Symmetry for Architectural Design

Niloy J. Mitra Mark Pauly

Figure 1: Various architectural models analyzed and modified by exploiting information on shape symmetries and regular repetitive patterns.

Abstract data [Pauly et al. 2005], [Thrun and Wegbreit 2005], [Pauly et al.
2008] or images [Müller et al. 2007], [Liu et al. 2008].
Symmetry and regularity abound in architectural models, often as
a result of economical, manufacturing, functional, or aesthetic con- In this paper, we summarize our previous work on symmetry de-
siderations. We show how recent work on symmetry detection and tection [Mitra et al. 2006], symmetrization [Mitra et al. 2007], and
structure discovery can be utilized to analyze architectural designs structure discovery [Pauly et al. 2008] with special emphasis on po-
and real-world artifacts digitized using 3D scanning technology. tential applications in architectural design.
This allows reverse engineering of procedural models that facili-
tate effective exploration of the underlying design space and the 2 Symmetries and Regular Structures
synthesis of new models by modifying the parameters of the ex-
tracted structures and symmetries. We demonstrate the effective- Our approach is based on the techniques for finding symmetry in-
ness of such an approach on a number of example designs. formation and repetitive structures introduced in [Mitra et al. 2006]
and [Pauly et al. 2008], respectively. We briefly describe the central
Keywords: shape analysis, symmetry, repetitive patterns, struc- ideas of these methods, but refer to the papers for a more detailed
tural regularity, procedural modeling discussion. Symmetry and structural repetitiveness can be formal-
ized using the notion of invariance under transformations. We say
1 Introduction that two parts A, B ⊆ S of a 3D model S are symmetric, if there
exists a transformation T , e.g., a rotation, reflection, or translation,
Architectural designs commonly exhibit significant symmetries or such that B = T (A). In general, we consider the space of similar-
contain repetitive patterns. These types of structural regularity are ity transformations composed of uniform scaling, rotation, transla-
not accidental, but often the result of economical, manufacturing, tion, and possibly reflection. To find symmetry transformations of
functional, or aesthetic considerations. Whether by evolution or de- a given shape, we apply a sampling approach illustrated in Figure 2
sign, symmetry implies certain economies and efficiencies of struc- that has been proposed in [Mitra et al. 2006] and, independently,
ture that make it universally appealing. Symmetry also plays an in [Podolak et al. 2006].
important role in human visual perception and aesthetics. Arguably The surface of the model is sampled uniformly with average sample
much of the understanding of the world around us is based on the spacing h. The user parameter h determines the scale of the small-
perception and recognition of shared or repeated structures, and so est symmetric elements that we want to detect. For every sample
is our sense of beauty [Thompson 1992].
Symmetry is also fundamental in the laws of physics, hence opti-
mality conditions in terms of statics often lead to symmetric config-
urations. In addition, structural regularity in architectural models y φ
allows pre-fabrication and mass-production of repetitive elements
and can thus lead to significantly reduced production costs. φ
d
Recent work in 3D shape analysis has focused on detecting sym- local evidence for
symmetry plane
metries and regular structures in geometric models [Martinet et al. x d
2006], [Mitra et al. 2006], [Podolak et al. 2006], [Simari et al.
2006], [Mitra et al. 2007], [Li et al. 2008], [Pauly et al. 2008]. Figure 2: To detect symmetries in geometric models, we uniformly
These research efforts offer a wealth of tools that can be employed sample the boundary of the shape (left). Every pair of samples with
to improve the architectural design process. In particular, explicit compatible local surface geometry provides local evidence for a
knowledge of symmetry and geometric regularity can be exploited symmetry transformation (center). In this example we consider re-
to facilitate reverse-engineering of design rules for procedural mod- flections that are parameterized by an angle φ and the distance d
eling or symmetry-aware shape optimization. Symmetry informa- to the origin. Accumulating such evidence using a clustering ap-
tion can also be beneficial for shape reconstruction from scanned proach yields the dominant symmetries of the model (right).
 rithm robustly finds two regular translational grids. The figure also
illustrates how the detected symmetry information can be utilized
Trans
for model repair.
Rot
Rot + Trans Rot + Scale Figure 6 illustrates the difference between top-down symmetry de-
Scale
tection according to [Mitra et al. 2006] and bottom-up structure dis-
covery using the method of [Pauly et al. 2008]. The former extracts
mostly pairwise symmetries, such as the global reflective symme-
try or the rigid motions mapping the towers or chimneys onto each
other. The latter detects translational and rotational grids of win-
dows and other structural elements, but ignores the chimneys, since
Trans × Trans Rot × Trans Rot × Scale
their spatial arrangement does not match any of the repetitive pat-
terns defined in Figure 3. On the other hand, this method is capable
of discovering and compactly representing structures composed of
Figure 3: Schematic illustration of regular structures. The helix very small elements such as the balustrade, which are not extracted
and spiral are generated by transformations that combine rotation by the top-down symmetry detection approach.
with translation and scaling, respectively. The bottom row shows
the three types of commutative 2-parameter groups that can be de- Procedural Modeling. A simple yet effective modeling opera-
tected with our approach. tion is part replacement. Structural elements can be replaced or
modified using standard modeling tools. The system then auto-
matically replaces all symmetric copies to preserve the structural
point we compute a local signature that compactly encodes local integrity of the model. This type of operation is illustrated in Fig-
geometric properties at that point that are invariant under transfor- ures 6 and 8. In addition, we can modify the parameters of the
mations of the specific transformation space under consideration. regular structures, e.g., the number of repetitions as illustrated in
Sample points with similar signatures are paired and a canonical Figures 7 and 8. This type of procedural design allows the user to
transformation that maps one sample to the other is computed and quickly create variations of an original design or scanned artifact
refined using local registration methods. that would be tedious to achieve with traditional modeling tools.
The key observation is the following: If a shape contains sym-
metries or repetitive structures, then the estimated transformations Symmetrization. The extracted symmetries are often not perfect
exhibit specific accumulation patterns when mapped to a suitable in the sense that the transformed part T (A) might not exactly match
transformation space. These patterns can be extracted using cluster- the corresponding part B. This occurs, for example, when scanning
ing methods and grid fitting techniques. While the method of [Mi- a real-world object due to the discrete sampling process, or when
tra et al. 2006] is mostly concerned with pairwise symmetries, the the model itself is not perfectly symmetric, e.g., a partly preserved
structure discovery method of [Pauly et al. 2008] in addition anal- ruin. In addition, many physical architectural prototypes or design
yses the spatial relations among different symmetries. The under- studies are often not build with high geometric accuracy, so that a
lying formulation is based on theory of transformation groups and digitized model might not possess all the intended symmetries. To
thus allows a rigorous mathematical treatment of the concept of enhance approximate symmetries we can employ the symmetriza-
structural regularity. Different types of regular repetitive structures tion approach of [Mitra et al. 2007]. As illustrated in Figure 9, this
that can be detected by this method are shown in Figure 3. method can be used to generate symmetric meshes, which can be
important if the mesh represents structural elements such as struts
The result of this analysis in transformation space is a set of sym- or beams, e.g., in a steel-glass construction.
metries and repetitive patterns that encode important medium and
large scale structural information of the processed shape. Symme-
tries can often be represented in a hierarchy, while repetitive struc- 4 Conclusion and Future Work
tures are described by a representative element, i.e., a patch P ⊂ S,
a set T of generating transformations, and the number of repetitions We discussed how symmetry and structure discovery algorithms
in each dimension (see Figure 6, lower left). can be exploited for shape analysis and synthesis in the context
of architectural design. These tools provide a first step towards a
more comprehensive framework for procedural modeling based on
3 Shape Analysis and Design reverse-enginerring of shape design rules. The analysis of symme-
try and repetitive structures can also be utilized in the classification
The analysis of digital 3D models using the methodology described of buildings from different historical periods and potentially pro-
above provides us with a compact representation of the symmetries vide insights into the style of a specific architect or designer.
and repetitive structures of a shape. We first show some examples
and then discuss how this information can be utilized to provide The modeling operations of the above examples solely rely on geo-
effective tools for shape exploration and manipulation in the context metric information and thus do not take into account semantic infor-
of architectural design. mation that might be important to adequately represent the underly-
ing design intent. An important avenue for future research concerns
Figure 4 shows the dominant symmetries detected in a digital model the development of a framework that allows combining symmetry
of the Sydney opera. The underlying transformation space is the information with other functional or semantic characteristics of dig-
seven-dimensional space of similarity transformations whose ele- ital 3D designs.
ments are composed of uniform scaling, rotation, and translation.
Figure 5 shows an application of the structure discovery algorithm Acknowledgements
to raw scanner output. The point cloud has been acquired with a
single-viewpoint laser scanner, which leads to gradually varying We thank our collaborators Leonidas Guibas, Helmut Pottmann,
sample spacing due to perspective distortion. Despite the low sam- and Johannes Wallner for important contributions to this work.
pling density and holes in the data caused by occlusion, the algo-
Figure 4: Large-scale symmetries detected in a digital model of the Sydney opera. The extracted symmetries include reflections, as well as
general similarities that involve uniform scaling, rotation, and translation.

Figure 5: Structure discovery and model repair on a laser scan of a complex outdoor scene. The algorithm fully automatically discovers
two translational grids within the acquired point cloud. Standard surface reconstruction yields an incomplete and inconsistent triangulation
shown in the zooms on the left. The models on the right have been created by augmenting the point set using replicated samples from the
representative elements prior to reconstruction.

Figure 6: Conceptual differences between bottom-up structure discovery and top-down symmetry detection. The transparent bounding boxes
(lower right) show the extracted symmetry hierarchy that can be utilized for symmetry aware part replacement as shown in the top right. The
repetitive translational and rotational structures shown on the lower left support more fine-grain edits to individual structural elements.
Figure 7: Structure discovery and procedural modeling on a building facade. The regularity patterns of the model on the left have been
extracted automatically and can be modified by the user to alter the facade design as shown in the middle. For comparison, the image on the
right shows the original model scaled along the horizontal axis.

original symmetrized

Figure 8: The input model in the top left corner has been analyzed Figure 9: Symmetrization illustrated on an architectural design
to reveal three dominant repetitive structures, illustrated in the top study. The top row shows how one of the symmetric elements
right. The zooms show the corresponding structural elements. A evolves during the optimization. After processing, the six-fold ap-
new design has been created by modifying the number of repetitions proximate symmetry of the original model is perfect both in terms
and replacing the repetitive elements with new geometry. of geometry and meshing.

References PAULY, M., M ITRA , N. J., G IESEN , J., G ROSS , M., AND
G UIBAS , L. 2005. Example-based 3D scan completion. In
L I , M., L ANGBEIN , F. C., AND M ARTIN , R. R. 2008. Detect- SGP ’05: Proceedings of the third Eurographics Symposium on
ing approximate symmetries of discrete point subsets. Computer Geometry Processing, 23–32.
Aided Design 40, 1, 76–93. PAULY, M., M ITRA , N. J., WALLNER , J., P OTTMAN , H., AND
G UIBAS , L. J. 2008. Discovering structural regularity in 3D
L IU , Y., B ELKINA , T., H AYS , J. H., AND L UBLINERMAN , R. geometry. ACM Trans. Graph. 27, 3.
2008. Image de-fencing. In Proceedings of CVPR 2008.
P ODOLAK , J., S HILANE , P., G OLOVINSKIY, A., RUSINKIEWICZ ,
M ARTINET, A., S OLER , C., H OLZSCHUCH , N., AND S ILLION , S., AND F UNKHOUSER , T. 2006. A planar-reflective symmetry
F. 2006. Accurate detection of symmetries in 3d shapes. ACM transform for 3D shapes. ACM Trans. Graph. 25, 3, 549–559.
Trans. Graph. 25, 2, 439–464.
S IMARI , P., K ALOGERAKIS , E., AND S INGH , K. 2006. Folding
M ITRA , N. J., G UIBAS , L. J., AND PAULY, M. 2006. Partial and meshes: hierarchical mesh segmentation based on planar sym-
approximate symmetry detection for 3D geometry. ACM Trans. metry. In SGP ’06: Proceedings of the fourth Eurographics
Graph. 25, 3, 560–568. Symposium on Geometry Processing, 111–119.
M ITRA , N. J., G UIBAS , L. J., AND PAULY, M. 2007. Symmetriza- T HOMPSON , D. W. 1992. On Growth and Form. Dover.
tion. ACM Trans. Graphics 26, 3, # 63. T HRUN , S., AND W EGBREIT, B. 2005. Shape from symmetry. In
M ÜLLER , P., Z ENG , G., W ONKA , P., AND VAN G OOL , L. ICCV ’05: Proceedings of the Tenth IEEE International Confer-
2007. Image-based procedural modeling of facades. ACM Trans. ence on Computer Vision.
Graph. 26, 3, # 85.