EUROGRAPHICS 2021 / N. Mitra and I.

Viola Volume 40 (2021), Number 2

(Guest Editors)

Quad Layouts via Constrained T-Mesh Quantization

M. Lyon1 , M. Campen2 , and L. Kobbelt1

1 RWTH Aachen University, Germany
2 Osnabrück University, Germany

Parametrization Aligned T-mesh Quad Layout

Figure 1: Given a seamless parametrization of a surface as input we construct a non-conforming T-mesh aligned with it. A modified version
of the motorcycle graph is employed for this purpose, in which traces do not stop at the first collision. We solve an integer linear program
to assign integers to the arcs of this T-mesh. By carefully constraining this quantization, the result implies a coarse conforming quad layout
whose separatrices deviate less than a user given bound from the directions prescribed by the seamless input parametrization.
Abstract
We present a robust and fast method for the creation of conforming quad layouts on surfaces. Our algorithm is based on the
quantization of a T-mesh, i.e. an assignment of integer lengths to the sides of a non-conforming rectangular partition of the
surface. This representation has the benefit of being able to encode an infinite number of layout connectivity options in a finite
manner, which guarantees that a valid layout can always be found. We carefully construct the T-mesh from a given seamless
parametrization such that the algorithm can provide guarantees on the results’ quality. In particular, the user can specify a
bound on the angular deviation of layout edges from prescribed directions. We solve an integer linear program (ILP) to find a
coarse quad layout adhering to that maximal deviation. Our algorithm is guaranteed to yield a conforming quad layout free of
T-junctions together with bounded angle distortion. Our results show that the presented method is fast, reliable, and achieves
high quality layouts.
CCS Concepts
• Computing methodologies → Computer graphics; Mesh models; Mesh geometry models; Shape modeling;

1. Introduction not explicitly promote is the global structure of the mesh, in partic-
ular the simplicity of the mesh’s so-called block structure or base
Quad meshes are the preferred discrete surface representation for complex [BLP∗ 13]§1.1. This aspect is closely related to the ques-
many shape modeling applications of design and engineering alike. tion how the irregular vertices are connected in the mesh by se-
Therefore, the automatic generation of such meshes has been an quences of edges. This connectivity constitutes the mesh’s quad
ongoing topic of research in computer graphics. For a high quality layout [Cam17]. If the quad mesh has a simple, i.e., a coarse layout
quad mesh, individual elements should have angles close to π/2 and it can be viewed as a “mesh of meshes”, a coarse quad partition
be aligned in certain ways, e.g. to the underlying surface’s prin- with finer regular quad grids inside each patch (Figure 1 right).
cipal curvature directions. A variety of approaches have been ex- This enables the construction of mesh hierarchies, the structured
plored for the generation of quad meshes [BLP∗ 13]. A class of parametrization of the mesh over simple domains, or the definition
algorithms with a particular focus on element shape and align- of spline spaces on top of the mesh [TPP∗ 11, MAB∗ 19, HSJ∗ 20].
ment quality is that of parametrization-based field guided methods
[KNP07, BZK09, KMZ11, BCE∗ 13, PPTSH14, ESCK16]. In this paper we present an algorithm for the creation of coarse
quad layouts on 2-manifold surfaces. Such layouts can then, for in-
An additional quality criterion that these methods, however, do stance, be refined to block-structured quad meshes, or be passed as

c 2021 The Author(s)

Computer Graphics Forum c 2021 The Eurographics Association and John
Wiley & Sons Ltd. Published by John Wiley & Sons Ltd.
 M. Lyon, M. Campen, & L. Kobbelt / Quad Layouts via Constrained T-Mesh Quantization

connectivity constraints to the above mentioned (natively layout- T-mesh with many T-junctions. It is obtained by tracing stream-
unaware) quad mesh generation techniques. The layout construc- lines of a surface cross field. In this layout so-called chord collapses
tion – akin to the above parametrization-based mesh generation are applied greedily, in order from narrow to wide, while excluding
methods – is built on top of a seamless surface parametrization. collapses that would result in too much directional deviation. Par-
This allows us to offer explicit control over the layout’s singu- ticularly on complex or closed surfaces, T-junctions may remain in
larity configuration and its directional alignment, to reliably yield the final layout, making it non-conforming.
valid layouts, and to share a common foundation with mesh gen-
These methods have in common that modifications are applied
eration techniques for seamless integration. As a particular feature,
iteratively in a greedy fashion.
our method offers explicit and precise control over the balance be-
tween the two key opposing objectives inherent to quad layouts:
coarseness and directional alignment. Layout from Separatrix Candidates. A common strategy to cre-
ate layouts from scratch is based on finding a set of separatrix can-
Problem Statement: Given a seamless surface parametrization didates, i.e., paths connecting pairs of prescribed (irregular) layout
with arbitrary singularities on a surface of arbitrary topology, gen- nodes in topologically distinct ways, from which a subset is then
erate a coarse quad layout with exactly these singularities as irreg- selected to define a complete layout.
ular nodes such that its arcs (also called separatrices) do not direc-
Razafindrazaka et al. [RRP15] trace isolines of a seamless pa-
tionally deviate from the parametrization’s isolines by more than a
rametrization, starting from its singularities (which form the lay-
user-given angular bound α.
out’s irregular nodes). Whenever two traces meet, this implies a
After reviewing previous methods for the creation of coarse quad separatrix candidate between their two origin singularities. Each
layouts in Section 2, we describe in Section 3 how we construct a candidate is associated with a cost, penalizing directional deviation
T-mesh by tracing parametric iso-lines of a seamless parametriza- from the parametric isolines. A binary problem is then solved to
tion taken as input. In our method we implicitly encode layout con- select a cost-minimizing subset that properly connects all singular-
nectivity via a discrete function on this T-mesh. Section 4 details ities without crossing in improper ways. In theory the candidate set
this encoding and shows how non-negative integers can be assigned is infinite; in practice one needs to restrict to a finite subset. This
to the T-mesh’s arcs such that a valid and high quality layout is im- may preclude the existence of a valid solution (or a high-quality so-
plied. Such an assignment, called quantization, can be found by lution). Tracing up to a maximum distance limit is reported to com-
solving an integer linear program as presented in Section 5, before monly work well, but the existence of a solution is not guaranteed
ultimately making the layout explicit in Section 6. We show the ef- (unless trial-and-error with increasing distance limit is performed).
fectiveness of our algorithm on a variety of examples in Section 7.
Pietroni et al. [PPM∗ 16] follow a similar approach but create
Figure 1 illustrates the process.
separatrix candidates based on a cross field [VCD∗ 16] rather than
a parametrization (which is harder to obtain with the required prop-
2. Related Work erties [CSZZ19]). Similar to [RRP15] they solve a binary linear
Early work that involved the generation of quad layouts as a sub- program to choose a non-conflicting subset of these. As a conse-
step produced rather unstructured layouts without any particular quence of not deriving directional guidance from a parametrization
form of shape-aware directional alignment [EH96, BMRJ04]. Later (which corresponds to an integrable cross field) but rather from
this problem received dedicated attention and subsequent methods a generic cross field, a complete conforming quad layout cannot
often take some form of directional guidance into account, whether be guaranteed; T-junctions have to be accepted, similar to the ap-
by means of the underlying surface’s principal curvature directions, proach of Viertel et al. [VOS19] discussed above. T-junctions may
or more flexibly and controllably by specifically designed or pre- be reduced by increasing the number of separatrix candidates es-
scribed cross fields, quad meshes, or seamless parametrizations. tablished for each singularity or by iterating the process with fixed
partial layout, but complete removal can only be achieved by in-
Layout Simplification. For instance, the method by Bommes et serting additional irregularities.
al. [BLK11] takes as input a quad mesh with a possibly dense base For both these methods, runtime, quality, and even the existence
complex and iteratively modifies it so as to remove certain helical of a solution depend on the precomputed set of separatrix candi-
connectivity patterns – which are one, though not the sole cause dates. If the set is chosen too small there may be no valid subset.
of low quality base complexes. A coarser quad layout can then be Increasing the candidate set size at the cost of increased runtime
extracted from the modified quad mesh. only increases the probability of program feasibility. In contrast,
Tarini et al. [TPP∗ 11] follow a similar strategy but enable more our formulation, instead of using a binary program which picks
general modifications by working directly on separatrices (i.e., the separatrices from a finite set of candidates, employs an integer pro-
paths forming the layout’s arcs, its patches’ borders). They itera- gram which can choose from an infinite set of separatrices and is
tively improve a layout energy which is based on length and di- always guaranteed to be feasible.
rection deviation of the separatrices by removing a separatrix and
[RP17] and [ZZY16] propose further candidate set based meth-
bringing the then incomplete layout back into a (coarser) complete
ods, but start from a quad mesh as input. This effectively enables
state by a series of separatrix reconnections and an insertion.
a fallback to the quad mesh itself or its base complex as a valid
Instead of starting with conforming dense quad layouts, Vier- (though commonly rather dense) output layout in case no other so-
tel et al. [VOS19] start from an initial non-conforming layout, a lution is found due to the involved restriction.

c 2021 The Author(s)

Computer Graphics Forum c 2021 The Eurographics Association and John Wiley & Sons Ltd.
 M. Lyon, M. Campen, & L. Kobbelt / Quad Layouts via Constrained T-Mesh Quantization

Dual Approach. Campen et al. [CBK12] tackle the problem in the

dual space and search for a set of dual loops which separate all sin-
gularities. While the dual setting allows for relatively simple con-
ditions guaranteeing a valid primal layout it does not allow explicit
control over the geometric quality of the implied separatrices.

Mesh as Layout. An alternative approach to create coarse quad

layouts is to use a generic quad mesh generation method (without
explicit layout considerations) and aim for a mesh with very large Figure 2: T-meshes on the B UNNY model created for angle devia-
elements – which can then be considered a coarse layout. While tion bound α = 5◦ , 15◦ , 35◦ , respectively.
some of these methods are robust enough to operate reliably under
this extreme requirement (e.g. [BCE∗ 13, CBK15]), it is typically
difficult for these methods to control the quality of the resulting αik ∈ [0, α] and αil ∈ [−α, 0], i.e., as soon as on
layout under these circumstances (cf. Section 7.1). both sides of trace ti a trace is found such that
the formed triangles are contained in the sector of
half-angle α around ti . With this stopping crite-
rion we ensure that we reach the traces of at least
3. T-Mesh Construction
one singularity on each side of the trace which, intuitively, could be
A motorcycle graph is a cell decomposition of a surface [EGKT08]. connected to singularity i by a separatrix which respects the maxi-
Campen et al. [CBK15, MC19] show how a rectangular partition mum angular deviation bound. We will see in detail in Section 4.3
of a surface can be constructed by tracing motorcycles starting at how this construction helps to guarantee a maximum separatrix de-
singularities along the iso-lines of a seamless parametrization. The viation. Figure 2 shows a few examples of the resulting T-meshes
resulting T-mesh T = (N , A, P) consists of nodes N for every for different α and how tighter bounds lead to longer traces.
singularity and intersection of traces, arcs A consisting of the seg-
ments of a trace between two nodes, and patches P representing 4. Quantization
rectangular regions bounded by arcs.
A quantization of a T-mesh is an assignment of a non-negative in-
Such a motorcycle graph forms the basis of our quantization teger qi to every arc ai ∈ A. These values are to be interpreted
(cf. Section 4) which assigns integer lengths to the arcs and thereby as parametric length specification for the arcs, virtually overrid-
defines the connectivty of the resulting layout. ing their length in the input parametrization. Since every arc is
associated with a parametric iso-line, the quantization implies an
We adopt a construction similar to [CBK15], with one important
assignment of integer parametric distances between singularities
difference: in our case motorcycles may survive a crash with an
[CBK15]. These distances can be used as constraints for a global
existing trace and continue driving. In such a case a regular valence
re-parametrization of the surface where every singularity is located
4 node is formed at the intersection. Before defining the criterion
on an integer position, enabling the extraction of a quad mesh
that determines this, we establish a few definitions and notation.
[CBK15, LCBK19]. This quad mesh’s base complex defines a quad
layout. The edges of the base complex – the separatrices – connect
Notation. Given two traces ti and t j starting in singularities singularities that lie on the same parametric iso-line. Effectively,
i and j, respectively, we refer to the node created at their the quantization implies which singularities are connected by sepa-
intersection as ni j ∈ N . (For simplicity of no- ratrices by assigning zero lengths to certain arcs. Depending on the
tation we ignore the fact that two traces may in- T-mesh structure and the quantization, singularities may be con-
tersect more than once; it will be clear from the nected by separatrices corresponding to different parametric direc-
context which intersection node is referred to.) tions (cf. Figure 3(left, middle)) or not (cf. Figure 3(right)).
We define Si j ⊂ A as the set of arcs between
the start of trace ti and node ni j , and li j ∈ R as While the above previous works aim to find a quantization which
the total parametric length of those arcs. The two arc sets Si j and approximates the arcs’ original parametric lengths, we detail in this
S ji form the legs of a right triangle. Let αi j ∈ (− π2 , π2 ) be the signed section how we instead find a quantization that promotes a coarse
(ccw) angle of that triangle at the start of trace ti . quad layout. We begin by discussing two important properties of a
quantization – validity and consistency – in the following section.
We further give a dedicated name, ni∗ , per trace ti to the We continue by presenting a sufficient condition which guarantees
intersection closest to the origin of ti which satisfies validity, and finally describe which additional constraints we set up
li∗ > l∗i , i.e. the first intersection of ti with a trace that to enforce a high quality quad layout.
starts inside the π/2-sector (blue) around ti (see inset).
In the classical motorcycle graph a trace ti would stop
4.1. Consistency and Validity
exactly at ni∗ . By contrast, we keep on tracing and only
stop based on the following criterion. In order to compute a valid (locally injective, fold-over free) para-
metrization that adheres to the singularity distance constraints de-
fined by the quantization, two properties need to be fulfilled. The
Stopping Criterion. Given a user defined angular bound α, a trace
quantization needs to be consistent and valid [CBK15].
is stopped as soon as it intersected two traces tk and tl such that
c 2021 The Author(s)
Computer Graphics Forum c 2021 The Eurographics Association and John Wiley & Sons Ltd.
 M. Lyon, M. Campen, & L. Kobbelt / Quad Layouts via Constrained T-Mesh Quantization

Proof. We show that every singularity in the π/2-sector centered

at a singularity i and around the positive u direction is separated
from i. The same argument then trivially holds for all other para-
metric directions as well.
Let j be the singularity at the start of t∗ . Then, j is separated from
i in u direction by at least 1 according to the lemma’s premise. Con-
sider another singularity k within the
π/2-sector and its trace tk in negative
(or positive) v direction towards ti . If
tk intersects the parametric iso-line of
Figure 3: Layouts resulting from different quantizations. Singular- ti it must do so further away than t∗
ities (dots) are separated in vertical direction leading to an angle since t∗ is the intersection closest to
α > π/4 compared to the desired direction (left). With our separa- i by definition. Since the intersection
tion constraint, singularities are separated in the direction corre- lies behind that of t∗ , k is separated by at least 1 in u direction
sponding to the larger difference in the seamless parametrization as well. If, on the other hand, the trace of k does not intersect the
always leading to separatrix angles smaller than π/4 (middle). If parametric iso-line of ti it must have been stopped before reaching
both separatrices form an unacceptably large angle, singularities it. Since at least one of the arcs of tk is quantized to at least one
need to be separated in both directions (right). Note that the ac- according to the condition in the lemma, k is separated from i in
tual quantization (beyond zero vs non-zero) is less important for positive (or negative) v direction.
the layout, consisting of the (bold) edges of the base complex only.
Note that Lemma 1 does not depend on our special T-mesh con-
struction and also holds if traces stop at the first intersection.
A quantization is consistent if pairs of opposite sides of each T-
mesh patch are quantized to the same length. This property ensures We conclude that it is sufficient to require one arc of Si∗ to be
that all patches remain rectangular in the parametric domain. quantized to a strictly positive number for each trace ti to ensure
A quantization is valid if the distance between any two singu- validity, i.e., one constraint per trace is sufficient.
larities is strictly positive, such that they do not collapse parametri-
cally. In terms of the quantization this means there must be no arc 4.3. Layout Constraints
path with total quantized length of 0 between any two singularities.
In the previous section we detailed how simple constraints on the
Let di j ∈ R2 be the (u, v)-difference between singularities i and j
minimal quantization of certain arcs guarantee a valid quantization.
measured in the seamless parametrization (in some path homotopy
Computing the smallest quantization adhering to these constraints
class) and qi j ∈ Z2 be the quantized difference vector. W.l.o.g. as-
results in a coarse quad layout. However, the separatrices of that
sume that duij > 0 and |duij | ≥ |dvij |. To ensure validity it is sufficient
layout may deviate up to π/4 from the directions of the seamless
to require quij > 0 ∨ qvij 6= 0 [CBK15]. We propose to use instead
parametrization (cf. Figure 3 b). In particular the separatrix corre-
the simpler sufficient condition of quij > 0, i.e. singularities are sep-
sponding to a trace ti will connect to the first singularity within the
arated by the quantization in the dominant separation direction in π/2-sector which is not separated in the direction orthogonal to ti .
the input parametrization. While this condition is stricter, what we
We propose a simple solution to create layouts in which the separa-
effectively exclude are quantizations with quij = 0 and qvij 6= 0 (Fig-
trices do not deviate from the intended directions more than a user
ure 3 left); thereby we already ensure a maximal separatrix devi-
specified maximum of α by enforcing the additional separation of
ation of π/4. Even more practically relevant, this disjunction-less
offending singularities in orthogonal direction (cf. Figure 3 c).
condition is easier to formulate and more efficient to handle.
Given two intersecting traces ti and t j with li j ≥ l ji originating
Campen et al. propose for their iterative quantization improve-
from singularites i and j, we defined in the previous sections con-
ment to test validity after each change by explicitly searching sin-
straints that ensure the arc set Si j is quantized to at least 1, sepa-
gularity connecting arc paths with quantized length of zero. In the
rating i and j in the direction of ti . If S ji will be quantized to 0, i
worst case this test is in O(n2 ) and in general not easily applicable
and j would lie on the same iso-parameter line and would therefore
in the context of solving an integer linear program. In the following
be connected by a separatrix (unless there is a singularity closer
section we present how separation of singularities can be guaran-
to i which also lies on the same iso-parameter line). If angle |αi j |
teed using only one linear constraint per trace.
of this separatrix is larger than the user defined bound α we need
to prevent such a quantization. By additionally separating j in the
4.2. Singularity Separation direction of t j it is ensured that the corresponding separatrix is not
To guarantee the separation of singularities in the quantization we formed. We therefore include the additional constraint that S ji is
propose a simple criterion based on the following lemma: quantized to at least 1 for every pair of intersecting traces ti and t j
Lemma 1. If for every trace ti one of the arcs in Si∗ is quantized with |αi j | > α.
to at least 1, the quantization is valid. In the light of this, we can now explain the rationale behind
Here, Si∗ is the set of arcs between the start of ti and ni∗ , i.e. the our choice of stopping criterion in the T-mesh construction, con-
first intersection of ti with an earlier trace (cf. Section 3). tinuing a trace ti until at least two traces tk and tl are intersected

c 2021 The Author(s)

Computer Graphics Forum c 2021 The Eurographics Association and John Wiley & Sons Ltd.
 M. Lyon, M. Campen, & L. Kobbelt / Quad Layouts via Constrained T-Mesh Quantization

with αik ∈ [0, α] and αil ∈ [−α, 0] (cf. Section 3). We want to To ensure validity we add the following validity constraints
guarantee that all separatrix deviations are below the user de- which ensure that for each trace ti one of its arcs between its origin
fined bound. With the constraints above, any separatrix created be- and the first intersection ni∗ with a trace starting in the π/2-sector
tween singularities whose traces intersect fulfills this. But what if around ti is quantized to at least 1 (cf. Section 4.2 ):
the quantization implies none of them, i.e. S ji is quantized to at
∑ qk ≥ 1 ∀ traces ti (3)
least 1 for all traces t j that intersect ti , which may be necessary ak ∈Si∗
due to a combination of separation and consistency constraints?
In that case the separatrix will lie in The layout constraints of Section 4.3 have a very similar form
the corridor (blue) between the two but are created for every intersection of two arcs forming a triangle
potential separatrices connecting the with angles larger than α to prevent the creation of separatrices with
start of ti to the starts of tk or tl . Since excessive deviation:
both these potential separatrices sat- l ji
isfy the bound, the actually implied separatrix (blue dotted line) ∑ qk ≥ 1 ∀ ni j with li j > tan α (4)
lying in between satisfies it as well. ak ∈S ji

Finally, to promote layout coarseness, we define the objective to

4.4. Feature Lines & Boundaries be minimized as

Some models contain sharp creases for which it is typically desir- E= ∑ li⊥ · qi → min (5)
able that they are represented by arcs in a layout. Similarly, bound- ai ∈A

aries of models should be represented by arcs. Both these cases, where li⊥ is half the parametric distance between
as well as arbitrary, user defined feature curves are supported by the two arcs opposite of arc ai (or half the para-
our method by aligning the parametrization with these features and metric distance between the one opposite arc and
tracing them with motorcycles to integrate them into our T-mesh ai itself if ai is boundary). Since the quantization
[CBK15]. To ensure that the traced separatrix is not diverted to- of an arc specifies the number of quad strips that
wards another singularity away from the feature curve it is enough pass through this arc orthogonally (in the quad mesh implied by
to simply add layout constraints (cf. Section 4.3) for all intersecting the quantized parametrization), this energy corresponds to the total
traces. This can be interpreted as prescribing a maximum separatrix length of the layout’s quad strips.
deviation of α = 0 for all traces that follow a feature curve. Figure 9
shows examples where boundary alignment is enforced this way. An integer linear program is feasible if an assignment of vari-
ables exists such that all constraints are satisfied. Campen et
al. [CBK15] show that a consistent quantization always exists in
5. Integer Linear Program which all arcs are quantized to at least 1. Such a baseline quantiza-
tion trivially fulfills all our constraints.
We established how a quantization can be constrained such that
no separatrix in the implied quad layout deviates more than a user
given bound from the intended direction. In the space of feasi- 5.2. ILP Size
ble quantizations respecting these constraints, we are looking for The size of the integer linear program described in the previous
a quad layout as coarse as possible. In this section we discuss how section depends largely on the size of the constructed T-mesh. For
such a quantization can be found efficiently by solving an integer every arc there is one integer variable representing its quantized
linear program (ILP). We first describe the basic integer linear pro- length (1), for every patch there are two consistency constraints (2),
gram, which can be constructed in a straightforward fashion follow- every trace adds a validity constraint (3), and for every node created
ing the previous discussion. After that we discuss how the program
size can be reduced for better performance of the solver.

5.1. Definition
Our ILP uses one integer variable qi ∈ Z for every arc ai ∈ A which
represents the quantization of this arc. Every arc ai ∈ A requires a
non-negative quantization:
qi ≥ 0 (1)
Figure 4: On the B UNNY mesh with 192 traces the number of arcs
For consistency (cf. Section 4.1) we add the following constraint quickly grows for decreasing α values to 189k at α = 0.25◦ but the
(analogous to previous work on consistent interval assignment for number of integer variables remains around 252 (left). Since motor-
non-conforming partitions [CBK15, ULP∗ 15]) for each pair of arc cycles need to be traced further and the resulting T-mesh consists
sets S and So which make up two opposite sides of a patch: of more elements for decreasing α, timings go up for both T-mesh
construction and setup of the ILP (to 2.9 s and 0.4 s, respectively, at
∑ qi − ∑ qj = 0 (2)
α = 0.25◦ ) but the time to solve the problem remains around 0.03 s.
ai ∈S a j ∈So

c 2021 The Author(s)

Computer Graphics Forum c 2021 The Eurographics Association and John Wiley & Sons Ltd.
 M. Lyon, M. Campen, & L. Kobbelt / Quad Layouts via Constrained T-Mesh Quantization

at an intersection a layout constraint (4) may be created. With de- In the context of the algorithm of [CBK15] our quantization
creasing angular bound α the traces get longer and the number of could directly be used, as a drop-in replacement, to compute an in-
arcs, patches, and nodes in the T-mesh increases quickly (cf. Fig- teger grid map in which singularities are constrained onto specific
ure 4 left). In the following we describe how the program can be integer locations resulting in a coarse quad mesh with one quad
simplified significantly; interestingly, its size ultimately is propor- per layout patch. Alternatively, a path search on the T-mesh can
tional to the number of traces n, which is a constant independent of be used to explicitly locate for each trace the closest singularity
parameter α. not separated in both coordinates. A separatrix between these two
singularities could then be traced in the seamless input parametri-
First of all, note that the number of T-junctions in the T-mesh is
zation as in [RRP15]§6.1, obtaining an explicit embedding of the
at most n since every trace can create at most one T-junction when
layout arcs.
it ends.
For reasons of flexibility and guaranteed reliability, we opted to
Now, we begin by looking at the number of variables and employ the recent re-parametrization approach of [LCBK19]§6 for
consistency constraints. Consider a simple strip of consecutive our experiments. In a first step, this algorithm integrates the T-mesh
patches bounded by two traces, one on each side. Such a strip into the underlying triangle mesh. Then, T-mesh re-embedding op-
ends at T-junctions leading to wider patches (see inset top) or erations are used to get rid of all T-mesh arcs which are quantized
narrower patches (bottom). Running across the strip to zero through collapsing. We then follow up with an additional
are individual arcs (blue) which need to be quantized step: we iteratively extend all T-junctions to the opposite sides of a
to the same length for consistency (Equation (2)).
By representing these arcs with the same variable,
the consistency constraint is trivially fulfilled. Con-
sistency constraints are then only needed for opposite
patch sides in which at least one contains more than
one arc which is only the case at T-junctions. Thus, the number of
required constraints is at most n. Further, while strips tend to get
longer for smaller values of α (cf. Figure 2) their total number, and
thus also the total number of integer variables, is bounded from
above by 32 n as every strip starts and ends with a T-junction and a
T-junction can be incident to at most three strips.
Finally, we consider the number of validity and layout con-
straints. Section 4.3 describes how the layout constraints are
generated for a given trace when it intersects another one.
These constraints define a set of arcs for which at least one
has to be quantized to at least one. The sets of arcs contain
all arcs between the start singularity and the intersection node.
Similarly, the validity constraints also de-
fine a set of arcs between the start of a
trace and an intersection which needs to be
quantized to at least one (cf. Section 4.2).
Thus, set Si∗ and sets {Si j } created for
trace i can be sorted by size, such that
S0 ⊂ S1 ⊂ . . . ⊂ Sk as illustrated in the in-
set. Obviously, if an arc in S0 is quantized
to one, all supersets S1 ... Sk are quantized
to at least one as well. Thus, a layout con-
straint needs to be created only from the
smallest set per trace. Therefore, the numbers of integer variables,
consistency constraints, validity constraints, and layout constraints
are all O(n).

6. Quad Mesh Extraction

A valid and consistent quantization such as produced by our al-
gorithm specifies which singularities should lie on the same para- Figure 5: Results of our algorithm on a variety of meshes. On the
metric iso-lines and thus defines which singularities should be con- lower right model (B RAIN) a huge number (3.7K) of singularitites
nected by separatrices. There are a couple of ways to make use (i.e. layout nodes) are prescribed as input; while this naturally lim-
of this layout specification. For instance, one can generate a quad its the level of coarseness that can be achieved, it serves to demon-
mesh adhering to this layout – in various ways. strate the reliability and efficiency of our approach.

c 2021 The Author(s)

Computer Graphics Forum c 2021 The Eurographics Association and John Wiley & Sons Ltd.
 M. Lyon, M. Campen, & L. Kobbelt / Quad Layouts via Constrained T-Mesh Quantization

α = 5◦ α = 10◦ α = 15◦ α = 25◦ α = 35◦ α = 45◦

#P = 516 #P = 268 #P = 162 #P = 129 #P = 77 #P = 56

Figure 6: Higher values of α lead to layouts with a smaller number of patches #P.

Model #Faces #Sing α #Traces #Arcs #Vars #P dmean dmax MSJavg tMCG tILP
ROCKERARM 20088 36 15◦ 144 2742 192 159 3.7◦ 14.7◦ 0.989 92 ms 30 ms
H EPTOROID 20000 166 10◦ 832 22822 1131 2051 2.3◦ 9.9◦ 0.982 348 ms 158 ms
S CULPT 7342 14 25◦ 88 936 107 54 5.5◦ 17.7◦ 0.982 29 ms 26 ms
K ITTEN 100000 68 35◦ 272 2072 364 322 8.1◦ 34.9◦ 0.971 152 ms 43 ms
B OTIJO 29994 72 45◦ 320 2116 400 121 7.2◦ 41.4◦ 0.979 87 ms 34 ms
M ASTER C YLINDER 100000 44 15◦ 192 3476 256 214 3.8◦ 14.2◦ 0.988 245 ms 42 ms
B LOCK 68352 48 35◦ 208 1604 291 76 6.6◦ 32.5◦ 0.985 116 ms 32 ms
B RAIN 100000 3721 45◦ 15332 122570 18722 23817 10.7◦ 45.0◦ 0.912 2214 ms 10522 ms
D UCK 19720 28 5◦ 104 5232 148 516 1.4◦ 3.8◦ 0.983 112 ms 37 ms
D UCK 19720 28 35◦ 104 852 138 77 9.8◦ 32.4◦ 0.976 38 ms 28 ms
T EST 1 16323 10 15◦ 38 106 17 28 0.4◦ 6.0◦ 0.996 26 ms 22 ms
S PRAYER 21381 4 5◦ 22 93 10 12 0.1◦ 0.8◦ 0.998 26 ms 2 ms
G LUEGUN 12186 50 25◦ 244 1293 181 209 3.2◦ 24.2◦ 0.969 57 ms 31 ms
C OGNIT 18934 54 25◦ 301 1631 199 181 2.1◦ 20.7◦ 0.988 66 ms 25 ms
C HAIN 5021 60 35◦ 303 1243 200 144 2.2◦ 34.6◦ 0.965 67 ms 27 ms
P UMP 2378 65 45◦ 320 1242 200 141 4.2◦ 40.7◦ 0.955 34 ms 28 ms
E NGINE 16502 24 15◦ 87 316 48 26 2.0◦ 10.1◦ 0.995 43 ms 25 ms
PART 29 10698 12 35◦ 43 105 20 20 0.4◦ 3.1◦ 0.998 11 ms 3 ms

Table 1: Statistical data for our results. From left to right: Model name, number of faces and singularities in the input, angular bound,
number of traces, number of arcs in the motorcycle graph, number of variables in the reduced problem, number of patches in the resulting
layout, mean and maximal separatrix deviation, average minimum scaled Jacobian of quads (of the depicted layout-aligned quad meshes),
time for motorcycle graph construction and for solving of the ILP.

patch, connecting opposing T-junctions if the quantization matches We show results of our method on a variety of models in Figure 5
or splitting the corresponding opposite arc if not. The result is a T- and summarize statistical data in Table 1. Our layouts are typically
mesh with no T-junctions left, explicitly representing all layout arcs well aligned and coarse – with α determining the balance. Both
(integrated into the triangle mesh as edge paths). Into each quad the construction of the T-mesh as well as solving the integer linear
layout patch region we map a regular quad grid (size chosen com- program typically take well below one second. Even on the B RAIN
patibly based on the patches’ parametric extent) by means of an op- model with 3721 singularities the T-mesh construction completes
timized harmonic parametrization as described in [LCBK19]§6.2 in about 2 seconds and the ILP is solved in less than 11 seconds on
for visualization purposes in the following. a commodity PC, showing good scalability of our formulation.

Our algorithm is controlled by one parameter α which defines

7. Results
the maximum acceptable separatrix deviation from the seamless
For the results in this section we created the input seamless pa- parametrization. As demonstrated in Figures 6 and 7 our layouts
rametrization by optimizing the energy proposed by Bommes et contain fewer patches at the cost of higher separatrix deviation for
al. [BZK09] which minimizes the difference between the parame- increasing values of α. Table 1 shows that the maximum separa-
trization gradient to a cross field. The cross fields were obtained trix deviation (dmax ) generally stays below the given bound α, and
using the method of Bommes et al. [BZK09] with directional con- the average deviation (dmean ) typically is significantly lower. Note
straints as proposed by [CIE∗ 16] (Sections 7 and 7.1), or using the that our formulation guarantees that the quad layout connectivity
method of Viertel et al. [VOS19] (Section 7.2). For models B OTIJO defined by the resulting quantization can always be embedded in
and E LK, as well as those in Section 7.3, we used the frame field the surface with deviation strictly bounded by α (e.g. by using the
provided in the supplemental material of [PPM∗ 16]. We use Gurobi re-tracing strategy of [RRP15], cf. Section 6); when further opti-
to solve the ILP and QEx [EBCK13] to extract the quad mesh from mizing or smoothing the embedding (e.g., following [LCBK19] or
the final parametrization. [CK14b]), this bound may be weakened of course.

c 2021 The Author(s)

Computer Graphics Forum c 2021 The Eurographics Association and John Wiley & Sons Ltd.
 M. Lyon, M. Campen, & L. Kobbelt / Quad Layouts via Constrained T-Mesh Quantization

Model Method `/α #P dmean dmax MSJavg

1% 3929 1.2◦ 33.3◦ 0.991
CBK15 5% 268 6.2◦ 71.6◦ 0.970
20% 138 9.7◦ 82.9◦ 0.862
E LK
5◦ 412 1.1◦ 4.8◦ 0.983
ours 10◦ 206 2.2◦ 8.7◦ 0.986
15◦ 134 3.2◦ 14.2◦ 0.982
1% 2109 1.3◦ 11.8◦ 0.994
CBK15 5% 161 5.9◦ 58.4◦ 0.980
20% 103 9.2◦ 62.8◦ 0.880
F ERTILITY
5◦ 337 1.3◦ 4.3◦ 0.985
ours 20◦ 161 3.3◦ 17.7◦ 0.984
30◦ 105 5.0◦ 29.3◦ 0.985
1% 7033 0.8◦ 8.3◦ 0.991
CBK15 5% 333 7.5◦ 83.3◦ 0.971
20% 146 10.1◦ 81.8◦ 0.891
B UNNY
5◦ 1088 1.5◦ 5.0◦ 0.977
ours 20◦ 344 4.3◦ 18.9◦ 0.971
35◦ 148 7.5◦ 34.2◦ 0.958

Table 2: Comparison with [CBK15]. In the third column we give

the target edge length ` in % of the bounding box diagonal for
[CBK15] and α for ours. #P, dmean , dmax , and MSJavg are num-
ber of layout patches, mean and max separatrix deviation, and the
average minimal scaled Jacobian of the quads, respectively.

Figure 8: Plotting number of layout patches #P against maximum

(left) and mean (right) separatrix deviation. Our method (orange)
consistently creates coarser layouts with less separatrix deviation
than QGP (blue).

Figure 7: Layout comparison. For QGP the target edge length ` is For small target edge lengths leading to larger quantized values
given in % of the bounding box diagonal. on the arcs QGP achieves results with uniformly sized quads but
typically very dense layouts (Figure 7 left). With larger target edge
lengths the base complex naturally becomes coarser as the quad
7.1. Comparison with Campen et al. mesh itself contains fewer quads (before subdivision). However, the
separatrix deviation quickly increases and is often close to π/2, es-
We compare our results to those of the Quantized Global Parame-
pecially if the mesh contains regions with denser layout vertex dis-
trization (QGP) algorithm of Campen et al. [CBK15] who also cre-
tribution where the distance between pairs of layout vertices may
ate quad meshes via a quantization of a T-mesh. However, their
be below the target edge length. Table 2 shows that QGP achieves
quantization tries to minimize the difference between the arc length
lower element quality, as indicated by the average minimum scaled
(rather than angle) in the input parametrization and the quantization
Jacobian, than our method except for the densest layout where the
which makes control over the resulting layout difficult. We gener-
higher element quality comes at the cost of an order of magnitude
ated quad meshes with their algorithm for target edge lengths ` of
more layout patches.
1%, 5%, and 20% of the bounding box diagonal. The two coarser
quad meshes are subdivided (via uniform scaling of the resulting Low layout complexity and small separatrix deviation are com-
parametrization) to achieve a similar number of quads as the result peting goals. In Figure 8 we plot the number of layout patches
for 1% for better visual comparison. We applied our algorithm with against the maximum and mean separatrix deviation for different
α between 5◦ and 45◦ in steps of 5◦ and show results with the best values of α and ` averaged over the models from Table 2. The plot
matching number of layout patches. In Table 2 we report the num- as well as the numbers in Table 2 show that our algorithm consis-
ber of layout patches, as well as average and maximum separatrix tently achieves a lower separatrix deviation with layouts of similar
deviation. or smaller complexity.

c 2021 The Author(s)

Computer Graphics Forum c 2021 The Eurographics Association and John Wiley & Sons Ltd.
 M. Lyon, M. Campen, & L. Kobbelt / Quad Layouts via Constrained T-Mesh Quantization

25°
20°
15°
10°
5°
0°
Figure 9: Our results on meshes with boundaries [VOS19]. Figure 10: Comparison with [PPM∗ 16]. The color coding shows
the angular deviation between input cross field and resulting para-
metrization, d is the area weighted average over the whole mesh.
7.2. Comparison with Viertel et al.
Our algorithm shares some similarity with the one presented by
Viertel et al. [VOS19] who create quad layouts (potentially non- ILP. Our algorithm is controlled by only a single, intuitive param-
conforming, containing T-junctions) by tracing a motorcycle graph eter defining the maximum separatrix deviation from the seamless
and iteratively collapsing quad strips. In our algorithm assigning a input parametrization. This could be especially useful in contexts
quantization of 0 to an arc is equivalent to collapsing it [LCBK19]. where a certain minimum quality needs to be guaranteed while
Our ILP formulation has the advantage of finding the global opti- coarseness of the layout is also desirable. Our experiments show
mum of a linear objective, Equation (5), and thus not suffering from that the presented method not only achieves good results but is also
running into local optima that a greedy optimization could reach computationally efficient, requiring only a couple of seconds for
and which Viertel et al. report may lead to T-junctions remaining in meshes with thousands of singularities.
the layout which could have been avoided with a different collapse
Our algorithm relies on a seamless parametrization given as in-
order. In addition, a valid and consistent quantization as produced
put to define layout singularities and desired separatrix alignment.
by our algorithm always defines a conforming quad layout free of
As we base all our measurements on that seamless parametriza-
T-junctions. We present the result of our method on the models used
tion, our algorithm is oblivious to any metric distortion potentially
in [VOS19] in Figure 9.
already present in that parametrization. In cases where the parame-
trization is computed from a cross field – which may then be seen
7.3. Comparision with Pietroni et al. as the actual, original alignment intent – it is certainly tempting to
omit the intermediate parametrization step. While a T-mesh simi-
In Figure 10 we compare our results with those of [PPM∗ 16]. Since
lar to ours could directly be traced based on a cross field, as done
Pietroni et al. create their layouts by searching separatrix candi-
in [MPZ14] and [PPM∗ 16], a key challenge is dealing with mat-
dates directly in a cross field, rather than a parametrization as in our
ters of non-integrability and non-parametrizable singularity config-
case, we present the average deviation of the two directions induced
urations in this setting. In any case, strict angle bounds will not
by the resulting patchwise parametrization to the direction of the
be possible in a cross field based approach; arbitrarily large devia-
input cross field for a fair comparison. Note that, while we guaran-
tion may be necessary to obtain any valid layout in theory – unless
tee a maximal deviation to our (field-guided) input parametrization,
T-junctions are acceptable. In that case rectangularity constraints
the deviation compared to the guiding cross field may deviate more
could be implemented in a soft manner. The quantized T-mesh will
than α. The results show that, when aiming for a similar number of
then contain non-rectangular patches which could be resolved by
patches, we achieve lower distortion. For complex models Pietroni
inserting singularities [MPZ14] or by leaving T-junctions in the
et al. generate layouts containing T-junctions. These can be con-
layout [PPM∗ 16, VOS19].
verted to a conforming layout by extending all T-junctions. Doing
this for the 78 T-junctions of the L ION model, however, leads to In scenarios where strict adherence to the given input configura-
a layout with over 82k patches. By contrast, our method yields a tion is not required many interesting opportunities open up. For in-
conforming layout with only 2.1k patches. stance it would be interesting to explore variations of our approach
that allow merging or splitting of singularities whenever beneficial
for layout quality. It would also be interesting if the layout algo-
8. Conclusion & Future Work
rithm could make use of the fact that singularities may be free to
We presented an algorithm for the creation of coarse quad layouts move – at least within a certain range. In that case, small move-
based on assigning integers to the arcs of a T-mesh by solving an ments of layout vertices could move separatrices into the acceptable

c 2021 The Author(s)

Computer Graphics Forum c 2021 The Eurographics Association and John Wiley & Sons Ltd.
 M. Lyon, M. Campen, & L. Kobbelt / Quad Layouts via Constrained T-Mesh Quantization

bound, opening new opportunities to form a potentially coarser or [EGKT08] E PPSTEIN D., G OODRICH M. T., K IM E., TAMSTORF R.:
better aligned layout. Motorcycle Graphs: Canonical Quad Mesh Partitioning. Computer
Graphics Forum 27, 5 (2008), 1477–1486. 3
Ultimately, for cases where the singularity configuration is en- [EH96] E CK M., H OPPE H.: Automatic reconstruction of B-spline sur-
tirely flexible, a holistic approach to layout creation would be de- faces of arbitrary topological type. In Proc. SIGGRAPH 96 (1996),
sirable that, instead of separating the creation of layout vertices and pp. 325–334. 2
the creation of their connecting separatrices in independent steps, [ESCK16] E BKE H.-C., S CHMIDT P., C AMPEN M., KOBBELT L.: Inter-
creates both of them together in one step. This perspective is taken actively controlled quad remeshing of high resolution 3d models. ACM
Trans. Graph. 35, 6 (2016), 218:1–218:13. 1
in [CK14a], albeit in a non-automatic approach, requiring user ex-
pertise and interaction. [HSJ∗ 20] H IEMSTRA R. R., S HEPHERD K. M., J OHNSON M. J., Q UAN
L., H UGHES T. J.: Towards untrimmed NURBS: CAD embedded
reparameterization of trimmed B-rep geometry using frame-field guided
global parameterization. Computer Methods in Applied Mechanics and
Acknowledgements Engineering 369 (2020). 1
The authors thank Alexandra Heuschling for help with the imple- [KMZ11] KOVACS D., M YLES A., Z ORIN D.: Anisotropic quadrangu-
mentation and Jan Möbius for creating and maintaining the ge- lation. Computer Aided Geometric Design 28, 8 (2011), 449–462. 1
ometry processing framework OpenFlipper [MK12]. Models were [KNP07] K ÄLBERER F., N IESER M., P OLTHIER K.: Quadcover – sur-
face parameterization using branched coverings. Computer Graphics Fo-
provided by [MPZ14, PPM∗ 16, VOS19]. This work was supported
rum 26, 3 (2007). 1
by the Gottfried-Wilhelm-Leibniz Programme of the Deutsche
[LCBK19] LYON M., C AMPEN M., B OMMES D., KOBBELT L.: Para-
Forschungsgemeinschaft (DFG) and in part funded by the Deutsche metrization quantization with free boundaries for trimmed quad mesh-
Forschungsgemeinschaft (DFG) - 427469366. ing. ACM Trans. Graph. 38, 4 (2019). 3, 6, 7, 9
[MAB∗ 19] M ARINOV M., A MAGLIANI M., BARBACK T., F LOWER J.,
BARLEY S., F URUTA S., C HARROT P., H ENLEY I., S ANTHANAM N.,
References F INNIGAN G. T., M ESHKAT S., H ALLET J., S APUN M., W OLSKI P.:
[BCE∗ 13] B OMMES D., C AMPEN M., E BKE H.-C., A LLIEZ P., Generative design conversion to editable and watertight boundary repre-
KOBBELT L.: Integer-grid maps for reliable quad meshing. ACM Trans- sentation. Computer-Aided Design 115 (2019), 194 – 205. 1
actions on Graphics 32, 4 (2013), 98:1–98:12. 1, 3 [MC19] M ANDAD M., C AMPEN M.: Exact constraint satisfaction for
truly seamless parametrization. Computer Graphics Forum 38, 2 (2019),
[BLK11] B OMMES D., L EMPFER T., KOBBELT L.: Global structure
135–145. 3
optimization of quadrilateral meshes. Computer Graphics Forum 30, 2
(2011), 375–384. 2 [MK12] M ÖBIUS J., KOBBELT L.: OpenFlipper: An open source ge-
ometry processing and rendering framework. In Curves and Surfaces,
[BLP∗ 13] B OMMES D., L ÉVY B., P IETRONI N., P UPPO E., S ILVA C., vol. 6920 of Lecture Notes in Computer Science. 2012. 10
TARINI M., Z ORIN D.: Quad-mesh generation and processing: A survey.
Computer Graphics Forum 32, 6 (2013), 51–76. 1 [MPZ14] M YLES A., P IETRONI N., Z ORIN D.: Robust field-aligned
global parametrization. ACM Transactions on Graphics 33, 4 (2014). 9,
[BMRJ04] B OIER -M ARTIN I. M., RUSHMEIER H. E., J IN J.: Parame- 10
terization of triangle meshes over quadrilateral domains. In Proc. SGP
’04 (2004), pp. 197–208. 2 [PPM∗ 16] P IETRONI N., P UPPO E., M ARCIAS G., S COPIGNO R.,
C IGNONI P.: Tracing field-coherent quad layouts. Computer Graphics
[BZK09] B OMMES D., Z IMMER H., KOBBELT L.: Mixed-integer quad- Forum 35, 7 (2016). 2, 7, 9, 10
rangulation. ACM Transactions on Graphics 28, 3 (2009), 77:1–77:10.
[PPTSH14] PANOZZO D., P UPPO E., TARINI M., S ORKINE -H ORNUNG
1, 7
O.: Frame fields: Anisotropic and non-orthogonal cross fields. ACM
[Cam17] C AMPEN M.: Partitioning surfaces into quadrilateral patches: Transactions on Graphics 33, 4 (2014), 134. 1
A survey. Computer Graphics Forum 36, 8 (2017), 567–588. 1 [RP17] R AZAFINDRAZAKA F. H., P OLTHIER K.: Optimal base com-
[CBK12] C AMPEN M., B OMMES D., KOBBELT L.: Dual Loops Mesh- plexes for quadrilateral meshes. Computer Aided Geometric Design 100,
ing: Quality Quad Layouts on Manifolds. ACM Transactions on Graph- 52-53 (2017), 63–74. 2
ics 31, 4 (2012), 110:1–110:11. 3 [RRP15] R AZAFINDRAZAKA F. H., R EITEBUCH U., P OLTHIER K.:
[CBK15] C AMPEN M., B OMMES D., KOBBELT L.: Quantized global Perfect matching quad layouts for manifold meshes. Computer Graphics
parametrization. ACM Transactions on Graphics 34, 6 (2015). 3, 4, 5, 6, Forum 34, 5 (2015), 219–228. 2, 6, 7
8 [TPP∗ 11] TARINI M., P UPPO E., PANOZZO D., P IETRONI N.,
[CIE∗ 16] C AMPEN M., I BING M., E BKE H.-C., Z ORIN D., KOBBELT C IGNONI P.: Simple quad domains for field aligned mesh parametri-
L.: Scale-Invariant Directional Alignment of Surface Parametrizations. zation. ACM Transactions on Graphics 30, 6 (2011). 1, 2
Computer Graphics Forum 35, 5 (2016). 7 [ULP∗ 15] U SAI F., L IVESU M., P UPPO E., TARINI M., S CATENI R.:
Extraction of the quad layout of a triangle mesh guided by its curve
[CK14a] C AMPEN M., KOBBELT L.: Dual strip weaving: Interactive de- skeleton. ACM Transactions on Graphics 35, 1 (2015), 6. 5
sign of quad layouts using elastica strips. ACM Transactions on Graphics
33, 6 (2014), 183:1–183:10. 10 [VCD∗ 16] VAXMAN A., C AMPEN M., D IAMANTI O., PANOZZO D.,
B OMMES D., H ILDEBRANDT K., B EN -C HEN M.: Directional field syn-
[CK14b] C AMPEN M., KOBBELT L.: Quad layout embedding via thesis, design, and processing. Computer Graphics Forum 35, 2 (2016).
aligned parameterization. Computer Graphics Forum 33, 8 (2014), 69– 2
81. 7
[VOS19] V IERTEL R., O STING B., S TATEN M.: Coarse quad layouts
[CSZZ19] C AMPEN M., S HEN H., Z HOU J., Z ORIN D.: Seamless para- through robust simplification of cross field separatrix partitions. Proc.
metrization with arbitrary cones for arbitrary genus. ACM Trans. Graph. 28th International Meshing Roundtable (2019). 2, 7, 9, 10
39, 1 (2019). 2
[ZZY16] Z HANG S., Z HANG H., YONG J.-H.: Automatic quad patch
[EBCK13] E BKE H.-C., B OMMES D., C AMPEN M., KOBBELT L.: layout extraction for quadrilateral meshes. Computer-Aided Design and
QEx: Robust quad mesh extraction. 168:1–168:10. 7 Applications 13, 3 (2016), 409–416. 2

c 2021 The Author(s)

Computer Graphics Forum c 2021 The Eurographics Association and John Wiley & Sons Ltd.