EUROGRAPHICS 2012 / P. Cignoni, T.

Ertl
(Guest Editors)
Volume 31 (2012), Number 2
Procedural Generation of Parcels in Urban Modeling
Carlos A. Vanegas
2,1
, Tom Kelly
3,1
, Basil Weber
1
, Jan Halatsch
4
, Daniel G. Aliaga
2,4
, and Pascal Mller
1
1
Esri R&D Center Zurich, Switzerland
2
Department of Computer Science, Purdue University, USA
3
University of Glasgow, UK
4
ETH Zurich, Switzerland
a
b
c
d
I
g
area
area dev
min width
min width dev
Observed Parcels Procedural Parcels
Area
Statistical Comparison
Visualization oI procedural
parcel aspect ratios
3D models on
procedural parcels
e
Aspect Ratio
procedural
observed
1:1
10:1
20:1
30:1
h
i
Aerial Image
Skeleton detail OBB detail
Figure 1: Procedural Parcel Generation. Our method creates parcels inside city blocks (f,i) using two dierent subdivision
techniques skeleton (g, shaded part of f) or OBB (h, unshaded part of f). The subdivision attributes are automatically
extracted from observed real-world cities (a,b,c) or determined by the user. The resulting parcel congurations closely resemble
real-world subdivisions, as shown by our statistical and visual comparison of procedural and observed parcel datasets (d,e).
Abstract
We present a method for interactive procedural generation of parcels within the urban modeling pipeline. Our
approach performs a partitioning of the interior of city blocks using user-specied subdivision attributes and
style parameters. Moreover, our method is both robust and persistent in the sense of being able to map individual
parcels from before an edit operation to after an edit operation - this enables transferring most, if not all, cus-
tomizations despite small to large-scale interactive editing operations. The guidelines guarantee that the resulting
subdivisions are functionally and geometrically plausible for subsequent building modeling and construction. Our
results include visual and statistical comparisons that demonstrate how the parcel congurations created by our
method can closely resemble those found in real-world cities of a large variety of styles. By directly addressing the
block subdivision problem, we intend to increase the editability and realism of the urban modeling pipeline and to
become a standard in parcel generation for future urban modeling methods.
Categories and Subject Descriptors (according to ACM CCS): I.3.5 [Computer Graphics]: Computational Geometry
and Object ModelingI.3.6 [Computer Graphics]: Methodology and Techniques
1. Introduction
Interactive large-scale urban modeling is becoming increas-
ingly popular in computer graphics research and in applica-

joint rst author

tions to several elds including gaming, urban planning, and
navigation tools. A key reason for the popularity of inter-
active urban modeling is the ability to quickly create large
complex models. The major steps of a typical modeling task
c 2011 The Author(s)
Computer Graphics Forum c 2011 The Eurographics Association and Blackwell Publish-
ing Ltd. Published by Blackwell Publishing, 9600 Garsington Road, Oxford OX4 2DQ,
UK and 350 Main Street, Malden, MA 02148, USA.
C. Vanegas, T. Kelly, B. Weber, J. Halatsch, D. Aliaga and P. Mller / Procedural Generation of Parcels in Urban Modeling
include (i) creating the underlying terrain and general land-
use pattern of the city or urban area, (ii) generating an in-
terconnected road network, (iii) subdividing the space in be-
tween roads (i.e., blocks) into parcels, and (iv) populating
parcels with buildings, parks, and other urban structures. In
this paper, we focus on the third major step in this pipeline:
providing a comprehensive and fully interactive approach
for subdividing blocks into parcels, a task which has been
largely ignored in previous computer graphics systems po-
tentially resulting in unrealistic and implausible results.
The subdivision of blocks and generation of parcels has
been partially addressed in previous works. Within computer
graphics research, Parish and Mller [PM01] pioneered a
procedural approach to urban modeling and many derivative
works have since been created. Vanegas et al. [VAW

10] pro-
vides a recent state-of-the-art report of related urban mod-
eling methods. While previous automatic block subdivision
methods take into account providing egress (i.e., ensuring
a parcel has street access), the methods do not always pro-
duce plausible parcel shapes, sometimes only support sim-
ple block shapes, and can yield areas within a block that are
not assigned to any parcel. Within urban design and plan-
ning research, the focus of parcel generation methods has
been on satisfying the major interests of real-estate investors
and complying with zoning and building law regulations
[PCDS04]. While automatic subdivision is well exploited in
computer graphics, in urban design blocks are partitioned
according to desired patterns (e.g., [TKD10, PPC08]) but the
labor is often performed manually, a task which does not
scale well to large urban modeling applications. Moreover,
in either graphics or urban design, altering an existing citys
geometry can cause unexpected changes in a blocks subdi-
vision, leading to dicult shape control and editing. Further,
this challenge is only exacerbated during interactive editing.
For example, a small change to road geometry can cause the
subdivision for an entire neighborhood of parcels to be un-
willingly altered both in number and in shape this lack of
parcel persistence can cause the loss of prior customizations.
The key motivation behind our work is to develop a gen-
eralized block subdivision method inspired by urban design
guidelines, suitable for intuitive large-scale interactive edit-
ing, and able to reproduce the parcel shapes and congura-
tions observed in many cities. In urban design, a block usu-
ally consists of one of two parcel varieties [Car03]:
The rst variety has parcels whose front-side is along a
street and rear-side is adjacent to another instance of the
same parcel variety.
The second variety includes parcels that may also be ad-
jacent to streets but can include a string of interior parcels
separated by small pathways or alleys, instead of formal
streets.
Moreover, parcels usually have a rather regular or uniform
structure that is typically a deep rectangle, wide rectangle,
approximate square, quadrilateral, or sometimes polygonal
[Cur97]. We seek to automate the subdivision of arbitrary
block shapes, ensure the aforementioned set of urban charac-
teristics are met, and provide user-controlled realistic parcels
that approximate the forms used in practice. This level of
automation gives more time to designers to concentrate on
high-level design decisions, including during virtual world
content creation.
Our approach for parcel generation uses a combination of
two subdivision methods to reproduce the aforementioned
two parcel varieties, including mixed-types, and to ensure a
set of subdivision attributes are satised. The input to our
method is a set of interconnected roads where an ordered se-
quence (loop) of road segments denes a block to be sub-
divided into parcels (Section 3). Any block can have po-
tentially one or both prototypical parcel varieties. The over-
all regularity and shape of all parcels is controlled by user-
specied subdivision attributes that ensure: (i) the parcels
collectively partition the block (i.e., there should be no un-
used/unassigned space), (ii) all parcels have the option of
street access (i.e., egress), (iii) parcels have a simple exterior
boundary, often nearly rectangular, and the parcels size and
aspect ratio is controllable, and (iv) parcels are aligned as
best as possible with the adjacent street segment, if any (Sec-
tion 4). Further, our solution includes a robust mechanism to
map individual parcels from before an interactive edit oper-
ation to after the edit operation this enables transferring
most, if not all, customizations, despite there being signi-
cant changes to the underlying road network and block ge-
ometry (Section 5). Our framework also supports the gener-
ation of urban models of up to half a million parcels of arbi-
trary shapes. As shown in our results (Section 6), the styles
aorded by our method combined with our expressive set of
attributes allow for a large variety in the subdivision results
and support the generation of many subdivision styles found
in real world urban layouts.
Succinctly, our solution improves the urban modeling
pipeline for all future methods by achieving:
realism we focus on providing an automatic subdivision
algorithm that is able to produce the patterns used by ur-
ban designers in real-world cities;
persistence we address the challenge of consistently la-
beling blocks and parcels so as to enable a best mapping
of parcels from before an edit operation to after an edit
operation, thereby enabling persistence of a priori cus-
tomizations; and
interactivity we support fully interactive editing (i.e.,
move, copy, transform) of intersection points, road geom-
etry, and parcel attributes at local and citywide scales.
2. Previous Work
We provide a brief review of previous and related work for
generating the 2D and 3D geometry of urban spaces. Parish
and Mller [PM01] introduced an initial approach in which
L-systems were adapted to resemble the growth of streets.
Subsequent block subdivision was implemented as an algo-
rithm that recursively divides the longest pair of approxi-
c 2011 The Author(s)
c 2011 The Eurographics Association and Blackwell Publishing Ltd.
C. Vanegas, T. Kelly, B. Weber, J. Halatsch, D. Aliaga and P. Mller / Procedural Generation of Parcels in Urban Modeling
mately parallel edges until parcel sizes are under a user-
specied threshold area. Parcels with no street access are dis-
carded. This simple algorithm does not necessarily produce
realistically shaped parcels and leaves odd-shaped empty ar-
eas inside city blocks that do not belong to any parcel.
Later work has built upon Parish and Mllers paper
and further developed dierent components. Several papers
have focused on providing realistic building content (e.g.,
[WWSR03,MWH

06,LWW08]). Other works have concen-

trated on road networks. For example, Chen et al. [CEW

08]
use hyper-streamlines and tensor elds to generate a road
network but do not provide a novel automatic block subdi-
vision algorithm. Galin et al. [GPMG10] provide a method-
ology to generate realistic roads between a source and des-
tination but do not address parcel generation. Vanegas et al.
[VABW09a] describe a block subdivision algorithm which
assumes parcels are mostly rectangular. Their method com-
putes the oriented bounding box of the parcel and uses the
middle (long) axis to optionally divide the block into two
areas, which are then partitioned into the same number of
parcels. While all parcels will have egress, other subdivision
styles are not supported, nor is parcel persistence addressed.
More recently, Lipp et al. [LSWW11] proposed a method
that enables editing an urban layout. While their approach is
interactive and does support a type of editing persistence, it
does not focus on block subdivision. They assume the sub-
division is present in the initial layout and use very simple
heuristics to subdivide small blocks that may appear dur-
ing an editing process (e.g., around the fringes of a street
or neighborhood that is moved or transformed interactively).
Several works have integrated urban simulation engines into
the urban modeling pipeline. For example, Weber et al.
[WMWG09] describe a geometrical simulation of a growth
of a city of overtime. Vanegas et al. [VABW09b] integrate
an urban behavioral simulator with procedural modeling.
Neither paper innovates block subdivision rather they re-
implement the methods in [PM01, AVB08] or [VABW09a].
In urban design and planning, automatic methods are
rarely used. The few available solutions (e.g., [WD11,
HKS08, Mar09, WCP

11]) do not go beyond basic imple-

mentations and have very limited stylistic control (Figure 2).
In contrast to previous work, we propose a guided space-
partitioning based scheme. Since we perform a space-
partitioning, we avoid the presence of unassigned areas. The
automatic processing of our method, which seeks to satisfy
a set of subdivision attributes, supports a range of block
subdivision styles, and produces regular and nicely-shaped
parcels despite partitioning arbitrarily shaped blocks (i.e.,
not only nearly rectangular blocks). Finally, our solution is
designed to support parcel persistence and fast interactive
editing for both local and city scale operations.
3. System Overview
In this section, we provide an overview of our interactive
editing system and underlying data structure. In particular,
Figure 2: Comparison to existing approaches. Given a street
network (a), several systems have been proposed to create
parcels. The Voronoi tessellation of points near the bor-
der of the street is one geometrical approach (b), while a
minimum area bounding box approach has also been sug-
gested [WCP

11] (c). Cube packing is another approach

that leads to badly formed parcels in concave areas (d). We
show our result (e) which gives a statistically realistic result.
A comparison to [WMWG09] is presented in Appendix C.
we explain our graph-based road network data structure and
our procedure for the extraction and updating of blocks from
a given road graph. To support (live) interactive editing of
road and block attributes, we also demonstrate how to trig-
ger updates to block subdivisions. Note that we do not ex-
plicitly discuss how to compute road geometry nor building
geometry, since this is not the objective of our work.
3.1. Data Structure
A city, or fragment of a city, is represented by road network
graph and its dual containing all city blocks. The road net-
work is a planar graph (V, E) with nodes V and edges E.
Each node n V is dened by its position in R
2
and log-
ically corresponds to a road intersection point. Each edge
e E stores the indexes of its source and target nodes and
logically corresponds to a road segment. In our implementa-
tion, the edge also stores the width of the street that it repre-
sents, and an interpolation type that determines whether the
edge is a straight line segment or a curve. In the latter case,
the adjacent nodes are interpolated using Bezier splines.
A city block B corresponds to exactly one face of the road
network graph. City blocks are connected by a shared road
segment. Thus the set of city blocks form the dual of the
road network graph. This dual graph can be stored explicitly
or recomputed each time it is needed from the road graph (as
is the case in our system - see Section 3.2). Each city block
B stores several attributes that determine how the block is
subdivided into parcels, including subdivision style, parcel
area and width bounds and split irregularity - see Section 4.1.
Some of these parameters are common to groups of blocks,
while others are specic to individual blocks.
3.2. Live Interactive Editing System
As with most urban modeling systems, our implementation
supports editing the road geometry and block parameters, ei-
ther for individual roads/blocks or for large areas of the city.
The editing system determines that the contour geometry of
the blocks needs to be updated (i) after road graph topology
changes and/or (ii) after any road/block attribute changes.
c 2011 The Author(s)
c 2011 The Eurographics Association and Blackwell Publishing Ltd.
C. Vanegas, T. Kelly, B. Weber, J. Halatsch, D. Aliaga and P. Mller / Procedural Generation of Parcels in Urban Modeling
If a topology change occurs (e.g., addition or removal of a
road segment), the set of blocks is updated through a pla-
nar face traversal method that eciently extracts the cycles
of the graph. For every new block the contour geometry is
computed. If an attribute change occurs (e.g. a node trans-
lation or a street width change), the set of aected blocks is
computed and the contour geometry of the respective blocks
is updated. Block subdivision (Section 4) is triggered for a
given block B whenever there is a change in either the geom-
etry of the contour C(B) of the block or in the values of the
subdivision attributes for the block. The parameters for indi-
vidual blocks are mapped to the corresponding set of blocks
as determined by our parcel consistency method (Section 5).
4. Block Subdivision
This section provides an overview of our block subdivision
method, describes the sought after subdivision attributes,
and explains our two main subdivision algorithms.
4.1. Algorithm Overview
Our approach consists of two main subdivision algorithms
that enable each of the two prototypical parcel varieties in-
spired from urban design work (see Section 1). The subdivi-
sion scheme to model the rst parcel variety is based on the
straight skeleton [AAAG95] of C(B). The straight skeleton
of a polygonal block ensures the following design require-
ments: (i) the resulting regions always have street access,
which is one of the main constraints for this parcel variety;
and (ii) the rear side of the generated parcels is always abut-
ting another parcel a characteristic of this parcel variety.
The rear-side neighbor parcels can be either parcels with
street access themselves, or a central patio parcel with no
street access. The latter case is known as perimeter block
(Figure 3 right), which according to urban planners is an
optimal design pattern to achieve very high urban densities
without high-rise buildings. Intuitively, such a perimeter lay-
out can be obtained by creating parcels around the me-
dial axis of the block contour, and we exploit the fact that
the straight skeleton is a linear approximation of the medial
axis. Our choice of the straight skeleton is further supported
by our observations of how urban planners often, out of in-
tuition, draw an approximate medial axis to homogeneously
subdivide blocks, especially in residential areas.
The second subdivision scheme is based on a recursive
splitting concept as introduced in [PM01] and performs an
adaptive spatial partitioning of C(B) using oriented bound-
ing boxes (OBBs). The result is parcels resembling the sec-
ond variety, namely the presence of parcels on the perimeter
of the block with egress and parcels interior to the block pre-
sumably separated by small pathways or alleys. Hybrid uses
of these two methods is also possible. For example, a large
block can be subdivided using skeleton-based subdivision to
yield parcels of a chosen maximumdepth around the perime-
ter and the interior space can be optionally partitioned using
OBB-based subdivision.
The algorithms are controlled by high-level user-specied
stylistic control parameters and also support mixed-type
congurations. In all cases, the partitioning process is per-
formed independently for each block. Let B be a block to be
subdivided. Let C(B) = {b
1
, b
2
, . . . , b
m
} be the ordered set of
m vertices describing the contour of B. The vertices are spec-
ied in a counter-clockwise direction. In general, our subdi-
vision algorithms partitions B into a set L = {l
1
, l
2
, . . . , l
n
} of
n regions such that
_
n
i=1
l
i
= B and
l
i
,l
j
L
(l
i
l
j
= ).
While our methods support an innite number of subdivi-
sions for a given block, only some such subdivisions are ad-
equate for mimicking land subdivision in real-world cities.
Hence, several constraints are considered when determining
suitable subdivisions, including a fairly homogeneous area
distribution among the resulting parcels, parcel areas, as-
pect ratios and angles that meet structure construction con-
straints, and access fromthe parcels to the road network. The
set of subdivision attributes that our system enforces for all
parcels is provided below.
Parcel area bounds (A
min
, A
max
): The upper and lower
bounds on the area of the resulting parcels. For all l
i
,
A
min
< A(l
i
) < A
max
.
Minimum parcel width (W
min
, W
max
): The upper and
lower bounds on the length of the sides of the oriented
bounding box of a parcel.
Split irregularity : The deviation of a split edge from its
default position, normalized in [0, 1]. Larger values result
in the split being further away from the mid-point, and
generally, in a higher variance in the parcel areas.
The detailed pseudocode for both styles is presented in Al-
gorithm 1 of Appendix A.
4.2. Skeleton-based Subdivision
This subdivision style creates parcels resembling the rst
prototypical parcel variety, i.e., parcels whose front-sides are
along a street, and rear-sides are adjacent to other parcels.
We made several observations which directed the design
of our algorithm. The rst was that every parcel should have
street access, and that the direction of this street access at
intersections varied. This lead us to the concept of strips of
parcels, which follow roads. Further observations were that
the initial split was frequently the common centerline be-
tween rows of parcels on either side of the block, and the
rear edges of the parcels were normally straight.
A polygonal skeleton supports these observations as it
identies both the potential strips and the centerline. We
found the straight skeleton the most realistic, since other
techniques, such as the medial axis, introduced curved seg-
ments. Therefore our st subdivision method computes the
initial split contour via the straight skeleton, identifying the
depth of the parcels via strips. Subsequent splits cast rays
from the street edges to position the sides of the parcels.
4.2.1. Initial Split
c 2011 The Author(s)
c 2011 The Eurographics Association and Blackwell Publishing Ltd.
C. Vanegas, T. Kelly, B. Weber, J. Halatsch, D. Aliaga and P. Mller / Procedural Generation of Parcels in Urban Modeling
Figure 3: Left: The straight skeleton of a polygon (black)
consists of arcs (dark blue) which dene faces adjacent
to each polygon edge. Right: A partial application of the
straight skeleton computes the oset contour C

(B) (dashed
lines). Note that the contour may split the innermost, or pa-
tio, region into several portions (a,b). The individual faces
of the straight skeleton conform to our denition of a strip,
and we may take the supporting edges to be portions of the
boundary of input polygon. We take these faces to be the ini-
tial set of -strips.
To dene the initial split contour, the user denes a perpen-
dicular distance d
o f f set
from the block contour C(B) to an
inwards oset contour C

(B). Intuitively, this value corre-

sponds to the maximum depth (distance from the road to the
rear) of the parcels. If the d
o f f set
value is suciently large
(e.g., innity), then the area enclosed by the initial split con-
tour collapses and the rear side of any resulting parcel will be
directly adjacent to another parcel. If the d
o f f set
value is suf-
ciently small, the initial split contour denes a closed inner
patio region, with no direct access to roads. This inner re-
gion may be disconnected if the initial block is non-convex,
and can be further partitioned using an arbitrary subdivision
style. While setting an innite value for d
o f f set
is typical and
better complies with the rst parcel variety, inner patios are
not uncommon and we designed our skeleton-based subdi-
vision to also support them. The contour C

(B) is calculated
(via the CGAL library [cga]) by a partial application of the
straight skeleton to C(B) (Figure 3), i.e., by computing the
intersection of the roof model of [AAAG95] with a horizon-
tal plane at a specic height.
The arcs of the skeleton application specify the division of
the region between C(B) and C

(B) into a set of strip poly-

gons. We initially refer to these as -strips, to dierentiate
them from the -strips, from which we have removed some
diagonal edges. These strips are an intermediate value in our
algorithm, representing a group of parcels with their primary
frontage on the same logical street. Collectively they form a
single connected region.
A strip, s
i
, is a simple polygonal area within B, such that
a single connected length of the polygons boundary forms
part of C(B). These lengths are the supporting edges, (s
i
),
of s
i
. A blocks cyclic list of strips, LS (B) = s
1
. . . s
n
, is such
that it covers the area between C(B) and C

(B) without over-

lap. The list LS (B) is ordered counter clockwise, such that
the last supporting edge of s
i
and rst of s
i+1
are adjacent
edges of C(B). Note that we take i +1 to mean (i +1) mod n
in the context of the cyclic list.
We initialize LS (B) from the faces of the straight skeleton
Figure 4: The -strips (solid colors, a) are recovered from
the skeleton and logical streets (bold lines, a). This leaves
undesirable diagonal edges (red dashed lines, a). Given
the classication T(v
i
) None, Previous (P) or Next (N),
we reassign regions (shaded, b), to create the set of -
strips. In the example (c-e) we use the classication scheme
S treetLength, specifying the direction Previous. The subse-
quent splits are computed over these -strips (f).
used to compute C

(B). We observe that these faces fulll the

strip properties bounded by the arcs of the skeleton, and
supported by an edge of C(B). Any strip in LS (B) may be
combined with either of its neighbors and the union retains
the strip properties; Therefore we may union adjacent faces
in LS (B) according to whether they lie on the same logical
street. In this manner we create a single -strip for every
logical street (Figure 4a).
4.2.2. Removing Diagonal Edges
The -strips computed from the skeleton faces suer from
diagonal edges at the intersection of logical streets (Figure 4,
a). To correct these edges, we modify LS (B) (Figure 4, b-e)
to transfer a near-triangular region from the strip on one side
of an oending edge to the strip on the other side. We refer
to these corrected strips as -strips.
Let the shared supporting vertex between each pair of -
strips s
i
and s
i+1
, be designated v
i
. The shared boundary
of these two strips forms the diagonal edges we are con-
cerned with, one end of which is v
i
. We provide a classi-
cation T(v
i
) {Previous, Next, None}, to specify which of
the pair of strips will gain the region, and which will lose
the same region. A vertex with the property Previous (re-
spectively Next), will assign a region to the previous (next)
strip, given the counter-clockwise vertex ordering. No action
is taken with a value of T(v
i
) = None.
The values of T(v
i
) may be assigned by one of two
schemes, determined by a per-parcel parameter specied by
the user. When the angle of the supporting edges at v
i
is re-
ex we always assign T(v
i
) = None. For the remaining v, we
choose between the following two schemes:
S treetWidth. If the average width of the supporting edges
of s
i
is greater (respectively lesser) than those of s
i+1
then
T(v
i
) = Previous (Next).
S treetLength. If the length of the supporting edges of s
i
is
greater (respectively lesser) than those of s
i+1
then T(v
i
) =
Previous (Next).
c 2011 The Author(s)
c 2011 The Eurographics Association and Blackwell Publishing Ltd.
C. Vanegas, T. Kelly, B. Weber, J. Halatsch, D. Aliaga and P. Mller / Procedural Generation of Parcels in Urban Modeling
Figure 5: The slice operation divides a -strips area (solid
color) into parcels using the supporting edges (bold lines)
to position rays (dashed lines) (a). An example showing the
skeleton arcs guiding the rays (b).
We nd that in most most urban environments the parcels
face the most important and thus widest street; Therefore we
use the S treetWidth scheme by default. There are situations
in which this assumption is not suited. For example, residen-
tial street patterns in which parcels prefer to face the quieter,
and longer, residential streets, rather than the wider access-
streets. In our experiments it proved dicult to make this
distinction automatically, so we allow the user to assign this
parameter manually.
Given the parameter T(v
i
), we calculate the direction in
which to reassign the region, (Figure 4b), creating the set of
-strips. The region is removed from strip s
x
, where s
x
= s
i
if T(v
i
) = Next, or s
x
= s
i+1
if T(v
i
) = Previous. The region
is removed by identifying the point on the boundary of both
s
i
and s
i+1
that is furthest from C(B), (Figure 4c, i), and cut-
ting to the nearest point on (s
x
), (Figure 4c, ii), which can
be reached by a straight line interior to the strip (Figure 4d).
The region is then unioned to the strip s
y
where s
y
= s
i+1
if T(v
i
) = Next, or s
y
= s
i
if T(v
i
) = Previous. Finally we
recompute the skeleton arcs to remain perpendicular to the
local edges (s
y
), (Figure 4e, iii), in order to ensure we are
able to guide the subsequent splits in the following stage
(Figure 4f). After processing each pair of adjacent strips in
LS (B), we are left with the list of -strips. Any strips of zero
area are removed (Figure 4a, cyan and brown).
The motivation behind this process is to (i) extend the
lots centerline such that it reaches C(B) as orthogonally as
possible, and (ii) allow a parameter controlling the direction
(street access) of the parcel at intersections.
4.2.3. Splitting Strips into Parcels
To subdivide the -strips into parcels, a set of points are
sampled approximately equidistantly on (s
i
) (Figure 5).
Rays from these points, perpendicular to the nearest edges
of (s
i
), split the -strip into parcels. The distance between
the points is normally distributed around (W
min
+W
max
)/2,
with
2
= 3, and clamped to the length of (s
i
). This pro-
cess adds a random displacement to the ray-origin points to
create less uniform parcels. To prevent local perturbations in
(s
i
) adversely aecting the parcel geometry, we limit the
rays to each skeleton face. If the ray crosses a skeleton edge,
it follows the edge to the boundary of the strip (Figure 5b).
There are several special cases that are handled indepen-
dently as post processing steps:
There are situations in which the block is too shallow to
accommodate the two rows of parcels assumed by the al-
gorithm. In this case we group shallow parcels and replace
them with parcels generated similarly to the skeleton sub-
sequent split technique. An example occurs in the detail
of Figure 1f, in an L-shaped lot in the bottom right corner.
If a parcel is deep, the area may be unacceptably large.
To mitigate this eect, blocks of area larger than A
max
are
split again, via a second ray cast from the street.
Triangular parcels and parcels with small areas are re-
peatedly unioned with their neighbors until they are either
larger than the minimum area (A
min
), neither small nor tri-
angular, or there is only one remaining parcel. We union
such parcels with the adjacent parcel with which it shares
the longest edge.
4.3. OBB-based Subdivision
This subdivision style creates parcels resembling the second
prototypical parcel variety, i.e., quadrilateral parcels with
and without street access, by using an adaptive spatial parti-
tioning. To determine the initial split line, the minimum-area
OBB of the block is computed. The pivot point of the split
line is given by the midpoint of the largest edge of the OBB,
translated by a random distance proportional to . The direc-
tion of the split line is given by the direction of the smallest
edge of the OBB. It is worth noting, that if the block shape
is near-rectangular, the middle axis of the OBB is roughly
equivalent to the straight skeleton but will not be curved, or
overly aected by reex corners in the street graph. Thus,
while we could use the straight skeleton to generate the ini-
tial split line, using the blocks OBB turns out in practice to
yield a more robust mechanism.
For subsequent recursive splits, we also found OBB-based
partitioning to yield very robust and well-behaved subdi-
visions. At each step, each parcel is split into two smaller
parcels, which are recursively split into more parcels until
the area of the resulting parcels is within (A
min
, A
max
) or the
width of the front side of the parcel is within (W
min
, W
max
).
If a parcel resulting from a split has no street access, the di-
rection of the split line is rotated 90 degrees about the normal
vector of the plane containing the block, with probability ,
where is a user-specied attribute in [0, 1] that indicates the
preference for parcels with street access. Larger values of
result in more parcels having street access: If = 1 street
access is always guaranteed, while if = 0 parcels with no
street access are frequent. See (Algorithm 2 in Appendix A)
for more details.
Our method takes into account the following additional
considerations during subdivision:
edge alignment - to increase the performance of subdi-
visions under interactive editing operations, we nd the
best-tting OBB by tting OBBs that align with at least
one of the parcel edges,
random seeds - to increase the stability of subdivisions
under interactive editing operations, the random seeds for
c 2011 The Author(s)
c 2011 The Eurographics Association and Blackwell Publishing Ltd.
C. Vanegas, T. Kelly, B. Weber, J. Halatsch, D. Aliaga and P. Mller / Procedural Generation of Parcels in Urban Modeling
a b c
d
e
f
g
Result
Figure 6: Oriented bounding box subdivision. This adap-
tive algorithm recursively splits a parcel into two smaller
parcels along the minor axis of the oriented bounding box
of the original parcel. The subdivision continues until user-
specied shape attributes are satised.
the child parcels of a given parcel are computed before the
recursive call to the subdivision function,
snap to block contour vertices - if the split line is within a
threshold distance from one of the vertices of the contour
of the original block, the pivot point of the split line is set
to that vertex.
This recursive approach considers only the geometry of
the parent shape and road access, instead of the geometry
of the entire block (Figure 6). As a result, recursive OBB
subdivision is very stable in the sense that localized changes
in the block geometry will typically only aect the parcels
nearby the parts of the geometry that are modied.
Notice that the inner region subdivisions could alterna-
tively be computed using the same slice operation as the
skeleton subdivision, since the two regions obtained upon
the initial OBB split conform to our denition of a strip.
While the slice operation is suitable for the relatively thin
strips generated by the initial split contours in skeleton sub-
divisions, applying this operation for subsequent splits in the
OBB style will likely result in parcels with atypical aspect
ratios. In practice, for inner regions, recursive use of OBB
splits has proven to be more suitable than slice operations
for obtaining subdivisions that comply with user-specied
shape parameters.
5. Parcel Consistency under Live Editing
This section presents the problem of consistent relative lo-
cation of any given parcel within a block and discusses its
relevance to live interactive editing of urban models. It also
proposes a solution to the parcel consistency problem that
has been implemented in our parcel generation system.
5.1. Consistent Relative Location
Changes in the geometry or the topology of a road net-
work require updating not only the road network itself (e.g.,
lanes, sidewalks, intersections), but also the shapes dened
by the road network, specically blocks and parcels. A sys-
tem that follows the urban modeling pipeline sequentially
typically proceeds by rst creating the road network, then
extracting the blocks as faces of a planar graph, and -
nally subdividing the extracted blocks into parcels (e.g.,
[PM01, VABW09b, WMWG09]). In this type of systems,
when an urban model is regenerated, the connection between
the entities of the previous model and the entities of the new
model has little importance. In live editing, however, edit-
ing operations include large-scale modications to the road
network, small displacements of road network vertices or
edges, and changes in the attributes of a subdivision. Many
of these edits are suciently small and localized for a cor-
respondence between the sets of parcels at two consecutive
states to be expected and desired. For this reason, any urban
modeling system that aims to support interactive editing of
road networks, blocks and parcels, must support parcel loca-
tion consistency. In spite of this requirement, existing urban
modeling systems have not addressed or proposed solutions
to the parcel consistency problem.
The consistent location of parcels under live editing op-
erations is crucial to guarantee persistence in the relative
location of specic objects in the urban models. Consider,
for instance, a parcel with an associated set of parameter-
ized rules that procedurally generates a building in the par-
cel. Assume that the rule parameters determine the geome-
try and textures of the building. If no matching between the
parcels of two consecutive subdivisions of a same block was
enforced, such parameters would either be lost or would be
inherited at random from the parcels in the previous subdi-
vision. By avoiding this randomness, the relative location of
a building with a distinctive style is preserved.
5.2. Parcel Matching and Inheritance
Our live editing method (section 3.2) proposes a solution to
the parcel consistency problem by computing a matching
between the sets of the parcels of a block at two consecu-
tive editing steps t and t +1 (Figure 7). The pairs of matched
parcels are used to establish inheritance of parcel-specic at-
tributes. Our matching approach uses the relative position of
a parcel inside its block and a metric to compute the distance
between instances of a single parcel at consecutive time steps
of the editing. The chosen metric is robust to changes in the
vertices of the block.
Let f be a user editing operation that modies one or
more of the following: the position of the vertices in the
contour C(B) of a block B; the number of vertices in C(B);
or the attributes used in the subdivision of B. Let B
t
and
B
t+1
be two instances of a block, before and after the edit-
ing operation f , respectively. Let L
t
=
_
l
t
1
, l
t
2
, . . . , l
t
m
_
be the
set of m parcels that results from subdividing B
t
, and L
t+1
=
_
l
t+1
1
, l
t+1
2
, . . . , l
t+1
n
_
be the set of n parcels that results from
subdividing B
t+1
. The values of m and n may or may not be
equal.
The goal of the consistent relative position step is to nd
a matching between the elements of L
t
and L
t+1
(Figure 8).
The matching aims to minimize the dierence between the
position of a parcel in L
t+1
and its matched parcel in L
t
, rel-
ative to the vertices of the block contour C(B). More speci-
cally, for each parcel l
t+1
i
L
t+1
, the matching function nds
c 2011 The Author(s)
c 2011 The Eurographics Association and Blackwell Publishing Ltd.
C. Vanegas, T. Kelly, B. Weber, J. Halatsch, D. Aliaga and P. Mller / Procedural Generation of Parcels in Urban Modeling
Figure 7: A street network (bottom left) may be interactively
modied in several ways (a-f). We use our matching func-
tion to compute the location of each parcel, and show the
location of a given parcel (blue house) after each editing
operation; moving a street graph vertex to enclose a block
(a); moving an edge (b); editing the curve of two edges(c);
changing the width of a street (d); changing the subdivision
style (e) or scaling the street network (f).
a parcel l
t
j
L
t
such that

l
t
h
L
t
_
d
_
l
t+1
i
, l
t
j
_
d
_
l
t+1
i
, l
t
h
__
where d : L
t
L
t+1
R is some distance function.
We need to nd an adequate distance function d robust
to rigid and non-rigid transformations applied to all or some
of the vertices C(B) =
_
b
1
, b
2
, . . . , b
|C(B)|
_
of the block. For
instance, translations, rotations and scaling of C(B), regard-
less of their magnitude, should not alter the matched pairs
__
l
t+1
i
, l
t
j
__
. Relatively small translations, rotations and scal-
ing of a subset of C(B) should also not alter the pairs.
Our solution denes a distance function inspired on a gen-
eralization of barycentric coordinates for irregular, n-sided
polygons, and is loosely based on [MLBD02]. This gener-
alization exploits that the barycentric coordinates of a point
on a triangle are invariable to rigid transformations, and that
they exhibit a relatively small change when the positions of
the triangle vertices are moved a small distance. The distance
function is given by
d(l
t+1
i
, l
t
j
) =
|C(B)|

k=1
_

l
t+1
i,k
l
t
j,k

_
where l
t+1
i,k
is the Euclidean distance from the center of l
t+1
i
to b
k
, divided by the sum of the Euclidean distances from
the center of l
t+1
i
to all b B. It follows from this denition
that
_
|C(B)|
k=1
_
l
t+1
i,k
_
= 1. Our distance function allows uniquely
identifying the position of a point relative to the positions of
the vertices of a convex or non-convex polygon.
An assumption made by our consistency mapping solu-
tion is that both the order and the number of the vertices of
B remain unchanged. One of the most common editing cases
is adding a new road segment with one endpoint splitting one
of the road segments enclosing B, forming a T-shaped junc-
tion and increasing the number of vertices in C(B) by one.
However, in our solution the new vertex is explicitly ignored
b
0

b
4

b
3

b
2

b
1

b
0

b
1

B
t

b
2

b
4

b
3

B
t+1

l
j
t
l
h
t
f
l
i
t+1
Figure 8: Relative Parcel Position Consistency. Live urban
layout editing requires computing a correspondence relation
between the parcels in a block before and after an editing op-
eration. Our approach uses a generalization of barycentric
coordinates and a distance function to estimate the relative
location of a parcel inside a block and a matching parcel.
during consistency computation while it remains collinear
to its adjacent vertices along the contour of B. While many
other edits are possible (e.g., a non-collinear vertex is added
or removed), it is often unclear, even to a human, how con-
sistency is dened and established between a block before
and after an editing operation. Thus, although arbitrary edits
can be performed with our GUI, we have not explicitly ad-
dressed all cases. Similarly, certain sequences of geometric
changes (e.g., scaling a block down and then back up) re-
sult in some parcels being removed and then added again. In
order to support subdivision consistency throughout the en-
tire editing sequence, comparison with several previous sub-
division states would be required. Our user experience and
system design decision is to support consistency only with
respect to the previous state.
6. Results
Our approach has been used to generate parcels with dier-
ent subdivision styles and attributes inside blocks of vary-
ing areas, aspect ratios and irregularity (Figure 9). Our al-
gorithms have been implemented within CityEngine, a large
software application for 3D city modeling [Cit]. All of the
parcels in our results can be generated from scratch in un-
der 3 seconds, and the subdivision of one block after local
editing can be done at interactive rates of between 1 and 10
milliseconds per block.
In order to evaluate how well our method reproduces par-
cel congurations of real-world urban spaces, we present for
each result a visual and statistical comparison to subdivi-
sions in areas of selected cities. The evaluation process con-
sists of (i) choosing a set of blocks from an existing city
for which GIS parcel data is available, (ii) automatically ex-
tracting per-block descriptive statistics of the subdivision at-
tributes of the observed parcels in each block, including the
mean and standard deviation of the parcel area (

A, s
A
) and
minimum width (

W, s
W
), (iii) loading the blocks into our
application and assigning the extracted set of per-block par-
cel attributes (Section 4.1) and a subdivision style to each
block, (iv) subdividing the blocks into parcels using the im-
plementation of our method, and (v) automatically extract-
c 2011 The Author(s)
c 2011 The Eurographics Association and Blackwell Publishing Ltd.
C. Vanegas, T. Kelly, B. Weber, J. Halatsch, D. Aliaga and P. Mller / Procedural Generation of Parcels in Urban Modeling
Figure 9: Examples of varying dierent attributes of OBB
(a,b) and skeleton, subdivision (c-g); a large and small dif-
ference between A
min
and A
max
(a); the eect of enforc-
ing street access (b); low or high lot-width (c); editing the
street widths, to change T(v
i
) (d); editing the criteria for
minimum-area lot removal (e); low or high value of d
o f f set
(f); a higher value of , and a variety of subdivision styles
in the patio region (g).
ing descriptive statistics of the attributes of the procedural
parcels in each block, including the area, minimum width,
aspect ratio and number of neighbors. The subdivision at-
tributes used in (iii) are assigned automatically to each block
and computed from the descriptive statistics extracted in (ii)
as follows: A
min
=

Aks
A
, A
max
=

A+ks
A
, W
min
=

Wks
W
,
W
max
=

W + ks
W
, where k is a positive constant that in
our examples was set to 2. The subdivision style for each
block or selected region of a city is determined by the user.
The comparison between the procedural parcels and the ob-
served parcels is made by visualizing the parcel congura-
tions and the aforementioned statistics, either in color-coded
maps of the parcels, or in superposed, non-normalized fre-
quency histograms. We also compare our results to those
obtained by [WMWG09] for the same set of input blocks
(Figure 12 in Appendix C).
We have applied our subdivision method to historic cities
with chaotic street and subdivision patterns (Figure 13 in
Appendix D), and have achieved high visual similarity be-
tween real and procedural parcel congurations by manually
adjusting attributes such as the split irregularity (). How-
ever, because of the stochastic component in the layouts of
these cities, we limit our automatic evaluation to planned
cities with street patterns of dierent geometric complexity.
Figure 1 shows the results of subdividing a set of blocks
with OBB and skeleton styles. The observed blocks are lo-
cated in a mixed use suburban area and are mostly rectan-
gular with both straight and warped edges as a result of the
geometry of the surrounding roads (Figure 1a,b). The set of
blocks exhibits signicant variability in both the area, the as-
pect ratio, and the minimum width of the parcels. This vari-
ability is visualized in the color-coded map (Figure 1b), and
in the black-and-white map (Figure 1c) showing on top of
each block four descriptive statistics of the observed parcels
in that block. In the color-coded maps, the parcels with as-
pect ratios close to one are shaded red, and the shading be-
comes closer to green as the aspect ratio increases. A simi-
lar distribution of colors in the observed and the procedural
parcels indicates high similarity. The parcels generated by
our method are shown in Figure 1f.
At a small scale, the similarity between the parcels gen-
erated by our method and the parcels observed in the real
world is evidenced by close visual inspection of the par-
cel conguration inside individual blocks, of either nearly-
rectangular or warped and irregular shapes. At a larger scale,
the similarity can be seen in the close resemblance between
the spatial distribution of colors in (Figure 1f), and in the
good match between the frequency histograms (Figure 1d,e)
of aspect ratio and area for the observed parcels (red) and
for the procedurally generated parcels (blue). All generated
parcels have dimensions and aspect ratios that are adequate
for containing buildings (Figure 1g). Following [MWH

06],
we used a shape grammar to create procedural trees and
buildings on the generated parcels (Figures 1i and 14 in Ap-
pendix E).
Figure 10 shows a set of blocks subdivided using skeleton
subdivision with oset. The observed blocks are located in
a low-density residential suburban area and have highly ir-
regular shapes as a result of the warped roads with frequent
loops and cul-de-sacs (Figure 10a,b). The large inner regions
in some blocks are covered by golf holes. The similarity be-
tween the observed subdivision and the procedural parcels
can be seen in the color-coded maps (Figure 10c,f) and in
the superposed histograms (Figure 10d,e). In this example,
the area under the blue (procedural) histograms is greater
than the area under the red (observed) histograms, which in-
dicates that the number of generated procedural parcels is
greater than the number of observed parcels. The reason for
this mismatch between parcel counts is that the number of
parcels is not directly specied by the user, but rather results
from the chosen values for the subdivision attributes (in this
case, parcel area and oset depth). Notice that one observed
parcel close the center of the map is a statistical outlier in
terms of area that is not captured by the extracted descrip-
tive statistics, and is thus not reproduced by our method.
Figure 11 in Appendix B shows the results of skeleton
subdivision applied to several blocks with nearly-rectangular
and irregular contours (Figure 11a,b). In this result, the col-
ors in the map indicate the parcel areas. In both the observed
and procedural parcel congurations the parcels in the mid-
dle blocks are generally larger than those in the side blocks.
7. Conclusion
We have presented an interactive method for procedural gen-
eration of parcels inside city blocks. Our approach generates
spatial congurations of parcels with high similarity to those
observed in real-world cities, and supports the consistent lo-
cation of parcels relative to their containing blocks under live
editing operations.
Our method can be extended in several ways. One of them
c 2011 The Author(s)
c 2011 The Eurographics Association and Blackwell Publishing Ltd.
C. Vanegas, T. Kelly, B. Weber, J. Halatsch, D. Aliaga and P. Mller / Procedural Generation of Parcels in Urban Modeling
Aspect Ratio
Area (m
2
)
observed
procedural
Parcels & road network
Procedural Parcels Statistical Comparison
Observed Parcels
Visualization oI procedural parcel neighbor number
P
Aerial Image
Observed parcel
neighbor number
0
3
6
9
12
Detail
a b c
d
e
I
Figure 10: Skeleton Subdivision with Osets. Residential
blocks are subdivided in oset style. The similarity between
the observed parcels (a,b,c) and the procedural parcels is
demonstrated in the color-coded maps showing the number
of neighbors of each parcel (f), and in the histograms of
other geometric attributes (d,e).
is to use machine learning techniques to automatically cap-
ture the subdivision style of a block. A second line of work is
to integrate our block subdivision algorithms with computer
vision methods that extract blocks and parcels from ortho-
graphic and oblique-angle aerial imagery. The computed in-
formation could guide the parcel extraction process towards
solutions that are statistically more likely to occur.
8. Acknowledgements
This research is partially supported by the European project
V-City (ICT-231199-V-CITY) and by NSF IIS 0964302. We
sincerely thank the reviewers for their insightful feedback.
References
[AAAG95] Aichholzer O., Aurenhammer F., Alberts D., G art-
ner B.: A novel type of skeleton for polygons. J. UCS 1, 12
(1995), 752761. 4, 5
[AVB08] Aliaga D. G., Vanegas C. A., Benes B.: Interactive
example-based urban layout synthesis. ACM Transactions on
Graphics 27, 5 (2008), 110. 3
[Car03] Carmona M.: Public places, urban spaces: the dimen-
sions of urban design. Architectural Press, 2003. 2
[CEW

08] Chen G., Esch G., Wonka P., M uller P., Zhang E.:
Interactive procedural street modeling. ACM Transactions on
Graphics 27, 3 (2008), 110. 3
[cga] Cgal, Computational Geometry Algorithms Library.
http://www.cgal.org. 5
[Cit] Esri CityEngine. http://www.esri.com/software/
cityengine/index.html. 8
[Cur97] Curdes G.: Stadtstruktur und Stadtgestaltung. Kohlham-
mer, 1997. 2
[GPMG10] Galin E., Peytavie A., Mar echal N., Gu erin E.: Pro-
cedural generation of roads. Comput. Graph. Forum29, 2 (2010),
429438. 3
[HKS08] Halatsch J., Kunze A., Schmitt G.: Using shape gram-
mars for master planning. In Design Computing and Cognition
08, Gero J. S., Goel A. K., (Eds.). Springer Netherlands, 2008,
pp. 655673. 3
[LSWW11] Lipp M., Scherzer D., Wonka P., Wimmer M.: Inter-
active modeling of city layouts using layers of procedural con-
tent. Comput. Graph. Forum 30, 2 (2011), 345354. 3
[LWW08] Lipp M., Wonka P., Wimmer M.: Interactive visual edit-
ing of grammars for procedural architecture. ACM Transactions
on Graphics 27, 3 (2008), 110. 3
[Mar09] Marshall S.: Cities design and evolution. Urban design
and planning. Routledge, 2009. 3
[MLBD02] Meyer M., Lee H., Barr A., Desbrun M.: General-
ized barycentric coordinates on irregular polygons. Journal of
Graphics Tools 7 (2002), 1322. 8
[MWH

06] M uller P., Wonka P., Haegler S., Ulmer A.,

Van Gool L.: Procedural modeling of buildings. In ACM SIG-
GRAPH 2006 (2006), pp. 614623. 3, 9
[PCDS04] Panerai P., Castex J., Depaule J., Samuels I.: Urban
forms: death and life of the urban block. Architectural Press,
2004. 2
[PM01] Parish Y. I. H., M uller P.: Procedural modeling of cities.
In Proceedings of ACM SIGGRAPH (2001), pp. 301308. 2, 3,
4, 7
[PPC08] Parolek D., Parolek K., Crawford P.: Form-based
codes: a guide for planners, urban designers, municipalities, and
developers. J. Wiley & Sons, 2008. 2
[TKD10] Thadani D., Krier L., Duany A.: The language of towns
& cities: a visual dictionary. Rizzoli, 2010. 2
[VABW09a] Vanegas C. A., Aliaga D. G., Benes B., Waddell P.:
Visualization of simulated urban spaces: Inferring parameterized
generation of streets, parcels, and aerial imagery. IEEE Trans-
actions on Visualization and Computer Graphics 15, 4 (2009).
3
[VABW09b] Vanegas C. A., Aliaga D. G., Bene s B., Waddell
P. A.: Interactive design of urban spaces using geometrical and
behavioral modeling. ACM Trans. Graph. 28 (December 2009),
111:1111:10. 3, 7
[VAW

10] Vanegas C. A., Aliaga D. G., Wonka P., M uller P.,

Waddell P., Watson B.: Modeling the appearance and behavior
of urban spaces. Computer Graphics Forum 29, 1 (2010), 2542.
2
[WCP

11] Wickramasuriya R., Chisholm L. A., Puotinen M.,

Gill N., Klepeis P.: An automated land subdivision tool for ur-
ban and regional planning: Concepts, implementation and test-
ing. Environmental Modelling & Software, 0 (2011), . 3
[WD11] Walker D., Daniels T.: The Planners Guide to Commu-
nityViz: The Essential Tool for a New Generation of Planning.
Orton Family Foundation Books. American Planning Associa-
tion, 2011. 3
[WMWG09] Weber B., M uller P., Wonka P., Gross M.: Inter-
active geometric simulation of 4D cities. Computer Graphics
Forum (Proceedings of Eurographics) 28, 2 (2009). 3, 7, 9, 13
[WWSR03] Wonka P., Wimmer M., Sillion F., Ribarsky W.: In-
stant architecture. ACM Transactions on Graphics 22, 3 (2003),
669677. 3
c 2011 The Author(s)
c 2011 The Eurographics Association and Blackwell Publishing Ltd.