A Spatial Data Structure for Fast Poisson-Disk Sample Generation

Daniel Dunbar Greg Humphreys

University of Virginia University of Virginia

Figure 1: A Poisson-disk distribution of 17,593 points generated in 80 ms.

Abstract particular, blue-noise patterns perform especially well in this

setting because of the low-energy annulus around the DC
spike in their frequency spectrum. High quality sampling
Sampling distributions with blue noise characteristics are patterns are especially important when sampling the image
widely used in computer graphics. Although Poisson-disk plane in a raytracer, not only because they do a better job of
distributions are known to have excellent blue noise char- capturing the continuous function being sampled, but also
acteristics, they are generally regarded as too computation- because in this setting the function being reconstructed is
ally expensive to generate in real time. We present a new displayed directly, so any sampling errors will be especially
method for sampling by dart-throwing in O(N log N ) time apparent to a viewer.
and introduce a novel and efficient variation for generating
Poisson-disk distributions in O(N ) time and space. Poisson-disk distributions have excellent blue noise spectra
and also mimic the distribution of photoreceptors in a pri-
CR Categories: I.3.3 [Computer Graphics]: Pic- mate eye [Yellot 1983]. These distributions have proven diffi-
ture/Image Generation—Antialiasing; I.4.1 [Image Process- cult to generate directly, so many alternate approaches have
ing and Computer Vision]: Digitization and Image been developed, few of which can guarantee the Poisson-
Capture—Sampling disk property. In this paper, we describe an O(N log N )
algorithm for directly generating maximal Poisson-disk dis-
Keywords: sampling, blue noise, poisson disk tributions identical to those produced by a dart-throwing
technique. We then present a variation of this algorithm
that both yields better spectral distributions and runs in
linear time and space. This algorithm generates point sets
1 Introduction and Background with excellent blue noise characteristics very quickly; it can
generate over 200,000 points per second on a modern CPU.
Almost all problems in computer graphics involve sampling.
It is well known that the properties of the sampling distri-
bution can greatly affect the quality of the final result. In 1.1 Previous Work

Copyright © 2006 by the Association for Computing Machinery, Inc.

Sampling theory is a well researched area of research in
Permission to make digital or hard copies of part or all of this work for personal or computer graphics, and it has even deeper roots in the sig-
classroom use is granted without fee provided that copies are not made or distributed nal processing and information theory literature. Stochastic
for commercial advantage and that copies bear this notice and the full citation on the sampling was first introduced to computer graphics by Dippé
first page. Copyrights for components of this work owned by others than ACM must be
honored. Abstracting with credit is permitted. To copy otherwise, to republish, to post on
and Wold [1985]. Cook analyzed the spectral properties of
servers, or to redistribute to lists, requires prior specific permission and/or a fee. various stochastic point processes [Cook 1986]. In that pa-
Request permissions from Permissions Dept, ACM Inc., fax +1 (212) 869-0481 or e-mail per, he extols the virtues of Poisson-disk distributions be-
[email protected].
© 2006 ACM 0730-0301/06/0700-0503 $5.00

503
cause of their blue noise properties and relationship to pho- 2 Dart-Throwing in O(N log N )
toreceptor distributions, but ultimately advocates the use
of jittered grids because the straightforward dart-throwing
The dart-throwing method for computing Poisson-disk dis-
algorithm for generating Poisson-disk distributions is pro-
tributions iteratively refines an existing point set by gen-
hibitively expensive. Since then, many algorithms have been
erating a series of random candidate points in the sample
proposed for generating point distributions; e.g. [Ulichney
domain and keeping only the first such point that is farther
1988; Shirley 1991; Hiller et al. 2001; Kollig and Keller 2002;
than the minimum distance 2r from all other points. Each
Kollig and Keller 2003].
sample effectively invalidates a disk of radius 2r centered
Mitchell’s O(N 2 ) “best-candidate” algorithm attempts to around that point.
mimic dart-throwing while providing a termination guaran- This algorithm is simple to implement and extends naturally
tee [Mitchell 1991]. Whenever a new sample is to be drawn, to any domain with a well-defined and computable distance
a number of candidate samples are randomly generated, and metric. However, the algorithm may not terminate, so in
the candidate that is farthest from the existing point set is practice the algorithm is stopped after some fixed number
accepted. This algorithm cannot guarantee the Poisson-disk of consecutive candidates have failed to be accepted.
property, but in practice it generates excellent point sets if
enough candidates are drawn. The primary drawback of this One side effect of this approximation is that the generated
technique is its long running time. point set is usually not maximal (there may be regions where
a point could be placed without violating the distance crite-
McCool and Fiume generated high-quality tile sets with a rion), so some regions of the domain may be undersampled.
toroidal distance function so that they could be repeatedly This problem is usually ignored, although Jones’ technique is
tiled across the plane [McCool and Fiume 1992] and intro- guaranteed to generate maximal distributions [Jones 2006].
duced the use of Lloyd’s relaxation to improve the blue noise Furthermore, since a large number of sample points is typ-
properties of the set. Lloyd’s relaxation transforms a point ically required, the algorithm is too slow to use directly.
set by moving each point to the center of its associated Consequently, several schemes for precomputing small dis-
Voronoi region and is typically applied iteratively or used tributions and tiling them have been proposed [Hiller et al.
to generate distributions directly [Hiller et al. 2001]. Once 2001; Cohen et al. 2003; Lagae and Dutré 2005].
the tiles have been generated, this method generates large
point sets very efficiently. However, the tiling process intro- We call the subdomain within which it is legal to add a point
duces easily recognizable structures. This is acceptable for the available subdomain. Let D(x, r) be the disk of radius r
most sampling applications as long as the tile size is large around a point x. For a domain X and existing point set P ,
enough, but is disastrous if the points are being used to dis- the available subdomain is given by
tribute objects or as part of a texture basis function. Several [
schemes have been proposed to solve this problem. AX = X − D(p, 2r).
p∈P
Multiple authors have proposed using Wang tiles to solve
this problem. Cohen et al. populated a set of Wang tiles
with small point distributions that were intended to tile the The key to emulating dart-throwing efficiently is the obser-
plane [Cohen et al. 2003]. Lagae et al. showed that cor- vation that we do not need to sample the entire available
rectly applying this technique requires careful attention to subdomain. Consider the annulus between radii 2r and 4r
how points are placed at the boundary of a tile and presented around some point. Every point in this annulus must be un-
Poisson-disk tiles to address these issues [Lagae and Dutré available in any maximal distribution and therefore within
2005]. Most recently, Kopf et al. have extended these tech- a distance of 2r from some other point. This means that
niques to allow generation of point sets with blue noise prop- there must be at least one point that lies inside the annu-
erties satisfying an arbitrary density function [Kopf et al. lus, and hence the union of all such annuli, intersected with
2006]. Their technique produces high quality point sets very the available subdomain, must contain at least one point.
efficiently once the tile set has been computed. Therefore, it is possible to emulate dart-throwing by sam-
pling from only this region, which we call the available neigh-
Ostromoukhov et al. described a fast technique for impor- borhood of a point set P . By carefully choosing the repre-
tance sampling a provided density function [Ostromoukhov sentation for this region, dart-throwing can be implemented
et al. 2004]. Their algorithm generates point sets with lo- in O(N log N ) time.
cal blue noise characteristics by using a clever modification
of Penrose tiles and exploiting the tilings’ aperiodic nature.
However, when used to generate many points from a con- 2.1 Representing the Available Neighborhood
stant density function, the resulting point sets have larger
angular anisotropy than techniques based on randomness. The problem of representing the available neighborhood can
be divided into two parts: developing a spatial structure
Jones presents an algorithm for generating 2D Poisson-disk for the available region of the annulus, and partitioning the
distributions in O(N log N ) time [Jones 2006]. Like our tech- available neighborhood so that these structures can be effi-
nique, Jones’ method builds a point set incrementally by ciently updated upon the insertion of a new point. In ad-
storing neighboring regions of points in a balanced tree and dition, it must be possible to quickly generate a uniformly
inserting points into these regions one by one. However, distributed random point inside the region.
neighbor regions are represented as Voronoi cells and Jones’
method requires an incremental Delaunay triangulation al- Our solution to the first part of this problem involves a new
gorithm with O(log N ) performance when adding a point. data structure, the scalloped region. This data structure can
This makes the implementation more complicated than that be used to efficiently represent arbitrary boolean operations
of our algorithm. Furthermore, in this paper we show a on 2D disks. We divide scalloped regions into a disjoint
variation of our algorithm that runs in linear time. union of scalloped sectors (Figure 2). A scalloped sector

504
 Although we do not have a closed form equation for A−1 , A
h(α) is the integral of a function that is both non-negative and
zero only at the angular endpoints, so A is monotonically
increasing and therefore invertible. In practice binary search
is sufficient to evaluate A−1 efficiently.
A(α)
The result of a boolean union, intersection, or difference op-
eration between a single scalloped sector and a disk can al-
ways be subdivided into a small number1 of new scalloped
C2 sectors. A scalloped region maintains a list of its constituent
r2 scalloped sectors and operations are performed by replac-
g(α) ing each sector with the result of applying the operation
to each individual sector. A complete description of how
α2 to locate the new scalloped sectors requires enumerating a
r1 C1 number of cases and is omitted here for brevity; the full de-
α1 tails can be found in our extended technical report [Dunbar
and Humphreys 2006].
The available neighborhood is partitioned into scalloped sec-
tors of outer radius 4r around each point in order to restrict
the number of sectors that must be updated after point in-
sertion to a small constant. Efficient sampling, however,
requires that all of these neighborhoods be disjoint. In gen-
Figure 2: A scalloped sector is a sector bounded above and eral, if an ordering relation is defined for a set S of sets it is
below by circular arcs. In the above diagram, α ranges from possible to derive a new set S 0 of disjoint sets where
α1 to α2 , and g(α) and h(α) are the distance functions from [ [
the sector’s apex to the near and far arcs, respectively. A(α) s0 = s,
is the partial area of the sector up to α, represented by the s0 ∈S 0 s∈S
light gray area in the figure.
by subtracting from each set the union of all members of S
that are less than it in the relation.
centered at the origin is defined as the sector lying between We use the generation order of the points as an ordering re-
angles α1 and α2 and bounded by near and far circular arcs. lation and then define the available neighborhood of a point
For convenience, the circular arcs are required to be non- p ∈ P as
intersecting, although they are allowed to meet at the edge [  D(p0 , 4r), p0 < p
of the sector. Np = D(p, 4r) − .
D(p0 , 2r), p0 ≥ p
The circular arcs are each described by a center C, radius r, p0 ∈P
and a sign k ∈ {−1, 1} which selects either the near or far
arc of the circle. The distance to this circular arc along a
S
The available neighborhood is N = p∈P Np (Figure 3).
ray at angle α is then given by Each disjoint Np is computed using boolean disk subtraction.
p
d cos(α − γ) + k r2 − (d sin(α − γ))2 ,
2.2 Algorithm Details and Complexity Analysis
where (d, γ) are the polar coordinates of C (taking the apex
of the sector to be the origin). If g(α) and h(α) represent Our algorithm A1 for efficient dart-throwing begins with an
the distance functions to the inner and outer bounding arcs initial set consisting of a single point randomly chosen in
respectively then the area of the sector up to an angle α is the domain. During sample generation, we maintain an as-
given by sociative map from candidate points (points with non-empty
Z α Z h(θ)
available neighborhoods) to their associated neighborhoods.
A(α) = r dr dθ,
α1 g(θ) A candidate point is then randomly chosen (using neigh-
borhood areas as a probability distribution) and a random
and the area of the full sector is A(α2 ). point within its neighborhood is added to the point set. The
The probability of a random point in a scalloped region available neighborhood for the new point is an annulus from
falling within a particular sector is the area of the sector radii 2r to 4r, minus a disk of radius 4r around the nearby
divided by the area of the region. A uniform random point points. The maximum distance required to search for neigh-
within the region can be generated by randomly choosing bors is 8r since the scalloped region and neighbor disk are
a sector using the sectors’ areas as a probability distribu- both bounded by 4r. All nearby neighborhoods are then up-
tion and then generating a uniformly distributed point inside dated by subtracting a disk of radius 2r around the newly
the chosen sector. Generating the point inside the sector is inserted point. This process continues until no candidate
done by transforming a uniformly distributed random point points remain.
(ξ1 , ξ2 ) in [0, 1]2 to a point (d, θ) in polar coordinates, where The maximum number of scalloped sectors in an available
p neighborhood is bounded by a constant. Furthermore, the
d = g(θ)2 + ξ1 (h(θ)2 − g(θ)2 ) 1 Bounded by a constant that depends only on the particular
−1
θ = A (ξ2 A(α2 )). operation (e.g, seven for disk subtraction).

505
 for applications such as randomized object placement, in
which it is desirable to tune the density of the point set.
Np1
A special case arises if r0 is taken to be the minimum value
2r. In this case, a point’s available neighborhood collapses
to a collection of circular arcs centered at the point. We
call these arcs the available boundary. By directly imple-
p1 menting boundary sampling, we no longer need to represent
the available neighborhood as scalloped regions; instead, the
available boundary is represented as a set of per-point angu-
lar ranges at which a point can be placed on the boundary.

p2 p3 Additionally, if we select the new candidate point at random

instead of according to the length of its available boundary
(similarly to how we obtained the linear algorithm A2 ), it
is no longer necessary to explicitly store the neighborhoods
Np 2 for every point already in the set. Once a candidate has
been chosen, its available boundary can be quickly computed
by intersecting the boundary circles of the candidate with
Np 3 its immediate neighbors. After the legal ranges have been
determined, we can repeatedly place new points at available
locations on its boundary until the available boundary is
empty. The addition of a new point only requires subtracting
Figure 3: A partial point set and its neighborhoods. The a single angular range from the candidate’s boundary.
dashed lines represent scalloped sector boundaries within a The resulting algorithm A3 , which we call boundary sam-
region. pling, is simple to implement and runs in O(N ) time and
space (pseudo code is given in our technical report [Dunbar
and Humphreys 2006] and an implementation is available
Poisson-disk distance condition bounds the number of neigh- from our website). Our implementation is approximately
bors within a fixed radius. We can therefore use a uniform 200 lines of C++ code and can generate over 200,000 points
grid to implement the neighbor search and update of the per second on a 3 GHz Pentium 4. Figure 1 shows an exam-
available neighborhoods in O(1) time. Similarly, picking an ple point set generated using our algorithm.
individual scalloped sector within an available neighborhood
and generating a point in that sector can be done in O(1)
time. By storing the available neighborhoods in a balanced
tree structure, we can choose an available neighborhood and 4 Results
update the tree within O(log N ) time, so the time complex-
ity of the entire algorithm is O(N log N ) where N is the In this section, we show results from the boundary sampling
number of generated points. The space complexity is O(N ). algorithm described in Section 3, and compare them to other
If we drop the requirement that the available neighborhoods methods for computing Poisson-disk distributions.
be sampled according to an area-weighted probability den- The best tiling schemes can generate point sets very ef-
sity function then this new algorithm A2 runs in linear time. ficiently and with spectra comparable to dart-throwing
This is a significant theoretical speedup, but in practice the or Lloyd’s relaxation, although there will be energy and
cost of maintaining the sectors, intersecting them with disks, anisotropy spikes associated with any tiling. Although we
and updating data structures dominates the running time. do not currently have access to an implementation of a so-
phisticated tiling method, we expect that the runtime per-
formance of our algorithm is comparable to that of a tiling
3 Boundary Sampling scheme, but our results will be artifact-free and require no
precomputation.
In this section, we show how the algorithms described in the We analyze the properties of two-dimensional noise distri-
previous section can be modified to avoid the complexity of butions in the style of McCool and Fiume, who compute
sector operations, thereby generating Poisson-disk point sets the radial power and anisotropy using the periodogram of a
in linear time extremely quickly. In particular, we generalize point set [McCool and Fiume 1992]. The primary charac-
either A1 or A2 by making the outer radius of the annulus teristic of a blue noise distribution is a low energy annulus
defining the available neighborhoods a parameter r0 , where around the central DC spike with energy returning to a rel-
2r < r0 ≤ 4r. This modification does not change the struc- atively constant value outside the annulus. The quality of a
ture or performance of either algorithm, and we will assume distribution depends on the magnitude of the difference be-
that they are parameterized by r0 for the rest of the paper. tween the DC spike, the low energy annulus, and the average
With this change, the available neighborhood is defined as energy in the high frequencies. Evaluating the distribution
in terms of radial power also requires analyzing the radial
0
[  D(p0 , r0 ), p0 < p anisotropy to ensure that the radial power spectrum is an
Np = D(p, r ) − . accurate representation of the pattern along all orientations.
D(p0 , 2r), p0 ≥ p
p0 ∈P
Figure 4 shows the radial power spectra of boundary sam-
Notice that the overall density of the generated point set will pling compared to both dart-throwing and linearized dart-
tend to be inversely proportional to r0 ; this can be exploited throwing. The graph shows that neither linearized sampling

506
 N 1000 10000 100000
Average Radial Power (dB) A1
Best Candidate 1.454 157.014 6.084h
A2
Dart-throwing (A1 ) 0.573 5.905 141.901
A3
O(N) Dart-throwing (A2 ) 0.667 6.442 61.186
Ostromoukhov et al. 0.015 0.095 1.546
Boundary Sampling (A3 ) 0.001 .058 0.496

Table 1: Timing results for generating point sets of varying

sizes. Times are in seconds except where otherwise noted.

O(N log N ) time and motivates a new algorithm for gener-

Radial Frequency at Center of Annulus ating 2D Poisson-disk point sets that runs in O(N ) time and
space and produces excellent blue noise patterns. Here we
Figure 4: The blue noise properties of the distribution are present three related avenues for future research.
preserved even if only the boundary of the available neigh-
borhood is sampled. Sampling without regard to the prob-
ability density of available neighborhoods also preserves the 5.1 Sampling on Arbitrary Manifolds
blue noise properties.
The boundary sampling method can be extended to sample
over any manifold with an associated distance metric. All
nor boundary sampling significantly change the blue noise that is required is an efficient method for finding the local
properties of the resulting distributions. neighbors of a point and for determining the intersection
of adjacent boundaries. As an example, consider generat-
Figure 5 displays averaged periodograms for 100 point sets ing a Poisson-disk-like point set on the surface of a sphere.
having a radius of 0.02, resulting in approximately 2000 The natural distance function is the length along arcs of
points each. The boundary sampling periodogram shows great circles, and the boundary of a point is a small circle
that the higher point density results in a greater magnitude (the circle resulting from the intersection of a plane with
difference between the low energy annulus and the peak tran- the sphere). Available portions of this boundary can still
sition energy. The periodograms also show that the method be represented as angular ranges and determined by small
of Ostromoukhov et al. is significantly less uniform than circle intersection. Sampling can still be done in linear time
that of our method. They address this issue by precom- by dividing the domain into a grid; for optimal performance,
puting relaxation vectors, but these precomputed tables are a grid should be used with roughly equi-areal cells.
only sufficient to improve the blue noise properties of small
local regions of the generated distribution. As the number of
points grows, the lookup table is no longer able to compen- 5.2 Generating Infinite Point Sets
sate for the inherent structure of the Penrose tiling. Because
we are interested in using these point sets for sampling the
image plane in high quality image synthesis, hundreds of mil- It is not always easy to predict the number of required points
lions of points will likely be required, and the lookup tables for an application. In these cases, it is convenient to be able
would become quite large. to generate points near an arbitrary position in the plane.
This ability also permits a Poisson-disk texture basis func-
Timing results for several methods of computing Poisson- tions as in Lagae and Dutré’s method [2005], where val-
disk distributions are shown in Table 1. For methods that ues such as a unique nearest-point identifier or distance to
require specification of a radius, one is chosen so that the the closest point are returned for any 2D location. Existing
number of generated points is approximately equal to the methods use precomputed tiles to solve this problem.
given value of N , and the time is computed as N times the
average number of points per second. The times for Ostro- We have successfully applied scalloped sectors to the prob-
moukhov’s method were generated with code provided by lem of sampling from infinite point sets. By judicious seed-
the authors of that paper, although they state that the pro- ing of a pseudorandom number generator, we are able to
vided code is not fully optimized. The results show that generate points in arbitrary subsets of the plane while guar-
although the linear approximation of dart-throwing (A2 ) is anteeing that the resulting point sets will not violate the
more efficient than true dart-throwing (A1 ) for large num- Poisson-disk property when placed next to neighboring point
bers of points, the computational overhead of sector subtrac- sets generated by our technique. This seems like a promising
tion still adds significant overhead compared to boundary way to compute infinite point sets without tiling artifacts;
sampling (A3 ). For small numbers of points, A2 may per- we have not yet implemented any applications that leverage
form more disk subtractions than A1 , making it less efficient this capability.
than its O(N log N ) counterpart.
5.3 Importance Sampling
5 Conclusion and Future Work Given an importance function over the domain, it is possi-
ble to derive a non-Euclidean distance metric for which our
We have described a new technique for efficiently imple- approach would yield a point set whose density was locally
menting the dart-throwing algorithm for the generation of proportional to that function. These point sets should have
Poisson-disk point sets, based on the manipulation of dis- local blue noise properties as in Ostromoukhov et al. [2004]
joint unions of scalloped sectors. This algorithm runs in and Kopf et al [2006].

507
 (a) A1 : Dart-throwing (b) Best Candidate (c) Ostromoukhov et al. (d) A3 : Boundary sampling

Figure 5: Averaged periodograms for several sampling methods. Boundary sampling generates the best blue noise spectrum
due to its extremely regular and dense sampling of the plane. Ostromoukhov et al.’s results are noisier because their technique
does not involve randomness, so averaging multiple runs does not provide smooth periodograms.

In general, this metric will not have a closed analytic form, Kollig, T., and Keller, A. 2002. Efficient multidimen-
and therefore the boundary cannot be directly determined. sional sampling. Computer Graphics Forum 21, 3, 557–
Nonetheless, it would be possible to solve for a piecewise 563.
linear (or higher order, if necessary) approximation of the
distance metric in the vicinity of a point and use that to Kollig, T., and Keller, A. 2003. Efficient illumination
represent the local boundary. By applying this technique by high dynamic range images. Rendering Techniques,
and the method in 5.1 we hope to be able to extend our 45–51.
technique to support efficient importance sampling. Kopf, J., Cohen-Or, D., Deussen, O., and Lischinski,
We intend to apply these results combined with those in the D. 2006. Recursive wang tiles for real-time blue noise.
previous section in order to be able to place objects with ACM Transactions on Graphics 25, 3.
varying radii across unbounded regions. We expect this to be Lagae, A., and Dutré, P. 2005. A procedural object
valuable for generating complex terrain and environments. distribution function. ACM Transactions on Graphics 24,
4, 1442–1461.
McCool, M., and Fiume, E. 1992. Hierarchical poisson
References disk sampling distributions. In Proceedings of the confer-
ence on Graphics interface ’92, Morgan Kaufmann Pub-
Cohen, M. F., Shade, J., Hiller, S., and Deussen, O. lishers Inc., San Francisco, CA, USA, 94–105.
2003. Wang tiles for image and texture generation. ACM
Transactions on Graphics 22, 3, 287–294. Mitchell, D. P. 1991. Spectrally optimal sampling for dis-
tribution ray tracing. In Computer Graphics (Proceedings
Cook, R. L. 1986. Stochastic sampling in computer graph- of ACM SIGGRAPH 91), ACM Press, New York, NY,
ics. ACM Transactions on Graphics 5, 1, 51–72. USA, 157–164.
Dippé, M. A. Z., and Wold, E. H. 1985. Antialiasing Ostromoukhov, V., Donohue, C., and Jodoin, P.-M.
through stochastic sampling. In Computer Graphics (Pro- 2004. Fast hierarchical importance sampling with blue
ceedings of SIGGRAPH 85), ACM Press, New York, NY, noise properties. ACM Transactions on Graphics 23, 3,
USA, 69–78. 488–495.
Dunbar, D., and Humphreys, G. 2006. Using scalloped Shirley, P. 1991. Discrepancy as a quality measure for
sectors to generate poisson-disk sampling patterns. Tech. sample distributions. In Proceedings of Eurographics, 183–
Rep. CS-2006-08, University of Virginia. 194.
Hiller, S., Deussen, O., and Kaufmann, A. 2001. Tiled Ulichney, R. A. 1988. Dithering with blue noise. In Proc.
blue noise samples. In VMV ’01: Proceedings of the Vision of the IEEE 76, 56–79.
Modeling and Visualization Conference 2001, Aka GmbH,
265–272. Yellot, J. I. 1983. Spectral consequences of photoreceptor
sampling in the rhesus retina. Science 221 , 382–385.
Jones, T. 2006. Efficient generation of poisson-disk sam-
pling patterns. Journal of Graphics Tools, to appear .

508