在東方題材相關的電影或遊戲場景中，通常會有大量的中國特色建築。由於
其結構複雜，即使使用3D 建模工具軟體，在建立模型時，仍需要大量時間與相
關背景知識。何況一個場景中往往含有大量的建築物，是一個需要極大人力成本
的工程，對於非專業人員更是一大難題。我們的目的是使用程序化生成大量中國
特色建築中之涼亭，且這些涼亭符合使用者喜好，對於非專業人員而言也好上手
的3D 建模方式。
本研究中，我們讓使用者選取涼亭樣本，其中樣本是以涼亭框架與控制點配
合參數建立而成。在取樣過程中，依照不同數量的參考樣本，給予不同的取樣策
略，讓目標樣本可以順利地被找到。用這種方式可以減少建構中國特色建築模型
時，開發人員所需的背景知識。更可以藉由過程中提供的不同樣本激發創造力。
透過對樣本的選取與喜好程度之評分後，我們利用類神經網路，根據樣本參數與
喜好分數之間的關係，產生出一個當前喜好風格的類神經網路模型(Model)。我們
將涼亭結構分為Roof、Body 與Platform 三個部分，將三個學習標的類神經網路模
型學習完成後，利用此三個類神經網路模型來推薦大量具有使用者喜好風格的3D
模型。
依據我們提出的程序化建模之實驗結果，系統可以加速取樣到使用者欲尋找
的目標樣本，且在過程中只需要使用視覺進行觀察與選取。再者使用類神經網路
進行風格學習，可以有效的學習出使用者喜好與樣本之間的關係。在學習完成後，
即可推薦大量使用者喜好的涼亭樣本，達到減少生成大量3D 模型的時間。在整
個使用過程中，有助於使用者了解中國特色建築與激發創造力，並且不需要大量
的背景知識。在使用者回饋結果中，皆能推薦出使用者喜好之風格涼亭樣本。

In Eastern movie or game scenes, there are usually use a lot of Chinese-characteristic
buildings. Due to its complex structure, even using 3D modeling software. It takes a lot
of time and related background knowledge to build these architectural models. What’s
more, a scene often contains a lot of buildings, it requires a lot of labor costs. For nonprofessionals
is a big problem. Our goal is to use programmatic generation of a large
number of Ting with Chinese characteristics. And these Ting meet user preferences, it is
also a good 3D modeling method for non-professionals.
In this study, we allow the user to select Ting samples. The sample is based on the
parameters of the Ting frame and control points. During the sampling process, according
to different number of reference samples, we give different sampling strategies. Let the
target sample can be found successfully. In this way, it is possible to reduce background
knowledge, when developer construction of architectural models with Chinese characteristics.
It can stimulate creativity through different samples provided in the process. After
the selection of the sample and give scores of preference, We generate a neural network
model based on the relationship between sample parameters and preference scores. We
divide the Ting structure into Roof, Body, and Platform. After these three neural network
models are be created by learning, we use these three models to recommend a large number
of 3D models with user preferences.
The experimental results of our procedural modeling method ,the system can accelerate
sampling to the target sample which is user looking for. In the process, only visual
observation and selection are needed. Using neural networks for style learning, which can
effectively learn the relationship between user preferences and samples. After learning is
completed, system can recommend a large number of Ting samples with user preferences.
Reducing time to generate large numbers of 3D models, and users don’t need a lot of background
knowledge. Learn more about characteristics of Chinese-characteristic buildings
in the process, and stimulate the effect of creativity.In the result of user study, we found
out that our system can recommend the user preferences style Ting samples.

[1] C.-Y. Huang and W.-K. Tai, “Ting tools: interactive and procedural modeling of
chinese ting,” The Visual Computer, vol. 29, no. 12, pp. 1303–1318, 2013.
[2] Y.-C. Chung and W.-K. Tai, “An icon-based procedural approach for chinese featured
buildings,” Master’s thesis, National Taiwan University of Science and Technology,
Republic of China, 2017.
[3] J. Liu, “Component-driven pocedural modeling for ancient chinese architecture of
the qing dynasty,” International Journal of Architectural Heritage, vol. 12, no. 2,
pp. 280–307, 2018.
[4] L. Krecklau, D. Pavic, and L. Kobbelt, “Generalized use of non-terminal symbols for
procedural modeling,” Computer Graphics Forum, vol. 29, no. 8, pp. 2291–2303,
2010.
[5] I. Demir, D. G. Aliaga, and B. Benes, “Proceduralization for editing 3d architectural
models,” 3D Vision (3DV), 2016 Fourth International Conference on, pp. 194–202,
2016.
[6] P. B. Silva, P. Müller, R. Bidarra, and A. Coelho, “Node-based shape grammar representation
and editing,” in Proceedings of the Workshop on Procedural Content
Generation in Games (PCG’13), 2013.
[7] F. Ricci, L. Rokach, B. Shapira, and P. B. Kantor, Recommender Systems Handbook.
New York, NY, USA: Springer-Verlag New York, Inc., 1st ed., 2010.
[8] K. Leimer, L. Gersthofer, M. Wimmer, and P. Musialski, “Relation-based
parametrization and exploration of shape collections,” Computers & Graphics,
vol. 67, pp. 127–137, 2017.
[9] Y. Zhang, Z. Liu, Z. Miao, W. Wu, K. Liu, and Z. Sun, “Single image-based datadriven
indoor scene modeling,” Computers & Graphics, vol. 53, pp. 210–223, 2015.
[10] M. Pirovano, R. Mainetti, and D. Loiacono, “Volcano: An interactive sword generator,”
in Games Entertainment Media Conference (GEM), 2015 IEEE, pp. 1–8, IEEE,
2015.
[11] A. Van den Oord, S. Dieleman, and B. Schrauwen, “Deep content-based music recommendation,”
Advances in neural information processing systems, pp. 2643–2651,
2013.
[12] S. H. Hsu, M.-H. Wen, H.-C. Lin, C.-C. Lee, and C.-H. Lee, “Aimed-a personalized
tv recommendation system,” in European Conference on Interactive Television,
pp. 166–174, Springer, 2007.
[13] S. Sedhain, A. K. Menon, S. Sanner, and L. Xie, “Autorec: Autoencoders meet collaborative
filtering,” in Proceedings of the 24th International Conference on World
Wide Web, pp. 111–112, ACM, 2015.
[14] C. Christakou, S. Vrettos, and A. Stafylopatis, “A hybrid movie recommender system
based on neural networks,” International Journal on Artificial Intelligence Tools,
vol. 16, no. 05, pp. 771–792, 2007.
[15] M. Dang, S. Lienhard, D. Ceylan, B. Neubert, P. Wonka, and M. Pauly, “Interactive
design of probability density functions for shape grammars,” ACM Transactions on
Graphics (TOG), vol. 34, no. 6, p. 206, 2015.
[16] J. H. Holland, Adaptation in natural and artificial systems: an introductory analysis
with applications to biology, control, and artificial intelligence. MIT press, 1992.
[17] Wikipedia, “Fitness proportionate selection — Wikipedia, the free encyclopedia.”
https://en.wikipedia.org/wiki/Fitness_proportionate_selection,
2003. [Online; accessed 29-March-2018].
[18] 羅哲文, 中國古代建築. Shanghai: 上海古籍出版社, 1990.
[19] 丁援, 一本就通：中國建築. Taipei: 聯經出版公司, 2015.
[20] Q. D. of Qing Dynasty, Qing Gong Bu Gongcheng Zuofa Zeli. China: Beijing, Qing
Dynasty of China, 1733.
[21] E. Catmull and R. Rom, “A class of local interpolating splines,” in Computer aided
geometric design, pp. 317–326, Elsevier, 1974.
[22] Wikipedia, “興慶宮— Wikipedia, the free encyclopedia.” https://zh.
wikipedia.org/wiki/%E5%85%B4%E5%BA%86%E5%AE%AB, 2006. [Online; accessed
29-March-2018].
[23] Wikipedia, “琵琶亭— Wikipedia, the free encyclopedia.” https://zh.
wikipedia.org/wiki/%E7%90%B5%E7%90%B6%E4%BA%AD, 2010. [Online; accessed
29-March-2018].
[24] E. Panadero and A. Jonsson, “The use of scoring rubrics for formative assessment
purposes revisited: A review,” Educational Research Review, vol. 9, pp. 129–144,
2013.
[25] Wikipedia, “Supervised learning — Wikipedia, the free encyclopedia.” https://
en.wikipedia.org/wiki/Supervised_learning, 2005. [Online; accessed 15-
March-2018].
[26] Wikipedia, “Nonlinear regression — Wikipedia, the free encyclopedia.” https://
en.wikipedia.org/wiki/Nonlinear_regression, 2007. [Online; accessed 15-
January-2018].
[27] Blueteak, “Unity-neural-network.” https://github.com/Blueteak/
Unity-Neural-Network, 2017. [Online; accessed 31-July-2017].
[28] Wikipedia, “Cross-validation (statistics) — Wikipedia, the free encyclopedia.”
https://en.wikipedia.org/wiki/Cross-validation_(statistics), 2003.
[Online; accessed 5-April-2018].
[29] A. L. Maas, A. Y. Hannun, and A. Y. Ng, “Rectifier nonlinearities improve neural
network acoustic models,” in International conference on machine learning, vol. 30,
p. 3, 2013.