LeNgocSang Thesis PDF

AUTOMATIC DESIGN OF
ORIGAMIC ARCHITECTURE
PAPER POP-UPS
LE NGOC SANG
NATIONAL UNIVERSITY OF SINGAPORE
2013
AUTOMATIC DESIGN OF
ORIGAMIC ARCHITECTURE
PAPER POP-UPS
LE NGOC SANG
B.Sc. (Hons.), NUS
A THESIS SUBMITTED FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
DEPARTMENT OF COMPUTER SCIENCE

NATIONAL UNIVERSITY OF SINGAPORE
2013
To my parents, my wife and my little daughter
Declaration
I hereby declare that this thesis is my original work and it has been
written by me in its entirety.
I have duly acknowledged all the sources of information which have
been used in the thesis.
This thesis has also not been submitted for any degree in any university
previously.
Le Ngoc Sang
30 April 2013
v
Acknowledgements
I am deeply grateful to Dr. Low Kok-Lim, my Ph.D. supervisor, for
his patient guidance, continued encouragement and constant support.
He always shares with me insightful advice, while giving me sufficient
room to grow. His wisdom, kindness, and wonderful personality will
always inspire me.
I would also like to thank Leow Su Jun, Le Nguyen Tuong Vu and
Conrado Ruiz Jr. for their prior work and consistent help in this research.
I really enjoy working with them.
Many thanks to Drs. Tan Tiow Seng, Cheng Ho-lun, Lee Yong Tsui
and Leong Hon Wai for their valuable feedback and suggestions on my
thesis. Thanks to the members of the graphics and geometry group for
their interesting discussions and sharings during our weekly meetings.
I can never forget my wonderful work environment in G3 lab and espe-
cially NUS, my second home.
I reserve the last words of thanks for my family, for their many years
of love, support and encouragement. I am lucky to have met and mar-
ried my wife, who has always been by my side during the last and tough-
est years of my Ph.D. program. This thesis is dedicated to my parents,
my wife and my lovely little daughter.
vii
Contents
Summary xiii
List of Figures xv
List of Tables xxiii
List of Publications xxv
1 Introduction 1
1.1 History of Paper Pop-up . . . . . . . . . . . . . . . . . . . . 1
1.2 Origamic Architecture . . . . . . . . . . . . . . . . . . . . . 3
1.3 Research Motivation . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Objectives and Contributions . . . . . . . . . . . . . . . . . 6
1.5 Thesis Overview . . . . . . . . . . . . . . . . . . . . . . . . 8
2 Literature Review 11
2.1 Paper Crafting . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.1 Paper Folding . . . . . . . . . . . . . . . . . . . . . . 12
2.1.2 Paper Cutting . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Paper Pop-Up and Origamic Architecture . . . . . . . . . . 14
2.2.1 General Paper Pop-Up . . . . . . . . . . . . . . . . . 14
2.2.2 Origamic Architecture . . . . . . . . . . . . . . . . . 16
2.2.3 Major Drawbacks in Previous OA Design Systems . 18
2.3 Model Simplification and Abstraction . . . . . . . . . . . . 20
2.4 Physical Strength Analysis . . . . . . . . . . . . . . . . . . . 21
2.4.1 General Structure Analysis . . . . . . . . . . . . . . 21
2.4.2 Thin Material Analysis . . . . . . . . . . . . . . . . . 23
3 Automatic Design of Origamic Architecture 27

3.1 Geometric Formulations . . . . . . . . . . . . . . . . . . . . 29
3.1.1 OA Components . . . . . . . . . . . . . . . . . . . . 29
3.1.2 Foldability of OA . . . . . . . . . . . . . . . . . . . . 31
3.1.3 Stability of OA . . . . . . . . . . . . . . . . . . . . . 33
ix
3.1.4 Comparison between the existing stability condi-
tions and our conditions . . . . . . . . . . . . . . . . 42
3.2 OA Design Algorithm . . . . . . . . . . . . . . . . . . . . . 43
3.2.1 Surface Segmentation . . . . . . . . . . . . . . . . . 45
3.2.2 Generation of Representative Patches . . . . . . . . 46
3.2.2.1 Slicing Orientation . . . . . . . . . . . . . . 47
3.2.2.2 Slice Positioning . . . . . . . . . . . . . . . 47
3.2.2.3 Surface Contour Projection . . . . . . . . . 50
3.2.3 Foldable OA Construction . . . . . . . . . . . . . . . 50
3.2.3.1 Pairwise Patch Connection . . . . . . . . . 51
3.2.3.2 Global Patch Connection . . . . . . . . . . 52
3.2.4 OA Stabilization . . . . . . . . . . . . . . . . . . . . 53
3.2.4.1 Path Sorting . . . . . . . . . . . . . . . . . 54
3.2.4.2 Path Stabilization . . . . . . . . . . . . . . 55
3.2.4.3 Path Update . . . . . . . . . . . . . . . . . 57
3.2.5 Comparisons between our stabilization algorithm
and Li et al. [68]’s . . . . . . . . . . . . . . . . . . . . 57
3.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.3.1 Visual Comparisons . . . . . . . . . . . . . . . . . . 60
3.3.2 Numerical Comparisons . . . . . . . . . . . . . . . . 63
3.4 Discussion and Conclusion . . . . . . . . . . . . . . . . . . 66
4 Origamic Architecture with Non-parallel Folds 69

4.1 Geometric Formulations . . . . . . . . . . . . . . . . . . . . 71
4.1.1 General OA Components . . . . . . . . . . . . . . . 71
4.1.2 Foldability of General OA . . . . . . . . . . . . . . . 73
4.1.2.1 Foldability of a Path of v-Structures . . . . 73
4.1.2.2 Foldability of a Combination of Structures 76
4.1.3 Stability of General OA . . . . . . . . . . . . . . . . 78
4.2 General OA Design Algorithm . . . . . . . . . . . . . . . . 85
4.2.1 Potential Surface Segments for V -Structure Gener-
ation . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.2.2 Generation of Foldable V -Structures . . . . . . . . . 88
4.2.3 Stabilization of V -Structures . . . . . . . . . . . . . 90
4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5 Strengthening Origamic Architecture Pop-Ups 97

5.1 Formulations . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.1.1 Governing Equation of Plate Bending . . . . . . . . 99
5.1.2 FDM-based Numerical Solution . . . . . . . . . . . 101
5.1.3 Boundary Conditions . . . . . . . . . . . . . . . . . 103
x
5.2 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . 112
5.2.1 Bending Approximation for Paper Structures . . . . 112
5.2.2 Weak Patch Detection and Correction . . . . . . . . 113
5.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.3.1 Comparisons with Analytical Solutions . . . . . . . 114
5.3.2 OA Structural Strength Analysis and Correction . . 116
6 Conclusion and Future Directions 121

6.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
6.2 Future Directions . . . . . . . . . . . . . . . . . . . . . . . . 123
6.2.1 Artistic Design . . . . . . . . . . . . . . . . . . . . . 123
6.2.2 Functional Design . . . . . . . . . . . . . . . . . . . 125
Bibliography 129
xi
Summary
Origamic architecture (OA) is a paper art form that involves cutting
and folding a single piece of paper to resemble three-dimensional ob-
jects. Because of the geometric and physical constraints, OA design is
labor-intensive and requires considerable skills. While similar to pop-up
books, OA is created with no gluing, which puts additional constraints
to the design process.
A number of computer tools have been developed to assist the OA

design process. However, in these tools, the user still needs to manu-
ally determine where and how the cuts and folds should be positioned.
Automatic design of OA has not been well-studied.
In this thesis, we present an algorithm for automatic design of an OA

that closely depicts an input 3D model. Our algorithm is grounded on a
general set of geometric conditions to ensure the stability and foldability
of the pop-ups. The generality of the conditions allows our algorithm
to generate valid structures that were excluded by previous algorithms.
Moreover, our method uses an image-domain approach that allows us
to capture important shapes using image segmentation. Our algorithm
is significantly better than the existing methods in the preservation of
contours, surfaces and volume. The designs have also been shown to
resemble those created by real artists.
In addition, we propose a simple yet effective approach for analyz-

ing the physical strength of OA structures. Our physical formulation is
based on Kirchhoff-Love theory of plate and Finite Difference Method.
It allows our system to automatically detect and correct physically weak
paper structures in real time.
By combining both aspects, we guarantee that our final OA designs

are both geometrically valid and phyiscally strong.
xiii
List of Figures
1.1 (a) A volvelle for astronomical illustration. (b) A flap demon-

strating a baby delivery. (c) A Bookano book. (d) Pop-up
from the Alice’s Adventures in Wonderland. . . . . . . . . 2
1.2 An origamic architecture (left) and a 360◦ pop-up (right). . 3
1.3 (a) A computer-aided tool for designing OA by Mitani and

Suzuki [77]. (b) OA of a spherical model obtained from the
voxel-based tool by Li et al. [68]. . . . . . . . . . . . . . . . 5
1.4 (a) Input 3D models. (b) The actual OA pop-ups cut and
fold from our designs. . . . . . . . . . . . . . . . . . . . . . 8
2.1 (a) Automatic origami design by tucking folded molecules

[97]. (b) Curved folding of a single planar sheet of material
without tearing or cutting [57]. . . . . . . . . . . . . . . . . 12
2.2 (a) Strip modeling for making papercraft toys from 3D

models [78]. (b) Chinese paper-cutting to illustrate a por-
trait of Abraham Lincoln [70]. . . . . . . . . . . . . . . . . . 14
2.3 (a) Simulation of the opening and closing of pop-up cards

by Iizuka et al. [49]. (b) A v-style pop-up generated by Li
et al. [69]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4 (a) An invalid OA with floating part is rejected by [77, 98].

(b) A fully-connected OA is accepted by [77, 98] but is un-
stable in practice. . . . . . . . . . . . . . . . . . . . . . . . . 16
2.5 (a) A sample input model. (b) The OA created based on

the stability formulation and voxelization technique by Li
et al. [68]. (c) The OA created based on our novel formu-
lation and slicing technique. . . . . . . . . . . . . . . . . . . 17
xv
List of Figures
2.6 (a) A simple unstable OA is made stable by (b) Li et al.
[68]’s method and by (c) ours. . . . . . . . . . . . . . . . . . 18
2.7 (a) A spherical model and (d) a church model are abstracted
(b, e) using Li et al. [68] and (c, f) using our method. Note
how the crosses on the church model appear in [68] and in
our work. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.8 (a) A feature-sensitive segmentation that can be used for

shape abstraction [59]. (b) Notable results from the slice-
based shape abstraction method [75]. . . . . . . . . . . . . . 21
2.9 Physically weak structures are detected and modified for

improved strength [106]. . . . . . . . . . . . . . . . . . . . . 22
2.10 The structural strength analysis in [94]. A weak model of a

bird is strengthened by thickening its neck (blue). Another
weak model is hollowed to reduce its load (gray). . . . . . 23
2.11 (a) Simulation of hanging cloth using the popular mass-

spring mechanism. (b) A failed attempt to produce paper-
like appearance by re-structuring the springs and adjust-
ing their coefficients. (c) Modeling of paper bending using
developable surface [12]. . . . . . . . . . . . . . . . . . . . . 24
3.1 Our automatic OA design pipeline. Note that the whole

process can be done in a 2D image space. In the last row
of this figure, we show the 3D popped-up models corre-
sponding to the 2D design layouts for illustration. . . . . . 28
3.2 Top: an OA plan with cuts (solid lines) and folds (dashed
lines) (a), which pops up at an arbitrary angle (b). Bottom:
a non-foldable OA plan (c), which is stuck during folding
(d). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.3 Projecting vector y along w, which bisects y and z, results

in vector y = −z. . . . . . . . . . . . . . . . . . . . . . . . . 32
xvi
List of Figures
3.4 1-paths (a) and 2-paths (b) were shown to be stable in

[4, 68]. We find out that 3-paths (c) and 4-paths (d) are
also stable if they are supported by appropriate patches.
From (a) to (e), blue bars show the side view of a 1-path, 2-
path, 3-path and 4-path; black bars show the known stable
patches; red bars show the supporting patches that make
the 3- and 4-paths stable. (e) illustrates a 3D view of the
stable 4-path in (d). . . . . . . . . . . . . . . . . . . . . . . . 34
3.5 An OA containing two double connections (left) and its

side view (right). . . . . . . . . . . . . . . . . . . . . . . . . 36
3.6 (a) An OA with a monotonic path, where d0 < d2 < d4 <

d6 , where d2i is the distance from patch p2i to the back
patch p0 . (b) An OA with a near-monotonic path, where
d0 < d4 < d2 . The actual OA structures are shown together
with their input models. . . . . . . . . . . . . . . . . . . . . 37
3.7 Side views of a monotonic B-path (a) and a near-monotonic

B-path (b) connecting stable patches p0 and pn+1 . At each
opening angle, the parallel configuration (above) is shown
to be the only possible configuration. Non-parallel config-
urations (below) are shown to be invalid. . . . . . . . . . . 40
3.8 Two structures with the outer patches (marked as red crosses)
known to be stable. These structures can be proven to be
wholly stable by our Proposition 3.2. Existing conditions
in the previous studies cannot prove the stability of any
part in these structures. . . . . . . . . . . . . . . . . . . . . . 42
3.9 A house model (top left) is approximated using OA: (a)

its depth map D, (b) normal map N and (c) OA patches in
the 45◦ view. (d) The cross section of the patches along the
purple line in (c). Note that the x-axis is along the main
fold line. The 45◦ image plane intersects the xy-plane and
xz-plane at lines (t, 1, 0) and (t, 0, 1), repectively. . . . . . . 44
xvii
List of Figures
3.10 Slice positioning and contour projection: An input sur-

face (a) is represented by a set of vertical slices (b), which
are first positioned to satisfy the minimum patch width
threshold and constant gap-gradient ratio. The slicing po-
sitions are then optimized to minimize the contour dis-
continuity, while maintaining the gap-gradient ratio (c).
Finally, the contour of the original hole is projected onto
the corresponding patch (d). Bottom row shows the side
view for each step. . . . . . . . . . . . . . . . . . . . . . . . 48
3.11 Possible relative positions of pa and pb . The top row shows

connectable pairs of patches. Red segments represent the
parts on existing patches that need to be removed. Black
and dashed blue segments represent the remaining parts
and new connections to be added, respectively. The bot-
tom row shows non-connectable pairs. The rightmost col-
umn illustrates the connecting result for case (f) in the 45◦
view. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.12 A 3D model (a) is approximated by a set of representa-

tive patches (b), which then go through the connecting-
merging process to form a foldable OA (c). . . . . . . . . . 53
3.13 Two B-paths (black thick patches, above and below) are
made monotonic by merging forward (above) and back-
ward (below). The dotted arrows show the merging di-
rection. Other patches not belonging to these paths (red
thin patches) may also be involved in the merging process. 56
3.14 (a, d) Unstable OA structures are made stable by (b, e) Li

et al. [68]’s stabilization method and by (c, f) ours. Our
stable OA is significantly closer to the original OA than [68]. 58
3.15 OA pop-ups of the Rialto bridge (left) and Colosseum (right),

as designed by artists [34] (top row), our system (middle
row) and the voxel-based OA tool in [68] (bottom row). . . 59
3.16 OA designs of some experimented models created by our

method (blue) and by [68] (green): a floating box, a torus,
a trefoil knot and an elephant models. We choose resolu-
tion 64×64×64 for [68] because higher resolutions are not
feasible for cutting and folding. . . . . . . . . . . . . . . . 61
xviii
List of Figures
3.17 OA designs of some experimented models created by our

method (blue) and by [68] (green): a triangular block, a
quarter-sphere, a complete sphere, and a Taj Mahal mod-
els. We choose resolution 64×64×64 for [68] because higher
resolutions are not feasible for cutting and folding. . . . . 62
3.18 OA pop-ups of the Stanford bunny and a church model.

From top to bottom are our designs (blue), and [68]’s de-
signs at 64×64×64 (green) and 256×256×256 (red). . . . . 63
3.19 Our OA design for the U.S. Capirol Building (left) and
[68]’s design at 64×64×64 (right). The bottom row shows
close-ups of the staircase. . . . . . . . . . . . . . . . . . . . 64
3.20 Input model that is not well-aligned with the back and
floor bases may not be converted nicely into an OA pop-up. 68
4.1 Pop-up designs with non-parallel folds. . . . . . . . . . . . 69
4.2 A series of parallel structures can be replaced by a single

non-parallel one for easier cutting and folding. . . . . . . . 70
4.3 (a) The components in a v-structure. (b) A v-structure with

two 90◦ angles may rotate freely when opened at 180◦ . . . . 72
4.4 A foldable path of v-structures and its angles. . . . . . . . . 74
4.5 (a) A v-structure in its closed configuration illustrates the

relationship between its angles. (b) The points in a v-structure
when it is opened. . . . . . . . . . . . . . . . . . . . . . . . . 75
4.6 (a) OA plan of a parallel structure (top) and its coverage

when fully closed (bottom). (b) OA plan of a v-structure
(top) and its coverage when fully closed (bottom). . . . . . 76
4.7 Three doubly connected v-structures (p1 , p2 , p3 , p4 ), (p2 ,

p5 , p6 , p7 ) and (p5 , p8 , p9 , p10 ), as described in Definition
4.2. For instance, in the first double connection, we have
δ34/1 = δ34/2 and δ12/3 = δ12/4 . . . . . . . . . . . . . . . . . . . 79
xix
List of Figures
4.8 (a) A path of n v-structures can be considered a combina-

tion of n single v-structures based on the floor patch. (b)
Extra patches can be added to a single v-structure to form
a foldable, doubly-connected v-structure. . . . . . . . . . . 80
4.9 A simulated OA containing a path of 3 doubly connected

v-structures. The closing motion of the OA is captured
from top to bottom and left to right. . . . . . . . . . . . . . 82
4.10 A real OA paper pop-up containing a path of 4 doubly

connected v-structures. . . . . . . . . . . . . . . . . . . . . . 83
4.11 The points on a line in a non-parallel OA plan (top) will

move along non-coplanar and non-parallel circles during
the opening and closing process (bottom). . . . . . . . . . . 84
4.12 (a) A model with a parallel block basing on a non-parallel

block is not a good candidate for v-structure generation,
because the resulting OA will not be foldable. (b) A model
with a non-parallel block basing on another one can still
be converted into a foldable OA containing two foldable
v-structures. . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.13 (Left) A selected path of segments for v-structure genera-

tion. (Right) The generated path of v-structures. The an-
gles along the v-path {ωα , ωβ , αi , βi } are computed based
on the angles along the segment path {ωα0 , ωβ0 , αi0 , βi0 , ...}. 89
4.14 Overlapping v-paths can be divided into shorter, separate

paths, and the angles along each path can be computed
independently of those on other paths. . . . . . . . . . . . . 90
4.15 A doubly-connected v-structure. . . . . . . . . . . . . . . . 91
4.16 An arbitrary series of triangular blocks and its correspond-

ing OA. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.17 A series of triangular blocks heading in different direc-

tions and its corresponding OA. . . . . . . . . . . . . . . . . 92
4.18 Non-parallel OAs can be used nicely to illustrate input

models that are not aligned to the back and floor bases. . . 93
xx
List of Figures
4.19 User-defined triangular blocks with arbitrary angles can

be easily converted into a fully foldable OA. . . . . . . . . 93
4.20 Non-parallel OA may not always be better than parallel

OA for preserving the visual appearance of slanted surfaces. 94
5.1 Bendings in OA structures may occur due to many rea-

sons: (a) Gravity on a long part. (b) External forces during
folding and storing. (c) Not being well-supported. (d) Big
size. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.2 The setup of points for computing of the plate deflection

at (m, n). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5.3 The mesh of grids is set up to completely cover the shape

of the patch. . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
5.4 The computed bending of a patch with two fixed edges

and two free edges. Left: Analytical solution of the gov-
erning differential equation. Right: Our solution of the
FDM-based governing equation. . . . . . . . . . . . . . . . 114
5.5 The computed bending of a patch with one fixed edges

and three free edges. Left: Analytical solution of the gov-
erning differential equation. Right: Our solution of the
FDM-based governing equation. . . . . . . . . . . . . . . . 114
5.6 The weak structures in Fig. 5.1 after being corrected. . . . . 115
5.7 A patch in the bunny OA is not well-supported (a), and is

corrected by extending (b). . . . . . . . . . . . . . . . . . . . 116
5.8 The trunk of the elephant OA bends down due to the heavy
weight (a), and is corrected (b). . . . . . . . . . . . . . . . . 117
5.9 The cross in the Capitol OA is strengthened. . . . . . . . . 117
5.10 The cross in the Taj Mahal OA is strengthened. . . . . . . . 118
5.11 A long thin patch made from a flexible material is bent

due to the weight acting along its longitudial axis. . . . . . 119
xxi
List of Figures
6.1 The multi-piece paper pop-ups produced by our automatic

design systems presented in [88] and [89]. . . . . . . . . . . 124
6.2 (a) A paper sculpture designed by [22]. (b) Foldable and

compact furniture can be designed automatically and 3D-
printed in the future. . . . . . . . . . . . . . . . . . . . . . . 126
xxii
List of Tables
3.1 Percentage differences of our OA structures and [68]’s at

various resolutions, as compared to the original input mod-
els. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.2 Comparison of number of cuts and folds between our

OA designs and those created by [68]. . . . . . . . . . . . . 65
3.3 Time (in minutes) to cut and fold the actual OA pop-up
from our design layout. . . . . . . . . . . . . . . . . . . . . 66
xxiii
List of Publications
[1] S. N. Le, S. J. Leow, T.-V. Le, and K.-L. Low. Surface and contour-
preserving origamic architecture paper pop-ups. IEEE Transactions on
Visualization and Computer Graphics, 20(2): 276–288, February 2014.
This publication contains the work on origamic architecture paper

pop-up in Chapter 3.
[2] T.-V. Le, K.-L. Low, C. Ruiz Jr., and S. N. Le. Automatic paper slice-
form design from 3d solid models. IEEE Transactions on Visualization and
Computer Graphics, 19(11): 1795–1807, November 2013.
This publication partially contains a similar formulation to the ge-

ometric stability of paper pop-up in Chapter 3.
[3] C. J. Ruiz, S. N. Le, and K.-L. Low. Generating multi-style paper pop-
up designs using 3d primitive fitting. In ACM SIGGRAPH Asia Technical
Briefs, Hong Kong, 2013.
This publication partially contains the geometric formulation for

the foldability of v-structure paper pop-up in Chapter 4.
xxv
[4] C. J. Ruiz, S. N. Le, J. Yu, and K.-L. Low. Multi-style paper pop-up
designs from 3d models. Computer Graphics Forum, 3(2):, 2014.
This publication partially contains the geometric formulation for

the foldability of multi-structure paper pop-up in Chapter 4.
xxvi
Chapter 1
Introduction
1.1 History of Paper Pop-up
Pop-up books and greeting cards have been fascinating people of all ages
with their beautiful and playful mechanics since their mass production
in the 1970s. But the existence of pop-ups dates back hundreds of years.
In their early days, they were also used greatly for scientific illustrations.
Montanaro [79] and Rubin [87] presented a thorough history of pop-up
and movable books since the 13th century.
The earliest known examples of “movable books” are Matthew Paris’

Chronica Majora (c. 1250) and Ramón Llull’s Ars Magna (1305), which
used volvelles (turnable disks) to represent calendars and used flaps to il-
lustrate maps (Fig. 1.1 (a)). A more recent scientific example was George
Spratt’s Obstetric Tables (1850), which used flaps to demonstrate the pro-
cedures for delivering babies (Fig. 1.1 (b)). In the same year, Dean &
Sons, one of the best-known publishers through the turn of the 19th cen-
tury, published the Little Red Riding Hood (1850). It was the first known
movable book in which a flat page rose into a 3D scene, though it was
actuated by pulling a string.
1
Chapter 1 Introduction
(a) (b)
(c) (d)
F IGURE 1.1: (a) A volvelle for astronomical illustration. (b) A flap demonstrat-
ing a baby delivery. (c) A Bookano book. (d) Pop-up from the Alice’s Adven-
tures in Wonderland.
One of the first self-erecting pop-ups, in which the rise into 3D was
actuated by opening only the page, was a card promoting the Trinity
Buildings in New York City (c. 1908), and S. Louis Girand’s Bookano
Books (c. 1930s) (Fig. 1.1 (c)). Modern pop-ups have taken these prin-
ciples to new heights, often employing linkage-like mechanisms to form
elaborate 3D shapes and motions. Some useful guides for designing pop-
ups are [8, 10, 16, 45, 51]. In recent years, pop-up books have risen to an
attractive art form with such notable books as Caroll and Sabuda’s Al-
ice’s Adventures in Wonderland [14] (Fig. 1.1 (d)), Carter’s series of dot
and spot books [15], Pelhem’s poetic pop-up book [85], and Bataille’s
ABC3D [9].
2
1.2 Origamic Architecture
Origamic Architecture (OA) is a striking form of paper pop-up. It was

developed by Masahiro Chatani in the early 1980s to reproduce archi-
tecture and other geometric structures [17, 18]. He combined techniques
of kirigami (Japanese papercutting) and origami (Japanese paper folding)
with his experience in architectural design. Though origamic architec-
ture and pop-up are often used interchangeably, they can be easily dis-
tinguished under strict definitions. OA is made by cutting and folding
only a single piece of paper, without the need for gluing. On the other
hand, most pop-ups involve two or more pieces of paper. OA are com-
monly seen as greeting cards that open into the desired structure at 90◦
(Fig. 1.2 (a)). Other types of pop-up may open to 180◦ (as seen in most
pop-up books such as Fig. 1.1 (c,d)), and 360◦ (as seen in some cards such
as in Fig. 1.2 (b)). A thorough bibliography of origamic architecture and
pop-up books is available at [2].
(a) (b)
F IGURE 1.2: An origamic architecture (left) and a 360◦ pop-up (right).
Even with the constraint of a single piece of paper, artists have been
able to create appealing and intricate OA. It is nowadays an indispens-
able part of the greeting card industry, which generates more than $7
billion annually in the United States alone [40].
3
Not only that, OA and paper pop-up also have practical scientific
applications. In manufacturing and fabrication, notably at micro and
nano levels, complex structures can be produced by cutting and folding
two-dimensional sheets into three-dimensional structures. Such process
is inspired by the techniques in origami and paper pop-up, especially
origamic architecture [30, 47, 107]. As shown in these studies, this fabri-
cation technique by cutting and folding is potentially easier, less expen-
sive, and more portable to transport than the more traditional machining
methods, including 3-D printing.
1.3 Research Motivation
Similar to other popular paper art forms, origamic architecture has mainly
been designed manually. However, unlike origami, algorithmic studies
in OA have been scarce.
In general, OA and other types of pop-up share a common property

that they have to be both foldable and stable. Being foldable means the
structure must not only pop up when opened, but also fold completely
flat when closed. On the other hand, stability of the pop-up only allows
the user to hold the two outer halves of the paper layout to open and
close it. A number of studies have been done on the folding mechanism
of paper pop-up [36, 37, 67]. However, these studies did not investigate
its stability conditions. In addition, they did not limit the number of
pieces of paper in the design process. As OA is made from a single piece
of paper without gluing or splicing, additional constraints will need to
be considered.
A number of computer-aided tools have been developed to pro-

vide virtual design environments for OA [44, 49, 77]. However, in these
4
systems, the user is ultimately responsible for positioning the patches

on the paper layout (Fig. 1.3 (a)). This trial-and-error process remains
labor-intensive, time-consuming and highly skill-demanding. Therefore,
a system that can automatically design an OA for novice users is desir-
able.
In a recent study, Li et al. [68] proposed a fully automatic method

for designing OA from input 3D model using voxel grid. However, their
method is based on a narrow set of geometric conditions for valid OA,
which greatly limit the types of structures that can be preserved. In ad-
dition, due to the nature of voxelization technique, they require a dense
grid to approximate the surface curvatures and contours, which may
make their designs difficult to cut and fold in practice (Fig. 1.3 (b)).
In general, despite the popularity and usefulness of origamic archi-

tecture, research in the abovementioned aspects is still very limited. It
motivates us to develop a more general and comprehensive framework
for automatic design of OA.
(a) (b)
F IGURE 1.3: (a) A computer-aided tool for designing OA by Mitani and Suzuki
[77]. (b) OA of a spherical model obtained from the voxel-based tool by Li et al.
[68].
5
1.4 Objectives and Contributions
Our main goal in this research project is to develop a framework for auto-
matically designing origamic architecture from an input 3D model. The
resulting OA pop-up needs to achieve a few important objectives.
First, it needs to be geometrically valid. From artists’ creations [34]

and by definitions in the previous studies [4, 68], a valid pop-up needs to
be both foldable and stable, as described in Section 1.3. The former means
if we hold and turn the two backing sheets of a pop-up, we should be
able to open it flat and close it completely. The latter means if we hold
the two backing sheets stationary, all other parts of the pop-up must not
be movable.
Second, we want to make the designed OA a good resemblance of the

input model, not only in terms of volume preservation, but also as an artistic
abstraction. In [68], although the volume of the input model is preserved
quite well using voxels, the resulting OA does not look similar to com-
mon artists’ creations, which use both smooth and sharp contours effec-
tively to depict a desired shape. We aim to find an abstraction method to
preserve the surfaces and contours of the input model nicely.
Third, all parts of an OA should be physically strong. A weak part

may collapse or bend significantly under gravity or external forces. We
aim to develop a method to detect and fix such weak parts in an OA. To
the best of our knowledge, our study is the first one that examines the
structural strength of a paper pop-up.
As mentioned earlier in Section 1.3, the most notable methods for

OA design [68, 77, 98] only achieve part of the first and second objec-
tives, and does not consider the third objective. We will investigate the
drawbacks of these methods in more details in Chapter 2.
6
We make important contributions to the domain of paper pop-up

research by presenting: (1) a comprehensive set of conditions for geo-
metrically valid OA, (2) A novel technique for checking and fixing the
stability of an arbitrary OA, (3) an image-based algorithm for design-
ing an OA that abstract an input 3D model nicely, and (4) an effective
method for detecting and fixing physically weak parts in an OA.
Our contributions can be elaborated further as follows.
1. We formulate a comprehensive set of geometric conditions for an

OA design to be folded flat in a stable manner. These conditions
cover a wide range of OA structures, including those created by
artists and previous computer tools, as well as those not considered
by these tools.
2. Based on the new formulation, we propose a novel technique for

checking and fixing the stability of an arbitrary OA. Our technique
preserves the shape of an OA significantly better than the existing
technique did. It can also be conveniently embedded in other OA
systems to improve their results.
3. We present an image-based method for abstracting an input 3D

model into an OA structure that preserves not only the input vol-
ume but also its smooth curvatures and subtle details. In addition,
the resulting OA can be cut and folded more easily than those ob-
tained from the existing automatic system.
4. We present an algorithm for analyzing and improving the struc-

tural strength of paper OA. Our algorithm is the first that analyzes
the structural strength of paper pop-up, and can be implemented
easily for incorporating into other OA systems.
7
(a) (b)
F IGURE 1.4: (a) Input 3D models. (b) The actual OA pop-ups cut and fold from
our designs.
1.5 Thesis Overview
First, to understand the difficulties and current status in this research

domain, it is necessary to study the existing related work and state-of-
the-art results (Chapter 2). Various types of paper craft are visited. We
also review some works in model simplification and abstraction, as OA
design can be considered an artistic way to abstract 3D models. Previous
studies on physical strength analysis for various types of structures are
also discussed.
8
In Chapter 3, we present our geometric conditions for the validity

of traditional parallel OA. Based on those formulations, we introduce
a novel approach for generating OA that is both foldable and stable. It
consists of three main stages: generating representative patches to depict
the input surface, making the patches foldable, and finally stable. We
compare our designs with artists’ creations and designs from the existing
OA tool, both visually and numerically.
Chapter 3 mainly deals with traditional parallel OA. In Chapter 4,

we present our geometric conditions for foldable and stable non-parallel
OA. We successfully formulate a stability condition that is analogous to
that of a parallel OA. Our non-parallel OA design is then introduced as
an extension that works smoothly with the original design pipeline for
parallel OA.
In Chapter 5, we present the physical strength analysis for OA

structures. Our analysis is based on Kirchhoff-Love theory of plate. By
using Finite Difference Method (FDM), we formulate a set of intuitive
numerical equations that can be solved in real time. Based on these
equations, we compute the possible bendings in the paper structures,
and correct the weak patches by adding new supporting connections.
In Chapter 6, we conclude our work and propose a number of pos-

sible future directions in our research.
9
Chapter 2
Literature Review
Paper craft is a collection of art forms employing paper or card to create

beautiful three-dimensional objects. Unlike other more complicated me-
dia, such as wood and ceramics, paper allows easy cutting and folding.
Many forms of paper craft share with origamic architecture these two
key techniques. In this chapter, we will first review some of those forms.
After that, we will analyze the notable results in algorithmic studies of
paper pop-up and origamic architecture.
The designing of OA can be considered a process of abstracting the

input 3D models. In order to keep the OA layout feasible in practice,
we need to generate only a reasonable number of cuts and folds. As the
input models are highly detailed, it is necessary to simplify them while
maintaing their meaningful visual appearance.
Another essential aspect we need to consider is the structural strength

of the final OA pop-up. Although existing studies of OA mainly concen-
trate on its geometric validity, it is also important to analyze whether the
structure is strong given the actual physical properties of paper.
2.1 Paper Crafting
Besides paper pop-up in general and origamic architecture in particular,
11
Chapter 2 Literature Review
other related types of paper craft have been studied mathematically. In

this section, we review two paper crafting techniques, paper folding and
paper cutting, as origamic architecture creation is a combination of these
techniques.
2.1.1 Paper Folding
Origami is probably the most well-known Japanese art of paper folding.

Traditional origami uses a single piece of paper without cutting or gluing
[48]. Folding algorithms for origami have been extensively studied in
computational geometry. Notable books on the geometric formulations
of origami include [26, 84]. Common mechanisms in origami have been
presented in numerous books [32, 56, 61, 80].
Some recent research works also aimed to find novel techniques

to create an origami. For reproducing an arbitrary input 3D model us-
ing origami, Tachi [97] proposed an algorithm to automatically convert a
polyhedral surface to molecules than can be folded and tucked together
(Fig. 2.1 (a)). Kilian et al. [57] considered curved folding as a new way to
reconstruct wavy surfaces from a single planar sheet of material without
tearing or cutting (Fig. 2.1 (b)).
(a) (b)
F IGURE 2.1: (a) Automatic origami design by tucking folded molecules [97]. (b)
Curved folding of a single planar sheet of material without tearing or cutting
[57].
12
Since the traditional origami does not allow cutting, reproduction

of an input 3D model is sometimes very challenging. Paper pop-up re-
lieves this constraint and may be more suitable for many beginners who
want to explore basic paper arts. On the other hand, paper pop-up re-
quires a different set of foldability and stability conditions than origami
does. For instance, the closing process of a pop-up is more restricted
than origami, as it requires all the pieces to flatten simultaneously and
with only external forces on the two outermost pieces. This requirement
is commonly seen in pop-up books. As a result, the theories related to
these two art forms cannot be easily ported to each other.
2.1.2 Paper Cutting
Kirigami, the Japanese art of cutting and fastening pieces of paper, has
also been the subject of a number of studies. Mitani and Suzuki [78],
Shatz et al. [93] and Massarwi et al. [72] proposed interactive methods
for approximating developable surfaces using strips or patches of pa-
per. With the use of cutting and splicing, strip and patch modeling can
achieve complex and even knotted geometry that is otherwise infeasi-
ble in other paper art forms. These interesting approaches are useful for
manufacturing paper-craft toys from 3D models (Fig. 2.2 (a)).
A similar art form is the Chinese paper-cutting, which is often used

for designing decorative patterns and figures. The main features and
contours of a desired image is normally cut out from a piece of paper
to create a strong visual effect (Fig. 2.2 (b)). A simple algorithm for
automatic paper-cutting from input images was proposed in Xu et al.
[108]. [70] also presented an interactive design for 3D paper-cutting and
animated paper-cut patterns.
13
(a) (b)
F IGURE 2.2: (a) Strip modeling for making papercraft toys from 3D models [78].
(b) Chinese paper-cutting to illustrate a portrait of Abraham Lincoln [70].
Although kirigami and the Chinese paper-cutting may illustrate

various interesting objects, most patterns tend to have fixed shapes af-
ter being created. For decorative purpose, some patterns may rise when
being pulled by a string. However, their mechanisms are totally different
from the opening and closing in pop-up and OA structures.
2.2 Paper Pop-Up and Origamic Architecture
2.2.1 General Paper Pop-Up
A few computer-aided environments for designing paper pop-up have

been presented. Based on a set of simple geometric rules, Glassner [36,
37] introduced a system in which the user can interactively perform dif-
ferent folding mechanisms. Hendrix and Eisenberg [44] designed an ap-
plication called “Popup Workshop” to introduce children to the crafting
and engineering disciplines of paper pop-up. However, these tools did
not completely guarantee the non-intersection between pieces of paper
during the folding process. Such issues had to be resolved manually by
the user.
14
Lee et al. [67] and Iizuka et al. [49] presented interactive systems
that simulated the opening and closing of pop-up cards, hence allowed
the user to detect collisions and protrusions between the pieces of pa-
per (Fig. 2.3 (a)). Hara and Sugihara [41] and Li et al. [69] also devel-
oped a geometric formulation to pre-determine whether a v-style pop-
up can be folded flat (Fig. 2.3 (b)). It is also a common class of pop-ups
consisting of paper patches falling into four parallel groups. Okamura
and Igarashi [83] developed an interface to assist the user in the design
and production of a pop-up card, by simulating the possible protrusions
and collisions. Recently, Abel et al. [4] proposed a polynomial-time algo-
rithm that creates pop-ups by subdividing a polygon into single degree-
of-freedom linkage structures.
(a) (b)
F IGURE 2.3: (a) Simulation of the opening and closing of pop-up cards by Iizuka
et al. [49]. (b) A v-style pop-up generated by Li et al. [69].
15
2.2.2 Origamic Architecture
Only a few studies particularly focus on origamic architecture, which

requires the structure to be created from only a single piece of paper.
Under this constraint, it is not easy to design a structure that can be both
folded flat and fully opened to a 2D layout. It is even more challenging
to resemble a given input model using a limited constant amount of pa-
per. The first notable computer-aided tool for OA design was presented
by Mitani and Suzuki [77]. Their pioneering approach ensures the de-
sign is fully-connected, and hence, no part is dangling or floating during
the folding process as in Fig. 2.4 (a). It was implemented in the Pop-
Up Card Designer by Tama Software [98]. This system allows the user
to position the vertical and horizontal faces in the desired pop-up, and
automatically detect whether the faces are connected. This is a neces-
sary condition for a design to pop up. Note that, in general, determining
whether a design can be fully opened or closed is an NP-hard problem
[103].
(a) (b)
F IGURE 2.4: (a) An invalid OA with floating part is rejected by [77, 98]. (b) A
fully-connected OA is accepted by [77, 98] but is unstable in practice.
16
All the mentioned approaches for paper pop-up and OA design re-
quire adequate knowledge and heavy user interaction to transform each
part of the input model into valid OA faces. This task can be challeng-
ing, even for simple shapes. In addition, most of them do not guarantee
that the design layout may pop up in a stable manner. In other words,
while the user is holding the two outer halves of the layout stationary,
other inner parts of the pop-up may still move or even collapse. Fig. 2.4
(b) illustrates an OA accepted by [77, 98], but is in fact unstable, both
theoretically and in practice. However, it may not be trivial for a user
to position the faces such that the OA is stable. Hence, it is desirable to
have a system that can design stable OA automatically.
The only available system for automatic OA design was developed

by Li et al. [68]. Their work provided a set of geometric conditions for
foldabilty and stability of OA. Based on these formulations, they used
3D voxels to approximate the input model. However, due to the highly
discrete nature of voxelization technique, a dense grid is needed to pre-
serve the smooth surfaces and small details of the input. As a result, their
designs can be difficult to cut and fold in practice. In addition, although
their algorithm guarantees that the OA pops up in a stable manner, it uti-
lizes a very narrow set of stability conditions, which significantly limit
the possible appearance of the generated structures (Fig. 2.5 (b)).
(a) (b) (c)

F IGURE 2.5: (a) A sample input model. (b) The OA created based on the stabil-
ity formulation and voxelization technique by Li et al. [68]. (c) The OA created
based on our novel formulation and slicing technique.
17
In the following section, we elaborate the drawbacks of the exist-

ing interactive and automatic methods according to the three OA design
objectives that were described in Section 1.4.
2.2.3 Major Drawbacks in Previous OA Design Systems
As mentioned in Section 1.4, there are three important objectives that an

OA design needs to satisfy: geometric foldability and stability, preserva-
tion of the input surfaces and contours, and sufficient physical strength.
The existing interactive and automatic OA design systems do not con-
sider the third objective, physical strength of the paper pop-up. Not only
that, their solutions for the first two objectives produce very limited re-
sults.
Geometric foldability and stability. This criteria indicates the va-

lidity of a pop-up. The existing methods ensures that the designed pop-
ups are fully foldable. However, as described in Section 2.2.2, the most
notable interactive system for OA design, which was developed by Mi-
tani and Suzuki [77][98], does not examine the stability of the OA, but
leave this task to the user. On the other hand, the latest automatic OA
system by Li et al. [68] proposes a simple set of geometric conditions
(a) (b) (c)

F IGURE 2.6: (a) A simple unstable OA is made stable by (b) Li et al. [68]’s
method and by (c) ours.
18
to guarantee that the resulting design is stable. However, these narrow

conditions require modifying a significant part of the existing unstable
OA to make it stable, affecting its boundary and appearance drastically
(Fig. 2.6 (b)). Our proposed geometric conditions preserve the original
shape of the OA notably better (Fig. 2.6 (c)).
Preservation of the input surfaces and contours. This criteria is

meant for retaining the meaningful appearance of a given input model.
When no input is available, Mitani and Suzuki [77][98] lets the user po-
sition horizontal and vertical faces to form a shape he desires. This ap-
proach does not allow the user to generate rounded faces easily. If he
wants to design curves, he needs to approximate them as many poly-
lines. In Li et al. [68], the approximation is done automatically by fitting
3D voxels to the surface of the given input 3D model. This approach, al-
though preserving the input volume quite accurately, still does not keep
the important curvatures on the input surface. In addition, for surfaces
with smoothly changing gradients, excessively many voxels are needed,
which make the design infeasible for actual cutting and folding (Fig. 2.7
(a) (b) (c)
(d) (e) (f)

F IGURE 2.7: (a) A spherical model and (d) a church model are abstracted (b, e)
using Li et al. [68] and (c, f) using our method. Note how the crosses on the
church model appear in [68] and in our work.
19
(b)). In order to reconstruct arbitrary contours better, Li et al. [68] also

smooth the cut lines along the voxels after generating them. However,
this postprocessing step sometimes makes the output worse, as it loses
the sharp details on the input model (Fig. 2.7 (e)). In our work, we
present a completely different method that preserves both rounded con-
tours and sharp details on the input surface nicely (Fig. 2.7 (c, f)).
2.3 Model Simplification and Abstraction
In order to preserve the meaningful appearance of an input model nicely,

it first needs to be transformed into a simpler shape or set of shapes.
Model simplification is a well-studied but still active research area [23,
33, 50, 74].
Many recent techniques not only simplified the model, but pre-
served the notable features of its shape as well. They are also known
as shape abstraction techniques. Eisemann et al. [27] developed a view-
dependent method for abstracting 3D models using 2D layers. Mehra
et al. [76] simplified and abstracted 3D models using a small set of char-
acteristic curves. The effectiveness of shape abstraction can be improved
by segmenting the surface into meaningful parts (Fig. 2.8 (a)). Lai et al.
[59, 60] used feature-sensitive remeshing to produce a hierarchy of meshes,
so that they can construct a hierarchical segmentation. Kalogerakis et al.
[54] proposed a data-driven approach to simultaneously segment and la-
bel the parts in a 3D model. A survey of segmentation and partitioning
techniques is available in [92].
Our approach is inspired by promising results in slice-based shape

abstraction [75], as shown in Fig. 2.8. Perception and vision research also
strongly emphasized the impact of contours created by slicing smooth
surfaces in interpreting 3D models [95, 102]. These contours have been
shown to capture important shape information of the objects [29, 58].
20
(a) (b)
F IGURE 2.8: (a) A feature-sensitive segmentation that can be used for shape ab-
straction [59]. (b) Notable results from the slice-based shape abstraction method
[75].
They may even appear more appealing than the original models, which
can be visually cluttered [76]. This is our motivation for using slicing as
the main foundation in our OA creation algorithm.
2.4 Physical Strength Analysis
In the previous OA studies, paper is considered a rigid material that has

no weight and does not bend, except along the fold lines. However, this
assumption does not hold in practice. Hence, in an OA design process,
it is important to examine whether the resulting structure is physically
strong, and how to strengthen the weak parts. In origami design, the
structures are often strong, as the pieces of paper overlap significantly.
This cannot be achieved in OA, which is cut and folded from a single
piece of paper. To the best of our knowledge, no study has been done on
the physical strength analysis of OA or general paper pop-up.
2.4.1 General Structure Analysis
In mechanical engineering, structural properties are a well-studied sub-

ject. Commercial CAD modeling systems, such as CATIA, provide visual
feedback from finite element analysis that indicates the current state of
stress in the structure. Nevertheless, it does not guide the user on how
21
to modify the designs for improved stability, and manual model adjust-
ment is still required. This software is suitable for simulation, but can be
ineffective for designers lacking intuition in mechanics.
In computer graphics, computational research on structural strength

is limited. Procedural methods have been coupled with physics-based
analysis for blocks and buildings, both in 2D [11, 35], and 3D contexts
[106]. In these approaches, structures are treated as sets of disjoined
blocks and the feasible equilibrium of those blocks is computed (Fig.
2.9). Although achieving remarkable results, their types of structures
and modeling primitives are very different from ours.
F IGURE 2.9: Physically weak structures are detected and modified for im-
proved strength [106].
Structural analysis has also been used for creating continuous mod-
els, such as plants [42, 52]. In general, a structurally strong object can be
obtained using methods based on shape optimization [6, 43], where the
geometry of a 3D model is modified to minimize certain physics-based
objective function. However, this approach is computationally very ex-
pensive for even little complex input models.
One of the first automatic methods that can detect structural issues
was proposed by Telea and Jalba [99], which was used for 3D printing.
The method detected thin and thick parts and used a set of basic geo-
metric rules to determine whether a given part can support its attached
22
parts. This work, however, did not address the correction of the struc-
tural issues. Stava et al. [94] extended it by giving user feasible choices
for improving the strength of the structures (Fig. 2.10). Suggested solu-
tions include thickening the thin parts and hollowing some inner parts
of the object to reduce its load. These techniques are not applicable to
paper structures.
F IGURE 2.10: The structural strength analysis in [94]. A weak model of a bird
is strengthened by thickening its neck (blue). Another weak model is hollowed
to reduce its load (gray).
2.4.2 Thin Material Analysis
Cloth is a commonly studied material in computer graphics. Most mod-

els used for cloth simulation are based on particles connected in a mesh.
One of the very first works on cloth modeling was presented by Weil
[105]. He used a geometric method, fitting catenary curves between
hanging points to recreate the look of cloth, but not its actual properties.
To bring a more realistic appearance, physics-based techniques have also
been developed. Feynman [31], Ng and Grimsdale [82] used equilibrium
equations that computed the position of a particle based on the gravity,
elastic forces and bending forces.
23
Provot [86] used similar equations but constrained the cloth parti-
cle using springs. The mass-spring mechanism has been used widely for
physics simulation in computer graphics (Fig. 2.11 (a)). However, find-
ing a suitable set of coefficients or models for the springs to resemble
paper material is not trivial and requires further study (Fig. 2.11 (b)). In
addition, these mass-spring coefficients are mainly for visualization and
may not be equivalent to the actual physical properties of real-life mate-
rials. Another notable early works was by Terzopoulos et al. [100], who
proposed using elastically deformable bodies to model cloth and other
flexible materials.
Research on simulation of cloth and other thin materials is still very

active nowadays [7, 19, 20, 38, 53, 55, 101, 104]. Nevertheless, actual
physics-based simulation of paper bending, especially in paper struc-
tures, is scarce. English and Bridson [28], Bo and Wang [12], and Kilian
et al. [57] modeled paper bending using developable surfaces (Fig. 2.11
(c)). A notable recent work on simulation of plastic deformation was
done by [81], who used Finite Element Method to align mesh edges with
folds and creases.
(a) (b) (c)
F IGURE 2.11: (a) Simulation of hanging cloth using the popular mass-spring
mechanism. (b) A failed attempt to produce paper-like appearance by re-
structuring the springs and adjusting their coefficients. (c) Modeling of paper
bending using developable surface [12].
24
Although very commonly used in physics-based simulation, Finite

Element Method (FEM) is computationally expensive and requires elabo-
rate implementation due to its usage of various physical discretizations
[62]. In our work, the bending of each piece of paper can be considered
vertical displacements of points on a horizontal plane. Under this setup,
we are able to efficiently utilize Finite Difference Method (FDM), a simpler
mathematical discretization method that is also used widely in mechan-
ical engineering. It only requires a simple set of linear equations; and
hence, is particularly suitable for the implementation of our lightweight
OA design system.
In our attempt to analyze the physical strength of paper structures,

we are inspired by the field of plate analysis in mechanical engineering
[96]. On one hand, a static plate can be treated as a two-dimensional
structure. On the other hand, it is weaker than thick structures and may
go under a wide range of deformations. As a result, it requires a differ-
ent set of governing equations. In normal construction, plate analysis is
used to examine the strength of floors, walls, slabs, gates and bridges,
to name a few. When applied on paper-like material, it is often used for
computing the bending stiffness of the material, based on the measured
deformations.
We apply FDM to approximate the possible bendings in our gener-

ated paper structures. Based on this computation, we are able to deter-
mine which parts of the structures are weak, then correct them by adding
supporting connections.
25
Chapter 3
Automatic Design of Origamic

Architecture
In this chapter, we present our approach for automatic design of geomet-

rically valid origamic architectures (OA). We first formulate the validity
conditions of OA, including both its foldability and its stability. Based on
our formulation, we propose an image-domain algorithm for generating
a foldable and stable OA from an input 3D model. Our OA is guaran-
teed to pop up fully when opened and fold flat when closed. In addition,
our novel geometric conditions allow us to make an arbitrary OA stable
without changing its shape much. Fig. 3.1 illustrates the following main
steps in our algorithm for automatic design of geometrically valid OA.
1. Depth and normal maps of the input 3D model are rendered in the
45◦ orthographic view.
2. By thresholding these maps, we divide the input surface into seg-

ments with distinct gradients.
3. Representative patches are generated to abstract each surface seg-

ment.
27
Chapter 3 Automatic Design of Origamic Architecture
4. The representative patches are connected to form a foldable origamic

architecture.
5. The OA is finally made stable using our novel geometric condi-

tions.
Although we consider physical properties of paper in our project,

in this chapter, we temporarily assume that paper has no mass and does
not bend, except along the fold lines.
In this chapter, we construct the traditional type of parallel OA, in

which all the patches are parallel to one of the two outermost bases. The
F IGURE 3.1: Our automatic OA design pipeline. Note that the whole process
can be done in a 2D image space. In the last row of this figure, we show the 3D
popped-up models corresponding to the 2D design layouts for illustration.
28
promising results in Section 3.3 demonstrate how our generated OA can

resemble a wide range of models. The work in this chapter has also been
presented in [65, 66].
3.1 Geometric Formulations
3.1.1 OA Components
An OA is created by cutting and folding a rectangular sheet of paper.

Fig. 3.2 (a, c) shows the basic elements of an OA. Essentially, the sheet
of paper is divided into various non-overlapping regions called patches,
whose boundaries are marked with cuts and straight fold lines. Among
these patches, the back patch pB and the floor patch pF are two special,
outer patches that share a central fold line and contain all the other patches
within their boundaries. These special patches serve as the outermost
bases of the OA and can be moved by the user to make other patches
pop up into the desired structure. In traditional parallel OA that we aim
to generate, all the fold lines are parallel to the central fold.
We position the OA in a right-handed coordinate system, in which

the x-axis is parallel to the central fold line, the y-axis is perpendicular
to the x-axis and parallel to the back patch, and the z-axis is perpendic-
ular to the x-axis and parallel to the floor patch. We also call the angle
between pF and pB the opening angle θ (Fig. 3.2). By turning the back and
floor patches from fully opened state (θ = 180◦ ), the OA can be made
“popped-up” and folded flat (θ = > 0◦ ). Note that the opening angle
cannot be exactly 0◦ , because the patches are not allowed to overlap.
As described above, in an OA plan, a fold line shared by two patches

allows them to be folded up or down, while maintaining their adjacency.
29
(a) (b)
(c) (d)
F IGURE 3.2: Top: an OA plan with cuts (solid lines) and folds (dashed lines) (a),
which pops up at an arbitrary angle (b). Bottom: a non-foldable OA plan (c),
which is stuck during folding (d).
To define an OA plan properly, we first define the adjacency and connect-

edness of OA patches as follows.
Definition 3.1. Two patches pa 6= pb are said to be adjacent if and only if they
share a common fold line.
Definition 3.2. Let P = {p0 =pa , p1 , . . . , pn−2 , pn−1 =pb } be an ordered set of n
distinct patches. We say that pa and pb are connected and P is an n-path from
pa to pb , if pi , pi+1 are adjacent for all 0 ≤ i < n − 1.
We now formally define an OA plan as
Definition 3.3. An OA plan is a set of patches where
30
1. All the patches are co-planar and form a rectangular domain with possible
holes.
2. They are non-intersecting, except at their boundaries.
3. For every patch p, there exists a path traversing from the back patch pB to
the floor patch pF and containing p.
The first two properties are conventional for general origamic ar-
chitecture plans, as used in [68, 77]. However, such general plan may
contain “floating” patches, which are not adjacent to any other patch, or
“dangling” patches, which are not connected to pB or pF . These patches
may move unwantedly due to their disconnection. Hence, property 3 is
defined to keep all the patches connected to both pB and pF .
3.1.2 Foldability of OA
An OA plan as defined above may still not be completely opened or

closed. Fig. 3.2 shows such an OA plan (c), which is stuck during its
folding process (d). We require an OA plan to be foldable from θ =
180◦ to θ = without affecting the shapes, pairwise adjacency and non-
intersection of the patches.
At each opening angle, a possible set of relative positions of the

patches, with respect to pF and pB , is called a configuration. When all the
patches are parallel to pF or pB , we have a parallel configuration. By con-
sidering this special configuration, we have the condition for a foldable
OA plan as follows.
Proposition 3.1. An OA plan is foldable if and only if this plan is the projec-
tion of a parallel configuration along vector w onto the xz-plane, where w is
perpendicular to the x-axis and bisecting the corresponding opening angle.
31
F IGURE 3.3: Projecting vector y along w, which bisects y and z, results in vector
y = −z.
Proof.
Sufficiency. Consider an OA plan that is the projection of a parallel con-

figuration C along vector w, where w is perpendicular to x and bisecting
the opening angle θ of C. We will show that this OA plan is foldable,
which means it can be opened and closed between θ and any arbitrary
angle θ0 ∈ (0◦ , 180◦ ].
We treat an OA configuration as an infinite set of points. At open-

ing angle θ, each point p has coordinate (i, j, k) in the (x, y, z) system of
that configuration. We can open or close the OA from θ to θ0 by rotating
y about x to the corresponding y0 . Each point p is then transformed to p0
but represented using the same coordinate (i, j, k) in the new system.
This transformation maintains the distance between every two points

on the same patch. Consider two such points p1 and p2. If they both lie
on a patch parallel to the back patch, then their z-coordinates are equal
p
and the distance between them is a constant (i1 − i2 )2 + (j1 − j2 )2 , be-
cause x and y are always perpendicular. Similarly, the distance between
every two points on a patch parallel to the floor patch will remain un-
changed under this folding transformation. As a result, we can open or
close the OA from configuration C at angle θ to any other angle while
maintaining the shapes and non-intersection of the patches.
Additionally, the 180◦ configuration obtained from the above open-

ing transformation is also identical to the given OA plan. In fact, as
shown in Fig. 3.3, projecting an OA configuration along vector w that
32
bisects y and z onto the floor is equivalent to rotating y-axis into a new
y0 -axis equal to −z, which in turn is equivalent to an opening transfor-
mation at 180◦ angle.
Necessity. The necessity of Proposition 3.1 can be proven more easily.

Since no patch is floating or dangling, as mentioned in Definition 3.3,
the opening of a parallel configuration from an arbitrary angle to 180◦ is
equivalent to projecting its patches along w onto the OA plan (Fig. 3.3).
An alternative proof for the sufficiency of Proposition 3.1 can be

seen in [68]. It guarantees that the representative patches created in Sec-
tion 3.2 can fold flat and pop up fully. In addition, the necessity of Propo-
sition 3.1 can be utilized to detect non-foldable OA structures easliy.
3.1.3 Stability of OA
In practice, an OA pop-up can be opened simply by turning the back and

floor patches without requiring any extra forces. If these two patches are
held stationary, other patches must not be movable. In other words, the
OA needs to be stable. In this section, we present a general stability con-
dition for parallel OA, which covers all the stable OA structures that we
have tested, both created by artists and by other computer-aided design
tools. First, we formulate the stability of OA components as
Definition 3.4. A fold line is said to be stable if it is not movable at each

opening angle of the OA.
A patch is said to be stable if at least two non-colinear fold lines on it are
stable.
An OA plan is said to be stable if all of its patches are stable.
33
Interestingly, previous studies only presented a narrow set of sta-

bility conditions for parallel OA. In particular, Li et al. [68] and Abel et al.
[4] showed that a patch is stable if it lies on a 1-path or 2-path connecting
two stable patches, as shown in Fig. 3.4 (a, b). These conditions limit the
possible arrangements of patches, and significantly affect the preserva-
tion of the input surface, as demonstrated later by the results in Section
3.3.
(a) (b) (c) (d) (e)
F IGURE 3.4: 1-paths (a) and 2-paths (b) were shown to be stable in [4, 68]. We
find out that 3-paths (c) and 4-paths (d) are also stable if they are supported by
appropriate patches. From (a) to (e), blue bars show the side view of a 1-path,
2-path, 3-path and 4-path; black bars show the known stable patches; red bars
show the supporting patches that make the 3- and 4-paths stable. (e) illustrates
a 3D view of the stable 4-path in (d).
In our first attempt to extend the previous stability conditions, we

came up with the conditions for the stability of 3-paths and 4-paths,
which are illustrated in Fig. 3.4 (c, d).
We later found out that many complex OA structures are geomet-

rically stable, but are not accepted by any conditions in Fig. 3.4. In the
following discussion, we will present a novel condition, which can check
for the stability of a general n-path.
To facilitate our formulation, let us first call the parallel configura-

tion at θ = 90◦ the orthogonal configuration, because every two adjacent
patches are orthogonal. It is important because the patches in this con-
figuration need to resemble the input model.
34
We now consider an OA plan initially constructed as the projection

of an orthogonal configuration in (0, −1, −1) direction onto the xz-plane.
This plan has been proven foldable by Proposition 3.1. We will explore
further conditions that make it stable.
Let P = {p0 = pB , p1 , . . . , pn , pn+1 = pF } be a path traversing from

the back patch to the floor patch. Along P, we mark p2k as even and p2k+1
as odd patches, where 0 ≤ k ≤ bn/2c. In a parallel configuration, even
and odd patches are respectively parallel to pB and pF . As a result, if
a patch is even (odd) on one path, it is also even (odd) on all the other
paths. Next, we define two special sets of patches.
Definition 3.5. B-set (F -set) is the set of patches that are always parallel to
pB (pF ) at any opening angle.
If a stable OA has a parallel configuration, then its even (odd) patches

are always parallel to the back (floor) patch, and thus, form the B-set
(F -set). Hence, to construct such stable OA, we aim to make all of its
patches belong to these two sets. For that purpose, we define a new type
of connection called double connection.
Definition 3.6. Two non-coplanar patches p1 and p2 are said to be doubly-

connected if there exist non-coplanar patches q1 , q2 such that pi and qj are
adjacent for all i, j ∈ {1, 2}.
Patches q1 and q2 are also doubly-connected and the structure {p1 , q1 , p2 , q2 }

is called a double connection.
Fig. 3.5 illustrates an OA containing two double connections and a

B-set. In general, for an arbitrary path, we can construct double connec-
tions along it to make all of its patches belong to B-set or F -set. This is
proven in Lemma 3.1 below.
35
Lemma 3.1.
If an n-path P satisfies
1. bn/2c pairs of even (odd) patches are doubly-connected, and
2. One of its even (odd) patches is in B-set (F -set),
then all the even (odd) patches in P belong to B-set (F -set).
Proof.
If path P contains n patches, its number of even patches and number of
odd patches are both at most bn/2c + 1.
On the other hand, if two even (odd) patches are doubly-connected,

then they both belong to B-set (F -set). Since there are bn/2c pairs of
doubly-connected even (odd) patches, and one of those patches is al-
ready in B-set (F -set), there are at least bn/2c+1 patches in B-set (F -set).
From the two observations above, it is clear that all the even (odd)
patches must belong to B-set (F -set).
F IGURE 3.5: An OA containing two double connections (left) and its side view
(right).
36
We call a path whose even (odd) patches all belong to B-set (F -

set) a B-path (F -path). It means all the even (odd) patches on a B-path
(F -path) are parallel to pB (pF ). We can make such path stationary by
making its odd (even) patches also parallel to pF (pB ). To do so, we utilize
monotonic paths and near-monotonic paths, which are defined as follows.
Definition 3.7. A B-path (F -path) is said to be monotonic if the Euclidean

distances from its even (odd) patches to pB (pF ) form a monotonic function.
Fig. 3.6 (a) illustrates an OA with a monotonic B-path, which can

be used to approximate staircase-like models efficiently.
(a) (b)
F IGURE 3.6: (a) An OA with a monotonic path, where d0 < d2 < d4 < d6 ,
where d2i is the distance from patch p2i to the back patch p0 . (b) An OA with a
near-monotonic path, where d0 < d4 < d2 . The actual OA structures are shown
together with their input models.
37
Definition 3.8. A B-path P = {p0 , p1 , . . . , p2k+1 } is said to be near-monotonic

if d0 , d2 , . . . , d2k , the Euclidean distances from its even patches to pB , satisfy
1. d0 < d2k < d2k−2 < . . . < d4 < d2 , or
2. d0 > d2k > d2k−2 > . . . > d4 > d2
An F -path P = {p0 , p1 , . . . , p2k+1 } is said to be near-monotonic if

d1 , d3 , . . . , d2k+1 , the Euclidean distances from its odd patches to pF , satisfy
1. d1 < d2k+1 < d2k−1 < . . . < d5 < d3 , or
2. d1 > d2k+1 > d2k−1 > . . . > d5 > d3
Fig. 3.6 (b) illustrates an OA with a near-monotonic B-path, which

can be used to preserve sunken surfaces.
We now present the condition for stable patches.
Proposition 3.2. If a path connects two stable patches and is monotonic or

near-monotonic, then all of its patches are stable.
Proof.
Here we prove Proposition 3.2 for B-paths. We can treat F -paths like-
wise. Fig. 3.7 illustrates the monotonic (a) and near-monotonic (b) paths
as seen from the side view. In this view, the patches are represented as
2D bars at cross sections parallel to the yz-plane.
Consider a B-path P = {p0 , p1 , . . . , pn , pn+1 } connecting the stable

patches p0 and pn+1 . Without loss of generality, we assume that the first
and the last even patches on P are p0 and pn . At each opening angle, there
exists a parallel configuration of P, because the OA plan is created from
the (0, −1, −1) projection of the orthogonal configuration. We will show
38
that this parallel configuration is also the only possible configuration,

which means all the patches on P are stable.
Let lk be the length of pk , and Ok be the fold between pk and pk+1 .

We also denote u (v) as the unit vector parallel to pn+1 (p0 ) and point-
ing away from On (O0 ). In a general configuration, let u2i+1 be the unit
vector parallel to p2i+1 , pointing in the positive direction of u. Note that,
because P is a B-path, p2i are always parallel to v.
We consider the two types of paths.
(a) If P is monotonic, we assume d0 < d2 < . . . < dn (Fig. 3.7 (a)).

From the parallel configuration, we have
bn/2c−1 bn/2c
X X
On = O0 + (l2i+1 u) − (l2i v)
i=0 i=1
In a general configuration, we also have
bn/2c−1 bn/2c
X X
On = O0 + (l2i+1 u2i+1 ) − (l2i v)
i=0 i=1
which leads to
bn/2c−1 bn/2c−1
X X
(l2i+1 u2i+1 ) = (l2i+1 u) (3.1)
i=0 i=0
The right-hand side of Eq. 3.1 represents a straight line segment

parallel to u and having length bn/2c−1
P
i=0 l2i+1 . As the left-hand side also
comprises the same set of length coefficients, the equality can occur if
and only if u2i+1 and u are parallel for all 0 ≤ i ≤ bn/2c − 1. In other
words, all the odd patches must be parallel to pn+1 .
39
(a) (b)
F IGURE 3.7: Side views of a monotonic B-path (a) and a near-monotonic B-
path (b) connecting stable patches p0 and pn+1 . At each opening angle, the
parallel configuration (above) is shown to be the only possible configuration.
Non-parallel configurations (below) are shown to be invalid.
(b) If P is near-monotonic, we assume d2 > d4 > . . . > dn−2 > dn >

d0 (Fig. 3.7 (b)). From the parallel configuration, we have
O1 = O0 + l1 u
bn/2c−1
X bn/2c
X (3.2)
= On + (l2i+1 u) + (l2i v)
i=1 i=1
In addition,
bn/2c−1
X
l1 = l0n + l2i+1 (3.3)
i=1
where l0n is the difference between O0 and On ’s coordinates along the

u-axis.
40
Eqs. (3.2) and (3.3) lead to
bn/2c
X
O0 = On + (l2i v) − l0n u (3.4)
i=1
In a general configuration, we also have
O1 = O0 + l1 u1
bn/2c−1
X bn/2c
X (3.5)
= On + (l2i+1 u2k+1 ) + (l2i v)
i=1 i=1
Substituting (3.3) and (3.4) into (3.5), we obtain
bn/2c−1 bn/2c−1
X X
l0n u1 + (l2i+1 u1 ) = l0n u + (l2i+1 u2k+1 )
i=1 i=1
Similar to Eq. 3.1, this equality occurs if and only if u is parallel to

u1 , and u2k+1 is parallel to u1 for all 1 ≤ i ≤ bn/2c − 1, which means all
the odd patches, including p1 , are parallel to pn+1 . Since the even patches
are also parallel to p0 , all the patches on P are stable.
Proposition 3.2 leads to an effective approach for making an OA

stable, as presented in Section 3.2.4. Unlike in [68], where the OA is
created in a restricted way to simultaneously guarantee foldability and
stability, we can take as input an arbitrary foldable OA and make it sta-
ble. As a result, our stabilization technique can be embedded into other
systems to allow independent improvements of other steps.
41
3.1.4 Comparison between the existing stability conditions

and our conditions
There exist a number of similar studies on the stability of OA, notably

[4] and [68]. Meanwhile, our stability conditions are a comprehensive
generalization of the existing ones. The common basic idea is to check
for the stability of an OA patch by considering a path that it belongs to.
However, our conditions accepts infinitely more types of paths than the
existing ones do.
Specifically, the earlier conditions can only check for the stability of
1- or 2-paths connecting two known stable patches. If a path is longer
than 2 patches, it needs to be divided into 1- or 2-paths that start and
end with stable patches. However, this is not always possible. As such,
the existing conditions can only examine and accept a very small set of
stable strutures. Many structures need siginificant modifications before
stability checking can be done.
In contrast, our conditions can check for the stability of n-paths,

with arbitrary large n. We observe that many stable structures accepted
F IGURE 3.8: Two structures with the outer patches (marked as red crosses)
known to be stable. These structures can be proven to be wholly stable by our
Proposition 3.2. Existing conditions in the previous studies cannot prove the
stability of any part in these structures.
42
by our conditions are also used in artists’ creations. Some of such struc-
tures are shown later in Section 3.3.1. Note that, while our conditions in
Proposition 3.2 accept staircase-like structures as shown in Fig. 3.6 on
page 18, they also accept many other complex types of structures (Fig.
3.8). The flexibility of our conditions significantly reduces the redundant
modifications required before stability checking. As a result, our novel
conditions provide notably more freedom for OA design.
3.2 OA Design Algorithm
In this section, we present our framework for automatic OA design. We

start with a 3D input model, represented as a triangular mesh and posi-
tioned by the user between the two orthogonal outer patches, which lie
on the xy- and xz- planes. The output of our system is a foldable and
stable OA plan in which the patches pop up into the desired structure at
90◦ opening angle. We achieve this OA plan by constructing the patches
in the orthogonal configuration, then projecting them in (0, −1, −1) di-
rection onto the xz-plane. Our method is carried out in four stages.
1. Segmenting the input surface: We divide the input surface into

non-overlapping, smooth segments. The surface segmentation serves
as a pre-processing step to facilitate the patch generation. More de-
tails are described in Section 3.2.1.
2. Generating representative patches: By slicing each segment in a

suitable orientation at computed positions, we generate a set of
parallel patches that represent the curvatures and details on each
segment nicely. More details are described in Section 3.2.2.
3. Constructing a foldable OA plan: We connect the representative

patches so that their projections along (0, −1, −1) onto the xz-plane
43
form an OA plan, and each patch is on a path from pB to pF . Ac-

cording to Proposition 3.1, the resulting OA plan is foldable. More
details are described in Section 3.2.3.
4. Stabilizing the OA plan: Utilizing Proposition 3.2, we check whether

the patches are stable. If they are not, we make them stable by cre-
ating double connections to form monotonic and near-monotonic
paths. More details are described in Section 3.2.4.
An overview of our method can be seen in Fig. 3.1 on page 28.
To facilitate the design process, we set up a 45◦ orthographic view,

which looks at the central fold line along the projection vector (0, −1, −1).
The image plane is placed perpendicular to this vector so that it intersects
with the xz- and xy-planes at (t, 0, 1) and (t, 1, 0) lines, respectively. We
define i and j as the two orthonormal axes of the 45◦ image plane, which
are parallel to (0, 1, −1) and (1, 0, 0), respectively. Our OA plan can be
designed completely in the 45◦ orthographic view (Fig. 3.9 (a,b,c)). How-
ever, the figures still show our process in 3D view and side view, such as
in Fig. 3.9 (d), for illustrative purpose.
(a) (b) (c) (d)
F IGURE 3.9: A house model (top left) is approximated using OA: (a) its depth
map D, (b) normal map N and (c) OA patches in the 45◦ view. (d) The cross
section of the patches along the purple line in (c). Note that the x-axis is along
the main fold line. The 45◦ image plane intersects the xy-plane and xz-plane at
lines (t, 1, 0) and (t, 0, 1), repectively.
44
In the 45◦ view, we render the normal map N and the depth map D
of the visible input surface enclosed by two orthogonal outer patches pF
and pB . The normal vector at each pixel in N is scaled to unit length. The
depth values in D are measured from the camera plane and vary from
√
0 (for points on the camera plane) to 2/2 (for points along the central
fold line). The details of the four stages are elaborated in the following
subsections.
3.2.1 Surface Segmentation
We observe that some parts on the input surface are represented better
using vertical patches, which are parallel to the xz-plane, while some
other are better with horizontal patches, which are parallel to the xy-
plane. From that observation, we divide the input surface into different
segments first, before generating the representative patches for each seg-
ment.
Our criterion is to divide the input surface into non-overlapping,

smooth segments such that each segment can be represented by a single
set of parallel patches. Although surface segmentation itself is challeng-
ing, we only perform it in the 45◦ view, and hence, the task becomes
image segmentation.
In other types of paper pop-up where multiple pieces of paper are

used, a single view may not be sufficient for the design process, and a
full mesh segmentation may be required [5, 92]. If the input model is
highly noisy, the simplification technique by Mehra et al. [76] may also
be used as a pre-processing step.
45
In our work, we allow over-segmentation, which means a surface

region is divided into multiple segments while it could have been con-
sidered one segment. In such case, because those segments have rela-
tively similar gradients, they will still be represented by patches in the
same orientation. With this flexibility, a simple segmentation by thresh-
olding is sufficient for our work.
We find the segments by flood-filling the depth map D and normal

map N. We determine whether a pixel p should remain in the current
segment by thresholding f (p) − q(p), where f (p) takes the depth value
and x-, y-, z-components of the normal vector at p in turn, and q(p) is
the quadratic approximation from the previously flooded pixel p0 .
1
q(p) = f (p0 ) + f 0 (p0 )(p − p0 ) + f 00 (p0 )(p − p0 )2
2
The derivatives of f are estimated as the average difference be-

tween the value from the current pixel and those from the neighbour-
ing pixels. We make pixel p belong to a new segment if f (p) − q(p)
exceeds the predefined thresholds in D or N. As described earlier, over-
segmentation may occur in our work due to noise. It is, however, tol-
erable because the slicing step in 3.2.2 will produce patches in the same
orientation if the segments have similar gradients.
3.2.2 Generation of Representative Patches
After the input surface is divided into distinct smooth segments, we gen-
erate a set of parallel representative patches to preserve both the overall
curvature and the detailed contours of each segment.
In order to achieve that goal, we first determine whether the patches
46
should be horizontal or vertical, then create them by “slicing” each seg-

ment using a number of parallel planes in the selected orientation. Each
patch is a continuous cross section, fully visible in the 45◦ view and
bounded by two consecutive slices in this view. The process of gener-
ating representative patches consists of three main steps: determining
the slicing orientation, positioning the slices, and projecting surface con-
tours on the slices.
3.2.2.1 Slicing Orientation
We observe that a segment can be represented better by vertical patches,

which are parallel to the xy-plane, if its gradient changes more signif-
icantly along the y-axis than along the z-axis. Otherwise, it should be
represented by horizontal patches parallel to the xz-plane.
From that observation, we determine the slicing orientation by com-

paring dz/di(p) and dy/di(p), the average z-gradient and y-gradient along
the i-axis of all pixels p in the current segment. The gradients are com-
puted from the depth map D. Note that all the pixels that are not visible
in the 45◦ view do not need to be considered.
We slice using vertical planes if dz/di(p) > dy/di(p), and horizontal

planes otherwise. For uniform surfaces such as a quarter-sphere, both
directions produce reasonable results.
3.2.2.2 Slice Positioning
Having determined the orientation of the slices, we proceed to find their

positions along the y-axis (horizontal slices) or z-axis (vertical slices).
The positions are computed according to three criteria.
47
F IGURE 3.10: Slice positioning and contour projection: An input surface (a) is
represented by a set of vertical slices (b), which are first positioned to satisfy
the minimum patch width threshold and constant gap-gradient ratio. The slic-
ing positions are then optimized to minimize the contour discontinuity, while
maintaining the gap-gradient ratio (c). Finally, the contour of the original hole
is projected onto the corresponding patch (d). Bottom row shows the side view
for each step.
Firstly, we aim to make the OA plan easy to cut. We keep the

patches bounded by consecutive slices not thinner than wmin , a threshold
proportional to the segment area. We constrain the width of a patch in
the 45◦ image space by using dilation technique in morphological image
processing [39]. At each slicing position s, we estimate the minimal gap
gmin (s) such that the width of the patch bounded by two slices at s and
s + gmin (s) is at least wmin .
Secondly, each slicing gap is kept proportional to the square root

of the average surface gradient perpendicular to the slice contour. For
each slice s, let g(s) be the slicing gap at s and grad(s) be the average
surface gradient perpendicular to s. We observe that, for each segment,
p
maintaining a constant ratio r = g(s)/ grad(s) over all the slices results
in a pleasant appearance.
Thirdly, we avoid slicing through holes on the input surface by

minimizing the total discontinuity along the slice contours. We measure
contour discontinuity as the distance between two contour pixels that
48
are both adjacent to another segment. We denote it as dis(s), a function

of slice s.
An illustration of our slice positioning is shown in Fig. 3.10. To

achieve criteria 1 and 2, we start from an initial slice position s∗ where
the maximal surface gradient is greatest. We place two slices at s∗ and
s∗ +gmin (s∗ ), and compute the desired ratio between slice gap and sur-
p
face gradient, r = gmin (s∗ )/ grad(s∗ ). We then estimate the positions
of other slices such that the corresponding ratio at every slice is equal to
r. These initial positions guarantee that all the widths of all the patches
are not smaller than wmin , and a constant ratio r is maintained over the
segment (Fig. 3.10, (b)).
To optimize the slice positions for criterion 3, we use dynamic pro-

gramming to minimize the discontinuity on the slices as follows. Let s1 ,
s2 , . . ., sn be the initial positions that satisfy criteria 1 and 2, and [s1 ], [s2 ],
. . ., [sn ] be small ranges around them to allow flexibility in slicing. We de-
fine Dis(k, s) as the total contour discontinuity up to slice k at position
s ∈ [sk ].
We accumulate the minimal total discontinuity using the following

formula.
Dis(k, s) = Dis(k−1, sprev (k, s)) + dis(s)
where sprev (k, s) is the position of slice k−1 that produces the minimal
discontinuity up to that slice, and is computed as
sprev (k, s) = arg min(Dis(k−1, s0 ))

s0
To maintain the gap-gradient ratio, s0 can only lie within the range
[sk−1 ] ∩ [s−r×grad(s)]. Eventually, after obtaining the minimal Dis(n, s)
49
for the last slice, we use sprev to trace the optimal positions for all the
slices (Fig. 3.10, (c)).
In our slice positioning algorithm, the slices may still cut through
holes on the input surface if they are big. Such case can be handled inter-
actively by allowing the user to adjust wmin and the affected slices based
on their positions suggested by the system.
3.2.2.3 Surface Contour Projection
After slicing the segments to generate the representative patches, we re-

construct the holes inside each segment by projecting their contours onto
the corresponding patch (Fig. 3.10, (d)). Consider point p on a contour
of the input surface. If the surface is projected onto a patch p, then p’s co-
ordinate in the 45◦ orthographic view, (ip , jp ), is transformed into (i0p , jp0 )
using the following simple formula.
√

i0p = ip + (−1)k (d(p) − d(p))/ 2

j 0 = jp

p
If p is a vertical patch, d is the distance measured from the xy-plane

and k = 1. Otherwise, if p is horizontal, d is measured from the xz-plane
and k = 0. Note that we only need to reconstruct the parts that are visible
in the 45◦ orthographic view.
3.2.3 Foldable OA Construction
The representative patches have been created as continuous, non-overlap-

ping regions in the 45◦ view. Hence, their projections along the (0, −1, −1)
50
direction onto the xz- plane satisfy properties 1 and 2 of an OA (Defini-

tion 3.3). We now connect the patches to form paths from pB to pF and
fulfill property 3.
3.2.3.1 Pairwise Patch Connection
Consider pa and pb , two patches sharing a non-vertical contour projec-

tion in the 45◦ image space. Without loss of generality, we assume that
pa and pb lie respectively above and below their shared contour in this
space. These two patches are adjacent if they also touch each other at
their actual contour in the 3D space. If they are not adjacent, we may
need to connect them to form a path from pF to pB . However, not all
such pairs of patches can be connected, because they must all be visible
in the 45◦ view. For pb and pa to be connectable, they need to satisfy

m(y(pb )) + δ ≤ m(y(pa ))

m(z(pa )) + δ ≤ m(z(pb ))

If pb is vertical and pa is horizontal, we use m = min (Fig. 3.11

(a,b)). Otherwise, m = max (Fig. 3.11 (c,d,e,f)). From the side view, δ is
the minimal length allowed for each patch.
If two connectable patches are not parallel, we create a new con-

nection by extending one of them along its contour (Fig. 3.11 (a,b,c,d)).
If they are both horizontal (vertical), their orthogonal connecting patch
is created to pass the bottom (top) contour of pa (pb ) (Fig. 3.11 (e,f)). After
positioning the new patch, we give it the maximal width such that the
gap between the new fold line and the original contour is smaller than a
51
F IGURE 3.11: Possible relative positions of pa and pb . The top row shows con-
nectable pairs of patches. Red segments represent the parts on existing patches
that need to be removed. Black and dashed blue segments represent the re-
maining parts and new connections to be added, respectively. The bottom row
shows non-connectable pairs. The rightmost column illustrates the connecting
result for case (f) in the 45◦ view.
predefined threshold.
3.2.3.2 Global Patch Connection
Besides knowing whether two patches are connectable and how to con-
nect them, we need to determine the order of the pairs to be connected.
It may be inefficient if we connect two patches adjacent in the 45◦ view
but are in fact very far in the 3D space.
We define the connection cost between pa and pb as the total length

of the parts added and removed when connecting pa and pb . This cost is
zero if they are adjacent.
Given the connectability of the patches, we keep constructing the

least-cost paths from pF to pB until no more patch needs to be added. The
cost of a path is evaluated based on the total connection cost to construct
it.
If a remaining patch cannot be added to any path from pF to pB , we

merge it into the nearest parallel patch that shares with it a contour in
the 45◦ view. If it cannot be merged, we simply discard it.
52
(a) (b) (c)
F IGURE 3.12: A 3D model (a) is approximated by a set of representative patches

(b), which then go through the connecting-merging process to form a foldable
OA (c).
The connecting-merging process as described above may not pre-

serve the input surface well. For instance, in Fig. 3.12, patch 1 is merged
into patch 2, which fills up the original alcove. In other cases, a patch
corresponding to a concave surface may be totally discarded, causing
a hole in the final structure. However, we find such results reasonable
because the remaining patches satisfy both Definition 3.3 and Proposi-
tion 3.1. Their projection along (0, −1, −1) on the xz-plane still forms a
foldable OA plan.
3.2.4 OA Stabilization
When we make a path, a patch or an OA stable, we say that we “stabi-

lize” it. In order to stabilize an OA plan, we make its paths of patches
monotonic or near-monotonic. The paths are repeatedly stabilized in
three steps: sorting the paths, stabilizing the best one, and updating
them. Before each iteration, we divide the paths so that each of them
starts and ends with stable patches, but does not visit any other stable
ones. When no more such path can be found, the stabilization process
stops.
53
3.2.4.1 Path Sorting
First, we need to determine the criteria for path comparison. If an n-

path P contains nB pairs of doubly-connected even patches that belong
to the B-set, we can make the other n∗B = bn/2c−nB even pairs doubly-
connected by adding n∗B odd patches to them. All the even patches along
P will then belong to B-set and P will become a B-path, according to
Lemma 3.1. Similarly, we denote n∗F as the number of even patches we
can add to make P an F -path.
These values give us n∗P = min(n∗B , n∗F ), a measurement for path

comparison. Let P1 and P2 be two paths that we want to compare, which
can become B-paths or F -paths by adding to them n∗P1 and n∗P2 patches,
respectively. We say that P1 is better than P2 if one of the following con-
ditions occurs.
1. n∗P1 < n∗P2 .
2. n∗P1 = n∗P2 , and P1 is monotonic but not P2 .
3. n∗P1 = n∗P2 , neither P1 nor P2 is monotonic, and P1 is near-monotonic

but not P2 .
4. n∗P1 = n∗P2 , P1 and P2 are both monotonic or both near-monotonic,

or neither of them is monotonic or near-monotonic. However, P1
visits more patches than P2 .
These conditions are used because they guarantee that a better path will
need fewer extra patches to become a monotonic or near-monotonic path.
Based on these conditions, we sort the paths from the best to the worst.
54
3.2.4.2 Path Stabilization
We stabilize the patches along the best path P by making it monotonic or

near-monotonic. First, we need to ensure that P is a B-path or an F -path.
Assume that n∗P = min(n∗B , n∗F ) = n∗B > 0. We consider n∗B pairs of non-
doubly-connected even patches with the smallest pairwise distances. We
add a new connection for each pair to make it doubly-connected. As a
result, P will then become a B-path.
To add a new connection between a pair pa and pb , we first assume

pb ’s coordinate is greater. Let pc be their existing connection and pd be the
new one to be added. We choose a set of potential positions for pd along
pb ’s upper contour where the gradients are smallest. This is to keep the
horizontal fold line between pb and pd from altering pb ’s contour signif-
icantly. Among the potential positions for pd , we pick the one farthest
from pc , so that the double connection formed by pa , pb , pc and pd does
not become too thin and affect the parallelism between pa and pb .
With the chosen position of pd , we give it the maximal width such

that the gap between the new fold line and the original contour is smaller
than a given threshold. However, in staircase-like structures (Fig. 3.6
(a)), pa , pb and pc have equal j-coordinates in the 45◦ view, at which pd is
also created. In such cases, we limit pd ’s maximal width to 1/6 of pc ’s.
After double connections are created to make P a B-path, if it is also

monotonic or near-monotonic then all of its patches are stable, according
to Proposition 3.2. Otherwise, we keep merging the patches until P be-
comes monotonic or near-monotonic.
Similar to the foldable OA construction, merging of a patch p is

done towards the nearest patch p0 parallel to p. When zp < zp0 ,the merg-
ing is forward. Otherwise, the merging is backward.
55
In details, let P = {p0 , p1 , . . . , pn } be a B-path with p2i in B-set. We

assume the last even patch has greater z-coordinate than the first one.
For each index i starting from 0, we find the first index j > i such that
zp2j > zp2i . If i+1 < j ≤ bn/2c, along the sub-path {p2i+2 , . . . , p2j−2 }, we
repeatedly merge the even patch with the smallest z-coordinate forward
until it reaches p2i or a stable patch (Fig. 3.13, 1st row). If j cannot be
found, we repeatedly merge the pi backward until it reaches p2bn/2c or a
stable patch (Fig. 3.13, 2nd row).
As no patch in P is stable, except the first and the last ones, and the
merging stops when we reach a stable patch, this process does not break
any existing stable paths while making P monotonic.
F IGURE 3.13: Two B-paths (black thick patches, above and below) are made
monotonic by merging forward (above) and backward (below). The dotted ar-
rows show the merging direction. Other patches not belonging to these paths
(red thin patches) may also be involved in the merging process.
56
3.2.4.3 Path Update
After the best path P is stabilized, we remove it from the list of paths.
For each remaining path that contains the newly stabilized patches, we
divide it into shorter paths, each starting and ending with stable patches,
but not visiting any other stable ones. If there is no unstable path left,
the algorithm ends. Otherwise, we go back to the first step and sort the
remaining paths. As the number of patches is finite, this process will
finally stabilize the whole OA.
During our OA creation, the patches are created in the 45◦ image
plane, while their plan lies on the xz-plane. Therefore, we need to scale
√
the image space by 2 along its i-axis to obtain the final OA plan.
3.2.5 Comparison between our stabilization algorithm and

Li et al. [68]’s
Our stabilization technique, as described in section 3.2.4, can fix arbitrary

unstable structures by adding double connections. In this section, we
will show some main differences between our stabilization results and
those from Li et al. [68].
As mentioned in 3.1.4, Li et al. [68]’s conditions can only examine

the stability of 1-paths or 2-paths. In order to stabilize a longer path, their
method needs to connect the unstable patches to the existing stable ones
to form 1- or 2- paths between these stable patches. Since most unstable
patches are not adjacent to the stable ones, this process normally causes
an extensive change to the original OA. In particular, in order for new
connecting patches to be created, the widths of the original patches will
become significantly narrower.
On the other hand, our algorithm stabilizes arbitrary paths by con-

necting pairs of unstable patches to form a monotonic or near-monotonic
57
paths. Since we have a wide range of choices for patches that can be con-
nected, our stabilization step does not need to alter the overall shape of
the OA much.
(a) (b) (c)
(d) (e) (f)

F IGURE 3.14: (a, d) Unstable OA structures are made stable by (b, e) Li et al.
[68]’s stabilization method and by (c, f) ours. Our stable OA is significantly
closer to the original OA than [68].
3.3 Results
We show a number of OA designs automatically generated by our sys-

tem and compare them with the results from Li et al. [68], the only auto-
matic OA design system available. In their implementation, an OA can
be generated at several voxel grid resolutions, ranging from 8×8×8 to
256×256×256. For comparison, we choose their results at 256×256×256
58
for the most detailed approximation, and 64×64×64 for the most bal-
anced between contour preservation and voxel size. We perform both
visual and numerical comparisons between the results from the two sys-
tems.
F IGURE 3.15: OA pop-ups of the Rialto bridge (left) and Colosseum (right), as
designed by artists [34] (top row), our system (middle row) and the voxel-based
OA tool in [68] (bottom row).
59
3.3.1 Visual Comparisons
In Fig. 3.15, the designs from both systems are compared with artists’
creations [34]. We observe that [68] is unable to preserve the archways
in the Rialto bridge model, and requires the user to manually add them,
unlike in our system where we can preserve them automatically. Fur-
thermore, [68] cannot produce a good approximation of the Colosseum
model. This may be due to their stabilization step, which does not han-
dle concave regions well.
Fig. 3.16 - 3.19 demonstrate the designs from our method and from
[68] for various models. We smooth the highly bumpy bunny model as a
pre-processing step for both systems. The figures illustrate the ability of
our algorithm in approximating smooth surfaces nicely (e.g. the bunny,
the sphere, the torus). On the other hand, we can also approximate the
sharp contours (e.g. the cross of the chapel), as well as the important
creases well (e.g. the edge between two slopes of the triangular block).
Our novel stabilization can be best exemplified in the U.S. Capi-

tol model (Fig. 3.19). Its main staircase is stabilized using a monotonic
structure, which minimizes the changes in volume and shape, as com-
pared to the technique used in [68]. In our implementation, we construct
double connections on both sides of the staircase to maintain symmetry.
Monotonic structures are also used for the Rialto bridge (Fig. 3.15),
the bunny (Fig. 3.18) and the trefoil knot (Fig. 3.16), while near-monotonic
structures are used for the gate model (Fig. 3.6) and the elephant models
(Fig. 3.16).
60
F IGURE 3.16: OA designs of some experimented models created by our method

(blue) and by [68] (green): a floating box, a torus, a trefoil knot and an elephant
models. We choose resolution 64×64×64 for [68] because higher resolutions are
not feasible for cutting and folding.
61
F IGURE 3.17: OA designs of some experimented models created by our method

(blue) and by [68] (green): a triangular block, a quarter-sphere, a complete
sphere, and a Taj Mahal models. We choose resolution 64×64×64 for [68] be-
cause higher resolutions are not feasible for cutting and folding.
62
F IGURE 3.18: OA pop-ups of the Stanford bunny and a church model. From
top to bottom are our designs (blue), and [68]’s designs at 64×64×64 (green)
and 256×256×256 (red).
3.3.2 Numerical Comparisons
Besides qualitative visual comparison, we compare our results quanti-

tatively by using the volumetric percentage difference of our structures
63
F IGURE 3.19: Our OA design for the U.S. Capirol Building (left) and [68]’s de-
sign at 64×64×64 (right). The bottom row shows close-ups of the staircase.
and the input models. This is computed by accumulating of the abso-

lute differences in 45◦ depths at all pixels, divided by the original input
volume. A comparison of the percentage differences of the structures
designed by [68] and our own results is shown in Table 3.1. Even though
we use considerably fewer cuts and folds, our volumetric percentage dif-
ferences are smaller than those of [68] at their highest resolution in many
cases. Especially, our results are significantly better for models with big
concave surfaces, since our algorithm can preserve them well.
For the Quarter-Sphere, the Stanford Bunny and the Triangular Block
models, our volumetric preservation is slightly less accurate than [68]’s,
because our system produces remarkably fewer folds (11 times, 3 times
and 9 times, respectively, as shown in Table 3.2). As a result, our pop-ups
64
Percentage Li et al. [68]

Ours
Difference (%) 32 64 128 256
Chapel 13.78 9.41 4.46 2.63 2.4
Colosseum 44.03 45.3 41.93 40.9 5.59
Quarter-Sphere 25.72 10.1 8.72 8.12 8.7
Rialto Bridge 52.88 23.59 15.64 14.57 12.27
Stanford Bunny 15.44 10.77 9.95 8.17 11.6
Torus 139.49 91.28 84.38 81.19 21.57
Triangular Block 13.66 6.63 3.95 2.16 3.32
US Capitol 26.85 14.28 9.65 5.9 4.93
TABLE 3.1: Percentage differences of our OA structures and [68]’s at various

resolutions, as compared to the original input models.
can be folded much more easily.
Li et al. [68] Ours

Model
Cuts Folds Cuts Folds
Church 19 35 13 23
Stanford Bunny 77 188 29 57
Triangular Block 246 517 25 55
Quarter Sphere 223 496 22 44
Torus 93 202 9 18
Trefoil Knot 12 27 8 17
Taj Mahal 43 85 34 53
Elephant 29 33 14 19
TABLE 3.2: Comparison of number of cuts and folds between our OA designs
and those created by [68].
Table 3.2 shows the number of cuts and folds in the OA designs, as a
measurement of their complexity. Except the Rialto bridge model, whose
archways are not reconstructed, and the Colosseum model, whose shape
is not preserved by [68], other experimented models show that our re-
sulting OAs require significantly fewer cuts and folds than the corre-
sponding OAs from [68] (Table 3.2).
65
In some cases, the cut lines may have arbitrary shapes. As such, the
number of cuts and folds may not be the only indicator of how complex
an OA design is. Hence, we also measure the time required to construct
the actual structures from the 2D designs as another basis for determin-
ing their complexity. Table 3.3 shows the approximate time in minutes
for a novice to cut and fold the pop-ups. All the presented OA plans
were generated by our system in at most 10 seconds on a Pentium (R)
Dual-Core 2GHz computer with 4GB RAM.
Model Time Model Time

Stanford Bunny 35 Torus 9
Rialto Bridge 36 Complete Sphere 24
Colosseum 35 Trefoil Knot 10
Church 10 Taj Mahal 29
Triangular Block 14 Elephant 26
TABLE 3.3: Time (in minutes) to cut and fold the actual OA pop-up from our
design layout.
3.4 Discussion and Conclusion
In this chapter, we have presented a practical method for automatic OA

generation grounded on a set of comprehensive geometric formulations.
Sufficient and necessary foldability conditions for parallel OA were for-
mulated. They support our novel image-based slicing approach for cre-
ating 2D plans that fully pop up into parallel patches, whose positions
and contours closely resemble the overall curvatures, as well as the sub-
tle details of the 3D input models. By using slices with exact contours
rather than discrete voxels, we can nicely preserve the meaningful shapes
of the input models.
66
In our formulation, we also introduced a new set of conditions for

stabilizing OA structures. Our stability conditions are more general than
those in recent studies [77], [68], [4] and cover all the structures created
by pop-up artists that we have seen [34], [8], [10]. Utilizing our novel
double connections, monotonic paths and near-monotonic paths, we can
stabilize OA structures without affecting their shapes or volumes signif-
icantly.
Visual and numerical analyses show the ability of our approach to

design foldable and stable OA that both approximate the desired struc-
tures well and are feasible in practice. Since we carefully compute the
slice positions and contours, our designs can preserve the input surfaces
while creating significantly fewer cuts and folds, as compared to the pre-
vious automatic OA tool [68]. A few comparisons with artists’ works
also show the potential of our system in actual production.
Limitations
Although our algorithm produces visually pleasing designs, it may not

always resemble some of the designer’s artistic choices. For instance, in
the Rialto bridge model (Fig. 3.15), the sharp tip of the roof is depicted
by artists, but not in our design, because it is projected to the horizontal
fold line. Similarly, in our implementation, the slicing width thresholds
computed automatically may be subjective and not conformable to the
cutting and folding skills of all users. A possible solution is an interactive
system, in which users can indicate the details they want to preserve, and
the minimal patch widths that are sufficient for them.
In addition, due to the OA constraint of a single piece of paper, we

are not always able to preserve highly concave regions, as shown with
the floating cube, complete sphere and trefoil knot models (Figs. 3.16 and
67
F IGURE 3.20: Input model that is not well-aligned with the back and floor bases
may not be converted nicely into an OA pop-up.
3.17). We believe these models can be better approximated by combining

OA with other types of paper pop-ups.
Another open problem in the design of traditional parallel OA is

how to align the input models automatically. As shown in Fig. 3.20, a
badly aligned model may result in an undesired OA pop-up. This ar-
tifact is particularly visible for architectural models. An alternative so-
lution is to use non-parallel patches to approximate the slanted input
surfaces. We will present that possibility in Chapter 4.
In our current geometric formulation, we do not take into account

the physical characteristics of paper. However, in practice, the thickness,
mass and elasticity of paper are important considerations for OA design.
We will analyze them in Chapter 5.
68
Chapter 4
Origamic Architecture with

Non-parallel Folds
In Chapter 3, we presented the theory and algorithm for the automatic

design of traditional parallel origamic architecture. Although the formu-
lated OAs only consisted of parallel fold lines, they were able to feature
numerous types of 3D shapes, from architectural to organic objects.
However, we also observe that in practical OA design, artists oc-

casionally use non-parallel folds (Fig 4.1). Since such folds are not con-
strained in a fixed orientation, they may be used to approximate slanted
surfaces better. In addition, a non-parallel fold may sometimes be used
to replace a series of parallel folds, which reduces the difficulty of the
actual pop-up creation (Fig. 4.2).
Hence, in this chapter, we explore an extended solution for OA

design that take into account both parallel and non-parallel folds. We
F IGURE 4.1: Pop-up designs with non-parallel folds.
69
Chapter 4 Origamic Architecture with Non-parallel Folds
investigate the conditions for geometrically valid OA with non-parallel

folds, as well as an algorithm to generate such folds. For convenience,
from this chapter, we will call an OA with only parallel folds, as de-
scribed in Chapter 3, parallel OA, and an OA with non-parallel folds gen-
eral OA. The work on general OA in this chapter has also been presented
partially in [88, 89].
While non-parallel folds can be used to capture slanted surfaces,

OA still mainly consist of parallel folds, because they create nicely uni-
form shadowing effects when popping up. Hence, we still use the algo-
rithm described in Chapter 3 for parallel OA as the backbone for design-
ing general OA. In this chapter, we will extend it to design non-parallel
folds for suitable slanted surfaces. Note that we only reconstruct sur-
faces that are visible in the 45◦ orthographic view, because our abstrac-
tion method is applied in this view.
First, in order to describe the algorithm for non-parallel fold con-

struction, we need to investigate the geometric conditions for its validity,
including both foldability and stability.
F IGURE 4.2: A series of parallel structures can be replaced by a single non-

parallel one for easier cutting and folding.
70
4.1 Geometric Formulations
4.1.1 General OA Components
The components of a general OA are similar to those of a parallel

OA defined in Section 3.1.1 on page 29, except that the fold lines do not
need to be parallel to the central fold. In addition, we denote a popped-
up structure that contains two patches connected by a non-parallel fold
as a v-structure, since these two patches form a v-shape structure. To fa-
cilitate our further discussion, we distinguish between concave folds and
convex folds (Fig. 4.3 (a)). We also call the fold line between the two
patches that a popped-up structure is based on a base fold line, and the
angles formed by this line and the two convex folds base angles. Note
that when we handle an angle between two fold lines, we consider their
extensions, not just the actual fold segments. This is because some v-
structures may be cut off at their tips, as shown in Section 4.3 on the
results.
In our study, we only consider the v-structures whose convex fold

lines intersect along the base fold line. This is similar to the type of struc-
tures considered in [67] and [36]. Specifically, the concurrency of the fold
lines gives the v-structure a potential to pop up. Additionally, in each
v-structure, we assume the angles between its convex and concave fold
lines to be smaller than 90◦ . Otherwise, they form a straight line and two
patches of the structure may rotate freely when it is open at 180◦ (Fig 4.3
(b)).
The definition of a general OA plan inherit the properties of a par-

allel OA plan, which were described in Definition 3.3 on page 30. We
add two more properties to incorporate non-parallel folds as follows.
71
(a) (b)
F IGURE 4.3: (a) The components in a v-structure. (b) A v-structure with two 90◦
angles may rotate freely when opened at 180◦ .
Definition 4.1. A general OA plan is a set of patches where
1. All the patches are co-planar and form a rectangular domain with possible
holes.
2. They are non-intersecting, except at their boundaries.
3. For every patch p, there exists a path traversing from the back patch pB to
the floor patch pF and containing p.
4. For every v-structure, its concave fold line and convex fold lines intersect
at one point along the base fold line.
5. For every v-structure, the inner angles between its convex fold line and
concave fold lines are smaller than 90◦ .
General OA plans share the first 3 common properties with tradi-

tional parallel OA plan. Properties 4 and 5 correspond to the assump-
tions on v-structures described earlier.
72
4.1.2 Foldability of General OA
The foldability of a general OA is defined similarly to that of a parallel

OA, which means the OA plan needs to be foldable from θ = 180◦ to θ =
without affecting the shapes, pairwise adjacency and non-intersection of
the patches.
In Proposition 3.1 on page 31, OA foldability is examined by pro-

jecting the popped-up structure onto the xz-plane and check whether it
matches the OA plan. Such condition is useful for parallel OA design,
because its abstraction utitilizes a 45◦ orthographic projection.
However, non-parallel patches do not project onto the OA plan in

one fixed direction. Hence, we may not re-use Proposition 3.1 for check-
ing the foldability of a general OA. In this section, we present another
set of foldability conditions for v-structures and general OA.
4.1.2.1 Foldability of a Path of v-Structures
First, we consider an OA plan that only consists of a series of n v-structures

forming a 2n-path between two bases p1 and p2 . For convenience, we call
this series a v-path. Let us denote the outermost base angles as ωα and
ωβ . The angles between alternating convex and concave fold lines on the
v-structures are denoted as α1 , β1 , α2 , β2 , ..., αn , βn . We now present an
angle condition for the foldability of a v-path.
Proposition 4.1. An OA plan consisting of only a path of n v-structures is

foldable if and only if ωα = β1 + β2 + ... + βn , and ωβ = α1 + α2 + ... + αn .
An illustration of Proposition 4.1 is shown in Fig. 4.4. The necessity

of this proposition for a single v-structure has been described in some
73
F IGURE 4.4: A foldable path of v-structures and its angles.
previous studies, such as [67] and [36]. In this section, we prove the
sufficiency and necessity of this proposition for multiple v-structures.
Proof.
Sufficiency. Assume that we have an OA plan of n v-structures satisfying

ωα = β1 + β2 + ... + βn , and ωβ = α1 + α2 + ... + αn , which lead to
ωα + α1 − β1 + ... + αn − βn − ωβ = 0. We will show that this OA is fully
foldable.
First, consider an OA plan consisting of only one v-structure. If

ωα0 + α0 − β0 − ωβ0 = 0, then it is possible to put the OA in the fully
closed configuration (Fig. 4.5 (a)). Now we can also prove that the OA is
foldable to any other opening angle. As shown in Fig. 4.5 (b), when the
back patch is rotated, point P remains the intersection of three spheres
centered at O, Oα and Oβ , with radii OP , Oα P Oβ P , respectively. This
intersection can always be found at any arbitrary opening angle, because
OOα < OP + Oα P and OOβ < OP + Oβ P .
To fold an OA with n v-structures, we can fold each of them succes-

sively. When we completely fold the first v-structure that is adjacent to
the floor patch, the remaining n − 1 v-structures will form with the floor
patch an angle wα + α1 − β1 . By continuing to fold completely each of the
first n − 1 v-structures, the last v-structure will form with the floor patch
an angle wα + α1 − β1 + ... + αn−1 − βn−1 . Since ωα + α1 − β1 + ... + αn − βn −
74
ωβ 0
Oβ P
β0
Oα
ωα α0
0
O
(a) (b)
F IGURE 4.5: (a) A v-structure in its closed configuration illustrates the relation-
ship between its angles. (b) The points in a v-structure when it is opened.
ωβ = 0, the last v-structure satisfies the foldable condition for a single v-

structure, ωα0 +α0 −β0 −ωβ0 = 0, where ωα0 = wα +α1 −β1 +...+αn−1 −βn−1 ,
ωβ0 = ωβ , α0 = αn and β0 = βn . Hence, the OA is fully foldable.
Necessity. Assume that we have a foldable OA plan consisting of n v-

structures. We will show that their angles satisfy ωα = β1 + β2 + ... + βn ,
and ωβ = α1 + α2 + ... + αn .
When it is fully closed, as shown in Fig. 4.5 (a), we have ωα + α1 −

β1 + α2 − β2 + ... − ωβ = 0, which means ωα + α1 + α2 + ... + αn =
ωβ + β1 + β2 + ... + βn . On the other hand, when the OA plan is opened
flat, we also have ωα + ωβ = α1 + β1 + α2 + β2 + ... + αn + βn (Fig. 4.4).
Hence, ωα = β1 + β2 + ... + βn , and ωβ = α1 + α2 + ... + αn .
Our generalization from a single v-structure to multiple v-structures

allows more complex OA. However, the foldability of v-structures is not
the only consideration in the design. We also need to guarantee that
different structures do not overlap or intersect during the opening and
closing process.
75
4.1.2.2 Foldability of a Combination of Structures
When there are only parallel fold lines, as discussed in Chapter 3, no

inter-structure overlapping occurs, since the projections of all the points
in the pop-up along the central fold are fixed. In addition, the patches
always remain parallel to the two outermost bases and do not intersect,
except along the fold lines. Fig. 4.6 (a) shows the coverage of a parallel
structure at 180◦ and 0◦ opening angles.
(a) (b)
F IGURE 4.6: (a) OA plan of a parallel structure (top) and its coverage when fully
closed (bottom). (b) OA plan of a v-structure (top) and its coverage when fully
closed (bottom).
76
On the other hand, when a v-structure is used, its projection on the

base fold line move during the opening and closing process. We need to
compute the range of movement of a v-structure to avoid inter-structure
overlapping.
Consider a v-structure whose fold lines form two angles α and β,

as shown in Fig. 4.6 (b, top). In order for the structure to be foldable, α,
β, and the base angles ωα , ωβ must satisfy ωα = β and ωβ = α.
If the fold lines of the v-structure intersect at fmin along the base
fold, and its projection lie from fmin to fmax at 180◦ opening angle, then
the actual length of its convex fold line is kf k = (fmax − fmin )/cos(γ),
where |γ| = |β − ωα | = |α − ωβ | = |β − α|. Then at 0◦ angle, the projec-
tion of this convex fold line on the base fold has length kf kcos(α + β).
Additionally, its projection on the axis perpendicular to the base fold has
length kf ksin(α + β).
Hence, at 0◦ opening angle, the v-structure lies within [fmin −(fmax −

fmin )cos(α + β)/cos(α − β), fmax ] along the base fold, and [0, (fmax −
fmin )sin(α + β)/cos(α − β)] along its perpendicular axis Fig. 4.6 (b, bot-
tom).
Knowing the coverage of each structure when folded from 180◦ to

0◦ , we present the conditions for the foldability of a general OA plan as
follows.
Proposition 4.2. A general OA plan is foldable if it satisfies
1. All v-structures satisfy the angle condition in Proposition 4.1.
2. No parallel structure is based on a v-structure.
3. For each base fold line, there is no overlapping between the coverages of
structures lying on it when they are folded.
77
The first condition follows Proposition 4.1. The second condition is

required for every parallel structure to pop up, because its convex fold
line has to be always parallel to the base fold. Finally, the third condition
is based on the coverage calculation earlier.
4.1.3 Stability of General OA
The stability of a general OA is defined similarly to Definition 3.4

on page 33 for a parallel OA. In brief, it needs to be able to fully pop up
when the user turns only the back and floor patches, but not any other
patch.
As discussed in Section 3.1.3 on page 33, previous studies only con-

sidered 1-paths and 2-paths for stable parallel OAs. Similarly, they con-
sider only 2-paths for stable v-structures. This may greatly limit the pos-
sibility for designing v-structures, especially when we have proven in
Section 4.1.2 that an arbitrary path of v-structures is foldable, as long as
it satifies the angle condition.
In Section 3.2.4 on page 53, parallel structures can be made sta-

ble by creating double connections. We observe that an equivalent ap-
proach can also be used for v-structures. However, since the fold lines
and patches in v-structures are not parallel, we define double connec-
tions slightly differently and based on the angles between the fold lines.
Definition 4.2. Consider 4 mutually non-coplanar patches p1 , p2 , p3 , p4 , where

p1 and p2 intersect p3 and p4 at l13 , l14 , l23 , and l24 , respectively. The angles
between (l13 , l14 ), (l23 , l24 ), (l13 , l23 ) and (l14 , l24 ) are denoted as δ34/1 , δ34/2 ,
δ12/3 and δ12/4 . We say (p1 , p2 ) and (p3 , p4 ) are doubly-connected if δ34/1 =
δ34/2 and δ12/3 = δ12/4 . In such case, the structure (p1 , p2 , p3 , p4 ) is also called
a double connection.
78
p
10
8=p
p B
p
9
p
5 p
7
p
6
p p
4 δ
2
p
δ34/
2 12/
4
3 δ12/
3 δ34/1
p
1
0=p
p F
F IGURE 4.7: Three doubly connected v-structures (p1 , p2 , p3 , p4 ), (p2 , p5 , p6 , p7 )

and (p5 , p8 , p9 , p10 ), as described in Definition 4.2. For instance, in the first
double connection, we have δ34/1 = δ34/2 and δ12/3 = δ12/4 .
Fig. 4.7 illustrates double connections of v-structures. We utilize

this type of connection to introduce a new stability condition for a path
of v-structures.
Proposition 4.3. A path of v-structre is foldable and stable if all pairs of even
(odd) patches along the path are doubly connected.
The odd and even patches are defined similarly to Section 3.1.3 on
page 33. If P = {p0 = pB , p1 , . . . , pn , pn+1 = pF } be a path traversing from
the back patch to the floor patch, then along P, p2k is called an even patch
and p2k+1 is called an odd patch, where 0 ≤ k ≤ bn/2c.
A sample foldable and stable path as described in Proposition 4.3

is shown in Fig. 4.7. The stability conditions for parallel structures and
v-structures are analogous. However, while the foldability of parallel
79
double connection is intuitive due to parallelism, that of non-parallel

double connection is not straightforward. We prove the foldability of
doubly-connected v-structures as follows.
β3
α3
ωβ
β2
α'
3 α2
β1
ωα3 α' α1
ωα2 2
ωα ωα1
(a)
p
B
Oβ
’
O
P’ p
F
Oβ
P
Q
Oα
(b)
F IGURE 4.8: (a) A path of n v-structures can be considered a combination of n
single v-structures based on the floor patch. (b) Extra patches can be added to a
single v-structure to form a foldable, doubly-connected v-structure.
Proof.
Consider a foldable OA consisting of only a v-path. As shown in

Fig. 4.8 (a), at any opening angle, we can hold the patches so that the
intersections of their extensions with the base patches satisfy ωαi = βi for
80
0
all i ∈ [1, n], ωβ = αn + αn0 , and αi+1 = αi + αi0 for all i ∈ [1, n − 1]. This is
Pn Pn
possible because ωα = βi , and ωβ = αi .
1 1
Hence, the OA can be treated as a vertical accumulation of n simple

OAs, each consisting of only one foldable v-structure. In order to show
that the doubly-connected v-path is foldable, now we only need to show
that doubly connecting each foldable v-structure still allows it to fold
completely.
Consider a foldable v-structure containing patches pB , pF , OOα P

and OOβ P (Fig. 4.8 (b)). We add patches OP P 0 and OP 0 Oβ0 on top of
0 0
this v-structure such that P
\ OP 0 = O\ 0
\
β OOβ and P OOβ = P OOβ . By
\
Proposition 4.1, the v-structure (OOβ P, pB , OP P 0 , OP 0 Oβ0 ) is also fold-

able, making the structure (pB , pF , OOα P 0 , OP 0 Oβ0 ) foldable and leading
to O\ 0 \0 \0 \0
α OP = QOOβ . Note that we also have QOOβ = QOOβ + Oβ OOβ =
\
O
\ \0 \0 \0
α OP + P OP . Hence, Oα OP = Oα OP + P OP . In other words, three
\
patches Oα OP , Oα OP 0 and P OP 0 are co-planar. Hence, the resulting

structure is a foldable doubly-connected v-structure, according to Defi-
nition 4.2.
Besides the foldability of doubly-connected v-structures, we also

examine its stability by using both simulated models and actual paper
pop-ups. We create a number of OAs consisting of up to 5 v-structures.
The angles between the fold lines are randomized such that they satisfy
Proposition 4.1 and Definition 4.2.
Figs. 4.9 and 4.10 show a simulated model and an actual paper
pop-up that make use of double connections to create foldable and sta-
ble v-structures. The simulation is done using V-REP software [24] and
rendered with OpenGL. We observe that no patch is stuck or collapses,
81
F IGURE 4.9: A simulated OA containing a path of 3 doubly connected v-

structures. The closing motion of the OA is captured from top to bottom and
left to right.
and all the patches move in a stable manner when we hold and turn only
the two outermost patches.
Our experiments with both simulated and actual pop-ups empiri-

cally show the stability of doubly-connected v-structures. It is also ob-
servable that, for every doubly-connected v-structure, the possible ori-
entation of each patch is constrained by that of the opposite patch. In a
82
F IGURE 4.10: A real OA paper pop-up containing a path of 4 doubly connected

v-structures.
v-path that consists of doubly-connected patches, their orientations are

mutually constrained. Together with the positional constraints of the
first and last patches, it may be possible to formulate a mathematical
proof for the stability of doubly-connected v-structure.
Presently, such formal proof has yet been achieved. The main chal-
lenge is to find an appropriate relationship between the patches, the fold
lines, or the points on the fold lines, such that it remains unchanged dur-
ing the opening and closing of the v-structures. Earlier, for parallel OA,
the parallelism between the patches holds true at any opening angle and
guarantees their stability. An equivalent property is difficult to find for
non-parallel folds. In particular, the points on the fold lines move along
non-coplanar and non-parallel circles (Fig. 4.11). As a result, the orien-
tations of the vectors connecting these points vary in a complex manner,
83
F IGURE 4.11: The points on a line in a non-parallel OA plan (top) will move
along non-coplanar and non-parallel circles during the opening and closing
process (bottom).
and investigating the constraints between them is not trivial.
Based on the empirical results, in this work, we use double con-

nections as a means to stabilize paths of v-structures. We discuss our al-
gorithm for designing general OAs that contain v-structures in the next
section.
84
4.2 General OA Design Algorithm
We presented our algorithm for designing parallel origamic archi-

tecture in Section 3.2 on page 43. In order to design a general OA con-
taining non-parallel structures, we extend that algorithm to create v-
structures for appropriate slanted regions, instead of using parallel struc-
tures for all regions.
As described in Section 3.2.1 on page 45, the input surface is di-

vided into distinct non-overlapping, smooth segments before patch gen-
eration. This pre-processing step is still performed for general OA de-
sign.
In our pipeline, the generation of v-structures is done after the sur-

face segmentation and before the generation of parallel patches. By de-
signing the OA in this order, we do not need to construct multiple paral-
lel patches for a slanted surface if simpler v-structures can be used.
The design of v-structures starts with the surface segments pro-

duced from the pre-processing step. It then follows three main steps.
1. Finding the potential surface segments that can be approximated

using v-structures: We select a list of surface segments that are po-
tential for v-structure construction based on property 5 of Defini-
tion 4.1 on general OA plans, and conditions 2 and 3 of Proposition
4.2 on general OA foldability. More details are described in Section
4.2.1.
2. Constructing foldable v-structures from the selected surface seg-

ments: We compute the angles of the v-structures so that they ap-
proximate the selected surface segments closely, while satisfying
property 4 of Definition 4.1, Proposition 4.1 and condition 1 of
Proposition 4.2. We separate this step from the segment selection as
85
it requires exact calculation of the angles between the folds. More

details are described in Section 4.2.2.
3. Stabilizing v-structures: We generate extra patches to form double

connections, which are used to support the stability of v-structures,
according to Proposition 4.3. More details are described in Section
4.2.3.
As the generation of non-parallel structures is a component of a

unified pipeline for general OA design, in this algorithm, we still process
the input surface using a 45◦ orthoggraphic view. V -structures are only
generated for appropriate surface segments that are visible in this view.
The details of our v-structure generation are described in the following
subsections.
4.2.1 Potential Surface Segments for V -Structure Genera-

tion
Since each v-structure contains only two patches sharing a straight

fold line, to construct it we first need to detect straight boundaries be-
tween adjacent surface segments. Similar to the approach in Section 3.2.1
on page 45, the adjacency of the segments is examined by thresholding
the changes in depth and normal values in the segments. Then straight
boundaries between adjacent segments are detected using Canny edge
operator [13] and Progressive Probabilistic Hough Transform [73] tech-
niques. The segments containing only straight fold lines that are not par-
allel to the base patches are marked for the next processing steps, which
check whether the segments can be approximated using v-structures.
From the pairs of adjacent segments that share a straight boundary,

we select those that do not have any two consecutive fold lines forming
86
(a) (b)
F IGURE 4.12: (a) A model with a parallel block basing on a non-parallel block is
not a good candidate for v-structure generation, because the resulting OA will
not be foldable. (b) A model with a non-parallel block basing on another one
can still be converted into a foldable OA containing two foldable v-structures.
a 90◦ or larger angle. This criteria is to satisfy property 5 in Definition 4.1

(Fig. 4.3 (b)).
Next, we refine the list of selected segments so that they meet con-
ditions 2 and 3 of Proposition 4.2. For condition 2, we remove from the
list any non-base segment that is adjacent to another segment not po-
tential for v-structure generation (Fig. 4.12 (a)). By doing this, we will
not have any parallel structures basing on a v-structure. Note that the
fold lines in a parallel structure need to be parallel to the base patches in
87
order for it to be foldable.
Finally, to fulfil condition 3 of Proposition 4.2, we compute the pos-

sible range of movement of the segments if the two outermost bases close
fully at 0◦ . This computation is based on section 4.1.2.2 on page 76. From
the coverage of the folded segments along each base fold line, we find
out the overlapping segments. We then remove the segments that over-
lap with the most other segments until there is no more overlapping.
4.2.2 Generation of Foldable V -Structures
After finalizing the surface segments that can be approximated us-

ing v-structures, we proceed to create the actual pop-up. We do this by
generating a path of v-structures for each path of potential surface seg-
ments. We first generate each v-path such that their angles between the
fold lines are equal to the angles on the corresponding input segments.
We then adjust the angles of the v-structures to satisfy the foldability
condition in Proposition 4.1.
Let us use α10 , β10 , α20 , β20 , ..., αn0 , βn0 to denote the angles between
the edges of the input segments, and ωα0 , ωβ0 to denote their base angles.
Each path of v-structures is generated with the initial angles ωα = ωα0 ,
ωβ = ωβ0 , αi = αi0 and βi = βi0 for all i ∈ [1, n], where n is the number of
pairs of segments forming a potential path.
In order for the v-path to be foldable, according to Proposition 4.1,

n
P n
P
the angles must satisfy ωα = βi and ωβ = αi , for all i ∈ [1, n].
1 1
To achieve this, we compute the average amount that each angle

ωα , ωβ , αi and βi needs to be adjusted to achieve foldability, ∆α = (ωα −
P P
βi0 )/(n + 1) and ∆β = (ωβ − αi0 )/(n + 1). We recompute the angles
of the v-structures as ωα = ωα0 − ∆β , ωβ = ωβ0 − ∆α , αi = αi0 + ∆α and
88
βi = βi0 + ∆β . By doing this, the constructed patches form a foldable v-

path, while still approximating of the original segments reasonably (Fig.
4.13).
If there are more than one v-paths, they may share some of the
patches. However, since each path is distinct, it contains at least one
patch that is not shared by any other paths. Hence, we can divide it
into shorter paths and adjust the angles on each individual path inde-
pendently to make it foldable. Fig. 4.14 shows a simple example of such
situation. Paths P1 = {p1 , p2 , p3 , p4 } and P2 = {p3 , p4 } are overlapping.
If we compute the angles for these two paths, α2 , α3 and β2 will be con-
strained by the foldability condition of both paths. However, since path
P1 can be divided into {p1 , p2 } and {p3 , p4 }, we can compute two inde-
pendent sets of angles, {α3 , β2 } based on {ωα2 , ωβ2 } and {α1 , β1 } based
on {ωα1 , ωβ1 }.
By generating the paths of v-structures as above, we allow the whole

OA to be foldable, while obtaining a reasonable approximation of the in-
put surface segments.
β20 ωβ
ωβ0 α20 β2α2
β10 β1
α10 α1
ωα0 ωα
F IGURE 4.13: (Left) A selected path of segments for v-structure generation.
(Right) The generated path of v-structures. The angles along the v-path
{ωα , ωβ , αi , βi } are computed based on the angles along the segment path
{ωα0 , ωβ0 , αi0 , βi0 , ...}.
89
4.2.3 Stabilization of V -Structures
The next step of the v-structure generation is to make them stable.

If a v-path has only one structure, it is readily stable. If it contains more
than one v-structures, we stabilize it by simply adding double connec-
tions to each of the v-structures. Fig. 4.15 shows such a doubly-connect
v-structure. We set the angle on the new patch, P
\ OP 0 , to range from
1/5 to 1/3 of the angle on the original patch, O
\α OP . In addition, similar
to the case of parallel double connections, the width of the new patch,
||P Q||, is set to 1/6 of that of the original patch, ||OP ||.
4.3 Results
In practice, that parallel structures and v-structures are not com-

monly used together in a single OA. From our observation, it is possibly
because parallel structures can create nicely uniform shadowing effects.
Nevertheless, v-structures are often used to illustrate interesting slanted
surfaces.
p
B
α1 p
1
p
ωα1 p
2
F
β
ωβ1 1
α2
p β2
3
α3
ωα2 p
4
ωβ2
F IGURE 4.14: Overlapping v-paths can be divided into shorter, separate paths,
and the angles along each path can be computed independently of those on
other paths.
90
O
P’
Q
P
Oα
F IGURE 4.15: A doubly-connected v-structure.
In this section, we demonstrate a few results of our method for gen-

erating v-structures from input 3D models. As described in section 4.2,
we choose to use v-structures only for appropriate planar surfaces that
share straight fold lines. In general, given an arbitrary surface, deter-
mining whether it should be approximated using parallel structures or
v-structures is a challenging problem that will require further compre-
hensive studies.
We experimented our v-structure generation on input models with
F IGURE 4.16: An arbitrary series of triangular blocks and its corresponding OA.
91
F IGURE 4.17: A series of triangular blocks heading in different directions and

its corresponding OA.
various ranges of angles. Figs. 4.16 and 4.17 illustrate two such cases.
While the double connections lie on the same side in Fig. 4.16, they some-
times need to be constructed on different sides of the v-path, as shown
in Fig. 4.17, due to the limited space between consecutive v-structures.
Note that, although the angles in the input models do not satisfy
the foldability condition of non-parallel OAs, we do not need to alter
them significantly in our generated structures. In fact, the differences in
angles between the input models and the generated OAs are hardly no-
ticeable. This is the result of the simple but effective averaging approach
that we use for v-structure generation in section 4.2.2.
Figs. 4.18 and 4.19 demonstrate our generated OAs that resemble
interesting shapes. In Fig. 4.18, a v-structure is used to approximate
a house model that is intentionally misaligned with the back and floor
patches. Note that such model cannot be approximated nicely using par-
allel structures, as shown in Fig. 3.20 on page 68. In contrast, the slanted
house model in Fig. 4.18 can be easily approximated using a simple v-
structure.
Fig. 4.19 shows our attempt to create a simple foldable OA that ar-
tistically illustrate a pine tree, which is similar to the card in Fig. 4.1 on
92
F IGURE 4.18: Non-parallel OAs can be used nicely to illustrate input models
that are not aligned to the back and floor bases.
page 69. The input model we use is created in Blender software by plac-
ing three simple tetrahedra, without the need to constrain any specific
angles for them. Our v-structure algorithm then automatically generates
the corresponding patches with appropriate angles so that they are fully
foldable.
F IGURE 4.19: User-defined triangular blocks with arbitrary angles can be easily
converted into a fully foldable OA.
93
F IGURE 4.20: Non-parallel OA may not always be better than parallel OA for
preserving the visual appearance of slanted surfaces.
Note that, while it is relatively easy to create triangular model in 3D

modeling software, such as Maya or Blender, it is not trivial to use the
correct angles on the faces of the model, so that they can be folded com-
pletely when being converted into origamic architecture pop-ups. Our
method solves this issue by adjusting each input angle by a reasonable
amount so that they satisfy the foldability condition. As shown from
Figs. 4.16 to 4.19, the resulting OA structures still approximate the input
94
models closely.
However, v-structures may not always be the optimal choice for

slant-ed surfaces. Fig. 4.20 illustrates a case where parallel structures
might possibly be preferred. In this case, the connectivity of the two tri-
angular blocks are not preserved when v-structures are employed. Since
only a single piece of paper can be used, the v-structures leave a signifi-
cant gap between opposite structures when the OA is opened. In paral-
lel structures, the connecting line between two blocks lies on the central
patches, and hence no unwanted gap appears during opening and clos-
ing.
In this chapter, we have presented a method for designing non-

parallel structures in origamic architecture pop-ups, which is also known
as v-structures. This method is grounded on our formulated foldability
and stability conditions for paths of v-structures. Our simple yet effec-
tive algorithm allows creating v-structures as an extended stage in the
existing pipeline for OA design. In our extended pipeline, the genera-
tion of v-structures comes after the input surface segmentation and be-
fore the generation of parallel structures. By selecting appropriate sur-
face segments for approximation using v-structures, we avoid employ-
ing over-complicated parallel structures for those segments. Addition-
ally, because the v-structure creation is performed in a separate stage, it
does not affect the generated parallel structures later.
Our algorithm first examines the potential surface segments for

v-structure generation, then create those structures so that they satisfy
our foldability condition, while still approximating the input segments
closely. Finally, the generated paths of v-structures are made stable using
double connections.
95
Our stabilizations for parallel structures and for v-structures are

strikingly analogous. In both cases, we successfully utilize double con-
nections to stabilize the paths of structures. For v-structures, double con-
nections are defined based on the angles between the fold lines. For par-
allel structures, double connections are defined based on the distances
between the fold lines.
In this chapter, we have proven that doubly-connected v-structures

are foldable. From our empirical study, they are also stable during the
opening and closing process. However, a formal mathematical proof
for the stability of doubly-connected v-structures has yet been achieved.
Such proof would strengthen our formulation significantly.
In terms of visual appearance, determining whether v-structures or

parallel structures should be used to approximate slanted surfaces with
arbitrary fold lines is a challenging question. If v-structures are used, it
is not trivial to compute the best angles that are both foldable and closely
resemble the non-straight fold lines. In contrast, using parallel structure
will lead to an over-complicated and undesired pop-up, as shown in Fig.
3.20. An optimal solution for this problem may require more elaborate
shape abstraction techniques.
96
Chapter 5
Strengthening Origamic
Architecture Pop-Ups
In the previous chapters, to facilitate our geometric formulation, we have

temporarily assumed that paper has no weight and does not bend. Un-
der that assumption, we were able to construct OA structures that are
geometrically both foldable and stable. The latter means when the user
holds the two outer patches stationary, no other patches may move.
However, the geometric stability in both our work and [4, 68] does
not always hold in practice, when physiscal properties exist. In some
cases, most area of a patch is stable, but a small part of it bends because
it is too long or not well-supported (Fig. 5.1 (a, c)). In other cases, a
horizontal patch may be too big and only supported at the two ends,
causing itself to bend down in the middle (Fig. 5.1 (d)).
The instability of patches may not always be because of their weight.

In many cases, they may bend due to external forces during the making
and holding process (Fig. 5.1 (b)). Hence, such impact also needs to be
taken into account in our structural strength analysis.
In computer graphics, much research has been done on the visual-

ization of thin material deformation, notably cloth. A few studies of pa-
per modeling were done based on developable surfaces [12, 57]. These
97
Chapter 5 Strengthening Origamic Architecture Pop-Ups
(a) (b)
(c) (d)
F IGURE 5.1: Bendings in OA structures may occur due to many reasons: (a)
Gravity on a long part. (b) External forces during folding and storing. (c) Not
being well-supported. (d) Big size.
approaches were mainly for visualization and may not correspond to the
actual physical properties of paper, such as mass density and bending
stiffness. More recent approaches took into account paper properties,
and solved the mesh deformation using Finite Element Method (FEM)
[81]. However, this numerical method is computationally expensive. It
also involves various types of physical discretization; and hence, is not
easy to implement or be readily embed into our design system.
As paper is a thin material, we consider our structural strength

analysis a subset of plate analysis, a well-known field in mechanical en-
gineering. In this field, Kirchoff-Love’s governing equation [71] is com-
monly used to compute the plate deflection without the need of carrying
98
out a full three-dimensional stress analysis as for other types of struc-

tures. Although the equation involves differentiation of fourth order,
we can find quite accurate results by using Finite Difference Method
(FDM). It is a simple method for approximating the solutions of differ-
ential equations. FDM is suitable for the scope of our project, as it is very
intuitive and easy to implement.
In this study, it is not our goal to produce a highly accurate defor-

mation model of paper pop-ups. We mainly aim to obtain a simple and
efficient method to examine whether any parts in an OA paper pop-up
may be weak and require fixing. In other words, the method should
work efficiently as a post-processing stage of the OA design pipeline.
Since the current pipeline takes less than 10 seconds for most input mod-
els (section 3.3.2), the method for OA structural strength analysis should
be appropriately lightweight. Moreover, we believe that the FDM-based
method will be highly reproducible for other interested researchers and
developers.
In the following sections, we present our analysis on the physi-

cal strength of paper structures using Kirchhoff-Love theory and FDM.
From the analysis, we can find out the weak parts and correct them by
adding extra supports. This physical analysis is the first of its kind in
paper pop-up research.
5.1 Formulations
5.1.1 Governing Equation of Plate Bending
Before analyzing a plate, we need to know how flexible it is. In me-

chanical engineering, it is called the bending stiffness of the plate, and is
computed as
99
Eh3
D= (5.1)
12(1 − ν 2 )
where h is the plate thickness, E is the Young’s modulus and ν is the

Poisson’s ratio of the material [21]. Young’s modulus measures the stiff-
ness of a material. Poisson’s ratio indicates how a material tends to com-
press (or expand) in one direction when being expanded (or compressed)
in the other two directions perpendicular to it.
Kirchhoff-Love theory treats a plate originally as a grid of points on

a 2D plane, and computes the deflection at each point on the plate. Let
us assume the plate lies in the xy-plane and w(x, y) is the transverse de-
flection of point (x, y). The theory computes w(x, y) using the following
governing differential equation.
∂ 4w ∂ 4w ∂ 4w pz (x, y)
+ 2 + = (5.2)
∂x4 ∂x2 ∂y 2 ∂y 4 D
where pz (x, y) is the external lateral load at each point.
By using the two-dimensional Laplacian operator
2 ∂ 2 (•) ∂ 2 (•)
∇ (•) = + (5.3)
∂x2 ∂y 2
we can rewrite Eq. (5.2) in a more condensed form as
pz (x, y)
∇2 ∇2 w(x, y) = (5.4)
D
In mechanical engineering, when a plate bends significantly, more

terms are involved to compute an accurate deformation. However, in
our context of paper pop-up, our goal is not to obtain such accurate de-
formation, but to achieve an approximate amount of bending that may
100
occur with each patch of paper. From that knowledge, we can determine
the patches that may be weak and require more support.
5.1.2 FDM-based Numerical Solution
When a plate has primitive shape, such as round or square, it is possible

to find the exact solutions for Eq. 5.4. However, if it has an arbitrary
shape, an analytical solution is hard to achieve.
Hence, numerical approaches have been employed to much suc-

cess. Commonly used methods include Finited difference method (FDM),
Finite element method (FEM), Grid-Work Method (GWM) and Bound-
ary Element Method (BEM), to name a few [96].
We choose FDM as its formulations are well-developed and en-

tirely transparent. Unlike FEM, which employs various types of phys-
ical discretization, FDM only involves simple mathematical discretiza-
tion. The number of equations is small, thus they can be conveniently
implemented and embedded into our system.
In general, FDM approximates the derivatives using simple lin-

ear equations. Consider function y = f (x), where x are discrete points
placed equally at the interval ∆x = xi+1 − xi . The derivatives of y at
point k can be approximated as follows.

dy 1
≈ (yk+1 − yk−1 )
dx k 2∆x
2
dy 1
≈ (yk+1 − 2yk + yk−1 ) (5.5)
dx k (∆x)2
2
4
dy 1
≈ (yk+2 − 4yk+1 + 6yk − 4yk−1 + yk−2 )
dx4 k (∆x)4
We can also apply FDM on a bivariate function, as in the case of
101
F IGURE 5.2: The setup of points for computing of the plate deflection at (m, n).
plate analysis. We assume the points on the plate are equally spaced and
∆x = ∆y = δ (Fig. 5.2). Then the terms on the left-hand side of Eq. 5.2 at
point (m, n) can be approximated using the following equations.
∂ 4w

1
≈ (wm+2,n − 4wm+1,n + 6wm,n − 4wm−1,n + wm−2,n )
∂x4 m,n δ4
(5.6)
∂ 4w

1
≈ 4 (wm,n+2 − 4wm,n+1 + 6wm,n − 4wm,n−1 + wm,n−2 )
∂y 4 m,n δ
and
∂ 4w ∆2
2
∆w
≈
∂x2 ∂y 2 m,n (∆y) (∆x)2 m,n
2
( )
∆2 w
2 2
1 ∆w ∆w
= −2 +
(∆y)2 (∆x)2 m,n+1 (∆x)2 m,n (∆x)2 m,n−1
1
= (4wm,n − 2(wm+1,n + wm−1,n + wm,n+1 + wm,n−1 )
δ4
+ wm+1,n+1 + wm+1,n−1 + wm−1,n+1 + wm−1,n−1 )
(5.7)
The points involved in approximating the derivatives of w(m,n) are

shown in Fig. 5.2. From Eq. 5.6 and 5.7, we can rewrite Eq. 5.4 into the
102
following form.
pz (m, n)
= 20wm,n − 8(wm+1,n + wm−1,n + wm,n+1 + wm,n−1 )
D
+ 2(wm+1,n+1 + wm−1,n+1 + wm+1,n−1 + wm−1,n−1 ) (5.8)
+ wm+2,n + wm−2,n + wm,n+2 + wm,n−2 ]
This is the main governing equation we use for computing the

bending at every point on the paper plate. When we combine these
equations for all the points into a linear system, we obtain a very sparse
matrix. If the dimension of the plate is M × N , the left-hand side ma-
trix of the system will have dimension M N × M N , but each row only
contains at most 13 nonzero coefficients. Hence, we can effectively use
available linear system solvers for sparse matrix in our implementation.
For convenient discussion, we also use the following grid format to

denote the full form of the governing equation in an intuitive way.
+1
+2 -8 +2
+1 -8 +20 -8 +1
+2 -8 +2
+1
5.1.3 Boundary Conditions
A solution of the discrete governing equation 5.8 must simultaneously
103
satisfy the equation and the boundary conditions of the plate. Each equa-
tion involves 13 points, but these points do not always physically exist.
Along the plate boundaries, we must introduce fictitious points outside
the plate and express them in terms of the existing ones.
In the scope of paper pop-ups, we consider two types of bound-

aries. A fixed edge is the fold line between two patches, and a free edge is
along the contour of a cut line. Note that in general structures, we may
also have other types of boundaries, such as simple supporting edge,
where a plate is placed on top of a supporting point but not fixed to that
point.
For each type of boundaries, we need to consider fictitious points

outside the plate at distance δ and 2δ from the edge, respectively. As-
sume the considered edge is parallel to the y-axis, and on the right-hand
side of the plate.
On a fixed edge, the bending and its gradient are both zero. Hence,
we have the following equations.
wm,n = 0

∂w

1 (5.9)
≈ (wm+1,n − wm−1,n ) = 0
∂x m,n 2δ
which give the boundary conditions for a fixed edge.
wm+1,n = wm−1,n
(5.10)
wm+2,n = wm−2,n
104
On a free edge, the force and moment about the edge axis have
been proven to be zero [96]. They are expressed as
∂ 2w ∂ 2w

(my )m,n = + ν =0
∂y 2 ∂x2 m,n
(5.11)
∂ 3w 3

∂ w
(vy )m,n = + (2 − ν) =0
∂y 3 ∂y∂x2 m,n
Again, using FDM to approximate the derivatives, we can trans-

form Eq. 5.11 into
(my )m,n ≈ −(2 + 2ν)wm,n + wm,n+1 + wm,n−1 + ν(wm+1,n + wm−1,n ) = 0
(vy )m,n ≈ (6 − 2ν)(wm,n−1 − wm,n+1 )
+ (2 − ν)(wm+1,n+1 + wm−1,w+1 − wm−1,n−1 − wm+1,n−1 )
− wm,n−2 + wm,n+2 = 0
(5.12)
From these approximations, we obtain the following linear bound-

ary conditions for a free edge.
wm+1,n = (2 + 2ν)wm,n − wm−1,n − ν(wm,n−1 + wm,n+1 )
wm+2,n = (−6ν 2 + 12ν + 12)wm,n + (4ν 2 − 8ν − 4)(wm,n−1 + wm,n+1 )
+ (4ν − 12)wm−1,n + (−2ν + 4)(wm−1,n−1 + wm−1,n+1 )
+ (−ν 2 + 2ν)(wm,n−2 + wm,n+2 ) + wm−2,n

(5.13)
From the boundary conditions in Eq. 5.10 and 5.13, we are able to
construct the bending equations for every point on the plate, including
those along the boundaries.
An arbitrary point may lie on a fixed edge, or a free edge, or it

may not lie near any edge. We categorize the position of a point into the
105
following 11 possible cases.
0. (m, n) has at least distance 2δ from all the edges. In this case, all
the involved points in Eq. 5.8 are available, and we do not need to
compute the boundary conditions.
1. (m, n) is on a fixed edge.
2. (m, n) is at distance δ from a fixed edge, and is not near any other.
3. (m, n) is at distance δ from two fixed edges.
4. (m, n) is on a free edge, and is not near any other edge.
5. (m, n) is at distance δ from a free edge, and is not near any other.
6. (m, n) is at a corner of two free edges.
7. (m, n) is on a free edge, and at distance δ from another free edge
8. (m, n) is at distance δ from two free edges.
9. (m, n) is on a free edge, and at distance δ from a fixed edge
10. (m, n) is at distance δ from a free edge, and δ from a fixed edge.
An illustration of the points that fall into each of these 11 cases can
be seen in the following grid.
1 1 1 1 1
1 3 2 10 9
1 2 0 5 4
1 10 5 8 7
δ
1 9 4 7 6
δ
106
Besides case 0, in which we have the complete governing equation

(Eq. 5.8), we need to formulate the equations for the other 10 boundary
cases. By using suitable conditions for the fictitious points in each of
these cases, we achieve valid equations for all the actual points on the
plate. The resulting coefficients of the actual points in each case are as
follows.
Case 1: Point (m, n) is on a fixed edge.
The bending at (m, n) is wm,n = 0. Hence, no further governing

equation is needed.
Case 2: Point (m, n) is at distance δ from a fixed edge but is not near any
other edge.
Without loss of generality, we can assume the fixed edge is on the

left side of the considered point. From the boundary conditions in Eq.
5.10, we obtain the following governing equation, illustrated in grid form.
Note that the bending of points along the fixed edge is zero. Thus, we
can ignore their coefficients in the equation.
+1
-8 +2
+21 -8 +1
-8 +2
+1
Case 3: Point (m, n) is at distance δ from two fixed edges.
107
We assume the fixed edges are above and on the left side of the
considered point. From Eq. 5.10, we obtain the following governing
equation.
+22 -8 +1
-8 +2
+1
Case 4: Point (m, n) is on a free edge but is not near any other edge.
Assume the considered point is at the bottom of the patch. From

Eq. 5.13, we have the following governing equation.
+2
−2ν + 4 4ν − 12 −2ν + 4
−ν 2 + 1 4ν 2 + 4ν − 8 −6ν 2 − 8ν + 16 4ν 2 + 4ν − 8 −ν 2 + 1
Case 5: Point (m, n) is at distance δ from a free edge but is not near any
other edge.
Assume the free edge is below the considered point. From Eq. 5.13,
the governing equation becomes
108
+1
+2 -8 +2
+1 -8 +19 -8 +1
−ν + 2 2ν − 6 −ν + 2
Case 6: Point (m, n) is at the corner of two free edges.
Assume the considered point is at the bottom right corner of the

patch. The governing equation becomes
−2ν 2 + 2
−8ν + 8 4ν 2 + 8ν − 12
−2ν 2 + 2 4ν 2 + 8ν − 12 −4ν 2 − 8ν + 12
Case 7: Point (m, n) is on a free edge and at distance δ from another free
edge.
Assume the considered point is on the right edge and near the bot-
tom edge of the patch. The governing equation becomes
109
−ν 2 + 1
−2ν + 4 4ν 2 + 4ν − 8
+2 4ν − 12 −5ν 2 − 8ν + 15
−2ν + 4 2ν 2 + 4ν − 6
Case 8: Point (m, n) is at distance δ from two free edges.
Assume the considered point is near the bottom right corner of the
+1
+2 -8 −ν + 2
+1 -8 +18 2ν − 6
−ν + 2 2ν − 6 −2ν + 2
Case 9: Point (m, n) is on a free edge and at distance δ from a fixed edge.
Assume the considered point is near the bottom left corner of the
patch. We combine the boundary conditions for both free edge and fixed
edge (Eqs. 5.10 and 5.13), which lead to
110
+2
−2ν + 4 4ν − 12 −2ν + 4
4ν 2 + 4ν − 8 −7ν 2 − 8ν + 17 4ν 2 + 4ν − 8 −ν 2 + 1
Case 10: The pivotal point is at distance δ from both a free edge and a
fixed edge.
Assume the considered point is near the bottom left corner of the
+1
+2 -8 +2
-8 +20 -8 +1
−ν + 2 2ν − 6 −ν + 2
By setting up the equations for all the possible cases of the grid
points, we can easily implement the bending of paper patches in our
system. Further implementation details are described in the next section.
111
5.2 Implementation
5.2.1 Bending Approximation for Paper Structures
We assume the generated OA plan has the same dimension as an

A4 sheet. For each patch, we set up a mesh of grid to completely cover
the patch (Fig. 5.3). We place the grid points regularly at distance δ =
∆x = ∆y = 5×10−3 m. The physical properties of paper used in our im-
plementation are obtained from the previous literatures. We use card
stock density 250g/m2 [3], bending stiffness D = 5×10−2 N m [1] and
Poisson’s ratio ν = 0.3 [90]. In the governing bending equation (Eq.
5.8 on page 103), the unit for the load at each point is N ewton (N ), and
that for the bending amount is metre (m). We use CSparse library [25]
to solve our sparse linear system. All the experimented bendings were
computed in real time.
F IGURE 5.3: The mesh of grids is set up to completely cover the shape of the
patch.
Currently we only take into account the gravity acting perpendicu-

lar to the paper patches. To examine both vertical and horizontal patches,
we consider the forces applied on them when the pop-up stands on its
floor patch, and when it stands on its back patch. The mass of each patch
112
is distributed uniformly at every grid point. In the future, we may ex-

amine other types of forces, such as rotational forces during the closing
and opening process.
5.2.2 Weak Patch Detection and Correction
From the gravity, we compute the possible bending of each patch when
the pop-up lies on its floor patch, and when it lies on its back patch. For
a pop-up made from an A4 piece of paper, we threshold the bending
amount for a weak patch to be at least 0.5cm. If all the points on a patch
move less than this amount, it is still considered physically strong.
We also assume that the deflections of the patches are independent.

For example, a patch does not bend significantly enough to touch other
patches that are originally not adjacent to it. Under this assumption,
we iteratively detect the weakest patch, correct it and update the whole
structure. The process stops when no more correction is required or can
be done.
The correction is achieved by connecting the considered weak part

to a strong patch that is connectable to it. This process is similar to the
patch connection in Section 3.2.3.2 on page 52, in which we look for
the new connection with lowest cost. By doing this, we minimize the
changes in the structure when correcting the patches.
As the OA was originally generated with only patches connectable

to each other, we can always find a new connection to strengthen a weak
patch. Although a global correction method for optimal visual and nu-
merical results is not yet available, our current greedy approach pro-
duces acceptable solutions for the tested models, as shown in the next
section.
113
5.3 Results
5.3.1 Comparison with Analytical Solutions
We examine our FDM-based governing equation (Eq. 5.8) by per-

forming it on simple rectangular patches. We then compare our solu-
tions with the analytical solutions for the original differential equation
(Eq. 5.4). Figs. 5.4 and 5.5 show the visualizations of the computed
bending. In both solutions, we use a patch of size 0.2m × 0.4m, with the
grid points regularly placed at distance ∆x = ∆y = 10−2 m. The physical
properties of the patch are as described in section 5.2.1.
F IGURE 5.4: The computed bending of a patch with two fixed edges and two
free edges. Left: Analytical solution of the governing differential equation.
Right: Our solution of the FDM-based governing equation.
F IGURE 5.5: The computed bending of a patch with one fixed edges and three
free edges. Left: Analytical solution of the governing differential equation.
Right: Our solution of the FDM-based governing equation.
114
In the first case, the patch has two fixed edges and two free edges
(Fig. 5.4). In the second case, the patch has one fixed edge and three free
edges (Fig. 5.5). We compute the average percentage difference from all
the grid points of our FDM-based solutions, as compared to the analyti-
cal solutions. For the first patch, the percentage difference is 18.22%. For
the second patch, the percentage difference is 10.84%. In both cases, the
amount of bending computed from the FDM-based solutions is slightly
smaller than that from the analytic solutions. However, the positions of
the weakest areas are identical in the FDM-based solution and the ana-
lytical solution.
The maximal amount of bending of the first patch is 0.36cm when

using the FDM-based approach, and 0.53cm when using the analyti-
cal approach. For the second patch, the maximal bending amount for
the second patch is 0.64cm using the FDM-based approach, and 0.82cm
(a) (b)
(c) (d)
F IGURE 5.6: The weak structures in Fig. 5.1 after being corrected.
115
(a) (b)
F IGURE 5.7: A patch in the bunny OA is not well-supported (a), and is corrected
by extending (b).
when using the analytical approach. This patch is likely to be weak and
require extra support along its weakest edge.
5.3.2 OA Structural Strength Analysis and Correction
Although we did not consider physical properties of paper in the

previous chapters, our investigation indicates that most of the generated
OAs are sufficiently strong. It may be because most patches are sup-
ported by at least two other patches, and the input models we use are
mostly structurally strong in real life.
Nevertheless, some of the OAs we create contain weak parts that

need to be fixed. Figs. 5.1, 5.8, 5.9 and 5.10 show a number of such cases
and the corrected designs.
116
(a) (b)
F IGURE 5.8: The trunk of the elephant OA bends down due to the heavy weight
(a), and is corrected (b).
F IGURE 5.9: The cross in the Capitol OA is strengthened.
Note that, in order to make a patch physically strong, we may have

to alter its originally desired shape. An example is the trunk of the ele-
phant OA, which has to be widen and connected to the floor patch. An
altenative solution in such a situation may be to give the user an option
to leave the structure as designed. We can then provide a list of paper
117
F IGURE 5.10: The cross in the Taj Mahal OA is strengthened.
materials in the system, and allow the user to choose a stronger mate-
rial for weak pop-ups. In order to do so, more studies on the physical
properties of other types of paper will be required.
In this chapter, we have presented a physical strength analysis for

paper structures using Kirchhoff-Love theory and FDM discretization.
The strength and weakness of each patch is determined by the amount
of its bending under gravity when the pop-up lies on the floor patch,
and when it lies on the back patch. With its simplicity, the approach is
easy to implement, yet effective. Preliminary experiments also show the
potential of this approach in analyzing arbitrary structures made of thin
materials.
Similar to our stabilization technique in Chapter 3, our physical

str-ength analysis can also be readily embeded into other systems for
designing paper pop-up, or other types of paper craft. It may also be
considered for interactive bending in virtual reality. For instance, a piece
118
of paper may bend when being pressed or waved by a virtual character.

Such object that can respond interactively in a physically-correct manner
will be very useful for educational purposes.
Limitations
From our observation in creating pop-ups, the patches seldom bend

under longitudinal forces. This may be due to high longitudinal bending
stiffness of the card stock material often used for paper pop-ups. Theo-
retically, longitudinal stiffness is also significantly higher than transverse
stiffness that we consider in this work [96]. Thus, to keep the formulation
simple, we have yet taken into account longitudinal forces in our work.
Nevertheless, when dealing with long thin patches, or more flexible ma-
terials (Fig. 5.11), we will need a more elaborate governing equation.
F IGURE 5.11: A long thin patch made from a flexible material is bent due to the
weight acting along its longitudial axis.
Our current method detects and corrects one weak patch at a time
until all the patches are strong. However, such approach may not be ef-
fective for complex structures. For instance, there may be weak patches
that can be corrected altogether using a single new connection. Our
119
method may need to add several connections, and eventually modify

the structures more than necessary. To deal with this problem, a global
method for structural correction will be needed.
120
Chapter 6
Conclusion and Future

Directions
6.1 Conclusion
Origamic Architecture (OA) not only is a paper art form but also
has pratical applications, such as in nano and micro fabrication [30, 47,
107]. As a special type of paper pop-ups, OA has the beauty of using
only a single piece of paper, yet inheriting the ability to resemble many
structures and daily objects.
Despite its popularity, OA creation requires considerable time and

skills. Designing the 2D layout of an OA pop-up is already challenging
itself, because it requires careful considerations in both geometric and
physical aspects. The design has to pop up fully into a desired shape,
while being stable at each opening angle. In addition, it has to be physi-
cally strong.
Existing works on computer-aided and automatic OA design are

still very limited. In commonly used voxel-based methods, the design
process requires high resolution of voxel grid to approximate the input
3D model closely. As a results, the number of cuts and folds caused by
the small voxels is significantly high and the voxel-based designs are
121
Chapter 6 Conclusion and Future Directions
hard to use in practice. Even so, the voxels are not able to preserve the
meaningful contours on the input surface.
In addition, previous studies only defined a very narrow set of sta-

bility conditions for OA structures. As we have examined, those con-
ditions ignore many commonly seen structures, and limit the possible
designs significantly.
In this thesis, we have presented a set of geometric formulations

and a novel algorithm for automatic OA design. Our comprehensive
foldability and stability conditions allow us to utilize an image-based
slicing approach for artistically abstracting the input model. We are able
to generate foldable and stable pop-up structures that were previously
excluded by other algorithms. Visual and quantitative comparisons of
results have shown that our algorithm is significantly better than the ex-
isting methods in the preservation of contours, surfaces and volume, as
well as the ease of actual creation. Our designs have also been shown to
resemble those created by real artists.
In addition to the geometric foldability and stability conditions, we

formulate a set of linear equations for analyzing the physical strength
of the generated OA structures. We utilize Finite Difference Method to
discretize Kirchhoff-Love’s differential plate equations. By solving the
obtained sparse linear system, we simulate the possible bendings in a
paper structure in real time. The weak parts can be corrected by adding
new supporting connections. Our physical analysis and fixing method
is the first of its kind in paper pop-up studies.
All the approaches presented in this thesis, including our novel slic-
ing method, foldability check, stabilization, and physical strength analy-
sis can be easily integrated into other design systems, such as [36, 44], to
122
name a few. They can help to reduce the manual work for the user, while
keeping the pop-ups valid.
6.2 Future Directions
Our study is part of computational design, a research area that is be-

coming more and more active. In this area, our work can be considered
artistic design, as its goal is to generate results that resemble artists’ cre-
ations. Another aspect that we plan to study is functional design, in which
the most important goal is to design objects that function according to
user requirements. We aim to continue our research in these two aspects
of computational design.
6.2.1 Artistic Design
Paper Pop-Up Our study offers interesting possibilities for future

research in general paper pop-up. Theoretically, we have proven the suf-
ficient and necessary condition for foldable OA parallel structures and
v-structures, and the sufficient condition for stable parallel structures.
However, it is still unknown whether a necessary condition for stable
parallel structures is achievable. In addition, the stability condition for v-
structures has only been studied empirically, but not been proven math-
ematically. If a formal proof can be obtained, we will have a stronger
theoretical foundation to support a unified framework for both parallel
and v-structures.
While origamic architecture only allows a single piece of paper, we

may convert it to a general pop-up by adding multiple pieces of paper
to preserve the concave surfaces more easily. To do so, we will need to
capture a complete shape of the input 3D model by extending our single-
view image-based abstraction to multiple views. We have achieve some
preliminary results for multi-piece paper pop-ups in [88], [89] (Fig. 6.1).
123
F IGURE 6.1: The multi-piece paper pop-ups produced by our automatic design
systems presented in [88] and [89].
124
Currently, our abstraction requires as input a 3D model. However,

such data may not always be available to all users. We plan to extend our
work to allow other types of input representations, for instance draw-
ings and photographs. An OA design system using such inputs will be
exciting, and also requires single-view reconstruction techniques. One
notable work on single-view pop-up generation is [46]. However, its
type of pop-up is still very simple.
Other Paper Art Forms Besides paper pop-up, there are other forms
of paper arts such as origami (paper folding) and paper sculpting [22].
For origamic design, although there have been numerous mathematical
studies, most of them do not propose an automatic approach, or only
generate a complex folding pattern that requires very good skills [97].
For origami learners and beginners, multi-piece origami is a more feasi-
ble choice, in which two or more sheets of paper are folded into origamic
structures and locked together to form a desired 3D shape. With our ex-
perience in origamic architecture and multi-style paper pop-up [88], we
believe it is possible to develop an algorithm that searches for a combi-
nation of folding patterns to form a multi-piece origami.
Aside from origami, paper sculpting has also been studied recently
[72]. However, existing algorithms can only use printed textures to de-
pict subtle details like hair and clothes. In practice, with the flexibility of
paper, those details can be abstracted very lively [22] (Fig. 6.2 (a)). We
believe an automatic design of detailed and artistic paper sculptures can
be achieved by utilizing a multi-view image-domain abstraction method,
and studying the effects of physical paper bending for representing dif-
ferent shapes.
6.2.2 Functional Design
Computational design of daily items is becoming more and more
125
a (b)
F IGURE 6.2: (a) A paper sculpture designed by [22]. (b) Foldable and compact
furniture can be designed automatically and 3D-printed in the future.
feasible, especially with the increasing availability of 3D scanning tech-

nologies. We are also interested in the automatic design of household
and office items, like furniture, that function according to user require-
ments. The items will be able to perform user-defined mechanical tasks,
have a desired appearance and fit in a given space. With the spatial con-
straints in houses and offices nowadays, it is useful to have foldable and
portable items. Similar recent studies, such as [91], do not consider the
foldability of the items or their physical balance when the constituent
parts are moving. Our earlier geometric formulation for the foldabil-
ity and stability treats each patch of paper as a rigid plane, and hence,
may share similarities with the corresponding geometric study of fold-
able furniture. However, we will need to take into account the thick-
ness of each part of the furniture, which was not an issue in paper struc-
tures. An automatic design system for functional, physically feasible and
portable furniture will be of much interest and may open a new horizon
for product design (Fig. 6.2 (b)).
Another research direction that requires more elaborate investiga-

tion is the computational design of soft and elastic objects, such as clothes.
Computational fashion design has not become an active research area,
126
but will definitely attract much attention. Although cloth simulation

has been studied extensively, designing how pieces of cloths can be cut,
pleated and sewed together to make nice garments is not easy for most
people. It is even more challenging if the garments need to fit the body
measurements of a certain person. In this research direction, we can cre-
ate a system that designs stylish clothes for a person simply from his or
her appearance and choices. Such system will require reconstructing hu-
man body shape and pose from images or simple measurements, which
we have attempted in an earlier project [63, 64]. Moreover, in order to
determine the patterns that look good on a person for a specific activity
or occasion, it is important to understand the psychological choices in
fashion design. This study may benefit from our experience in observ-
ing artists choices for paper craft design.
127
Bibliography
[1] Pulp & paper resources site. http://www.paperonweb.com/,

2010.
[2] Pop-up card books. http://www.evermore.com/oa/books.php3,

2011.
[3] Business card stock. http://standardbusinesscardsize.net/business-

card-stock-explained/, 2012.
[4] Z. Abel, E. D. Demaine, M. L. Demaine, S. Eisenstat, A. Lubiw,

A. Schulz, D. L. Souvaine, G. Viglietta, and A. Winslow. Univer-
sality results for pop-up cards. Preprint, 2012.
[5] E. Agathos, I. Pratikakis, S. Perantonis, N. Sapidis, and P. Azari-

adis. 3d mesh segmentation methodologies for cad applications.
Computer-Aided Design and Applications, 4(6):827–842, 2007.
[6] G. Allaire and F. Jouve. Minimum stress optimal design with the
level set method. Engineering Analysis with Boundary Elements, 32
(11):909918, 2008.
[7] D. Baraff and A. Witkin. Large steps in cloth simulation. In Proceed-

ings of the 25th annual conference on Computer graphics and interactive
techniques, SIGGRAPH ’98, pages 43–54, New York, NY, USA, 1998.
ACM.
129
Bibliography
[8] C. Barton. The Pocket Paper Engineer: How to Make Pop-Ups Step-by-
Step, volume Two volumes. Popular Kinetics Press, 2005-2008.
[9] M. Bataille. ABC3D. Roaring Brook Press, 2008.
[10] D. Birmingham. Pop-Up Design and Paper Mechanics: How to Make

Folding Paper Sculpture. Guild of Master Craftsman, 2011.
[11] P. Block, T. Ciblac, and J. Ochsendorf. Real-time limit analysis of

vaulted masonry buildings. Computers & Structures, 84(2930):1841–
1852, 2006.
[12] P. Bo and W. Wang. Geodesic-controlled developable surfaces for

modeling paper bending. In Computer Graphics Forum, volume 26,
pages 365–374. Wiley Online Library, 2007.
[13] J. Canny. A computational approach to edge detection. IEEE Trans.

Pattern Anal. Mach. Intell., 8(6):679–698, June 1986.
[14] L. Carroll and R. Sabuda. Alice’s Adventures in Wonderland: A Pop-

up Adaptation. Little Simon, 2003.
[15] D. A. Carter. One Red Dot: A Pop-Up Book for Children of All Ages.
Little Simon, 2005.
[16] D. A. Carter and J. Diaz. Elements of Pop-Up. Little Simon, 1999.
[17] M. Chatani. Origamic Architecture of Masahiro Chatani. Shokokusya,

Tokyo, 1984.
[18] M. Chatani. Pop-Up Origamic Architecture. Ondorisha Publications,

1985.
[19] M. Chen and K. Tang. A fully geometric approach for developable

cloth deformation simulation. The Visual Computer, 26(6-8):853–
863, 2010.
130
Bibliography
[20] K.-J. Choi and H.-S. Ko. Stable but responsive cloth. In ACM SIG-
GRAPH 2005 Courses, SIGGRAPH ’05, New York, NY, USA, 2005.
ACM.
[21] P. C. Chou. Elasticity: Tensor, Dyadic, and Engineering Approaches

(Dover Civil and Mechanical Engineering), volume 1. Dover Publica-
tions, 1992.
[22] S. Christopher. Paper Sculptor.

http://www.sherchristopher.com/, 2013.
[23] J. Cohen, M. Olano, and D. Manocha. Appearance-preserving sim-

plification. In Proceedings of the 25th annual conference on Computer
graphics and interactive techniques, SIGGRAPH ’98, pages 115–122,
New York, NY, USA, 1998. ACM.
[24] CoppeliaRobotics. V-REP.

http://www.coppeliarobotics.com/, 2013.
[25] T. A. Davis. Direct Methods for Sparse Linear Systems (Fundamen-

tals of Algorithms). Society for Industrial and Applied Mathematic,
2006.
[26] E. Demaine and J. O’Rourke. Geometric Folding Algorithms: Link-

ages, Origami, Polyhedra. Cambridge University Press, 2007.
[27] E. Eisemann, S. Paris, and F. Durand. A visibility algorithm for

converting 3d meshes into editable 2d vector graphics. In ACM
SIGGRAPH 2009 papers, SIGGRAPH ’09, pages 83:1–83:8, New
York, NY, USA, 2009. ACM.
[28] E. English and R. Bridson. Animating developable surfaces using

nonconforming elements. In ACM SIGGRAPH 2008 papers, SIG-
GRAPH ’08, pages 66:1–66:5, New York, NY, USA, 2008. ACM.
131
Bibliography
[29] J. Feldman and M. Singh. Information along contours and object

boundaries. Psych. Review, 112:243–252, 2005.
[30] S. M. Felton, M. T. Tolley, B. Shin, C. D. Onal, E. D. Demaine,

D. Rus, and R. J. Wood. Self-folding with shape memory com-
posites. Soft Matter, 9:7688–7694, 2013.
[31] C. R. Feynman. Modeling the appearance of cloth. PhD thesis, Mas-

sachusetts Institute of Technology, 1986.
[32] T. Fuse. Unit Polyhedron Origami. Japan Publications Trading, 2006.
[33] M. Garland and P. S. Heckbert. Surface simplification using

quadric error metrics. In Proceedings of the 24th annual conference on
Computer graphics and interactive techniques, SIGGRAPH ’97, pages
209–216, New York, NY, USA, 1997. ACM Press/Addison-Wesley
Publishing Co.
[34] M. Garrido and I. Siliakus. The Paper Architect: Fold-It-Yourself

Buildings and Structures. Potter Craft, 2009.
[35] M. Gilbert. Ring: A 2d rigid-block analysis program for masonry

arch bridges. In Third International Arch. Bridges Conference, pages
459–464, 2001.
[36] A. Glassner. Interactive pop-up card design, part 1. In Computer

Graphics and Application, volume 22, pages 79–86, 2002.
[37] A. Glassner. Interactive pop-up card design, part 2. In Computer

Graphics and Application, volume 22, pages 74–85, 2002.
[38] R. Goldenthal, D. Harmon, R. Fattal, M. Bercovier, and E. Grin-

spun. Efficient simulation of inextensible cloth. In ACM SIG-
GRAPH 2007 papers, SIGGRAPH ’07, New York, NY, USA, 2007.
ACM.
132
Bibliography
[39] R. C. Gonzalez and R. E. Woods. Digital Image Processing. Prentice

Hall, 2 edition, 2002.
[40] D. Grant. The Business of Being an Artist. Allworth Press, 2010.
[41] T. Hara and K. Sugihara. Computer-aided design of pop-up books

with two- dimensional v-fold structures. In In Proc. 7th Japan Con-
ference on Computational Geometry and Graphs, 2009.
[42] J. C. Hart, B. Baker, and J. Michaelraj. Structural simulation of tree

growth and response. The Visual Computer, 19:151–163, 2003.
[43] J. Haslinger and R. A. E. Makinen. Introduction to Shape Optimiza-

tion: Theory, Approximation, and Computation. Society for Industrial
and Applied Mathematics, 2003.
[44] S. L. Hendrix and M. A. Eisenberg. Computer-assisted pop-up

design for children: computationally enriched paper engineering.
Adv. Technol. Learn., 3(2):119–127, April 2006.
[45] M. Hiner. Paper Engineering for Pop-Up Books and Cards. Parkwest
Pubns, 1986.
[46] D. Hoiem, A. A. Efros, and M. Hebert. Automatic photo pop-up.

In ACM SIGGRAPH 2005 Papers, pages 577–584, 2005.
[47] E. E. Hui, R. T. Howe, and M. S. Rodgers. Single-step assembly of

complex 3-d microstructures. In The Thirteenth Annual International
Conference on Micro Electro Mechanical Systems, pages 602–607, 2000.
[48] T. Hull. On the mathematics of flat origamis. Congressus Numeran-

tium, 100:215–224, 1994.
[49] S. Iizuka, Y. Endo, J. Mitani, Y. Kanamori, and Y. Fukui. An inter-

active design system for pop-up cards with a physical simulation.
Vis. Comput., 27(6-8):605–612, June 2011.
133
Bibliography
[50] M. Isenburg, P. Lindstrom, S. Gumhold, and J. Snoeyink. Large

mesh simplification using processing sequences. In Proceedings of
the 14th IEEE Visualization 2003 (VIS’03), VIS ’03, pages 61–, Wash-
ington, DC, USA, 2003. IEEE Computer Society.
[51] P. Jackson. The Pop-Up Book: Step-by-Step Instructions for Creating

Over 100 Original Paper Projects. Holt Paperbacks, 1993.
[52] C. Jirasek, P. Prusinkiewicz, and B. Moulia. Integrating biome-

chanics into developmental plant models expressed using l-
systems. In Plant biomechanics 2000, Proceedings of the 3rd Plant
Biomechanics Conference, pages 615–624. Georg Thieme Verlag,
2000.
[53] J. M. Kaldor, D. L. James, and S. Marschner. Efficient yarn-based

cloth with adaptive contact linearization. In ACM SIGGRAPH 2010
papers, SIGGRAPH ’10, pages 105:1–105:10, New York, NY, USA,
2010. ACM.
[54] E. Kalogerakis, A. Hertzmann, and K. Singh. Learning 3d mesh

segmentation and labeling. In ACM SIGGRAPH 2010 papers, SIG-
GRAPH ’10, pages 102:1–102:12, New York, NY, USA, 2010. ACM.
[55] L. Kavan, D. Gerszewski, A. W. Bargteil, and P.-P. Sloan. Physics-

inspired upsampling for cloth simulation in games. In ACM SIG-
GRAPH 2011 papers, SIGGRAPH ’11, pages 93:1–93:10, New York,
NY, USA, 2011. ACM.
[56] E. Kenneway. Complete Origami: An A-Z of Facts and Folds, with

Step-by-Step Instructions for Over 100 Projects. St. Martin’s Press,
1987.
134
Bibliography
[57] M. Kilian, S. Flöry, Z. Chen, N. J. Mitra, A. Sheffer, and

H. Pottmann. Developable surfaces with curved creases. In Ad-
vances in Architectural Geometry, pages 33–36, 2008.
[58] J. Koenderink. Solid shape. MIT Press, 1990.
[59] Y.-K. Lai, Q.-Y. Zhou, S.-M. Hu, and R. R. Martin. Feature sensitive
mesh segmentation. In Proceedings of the 2006 ACM symposium on
Solid and physical modeling, SPM ’06, pages 17–25, New York, NY,
USA, 2006. ACM.
[60] Y.-K. Lai, S.-M. Hu, R. R. Martin, and P. L. Rosin. Rapid and ef-
fective segmentation of 3d models using random walks. Comput.
Aided Geom. Des., 26(6):665–679, August 2009.
[61] R. J. Lang. Origami in Action : Paper Toys That Fly, Flap, Gobble, and
Inflate. St. Martin’s Griffin, 1997.
[62] L. Lapidus and G. F. Pinder. Numerical solution of partial differential

equations in science and engineering. Wiley-Interscience, 2011.
[63] S. N. Le, J. Karlekar, and A. C. Fang. Articulated registration of 3d

human geometry to x-ray image. In Image Processing, 2008. ICIP
2008. 15th IEEE International Conference on, pages 1108–1111, 2008.
[64] S. N. Le, M. K. Lee, S. Banu, and A. C. Fang. Volumetric recon-

struction from multi-energy single-view radiography. In Computer
Vision and Pattern Recognition, 2008. CVPR 2008. IEEE Conference on,
pages 1–8, 2008.
[65] S. N. Le, S.-J. Leow, T.-V. Le-Nguyen, C. Ruiz, and K.-L. Low. Sur-
face and contour-preserving origamic architecture paper pop-ups.
IEEE Transactions on Visualization and Computer Graphics, 20(2):276–
288, February 2014.
135
Bibliography
[66] T.-V. Le, K.-L. Low, C. Ruiz Jr., and S. N. Le. Automatic paper
sliceform design from 3d solid models. IEEE Transactions on Visu-
alization and Computer Graphics, 19(11):1795–1807, November 2013.
[67] Y. Lee, S. Tor, and E. Soo. Mathematical modelling and simulation

of pop-up books. Computers & Graphics, 20(1):21–31, 1996.
[68] X.-Y. Li, C.-H. Shen, S.-S. Huang, T. Ju, and S.-M. Hu. Popup: au-
tomatic paper architectures from 3d models. In ACM SIGGRAPH
2010 papers, SIGGRAPH ’10, pages 1–9, 2010.
[69] X.-Y. Li, T. Ju, Y. Gu, and S.-M. Hu. A geometric study of v-style
pop-ups: theories and algorithms. In ACM SIGGRAPH 2011 pa-
pers, SIGGRAPH ’11, pages 98:1–98:10, New York, NY, USA, 2011.
ACM.
[70] Y. Li, J. Yu, K.-l. Ma, and J. Shi. 3d paper-cut modeling and anima-
tion. Comput. Animat. Virtual Worlds, 18(4-5):395–403, September
2007.
[71] A. E. H. Love. On the small free vibrations and deformations of

elastic shells. Philosophical trans. of the Royal Society (London), serie
A(17):491549, 1888.
[72] F. Massarwi, C. Gotsman, and G. Elber. Papercraft models using

generalized cylinders. In Proceedings of the 15th Pacific Conference
on Computer Graphics and Applications, pages 148–157, 2007.
[73] J. Matas, C. Galambos, and J. Kittler. Robust detection of lines

using the progressive probabilistic hough transform. Comput. Vis.
Image Underst., 78(1):119–137, April 2000.
[74] O. Matias van Kaick and H. Pedrini. A comparative evaluation of

metrics for fast mesh simplification. Computer Graphics Forum, 25
(2):197–210, 2006.
136
Bibliography
[75] J. McCrae, K. Singh, and N. J. Mitra. Slices: a shape-proxy based on

planar sections. ACM Trans. Graph., 30(6):168:1–168:12, December
2011.
[76] R. Mehra, Q. Zhou, J. Long, A. Sheffer, A. Gooch, and N. J. Mitra.

Abstraction of man-made shapes. In ACM SIGGRAPH Asia 2009
papers, pages 137:1–137:10, 2009.
[77] J. Mitani and H. Suzuki. Computer aided design for origamic ar-
chitecture models with polygonal representation. In Proceedings of
the Computer Graphics International, pages 93–99, 2004.
[78] J. Mitani and H. Suzuki. Making papercraft toys from meshes us-
ing strip-based approximate unfolding. ACM Trans. Graph., 23(3):
259–263, August 2004.
[79] A. Montanaro. A concise history of pop-up and movable books.

http://www.libraries.rutgers.edu/rulflibs/scua/montanar/p-
intro.htm, 2005.
[80] M. Mukerji. Marvelous Modular Origami. A K Peters/CRC Press,

2007.
[81] R. Narain, T. Pfaff, and J. F. O’Brien. Folding and crumpling adap-

tive sheets. ACM Transactions on Graphics, 32(4):xxx:1–8, July 2013.
Proceedings of ACM SIGGRAPH 2013, Anaheim.
[82] H. N. Ng and R. L. Grimsdale. Computer graphics techniques for

modeling cloth. IEEE Comput. Graph. Appl., 16(5):28–41, September
1996.
[83] S. Okamura and T. Igarashi. An assistant interface to design and

produce a pop-up card. International Journal of Creative Interfaces
and Computer Graphics, 1(2):40–50, 2010.
137
Bibliography
[84] J. O’Rourke. How to Fold It: The Mathematics of Linkages, Origami,

and Polyhedra. Cambridge University Press, 2011.
[85] D. Pelham. Trail: Paper Poetry Pop-Up. Little Simon, 2007.
[86] X. Provot. Deformation constraints in a mass-spring model to de-

scribe rigid cloth behaviour. In Graphics interface, pages 147–147.
Canadian Information Processing Society, 1995.
[87] E. G. Rubin. A history of pop-up and movable books: 700 years of

paper engineering.
http://www.youtube.com/watch?v=iDJJOaZ1myM, November
2010. Public lecture.
[88] C. J. Ruiz, S. N. Le, and K.-L. Low. Generating multi-style paper

pop-up designs using 3d primitive fitting. In ACM SIGGRAPH
Asia Technical Briefs, SA ’13, pages 4:1–4:4, 2013.
[89] C. J. Ruiz, S. N. Le, J. Yu, and K.-L. Low. Multi-style paper pop-up
designs from 3d models. Computer Graphics Forum, 3(2):–, 2014.
[90] K. Schulgasser. The in-plane poisson ratio of paper. Fibre Science

and Technology, 19(4):297–309, 1983.
[91] Y. Schwartzburg and M. Pauly. Fabrication-aware design with in-

tersecting planar pieces. Computer Graphics Forum, 32(2pt3):317–
326, 2013.
[92] A. Shamir. A survey on mesh segmentation techniques. Computer

Graphics Forum, 27(6):1539–1556, 2008.
[93] I. Shatz, A. Tal, and G. Leifman. Paper craft models from meshes.
Vis. Comput., 22(9):825–834, September 2006.
138
Bibliography
[94] O. Stava, J. Vanek, B. Benes, N. Carr, and R. Mech. Stress relief: Im-
proving structural strength of 3d printable objects paper abstract.
In In ACM SIGGRAPH 2012 papers, 2012.
[95] K. A. Stevens. The visual interpretation of surface contours. Arti-

ficial Intelligence, 17(13):47–73, 1981.
[96] R. Szilard. Theories and Applications of Plate Analysis: Classical Nu-

merical and Engineering Methods. Wiley, 2004.
[97] T. Tachi. Origamizing polyhedral surfaces. IEEE Transactions on

Visualization and Computer Graphics, 16(2):298–311, March 2010.
[98] TamaSoftware. Pepakura designer.

http://www.tamasoft.co.jp/pepakura-en/, 2007.
[99] A. Telea and A. Jalba. Voxel-based assessment of printability of 3d

shapes. In Proceedings of the 10th international conference on Math-
ematical morphology and its applications to image and signal process-
ing, ISMM’11, pages 393–404, Berlin, Heidelberg, 2011. Springer-
Verlag.
[100] D. Terzopoulos, J. Platt, A. Barr, and K. Fleischer. Elastically de-

formable models. In ACM Siggraph Computer Graphics, volume 21,
pages 205–214. ACM, 1987.
[101] B. Thomaszewski, M. Wacker, and W. Straß er. A consistent bend-

ing model for cloth simulation with corotational subdivision finite
elements. In Proceedings of the 2006 ACM SIGGRAPH/Eurographics
symposium on Computer animation, SCA ’06, pages 107–116, Aire-la-
Ville, Switzerland, Switzerland, 2006. Eurographics Association.
[102] J. Todd and F. Reichel. Visual perception of smoothly curved sur-

faces from double-projected contour patterns. J. of Exp. Psych.: Hu-
man Percep. and Perform., 16:665–674, 1990.
139
Bibliography
[103] R. Uehara and S. Teramoto. The complexity of a pop-up book. In

In Proc. 18th Canadian Conference on Computational Geometry, 2006.
[104] P. Volino, N. Magnenat-Thalmann, and F. Faure. A simple ap-

proach to nonlinear tensile stiffness for accurate cloth simulation.
ACM Transactions on Graphics, 28(4):105:1–105:16, 2009.
[105] J. Weil. The synthesis of cloth objects. In Proceedings of the 13th an-
nual conference on Computer graphics and interactive techniques, SIG-
GRAPH ’86, pages 49–54, New York, NY, USA, 1986. ACM.
[106] E. Whiting, J. Ochsendorf, and F. Durand. Procedural modeling of

structurally-sound masonry buildings. In ACM SIGGRAPH Asia
2009 papers, SIGGRAPH Asia ’09, pages 112:1–112:9, New York,
NY, USA, 2009. ACM.
[107] J. Whitney, P. Sreetharan, K. Ma, and R. Wood. Pop-up book mems.

Journal of Micromechanics and Microengineering, 21(11):115021, 2011.
[108] J. Xu, C. S. Kaplan, and X. Mi. Computer-generated papercutting.

In Proceedings of the 15th Pacific Conference on Computer Graphics and
Applications, pages 343–350. IEEE Computer Society, 2007.
140

LeNgocSang Thesis PDF

Încărcat de

Informații document

Titlu original

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

LeNgocSang Thesis PDF

Încărcat de

Drepturi de autor:

Formate disponibile

AUTOMATIC DESIGN OF

NATIONAL UNIVERSITY OF SINGAPORE

A THESIS SUBMITTED FOR THE DEGREE OF

DEPARTMENT OF COMPUTER SCIENCE

written by me in its entirety.

I have duly acknowledged all the sources of information which have

been used in the thesis.

I am deeply grateful to Dr. Low Kok-Lim, my Ph.D. supervisor, for

his patient guidance, continued encouragement and constant support.

He always shares with me insightful advice, while giving me sufficient

room to grow. His wisdom, kindness, and wonderful personality will

always inspire me.

I would also like to thank Leow Su Jun, Le Nguyen Tuong Vu and

I really enjoy working with them.

their interesting discussions and sharings during our weekly meetings.

I can never forget my wonderful work environment in G3 lab and espe-

cially NUS, my second home.

of love, support and encouragement. I am lucky to have met and mar-

est years of my Ph.D. program. This thesis is dedicated to my parents,

my wife and my lovely little daughter.

List of Tables xxiii

List of Publications xxv

3 Automatic Design of Origamic Architecture 27

4 Origamic Architecture with Non-parallel Folds 69

5 Strengthening Origamic Architecture Pop-Ups 97

6 Conclusion and Future Directions 121

A number of computer tools have been developed to assist the OA

In this thesis, we present an algorithm for automatic design of an OA

In addition, we propose a simple yet effective approach for analyz-

By combining both aspects, we guarantee that our final OA designs

1.1 (a) A volvelle for astronomical illustration. (b) A flap demon-

1.2 An origamic architecture (left) and a 360◦ pop-up (right). . 3

1.3 (a) A computer-aided tool for designing OA by Mitani and

2.1 (a) Automatic origami design by tucking folded molecules

2.2 (a) Strip modeling for making papercraft toys from 3D

2.3 (a) Simulation of the opening and closing of pop-up cards

2.4 (a) An invalid OA with floating part is rejected by [77, 98].

2.5 (a) A sample input model. (b) The OA created based on

2.8 (a) A feature-sensitive segmentation that can be used for

2.9 Physically weak structures are detected and modified for

2.10 The structural strength analysis in [94]. A weak model of a

2.11 (a) Simulation of hanging cloth using the popular mass-

3.1 Our automatic OA design pipeline. Note that the whole

3.3 Projecting vector y along w, which bisects y and z, results

3.4 1-paths (a) and 2-paths (b) were shown to be stable in

3.5 An OA containing two double connections (left) and its

3.6 (a) An OA with a monotonic path, where d0 < d2 < d4 <

3.7 Side views of a monotonic B-path (a) and a near-monotonic

3.9 A house model (top left) is approximated using OA: (a)

3.10 Slice positioning and contour projection: An input sur-

3.11 Possible relative positions of pa and pb . The top row shows

3.12 A 3D model (a) is approximated by a set of representa-

3.14 (a, d) Unstable OA structures are made stable by (b, e) Li

3.15 OA pop-ups of the Rialto bridge (left) and Colosseum (right),

3.16 OA designs of some experimented models created by our

3.17 OA designs of some experimented models created by our

3.18 OA pop-ups of the Stanford bunny and a church model.

4.1 Pop-up designs with non-parallel folds. . . . . . . . . . . . 69

4.2 A series of parallel structures can be replaced by a single

4.3 (a) The components in a v-structure. (b) A v-structure with

4.4 A foldable path of v-structures and its angles. . . . . . . . . 74

4.5 (a) A v-structure in its closed configuration illustrates the

4.6 (a) OA plan of a parallel structure (top) and its coverage