Fast Computation of Generalized Voronoi Diagrams Using Graphics Hardware

Kenneth E. Hoff III, Tim Culver, John Keyser, Ming Lin, Dinesh Manocha
University of North Carolina at Chapel Hill
Department of Computer Science

Abstract: We present a new approach for computing generalized

2D and 3D Voronoi diagrams using interpolation-based polygon
rasterization hardware. We compute a discrete Voronoi diagram
by rendering a three dimensional distance mesh for each Voronoi
site. The polygonal mesh is a bounded-error approximation of a
(possibly) non-linear function of the distance between a site and a
2D planar grid of sample points. For each sample point, we
compute the closest site and the distance to that site using polygon
scan-conversion and the Z-buffer depth comparison. We construct
distance meshes for points, line segments, polygons, polyhedra,
curves, and curved surfaces in 2D and 3D. We generalize to
weighted and farthest-site Voronoi diagrams, and present efficient
techniques for computing the Voronoi boundaries, Voronoi
neighbors, and the Delaunay triangulation of points. We also show Cover Plate: Discrete approximation of the generalized Voronoi
how to adaptively refine the solution through a simple windowing diagram of four points, a line, a triangle, and one cubic Bézier curve
operation. The algorithm has been implemented on SGI computed interactively on a PC.
workstations and PCs using OpenGL, and applied to complex
datasets. We demonstrate the application of our algorithm to fast 1970s, algorithms for computing Voronoi diagrams of geometric
motion planning in static and dynamic environments, selection in primitives have been developed in computational geometry and
complex user-interfaces, and creation of dynamic mosaic effects. related areas.
CR Categories: I.3.5 [Computer Graphics]: Computational Good theoretical and practical algorithms are known for
Geometry and Object Modeling; I.3.3 [Computer Graphics]: computing ordinary Voronoi diagrams of points in any dimension.
Picture/Image Generation. Ordinary Voronoi diagrams can be generalized in many different
Additional Key Words: Voronoi diagrams, graphics hardware, ways by using different distance functions and site shapes. A
polygon rasterization, interpolation, motion planning, proximity common generalization is to compute the diagram for higher-order
query, medial axis, OpenGL, framebuffer techniques. sites, such as lines and curves. This greatly increases the
complexity since the boundaries of the diagram are composed of
high-degree algebraic curves and surfaces, and their intersections;
1 INTRODUCTION the boundaries of an ordinary point Voronoi diagram are linear.
No practically efficient and numerically robust algorithms are
Given a set of primitives, called Voronoi sites, a Voronoi diagram known for constructing a topologically consistent, continuous
partitions space into regions, where each region consists of all representation of generalized Voronoi diagrams.
points that are closer to one site than to any other. Voronoi
diagrams have been used in a number of applications including Given the practical complexity of computing an exact generalized
visualization of medical datasets, proximity queries, spatial data Voronoi diagram, many authors have proposed approximate
manipulation, shape analysis, computer animation, robot motion algorithms. Interesting approaches include computing the Voronoi
planning, modeling spatial structures and processes, pattern diagram of a point-sampling of the sites, adaptively subdividing
recognition, and locational optimization. The concept of Voronoi space to locate the Voronoi boundary, and point-sampling the
diagrams has been around for at least four centuries, and since the space to form a volumetric representation of the diagram. In
practice, these previous algorithms take considerable time and
e-mail: {hoff,culver,keyser,lin,dm}@cs.unc.edu memory on large numbers of input sites, or are restricted in
WWW: http://www.cs.unc.edu/~geom/voronoi/
generality.
Main Contributions: In this paper, we present an approach that
computes discrete approximations of generalized Voronoi
diagrams to an arbitrary resolution using polygon rasterization
hardware. Our contributions include:
1. Efficient methods to approximate the distance function, with
bounded error, for points, lines, polygons, polyhedra, curves,
and curved surfaces using a polygonal mesh that is linearly
interpolated by graphics hardware.
2. Efficient algorithms to find Voronoi boundaries and neighbors,
and to construct Delaunay triangulations.
3. Techniques to construct weighted and farthest-site generalized 2.2 Generalized Voronoi Diagrams
Voronoi diagrams in 2D and 3D.
Algorithms have been proposed for constructing Voronoi
4. Demonstration of the effectiveness of our approach to the diagrams of higher order sites. Two broad approaches based on
following applications: incremental and divide-and-conquer techniques have been
• Fast motion planning in static and dynamic environments summarized in [Okabe92]. The set of algorithms includes divide-
and-conquer algorithms for polygons [Lee82, Held97], an
• Selection in complex user-interfaces
incremental algorithm for polyhedra [Milen93b], and 3D tracing
• Generation of dynamic mosaics
for polyhedral models [Milen93, Sherb95, Culve99]. Curved sites
The resulting techniques have been effectively implemented on and CSG objects are handled in [Chian92, Dutta93, Hoffm94]. In
PCs and high-end SGI workstations using the OpenGL graphics all these cases, the computation of generalized Voronoi diagrams
library. A 2D example computed in real-time is shown in the involves representing and manipulating high-degree algebraic
cover plate. Our techniques improve upon the state of the art in curves and surfaces and their intersections. As a result, no efficient
following ways: and numerically robust algorithms are known for computing them.
• Generality: We make no assumption with respect to input
primitives. We only need to mesh the distance function of a site
2.3 Approximate Voronoi Diagrams
over a grid of point samples. Many authors compute approximations of generalized Voronoi
• Efficiency: We show that our approach is quite fast. Its speed diagrams based on the Voronoi diagram of a point-sampling of the
sites [e.g. Sheeh95]. However, deriving any error bounds on the
arises from using coarse polygonal approximations of the
output of such an approach is difficult, and the overall complexity
distance functions while still maintaining a specified error
is not well understood.
bound, using polygon rasterization hardware to reconstruct the
distance values, and using the Z-buffer depth comparison to [Vleug95] and [Vleug96] have presented an approach that
perform distance comparisons. We demonstrate the 2D adaptively subdivides space into regular cells and computes the
approach on models composed of nearly 100K triangles in a Voronoi diagram up to a given precision. [Laven92] uses an octree
real-time motion planning application through a complex representation of objects and performs spatial decomposition to
dynamic scene. We derive efficient meshing strategies for compute the approximation. [Teich97] computes a polygonal
polygonal models in 3D, and show the results of a prototype approximation of Voronoi diagrams by subdividing the space into
implementation that demonstrates its potential. tetrahedral cells. All these algorithms take considerable time and
memory for large models composed of tens of thousands of
• Tight Bounds on Accuracy: Although our approach produces
triangles, and cannot easily be extended to directly handle
a discretized Voronoi diagram, all sources of error are
dynamic environments.
enumerated and techniques are given to produce output within
any specified tolerance. The idea of using polygon rasterizing hardware and rendering of
cones to construct 2D Voronoi diagrams of points is suggested in
• Ease of Implementation: The approach can be easily
[Haebe90] and in the OpenGL 1.1 Programming Guide [Woo97].
implemented on current graphics systems. The special cases are
limited and the problem reduces to simply meshing a distance
function for any new site. 2.4 Graphics Hardware
Polygon rasterization graphics hardware has been used for a
2 RELATED WORK number of geometric computations, such as visualization of
constructive solid geometry models [Rossi86, Goldf89] and
The concept of Voronoi diagrams has been around for at least four interactive inspection of solids, including cross-sections and
centuries. In his treatment of cosmic fragmentation in Le Monde interferences [Rossi92]. Algorithms for real-time motion planning
de Mr. Descartes, ou Le Traite de la Lumière, published in 1644, using raster graphics hardware have been proposed by [Lengy90].
Descartes uses Voronoi-like diagrams to show the disposition of
matter in the solar system and its environment. The first
presentations of this concept appeared in the work of [Diric50] 3 OVERVIEW
and [Voron08]. Algorithms for computing Voronoi diagrams have In this section, we present the basic concepts important to our
been appearing since the 1970s. See the surveys by [Auren91] and approach. We give a formal definition of generalized Voronoi
[Okabe92] on various algorithms, applications, and diagrams and present a simple brute-force strategy for computing
generalizations of Voronoi diagrams. a discrete approximation. We then show how we may greatly
accelerate this using graphics hardware.
2.1 Voronoi Diagrams of Points
Among the algorithms known for computing Voronoi diagrams of 3.1 Generalized Voronoi Diagrams
points in 2D, 3D, and higher dimensions are the divide-and- The set of input sites is denoted as A1, A2, …, Ak. For any point p
conquer algorithm proposed by [Shamo75] and Fortune’s in the space, dist(p, Ai) denotes the distance from the point p to the
sweepline algorithm [Fortu86]. Numerically robust algorithms for site Ai. The dominance region of Ai over Aj is defined by
constructing topologically consistent Voronoi diagrams have been
proposed by [Inaga92, Sugih94]. A number of implementations in Dom(Ai, Aj) = { p | dist(p, Ai) ≤ dist(p, Aj) }
exact and floating-point arithmetic are also available.
For a site Ai, the Voronoi region for Ai is defined by

V(Ai) = j≠i Dom(Ai , Aj)

The partition of space into V(A1), V(A2), …, V(Ak) is called the values across the triangle’s surface. All raster samples covered by
generalized Voronoi diagram. The (ordinary) Voronoi diagram a triangle have an interpolated z-value.
corresponds to the case when each Ai is an individual point. The
boundaries of the regions V(Ai) are called Voronoi boundaries. For 3.4 Our Approach
primitives such as points, lines, polygons, and splines, the Voronoi
boundaries are portions of algebraic curves or surfaces. A key concept for our approach is that of the distance function for
a site, which gives, for any point, the distance to that site. The
3.2 Discrete Voronoi Diagrams main idea of our approach is to render a polygonal mesh
approximation to each site's distance function. Each site is
Perhaps the simplest way to compute a discrete Voronoi diagram assigned a unique color ID, and the corresponding distance mesh
is to uniformly point-sample the space containing Voronoi sites. is rendered in that color using a parallel projection. We make use
For each sample point, we find the closest site and its distance. of two components of the graphics hardware: linear interpolation
Associating each point in space with its closest sample point across polygons and the Z-buffer depth comparison operation.
induces a uniform subdivision into rectangular cells. For any When rendering a polygonal distance mesh, the polygon
point, we know the distance to the closest site to within the rasterization reconstructs all distances across the mesh. The Z-
maximum distance between a point in space and a sample point, buffer depth test compares the new depth value to the previously
i.e. half the diagonal length of a cell. stored value. If the new value is less, the Z-buffer records the new
distance, and the color buffer records the site’s ID. In this way,
A simple brute-force approach to find the closest sites is to iterate each pixel in the frame buffer will have a color corresponding to
through all sample points, computing distances to all sites and the site to which it is closest, and the depth-buffer will have the
recording the closest site and distance. The algorithm can be distance to that site. In order to maintain an accurate Voronoi
rearranged to iterate through the sites: for each site, compute diagram, we bound the error of the mesh to be smaller than the
distances to all sample points and update the current closest site distance between two sample points.
and distance. The second arrangement is amenable to an
implementation in graphics hardware. Our approach is inspired by an interesting sidenote in the OpenGL
1.1 Programming Guide [Woo97]. In the Section “Now That You
Know” on “Dirichlet Domains”, the authors briefly discuss a
simple method to construct discretized 2D Voronoi diagrams for
points using OpenGL graphics hardware. The authors mention the
use of cones for Voronoi diagrams of points in 2D, but warn that
the technique “might require thousands of polygons.” We show
that we can render cones using fewer than 100 triangles for a
1K×1K resolution grid and achieve the same level of accuracy. In
addition, we generalize this approach to higher-order sites in both
two and three dimensions.

Figure 1: Image of the sampled distance functions for two point 4 THE DISTANCE FUNCTIONS
sites. Uniform point sampling induces a rectangular cell subdivision
of space. For both 2D and 3D, our discrete Voronoi diagram computation
has been reduced to finding a 3D polygonal mesh approximation
to the distance function of a Voronoi site over a planar 2D
rectangular grid of point samples. The error in the approximation
must be bounded so that by rendering this mesh using graphics
hardware, we can efficiently and accurately compute the distances
between the site and all of the point samples.
In this section, we describe the distance functions associated with
various sites, and provide efficient methods for meshing these
functions within a specified error tolerance.

4.1 2D Voronoi Diagrams

Figure 2: The two distance images are composited through a
Denote the distance from a site A to each pixel location (x,y) by
distance comparison operation. The current closest site and the dist(A,(x,y)). The distance function of A is given by
distance to each site is updated based on the lesser distance value. d(x,y)=dist(A,(x,y)). Meshing this function corresponds to
The resulting Voronoi diagram is composed of a distance image approximating the graph of d(x,y) with a polygonal model.
(left) and an closest-site ID image (right).
The three basic types of 2D sites are points, lines, and polygons.
Their corresponding distance functions are shown in the table. In
3.3 Polygon Rasterization Hardware this section, we present algorithms for computing distance meshes
for each of them.
Our approach makes use of standard Z-buffered raster graphics
hardware for rendering polygons. The frame buffer stores the
attributes (intensity or shade) of each pixel in the image space; the
Z-buffer, or depth buffer, stores the z-coordinate, or depth, of
every visible pixel. Given only the vertices of a triangle, the
rasterization hardware uses linear interpolation to compute depth
 2D site Shape of Distance Function Figure between them. By rendering the distance meshes for different
Point Right circular cone 3a features in different colors, we obtain a discrete approximation of
Line segment “Tent” 3b a per-feature Voronoi diagram. Such diagrams are useful in
Polygon Cones and tents 5 several contexts: for example, the computation of a medial axis of
Table 1: Shape of Distance Functions for 2D Sites a polygon. A picture of a per-feature Voronoi diagram for a
polygon is given in Figure 5-left.
Y

a b

Figure 3: The distance meshes used for a point (left) and a line
segment (right). The XY-plane containing the site is shown above
each mesh.

4.1.1 Points in 2D
The distance function for a point in the plane is a right circular Figure 5: The per-feature Voronoi diagram of a quadrilateral (left).
cone. We approximate cones as a triangle fan proceeding radially The corresponding distance mesh (right).
outward from the apex (Figures 3a and 4-left). A point’s Voronoi
region can potentially extend to any portion of the region of Polygons are rendered as a series of linear segments connected at
interest, and thus the radius at the cone’s base must be of size M√2 the vertices. Each edge and vertex is a feature. For the vertices,
if the scene is contained in an M×M square.The mesh’s radial lines rendering a triangle fan connecting two adjacent edges, rather than
lie on the cone. The maximum error in distance occurs at the cone a full point distance mesh cone, saves on the total number of
base between adjacent vertices. Because the cone is right circular, triangles computed and ensures that the distance meshes for
the error in approximating the circular base as viewed from above adjacent features join smoothly. See Figure 5-right for an
is equal to the error in distance. illustration.

4.2 3D Voronoi Diagrams

α R ε Our algorithm computes a 3D discrete Voronoi diagram slice-by-
R R-ε
slice. Each slice is parallel to the (x,y)-plane and is computed
independently.
α/2 Consider the slice z=z0. To construct the intersection of the
Figure 4: A single triangle of the meshed point distance function
Voronoi diagram with this slice, consider the distance function for
cone. α is the angle we wish to maximize, R is the radius of the a site A, restricted to the slice. Denote the restricted distance
cone (max dist between site and sample pt), and ε is the max error. function by dist(x,y)=dist(A,(x,y,z0)). In this section, we describe
dist(x,y) for polygon, line segment, and point sites. As in the 2D
case, computing the discrete Voronoi diagram is a matter of
From this formulation (see Figure 4), we compute the maximum meshing the distance function d=dist(x,y) for each site and
angle as: rendering these meshes.
cos(α ) =
2
R −ε
R
Ð  R −ε 
α = 2 cos−1 
 R 
 The distance meshes we give for the 3D problem are for a per-
feature Voronoi diagram. Thus, a detached triangle site is treated
For example, for a maximum distance error of no more than one as seven features: a polygon, three line segments, and three points.
pixel's width, a cone mesh for a 512×512 grid will require only 60 As in 2D per-feature diagrams, some features have a restricted
triangles. A 1024×1024 grid will require 85 triangles. region of influence.
4.1.2 Line Segments in 2D
3D site Shape of distance function Figure
The distance function for a line segment is composed of three Polygon Plane 6
parts: one for the segment itself and one for each endpoint. The Line segment Elliptical cone 7
endpoints are treated the same way as points. The distance Point 1 sheet of a hyperboloid of 2 sheets 8
function for the line segment (excluding the endpoints) is just a Table 2: Shape of Distance Functions for 3D Sites
“tent” (Figure 3b); its distance mesh is composed of two
quadrilaterals. These represent the distance function exactly, so
there is no error in the distance mesh representation. The only 4.2.1 Polygons in 3D
error for the line segment is in the cone mesh for the endpoint
The influence of this site in 3D is confined to the region formed by
distance functions, as described in the previous section.
sweeping the polygon orthogonally through space, since points
4.1.3 Polygons and Per-feature Voronoi Diagrams outside this region are considered to be closer to an edge or vertex
of the polygon. In the slice, this region is a polygon, and dist(x,y)
It is often useful to consider sites as a collection of features, rather is linear within this region, as illustrated in Figure 6. The distance
than as a single entity. For example, a line segment would be to the site is computed at the vertices of the region, and a distance
considered as three features: the two endpoints and the linear edge mesh composed of a single polygon is rendered. No meshing error
is incurred. If the polygon intersects the slice, the intersection is using table-lookup. Examples of the meshes produced by this
computed and the polygon is decomposed into two sub-polygons. method are shown in Figure 9.
Each sub-polygon is treated as above.

z d
y y
x x

Figure 6: A polygonal site and its region of influence in a slice (left). y

y
The corresponding linear distance function (right).
x
x

4.2.2 Line Segments in 3D Figure 9: A bounded-error distance mesh for the line-segment site
(left) and the point site (right).
The graph of the distance function for a line segment site is an
elliptical cone (Figure 7). The apex of the cone lies at the
intersection of the segment's line with the slice, and the cone’s 4.3 Generalization to Curved Sites
eccentricity is determined by the relative angle of the line and the
slice. The 3D region of influence of a line segment lies between The exact distance function for a curved site can be rather
two parallel planes through the endpoints, since a point outside complicated, and for splines or algebraic curves is a high-degree
these planes is closer to one of the endpoints than to the segment. algebraic function. We simplify this by creating a linear
tessellation of the curved site, and then meshing the distance
function of this approximation. We can use algorithms such as in
[Filip87] and [Kumar96] to obtain bounded-error tessellations.
Figure 10 shows the mesh for a Bézier curve. Since the mesh for a
linear segment is exact, the distance error for any of the linear
segments is just the error in the deviation of the line from the
z d
y y
original curve. The endpoints of the curve must be treated as
x x points, just as for the line segment. The distance mesh for the
“joints” between linear segments is a portion of the radial mesh of
Figure 7: A line-segment site and its region of influence in a slice triangles. An overall maximum error bound of ε can be obtained
(left). The corresponding conical distance function (right). for the entire curve by:
• tessellating the curve into linear segments with maximum error
4.2.3 Points in 3D bound of ε;
The distance function for a point site is shown in Figure 8. Its
• rendering the distance mesh for the linear segments; and
graph is one sheet of a hyperboloid of revolution of two sheets. If
the point lies in the slice, the distance function is a cone rather • treating the endpoints and joints as points, and rendering each
than a hyperboloid. The region of influence for a single point is point distance mesh with maximum error bound of ε.
the entire slice.
This approach generalizes to 3D surfaces, which can be tesselated
into a polygonal mesh. The error is bounded in a similar way.

z d
y y
x x

Figure 8: A point site and its region of influence in a slice (left). The
corresponding hyperbolic distance function (right).

4.2.4 Meshes for Line Segments and Points in 3D Figure 10: The Voronoi diagram of a Bézier curve and 5 points
The construction of bounded-error meshes for the line-segment (left). The distance mesh for the Bézier curve that has been
tessellated into 16 segments (right).
and point distance functions is detailed in [Hoff99]. The method
attempts to minimize the complexity of the mesh by committing
the maximum allowable error ε in each mesh cell. The structure of 4.4 Weighted and Farthest-site Diagrams
the mesh depends only on the resolution of the Voronoi diagram,
defined by the ratio of the diameter M of the model to the In a weighted Voronoi diagram, the distance functions are
maximum meshing error ε. The mesh structure is precomputed; additively or multiplicatively weighted [Okabe92]. Translation of
during the Voronoi diagram construction, the mesh is constructed a distance mesh along the distance axis accounts for additive
weights. Linear scaling along the distance axis accounts for neighbors that are a different color from the current pixel's. All
multiplicative weights. In 2D, this is equivalent to changing the visited pixels are marked and avoided in the recursion.
angle of the cone or tent. Scaling the distance mesh also scales the
meshing error. This algorithm also finds the Voronoi neighbors–pairs of sites that
share a Voronoi boundary. This concept is useful in a wide variety
In a farthest-site Voronoi diagram, the farthest site from each of applications, including computing the dual of the ordinary
point is found. Unlike in the nearest-site diagram, the distance Voronoi diagram–the Delaunay triangulation. The boundary
function monotonically decreases as we move away from the site. finding algorithms find pairs of adjacent pixels with different
We obtain the proper distance relationships by negating the colors. The sites corresponding to those two colors are reported to
distance functions. In practice, however, we need only reverse the be Voronoi neighbors. Connecting Voronoi neighbors with line
depth-test (less-than to greater-than) and change the depth segments constructs the Delaunay triangulation.
initialization from ∞ to 0.

Figure 12: Voronoi diagram of set of 2D points (Left). Boundaries

Figure 11: Standard nearest-site Voronoi regions (left). Farthest found with continuation-based approach (middle). Delaunay
regions for the same sites (middle). Weighted regions (right). triangulation by connecting neighboring sites (right).
Weights: line, 2; dark point, 1; light point, 0.5.

6 SOURCES OF ERROR
5 BOUNDARIES AND NEIGHBORS
In this section we analyze all sources of error in our approach, and
A continuous Voronoi diagram representation usually specifies the discuss how to reduce this error. We consider two broad
Voronoi boundaries that separate the set of Voronoi regions. In categories: error in distance approximation and combinatorial
our discrete representation, we must search for the boundaries error.
using approaches similar to iso-surface extraction and root-finding
techniques [Bloom97]. However, instead of trying to bracket zero-
crossings between sample points where iso-surface functions 6.1 Distance Error
evaluate to values of opposite sign, we simply find the boundaries Distance error is the error in the distance computed from a pixel to
in the space between pixel samples of different color. Using the a site. There are three sources of distance error:
same approaches, we can either point-sample the boundary or
compute an approximate mesh representation. In order to increase • Meshing error, from approximating the true distance function
the precision, we must either use a higher overall resolution or by the distance mesh. We discussed how to bound this error in
adaptively refine. Section 4.
One approach is to examine each pair of adjacent cells in 2D or • Tessellation error, from tessellating a curved site into a number
3D. If the colors are different, the location between the samples is of linear sites. The tessellation algorithms presented in [Filip87,
marked as a point on the Voronoi boundary. The operation is very Kumar96] give tight bounds. Tessellation error is reduced by
simple and can be accelerated through image operations in using a finer approximation to the site.
graphics hardware.
• Hardware precision error, from the use of fixed-precision
Another approach is based on a continuation method that starts at arithmetic (integer or floating-point) during rasterization.
a point known to be on the boundary and walks along the Hardware precision error cannot be removed without resorting
boundary until all boundary points have been found [Bloom97]. to multiple-precision arithmetic, but hardware error is usually
Since we only compare locations near known boundaries, it is negligible compared to meshing error.
output sensitive. The correctness of the continuation method
depends on whether the Voronoi boundaries are connected. The These errors are additive–i.e. the error from one source is not
boundaries of a generalized Voronoi diagram of a collection of magnified by the other sources. The total distance error is at most
convex sites are always connected, so the method is correct for the sum of the errors from these three sources.
inputs consisting of point, line-segment, or convex polygonal
sites. The method may fail in the presence of curves, curved 6.2 Combinatorial Error
surfaces, or concave sites where the generalized Voronoi diagram
may have isolated components. Combinatorial error refers to qualitative error as opposed to
quantitative. For example, a pixel is assigned the wrong color, or
In this approach, at least one boundary point must be known as a the algorithm reports an incorrect pair of Voronoi neighbors.
“seed” value. Assuming convex sites, some Voronoi boundary There are three sources that contribute to combinatorial error:
passes through the edge of the bounded region in which we are
computing the diagram, so the method begins by examining every • Distance error, as described in the previous section. With
window border pixel. When all Voronoi boundaries are connected significant distance error, depth comparison at a pixel may
only one seed point is needed since all others can be reached from make a farther site appear closer, causing the pixel to be
that first point. Starting from a seed point, we recursively check all colored incorrectly.
• Resolution error, a result of discrete sampling. If this sampling right color. Assume that there is no Z-buffer precision error, and
is too coarse, we may miss some Voronoi regions or find that we can bound the maximum distance error by ε, as described
spurious neighbors. Handling resolution error is described earlier. For a pixel P colored with the ID of site A and with a
below. computed depth buffer value of D, we know that:
• Z-buffer precision error, the limitations of the number of bits of D - ε ≤ dist(P,A) ≤ D + ε
precision provided by the Z-buffer. Current graphics systems
have 24 bits or 32 bits of precision for each pixel in the Z- Furthermore, we know that for any other site B,
buffer, which is more than the 23 bits provided in standard D - ε ≤ dist(P,B)
floating-point. If the distances between two pixels cannot be
determined within that precision, the Z-buffer cannot From this information, we easily determine that
accurately choose the correct color. This effect is small when dist(P,A) ≤ dist(P,B) + 2ε
compared to the other two, but can be significant at very high
resolutions with very little distance error. A higher-precision Z- where dist(X,Y) means the distance from the center of pixel X to
buffer can be simulated in software at a significant loss in site Y. That is, if a pixel is colored with the ID of A, then site A is
efficiency. no more than 2ε farther from the pixel center than any other site.
The same bound holds in 3D.
Adaptive resolution allows us to “zoom in” on a region of interest,
reducing potential resolution error. This involves identifying a
window of interest and applying the appropriate linear 7 APPLICATIONS
transformation for zooming into that region. Figure 13 shows an
example. Note that when zooming in, sites outside of the viewing There are many applications that benefit from fast computation of
region can still have Voronoi regions inside the region. Thus, the a discrete Voronoi diagram, an approximation to the distance
“maximum distance to a site” must be adjusted appropriately when function, or both. We describe three that we have implemented.
computing the distance error bounds.
7.1 Motion Planning
Motion planning is a fundamental problem in robotics and
computational geometry, with applications to the animation of
digital actors, maintainability studies in virtual prototyping, and
robot-assisted medical surgery. The classic Piano Mover’s
problem involves finding a collision-free path for a robot moving
from one location (and orientation) to another in an environment
filled with obstacles. Numerous approaches to this problem have
Figure 13: Adaptive resolution allows us to zoom in on features that been proposed, some of which are based on generalized Voronoi
could otherwise be missed. diagrams [Latom91]. The underlying idea is to treat the obstacles
as sites. The Voronoi boundaries then provide paths of maximal
Resolution error can cause a number of combinatorial problems, clearance between the obstacles. Due to the practical complexity
such as missing the entire Voronoi region of a site. One such of computing generalized Voronoi diagrams, the applications of
example is shown in Figure 14 (left two images). When no cell has such planners have been limited to environments composed of a
the color of a particular site, we can separately render the site few simple obstacles.
itself, computing the pixels covering that site. By zooming around
Our discrete Voronoi computation algorithm can be applied to
those pixels, we will find pixels in the Voronoi region of that site.
motion planning in both static and dynamic environments. The
The same technique can be applied to cells in 3D. Another
Voronoi algorithm computes the approximate distance to the
problem arising from resolution error is incorrectly finding
nearest obstacle. The basic approach we implemented is based on
Voronoi neighbors (shown in Figure 14 – right two images). This
the potential field method, which repels a robot away from the
problem (when due solely to resolution error) can be alleviated by
obstacles and towards the goal using a carefully designed artificial
adaptively zooming in on all boundary pixels.
potential function. Other Voronoi diagram or distance-based
approaches are also possible. The details of our motion planning
algorithm are provided in [Hoff99].
We demonstrate our planner’s effectiveness in a complex
environment: the interior of a house, composed of over 100,000
triangles. We use the x- and y-components of the polygons to give
Figure 14: Problems caused by resolution error. An entire region in the 2D input primitives for our algorithm. The robot has three
the center will be missed since it does not hit any pixel centers (left degrees of freedom: x- and y-translation along the ground and
two images). The left and right regions, which should meet, become rotation about the z-axis. Color plate 2 and the video show a
disconnected after rasterization (right two images). sequence of piano motions automatically generated by our motion
planner in a static environment. Color plate 2 also shows an image
of the distance function for the house. We also apply our planner
6.3 Error Bounds to environments with moving obstacles. Our video demonstrates
the movement of a music stand through a house filled with
Distance error occasionally causes a pixel to be colored moving furniture. The entire potential field and the motion
incorrectly. However, in a certain sense, the pixel is “almost” the planning sequence are computed in real time.
 400 Mhz Intel Pentium II PC with an Intergraph Intense 3D Pro
3410-T graphics accelerator was comparable to the SGI
performance. In fact, in boundary finding, neighbor finding, and
particle motion planning applications, the performance exceeded
the high-end SGI. This was mainly due to intense buffer readback
requirements. Each distance mesh must cover every pixel, so
performance is bounded by the graphics hardware’s pixel fill-rate.
For large numbers of input sites, therefore, the SGI outperforms
the PC.
When the distance-error tolerance is relaxed, the amount of
geometry rendered for each site can be reduced, slightly
Figure 15: Motion planning of falling particles. Sites are avoided by improving performance. However, the biggest gains are achieved
using the Voronoi diagram’s distance buffer (right) to create a by reducing the number of pixels filled. In many practical cases,
potential field. This same principle is used in the rigid-body planner. we can increase the performance significantly by bounding the site
distance functions to a maximum distance. This allows reduction
of the size of the distance meshes drawn so that only a portion of
7.2 Selection in Complex User Interfaces the screen is covered for each site. We exploit this observation to
obtain interactive rates in the 1,000-point example shown in color
Complex 2D user interfaces sometimes require quick plate 1, in the 10,000-point example shown in the video, and in the
determination of the object nearest to the cursor. The Voronoi general case for the computation of the potential field used in the
diagram of the interface can be used as a nearest-object lookup motion-planner. For closed higher-order primitives, such as
table indexed by sample points. Given the cursor position, it is polygons, we can further increase performance by restricting the
simple to find the nearest sample point, and thus the nearest distance function to only the inside or outside regions. This is
object. In some interfaces it may be desirable to know the distance useful in computing potential fields and medial axes.
to the selected object as well. We used this technique in our 2D
implementation to allow the user to interactively move sites with
the mouse. 9 CONCLUSIONS AND FUTURE WORK
We have presented a method for rapid computation of generalized
7.3 Mosaics discrete Voronoi diagrams in two and three dimensions using
We can use our approach for generating Voronoi diagrams to graphics hardware. We have presented techniques for creating a
create an interesting artistic effect called mosaicing. A mosaic is a mesh of the distance function for each site with bounded error, and
tiled image, where each tile has a single color. The Voronoi described how this distance mesh allows us to compute the
diagram of a point set can be used as a tiling [Haebe90]. Each Voronoi diagram rapidly. We have analyzed various sources of
Voronoi tile is colored with a color taken locally from the image. error, as well as how to bound or reduce those errors. Finally, we
In our implementation, each tile is colored by the image pixel have demonstrated a few applications using our approach.
closest to the point site (see color plate 1). Our algorithm can In the future, we would like to extend this work in the following
perform this operation very quickly, allowing dynamic mosaics in ways: generalizations of distance functions and site geometry,
which the mosaic tiling, the source image, or both may change in further applications, other distance meshing strategies, and more
real time. acceleration techniques for the 3D Voronoi volume computation.
By randomly distributing point sites across an image, we obtain an
effect similar to many mosaic filter effects seen in image editing ACKNOWLEDGEMENTS
programs. By clustering point sites around areas of higher detail,
we obtain a classic tiling seen in many real-life mosaics where Supported in part by ARO Contract DAAH04-96-1-0257 and
smaller tiles are used in areas of greater detail. DAAG55-98-1-0322, NSF Career Award CCR-9625217, NSF
grants EIA-9806027 and DMI-9900157, NIH Research Resource
Award 2P41RR02170-13, ONR Young Investigator Award and
8 IMPLEMENTATION Intel. We would also like to thank Sarah Hoff for extensive help
For the 2D case, we implemented a complete interactive system with editing and color plates, Chris Weigle for suggesting
incorporating all of the features and applications described here. mosaicing using 2D point Voronoi diagrams, the UNC-
Example output is shown throughout the paper. The video walkthrough group for the house model, and the reviewers for
demonstrates interactive computation of more complex diagrams. their helpful comments.
In 3D, we show results from a prototype system that uses a
simpler distance meshing strategy (see color plate 3 and the video REFERENCES
for example output).
[Auren91] F. Aurenhammer. Voronoi Diagrams: A Survey of a
We implemented the 2D and 3D systems in C++ using the Fundamental Geometric Data Structure. ACM Computing
OpenGL graphics library and the GLUT toolkit. Any graphics API Surveys, 23:345−405, 1991.
[Bloom97] J. Bloomenthal, C. Bajaj, J. Blinn, M-P. Cani-Gascuel, A.
specification that uses a standard Z-buffered interpolation-based Rockwood, B. Wyvill, and G. Wyvill. Introduction to Implicit
raster graphics system is sufficient to support the Voronoi diagram Surfaces. Morgan Kaufmann Publishers, Inc. San Francisco,
computation. Motion planning and the basic operations of CA. 1997.
boundary and neighbor finding require reading back of the color [Chian92] C–S. Chiang. The Euclidean Distance Transform. Ph. D. thesis,
Dept. Comp. Sci., Purdue Univ., West Lafayette, IN, August
and depth buffers. Our system runs, without source modification, 1992. Report CSD-TR 92−050.
on both an MS-Windows-based PC and a high-end SGI Onyx2
with InfiniteReality Graphics. Surprisingly, the performance on a
[Culve99] T. Culver, J. Keyser, and D. Manocha. Accurate Computation [Vleug95] J. Vleugels and M. Overmars. Approximating Generalized
of the Medial Axis of a Polyhedron. Proc. of the Fifth Symp. on Voronoi Diagrams in Any Dimension. Technical Report UU-
Solid Modeling and Applications. 1999. CS-1995-14, Dept. of Comp. Sci., Utrecht University, 1995.
[Dutta93] D. Dutta and C.M. Hoffmann. On the Skeleton of Simple CSG [Vleug96] J. Vleugels, V. Ferrucci, M. Overmars, and A. Rao. Hunting
Objects. Journal of Mechanical Design, ASME Transactions, Voronoi Vertices. Comp. Geom. Theory Appl., 6:329−354,
115(1):87−94, 1993. 1996.
[Diric50] G.L. Dirichlet. Uber die Reduktion der Positiven [Voron08] G.M. Voronoi. Nouvelles Applications des Paramètres
Quadratischen Formen mit Drei Unbestimmten Ganzen Continus à la Théorie des Formes Quadratiques. Deuxième
Zahlen. J. Reine Angew. Math., 40:209−27, 1850. Mémoire: Recherches sur les Parallélloèdres Primitifs. J.
[Filip87] D. Filip and R. Goldman. Conversion from Bézier-rectangles to Reine Angew. Math., 134:198−287, 1908.
Bézier-triangles. CAD, 19:25−27, 1987. [Woo97] M. Woo, J. Neider, and T. Davis. OpenGL Programming
[Fortu86] S. Fortune. A Sweepline Algorithm for Voronoi Diagrams. In Guide, Second Edition. Addison Wesley, 1997.
Proc. 2nd Annual ACM Symp. on Comp. Geom., pages
313−322, 1986.
[Goldf89] J. Goldfeather, S. Molnar, G. Turk, and H. Fuchs. Near Real-
time CSG Rendering Using Tree Normalization and Geometric
Pruning. IEEE Computer Graphics and Applications,
9(3):20−28, May 1989.
[Haebe90] P. Haeberli. Paint by Numbers: Abstract Image Representation.
Computer Graphics (SIGGRAPH ’90 Proc). vol. 25. pgs 207-
214.
[Held97] M. Held. Voronoi Diagrams and Offset Curves of Curvilinear
Polygons. Computer-Aided Design, 1997.
[Hoff99] K. Hoff, T. Culver, J. Keyser, M. Lin, and D. Manocha. Fast
Computation of Generalized Voronoi Diagrams Using
Graphics Hardware. Technical Report TR99-011, Dept. of
Comp. Sci., University of North Carolina at Chapel Hill, 1999.
[Hoffm94] C.M. Hoffmann. How to Construct the Skeleton of CSG
Objects. In A. Bowyer and J. Davenport, editors. Proc. of the
Fourth IMA Conference, The Mathematics of Surfaces,
University of Bath, UK, Sept. 1990. Oxford University Press,
New York, 1994.
[Inaga92] H. Inagaki, K. Sugihara, and N. Sugie. Numerically Robust
Incremental Algorithm for Constructing Three-dimensional
Voronoi Diagrams. In Proc. 4th Canad. Conf. Comp. Geom.,
pgs 334-339, 1992.
[Kumar96] S. Kumar, D. Manocha, and A. Lastra. Interactive Display of
Large NURBS Models. IEEE Trans. on Vis. and Computer
Graphics. vol 2, no 4, pgs 323-336, Dec 1996.
[Latom91] J.C. Latombe. Robot Motion Planning. Kluwer Academic
Publishers, 1991.
[Laven92] D. Lavender, A. Bowyer, J. Davenport, A. Wallis, and J.
Woodwark. Voronoi Diagrams of Set-theoretic Solid Models.
IEEE Computer Graphics and Applications, 12(5):69−77, Sept
1992.
[Lee82] D.T. Lee. Medial Axis Transformation of a Planar Shape.
IEEE Trans. Pattern Anal. Mach. Intell., PAMI−4:363−369,
1982.
[Lengy90] J. Lengyel, M. Reichert, B.R. Donald, and D.P. Greenberg.
Real-time Robot Motion Planning Using Rasterizing Computer
Graphics Hardware. Computer Graphics (SIGGRAPH ‘90
Proc.), vol. 24, pgs 327−335, Aug 1990.
[Milen93] V. Milenkovic. Robust Construction of the Voronoi Diagram of
a Polyhedron. In Proc. 5th Canadian. Conference on Comp.
Geom., pgs 473-478, 1993.
[Milen93b]V. Milenkovic. Robust Polygon Modeling. Computer Aided
Design, 25(9), 1993. (special issue on Uncertainties in
Geometric Design).
[Okabe92] A. Okabe, B. Boots, and K. Sugihara. Spatial Tessellations:
Concepts and Applications of Voronoi Diagrams. John Wiley
& Sons, Chichester, UK, 1992.
[Rossi92] J. Rossignac, A. Megahed, and B. Schneider. Interactive
Inspection of Solids: Cross-sections and Interferences.
Computer Graphics (SIGGRAPH ‘92 Proc.), vol. 26, pgs
353−360, July 1992.
[Rossi86] J.R. Rossignac and A.A.G. Requicha. Depth-buffering Display
Techniques for Constructive Solid Geometry. IEEE Computer
Graphics and Applications, 6(9):29−39, 1986.
[Sheeh95] D.J. Sheehy, C.G. Armstrong, and D.J. Robinson. Computing
the Medial Surface of a Solid from a Domain Delaunay
Triangulation. In Proc. ACM/IEEE Symp. on Solid Modeling
and Applications, May 1995.
[Shamo75] M.I. Shamos and D.Hoey. Closest-point Problems. In Proc.
16th Annual IEEE Symposium on Foundations of Comp. Sci.,
pages 151−162, 1975.
[Sugih94] K. Sugihara and M. Iri. A Robust Topology-oriented
Incremental Algorithm for Voronoi Diagrams. International
Journal of Comp. Geom. Appl., 4:179−228, 1994.
[Sherb95] E.C. Sherbrooke, N.M. Patrikalakis, and E. Brisson.
Computation of the Medial Axis Transform of 3D Polyhedra. In
Solid Modeling, pages 187−199. ACM, 1995.
[Teich97] M. Teichmann and S. Teller. Polygonal Approximation of
Voronoi Diagrams of a Set of Triangles in Three Dimensions.
Tech Rep 766, Lab of Comp. Sci., MIT, 1997.