Developing aesthetic measures for 3D head models (Cameron ...

broadbeansromanceΤεχνίτη Νοημοσύνη και Ρομποτική

18 Νοε 2013 (πριν από 3 χρόνια και 9 μήνες)

115 εμφανίσεις

Project Title
:

Developing Aesthetic Measure
s

for 3D Head Models

Big Question
s
:


1.

What constitutes aesthetic judgment?

2.

How do we make value judgments?

3.

Are these judgments innate or
acquired?

Research Question:

Can we devise a
goodness
-
of
-
fit

function that measures the “visual
a
esthetics” of generated three
-
dimensional meshes?

Project Team
:

Marietta E. Cameron, Ph.D., Associate Professor of Computer Science




Birmingham
-
Southern College, University of North Carolina at Asheville



Reed Milewicz, Class of 2011
, Computer Science


Birmingham
-
Southern College





Brandon Shewmake
, Sophomore, Computer Science and Mathematics


U
niversity of Alabama



Introduction
:

Our ultimate goal is to produce an automated character generator. Our envisioned virtual
modeler would produce a selection of characters that embod
y
various attributes given by the
system’s users. There are multiple
systems currently available that create three
-
dimensional
characters with substantial interactive guidance. Many of the current popular games provide
templates of character attributes and allow gamers to mix and match to create new characters.
The graphi
cal social network site, Second Life, provides a similar system that allows each user to
create his/her own three dimensional avatar. We seek a system that would create fully developed
characters and make “a judgment” on which characters to present by usin
g a broad feature list
provided by the user.

This project falls with
in

the recently defined field of “Computational Aesthetics.” Computational
Aesthetics integrates concepts and ideas from art, philosophy, psychology, and computer
science
(Hoenig, 2005)
. The discipline supports efforts in developing a formalized notation of
how we make aesthetic judgments. Researchers and designers in the area seek to create virtual
artists that assist
their human counterparts

in evaluating wor
ks of art.
(McCormack, 2008)

A
review of the literature reveals a formalized definition of interactive systems utilizing human
evaluation
(Sims, 1991)
,
a plethora of functions designed to understand,

formalize, and measure
the influence of aesthetics on website traffic
(Faria & B. S. de Oliveira, 2006)
, and a number of
attempts to evaluate the intrinsic beauty of two
-
dimensional images
(Machado, Romer
o, &
Manaris, 2008)
. While there are numerous published techniques of generating three dimensional
meshes, articles defining a
“goodness of fit” function for these meshes are limited.

In this work
w
e
establish

a

working


definition of aesthetics
,
explore the role of
aesthetics
in
producing a work of art, and
seek

to develop functions that measure the “visual aesthetics” of
the
se

generated models.

Developing Aesthetic Measures for Three
-
Dimensional Head
Models




Page
2

of
11



Project Activities:

Our current system is limited to three
-
dimensional head models. We designed a template
tria
ngular mesh of 3204 points and
manually
annotated the mesh with
36 biological landmark
points.
Appendix A contains

a detailed description

of each landmark
.
Since we assume facial
symmetry,
our template mesh is of a half
-
head

(
Figure
)
. To generate new head models, we
randomly p
erturb the
landmarks, create

a movement function utilizing the translation between
the template landmarks and
the new positions, and then use

the movement function to move the
other template points
(
Figure
)
.


The following is a list of project activities conducted June 2010 through May 2011.

1.


Project Website
(Summer and Fall 2010
)


Undergraduate Brandon
Shewmake

s
erved as the webmaster for the site until February
2011. As of May 2011, the website will be re
formatted and transferred to a G
oogle site
sponsored by the University of North Carolina at Asheville. The URL is


H
alf
-
head
template annotated with
landmark points.

Full head with symmetry applied.

Figure 1: Head Template


Figure 2: Sample Generated Heads. Insert is of template.

Developing Aesthetic Measures for Three
-
Dimensional Head
Models




Page
3

of
11



https://sites.google.com/a/unca.edu/computational
-
aesthetics/


2.


Discussion Seminars (June 2010


Spring 2011
)

During the summer of 2010 the research team met Mondays, Tuesdays, and Thursdays
to discuss research
articles, soft
ware development issues, our

ever
-
changi
ng definition
of aesthetics, and demographic requirements for our questionnaire participants.


During
the academic year, the frequency of the meetings
was
reduced to once a week.


3.


Software

Implementations and Revisions

(June 2010


Spring 2011)

1)

3D Head Generator:
Our previous work had been implemented in C++ and
required Cinema 4D for some of our visuals.
To make our work more accessible
via the web,
we
have
converted our programs
to Java

and utilize Java 3D
for our
visualizer. Our previous version of head generator produced several non
-
human
heads. We dedicated considerable time
to
modifying our algorithms to recognize
geometric

constraints.


2)

Questionnaire Applet:
This applet was
designed so that questions and expected
responses could be easily modified.


4.


Face
Recognition
Database

(Spring 2011)

We requested and received access to the Texas 3D Face Recognition Database
(Gupta,
Castleman, & Bovik, Texas 3D
Face Recognition Database, 2010)
. This database
contains 1149 range and portrait images pairs of
118
adult human subjects.
The
database provides facial landmarks, gender
, age, and ethnicity for each subject. While
this database provides only a subset of the landmarks we use, the data will
prove useful
in testing our
head generation method. We are investigating the idea of generating an
“average” head from this database
to replace our template model.



5.


Que
stionnaire (June 2010


Spring 2011)

The objective of our questionnaire is to determine which aesthetic properties (eg.
attractiveness, likability, dominance) human
participants would associate

with
a subset
of
models

generated from our system.

The results would be used to develop an
aesthetic system to evaluate any individual model.
Currently, we propose
a closed
format
questionnaire to present to
participants a series of twenty generated model
s
. For
each mo
del, a
participant would rate using

a scale from 0 to 10 the degree that a
particular aesthetic property

(e
.
g. attractiveness, likeability,

cheerfulness
,
trustworthiness, dominance
)

is featured

in the model.
We will extend invitations to at
least one hundred p
articipants. Participants will be encouraged to share demographic
i
nformation (gender, age, ethni
city) so that we can examine the extent of cultural bias
in making aesthetic evaluations.

We must resolve the following questionnaire issues:


a.

Determination
of a
esthetic properties to be evaluated.

One option is
to narrow the properties to likability and dominance. Another option is
Developing Aesthetic Measures for Three
-
Dimensional Head
Models




Page
4

of
11



to change the question
naire to open format in which a

participant list
s

and rank
s

adjectives that he/she feels best describe th
e model.

While
this option reduces the “leading question”


effect, statistical analysis
becomes more difficult.

Open format would also require a greater time
commitment from the participant.


b.

Selection of head model subset.
In selecting a subset, we are making
initial aesthetic judgments that will greatly skew the results from our
participants.
Presenting twenty
models

randomly generate
d

for each
participant

is infeasible since the method would produce approximately
2000 mo
dels each evaluated by only one user.



c.

Participant invitations.
Initially, we proposed to invite students who
were enrolled in art, psychology, and computer science classes at
Birmingham
-
Southern College. Since this project is to be continued at
another

institution, we have decided to extend the inter
-
institutional
invitations.


d.

Participant

rewards.
Our function depends heavily on each
questionnaire respondent’s commitment to completing the entire
questionnaire.

To encourage completion of the questionnaire we
propose offering the participant an Amazon gift card. We must locate
funding for this idea.


6.


Honor’s Day Presentation

(April 28, 2011)

Senior Reed Milewicz

presented the current status of the project during the Honor’s
Day Conference of Birmingham
-
Southern College.


Relationship between Research Question and BIG Questions
:






Knuth’s statement above sums up our thoughts on how our research question actually contributes
to a better understanding o
f
possible answers to the big questions. If we are successful in
creating a function to measure aesthetic value, then we will have shown aesthetic judgments to
be acquired since machines
--

according to our current understanding
--

do not innately possess
It has often been said that a person doesn’t really understand something until
he
teaches it to someone else. Actually, a person doesn’t really understand something
until he can teach it to a computer, i.e., express it as an algorithm… The attempt to
formalize things as algorithms leads to a much deeper understanding than if we simp
ly
try to understand things in the traditional way.

(Knuth, 1973)


Developing Aesthetic Measures for Three
-
Dimensional Head
Models




Page
5

of
11



a
esthetic tastes.
In attempting to answer
our research

question
,

we investigate what
constitutes
“aesthetic judgment


and how we make these judgments.
Many of the project team meetings
featured lively and heated discussions among team members concerning the definition of an
aesthetically pleas
ing head.
We all agreed on a base criterion that the model must appear
decidedly human
possessing no more than a

few
“minor” biological blemishes.
For example, we
all agreed that the models in
Figure
3

are not aesthetically pleasing. In viewing models without
obvious deformat
ions agreement came less frequently. With only three people, we quickly
discovered that gender
, age, experiences,

and culture play a role in attributing aesthetic value.
From o
ur
preliminary
review of
philosophy’s aesthetic literature, we learned of four

approaches
in determining aesthetic value: expression, representation, form, and transparency
(Goldman,
1995)
,
(Manns, 1998)
,
(Stiny & Gips, 1978)
. We decided to concentrate on
the
expression

approach

and
on the
aesthetic properties.



Aesthetic
Properties and
Expression

Consider our working definition of aesthetic
value:

A
esthetic value is a measure of information content. Something that has aesthetic value
carries information that evokes a response from a viewer when perceived.

T
he phrase “information content” is intentionally general.
While we agreed that aesthe
tic
properties featured information beyond simply physical descriptions, w
e were unable to settle on
a subset of specific aesthetic properties.
In our initial discussions, we considered how we could
measure a model’s degree of attractiveness. We all quic
kly agreed that attractiveness was
overly
broad, encompassing a large subset of properties that could be considered attractive in different
circumstances. For example, a cheerful person could be consider “extremely attractive” during a
dinner conversation

but “extremely disturbing” during a funeral
.

Table
1

offers

examples of
various
aesthetic properties

we discussed.


The features are presented with classifications and
associations

as
described in
(Goldman, 1995)
.




Figure
1
: Examples of Problematic Head Models

Developing Aesthetic Measures for Three
-
Dimensional Head
Models




Page
6

of
11




Our

working
definition
embraces the idea that aesthetic judgment is based on expression
. In this
approach,
our head models are to be evaluated according to the feelings and emotions invoked in
the user.

Stiny and Gips’

survey
(Stiny & Gips, 1978)

mentions four types of artistic expression:

E
1.) emotions/feelings of the

artist translated to the work

E
2.) the emotions/feelings invoked in the observer

by the work

E
3.) the work’s communication of emotion/feeling betwe
en the artist and observer

E
4.)
neither the artist nor the observer has the emotions/feelings.

Since our project seeks to simulate the aesthetic tastes of o
thers and does not focus on
implementing a
system
with
a sense of its on
aesthetic preferences
, we do not consider the work
to mod
el type E1 and E3.
The
re

is some argument on whether o
ur current work
models
E
2 or
E
4
.

Our work
ing

definition expli
citly matches the description given in

E
2.

Our proposed
que
stionnaire surveying prospective users justifies the E2 classification.
However, consider a
head model that depicts a

cheerful
persona. The virtual artist is not necessarily happy. The
model may or may not invoke happiness in the user viewing the model. The expression of
happiness is attrib
uted to the head model itself. Thi
s perspective offers a rationale for the E4
classification.

Continuing Work

As previously noted, we spent considerable time attempting to insure that the current system
does not produce extremely deformed head models

by implementing

geometric constraints
.

Our
next immediate goal is to implement

biological constrain
ts based on landmark locations. The
field of morphometrics provides a collection of shape tools based on landmarks
(Bookstein,
1991)
. In our continuing work, we will

implement this collection and investigate its relationship

Classification

Terms

evaluative

beautiful, ugly, sublime, dreary, “good”, evil, likable, trustworthy

formal

balanced,
graceful, concise, clumsy

emotion

sad, angry, joyful, serene

evocative

powerful, stirring, amusing, hilarious, boring

behavioral

sluggish, bouncy, jaunty, dominant, submissive

representational

Realistic, distorted, erroneous

perceptual

vivid, dull,
muted, steely, mellow

historical

original, derivative, daring, bold, conservative

Table
1
: Aesthetic Properties (classifications Alan Goldman
(Goldman, 1995)

page 17)

Developing Aesthetic Measures for Three
-
Dimensional Head
Models




Page
7

of
11



to aesthetics. It is also our intent to narrow our focus on designing “mini
-
functions” instead of a
general aesthetic value function. Each “mini
-
function” would measure the amount of a specific

aesthetic property in an individual head model.

The following is a list
of activities to be completed
:

1.


Transfer of project website to new host institution.

2.


Implementation of biological constraints based on landmark proportions.

3.


Deployment of
questionnaire.

4.


Reconciliation of landmark information from Texas Face Recognition Database to
landmark information in Aesthetic System.

5.


Verify thin
-
plate splines algorithm with data from Texas Face Recognition Database.

6.


Investigation and implementation of shape variables based on landmark positions.

7.


Exploration of the relationship between shape variables and the results of the
questionnaire.







Developing Aesthetic Measures for Three
-
Dimensional Head
Models




Page
8

of
11



Appendix A: Landmarks

Name of Landmark

Mesh
Point
Number


Description

Alare (al)

894

The most lateral point on the nasal ala.

Alare curvature

point
(ac)

733

The most posterolateral point of the curvature of the base of the
nasal alae, the lateral flaring walls of the nostrils.

Cheilion (ch)

36

The outer corner

of the mouth where the outer edges of the
upper and lower vermilions meet.

Crista philtre (cph)

122

The point on the crest of the philtrum, the vertical groove in

the
median portion of the upper lip, just above the vermilion

border.

Endocanthion (en)

11
74

The inner corner of the eye fissure where the eyelids meet, not
the caruncles (the red eminences at the medial angles of the
eyes).

Euryon (eu)

685

The most lateral point on the head.

Exocanthion (ex)

1133

The outer corner of the eye fissure where the

eyelids meet.

Frontotemporal (ft)

1038

The most medial point on the temporal crest of the frontal bone.

Frontozygomaticus (fz)

1579

The most lateral point on the frontozygomatic suture

Glabella (g)

1659

The most prominent point in the median sagital

plane between
the supra
-
orbital ridges.

Gnathion (gn)

236

The lowest point in the midline on the lower border of the chin.

Gonion (go)

1445

The most lateral point at the angle of the mandible.

Labial inferius (li)

9

The mid point of the vermilion borde
r of the lower lip.

Labial superius (ls)

215

The mid point of the vermilion border of the upper lip.

Labiale superius

lateralis (ls’)

122

The point on the upper vermilion b
order directly inferior to the
s
ubalare (sbal)

Developing Aesthetic Measures for Three
-
Dimensional Head
Models




Page
9

of
11



Maxillofrontal (mf)

998

The anterior lacrimal crest of the maxilla at the frontomaxillary
suture.

Nasion (n)

1354

The midpoint of the nasofrontal suture.

Ophyron (on)

207

The point at the mid
-
plane of a line tangent to the upper limits of
the eyebrows (sci
-
sci).

Opisthorcranion (op)

274

The most prominent posterior point of the occiput.

Orbitale (or)

1661

The lowest point on the margin of the orbit.

Orbitale superius

(os)

1666

The highest point on the margin of the orbit.

Palpebrale

superius (ps)

1662

The
highest point on the upper margin of the middle portion of
the eyelid.

Palpebrale

inferius (pi)

1660

The lowest point in the middle of the margin of the lower eyelid.

Pogonion (pg)

169

The most anterior point in the middle of the soft tissue chin.

Prona
sal (prn)

136

The most protruded point of the nasal tip.

Sellion (s)

1358

The deepest point of the nasofrontal angle.

Stomion (sto)

point buccal

16

The mid point of the labial fissure when the lips are closed
naturally.

Subalare (sbal)

879

The point on
the lower margin of the base of the nasal ala where
the ala disappears into the upper lip skin.

Subaurale (sba)

1025

The lowest point of the ear lobe.

Sublabial (sl)

159

The midpoint of the Labiomental sulcus.

Subnasal (sn)

174

The junction between the lower border of the nasal septum, the
partition that divides the nostrils, and the cutaneous portion of
the upper lip in the midline.

Superaurale (sa)

1206

The highest point of the free margin of the ear.

Superciliare (sci)

412

T
he highest point on the upper margin of the middle portion of
the eyebrow.

Developing Aesthetic Measures for Three
-
Dimensional Head
Models




Page
10

of
11



Tragion (t)

141

Located at the notch above the tragus of the ear, the
cartilaginous projection in the front of the external auditory
canal, where the upper edge of the cartilage di
sappears into the
skin of the face.

Trichion (tr)

213

Midpoint of the hairline.

Vertex (v)

334

The highest point of the head with the subject in the Frankfurt
horizontal plane.

Zygion (zy)

1632

The most lateral point on the zygomatic arch




Developing Aesthetic Measures for Three
-
Dimensional Head
Models




Page
11

of
11



Bibliography


Bookstein, F. L. (1991). Principal and Partial Warps: Components of Bending Energy. In
Morphometric
Tools for Landmark Data: Geometry and Biology

(pp. 315
-
338). New York: Cambridge University Press.

Faria, A. C., &

B. S. de Oliveira, J. (2006). Measuring aesthetic distance between document templates
and instances.
Proceedings of the 2006 ACM symposium on Document engineering

(pp. 13
-
21).
Amsterdam, Netherlands: ACM.

Gips, J. (1975). Shape Grammars and Their Uses: Ar
tificial Perception, Shape Generation, and Computer
Aesthetics.
Ph.D. Dissertation

. Birkhauser Verlag.

Goldman, A. H. (1995).
Aesthetic Value.

Boulder, Colorado: Westview Press.

Gupta, S., Castleman, K. R., & Bovik, A. C. (2010). Texas 3D Face Recognition

Database.

Gupta, S., Markey, M. K., & Bovik, A. C. (2010, June 05). Anthropometric 3D Face Recognition.
International Journal of Computer Vision

.

Hoenig, F. (2005). Defining Computational Aesthetics.
Computational Aesthetics in Graphics,
Visualization, a
nd Imagining (2005).

The Eurographics Association 2005.

Knuth, D. E. (1973). Computer Science and Mathematics.
American Scientist

, 61

(6), 709.

Machado, P., Romero, J., & Manaris, B. (2008). “Experiments in Computational Aesthetics: An Iterative
Approach
to Stylistic Change is Evolutionary Art”. In
The Art of Artificial Evolution: A Handbook on
Evolutionary Art and Music

(pp. 381
-
415). New York : Springer.

Manns, J. W. (1998).
Aesthetics.

Armonk, New York: M. E. Sharpe.

McCormack, J. (2008). Facing the Future: Evolutionary Possibilities for Human Machine Creativity. In J.
Romero, & P. Machado,
The Art of Artificial Evolution.

New York: Springer.

Sims, K. (1991). Artificial evolution for computer graphics.
Computer Graphic
s

(pp. 319
-
328). New York,
New York: ACM Press.

Stiny, G. S., & Gips, J. (1978).
Algorithmic Aesthetics: Computer Models for Criticism and Design in the
Arts.

Berkley, Ca: University of California Press.