Volume Graphics

Arie Kaufman, State University of New York at Stony Brook

Daniel Cohen, Ben-Gurion University

Roni Yagel, Ohio State University

wift advances i n hardware - p arti c ul a rly faster, l arger, and c h e ape r

memories - have been transforming re v ol utio nar y approac hes in com­
puter grap hics into re ali ty. A typical example is the revolution of the
seventies. when hardware i nn ova t ions enabled the transition from vector grap hics
to raster g r ap hi (;" . Another transition with similar p ot en tial is shaping up in
volume graphics, an emerging subfield of computer graphics. This trend is rooted
in the extensive research and development effort in scientific visualization in
general and in volume visualization in particular.
Volume visualization is a method for extracting meaningful information from
volumetric data sets through the use of interactive graphics and i m ag ing . It is
concerned with the representation, manipulation, and re nde ri ng of volumetric
Just as 2D raster data sets.l The objective is [0 p r ov i de mechanisms for peering inside volumetric
graphics superseded data sets and for probing into voluminous and complex structures and dynamics.
V olume visualization encompasses an array of techniljues for projecting and
vector graphics, shading a volumetric data set, or properties thereof, and for interactively extract­
ing from it meaningful information using transformations, cuts, segmentation,
volume graphics has
translucency (;ontrol. me asurements, and so forth. Typically, the volumetric data
the potential to set is represented as a 3D discrete regular grid (a 3D raster) o f voxels (volume
elements) and is commonly stored in a volume buffer (also called a cubic frame
supersede surface buttrr). which is a large 3D array of voxcls (see '·Glossary"). Alternatively, othcr
data structures and formats have been used for storing and m an ipu la tin g the data
graphics for 3D
set. such as cell decomposition as in odrees,' sparse voxel matrices, semibound­
geometric scene aries. voxel runs, irregular grids, and surfaces of objects.
A voxel is the cubic unit of volume centered at the integral grid point. As a unit
representation, of v olu m e. the voxcl is the 3D counterpart of the 20 pixel, which represents a unit

manipulation, and of area. Thus, we can regard the volume buffer 01 voxels as the 3D counterpart of
the 2D frame buffer of pixels. Each voxel has numeric values associated with it:
rendering. these repres ent some measurahle properties or independent variables (for exam-

July 1993 5l

Authorized licensed use limited to: Imperial College London. Downloaded on May 30,2024 at 11:54:32 UTC from IEEE Xplore. Restrictions apply.
 Glossary

The terms in this glossary are defined in the context of Vector graphics - A subfield of computer graphics
volume graphics; their general and formal definition is not that represents the scene by a set of lines (vectors) that
necessarily reflected in the definition given here. Words are repeatedly redrawn to the screen by a random vec­
italicized in the definitions are defined under their own en­ tor generator.
tries in this glossary.
Visualization - A method of extracting meaningful in­
Adjacency - A relation between two voxels that are said formation from complex data sets through the use of in­
to be 6-adjacent if they share a face, 18-adjacent if they teractive graphics and imaging.
share a face or an edge, and 26-adjacent if they share a
face, an edge, or a vertex (see Figure 8 in the sidebar Volume buffer (3� raster, cubic frame buffer) A -

"Fundamentals of voxelization"). 3D array of voxels used for storing a regular volumetric

data set.
Connectivity - A relation between two voxels that are
said to be 6-connected (18-connected, 26-connected) if Volume graphics - The subfield of computer graphics
there exists a 6-path (18-path, 26-path) between them. that employs a volume buffer for scene representation
and is concerned with synthesizing, manipulating, and
Hybrid (intermixed) voxel model - A volumetric data set rendering such scenes. Volume graphics is the 3D
in which a synthetic voxelized model is intermixed with counterpart of raster graphics.
sampled or computed data.
Volume modeling -The synthesis, analysis, and ma­
Neighborhood -The 6-neighborhood (18-neighborhood, nipulation of sampled, computed, and synthetic objects
26-neighborhood) of a voxel is the set of all voxels that are contained within a volumetric data set.
6-adjacent (18-adjacent, 26-adjacent) to it.
Volume rendering - A direct technique for visualizing
Path - A sequence of voxels is a 6-path (18-path, 26-
volume primitives without any intermediate conversion
path) if every two consecutive voxels along the sequence
of the volumetric data set to surface representation (see
are 6-adjacent (18-adjacent, 26-adjacent).
surface rendering).

Raster graphics - A subfield of computer graphics that

Volume visualization - A visualization method con­
represents the scene by a 2D array (raster) of pixels stored
cerned with the representation, manipulation, and ren­
in a 2D frame buffer.
dering of volumetric data.
Ray casting - A volume-viewing algorithm in which sight
Volumetric data set The aggregate of voxels tessel­
rays are cast from the viewing plane through the volume.
-

The tracing of the ray stops when the visible voxels are de­ lating the volume.
termined by accumulating or encountering opaque value.
Voxblt (voxel block transfer) - A set of operations by
Ray tracing (3D raster ray tracing, discrete ray tracing, which a rectangular subvolume of voxels (called a
volumetric ray tracing) - A volume-viewing algorithm in "room") can be copied within the volume buffer with ar­
which light behavior is simulated by recursively tracing indi­ bitrary write modes and maskings. Voxbit is the 3D
vidual imaginary rays of light as discrete lines through the counterpart of the 2D bitblt (bit block transfer).
scene represented in a volume buffer (see the sidebar "Dis­
crete ray tracing"). Voxel - An abbreviation for "volume element" or "vol­
ume cell." Each voxel is a unit of volume and has a nu­
Surface rendering - An indirect technique used for visu­ meric value (or values) associated with it that repre­
alizing volume primitives by first converting them into an in­ sents some measurable properties or independent
termediate surface representation and then rendering them variables of a real object or phenomenon. The voxel is
to the screen using conventional computer graphics tech­ the 3D conceptual counterpart of the 20 pixel.
niques (see volume rendering).
Voxel space -A 3D integer grid or lattice of voxels in
Three-dimensional discrete topology - The formal topo­ which the volumetric data set (the object) resides.
logical characterizations of the digital voxel space that con­
form as closely as possible to the corresponding character­ Voxelization (volume synthesis, 3D scan conver­
izations used in the topology of the continuous 3D space. sion) - The process of converting a geometric repre­
sentation of a synthetic model into a set of voxels that
Tunnel - A 3D path of certain connectivity that penetrates best represents that synthetic model within the discrete
a discrete surface. Surface thickness is defined by the ab­ voxel space (see the sidebar "Fundamentals of voxel­
sence of particular tunnel$ in the surface. ization").

52 COMPUTER

Authorized licensed use limited to: Imperial College London. Downloaded on May 30,2024 at 11:54:32 UTC from IEEE Xplore. Restrictions apply.
 pie. color. opacity. density. material,
coverage proportion. refractive index. Sampled/computed data GeometriC model
velocity. ,trength. and time) of thc real
phenomenon or ohj ect residing in the
unit volume represented by that voxel. F(t)=TMG
The aggregate of voxels tessellating the 0<1<1
volume bu ffer forms the volumetric data
set.l
The source of volume data is sampled
: 3D reconstruction Surface
data of real objects or phenomena. com­ Volume: construction Surface
puted data produced by a computer sim­ representation
ulation, or modeled data generated from
a geometric model. Examples of appli­
cations generating sampled data occur
in medical imaging (computed tomog­
raphy (eT). magnetic resonance imag­
ing, ultrasonography). biology (confo­
cal microscopy), geoscience (seismic
measurements), industry (industrial CT ,,"
Volume rendering Surface
inspection), and molecular systems (elec­ rendering
Figure 1. Taxono­
tron density maps).l Examples of appli­

•
my and dataflow
cation> that generate computed data
of volume visual­
sets. typically hy running a simulation
ization and volume
on a supercomputer. occur in meteorol­
graphics. The solid
ogy (storm prediction), computational
arrows mark the
fluid dynamics (water flow), and com­
primary volume
putational chemistry (new materials).
graphics dataflow.
Although 3D raster representation
seems to he more natural for empirical
imagery because it can represent interi­
ors and digital samples, its advantages greatly advance the field of 3D graphics tional surface-rendering algorithm.
are abo attracting traditional surface­ by offering a comprehensive alterna­ HoweVer, the preferred alternative for
based applications that deal with the tive to traditional surface graphics. volume viewing is for the volume prim­
modeling and rendering of synthetic Figure 1 shows the taxonomy and the itives to be projected directly onto a
scenes represented by geometric mod­ dataflow of volume visualization and 2D pixel buffer. This process, called
els. Some examples are the rendering of volume graphics. In this figure, the use volume rendering, involves both the
fractals. hypertextures,! fur.' and gas­ of volume graphics techniques in vari­ viewing and the shading of the volume
es.' and CAD models and terrain mod­ ous stages of volume visualization is image and can be accomplished by for­
els for flight simulators.' Furthermore, m arkcd with solid lines. The major ward projection.' ray casting,' or dis­
in many applications i nvolving sampled sources of volumetric data are sampled/ crete ray tracing (see the sidebar "Dis­
data - for example, surgical planning computed data ( t op left) and geometric crete ray trac i n g" a n d Kaufman,'
and radiation therapy planning - the models (top right). The sampled/com­ Chapters � and 4).
data need to be visus lized along with puted input is 3D reconstructed to fill After explaining the primary reasons
synthetic objects that may not be avail­ gaps of missing information and is then hehind the transition in 2D graphics
able in digital form, such as prosthetic stored in the volume buffer. The geo­ from the vector-based to the raster-based
devices. scalpels, injection needles, iso­ metric model in 3D continuous space is approach, we will focus on volume graph­
dose surfaces. and radiation beams. The represented by a geometric formula that ics and show that it has the potential to
geometric objects can be converted into is 3D scan converted (vu_telized) into a promote a similar transition in 3D graph­
voxel representations (voxelized) and set of voxels that best approximate the ics as an alternative to traditional sur­
intermixed with the sampled organ in model and are stored in the volume face graphics. Then we will elahorate
the voxel buffer.- huffer (see Kaufman,l Chapter 5). For on the weaknesses and advantages of
Volume graphics is concerned with the fundamentals of voxelization and volume graphics.
the synthesis. manipulation, and ren­ the related 3D discrete topology issues.
dering of volumetric objects stored in a see the sidebar "Fundamentals of vox­
volume huffer of voxels. Unlike volume elization." From vector graphics to
visualization. which focuses primarily To visualize the data set in the vol­ raster graphics
on sampled and computed data sets. ume huffer. the sampled/computed data
volume graphics is concerned primarily can be converted into a geometric
with m odeled geometric scenes. and model by fitting geometric primitives to Graphics displays in the sixties and
particularly with those represented i n a surfaces that have been detected in the seventies were based on vector drawing
regular volume buffer. The volume volume. This set of primitives is then devices and on an object-based approach
graphics approach has the potent i al to rendered to the screen using a (radi- to scene representation, manipulation.

Jul) 1 993 53

Authorized licensed use limited to: Imperial College London. Downloaded on May 30,2024 at 11:54:32 UTC from IEEE Xplore. Restrictions apply.
 and d isplay. A geometric representa­ list. The major advantages of vector bited no aliasing. This technology. how­
tion ofthe objects constituting the scene graphics were its capability for perform­ ever, offered calligraphic drawing onl y ,

was stored in a display list. The screen ing object-related operations on the dis­ while the interior shaded areas were
was refreshed by redrawing the vectors play list and the fact that the vectors it extre m ely hard to render.
constituting the objects in the display drew were cuntinuous and thus exhi- The alternative approach - ra,ter

Fundamentals of voxelization

The synthesis of voxel-represented objects is an indis­ Unfortunately, the extension of the 20 definition of sepa­
pensable stage in volume graphics. This stage, called vox­ ration to the third dimension and to surfaces is not straight­
elization, is concerned with converting geometric objects forward, since voxelized surfaces cannot be defined as an
from their continuous geometric representation into a set of ordered sequence of voxels and a voxel on the surface
voxels that best approximates the continuous object. Since does not have a specific number of adjacent voxels. Fur­
this process mimics the scan-conversion process that pixel­ thermore, there are important topological issues, such as
izes (rasterizes) 20 geometric objects, it is also referred to the separation of both sides of a surface, which cannot be
as 3D scan conversion. In 20 rasterization, the pixels are well defined using 20 terminology. The theory that deals
drawn directly onto the screen to be visualized, and filtering with these topological issues is called 3D discrete topology.
is applied to reduce the aliasing artifacts. However, voxel­ Below we sketch some basic notions and informal defini­
ization does not render the voxels; it merely generates a tions used in this field.
database of the discrete digitization of the continuous ob­
ject. Fundamentals of 3D discrete topology. The 3D dis­
Intuitively, it would seem that a proper voxelization simply crete space is a set of integral grid points in 3D Euclidean
"selects" all voxels that are met (if only partially) by the ob­ space defined by their Cartesian coordinates (x, y, z). A
ject body. Although this ap­ voxel is the unit cubic volume centered at the integral grid
proach could be satisfactory point. The voxel value is mapped onto {O,l}: The voxels as­
in some cases, the objects signed "1" are called the "black" voxels representing
generated are commonly too opaque objects, and those aSSigned "0" are the "white" vox­
coarse and include more els representing the transparent background. Outside the
voxels than necessary. For scope of this article are nonbinary approaches in which the
example, in Figure A a 20 voxel value is mapped onto the interval [0,1] representing
Figure A. An example of a curve is rasterized into a either partial coverage, variable densities, or graded opaci­
20 discrete curve (shaded connected sequence of pix­ ties. Because of its larger dynamic range of values, this ap­
pixels) that intuitively sep­ els. Although the discrete proach may support higher quality rendering.
arates its two sides even curve does not "cover" the Two voxels are 26-adjacent if they share either a vertex,
without containing all entire continuous curve, it is an edge, or a face (see Figure B). Every voxel has 26 such
those pixels pierced by connected and concisely adjacent voxels: Eight share a vertex (corner) with the cen­
the continuous line. and successfully separates ter voxel (Figure Bl), 12 share an edge (Figure B2) and six
,

both sides of the curve. share a face (Figure 83). Accordingly, face-sharing voxels
One practical meaning of separation is apparent when a are defined as 6-adjacent, and edge-sharing and face·shar­
voxelized scene is rendered by casting discrete rays from ing voxels are defined as 18-adjacent.
the image plane to the scene. The penetration of the back­ The prefix N is used to define the adjacency relation,
ground voxels (which simulate the discrete ray traversal) where N 6, 18, or 26. We say that a sequence of voxels
=

through the voxelized surface causes the appearance of a having the same value (for example, "black") is an N-path if
hole in the final image of the rendered surface. Another all consecutive pairs are N-adjacent. A set of voxels A is N­
type of error might occur when a 3D flooding algorithm is connected if there is an N-path between every pair of vox­
used either to fill an object or to measure its volume, sur­ els in A. An N-connected component is a maximal N-con­
face area, or other properties. In this case, the nonsepara­ nected set.
bility of the surface causes the flood to leak through the dis­ Figure C shows a 20 discrete 8-connected black curve
crete surface. and a sequence of 8-connected white pixels (B-component)

Figure B. The three types of voxel adjacencies in 3D discrete space: Figure C. An example of a 20 8-curve
(1) the six voxels thai are 6-adjacent to the voxel al the center (not seen), that does not separate its two sides.
(2) the 18 voxels that are 18-adjacent to the voxel at the center, and (3) the The white 8-curve penetrates from
26 voxels that are 26-adjacent to the voxel at the center. one side to the other.

Authorized licensed use limited to: Imperial College London. Downloaded on May 30,2024 at 11:54:32 UTC from IEEE Xplore. Restrictions apply.
 graphics - has been predominant since correspond to the discrete representa­ Table 1 contrasts vector graphics with
the late seventies. Raster graphics uses tion ofthe geometric objects. The screen raster graphics. Unlike vector graphics,
a 2D frame buffer (a raster) of pixel s for is refreshed by a video controller, which raster graphics provides the capability
scene representation and a point-based repeatedly displays the frame buffer on to present realistic, shaded, and tex­
renderer for coloring those pixels that the screen. III tured surfaces in full color, as well as

should be directly voxelized, preferably generating an N­

separating, an N-minimal, and a covering set, where N is
application dependent. The voxelization algorithms should
follow the same paradigm as the 2D scan-conversion algo­
rithms; they should be incremental and accurate, use sim­
ple arithmetic (preferably integer only), and have a com­
plexity that is not more than linear with the number of
(1) voxels generated.
Figure D. (1) a 6-connected surface that is not 6-separat­ The literature of 3D scan conversion is relatively small.
ing; (2) a 6-connected surface that is not l8-separatlng. Danielsson3 and Mokrzycki4 developed similar 3D-curve al­
gorithms in which the curve is defined by the intersection of
two implicit surtaces. Voxelization algorithms have been de­
that passes from one side of the black component to its veloped for 3D lines, 3D circles, and a variety of surfaces
other side without intersecting it. This phenomenon is a dis­ and solids, including polygons, polyhedra, and quadric ob­
crete disagreement with the continuous case wherein there jects.s Efficient algorithms have been developed for voxeliz­
is no way of penetrating a closed curve without intersecting ing polygons using an integer-based decision mechanism
it. To avoid such a scenariO, it has been the convention to embedded within a scan-line filling algorithm"; for paramet­
define "opposite" types of connectivity for the white and ric curves, surfaces, and volumes using an integer-based
black sets. Opposite types in 2D space are 4 and a, while forward-differencing technique7; and for quadric objects
in 3D space 6 is opposite to 26 or to 1a. such as cylinders, spheres, and cones using "weaving" al­
Assume that a voxel space, denoted by 1, includes one gorithms by which a discrete circle/line sweeps along a dis­
subset of black voxels S. If 1- S is not N-connected, that crete circle/line.' Figures 2-12 in the main text consist of a
is, 1 S consists of at least two white N-connected compo­
- variety of objects voxelized using the above methods.
nents, then S is said to be N-separating in 1. Loosely These pioneering attempts should now be followed by en­
speaking, in 2D, an a-connected black path that divides the hanced voxelization algorithms that in addition to being effi­
white pixels into two groups is 4-separating and a 4-con­ cient and accurate will also adhere to the topological re­
nected black path that divides the white pixels into two quirements of separation, coverage, and minimality.
groups is a-separating. There are no analogous results in
3D space. For example, Figure D shows a 6-connected set References
of voxels that is not 6-separating (Figure 01) and a 6-con­
1. A.S. Glassner, "Space Subdivision for Fast Ray TraCing," IEEE
nected set of voxels that is not la-separating (Figure D2). Computer Graphics and Applications, Vol. 4, No. 10, Oct. 1984. pp.
Let A be an N-separating surtace. A voxel pE A is said to 15-22.

be an N-simpJe voxeJ if A - P is still N-separating. An N­ 2. Y.T. Lee and AAG. Requicha, "Algorithms lor Computing the Vol­
separating surface is called N-minimaJ if it does not contain ume and Other Integral Properties of Solids: I - Known Methods and
any N-simple voxels. A cover of a continuous surface is a Open Issues; II - A Family 01 Algorithms Based on Representation
Conversion and Cellular Approximation," Comm. ACM, Vol. 25,
set of voxels such that every point of the continuous sur­ No.9, Sept. 1982, pp. 635-650.
face lies in a voxel of the cover. A cover is said to be a min­
3. P.E. Danielsson, "Incremental Curve Generation," IEEE Trans.
imal cover if none of its subsets is also a cover. The cover Computers, Vol. C-19, 1970, pp. 783-793.
property is essential in applications that use space subdivi­
4. W. Mokrzycki, "Algorithms of Discretization 01 Algebraic Spatial
sion for fast ray tracing. 1 The subspaces (voxels) that con­
Curves on Homogeneous Cubical Grids," Computers and Graphics,
tain objects must be identified along the traced ray. Note Vol. 12, Nos. 3/4, 1988, pp. 477-487.
that a cover is not necessarily separating, while on the other
5. A. Kaufman and E. Shimony, "3D Scan-Conversion Algorithms for
hand, as mentioned above, it may include simple voxels. In Voxel-Based Graphics," Proc. ACM WorkshOp Interactive 3D
fact, even a minimal cover is not necessarily N-minimal for Graphics, ACM, New York, 1986, pp. 45-76.

any N. 6. A. Kaufman, "An Algorithm for 3D Scan Conversion 01 Polygons,"

Proc. Eurographics 87, North Holland, Amsterdam, Aug. 1987, pp.
197-208.
Voxelization algorithms. In the past, digitization of sol­
ids was pertormed by spatial-enumeration algorithms that 7. A. Kaulman, "Efficient Algorithms for 3D Scan Conversion of Para­
use point or cell classification methods in either an exhaus­ metric Curves, Surfaces, and Volumes," Computer Graphics (Proc.
Siggraph 87), Vol. 21, No.4, July 1987, pp. 171-179.
tive fashion or by recursive subdivision.2 However, subdivi­
sion techniques for model decomposition into rectangular 8. D. Cohen and A. Kaufman, "Scan-Conversion Algorithms for Linear
and Quadratic Objects," in Volume Visualization, A. Kaufman, ed.,
subspaces are computationally expensive and thus inappro­
IEEE CS Press, Los Alamitos, Calif., Order No. 2020, 1990, pp.
priate for medium- or high-resolution grids. Instead, objects 280-301.

Authorized licensed use limited to: Imperial College London. Downloaded on May 30,2024 at 11:54:32 UTC from IEEE Xplore. Restrictions apply.
 Table 1. A comparison between vector graphics and raster graphics.

Capability Vector Graphics Raster Graphics

1. R ende rin g and screen re fresh R en de ring is embed de d in screen Scan conversion is decoupled from
refresh screen refresh

2. Rende rin g performance Sensitive to scene and object Insensitive to scene and object
comp lexity complexity

3. Memory and processing Variable: depends on scene and Large but constant
re q uire men t object complexity

4. Screen-space alias in g :'-lone Frequent

5. Transformation Continuous: performed on the Discrete: performed on pixel

geometric definition of objects blocks (windows)

6. Boolean and block operations Difficult: must be performed Trivial: by using bithlt. pixel -b y­
analytically pixel operation, aggregation,
quadtrees

7. Rendering of interior :--10: boundary only Yes: colored, shaded, and textured
surfaces

8. Adequacy for sampled :--10 Yes

digital images

9. Measurements (for example, Analytical. but often complex Discrete approximation.

di stance . are a ) but simple

line drawings (row 7 in Table 1). The th u s p rovides the ideal environment for ished in the past decade. making sur­
main d is advanta ges of this approach are mixing digital im age s with synthetic face graphics the state o f the art in 3D
thc aliasing present in the image. result­ graphi cs (row 8). These advantages. graphics. III
ing from the discrete nature of the rep­ coupled with advances in hardware and Surface graphics strikingly resembles
resentation (row 4). and the l a rge mem­ the de velopmen t of antialiasing m eth­ vector graphi cs in many way s. Both rep­
ory and processin g power required (row ods. have made raster graphics the pri­ resent the scene as a sct of geometric
3). These two difficulties delayed the mary technology for computer graph­ primi tives kept in a display list. In sur­
full acceptance of raster graphics until ics. face graphics. these primitives are tran s ­
cheaper and faster memory and hard­ formed, mapped to the screen coordi­
ware were available to support the de­ nates, and converted by scan-conversion
mand, of this appro a ch . In addition. the From surface graphics al gori thm s into a discret e set of pixels,
discrete nature of rasters makes them to volume graphics which is stored in the frame buffer. This
less suitable for geometric operations digitization process is also called raster­
such as transformations (row 5) and izatio/1 or pixeLization. Any change to
accurate measurements (row 9). The object-based approach of vector the sccne. viewin g parameters, or shad­
On the other hand, raster graphics graphics has been adapted for 3D graph­ in g parameters requires the i m a ge gen­
has the advantage that it dec oup les im­ ics at the expense of m ain taining and eration system to repeat these opera­
a ge g eneration from screen refresh (row manipulating a display list of ge om etric tions a nd reprocess the comp lete scene
1). thus m akin g the refresh task insensi­ obj ec ts a n d re ge n e rati ng the frame description. Surface graphics gener­
tiv'e to scene complexity (row 2). In buffer after every change in the ,cene or ates mer ely the surfaces of 3D solid
addition. the raster representation lends v i ewi ng parameters. This approach. obje cts viewed from a given direction:
itself to block operations, such as bitblt called surface graphics. combines raster subject to limitations similar to those of
(bit block transfer). in which a WIndow technology for the di sp l ay and an ob­ vector graphICS, It does not support ren­
or a r ectang ular block of pixels can be ject-based approach for thc rcpresenta­ dering of the interior of these 3D ob­
rapidlv transferred with a variety of tion. manipulation, and rende rin g of jects.
pixel-by-pixel operations between the 3D scenes. The method is supported by Instead of a list of geometric objects,
source and destination blocks (row 6)1l powerful geometry engines, which con­ volume graphics uses a 3D volume buff­
Raste r graphics is al so suitable for dis­ stitute the present hardware for poly­ er as a me dium for the re pre se ntation
playing 2D sample d digital im ages. and gon r end erin g. Such engi nes have f1our- and manipulation of 3D scenes. A 3D

56 COMPUTER

Authorized licensed use limited to: Imperial College London. Downloaded on May 30,2024 at 11:54:32 UTC from IEEE Xplore. Restrictions apply.
Table 2. A comparison between surface graphics and volume graphics.

Capability Surface Graphics Volume Graphics

I. Rendering performance Sensitive to scene and object Insensitive to scene and ohjeet
complexity complexity

2. Memory and processing Variable: depends on scene and Large but constant
requirement object complexity

.'. Object-space aliasing None Frequent

4. Transformation Continuous; performed on the Discrete; performed on voxel

geometric definition of objects subvolumcs

5. Scan conversion and rendering Pixelization is embedded in viewing Voxelization is decoupled from
viewing

6. Hookan and block operations Difficult: must be performed Trivial; by using voxblt. voxel-by­
analytically voxel operation. aggregatioll, octrees

7. Rendering of interior and No: surfaces only Yes; rendering of inner structures
amorphous phenomena as well as surfaces

8. Adequacy for sampled data and Partially and indirectly (fitting Supports representation and direct
intermixing with geometric data followed by surface rendering) rendering

9. M easurements (for example. Analytical. but may be complex Discrete approximation.

distance. area. volume. normal) hut simple

10. Viewpoint dependency Requires rccalculation for every Precomputes and stores viewpoint­
viewpoint change independent information

scene is discretized earlier in the image and information on both the interior Disadvantages of
generation pipeline. and the resulting and exterior of 3D objects.
volume graphics
3D discrete form is used as a database of Several disadvantages of thi;, approach
the scene for manipulation and render­ are related to the discrete nature of the
ing purpo;,es. which in effect decouplcs representation . namely. that transfor­ Table 2 contrasts volume graphics with
discretization from rendering (viewing mations and shading are performed in surface graphics. The disadvantages of
and shading). Furthermore. all objects discrete space. In addition. this approach volume graphics are due to its discrete
are converted into one uniform meta­ requires substantial amounts of storage form (rows 3 and 4 in Table 2). the loss
object - the voxel. Each voxel is atom­ space and specialized processing. The of geometric information (row 9). and
ic and represents the information about advantages and disadvantages, as com­ the memory and processing power it
no more than one object that resides in parcd with surface graphics. are dis­ requires (row 2).
that voxel. cussed in detail in the following sec­
V olume graphics offers the same ben­ tions. Discrete form. Unlike surface graph­
efits as surface graphics. with several The same forces that drove the evolu­ ics. volume graphics represents the 3D
advantages due to the decoupling. uni­ tion of computer graphics from vector scene in discrete form. This causes many
formity. and atomicity features. The graphics to raster graphics. once the of the maladies ofvoxcl-bascd graphics.
rendering phase is viewpoint indepen­ memory and processing power became which are similar to those of 2D ras­
dent and insensi tive to scene and object available. are starting to drive a variety ters.11 The finite resolution of the raster
complexity. It supports Boolean and of applications from surface-based rep­ poses a limit on the accuracy of opera­
block operations and constructive solid resentation of 3D scenes to voxel-based tions that are based on voxel counting
modeling. When 3D sampled or simu­ representation. Naturally. this trend first (row Y in Table 2), for example. volume
lated data is available. such as that gen­ appeared in applications involving sam­ and area measurements. It is difficult to
erated by medical scanners (for exam­ pled 3D data. such as medicine and sci­ manipulate and transform the discrete
ple. i n compute d tomography and cntific visualization. in which the data volume without degrading the image
magnetic resonance imaging) or scien­ sets are in volumetric form. The diverse quality or losing some information (row
tific simulations (for example. compu­ empirical im agery applications of vol­ 4). Rotation of rasters by angles other
tational fluid dynamics). volume graph­ ume visualization still provide a major than 90 degrees is especially problem­
ics is suitable for its representation. It driving force for advances in volume atic. since a sequence of consecutive
can represent amorphous phenomena. graphics. rotations will distort the image.

July 1993 57

Authorized licensed use limited to: Imperial College London. Downloaded on May 30,2024 at 11:54:32 UTC from IEEE Xplore. Restrictions apply.
 2). However. since computer memory
prices are decreasing significantly while
compactness and speed are improving.
Figure 2. A T62
such large memories arc becoming more
and more feasible. A similar argument
tank made of abont
40,000 voxelized
occurred whcn raster graphics emerged
polygons, partially
as a technology in the m id-seventies.
occluded by a tree
With greater compactness accompany­
made of 20,000
ing the rapid decline in memory price. it
fractally grown,
is safe to predict that. just as with raster
voxelized polygons.
graphics. memory will soon cease to bc
The tank resolution
a stumbling block for volume graphics.
is about 1.5 inches
Nevertheless, the extremely large
per voxel i n the
throughput that must be handled re­
foreground, which
quires a special architecture and extra
translates to about
attention to processing. Volume engines.
a 256'·voxel model.
analog, to the currently available ge­
ometry engines. are emerging (Kauf­
man.1 Chapter 6). Because of the "pre­
Since the continuous object is recon­ dering discrete surfaces. An essential sortedness" of the volume buffer and
structed by sampling the discrete data requirement for most shading methods the fact that only a single type of ob­
duri ng rendering. a low-resolution vol­ is the capability to calculate the normal ject - the voxel - must be handled.
ume yields high aliasing artifacts (row 3 vector to the surfaces making up the 3D volume engines are conceptually sim­
in Table 2). This becomes especially 'cene. In traditional surface graphics, pler to implement than current geome­
apparent when zooming in on the 3D normal vectors are either analytically try engines. Consequently. we predict
raster. When naive rendering algorithms calculated from the surface representa­ that the volume engines that will mate­
are used. the 3D discrete points may tion or stored as part of the surface rialize i n the near future will have capa­
appear to be parted from each other representation. In voxel-ba,ed models. bilities to synthesize. load. store. ma­
and may cause the appearance of holes. a discrete shading method is used to nipulate. and render volumetric scenes
This can be alleviated to some extent by estimate the normal from a context of in real time (30 frames per second) and
methods similar to those adopted in 2D voxels. A variety of image-based and may be configured as accelerators or
raster graphics. such as using either re­ object-based methods for normal esti­ cosystems to existing geometry engines.
construction techniques (supersampling mation [rom volumetric data have been
or filtering. for example) or a high-res­ devised (Kaufman.1 Chapter 4): most of
olution volume buffer. these methods are based on fitting some Advantages of volume
type of surface primitive to a small neigh­ graphics
Loss of geometric information. In borhood of voxels. Nevertheless. this
volume graphics we allow each voxel to subject is still an active field of research.
m aintain only local information per­ A partial integration between sur­ The advantages of volume graphics
taining to the volume unit it represents. face an d volume graphics is conceivable include its insensitivity to scene and
After a surface obj ect has been voxel­ as part of an object-based approach in obje ct complexity (row I in Table 2). its
i7ed. the voxels constituting the dis­ which an auxiliary object table. consist­ viewpoint independence (rows 5 and
crete object do not retain any geometric ing of tbe geome tric definition and glo­ 10). its capability to represent sampled
information regarding the object's bal attributes of each object. is main­ and simulated data sets (row 8). its ca­
surface definition. Thus. when exact tained in addition to the volume buffer. pability to represent inner information
measurements arc required (say. for Each voxel consists of on ly an index to and amorphous phenomena such as
distance. volume. or area). it is advanta­ the object table. allowing exact calcula­ clouds and smoke (row 7). and its capa­
geom to use a surface-based model. tion of normal. exact measurements. bility to support various block opera­
where the geometric surface definition and intersection verification for discrete tions (row 6).
of the object is available. A voxel-based ray tracing (sec "Discrete ray tracing"
object is only a discrete approximation sidebar). The auxiliary geometric infor­ Insensitivity to scene complexity. One
of the original continuous object: the mation might be useful for revoxelizing of the most appealing attributes of vol­
volume buffer resolution determines the the scene in case of a change in the ume graphics is its insensitivity to the
precision of such measurements. On the scene itself. eomplexitv of the scene. since all ob­
other hand. several mea surement types jects have been preconverted into a
are more easily computed in voxel space: Memory and proressing. A typical finite-size volume buffer (row I in Ta­
for example. mass property. adjacency volume buffer occupies a large amount ble 2). Although the performance of the
detection. and volume computation (row of memory: for example. for a moderate voxelization phase is influenced by the
9 in Table 2). resolution of 5 12·'. the volume buffer scene complexity. rendering perfor­
The lack of geometric information in consists of more than I O� voxels. Even if mance depends mainly on the constant
the voxel may lead to other difficulties. we allocate only one byte per voxel. 128 resolution of the volume buffer and not
such as those cncountered whcn ren- Mbytes will be required (row 2 in Table on the number of objects in the scene.

58 COMPUTER

Authorized licensed use limited to: Imperial College London. Downloaded on May 30,2024 at 11:54:32 UTC from IEEE Xplore. Restrictions apply.
 Figure 3. A teapot modeled by 32
Rezier patches and voxclizcd to a
5 123-resolution voxel model. The gen­
erated voxels were assigned color dur­
ing the voxelization process according
to a 2D texture map, while the shad­
ing was applied during the rendering
process. Figure 4. A camouflaged T62 tank voxelized from II 50,OOO-polygon model into
a 2-inch-per-voxel resolution. Voxels were given color during the voxelization
process according to a spherical texture map of pseudocamouflage. The tread is
This contrasts with volume representa­ invisible beclluse it blends witb the background except over the back sprocket.
tion via an octree whose size varies ac­
cording to the scene complexity.' Insen­
sitivity to scene complexity makes the technique known as texture mapping, value. texture mapping n eed not be rc­
volumetric approach especially attrac­ whereby object realism is i ncreased by peated.
tive for scenes consisting of many ob­ simulating surface details. Texture map­
jects. such as thosc generated by fractal ping is commonly implemented during Viewpoint independence. A main dif­
systems (see Figure 2). Another exam­ the last stage of the rendering pipeline, ference between voxel-based graphics
ple of such a scene is a curved surface where the texture is extracted from a and conventional surface graphics is that
represented by a large polygon mesh 20 texture image and mapped onto the in the former the scene is discretized
that is generated by a polyhedral­ surface to be rendered: its complexity is (voxelized) once for multiple viewing
smoothing or -fitting algorithm. A poly­ proportional to the object complexity 111 conditions. while in the latter the scene
gon mesh can approximate a curved In volume graphics. texture mapping is is repeatedly scan-converted after ev­
surface. and the approximation preci­ performed only once, during the voxel­ ery change i n the viewing parameters,
sion and presentation quality increase ization stage, where the texture color is causing a performance hottleneck in its
with the number of polygons in the mesh. calculatcd and stored in each voxel. Solid rcndcring pipeline (row 10 in Table 2).
However. using a very fine mesh in con­ texturing. which uses a 30 texture im­ This attractive advantage of volume
ventional surface graphics is expensive age. also has a high complexity similar graphics exists partly because in the
in terms of space and display time. to that of texture mapping. III In volume volumetric representation a u n i t of
graphics. however. solid texturing, like memory is allocated for each unit of
Insensitivity to obj ect complexity. In texture mapping. is performed during space: in surface graphics, memory is
vulume graphics. rendering (viewing and the voxclization stage. assigned only to complete surface
shading) is deeoupled from digitization The textured objects in Figures 3 and patches. This enables volume graphics
(voxelization). and all objects are first 4 have heen assigned texture during the to store view-independent attrihutes at
converted into one metaobject. the voxelization stage by mappingeach voxel each volume unit, while surfacc graph­
voxel. which makes the rendering pro­ back to the corresponding value on a ics is not capable of providing storage
cess insensitive to the complexity of the texture surrounding the ohject. Figures for attrihutes that vary across its hasic
objects (row 5 in Table 2). Thus, volume 10-12 show voxclizcd terrain that has surfacc clcmcnts.
graphics is particularly attractive for been mapped with satellite or aerial In anticipation of repeated access to
objects that are hard to render using photos during the vuxeliLatiun stage. the vulume buffer (in animation, for
conventional graphics systems. Exam­ We have also implemented a photo­ cxample), all viewpoint-independent
ples of such objects include curved sur­ mapping technique whereby six pho­ attributes can be precomputed during
faces of high order and fractals. which tographs of the real object are project­ the yoxeliwtion stage and stored with
require the expensive computation of ed back onto the voxelized object (see the yoxel. making them readily accessi­
an i terative function for each volume Figure 6 and the building in figure 1 2). ble for speeding up the rendering. The
unit (see Figure 2). Constructive solid Once this mapping is applied. i t is voxelization algorithm can generate for
models are also hard to render hy con­ stored with the voxels themselvcs dur­ e ach objcct voxel its color, its texture
ventional methods but arc straightfor­ ing the voxelization stage: thus, the ren­ color. its normal vector (for visible vox­
ward to render in volumetric represen­ dering performance is not degraded. els). and information concerning the
tation. (A separate discussion appears In addition, texture mapping and pho­ visibilitv of the light sources from that
later under "Block operations. ) " tomapping are also viewpoint-indepen­ voxel. Actually, the viewpoint-indepen­
Another type of object complexity dent attributes, i mplying that once the dent parts of the illumination equation.
involves objects enhanced through a texture is stored as part of the voxel that is. the amhient illumination and the

July 1993 S9

Authorized licensed use limited to: Imperial College London. Downloaded on May 30,2024 at 11:54:32 UTC from IEEE Xplore. Restrictions apply.
 Discrete ray tracing

Ray traci ng' is an image generation technique that tio n , the discretization step is, of course, u n n ecessary.
s i mulates light's behavior in a scene by following sight I n the second phase, a discrete variation of the con­
rays from the observer's eye as they interact with the ventional recursive ray tracer is used, R RT differs from
scene a n d the light sources. The refore, the basic compu­ conventional ray-tracing algorithms, in which analytical
tation performed by a ray tracer is the calculation of the rays are intersected with the object list to find the clos­
intersection points between rays and objects. This com­ est intersection. I n RRT, 3D d i screte rays (which are
putationally expensive operation accounts for ray trac­ essentially voxelized Imes) are traversed through the
i ng's reputation as a costly method for image generation, 3 D raster to find the fi rst s u rface voxel. Encountering a
though it produces superior image quality. nontransparent voxel i n d icates a ray-surface hit. The
A major strategy for reducing the cost of ray tracing is view-independent attributes stored i n the voxel (for ex­
to divide the world space into a set of rectangular subvol­ ample, normal) are readily available for calculating
umes called cells 2 A cell consists of a list of all objects spawning offspring rays and computing the i l l u m i nation
(partially) residing i n that region of the world. Rays are at that point. Figures 7 and 8 i n the main text show the
then traced as they travel from cell to cel l . Only the lists R RT rendering of both a sampled data set intermixed
of objects residing in the visited cells are cand idates for With geometric objects and manipulated with block op­
ordered ray-object intersection calculation . The approach erations (Figure 7) and a voxelized geometric model
taken in this sidebar follows a regular space-subdivision (Figure 8 ) .
scheme; that is, all cells are equal-sized cubes, Alterna­ I n conventional ray tracing, computation t i m e grows
tively, the octree approach divides the space into equal­ with the n u mber of objects, and performance is greatly
sized octants. Each octant is independently and recur­ influenced by the type of objects constituting the scene,
sively subdivided into additional octants until some I ntersection calculation between a ray and a parametric
measure of u niformity or simpl icity of the scene enclosed surface is significantly more complex than intersecting
by the octant is satisfied. the ray with a sphere or a polygon. I n contrast, RRT
When considering any subdivision approac h , two is­ elimi nates the com putationally expensive ray-object in­
sues must be resolved: f i rst, how to assign each cell its tersection calculation and i n stead relies solely on a fast
list of objects, that is, how to effiCiently compute cell­ discrete ray-traversal mechanism a n d a single simple
object intersection; and second, how to step along the type of object - the voxel. Consequently, RRT perfor­
ray as it crosses cell boundaries, that is, how to efficient­ mance is effectively independent of the n u m ber of ob­
ly compute ray-cell intersection. Although in some appli­ jects i n the scene or the objects' complexity or type,
cations the octree presents a compact representation The refore, for a given resolution, ray-tracing time is
scheme, it does not perform as well i n the case of sam­ nearly constant and can even decrease as the number
pled data. Moreover, because it lacks u n iformity, it does of objects i n the scene increases, since less stepping is
not perform the above two tasks as efficiently as the uni­ necessary before a n object is encou ntered.
form subdivision approach . RRT can precompute the view-independent attributes
Armed with a large memory, we can take the space­ during the voxelization phase, is attractive for ray trac­
subdivision approach to an extreme by subdividing i nto a ing 3 D sampled data sets and computed data sets: a n d
high -resolution u n iform grid. This, in turn, can allow us to can su pport ray tracing of constructive s o l i d geometry
transform o u r internal scene representation from a 3D models that are efficiently synthesized and constructed
array of cells into a vol ume buffer of voxels. The major during the voxelization phase. I n short, RRT is a typical
distinctions between a cell and a voxel lie i n the assump­ example of a vol ume graphics technique and demon­
tions made concerning voxels. First, it is assumed that strates the attractive features intrinsic to the volumetric
the voxel grid resolution is adequate: that is, there is no approach to 3D graphics.
large variation of surface attributes across the voxel ex­
tent. In addition, un like a cel l , Which consists of a list of
objects, the voxel represents the smallest space-enu­ References
meration entity and therefore consists of i n formation on a
single object, much like its 2D counterpart, the pixel. 1 A.S, Glassner. ed., An Introduction to Ray Tracing, Ac adem ic
Press, London, 1 989.
A ray-traci n g approach that employs a volume buffer is
called 3D raster ray tracing (RRT).' It operates i n two 2. A. Fujimoto, T, Tanaka, and K. Iwata. "ARTS: Accelerated Ray­
phases: a preprocessing voxelization phase and a dis­ Tracing System." IEEE Computer Graphics and Applications,
crete ray-tracing phase. In the voxelization phase the Vol. 6, N o . 4 , Apr. 1 986, pp, 1 6-26,

geometric model is digitized using voxelization algo­

3. R , Yagel, D . Cohen, and A , Kaufman, "Discrete Ray Traci ng, '
rithms, which convert the continuous representation of IEEE Computer Graphics and Applications, Vol. 1 2, NO. 5 .
the model into a discrete representation that is stored in Sept . 1 992. p p . 1 9 -28.

the volume buffer (see the sidebar "Fundamentals of

4. R , Yagel, A , Kaufman, and Q . Zh a n g , "Realistic Volume Imag­
voxelization"), For data sets that are already digitized, as
ing," Proc. Visualiza tion 91, I E E E CS Press, Los Alamitos, Ca­
in 3D medical i m ag i n g and 3D computational visualiza- lif" Order No. 2245, 1 991 , pp. 226-231

6U CO!vlPl. l E R

Authorized licensed use limited to: Imperial College London. Downloaded on May 30,2024 at 11:54:32 UTC from IEEE Xplore. Restrictions apply.
 sum of the attenuated diffuse illumina­
t i o n of all
the visible light sources, 1 0 can
also be precomputed and stored as part f'igure 5. The nu­
of the voxel value. clear power plant
Once a volume buffer with precom­ from San Onofre,
puted view-independent attributes is California, embed­
available, a rendering algorithm such as ded in a terrain
a discrete ray-tracing algorithm can b e database of up­
engaged. Discrete r a y tracing is based state New York.
on traversing 3D discret e rays through This image was
the volume buffer (see the sidebar "Dis­ voxelized from a
crete ray tracing"). The discrete ray­ plant model con­
tracing approach is especially attractive sisting of about
for ray tracing complex surface scenes 45,000 polygons
and constructive solid models, as well as into 2,000 x 700
3D sampled and computed data sets voxels horizon­
( see below ) . Figures 7 a n d 8 show exam­ tally with a I-foot­
pies of objects that were voxelized and per-voxel resolu­
thcn ray traced in discrete voxel space. tion. The terrain
In spite of the complexity of these scenes. resolution is about
ray t racing time was approximately the
- 6 inches per voxeI.
same as for much simpler scenes and
significantly shorter than in traditional
space-subdivision ray-tracing methods. fluid dynamics, are often reconstructed Inner information. A central feature
Moreover. in spite of the discrete na­ from the acquired sampled or simulated of volumetric rcpresentation is that un­
turc of the volume buffer representa­ points into a regular grid of voxels and like surface representation. it can rep­
tion. images indistinguishable from those stored in a volume buffer. Such data resent inner structures of objects, which
produced by conventional surface-based sets provide for the majority of applica­ can be revealed and explored with the
ray tracing can be generated by using tions using the volumetric approach. appropriate manipulation and render­
auxiliary object tables and screen su­ Unlike surface graphics, volume graph­ ing techniques (row 7 in Table 2). Nat­
persampling techniques, for example, ics narurally and directly supports the ural objects as well as synthetic objects
by casting several rays per pixel (see representation, manipulation, and ren­ are likely to be solid rather than hol­
Figure 8). dering of such data sets (row 7 in Table low. The i nner structure is thus an i m­
2). as well as providing the volume buff­ portant aspect of image complexity and
Sampled and simulated data sets. Sam­ er medium for intermixi ng d a t a sets. is easily explored using volume graph­
pled data sets. like those i n 3D medical such as ,ampled or simulated data sets. ics (see Figure 7). Moreover. while
imaging (see Figure 7), and simulated with geometric objects (row S),' as shown translucent objects can be represented
data sets. like those in computational in figures 5, 7, and 9. by surface methods, these methods can-

Figure 7. A 2563-resolution volume of

a reconstructed MRI head reflected in
a voxelized mirror. The cut was per­
formed by a constructive solid model­
Figure 6. A Tomcat aircraft voxelized in 3-inch-per-voxel resolution. Note the ing subtraction (voxel by voxel) dur­
pilots' helmets visible through the translucent canopy. ing the voxelization stage.

July 1 993 61

Authorized licensed use limited to: Imperial College London. Downloaded on May 30,2024 at 11:54:32 UTC from IEEE Xplore. Restrictions apply.
 Figure 8. Turner Whitted's spheres
and a checkerboard floor voxelized
into a 320' volume and ray traced in
377 seconds on a 20-MIPS machine
with geometric intersection verifica­
tion and supersampling by casting
four rays per pixel (see the sidebar
"Discrete ray traciug").
Figure 9. A voxelized Tomcat (see Figure 6) flying over Yosemite Valley (see
Figure 11) in a flight simulation of a dogfight.

not efficiently support the modeling

and rendering of amorphous phenom­
ena (for example. clouds. fire. and
smoke) thst are volumetric in nature
and do not actually have tangible sur­
faces." Figure 10 exemplifies the ren­
dering of haze as part of a voxel-based
terrain model.

Block operations. An intrinsic char­

acteristic of rasters is that adjacent ob­
j ects in the scene are also represented
by neighhoring memory cells. There­
fore. rasters lend themselves to various
meaningful grouping-based operations.
such as bitblt operations. or their 3D
counterpart. voxblt (voxel hlock trans­
fer ) operations. I' Thesc operations
support transfer of cuboidal voxel blocks
with a variety of voxel-hy-voxel opera­
tions between source and destination
blocks (row 6 in Table 2). Such block
operations add a variety of modeling
capabilities that aid i n the task of image
synthesis. Moreover, the volume buffer
lends itself to Boolean operations that Figure 10. View west over Camp Pendleton, California, with moderately dense
can be performed on a voxel-by-voxel haze. Resolution is about 16 feet per voxel.
basis during the voxelization stage.
This property is very advantageous
when constructive solid geometry is the ean operation between pairs of voxels structed M R I head with a CSG subtrac­
modeling paradigm. CSG operations during a preprocessing stage. Once a tion and a back mirror that has been
such as subtraction. union. and inter­ CSG model has been constructed in generated by a polygon voxelization al­
section between two vox eli zed ohjects voxel representation, it is rendered like gorithm (see the sidebar "Fundamen­
are accomplished at the voxel level': any other volume b utler. This makes tals of voxelization").
thus. the original problem of evaluating discrete ray tracing of constructive solid The spatial presortedness of the vol­
a CSG tree of such operatiom d uring models straightforward. Figure 7 shows ume buffer voxels lends itself to other
rendering time is reduced to a I D Bool- a volumetric ray tracing of a 2563 recol1- types of grouping or aggregation ot

()2 COMPUTER

Authorized licensed use limited to: Imperial College London. Downloaded on May 30,2024 at 11:54:32 UTC from IEEE Xplore. Restrictions apply.
Figure 1 1 . A typical view of a voxelized terrain model in Figure 12. A typical view of voxelized terrain with an em­
flight simulation. This image is a view dowu Yosemite Val­ bedded vllxclizcd hotcl building and voxelized trees.
ley showing Cathedral Rocks; each voxcl represents 8 cu­
bic feet.

A
neighboring voxels. For example. the s Tahle 2 s h ow s v o l u m e
. rendering synthetic scenes composed of
terrain I mages shown i n Figures 2. 5, 9. graphics h a s advantages over surfaces • .

10, and 11 were generated by the Hughes surface graphics by b e i ng

A ircraft RealScene flight simulator' viewpoint i n d e p e nd e nt i n sensitive to
,

and that in Figure 12 by the Tiltan Sys­ scene and object complexity. and suit­
tem Engineering Scene Generator. Both able for the representation of sampled Acknowledgments
systems simulate a flight over voxel­ and simulated data sets. Moreover,
represented terrain enhanced by satel­ geometric objects can be mixed with This work has b e e n s u pported by the Na­
tional Science Foundation under grants IRI-
lite photomapping with addition al syn­ these data sets. Volume graphics sup­
9008 1 09 . CCR-9205047, and CCR-92 1 1 288.
thetic raised objects, such as buildings, ports the visualization of internal struc­ and grants from Hughes A ircraft Company
trees. vehicles, aircraft, and clouds (see tures and lends itself to the realization and Hewlett·Packard. Figure 12 is courtesy
Figures 5 a nd 1 2 ) . Since i n t h i s a p pl ic a ­ of hlock operations, CSG modeling. of Tiltan System E ngine eri ng. Figures 2. 4.
tion th e information below the terrain ) , ii, 9, 10, and I I are courtesy of H ugh e s
irregular voxel sizes. and h i e rarc h i ­
Aircraft Company. These images have been
surface is invisible . terrain voxels can c a l representation. T h e prohlcms as­
voxelized using voxelization al gori t h m s . a
be represented as tall cuhoids extend­ sociated with the volume buffer rep­ voxel-bascd modeler, and a photomapper
ing from se a level to the terrain height. resentation are remin i sc en t of those en­ devcloped a t the State U n iversity of :-..l e w
This representation saves storage space c o u n t e r e d w h e n rast e r graph ics York at S t on y B rook .

as well as retrieval and proces si n g time. emerged as an alternative to vector

The raised objects. however, m ust be graphics, and they can be alleviated in
rep re s e n t ed in a conventional voxeI­ similar ways.
based form. Progress i n computer hardware and References
Simi larly. voxe l s ca n be aggregated memory systems, coupled WIth the de­
into supervoxels in a pyramid-like hier­ sire to reveal the inner structures of I . A . Kaufman. Volume Visualization, IEEE

arc h y For example. in a voxel-hased CS Press. Los A lamitos, Calif., Order

. volumetric objects. i mpl i es an expand­
No. 202U. 1 990.
flight simulatur. voxels representing I i ng role for volume graph ics. In f a c l , the
cuhie foot can be used for takeoff and striking s i m i l a r i t i e s between Tables I 2. D . l . Meagher. " Geometric Modeling
la n d i ng . As the aircraft ascends, fewer a n d 2 suggcst that volume graphics has U si n g O c t r e e E nc o d i n g . " Computer
and fewer details need to be processed the potential to revolutionize computer Graphic; and Image Processing. Vul. 19.

and visualized. so a 1 6-cubic-foot or No. 2, June 1 982, pp. 1 29- 1 47 .

graphics. I n t h e seventies, raster graph­
larger voxel size will suffice. A h i e r ar­ ics superseded vector graphics for visu­
3. K . Perlin a n d E . M . Hoffert, ··Hypertcx­
chical volume b u ffer can be prepared in alizing surfaces; similarly. volume graph­ ture," CompulerGraphics (Proc. Siggraph
advance or on-the-f1y by subsampling ics has the potenlial to supersede surface 8 9 ) , Vol. 23. :-..1 0. 3. July 1 9S9. pp. 253-262.
or averaging the appropriate size ne i gh ­ gra p h ics for handling and v i sual i zing
borhoods of voxels. volumes as well as for modeling and 4 . .T.T. Kajiya and T.L Kay. "Rendering

Julv 1 993 63

Authorized licensed use limited to: Imperial College London. Downloaded on May 30,2024 at 11:54:32 UTC from IEEE Xplore. Restrictions apply.
 Fur with Three-Dimensional Text ures ," 1 2. A K a ufman , "The Voxblt Engine: A
Cumputer Graphics (Proc. Siggraph 89), V ox e l Frame Buffer Processor," in A d­
Vol. 23. No. 3, J u ly 1989. pp. 271 -280. vances in Graphics Hardware lII. A.A.M.
Ku ijk , ed . • S p ri n ger- Verlag. Berlin. 1992,
pp. 85-102.
5. D . S . Ebert and R . E . Pa rent, " R end ering
and Animatio n of Gaseous Phenomena
by Com b i ning Fast Volume and Sc an li ne
A-bufferTechniques," Computer Graph­
ics (Proc. Siggrap h 90), Vol. 24. No. 4,
Aug. 1 990. pp. 367-376.
Daniel Cohen is a le ctu rer in t h e Depart­
ment of Computer Science at B e n-G uri on
6. J. Wright and J. Hsieh, "A Voxel-Based.
Unive rsity. Beer-Sheva, and at the School of
Forw ard P roj ecti on A lgorith m for Ren­
Mathematics a t Tel-Aviv U n i v e rs ity, Israel.
deri ng Surface and Vo lumetri c Data,"
Hc is also devel opin g a real-time ray tracer
Pmc. Visualization 92, IEEE CS Press,
of te rrain systems at M i l i ko n Ltd. Previous­
Los Alamitos, Calif.. Order No. 2897,
ly. he was a softwa re engin e er working on
1 992, pp. 340-348.
bitmapped gra p hic s . His research interests
include re n d eri ng techni q ues , volume visu­
7. A. Kaufman, R. Yagel , and D. C oh e n, alization. an d architectures and a lgori t h ms
" In termixing Surface and Vo l um e Ren­ for voxcl-based grap h ics.
deri ng ," in 3D
Im aging in Medicine: Al­ Cohen received a BSc cum laude in math­
gorithms, Systems, A pplicatio n s , K.H. em a ti cs and computer science ( 1 985) a n d an
H o e hn e . H . Fuchs, and S . M . Pizer, eds . , MSc cum laudc in computer s cie nce ( 1 986) .
Springer-Verlag, Berlin. June 1990. pp. both from Ben-Gurion University. and a PhD
2 1 7-227. in comp uter science from t h e State Universi­
ty of New York at Stony Brook ( 1 99 1 ) .
8. R.A. Drebin. L. Carp en ter , a n d P. Han­
rahan. "Volume Rendering." Computer
Graphics (Proc. Sigg r ap h 88). Vol. 22.
Arie Kaufman is a p rofesso r of computer
NO. 4. Ang. 1 988. pp. 65-74. sc i enc e and director of the C ub e proj ect for
vol umc vi s u al ization at th e S tate U n iv ersit y
9. M. Levoy. "Display of Surfaces from of New York a t Stony Brook. He h as con­
Volume Data." IEEE Cumputer Grap h ­ du ct e d res ear ch an d con sulted i n compute r
ics a n d Applications. V o L 8. N o . 3 , May grap h ics for more than 20 years, special iz i ng
1988. pp. 29-37. in volume visualization; gra ph ics arch i te c ­
ture s. a l go rithm s . and languages; u se r inter­
faces: an d mu lt imedia.
10. J . F oley ct al.. Comp uter Graphics: Prin­
Kaufman received a BS in mathematics
ciples and Practice. 2nd ed.. A ddiso n ­
and p h ysi cs from the Hebrew U n iv ersity of
W esle y. Re ad ing, Ma"., 1990.
Je rus a lem in 1969, an M S i n com p uter sci­
ence from th e Weizmann Institute of Sci­ Roni Yagel is an assistant profe ss or in the
1 1 . C..vL E ast man. "Vector versus Raster: A ence. Rehovot . in 1973, and a PhD i n com­ D epa rtment of Com p uter and Information
Functional Com p arison of Drawing Tech­ puter science from Ben-Gurion University . Science at Ohio State Un iversit y . Previous­
nologics . " IEEE Computer Graphics and Israel. in 1977. He ch ai rs the IEEE C omp u t ­ ly. he was a researcher in t h e De p a rt m e n t s of
Applications. VoL 10, N o . 5 , Sept. 1990, er Soci etv ' s Technical Committee on Com­ Comp uter Science. A natomy. and Physiol o­
pp. 68-80. puter G r� ph ics . gy and Biophysics at the State U niversity of
Ne w York at Stony Brook. His resea r c h in­
terests i n cl u d e algorithms for voxel-based
gra p hics . imaging. animation. three-dimen­
,ion al user interfaces. hardware for vo l u m e
viewi ng . and visualilation tool s for bi o med­
ical applicati o ns .
Yagel re cei ved B S c cum laude and MSc
cum lau d e d egre es in mathematics a n d com­
puter science at Ben-Gurinn U niversi t y of
the N e gev. Israel, i n 1 986 and 1987. respec­
tively. He received a PhD from the State
Name (please Print) U niversi ty of "ew Y ork at Stony Brook in
1 99 1 . H e is a membc r o f the I E E E Compute r
PLEASE NOTIFY So ci e ty.
US 4 WEEKS IN
New Address
ADVANCE

Readers can contact Kaufman at t he De­

City State/Country Zip partment of Computer Science. State Uni­
ve rsi ty u f New Y ork at Stonv Brook. Stony
Brook . NY 1 1 7 9 4 -4400. e - m a i l a n @'cs.
su n y sb . e d u .

Bruce Shriver, formerly Complller's ad­

MAIL TO:
v a nc e d te ch n o l o gy area e d i t o r . coordi­
IEEE Service Center
445 Hoes Lane nated and recommended thIS a rt i c l e for
Piscataway, NJ 08854 publicat i o n . His e-mail address is b . s h river0'
compm a i l .com.

64 CO.v!PUTER

Authorized licensed use limited to: Imperial College London. Downloaded on May 30,2024 at 11:54:32 UTC from IEEE Xplore. Restrictions apply.