Drawing Architecture using Manga Techniques

Y. Qu
Department of Computer Science and Engineering, The Chinese University of Hong Kong, Hong Kong
M.A. Schnabel
Department of Architecture, The Chinese University of Hong Kong, Hong Kong

ABSTRACT: Manga is a type of stylised bitonal drawings pleasing in terms of screen variety and tidy lines.
In this paper, we present a method to draw architecture, spatial compositions and their details, including col-
ours, textures, and tones, using manga drawing techniques. Given a colour image, such as photography, or
drawing, our method finds the optimal solution for screening during the bi-tonal transformation, as well as a
set of ideally abstracted line drawings. This offers a novel rendering appearance for the communication of ar-
chitectural design. In our results we demonstrate successfully that our method generates manga-like drawings
from an architectural image that preserve architectural key-elements, such as surface or material properties.

1 INTRODUCTION cording to texture, material property, chromaticity or

even function of the underlying architecture that is
Digital techniques introduced in the past decades being depicted and expressed.
have led a cultural shift in architectural design and
its communication; new art forms and techniques in-
fluence the expression of architects’ thoughts and
ideas. However, continuing emphasis remains that
hand-drawn sketches allow for a deeper understand-
ing of design, which computer generated visualisa-
tions do not offer. Inevitably media play influential
roles in the expression and communication of de-
sign. Subsequently the act of sketching allows for its
own experience and interpretation of spatial ar-
rangements (Goldschmidt 1991). Yet digitally aided
sketching merges both realms into a novel expres-
sion of architectural communication, which we pre-
sent in this paper.
Manga, a popular Japanese comic art form and
medium of entertainment, is increasingly becoming
known around the world. The graphical quality is
unique in its elegant use of rich set of screens, and
tidy and fine drawing styles. This novel way of vis-
ual communication of manga provides another realm
for representing architectural drawings or photogra-
phy, being real or virtual. Although the manga draw-
ings depends on pure black and white (b/w) patterns
and lines, or is bi-tonal represented, it is still possi-
ble to perceive the variety of textures or materials of
the surfaces of buildings. During the screening proc-
ess manga artists lay multifarious screens to express
different semantics. Here the ‘screening’ refers to
Figure 1. Generation of manga background from a real photo-
the process of laying pre-printed b/w patterns over a graph (a). Commercial software (b) adopts halftone technique
certain area of the image. Screens are selected not and present uniform screening result. Our results (c) & (d) pre-
solely according to the shading, or tone, but also ac- serves colour distinguishability, texture similarity, and tone.
 Such a screening process is a very labour inten- ing, communicating and construing. It incorporates
sive undertaking. Once an architectural design is abstraction, fuzziness as well as clarity and sharp-
rich in details and elaborated in a high level, the ness. As that the expressions of lines and hatches
process of creating a manga becomes even more themselves become the objects of meaning and in-
complex. Speeding up the production some artists terpretation. A sketch explores from overall relation-
already employ computer techniques to convert im- ships to fine resolution of detailed attributes of the
ages to manga backgrounds, using techniques such depicted design. In order to make the communica-
as ‘half-toning’ or ‘hatching’ (Figure 1a & 1b). tion of space more precise the drawing makes use of
However, these digital methods produce only mo- graphic modifiers to convey information that go be-
notonous patterns and unsatisfactory results that are yond the pure outlines or dimensions of the spatial
not in line with the rich expression of manga graph- arrangement (Robbins 1994).
ics and depicted architecture.
In this paper, we introduce a novel application
2.2 Bitonal Image Generation
that provides the user with an automatic screening
tool, which is useful in architectural representation. A straightforward way to produce computational bi-
Given an arbitrary colour image of an architectural tonal images from greyscale/colour photographs is
arrangement, our novel computational method can halftoning (Figure 1c). It exploits the spatial integra-
instantly transfer the input into a bi-tonal multi- tion of human vision to approximate the intensity
screening manga representation. over a small local region with only b/w pixels
Our technique can deal with an arbitrarily com- (Knuth 1987, Floyd & Steinberg 1974, Ulichney
plex image or colour palette. Rich sets of pat- 1987, Jarvis et al. 1976). As artefact patterns may
terns/screens are employed to represent the original appear, some techniques (Velho & Gomes 1991,
image with the goal of preserving three key-factors Naiman & Lam 1996) try to reduce the artefact pat-
of manga graphical representation: ‘chromaticity terns by adjusting the scanning path over the image.
distinguishability’, ‘texture similarity’, and ‘tone To preserve the edges during halftoning, Knuth
similarity’. While the tone is preserved by matching (1987) enhances edges in a pre-processing step.
the density of tone with the one of each pattern, our Velho & Gomes (1995) and Buchanan & Verevka
core contributions of our application are the preser- (1995) approximate edge regions by rearranging the
vation of chromaticity distinguishability and texture clusters around edges. Recently, Pang et al. (2008)
similarity. We develop a novel colour-to-pattern preserve the fine structure by optimizing the struc-
mapping based on the Multi-Dimensional Scaling ture similarity.
(MDS) technique. MDS enables the relative distance Hatching is another technique to produce bitonal
to preserve the mapping from a ’24 dimensional tex- images (Figure 1b). Winkenbach et al. (1994) gener-
ture feature space’ (of the screen patterns) to the ated pen-and-ink illustrations by rendering a geo-
‘two-dimensional chromaticity space’. As shown in metric scene with prioritized stroke textures. A user-
Figure 1d, our results not only preserve the original specified “detail indication” can be used to highlight
tone, but also differentiate the chromaticity of the complex textures and minimize clutter. Salisbury et
original image. Our system allows designers to in- al. (1997) use direction fields to guide the orienta-
teractively control and fine-tune their results. tion of strokes. Durand et al. (2001) suggest a
thresholding model of strokes that can express a rich
set of stroke styles to assist artists in achieving vari-
2 RELATED WORK ous styles. Grabli et al. (2004) improve the clutter
control of lines by estimating the line density. This
2.1 Architectural Sketch
predicts the visual complexity before rendering and
There is much written about manual or digital archi- allows simplification.
tectural sketches and the depiction of spatial ar- Images generated by halftoning and hatching,
rangements using a two-dimensional medium. Since containing more or less uniform patterns, may ap-
the 16th century the tradition of drawing as an act of pear monotonous to the reader; traditional manga in-
designing freed the architect from working on site at stead uses a wide variety of screens to enrich the
a scale of 1:1 and subsequently design became also viewing experience. Qu et al. (2008) proposed a
an intellectual discipline engaged in the plane of pa- novel richness-preserving manga screening method
per. Goldschmidt (1991) argues that the architectural utilizes the variety of pattern to preserve the chroma-
sketch is a mode of visual thinking and communica- ticity distinguishability as well as the tone and tex-
tion, which is crucial to a conceptual framework of ture similarities. Based on their work on pure
the depicted architectural design. Despite the ad- screening part, we propose a complete system to
vancement of digital media, there remains a certain generate manga style architectural drawings includ-
quality of a hand drawing and the architectural ing screening, line drawings, and the final rendering.
sketch remains the predominant medium of design-
 Figure 2. The typical workflow of drawing a manga frame

3.3 Conventional Architectural Screens

3 CONVENTIONAL DRAWING SCHEMES
Drafting is a universal convention that uses
3.1 Traditional Manga Production screens, symbols, hatches, line-types, -styles and -
The notion of manga techniques in this work is widths to describe elements of a design, its proper-
based on the schemes used in traditional manga pro- ties, functions or materials. It is filled with repre-
duction (Nagatomo 2003), which is widely adopted sented knowledge and key information for a specific
in manga creative industries. Figure 2 illustrates the spatial composition. In addition to that, Architectural
typical five drawing steps for producing one mange drafting is flexible and allows the creativity of the
image-frame. First, artists decide the perspective and individual architect to influence the overall style of
roughly sketch the major structure of the scene with the screens and their variations. Mechanical engi-
pencils (Step 1). Next, they finalize the drawing with neering and Computer Aided Architectural Drafting
ink (Step 2). With the precise lines in place, artists (CAAD) typically have no personal style and depict
then begin the screening procedure. Based on the neutral and standard screens following norms of the
inked lines, manga artists usually select appropriate architectural, engineering and construction (AEC)
pre-print screen sheets to fill regions in order to ex- industries. These drawing styles are defined e.g. in
press shading, tone, texture, or atmosphere. The se- ISO 128 (2002).
lected screen paper, which is semi-transparent with The architectural sketch however, to which we refer
pre-printed patterns, is then overlaid on each of the here in this paper, does not employ these conven-
target regions (Step 3). The artist usually uses a tions. Similarly to an artistic drawing, various
knife to carefully carve out screen paper along the graphical elements and drawing techniques express
boundary and paste it on the target region (Step 4). visual clues of the depicted scene. Schön and Wig-
The manuscript is ready for print when all the re- gins (1992) describe the architectural drawing as a
gions are overlaid with selected screens (Step 5).
‘reflective conversation with materials of a design
The task is rather tedious, time- and labour-
intensive, especially when large amount of irregular situation’. ‘A designer sees, moves and sees again.
regions exist. Working in some visual medium the designer sees
what is ‘there’ in some representation of a site,
draws in relation to it, and sees what has been
3.2 Screen Library drawn, thereby informing further designing.’
A traditional screen paper is a transparent layer of
film, with a printed pattern on one side and adhesive
on the back. Screen papers are produced by different
companies in a variety of patterns and tones. Our
digital screens consist of repeated patterns generated
by texture synthesis methods as well as scanned pa-
per screens. In the screen library, each type of screen
(e.g. dot or hatch) has a range of density (tones) as
shown in Figure 3. Note that the bitonal screens of-
fer not only the one-dimensional tone density, but
also the variety of patterns. In other words, the
screen space is two-dimensional in nature.
Figure 3. The screen space. Apart from tone variety, another
dimension is the texture or pattern.
4 OVERVIEW OF THE MANGA DRAWING the input reference image. Previous image segmenta-
SYSTEM tion methods have used various schemes to deter-
mine the partitioning in a digital image, so that each
Richness preserving manga screening (Qu et al. of the pixels in a region are similar with respect to
2008) offers a solution to generating bitonal manga some characteristic or computed property, such as
screening for an input colour image. Their goal is to colour, intensity, or texture. We perform the seg-
preserve the visual richness in the original photo- mentation using mean-shift, which is a non-
graph by utilizing not only screen density, but also parametric clustering technique based on the analy-
the variety of screen patterns. Base on this work, we sis of a complex multi-modal feature space and de-
build a system to automatically draw architectural lineation of arbitrarily shaped clusters (Comaniciu &
depictions using manga techniques. Our system con- Meer 2002). With a kernel measure of the distance
sists of two major components, screening and line between pixels, it robustly produces noise-reduced
drawing, which are similar to the conventional segments. The meanshift segmentation result is usu-
manga production. For the line drawing part, we ally over-segmented, which is obviously too frag-
propose a line importance model to rank each line. mented for direct screening. Thus, we re-group seg-
With this ranking, architects can control the detail ments according to their colour difference and
level of lines needed in their architectural manga, proximity. Near segments with similar colours are
via a simple threshold. For the screening process, we grouped first. Each segment is then referred as the
utilize the solution of Qu et al. (2008) that automati- basic unit of the screen matching (Figure 4).
cally selects appropriate bitonal screens to represent
(or fill up) different regions in the image. Our aim is
to automate the screen selection process with the
goal of richness preservation and style consistency
to traditional manga.
Each type of screen supports a range of densities
to represent different intensity. Therefore, whenever
a screen is selected for a region, the intensity of this
region can be approximated by matching the screen
density. The major difficulty is to select an appro-
priate type of screen for each region, which can pre-
Figure 4. (a) an image with rich content will have thousands of
serve the texture similarity and the colour distin- segments. (b) Semi-automatic segmentation plus manual screen
guishability of the target regions in order to preserve assignment. (c) Our automatic screen assignment significantly
the architectural content. relieves user intervention producing reasonable results.
The system starts by segmenting the image into
regions (Section 5.1). Then, the system can intelli-
gently provide the optimal assignment of screens for 5.2 Screen Matching
different regions (Section 5.2). This is done by first After segmentation, each segment is matched with
projecting the available patterns from high- an appropriate screen from the screen library. Each
dimensional texture feature space to the low dimen- type of screen (e.g. dot or hatch) in this library has a
sional colour space, and then optimizing texture range of density (tones) as shown in Figure 2 above.
similarity. In addition, users are still able to control The screens are organized in a two-dimensional ma-
and/or override the selection of screens via simply trix with columns of varying density and rows of
tuning a few parameters. Finally lines are detected varying screen type.
and overlaid on the screens to finish the result (Sec- The main objective of screen matching is to pre-
tion 6). serve the rich content of the original reference im-
age. Here, we aim to preserve three classes of con-
tents including tone (or luminance), texture and
5 SCREENING chromaticity distinguish-ability. Referring to the two
available dimensions in the screen space (Figure 2),
Screening is a technique for applying textures and tone can be straightforwardly matched with the hori-
shades to drawings, used as an alternative to hatch- zontal screen density axis. However, matching the
ing. In the conventional process, patterns are trans- tone alone does not produce satisfactory results as
ferred to paper from pre-printed sheets, but the tech- demonstrated in Figure 6b. This screening result ig-
nique is also simulated in computer graphics. nores texture and chromaticity, and more impor-
tantly, its style is not consistent with the artworks of
5.1 Segmentation professional manga artists. Although the screening
style of artists varies, they all enrich the drawing by
As screens are laid over regions rather than individ- introducing appropriate pattern varieties.
ual pixels, we first identify regions, or segments, in
Figure 5. Objectives of the screen matching: (a) Texture preservation, (b) chromaticity distinguishability, and (c) tone preservation.
In (b), we assign the red and orange regions with two close patterns in below, as they are closer in colour; on the other hand, the
blue region will be screened with a substantially different pattern.

The screening method consists of the following 5.2.2 Colour-to-Screen Mapping

three steps: After the above texture-based matching, only tex-
− Texture-based matching. Segments containing tural regions are assigned with screens for the sake
apparent texture are first matched with screen of texture preservation. There are still many seg-
patterns based on the texture similarity, as shown ments may still not be assigned with any screen due
in Figure 5a. to their textureless characteristics. In this section, we
− Colour-to-pattern mapping. Unmatched segments assign these non-textural regions according to their
are then mapped to different pattern types with chromaticity difference among segments.
the goal of maintaining colour distinguishability, The solution proposed for chromaticity
via a colour-to-pattern mapping. As shown in distiguishability is to maintain the perceived dis-
Figure 5b, suppose we are assigning the bottom tances among different segment colours during the
patterns to the above three colours. It is desirable screen type assignment. In other words, we want to
to assign the red and orange regions with two maintain the same distance relationship among dif-
close patterns, as they are closer in colour. On the ferent colour segments after mapping to the screen
other hand, the blue region will be screened with space.
a substantially different pattern. The problem is that the screen feature is calcu-
− Tone matching. Finally, segments assigned to pat- lated in a high-dimensional space (24D), while the
tern types are screened by matching the tone. An colour is in a low-dimensional space (2D). And
illustration is shown in Figure 5c. there is no direct mapping between the different di-
mensional spaces. Note that we adopt the ‘LAB’
5.2.1 Texture-based matching colour space that consists of one luminance (L) and
During the screen mapping, we give high priority to two chrominance channels (AB). Only the two-
the texture feature preservation. That means seg- dimensional chrominance is considered as the colour
ments exhibiting apparent texture characteristics are feature during the screen type assignment. As men-
first assigned a screen pattern based on texture simi- tioned before, the luminance (tone) can be easily
larity. And such assignment will not be modified matched by selecting the appropriate screen density,
during the rest of the screening procedure. once the screen type is selected.
To quantify the texture characteristics, we com- We employ MDS (Cox & Cox 1994) to maintain
pute the texture features on both segments and such distance relationship during the assignment of
screens using Gabor wavelets (Manjunath & Ma high dimensional screens to the low dimensional
1996). It has been demonstrated as an effective tex- colours. The MDS reduces the dimensionality of the
ture identification technique for manga screens (Qu 24D texture feature to 2D, so that we can do the
et al. 2006). Consider an image we compute its Ga- mapping of texture to colour in the 2D space (Figure
bor wavelet feature in a per-pixel manner. In other 6b). This visual richness preserving screen assign-
words, each Gabor wavelet feature is represented as ment is called colour-to-screen mapping.
a 24-dimensional vector. If we have computed the MDS for all screens,
For each segment in the reference and each then we process each segment by first computing its
screen in the screen library, we compute their aver- average colour in the segment and then match a
age Gabor wavelet features as the representative. screen in the colour space. Note that the total num-
The matching can be done by measuring the Euclid- ber of distinct segment colours (in the magnitude of
ean distance of the representative texture feature hundreds) is usually much larger than the total num-
vectors between the segment and the candidate ber of screen types used for a single manga. Manga
screen. Figure 5c shows an example matching of the artists seldom use hundreds of types of screen in one
textured roof with the screen “grid”. Segments with- artwork. To maximize the usage of the available
out such a proper match either contain no apparent screens, we first perform the means clustering on the
texture characteristics or no suitable screen available segment colours. The centroid of each cluster is used
in the library. Hence, they are left to the following as a representative colour to match with the nearest
colour-to-pattern mapping. screen coordinate in colour space.
 istic of the cyan region. By maximizing the texture
similarity during adjustment of transformation, a
better screen can be matched.
Our target of optimization is the minimal texture
distance between the pre-assigned screens and the
new assignments. For the purpose of relative dis-
tance preservation, only rotation and flipping are
permitted. In specific, we quantize the rotation to
359 intervals, then the total number of possible
transformation is only 718 (359 x 2). Even an ex-
haustive search can solve the optimization in sec-
onds. Note that it is possible that the screens selected
via the above optimization do not completely coin-
cide with the pre-assigned screens from the first
stage. In that case, we simply use the pre-assigned
screens, as we give higher priority to the texture
similarity.

5.2.3 Tone Matching

After the first two stages, each segment has been as-
signed a screen type. The final step is to perform a
tone matching by selecting the appropriate screen
density. We match the overall greyness of the screen
with the average luminance of the segment. Next,
we need to rotate the chosen screen (with specific
density) in order to align with the texture character-
istics in the segment. Note that the Gabor wavelet
texture feature we employed is rotational invariant.
Such matching is performed on each segment by the
one by one manner until all regions are processed.

6 LINE DRAWING AND ABSTRACTION

Figure 6. The MDS reduce the 24D texture feature into 2D,
while maintaining the relative distance, so that the 24D tex- The lines drawn by manga artists are usually tidy
tures can be mapped into 2D colour space. without being too crowded or chaotic. They inten-
tionally omit many detail lines in order to maintain
Notice that such mapping is not unique, as the the tidiness of the manga. However, structural lines
MDS only maintains the relative distances. We can are retained, meaning that some lines are more im-
still transform the mapped coordinates of the colour portant than others, and these lines usually form the
space by rotation, scaling, translation, and flipping. main structure of a building or spatial arrangement.
Both scaling and translation can be excluded as the Edge detection extracts lines purely based on the
coordinates segment colours can be normalized and intensity gradient in the image. No attention is paid
zero-meaned. However, there are still two degrees of to the importance of a line. In many cases, edges ob-
freedom left. To determine the best transformation, tained by automatic edge detection methods are far
we actually perform colour-to-screen mapping on all from satisfactory as demonstrated in Figure 7b.
segments including those already assigned with
screens in the previous texture-based matching (Fig-
ure 6d), the cyan colour in the underlying colour
space is pre-assigned with the screen “line”. The ra-
tionale is that texture-based matching does not ac-
count for colours and, more importantly, we can util-
ize these pre-assigned segments to anchor the
transformation. Hence, we optimize the transforma-
tion with an objective of maximizing texture similar-
ity of the screens and segments. Figure 6c shows the
initial transformation with the cyan colour being as-
Figure 7. Edge detection result (Meer & Georgescu 2001):
signed with a “cross hatching” pattern which obvi- (a) Original image (b) Edge detection result
ously does not match well with the texture character-
 These noisy edges cannot be easily eliminated by decided by the overall curliness, length, visibility
parameter tuning of the edge detection, as important level, and the size of its largest associated segment.
structural lines may also be removed indifferently Figure 7b shows the unstructured edges from the
during the tuning. Therefore, we propose a model to raw edge detection result. With the proposed line
measure the importance of a line and unimportant importance model, we can generate the lines with a
lines are filtered out by adjusting a single threshold. desired degree of abstraction to simulate manga
style architectural drawings with different level of
complexity or line abstraction. Figure 8 shows five
6.1 Line Formation
different levels of abstraction of lines.
We begin our line formation from the edge detection
on the original image. We adopt an improved edge
detection method (Meer & Georgescu 2001), which 7 RESULTS AND DISCUSSION
shows the ability to detect consistent results, even on
weak edges. The detected edge pixels however are Variety of input examples, including architectural
only a set of unstructured entities, so we first con- photographs, colour drawings and artworks of spa-
nect them using 8-connectivity. To resemble lines tial descriptions have been testified with our auto-
drawn by the artist, which usually contains no matic manga style architectural drawing system. We
branching or self-intersection, we break the lines have implemented the described system on a
into pieces at the point with abrupt change in tan- 2.6GHz Intel Core2 CPU computer with 3GB mem-
gent. We refer the result as line map. ory. The solution time for an 800x800 image is only
about two minutes, where the time includes segmen-
tation, texture matching, colour-to-pattern mapping,
6.2 Line Importance Modelling
tone matching, and the minimal user control.
To tell structural lines from details, the key is to de- In all the resulting images, we laid the same set of
termine the importance of a line, so that a high im- lines, determined in Section 5 above, over all the
portance value indicates structural lines while lower screen renderings compared in this paper, including
value indicates detail lines. the manga screening, which is determined in Section
Architects usually first draw the long and straight 4, and the halftoning and hatching results, for the
structural lines during sketching. The short and curly sake of fair comparison. As illustrated in Figure 9,
lines are drawn at later stages or even omitted the result (d) is generated by blending the screening
(Cheng 2006). Hence, the curliness and the length of result (b) with the line abstraction result (c). This
lines suggest its importance. We also believe that technique allows communicating architectural de-
lines associated with the major segment in the pho- sign and features in a novel way, whereby thresholds
tograph should be more important. This suggests and details can be emphasised and distinct elements
four factors are to be considered, the length, the highlighted. This is of particular interest in a context
curliness of a line, the size of largest associated of design-learning or explanation of architectural in-
segment, and the visibility value. tensions.
We measure the curliness by computing the de- In Figure 10 we show an example that demon-
viation of curvatures of each edge pixel on the line. strates the chromaticity distinguishability of the pro-
The curvature at each pixel can be estimated from posed system. Unlike the monotonous pattern ap-
the fitted polylines. The length can be obtained by peared in the halftone result in Figure 10b, the
counting the number of pixels on a line. The visibil- colour variety in the input photo is well preserved by
ity level is computed by first forming an image with screen variety as shown in 10d. We used five
pyramid, and lines with a counterpart in higher level screens to represent five distinctive input colours
have a larger visibility value. We identify the struc- (Figure 10c). By blending our result with the input
tural lines (lines associated with the major compo- image, such result illustrates the faithfulness of our
nents in the photograph) by measuring also the size textural presentation and emphasises the relationship
of its largest associated segment. The size of a seg- of building elements with each other as well as to
ment can be obtained from our segment map (Sec- the whole architectural design.
tion 4.2.1). The final line importance value will be

Figure 8. Level of line abstraction controlled by the line importance values.

Figure 10. Unlike the monotonous pattern in halftone technique (b), the colour variety in a colour input photo or drawing will be
well preserved with screens variety, as shown in (c), by our colour-to-pattern mapping (d).

This is particular useful to communicate large

concepts that feature a variety architectural topolo-
gies, material, form, details as well as additional ab-
stract information, such as design intent, light quali-
ties, movement, and depth. Subsequently manga can
enhance the communication of architectural repre-
sentations. Our technique allows designers to ab-
stract spatial relationships while at the same time
preserve properties of material, surface or detail.
The narrative of a manga has long been defined
by the figure-ground synergy since its conception
that is the depiction of character or protagonist (fig-
ures) going through a certain action or event in a set-
ting (ground). The idea of a spatial character starts
by placing the emphasis onto seeing the background,
the setting as the focus for narrative, with or without
the figures as supporting actors (Ng et al. 2006).
Subsequently the protagonist in our drawings shifts
from a character or person to the architecture itself.
Our technique allows for the development of a
new gene of architectural design communication that
reaches beyond graphical depiction of architecture.
Figures 1, 9, 11 and 12 demonstrate the effect of
the global optimization, which considers all the tex-
ture preservation, chromaticity distinguishability, as
well as tone preservation. During the bitonal proc-
ess, images with large variety of colours appear
more pleasing by our method, comparing with the
start-of-the-art halftoning and hatching techniques.
The strengths of architectural drawings lay in the
ambiguity and abstract translations of the original
depicted spaces and designs. Our technique allows
architects to blend from a photorealistic to an ab- Figure 9. Architectural drawing using manga techniques. The
stract depiction of architecture using a distinct ex- result (d) is generated by blending screening result (b) with the
line abstraction result (c).
pression of manga style visual communication.
Figure 11. “Tea house”. Our result (b) properly preserves the
multiple textures and colour distinguishability in the original
photo (c).

8 CONCLUSIONS

We have presented a novel bi-tonal image drawing

technique for architectural design that utilize manga
drawing techniques, and provide an optimal screen-
ing scheme by using multiple screens and level-of-
detail line abstractions. The high dimensional
screens enable the texture and tone preservation, as
well as the chromaticity distinguishability of the de-
picted original. Our results give manga-like draw-
ings comparable to those manually prepared by art- Figure 12. “Colour house”. (a) Original image, (b) halftone
ists, and the system is efficient and convenient in use method, and (c) our method.
in architectural design and communication. Addi-
tionally we successfully demonstrated how the spe-
cific drawing technique of manga can be utilized to REFERENCES
design and communicate spatial arrangements. The
intersections of manual and digital instruments mir- Buchanan, J.W. & Verevka, O. 1995. Edge preservation with
ror the working styles of architects, who deal with a space-filling curve halftoning. Proceedings of Graphics In-
terface 95: 75–82.
variety of realms, instruments and stakeholders. Cheng, N.Y.-W. 2006. Learning design sketching from anima-
Hereby manga architectures aid not only users to tions and storyboards. International Journal of Architec-
understand larger architectural concepts, properties tural Computing 4(1): 1-17.
of form, material, light, etc, but aids in story telling Comaniciu, D. & Meer, P. 2002. Mean shift: A robust ap-
and consultancy of professionals (Schnabel & Howe proach toward feature space analysis. IEEE Trans. Pattern
Anal. Mach. Intell. 24(5), 603–619.
2010). Laypersons easily can access the architects’ Cox, T.F. & Cox M.A.A. 1994. Multidimensional Scaling.
intension of design and function. London, UK: Chapman & Hall
In future work, we would like to explore other Decarlo, D. & Santella, A. 2002. Stylization and abstraction of
options of colour spaces and texture spaces. Cur- photographs. SIGGRAPH ’02: Proceedings of the 29th an-
rently, we measure the chromaticity similarity in nual conference on Computer graphics and interactive
CIE L*a*b* space and pattern similarity in Gabor techniques. New York, NY. ACM Press. 21: 769–776.
Durand, F., Ostromoukhov, V., Miller, M., Duranleau, F. &
wavelet feature space. It is still an open problem to Dorsey, J. 2001. Decoupling strokes and high-level attrib-
understand whether these two kinds of similarities utes for interactive traditional drawing. Proceedings of the
are linear to our human perceptual distance. This 12th Eurographics Workshop on Rendering Techniques,
will lead to a novel understanding of how architec- London, UK. Springer-Verlag: 71–82.
tural design is communicated and understood. Floyd, R.W. & Steinberg, L. 1974. An adaptive algorithm for
spatial grey scale. SID International Symposium Digest of
Technical Papers, Washington, D.C. Society for Informa-
tion Display: 36–37.
Goldschmidt, G. 1991. The Dialectics of Sketching. Creativity Pang, W.-M., Qu, Y., Wong, T.-T. Cohen-Or, D. & Heng, P.-
Research Journal. 4(2):123–143. A. 2008. Structure-aware halftoning. ACM Trans. Graph.
Gooch, A.A., Olsen, S. C., Tumblin, J. & Gooch, B. 2005. 27(3): 1–8.
Color2gray: saliencepreserving colour removal. ACM Qu, Y., Pang, W.-M., Wong, T.-T. & Heng, P.-A. 2008. Rich-
Trans. Graph. 24(3): 634–639. ness-preserving manga screening. ACM Transactions on
Grabli, S., Durand, F. & Sillion, F. 2004. Density measure for Graphics (SIGGRAPH Asia 2008 issue) 27: 155:1-155:8.
line-drawing simplification. Proceedings of Pacific Graph- Qu, Y., Wong, T. T. & Heng, P. A. 2006. Manga colorization.
ics. ACM Transactions on Graphics, SIGGRAPH 2006. 25:
ISO 128 (2002). Standards Handbook: Technical drawings 1214–1220.
Volume 1: Technical drawings in general. Robbins, E. 1994. Why Architects Draw. Cambridge, MA:
Jarvis, J. F., Judice, C.N. & Ninke, W.H. 1976. Survey of MIT Press.
techniques for the display of continuous tone pictures on Salisbury, M.P., Wong, M.T., Hughes, J.F. & Salesin, D.H.
bilevel displays. Computer Graphics Image Process 5(1): 1997. Orientable textures for image-based pen-and-ink il-
13–40. lustration. SIGGRAPH ’97: Proceedings of the 24th annual
Knuth, D.E. 1987. Digital halftones by dotdiffusion. ACM conference on Computer graphics and interactive tech-
Transactions on Graphics 6(4): 245–273. niques, New York, ACM Press/Addison-Wesley Publishing
Li, Y., Sun, J., Tang, C.-K. & Shum, H.-Y. 2004. Lazy snap- Co: 401– 406.
ping. ACM Transaction on Graphics 23(3): 303–308. Schnabel, M.A. & Howe, E.L.C. 2009. Client design literacy
Manjunath, B.S. & Ma, W.Y. 1996. Texture features for and consultancy skills in architectural education, N. Gu, M.
browsing and retrieval of image data. IEEE Transactions Ostwald & A. Williams (eds.), Computing, Cognition and
on Pattern Analysis and Machine Intelligence 18(8): 837– Education - Recent Research in the Architectural Sciences.
842. Newcastle, Australia: ANZAScA & Uni. of Newcastle.
Mantiuk, R., Myszkowski, K. & Seidel, H.-P. 2006. A percep- Schön, D.A. & Wiggins, G. 1992. Kinds of seeing and their
tual framework for contrast processing of high dynamic functions in designing. Design Studies. 13(2): 135-156.
range images. ACM Trans. Appl. Percept. 3(3): 286–308. Ulichney, R. A. 1987. Digital Halftoning. Cambridge, MA:
Meer, P. & Georgescu, B. 2001. Edge detection with embed- MIT Press.
ded confidence. IEEE Transactions on Pattern Analysis Velho, L. & Gomes, J.M. 1991. Digital halftoning with space
and Machine Intelligence 23(12): 1351–1365. filling curves. Computer Graphics 25(4): 81–90.
Nagatomo, H. 2003. Draw Your Own Manga: All the Basics. Velho, L. & Gomes, J.M. 1995. Stochastic screening dithering
Japan: Kodansha International Ltd. with adaptive clustering. Computer Graphics (Proceedings
Naiman, A. & Lam, D. 1996. Error diffusion: wavefront tra- of SIGGRAPH’ 95): 273–276.
versal and contrast considerations. Proceedings of Graph- Winkenbach, G. & Salesin, D. H. 1994. Computer-generated
ics Interface 96: 78–86. pen-and-ink illustration. Proceedings of SIGGRAPH 1994:,
Ng, K, Schnabel, M.A. & Kvan,T. 2006. Architectural Anima- 91–100.
tion becomes Alive - Creating Spatial Narrative with Spa-
tial Characters for Animations, Communicating Space(s)
Volos, Greece. eCAADe: 598-603