How2Sketch: Generating Easy-To-Follow Tutorials for Sketching 3D Objects

James W. Hennessey1 Han Liu2 Holger Winnemöller3 Mira Dontcheva3 Niloy J. Mitra1

1 2 3
University College London KAUST Adobe Research

step 6/28 step 7/28 step 9/28

arXiv:1607.07980v1 [cs.GR] 27 Jul 2016

step 15/28 step 24/28 step 28/28

(a) easy-to-follow tutorial steps (only a few shown) (b) user sketch

Figure 1: (a) We present How2Sketch, a system that automatically generates easy-to-follow tutorials for drawing 3D models. Each generated
tutorial comes as a list of steps for drawing scaffolding primitives that help the user draw the object in correct perspective. To help the user
draw the scaffolding, the tutorial shows how to construct guidelines that anchor objects parts relative to one another. User feedback on the
tutorials was positive as many created what they felt were more accurate drawings (b) User sketch from alternative viewpoint.

Abstract than what you know [Edwards 1999]. A simple cylinder, for exam-
ple, is known to have a circular cross-section with equal widths at
Accurately drawing 3D objects is difficult for untrained individuals, the top and bottom. However, when we actually see a cylinder, it
as it requires an understanding of perspective and its effects on ge- is subject to perspective distortion: circles become ellipses while
ometry and proportions. Step-by-step tutorials break the complex projected radii diminish with distance from the viewer.
task of sketching an entire object down into easy-to-follow steps
Art books and tutorials provide step-by-step instructions to teach
that even a novice can follow. However, creating such tutorials re-
sketching [Tizon 2007; Eissen and Steur 2011]. To help with pro-
quires expert knowledge and is a time-consuming task. As a result,
portions and relative positioning, they introduce geometric con-
the availability of tutorials for a given object or viewpoint is lim-
structions, such as scaffolding primitives, and follow a coarse-to-
ited. How2Sketch addresses this problem by automatically gener-
fine approach, using prior steps as basis for subsequent ones.
ating easy-to-follow tutorials for arbitrary 3D objects. Given a seg-
mented 3D model and a camera viewpoint, it computes a sequence Authoring such tutorials requires significant expertise and time
of steps for constructing a drawing scaffold comprised of geometric commitment even for trained artists. Consequently, objects and
primitives, which helps the user draw the final contours in correct viewpoints in existing tutorials tend to be limited and are chosen
perspective and proportion. To make the drawing scaffold easy to by the expert, rather than the users of the tutorials. To address these
construct, the algorithm solves for an ordering among the scaffold- issues, we present an approach for automatically generating easy-
ing primitives and explicitly makes small geometric modifications to-follow tutorials for drawing 3D models from user specified view-
to the size and location of the object parts to simplify relative po- points. Figure 1 shows parts of a tutorial generated by our system
sitioning. Technically, we formulate this scaffold construction as a and the drawing by one of our study participants based on that tuto-
single selection problem that simultaneously solves for the ordering rial from an alternative viewpoint. Our algorithm targets man-made
and geometric changes of the primitives. We demonstrate our algo- objects where part relations and proportions tend to be crucial for
rithm for generating tutorials on a variety of man-made objects and accurate depiction.
evaluate how easily the tutorials can be followed with a user study.
Inspired by instructional books and online tutorials, we take ex-
plicit steps to make a sketching tutorial easy-to-follow: (i) focus
1 Introduction on accurate inter-part proportions and relations via a drawing scaf-
fold, followed by detailing of the object contour; (ii) proceed in a
The ability to draw real-world objects is a useful and important skill coarse-to-fine fashion, where object parts are abstracted as primi-
across many disciplines. Product designers draw daily as they gen- tives (e.g., cuboids, cylinders) over several levels of detail to build
erate and refine product ideas, fine artists may spend hours in fig- up said scaffold; (iii) propose a particular drawing order among the
ure drawing classes learning how to replicate a shape from the real scaffolding primitives such that those sketched later can be easily
world, while hobbyists use sketches for visual expression. Still, anchored (i.e., drawn with guidance) off already drawn primitives;
sketching requires skill and practice. One of the major challenges in and (iv) provide explicit steps for the construction of guidelines to
drawing real-world objects is learning to draw what you see rather accurately anchor the scaffolding primitives.
One of our key observations is that in easy-to-follow tutorials the construct guides for accurate depiction of perspective and propor-
dimensions and arrangements of object-parts tend to have ratios that tions. Stroke correction or beautification is orthogonal to our main
are easy to construct. For example, it is easier to construct the cen- contribution and may be used to complement the contour drawing
ter line of a rectangular face compared to its one fifth line. Tutorial phase of our tutorials. Other assisted sketching systems take as in-
authors choose to construct with such ‘easy ratios’ to simplify the put 2D sketches and interpret them as 3D curve networks [Xu et al.
drawing process and to focus on the procedure, rather than inciden- 2014]. More advanced methods [Shao et al. 2012; Iarussi et al.
tal and arbitrary measurements (see Figure 2). To apply this tech- 2015; Pan et al. 2015] use 2D input to infer 3D geometry or surface
nique to existing objects, our system How2Sketch proposes small normals for complex shading. We focus on the automatic genera-
geometric changes while keeping overall deviations from the source tion of sketching tutorials, rather than automatic inference based on
model minimal. Since in each step new primitives and guidelines the sketched curves.
are anchored with respect to those drawn in previous steps, the or-
dering of steps significantly affects the simplicity of ratios that can Tutorials. A good tutorial greatly facilitates understanding. Many
be employed, and the geometric approximations that this incurs. attempts have been made to automatically generate high-quality tu-
This tight interdependence between ordering of primitives and their torials for different applications. A digital drawing tutorial system
geometric changes makes the problem non-trivial. A further chal- was proposed by Fernquist et al. [2011] that allows an expert to
lenge is to preserve the original inter-part relationships of objects, create tutorials for novices. Tutorial generation systems [Takagi
even under geometric perturbations. For example, in Figure 1 the et al. 2003; Cummmings et al. 2012] for specific sketching tasks
coaxial relationship between the mixer bowl and mixer blade is pre- have also been proposed, for example drawing a single scene with
served. pre-defined objects, or ‘eyes.’ Grabler et al. [2009] developed a tu-
torial system for photo manipulation tasks. How2Sketch produces
Technically, we map the geometric adjustment and ordering of parts sketching tutorials for man-made objects automatically, rather than
to a single selection problem. We first generate a set of potential relying on an experienced artists to create them.
candidate primitives by enumerating different anchoring possibili-
Drawing expertise. Tchalenko [2007] found that novices and pro-
ties. Since anchoring requires drawing guidelines, and some guide-
fessional artists have comparable accuracy when perform basic line
lines are easier to construct than others, the algorithm prefers an-
drawing tasks (straight lines and simple curves). However, in a
choring possibilities that rely on easy-to-construct guidelines, such
follow-up study [Tchalenko 2009], he showed that when copy-
as the top edge, bottom edge, center line, etc., of existing primi-
ing complex artworks, novices made significantly more errors than
tives. Our key insight is that the problem of geometric adjustment
artists. The main difference in drawing strategy was that experts
and ordering of parts can be simultaneously solved by selecting an
divided complex lines into easy-to-draw short segments. Schmidt
appropriate subset from the candidate primitives, in order to bal-
et al. [2009a] found that experts made qualitatively similar errors
ance between geometric changes and ease of constructing necessary
to non-artists, indicating that perspective drawing is hard, even for
guidelines.
trained users. Particularly for off-axis viewing angles, drawing er-
We test our algorithm on a range of examples and evaluate our al- ror increased significantly. In an observational study, Grimm [2011]
gorithmically generated, easy-to-follow tutorials with a user study. found that artists commonly used a coarse-to-fine strategy starting
with blocking shapes and finishing by drawing detailed items at the
end. How2Sketch assists the user by breaking the drawing process
2 Related Work up into basic steps that are easy to execute and by explicitly indi-
cating vanishing line directions.
Assisted drawing. Various applications have been proposed to as-
sist a user in sketching. Some correct the user input based on ge- Line drawings. Many methods for generating stylized artistic ren-
ometric analysis of the users input strokes [Igarashi and Hughes derings of objects have been proposed (see [Kyprianidis et al. 2013]
2001; Bae et al. 2008; Schmidt et al. 2009b], others rely on an for a survey). We leverage stylization to visually distinguish the
underlying image to guide the user [Iarussi et al. 2013; Xie et al. various line types of our tutorials (perspective lines, guides, con-
2014; Benedetti et al. 2014; Fernando et al. 2014], and yet others tours, etc). Other researchers investigated which features artists
on crowdsourced data (e.g., many sketches) to improve the users typically draw to convey 3D shape [DeCarlo et al. 2003; DeCarlo
drawing [Dixon et al. 2010; Lee et al. 2011; Gingold et al. 2012; et al. 2004; Burns et al. 2005; DeCarlo and Rusinkiewicz 2007]. We
Limpaecher et al. 2013; Simo-Serra et al. 2016]. These methods display suggestive contours computed on the optimized object to
aim to improve the quality of the users’ strokes at a local level. aid the user in adding final details to their drawing. Fu et al. [2011]
Our focus is on suggesting a meaningful drawing order and easy-to- and Liu et al. [2014] infer plausible contour ordering from 2D and
3D inputs, respectively. While the derived sequences are plausi-
ble, they are not tailored for tutorials and do not provide specific
guidelines to make them easy to follow.

3 Learning How to Sketch

To inform the design of How2Sketch we studied several drawing
books [Edwards 1999; Eissen and Steur 2007; Eissen and Steur
2011], visited a number of sketching websites (e.g., Sketch-a-Day),
carried out an expert interview with a professional artist, and took
a drawing course.
Through this process we found that effective tutorials for drawing
3D objects typically include the following:
Figure 2: A step-by-step sketching tutorial for drawing a • Parts are approximated by geometric primitives: Plane, cubes
car ©Czajkowski. The task is made simpler by breaking it into steps and cylinders are heavily used to approximate shapes. They
and by providing guidance about part proportions and alignments. are easy to construct and verify visually.
 segmented
input model

generate create select create

primitives & relations primitive candidates valid + desirable cand. tutorial view point

Figure 4: System Overview. Starting from an input part-level segmented model and a user-specified viewpoint, How2Sketch generates
easy-to-follow step-by-step tutorials. The system automatically makes subtle geometric modifications to simplify the resultant tutorial.

• Steps are coarse-to-fine: First, the overall object is scaffolded where ABCD is extended by reflection to form BCLK such that
with approximate shapes, and then, finer contour details are AB = AK. Finally, we also support alignment, as in Figure 3e,
added. Primitives are drawn sequentially, in a particular order. M 0 N 0 is aligned with M N .
• Anchor shapes to each other: Shapes are drawn with respect (d) Perspective. To provide perspective information, we show the
to previously drawn shapes, to aid with correct placement and vanishing points (if within the drawing area) and also show the
proportions. Instructions for positing shapes relative to each vanishing lines leading to them (Figure 3f). How2Sketch supports
other use simple measurements (e.g., draw box half way down sketching in 2-point and 3-point perspective.
the side, draw circle in the center of the rectangle), etc.
• Vanishing lines for perspective: Vanishing points are explic- 4 Generating Sketch Sequences
itly indicated to aid the user to draw correctly.
Given a 3D object (S) segmented into parts and a desired viewpoint,
How2Sketch supports the above tutorial features as follows: our goal is to establish an easy-to-follow sequence for drawing the
object, starting with the scaffolding and progressing to the contour
(a) Scaffolding Primitives. How2Sketch utilizes scaffolding prim- details. We make it easier to draw the scaffold by actively making
itives to geometrically approximate each segmented object part. small part-level geometric changes to facilitate relative anchoring
The system supports planes, cuboids, cylinders, and truncated pyra- using a set of guidelines.
mids, as they allow for planar guidelines to be used, which are sim-
ple to construct, and cover a wide range of shapes. In addition to As described in Section 3, we have adopted simple procedures to
scaffolding, we guide users in drawing ellipses to better approxi- accurately draw guidelines at easy-to-construct ratios (1⁄2, 1⁄3, 1⁄4, 1×,
mate some shapes. 2×, etc). Object part placements and sizes in the original mod-
els, however, rarely conform to such ratios. Hence, we propose to
(b) Ordering. Our algorithm provides the relative ordering of the modify object parts, so that they end up with part relationships that
scaffolding primitives. Further, How2Sketch offers detailed, se- are easy to draw. We motivate this choice twofold: (i) Scaffolding
quenced instructions for constructing primitives. primitives in tutorials like those generated by How2Sketch are al-
ready approximations of real geometry and thus contain a measure
(c) Placement, Alignment, and Proportions. We support a set of error. Some of this error can actually be compensated by adjust-
of coplanar guidelines (see Figure 3). Given a face ABCD, its ing the fit of contours within the scaffold. (ii) Accurate estimation
diagonals help construct the 1⁄2 line EF (Figure 3a). Two levels of lengths and ratios is difficult, even for experts, so errors are al-
of 1⁄2 lines produce a 1⁄4 line GH (Figure 3b); while intersecting a most unavoidable. By enforcing that parts relate via simple ratios
diagonal BD with line CE produces a 1⁄3 line IJ (Figure 3c). Simi- for which reasonable geometric constructs can guide the user, the
larly, we support extrusion towards a vanishing point as in Figure 3d overall drawing error is unlikely to increase significantly beyond
A
E
A
E A E I
K
A the unguided case.
G
B B B B
Our algorithm proceeds in three main stages (see Figure 4): (i) gen-
erating part-level primitives and encoding inter-primitive relations;
L (ii) creating primitive candidates based on various inter-primitive
D D D D anchorings strategies; and (iii) selecting a valid and desirable set of
F
F
C
F
H C J C C
primitives among the candidate selections. The result implicitly en-
(a) 1/2 guide (b) 1/4 guide (c) 1/3 guide (d) extend guide

Table 1: Notation table.

N’
symbol denotes
M’ S input part-segmented model
N Pi primitive corresponding to the i-th part of S
M Ri,j relation between primitive pairs (Pi , Pj )
(e) alignment guide (f) perspective guide k
Cj→i candidate for the i-th part primitive with (anchoring) parent from the
j-th part primitive, where k denotes the k-th such instance
Figure 3: Our system supports different forms of guidelines for C∗i set of all the candidate primitives generated for part primitive Pi
drawing coplanar proportions (a-d), for anchoring alignments (e), χ(X) indicator variable corresponding to the selection of X
and for previewing 2-point perspectives (f). See Section 3. Λ assignment of indicator variables denoting a set of selected candidates
 cuboid (a) 1 4 (b)
truncated
pyramid 3

common
cylinder bisector plane
2
plane

(c) (d)
(a) (b)

Figure 5: Given a part-segmented input model S (top-left inset),

we abstract the parts as different primitives (a) and identify inter-
part relations. For example, here the mixer bowl and mixer head
primitives share a common bisector plane.

codes how to geometrically modify each part (both their dimension Figure 6: Starting from initial primitives P1 , P2 , P3 , P4 , for each
and placement), and in which order to draw them. Intuitively, our pair of primitives we generate several candidate primitives. For
algorithm produces an easy-to-follow primitive drawing sequence example in (b), we show the primitives generated for P3 using P4
k
at the cost of deviating from the original geometry in a controlled as parent, indicated as different instances P4→3 shown in different
fashion. We now elaborate each step. Please refer to Table 1 for dotted/solid brown. Relations are restored leading to further new
symbols used in the following. primitives, for example in (c) the green primitive was lifted to re-
store coplanarity with the brown primitive. Finally, we also have
second level primitives. As shown in (d), the new brown primitive
4.1 Generating Primitives and Inter-part Relations leads to a new primitive for P1 . This is an illustrative figure in 2D
with only some of candidate primitives shown for simplicity.
We abstract model parts by primitive shapes. In our implementation
we support planes, cuboids, cylinders, and truncated pyramids (see form Ri,j , we add additional primitives to their candidate sets to
Figure 5a). (Note that in our visualization we show axis-aligned restore the relations. Specifically, corresponding to a candidate of
bounding box for cylinders as the box faces are used for providing k
the form Cj→i (created in stage (i)), we create a new candidate of
guidance for drawing ellipses.) For each part of the input model S, k0 k k0
the form Ci→j such that Cj→i ↔ Ci→j are similarly related as in
we use least-squares to fit (axis-aligned) different primitive types
Pi ↔ Pj . We append all such relation-based additional candidate
and take the one with the least residue. In case of ties, we prefer k0
the simpler primitive. We denote the primitive for the i-th part as primitives to the respective candidate sets, i.e., C∗i ← C∗i ∪ Ci→j
Pi (type of primitive is not explicitly indicated in this notation). (see Figure 6c).

Man-made objects, which are our target objects, often have domi- Note that in the above a candidate is allowed to be anchored from
nant inter-part relations. We found it highly desirable to preserve one or multiple parents, as each axis can be independently an-
such relations in the generated tutorials. Hence, we first detect such chored. Additionally a candidate can be partially unguided (e.g.,
inter-part relations and later preserve them in the generated tutori- the width and length of cuboid is guided but the height is not) or
als. We simply test (see [Mehra et al. 2009]) each pair of primitives completely unguided (e.g., it is simply the input primitive) (see Fig-
Pi and Pj for any relations. In our implementation, we consider ure 6). We defer further details to the implementation section.
coplanar, coaxial, and common bisector plane relations. In case of (iii) We allow second-level anchors, i.e., candidate primitives as
multiple relations between a pair of primitives, we prefer common generated above are allowed to act as anchors for other primitives.
bisector plane over coaxial over coplanar. We represent a relation To this end, we simply iterate one more time stage (i) and (ii) (see
using a binary variable Ri,j where i and j respectively denote the Figure 6d). Note that before starting this step, we remove the can-
primitives Pi and Pj (type of relation is not explicitly indicated didate primitives with large changes in geometry or relative place-
in this notation). If a relation is present, we mark Ri,j = 1, and ments (more details in the implementation section).
Ri,j = 0 otherwise. Figure 5 shows some examples.
At the end of this stage, we have a set of candidates for each part of
4.2 Creating Candidate Primitives the input model resulting in the super set of candidate primitives of
the form {C∗i } (see Figure 7).
We now describe the candidate primitive generation step that cre-
ates additional primitives based on possible anchoring strategies. 4.3 Selecting Candidate Primitives
We use C∗i to denote the set of all the candidate primitives gen-
erated corresponding to primitive part Pi . Since the the original Having generated multiple candidates, our remaining task is to se-
primitive is always a candidate, we start by C∗i := {Pi }. We gen- lect a set valid and desirable candidates, as explained next.
erate candidate primitives in three stages:
Valid candidate sets. We first characterize the notion of valid se-
(i) For each pair of primitives Pi and Pj , we generate candidates of lections. We use indicator variables χ(X) to denote if a candidate
k primitive X is selected (i.e., χ(X) = 1) or not (i.e., χ(X) = 0).
the form Cj→i , where j → i indicates that a candidate is generated
k k
for primitive part Pi and is anchored off Pj with k denoting dif- We have χ(Cj→i ) ∈ {0, 1} for each Cj→i ∈ C∗i . Let Λ denote
ferent anchoring possibilities. For example, parts can be anchored a particular assignment for the indicator variables for all the candi-
based on different guidelines described in Section 3 for different date primitives.
face- or plane-based anchors. We append these candidates to the
respective candidate sets as: C∗i ← C∗i ∪ {Cj→i 1 2
, Cj→i , . . . } (see Among the various possible selections, not all the subsets of candi-
Figure 6b). dates of the form Λ constitute valid selections. A valid selection of
candidates should satisfy three conditions:
(ii) For each pair of primitives (Pi , Pj ) sharing a relation of the (1) for each part of S, only one candidate primitive should be se-
 Selecting any primitive, however, incurs an associated error that
k
we indicate as Ed (Cj→i ) due to deviation from original geometry
(see later describe how we measure Ed ). So, the total data cost of
selecting a set of primitives is:
X k k
Eadjust (Λ) := χ(Cj→i )Ed (Cj→i ) (5)
i,j,k

with a higher cost indicating larger geometric deviations from the

original parts.
Figure 7: From a set of candidate primitives (left), our algorithm
selects a subset of primitives that is valid and desirable as shown Thus, we arrive at the final formulation for desirable selection as,
on the right. The selection implicitly encodes in which order to
draw the primitives and also how to change each primitive (size min(Eadjust (Λ) + Edifficulty (Λ)) (6)
Λ
and/or placement) such that the resulting tutorial is easy to con-
struct. Please refer to the text for details. subject to Equations (1)-(3) to ensure a valid selection. Thus, we
have formulated our problem as a quadratically constrained linear
lected; program.
(2) if a selected candidate primitive is anchored off one or more
parent (candidate) primitives, then its parent primitive(s) must also Error functions. The above formulation requires metrics for Ee
be selected; and Ed . While various metrics may be substituted, we used the
(3) if any two primitives Pi and Pj share a relation, then their corre- following in our implementation.
sponding selected candidate primitives should also respect the same k
For the difficulty of drawing term Ee (Cj→i ), we associate a higher
relation. (Note that by (1) implies that each primitive should have a cost for anchors that are harder to replicate (e.g., requiring more
unique selected candidate associated with it.) construction lines). Specifically, we set the cost to the number of
We now express the above conditions in terms of the indicator vari- guidelines divided by the area of the parent plane where construc-
ables in Λ. tion lines are to be drawn. This encourages fewer guides but also
(a) We encode (1) as using planes/faces with larger areas for drawing sketch guides. (The
X effect of viewpoint is only considered at runtime as discussed in
k Section 5).
χ(Cj→i )=1 ∀i. (1)
j,k k
For the data error Ed (Cj→i ), we sum the changes in length along
(b) We encode (2) as a quadratic constraint involving the binary each axis, normalized by the original axis length, with the transla-
selection variables as tion of the midpoint of each axis, again normalized by the input axis
length. For an unguided axis we set the data error to the maximum
k k of 2 to discourage unguided candidates.
χ(Cj→i )χ(Cj ) − χ(Cj→i )≥0 (2)
k Final drawing order. The solution to the above optimization di-
for each dependent pair Cj→i ∈ C∗i and its parent Cj . Note that
k rectly gives us both the ordering and the modifications of the parts.
this condition disallows χ(Cj→i ) = 1 AND χ(Cj ) = 0, but al- Note that the above solution may only return a partial ordering
k
lows any of the other three assignments involving χ(Cj→i ) and among the candidates primitives. This implies that the relative
χ(Cj ). drawing order among of certain primitives are not specified. We
(c) We now encode (3). Let two primitives Pi and Pj share a rela- break such ties only at runtime once the user selects a view as de-
1 2
tion, i.e., Ri,j = 1. Let C∗i = {C∗i , C∗i , . . . } be all the generated scribed in Section 5.
1 2
candidates for primitive Pi and similarly C∗j = {C∗j , C∗j ,...}
k
for primitive Pj . Then for each pair of the form C∗i ∈ C∗i and A note about greedy alternative: An alternative algorithm is to
k0 greedily consider primitive pairs and snap them relative to nearest
C∗j ∈ C∗j that does not share the same relation as Ri,j , we require guidelines. Such an adhoc strategy fails to take care of the sec-
k k 0 ond level candidates, and produces an inferior solution. More im-
χ(C∗i )χ(C∗j ) = 0. (3) portantly, this approach does not provide any meaningful ordering
This condition disallows k
χ(C∗i )
= 1 AND k0
χ(C∗j )
= 1, i.e., they based on ease of drawing, which is a key focus of our problem.
cannot be jointly selected as these candidate primitives do not share
the same relations as of their respective primitive parts. 4.4 Implementation details
Thus, a selection Λ is valid if Equations 1-3 are all satisfied. Among We now clarify some additional implementation details. Most of
all such valid selection sets, we next determine which one is most these are choices we made in our implementation, and can be re-
desirable. Figure 7 shows a set of candidate primitives and a valid placed by multiple comparable alternatives.
selection.
Primitives that are not well approximated by one of the currently
Sequencing sketching as a selection problem. We balance the er- supported ones can be represented as a custom primitive (e.g., line)
ror due to making changes to the geometry with difficulty of draw- but such primitives cannot be part of our optimization step. Instead
ing arising from anchoring. In other words, an unanchored part is after our optimisation step their relative positioning is updated and
more difficult to draw compared to an anchored part – this is a key when sketched they are unguided. Alternatively, the user can ap-
hypothesis of this work. We indicate this difficulty of drawing cost proximate them by their bounding box primitive.
k
as Ee (Cj→i ) with a lower cost denoting easier to draw (see later
for details). The total cost is expressed as: The candidate primitive generation works in two steps: first, we use
X the coplanar relations to generate candidate planes cki→j , and then
k k depending on the primitive type we combine the planes to create
Edifficulty (Λ) := χ(Cj→i )Ee (Cj→i ). (4) k
i,j,k a complete primitive Ci→j (here, lowercase c for candidate plane
rather than complete primitive that uses uppercase C). This choice
unifies candidate primitive generation across primitive types (recall
cylinders are processed based on their axis-aligned bounding box).
For each pairwise coplanar relation Ri,j we have two participating
planes in Pi and Pj : at this stage the relation is undirected and Step 1 Step 2 Step 3
we produce candidate planes using both combinations ckj→i and
cki→j . To generate a candidate plane, each axis is considered in- Figure 9: Guide lifetime. Guides first appear in orange (left). In
dependently then all combinations are combined to create planes subsequent steps guides that are no longer required are removed,
cki→j . An axis can be anchored by the parent plane using the end while those that are to be reused are marked in blue (middle, right).
points of the same axis. This means there are several anchoring
possibilities. For example, anchoring the vertical axis of Pi on Pj tive falls below this threshold, we ask the user to simply ‘eyeball’
might involve anchoring the top edge of Pi to the 1⁄3 line of Pj and the primitive without drawing intermediate guides. (iii) Finally, a
the bottom edge Pi to the bottom edge of Pj . An alternative might segment that is occluded and its primitive does not help anchor any
be to anchor the top edge of Pi to the 1⁄3 line of Pj and the bottom other visible primitive is deemed unnecessary and hence is left out
edge Pi to 1⁄4 line of Pj . We initially generate all candidates but from the tutorial.
to reduce the number of candidates to select from we discard those
where an axis length or translation change of more than 10% of the User ability. We adapt our tutorials to different sketching abilities
input length. by classifying the various guidelines as suitable for novice, appren-
tice, or master users. For example, dividing a face of a primitive
With all the candidate planes generated using all the pairwise rela- into halves requires three guidelines. A novice is shown all the
tions we generate complete primitives by combining the different three, an apprentice only the 1⁄2 line itself, and a master is not pro-
planes based on the primitive type. To generate a complete cuboid vided with any intermediate guidance. Note that in all cases, the
primitive, for example, we find the missing height axis from one of user is instructed to divide the highlighted face into half by a text
the other planes to complete the primitive. For truncated pyramids label in the viewer (see Figure 8).
we combine top and bottom planes with a height axis to make a
truncated pyramid. To generate second level candidates, we repeat Guide lifetime. In order to reduce the amount of guidelines on
this process but use the first level candidates as the parent primi- a sketch at any point in time, we determine each guide’s lifetime
tives. to inform the users when a guide can be safely erased. To this
end, we first go over the list of generated guidelines to identify the
We use the Gurobi linear Solver [Gurobi Optimization 2015] to equivalent ones, and store their lifetime, i.e., when they first appear
solve the quadratically constrained LP as described above. Typi- and when they are last used. During the tutorial, a guideline is
cally the solver takes 1-2 minutes in the presented examples. drawn in orange when it first appears. If the guideline is used in
any later step, it is changes to blue. After the last step a guide
is used, it is no longer shown. As a result, users do not have to
5 Presenting Sketch Sequences unnecessarily erase/redraw guides, and thus reduce clutter as they
sketch (see Figure 9).
The sequence generated in Section 4 provides primitive ordering,
sketching guidelines, and adjusted part geometry for drawing the Vanishing points and ellipses. Vanishing lines and vanishing
scaffolding of the object. How2Sketch tutorials can be adapted fur- points are indicated with respect to the paper boundary (shown
ther based on the user chosen viewpoint and user indicated drawing as green corners) to help users better position the lines. We ad-
level (novice/apprentice/master), which can be controlled interac- ditionally guide users in sketching ellipses on a primitive face by
tively. Our custom viewer indicates when guidelines can be erased using guides to the vanishing points. These guides intersect with
and provides hints for drawing in perspective and object contours. the edges of the face at the perspective mid-points, which are the
points where the ellipse should touch the face of the primitive.
Viewpoint. We use the specified viewpoint to customize the tutorial
as follows: (i) Although primitive ordering is determined based on Contour ordering. Once the user has sketched the scaffolding and
anchoring strategies, multiple primitives can anchor from the same ellipses, we guide them to sketch the contours. We use suggestive
parent, resulting in a tie. We break such ties in ordering by first contours [DeCarlo et al. 2003] computed on the modified underly-
choosing the primitive that is closest to the user from the indicated ing model segments (the contours are computed at a segment level).
viewing position. (ii) The selected viewpoint can make some guide- The contour segments are progressively displayed per segment, fol-
lines cumbersome to draw because of limited space on the projected lowing the order determined by the primitives. Already drawn parts
area of a primitive face. We identify such instances by threshold- are used to determine occlusion for the new primitives, thus reduc-
ing based on Ap /k, where Ap indicates the projected area and k ing clutter (see Figure 1).
the number of guides necessary to draw the primitive. If a primi-
Interface. How2Sketch tutorials can be presented in many different
Ext
endby1/
2 forms. They can be navigated manually using an our interface; they
can be printed (see supplementary material), or sequenced into a
video.

6 Results and Discussion

Novi
ce Appr
ent
ice Mast
er

We used How2Sketch to generate sketching tutorials for four man-

Figure 8: User ability. Users specify drawing level made objects - a Digital SLR Camera, Kitchen Mixer, Train and
(novice/apprentice/master) which determine the number of inter- Paint Roller. For these models, numerous tutorials depending on
mediary guides presented for each step. For the ‘extend by 1/2’ viewpoint and user ability can be generated. Parts of the tutori-
step, novices (left) are shown 9 guidelines, apprentices (center) 6 als are shown in Figure 10 (see supplementary material for full se-
guidelines, and masters (right) 3 guidelines. quences). Each tutorial takes between 15 and 45 minutes to com-
 step 4/24 step 2/14 step 5/27 step 2/34

step 8/24 step 4/14 step 7/27 step 5/34

step 12/24 step 5/14 step 9/27 step 6/34

step 15/24 step 6/14 step 11/27 step 9/34

step 17/24 step 7/14 step 19/27 step 14/34

step 21/24 step 10/14 step 23/27 step 23/34

(a) camera (b) roller (c) mixer (d) train

Figure 10: Example step-by-step tutorials generated by our system: (a) and (b) were generated in the master-user setting, while (c) and (d)
were generated in the novice-user setting. Please refer to the supplementary materials for complete examples.
 the length of the first carriage and the top edge of the wheels using
the 1⁄4 guide on the vertical axis of the first carriage. The driver’s
compartment is unguided.
Limitations. How2Sketch only makes small changes to the in-
put geometry. However, small gaps between object parts can have
important semantic meaning. An example of this can be seen in
Figure 11 where the main body of the mixer and the stand sepa-
rate slightly in the adjusted version. We know these two segments
would be joined by a hinge making such an adjustment unrealis-
tic. Symmetry or regular structure can similarly be lost from the
small geometry changes. An example of this is the roller in Figure
11, which ceases to be a perfect cylinder. Note that most of these
violations are difficult to spot unaided and tend to get masked by
drawing inaccuracies. Finally we find relations from the input seg-
ments but do not allow adjustments in geometry to create a relation
that was not already present. In the future, we might enable such
changes to allow for an even wider range of candidates.

7 Evaluation
To evaluate the effectiveness of the How2Sketch tutorials, we com-
pare to a simple step-by-step tutorial that shows scaffolding prim-
itives for each part of the object but does not simplify the sizes or
locations of the primitives to make them easier to draw. In this Ba-
sic tutorial type, the scaffolding primitives are shown in order from
largest to smallest with a base primitive anchored to the ground
plane. No guidelines are shown. Please see the supplemental mate-
rials for the complete tutorials used in the study.

(a) original models (b) modified models Preliminary Study. We conducted a preliminary study with 8 par-
ticipants comparing an earlier version of How2Sketch with the ba-
sic tutorial type. User responses to questions about satisfaction
Figure 11: (Left) Original models. (Right) Subtle changes pro-
with their drawing, perceived accuracy of their drawing, and experi-
posed by our algorithm in order to make the objects easier to draw.
ence with the tutorial were significantly higher for the How2Sketch
plete due to their varying complexity. The small changes made tutorial compared to the Basic tutorial. Additionally an ANOVA
to the input geometry by the method are illustrated in Figure 11. across tutorial type and object drawn revealed a significant effect
As desired, the alterations to geometry are subtle but now enable of tutorial type on satisfaction, accuracy, and experience ratings (p
simple anchoring strategies based on the altered segment bounding < 0.022 in all cases). Despite this positive feedback we did not
boxes (also shown). observe an improvement in drawing quality across tutorial type.
Based on the preliminary study observations and user feedback, the
As demonstrated in Figure 10, our tutorials follow a coarse-to-fine current version of How2Sketch introduces a wider range of primi-
strategy, starting with a single primitive that can be used to anchor tives, indicates guide lifetime, adapts tutorials based on user ability,
subsequent primitives. Figure 10a shows excerpts from a tutorial and uses relations for candidate generation.
sequence with master user ability. Here, the grip is anchored on
the edges of the camera body and a 1⁄4 guide. Additionally, the grip Participants. We recruited 10 participants (ages 18-55+, 6 men)
and flash are both extended by one half the depth of the main body. with varied expertise in drawing. Two participants reported never
The lens, an example of a second level anchoring, uses the flash for drawing. Four reported drawing once in a while. Six reported draw-
anchoring by extruding 1×. Guides for ellipses are provided before ing at least once a month. Three had taken college-level art classes
contours are drawn. or private/non-accredited art classes. When asked (free-form) what
they found most challenging about drawing, 4 mentioned perspec-
In the paint roller tutorial in Figure 10b the handle anchors the roller tive, proportions, scale, and relative positions. When asked to rate
using the common bisector plane. The top edge of the roller is 1× their drawing skills on a scale of 1 (poor) to 5 (great), only 4 people
the length of the handle. The bottom edge is 1⁄2× the length of the rated their drawing skills above 2.
handle but due to the limited projected area and number of guides
otherwise required, the step is unguided (as per Section 5). Methodology. In advance, each participant filled out an introduc-
tory online questionnaire about their experience with drawing real
Figures 1 and 10c both show novice ability tutorials for the food objects. Upon arrival, each participant was told that they will be
mixer but from different viewpoints. The plane primitive for the asked to draw two objects, a camera and a mixer, using two differ-
base of the mixer anchors the bowl using a planar relation and 1⁄2 ent tutorials. Participants always followed a How2Sketch tutorial
guide. The common bisector plane between the base and the main first to disadvantage How2Sketch to any learning effect. The two
body of the mixer is used for anchoring the length of the main body. objects (camera and mixer) counter-balanced with half of the par-
The bisector plane is first drawn before being extended in both di- ticipants using the Basic tutorial type for the camera and half using
rections to create the cuboid primitive. The Mixer’s stand is an the How2Sketch tutorial for the mixer. The study had 4 conditions
example of a primitive with two parents, being anchored off both (2 objects x 2 tutorial types). The How2Sketch tutorial was set to
the main body and base. the novice ability for all participants.
The train example, Figure 10d, anchors the second carriage as 1× Before the How2Sketch tutorial, participants were given a written
Figure 12: Average user ratings for satisfaction, perceived ac-
curacy, enjoyment, and ease of drawing were all higher for the
How2Sketch tutorials than for the basic tutorials. Showing stan-
dard error of mean (SEM) bars for N = 10.

handout (see supplemental material) that described how to draw

construction lines for 1⁄2, 1⁄4, and 1⁄3 guidelines and extending planes
(see Figure 3). This written tutorial was designed to give them con-
text for what they would encounter in the How2Sketch condition.
Both the Basic and How2Sketch tutorials were followed using a
Macbook Pro 13” laptop; participants used the trackpad to advance
forward and backward through the tutorial. All drawings were done Figure 13: User Study Sketches: User sketches all overlaid on the
on paper. Each participant was given two pencils (HB, 0.3mm and object they drew sketches from following basic tutorial (left) show
0.7mm). They were allowed to use a provided straight-edge and much greater variation in proportions and alignment than sketches
eraser. For creating each drawing, the participants were given a from following How2Sketch tutorials (right).
sheet of paper that included the vanishing points and the ground
plane of the first primitive. This initial calibration allowed us to sketch using the blocks (scaffold) for guidance”. Users also had
more easily compare drawings across users. All users drew the suggestions for improvements to the system with several asking for
scaffolding primitives first on the calibrated paper. For drawing further guidance with perspective.
the final contours of the object, the moderator attached a transpar-
Sketch quality. Figure 13 overlays the registered user sketches
ent sheet to the paper with the scaffolding. This allowed us to more
from the different conditions on the original model for the condition
easily compare both the contour drawings and the scaffolding prim-
(e.g How2Sketch model after part level adjustments). While there
itives across users.
appears to be variation in contour placement in both tutorial types
Participant filled out a questionnaire after drawing each object, in- the variation in the basic tutorial sketches is greater.
dicating their level of satisfaction with their drawing (1 - not at all,
In the camera tutorials the basic version starts with the ground plane
5 - very much), perceived accuracy of their drawing (1 - not at all
for the lens and How2Sketch with the ground plane of the main
accurate, 5 - very accurate), enjoyment with the tutorial experience
body. With this anchoring in the basic tutorial sketches, the width
(1 - not at all, 5 - very much), and how ease the tutorial steps were
and length of the lens are accurate. However, the lens height and the
to follow (1 - not at all easy, 5 - very easy). They also gave free-
other three primitives have a variety of errors in proportion and part
form responses about what they liked about each tutorial type and
placement. Comparing with How2Sketch sketches, there are simi-
how it could be improved. At the end of the study, subjects were
lar variations in the height of the main body. However, the guided
asked which tutorial type they preferred (Basic or How2Sketch). In
steps for the grip and lens shows decent convergence in position-
the study itself we referred to the Basic tutorial type as the tuto-
ing across users. For the Mixer sketches - where both tutorials start
rial without guides and the How2Sketch tutorial type as the tutorial
with the base plane of the mixer - there is much more consistency
with guides.
of object part placements across users.
All participants were given a $25 gift card for their time.
To further validate these findings, we conducted an additional user
Feedback. Nine out of ten participants preferred the How2Sketch study using Amazon Mechanical Turk (AMT). In the study, we pre-
tutorial over the Basic tutorial. User responses to questions about sented users with two sketches of the same object type overlaid and
satisfaction with their drawing, perceived accuracy of their drawing, registered to their condition specific model (see supplementary ma-
enjoyment of the drawing process, and ease of following tutorials MIXER object CAMERA object
were all higher for How2Sketch than the Basic tutorial (Fig. 12). basic basic
number of pairwise comparisons
number of pairwise comparisons

How2Sketch How2Sketch
An ANOVA across tutorial type and object drawn reveals a strong
significant effect of tutorial type on accuracy and ease of following
tutorial (p<0.003), significant effect on enjoyment (p<0.034), and
marginally significant effect on satisfaction (p<0.058). The object
drawn did not have an effect on any measure, despite their varying
difficulty, and there was no interaction between tutorial type and
object drawn.
probability of being judged more accurate probability of being judged more accurate

Freeform feedback echoed the ratings. Participants enjoyed using

the guides as it gave them more confidence in their accuracy. One Figure 14: Bradley-Terry Model for the Mixer and the Camera
participant said “It was satisfying drawing the guides and getting Sketches produced by users of our tutorials and evaluated by an-
the proportions right. It was then so much easier to draw the final other user study with Amazon Mechanical Turk rankers.
terial). The two pairs of object could be from the same or different D E C ARLO , D., AND RUSINKIEWICZ , S. 2007. Highlight lines
tutorial types. We asked participants to “Please select the sketch for conveying shape. In International Symposium on Non-
that is more accurate to the underlying model. Do consider the pro- Photorealistic Animation and Rendering (NPAR).
portions, alignment and perspective. Please ignore style or shad-
ing”. The studies for the two objects were run independently, so D E C ARLO , D., F INKELSTEIN , A., RUSINKIEWICZ , S., AND
did not have the same responders. Each participant evaluated the S ANTELLA , A. 2003. Suggestive Contours for Conveying
45 unique image pairs for one of the objects. Each study had 225 Shape. ACM Transactions on Graphics (Proc. SIGGRAPH) 22,
sets of responses. AMT users we compensated $0.01 per compari- 3 (July), 848–855.
son. D E C ARLO , D., F INKELSTEIN , A., AND RUSINKIEWICZ , S. 2004.
Interactive rendering of suggestive contours with temporal co-
To evaluate the results, we performed pairwise comparisons using herence. In NPAR, 15–24.
the Bradley-Terry model. We plot probability histograms for both
objects in Figure 14. The bottom axis is the probability of being D IXON , D., P RASAD , M., AND H AMMOND , T. 2010. iCanDraw:
judged more accurate in a pairwise comparison. The left axis is the Using Sketch Recognition and Corrective Feedback to Assist a
number of pairwise comparisons that has this probability. Visually User in Drawing Human Faces. 897–906.
you can see the trend that in a pairwise comparison sketches from
following How2Sketch tutorials are more likely to be judged more E DWARDS , B. 1999. The New Drawing on the Right Side of the
accurate. This evidence is clear across both object types. Brain. Jeremy P. Tarcher/Putnam.
E ISSEN , K., AND S TEUR , R. 2007. Sketching: Drawing Tech-
niques for Product Designers. Bis B.V., Uitgeverij(BIS Publish-
8 Conclusion ers).

We presented How2Sketch, a system that automatically generates E ISSEN , K., AND S TEUR , R. 2011. Sketching: The Basics. BIS.
easy-to-follow tutorials for drawing man-made 3D objects from se- F ERNANDO , P., P EIRIS , R. L., AND NANAYAKKARA , S. 2014.
lected views. We evaluated our system using a user study, and I-draw: Towards a freehand drawing assistant. OzCHI ’14, 208–
found that sketches made by following How2Sketch tutorials had 211.
more accurate proportions and relative part placements compared
to a basic step-by-step tutorial with scaffolding primitives. Addi- F ERNQUIST, J., G ROSSMAN , T., AND F ITZMAURICE , G. 2011.
tionally, users preferred the How2Sketch tutorials over the basic tu- Sketch-sketch revolution: An engaging tutorial system for
torial, giving significantly higher ratings for satisfaction, accuracy, guided sketching and application learning. In Proceedings of the
and enjoyment. 24th Annual ACM Symposium on User Interface Software and
Technology, ACM, UIST ’11, 373–382.
One possible future direction is to provide stroke level support to
help users draw the final object contours, possibly by explicitly pro- F U , H., Z HJOU , S., L IU , L., AND M ITRA , N. 2011. Animated
viding guidelines with respect to the scaffold primitives. Another construction of line drawings.
direction would to to explore new types of guidelines that can help G INGOLD , Y., VOUGA , E., G RINSPUN , E., AND H IRSH , H. 2012.
reduce the number of unguided steps in a tutorial. Diamonds from the rough: Improving drawing, painting, and
singing via crowdsourcing. In Proceedings of the AAAI Work-
A very interesting future question is to investigate if How2Sketch
shop on Human Computation (HCOMP).
really teaches users to sketch better by drawing “what you know.”
While this is the ultimate goal of any sketching tutorial, answering G RABLER , F., AGRAWALA , M., L I , W., D ONTCHEVA , M., AND
this question will require a much more involved user study where I GARASHI , T. 2009. Generating photo manipulation tutorials by
we have to track and quantify user-specific improvements, if any, demonstration. In ACM SIGGRAPH 2009 Papers, ACM, SIG-
as observed over a substantial timeframe. GRAPH ’09, 66:1–66:9.
G RIMM , C. 2011. Results of an observational study on sketching
References results of an observational study on sketching.
G UROBI O PTIMIZATION , I., 2015. Gurobi optimizer reference
BAE , S.-H., BALAKRISHNAN , R., AND S INGH , K. 2008. ILoveS- manual.
ketch: As-natural-as-possible Sketching System for Creating 3D
Curve Models. In Proceedings of UIST, UIST ’08, 151–160. I ARUSSI , E., B OUSSEAU , A., AND T SANDILAS , T. 2013. The
drawing assistant: Automated drawing guidance and feedback
B ENEDETTI , L., W INNEM ÖLLER , H., C ORSINI , M., AND from photographs. In ACM Symposium on User Interface Soft-
S COPIGNO , R. 2014. Painting with Bob: Assisted Creativity ware and Technology (UIST), ACM.
for Novices. In Proceedings of the 27th Annual ACM Sympo-
sium on User Interface Software and Technology, ACM, UIST I ARUSSI , E., B OMMES , D., AND B OUSSEAU , A. 2015. Bend-
’14, 419–428. fields: Regularized curvature fields from rough concept sketches.
ACM Transactions on Graphics.
B URNS , M., K LAWE , J., RUSINKIEWICZ , S., F INKELSTEIN , I GARASHI , T., AND H UGHES , J. F. 2001. A Suggestive Interface
A., AND D E C ARLO , D. 2005. Line drawings from volume for 3D Drawing. In Proceedings of the 14th Annual ACM Sym-
data. ACM Transactions on Graphics (Proc. SIGGRAPH) 24, 3 posium on User Interface Software and Technology, ACM, UIST
(Aug.), 512–518. ’01, 173–181.
C UMMMINGS , D., V IDES , F., AND H AMMOND , T. 2012. I don’t K YPRIANIDIS , J. E., C OLLOMOSSE , J., WANG , T., AND I SEN -
believe my eyes!: Geometric sketch recognition for a computer BERG , T. 2013. State of the ‘art’: A taxonomy of artistic styl-
art tutorial. In Proceedings of the International Symposium on ization techniques for images and video. IEEE Transactions on
Sketch-Based Interfaces and Modeling, SBIM ’12, 97–106. Visualization and Computer Graphics 19, 5, 866–885.
L EE , Y. J., Z ITNICK , C. L., AND C OHEN , M. F. 2011. Shad-
owdraw: Real-time user guidance for freehand drawing. ACM
SIGGRAPH, 27:1–27:10.
L IMPAECHER , A., F ELTMAN , N., T REUILLE , A., AND C OHEN ,
M. 2013. Real-time drawing assistance through crowdsourcing.
ACM SIGGRAPH 32, 4 (July), 54:1–54:8.
L IU , J., F U , H., AND TAI , C.-L. 2014. Dynamic sketching: Sim-
ulating the process of observational drawing. In Proceedings of
the Workshop on Computational Aesthetics, CAe ’14, 15–22.
M EHRA , R., Z HOU , Q., L ONG , J., S HEFFER , A., G OOCH , A.,
AND M ITRA , N. J. 2009. Abstraction of man-made shapes.
ACM SIGGRAPH Asia 28, 5, #137, 1–10.
PAN , H., L IU , Y., S HEFFER , A., V INING , N., L I , C., AND
WANG , W. 2015. Flow aligned surfacing of curve networks.
ACM Trans. Graph. (SIGGRAPH) 34, 4.
S CHMIDT, R., K HAN , A., K URTENBACH , G., AND S INGH , K.
2009. On Expert Performance in 3D Curve-drawing Tasks.
In Proceedings of the 6th Eurographics Symposium on Sketch-
Based Interfaces and Modeling, ACM, New York, NY, USA,
SBIM ’09, 133–140.
S CHMIDT, R., K HAN , A., S INGH , K., AND K URTENBACH , G.
2009. Analytic Drawing of 3D Scaffolds. In ACM SIGGRAPH
Asia 2009 Papers, ACM, SIGGRAPH Asia ’09, 149:1–149:10.
S HAO , C., B OUSSEAU , A., S HEFFER , A., AND S INGH , K.
2012. Crossshade: Shading concept sketches using cross-section
curves. ACM Transactions on Graphics (Proceedings of ACM
SIGGRAPH 2012) 31, 4.
S IMO -S ERRA , E., I IZUKA , S., S ASAKI , K., AND I SHIKAWA , H.
2016. Learning to Simplify: Fully Convolutional Networks for
Rough Sketch Cleanup.
TAKAGI , S., M ATSUDA , N., S OGA , M., TAKI , H., S HIMA , T.,
AND YOSHIMOTO , F. 2003. A learning support system for be-
ginners in pencil drawing. In Proceedings of the 1st international
conference on Computer graphics and interactive techniques in
Australasia and South East Asia, ACM, 281–282.
T CHALENKO , J. 2007. Eye movements in drawing simple lines.
PERCEPTION 36, 8, 1152.
T CHALENKO , J. 2009. Segmentation and accuracy in copying and
drawing: Experts and beginners. Vision Research 49, 8, 791 –
800.
T IZON , N. 2007. Art of Sketching. Sterling Publishing Company,
Inc.
X IE , J., H ERTZMANN , A., L I , W., AND W INNEM ÖLLER , H.
2014. PortraitSketch: Face Sketching Assistance for Novices.
In Proceedings of the 27th Annual ACM Symposium on User In-
terface Software and Technology, ACM.
X U , B., C HANG , W., S HEFFER , A., B OUSSEAU , A., M C C RAE ,
J., AND S INGH , K. 2014. True2form: 3d curve networks from
2d sketches via selective regularization. Transactions on Graph-
ics (Proc. SIGGRAPH 2014) 33, 4.