Honeycomb Subdivision

Pxd5284, hxp9515
smooth any 2-manifold (or 2manifold with boundary) mesh [20, 21]. Subdivision surfaces assume that users first provide an ir- regular 2-manifold control mesh, M0 . By applying a set of subdivision rules, a sequence of finer and finer 2-manifold meshes M1 , M2 , . . . , Mn , . . . are created. These meshes eventually converge to a smooth limit surface M [10]. One way to classify subdivision schemes is based on what kind of regularity emerges with the application of the scheme [16]. In other words, the pattern of regular regions can be used to characterize the scheme. Based on regular regions, existing subdivision schemes can be classified into three major categories:

Abstract
In this paper, we introduce a new subdivision scheme which we call honeycomb subdivision. After one iteration of the scheme each vertex becomes exactly 3-valent and with consecutive applications regular regions strongly resembles a honeycomb. This scheme can be considered as a dual for triangle schemes. The major advantage of the new scheme is that it creates a natural looking mesh structure.

1. Introduction
Although subdivision surfaces were introduced more than 20 years ago by Doo, Sabin, Catmull and Clark[5, 8], they were ignored by the computer graphics industry until they were used in 1998 Academy Award-winning short film Geris Game by Pixar [7, 20]. Since then subdivision surfaces have become increasingly popular in the computer graphics and modeling industry. This is not a surprise since subdivision methods solve the fundamental problem of ten- sor product parametric surfaces [12, 13] without sacrificing the speed of shape computation [5, 8, 11, 9, 18, 7]. Unlike tensor product surfaces, with subdivision surfaces, control meshes do not have to have a regular rectangular structure. Subdivision algorithms can

 Corner cutting schemes such as Doo-Sabin [8, 3]: after one iteration all vertices 1 become 4-valent and the number of non-4-sided faces remains invariant after the first refinement. Vertex insertion schemes such as Catmull-Clark [5]: after one iteration all faces become 4-sided and the number of non-4-valent vertices is constant after the first refinement. Triangular schemes such as 3-subdivision [10, 11]: after one iteration all faces become 3-sided and the number of non-6-valent vertices is constant after the first refinement. Notice that in this list vertex insertion and corner cutting schemes are dual, i.e., one of them makes every face 4-sided, the other one makes every vertex 4-valent. There is currently no dual for triangular schemes. The scheme we present in this paper provides the missing dual for triangular schemes:
Corresponding Author: Address: 216 Langford Center, College Station, Texas 77843-3137. email: ergun@[Link]. phone: +(979) 8456599. fax: +(979) 845-4491.

Honeycomb schemes: after one iteration all vertices become 3-valent and the number of non-6-sided faces is constant after first refinement. We call such schemes honeycomb since the resulting meshes strongly resemble honeycombs, which our dictionary defines as (1) A structure of hexagonal, thin-walled cells constructed from beeswax by honeybees to hold honey and larvae, (2) Something resembling this structure in con- figuration or pattern. Figure 1 shows five iterations of our honeycomb scheme. In this example, the control mesh M0 is a dodecahedron, M1 is a truncated icosahedron or soccer ball [19]. More interestingly, the mesh strucures from M2 to M5 can be found in virus structures [17]. As seen in this example, the most important property of the new scheme is that the re- sulting mesh structures strongly resemble natural structures such as cells or honeycombs. It is also interesting to note that their structure looks similar to Voronoi diagrams.
1 Vertices and faces are also called vertets and facets in order to avoid confusion with the vertices and faces of a solid model [16].

The remainder of this paper is organized as follows. In

vertex v 0 (See Figure 2.B) (we assume that the vertex and edge indices are the next section, given in the we provide order of a face refinement rules traversing). for our general Compute the honeycomb position of scheme. In each new verSection 3, we tex as a linear introduce the combination of coeffi- cients of the old vertex our honeycomb positions algorithm that n are used to n {v , v , . . . , create the v 0, . . 1 .,v meshes shown in 1 }.n Note N that Figure 1. since v 0 s corSection 4 explains how to respond to implement a edges instead of vertices, in remeshing order to derive algorithm for the coefficients honeycomb of a subdivision. honeycomb Section 5 scheme, it will discusses be use- ful to implementation use the and results. Fifollowing nally, our linear conclusion is equation: given in Section 6.
0

v = X
n t) v + tv

(1) m

Figure 1. Five iterations of our honeyco mb scheme over a dodecah edron.

2. Refinement Rules for Honeycomb Schemes

We give the refinement rules for a general honeycomb

algorithm as follows (The algorithm is also illustrated in Figure 2, where the black vertices and edges are new and constitute the new mesh, and the gray vertices and edges are in the old mesh and are removed by the algorithm.): Step 1 : For each edge en of a face f = {e , e , . . . , e , . . . 0, e 1 } of n the N 1 mesh, create a new

m=0

where the an,m s are real coefficients, N is the valence of the face and t is a real number between 0 and 1.

g e n e r a t e d f o r

t h a t e d g e ( S e e

(A) Initial mesh

(B) New vertices are created (C) New faces are created

(D) New edges are inserted (E) Final mesh

Figure 2. Illustration of honeycomb process. Remark 1. Note that (1 t)vm + tv(m+1) is a point on (mod N ) the edge em . In our examples we generally use t = 0.5. Step 2 : For each face, create a new face by connecting all the new points that have been generated by that face (See Figure 2.C); Step 3 : For each edge of the mesh, connect the two new points that have been

Figure 2.D); Step 4 : Remove all old vertices and edges (See Figure 2.E). Remark 2. Let Fn , En and Vn denote the number of faces, edges and vertices created by nth iteration. It is straightforward to show that Fn = Fn1 + Vn1 , En = 3En1 and Vn = 2En1 . Based on these equa- tions, it follows that

In these equations, the parameter a in equation (2) is provided by the users and is as a tension parameter [4]. Note that the value of a controls the speed of convergence. We suggest making a a decreasing function of N , i.e, for higher-sided faces (larger values of N ) the algorithm must converge faster. In all the examples in this paper we use the following function for a a = 0.45 + lim Fn n 3.

1.25 . N

F
n 1

3. The Coefficients of Our Honeycomb Algo- rithm

Like any subdivision scheme relying on approximation, in honeycomb subdivision the coefficients an,m in step 1 in the above algorithm must satisfy the following conditions: 1. an,m 0 for all n and m and
N 1

In other words, this honeycomb subdivision scheme increases the number of faces 3fold at each iteration. Therefore, regular haneycomb scheme can also be called a 3 honeyco mb algorithm [10]. Remark 3. The support region of our honeycomb scheme (the region that is influenced by a given control vertex [16]) has a fractal boundary, a Koch island [15]. Remark 4. It is difficult to identify the precision set of any honeycomb scheme (comparing the scheme with a parameteric surface) since the regular regions of the honeycomb scheme do not correspond correspond known to any parameterization scheme such as the ten- sor product or box-spline parameterizations.

2 . f o r a l l n X a
n , m

= 1 .
m = 0

These conditions

guarantee convergence of the algorithm and provide C 0 continuity and affine invariance properties. The coefficients of our scheme are inspired by the coeffi- cients of the Doo-Sabin scheme. Let vertex indices in a face be given in rotation order. We use the following formula to compute the coefficients an,m which are needed to compute the position of v 0 . ( a if n =m

It is easy to verify that in this scheme, each coefficient an,m is always greater than zero if 1 > a > 0, and the co- efficients do add up to 1. By using Fourier analysis, sim- ilar to Doo-Sabins approach, we can show that the new scheme also has three distinct but complex eigenvalues2 . The complex part comes from the phase shift introduced by the term vm + v(m+1) (mod ()N ) . Absolute values of the other eigenvectors will be smaller than one if a < 1, regard- less of the value of N . Similar to DooSabins scheme, only one eigenvector corresponds to the eigenvalue 1, two eigen- vectors correspond to the second largest (in absolute value) eigenvalue and the rest of the eigenvectors correspond to the smallest eigenvalue. Because of this property our scheme provides tangent plane continuity.

the honeycomb algorithm creates a very natural looking mesh structure. It therefore can be used after other subdivision schemes to create natural looking meshes. The honeycomb algorithm can create lateral artifacts [16]. Figure 4.(B) shows such an artifact. Although the hexagonal prism control mesh shown in Figure 4. (A) is a convex shape, we see smooth but periodic a 1 otherwise (2) n , w m a h = e r 3N e 5 2(n = m) 3 + .N 2 c o s

lumps and bumps across the height as shown in Figure 4.(B). These periodic bumps are created during the first iteration hexagonalization and smoothed after one more iteration. We call this periodic bumping Bricklayers Problem since a straight boundary cannot be obtained with hexagonal tiles. Bricklayers solve this problem quite simply: they cut hexagonal
We want to point out that it is possible to derive coefficients that give real eigenvalues. In fact, we tested various sets of coefficients that gave real eigenvalues. However, the visual quality of the shapes turned out better with complex eigenvalues. Based on this information, we deduce that it is better to use complex eigenvalues because of the rotation inherent in the scheme.
2

4. Impl eme ntati on and Resu lts

The remeshing algorithm explained in the section 2 has been implemented in a C++ program. We applied the hon- eycomb algorithm to the various control shapes. Figure 3 shows examples of applying our honeycomb scheme. The images at the left in Figure 3 show the control meshes and images at the right show the mesh after four iterations of our honeycomb algorithm. It is interesting to note that

tiles along the borders. To implement their solution: (1) convert the faces of the control mesh into hexagons, (2) cut hexagonal faces at the boundaries as bricklayers. As shown in Figure 4.(D), when we use a control shape shown in Fig- ure 4.(C) periodic bumps disappear.

expect that this problem can be solved with non-stationary planarization schemes. We are currently working on the de- velopment of such a scheme. Another problem we face is that it is very difficult to identify the precision set for any given honeycomb scheme. Unfortunately, we still do not have a promising approach to solve this problem.

Figure 4. Lateral artifacts and bricklayers so- lution.

Figure 3. Examples of honeycomb subdivision. The images at the left show the control meshes and images at the right show the mesh structures after four iterations of our honeycomb algorithm.

6. Acknowledgments
We are very grateful to the Malcolm Sabin who gave us encouragement when we first identified the honeycomb remeshing scheme. We are thankful to Gary Greenfield for his helpful suggestions. This work was partially funded by the Research Council of College of Architecture and the In- terdisciplinary Program of Texas A&M University .

5. Discussion, Conclusion and Future Work

In this paper, we introduced a new subdivision algorithm. The algorithm provides a tension parameter to control the shape of subdivided surface. We implemented a honey- comb algorithm, over the DLFL structure. We observed that honeycomb schemes can create natural looking mesh structures. The main problem with honeycomb schemes are the lateral artifacts. Our current solution requires user intervention which may not be desirable for many applications. We

References
[1] E. Akleman and J. Chen, Guaranteeing the 2Manifold Property for meshes with Doubly Linked Face List, International Journal of Shape Modeling Volume 5, No 2, pp. 149-177. [2] E. Akleman, J. Chen, and V. Srinivasan, A New Paradigm for Changing Topology During Subdivision Modeling, Pacific Graphics 2000, October 2000, pp. 192-201.

[3] E. Akleman, J. Chen, F. Eryoldas and V. Srinivasan, Handle and Hole Improvement with a New Corner Cutting Scheme with Tension, Shape Modeling 2001, May 2001, pp. 183-192. [4] R. H. Bartels, J. C. Beatty, and B. A. Barsky, An Intro- duction to Splines for use in Computer Graphics and Geometric Modeling, Morgan Kaufmann Publishers, Los Altos, CA, 1987. [5] E. Catmull and J. Clark, Recursively Generated Bspline Surfaces on Arbitrary Topological Meshes, Computer Aided Design, No. 10, September 1978, pp. 350-355. [6] J. Chen, Algorithmic Graph Embeddings, Theoreti- cal Computer Science, No. 181, 1997, pp. 247-266.

[18] T. W. Sederberg, D. Sewell, and M. Sabin, NonUniform recursive Subdivision Surfaces, Computer Graphics, No. 32, August 1998, pp. 387-394. [19] R. Williams, The Geometrical Foundation of Natural Structures, Dover Publications, Inc., 1972. [20] D. Zorin and P. Schro der, co-editors, Subdivision for Modeling and Animation, ACM SIGGRAPH99 Course Notes no. 37, August, 1999. [21] D. Zorin and P. Schro der, editor, Subdivision for Mod- eling and Animation, ACM SIGGRAPH2000 Course Notes no. 23, July, 2000.

[7] T. DeRose. M. Kass and T. Truong, Subdivision Sur- faces in Character Animation, Computer Graphics, No. 32, August 1998, pp. 85-94. [8] D. Doo and M. Sabin, Behavior of Recursive Subdivision Surfaces Near Extraordinary Points, Computer Aided Design, No. 10, September 1978, pp. 356-360. [9] M. Halstead, M. Kass, and T. DeRose, Efficient, fair interpolation using Catmull-Clark surfaces, Computer Graphics, No. 27, August 1993, pp. 35-44. [10] Kobbelt L., 3-Subdivision, Computer Graphics, No. 34, August 2000, pp. 103-112. [11] C. Loop, Smooth Subdivision Surfaces Based on Tri- angles, Masters Thesis, Department of Mathematics, University of Utah, 1987. [12] C. Loop and T. DeRose, Generalized B-spline surfaces with arbitrary topology, Computer Graphics, No. 24, August 1991, pp. 101-165. [13] C. Loop, Smooth spline surfaces over irregular meshes, Computer Graphics, No. 28, August 1994, pp. 303-310. [14] M. Mantyla, An Introduction to Solid Modeling, Com- puter Science Press, Rockville, Ma., 1988. [15] B. Mandelbrot, The Fractal Geometry of Nature, W. H. Freeman and Co., New York, 1980. [16] M. Sabin, Subdivision: Tutorial Notes, Shape Mod- eling International 2001, Tutorial, May 2000. [17] I. Stewart, Game, Set and Math: Enigmas and Conun- drums, Penguin Books, London, 1991.