ICVG : Practical Constructive Volume Geometry for Indirect Visualization

International Journal of Computer Graphics & Animation (IJCGA) Vol.7, No.3/4, October 2017
DOI : 10.5121/ijcga.2017.7201 1
ICVG: PRACTICAL CONSTRUCTIVE VOLUME
GEOMETRY FOR INDIRECT VISUALIZATION
Mark Laprairie, Howard J. Hamilton, and Andrew Geiger
Department of Computer Science, University of Regina, Regina, Canada
ABSTRACT
The task of creating detailed three dimensional virtual worlds for interactive entertainment software can be
simplified by using Constructive Solid Geometry (CSG) techniques. CSG allows artists to combine
primitive shapes, visualized through polygons, into complex and believable scenery. Constructive Volume
Geometry (CVG) is a super-set of CSG that operates on volumetric data, which consists of values recorded
at constant intervals in three dimensions of space. To allow volumetric data to be integrated into existing
frameworks, indirect visualization is performed by constructing and visualizing polygon meshes
corresponding to the implicit surfaces in the volumetric data. The Indirect CVG (ICVG) algebra, which
provides constructive volume geometry operators appropriate to volumetric data that will be indirectly
visualized is introduced. ICVG includes operations analogous to the union, difference, and intersection
operators in the standard CVG algebra, as well as new operations. Additionally, a series of volumetric
primitives well suited to indirect visualization is defined.
KEYWORDS
Constructive solid geometry, constructive volumetric geometry, surface construction, isosurface,
virtual volume sampling function
1. INTRODUCTION
Volumetric data consists of values recorded at constant intervals in three dimensions of space. In
mathematics, volumetric data can be treated as a 3D scalar field. In practise, these data sets are
stored in three dimensional array structures. Such data are utilized in a variety of tasks, including
scientific visualization, medical imaging, and interactive entertainment. A software system that
takes volumetric data, processes them, and produces visualization is called a voxel system. One
method for visualizing volumetric data is to first perform surface construction and then visualize
the resulting surface [1]. Surface construction, also called isosurface extraction or surface
reconstruction, extracts information about surfaces that are implicitly present in the volumetric
data [2] and uses this information to build geometric meshes corresponding to the surfaces.
Visualizing volumetric data through surface construction is referred to as indirect visualization.
In computer graphics, Constructive Solid Geometry (CSG) is a modeling technique that facilitates
the rapid construction of geometric data [3]. Complex scenes can be constructed by applying a
series of Boolean operations, such as union, intersection, and difference, to small sets of 3D
objects. Constructive Volume Geometry (CVG) is an extended version of the CSG technique that
operates on volumetric data instead of conceptually solid objects [4]. The naive approach to
implementing CVG leads to visual artifacts when indirect visualization is performed on the
resulting volumetric data.
This paper focuses on techniques for performing CVG that reduce the number of visual artifacts
when indirect visualization is performed. The discussion is restricted to volumetric data where all
values are non-negative integers. This restricted domain is appropriate for indirect visualization
techniques because they primarily work with density data. The Indirect CVG (ICVG) algebra,
which provides constructive volume geometry operators appropriate to volumetric data that will
be indirectly visualized, is introduced. ICVG includes operations analogous to the union,

2
intersection, and difference operators in the standard CVG algebra, as well as several new
operations. Additionally, a series of volumetric primitives well suited to indirect visualization is
defined.
The remainder of this paper is organized as follows. Section 2 describes background material on
CSG, volumetric visualization, CVG, and voxelization. Section 3 discusses practical issues
related to developing an indirect CVG system, as well as the changes made to the CVG algebra to
create the ICVG algebra. Section 4 presents experimental results. Section 5 concludes and
summarizes the findings. Possible future extensions to the approach are also described.
2. BACKGROUND
This section surveys relevant previous work. Section 2.1 presents the main concepts of CSG and
Section 2.2 describes Boolean operations for CSG. Section 2.3 discusses techniques for
visualizing volumetric data. Section 2.4 describes the CVG algebraic framework. Finally,
Section 2.5 describes methods of voxelization.
2.1. Constructive Solid Geometry
As mentioned previously, CSG is a modeling technique that allows for the rapid construction of
geometric data [3]. The technique's effectiveness depends on the idea of combining well-defined,
relatively simple 3D objects called primitives through the use of Boolean operators. Since well-
defined objects are used as the basis for construction of a 3D geometric object, here called
geometry, the resulting compound object retains the same well-defined nature as the primitives
from which it was made.
The conceptual simplicity of working with a few, relatively simple, primitives enable users to
create models quickly, as is required in the interactive entertainment industry. Having only a few
primitive objects, an artist can easily conceptualize the construction of a scene through its parts.
This conceptualization is similar to how a cartoon artist first defines the main geometric form of a
character and then fills in details.
CSG primitives are easily combined into more complex and interesting shapes called compound
objects. In Figure 1, three common primitives are shown: a sphere, a cube, and a cylinder. These
primitives are easily defined mathematically, and thus the mathematical properties of any
resulting compound object can be readily calculated.
Figure 1. Constructive Solid Geometry primitives
The well-defined nature of CSG is suited to engineering tasks [5]. Using this method of
construction allows artists and engineers to produce geometry that is guaranteed to be solid [3].
For example, if an engineer wishes to design an engine compartment that is water tight in a
Computer Aided Design (CAD) environment, CSG can help. The engineer can start with a solid
cube, and then use other primitives to carve cavities within it. If no action is performed that
pierces the surface of the cube, then the resulting cavity is guaranteed to be fully enclosed and
thus water tight.

3
In interactive entertainment, CSG has gained popularity as a method for modelling indoor
environments. Valve Software's Hammer editor implements CSG [6]. Hammer has been used
successfully in commercial computer games, including Half-Life 2, Portal, Team Fortress 2, and
Counter-Strike Source. The editor provides wedge, torus, spike, sphere, cylinder, block, and arch
primitives, which can be combined using CSG operations into more complex shapes.
2.2. Boolean Operators for CSG
The operators used in CSG have traditionally been called Boolean operators [3]. Boolean
operators allow CSG to describe novel shapes through a series of relations between two objects.
The most common operators are: union, intersection, and difference.
2.2.1. Union (∪)
The union of primitives A and B produces a new piece of geometry that encompasses the volume
of both A and B. Figure 2 shows A ∪ B, the union of a cube A and a sphere B, for a particular
combination of locations and sizes of these primitives. Union is symmetric, i.e., A ∪ B = B ∪ A.
Figure 2. Union
2.2.2. Intersection (∩)
The intersection of primitives A and B results in a new piece of geometry that contains only the
volume where A and B overlap. Figure 3 shows A ∩ B, the intersection of a cube A and a sphere B;
the removed parts of the cube and sphere are shown with transparent shapes. Intersection is also
symmetric, i.e., A ∩ B = B ∩ A.
Figure 3. Intersection
2.2.3. Difference (−)
The difference of primitives A and B results in a new piece of geometry that encompasses the
volume of A, less the shared volumes of A and B. In practice, this operator is often referred to as

4
the carve operator, because it allows artists to carve into a piece of geometry [7]. Figure 4 shows
A − B, the difference of a cube A and a sphere B; the location of the sphere is shown with a
transparent shape. Difference is asymmetric, i.e., in general − ≠ − .
Figure 4. Difference
2.3. Volumetric Visualization
Volumetric data is comprised of an array of values, each representing a voxel. A voxel is a
volumetric primitive that occupies a single unit of three-dimensional space. A voxel can have
various associated values, such as density or temperature. Volumetric density data is comprised
of an array of values, each of which stores a single value for density. In this array representation
for volumetric data, the position of a voxel in three-dimensional space is inferred from its location
in the array.
To visualize volumetric data, two broad approaches are used. In the direct approach, each value
is directly visualized using a process called volume rendering [8]. Volume rendering works
directly with the data to produce visualization through a variety of techniques, including ray
casting [9], splatting [10], and shear warp [8]. In the indirect approach (or polygon approach), a
three-dimensional polygon mesh representing the surface is constructed and then visualized.
Density values at or above an arbitrary surface threshold (or extraction threshold) value are said
to describe parts of solid objects. Similarly, values below this surface threshold are said to be
outside the solid objects. The polygon mesh typically consists of a list of vertices and edges,
which can be displayed in a three-dimensional rasterization system such as OpenGL or DirectX.
A polygon mesh built from triangles can be readily and efficiently visualized and manipulated
with modern graphics processors. This paper focuses on CVG techniques and primitives well-
suited to the indirect visualization of volumetric density data.
The best-known method for surface construction from volumetric data is the marching cubes
algorithm [11, 12]. The algorithm processes eight neighbouring locations in the voxel data to
form an imaginary cube called a cell. The algorithm then iterates through all possible cells in the
voxel data. For each cell, the algorithm decides whether or not it partitions the object from the
outside world. In other words, it detects cells that form the boundary of the object. If a cell does
so, then it is a part of the surface, and the algorithm constructs a small number of triangles that
form the portion of the surface corresponding to the cell. The collection of all such triangles
forms the polygon mesh.
To construct the triangles corresponding to a cell, the algorithm utilizes a look up table of polygon
shapes corresponding to all possible partitions. Each cell is examined to determine which of the
eight sampling points are within the surface. There are a total of 256 (28
) possible configurations
for each cell. The total number of configurations can be reduced by considering reflections and
symmetrical rotations of each configuration. Figure 5 illustrates the fifteen unique cell
configurations shown in the original marching cubes paper [11]. The cube configurations are
designed in such a way that, when placed together, they produce a visibly solid surface.

Figure 5. Unique cell configuration
To create more accurate surfaces, the vertices of the polygon mesh are manipulated. Each vertex
of the polygon mesh is positioned on its cell's edge such that it is linearly interpolated between the
two voxel values connected to that edge. The final position of the vertex on a cell's edge
corresponds to the surface threshold position as determined by interpolation. This step modifies
the visibly solid surface, produced by converting the volume data into the correspondin
configurations, to better fit the surfa
Voxel systems for indirect visualization are able to use external colouring information through
texture mapping. Texture mapping applies surface detail supplied by an external
onto the constructed surface. This mapping can be done through a variety of methods, often using
a method called triplaner mapping
material identifier parameter can be added
performing the mapping. Williams proposed the use of discrete material identifiers for the voxel
values to allow for multiple such mappings ins
the identifier tells which material should be placed on a given portion of surface. The use of a
texture map makes the number of
volumetric data. This independence allows for high quality colouring
volumetric data without the overhead of storing p
2.4. Constructive Volume Geometry
Chen and Tucker devised the Constructive Volume Geometry
algebraic operations on volumetric data
describe the interior of the objects it is defining. CVG rectifies this deficiency by working with
objects defined by mathematical scalar fields, which
The CVG algebra provided by Chen and Tucker was intended to
rather than serve as a final standard
In their algebra, Chen and Tucker considered voxels
(here analogous to density) and three colour fields representing red
Chen and Tucker defined versions of these operators suited for
each voxel is composed of four components, opacity, red, gree
information is assumed to be embedded in the volumetric data.
Unique cell configurations for the marching cubes algorithm [11]
ected to that edge. The final position of the vertex on a cell's edge
the visibly solid surface, produced by converting the volume data into the correspondin
configurations, to better fit the surface implicit in the volume data.
. Texture mapping applies surface detail supplied by an external raster image file
triplaner mapping [11]. To allow for multiple mappings of different textures, a
material identifier parameter can be added to each voxel to specify which raster file to use when
Williams proposed the use of discrete material identifiers for the voxel
values to allow for multiple such mappings instead of colouring information [13]. The value of
ntifier tells which material should be placed on a given portion of surface. The use of a
texture map makes the number of texture elements (texels) independent of the resolution of the
volumetric data. This independence allows for high quality colouring of visualizations of
volumetric data without the overhead of storing per-voxel colouring information.
Constructive Volume Geometry
Constructive Volume Geometry (CVG) framework for describing
tric data [4, 14]. The primary deficiency of CSG is its inability to
objects defined by mathematical scalar fields, which are treated as analogues to volumetric data.
CVG algebra provided by Chen and Tucker was intended to encourage further discussion
serve as a final standard [4].
In their algebra, Chen and Tucker considered voxels with four distinct parameters:
ere analogous to density) and three colour fields representing red (R), green (G)
Chen and Tucker defined versions of these operators suited for a voxel system where the value for
each voxel is composed of four components, opacity, red, green, and blue. Thus, colour
information is assumed to be embedded in the volumetric data.
5
ected to that edge. The final position of the vertex on a cell's edge
the visibly solid surface, produced by converting the volume data into the corresponding cube
raster image file
o allow for multiple mappings of different textures, a
to each voxel to specify which raster file to use when
Williams proposed the use of discrete material identifiers for the voxel
. The value of
ntifier tells which material should be placed on a given portion of surface. The use of a
independent of the resolution of the
of visualizations of
) framework for describing
. The primary deficiency of CSG is its inability to
analogues to volumetric data.
further discussion
with four distinct parameters: opacity (O)
and blue (B).
voxel system where the value for
n, and blue. Thus, colour

6
Chen and Tucker define a spatial object as a tuple = , … of scalar fields defined in
ℝ , including an opacity field ∶ ℝ
→ 0, 1 specifying the “visibility” of every point p in
ℝ
and possibly other attribute fields … ∶ ℝ
→ ℝ, > 0 [4]. They define the traditional
CSG operations of union, intersection, and difference for objects and as functions over their
component opacity fields and as follows:
union: ∪ , = ! , (1)
intersection: ∩ , = "# , (2)
difference: − , = − (3)
where MAX and MIN are as specified in the Appendix. Chen and Tucker showed that CVG is a
superset of CSG [14]. They defined additional operators for CVG other than those available in
CSG. These operators included cap and trim, which are as follows:
cap: $%& , = ' ( − , ' ( ) − ) ,
' ( * − * , ' ( + − +
(4)
trim: ,-./ , = 0)" − , 0)" ) − ) ,
0)" * − * , 0)" + − +
(5)
where CAP and TRIM are as specified in the Appendix.
The opacity class provided by Chen and Tucker was restricted to the range [0, 1]. Opacity values
from 0 to 1 are well-suited to visualization. Based on the work of Chen and Tucker, Johnson and
Tucker subsequently developed a formal approach to specifying a spatial object as a data type
[15, 16]. They gave a useful algebra of continuous functions that has operations derived from
operations on space and data, and is equipped with an appropriate topology. To the best of our
knowledge, no previous CVG framework optimized for indirect visualization or using material
identifiers has been described in the literature.
2.5. Voxelization
Other related research concerns voxelization (also known as 3D scan conversion), which is the
process of taking a geometric representation of a continuous 3D object and converting it to a
voxel representation that approximates the continuous object. A naive approach would either test
every voxel exhaustively for an intersection with the object or else perform recursive subdivision
of the object. Unfortunately, the voxelizations generated by this approach are often too course
and include more voxels than necessary [17]. A more refined approach was devised by Kaufman
et al. [17]. They define two voxels to be 6-adjacent if they share a face, 18-adjacent if they share
an edge or a face, and 26-adjacent if they share a vertex, an edge, or a face. Based on these
definitions, they define notions of N-separating and covering and then propose that an effective
voxelization should provide a minimal cover of the object. A more recent approach, called
topological voxelization, is based on the topological properties of the objects [18]. Every voxel is
given an associated geometric intersection target and the voxel is marked as solid if any input
primitive intersects this target. Laine proved that selecting appropriate intersection targets results
in voxelizations with desirable connectivity and separability properties. The various voxelization
approaches were compared by Nourian and Zlatanova [19].
Related research addressed the challenge of performing fast voxelization using a Graphics
Processing Unit (GPU). Schwarz and Seidel introduced a conservative surface voxelization
technique, which marks all voxels as solid if they are partially or fully overlapped by a mesh’s
triangles [20]. The technique uses a triangle/box overlap test that can be adapted to yield a 6-
separating surface voxelization with two important properties: it is thinner than the naive
voxelization and it is still connected and gap-free. Recently, Baert et al. have devised an
approach suited to objects represented as very large 3D meshes [21]. Their approach allows the
input triangle mesh, the intermediate 3D voxel grid, and the voxelized output to all be larger than
available memory.

7
Although the previous voxelization approaches are fast and effective, none of them address the
problem of creating visual artefacts that is overcome by the method proposed in this paper.
3. CONSTRUCTIVE VOLUME GEOMETRY FOR INDIRECT VISUALIZATION
This section discusses the ICVG approach to implementing a voxel system that performs CVG
for indirect visualization of volumetric density data. CVG was implemented for indirect
visualization in a voxel system named Isovox. Isovox manipulates volumetric data and produces
polygon meshes for visualization. Surface colours for the meshes are obtained by texture
mapping from a texture external to the volumetric data. Section 3.1 presents assumptions for this
work. Section 3.2 describes techniques for developing visually correct volumetric primitives for
indirect visualization. Section 3.3 defines the ICVG algebra designed for indirect volumetric
visualization.
3.1. Assumptions
All values in the volumetric data are assumed to be discrete and non-negative integers in the
range [L, H], where L is the lowest possible value and H is the highest possible value.
Throughout this paper, H, L, and S are used as values rather than variables. Such values are well-
suited for volumetric density data, which are always non-negative. Floating point data would
need to be converted to integers. For example, suppose a value is represented as a single byte of
data. The use of a byte data type, as opposed to an unsigned integer, reduces memory
requirements. For one byte voxels, the L is 0, and the H is 255. Following the 2011 release of the
commercial volume editing tool 3D-Coat [22], the surface threshold (S) is defined to be the
midpoint between L and H, as shown in Equation 6.
S = 2
345
6 (6)
For example, S = 2
7748
6 = 127.
Throughout this paper, a right-handed coordinate system is used with right, up, and out
corresponding to the positive X, Y, and Z axis. When performing indirect visualization, a voxel is
visualized if its associated value in the volumetric data is greater than or equal to S.
3.2. ICVG Primitives
When defining primitives for indirect visualization of volumetric data, several factors need to be
considered. The straightforward approach to specifying primitives may produce undesirable
artifacts when indirect visualization is performed. Figure 6 illustrates the desired appearance of
four common volumetric primitives, namely the cuboid, the ellipsoid, the cylinder and the cone.
Within this section, Section 3.2.1 defines virtual volume sampling functions; Section 3.2.2
introduces clipping and padding; and Sections 3.2.3 through 3.2.6 describe the cuboid, ellipsoid,
cylinder, and cone primitives, respectively.
Figure 6. Primitives constructed from volumetric data sets

8
3.2.1. Virtual Volume Sampling Functions
A virtual volume sampling function is now defined. Let ℕ represent the non-negative integers.
Let +
<= = +>, +?, +@ ∈ ℕ represent the upper bounds on a 3D space, called the extension space
(also known as the volume), with origin (0, 0, 0). Thus, the width, height, and depth of the
extension space correspond to Bx, By and Bz, respectively. Let B
<= = B
>, B
?, B
@ ∈ ℕ , such that
B
> < +>, B
? < +?, B
@ < +@, represent a position in the extension space. Thus, the positions in the
extension space range from (0, 0, 0) to +> − 1, +? − 1, +@ − 1 . One voxel is located at each
position. A virtual volume sampling function, D: B
<= ⟶ L, H , specifies the density value for any
position (i.e., voxel) in the extension space. The function D takes a position, called the sampling
position, in an arbitrarily large extension space as input and returns as output a value in the range
[L, H]. This output value represents a density.
Virtual volume sampling functions are resolution independent, meaning that a virtual volume can
be sampled at any level of granularity by changing the resolution of the extension space. To fully
reconstruct the virtual volume at a specified resolution, the virtual volume function is sampled at
every point in an extension space at that resolution.
3.2.2. Clipping and Padding
As previously mentioned, the marching cubes algorithm operates on eight data points at a time.
Thus, data points along the edge of the extension space cannot be processed in the same fashion
as the interior data points. Some implementations of the marching cubes algorithm handle this
difficulty by pretending the extension space is surrounded by voxels with L or H values. Suppose
H values are used. If the surface of the object extends to any edge of the extension space, then
holes will appear where the surface touches the edge. This artifact is referred to as clipping.
To address the clipping issue, it would seem sufficient to assign L values to voxels outside of the
extension space, but doing so may make CVG operations on some objects less efficient. The
constraints of an application may ensure that some portions of a constructed surface can never be
viewed. If there are constraints of this fashion, then time will be wasted performing surface
construction and visualization of these unseen boundaries. Suppose that mountainous terrain is
created for a first person shooter game but that the player can never see the mountains from below
the surface. Therefore any visualization of the bottom of these mountains is wasted.
(a) View from above (b) View from below
Figure 7. Hollow island surface displayed by Isovox
A more flexible solution to this problem is to provide additional voxel values called padding
voxels surrounding the extension space of interest. In the case of the cuboid primitives, these

padding voxels are set to L, to ensure the creation of the expected surface mesh. In the
mountainous terrain example, no such padding voxels would be placed
extension space, resulting in a visually correct, hollow mesh. If
the mesh to be filled in with polygons, a set of padding L values could be added.
shows a visually correct unpadded terra
view of the same surface.
3.2.3. Cuboid
The cuboid volumetric primitive is used for representing rectangular solids. The construction of a
cuboid primitive for CVG may seem trivial, but two
designing such a primitive for indirect visualization:
The naive approach to defining a cuboid primitive is to simply fill a volume uniformly with H
values; this approach suffers from clipping. The clippin
outside of the volume, where needed, with L values, as described in Section 3.2.2.
The second problem is beveling. When a cuboid primitive is visualized, it should have sharp
corners. However, the naive definition of the cuboid primitive allows beveling artifacts to appear
on the edges of the visualized surface.
Figure 8 (a) shows how a beveling artifact can have an undesirable smoothing effect on the
cuboid primitive, shown here with a checkerboard texture to facilitate seeing the edges. Observe
how the edges of the cuboid have a smooth, rounded appearance.
(a) With beveling artifact
Figure 8. Cuboid primitive with
This problem can be easily understood by considering the two
three-dimensional one. Figure 9
a dotted line, while the solid line shows how the actual
(a) Corner with beveling artifact
Figure 9
voxels are set to L, to ensure the creation of the expected surface mesh. In the
mountainous terrain example, no such padding voxels would be placed at the bottom of the
extension space, resulting in a visually correct, hollow mesh. If one desired the bottom portion of
the mesh to be filled in with polygons, a set of padding L values could be added.
shows a visually correct unpadded terrain example and Figure 7 (b) shows the hollow underside
cuboid primitive for CVG may seem trivial, but two problems need to be addressed when
tive for indirect visualization: clipping and beveling.
pproach suffers from clipping. The clipping problem can be avoided by padding the
outside of the volume, where needed, with L values, as described in Section 3.2.2.
is beveling. When a cuboid primitive is visualized, it should have sharp
on the edges of the visualized surface.
shows how a beveling artifact can have an undesirable smoothing effect on the
how the edges of the cuboid have a smooth, rounded appearance.
(a) With beveling artifact (b) Without beveling artifact
Cuboid primitive with and without beveling artifact
can be easily understood by considering the two-dimensional case rather than the
Figure 9 (a) illustrates the desired appearance of a square primitive with
while the solid line shows how the actual beveled surface is constructed.
(a) Corner with beveling artifact (b) Corner with artifact removed
9. Two-dimensional view of beveling artifact
9
voxels are set to L, to ensure the creation of the expected surface mesh. In the
at the bottom of the
the bottom portion of
Figure 7 (a)
(b) shows the hollow underside
need to be addressed when
can be avoided by padding the
is beveling. When a cuboid primitive is visualized, it should have sharp
shows how a beveling artifact can have an undesirable smoothing effect on the
dimensional case rather than the
illustrates the desired appearance of a square primitive with
constructed.

10
The beveling of the corner of the square primitive is an artifact of the marching cubes algorithm.
When the corner cell is sampled, two points out of eight are within the boundary of the volume,
corresponding to the third case in Figure 5. The look up table selects a diagonal corner piece to
be placed in the cell. The vertices are positioned on their related cell edges using linear
interpolation. With the naive cuboid definition, the interpolated surface needs to be between H
and L density, which forces the vertices of the constructed triangles to be placed on the midpoints
of the edges.
In ICVG, beveling is addressed by adding a thin layer (one voxel wide) of surface S value voxels
surrounding the interior H values of the cuboid. Figure 9 (b) shows how adding these surface
value voxels around a square can remove the beveling artifact. Although the look up function
chooses a diagonal corner piece to be placed in the cell, as before, when the location of the
vertices are interpolated between L and S, they are placed touching the S side. Figure 8(b) shows
a cuboid primitive with the beveling artifact removed, resulting in sharp corners, as desired.
The ICVG cuboid sampling function D
IJ is defined as follows:
D
IJKB
<=L =
M
N
O
N
P
L if B
> = 0 ∨ B
> = +> − 1 ∨ KB
? = 0L ∨ KB
? = +? − 1L
∨ B
@ = 0 ∨ B
@ = +@ − 1 ,
S otherwise if B
> = 1 ∨ B
> = +> − 2 ∨ KB
? = 1L
∨ KB
? = +? − 2L ∨ B
@ = 1 ∨ B
@ = +@ − 2 ,
H otherwise
[ (7)
z
y
x B
B
B ,
, are restricted to values greater than 2. Given a position V
r
bounded within the volume
dimensions B
r
, the cuboid sampling function D
IJ returns the voxel value at that position, which is
L, S, or H. If the position is touching the bounding volume, then it is assigned L. Positions that
are one voxel away from the edges of the bounding volume are assigned S. All other positions
are assigned H. This method of defining the cuboid sampling function overcomes the clipping
and beveling problems described above.
An undesirable side effect of having the cuboid primitive contain three possible values instead of
two (i.e., it contains L, H, and S values instead of only H and L values) is that the potential for
compression is reduced. Generally, having less homogenous volumetric data reduces the
expected compression ratio.
3.2.4. Ellipsoid
The ellipsoid volumetric primitive is used for representing spherical and other rounded shapes.
The naive approach to the ellipsoid primitive would be to linearly interpolate values from H to L
radically from the midpoint of the volume. Figure 10 (a) shows the density function along some
axis where the ellipsoid has radius R that results from this approach. This approach produces a
visually correct surface mesh, but it suffers from three flaws. First, much of the volume will be
filled with wasted space, i.e., values outside of the surface mesh that do not affect construction.
This wasted space means that the ellipsoid surface is relatively small compared to the size of the
volume. Secondly, these invisible values may have undesired effects on CVG operations.
Finally, the volumetric data resulting from this approach will be inefficient to compress because
the values inside the surface and outside the surface, which do not affect surface construction,
vary widely. Figure 10 (b) shows the density function corresponding to the ICVG virtual volume
sampling function, which addresses the problems with the naive approach. Only a small amount
of extension space is allocated past the radius R, and values well below this threshold are assigned
values of H.

(a) Naive density function
Figure 10
Let )
<= = )>, )?, )@ ∈ ℕ , where
the desired radius in the X, Y, and Z axes. Let,
represent the position in the volume translate
to the center of the ellipsoid. Additionally
]].&^_.`+_`aK(
<
bL = c
de
f
ge
f h
di
f
gi
f h
The ellipsoid sampling function D
DjkKB
<=L =
M
N
O
N
P H
L
l
m
nkkopqrostrs?
where 0
u = 0
<b/w0
<bw represents the unit vector in the same direction as
volumetric data that can be visualized as an ellipsoid. This function reduces the number of
unique values in the volumetric data by
construction because they are definitely inside or definitely outside the ellipsoi
assigned H values if they are inside the ellipsoid and L values if they are outside. Voxels near the
surface of the ellipsoid are assigned the value of one divided by the ellipsoid body function,
scaled by S. By calculating the near s
from the data will appear to have a
will be highly compressible due to the reduced number of unique values.
An ellipsoid primitive constructed with the
center point )
 = )? = )@, the resulting primitive is a sphere.
The visual quality of an ellipsoid primitive is limited by the dimensions of the volume
which it is defined, as well as the range of possible
B
<b, the primitive is less smooth in appearance. When the range of values for the voxel is
insufficient, a distinctive banding
(a) shows the banding artifact on an ellipsoid visualized using a volumetric ellipsoid primitive
with B
<b = 32,32,32 and a voxel range of 256. This artifact can be reduced in severity by
increasing the number of possible va
artifact on an ellipsoid with B
 y 2
te
6, 1 y )? y 2
ti
6, and 1 y )@ y 2
the desired radius in the X, Y, and Z axes. Let, 0
, 0?, 0@ ∈ ℕ , such that 0
represent the position in the volume translated so that (0, 0, 0) corresponds as nearly as possible
Additionally the ]].&^_.`+_`a: (
 1,
nkkopqrostrs?K|
<=L
} otherwise
[
represents the unit vector in the same direction as 0
<b. Function
data that can be visualized as an ellipsoid. This function reduces the number of
unique values in the volumetric data by identifying voxels that will not influence surface
construction because they are definitely inside or definitely outside the ellipsoid. These voxels are
scaled by S. By calculating the near surface values in this fashion, the polygon mesh produced
from the data will appear to have a relatively smooth surface. Nonetheless the volumetric data
will be highly compressible due to the reduced number of unique values.
An ellipsoid primitive constructed with the ellipsoid sampling function Djk is centered around the
extension space. Function Djk can be used to create a variety of ellipsoid
altering the separate components of R. In cases where an ellipsoid is defined in a volume such
, the resulting primitive is a sphere.
The visual quality of an ellipsoid primitive is limited by the dimensions of the volume
which it is defined, as well as the range of possible values for each voxel. For smaller values of
, the primitive is less smooth in appearance. When the range of values for the voxel is
banding artifact appears on the constructed polygon mesh.
rtifact on an ellipsoid visualized using a volumetric ellipsoid primitive
and a voxel range of 256. This artifact can be reduced in severity by
of possible values for each voxel. Figure 11 (c) shows the roughness
b = 8,8,8 and the same voxel range. Figure 11 (b) shows no
11
2
tz
6 represent
0
<b = B
<b − )
<b ,
0) corresponds as nearly as possible
function is defined as
known ellipsoid function.
(8)
. Function Djk produces
data that can be visualized as an ellipsoid. This function reduces the number of
oxels that will not influence surface
d. These voxels are
the polygon mesh produced
smooth surface. Nonetheless the volumetric data
is centered around the
can be used to create a variety of ellipsoids by
where an ellipsoid is defined in a volume such
The visual quality of an ellipsoid primitive is limited by the dimensions of the volume B
<b for
values for each voxel. For smaller values of
, the primitive is less smooth in appearance. When the range of values for the voxel is
tructed polygon mesh. Figure 11
rtifact on an ellipsoid visualized using a volumetric ellipsoid primitive
and a voxel range of 256. This artifact can be reduced in severity by
(c) shows the roughness
(b) shows no

distinctive artifacts on an ellipsoid with
roughness artifact, but small enough to not r
(a) B
<b = 32,32,32
Figure
3.2.5. Cylinder
The cylinder sampling function D
D
I?KB
<=L =
M
N
N
N
O
N
N
N
P L
S

[
H
l
m
nkkopqjtrs?
where ]].&^•+_à: (
<b → 0, ∞
represents the well-known two-
definition, the cylinder primitive is related to the ellipsoid primitive. The ba
two-dimensional ellipse, defined on the X/Z plane. Surface detail is calculated in the same
fashion as for the ellipsoid primitive, with the ellipsoid body function replaced by a two
dimensional ellipse body function. The ellipse
for the ends, which must be treated differently in order to cap the primitive.
position height B
? is used as a parameter to determine if the position is vertically within the
cylinder, i.e., between 0 and +?
form the surface, followed by L values for padding.
3.2.6. Cone
The cone volumetric primitive is used for representing
shape is defined in terms of its height and the radiuses
h is defined as c1 −
€i
ti
{, which linearly interpolates fro
extension space, relative to the height of the
linearly interpolates from 1 to 0 as the position height value
of the extension space. Let '_••+_à
distinctive artifacts on an ellipsoid with B
<b = 16,16,16 , which is large enough to minimize the
roughness artifact, but small enough to not require additional voxel range.
(b) B
 1L ∨ KB
? = 0L
∨ KB
? = +? − 1L,
if KB
? = 1 ∨ B
? = +? − 2L
∧ K]].&^•+_àK0
<= h 0
uL < 1L,
[
if ]].&^•+_àK0
<= h 0
uL < 1,
nkkopqjtrs?K|
<=L
} otherwise
[
[
is a function such that ]].&^•+_àK(
<bL = „
de
f
ge
f h
-dimensional ellipse function. As can be seen from the above
definition, the cylinder primitive is related to the ellipsoid primitive. The base of the cylinder is a
dimensional ellipse, defined on the X/Z plane. Surface detail is calculated in the same
fashion as for the ellipsoid primitive, with the ellipsoid body function replaced by a two
dimensional ellipse body function. The ellipse base is used at each level of the cylinder, except
for the ends, which must be treated differently in order to cap the primitive. T
as a parameter to determine if the position is vertically within the
− 1. At the vertical ends of the cylinder, S values are placed to
llowed by L values for padding.
volumetric primitive is used for representing cones and other conical shapes
erms of its height and the radiuses of its elliptical base. The sampling height
, which linearly interpolates from 0 to 1 based on the position in the
, relative to the height of the extension space. More precisely, the
linearly interpolates from 1 to 0 as the position height value B
? varies from the bottom to the top
'_••+_à: (
<b → 0, ∞ , where '_••+_àK(
<bL = „
d
ge
12
, which is large enough to minimize the
8,8
L
[
[ (9)
„ h
dz
f
gz
f…, which
dimensional ellipse function. As can be seen from the above
se of the cylinder is a
dimensional ellipse, defined on the X/Z plane. Surface detail is calculated in the same
fashion as for the ellipsoid primitive, with the ellipsoid body function replaced by a two-
base is used at each level of the cylinder, except
The sampling
as a parameter to determine if the position is vertically within the
. At the vertical ends of the cylinder, S values are placed to
cones and other conical shapes. A conical
The sampling height
position in the
More precisely, the value of h
bottom to the top
L „
de
f
e† f h
dz
f
gz† f…,

13
represent the ellipse function, modified by the height-dependent h variable. The cone sampling
function D
Ir for a volume B
 1,
S otherwise if KB
? = 1L ∧ K'_••+_`aK0
<= h 0
uL < 1L,
[
H otherwise if '_••+_`aK0
<= h 0
uL < 1,
l
m
‡rˆjtrs?K|
<=L
} otherwise
[
[ (10)
As with the cylinder, the sampling height h is used to determine if a position is vertically outside
of the primitive, but it is also used as a parameter for altering the desired radiuses of the cone.
Specifically, h reduces the ellipse radiuses as the positions progress up the cone. L and S padding
are applied at the bottom of the cone.
With the above cone sampling function, it is important to consider where in the volume the
primitive's center point on the X/Z plane is to be located. If this center point is located in the true
middle of volume, the tip of the resulting cone shape may contain artifacts. Error! Reference
source not found. shows the broken tip artifact, where the topmost portion of the primitive
contains undesirable surface deformation. This artifact is a result of the logical center being
defined between voxels in the extension space. If the center of the cone is forced to be on a
voxel, this artifact will not occur.
Figure 12. Cone primitive with the broken tip artifact
3.3. The ICVG Algebra
Chen and Tucker's definition of CVG operators was purposely general, and allowed for real and
negative values. The modified algebra, called ICVG, concentrates on the use of primitives and
operations optimized for indirect visualization. Consequently, the algebra is limited to a restricted
range non-negative integer values. The ICVG union, intersection, trim, and cap operations are the
same as those of CVG for opacity, except that the input and output values are restricted to discrete
positive volumes:
union: ∪ ‰ , ‰ = ! ‰ , ‰ (11)
intersection: ∩ ‰ , ‰ = "# ‰ , ‰ (12)
trim: ,-./ ‰ , ‰ = 0)" ‰ , ‰ (13)
cap: $%& ‰ , ‰ = ' ( ‰ , ‰ (14)

14
The difference operator is defined differently in ICVG than in CVG to ensure that the values in
the resulting volume are always non-negative. The following three possible operators were
considered:
differenceA: − ‰ , ‰ = ! ‰ − ‰ , L (15)
differenceB: − ‰ , ‰ = Š
! ‰ − ‰ , L if ‰ ≥ S,
‰ otherwise
[ (16)
difference: − ‰ , ‰ = Š
! ‰ − ‰ , S if ‰ ≥ S ∧ ‰ ≥ S ,
! ‰ − ‰ , L otherwise
[ (17)
The differenceA operator represents the range-restricted version of the difference operator
described by Chen and Tucker [4]. Figure 13 (a) shows that the differenceA operator produces
undesirable artifacts on the surface of the resulting object when applied with a cuboid for ‰ and
an ellipsoid for ‰ . Observe that the cavity created in ‰ is smooth, but every edge of this cavity
has an apparent beveling artifact. Values in the ellipsoid primitive that are below S are not
visualized, but they affect the cuboid primitive when the differenceA operator is applied. These
problems are typical of those encountered when combining any two primitives with differenceA.
The differenceB operator is defined similarly to differenceA but ignores the values that are below
S, i.e., values corresponding to parts of the ellipsoid primitive that are outside the surface of the
ellipsoid. Figure 13 (b) shows how ignoring these invisible values in the ellipsoid primitive
results in a jagged surface on the carved areas of the cuboid. This artifact appears because the
ellipsoid primitive owes part of its smoothness to these invisible values. As shown in Figure 13
(c), the difference operator combines the desirable features of the differenceA and the differenceB
operators. In cases where the surface of the source object is visible and the surface of the operand
object is not, this operator performs a subtraction operation that is bounded between S and H.
Otherwise, it acts in the same manner as differenceA. The difference operator does not suffer
from the beveling artifact that occurs with differenceA, and it does not produce the jagged interior
surface that occurs with difference B.
(a) differenceA (b) differenceB (c) difference
Figure 13. Three possible difference operators
Additionally, ICVG defines the add and modulate operators as follows:
add: h ‰ , ‰ = "# ‰ h ‰ , H (18)
module: ∗ ‰ , ‰ = "# ‰ ∗ ‰ , H (19)
These two operators may produce compound objects with surfaces not found in the source
primitives because when two densities that are lower than the surface threshold are combined the
result may be a density that is higher than the surface threshold. Thus these two operations have
less predictable behavior than the previously defined operations.

15
Two possible divide operators, called divideA and divide, were also considered. The naive
divideA operator is defined as follows:
divideA: ÷ ‰ , ‰ = Ž
‰ ÷ ‰ if ‰ ≥ 0,
H otherwise
[ (20)
and the more useful divide operator is defined as follows:
divide: ÷ ‰ , ‰ , •• = Ž
‰ ÷ ‰ ∗ •• if ‰ ≥ 0,
H otherwise
[ (21)
where SL is a real number in the range 0 to 1.
Since an object's surface is the middle of the valid range of possible voxel values, the divideA
operator produces a compound object with no visible surface wherever ‰ ≥ 2. The divide
operator is similar to divideA, but it multiplies the divisor by a real-valued scalar during each
operation, providing additional control. The divide operator does not require ‰ to be specified as
a value in the real domain in order to provide small adjustments. The divide operator can be
implemented using floating point operations during calculations, allowing for a greater range of
results, but only integers are used as input and output. Thus, only integers need to be stored.
In addition to the binary operators just discussed, ICVG also includes scalar and unary operations.
A scalar operation is performed by applying the same arithmetic operation to every voxel in
volumetric data. An example of a scalar operation is addScalar, which adds a single number to a
data set in order to uniformly increase, or decrease, the densities. A unary operator, such as
inverse, requires a single operand. The inverse operation creates an extension space where
solidness and emptiness are reversed. Taking the inverse of an object ‰ is equivalent to applying
the difference operator to a completely solid extension space and ‰ . The inverse operator can be
implemented by computing the difference between H and each voxel. The inverse operator is
defined as follows:
inverse: " ‰ = H − ‰ (22)
4. EXPERIMENTAL RESULTS
All sampling functions described in Section 3.2 were implemented in both CPU and GPU
versions in a common test bed, with the only differences being those that are essential to invoke
CPU versus GPU operations. Experiments were run on a Dell Alienware Aurora computer with an
Intel Core i7-3820 CPU and an NVIDIA GeForce GTX 560 GPU. Every result reported is the
average (mean) of the results for 100 separate runs, each of which involved generating 100
identical volumetric solids.
The four graphs shown in Figure 14 and Figure 15 depict the performance of volume generation
for four primitives (cuboid, ellipsoid, cylinder, and cone) using the sampling functions defined in
ICVG. Figure 14 (a) and Figure 15 (a) show the execution times in milliseconds on the CPU
and GPU, respectively. For example, the first set of eight lines in Figure 14 (a) shows the
execution time on the CPU while generating cuboids in various sizes of extension spaces. The
"EllipsoidB" and "CylinderB" results are for alternative implementations that are not restricted to
aligning precisely with the grid. Figure 14 (b) and Figure 15 (b) represent the same timing data,
but with the time (in nanoseconds) per voxel in the extension space. For example, for generating a
cuboid on the CPU in an 8 x 8 x 8 extension space, the time of 0.005 milliseconds is divided by
512 (83
) voxels, giving 9.7 nanoseconds per voxel.
The results in Figure 14 (a) show that, for each primitive, the execution time on the CPU
increases as the primitive is generated in larger extension spaces. Figure 14 (b) shows that the
execution time per voxel was roughly constant, except for the cone. When generating a cone, a
fixed overhead occurred, which increased the cost per voxel for small sizes of extension spaces.

16
(a) Time per primitive
(b) Time per voxel
Figure 14. CPU execution times for generating primitives with varying sizes of extension spaces
Timing results for generating the primitives on the GPU were also obtained. The version of
DirectX that was used for our implementation did not provide the ability to query the time that it
took for a dispatch call (which in this case causes a single primitive to be generated on the GPU).
To counter this absence, the results of the dispatch calls were immediately pulled back into CPU
memory after the dispatch call was made and the timer was stopped when the transfer completed.
If a request is made to bring GPU data back into CPU memory, then all queued operations (such
as the dispatched generation procedures) on that data are performed as soon as possible and the
data afterwards is transferred back to the CPU. To obtain timing information for only the GPU

17
calculations, the fixed time cost of transferring the data from the GPU was subtracted, which
otherwise would have distorted the results for small extension spaces. However, this introduced
some inaccuracy for these small extension spaces. Thus, overall the timing mechanism for our
experiments on the GPU was less accurate than the one for those on the CPU.
(a) Time per primitive
(b) Time per voxel
Figure 15. GPU execution times for generating primitives with varying sizes of extension spaces
The results in Figure 15 (a) show that execution was much faster on the GPU than the CPU, e.g.,
0.43 ms instead of 36 ms for the largest cone. For the cuboid, the execution time on the GPU was
roughly constant as the cuboid is generated in larger extension spaces. For the other primitives,
small linear increases in total execution time on the GPU were observed as the size of the
extension space increased, possibly because of timing inaccuracies, as described above. Given
the timing inaccuracies, the execution time per voxel was roughly constant for all primitives (see
Figure 15(b)).

18
Overall, acceptable performance was obtained with our implementations of ICVG on both the
CPU and GPU. The time required to generate any 72 x 72 x 72 shape was less than 40 ms on the
CPU and less than 0.45 ms on the GPU.
5. CONCLUSIONS AND FUTURE WORK
The ICVG algebra was described and descriptions of primitives for volumetric data presented
through indirect visualization were provided. The concept of a volume sampling function was
introduced, and examples were provided for cuboids, ellipsoids, cylinders, and cones. Potential
banding, beveling, and clipping artifacts were identified and methods for avoiding them were
provided. Three alternative definitions for the difference operator were described, and the visual
results for indirect visualization were compared. As well, the divideA, divide, modulate and add
CVG operations were defined. The inverse unary operation was also specified.
In future work, experiments could be conducted on alternative methods for reducing the
remaining artifacts due to the difference operation. Alternative values of S should be investigated
to determine whether the visual impact of the artifact can be reduced. Modified versions of the
marching cubes algorithms could be investigated to see if their changes to the resulting meshes
affect the presence of the identified artifacts. Additional primitive objects could be defined,
including torus, pyramid, capsule, and prism objects [23].
REFERENCES
[1] L. Kobbelt, M. Botsch, U. Schwanecke, and H. Seidel, "Feature Sensitive Surface Extraction from
Volume Data," in 28th Annual Conference on Computer Graphics and Interactive Techniques
(SIGGRAPH'01), Los Angeles, 2001.
[2] R. Geiss, "Generating Complex Procedural Terrains Using the GPU," in GPU Gems, Boston,
Pearson Education, 2008, pp. 5-37.
[3] A. Requicha and H. Voelcker, "Constructive Solid Geometry," Production Automation Project,
University of Rochester, Rochester, 1977.
[4] M. Chen and J. Tucker, "Constructive Volume Geometry," Computer Graphics Forum, vol. 19, no. 4,
pp. 281-293, 2000.
[5] M. Schifko, B. Jüttler, and B. Kornberger, "Industrial Application of Exact Boolean Operations for
Meshes," in 26th Spring Conference on Computer Graphics (SCCG'10), Budmerice, Slovakia, 2010.
[6] Hammer Editor, Valve [computer software], 2017.
[7] N. Zhang, H. Qu, and A. Kaufman, "CSG Operations on Point Models with Implicit Connectivity,"
in Computer Graphics International Conference (CGI'05), Stony Brook, NY, 2005.
[8] P. Lacroute and M. Levoy, "Fast Volume Rendering Using a Shear-warp Factorization of the Viewing
Transformation," in 21st Annual Conference on Computer Graphics and Interactive Techniques
(SIGGRAPH'94), Orlando, 1994.
[9] B. Kainz, M. Grabner, A. Bornik, S. Hauswiesner, J. Muehl, and D. Schmalstieg, "Ray Casting of
Multiple Volumetric Datasets with Polyhedral Boundaries on Manycore GPUs," ACM Transactions
on Graphics, vol. 28, no. 5, pp. 1-9, 2009.
[10] D. Laur and P. Hanrahan, "Hierarchical Splatting: A Progressive Refinement Algorithm for Volume
Rendering," in 18th Annual Conference on Computer Graphics and Interactive Techniques
(SIGGRAPH'91), Las Vegas, 1991.
[11] W. Lorensen and H. Cline, "Marching Cubes: A High Resolution 3D Surface Construction
Algorithm," in 14th Annual Conference on Computer Graphics and Interactive Techniques
(SIGGRAPH'87), Anaheim, 1987.
[12] T. Newman and H. Yi, "A Survey of the Marching Cubes Algorithm," Computer & Graphics, vol.
30, no. 5, pp. 854-879, 2006.
[13] D. Williams, "Volumetric Representation of Virtual Environments," in Game Engine Gems, Sudbury,
Jones and Bartlett, 2010, pp. 39-60.

19
[14] M. Chen and J. Tucker, "Constructive Volume Geometry," Department of Computer Science,
University of Wales, Swansea, 1998.
[15] K. Johnson and J. V. Tucker, "The Data Type of Spatial Objects," Formal Aspects of Computing, vol.
25, no. 2, pp. 189-218, 2013.
[16] K. H. A. Johnson, The Data Type of Spatial Objects with Applications to Constructive Volume
Geometry, Doctoral thesis, University of York, UK, 2007.
[17] A. Kaufman, D. Cohen, and R. Yagel, "Volume Graphics," IEEE Computer, vol. 26, no. 7, pp. 51-64,
1993.
[18] S. Laine, "A Topological Approach to Voxelization," Computer Graphics Forum, vol. 32, no. 4, pp.
77-86, 2013.
[19] P. Nourian Ghadi Kolaee and S. Zlatanova, "Voxels and Voxel Algorithms; Two Methods of
Rasterization, Raster Specific Operations," 31 12 2015. [Online]. Available:
https://repository.tudelft.nl/islandora/object/uuid:7954d601-5e37-47bc-b73e-
1df435efea6d?collection=research. [Accessed 2 10 2017].
[20] M. Schwarz and H.-P. Seidel, "Fast Parallel Surface and Solid Voxelization on GPUs," in 3rd ACM
SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia
(SIGGRAPH Asia 2010), Seoul, South Korea, 2010.
[21] J. Baert, A. Lagae, and P. Dutré, "Out-of-Core Construction of Sparse Voxel Octrees," Computer
Graphics Forum, vol. 33, no. 6, pp. 220-227, 2014.
[22] 3D Coat, Pilgway [computer software], 2011.
[23] 3D Studio Max, Autodesk / Discreet [computer software], 2017.
APPENDIX
The following functions are defined on scalars in the range of [L, H]:
! ! , ! = Š
! if ! > ! ,
! otherwise
[ (A1)
"# ! , ! = Š
! if ! < ! ,
! otherwise
[ (A2)
0)" ! , ! = Š
! if ! ≥ ! ,
! otherwise
[ (A3)
' ( ! , ! = Š
! if ! y ! ,
! otherwise
[ (A4)

ICVG : Practical Constructive Volume Geometry for Indirect Visualization

More Related Content

What's hot (16)

Similar to ICVG : Practical Constructive Volume Geometry for Indirect Visualization (20)

Recently uploaded (20)

ICVG : Practical Constructive Volume Geometry for Indirect Visualization