCS): I.3.7 [Computer Graphics]: Three-Dimensional
Graphics and Realism

1. Introduction
Motion blur is essential for producing high-quality anima-
tions. The frame rate of most films and videos is either 24
or 30 Hz, whereas human vision is reported to be sensitive
up to 60 Hz [Wan95, CJ02]. Due to the lower frame rate of
film and video, when each frame is drawn as a simple in-
stantaneous sampling of the dynamic phenomena, artifacts
such as temporal strobing can occur. The graphics commu-
nity has been aware of this problem, and several motion blur
techniques have been proposed to solve this problem.
Fluids are often important elements of a dynamic scene,
and for the artifact-free production of such a scene, fluids
need to be rendered with motion blur. Since the graphics
field already has an abundance of motion blur techniques,
one may consider applying existing techniques to the mo-
tion blur of fluids. Unfortunately, existing techniques do not
produce satisfactory results. This paper describes why the
existing solutions do not work for fluids and how to modify
existing motion blur techniques to make them applicable to
fluids. Figure 1: A motion blurred image (left) produced with the
algorithm presented in this paper and an unblurred image
Motion blur techniques developed so far are intended for (right): A slice of water is falling along the wall, which
hits the logo and makes the splash. To factor out the effects
caused by the transparent material, we rendered the water
kim@graphics.snu.ac.kr as opaque.
ko@graphics.snu.ac.kr

c The Eurographics Association 2007.

 Doyub Kim & Hyeong-Seok Ko / Eulerian Motion Blur

rendering simulations that are performed in the Lagrangian global illumination and caustics using ray tracing and photon
framework. We will call this type of motion blur techniques mapping.
Lagrangian motion blur (LMB). The majority of objects en-
The third class of motion blur is known as image-based
countered in 3D graphics scenes (including rigid bodies, ar-
methods. Max and Lerner [ML85] proposed an algorithm
ticulated figures, deformable solids, and clothes) are simu-
to achieve motion blur effect by considering the motion on
lated in the Lagrangian framework; thus their motion blur
the image plane. Brostow and Essa [BE01] also proposed
can be readily rendered with LMB.
an entirely image-based method which can create motion
Simulation of fluids, however, is often carried out in the blur from stop motion or raw video image sequences. These
Eulerian framework. Considering the high quality and broad methods are suited to cases where the 3D motion is not avail-
applicability of LMB, and considering there is no specific able or the motion is already rendered. A more complete
reason to discriminate fluids from other 3D objects, one may survey of motion blur techniques can be found in Sung et
consider employing LMB for rendering fluids. An interest- al. [SPW02]
ing finding of this paper is that LMB is not suitable for ren-
We assume in this work that the 3D data of the fluid at
dering the results generated by an Eulerian simulation. So
every frame are available, but the data are not given in a pa-
far no algorithm has been proposed that can properly render
rameterized forms. Therefore the temporal super-sampling
motion blur of fluids that are simulated using the Eulerian
method seems to fit to the situation. In this paper, we de-
framework. In this paper, we explain why Lagrangian mo-
velop a motion blur technique based on the temporal super-
tion blur should not be used for rendering Eulerian simula-
sampling method.
tions. Insight obtained during this process led us to develop
a simple step that can be added to existing motion blur tech- Realistic rendering of fluids has been studied as well
niques to produce motion blur techniques that are applicable as the fluid simulation itself in the graphics community.
to Eulerian simulations (i.e., Eulerian motion blur (EMB)). Fedkiw et al. [FSJ01] visualized smoke simulation using a
Monte Carlo ray-tracing algorithm with photon mapping,
and Nguyen et al. [NFJ02] presented a technique based on
2. Previous Work Monte Carlo ray tracing for rendering fire simulations. Tech-
Motion blur was first introduced to the computer graph- niques for rendering liquids were also developed by Enright
ics field by Korein and Badler [KB83], and Potmesil and et al. [EMF02]. However, motion blur was not considered in
Chakravarty [PC83]. Korein and Badler proposed a method those studies.
that works on an analytically parameterized motion and cre- Mller et al. [MCG03] used blobby style rendering for vi-
ates a continuous motion blur. Potmesil and Chakravarty sualizing water represented with particles, and their method
proposed another method that creates continuous motion was subsequently improved by Zhu and Bridson [ZB05]
blur by taking the image-space convolution of the object to have smoother surfaces. For the visualization of La-
with the moving path. We will classify the above sort of mo- grangian particles, Guan and Mueller [GM04] proposed
tion blur techniques as analytic methods. point-based surface rendering with motion blur. Geundel-
The next class of motion blur we introduce is the tempo- man et al. [GSLF05] and Lossaso et al. [LIG06] attempted
ral super-sampling methods. Korein and Badler [KB83] pro- to include the rendering of the escaped level-set particles to
posed another method that renders and accumulates whole create the impression of water sprays.
(not partial) images of the object at several super-sampled Motion blur of Eulerian simulation has rarely men-
instants, resulting in a superimposed look of the object. tioned/practiced before; To our knowledge, there have been
The distributed ray tracing work of Cook et al. [CPC84] only two reports on motion blur of Eulerian simulations in
brought improved motion blur results. Their method suc- computer graphics thus far. In rendering water simulation,
cessfully increased the continuity of the motion blur by Enright et al. [EMF02] mentioned that a simple interpolation
retrieving pixel values from randomly sampled instants in between two signed distance volumes can be applied in order
time. Recently, Cammarano and Jensen [CJ02] extended this to find ray and water surface intersection. A few years later,
temporal super-sampling method to simulate motion-blurred Zhu and Bridson [ZB05] mentioned that the method will de-
stroy surface features that move further than their width in
one frame.
We note the intrinsic differences of the physical quantities used in
the Lagrangian and Eulerian frameworks. In the Lagrangian frame-
work, the simulator deals with the quantities carried by the moving 3. Computing Motion Blur
objects (e.g., the position, velocity, acceleration of the objects). In
The basic principle of motion blur is to add up the radiance
the Eulerian framework, on the other hand, the domain is discretized
into grids and the simulator deals with the quantities observed from
contributions over time, which can be expressed as
Z Z
fixed 3D positions (e.g., the velocity and density of the fluid at a
fixed grid point).
Lp = L(x0 ,~,t)s(x0 ,~,t)g(x0 )dA(x0 )dt, (1)
ts A

c The Eurographics Association 2007.

 Doyub Kim & Hyeong-Seok Ko / Eulerian Motion Blur

tual ray tracer. Usually multiple rays are shot for each pixel
for better results (Figure 2 (f)), which can be easily done by
t1 t2 t3 t4 associating multiple time samples to a pixel.

4. Lagrangian Motion Blur

(a) (b)
LMB is used for rendering objects that have explicit sur-
faces such as rigid bodies, deformable solids, and clothes.
The core part of the LMB approach is to compute, from the
given 3D data of each frame, rayobject intersection at arbi-
trary super-sampled instants. In order to do this, the location
t3 t2 t4 t1 of the surface at an arbitrary moment has to be estimated. In
(c) (d)
LMB, the estimation is done by taking the time-interpolation
of the vertices of the two involved frames; When the po-
sitions x(t n ) and x(t n+1 ) of the vertices at t n and t n+1 are
given, the estimated position xL () at super-sampled time
is calculated by
tn t n+1
xL () = x(t n+1 ) + x(t n ). (2)
t n+1 t n t n+1 t n
(e) (f)
We now briefly consider the physical meaning of the esti-
Figure 2: Motion blur with temporal super-sampling mation given by rearranging Equation (2) into the form
x(t n+1 ) x(t n )
xL () = x(t n ) + ( t n ) . (3)
t n+1 t n
where g() is the filter function, s() represents the shut- This equation shows that the estimation is the result of as-
ter exposure, and L() is the radiance contribution from the suming the movement was made with a constant velocity
ray [CJ02]. The above principle applies to both Lagrangian (x(t n+1 ) x(t n ))/(t n+1 t n ). However, how valid is this as-
and Eulerian motion blurs. In the equation, x is the place sumption? Movement of any object with non-zero mass has
where the movement of objects jumps into the motion blur; the tendency to continue its motion, and thus has an iner-
for the evaluation of x0 , the locations of the objects at arbi- tial component. When specific information is not available,
trary (super-sampled) moments need to be estimated, which calculation of the object position based on the inertial move-
forms a core part of motion blur. ment turns out to give quite a good estimation in many cases,
judging from images rendered using LMB. The error of the
For the development of a motion blur technique based on
estimation is proportional to the acceleration.
temporal super-sampling, we use Monte Carlo integration.
It computes the integral in Equation (1) by accumulating the
evaluations of the integrand at super-sampled instants. 5. Eulerian Motion Blur
More specifically, imagine the situation shown in Figure 2 In developing Eulerian motion blur, we assume that the sim-
(a) in which a ball is moving horizontally. Suppose that we ulation result for each frame is given in the form of 3D
have to create a blurred image for frame t n . Let be the grid data. The grid data consists of the level-set (or density)
shutter speed. For each pixel, we associate a time sample and velocity fields. As in the Lagrangian motion blur, it is
picked within the interval [t n /2,t n + /2]; the samples necessary to know how a ray traverses the fluid at an ar-
are taken from both past and future. Figure 2 (b) shows that, bitrary super-sampled instant. However, rendering Eulerian
for example, the time samples t1 , t2 , t3 , and t4 (which do not simulations needs a different type of information: instead
need to be in chronological order) are associated with the of the ray-surface intersection, the required information is
four pixels in a row. the level-set (in the case of water) or density (in the case
of smoke) values at the cell corners of all the cells the ray
For each pixel, we now shoot the ray at the associated
passes.
time, test for intersection, and estimate the radiance contri-
bution. Shooting a ray at a certain time and testing for inter-
section implies that the location of the objects at that time When the fluid has a clear boundary, as is the case for water,
should be estimated. Figure 2 (c) shows the object locations the surface can be extracted from a Eulerian simulation using the
at t1 , t2 , t3 , and t4 . In this particular example, only the ray marching cube algorithm [LC87]. In such a case, rendering can be
shot at t3 hits the moving object. Figure 2 (d) shows the final done with raysurface intersections. However, this approach is not
result. Figure 2 (e) shows an image produced with an ac- applicable to surfaceless fluids such as smoke which do not have

c The Eurographics Association 2007.

 Doyub Kim & Hyeong-Seok Ko / Eulerian Motion Blur

= 0.58 = 2.54 T(Tt n )u()

? ?
tn tn+1 ? ?
(a) (b)

TI
1.36
tn
0
-0.29 Figure 4: Estimation of the level-set values for Eulerian mo-
= tn+0.4 tn tn+1 tion blur. The grid points marked with ?s are the locations
(c) (d) whose level-set values must be estimated. The short solid ar-
rows at those points represent the estimated velocity u(x, ).
Figure 3: Characterization of the level-set change in a sim-
ple example: (a) the snapshot at t n , (b) the snapshot at t n+1 ,
(c) the situation at t n+0.4 , (d) the level-set changes.
the error is remarkable; and (3) the error is not related to the
grid resolution.

5.1. Why Time-Interpolation Does Not Work We now investigate why the time-interpolation gives such
an incorrect result. When specific information about the
Since the grid data are available only for the frames, we must movement is not available, exploiting the inertial compo-
somehow estimate the level-set values at arbitrary time sam- nent of the movement works quite well. The reason the LMB
ple . For the estimation, Enright et al. [EMF02] presented method works so well for Lagrangian simulations can be at-
a method which interpolates the level set data between two tributed to the fact that the LMB-estimation of object loca-
frames. Note that this is just as same as LMB-style estima- tion exploits the inertia. We can adopt this idea of exploiting
tion. An LMB-style solution would be to make the estima- inertia in the development of Eulerian motion blur. A ques-
tion with tion that arises here is whether the time-interpolation T I is
(t n+1 ) (t n ) exploiting the inertia.
T I () = (t n ) + ( t n ) . (4)
t n+1 t n It is critical to understand that it cannot be assumed that
the level-set/density change at a grid point will continue to
Contrary to expectation, the above estimation gives incor-
happen at the current rate. The space in which the fluid ex-
rect results. Imagine the simple case shown in Figure 3, in
periences inertia in the conventional sense is the 3D space.
which a spherical ball of water is making a pure translational
The inertial movement of the fluid in 3D space is reflected to
movement along the horizontal direction at a constant veloc-
the level-set field by updating the level-set according to the
ity. Figures 3 (a) and 3 (b) show two snapshots taken at t n
equation
and t n+1 , respectively. At the marked grid point, the level-set
values are (t n ) = 0.58 and (t n+1 ) = 2.54. The question is
+ u = 0. (5)
what would be the level-set value () at = t n+0.4 at that t
position? Since the fluid movement is analytically known in This equation states that the level-set should be advected in
this example, we can find out the exact location of the water the direction u at the rate |u|.
ball at , as shown in Figure 3(c). At , the marked position
comes within the body of fluid; therefore () has a negative
value. In fact, we can find out the trajectory of (t) for the 5.2. Proposed Method
duration [t n ,t n+1 ], which is plotted as a solid curve in Fig- For the Eulerian motion blur to exploit the inertial movement
ure 3(d). On the other hand, the time-interpolated result is of fluids, therefore, we propose that the estimation of the
T I () = 1.36, which is far from what has happened. Varia- level-set values at arbitrary super-sampled instants be based
tion of T I (t) within the duration follows a straight line and on the level-set advection, rather than the time-interpolation
is plotted with a dashed line in Figure 3(d). Here, we note of the level set values. More specifically, we propose to es-
that (1) the time-interpolation gives an incorrect result even timate E (x, ) of the level set value at a 3D position x,
in such a simple, non-violent, analytically verifiable case; (2) at a super-sampled time with the semi-Lagrangian advec-
tion [Sta99, SC91]
a distinct boundary. Even for cases where the surface extraction is E (x, ) = (x ( t n ) u(x, ),t n ). (6)
possible (as in the water), when the topology changes over frames,
LMB is difficult because finding the vertex correspondence is a non- This equation states that E (x, ) takes the level-set value of
trivial process. t n at the back-tracked position x ( t n ) u(x, ). In the

c The Eurographics Association 2007.

 Doyub Kim & Hyeong-Seok Ko / Eulerian Motion Blur

above procedure, we have not yet described how to estimate

the 3D velocity u(x, ) at . However, the velocity can be es-
timated with another advection, but this time on the velocity:
u(x, ) = u(x ( t n ) u(x,t n ),t n ). (7)
The above procedure works regardless of whether is in the
past or future side of t n . If needed, we can increase the ac-
(a) (b)
curacy by adopting the 3rd-order CIP method for the advec-
tion [SSK05, YXU01].
It is clear when |( t n ) u(x, )|/x in Equation (6) gets
larger, error from the estimation also increases. The artifact
from this error can be quite visible when fluid has fast ro-
tational movements. We can simply cure this problem by
adopting sub-stepping technique which is commonly used
by particle tracing method [EMF02, ZB05], so the each sub- (c) (d)
step moves less than one grid cell.
Figure 5: Results of the Eulerian motion blur with different
shutter speeds: (a) 1/500 sec., (b) 1/250 sec., (c) 1/125 sec.,
5.3. Discussion (d) 1/60 sec.
Normally the time-step taken for the fluid simulation is
smaller than the frame duration. So, one may store the data
for every simulation time step and perform the Lagrangian
motion blur with those dense data, which will result in re-
duction of the interpolation error. However, the level-set or
density data are usually large. So this approach will require
a large storage space, and I/O process will critically slow
down the simulation. The approach can also be problematic
during the rendering stage: (1) For producing motion blurred
image of a frame, the renderer will have to load the dense re-
sults. (2) Rendering the scene with multiple sets of data can
be a lot more complex than the Eulerian motion blur. Those
are the practical reasons this paper develops a motion blur
technique which works on the sparse data.

6. Experimental Results Figure 6: Sequence of images taken from the simulation in

which chunks of water are dropped onto shallow water.
The technique presented in this paper was implemented on
a Power Mac with G5 2.5 GHz processors. We applied it to
several Eulerian simulations as described below. For all fluid
simulations, a CFL restriction was 3. That is, the time-step shallow water. This simulation was done by octree-based
t < 3x/||u||max was used. level-set method [LGF04] with an effective grid resolution
of 128 192 128. Frames were dumped every 1/30 second
6.0.0.1. A Slice of Water Hitting a Logo Figure 1 shows and the rendering was done with a shutter speed of 1/60 sec-
the result of the EMB (with E ) when it was applied to a ond. Figure 7 shows the (enlarged) results of motion blur for
scene in which a slice of water falling along the wall hits a particular frame. Figures 7 (b) and (c) are the results gen-
the logo. This simulation was done by the level-set method erated with the level-set time-interpolation T I and the EMB
with an grid resolution of 240 320 80. The degree of mo- E , respectively. Figure 7 (a) is the result without any motion
tion blur can be controlled by giving different shutter speeds. blur, i.e., the un-blurred version of the scene. Although Fig-
Figure 5 shows the results; A longer shutter speed generates ure 7 (a) looks crisper, the video may have temporal-strobing
more blurred results. To factor out the effects caused by the artifacts.
transparent material, we rendered the water in opaque.
6.1. Quality Comparison of T I and E
6.0.0.2. Chunks of Water Dropping onto Shallow Wa-
ter Figure 6 shows four (blurred) snapshots taken from The quality enhancement gained by employing E instead
the simulation in which chunks of water were dropped onto of T I is more clearly visible when they are applied to an

c The Eurographics Association 2007.

 Doyub Kim & Hyeong-Seok Ko / Eulerian Motion Blur

(a) (b) (c)

Figure 7: Enlarged images of motion blur for a particular frame: (a) no motion blur, (b) motion blur with T I , (c) motion blur
with E .

either the level-set or the density (scalar) field. On the other

hand, Eulerian motion blur additionally dumps the 3D ve-
locity field, resulting in four times the dumped storage of the
other two cases. In generating a blurred frame for t n , motion
(a) (b) (c) blur with T I refers the level-set fields of t n1 , t n , and t n+1 ;
therefore it loads three arrays (three units of space). Regard-
Figure 8: Quality comparison of T I and E in a simple less of performing motion blur with T I or E , the level-set
example: (a) with T I , (b) with E , (c) the ground truth. estimation for a cell requires only simple extra computations
(O(1)-complexity and almost negligible) in addition to the
basic rendering. However, when we measure the time taken
analytically characterized fluid movement such as the one to render a whole image (Table 1 (C)), motion blur required
shown in Figure 3. Figure 8(a) and (b) show the results of 103% more time compared to the case with no motion blur.
motion blur when T I and E are used, respectively. In this This was because the total volume that the fluid occupies
example, evolution of the fluid surface over time is known over a frame duration is significantly larger than the instan-
analytically. Therefore we can use conventional LMB for taneous volume, as summarized in Table 1 (D). The number
rendering the ball. The result, shown in Figure 8(c), can be of floating point operations for EMB was slightly larger than
regarded as the ground truth image in this experiment. It can that for motion blur with T I .
clearly be seen that the result produced with T I is far from
the ground truth image, whereas the result produced using 7. Conclusion
E properly generates the blurred accumulation of the hori-
zontal movement. In this paper we classified existing motion blur techniques as
the Lagrangian motion blur to emphasize that they are suited
to rendering Lagrangian simulations. We showed that using
6.2. Analysis of the Memory and Computation the Lagrangian-style level-set/density interpolation T I for
Requirements Eulerian simulations produces incorrect results. We showed
We summarize the amount of space and computation needed that the error is remarkably large and is not reducible by em-
for the EMB in Table 1, along with the amounts for the cases ploying finer-resolution grids. The proposed method prop-
of no motion blur and motion blur with T I , for side-by- erly accounts for the inertial component of the fluid move-
side comparison. The statistics were taken while the images ment, by utilizing the advection equation for the estima-
shown in Figure 7 were rendered. The statistics for motion tion of the level-set/density at super-sampled instants. As a
blur with T I are included here because it might be help- result, the proposed method successfully produces conven-
ful for judging the memory and computation requirements tional kind of motion blur for Eulerian simulations.
of the EMB method in comparison to those of existing mo-
tion blur techniques. In this experiment, the image resolution
Acknowledgment
was 640 480, and 64 samples were used per pixel. The
same number of samples were used even for the no motion We would like to thank Oh-young Song for his insightful
blur case, in order to see the incremental computation taken comments. This work was supported by the Korea Science
for the level-set/density estimation steps introduced in this and Engineering Foundation (KOSEF) grant funded by the
paper. No-motion blur or motion blur with T I dumps only Korea government (MOST) (National Research Laboratory

c The Eurographics Association 2007.

 Doyub Kim & Hyeong-Seok Ko / Eulerian Motion Blur

No Motion Blur Motion Blur with T I Motion Blur with E

(A) Total dumped space (for the fluid) 1 1 4
(B) Space loaded for generating one frame 1 3 4
(C) Time taken for rendering the frame (in min.) 29 58 59
(D) Number of grid cells the fluid occupied 244711 441177 441177

Table 1: Comparison of memory and computation requirements.

M10600000232-06J0000-23210), Ministry of Information [LIG06] L OSASSO F., I RVING G., G UENDELMAN E.:
and Communication, the Brain Korea 21 Project, and Au- Melting and burning solids into liquids and gases. IEEE
tomation and System Research Institute at Seoul National Transactions on Visualization and Computer Graphics 12,
University. 3 (2006), 343352.
[MCG03] M LLER M., C HARYPAR D., G ROSS M.:
References Particle-based fluid simulation for interactive appli-
[BE01] B ROSTOW G. J., E SSA I.: Image-based mo- cations. In Proceedings of the 2003 ACM SIG-
tion blur for stop motion animation. Computer Graphics GRAPH/Eurographics symposium on Computer anima-
(Proc. ACM SIGGRAPH 2001) 35 (2001), 561566. tion (2003), pp. 154159.

[CJ02] C AMMARANO M., J ENSEN H. W.: Time depen- [ML85] M AX N. L., L ERNER D. M.: A two-and-a-half-d
dent photon mapping. In EGRW 02: Proceedings of motion-blur algorithm. vol. 19, ACM Press, pp. 8593.
the 13th Eurographics workshop on Rendering (2002), [NFJ02] N GUYEN D. Q., F EDKIW R., J ENSEN H. W.:
pp. 135144. Physically based modeling and animation of fire. ACM
[CPC84] C OOK R. L., P ORTER T., C ARPENTER L.: Dis- Trans. Graph. 21, 3 (2002), 721728.
tributed ray tracing. In SIGGRAPH 84: Proceedings of [PC83] P OTMESIL M., C HAKRAVARTY I.: Modeling
the 11th annual conference on Computer graphics and in- motion blur in computer-generated images. In SIG-
teractive techniques (1984), pp. 137145. GRAPH 83: Proceedings of the 10th annual conference
[EMF02] E NRIGHT D., M ARSCHNER S., F EDKIW R.: on Computer graphics and interactive techniques (1983),
Animation and rendering of complex water surfaces. pp. 389399.
ACM Transactions on Graphics 21, 3 (2002), 736744. [SC91] S TANIFORTH A., C T J.: Semi-lagrangian inte-
[FSJ01] F EDKIW R., S TAM J., J ENSEN H. W.: Visual gration scheme for atmospheric model - a review. Mon.
simulation of smoke. Computer Graphics (Proc. ACM Weather Rev. 119, 12 (1991), 22062223.
SIGGRAPH 2001) 35 (2001), 1522. [SPW02] S UNG K., P EARCE A., WANG C.: Spatial-
[GM04] G UAN X., M UELLER K.: Point-based surface temporal antialiasing. IEEE Transactions on Visualization
rendering with motion blur. In Proceedings of the and Computer Graphics 8, 2 (2002), 144153.
2004 Eurographics Symposium on Point-Based Graphics [SSK05] S ONG O.-Y., S HIN H., KO H.-S.: Stable but
(2004). non-dissipative water. ACM Transactions on Graphics 24,
[GSLF05] G UENDELMAN E., S ELLE A., L OSASSO F., 1 (2005), 8197.
F EDKIW R.: Coupling water and smoke to thin de- [Sta99] S TAM J.: Stable fluids. Computer Graphics
formable and rigid shells. ACM Transactions on Graphics (Proc. ACM SIGGRAPH 99) 33, Annual Conference Se-
24, 3 (2005), 973981. ries (1999), 121128.
[KB83] KOREIN J., BADLER N.: Temporal anti-aliasing
[Wan95] WANDELL B. A.: Foundation of Vision. Sinauer
in computer generated animation. In SIGGRAPH 83:
Associates, 1995.
Proceedings of the 10th annual conference on Computer
graphics and interactive techniques (1983), pp. 377388. [YXU01] YABE T., X IAO F., U TSUMI T.: The con-
strained interpolation profile method for multiphase anal-
[LC87] L ORENSEN W. E., C LINE H. E.: Marching cubes:
ysis. J. Comp. Phys. 169 (2001), 556593.
A high resolution 3d surface construction algorithm. In
SIGGRAPH 87: Proceedings of the 14th annual confer- [ZB05] Z HU Y., B RIDSON R.: Animating sand as a fluid.
ence on Computer graphics and interactive techniques ACM Transactions on Graphics 24, 3 (2005), 965972.
(1987), vol. 21, pp. 163169.
[LGF04] L OSASSO F., G IBOU F., F EDKIW R.: Simulat-
ing water and smoke with an octree data structure. ACM
Transactions on Graphics 23, 3 (2004), 457462.

c The Eurographics Association 2007.