# The Simplified Surface Spline Erik Christopher Dyken ```University of Oslo
Department of Informatics
The Simplified
Surface Spline
Erik Christopher
Dyken
Candidatus
Scientiarum Thesis
January 24, 2003
ii
Abstract
This thesis studies the problem of using surface splines as an approximation basis.
The simplified surface spline basis, a slight simplification of the C 1 -surface spline
construction of Jörg Peters, is presented. The simplified surface does not guarantee a
C 1 surface, but has explicit formulas, explicit basis functions and known dimension.
In addition, it resides close to the C 1 -surface spline. The simplified surface spline
basis is employed in interpolation, least squares approximation and faired least squares
approximation. In addition, the mesh notation, a notation for writing algorithms on
polyhedral meshes in a concise and mathematical way, is presented.
iii
iv
ABSTRACT
Foreword
This thesis is part of my Candidatus Scientiarum studies at the Department of Informatics, Faculty of Mathematics and Natural Sciences, University of Oslo. These graduate
studies are composed of one semester of courses and two semesters of scientific research. This thesis is the result of my scientific research.
Initially, the scope of my scientific research was to do an approximation using
surface splines. This included triangle surface generation, control mesh generation,
parameterisation of the fine triangle mesh over the control mesh, blend ratio estimation,
and finally approximation. I quickly realized that in order to overcome this ordeal, I
had to treat all aspects of this process rather superficially. Thus, I decided to pick just
one aspect of this process instead, and study this in some detail.
It was not particularly difficult to choose which of these subjects to pick. Triangle surface generation is a well-populated scene, and I doubt I can give a significant
contribution to this topic. Control mesh generation, on the other hand, is a more open
question. This involves finding a base mesh for, among other uses, approximation.
Blend ratio estimation and surface spline approximation methods are both very open
questions. With inspiration from the approximation theory of B-splines presented in
“Spline Methods” (Mørken & Lyche 2002), a compendium for a graduate course in
spline methods, I decided to try doing something similar with surface splines. Thus,
my topic becomes approximation with surface splines.
My intention with this thesis is to give a respectable account of the scheme and
forms of evaluation, finding expressions for the basis functions, finding the dimension
of this basis, and finding out if in fact this basis always was of full rank. The reader may
be amused by my ambition. However, my intention was never to create a surface spline
theory comparable to spline theory. The spline theory was only a source of inspiration.
I find it a yielding practice to do things simple first, and when the simple thing
works, expand it to the full thing. Moreover, this was the strategy employed on the
surface spline. The simple thing is the simplified surface spline. However, the study of
approximation with the simplified surface spline basis grew into a topic large enough
to be covered alone in this thesis.
Thus, extending my work to the “full” surface spline scheme must be an ordeal left
to future generations of inquisitive minds — if anyone cares to bother.
I use the opportunity here to thank some of the people that have helped me with
this thesis, ranging from general inspiration to the trade of mathematical writing.
I wish to thank Professor Knut Mørken, my supervisor, for giving me the freedom to
mould this thesis into precisely what I wanted, as well as numerous valuable comments,
discussions tips, and constructive criticism.
I also wish to thank Martin Reimers, a Ph.D. fellow here at the department, for
friendship as well as countless discussions, comments and criticism — none necessarily limited to surface splines.
v
vi
FOREWORD
I also wish to thank my parents for encouragement and proofreading.
I also wish to thank the employees at Sim Surgery A/S, especially Jan Sigurd
Røtnes and Geir Westgaard. Røtnes initially posed the problem of my thesis, and Westgaard has given me a lot of useful information on the surface spline schemes, as well
as discussions, comments and constructive criticism.
Finally, I wish to thank my fellow students at “Parken” study room for the good
social environment.
Erik Christopher Dyken
Oslo, January 24, 2003
Contents
Abstract
iii
Foreword
v
1 Introduction and background
1.1 About surface splines and meshes . . . . . .
1.1.1 Surface splines in general . . . . . .
1.1.2 The simplified surface spline . . . . .
1.1.3 About mesh notation . . . . . . . . .
1.2 Organisation of thesis . . . . . . . . . . . . .
1.3 Literature review . . . . . . . . . . . . . . .
1.4 Background material . . . . . . . . . . . . .
1.4.1 Convex combinations . . . . . . . . .
1.4.2 Bézier splines . . . . . . . . . . . . .
1.4.3 Composite Bézier curves and surfaces
1.4.4 B-splines . . . . . . . . . . . . . . .
1.5 Summary . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
1
1
2
2
2
3
4
4
6
8
9
10
2 Meshes
2.1 Some notes on sets . . . . . . . . . . . . .
2.2 The mesh definition . . . . . . . . . . . . .
2.2.1 The vertices of a face . . . . . . . .
2.2.2 Positions in the mesh . . . . . . . .
2.2.3 The number of mesh elements . . .
2.3 Mesh queries . . . . . . . . . . . . . . . .
2.3.1 Relational concepts . . . . . . . . .
2.3.2 The structure . . . . . . . . . . . .
2.3.3 Dart selection and projection . . . .
2.4 Mesh modification and refinement . . . . .
2.4.1 Dual mesh . . . . . . . . . . . . .
2.4.2 Face split (polyhedral split) . . . .
2.4.3 Vertex split (mid-edge subdivision)
2.4.4 The Doo-Sabin split . . . . . . . .
2.4.5 Planar cut polyhedra . . . . . . . .
2.5 Summary . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
11
11
12
14
14
16
17
17
19
21
24
24
25
27
28
28
29
vii
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
CONTENTS
viii
3 The surface spline curve
3.1 The scheme . . . . . . . . . . . . . . . . . . . .
3.1.1 Refinement . . . . . . . . . . . . . . . .
3.1.2 Sub-interval midpoints . . . . . . . . . .
3.1.3 Bézier coefficients . . . . . . . . . . . .
3.1.4 Defining the Bézier curve segments . . .
3.1.5 Summary of construction . . . . . . . . .
3.2 Explicit formulas . . . . . . . . . . . . . . . . .
3.3 Exists a subdivision formula? . . . . . . . . . . .
3.3.1 Bézier segment subdivision . . . . . . .
3.3.2 Surface spline subdivision . . . . . . . .
3.3.3 Futile attempt no. 1: inserting one point .
3.3.4 Futile attempt no. 2: uniform refinement .
3.4 Summary . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
31
31
32
32
33
34
36
39
39
39
41
41
43
44
4 Surface spline schemes
4.1 The C 1 -surface spline . . . . . . . . . . . .
4.1.1 The scheme . . . . . . . . . . . . .
4.2 Two other surface spline schemes . . . . .
4.2.1 Surfaces . . . bicubics . . . . . . . .
4.2.2 Localised-hierarchy surface splines
4.3 Summary . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
45
45
47
50
50
52
53
5 The simplified surface spline
5.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . .
5.1.1 Topological terms of the simplified surface spline
5.1.2 The three meshes . . . . . . . . . . . . . . . . .
5.1.3 The intrinsic parameterisation . . . . . . . . . .
5.2 Layer I: The control mesh vertices . . . . . . . . . . . .
5.3 Layer II: The quad mesh points . . . . . . . . . . . . . .
5.4 Layer III: Quad midpoints (inner mesh vertices) . . . . .
5.5 Layer IV: Bézier coefficients . . . . . . . . . . . . . . .
5.6 Layer V: Bézier evaluation . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
55
55
56
57
58
58
58
60
62
65
6 The simplified surface spline basis
6.1 Overview . . . . . . . . . . . . . . . . . . . . . . . .
6.2 Layer II: Quad mesh points . . . . . . . . . . . . . . .
6.3 Layer III: Quad midpoints . . . . . . . . . . . . . . .
6.4 Layer IV: Bézier coefficients . . . . . . . . . . . . . .
6.5 Layer V: The surface in terms of basis functions . . . .
6.6 The simplified surface spline space . . . . . . . . . . .
6.6.1 The space . . . . . . . . . . . . . . . . . . . .
6.6.2 Linear independence . . . . . . . . . . . . . .
6.6.3 A local criterium . . . . . . . . . . . . . . . .
6.6.4 The local quad midpoint matrices . . . . . . .
6.6.5 The simplified surface spline space and its basis
6.7 Summary . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
67
68
70
71
75
77
79
79
80
81
82
84
86
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
CONTENTS
ix
7 Practical considerations
7.1 Evaluation of layer based schemes . . . . . . . . . . . . . .
7.1.1 Evaluation in matrix guise . . . . . . . . . . . . . .
7.1.2 Direct and indirect evaluation . . . . . . . . . . . .
7.1.3 Retained Bézier evaluation . . . . . . . . . . . . . .
7.2 OpenGL rendering of surfaces with local support . . . . . .
7.2.1 Efficient rendering of Bézier patches . . . . . . . . .
7.2.2 Frustum culling . . . . . . . . . . . . . . . . . . . .
7.2.3 Occlusion tests . . . . . . . . . . . . . . . . . . . .
7.3 Benefits of using composite polynomial patches . . . . . . .
7.3.1 Ray tracing . . . . . . . . . . . . . . . . . . . . . .
7.3.2 Mass, centre of gravity and moment of inertia tensor
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
87
87
87
88
88
89
89
90
92
93
93
94
8 Approximation experiment
8.1 Our choice of parameterisation . . . . .
8.2 The original models . . . . . . . . . . .
8.3 Error, approximation order and stability
8.4 Four approximation methods . . . . . .
8.4.1 Lazy approximation . . . . . .
8.4.2 Vertex interpolation . . . . . . .
8.4.3 Least squares . . . . . . . . . .
8.4.4 Faired least squares . . . . . . .
8.5 Application . . . . . . . . . . . . . . .
8.5.1 Overview . . . . . . . . . . . .
8.5.2 Nitty-gritty details . . . . . . .
8.6 Results . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
95
95
96
97
97
97
98
98
99
100
100
101
104
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
9 Conclusions and further work
121
References
122
Index
124
x
CONTENTS
List of Figures
1.1
Lay-out of Bézier control points . . . . . . . . . . . . . . . . . . . .
2.1
2.2
2.3
2.4
2.5
A mesh . . . . . . . . . . .
Positions in a mesh . . . . .
Overview of dart queries . .
A consecutive set of corners
Face split and vertex split . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
13
15
21
24
25
3.1
3.2
3.3
3.4
Layout of surface spline curve construction . . . .
Example of a surface spline curve . . . . . . . . .
A surface spline curve and a cubic Bézier segment .
Inserting one control point. . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
35
36
38
40
4.1
4.2
4.3
4.4
4.5
4.6
Different blend ratios . . . . . . . . . . . . . . . . .
The control mesh, the quad mesh, and the inner mesh
C 1 -surface spline mesh refinement . . . . . . . . . .
C 1 -surface spline edge cutting . . . . . . . . . . . .
C 1 -surface spline quadratic meshing . . . . . . . . .
Construction of “surfaces . . . bicubics” . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
46
47
48
49
50
51
5.1
5.2
5.3
5.4
5.5
5.6
5.7
The simplified surface spline concepts . . . . . . . . .
The simplified surface spline control mesh . . . . . . .
The simplified surface spline quad mesh . . . . . . . .
The simplified surface spline quad midpoints . . . . .
The simplified surface spline Bézier coefficients . . . .
The simplified surface spline boundary coefficients . .
The simplified surface spline Bézier coefficient layout .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
57
59
59
61
63
64
65
6.1
6.2
6.3
Deduction of basis support . . . . . . . . . . . . . . . . . . . . . . .
Construction of the local quad midpoint matrix. . . . . . . . . . . . .
The inequality of lemma 6.6.10 . . . . . . . . . . . . . . . . . . . . .
74
81
85
7.1
7.2
The view frustum. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
The vertex transformation pipeline of OpenGL. . . . . . . . . . . . .
90
91
8.1
8.2
8.3
8.4
8.5
Class diagram of application
Approximation: cube0 . . .
Approximation: cube1 . . .
Approximation: cube2 . . .
Approximation: cube3 . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
xi
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
8
102
106
107
108
109
LIST OF FIGURES
xii
8.6
8.7
8.8
8.9
8.10
8.11
8.12
8.13
Approximation: cube4 . .
Approximation: sphere0 .
Approximation: sphere1 .
Approximation: sphere2 .
Approximation: sphere3 .
Approximation: sphere4 .
Approximation: jittercube
Approximation: bunny . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
110
111
112
113
114
115
116
117
List of Tables
5.1
The simplified surface spline terms . . . . . . . . . . . . . . . . . . .
8.1
8.2
8.3
8.4
8.5
Data on original models . . . . . . . . . .
Results from lazy approximation . . . . .
Results from vertex interpolation . . . . .
Results from least squares approximation
Results from faired approximation . . . .
xiii
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
56
96
118
118
118
119
xiv
LIST OF TABLES
Chapter 1
Introduction and background
This is a thesis about surface splines. We present the simplified surface spline, a simplification of the “C 1 surface splines”-scheme by Jörg Peters (Peters 1995a). The thesis
can be divided into four subjects:
1. The mesh notation, a clean, concise and mathematical method of defining subsets of connectivity. This notation is not restricted to analysis of surface spline
schemes.
2. A thorough investigation of the simplified surface spline scheme. Much of the
structure of this investigation can be applied to other surface spline schemes.
In addition, a few approximation experiments are performed with the simplified
surface spline basis.
3. An overview of a few surface spline schemes.
4. A thorough discussion of practical advantages for using surface spline schemes.
Much of this discussion is applicable for any scheme based on convex combinations.
The contribution of this thesis with originality are items (1) and (2). Items (3) and (4)
are more of an overview of existing literature and knowledge.
1.1 About surface splines and meshes
We give a quick overview of surface spline schemes in general, and then present the
motivation of the simplified scheme.
1.1.1 Surface splines in general
Surface splines may be regarded as a macro spline, similar to the Clough-Tocher and
Powell-Sabin macro splines. The surface spline tries to have the same “conceptual”
framework as B-splines, where the connectivity and the blend ratios play the same
role as knot vectors do for B-splines. Given a control mesh, this mesh is refined and
weights called blend ratios are used to tune the location of the refined vertices. The
refined mesh and vertices are then used to construct a connected patchwork of tensor
product B-spline patches. These patches abut with the desired smoothness.
1
2
CHAPTER 1. INTRODUCTION AND BACKGROUND
The surface spline scheme generates a surface from control meshes of rather arbitrary topologies. The surface does satisfy standard smoothness criteria. The number of
patches output is usually kept to a minimum, and a closed form of the surface exists.
The parameterisation is simple and based on standard patches. Thus the surface is not
difficult to evaluate, render, differentiate, and integrate.
With this said, the surface spline does not possess the subdivision property essential
for wavelets. Further, the construction is not particularly simple and elegant — many
of the schemes are in fact rather hard to implement. In addition, the schemes are
non-trivial to generalise for arbitrary degree and smoothness, and there exists little
approximation theory for these schemes.
1.1.2 The simplified surface spline
The reason of this simplification is to make a simple prototype for experimenting with
methods of finding basis functions and properties.
When such a method is found, it can be applied to the full scheme. Investigating
the scheme in full is not in the scope of this thesis. However, we will examine the
simplified surface spline in detail: various forms of construction, the basis functions,
and how to obtain them, as well as some properties of the scheme.
The focus of the thesis is approximation properties of the simplified surface spline.
To shed light on this, we present four kinds of approximation schemes, and put these
to practice with some experiments.
Surface splines are defined over arbitrary topology meshes. We have found no useful
notation for dealing with arbitrary meshes that both is suitable for mathematical analysis as well as describing a potential interface between algorithms and the connectivity
kernel1 . We call this notation mesh notation. Mesh notation is the result of applying
the relational database paradigm on geometrical objects. Using relational databases for
geometric modelling per se is nothing new, relational databases are sometimes used in
GIS-systems.
However, we have never encountered the use of this paradigm to define a notation
to extract certain subsets of the connectivity. Connectivity information in the guise of
relational algebra, combined with mathematical set operations, yields some interesting
results. We will deal with this in chapter 2.
1.2 Organisation of thesis
We give a small overview of the chapters:
1. Introduction and background gives a rough overview of the problem and the
contents of this thesis, as well as a literature review. We give an overview of
convex combinations and some properties of these operations. Convex combinations are used to prove some properties of the simplified surface spline. The rest
of the chapter is an quick overview, to set the a context, of Bézier splines and
B-splines.
1 By connectivity kernel we mean whatever data structure containing the connectivity of the object as well
as interface function
1.3. LITERATURE REVIEW
3
2. Meshes defines the concepts of meshes and mesh queries. Mesh queries are used
extensively in the following chapters.
3. The surface spline curve serves as an introduction to surface spline schemes.
In addition, the boundary curve of the simplified surface spline is such a curve.
4. Surface spline schemes gives an overview of three existing surface spline schemes.
5. The simplified surface spline gives an account of the simplified surface spline.
6. The simplified surface spline basis gives a detailed account of how the basis
functions were found, and a few properties are deduced.
7. Practical considerations gives a discussion about practical employment of surface splines, with substantial weight on methods of evaluation and rendering.
8. Approximation experiment is an experiment to test the simplified surface spline
basis deduced in chapter 6.
9. Conclusions tries to summarise this thesis. In addition, a road map for further
research on approximation with surface splines is given.
1.3 Literature review
A relatively small amount of literature on the topic of surface splines exists. A few
books mention the scheme, in the fourth edition of (Farin 1996) a small presentation of
surface splines is given, and a few articles employ these schemes.
The Primary source of information on surface splines is found in the articles of
Jörg Peters. The earliest article from Peters on surface splines is (Peters 1994), and
this scheme is discussed in section 4.2.1. This scheme is improved in (Peters 1995a),
which is discussed in section 4.1. This is the scheme that the simplified surface spline
is based upon. In addition to a presentation of the construction, proofs for continuity
and the convex hull property are given in the paper.
In (Gonzalez & Peters 1999) the C 1 surface spline is extended to incorporate local
refinement and topology changes. However, the refinement procedure does not give the
exact same surface. Finally, in (Peters 1996) a G2 surface spline construction is given.
On application of surface splines, (Eck & Hoppe 1996) gives an account of employing surface splines when reconstructing B-splines from point clouds.
The article (Gonzalez-Ochoa, McCammon & Peters 1998) gives an account of how,
by using the great theorem of Gauss, to compute various physical properties of objects
defined by piecewise polynomials.
Very little is written on the topic of planar cut polyhedra. Planar cut polyhedra is
mentioned in some of Peters’ papers, especially in (Peters 1998) where an algorithm to
convert planar cut polyhedra to trimmed NURBS patches is given.
In (Farin 1996), a thorough account on continuity of composite spline patches is
given. Another account, probably more aimed at surface splines, are given in (Peters
2001). This paper reveals some of the motivation behind the steps of surface spline
constructions.
The mesh notation is based on G-maps. The articles (Halbwachs & Hjelle 1999)
and (Halbwachs, Courriouz, Renaud & Repusseau 1996) both give a good description
of the concept of G-maps, as well as relevant references.
CHAPTER 1. INTRODUCTION AND BACKGROUND
4
1.4 Background material
In this section we will look at Convex combinations, Bézier splines and B-splines and
surface splines. Convex combinations are used extensively in this text, and Bézier
patches are used in the surface spline construction.
1.4.1 Convex combinations
We assume that all points mentioned are points in a linear space. Thus, this will not
be stated explicitly, but nevertheless it is an important requirement. We begin with the
formal definition of a convex set.
Definition 1.4.1. A set S is convex if
αp0 + (1 − α)p1 ∈ S,
∀p0 , p1 ∈ S,
and α ∈ [0, 1]
(1.1)
A convex set is a set where any line segment between two elements of the set is contained in the set as well. We also define the convex hull.
Definition 1.4.2. The convex hull of the set of points P, CH(P), is the smallest convex
sets containing all of the points of P.
The linear interpolation of definition 1.4.1 is in fact a convex combination. We give
this concept a formal definition.
Definition 1.4.3. A convex combination of the points c1 , . . . , cn is a construction
c=
n
X
αi c i ,
(1.2)
i=1
where the weights αi satisfy the two requirements
1. αi ≥ 0 i = 1, . . . , n
Pn
2.
i=1 αi = 1.
If we can write an expression as a convex combination, the result has some useful
properties, in particular, the convex hull property.
Property 1.4.4. A convex combination always resides inside the convex hull of the
points used in the convex combination.
This follows from the definitions 1.4.1 and 1.4.2. Convex combinations of convex
combinations are in turn convex combinations themselves. This is useful when proving
that a construction is a convex combination.
Property 1.4.5. Given a set of points P, let Q be a set of points defined as convex
combinations of the points in P. Let the set R be a set of points defined as convex
combinations of the points in Q. Then, the points in R are convex combinations of the
points in P.
Proof. Let a convex combination of a convex combination be


nβ
nα
X
X
βij cj 
αi 
c=
i=1
j=1
(1.3)
1.4. BACKGROUND MATERIAL
5
Since both combinations are convex, all weights are positive. Any multiplication of two
of these weights, αi · βj will be positive as well; hence requirement (1) of definition
1.4.3 is satisfied. Further,


nβ
nα
nα
X
X
X


αi
βj =
αi · 1 = 1,
(1.4)
i=1
j=1
i=1
which fulfils requirement (2) of definition 1.4.3.
Instead of writing convex combinations as a sum, we can write it as a matrix product
between a row vector of weights and a column vector of the objects to be combined. If
we stack several of these convex combinations on top of each other, we get a convex
combination matrix.
Definition 1.4.6. A convex combination in matrix form is the construction
P = AC
 
p1
α11
 ..   ..
 . = .

pm
αm1
 
. . . α1n
c1
..   .. 
..
.
.  . 
. . . αmn
cn
(1.5)
where the elements of the weight matrix A satisfy
1. αij ≥ 0, 1 ≤ i ≤ m, 1 ≤ j ≤ n,
Pn
2.
1 ≤ i ≤ m.
j=1 αij = 1,
and
Each row of A consists of the weights of one convex combination. We will prove one
situation where the weight matrix is non-singular. In order to do so, we must define
diagonally dominant matrices.
Definition 1.4.7. If the entries aij of a matrix A possess the following property,
|aii | >
n
X
j=1,j6=i
|aij |
1 ≤ i ≤ n,
(1.6)
then A is a diagonally dominant matrix.
Diagonally dominant matrices possess two useful properties:
Property 1.4.8. If the matrix A is diagonally dominant, the following is true:
1. The matrix A is non-singular and has an LU-factorisation.
2. The system Ax = b can be solved by Gaussian elimination without pivoting.
Proofs for both properties can be found in (Cheney & Kincaid 1996, pp.188-190).
We will now show when the weight matrix of a convex combination is diagonally
dominant, and thus, possesses properties 1.4.8.
Property 1.4.9. The weight matrix of a convex combination in matrix form is diagonally dominant if
1. m = n, and
CHAPTER 1. INTRODUCTION AND BACKGROUND
6
2. αii > 1/2, for 1 ≤ i ≤ n,
that is, it is square and the diagonal entries are larger than 1/2.
Proof. Each row sum to one and each entry in the matrix is non-negative. Thus, for
each row, if the diagonal element is larger than 1/2, then the sum of the other entries
of the row will be smaller than 1/2, and therefore the matrix is diagonally dominant
according to definition 1.4.7.
We think it is good practice to define terms before using them. Thus, we boldly
ignore the potential peril of being regarded as pedantic, and start with convex combinations and Bernstein polynomials.
1.4.2 Bézier splines
Bézier splines are closely related to B-splines, in fact it is possible to transform any
B-spline to a Bézier spline by knot insertion, and thus it is always possible to write a
Bézier spline as a B-spline, it is just a matter of choosing a representation. The Bézier
curves and surfaces represent a fundamental building block for surface spline schemes.
Bernstein polynomials
Before defining Bézier curves and surfaces, we define the Bernstein polynomials.
Definition 1.4.10. The Bernstein polynomials of degree d are defined as
d i
Bid (u) =
u (1 − u)d−i , u ∈ [0, 1]
i
(1.7)
where i = 0, . . . , d.
The Bernstein polynomials exhibit some useful properties.
Property 1.4.11. The Bernstein polynomials are
1. a partition of unity,
2. all non-negative on the interval [0, 1], and
3. a basis for polynomials.
Proofs are found in (Farin 1996). From (1) and (2) we see that the set of Bernstein
polynomials for a given degree can be used as weights for a convex combination.
Bézier curves
The Bernstein polynomials form the basis for the Bézier curve:
Definition 1.4.12. The Bézier curve segment of degree d is defined as
f (u) =
d
X
i=0
Bid (u)ci ,
ci ∈ Rn
(1.8)
where Bjd (u) are the Bernstein polynomials and c0 , . . . , cd are the d + 1 control points
of the curve.
1.4. BACKGROUND MATERIAL
7
We will give an example of two Bézier curves, the linear interpolation and the quadratic
Bézier blend. These two curves will be used later in the text.
Example 1.4.1. The linear interpolation is a Bézier curve of degree 1,
f (u) =
1
X
Bi1 (u)ci
i=0
B01 (u)c0
(1.9)
=
+ B11 (u)c1
= (1 − u)c0 + uc1 .
The quadratic Bézier blend is a Bézier curve of degree 2,
f (u) =
=
2
X
Bi1 (u)ci
i=0
B02 (u)c0
2
+ B12 (u)c1 + B22 (u)c2
(1.10)
= (1 − u) c0 + 2(1 − u)uc1 + u2 c2 .
Tensor product Bézier surfaces
The tensor product of two Bézier curve basises creates a bivariate tensor product Bézier
basis. For details, see (Farin 1996) or (Gallier 2000).
Definition 1.4.13. A Bézier tensor product surface s(u, v) is defined as
s(u, v) =
dv
du X
X
ci,j Bidu (u)Bjdv (v),
i=0 j=0
ci,j ∈ Rn ,
(1.11)
where Bidu (u) and Bjdv (v) are Bernstein polynomials and ci,j are elements of a du ×dv
grid of control points.
We give an example of two tensor product surfaces (both will be used in the simplified
surface spline construction).
Example 1.4.2. The bilinear interpolation is a tensor product of two linear interpolations, thus the degrees are both 1 in the u and v directions.
s(u, v) =
1 X
1
X
ci,j Bi1 (u)Bj1 (v)
i=0 j=0
(1.12)
= (1 − u)(1 − v)c0 + (1 − u)vc1 + u(1 − v)c2 + uvc3 .
The biquadratic blend is a tensor product of two quadratic Bézier blends. In this case
the degrees are 2 in both the u and v directions.
s(u, v) =
2 X
2
X
ci,j Bi2 (u)Bj2 (v)
i=0 j=0
= (1 − u)2 (1 − v)2 c0,0 + 2(1 − u)2 (1 − v)vc0,1 + (1 − u)2 v 2 c0,2 +
2(1 − u)u(1 − v)2 c1,0 + 4(1 − u)u(1 − v)vc1,1 + 2(1 − u)uv 2 c1,2 +
u2 (1 − v)2 c2,0 + 2u2 (1 − v)vc2,1 + u2 v 2 c2,2 .
Figure 1.1 gives an idea of how the control points relate to the parameter space.
(1.13)
CHAPTER 1. INTRODUCTION AND BACKGROUND
8
v
1.0
v
c 0,1
c 1,1
1.0
0.5
0.0
0.5
c 0,0
c 1,0
u
0.0
0.5
1.0
0.0
c 0,2
c 1,2
c 2,2
c 0,1
c 1,1
c 2,1
c 0,0
c 1,0
c 2,0
u
0.0
0.5
1.0
Figure 1.1: Lay-out of the control points for the bi-linear interpolation (left) and the
Properties of Bézier curves and surfaces
Bézier curves and surfaces have numerous useful properties, most of which stem from
the fact that these shapes are convex combinations of the control points. We will present
some of them here, though without motivation and proofs — the reader is encouraged
to consult texts like (Farin 1996), (Gallier 2000) and (Foley, van Dam, Feiner & Hughes
1996).
Property 1.4.14. Bézier curves and surfaces exhibit the following properties:
1. The derivative of a Bézier curve or surface is another Bézier curve or surface of
lower degree, where the coefficients are differences of the original coefficients.
2. The Bézier curve or surface reside inside the convex hull of the control points.
3. Bézier curves or surfaces of degree 1 or higher exhibit linear precision (reproduces a line or plane).
4. Bézier curves interpolate its control points at the end, and the Bézier surfaces
interpolate the four corner control points.
1.4.3 Composite Bézier curves and surfaces
Polynomials are infinitely smooth; hence the continuity properties of schemes producing piecewise polynomials are only interesting at the joints. We will not treat this
subject with great rigour since it is not in the major scope of this thesis. The reader
is referred to (Farin 1996) and (Peters 2001) for discussions of continuity and geometric continuity, where (Peters 2001) is very focused on requirements for surface spline
schemes. We define the concepts of continuity at a joint very loosely:
Definition 1.4.15. A composite polynomial has the following continuity properties at
a joint if the following properties match for all the pieces meeting at the joint:
1. C 0 or G0 if the positions match.
1.4. BACKGROUND MATERIAL
9
2. C 1 if all of the derivatives match.
3. G1 if the tangent planes match.
In general, C n -continuity is dependent on the parametrisations: The speed an parameter space must match across the joint. Gn -continuity requires only such a reparametrisation to exists. Thus, there is a difference between C 0 and G0 . For a surface to be C 0 ,
the parameter spaces must match.
Contrary to intuition, G1 can be stricter than C 1 . G1 requires the tangent planes
to exists, and thus, the derivatives of the parametrisation can not drop to zero. We
call such a parameterisation a regular parameterisation. C n -continuity does not have
such a restriction. However, an irregular parameterisation where the “speed” of the
parameterisation drops to zero can be desirable; being a mean to achieve sharp corners,
as is the case with surface splines.
Bézier curves and surfaces are polynomials, and in order to make interesting shapes
with them, they must be “glued” together. This “glue” is smoothness constraints. The
derivatives of the Bézier splines are very simple, so these smoothness constraints can
often be reduced to co-linearity of coefficients etc. However, it is difficult to create a
single vector space for the composite curve. Surface splines, and B-splines in particular, solve this problem.
1.4.4 B-splines
A B-spline curve is composed of three elements: A degree, a knot vector and a set of
coefficients. The degree is the degree of the polynomial pieces, and provides an upper
bound for the smoothness at the knots. The knot vector is a vector of non-decreasing
values and defines the parameter space of the B-spline. The relative spacing between
the knots governs the tightness of the curve. If the spacing is small, the curve becomes
sharper and closer to the control point, and if two knots coincide, the curve loses one
degree of continuity at that parameter value. Thus, the knot vector provides a mean to
control the local shape of the curve.
The degree and the knot vector together form a spline space. The coefficients define
a spline of this spline space. The great advantage of B-splines is that smoothness of
segment joints are automatically taken care of, and thus no peripheral constraints are
needed.
Similar to Bézier surfaces, we can create a bivariate B-spline surface as the tensor
product of two univariate B-spline curves. With this construction, we have the flexibility of two knot vectors. In addition, it is required that the control points is organised
in a grid (and thus is not of arbitrary topology). To model a shape of a more arbitrary
topology, we must slice the surface into pieces, and by performing this slicing, the unified behaviour gained by using B-splines rather than Bézier splines is lost since we yet
again have some peripheral smoothness constraints.
There have been attempts to generalise the tensor product B-spline construction
over arbitrary topologies (arbitrary is probably a bit of an over-statement, since all
schemes do have some restrictions on the parameter domain. However, they are more
arbitrary than the parameter domain of B-splines). This is not new, two papers were
published in 1978 on this topic, (Catmull & Clark 1978) and (Doo & Sabin 1978).
Subdivision surfaces and surface splines are offsprings of the strategies presented in
these articles.
10
CHAPTER 1. INTRODUCTION AND BACKGROUND
1.5 Summary
This initial part of this chapter gave a very brief overview of what surface splines are,
as well as a literature review.
The latter part of this chapter gave some background information. The well known
convex combinations are used to prove miscellaneous properties of the simplified surface spline scheme. The treatment of Bézier splines and B-splines are superficial.
These topics are included to give a setting for the surface spline schemes in this text.
Chapter 2
Meshes
This chapter defines the mesh and mesh queries. A mesh is simply put the connectivity
of an object defined by polygons. The interesting part is the mesh queries, the result of
applying a relational database paradigm to mesh connectivity, and which enables us to
specify parts of the mesh in a concise way.
We borrow some terms and concepts from G-maps. Information on G-maps can be
found in (Halbwachs & Hjelle 1999) and (Halbwachs et al. 1996).
2.1 Some notes on sets
The mesh definition is based on sets. A set is a collection of elements. An element
could be any entity, be it a vertex, an edge or another set. The size of a set A, denoted
#A, is the number of elements the set contains. Thus, if A = {a, b, c}, then #A = 3.
An offset invariant ordered set is just as an ordered set, however with a different
equivalence relation attached such that only the relative ordering of elements is considered.
Definition 2.1.1. An offset invariant ordered set is an ordered set. Two offset invariant
ordered sets are considered equal if one can be made identical to the other by cyclic
shifting of the set’s elements.
Thus, only the contained elements and their relative order is of importance. If A =
{a, b, c, d} and B = {b, c, d, a} are offset invariant ordered sets, they’re considered
equal since B can be made identical to A by applying one cyclic shift to the right on
the set. However, the set C = {a, c, b, d} is equal to neither, since the relative order
between the elements is disturbed.
We use pairs to define edges.
Definition 2.1.2. A pair is a set of two distinct elements.
Circuits are used to define faces.
Definition 2.1.3. A circuit is a shift invariant ordered set of pairs where:
1. Each pair is used at most once.
2. Adjacent pairs contain one common element.
3. The elements of the pairs are used exactly twice by the pairs of the circuit.
11
CHAPTER 2. MESHES
12
The fact that a circuit is shift invariant and the requirement (2) of definition 2.1.3 implies that the first and the last pair of a circuit contains the same element. The length of
a circuit is the number of pairs from which the circuit is defined. This number happens
to be the number of distinct elements used by the pairs of the circuit as well.
2.2 The mesh definition
Usually, a vertex has a geometrical position, and we define faces by connecting the
vertices. In this context, the vertex serves two purposes. First of all, it is a “connectivity
thingy” from which we can define faces. In addition, each vertex has as mentioned a
position.
However, in the mesh paradigm we separate these two roles. A mesh is the connectivity of a shape. There is no geometric information in the mesh. However, we can
embed (much in the same way as embeddings of G-maps) the mesh into a space by
associating geometry with mesh elements (e.g. the vertices). If this embedded mesh
is to define a surface, we must also associate a scheme. The scheme describes how to
define a surface from the geometry and the connectivity.
There is an advantage of this. With surface splines, we let the mesh define a surface
spline space, and the geometry attached to the vertices define a specific surface in this
space. This is complementary to knot vectors and coefficients of B-splines.
We define the vertex, the edge and the face.
Definition 2.2.1. A vertex is a fundamental element, and has no properties. An edge
is a pair of two vertices. A face is a circuit of edges.
Thus, edges are the elements of faces, and vertices are the elements of edges. We use
the notational convention of letting vi denote vertex i, ej edge j, and fk face k. Then
we are ready to define the mesh.
Definition 2.2.2. A mesh is the collection of the three sets
M = (V, E, F) ,
(2.1)
where V is the set of vertices, E is the set of edges and F is the set of faces defining the
mesh. The elements of the three sets must collectively satisfy:
1. Every vertex in V must be used at least twice by edges in E.
2. The edges in E must be used at least once and at most twice by faces in F.
3. If an edge is used by two faces, the faces must visit the vertices of the edge in
opposing order.
4. The vertices in V must be used by exactly zero or two boundary edges.
Requirement (1) of definition 2.2.2 makes sure that no vertices are left unused. This
simplifies things when building approximation matrices. Further, requirement (2) assures that edges are used at least once, and not used by three or more faces. Three or
more faces sharing a common edge cannot define a two-manifold surface. Requirement
(3) assures a consistent ordering of the faces. Thus, a Möbius strip cannot be represented as a mesh. Requirement (4) assures a simple boundary. Faces are always of at
least length three due to the circuit definition (the reason why is left as a little puzzle
2.2. THE MESH DEFINITION
13
f1={{1,2},{2,6},{5,6},{1,5}}
f2={{1,5},{1,3},{3,7},{5,7}}
1
e1={1,2}
f3={{1,3},{3,4},{2,4},{1,2}}
2
e5={1,5}
5
e9={5,6}
e6={2,6}
6
e2={2,4}
e4={1,3}
e12={5,7}
e10={6,8}
3
e3={3,4}
4
e8={3,7}
7
e11={7,8}
e7={4,8}
8
f4={{5,6},{6,8},{7,8},{5,7}}
f5={{2,4},{4,8},{6,8},{2,6}}
f6={{3,7,{7,8},{4,8},{3,4}}
Figure 2.1: Example mesh. The numbers denote vertex numbers.
Example 2.2.1. We will now define the mesh of the unit cube (refer to figure 2.1). The
unit cube is composed of eight vertices which we simply enumerate from 1 through 8,
V = {v1 , v2 , v3 , v4 , v5 , v6 , v7 , v8 }.
(2.2)
The unit cube has twelve edges, each connecting a pair of vertices:
E = {e1 , e2 , e3 , e4 , e5 , e6 , e7 , e8 , e9 , e10 , e11 , e12 }
= {v1 , v2 }, {v2 , v4 }, {v3 , v4 }, {v1 , v3 }, {v1 , v5 }, {v2 , v6 },
{v4 , v8 }, {v3 , v7 }, {v5 , v6 }, {v6 , v8 }, {v7 , v8 }, {v5 , v7 } .
(2.3)
And there are six faces:
F = {f1 , f2 , f3 , f4 , f5 , f6 }
= {e1 , e6 , e9 , e5 }, {e5 , e4 , e8 , e12 }, {e4 , e3 , e2 , e1 }
{e9 , e10 , e11 , e12 }, {e2 , e7 , e10 , e6 }, {e8, e11 , e7 , e3 }
n
{v1 , v2 }, {v2 , v6 }, {v5 , v6 }, {v1 , v5 }
=
{v1 , v5 }, {v1 , v3 }, {v3 , v7 }, {v5 , v7 }
{v1 , v3 }, {v3 , v4 }, {v2 , v4 }, {v1 , v2 }
{v5 , v6 }, {v6 , v8 }, {v7 , v8 }, {v5 , v7 }
{v2 , v4 }, {v4 , v8 }, {v6 , v8 }, {v2 , v6 }
o
{v3 , v7 }, {v7 , v8 }, {v4 , v8 }, {v3 , v4 }
(2.4)
CHAPTER 2. MESHES
14
Note that the set of edges is in fact a set of sets of vertices, and the set of faces is a set
of sets of sets of vertices. From these three sets we define the mesh:
M = (V, E, F) .
(2.5)
A mesh can contain most kinds of polygons. However, if the mesh only contains triangles, the mesh is called a triangular mesh, and if the mesh only contains quadrilaterals
the mesh is called a quad mesh.
2.2.1 The vertices of a face
It is often useful to represent a face as a set of vertices instead of a circuit (which is
a set of edges). We use the “hat” to denote when we use the face as a set of vertices
instead of a set of edges.
Definition 2.2.3. The vertices of a face are
[
fˆ =
ei .
(2.6)
ei ∈f
Thus fˆ is the union of all the edges of f , and since the edges are sets of vertices, the
result is a large single set containing the vertices of the face.
Example 2.2.2. Using the mesh defined in example 2.2.1 we have
fˆ1 = {v2 , v6 , v5 , v1 }
fˆ2 = {v1 , v3 , v7 , v5 }
fˆ3 = {v1 , v3 , v4 , v2 }
fˆ4 = {v6 , v8 , v7 , v5 }
fˆ5 = {v4 , v8 , v6 , v2 }
fˆ6 = {v3 , v7 , v8 , v4 }
(2.7)
2.2.2 Positions in the mesh
We use some terms to specify a position in the mesh. Figure 2.2 gives an illustration
of some of them. These terms specifies positions in the connectivity (mesh) and does
not specify a geometrical position. The first two of them, vertices and edges, are rather
obvious. These two are the same concepts known from graph theory. In addition, a
mesh has faces.
Then we have some more uncommon terms. The corners is where two faces of an
edge connect.
Definition 2.2.4. A corner is a vertex-face pair (vi , fk ) where vi ∈ fˆk . A corner
represents a position between two edges in the face description. These two edges are
named ein or the incoming edge, which is the edge in fk that connects to vi , and the
outgoing edge eout which connects from vi in the circuit.
A face that is an n-gon has n corners. Each corner resides in the position where two
edges meet (at a vertex) in this face. Since the circuit definition enforces a vertex to be
visited either by none or by two edges, the corner definition causes no ambiguities. A
2.2. THE MESH DEFINITION
15
Figure 2.2: Positions in a mesh. The nodes (upper left), the edges (upper right) and the
faces (middle left) forms a mesh (middle right). In this mesh we find corners, the black
triangles, (lower left) and darts, the pointed arrows (lower right).
CHAPTER 2. MESHES
16
corner “has” one vertex (the vertex where the corner resides) as well as two edges, the
incoming edge and the outgoing edge. Which edge is incoming and which is outgoing
depends on the orientation of the faces.
At each end of each edge we have an edge-end.
Definition 2.2.5. A edge-end is a vertex-edge pair (vi , ej ) where vi ∈ eh . The edgeends are located on the end of edges.
Obviously, each edge has two edge-ends. Each edge also have a number of half-edges.
The half-edges are along the sides of the edges.
Definition 2.2.6. A half-edge is a edge-face pair (ej , fk ) where ej ∈ fk . The halfedges are located along an edge where a face abuts.
In this way, if m faces abut an edge, the edge has m half-edges. Since only one or
two faces may abut an edge, consequently an edge only has one or two half-edges
depending on whether or not the edge is a boundary edge. Boundary edges has only
one half-edge, while non-boundary edges has two.
A concept borrowed from G-maps are the darts. Each half-edge has two darts. In
contrast to the darts of the G-maps, our darts do not have any involutions defined. Our
darts is a simple vertex-edge-face triple.
Definition 2.2.7. A dart is a vertex-edge-face triple (vi , ej , fk ) where vi ∈ ej and
ej ∈ fk . Darts are located on the end of half-edges.
From the definition of darts we see that the dart is not an arbitrary triple. This requirement put a lot of topological information of the mesh into the darts.
We can look at these positions as cursors in the connectivity at different “resolutions”. We let vi ∈ V, ej ∈ E, fk ∈ F, and M = (V, E, F). We then have some
deduced sets (they are not explicitly defined). The set C is the set of all possible corner
specifications in M, the set D as the set of all possible darts, the set HE as the set of all
possible half-edges and EE as the set of all possible edge-ends.
Example 2.2.3. The mesh defined in example 2.2.1 has 24 corners, and we define the
set of corners C:
C = {v2 , f1 }, {v6 , f1 }, {v5 , f1 }, {v1 , f1 }, {v1 , f2 }, {v3 , f2 },
{v7 , f2 }, {v5 , f2 }, {v1 , f3 }, {v3 , f3 }, {v4 , f3 }, {v2 , f3 },
{v6 , f4 }, {v8 , f4 }, {v7 , f4 }, {v5 , f4 }, {v4 , f5 }, {v8 , f5 },
{v6 , f5 }, {v2 , f5 }, {v3 , f6 }, {v7 , f6 }, {v8 , f6 }, {v4 , f6 }
(2.8)
2.2.3 The number of mesh elements
Some observations can be made about the different kind of mesh elements. However,
we must first define some classifications of the mesh elements. We classify each edge
to be a boundary or a non-boundary edge.
Definition 2.2.8. An edge is boundary if only one face abuts the edge. Otherwise, two
faces abut the edge and the edge is non-boundary.
We let the two sets BE and IE be the sets of boundary edges and non-boundary edges
respectively.
2.3. MESH QUERIES
17
A vertex can have both a edge degree or face degree. The edge degree is the number
of edges connected to the vertex, and the face degree is the number of faces abutting
the vertex. If the vertex is interior (all connecting edges are non-boundary edges), the
edge and face degree is the same.
Then to the observations. If the mesh is purely triangular, the number of corners
#C = 3#F. Likewise, if the mesh is purely quadrilateral, then #C = 4#F. If the
mesh is mixed with different kinds of polygons, the number of corners must by found
by summation. Each triangle contributes with three corners, each quadrilateral with
four, etc. In addition we have some invariants of the sizes of the sets of corners, edges
and darts.
Property 2.2.9. From the mesh M the set of corners C and the set of darts D may be
deduced. We partition the set of edges in two, into the set of boundary edges BE and
the set of internal (non-boundary) edges IE. The following is true:
1. #C = #BE + 2#IE
2. #D = 2#C = 2#BE + 4#IE
Proof. Each internal edge is surrounded by four corners, and each boundary edge is
surrounded incompletely by two corners. A corner abuts two edges, and this gives (1).
Each corner contains two unique darts (one along the incoming edge and one along the
outgoing edge), and this gives (2).
Sometimes, in sub- and superscripts of sums, the number of vertices in a mesh are
denoted Nv , the number of edges are denoted Ne , the number of faces Nf , and the
number of corners Nc , etc. The reason for this is purely aesthetically reasons. It is
also worth noticing that even though both circuits and shift invariant ordered sets are
in fact sets, we do not use the set font when dealing with edges and faces. The reader
may think that this inconsistency can be a cause of confusion, however, in our humble
opinion, the opposite is the case. Therefore, vertices, edges and faces are denoted in
the same font.
2.3 Mesh queries
This sections presents the results from applying the paradigm of relational algebra
(databases) on connectivity of geometrical models. This is indeed an unorthodox approach. However, we will see that this approach let us reap useful concepts.
In computational geometry we often need to do operations “for each abutting face”,
“for each face defined by four vertices where one vertex is boundary” and so on. These
verbose descriptions are (sometimes) intuitive, however, they lack the conciseness and
compactness of a mathematical notation. Ambiguities is especially a problem when implementing algorithms. Mesh queries is supposed to be a notation to write algorithms
of computational geometry in a compact and concise form.
2.3.1 Relational concepts
The idea is very based on relational algebra (used in database design). Relational algebra in a database setting is not the same as a mathematical binary relation. For
CHAPTER 2. MESHES
18
comprehensive information on relational algebra the reader is referred to one of the numerous books dealing with this topic, for example (Elmasri & Navathe 1994). However
we will give a extremely brief overview of the concept used in this section.
A relation can be viewed as a table. Each column of this table represent a property,
and is called an attribute. The class of data that can appear in a column is called the
domain of that attribute. Each row represents a collection of related data values and is
called a tuple. The tuple relates specific occurrences of the attributes together.
At least one attribute of an relation should be a primary key. If an attribute is the
primary key, it is an unique identifier for tuples of the relation. An attribute can be a
foreign key as well. This restricts the domain of the attributes in such a way that only
tuples with a value in the attribute field that already exists in a referenced attribute in
another relation are allowed.
The CREATE TABLE operation in SQL is used for relation creation:
CREATE TABLE <name of relation>
( <attribute, domain>, ...
PRIMARY KEY(<attribute>),
FOREIGN KEY(<attribute>)
REFERENCES <other relation>(<attribute>)
We will use three fundamental operations on the relations. The result of these operations are a new relation:
1. The select, denoted σ<condition> (R), extracts tuples (rows) from the relation R.
The result is all tuples satisfying the condition, and this is a subset of the population of R.
2. The project, denoted π<attributes> (R), extracts attributes (columns) of the relation
R. All tuples remain, but with fewer attributes.
3. The Cartesian product, denoted R1 × R2 , of the two relations R1 and R2 results
a composite relation with all the attributes from both relations and all possible
combinations of the tuples from the two relations.
The SELECT operation in SQL is a composition of these three operations. The syntax
is:
SELECT <attributes>
FROM
<relations>
WHERE <conditions>
Conceptually, the Cartesian product of all the relations combined are made. All tuples
not satisfying <condition> are discarded, as well as attributes not in <attributes>. In
relational algebra this expression is:
σ<condition> (π<attributes> (R1 × R2 × · · · × Rm )) .
(2.9)
with all possible combinations of the tuples from the two relations. thus if relation
A has nA tuples, and relation B has nB tuples, the Cartesian product has nA × nB
tuples. Join operations are Cartesian products with a selection applied (thus, some of
2.3. MESH QUERIES
19
2.3.2 The structure
The idea is to represent a mesh as a relational database, and then we can use select and
project to extract different subsets of the mesh. In this way, we define exact a great
variety of useful subsets. In fact, the mesh used for mesh queries does not be as strict
as definition 2.2.2. In this way, mesh queries could be used to define the restrictions of
this definition. However, this would result in a chicken-and-egg-issue (a mesh must be
defined for the queries to be defined).
If the application is divided in a geometry kernel (which deals with handling the
geometry) and an algorithms which use this library. The mesh query is a suitable candidate for defining a concise and clear interface between the kernel and the algorithms.
G-maps is another such interface. Neither G-maps nor mesh queries impose a particular
design for the kernel implementation, only a interface the kernel must satisfy.
The domains
We define three domains, the vertices, the edges and the faces. These domains are used
to identify a vertex, edge or face. We use three relations (notice that these relations does
not relate anything since they have only one single attribute each). The corresponding
SQL-statements can be:
CREATE TABLE Vertices
(vertex, int NOT NULL,
PRIMARY KEY(vertex));
CREATE TABLE Edges
(edge, int NOT NULL,
PRIMARY KEY(edge));
CREATE TABLE Faces
(faces, int NOT NULL,
PRIMARY KEY(face));
These three relations define only three lists, the list of vertices, the list of edges and the
list of faces:
Vertices
vertex
v1
..
.
Edges
edge
e1
..
.
Faces
face
f1
..
.
We will not dwell on how to populate these relations.
Relating elements of the domains
We use two relations to relate elements in the domains defined in the previous section.
One relation, VertexEdge, relates vertices and edges, and the other, EdgeFace, relates
edges and faces.
The VertexEdge relation has two attributes: vertex and edge. These two attributes
are foreign keys of the single attributes of the Vertices and the Edges relations respectively. Each tuple in this relation is a “vertex is in edge” (or conversely, “edge has
vertex”) statement.
CHAPTER 2. MESHES
20
The EdgeFace relation has two attributes: edge and face. These two attributes are
foreign keys of the the single attribute in the Edges and Faces relations respectively.
Each tuple in this relation is an “edge is in face” (or conversely, “face has edge”)
statement. SQL-statements for the two relations can be:
CREATE TABLE VertexEdge
(vertex, int NOT NULL,
edge,
int NOT NULL,
PRIMARY KEY(vertex,edge),
FOREIGN KEY(vertex) REFERENCES Vertices(vertex),
FOREIGN KEY(edge) REFERENCES Edges(edge));
CREATE TABLE EdgeFace
(edge, int NOT NULL,
face, int NOT NULL,
PRIMARY KEY(edge,face),
FOREIGN KEY(edge) REFERENCES Edges(edge),
FOREIGN KEY(face) REFERENCES Faces(face));
Which creates the following structure:
VertexEdge
vertex edge
vi
ej
..
..
.
.
EdgeFace
edge face
ej
fk
..
..
.
.
Here we see that vi is one of the two edges of ej , and ej is one of the edges defining
fk . When combining the two relations we know that vi is a vertex of fk . These two
relations contain all the connectivity information needed.
The virtual dart relation
Is indicated in the previous section, we can combine the two relations. In this way
we define the darts. The darts are a virtual relation. The population of this relation is
deduced from other relations.
We find the darts as the Cartesian product of the VertexEdge and EdgeFace relations, followed by an select and project operation:
CREATE VIEW Darts
AS
SELECT VertexEdge.vertex,
VertexEdge.edge,
EdgeFace.face
FROM
VertexEdge, EdgeFace
WHERE VertexEdge.edge = EdgeFace.edge
A virtual relation (called “view” in SQL) can be used as a regular relation in queries.
The SQL-statements produces the following structure:
Darts = πa (σ(VertexEdge × EdgeFace)c )
vertex edge
face
vi
ej
fk
..
..
..
.
.
.
2.3. MESH QUERIES
21
1
2
3
4
5
6
7
8
Figure 2.3: Dart queries. Enumeration refers to list given on page 22. A black circle is
a selected vertex, a black edge is a selected edge and a dark face is a selected face.
where a are the attributes from the SELECT statement and c is the condition from the
WHERE statement. The triple (vi , ej , fk ) is a dart in the mesh. The virtual dart relation
is the core of the mesh queries.
2.3.3 Dart selection and projection
A mesh query is the result from a dart selection and a dart projection on the dart virtual
relation.
Dart selection
We start with the selection part, which we call dart queries. The dart query defines the
condition for the selection, and the query picks out the matching darts.
Definition 2.3.1. A dart query (vi , ej , fk ) is a triple of a vertex, an edge and a face
where none, some or all of these three elements can be replaced by a wildcard denoted
“∗”. The dart query matches all darts which is composed of vi , ej and fk or all
vertices, edges or faces in the case of a wildcard.
CHAPTER 2. MESHES
22
In SQL the dart query (vi , ej , fk ) becomes
SELECT
FROM
WHERE
AND
AND
*
Darts
vertex = v_i
edge = e_j
face = f_k;
If some elements of the query are wildcards, the corresponding WHERE statements are
removed. Thus, the dart query (vi , ∗, ∗) becomes
SELECT *
FROM
Darts
WHERE vertex = v_i;
There is in fact eight classes of these queries (refer to figure 2.3):
1. (∗, ∗, ∗) selects all darts of a mesh.
2. (vi , ∗, ∗) selects all darts around vertex vi (it is the 0-orbit of a G-map).
3. (∗, ej , ∗) selects all darts around edge ej (it is the 1-orbit of a G-map).
4. (vi , ej , ∗) ≡ (vi , ∗, ∗) ∩ (∗, ej , ∗) selects all darts around edge ej at the side of
vi .
5. (∗, ∗, fk ) selects all darts of face fk (it is the 2-orbit of a G-map).
6. (vi , ∗, fk ) ≡ (vi , ∗, ∗) ∩ (∗, ∗, fk ) selects all darts of face fk at vi .
7. (∗, ej , fk ) ≡ (∗, ej , ∗) ∩ (∗, ∗, fk ) selects all darts of face fk along ej .
8. (vi , ej , fk ) ≡ (vi , ∗, ∗) ∩ (∗, ej , ∗) ∩ (∗, ∗, fk ) selects a single dart or none.
The queries 1,2,3 and 5 always match something. The other queries are called composite (since they can be written as an intersection) does not necessarily match something.
For example, the query (vi , ej , ∗) does not match anything unless vi is one of the two
edges of ej . Thus, these tests can be used to check connectivity relations (if a vertex
belongs to a face, an edge connects a vertex etc.).
Projection
A projection picks out certain attributes, and removes duplicate tuples in the result. In
this case, the tuples are a vertex index, an edge index and a face index.
Definition 2.3.2. A dart projection is defined by a subset of the set {v, e, f }. The
projection extracts some (or all) of the three attributes of a dart and removes duplicate
tuples.
There exist seven projections of darts (in fact there are eight, but the void projection is
completely useless):
1. {v} picks vertices.
2. {e} picks edges
2.3. MESH QUERIES
23
3. {v, e} picks vertex-edge combinations (known as edge-ends).
4. {f } picks faces.
5. {v, f } picks vertex-face combinations (known as corners).
6. {e, f } picks edge-face combinations (known as half-edges)
7. {v, e, f } picks vertex-edge-face combinations (known as darts).
Each of these projections has a name (like corner or half-edge). We use these terms
when indexing the mesh (“for each corner in the mesh”, “for each half-edge”, “for each
vertex”).
Combining projection and selection
We combine a dart selection with a successive dart projection on the dart relation to
create a mesh query.
Definition 2.3.3. We query the mesh M with the following notation
M[projection; selection],
(2.10)
where selection is a dart selection (defined in definition 2.3.1) selecting a subset of
darts and projection is a dart projection (defined in definition 2.3.2) to extract the
desired kind of unique tuples.
The result is a set of tuples satisfying the requirement of the selection, with the attributes specified by the projection. In SQL the mesh query is the following select
statement:
SELECT DISTINCT <projection>
FROM Darts
WHERE <selection>;
Mesh query example
We conclude this chapter with some examples of mesh queries.
Example 2.3.1. We use the mesh defined in example 2.2.1. The sets of vertices (equation 2.2), edges (equation 2.3) and faces (equation 2.4) can be found by the following
mesh queries:
V = M[v; ∗, ∗, ∗],
E = M[e; ∗, ∗, ∗] and
F = M[f ; ∗, ∗, ∗].
(2.11)
Further, we may find the set of all the corners C (equation 2.8):
C = M[v, f ; ∗, ∗, ∗].
(2.12)
{f1 , f2 , f3 } = M[f ; v1 , ∗, ∗].
(2.13)
The faces abutting v1 :
CHAPTER 2. MESHES
24
Edge in
Edge out
Edge in
Edge out
Edge out
Edge out
Edge in
Edge in
Edge in
Edge in
Edge out
Edge out
Figure 2.4: Consecutive ordering of corners around a vertex (left) and of a face (right).
The edge in of one corner must be equal to the edge out of the next corner.
The vertices of edge e1 :
{v1 , v2 } = M[v; ∗, e1 , ∗].
(2.14)
We can define integrity constraints. The condition
0 < #M[f ; ∗, ej , ∗] ≤ 2,
∀ej ∈ E
(2.15)
requires all edges to be used, and by two faces at most. Further,
0 ≤ #M[e; vi , ∗, ∗] − #M[f ; vi , ∗, ∗] ≤ 1,
∀vi ∈ V
(2.16)
requires a simple boundary.
2.4 Mesh modification and refinement
A refinement procedure is always employed in subdivision surfaces or surface splines.
The two basic operations are known under various names, but we refer to them as the
face split and the vertex split. Before diving into the details of the different schemes,
we define a property, which we call consecutive ordering of corners (see figure 2.4).
This definition is just convenient when describing the splitting procedures.
Definition 2.4.1. A consecutive set of corners is an offset invariant ordered set where
the edge out of one corner is equal to the incoming edge of the next corner.
In addition, we assume that all meshes are closed in this section. Thus, we do not
present any rules for dealing with boundary elements.
2.4.1 Dual mesh
A mesh has a dual where faces and vertices occupy complimentary locations — analogous to duals of polyhedra. A point can be defined by the intersection of some planes,
and a plane can be defined by some points. In this sense, the point and the plane have
2.4. MESH MODIFICATION AND REFINEMENT
Face split
25
Vertex split
Figure 2.5: The two major splitting approaches. Face split (lower left): Each n-gon
is split into n quadrilaterals. Vertex split (lower right): Each vertex of n-degree is
removed and replaced with an n-gon.
a relationship of duality. The operations of defining these two kinds of objects are regarded as complementary dual operations. The dual of a polyhedron is defined by the
exchange of these two dual elements.
Definition 2.4.2. Vertices and faces are dual elements. Let a vertex be defined by
abutting faces, and a face defined by vertices along the perimeter. These two operations
are dual operations. The dual of a mesh is a mesh built with the dual elements and dual
operations of this mesh.
In this way, the vertices of one mesh corresponds to the faces of it’s dual and vice versa.
We can find the dual B of a mesh A by defining a face for each vertex and defining a
vertex for each face: For each face in A, we define a vertex in B. For each vertex in A,
we extract the fan of abutting corners in consecutive order, and use the vertices in B
corresponding to the face of the corners to define a new face in B. Each pair of adjacent
corners in this ordering defines an edge between the vertices of their respective faces.
We see from the definition that the corner of a mesh represents a kind of “pivot”
between the duals. The number of corners remains constant. The dual mesh of the
dual mesh is the mesh itself. We use the dual mesh of the quad mesh to define the inner
mesh for the simplified surface spline.
2.4.2 Face split (polyhedral split)
Most subdivision schemes can be split into two operations: a splitting step and a
smoothing step. The splitting step refines the mesh, that is, it produces a new mesh
with more and smaller faces. Ideally, the splitting step should not change the geometry
of the shape. This is the case if the embedded mesh is a polyhedron (which has planar
CHAPTER 2. MESHES
26
faces). The smoothing step generates a new set of points, where each point is an average of some of the points produced by the splitting step. We ignore the smoothing step
for now, and focus on the splitting step. For more information on both steps, (Stollnitz,
DeRose & Salesin 1996) is a good source.
We begin with a subdivision scheme known as both the “face split” and the “polyhedral split”. This scheme splits each face into several new faces. We call the original
mesh Mi and the refined mesh Mi+1 .
Vertex and point generation First, we define the vertices of Mi+1 and their positions.
1. For each vertex in Mi , create a vertex in Mi+1 with the same location.
2. For each edge in Mi , create a vertex in Mi+1 with the average of the endpoints
of the edge as location.
3. For each face in Mi , create a new vertex in Mi+1 with the average of the vertices
of the face as location.
The number of vertices in the new mesh is equal to the sum of vertices, edges, and
faces of the old mesh.
Defining the new edges The edges are defined implicitly by the face descriptions,
but for completeness, we define them here.
1. For each edge in Mi , create two edges in Mi+1 . These connect to the vertices in
Mi+1 corresponding to the end points of the edge as well as the edge midpoint.
2. For each edge in Mi , create an edge in Mi+1 for each abutting face. This edge
connects the edge midpoint to the face midpoint.
The number of edges is equal of the total sum of corners in the old mesh plus two times
the number of edges in the old mesh. If the mesh is closed, as we assume, the number
of corners are 2 · #E, and thus the total number of new edges are 4 · #E.
Defining the new faces For each corner in Mi , create a quadrilateral in Mi+1 with
the edges:
1. The edge from the vertex of the corner to the midpoint of the outgoing edge.
2. The edge from the midpoint of the outgoing edge to the face midpoint.
3. The edge from the face midpoint to the midpoint of the incoming edge.
4. The edge from the incoming edge to the vertex of the corner.
The number of faces in the new mesh is equal to the total sum of corners in the old
mesh. The subdivision scheme is visualised in figure 2.5. Meshes generated by the
face split scheme possess some useful properties:
2. The vertices from Mi keep their degree.
2.4. MESH MODIFICATION AND REFINEMENT
27
3. The degree of a midpoint of an n-gon in Mi , has a degree of n. Thus, midpoints
of quadrilaterals are of degree 4.
4. The degree of edge midpoints are 4.
Thus, after one application of this scheme the quads of the refined mesh may have more
than one vertex of degree not equal to 4.
If all vertices in the original mesh with a degree other than 4 are only surrounded
by quadrilaterals, or the converse, every non-quadrilateral face in the mesh are defined
only by vertices with degree 4, a quad in the refined mesh have at most one vertex
with degree not equal to 4. However, if these prerequisites are not fulfilled, two successive applications of the splitting procedure always generate a refined mesh with this
property.
2.4.3 Vertex split (mid-edge subdivision)
The face split splits faces, and the vertex split splits, as the name implies, vertices.
The vertex split (also known as mid-edge subdivision) is very simple, and the scheme
proceeds as follows:
Vertex and point generation For each edge in Mi , create a vertex in Mi+1 and
position this vertex at the midpoint of the edge. We call this point the midpoint of the
edge.
Defining the new edges For each vertex in Mi , find the set of corners abutting the
vertex. Create an edge in Mi+1 for each of these corners, connecting the midpoint of
the incoming edge to the midpoint of the outgoing edge.
Defining the new faces Each face and each vertex in Mi generates a face in Mi+1 .
1. For each vertex, extract the corners abutting that vertex and organise them as
a consecutive set of corners. Then define a face by the edges connecting the
midpoints of the incoming and outgoing edge each corner.
2. For each face in the unrefined mesh, extract the corners and organise them and
build a face as done in the previous step.
The number of vertices in the new mesh is thus equal to the number of edges in the old
mesh, the number of edges in the new mesh is equal to the number of corners in the old
mesh, and the number of faces in the new mesh is the sum of the faces in the old mesh
plus the number of vertices in the old mesh. The refined mesh has some interesting
properties:
1. All vertices of the refined mesh have a degree of 4 (two faces and two vertices
abut an edge).
2. An n-gon in the unrefined mesh becomes an n-gon in the refined mesh.
3. A vertex of degree n becomes an n-gon in the refined mesh.
28
CHAPTER 2. MESHES
2.4.4 The Doo-Sabin split
The Doo-Sabin split is a bit more elaborate than the two previous subdivision schemes.
It splits both vertices and edges, while keeping faces. The algorithm proceeds as follows:
Vertex generation For each corner in Mi , create a vertex in Mi+1 . The position
of this vertex is defined by a linear interpolation between the position of the corner’s
vertex (in Mi ) and the midpoint of the face (average of the face’s vertices in M i ). The
number of vertices in Mi+1 equals the number of corners in Mi . The parameter of this
linear interpolation defines the sharpness (analogous to blend ratios) when building
Doo-Sabin subdivision surfaces.
Defining the new edges Each edge in Mi is abutted by four corners. We connect the
midpoints (vertices in Mi+1 ) with four edges and the result is a quadrilateral centred
on the edge in Mi . The number of edges in Mi+1 is four times the number of edges in
Mi .
Defining the new faces Each vertex, each edge, and each face define a face in the
new mesh. The midpoint of the abutting corners define a new face, thus, a vertex in
Mi of valence n defines an n-gon in Mi+1 . Each edge will form a quadrilateral, while
each face will keep their number of vertices. The number of faces in M i+1 equals the
sum of the number of vertices, the number of edges, and the number of faces in M i .
There is a connection between the polyhedral split and the Doo-Sabin split. Both
schemes produces the same number of edges, and the Doo-Sabin split produces the
same number of vertices as the number of the polyhedral split produces faces, and vice
versa. This is indeed a hint that the results from the polyhedral split and the Doo-Sabin
split are duals of each other.
For a more extensive exposition of splitting procedures, as well as subdivision surfaces, some suggested books are (Farin 1996), (Gallier 2000) and (Stollnitz et al. 1996).
2.4.5 Planar cut polyhedra
Planar cut polyhedra (PCP) is a central component of surface spline schemes. Often,
the term “polyhedron” is used quite loosely about anything built from polygonal faces.
However, for such a shape to be a polyhedron, it is required that every face is planar.
This is of course not an issue when dealing with triangles, since they are planar when
not degenerate. To emphasise this planarity, we call such a shape a planar polyhedron
when this is of importance.
We get a planar cut polyhedron from a planar polyhedron by chopping off the shape
at each vertex with a plane. The amount of the shape cut away is called the depth of the
cut. The planar cut polyhedron can be viewed as a generalisation of the tensor product
control structure, where each planar cut represents a subdivision step.
If an embedded mesh is equivalent regarding connectivity (and not geometry) to a
planar cut polyhedron, we can sometimes make this shape into a planar cut polyhedron
by utilising some local projections. This is always possible if the mesh is local degree
bounded. This requires that at most one non-planar face abuts a vertex. If two or
more offending faces abut, then we get a dependency in the projections, which are not
necessarily solvable (one projection destroys the planarity of the other).
2.5. SUMMARY
29
Definition 2.4.3. A polyhedron is local degree bounded if all non-quadrilateral faces
of a polyhedron is only defined from vertices where only four faces abut, or conversely,
all vertices surrounded by more or less than four faces is only surrounded by quadrilateral faces.
If a polyhedron is not local degree bounded, it can be made local degree bounded by
refinement.
2.5 Summary
In this chapter we have defined the mesh. Meshes are a formal way to define connectivity of geometric objects. Vertices are the fundamental atom of meshes, an edge is a
set of two vertices and faces are sets of edges. We also defined some terms of positions
in the connectivity.
Then we gave the background and definition of the mesh queries. The qualities
of mesh queries will be obvious in chapters 5 and 6, in which mesh queries are used
extensively.
We concluded the chapter with an overview of mesh duals and some splitting
schemes. Both of these topics are essential in the construction of surface splines.
30
CHAPTER 2. MESHES
Chapter 3
The surface spline curve
What happens if we apply the surface spline paradigm to curves? Surface spline curves
are smooth, modelled by the control polygon, can interpolate data points with some
modifications, and have a simple construction. A study of the simple case of the curve
is a gentle introduction to the surface spline paradigm, and in addition, this curve defines the boundary curves of a simplified surface spline surface.
3.1 The scheme
Surface splines have a peculiar parameterisation, and the same is true for surface spline
curves. The curves are parameterised by two numbers: an integer index, and a scalar
running from zero to one. The index specifies in which sub-interval we are in, and the
scalar specifies the position in this sub-interval. For visualisation of the construction
we arrange the control points along the real axis.
Definition 3.1.1. A surface spline curve is defined by a vector of m control points,
c = c1 , c2 , . . . , cm ,
ci ∈ Rn ,
(3.1)
and a vector of 2m − 2 blend ratios,
− +
−
+
−
b = b+
1 , b2 , b2 , . . . , bm−1 , bm−1 , bm ,
bi ∈ [0, 1],
(3.2)
organised in the following manner:
interval 1
interval 2
interval 3
− •
+
− •
+
• b+
•
← b2
b2 →
← b3
b3 →
← b−
1 →
4
c1
c2
c3
c4
+
Each interior ci has two blend ratios associated with it, b−
i and bi , and c1 and cm
has one blend ratio each. The control point array defines m − 1 intervals, one interval
between each successive pair of control points: Interval i is between c i and ci+1 . Each
interval defines two sub-intervals, one odd and one even. The odd sub-interval is the
left half of the interval, and the even sub-interval is the right half. Blend ratios and
sub-intervals have a one-to-one correspondence, and use the same indices.
31
CHAPTER 3. THE SURFACE SPLINE CURVE
32
3.1.1 Refinement
For every interval i we introduce a midpoint, mi defined by
mi = 1/2(ci + ci+1 ).
(3.3)
These midpoints split every interval into the two sub-intervals.
•
•
?
?
•
?
•
interval 1
interval 2
interval 3
sub 1
sub 2
sub 3
sub 4
sub 5
sub 6
•
•
•
•
?
?
?
m1
m2
m3
c1
c2
c3
c4
We have so far done nothing extraordinary. The midpoints are the averages of the
control points at the extremities of the intervals, and the geometry of the curve is not
changed.
3.1.2 Sub-interval midpoints
We introduce yet another level of midpoints, the 2m − 2 sub-interval midpoints. These
are just logical midpoints, not necessarily geometrical. For each sub-interval i we
define the sub-interval midpoint q i as
+
+
q+
i = bi ci + (1 − bi )mi
−
−
q−
i = bi ci + (1 − bi )mi−1
(3.4)
3.1. THE SCHEME
33
Each inner control point and interval midpoint is surrounded by two sub-interval midpoints. We get a constellation looking like
?
•
•
?
?
•
•
interval 1
interval 2
interval 3
sub 1
sub 2
sub 3
sub 4
sub 5
sub 6
•
•
•
•
?
?
?
−
−
+ m
−
+ m
m
c1 q +
c
c
c
1 q2
q2
2 q3
q3
3 q4
2
3
4
1
If all blend ratios are equal to 12 , the q’s are the geometric averages of their two respective neighbours. If blend ratios are small, the sub-interval midpoints are located
closer to the control points, and conversely, when blend ratios are large, the sub-interval
midpoints are located closer to the interval midpoints. Regardless of blend ratios, the
sub-interval midpoints are just convex combinations of an interval midpoint and a control point, hence it lies on the line between these two points. The geometry is not
changed in this step either.
3.1.3 Bézier coefficients
Then we change the geometry; we generate new positions for the control points and
the interval midpoints. The collection of new control points, new interval midpoints
and the sub-interval midpoints form the Bézier coefficients of the curve. The new
control points are denoted ĉi and the new interval midpoints m̂i , and are defined by
the averages
ĉ1 = c1
+
ĉi = (q −
i + q i )/2
ĉm = cm
−
m̂i = (q +
i + q i+1 )/2
i = 2, . . . , m − 1
i = 1, . . . , m − 1
(3.5)
CHAPTER 3. THE SURFACE SPLINE CURVE
34
If we connect each subsequent coefficient with a line, we get
?
•
•
?
?
•
•
interval 1
interval 2
interval 3
sub 1
sub 2
sub 3
sub 4
sub 5
sub 6
•
•
•
•
?
?
?
+
−
+
−
+
−
ĉ1 q 1 m̂1 q 2 ĉ2 q 2 m̂2 q 3 ĉ3 q 3 m̂3 q 4 ĉ4
This step is important. If the last control point of one segment equals the first control point of another segment, the two segments join continously (C 0 ). In addition, if
the next-to-last and last control point of the first segment are coplanar with the first
and second control point of the second segment, they joint tangent continously (G 1 ).
Even further, if the distances between the next-to-last and last control points of the first
segment equals the distance between the first and second control point of the second
segment, the derivatives of the two segments match, and thus the segments join C 1 .
And the equation 3.5 makes the curve satisfy these requirements. The geometry is
changed in this step; we get a smoothed version of the original control curve, though it
interpolates the control curve at the first and at the last control points.
3.1.4 Defining the Bézier curve segments
Each sub-interval defines a quadratic Bézier curve segment, denoted s i (u). Each of
these curve segments are defined by three points.
(
2
B02 (u)cj + B12 (u)q +
j = 1 + (i − 1)/2 if i is odd
i + B2 (u)mj ,
+
2
2
2
if i is even,
B0 (u)mj + B1 (u)q i + B2 (u)cj+1 , j = i/2
(3.6)
2
where B0,1,2
are the Bernstein polynomials defined in definition 1.4.10. This produces
the curve
si (u) =
3.1. THE SCHEME
35
IV
III
II
I
c1
cˆ1
B02 (u)
s1 (u)
b+
1
q+
1
B12 (u)
B22 (u)
1
2
1−b+
1
1
2
m1
m̂1
B02 (u)
s2 (u)
1−b−
2
1
2
q−
2
B12 (u)
B22 (u)
1
2
b−
2
1
2
c2
cˆ2
s3 (u)
b+
2
1
2
B02 (u)
q+
2
B12 (u)
B22 (u)
1
2
1
2
1−b+
2
m2
m̂2
s4 (u)
1−b−
3
1
2
B02 (u)
q−
3
B12 (u)
B22 (u)
1
2
b−
3
1
2
c3
cˆ3
1
2
B02 (u)
s5 (u)
b+
3
q+
3
B12 (u)
B22 (u)
1
2
1−b+
3
1
2
m3
m̂3
s6 (u)
1−b−
4
1
2
B02 (u)
q−
4
B12 (u)
1
2
b−
4
B22 (u)
cˆ4
Figure 3.1: Layout of surface spline curve construction
c4
CHAPTER 3. THE SURFACE SPLINE CURVE
36
•
c2
•
q−
2
m1 = m̂1
q+
1
ĉ2
ˆ
•
?
q+
2
?
m2 = m̂2
q−
3
?
ˆ
•
ĉ3
c4 = ĉ4
q−
4
m3 = m̂3
q+
3
•
c3
c1 = ĉ1•
Figure 3.2: The surface spline curve of example 3.1.1.
?
•
•
?
?
•
•
interval 1
interval 2
interval 3
sub 1
sub 2
sub 3
sub 4
sub 5
sub 6
•
•
•
•
?
?
?
−
+
−
+
−
ĉ1 q +
ĉ
ĉ
ĉ
m̂
m̂
m̂
q
q
q
q
q
2
3
4
1
2
3
1
2
2
3
3
4
3.1.5 Summary of construction
A systematic overview of the complete scheme is given in figure 3.1. The scheme
consists of four distinct stages (from right to left). At stage I, the interval midpoints are
created. At stage II the subinterval midpoints are created, utilising the blend ratios. At
stage III smoothing is performed to assure the C 1 continuity requirement of composite
Bézier curves, and at stage IV conventional quadratic Bézier evaluation is done. The
dashed lines represent “keeping a value” to be used later in the scheme.
3.1. THE SCHEME
37
Example 3.1.1. Find the surface spline curve defined by the four points in the plane,
c = c1
c2
c4 =
c3
0
1
2
3
,
0
2
1
3
(3.7)
and the blend ratios
b = b+
1
1
= 2
b−
2
1
2
b+
2
1
2
b−
3
1
2
b+
b−
3
4
1
2 .
1
2
First, we find the interval midpoints
m = m1 m2 m3
= 21 (c1 + c2 ) 21 (c2 + c3 )
"" # " # " ##
1
1 21
2 21
2
=
.
1
1 21
2
1
2 (c3
(3.8)
+ c4 )
(3.9)
We then find the sub-interval midpoints, and since all blend ratios are equal to one half,
these become simple averages,
q = q+
q−
q+
q−
q+
1
2
2
3
3
 +
T
+
b1 c1 + (1 − b1 )m1
(1 − b− )m + b− c 
1

2
2 2
 +

+
b2 c2 + (1 − b1 )m2 


=
−
− 
(1 − b3 )m2 + b3 c3 
 +

b3 c3 + (1 − b+

1 )m3
−
(1 − b−
4 )m3 + b4 c4
1
T
2 (c1 + m1 )
 1 (m + c )
1
2 
2
1

 2 (c2 + m2 )


= 1

 2 (m2 + c3 )
1

 2 (c3 + m3 )
=
1
(m3 + c4 )
""2 # " # "
1
4
1
2
3
4
1 21
1 14
1 34
# "
1 43
1 41
q−
4
(3.10)
#
"
2 41
1 12
# "
2 43
2 21
##
.
and from this the smoothed control points,
ĉ = ĉ1 ĉ2 ĉ3 ĉ4
+
−
+
1
= c1 12 (q −
2 + q2 )
2 (q 3 + q 3 ) c4
"" # " # " # " ##
0
1
2
3
=
,
5
3
0
3
18
18
(3.11)
CHAPTER 3. THE SURFACE SPLINE CURVE
38
•c4
c2
•
cubic Bézier curve
surface spline curve
•
c3
c1•
Figure 3.3: Comparison between a four-point surface spline curve and a cubic Bézier
curve.
and the smoothed interval midpoints,
m̂ = m̂1 m̂2 m̂3
−
−
1
= 21 (q +
(q +
1 + q2 )
2 + q3 )
"" # " # 2 " ##
1
1 21
2 21
1
2
,
=
1
2
1
12
2
+
1
2 (q 3
+ q−
4 )
(3.12)
and this defines six quadratic Bézier segments,
2
s1 (u) = B02 (u)ĉ1 + B12 (u)q +
1 + B2 (u)m̂1
2
s2 (u) = B02 (u)m̂1 + B12 (u)q +
2 + B2 (u)ĉ2
2
s3 (u) = B02 (u)ĉ2 + B12 (u)q +
3 + B2 (u)m̂2
2
s4 (u) = B02 (u)m̂2 + B12 (u)q +
4 + B2 (u)ĉ3
(3.13)
2
s5 (u) = B02 (u)ĉ3 + B12 (u)q +
5 + B2 (u)m̂3
2
s6 (u) = B02 (u)m̂3 + B12 (u)q +
6 + B2 (u)ĉ4
The collection of these curve segments is the surface spline curve.
The curve is rather smooth and bears resemblance to a traditional spline. Figure 3.3
compares the surface spline curve to a cubic Bézier curve with the same coefficients.
The surface spline curve is closer to the control polygon.
3.2. EXPLICIT FORMULAS
39
3.2 Explicit formulas
For such a simple scheme as the surface spline curve scheme it is easy to deduce an
explicit formula for the curve.
Proposition 3.2.1. The Bézier coefficients of the surface spline curve can be found by
the expressions
cˆ1 = c1
cˆm = cm
1
1
1
1
−
+
1 + (b+
cˆi = (1 − b−
i )ci−1 +
i + bi ) ci + (1 − bi )ci+1
4
2
2
4
1
1
1
1
−
+
2 + (b−
m̂i = 1 + (b+
i − bi+1 ) ci +
i+1 − bi ) ci+1
2
2
2
2
1
1
+
+
+
q i = (1 + bi )ci + (1 − bi )ci+1
2
2
1
1
−
−
q i = (1 + bi )ci + (1 − b−
i )ci−1 .
2
2
(3.14)
The Bézier coefficients are convex combinations of the control points if the blend ratios
are values in [0, 1].
Proof. The formulas can easily be deduced from the construction (see figure 3.1). If
the blend ratios are kept within [0, 1], we see that none of the weights in the expressions become negative. Expanding the expressions, we see that the weights in each
expression sum to one.
3.3 Exists a subdivision formula?
The existence of a simple subdivision formula considerably increases the value of a
curve or surface algorithm. The “LeSS” (Gonzalez & Peters 1999) construction (see
section 4.2.2) does not give nested spaces. However, it is worthwhile to see whether
this is possible.
3.3.1 Bézier segment subdivision
Our starting point will be the refinement formulas for Bézier segments. From the literature (e.g. (Farin 1996) and (Gallier 2000)) we know that we can rewrite one single
Bézier curve as two new ones.
We split the quadratic Bézier segment defined by the Bézier control points {x1 , x2 , x3 }
into two equal halves, the two segments defined by the Bézier control points { x̂1,1 , x̂1,2 , x̂1,3 }
and {x̂2,1 , x̂2,2 , x̂2,3 }
x̂1,1 = x1
1
x̂1,2 = x1 +
2
1
x̂1,3 = x1 +
4
and
1
x2
2
1
1
x2 + x3 ,
2
4
(3.15)
CHAPTER 3. THE SURFACE SPLINE CURVE
40
Refined
..
.
..
.
Coarse
1
?
1
?
1
1
•
1
•
1
2
1
4
1
2
2
?
1
2
1
2
1
4
1
2
•
1
?
2
1
2
1
4
1
2
3
?
1
2
1
2
1
4
1
2
4
•
1
•
1
?
1
?
..
.
..
.
Figure 3.4: Inserting one control point.
3
3.3. EXISTS A SUBDIVISION FORMULA?
1
x1 +
4
1
= x2 +
2
= x3 .
x̂2,1 =
x̂2,2
x̂2,3
41
1
1
x2 + x3
2
4
1
x3
2
(3.16)
This will use this relation of Bézier segments to relate the coarse and the refined surface
spline curve.
3.3.2 Surface spline subdivision
To keep things simple, we let the blend ratios equal one half, and we write the problem
as a matrix equation. We let ci be the control points of the surface spline curve at
subdivision level i, and ci+1 be the control points at level i + 1.
The matrix A finds the Bézier coefficients from the surface spline curve control points. Thus, A represents the steps “refinement”, “sub-interval midpoints”, and
“Bézier coefficients” combined. We let A i be the matrix that transforms the control
points at level i to the Bézier coefficients at level i, and Ai+1 be the similar matrix at
level i + 1.
The matrix Ri refines the Bézier control points at level i to Bézier control points at
level i + 1 (this expression is visualised in figure 3.4). This gives us
Ai+1 ci+1 = Ri Ai ci ,
(3.17)
ci+1 = (Ai+1 )−1 Ri Ai ci ,
{z
}
|
(3.18)
which would give
subdivision rule
the subdivision rule we are looking for. However, A is not square, hence this rule
does not necessarily exists. We will not try to find (A i+1 )−1 Ri Ai , only see if this
expression in fact exists.
3.3.3 Futile attempt no. 1: inserting one point
Initially, let’s try to insert one single control point into the coarse curve. Given a curve
with n control points, exists a curve with n + 1 control points having the exact same
geometry? That is, does the curve with n control points live in the space of the curve
with n + 1 control points?
Then, let us start with the simplest scenario. Let the coarse curve consist of two
control points. We want to refine this curve by inserting a new control point into the
single interval of this curve. Using the relation
Ai+1 xi+1 = RAi ci ,
(3.19)
CHAPTER 3. THE SURFACE SPLINE CURVE
42
we get

1
3
4
1

2
1
4
1

8

0

0


0
0
0
1
4
1
2
3
4
3
4
3
4
1
2
1
4
0
0


1
1

0
2
1



0
4
  i+1  
0  c1
0


1   i+1 
0
=
c
8  2 


1
i+1
0

c
3
4

1
0
2



3
0
4
1
0
0
0
0
1
2
1
2
1
2
0
0
1
4
1
2
0
0
1
0
0
0
1
2
1
4
0
0
1
2
1
2
1
2
0
0
0
0
0


0

0
 1

0   34
1

0
2
1
0  4

1
4 0
1
2
1
0

1
4
 i
1  c1
2  ci .
2
3
4
(3.20)
1
We multiply together the matrices on the right-hand side

1
3
4
1

2
1
4
1

8

0

0


0
0
0
1
4
1
2
3
4
3
4
3
4
1
2
1
4
0
0


ci1

7 i
0
 8 c1 +

3 i
 c1 +
0

4
  i+1   5 i
0 c
 8 c1 +
 1
1 i
1   i+1 
=
8  c2
 2 c1 +
3 i
1  ci+1
3

 8 c1 +
4
1 i
1
 c +
2
4 1

1 i
3

 8 c1 +
4
1
ci2

1 i
8 c2 

1 i
4 c2 
3 i
8 c2 

1 i
2 c2 
5 i
8 c2 

3 i
4 c2 
7 i
8 c2 
(3.21)
and perform Gaussian elimination,

1
0

0

0

0

0

0

0
0
0
1
0
0
0
0
0
0
0



ci1
0
 1 ci1 + 1 ci2 
0
2

2




0
0 



i+1



0
0  c1


i+1 
i




1  c2
c2
=

i+1



0
0  c3





0
0





0
0
0
0
(3.22)
which reveals that indeed this expression has a solution! The solution is ci+1
= ci1 ,
1
i+1
i+1
1 i
1 i
i
c2 = 2 c1 + 2 c2 and c3 = c2 . This does not really come as a surprise because of
the convex hull property. A two point surface spline curve is just a line segment.
We increase the number of intervals in the coarse mesh to two, and insert a point
into one of these intervals. After right-hand side multiplication and Gaussian elimina-
3.3. EXISTS A SUBDIVISION FORMULA?
tion, we get

1
0

0

0

0

0

0

0

0

0

0

0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0



0
ci1
 1 ci1 + 1 ci2 
0
2


 2



0
0





0
0


i



c2
0  i+1  

c
1



0
0  i+1  

c2   i
i
i 


0  i+1  =  c1 − 2c2 + c3 
c3
 ci1 − 2ci2 + ci3 
0

 ci+1
1 i

 c1 − 1 ci2 + 2ci3 
4
1
2

2


ci1 − ci2
0



i
i



0
c1 − c2


i
i



c1 − c2
0
i
i
c1 − c2
0
43
(3.23)
The last four equations requires ci1 and ci2 to be equal, and thus this system does not
have a solution for arbitrary control points at level i.
We conclude that, while keeping blend ratios equal to 1/2, we cannot insert a new
control point and get the exact same geometry.
3.3.4 Futile attempt no. 2: uniform refinement
Inserting a point in one of the intervals failed, so instead we try to insert a point into
every interval. Given a surface spline curve with n control points, can we find a curve
with 2n − 1 control points with the exact same geometry? The initial scenario, refining
a curve with two control points, is exactly the same as in the previous section — the
curve has only one interval to be refined. And this problem did in fact have a solution.
Then, let us try to refine a curve with three control points, that is, a curve with two
intervals. We insert a new control point in the middle of each interval. After right-hand
side multiplication and Gaussian elimination we get




ci1
1 0 0 0 0
1 i
1 i

0 1 0 0 0

2 c1 + 2 c2





0 0 0 0 0

0





0 0 0 0 0

0




i

0 0 1 0 0

c
2





0 0 0 0 0

0
  i+1   i


 c1 − 2ci2 + ci3 
0 0 0 0 0 c1
 i
 i+1


0 0 0 0 0 c2   c1 − 2ci2 + ci3 
  i+1   1 i


0 0 0 1 0 c  =  c1 − 1 ci2 + ci3 
(3.24)
2
  3i+1   2 i


0 0 0 0 0 c   c1 − 2ci2 + ci3 
 i
 4


 c1 − 2ci2 + ci3 
0 0 0 0 0 ci+1
 i
 5


 c1 − 2ci2 + ci3 
0 0 0 0 0




−3ci1 + 6ci2 − 2ci3 
0 0 0 0 1
 i



 c1 − 2ci2 + ci3 
0 0 0 0 0
 i



 c1 − 2ci2 + ci3 
0 0 0 0 0
 i



 c1 − 2ci2 + ci3 
0 0 0 0 0
0 0 0 0 0
ci1 − 2ci2 + ci3
which requires 2ci2 to equal ci1 − +ci3 , and thus we do not have a solution for arbitrary
control points at level i.
44
CHAPTER 3. THE SURFACE SPLINE CURVE
We conclude therefore that, when blend ratios are equal to 1/2, we cannot insert
a new control point in between each adjacent pair of control points and get the exact
same geometry. However, allowing blend ratios to change could possibly give us a
subdivision formula.
Our conclusion is therefore that the surface spline curve does not have a general
subdivision formula, and this implies that that the simplified surface spline scheme
does not have a subdivision formula either.
3.4 Summary
The surface spline curve is the result of applying the surface spline paradigm on a
curve. We defined the steps of the curve and investigated the steps of the construction.
In addition, we tried to find a refinement rule for surface spline curves, however this
This chapter serves two purposes; primarily it is an intuitive introduction to surface
spline schemes, since the case of the curve are simpler than the case of the surface, and
requires substantially less book-keeping. Secondly, it is used to define the boundary
curve of the simplified surface spline curve.
Chapter 4
Surface spline schemes
Surface spline schemes are methods for obtaining a patchwork of traditional triangular
or rectangular B-spline patches, with smoothness constraints fulfilled over the joints,
from a control mesh of arbitrary topology. The surface spline is the resulting surface
defined by this patchwork. The surface spline scheme tries to:
1. Create a surface over an arbitrary parameter domain, where the surface satisfies
smoothness constraints.
2. Represent these patches as a patchwork of conventional patches, so that the
scheme can be easily integrated into existing frameworks and pipelines.
3. Achieve minimal polynomial degree for each patch.
4. Have a local construction, so that no large linear system has to be solved.
Blend ratios are used to control the tightness of the surface (see figure 4.1 on page
46). The blend ratios play the same role for surface splines as knot intervals do for
B-splines.
If the control mesh of the surface is a planar polyhedron (see section 2.4.5), we can
usually make a planar cut polyhedron from it, and from this we can always define a
C 1 surface of Bézier patches (Peters 1995b). Thus, the surface spline must, unless the
control polyhedron is a suitable polyhedron, perform a refinement (to isolate offending vertices) and perform a local planarisation. Then we can define the surface as a
composite set of patches.
4.1 The C 1-surface spline
The C 1 -surface spline scheme (Peters 1995a) is a generalisation of the biquadratic
tensor product B-spline. The input mesh topology gives the layout of the “knot vectors”
and the blend ratios describe the knot spacing. This defines a vector space, where the
points associated with the vertices of the control mesh are coefficients relative to some
basis, and thus specifies a particular spline inside this space. The resulting surface is
described as a collection of Bézier patches that are either rectangular, triangular or a
mix. If the mesh is composed of triangular or a mix of rectangular or triangular patches,
the surface is always G1 . The input of this scheme consists of three ingredients:
45
46
CHAPTER 4. SURFACE SPLINE SCHEMES
Figure 4.1: Different blend ratios. The simplified surface spline scheme was used to
generate these images. In upper left the blend ratios are all equal to 0. Upper right,
blend ratios are 0.25, lower left, 0.5, and lower right, 0.75.
4.1. THE C 1 -SURFACE SPLINE
47
Figure 4.2: The tree meshes of the C 1 -surface spline scheme. Left: The control mesh,
Middle: The quad mesh, and right: The inner mesh.
1. A control mesh. This mesh must satisfy the local degree boundedness property
(see definition 2.4.3 on page 29).
2. One blend ratio associated with each dart of the control mesh. The blend ratios
have a value in the range [0, 1]. The blend ratios describe a property analogous to
knot spacing. Smaller blend ratios give a sharper surface, and if all blend ratios
associated with a vertex is zero, we get the same effect as collapsing a knot
interval; a degree of smoothness is lost (and in this case we get interpolation of
that particular vertex).
3. One point in R3 associated with each vertex of the control mesh. The points
must satisfy the projective convexity property. The projective convexity property
states that for each face there exists a projection of the face such that the face is
the convex hull of it’s vertices. This assures that holes and bays of the mesh faces
are not covered by the surface. However, this means that some restrictions apply
to the coefficients. This makes it awkward to find a general basis. However, the
simplified surface spline scheme does not this restriction.
4.1.1 The scheme
We give a quick overview of the C 1 -surface spline scheme. For more elaborate explanations and formulas, see (Peters 1995a).
Three meshes are used in the construction (see figure 4.2). The first mesh is the
control mesh, the second is the quad mesh and the third is the inner mesh. The quad
mesh is the result of a polyhedral split of the control mesh, and the inner mesh is the
The scheme is composed of six steps: Mesh refinement, edge cutting, quadratic
Mesh refinement
The initial step of this scheme refines the control mesh. There are at least two reasons
for this step: to separate offending regions such that the planar cut polyhedron property
is achieved, and to convert the mesh to an all-quad mesh. This refinement procedure
is the polyhedral split, described in section 2.4.2. The quad mesh consists only of
quadrilaterals (see figure 4.3 and figure 4.2 left and middle).
48
CHAPTER 4. SURFACE SPLINE SCHEMES
Figure 4.3: Mesh refinement of the control mesh (left) resulting in the quad mesh
(right).
The refined mesh Mi+1 is called the quad mesh. Each face in the quad mesh is a
these edges are called quad edges. Two of the edges are the result from a split face, and
two of the edges are the result from a split edge. There is a one-to-one correspondence
between corners in Mi and faces in Mi+1 .
An irregular vertex is a vertex where more or less than four faces abut. Each quad
of the quad mesh abuts at most one irregular vertex. Each quad is defined from one
vertex which corresponds to a vertex in the control mesh, one vertex which corresponds
to a face midpoint, and two vertices corresponding to edge midpoints.
A vertex in the quad mesh is either a vertex from the control mesh, an edge midpoint
or a face midpoint. The edge midpoints always have four abutting quads (degree four).
If a face midpoint has a degree larger than four (the result of a face of more than four
vertices), then all of the surrounding vertices are either edge midpoints, or vertices
from the control mesh which has a degree of four or lower (due to the local degree
boundedness).
Edge cutting
This step performs the planar cut. Each quad has two exclusive blend ratios associated
with the two edges of the quad corresponding to edges in the control mesh (Remember
the one-to-one correspondence between corners of the control mesh and the quads).
The quad midpoint is a bilinear interpolation of the four vertices of the quad, where
the blend ratios are used as parameters. If the blend ratios are zero, the quad midpoint
ends up at the position of the one vertex of the quad which corresponds to a vertex of
the control mesh. If both blend ratios are one, then the quad midpoint ends up at the
face midpoint of the mesh.
Then, for each vertex in the quad mesh with a degree larger than four, all the quad
midpoints of the abutting quads are projected into a common plane. The projected quad
midpoints and quad midpoints not touched by a projection are kept, and the rest of the
From the quad midpoints we define the inner mesh (see figure 4.2 right). The inner
4.1. THE C 1 -SURFACE SPLINE
49
Figure 4.4: Edge cutting: The quad midpoints are bilinear interpolations of the vertices
mesh is the dual of the quad mesh, and we use the quad midpoints as geometry for this
inner mesh. The inner mesh is a planar cut polyhedra of the control mesh.
and quad edge vertices (see figure 4.5).
First we calculate the quad mesh vertices. The position of each vertex in the quad
mesh is set to the average of the midpoints of abutting quads.
Then we calculate the quad edge midpoint. A quad edge midpoint is the average
abutting a quad mesh vertex always lie in a common plane. If three quads abut, this is
obviously true. If four quads abut, this is also true (proof: pose it as a matrix problem,
and this matrix has rank 3). If more than four quads abut, the quad midpoints have been
projected into a common plane and thus averages are in the same plane.
Each edge in the quad mesh has three coefficients associated (two quad vertex
coefficients and one quad edge midpoint), and these coefficients can be interpreted as
the coefficients for a quadratic Bézier curve segment. The curve segments connecting
a quad mesh vertex have first order end derivatives in a common plane (the quad edge
midpoints of edges abutting a vertex resides in the same plane and the curves have the
same end point).
Each quad has nine coefficients associated: One quad midpoint, four quad edge midpoints, and four quad mesh vertices. One of these is from the split face, two is from
split edges, and one is from a control mesh vertex.
We can patch each of these quads with either one quadratic Bézier patch (using the
coefficients directly as Bézier coefficients) or alternatively as four triangular patches
where the coefficients are weighted averages of the nine coefficients. The result is a
surface that is C 1 in most areas.
50
CHAPTER 4. SURFACE SPLINE SCHEMES
Figure 4.5: Quadratic meshing. The vertices of the quad mesh are relocated to the
Degree raising
The term “irregular vertex” is dependent on context. In context of C 1 surface splines,
the following classifies whether a vertex is regular or irregular.
Definition 4.1.1. A regular vertex is a vertex where exactly four patches abut. A vertex
which is not regular is irregular.
The surface is not smooth in the neighbourhood of irregular vertices. The twist (mixed
derivatives) does not necessarily match. We can fix this, but we need some additional
degrees of freedom. To get these degrees of freedom we raise the degree of patches
abutting irregular vertices from quadratic to bicubic. This action does not change the
surface.
Then, we employ the new degrees of freedom to get the desired smoothness around
irregular vertices as well. We only tweak the coefficients residing in the interior of the
patch, and thus the boundary curve of each patch is kept. The procedure is dependent
on our choice of four sided or three sided patches in the step “quadratic patching”.
4.2 Two other surface spline schemes
We give an overview of two other surface spline schemes.
4.2.1 Surfaces . . . bicubics
The article (Peters 1994) is the predecessor of (Peters 1995a). To quote Peters: “For
general modelling I recommend C 1 -surface splines. Surfaces . . . bicubics is older, uses
more patches, uses more complicated formulas, and does not prove the convex hull
property. It is however the first of this type of construction.” The algorithm is rather
similar to that of C 1 surface splines, however a bit more laborious:
4.2. TWO OTHER SURFACE SPLINE SCHEMES
51
Figure 4.6: The scheme of surfaces . . . bicubics. Upper left: The initial control mesh.
Upper right: The refined mesh after one refinement. Lower left: The refined mesh
refined yet another time. Lower right: The patch structure. Each vertex of the refined
mesh is in the middle of one spline patch. The solid line renders the “hole” which is not
fitted with quadrilateral patches, the three patches inside this hole is of cubic degree.
52
CHAPTER 4. SURFACE SPLINE SCHEMES
1. The input mesh is refined twice with the Doo-Sabin refining and averaging procedure in order to separate irregular vertices (which become non-quadrilateral
mesh cells after the first refinement). This removes the need for the local degree
boundedness property. However, it results in more patches.
2. Every vertex in the double refined mesh where all neighbours are regular (and
thus is a 3 × 3 regular sub-mesh) is patched with a biquadratic B-spline.
The knots of the knot vectors are evenly spaced, and the knot vectors have no
multiple knots. As a result of this, the patch only covers the area around the
vertex and halfway to the neighbours (see figure 4.6). This knot vector makes it
possible for the B-spline to use the neighbouring vertices as coefficients, and the
patches will abut nicely and smooth.
Alternatively, the patches may be rewritten in Bézier form, but this increases the
number of coefficients.
3. An irregular vertex in the control mesh will represent a triangle after the DooSabin splits (see figure 4.6). The patches corresponding to the vertices of this
triangle is not yet patched, and thus around each irregular vertex in the control
mesh, we have a hole in the mesh in the shape of a convex n-gon built from n
rectangular patches, where n is the degree of the irregular node.
Due to two steps of the Doo-Sabin refinement, none of these holes touches each
other, and thus can be handled separately.
Each vertex corresponding to a patch in the hole (there is n of them) is patched
with a bicubic patch. The patch boundaries of these meshes that abuts the regular
patches are degree raised quadratics, and thus we get smooth transitions between
the regular and irregular patches.
The increased degree of these patches gives the required degrees of freedom
required to make a smooth transition between the irregular patches.
This scheme performs one more refinement step than C 1 surface splines. The gain
from this is that the input mesh can have a more arbitrary topology, but at the expense
of increased number of patches.
4.2.2 Localised-hierarchy surface splines
LeSS (Gonzalez & Peters 1999), an acronym for Localised-Hierarchy Surface Splines,
is an extension of (Peters 1995a). In essence, the surface is the same, but methods are
presented to remove a small part of the mesh and replace it with another mesh fragment.
The influence of this operation is local. An interesting effect of this scheme is that
we can not only make refinement, but actual topological changes to the mesh with
local influence. We can make bridges between two parts of the meshes, or introduce
topological holes.
However, it is important to note that this new surface is not exactly equal to the old
surface, only very similar. Thus, we cannot get nested spaces of surface splines with
this construction.
4.3. SUMMARY
53
4.3 Summary
This chapter begins with the motivation of surface spline scheme. The main idea is to
create a scheme behaving like the regular tensor product B-spline, but applicable for
arbitrary topologies. In addition, by defining the surface as a set of standard Bézier
patches, the scheme can easily be incorporated into existing geometry frameworks.
Three surface spline schemes are described, with extra emphasis on one particular
scheme, the C 1 -surface spline. This scheme is the “full” surface spline scheme of
which the simplified surface spline is derived from.
A regular vertex is a vertex surrounded by four faces. An irregular vertex is vertex
where this is not the case. The idea is to refine the control mesh to separate irregular
vertices. Then we associate a midpoint to each quad in the refined mesh. We can create
a dual of the refined mesh by using the quad midpoints as vertices. This dual is a planar
cut polyhedron after the local projection.
Around irregular vertices the following steps are necessary to produce a C 1 surface:
The local planar projection assures coplanarity of Bézier control points of relevant
coefficients from abutting patches. The raising of degree from quadratic to cubic on
the patches increases the number of degrees of freedom which are used by the twist
correction, which perturbs the inner control points to create matching cross-derivatives.
54
CHAPTER 4. SURFACE SPLINE SCHEMES
Chapter 5
The simplified surface spline
The simplified surface spline is a simplification of C 1 -surface splines (see section 4.1).
The construction is almost the same, however with the following differences:
• No planar projection is performed.
• Only rectangular biquadratic Bézier patches are used, thus, only one kind of
patches.
• No twist-correction is performed.
The simplified surface spline is not necessarily smooth in small neighbourhoods of
irregular vertices (see definition 4.1.1). We trade full C 1 -continuity with “almost C 1 continuity” for the benefit of a simpler scheme.
5.1 Overview
The scheme consists of five layers, and each layer has an array of points. A point in
one layer is a convex combination of points in the preceding layer. The layers are:
1. Layer p is filled by the control mesh vertices, which are indexed by vertices. The
polyhedral split combines the vertices of the control mesh to form the vertices
of the quad mesh (p → ps). The points in ps plays the same role as the refined
points of the surface spline curve (section 3.1.1).
2. Layer ps is filled by the vertices of the quad mesh, which are indexed by vertices,
edges and faces. The quad midpoint generation combines the vertices of the
equivalent with the sub-interval midpoints of the surface spline curve (section
3.1.2).
3. Layer qm is filled by the quad midpoints, which are indexed by corners. Bézier
coefficient calculation calculates the Bézier coefficients from the quad midpoints
(qm → bc). The Bézier coefficients corresponds to the Bézier coefficients of
the surface spline curve (section 3.1.3). This step is changes geometry such that
abutting patches often1 have coplanar control points at joints.
1 For
this always to happen, a planar projection must take place.
55
56
CHAPTER 5. THE SIMPLIFIED SURFACE SPLINE
Topological term
Original vertices
Edge midpoint
Face midpoint
Edge quartile
Inner edge quartile
Symbol
•
F
H
N
N
Indexing
(vi )
(ej )
(fk )
(vi , fk )
(vi , ej )
(vi , fk )
Used in layers
p,pv,bv
pe,be
pf ,bf
qm
beq
biq
Table 5.1: Terms in the construction with the symbols used in the figure. Note that the
edge quartiles and the inner edge quartiles have the same symbol. Indexing denotes
how the points are indexed.
4. Layer bc is filled by the Bézier coefficients. The indexing is too complex to be
stated in one sentence. The Bézier evaluation does a biquadratic Bézier evaluation of the relevant Bézier coefficients (bc → s). The biquadratic Bézier evaluation is equivalent to the quadratic Bézier segment evaluation of the surface spline
curve (section 3.1.4).
5. Layer s is the evaluated surface.
5.1.1 Topological terms of the simplified surface spline
All layers are indexed by topological terms, and these terms have symbols, used in the
subsequent figures, associated with them. As an example, the vertex term refers to a
coefficient attached to a vertex, and the inner-edge quartile term refers to a coefficient
attached to the midpoint of an edge in the quad mesh. Section 2.2.2 defines the different
positions in the mesh.
Since the control mesh can be more or less arbitrary, we cannot utilise regular sequential indices, which are used in the construction of B-splines. This can be confusing
(if conventional splines over regular topologies have a mess of indices, surface splines
over arbitrary topologies are even worse). For simplified surface splines, we use six
terms to define the coefficients:
1. Vertices are where the vertices of the control mesh are. These are indexed by the
vertices.
2. Edge midpoints are the middle of each edge of the control mesh. These are
indexed by the edges.
3. Face midpoints are the middle of each face of the control mesh. These are
indexed by the faces.
4. Quad midpoints are located in the middle of each quad of the refined control
mesh (the quad mesh). There is a one-to-one correspondence between each quad
in the quad mesh and each corner in the control mesh, and these are indexed by
the corners of the control mesh.
5. Edge quartiles are in the middle of the edges in the quad mesh which is the
result of a split edge. Every edge in the control mesh has two edge quartiles,
between the edge midpoint and their respective end vertex of the edge (thus their
5.1. OVERVIEW
57
Figure 5.1: Surface spline concepts. The individual patches are coloured in different
shades of grey. Upper left, the control mesh. Upper right, the quad mesh. Lower left,
the inner mesh, and lower right, the Bézier coefficients.
name quartiles). The edge quartiles are indexed by an edge and a vertex (one of
the two defining the edge).
6. Inner edge quartiles are in the middle of the edges in the quad mesh which is
a result of face splitting (Each edge in the quad mesh are either the result of a
split edge or a split face). Each edge in the control mesh about a face in the same
mesh gives raise to an edge in the quad mesh which is the result of a face split.
Thus, these are indexed by an edge and a face from the control mesh.
The reader is recommended to look at table 5.1 which is the key between terms and the
symbols used in figures, as well as the figures 5.2, 5.3, 5.4 and 5.5 where these symbols
are used.
5.1.2 The three meshes
There exist three meshes in the scheme (refer to figure 5.1. The first mesh is the control
mesh, which is the input to the scheme. The second mesh is a refinement of the control
mesh, which is known as the quad mesh. The vertices of the control mesh also exist
in the refined mesh. Hence, the set of vertices in the control mesh is a subset of the
58
CHAPTER 5. THE SIMPLIFIED SURFACE SPLINE
vertices in the refined mesh. The third mesh, not used explicitly, is the inner mesh. It
is the aim of the C 1 -surface spline scheme to make this inner mesh into a planar cut
polyhedron.
5.1.3 The intrinsic parameterisation
There is a one-to-one mapping between corners of the control mesh and the Bézier
patches of the surface. Thus, the corners are useful for indexing the patches. Each
patch has the unit square as its local domain.
Definition 5.1.1. A parameter point p on the simplified surface spline surface is a triple
p = (c, s, t),
(5.1)
where (s, t) ∈ [0, 1] × [0, 1], and c refers to a corner in the control mesh.
The patches do overlap at the patch boundaries. However, the abutting patches evaluate
to the same value, and thus this is just a matter of choosing which patch to evaluate.
The parametrisation is much influenced by the blend ratios. Blend ratios are in the
interval [0, 1]. If the blend ratios are outside this interval, the construction is no longer
entirely constructed from convex combinations. The blend ratios play the same role
as the knot spacing does for B-splines. From figure 4.1 we see that small blend ratios
produce the same effect as small knot spacing, sharper geometry, and large blend ratios
give smoother geometry. When blend ratios are reduced to zero, we loose a degree of
continuity.
5.2 Layer I: The control mesh vertices
The input to the scheme is a control mesh with a blend ratio associated with each dart
of this mesh and a control point associated with each vertex.
The initial layer is composed of the control points. Each vertex has a control point
associated, and thus, we index the control points with the vertices. The control points
must reside in a linear place such that convex combinations can take place. Otherwise,
there are no restrictions, and thus, the points can be of arbitrary dimension. However,
points from R3 are probably the most useful choice.
5.3 Layer II: The quad mesh points
The control mesh (figure 5.2) is split with the polyhedral split scheme (see section
2.4.2) to make the quad mesh (figure 5.3). The original vertices are kept, the edge
midpoint is the average of the two vertices the edge connects, and the face midpoint is
the average of the vertices of the face. We call the vertices obtained from this step the
ps-vertices, referring to the splitting scheme, the polyhedral split.
We use the notation |restriction to restrict the expression to a specific case. Thus, ps|V
is the polyhedral split in the case of a vertex.
Definition 5.3.1. We define the vertices of the layer ps as a combination of vertices in
the control mesh layer p. There are three flavours of these; vertices, edge midpoints
5.3. LAYER II: THE QUAD MESH POINTS
59
Figure 5.2: The initial control mesh. In this case: Three quadrilaterals and three triangles, fifteen edges and ten vertices.
Figure 5.3: The quad mesh. This mesh defines the patch structure; each quadrilateral
corresponds to one Bézier patch.
CHAPTER 5. THE SIMPLIFIED SURFACE SPLINE
60
and face midpoints:
ps|V (vi ) = p(vi )
1 X
ps|EM (ej ) =
p(vl )
2 v ∈e
i
l
1 X
p(vl ).
ps|F M (fk ) =
#fˆk
ˆ
(5.2)
v l ∈f k
The three functions give the position of the vertex corresponding to the old vertex,
position of the vertex corresponding to the edge midpoint and the position of the vertex
corresponding to the face midpoint respectively.
The notation is defined in section 2.2. Quad meshes (the result of a polyhedral split)
is treated in section 2.4. This operation is the equivalent of the refinement-step of the
surface spline curve construction (section 3.1.1).
One important observation can be made; all of the new vertices are convex combinations of vertices in the control mesh.
Lemma 5.3.2. All vertices of the quad mesh (layer II) are convex combinations of the
vertices of the control mesh (layer I).
Proof. The averages in definition 5.3.1 are obviously convex combinations.
Corollary 5.3.3. There are #V + #E + #F points in the quad mesh layer.
Proof. The points of the vertices are kept, and each edge get a midpoint as well as each
face.
5.4 Layer III: Quad midpoints (inner mesh vertices)
define the inner mesh vertices. The inner mesh is not explicitly used, but the inner
mesh is the planar cut polyhedron2 from which we define the faces.
A quad midpoint is a bilinear interpolation of the four vertices of the quad. The
parameters of this interpolation are the blend ratios, and the coefficients are the four
vertices of the quad. A overview of the bilinear interpolation is given in section 1.4.2.
Each quad corresponds to a corner in the control mesh. Each corner in the control
mesh “has” two non-exclusive edges associated, the edge in and the edge out. The combination of face and vertex (from the corner) in addition to an edge gives an exclusive
key (a dart in this case), and for each such key we have a blend ratio associated. Thus,
each corner has two exclusive blend ratios associated. This calculation is visualised in
figures 5.3 and 1.1.
At the boundary, the edge quartiles of the curve play an equal role as the quad
midpoints of the surface. To achieve this, we assign a value to the edge quartiles of
every boundary edge in this step. This is to achieve a border in the form of a surface
spline curve. No edge quartiles are defined for interior edges. Definition 5.4.1 uses
mesh queries, defined in section 2.3.3.
2 It
it not necessarily a planar cut polyhedron since the planar projection is skipped
5.4. LAYER III: QUAD MIDPOINTS (INNER MESH VERTICES)
61
Figure 5.4: The quad midpoints. Connecting these points together forms the inner
mesh.
Definition 5.4.1. Let the set BE be the set of edges in the mesh M which are boundary
edges,
BE = ej ∈ M[e; ∗, ∗, ∗] |#M[f ;∗,ej ,∗]=1
(5.3)
which is the set of edges where the number of abutting faces are one. Similarly, we may
define the set of boundary vertices
BV = vi ∈ M[v; ∗, ∗, ∗] |#M[f ;vi ,∗,∗]6=#M[e;vi ,∗,∗] ,
(5.4)
which is the set of vertices where the number of connecting edges are not the same as
the number of abutting faces.
One may read “if vi is a boundary edge” as “if vi ∈ BV” to make it more formal,
and less readable. The same goes for edges. The quad midpoints corresponds to the
sub-interval midpoints of the surface spline curve (section 3.1.1).
Definition 5.4.2. We define the vertices of the layer qm, the quad midpoints, as combinations of the vertices of layer ps. The vertices are indexed by the corners of the
control mesh. Given a corner (vi , fk ), we have two edges associated with it, ein and
CHAPTER 5. THE SIMPLIFIED SURFACE SPLINE
62
eout as well as two blend ratios, bin and bout , and qm is defined
qm|QM (vi , fk ) = (1 − bin )(1 − bout )ps|V (vi ) +
bin (1 − bout )ps|EM (ein ) +
(1 − bin )bout ps|EM (eout ) +
bin bout ps|F M (fk ),
qm|EQ (vi , ej ) =
(
(1 − b)ps|V (vi ) + bpsEM (ej )
undefined
if ej is a boundary edge
otherwise,
(5.5)
where b is a blend ratio. The query d = M[v, e, f ; vi , ej , ∗] will give one dart, d, if ej
is a boundary edge, and b is the blend ratio associated with this dart.
The expressions in equation 5.5 are a bilinear interpolation and a linear interpolation
respectively. If both blend-ratios are zero, the quad midpoint gets the position of the
original vertex of the patch. If both blend-ratios are one, the patch midpoint gets the
position of the face midpoint. Since all the Bézier coefficients depend on these patchmidpoints, we may utilise the blend-ratios to control the curvature of the resulting
surface. While we’re at it, patch midpoints are also convex combinations.
Lemma 5.4.3. All vertices in qm(. . .) are convex combinations of the vertices of the
control mesh.
Proof. The bilinear interpolation is a convex function. See section 1.4.2 for details.
Thus, in the quad midpoint case, the vertices are convex combinations of ps(. . .),
which in turn is a convex combination of the vertices of the control mesh. From property 1.4.5 we know that this results in a convex combination. In the edge quartile case,
the vertex is a linear interpolation of two points in ps(. . .) and the same argument
applies here as well.
Corollary 5.4.4. There are #C + 2#BE points in the quad midpoint layer.
Proof. There is a one-to-one correspondence between the quads in the quad mesh and
the corners of the control mesh. Each quad has one quad midpoint, and there are #C
quads. In addition we need the edge quartiles at the boundary. Each edge has two edge
quartiles, and thus this number is 2#BE.
5.5 Layer IV: Bézier coefficients
The next layer is the layer of Bézier coefficients, which are combinations of the quad
midpoints, and the Bézier coefficients are used by the biquadratic Bézier evaluation.
It is in this step the smoothing takes place. All Bézier coefficients are averages
of surrounding quad midpoints (which reside in the interior of patches). In this way,
the control points are coplanar for abutting patches along boundaries. The simplified
surface spline construction is equal to the full C 1 surface spline scheme except at the
neighbourhood of irregular vertices, and has thus the same smoothness properties as
the full scheme at non-irregular neighbourhoods.
The Bézier coefficients of the surface scheme are analogous to the Bézier coefficient step of the surface spline curve construction. Bézier coefficients come in six
flavours:
5.5. LAYER IV: BÉZIER COEFFICIENTS
63
Figure 5.5: The Bézier coefficients.
1. Vertex coefficients are the average of the quad midpoints abutting that particular
vertex in the control mesh. The number of abutting patches equals the number
of faces abutting the vertex.
2. Edge midpoint coefficients are the average of the four abutting quad midpoints.
Edge midpoint can also be the average of the two edge quartile coefficients of
that edge, which gives the exact same result.
3. Face midpoint coefficients are the average of the quad midpoints of that particular face. If the face is an n-gon, the n quad midpoints of this face are used.
4. Quad midpoints coefficients are just a copy of the quad midpoint of the layer
below.
5. Edge quartile coefficients are the average of the two quad midpoints abutting at
that end of the edge.
6. Inner edge quartile coefficients are the average of two succeeding patch midpoints in a face in between the inner edge lies.
Figure 5.5 shows the layout of these coefficients and definition 5.5.1 defines the formulas.
The coefficients have two major cases at this step: boundary and non-boundary.
In the case of a non-boundary coefficient, the coefficient is the average of the quad
midpoints. However, in the boundary case, the coefficient is the average of some edge
quartiles from the same layer. Consult figure 5.6 to get an idea of which coefficients
are boundary and which are not.
CHAPTER 5. THE SIMPLIFIED SURFACE SPLINE
64
Figure 5.6: The boundary coefficients are in black while the non-boundary coefficients
are in grey. Notice that no inner edge quartiles or face midpoints lie on the boundary.
Definition 5.5.1. All the Bézier coefficients are averages of a subset of the quad midpoints of the layer below. There are five types of Bézier coefficients: vertex (V ), edge
midpoint (EM ), face midpoint (F M ), edge quartile (EQ) and inner edge quartile
(IQ). The general formula is
bc(. . .) =
(
1 P
(v,f )∈C qm|QM (v, f )
#C
P
1
(v,e)∈EQ qm|EQ (v, e)
#EQ
if . . . is non-boundary
if . . . is boundary,
(5.6)
where . . . is the appropriate parameters. Obviously, only vertices, edge midpoints and
edge quartiles can be boundary. In the non-boundary case, the Bézier coefficients are
blends of the quad midpoints of the layer below. In the boundary case, the coefficients
are blends of the edge quartiles defined in the layer below. In the non-boundary case,
the set C is a subset of the quad midpoints specific to the case and parameters,
C|V (vi ) = M[v, f ; vi , ∗, ∗],
C|EM (ej ) = M[v, f ; ∗, ej , ∗],
C|F M (fk ) = M[v, f ; ∗, ∗, fk ],
C|QM (vi , fk ) = {(vi , fk )},
(5.7)
C|EQ (vi , ej ) = M[v, f ; vi , ∗, ∗] and
C|IQ (ej , fk ) = M[v, f ; ∗, ej , fk ],
In the boundary case, the set EQ is a subset of the edge quartiles defined in the layer
5.6. LAYER V: BÉZIER EVALUATION
65
t direction −→
s=0
t=1
s=0
t=0
s=1
t=1
bc|EM (ein )
bc|IQ (ein , fk )
bc|F M (fk )
bc|EQ (vi , ein )
bc|QM (vi , fk )
bc|IQ (eout , fk )
bc|V (vi )
bc|EQ (vi , eout )
bc|EM (ej , eout )
s=1
t=0
s direction −→
Figure 5.7: The layout of the coefficients used in the Bézier evaluation. From a corner
(vi , fk ), we find the two edges, the incoming and outgoing. From this we know which
vertices from the bc layer to use.
below,
EQ|V (vi ) =
[
∀e∈BE:vi ∈e
(vi , e),
EQ|EM (ej ) = M[v, e; ∗, ej , ∗]
EQ|EQ (vi , ej ) = {(vi , ej )}
(5.8)
Lemma 5.5.2. The Bézier coefficients are convex combination of the control mesh
vertices.
Proof. All the functions are simple averages, which are convex combinations, of either
(a) the quad midpoints of the layer below, or (b) the edge quartiles of the layer below.
We utilise lemmas 5.3.2, 1.4.5, and 5.5.2, and this concludes the proof.
Corollary 5.5.3. There are #V + 3#E + #F + 2#C points in the Bézier coefficient
layer.
Proof. Each vertex has a point, each edge has three points (one midpoint and two
quartiles), each face has the same number of inner-edge quartiles as corners and one
midpoint. In addition has each patch a patch midpoint.
Corollary 5.5.4. If the control mesh is closed, all Bézier coefficients can be calculated
Proof. If the control mesh is closed, it has no boundary edges, and edge quartiles are
not used in this step.
5.6 Layer V: Bézier evaluation
Each patch is defined from nine coefficients. All of these coefficients, except the quad
midpoints, are shared between neighbouring patches, which assures C 0 continuity. To
get a consistent orientation of the surface, we let s and t run along the outgoing and
the incoming edge respectively. If the control mesh is consistently oriented, the surface
will be consistently oriented as well.
66
CHAPTER 5. THE SIMPLIFIED SURFACE SPLINE
Definition 5.6.1. A patch is associated with each corner of the control mesh. The
corner c is defined by a vertex-face pair. In addition, a corner has two edges associated
with it, the incoming edge ein and eout .
s(c, s, t) = (1 − s)2 (1 − t)2 bc|V (v) +
2(1 − s)2 (1 − t)t bc|EQ (v, eout ) +
(1 − s)2 t2 bc|EM (eout ) +
2(1 − s)s(1 − t)2 bc|EQ (v, ein ) +
4(1 − s)s(1 − t)t bc|QM (v, f ) +
(5.9)
2
2(1 − s)st bc|IQ (eout , f ) +
s2 (1 − t)2 bc|EM (ein ) +
2s2 (1 − t)t bc|IQ (ein , f ) +
s2 t2 bc|F M (f ),
where (c,s,t) is a parameter point (definition 5.1.3), v,f ,e in and eout are the vertex,
face, incoming edge and outgoing edge of corner c = (v i , fk ) respectively.
The nine coefficients are blended with a biquadratic blend (see example 1.4.2). The
biquadratic blend blends coefficients laid out in a three-by-three grid (see figure 5.7).
Property 5.6.2. The expressions given in definition 5.6.1 are convex combinations of
the vertices of the control mesh.
Proof. The biquadratic Bézier blend is a convex combination of its coefficients (see
section 1.4.2). In the usual spirit, we utilise properties 5.3.2, 1.4.5, and 5.5.2, and this
concludes the proof.
With this, we have proved that any point on the simplified surface spline surface
is computed as convex combinations of the control mesh vertices. This gives a useful
property, the convex hull property.
Property 5.6.3. The complete simplified surface spline surface resides inside the convex hull of the control mesh vertices.
Proof. We know from property 5.6.2 that any point on the surface can be expressed as
a convex combination of the control mesh vertices, and thus by property 1.4.4 we get
property 5.6.3.
Chapter 6
The simplified surface spline
basis
The steps of chapter 5 are intuitive and rather straight-forward. However, this way
of representing the surface is not the best when dealing with qualitative aspects of
approximation. We are interested in how much a specific control vertex is contributing
to the surface at a given parameter point. Since all operations are convex combinations
of the layer below we can view the scheme as a series of matrices applied to the control
points. Each step defines a matrix. In addition, we introduce an additional matrix, the
select matrix. The select matrix picks out the correct Bézier coefficients to be used in
the Bézier evaluation. We define the following matrices:
1. The matrix S ps performs the polyhedral split. This matrix has #V columns, and
#V + #E + #F rows.
2. The matrix S qm calculates the quad midpoint matrices and copies the boundary
edge quartiles. This matrix has #V+#E+#F columns and #C+2#BE rows.
3. The matrix S bc calculates the Bézier coefficients. This matrix has #C + 2#BE
columns and #V + 3#E + #F + 2#C rows.
4. The matrix S sel (c) is dependent on which patch to be evaluated. It extracts the
correct set of Bézier coefficients to be used in the Bézier evaluation. This matrix
has #V + 3#E + #F + 2#C columns and 9 rows. This matrix contains nine
1’s and the rest of the entries are zero. Which entries are non-zero is dependent
on c.
5. The matrix S bz (s, t) performs the Bézier evaluation. The entries in this matrix
is the nine basis functions for the biquadratic Bézier patch. The matrix has 9
columns and 1 row.
Combining this, we get the matrix product
s(c, s, t) = S bz (s, t) · S sel (c) · S bc · S qm · S ps · p
|
{z
}
basis
=
Nv
X
B(vb ; c, s, t)p(vb ),
b=1
67
(6.1)
68
CHAPTER 6. THE SIMPLIFIED SURFACE SPLINE BASIS
and the matrices combined give the basis functions. A basis function is a function that
gives the amount of contribution a vertex has at a particular parameter value. It is in fact
rather easy to deduce the entries for these matrices. However, these matrices become
rather large, even for small models.
We will not use this matrix formulation directly when deducing the basis functions. Instead, we will write the matrix products explicitly as sums, and formulate the
following hypothesis:
Hypothesis 6.0.1. The simplified surface spline is a linear scheme (all steps are linear
combinations of values of the previous step), thus we can write the surface as a sum,
s(c, s, t) =
Nv
X
B(vb ; c, s, t)p(vb ),
(6.2)
b=1
where B(vb ; c, s, t) is the basis function associated with vi evaluated at the parameter
point (c, u, v).
Each vertex of the control mesh has its own basis function. Instead of writing “the
vertex of which the basis function corresponds to”, we will consistently use v b as this
vertex. When we have several basis vertices (as in the sum of hypothesis 6.0.1), we use
b as the summation variable.
6.1 Overview
Conceptually, we are sampling the scheme. A basis function is the response given by
the scheme when emitting a unit pulse at vb . The unit pulse is defined as
(
1 i=b
δib =
.
0 i=
6 b
(6.3)
We use the unit pulse to rewrite the control mesh vertices as a sampling of the control
mesh vertices,
p(vi ) =
Nv
X
δib p(vb ),
(6.4)
b=1
This cumbersome identity will be useful. We first unwrap the scheme of section 5, then
introduce our identity given in equation 6.4, and finally re-wrap the scheme with some
algebraic juggling. The result is an expression on the form given in hypothesis 6.0.1.
From definitions 5.6.1, 5.5.1, 5.4.2 and 5.3.1 we see that the scheme can be composed of convex combinations of a “layer below”. The scheme is composed of layers
of computation; first, the polyhedral split is performed, and is the layer above the control mesh vertices. Then the quad midpoints are calculated from the polyhedral split
vertices, and form the layer above this again, and so on (we skip most of the parameters
6.1. OVERVIEW
69
for simplicity here).
s(c, s, t) =
X
wbz bc(. . .)
X
wbc qm(. . .)
s(c, s, t) =
wbz
X
X
X
s(c, s, t) =
wbz
wbc
wqm ps(. . .)
X
X
X
X
s(c, s, t) =
wbz
wbc
wqm
wps p(vi )
X
(6.5)
i
s(c, s, t) =
X
wbz
X
wbc
X
wqm
X
wps
i
Nv
X
b=1
|
δij p(vb )
{z
from eq. 6.4
}
We will prove that we may move the sum over b outside the nesting of the computation
step by step,
ps(...)
s(c, s, t) =
X
wbz (u, v)
X
wbc (c)
X
zX X }|
{
wps δib p(vb )
wqm
b
|
i
{z
}
Bps (vb ;...)
s(c, s, t) =
s(c, s, t) =
s(c, s, t) =
X
X
wbz (u, v)
X
|
wbc (c)
zX X
b
zX X
b
XX
b
s(c, s, t) =
wbz (u, v)
X
|
|
qm(...)
}|
{
wqm Bps (vb ; . . .) p(vb )
{z
}
bc(...)
Bqm (vb ;...)
(6.6)
}|
{
wbc (c)Bqm (vb ; . . .) p(vi )
{z
}
Bbc (vb ;...)
wbz Bbc (vb ; . . .) p(vb )
{z
}
B(vb ;...)
B(vb ; . . .)p(vi ),
b
and we end up proving hypothesis 6.0.1. The un-nesting done in equation 6.5 are
proved in the opposite direction in the definitions of the previous section. To overcome
extremely tedious bookkeeping, we will perform the change of summation and renesting of equation 6.6 step by step, and not all steps at once.
We will also see that the simplified surface spline has good local support properties.
We will prove a corollary for each step determining the size of the support of the
expressions of the step, and combine this to find the support of the scheme in total.
70
CHAPTER 6. THE SIMPLIFIED SURFACE SPLINE BASIS
6.2 Layer II: Quad mesh points
Lemma 6.2.1. The vertices of the polyhedral split layer can be written as
ps(. . .) =
Nv
X
Bps (vb ; . . .)p(vb ),
(6.7)
b=1
where . . . are either a vertex, an edge or a face. There are three cases in this layer: the
vertex (V ), the edge midpoint (EM ) and the face midpoint (F M ),
Bps |V (vb ; vi ) = δib
1 X
Bps |EM (vb ; ej ) =
δlb
2 v ∈e
j
l
(6.8)
1 X
δlb ,
Bps |F M (vb ; fk ) =
#fˆk
vl ∈ej
where fˆk is the vertices of face fk as defined in definition 5.3.1.
Proof. From definition 5.3.1 we have
ps|V (vi ) = p(vi ).
(6.9)
We introduce an extra summation,
ps|V (vi ) =
Nv
X
δib p(vb ),
(6.10)
b=1
The equations 6.9 and 6.10 are equivalent. The expression for vertices in the layer of
the polyhedral split, in the case of a vertex, are then
Bps |V (vb ; vi ) = δib .
(6.11)
Now, let’s look at the edge midpoints of this layer, from definition 5.3.1
ps|EM (ej ) =
1 X
p(vl )
2 v ∈e
l
i
Nv
1 X X
δlb p(vl )
=
2 v ∈e
b=1
i
l
!
Nv
X
1 X
=
δlb p(vl )
2 v ∈e
b=1
=
Nv
X
b=1
l
(6.12)
i
Bps |EM (vb ; ej )p(vl ).
The change of summation in the two middle rows of equation 6.12 is appropriate, and
71
this enables us to find the expressions. Finally, the case of the face midpoints,
ps|F M (fk ) =
1 X
p(vl )
#fˆk
vk ∈fˆk
=
Nv
X
1 X
δlb p(vb )
#fˆk
=
Nv
X
b=1
(6.13)
vk ∈fˆk
b=1
Bps |F M (vb ; fk )p(vl ),
and this proves lemma 6.2.1.
Corollary 6.2.2. The functions Bps (vb ; . . . ) are
Bps |V (vb ; vi ) = 0 if vb 6= vi ,
Bps |EM (vb ; ej ) = 0 if vb ∈
/ ej and
Bps |F M (fk ) = 0 if vb ∈
/ fˆk .
(6.14)
Proof. This follows easily from the definitions of Bps given in lemma 6.2.1.
The expressions of the polyhedral split layer are non-zero at the vertex v b , the midpoint
of edges connected to vb , and at the face midpoints of faces abutting vb . Otherwise, the
functions are zero.
The vertices in the quad midpoint layer are convex combinations of the vertices in the
polyhedral split layer. There are two cases of a quad midpoint:
1. it is in face the midpoint of a quad, or
2. it is the midpoint of an edge quartile.
Item (2) can be a bit confusing. The quad midpoint layer consists of a midpoint of
each quad in the mesh as well as the midpoint of each boundary quartile. Remember
from chapter 3 that the sub-interval midpoints played an equivalent role as the quad
midpoints. The boundary of a simplified surface spline surface is defined by a surface
spline curve, and the edge quartiles on the boundary are in fact the sub-edge midpoints
of this boundary curve.
Lemma 6.3.1. The points of the quad midpoint layer can be written as
qm(. . .) =
Nv
X
Bqm (vb ; . . .)p(vb ),
(6.15)
b=1
where . . . are either a corner or an edge quartile. There are two cases in this layer,
the regular quad midpoint, and the edge quartiles for use in boundary-cases. The
CHAPTER 6. THE SIMPLIFIED SURFACE SPLINE BASIS
72
expressions are
Bqm |QM (vb ; vi , fk ) = (1 − bin )(1 − bout )
+bin (1 − bout )
(1 − bin )bout
bin bout
(
(
(
(
1 if vb = vi
0 otherwise
1/2 if vb ∈ ein
0
otherwise
(6.16)
1/2 if vb ∈ eout
0
otherwise
1/#f
0
if vb ∈ fˆk
otherwise,
and
Bqm |EQ (vb ; vi , ej ) = (1 − b)
(
(
1 if vb = vi
0 otherwise
(6.17)
1/2 if vb ∈ ej
+b
0
otherwise.
The edges ein and eout are the incoming and outgoing edge of c = (vi , fk ), and
bin is the blend ratio associated with M[v, e, f ; vi , ein , fk ] and similarly for bout .
We assume that Bqm |EQ are only evaluated when ej is a boundary edge, and thus,
M[v, e, f ; vi , ej , ∗] returns a single dart, and b is the blend ratio associated with this
dart.
6.2.1,
qm|QM (vi , fk ) = (1 − bin )(1 − bout )
bin (1 − bout )
(1 − bin )bout
bin bout
Nv
X
vb =1
Nv
X
vb =1
Nv
X
vb =1
Nv
X
vb =1
Bps |V (vb ; vi )p(vb ) +
Bps |EM (vb ; ein )p(vb ) +
(6.18)
Bps |EM (vb ; eout )p(vb ) +
Bps |F M (vb ; fk )p(vb ).
We may move the summation of vb outside the expression, and insert the expressions
73
for Bps ,
qm|QM (vi , fk ) =
Nv
X
vb =1
(1 − bin )(1 − bout )δbi +
X δbk
+
2
vk ∈ein
X δbk
(1 − bin )bout
+
2
vk ∈eout
X δbk
bin bout
#fk
p(vb ).
1
bin (1 − bout )
(6.19)
vk ∈fˆk
Only one term of the sums may be nonzero for each vb , and thus, we get equation 6.16.
We do the same for the edge quartiles. We assume ej to be a boundary edge, and from
the same definition and lemma we get
qm|EQ (vi , ej ) = (1 − b)
=
vb =1
Bps |V (vb ; vi )p(vb ) + b
Nv
X
vb =1

X δik
 p(vb ).
(1 − b)δib + b
2
v ∈e
=1
Nv
X
vb
Nv
X

k
Bps |EM (vb ; ej )p(vb )
j
(6.20)
Here as well, only one term of the sum is nonzero, and thus we get equation 6.17 of
lemma 6.3.1.
From the nature of the control mesh, we see that the quad midpoints of this layer can
be in four different situations in respect of vb . Either
1. The vertex vb = vi , and thus vb ∈ ein , vb ∈ eout and vb ∈ fˆk . Thus all four of
the terms in the bilinear blend are nonzero.
2. The vertex vi 6= vb but either vb ∈ ein or vb ∈ eout . Then vb ∈ fˆk and two of
the terms are nonzero.
3. The vertex vb ∈
/ ein or vb ∈
/ eout , and thus vb 6= vi but vb ∈ fˆk . Then only the
term of the face midpoint is nonzero.
4. If vb ∈
/ fˆk , all of the terms are zero.
Situation (4) is especially interesting, and we will record this in a corollary. We call the
set of nonzero patch midpoints for a particular vertex vb the quad midpoint influence
of vb (see figure 6.1). This will be used when deducing the support of the expressions
of other kinds of coefficients.
Corollary 6.3.2. The functions Bqm (vb ; . . .) are zero when
Bqm |QM (vb ; vi , fk ) = 0 if vb ∈
/ fˆk
Bqm |EQ (vb ; vi , ej ) = 0 if vb ∈
/ ej .
(6.21)
74
CHAPTER 6. THE SIMPLIFIED SURFACE SPLINE BASIS
Figure 6.1: Deduction of support from the influence of vb and sources of computation
for the various coefficients. Upper left: The influence of vb on the quad midpoints H.
The gray quad midpoints are zero and the black quad midpoints are not necessarily
nonzero. Upper right: The source of computation for original vertices (white square).
The black quad midpoints are used in the calculation of this coefficient, while the
gray ones are not used. Middle left: The source of computation for edge midpoints.
Middle right: The source of computation for face midpoints. Lower left: The source of
computation for edge quartiles. Lower right: The source of computation for inner edge
quartiles.
6.4. LAYER IV: BÉZIER COEFFICIENTS
75
Proof. From the fundamental concept of a mesh, the incoming and outgoing edge must
be a part of the face the corner belongs to. Also, both of these two edges must connect
to the vertex of the corner. Thus, if a vertex is not part of the face of the corner, the
vertex cannot be the vertex of the corner or one of the end vertices of the incoming
and outgoing edge. This if vb ∈
/ fˆk , then all four terms are zero, and therefore the
result is zero. Similarly, if a vertex is not one of the two end vertices of an edge, then
it obviously cannot be one of the two end vertices of the same edge. Thus, if v b ∈
/ ej ,
then both terms are zero in the case of edge quartiles.
6.4 Layer IV: Bézier coefficients
The Bézier coefficients are convex combinations of either a set of quad midpoints of
the layer below, or a set of edge quartiles.
Lemma 6.4.1. The points of the Bézier coefficients layer can be written as
bc(. . .) =
Nv
X
Bbc (vb ; . . .)p(vb ),
(6.22)
b=1
where
Bbc (vb ; . . .) =
(
P
if . . . is non-boundary
(v,f )∈C Bqm |QM (vb ; v, f )
(6.23)
P
(v,e)∈EQ Bqm |EQ (vb ; v, e) if . . . is boundary,
1
#C
1
#EQ
that is, the expressions are averages of quad midpoints in the non-boundary case, and
averages of the edge quartiles in the boundary case. The sets of which the averages
are taken are defined in definition 5.5.1.
Proof. From definition 5.5.1 and lemma 6.3.1 we know that
(
P Nv
1 P
if . . . is non-boundary
(v,f )∈C
vb =1 Bqm |QM (vb ; v, f )p(vb )
#C
bc(. . .) =
P
P
Nv
1
(v,e)∈EQ
vb =1 Bqm |EQ (vb ; v, e)p(vb ) if . . . is boundary
#EQ
(6.24)
In both cases we can swap the two summations and,
P
P
1
 Nv
non-boundary
(v,f )∈C Bqm |QM (vb ; v, f ) p(vb )
vb =1 #C
bc(. . .) = PNv P
1

vb =1 #EQ
(v,e)∈EQ Bqm |EQ (vb ; v, e) p(vb ) boundary
(6.25)
and by defining the expressions inside the big parentheses as the expressions, we get
lemma 6.4.1.
We let Q and EQ be the sets of quad midpoints or edge quartiles used in the computation of a point, and let Iqm and Ieq be the influenced quad midpoints and edge quartiles
respectively for a given vertex. In general, we have
(
Q ∩ Iqm (vb ) = ∅ if non-boundary
Bbc (vb ; . . .) = 0 if
(6.26)
EQ ∩ Ieq (vb ) = ∅ if boundary.
The specific cases are given in the following corollary.
CHAPTER 6. THE SIMPLIFIED SURFACE SPLINE BASIS
76
Corollary 6.4.2. The expressions used in the computation of a point in the Bézier
coefficient layer are zero for a basis vertex vb when
1.
Bbc |V (vb ; vi ) ≡ 0 if

/ M[v, ∗, ∗, fk ] ∀fk ∈ M[f ; vb , ∗, ∗]

 vi ∈
m
(6.27)


vb ∈
/ M[v, ∗, ∗, fk ] ∀fk ∈ M[f ; vi , ∗, ∗].
These two tests are equivalent.
2. Both
)
Bbc |EM (vb ; ej ) ≡ 0
Bbc |EQ (vb ; vi , ej ) ≡ 0
3. Both
if ej ∈
/ M[e, ∗, ∗, fk ] ∀fk ∈ M[f ; vb , ∗, ∗].
)
Bbc |F M (vb ; fk ) ≡ 0
Bbc |IQ (vb ; ej , fk ) ≡ 0
These two tests are equivalent.
(6.28)
if

/ M[f ; vb , ∗, ∗]

 fk ∈
m


vb ∈
/ fˆk .
(6.29)
The proof is rather argumentative, and figure 6.1 can be of good help when examining
this proof.
Proof. We will begin with the non-boundary case. We make some observations of
Iqm (vb ): The corners in this set form full faces, that is, either all or none of the corners
of a face is in the set.
This implies that if a face is “in” the influence region, all of the inner edge quartiles and the face midpoint as well are possible nonzero. However, if the face is not
inside, which means that none of the corners use this face, all of the before-mentioned
coefficients are exactly equal to zero. Thus, the test for face midpoints and inner edge
quartiles are the same.
This also implies that an edge can either (i) be inside the region, that is, both abutting faces are in the region, or (ii) the edge is on the border of the region, that is one
of the abutting faces are in the region while the other is not or (iii) the edge is not in
the region at all. Thus, if the edge is either (i) or (ii), the edge midpoint is influenced
by quad midpoints in the influence region. If the edge is (iii), the edge cannot be influenced since both of the two faces used in all corners around the edge are not present
in the influence region. Further, since both of the corners on one side of the edge are
either both in or both out of the influence region, the test for edge quartiles are the same
as for edge midpoints.
Test (1a) tests whether vi is one of the vertices of the faces abutting vb . If this is
not the case, then obviously none of the surrounding corners of v i is in the influence
region.
To verify that (1a) and (1b) are equivalent, we remark that if vi is one of the vertices
of faces abutting vb , then vb must also be a vertex of this same face abutting vi . Thus,
this test is symmetric.
Test (2) tests whether ej is one of the edges bounding the faces abutting vb , thus
this test checks if the edge ej is of type (iii) or not.
6.5. LAYER V: THE SURFACE IN TERMS OF BASIS FUNCTIONS
77
Test (3a) tests whether fk is one of the faces abutting vb . If this is not the case, then
as implied, both the face midpoint and the inner edge quartiles associated with f k are
equal to zero. Test (3b) is equivalent since if a face is one of the faces abutting a vertex,
then this vertex must be in this face.
Let us turn to the boundary case. In fact, we can make some sharper tests at the
boundary, but we will use the same test for simplicity.
For a boundary vertex vi to be nonzero, the basis vertex vb must satisfy either (i)
vb = vi or (ii) vb must be in one of the two boundary edges connecting to vi . Thus, if
any of the tests in (1) are true, then neither of (i) or (ii) could be true.
For an edge midpoint (ej ) or an edge quartile (vi , ej ) to be nonzero, the basis vertex
vb must be either of the two vertices in the edge. Thus, if ej is not in any face abutting
vb , then this cannot be the case.
6.5 Layer V: The surface in terms of basis functions
Theorem 6.5.1. The surface can be written as a linear combination of the basis functions, that is, a linear combination of the control points with the basis functions as
weights,
s(c, s, t) =
Nb
X
B(vb ; c, s, t)p(vb ),
(6.30)
b=1
where
B(vb ; c, s, t) = (1 − s)2 (1 − t)2 Bbc |V (vb ; v) +
2(1 − s)2 (1 − t)t Bbc |EQ (vb ; v, eout ) +
(1 − s)2 t2 Bbc |EM (vb ; eout ) +
2(1 − s)s(1 − t)2 Bbc |EQ (vb ; v, ein ) +
4(1 − s)s(1 − t)t Bbc |QM (vb ; v, f ) +
(6.31)
2
2(1 − s)st Bbc |IQ (vb ; eout , f ) +
s2 (1 − t)2 Bbc |EM (vb ; ein ) +
2s2 (1 − t)t Bbc |IQ (vb ; ein , f ) +
s2 t2 Bbc |F M (vb ; f ),
and where v,f ,ein and eout are the vertex, face, incoming edge and outgoing edge of
corner c = (vi , fk ) respectively.
Proof. From definition 5.6.1 and lemma 6.4.1 we have
= B(vb ; c, s, t)p(vb ),
(6.32)
78
CHAPTER 6. THE SIMPLIFIED SURFACE SPLINE BASIS
where
s(c, s, t) = (1 − s)2 (1 − t)2
2
2(1 − s) (1 − t)t
(1 − s)2 t2
2(1 − s)s(1 − t)2
4(1 − s)s(1 − t)t
2(1 − s)st2
s2 (1 − t)2
2
2s (1 − t)t
s 2 t2
Nb
X
b=1
Nb
X
b=1
Nb
X
b=1
Nb
X
b=1
Nb
X
b=1
Nb
X
b=1
Nb
X
b=1
Nb
X
b=1
Nb
X
b=1
Bbc |V (vb ; v)p(vb ) +
Bbc |EQ (vb ; v, eout )p(vb ) +
Bbc |EM (vb ; eout )p(vb ) +
Bbc |EQ (vb ; v, ein )p(vb ) +
Bbc |QM (vb ; v, f )p(vb ) +
(6.33)
Bbc |IQ (vb ; eout , f )p(vb ) +
Bbc |EM (vb ; ein )p(vb ) +
Bbc |IQ (vb ; ein , f )p(vb ) +
Bbc |F M (vb ; f )p(vb ),
The sum is common for all the terms, and thus can be moved outside the expressions.
This applies to p(vb ) as well. We then get
s(c, s, t) =
Nb
X
(1 − s)2 (1 − t)2 Bbc |V (vb ; v) +
b=1
2(1 − s)2 (1 − t)t Bbc |EQ (vb ; v, eout ) +
(1 − s)2 t2 Bbc |EM (vb ; eout ) +
2(1 − s)s(1 − t)2 Bbc |EQ (vb ; v, ein ) +
4(1 − s)s(1 − t)t Bbc |QM (vb ; v, f ) +
(6.34)
2(1 − s)st2 Bbc |IQ (vb ; eout , f ) +
s2 (1 − t)2 Bbc |EM (vb ; ein ) +
2s2 (1 − t)t Bbc |IQ (vb ; ein , f ) +
s2 t2 Bbc |F M (vb ; f ) p(vb ).
The expression inside the brackets are indeed the basis functions defined in theorem
6.5.1, and thus this theorem is proved.
The parameter values (c, u, v) where a basis function for a particular vertex of the
control mesh is nonzero is called the support of vb . Actually, we will not be that
specific, we will just denote the set of patches (indexed by the corner) which is nonzero.
6.6. THE SIMPLIFIED SURFACE SPLINE SPACE
Corollary 6.5.2. A patch c = (v, f ) is outside the support of v b if




[
vi ∈
/
M[v; ∗, ∗, fk ]


79
(6.35)
fk ∈M[f ;vb ,∗,∗]
that is, if vb is not in any face abutting vi , or equivalently




[
vb ∈
/
M[v; ∗, ∗, fk ]


(6.36)
fk ∈M[f ;vi ,∗,∗]
Otherwise, the patch may have at least one non-zero coefficient.
Proof. If vb is not in any face abutting vi , then neither is any face abutting one of the
two associated edges (since these faces abuts vi as well as the edges), and vb is not in
f (since f by definition abuts vi ), and finally, vi 6= vb . With this information, we know
from corollary 6.4.2 that none of the blended expressions are nonzero. The symmetry
of the condition is proved in corollary 6.4.2.
6.6 The simplified surface spline space
In the previous section we found a set of functions which can be used as weights when
performing the surface evaluation as one single linear combination. We called these
functions basis functions, since they are suitable candidates for a simplified surface
spline basis.
6.6.1 The space
We let Bi denote the basis function of vertex vi . We put all the n = Nv basis functions
of the mesh M into a row vector,
B(c, s, t) = B(v1 ; c, s, t) B(v2 ; c, s, t) . . . B(vn ; c, s, t) ,
(6.37)
where B(vb ; c, s, t) is defined in theorem 6.5.1, and the control mesh vertices in a
column vector,


p(v1 )
 p(v2 ) 


(6.38)
p =  . .
 .. 
p(vn )
The simplified surface spline is then the vector product of the two vectors,
s(c, s, t) = B(c, s, t)p.
(6.39)
Definition 6.6.1. A simplified surface spline space S is defined by the connectivity
and the blend ratios of a mesh, and is populated by all simplified surface splines with
this mesh. An object in this space is specified by a vector of control points, one control
point associated with each vertex in the mesh. Addition and scalar multiplication of
object in this space is defined by addition of the control point vectors of the objects.
Theorem 6.5.1 states that any object in S, can be written on the form of equation 6.39.
Lemma 6.6.2. The simplified surface spline space S is a linear vector space.
80
CHAPTER 6. THE SIMPLIFIED SURFACE SPLINE BASIS
6.6.2 Linear independence
We define what we mean by linearly independent basis functions, and how this relates
to the points in the quad midpoint layer. If the surface is closed, the quad midpoint
layer consists only of quad midpoints, otherwise, the surface has a boundary, and each
boundary edge contributes two edge quartiles to this layer.
Definition 6.6.3. A set of basis functions are linearly independent if and only if the
zero vector gives the zero surface,
B(c, s, t)p = 0 ∀(c ∈ M[v, f ; ∗, ∗, ∗], s ∈ [0, 1], t ∈ [0, 1])
m
T
p = 0 0 ... 0 .
(6.40)
The zero surface is a surface that is zero for all possible parameter points. We can relate
definition 6.6.3 to the points in the quad midpoint layer.
Lemma 6.6.4. The patches of a simplified surface spline surface can only be exactly
equal to zero if and only if all the points of the quad midpoint layer are equal to zero.
Proof. The quad midpoint layer consists of the quad midpoint and the edge quartiles
for the boundary edges. The weights used in the Bézier blending on each patch are
parametric polynomials, and thus all coefficients must be zero for this to be the zero
surface and vice versa.
Thus, if all the points in the quad midpoint layer are zero, all Bézier coefficients
are zero, and thus the surface is the zero surface. Conversely; each patch has exactly
one exclusive quad midpoint each. If at least one quad midpoint or edge quartile are
non-zero, some of the Bézier coefficients are non-zero, and thus, the surface cannot be
the zero surface.
We can reduce the problem to examine for what conditions all the quad midpoints are
equal to zero. Lemma 6.6.4 is the connection between the quad midpoint layer and
the zero surface. We will define the quad midpoint matrix, a matrix which transform
control mesh vertices into quad midpoints, to examine the connection between the
control points and the quad midpoint layer.
Definition 6.6.5. The quad midpoint matrix B qm is defined by


Bqm |QM (v1 ; c1 ) . . . Bqm |QM (vn ; c1 )


..
..
..


.
.
.


Bqm |QM (v1 ; cm ) . . . Bqm |QM (vn ; cm )

,
B qm = 

 Bqm |EQ (v1 ; g1 ) . . . Bqm |EQ (vn ; g1 ) 


..
.
.
..
..


.
Bqm |EQ (v1 ; gl ) . . . Bqm |EQ (vn ; gl )
(6.41)
where ci , 1 ≤ i ≤ m are the corners, vj , 1 ≤ j ≤ n are the vertices, and gj , 1 ≤ l ≤
n are the boundary edge-ends of the control mesh. The expression B qm |QM is defined
in lemma 6.3.1.
The matrix B qm is the product of the two matrices S qm · S ps in equation (6.1). Each
edge has two edge-ends, and thus B qm has #V columns and #C + 2#BE rows. A
vertex has at least one corner, and thus, this matrix has never less rows than columns.
6.6. THE SIMPLIFIED SURFACE SPLINE SPACE
81
Vertives of f
v3 v4
v7
v9
v3 v4 v7 v9
Entry extraction
{f,v3 }
{f,v7}
{f,v9 }
{f,v4 }
{f,v3 }
{f,v7}
Row permutation
Corners of f
{f,v9 }
{f,v4 }
v3 v4 v7 v9
{f,v3 }
{f,v4}
{f,v7 }
{f,v9 }
Local quad midpoint matrix of f
Figure 6.2: Construction of the local quad midpoint matrix.
Lemma 6.6.6. The simplified surface spline basis functions are linearly independent
if and only if the quad midpoint matrix B qm of definition 6.6.5 has full rank.
Proof. This follows from lemma 6.6.4. If it is of full rank, only the zero vector give
zero quad midpoints. If it isn’t of full rank, it has a non-trivial null-space.
For the rest of this chapter we will use a one-way implication instead of the two way
implication of lemma 6.6.6. In other words, we settle for finding one criterium implicating linear independent basis functions, a criterium not necessarily implied by linear
independent basis functions.
Each vertex has at least one abutting corner, and thus, the number of corners is
larger or equal to the number of vertices. Each corner corresponds to a row, and thus,
we can prove full rank of the quad midpoint matrix by examining the B qm |QM -rows
only. For simplicity, we let B qm |QM be a sub-matrix of the quad midpoint matrix
where all Bqm |EQ -rows are removed.
6.6.3 A local criterium
We will now show that diagonally dominant local quad midpoint matrices result in a
non-singular quad midpoint matrix. Each face has a local quad midpoint matrix, and
the entries of this matrix is a subset of the entries of the quad midpoint matrix.
Definition 6.6.7. The local quad midpoint matrix for face fk , denoted B L
qm |QM (fk ),
is a subset of the entries of the quad midpoint matrix. First, remove all rows not corresponding to corners of fk . Then, remove all columns not corresponding to vertices
of fk . Then permute the remaining rows such that the entry of vertex v i in row {f, vi }
lies on the diagonal.
82
CHAPTER 6. THE SIMPLIFIED SURFACE SPLINE BASIS
The construction is depicted in figure 6.2.
Lemma 6.6.8. The local quad midpoint matrix for face fk is a square matrix of size
#fk × #fk , and all the removed columns have zero as entries in the remaining rows.
Proof. There is a one-to-one correspondence between quad midpoints and corners, and
a corner can only belong to one face. There are #fk vertices in face fk , and thus the
same number of corners. Hence, #fk rows and #fk columns are retained and the
result is a square matrix of size #fk , see figure 6.2.
Each quad midpoint is a bilinear blend of only the vertices of the face, thus entries
of vertices not belonging to fk on rows belonging to fk are zero.
Each row of the quad midpoint matrix “belongs” to one single face, and thus, the rows
of the local quad midpoint matrices do not overlap. We will now prove that if all the
local quad midpoint matrices are diagonally dominant, then the quad midpoint matrix
B qm |QM is of full rank, and thus, B qm is of full rank as well.
The idea is to reverse the construction of definition 6.6.7. First, we pick N v rows
from the local matrices and pad them (the opposite of the column removal). In this way
we can build a Nv × Nv diagonally dominant matrix. By adding the unused rows from
the local matrices, and permuting rows this square matrix can be made identical to the
Lemma 6.6.9. If all local quad midpoint matrices are diagonally dominant, the quad
midpoint matrix is of full rank.
L
Proof. For each face fk , we define a padded local quad midpoint matrix B̂ qm (fk ).
This matrix is identical to B L
qm (fk ) except columns of zeros are insert into appropriate
places such that the matrix ordering along the columns match B qm . This matrix is of
size #fk × Nv .
For each vertex vi pick one of the abutting faces, which we call fk . Pick the row
L
corresponding to corner (fk , vi ) of B̂ qm (fk ) and let this be row i of a matrix which
we call D.
The matrix D is a Nv × Nv diagonally dominant matrix. By adding the remaining
unused rows from the padded local quad midpoint matrices, we can obtain B qm |QM
from D. The matrix D is of rank Nv , and is a sub-matrix of the quad midpoint matrix
B qm |QM . Hence B qm |QM is of rank Nv .
Note that the converse of lemma 6.6.9 is not necessarily true.
6.6.4 The local quad midpoint matrices
Then, if the local quad midpoint matrices are diagonal dominant, then the quad midpoint matrix is of full rank, and further, the basis functions are linearly independent.
Initially, we try blend ratios of 1/2 and see if we get diagonally dominant local quad
midpoint matrices.
We deduce an expression for the local quad midpoint matrix. For simplicity, we
rename the vertices of the face fk as v1 . . . vn , where n = #fk . We let indices “wrap”
such that vn+1 = v1 . Further, pi is the geometric position associated with vi . We
combine the steps of polyhedral split (definition 5.3.1) and quad midpoint generation
(definition 5.4.2) and after some get:
6.6. THE SIMPLIFIED SURFACE SPLINE SPACE
83
n
qmi =
1
1
1
1 X
p
pi + (pi + pi+1 ) + (pi−1 + pi ) +
4
8
8
4n j=1 j
n
1
1
1 X
1
= pi + pi+1 + pi−1 +
p .
2
8
8
4n j=1 j
(6.42)
In the case of a triangle we get three expressions
1
(14p1 + 5 (p2 + p3 ))
24
1
qm2 =
(14p2 + 5 (p3 + p1 ))
24
1
qm3 =
(14p3 + 5 (p1 + p2 )) .
24
qm1 =
(6.43)
which we can pose in matrix form,


 
qm1
p1
qm2  = B qm |QM (fk ) p2 
qm3
p3
 



p1
14 5 5
qm1
1
 5 14 5  p2  .
qm2  =
24 5 5 14 p
qm
3
(6.44)
3
The local quad midpoint matrix is of size 3 × 3, and 14 is larger than 5 + 5, and thus
this matrix is diagonally dominant. If we do the same for a quadrilateral, we get



 
qm1
9 3 1 3 p1
qm2 
 
1 


3 9 3 1 p2 
(6.45)
qm3  = 16 1 3 9 3 p3  ,
qm4
3 1 3 9 p4
which is diagonally dominant as well.
By investigating equation (6.42), we can obtain a general formulation for the elements of the local quad midpoint matrix. We let aji denote the element at position j, i
and get
aj,i

1

2 +
= 81 +

1
4n
1
4n
1
4n
if i = j
if i = mod(j − 1, n) or i = mod(j + 1, n),
otherwise.
(6.46)
1
, larger than
The rows are convex combinations, and the diagonal element is 21 + 4n
1/2, and thus each row is diagonally dominant.
However, requiring all blend ratios to be equal to 1/2 is an unnecessary rigid condition, so let us consider what happens in the general case.
CHAPTER 6. THE SIMPLIFIED SURFACE SPLINE BASIS
84
Lemma 6.6.10. The local quad midpoint matrix is diagonally dominant if the blend
ratios for each corner of the face satisfy
bin bout
1
(1 − bin − bout ) +
> 0,
2
#f
(6.47)
where f is the face of the corner.
Proof. We set n = #f . We combine definitions 5.3.1 and 5.4.2, and organise the
coefficient in a cyclic order. The expression for the rows of the local quad midpoint
matrix is
1
pmi = (1 − biin )(1 − biout )pi + (1 − biin )biout (pi−1 + pi )+
2
n
bi bi X
1 i
bin (1 − biout )(pi + pi+1 ) + in out
p .
2
n j=1 j
(6.48)
The weight of pi is the diagonal entry in the expression of pmi . If this is larger than
1/2, then the row is diagonal dominant. With some algebraic juggling we get
pmi =
1
1
bi bi
(1 − biin )(1 − biout ) + (1 − biin )biout + biin (1 − biout ) + in out
2
2
n
i i
b
b
1
(1 − biin )biout + in out pi−1 +
2
n
i i
1 i
b b
bin (1 − biout ) + in out pi+1 +
2
n
i i
X
bin bout
p ,
n j=1,...,i−2,i+2,...,n j
pi +
(6.49)
and we pose the following inequality
1
1
bi bi
1
(1 − biin )(1 − biout ) + (1 − biin )biout + biin (1 − biout ) + in out >
2
2
n
2
biout
biin biout
1
biin
−
+
− >0
1−
2
2
n
2
biin biout
1
i
i
1 − bin − bout +
>0
2
n
(6.50)
Thus, if this requirement is fulfilled, the local quad midpoint matrix is diagonally dominant.
Figure 6.3 gives an indication of where the inequality of lemma 6.6.10 is satisfied.
6.6.5 The simplified surface spline space and its basis
We combine the results of this section to define when the basis functions form a basis
for S, the space of simplified surface splines for a given connectivity.
6.6. THE SIMPLIFIED SURFACE SPLINE SPACE
85
n=3
n=4
1
1
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0.2
0.4
0.6
0.8
1
0.2
0.4
n=8
1
1
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0.2
0.4
0.6
0.8
1
0.6
0.8
1
n=100
0.6
0.8
1
0.2
0.4
Figure 6.3: The domain where the inequality of lemma 6.6.10 is satisfied for a triangle (upper left), a quadrilateral (upper right), a octagon (lower left) and a 100-sided
polygon (lower right). Notice that the domain converges towards u + v < 12 when the
vertex count increases.
86
CHAPTER 6. THE SIMPLIFIED SURFACE SPLINE BASIS
Theorem 6.6.11. The simplified surface spline basis functions form a basis for S if
(but not necessarily only if) all blend ratios satisfy
bin bout
1
(1 − bin − bout ) +
> 0.
2
#fk
(6.51)
In this case, the dimension of S is Nv , the number of control vertices.
Proof. This follows from combining lemmas 6.6.9 and 6.6.10 and definition 6.6.1. The
basis functions are linearly independent and therefore form a basis for the space S of
simplified surface splines.
6.7 Summary
In this chapter we have found a set of basis functions for the simplified surface spline.
These functions were found by inserting an extra summation over the vertices inside
the scheme, and then moving this summation step by step to the outside of the scheme.
The result is a sum over the basis functions. We also investigated the support of these
basis functions.
In addition, we investigated when this set of basis functions formed a basis. We
defined linear independence between basis functions, and linked this definition to the
quad midpoints. Further, we found a local criterium for the blend ratios which gives
linearly independent basis functions, which in turn forms a basis.
Chapter 7
Practical considerations
In this chapter we investigate some aspects of putting the simplified surface spline into
practice. Many of the ideas presented in this chapter apply to C 1 -surface splines, and
some ideas apply to any scheme based on convex combinations, for example B-splines
and subdivision surfaces.
7.1 Evaluation of layer based schemes
The discussion presented in this section applies to most layer based schemes with a
finite number of layers. A layer based scheme is a scheme where each layer represents
a set of intermediate points, where each point in one layer is a convex combination of
points in the layer below. The initial layer is populated with the control points, and the
final layer are the evaluated point on the surface. Both surface splines and B-splines
are such schemes.
There exist two major strategies when evaluating the simplified surface spline. The
first strategy, called direct evaluation, begins with the control points, applies the layers
of the scheme, and ends up the position of the surface at the given parameter point.
The other strategy, called indirect evaluation, applies the layers in the opposite
order. We find the amount of influence each control point has for the given parameter
value, and then multiply the this amount of influence with the position of the control
point. The sum of this over all the control points is the position of the surface at the
given parameter point.
These two strategies are analogous to the two major strategies used to evaluate Bsplines. Which to use of these two strategies depends on the context. As a rule of
thumb, if we want to evaluate the surface at the same parameter point for many sets of
control vertices (for example when performing animation), indirect evaluation is a natural choice. On the other hand, if evaluating the surface at numerous parameter points,
and the positions of the control vertices remains fixed, direct evaluation is probably
appropriate. In addition, direct evaluation implies less book-keeping, and is thus easier
to implement.
7.1.1 Evaluation in matrix guise
Each layer of the simplified surface spline scheme consists of convex combinations
of the results of the layer below, and each layer only depends on the layer below. In
87
88
CHAPTER 7. PRACTICAL CONSIDERATIONS
chapter 6 we stated that the simplified surface spline evaluation could be posed as a
series of matrix multiplications. We have
s(c, s, t) = S bz (s, t) · S sel (c) · S bc · S qm · S ps · p
(7.1)
Each matrix is a matrix of convex combinations.
7.1.2 Direct and indirect evaluation
There are in essence two ways to carry out these matrix multiplications. Direct and
indirect evaluation corresponds to whether the matrices are multiplied together from
the left or from the right.
The strategy of direct evaluation applies the matrix S ps to the control points p. The
result of this multiplication is then multiplied with S qm and so on. This corresponds to
!
s(c, s, t) = S bz (s, t) · S sel (c) · S bc · S qm · S ps · p
.
(7.2)
The parentheses reveal the order of evaluation.
The strategy of indirect evaluation works in the opposite direction. The matrices
S bz (s, t) and S sel (c) are first multiplied. The result of this is multiplied with S bc and
so on. In the last step the intermediate result is multiplied with p, the control points.
Written as a matrix product this is
!
s(c, s, t) =
S bz (s, t) · S sel (c) · S bc · S qm · S ps · p.
(7.3)
The last intermediate result is a row vector. This vector reveals the contribution of
each control point to the surface point. In fact, the entries of this vector are the basis
functions evaluated at a specific parameter point.
This matrix factorisation of the scheme is of more theoretical importance than of
practical use. The matrices are seldom calculated explicitly. However, comparing with
the two strategies given in chapter 4 we see that the strategy of section 5 (the scheme)
is in fact direct evaluation, while the strategy of section 6 (basis functions) is indirect
evaluation.
7.1.3 Retained Bézier evaluation
In the matrix product (equation 7.1) we see that only the two leftmost matrices of the
expression are dependent on the parameter value. The rest of the matrices are constant
for any parameter point. Thus it may be a good idea to split the evaluation in two; first
do the constant part, and then for each parameter point carry out the two multiplications
on the left. We get
s(c, s, t) = S bz (s, t) · S sel (c) · S bc · S qm · S ps · p .
{z
} |
|
{z
}
parameter dependent
(7.4)
constant
This expression can be split into a parameter dependent part and a constant part. The
constant part of the evaluation is performed once for every set of control points. The
parameter dependent part is performed once for every evaluation.
7.2. OPENGL RENDERING OF SURFACES WITH LOCAL SUPPORT
89
By retaining the Bézier evaluation, we can use the constant part to find the set of
patches defining the surface. Having the surface represented as a set of patches is often
useful for rendering, differentiation, integration, etc.
We can use both the direct and indirect approaches to find the constant part. Usually, the direct approach is the simplest. We can also find the basis functions expressed
as a set of Bézier coefficients,
B bc = S bc · S qm · S ps .
(7.5)
This is simply the constant part where the multiplication with p is omitted. The matrix
B bc is of size (Nv + 3Ne + Nf + 2Nc ) × Nv , and is thus a pretty large matrix.
However, from the local support property of the scheme, most of the entries are zero,
and thus a full storage and matrix product is not necessary. Each time p changes, we
can find all the new Bézier coefficients by multiplying B bc and p, and with a smart
sparse representation of the basis matrix, this can be quite effective.
7.2 OpenGL rendering of surfaces with local support
This section deals with OpenGL at a rather detailed level, and many terms specific to
rendering are used without definition. Books like (OpenGL ARB, Woo, Neider, Davis
& Schreiner 1999) and (Foley et al. 1996) give thorough definitions and discussion of
these terms. Even though this section is specific to OpenGL, the discussion applies to
Direct3D (MSDN n.d.) as well.
7.2.1 Efficient rendering of Bézier patches
OpenGL has support for Bézier patches of arbitrary degree at core level. The application developer has the choice between sending the coefficients to OpenGL, or to
tessellate the patch to triangles in the application. Using OpenGL for patch evaluation
is straightforward. Usually, the implementation is quite good, and the speed acceptable. In some cases, one can write specialised code that run faster. On the other hand,
due to the arrival of geometry engines on off-the-shelves consumer graphics card, the
Bézier evaluation has the potential to be carried out in hardware — quite superior to
software calculation in terms of speed. If OpenGL does the evaluation, this will be
done in hardware when this is possible.
It is the application’s responsibility to calculate the static part of equation 7.4 and
perform the actions of the matrix S sel (c), that is, shoving the correct coefficients for
each patch to OpenGL. The best way to obtain the Bézier coefficients varies among
different applications. We will give some thoughts on this question.
If the surface is constant, that is, the vertices are not changed except for a rigid
transformation, direct evaluation is probably most appropriate. First, calculate all the
Bézier coefficients. Then, for each patch, create a display list1 for the evaluation of
each patch. Rendering of the visible patches then become a simple call to the respective
display list. However, the number of display lists is limited, and therefore, if rendering
a high amount of patches, it can be a sound idea to cluster several patches into each
display list.
1 Display lists are a mechanism in OpenGL to store a set of OpenGL commands on the server. For
example, the rendering of an object could be encapsulated into a display list, and then this object would be
rendered by simply calling this display list.
CHAPTER 7. PRACTICAL CONSIDERATIONS
90
Top plane
Far plane
Left plane
Right plane
Near plane
Bottom plane
Viewpoint
Figure 7.1: The view frustum.
If the model is dynamic, that is, the points in p are changed over time (e.g. by animation), display lists are not appropriate. Much effort must then be put into extracting
the coefficients in the quickest way possible. Not all coefficients need to be calculated
— a large amount of the model could be culled (culling is elaborated on in the next
section), and coefficients used exclusively by culled patches are never used. Thus, we
must find which coefficients are in need of re-computation, and then pass the visible
patches to OpenGL.
7.2.2 Frustum culling
Culling of invisible parts of the scene are the bread and butter of fast rendering code.
OpenGL has highly optimised clipping code, and employs hardware to perform this
task if this is present. Thus, there is no need to cull at a very detailed level. Outperforming the clipping code of OpenGL is probably quite a challenge.
However, if it can be done without much computation, it is a good idea to cull at
object level. This keeps some pressure off the rendering pipeline. We will present an
algorithm to cull patches which is definitely outside the view frustum.
The view frustum
The view frustum is a convex set with the shape of a trapezoid. It is defined by either
the intersection of half-spaces (each defined by a clipping plane) or as the convex hull
of the eight points where three of the six before-mentioned clipping planes intersect,
see figure 7.1.
The view frustum is defined in clip coordinates, and our model is defined in world
coordinates. Thus, either the model must be transformed to clip coordinates, or the
view frustum must be transformed to world coordinates. The latter is the natural
choice. For this to make sense, we must take a superficial look at the vertex transformation pipeline. OpenGL does some transformations on the points the application
7.2. OPENGL RENDERING OF SURFACES WITH LOCAL SUPPORT
Normalized device coordinates
Eye coordinates
X
Y
Z
W
Vertex
Modelview
matrix
91
Projection
matrix
Perspective
division
Clip coordinates
Viewport
transform
X
Y
Z
Window coordinates
Figure 7.2: The vertex transformation pipeline of OpenGL.
passes into it, and these transformations are known collectively as the vertex transformation pipeline, see figure 7.2.
The vertex transformation pipeline is composed of three matrix multiplications and
a division. First, points expressed in world coordinates are passed from the application. These points are in homogenous coordinates. The points are converted to eye
coordinates by the model view matrix. The model view matrix defines the orientation
of the camera in the world. The eye coordinates are then transformed to clip coordinates by the projection matrix. Clipping takes place in clip coordinates, and after this,
the perspective division is done. This division transforms the point from homogenous
coordinates to non-homogenous coordinates. Finally, these points are transformed by
the view port transform, which scales and translates coordinates to match the pixel
resolution of the frame buffer.
The model-view matrix and the projection matrix combined define the transformation from world space to clip space, and thus the inverse transforms the view frustum to
world coordinates. However, this inversion is not necessary; we can find the view frustum directly from the product of the model-view matrix and the projection matrix. The
web site (Gribb 1999) gives an an overview on how to extract the half spaces defining
the frustum directly.
The view frustum defines in world space the part of the world is visible through the
current view. Thus, everything outside the view frustum is invisible.
Culling
The key element to our culling strategy is the local support property of the simplified
surface spline, as well as the fact that the scheme is composed of convex combinations. Thus, a patch resides inside the convex hull spanned by the control points which
have a non-zero contribution to that particular patch. Therefore, if this convex hull is
completely outside the view frustum, the patch is not visible.
All patches of one face (in the control mesh) have the same set of control points
with non-zero contribution. Thus, we can deal with all the patches of a face combined.
We test if a vertex is inside or outside the view frustum by checking its position
relative to the six half-spaces defining the view frustum. Each half-space is defined by
a plane,
Ax + By + Cy + D < 0,
(7.6)
where A, B, C and D are the coefficients from the plane equation. If the inequality is
satisfied, the vertex is inside the half space. This check requires three multiplications
and three additions and a comparison per half-space.
CHAPTER 7. PRACTICAL CONSIDERATIONS
92
The algorithm consists of two stages. We let ⊕ denote a bitwise OR, and % be a
prefix of numbers written in binary form. Further, we define the set A(f k ) to be the set
of control vertices contributing to the patches of fk (This set can be found with the aid
of corollary 6.5.2).
1. For each vertex vi : Check the position of the vertex against the six half-spaces,
and store the result as a bit-mask of six bits, denoted bm(vi ). If the vertex is
inside a half-space, the corresponding bit is set, otherwise it is not set.
2. For each face fk : We define the bitmask for the face bm(fk ),
bm(fk ) =
M
bm(vi ).
(7.7)
vj ∈A(fk )
If bm(fk ) = %111111 then render the patches of the face, otherwise cull the
patches.
Thus, bm(fk ) is the result from combining all the bitmasks of contributing vertices with
a bitwise OR. If at least one of the bits are not set, all of the patches are completely
outside a half-space.
We need the bit-mask to store the results from the six tests. Even though all the
supporting vertices are completely outside the view frustum, the patch does not necessarily reside completely outside the view frustum. This is why reason for all half space
checks are done for each patch.
7.2.3 Occlusion tests
Surfaces inside the view frustum can be occluded by other surfaces. This is usually dealt with either by using the painter’s algorithm or a depth buffer. Evaluating and triangulating patches invisible in the image is a waste of precious processor time. Two vendor specific extensions exists to deal with these problems: the
HP_occlusion_test and NV_occlusion_query extensions. Information on
extensions to the OpenGL standard can be found in (OpenGL Extension Registry n.d.).
These two extensions provide the means to do a “dry run” of the shape (or a bounding shape), to check if any pixels of this shape will be visible in the image. The dry run
do not alter the image, but the full rasterisation process is done, and therefore, the tests
have a rasterisation cost.
Rasterising a bunch of pixels, to find out whether to rasterise even more pixels, isn’t
always a wise choice — especially if the rasterisation is a bottleneck in the rendering
pipeline. Anyway, let’s assume that rasterisation is not the bottleneck. Then, to reduce
the number of potentially expensive Bézier evaluation and tessellations, we need a
simple bounding object to use in the dry run.
Examples of simple occlusion detection shapes are the following: The first and
obvious candidate is the convex hull of the control points contributing to the patches of
the face. However, the 3D convex hull is not particularly easy and fast to find.
An alternative is to use a coordinate system aligned bounding box. This is not an
optimal shape, but we just need the three pairs of minimum and maximum values, and
the dry run is the rasterisation of six quadrilaterals.
Instead of using polygons, we may use lines, lowering the pressure at the rasterisation stage. However, testing with lines can fail in pathological cases. With n vertices,
7.3. BENEFITS OF USING COMPOSITE POLYNOMIAL PATCHES
93
we need n2 lines to connect them all. The number of lines grows quite rapidly when
n increases, but usually n is rather low.
We conclude with an important, but not necessarily obvious note. When using
occlusion tests, it is important to render the scene from front to back (the opposite of
painter’s algorithm). This results in a maximum number of objects being culled by the
occlusion tests.
7.3 Benefits of using composite polynomial patches
The surface of the surface spline schemes is defined by a set of polynomial patches.
We will now discuss two useful aspects of this.
7.3.1 Ray tracing
Raytracing consists of intersecting an object with a parameterised line. The use of
raytracing is not limited to the image generation technique called raytracing: Picking
and simple collision detection, as an example, may make use of ray tracing. In an
interactive application, picking is the task if finding which object is pointed to by the
mouse. A strategy for this is to transform the pixel coordinates of the mouse pointer
into the world space and shoot a ray from this pixel and outwardly away from the eye
point and see which object this ray intersects.
An example of simple collision detection is the following: Assume for a moment
we are creating a car simulation, in where the world is represented as surface splines.
We need to know where the tires touch the ground, what kind of surface this is (asphalt,
gravel, grass) to determine the friction, and we must know the normal plane of the
surface at each wheel. A strategy to solve these problems is to employ ray tracing.
We begin by positioning the object in an approximate position. Then, for each point
of contact (for example, one point per wheel), we shoot a ray from this point and in
the direction of the gravity pull (usually straight downward). From this, we know
whether the points of contact are above or have pierced through the ground, and we
then reposition the car.
Thus, raytracing is a common task. Several strategies to raytrace polynomial patches
are given in the book (Arvo, Cook, Glassner, Hanrahan, Heckbert & Kirk 1989), in
which Hanrahan gives an extensive survey of ray-surface intersection algorithms:
1. We can tessellate the surface into planar triangles. Ray-triangle intersections are
easily dealt with.
2. We can convert our patches to implicit functions, and in our case we then get an
implicit function of degree eight.
3. We can rewrite the ray as an intersection between two planes and insert this into
the expression for the patch. We then get a pair of implicit equations, and by
elimination, we can write this as a single equation, and this polynomial will be
of degree four.
Algorithm (1) is the most common. However, if the number of rays are few, and we
do not need the tessellated surface for anything else, one of the other methods can be
suitable.
Algorithms (2) and (3) can be solved by some numerical solver like Newton iteration. Algorithm (3) solves the equation for the parameter coordinates, and this can
94
CHAPTER 7. PRACTICAL CONSIDERATIONS
be useful if we need additional information about the surface at that point besides the
point of intersection (e.g. derivatives and the normal plane).
7.3.2 Mass, centre of gravity and moment of inertia tensor
Often it is useful to find different physical properties of an object. When dealing with
rigid body dynamics, as an example, we need the mass, centre of gravity and the moment of inertia tensor to pose a suitable differential equation.
The mass, centre of gravity and moment of inertia tensor can often be calculated
quite easily for bodies defined by a simplified surface spline. All of these integrals are
volume integrals, and thus we need some way of specifying the limits of the integral
from the extent of the object — which is rarely a trivial exercise.
Depending on the density function, these integrals can be simplified from volume
integrals to surface integrals by the Gauss divergence theorem. The article (GonzalezOchoa et al. 1998) gives a thorough exposition of this. The key element is that the
surface spline surface is defined as a collection of standard Bézier patches, and these
are rather simple to perform surface integrals on.
Chapter 8
Approximation experiment
In this chapter we will put the simplified surface spline to practice; we will state some
approximation methods using this basis. To our knowledge, surface splines have not
been used extensively for approximation, and thus this experiment should be interesting.
8.1 Our choice of parameterisation
We have employed a very simple parameterisation in this experiment. The original,
that is the object which we try to approximate, is parameterised in the same manner
as the simplified surface spline. All of the originals have the same connectivity as
the surface spline used to approximate that particular original. Thus, in essence, the
originals and the approximation have the same mesh, but have a different scheme and
different control points.
The originals are polyhedral meshes, and they one of two surface schemes. The
first surface scheme refines the mesh once with polyhedral split, such that all faces in
the refined mesh are quads. Each of these quads are then patched with a bilinear Bézier
interpolation. The result is a method to define a surface for polyhedral meshes, which
gives flat faces if the polyhedron is planar, and can give a well-defined surface if the
polyhedron is not planar. Further, since the original is refined once with the polyhedral
split, they can be parameterised with the same parameterisation used for simplified
surface splines (see section 5.1.3).
The other surface scheme is almost identical to the first surface scheme, except that
evaluated points are normalised. The result of this is that all points on the surface is on
the unit sphere (a point on the surface is defined as the point where a line from the point
returned from the first scheme and the origin pierces the unit sphere). However, to “fill”
the complete sphere with surface, the polyhedral mesh used to define the surface must
enclose the origin entirely.
There are three types of parameter points we will use extensively, the parameter
point of a vertex, the parameter point of an edge midpoint and the parameter point
of a face midpoint. When evaluating the surface spline at such parameter points, we
can choose one among several abutting patches to evaluate. Since the surface is C 0 continuous, we get the same result regardless of which of these patches we choose.
Further, both the biquadratic and bilinear Bézier blends used to define the surfaces
interpolate their corners, and the corners are located at these three types of parameter
95
CHAPTER 8. APPROXIMATION EXPERIMENT
96
Model
cube0
cube1
cube2
cube3
cube4
sphere0
sphere1
sphere2
sphere3
sphere4
jittercube
bunny
Nv
8
26
98
386
1538
8
26
98
386
1538
26
473
Ne
12
48
192
768
3072
12
48
192
768
3072
48
1194
Nc
24
96
384
1536
6144
24
96
384
1536
6144
96
2388
Nf
6
24
96
384
1536
6
24
96
384
1536
24
723
Table 8.1: Miscellaneous data on the original models. The symbol N v denote number
of vertices, Ne denote number of edges, Nc denote number of corners, and Nf denote
number of faces.
points. Thus, we can use the Bézier coefficients directly, and skip the Bézier evaluation
altogether.
8.2 The original models
Four different models with different characteristics were employed to test out the four
approximation methods. Two of the models, the cube model and the sphere model,
were refined into four levels. Thus, twelve models were employed in total.
The blend ratios of all models were set to 1/2. Probably a lot of approximation
power could be gained by using more sensible choices for the blend ratios. However,
we chose to keep things simple, and these blend ratios makes the system at least solvable.
The first five of the models have their origin in the standard cube, “cube0”. This
cube was refined four times with the polyhedral split method (see section 2.4.2) producing “cube1”, “cube2”, “cube3” and “cube4”. A cube has sharp corners, and these
sharp corners were a challenge for the smooth simplified surface spline.
We also wanted to test smooth surfaces without sharp corners. We created five
spheres of different level of resolution by normalising the surface (see class ControlSphere in section 8.5) of the five cube models. Thus, the sphere and cube models have
equal connectivity at each subdivision level.
The “jittercube” is a perturbed version of “cube2”. This model was included to test
how the schemes would handle noisy and extreme data.
To test the method on “real-world” data, we used the “bunny”. The “bunny” is
a heavily decimated version of the bunny model from Stanford University Computer
Graphics Laboratory. This decimation resulted in quite a coarse model, and was therefore a challenge for approximation with a smooth surface. In addition to decimation,
many of the triangle pairs forming quadrilaterals have been exchanged with quadrilaterals.
8.3. ERROR, APPROXIMATION ORDER AND STABILITY
97
8.3 Error, approximation order and stability
The basis of our error estimate is the geometric distance between the surface of the
original and the surface of the approximation for a given parameter point. This distance
is calculated at several parameter points, and from these samples the mean and the
maximum value are calculated. The mean and the maximum value are used as two
estimates of the approximation error.
We will use these two error approximates to compare the different approximation
methods, and we will investigate if we can find some information on the approximation
order. We assume the error is on the form
Eh (f ) ∼ hn k f (n) k,
(8.1)
where f is the original, h is a measurement of the size of parameter domain resolution
and n is the approximation order. We let h be the maximum length of edges in the
and thus halves the length of each edge. We therefore assume that h is halved for each
refinement.
Thus, if the surface is smooth, we assume that
hn k f (n) k
Eh (f )
∼
≈ 2n .
Eh/2 (f )
(h/2)n k f (n) k
(8.2)
The simplified surface scheme has the convex hull property, and thus, should have an
approximation order of at least 2. However, the term k f (n) k grows rather large for the
cube models. The cube has sharp edges, and thus the first derivative is discontinuous,
and the second derivative is therefore infinite. Thus, we anticipate a behaviour of order
1 for the cube approximation experiment and a behaviour of order 2 for the sphere
experiment.
To give a notion of the stability of the simplified surface spline basis, we have
computed the condition number and the number of iterations of the linear systems to
be solved in the approximation experiments.
8.4 Four approximation methods
We have employed four different approximation methods in our experiment. The first,
“lazy approximation” is extremely simple. The next “vertex interpolation” is interpolation at each control vertex. The “least squares” increases the number of interpolation
constraints, while “faired least squares” adds a number of fairness constraints. The two
last methods produce over-determined systems, and we use a least squares construction
to minimise the residuals and thus find the best solution.
8.4.1 Lazy approximation
The name of the lazy approximation scheme was inspired by the lazy wavelets. The
reason for this name is that nothing at all has to be done. For each vertex, the original is
evaluated at the parameter point of this vertex, and the result is simply used as a control
point.
It is worth to note that this is indeed the same scheme as the variation diminishing approximation scheme for B-splines, however no variation diminishing properties
CHAPTER 8. APPROXIMATION EXPERIMENT
98
have been proved, except for the convex hull property. But the non-existence of these
properties hasn’t been proved either, and thus, they may exist.
8.4.2 Vertex interpolation
Vertex interpolation is a special case of interpolation. As mentioned in section 8.1, we
can use the Bézier coefficients at the vertices directly, and therefore skip the quadratic
Bézier evaluation. The linear system to be solved is
BV · p = F V ,
(8.3)
where B V is the matrix of basis functions for the Bézier coefficients at vertices, p is
the control points we’re trying to find, and finally, F V is the original evaluated at the
vertices. To simplify the notation, we let n = Nv , which is the number of vertices in
the control mesh. If we write out the matrices, we get


 

Bbc |V (v1 ; v1 ) . . . Bbc |V (vn ; v1 )
p(v1 )
f (v1 )

  ..   .. 
..
..
(8.4)

 .  =  . .
.
.
...
Bbc |V (v1 ; vn ) . . . Bbc |V (vn ; vn )
p(vn )
f (vn )
The matrix B V is a Nv × NV matrix and both the column vectors p and F V have Nv
entries. It is essential to know if and when this system has a solution.
Lemma 8.4.1. The linear system of equation 8.4 has a solution if B is a basis (theorem
6.6.11).
Proof. If theorem 6.6.11 is satisfied, the expressions for quad midpoints of corners
abutting the vertex of which the basis functions belongs are all larger than 1/2 (see the
proof of lemma 6.6.10). The Bézier coefficient at the vertex are the average of all the
From this we see that all the diagonal entries are averages of values larger than
1/2, and therefore this average is larger than 1/2. Proposition 5.5.2 states that the
Bézier coefficients are convex combinations of the control mesh vertices, and therefore
the matrix B V is diagonally dominant, and according to property 1.4.8 this matrix is
non-singular.
The system is not symmetric and positive definite, and our choice of solver fell on the
Gauss-Seidel algorithm.
8.4.3 Least squares
The vertex interpolation method of section 8.4.2 uses only the vertex positions of the
original. Each row of the vertex interpolation matrix is an interpolation constraint for
a vertex. To capture more of the surface shape, we can use a least squares construction
We add rows of further interpolation constraints, but adding more rows to the system results in an over-determined system. Such a system has more more constraints
than free variables. Hoping for a solution of such as system is rather futile. However,
we can find the “best” solution of this problem by using a norm.
With this motivation, we introduce our next method, the least squares method. The
least squares method is actually just a method to produce a specific minimum of an
8.4. FOUR APPROXIMATION METHODS
99
over-determined system, not how to build this system. We will choose a rather simple
strategy; we will try to interpolate the original at the vertices, the edge midpoints and
the face midpoints. This results in Nv +Ne +Nf equations and Nv degrees of freedom.
To minimise the error of the system, we will use the 2-norm (see (Trefethen & Bau
1997, p. 18)).
Property 8.4.2. Given an over-determined set of linear equations Ax = b, the solution of AT Ax = AT b minimises the 2-norm residual r =k b − Ax k2 . This problem
has a solution if and only if A has full rank.
The proof is given in any text on linear algebra, e.g. (Trefethen & Bau 1997).
To build the system, we define two matrices and two vectors in addition to B V and
F v , defined in section 8.4.2. We let n = Nv , m = Ne ,l = Nf and define:




Bbc |EM (v1 ; e1 ) . . . Bbc |EM (vn ; e1 )
f EM (e1 )




..
..
..
..
B EM = 

 F EM = 
.
.
.
.
BF M
Bbc |EM (v1 ; em ) . . . Bbc |EM (vn ; em )

Bbc |F M (v1 ; f1 ) . . . Bbc |EM (vn ; f1 )


..
..
=

.
.
...

Bbc |F M (v1 ; fl )
...
Bbc |EM (vn ; fl )
F FM
f EM (em )

f F M (f1 )


..
=
.
.

f F M (fl )
The entries of the two matrices are the expressions for the edge midpoints and face
midpoints of lemma 6.4.1.
To build the complete system, we stack these matrices and vectors on top of each
other, and we get the following over-determined linear system of equations:

  p(v )  

1
FV
BV
. 
B EM  
(8.5)
 ..  = F EM  .
F
BF M
F
M
p(vn )
In this system, there are n+m+l equations and n unknowns. If the blend ratios satisfy
the requirements of theorem 6.6.11, we know from lemma 8.4.1 that the matrix B V is
non-singular, and thus of full rank. Adding rows does not lower the rank of the matrix,
and thus, the complete stack of matrices has full rank. By property 8.4.2 we know
that we may minimise the residual of equation 8.5. When applying the transposed
matrices, the system becomes symmetric and positive definite, and thus, we may solve
this system with the conjugate gradient method.
When building these systems, the stacked matrix can become quite large. It is rather
easy to build the products AT A and AT b directly. Each element in the product of two
matrices is a regular inner product. Thus the entry âij of AT A is the inner product
of column i and column j of A. Similarly, the entries of AT b are inner products of a
column of AT and b. The matrices of basis functions are relatively sparse; all entries
corresponding to vertices outside the support of the basis vertex are zero. Using this
fact can speed up the inner product calculations.
8.4.4 Faired least squares
In the previous section we stacked a lot of interpolation conditions into a composite
system matrix to create an over-determined system, which we minimised with the least
100
CHAPTER 8. APPROXIMATION EXPERIMENT
interpolation, we will in this method add Nv fairing constraints.
A faired surface is a surface where smoothness has been minimised in some way,
resulting in a nice and smooth surface. We will use a rather naive fairing constraint in
this method. We want each vertex to lie where the averages of the neighbours of that
vertex is. We define


if vi = vj
−1
(8.6)
g(vi ; vj ) = 1/#M[e; vi , ∗, ∗] if {vj , vi } ∈ M[e; vi , ∗, ∗]


0
otherwise
to formulate this mathematically. The expression g(vi ; vj ) is equal to −1 if the vertices
are equal, and is one over the number of edges connecting v i if vi and vj are connected.
Otherwise, this expression is zero. The sum of all g(vi ; vj ) for a vertex vi sums to zero.
With this function we define the matrix of fairing conditions,


g(v1 ; v1 ) . . . g(vn ; v1 )


..
..
..
G = α
(8.7)
.
.
.
.
g(v1 ; vn ) . . . g(vn ; vn )
The parameter α defines the amount of contribution the fairing conditions give to the
system. Setting α = 0 results in an interpolation problem. Each of the rows in G sum
to zero, and the matrix is square and has Nv columns and rows.
By stacking the vertex interpolation conditions and the fairing conditions, and letting 0 denote a column vector of Nv zeros (the fairing conditions equals zero), we get
the system


p(v1 )
FV
BV  . 
.
(8.8)
=
0
G  . 
p(vn )
In this system we have 2Nv equations and Nv degrees of freedom. To solve this system,
we apply the least squares method of the previous section. As always with these kinds
of least squares systems, the linear problem to solve is positive definite and symmetric,
and thus the conjugate gradient method is a natural choice for solving the system.
8.5 Application
In this section we will briefly describe the application developed to perform the approximation experiments.
8.5.1 Overview
The aim of this program is to test out the four approximation methods. A class diagram
is given in figure 8.1.
All surfaces are subclasses of the common surface interface ControlMeshEmbedding. A ControlMeshEmbedding has the ability to be evaluated in terms of control
mesh parameter points and has the ability to render itself. A control mesh parameter
point is a corner index plus a point in the unit square (the same parametrisation used
by the simplified surface spline surface, see chapter 5.1.3). In addition, just for plain
8.5. APPLICATION
101
convenience, a ControlMeshEmbedding can also be evaluated at a vertex, at an edge
midpoint and at a face midpoint.
When rendering, the render function of ControlMeshEmbedding is called with a
valid OpenGL context. This function is virtual, and it is thus the responsibility of the
subclass to implement this. However, the ControlMeshEmbedding class provides
some help: It allocates textures and fills them with distance information (as well as calculating statistics on this information). Thus, the subclass only has to bind the correct
texture when rendering.
The SimplifiedSurfaceSpline class is implemented as a subclass of ControlMeshEmbedding. In addition, the class has an instance of a SimplifiedSurfaceSplineBasis, the basis in which the surface is expressed. In addition to fulfilling the ControlMeshEmbedding interface, the simplified surface spline has an approximation constructor. Instead of specifying the surface with control points, one can give the constructor another ControlMeshEmbedding and an approximation scheme. Four different approximation schemes are implemented: Lazy, interpolation, least squares and
faired least squares.
The application is very simple. First, it reads a control mesh (by using the SSSReader class), and define an original from this (by using either ControlPolyhedron
or ControlSphere) data. Then, a SimplifiedSurfaceSpline is created for each of the
four scheme by by passing the original and the appropriate scheme to the constructor. Finally, the rest of the program is a simple GLUT-program opening a window and
setting up an OpenGL context.
8.5.2 Nitty-gritty details
We will dive into the heart of the matter; the details of the classes. We divide the classes
in two: core classes and helper classes. The core classes represent the conceptual flow
of the program, while the helper classes, are, as the name implies, helper classes.
Core classes
• ControlMesh. This class is used to hold the connectivity and blend ratios (blend
ratios are only used by the simplified surface spline surface) for surfaces, and
answer mesh queries. However, the mesh query paradigm is not strictly followed
(the reason for this is that it was programmed prior to the development of this
notation, and the awkwardness of this class interface was one of the reasons
for developing this notation). It is also rather bloated, much information which
would be easily deducible with a more clever data structure is explicitly stored,
and therefore it has a large memory footprint. To summarise: It is not an ideal
implementation, but even so, it works.
• ControlMeshEmbedding. This is the superclass for the surfaces parameterised
by a control mesh. This class serves several purposes. Mainly, it is an abstract
interface for surfaces parametrised by control meshes. The classes sub-classing
this class must implement a few methods for evaluating itself and a method for
OpenGL rendering. In addition, as mentioned earlier, the class implements some
functionality to calculate distances between two surfaces and generating texture
maps of this information.
• SimplifiedSurfaceSplineBasis. This class has functionality to evaluate the basis functions defined by its control mesh. The implementation follows the math-
CHAPTER 8. APPROXIMATION EXPERIMENT
102
ControlMesh
Statistics
Original
Topological Queries
Min/max
Mean/variance
ControlMeshEmbedding
Calc statistics
virtual: OpenGL rendering
virtual: Evaluation
Base mesh
SimplifiedSurfaceSplineBasis
Basis function evaluation
ControlPolyhedron
ControlSphere
SimplifiedSurfaceSpline
OpenGL rendering
Evaluation
OpenGL rendering
Evaluation
OpenGL rendering
Evaluation
Approximation
Base mesh
linearsolver
Parse SSS files
Surfaces
Gauss−Seidel
Input
Base mesh
SSSWriter
Application
Output
Write SSS files
Figure 8.1: ]
Class diagram of application. Arrows denote class inheritance, and dotted lines with
diamond denotes a ’has-a’ relationship, where the class with diamond has one or more
objects of the class without the diamond.
8.5. APPLICATION
103
ematical definition rather closely. Thus, each evaluation results in a few calls
down the layers. A speed-up could be obtained by joining all or some of the
layers in the evaluation process, but this would give increased code-size and
complexity.
The class has functionality to determine whether or not a parameter point is
inside the support of the basis function of a vertex. This is used extensively
when building the approximation matrices, and gives a significant speed-up.
• SimplifiedSurfaceSpline. This is the implementation of the simplified surface
spline surface. It has a ControlMesh and a SimplifiedSurfaceSplineBasis defined by this ControlMesh. In addition, it has a set of points, one for each of
the vertices in the ControlMesh. The set of points defines the surface in the
SimplifiedSurfaceSplineBasis.
The approximation methods reside in the constructor. Instead of explicitly defining the surface with a set of points (which is done with the ControlPolyhedron
and ControlSphere), the approximation constructor is given a ControlMeshEmbedding and a choice of approximation. We elaborate on the approximation
methods in the section 8.4.
To evaluate the surface at a parameter point, all the basis functions with this
parameter point in its support are evaluated and multiplied with the point of their
associated vertex, and then added together. The result of this is the point on the
surface.
• ControlPolyhedron. This is in essence just a plain polyhedral mesh. However,
since control meshes allows n-gons, and they are not necessarily planar, some
tricks must be employed. The trick here is to refine the mesh once with the face
splitting scheme (similar to the surface spline), and use a bilinear interpolation
over each quad. The resulting surface is flat where the control mesh facets are
flat, and in addition is well-defined where the facets are not flat. A desired sideeffect of this is that it is possible to parameterise the surface with the control
mesh parameter points.
• ControlSphere. This class was implemented to test the methods on smooth
convex objects. This class is almost the same as a ControlPolyhedron, but all
evaluated points are normalised, and thus, the surface is “shrink-wrapped” to the
unit sphere (see section 8.1).
Helper classes
• Linearsolver. This is in fact not a class, but just some C-functions to build a
sparse matrix representation and solve these linear systems. Two solvers are
available, the Gauss-Seidel method, useful for asymmetric matrices, and the
conjugate gradient method, useful for symmetric and positive definite matrices.
These are straight-forward implementations of the pseudo-code given in (Cheney
& Kincaid 1996).
• Statistics. This is a simple class to calculate the mean, variance, minimum and
maximum of a series of numbers.
• SSSReader. Class reading the SSS-file format, a simple ASCII-based file format for simplified surface splines.
104
CHAPTER 8. APPROXIMATION EXPERIMENT
• SSSWriter. The class for writing SSS-files.
The approximation schemes
All of the four approximation methods are built as a linear problem represented as a
matrix equation. In the interpolation scheme, the A matrix is built, while in the least
squares-based schemes AT A is built directly. The systems are solved separately for
X, Y and Z. Most entries of the matrices are zeros, and advantage of this is taken by
examining the support of each basis function.
Most of the computation time is spent building the matrices. The evaluation of the
simplified surface spline basis is rather expensive since all layers are implemented explicitly, and not in a composite fashion. This results in the overhead of several function
calls per evaluation.
There are two major strategies for evaluation of the basis functions (see section
7.1). The easiest is to run the full scheme for each vertex, setting the value of the
current vertex to one, and the rest to zero. In this way, the resulting surface is the basis
function for that particular vertex. The other approach is to find expressions for the
basis functions analyticly, and then use these functions. This implementation uses the
latter approach.
8.6 Results
As a practical test for the simplified surface spline, we applied the four approximation
methods on several models. Images of the results can be found in figures 8.2 to 8.13. In
the upper left corner, there is an image of the approximation subject. In the upper right
corner the errors of the different methods are depicted. The black column represents
the mean value of the errors and the grey block is the maximum error for the method
(see section 8.3). In the middle and bottom rows the output of the four approximation
methods are depicted. The surface is coloured according to their local error. White is
zero error, and black is maximum error.
Data on the approximation experiments can be found in tables 8.2 to 8.5. In this
tables Eiµ is the mean error of model i, Ei∞ is the maximum error, EEi−1
is the ratio
i
between the error of the refined and unrefined version of a model. The number of
iterations used is in column “#it”, and the condition number of the matrix used in
the linear systems based on the 2norm and infitity norm is in column C 2 and C ∞
respectivly.
From the experiments, we see that the lazy approximation consistently gives an
approximation smaller than the original. Because of the convex hull property, this
comes as no surprise, and is particularly visible in the sphere2 approximation. The
lazy approximation has the largest maximum error for all models except the bunny
model, where the faired approximation has a slightly larger error. The mean error of the
lazy approximations are small for cube models, but large for sphere models. The local
convex hull property can be used to interpret this: When a local region is flat, the local
convex hull of that region is thin and the approximation resides inside this volume. In
curved areas, the local convex hull grows fatter, and the deviation increases. However,
the lazy approximation gives smooth and well-shaped control meshes.
Except for the sphere2 case, the vertex interpolation has larger average error than
the other methods. Another drawback of the vertex interpolation is the oscillations,
and thus, non-smooth and badly shaped control meshes. This is particularly obvious
8.6. RESULTS
105
in the bunny model. In addition, at sharp corners, the approximation has some major
artefacts that are not present in the original: Instead of smoothing out corners like the
lazy approximation, the vertex interpolation creates an artificial wobble.
The least squares approximation remedies the artefacts of vertex interpolation to
some extent. Oscillations are still present, but on a smaller scale. However, the control
mesh is not particularly well formed. The average error, as well as the maximum error,
is in general smaller compared with the other methods.
When measuring average and maximum error, the faired approximation has the
worst approximation performance on average. However, it does not make the models
too small (which lazy approximation did on the sphere2 model), and the oscillations
are moderate. In fact, the control mesh of this scheme is smoother than the control
meshes from the lazy approximation.
All of the approximations were of rather good condition. The systems solved by
the conjugate gradient method behaved very well, and a good solution was found after
few iterations.
As a conclusion, the least squares method had the best approximation properties,
but at the cost of an oscillating surface. The fair approximation had moderate approximation properties for all cases, but suffered little of the problems of the other schemes
(oscillations and making the model too small). Thus, a least squares method with some
fairing criteria is probably a suitable approximation scheme for the simplified surface
spline.
Introducing blend ratio estimation and better parameterisation could on the other
hand give completely different results. Moreover, as a critique to the experiment is the
fact that the models were specially crafted to expose weaknesses in the approximation
methods. Thus, real-life situations could give other results.
CHAPTER 8. APPROXIMATION EXPERIMENT
106
0.5
0.4
0.3
0.2
0.1
lazy
vertex interp. least squares
Original
Error plot
Lazy
Interpolation
Least squares
Faired
fair
Figure 8.2: Original: “cube0”, Nv = 8, Ne = 12, Nc = 24 and Nf = 6
8.6. RESULTS
107
0.25
0.2
0.15
0.1
0.05
lazy
vertex interp. least squares
Original
Error plot
Lazy
Interpolation
Least squares
Faired
fair
Figure 8.3: Original: “cube1”, Nv = 26, Ne = 48, Nc = 96 and Nf = 24
CHAPTER 8. APPROXIMATION EXPERIMENT
108
0.14
0.12
0.1
0.08
0.06
0.04
0.02
lazy
vertex interp. least squares
Original
Error plot
Lazy
Interpolation
Least squares
Faired
fair
Figure 8.4: Original: “cube2”, Nv = 98, Ne = 192, Nc = 384 and Nf = 96
8.6. RESULTS
109
0.07
0.06
0.05
0.04
0.03
0.02
0.01
lazy
vertex interp. least squares
Original
Error plot
Lazy
Interpolation
Least squares
Faired
fair
Figure 8.5: Original: “cube3”, Nv = 386, Ne = 768, Nc = 1536 and Nf = 384
CHAPTER 8. APPROXIMATION EXPERIMENT
110
0.035
0.03
0.025
0.02
0.015
0.01
0.005
lazy
vertex interp. least squares
Original
Error plot
Lazy
Interpolation
Least squares
Faired
fair
Figure 8.6: Original: “cube4”, Nv = 1538, Ne = 3072, Nc = 6144 and Nf = 1536
8.6. RESULTS
111
0.4
0.3
0.2
0.1
lazy
vertex interp. least squares
Original
Error plot
Lazy
Interpolation
Least squares
Faired
fair
Figure 8.7: Original: “sphere0”, Nv = 8, Ne = 12, Nc = 24 and Nf = 6
CHAPTER 8. APPROXIMATION EXPERIMENT
112
0.14
0.12
0.1
0.08
0.06
0.04
0.02
lazy
vertex interp. least squares
Original
Error plot
Lazy
Interpolation
Least squares
Faired
fair
Figure 8.8: Original: “sphere1”, Nv = 26, Ne = 48, Nc = 96 and Nf = 24
8.6. RESULTS
113
0.05
0.04
0.03
0.02
0.01
lazy
vertex interp. least squares
Original
Error plot
Lazy
Interpolation
Least squares
Faired
fair
Figure 8.9: Original: “sphere2”, Nv = 98, Ne = 192, Nc = 384 and Nf = 96
CHAPTER 8. APPROXIMATION EXPERIMENT
114
0.014
0.012
0.01
0.008
0.006
0.004
0.002
lazy
vertex interp. least squares
Original
Error plot
Lazy
Interpolation
Least squares
Faired
fair
Figure 8.10: Original: “sphere3”, Nv = 386, Ne = 768, Nc = 1536 and Nf = 384
8.6. RESULTS
115
0.007
0.006
0.005
0.004
0.003
0.002
0.001
lazy
vertex interp. least squares
Original
Error plot
Lazy
Interpolation
Least squares
Faired
fair
Figure 8.11: Original: “sphere4”, Nv = 386, Ne = 768, Nc = 1536 and Nf = 384
CHAPTER 8. APPROXIMATION EXPERIMENT
116
0.6
0.5
0.4
0.3
0.2
0.1
lazy
vertex interp. least squares
Original
Error plot
Lazy
Interpolation
Least squares
Faired
fair
Figure 8.12: Original: “jittercube”, Nv = 26, Ne = 48, Nc = 96 and Nf = 24
8.6. RESULTS
117
0.005
0.004
0.003
0.002
0.001
lazy
vertex interp. least squares
Original
Error plot
Lazy
Interpolation
Least squares
Faired
fair
Figure 8.13: Original: “bunny”, Nv = 473, Ne = 1194, Nc = 2388 and Nf = 723
CHAPTER 8. APPROXIMATION EXPERIMENT
118
µ
cube0
cube1
cube2
cube3
cube4
sphere0
sphere1
sphere2
sphere3
sphere4
jittercube
bunny
#it
C2
C∞
−1
∞
Ei−1
Ei∞
5.774 · 10
2.887 · 10−1
1.443 · 10−1
7.217 · 10−2
3.608 · 10−2
—
2.000
2.000
2.000
2.000
—
—
—
—
—
—
—
—
—
—
—
—
—
—
—
4.226 · 10−1
1.445 · 10−1
5.200 · 10−2
1.484 · 10−2
7.003 · 10−3
—
2.925
2.779
3.503
2.120
—
—
—
—
—
—
—
—
—
—
—
—
—
—
—
6.622 · 10−1
4.264 · 10−3
—
—
—
—
—
—
—
—
Ei−1
Eiµ
model
Ei∞
µ
Ei
−1
2.138 · 10
5.666 · 10−2
1.457 · 10−2
3.693 · 10−3
9.301 · 10−4
—
3.773
3.890
3.944
3.971
3.827 · 10−1
1.270 · 10−1
3.497 · 10−2
9.223 · 10−3
2.389 · 10−3
—
3.013
3.632
3.791
3.861
8.001 · 10−2
5.775 · 10−4
—
—
Table 8.2: Results from lazy approximation
model
cube0
cube1
cube2
cube3
cube4
sphere0
sphere1
sphere2
sphere3
sphere4
jittercube
bunny
µ
Ei−1
Eiµ
µ
Ei
−1
3.273 · 10
1.194 · 10−1
3.414 · 10−2
9.398 · 10−3
2.439 · 10−3
—
2.742
3.497
3.633
3.853
8.861 · 10−2
3.894 · 10−2
6.956 · 10−3
1.983 · 10−3
5.765 · 10−4
—
2.276
5.598
3.508
3.439
1.452 · 10−1
1.064 · 10−3
—
—
#it
C2
C∞
−1
∞
Ei−1
Ei∞
5.000 · 10
2.065 · 10−1
1.028 · 10−1
5.141 · 10−2
2.570 · 10−2
—
2.421
2.009
2.000
2.000
100
100
100
100
100
4.000
4.026
4.007
4.002
4.000
4.000
4.000
4.000
4.000
4.000
1.341 · 10−1
7.959 · 10−2
1.622 · 10−2
9.494 · 10−3
5.415 · 10−3
—
1.685
4.907
1.708
1.753
100
100
100
100
100
4.000
4.026
4.007
4.002
4.000
4.000
4.000
4.000
4.000
4.000
4.362 · 10−1
4.027 · 10−3
—
—
100
100
4.026
3.087
4.000
3.840
Ei∞
Table 8.3: Results from vertex interpolation
µ
model
Eiµ
Ei−1
cube0
cube1
cube2
cube3
cube4
1.611 · 10−1
6.235 · 10−2
1.981 · 10−2
5.964 · 10−3
1.643 · 10−3
—
2.584
3.148
3.321
3.631
8.221 · 10−2
5.183 · 10−4
—
—
sphere0
sphere1
sphere2
sphere3
sphere4
jittercube
bunny
6.103 · 10−2
3.451 · 10−2
5.983 · 10−3
1.521 · 10−3
4.263 · 10−4
µ
Ei
—
1.768
5.769
3.933
3.569
#it
C2
C∞
3.260 · 10−1
1.160 · 10−1
5.863 · 10−2
2.949 · 10−2
1.475 · 10−2
∞
Ei−1
Ei∞
—
2.809
1.979
1.988
1.999
3
4
12
37
60
—
1.200
3.856
1.858
1.952
4
9
15
39
59
5.200 · 10
6.128 · 10
6.336 · 10
6.385 · 10
6.397 · 10
5.200 · 10
6.625 · 10
6.481 · 10
6.469 · 10
6.469 · 10
2.789 · 10−1
2.874 · 10−3
—
—
14
42
6.128 · 10
2.661 · 10
6.625 · 10
7.224 · 10
Ei∞
8.369 · 10−2
6.973 · 10−2
1.808 · 10−2
9.734 · 10−3
4.987 · 10−3
5.200 · 10
6.128 · 10
6.336 · 10
6.385 · 10
6.397 · 10
Table 8.4: Results from least squares approximation
5.200 · 10
6.625 · 10
6.481 · 10
6.469 · 10
6.469 · 10
8.6. RESULTS
model
cube0
cube1
cube2
cube3
cube4
sphere0
sphere1
sphere2
sphere3
sphere4
jittercube
bunny
119
µ
Ei−1
Eiµ
µ
Ei
−1
1.173 · 10
1.035 · 10−1
3.423 · 10−2
9.484 · 10−3
2.499 · 10−3
—
1.133
3.023
3.610
3.796
2.886 · 10−1
4.350 · 10−2
1.445 · 10−2
2.027 · 10−3
5.497 · 10−4
—
6.634
3.010
7.130
3.688
1.418 · 10−1
8.919 · 10−4
—
—
#it
C2
C∞
−1
∞
Ei−1
Ei∞
3.997 · 10
1.571 · 10−1
7.956 · 10−2
3.979 · 10−2
1.990 · 10−2
—
2.544
1.975
1.999
2.000
2
2
7
11
12
2.185
2.341
2.312
2.312
2.314
2.846
3.438
3.477
2.312
3.503
3.338 · 10−1
7.509 · 10−2
2.459 · 10−2
1.245 · 10−2
6.412 · 10−3
—
4.446
3.053
1.975
1.942
2
2
7
10
11
2.185
2.341
2.312
2.312
2.314
2.846
3.438
3.477
2.312
3.503
5.749 · 10−1
5.166 · 10−3
—
—
9
9
2.341
1.779
3.438
3.590
Ei∞
Table 8.5: Results from faired approximation
120
CHAPTER 8. APPROXIMATION EXPERIMENT
Chapter 9
Conclusions and further work
The aim of this thesis is to explore the approximation properties of the surface spline
schemes. We will in this chapter try to state some conclusions as well as suggesting
topics for further research.
To explore the approximation properties of the surface spline schemes, we have developed the simplified version of the C 1 -surface spline, the simplified surface spline.
This scheme has explicit formulas for both surface evaluation, as well as basis functions, in addition to a known dimension of the space. Thus it should be quite a good
basis for approximation purposes.
However, the results were not particularly impressive. There can be several reasons
for this. First of all, the conditioning of the basis is not known. The upper bound in the
condition inequality is one (since the scheme is composed of convex combinations),
but the lower bound is unknown. Obtaining this lower bound can reveal whether or not
this basis is stable. However, this can be tested numerically, and the condition numbers
from the approximation experiments indicate stability.
Further, an approximation of a model where large and small faces intermixed would
be interesting, and could reveal more of the stability properties of the basis. In addition,
it would be interesting to run the tests with better parameterisations as well as a more
sensible utilisation of the blend ratios. It is probably easy to create some heuristics to
determine suitable blend ratios. Another interesting approximation project would be to
check if the condition number of the approximation problem is somewhat linked to the
degree of vertices.
Extending the scheme to the full C 1 -surface spline scheme will probably not improve the condition, since an increased number of operations are performed as well as
the planarisation-projections (which could have an effect on the dimension of the basis
as well) that transmogrify the mesh into a planar cut polyhedron. But, on the other
hand, the it is possible that the full scheme will deal with irregular corners in a better
way.
The importance of parameterisation cannot be emphasised too much. It is common
knowledge that oscillations often occur where excess parameter space is present. In
addition, the speed of the parameterisation is of high importance; sudden changes in
parameter speed can give strange results. Hence, an disproportion between geometric
size of the surface and the size of the corresponding parameter space is therefore indeed
a peril. The parameterisation used in the approximation experiments were particularly
simple, and thus a better parametrisation could improve the results.
In addition, the power of the blend ratios were not utilised either. The blend ratios
121
122
CHAPTER 9. CONCLUSIONS AND FURTHER WORK
have the ability to shape the parameter space, and thus make sharp corners and creases
more “natural” for the surface spline.
With this said, we have found a useful method of dealing with polygonal meshes
(and indeed surface splines), the mesh notation. In our opinion, the formulas are nice,
and hopefully equally understandable. It is our hope that the framework as well as the
results of this thesis can serve as an inspiration for further exploration of the approximation properties of the surface spline schemes.
Finding some bound on the distance between a C 1 -surface spline surface and a
simplified surface spline surface of the same coefficients would be interesting. If this
is small, the simplified surface spline basis would be a suitable approximation basis for
obtaining coefficients to be used with the full scheme.
The surface splines share a fundamental problem with subdivision surfaces, and this
is how to find the control mesh (base mesh). Current decimation algorithms are well
suited for coarse polygonal representations. Polygonals are inheritly flat, but surface
splines are curved, and thus, most standard decimation strategies are not necessarily
optimal for finding a control mesh. Decimation with a more suitable criteria based
on some intrinsic properties of the surface spline like “flatness” or curvature could be
interesting.
Bibliography
Arvo, J., Cook, R. L., Glassner, A., Hanrahan, P., Heckbert, P. S. & Kirk, D. (1989),
An Introduction to Ray Tracing, Morgan Kaufmann Publishers, Inc.
Catmull, E. & Clark, J. (1978), ‘Recursively generated B-spline surfaces on arbitrary
topological meshes’, Computer-aided Design 10(6), 350–355.
Cheney, W. & Kincaid, D. (1996), Numerical Analysis, 2. edn, Brooks/Cole Publishing
Company.
Doo, D. & Sabin, M. (1978), ‘Behaviour of recursive subdivision surfaces near extraordinary points’, Computer-aided Design 10(6), 356–360.
Eck, M. & Hoppe, H. (1996), Automatic reconstruction of b-spline surfaces of arbitrary
topological type, in ‘SIGGRAPH 96 Conference Proceedings’, Annual Conference Series, SIGGRAPH, ACM Press, pp. 325–334.
Elmasri, R. A. & Navathe, S. B. (1994), Fundamentals of database systems, 2nd edn,
The Benjamin/Cummings Publishing Company, Inc.
Farin, G. (1996), Curves and Surfaces for CAGD, a practical guide, 4. edn, Academic
press.
Foley, J. D., van Dam, A., Feiner, S. K. & Hughes, J. F. (1996), Computer Graphics:
Principle and practice, The systems programming series, 2. edn, Addison-Wesley
Publishing Company. in C.
Gallier, J. (2000), Curves and Surfaces in Geometric Modeling, theory and algorithms,
Morgan Kaufmann Publishers, Inc.
Gonzalez, C. & Peters, J. (1999), Localized hierarchy surface splines, in S. S.
J. Rossignac, ed., ‘ACM Symposium on Interactive 3D Graphics’, pp. 7–15.
Gonzalez-Ochoa, C., McCammon, S. & Peters, J. (1998), ‘Computing moments of objects enclosed by piecewise polynomial surfaces’, ACM Transactions on Graphics
17(3), 143–157.
Gribb, G. (1999), ‘Re: Getting the modelview matrix — how bad?’, OpenGL Gamedeveloper Mailinglist. http://apollo.iwt.uni-bielefeld.de/˜ ml_robot/OpenGL-121999/0347.html.
Halbwachs, Y., Courriouz, G., Renaud, X. & Repusseau, P. (1996), ‘Topological and
geometric characterization of fault networks using 3-dimensional generalized
maps’, Mathematical Geology 28(5), 625–656.
123
124
BIBLIOGRAPHY
Halbwachs, Y. & Hjelle, Ø. (1999), Generalized maps in geological modeling: Object
oriented design of topological kernel. To appear in the book Advances in Software
Tools for Scientific Computing edited by H.P. Langtangen and E. Quak.
Mørken, K. & Lyche, T. (2002), Spline methods. IN-329 Lecture notes.
MSDN (n.d.), ‘Direct3D architecture’, Microsoft Developer Network website.
http://msdn.microsoft.com/library/.
OpenGL ARB, Woo, M., Neider, J., Davis, T. & Schreiner, D. (1999), OpenGL programming guide: the official guide to learning OpenGL, version 1.2, 3rd edn,
OpenGL Extension Registry (n.d.), SGI web site.
sample/registry/.
http://oss.sgi.com/projects/ogl-
Peters, J. (1994), Surfaces of arbitrary topology constructed from biquadrics and bicubics, in N. S. Sapidis, ed., ‘Designing fair curves and surfaces: shape quality
in geometric modelling and computer aided design’, Society for Industrial and
Applied Mathematics, pp. 277–293.
Peters, J. (1995a), ‘C 1 -surface splines’, SIAM Journal on Numerical Analysis
32(2), 645–666.
Peters, J. (1995b), ‘Smoothing polyhedra made easy’, ACM Transactions on Graphics
14(2), 161–169.
Peters, J. (1996), ‘Curvature continous spline surfaces over irregular meshes’,
Computer-Aided Geometric Design 13(2), 101–131.
Peters, J. (1998), ‘Pcp2nurb — smooth free-form surfacing with linearly trimmed bicubic b-splines’, ACM Transactions on Mathematical Software 24(3), 261–267.
Peters, J. (2001), Handbook of Computer Aided Design, Springer-Verlag, chapter Geometric Continuity. In press.
Stollnitz, E. J., DeRose, T. D. & Salesin, D. H. (1996), Wavelets for Computer Graphics, theory and applications, Morgan Kaufmann Publishers.
Trefethen, L. N. & Bau, D. I. (1997), Numerical Linear Algebra, Society for Industrial
and Applied Mathematics.
Index
Bbc , 98, 99
C 1 surface spline, 87
C 1 -surface spline, 55
R1 × R2 , 18
M, 12
fˆ, 14
π(R), 18
σ(R), 18
C, 16
D, 16
EE, 16
HE, 16
B V , 98, 100
B EM , 99
B F M , 99
F V , 98, 100
F EM , 99
F F M , 99
G, 99, 100
ej , 12
ein , 14
eout , 14
fk , 12
g, 99
vi , 12
2-norm, 98, 99
coefficient, 98
Bézier curve
control points, 6
definition of, 6
Bézier curve segment, 49
Bézier patch, 45
Bézier surfaces
tensor product construction, 7
Basis
edge midpoint function, 99
face midpoint function, 99
function, 98
support, 99
vertex basis function, 98
Basis function, 2, 79–81, 86, 101
evaluation strategy, 104
Linearly independent, 81
linearly independent, 80, 86
support, 103, 104
Bernstein polynomial
definition of, 6
Bilinear interpolation, 103
Blend ratio, 1, 31–33, 45, 47, 58, 60,
96
criterium, 86
Boundary curve, 31
of patch, 50
Boundary edges, see mesh
Animation, 87
Approximation, 2, 95, 103, 104
faired least squares, 97
lazy, 97
least squares, 97
variation diminishing, 97
vertex interpolation, 97, 98
Arbitrary topology, 2
Attribute, see relational algebra
Cartesian product, see relational algebra
Circuit, see set
Clough-Tocher, see macro splines
Code
complexity, 101
size, 101
Coefficient, 97
Computation time, 104
Connectivity, 1, 12
kernel, 2
B-spline, 1, 45, 56, 58, 87
knot vector, 1
B-splines, 97
Bézier
125
INDEX
126
subset, 2
Constraints
fair, 99, 100
Constructor, 103
Continuity, 55
at a joint, 8
Control Mesh, 100, 101
Control mesh, 47, 48, 55–58, see mesh,
control
embedding, 95
Control point, 31–33
ControlMesh (class), 101, 103
ControlMeshEmbedding (class), 100, 101,
103
ControlPolyhedron (class), 101, 103
ControlSphere (class), 96, 101, 103
Convex combination, 1, 4, 33, 83, 98
definition of, 4
matrix form, 5
of a convex combination, 4
Convex hull, 4, 92
Convex set, 4
Core class, 101
Corner, see mesh
Dart, see mesh
Data structure, 101
Decimation, 96
Degree raising, 47
Direct3D, 89
Domain, see relational algebra
Doo-Sabin split, 52
Doo-Sabin subdivision, see Mesh, subdivision scheme
Dual mesh, see Mesh, dual, see mesh
Edge, see mesh
Edge cutting, 47
Edge midpoint, 48, 56
Edge quartile, 56
Edge-end, see mesh
Element, see set
Embedded mesh, see mesh
Evaluation
direct, 87, 88
in hardware, 89
indirect, 87, 88
retained, 88
Face, see mesh
refinement, 57
Face midpoint, 48, 56
Face split, see Mesh, subdivision scheme
Fair
constraints, 99, 100
Faired least squares, 101
Foreign key, see relational algebra
Front to back rendering, 92
Function call
G-map, 11, 12, 16, 19
Gauss-Seidel solver, 98
Geometrical midpoint, 32
Geometry engine, 89
Geometry kernel
interface, 19
GIS, 2
Half-edge, see mesh
Half-space, 91
Helper class, 101
Incoming edge, see mesh, corner
Inner edge quartile, 57
Inner mesh, 47, 57, see Mesh, inner
Inner product, 99
Interpolation, 101
bilinear, 60
constraint, 98
edge midpoint, 98
face midpoint, 98
vertex, 98
Interval
midpoint, 33
Interval midpoint, 32, 33
Irregular vertex, 48, 50
Knot spacing, 47, 58
Knot vector, 45
Layer based schemes, 87
Lazy approximation, 101
Least squares, 101
faired, 101
LeSS, 52
Linear system, 98
building, 103
composite, 99
103
INDEX
fair, 99
Gauss-Seidel solver, 98, 103
least squares, 98, 99
minimisation, 98
norm, 98
over determined, 98, 99
positive definite systems, 103
rank, 99
residual, 99
Linearsolver (class), 103
Local construction, 45
Local degree bounded, see Mesh
Local degree boundedness, 45, 48
Local planarisation, 45
Local support property, 89
Local topology change, 52
Logical midpoint, 32
Möbius strip, 12
Macro splines
Clough-Tocher, 1
Powell-Sabin, 1
Matrix
composite, 99
diagonal element, 83
diagonally dominant, 5, 81, 83, 98
factorisation, 88
multiplication, 99
non-singular, 81, 98
of convex combinations, see convex combination
positive definite, 98–100
rank, 82, 99
sparse, 89
square, 82
stacked, 5, 99, 100
symmetric, 98–100
transpose, 104
Memory footprint, 101
Mesh, 12, 19, 23, 79
boundary, 24
edges, see mesh, edge
simple, 24
closed, 24
connectivity, 20
control, 1
corner, 14, 16, 23, 25, 58, 83
consecutive ordering, 24
definition of, 14
edge in, 60
127
edge out, 60
example of, 16
incoming edge, 14
outgoing edge, 14
dart, 20
definition of, 16
projection, 22
projection definition, 22
query, 21, 22
query definition, 21
definition of, 12
dual, 24, 28
definition of, 25
edge, 12, 14, 19, 23, 56, 57
boundary, 16
definition of, 12
incoming, see mesh, corner
non-boundary, 16
outgoing, see mesh, corner
edge-end, 16
definition of, 16
edges, 12
embedded, 12, 25
face, 12, 14, 19, 23–25, 56, 57, 83
definition of, 12
half-edge, 16, 23
definition of, 16
inner, 25
local degree bounded, 28, 29
position, 14
query, 19, 21
example, 23
refinement, see Mesh, subdivision
scheme
scheme, 12
simple boundary, 12
splitting, see Mesh, subdivision scheme
subdivision scheme
Doo-Sabin, 28
polyhedral, 24–26, 28
vertex, 24, 27
triangle, 83
triangular mesh, 14
vertex, 12, 14, 19, 23–25, 27, 56
definition of, 12
degree, 27
vertices
of face, 14
128
Mesh notation, 1, 2
Mesh query
definition of, 23
Mesh refinement, 45, 47
Mid-edge subdivision, see Mesh, subdivision scheme
Midpoint
of interval, 33
of sub-interval, 33
Minimal degree, 45
Mixed derivatives, 50
Model
bunny, 96
Occlusion, 92
Offset invariant ordered set, see set
OpenGL, 89, 90, 100
HP_occlusion_test, 92
NV_occlusion_query, 92
context, 100, 101
coordinate system
clip, 90, 91
normalised device, 91
window, 91
world, 90, 91
display list, 92
GLUT, 101
occlusion test, 92
rasterisation, 92
texture, 101
textures, 100
transformation
modelview matrix, 90, 91
perspective division, 90, 91
projection matrix, 90, 91
viewport transform, 90, 91
vertex transformation, 90
view frustum, 90
Original, 95
Outgoing edge, see mesh, corner
Pair, 11, see set
Parameter point, 95, 103
edge midpoint, 95
face midpoint, 95
vertex, 95
Parameterisation, 95
Patch boundary, 58
PCP, see Polyhedron, planar cut
Pixel count, 92
INDEX
Planar cut polyhedron, 45, 47, 57
Plane equation, 91
Polyhedral split, 47, 96, see Mesh, subdivision scheme
Polyhedron, 25, 28, 29, 45
planar, 28
planar cut, 28
Powell-Sabin, see macro splines
Primary key, see relational algebra
Project operator, see relational algebra
Projective convexity, 47
Quad mesh, 47, 48, 55, 57, see mesh
matrix, 80–83
local, 81–83
Rectangular patch, 45
Refinement, 96, see Mesh, subdivision
scheme
Regular parameterisation, 9
Relation, see relational algebra
Relational algebra, 2
attribute, 18
Cartesian product, 18, 20
domain, 18, 19
foreign key, 18, 19
join operator, 20
primary key, 18
project operator, 18, 20, 22
relation, 18, 19
virtual, 20
select operator, 18, 21
tuple, 19
Relational database, 2, 19
Rendering, 100, 101
constant model, 89
dynamic model, 89
Scheme, see mesh
Select operator, see relational algebra
Set
circuit, 12
definition of, 11
length of, 12
element, 11
INDEX
offset invariant ordered, 11
definition of, 11
pair, 11, 12
definition of, 11
size of, 11
Simple boundary, see mesh
Simplified surface spline, 1, 2
parameter point, 58
SimplifiedSurfaceSpline (class), 101, 103
SimplifiedSurfaceSplineBasis (class), 101,
103
Smoothness, 99
Smoothness constraint, 45
Space
simplified surface spline, 79
dimension, 86
Splitting, see Mesh, subdivision scheme
SQL, 20
AND, 22
AS, 21
CREATE
TABLE, 18–20
VIEW, 21
FOREIGN KEY, 18, 20
FROM, 18, 21, 22
PRIMARY KEY, 18–20
REFERENCES, 18, 20
SELECT, 18, 21, 22
WHERE, 18, 21, 22
SSSWriter (class), 103
Stanford University, 96
Statistics
average, 103
mean, 103
Statistics (class), 103
Sub-interval, 32, 33
midpoint, 33
Subdivision
property, 2
Subdivision surface, 87
Surface
fair, 99
smoothness, 99
Surface spline, 1, 2
C 1 , 45
simplified, see Simplified surface
spline
Surface spline curve, 31
definition of, 31
129
Surface splines
C 1, 1
Tensor product
Bézier surface example, 7
topological 2-manifold, 12
Topological holes, 52
Triangular mesh, see mesh
Triangular patch, 45, 49
Vector space
linear, 79
Vertex, see mesh
Vertex classification, 91, 92
View frustum, 92
Wavelets
lazy, 97
Zero surface, 80
```