Computational Aesthetics of Character Design
NUS School of Computing

Objectives

The impact of character design can be paraphrased as the alignment of artist-intent and audience-perception of key character features (e.g. traits, role, and actions). Our goal is to capture expert domain knowledge and help novices create impactful characters in an intuitive and simple manner. Objectives:

1) Given an input character figure, computer should be able to classify its stereotype;

2) Given any doodle, the computer should be able to retarget a drawing or 3D model from a library to that doodle;

3) Given very simple inputs, like stereotype labels, the computer should be able to generate characters;

4) Given an input image, and some target rules or a template, the computer should be able to cartoonify a face.

Achievements

We have contributed several key technologies and data that take us closer to these aims:

Novel vector representation of blended circle, triangle and square primitive shapes that helps us intuitively describe humanoid body shapes
Vector based image and geometry warping that allows retargeting of existing character art to input doodles or machine generated scaffolds
Online corpus of annotated popular cartoon human body shapes available to the scientific community
Shape perception model for humanoid cartoon characters harnessed from online rating games
Intelligent generation of shape scaffolds with Bayesian models and hybrid classifiers, that help in procedural character generation
Shape stylized Face Cartoonification

Research Output

	We share the NUS Humanoid Cartoon Corpus here. It comprises of 143 simplified silhouettes, About 140 characters were sourced from the internet, or drawn in-house (some at GAMBIT MIT). Three CSV files accompany these images: body-part point-contour data, body-part vector shape data, and perception labels aggregated from laymen via an online game. The character images were annotated through our online annotation tool. We will continue to update this database, and hope that media and informatics researchers test new algorithms. Download Corpus (to be released soon)
	We have proven that our shape vector is more intuitive and performs better than low level point features in learning and deformation applications. We propose a continuous blend between the three distinct primitive shapes: circle, triangle and square, implementing consistent interpolation between 8 quadratic Bezier segments. We invite researchers to use our Humanoid cartoon corpus to verify this, and hope that we can soon put up our vector fitting code as well. Learning Primitive Shapes in Cartoon Designs, Trans. MLDM 2011
	Our game-inspired layman perception modeling methodology will impact the media industry, which we know collects audience consumption data for trend analysis. Shadow Shoppe Game, 2010
	We have contributed a general perception classifier with implementation details that helps drawing applications incorporate intelligent character doodle completion/classification tools. Informed character pose and proportion design, Visual Computer 2011 Learning Shape-Proportion Relationships from Labeled Humanoid Cartoons, IDC 2010
	We have shown better face cartoonification with identity preservation and shape inspired exaggeration, opposed to existing point and distance feature based methods. This could help artists exploit our novel shape based image warping and create better cartoon faces from real/neutral images. Shape Stylized Face Caricatures, MMM 2011
	Procedural retargeting of high quality content is now possible for generating characters with different shape perception probabilities. Drafting 2D Characters with Primitive Shape Scaffolds, Computation World 2010 Sketch Based 3D Character Deformation, MLDM 2010

Awards/ Achievements

1) Best Paper Award for "Drafting 2D Characters with Primitive Shape Scaffolds", Content 2010 (Computation World 2010), Lisbon, Nov 2010

2) Distinction award at the Youth Science Conference 2010, for Flash Character Design Game by Raffles Secondary Boys School

3) Bronze award at the Youth Science Conference 2009, for Flash Game and data analysis by Nanyang Girls Secondary School

Papers

(Journal)

1) Md. Tanvirul Islam, Kaiser Md. Nahiduzzaman, Why Yong Peng, Golam Ashraf, Informed character pose and proportion design, The Visual Computer, Springer, 27:4, pp 251-261, 2011.

2) Md. Tanvirul Islam, Kaiser Md. Nahiduzzaman, Why Yong Peng, Golam Ashraf, Learning Primitive Shapes in Cartoon Designs, Transactions on Machine Learning and Data Mining, IBai Publishing, 4:1, pp 19-31, 2011.

3) Li Mo, Golam Ashraf, Sketch Based 3D Character Deformation, Transactions on Edutainment (special issue of 5th International Conference on Digital Media and Digital Content Management), LNCS, Volume 6530/2011, pp. 177-188, DOI: 10.1007/978-3-642-18452-9_14, Chongqing, China, Dec. 2010.

4) Golam Ashraf, Why Yong Peng, Md. Tanvirul Islam, Mining human shape perception with role playing games, GSTF International Journal on Computing, ISSN: 2010-2283, Vol. 1, Aug. 2010.

(Conference)

1) Nguyen Kim Hai Le, Why Yong Peng, Golam Ashraf, Shape Stylized Face Caricatures, 17th International Conference on MultiMedia Modeling, Advances in Multimedia Modeling, LNCS, 6253:536-547, Taipei, Taiwan, Jan. 2011

2) Golam Ashraf, Kaiser Md. Nahiduzzaman, Le Nguyen Kim Hai, Li Mo, Drafting 2D Characters with Primitive Shape Scaffolds, Second International Conference on Creative Content Technologies, Computation World, Lisbon , Nov. 2010. (Best Paper)

3) Md. Tanvirul Islam, Why Yong Peng, Golam Ashraf, Learning Shape-Proportion Relationships from Labeled Humanoid Cartoons, 6th International Conference on Digital Content, Multimedia Technologies and its Applications, pp. 416-420, Seoul, Aug. 2010

4) Md. Tanvirul Islam, Kaiser Md. Nahiduzzaman, Why Yong Peng, Golam Ashraf, Learning character design from experts and laymen, Cyberworlds 2010, pp 134-141, Singapore, Oct 2010

5) Md. Tanvirul Islam, Why Yong Peng, Golam Ashraf, Learning from Humanoid Cartoon Designs, Advances in Data Mining, Applications and Theoretical Aspects (LNCS Springer), 6171:606-616, Berlin, 2010

Manpower

Rubiat Habib	Graduate Intern	Jan-July 2009	Shape fitting, data analysis
Sarika Mohapatra	Undergrad Intern	July-Aug 2008	Vector shapes, envelop deformation
Kechit Goyal	Undergrad Intern	July-Aug 2008	Expressive vector face
Naval Chopra	Undergrad Intern	July-Aug 2009	Vector field image warping
Nikhil Vij	Undergrad Intern	July-Aug 2009	Sketch fitting
Zibran Shaikh	Undergrad Intern	July-Aug 2009	Body part deduction from sketches
Li Bo	Graduate Student	2008-2009	Voronoi Skin Diagrams
Qian Kun	Graduate Student	2008-2009	Bayesian Shape Inference from labels
Li Mo	Masters Student	2009	3D mesh analysis, deformations
Md. Tanvirul Islam	Research Assistant	2009/10	Data analysis and modeling
Kaiser Md. Nahiduzzaman	Research Assistant	2009/10	Annotation, shape fitting, warping
Le Nguyen Kim Hai	Research Assistant	2010	Mobile graphics, face cartoonification
Le Truong Vinh Phu	Research Assistant	2010	Shape fitting, perception experiments

Administrative Details

WBS:		R-252-000-357-490
Project Title:		Computational Aesthetics: Shape Psychology in Character Design
University:		National University of Singapore
Approved budget:		S$ 225,000
Period of funding		06/2008 to 12/2010
Funding Agency		Media Development Authority of Singapore, GAMBIT
Title, Name	Role	Department, Faculty	University
Dr. Golam Ashraf	PI	Computing	NUS
Dr. Why Yong Peng	Collaborator	Psychology	NUS
A/P Fredo Durand	Co-PI	CSAIL	MIT

Activity Log: visits since Aug. 2011

Computational Aesthetics of Character Design NUS School of Computing

Computational Aesthetics of Character Design
NUS School of Computing