title | date | author | abstract | bibliography |
---|---|---|---|---|
Planar Quads in Architecture Free-Form Surfaces |
September 2018 |
Christian Dimitri, UPC BarcelonaTech |
This paper will cover the preprocessing techniques for planar quad meshes in architecture free-form surfaces. As a first step, we will be covering the problems and objectives behind planar quads for construction, their benefits, their metrics as well as their goals, considering their constraints for a better optimization of the candidate *PQ mesh*. Secondly, we will explain the several preprocessing algorithms that generate a candidate *PQ mesh* ready for optimization and apply them on four different type of surfaces. In addition to that, the output will be optimized according to it's properties qualifying it to be PQ meshes. The last-mentioned are based on scientific papers, and were applied to concrete architectural projects. Combining chapter two and chapter three iteratively, we will be hitting the last chapter of this paper; generating subdivision method algorithm and a quad planarization in order to have a PQ mesh. |
paper.bib |
Planar Quad meshes have been nearly ubiquitous in architecture and construction. A large body of data structures and geometry processing algorithms based on them has been developed in the literature and adapted in construction of free-form surfaces. This type of re-meshing has many advantages especially the semi-regular ones, and significant progresses were made in quadrilateral mesh generation and processing during the last years. In this paper, we will study four algorithms behind planar quad meshes and their goals in order to fulfill the objectives. We will apply them on four input surfaces having different curvatures.
In construction, planar quads should always be planar and their distribution on the mesh is preferably equidistant so that their size does not vary a lot. In the first section, the geometric properties of PQ meshes are introduced as well as their benefits over other ones. Therefore, the metrics and measures are split by type and explained graphically and mathematically.
A polygon face is planar if and only if its vertices
Planar quad meshes may be preferred over triangular meshes for construction reasons. In addition, planar quads have the same fabrication and assembling benefits as triangles. The advantages of planar quads meshes for construction over other meshes is that: PQ meshes have higher surface to edge ratio than triangles, thus, a lower mullion cost. PQ meshes consumes less energy during fabrication.
To have planar quads, several Figure_3 are mentioned below. For a better quality, the mathematical Figure_3 and the conditions are classified by face and by mesh [@fig:Figure_3 and @fig:Figure_4}. In addition to that, some conditions are translated to custom goals that improve the quality of the mesh.
The measurements and conditions applied to the mesh itself are:
The measurements and conditions applied to the elements of the mesh are:
Given four different meshes as inputs with different curvatures, several pre-processing techniques will be adapted in order to generate a PQ mesh with planar faces. The used techniques will depend on the surface type. Translation surfaces is an easy and fast algorithm to generate specific surfaces. However architecture free-form surfaces with high curvature require more complex algorithms to generate PQ meshes see [@fig:Figure_5].
Translation surfaces are limited and easy to generate. The quads generated are the proof that it is generated through a set of parallel vectors that result in a planar face. In addition to that they are homogeneous because adding the same length vector as a constraint leads to have evenly spaces faces and reduce the variance. If the sectional curves are plane and the vectors are parallel having the same length the result will respond to the design principle of a translation surface. Assuming that one direction of the quad mesh net to be the sectional curve, two design principles can appear:
- The row of longitudinal sectional curves form parallel vectors.
- The row of lateral sectional curves form parallel vectors.
The family of sectional curves
Several geometrical shapes have been developed in architecture during the history using translation surfaces. The elliptical paraboloid is the most familiar shape found in architecture. It is generated using the same principle, translating one parabolic curve against another.
In transition surfaces, some geometrical shapes admit boolean and joining operations, for example, the hyperbolic paraboloid is a type of translation surface that acknowledge such operations. By translating a parabolic curve over a hyperbolic the result is as seen in [@fig:Figure_8]
Scale Translation surfaces are generated by adding a scale parameter to the output curves
Some curve networks are a robust and efficient method to extract PQ meshes [@liu2006geometric]. Such method admits a huge variety of free-form surfaces. The advantage of designing a conjugate direction field is that the user possesses total freedom in controlling the PQ mesh layout [@zadravec2010designing]. Thus, the panels are flat and discretize the principal curvature lines see [@liu2006geometric].
In addition to that, it can admit free torsion node while aligning the curve networks with the stress and curvature field see [@fig:Figure_10] for more information on statics sensitive layout [@schiftner2010statics].
As seen in [@fig:Figure_11], conjugate curve networks are families of curves
Examples of Conjugate Curve Networks on Surfaces:
-
Suited for PQ meshes: [@liu2006geometric]
- The network of principle curvature lines see (@fig:Figure_13).
- In a translation surface of the form
$p(u,v)$ $\mathbf{p}(u)$ a sectional curve is translated along another curve generatrix$\mathbf{p}(v)$ and vice versa see @fig:Figure_6.
-
Less suited for PQ meshes:
-
Epipolar curves: The translation of a point
$p$ along a line$l$ and the intersection of the planes through the points$p(i)$ with that surface$\Phi$ generates asymptotic curves that are not suited for such meshing see (@fig:Figure_13). - Isophotic curves are conjugate to the system of the steepest descent curves respecting the z-axis see (@fig:Figure_13).
-
Epipolar curves: The translation of a point
On a smooth surface
- Four vectors {$\mathbf{{v}{i},{w}{i}.{-v}{i}{-w}{i}}$}
- Two scalar parameters {$\theta_{i},\alpha_{i}$}:
-
$\theta_{i}$ oriented angle between$e_{i,1}$ and$\mathbf{v_{i}}$ -
$\alpha_{i}$ oriented angle between$\mathbf{v_{i}}$ and$\mathbf{w_{i}}$ - They define the following:
$\mathbf{{v_{i}}=(cos\theta_{i},sin\theta_{i})}{^T}$ and$\mathbf{w_{i}=(cos(\theta_{i}+\alpha_{i}),sin(\theta_{i}+\alpha_{i})}{^T}$
-
When the input is a mesh and not a surface, it is preferable to imply isotropic re-meshing. In this case, the re-meshing tool mesh machine is used [@MeshMachine]. After re-meshing the given input meshes
The quality of the mesh is always better when the panels are aligned with the curvature or the stress lines. Given four different meshes
In [@fig:Figure_15], it is clear that the smoothed vector field and the parallel transport have been well generated. In order to find a smooth and aligned vector field
After smoothing the vector field in the previous step, a quad mesh can be computed after generating the conjugate networks [@liu2011general]. From the previous step a conjugate vector field
If the mesh possess negative curvature, the parametrization has to be done by patches, see @fig:Figure_17 otherwise the parametrization can be done on a single patch see @fig:Figure_18. The algorithm succeeds on all the meshes except for the last one $\mathbb{R}^{3}{4}$ where collisions appear. The global parameterization using frame fields [@fig:Figure_17] is computed at the index i to shape the new mesh in a 2D topology. For each of the given 3D meshes, align the topology with the given vector fields $[e{1},e_{2}]$ at index i. Therefore, such field can be easily manipulated by the user.
The streamlines [$Pl_{i}$] are traced on the 2D maps after integrating the Vector field
Given the conjugate filed
This method is different from the previous one. The network of curves
The principle curvature networks
The curve networks
After mapping the curve networks and rebuilding the quad mesh on the unit plane, it is now possible to remap the meshes on input geometry.The surface with a double curvature
Conical meshes are planar quad meshes which discretize principle curvature lines, possess an offset at a constant distance as well as planar connecting elements [@liu2006geometric] see [@fig:Figure_22]. A conical mesh is conical if and only if all of its vertices
Assuming that the sum of the opposite angles on a vertex
Triangular meshes are missing the offset property at a constant distance, while conical meshes answer to this property [@liu2006geometric]. The faces of a conical mesh are incident to a common vertex $\mathbf{v}{i,j}$ and tangent to a cone with an axis $Q{i,j}$. After offsetting, the axis remains the same and the faces are still tangent to the cone [@liu2006geometric]. The Languerre transformation [@liu2006geometric] contains one of the instances for offsetting planes by a fixed distance along their normal vector. The Languerre transformation preserves the conical meshes at the offset.
The spherical image is a fact where the vertices $\mathbf{v}{ij}$ of a PQ mesh built on a unit sphere are converted to the normal vectors of $Q{i,j}$. As the four faces incident to a common vertex $\mathbf{v}{ij}$ tangent to the same cone see [@fig:Figure_24], the normal vectors ${n{i,j}}$ on each of the four faces share the same angle with the cone's axis
PQ meshes generated after computing the principle curve networks are well suited to be optimized using conical meshes conditions. In order to do that, the angles and normals are measured and visualized with a gradient color that varies in a range between the common meshes see @fig:Figure_25 and @fig:Figure_26.
Left: The sum of the opposite angles
For each vertex $\mathbf{v}{i,j}$ on the mesh $\mathbb{R}^3$ minimize the sum of the opposite angles equals to zero $\omega{1}+\omega_{3}-\omega_{2}-\omega_{4}=0$ using [@Kangaroo3d] solver.
In order to optimize the panels, the planarity
After analyzing the panels under their required goals for them to be planar, the elements are optimized using kangaroo2 solver [@Kangaroo3d]. The results have been reduced noticeably see in [@fig:Figure_29 and @fig:Figure_30].
A coarse mesh that approximates the topology of a input surface can be subdivided using the catmull-clark algorithm [@Weaverbird]. For PQ meshes, the valence of the each vertex should be four, vertices with a valence more then four are considered as singularities. After applying the subdivision on the coarse mesh, singularities with negative indices take a negative curvature and singularities with a positive indices take a positive curvature see [@fig:Figure_31].
On the given input meshes, the curvature
Subsequently to the previous step, a 2D map by patches is generated. Such a method can help out predicting the pre-networking between singularities and avoiding unexpected ones. Therefore it is now possible to generate the coarse mesh following the 2D map see [@fig:Figure_33].
The catmull-clark algorithm is applied to the coarse meshes. Using kangaroo2 [@Kangaroo3d] the coarse mesh is pulled by constraining the latter's points on the input meshes. The returning outputs are the candidate PQ meshes that need iterative optimization [@fig:Figure_34].
After analyzing the panels under their required goals for them to be planar, the elements are optimized using kangaroo2 solver [@Kangaroo3d]. The results have been reduced noticeably see in [@fig:Figure_34 and @fig:Figure_35].
PQ meshes must show different results from mere geometry. The planarity of the faces should obey the goals in order to fulfil the basis of the planar quad meshes [@zadravec2010designing]. They are very hard to deal with when the input surface is a free-form. However some algorithms have shown the differences between them and their results. Having a conjugate direction field as a tool to control the mesh layout is very useful. Thus, generating PQ meshes from curve network is strongly accurate. The two different methods are almost planar after generation since they are extracted from the principal directions. The conical optimization has proven its effectiveness over planar quad meshes. By optimizing and combining the methods, the last one was to generate planar quads by subdividing a coarse mesh and then optimizing it to planar. The boundary condition has been neglected in these methods, however, we recommend the first method due to its smooth results and its manipulation liability.
For further research the boundary will be taken in consideration while generating the PQ meshes. The fourth mesh