Use of Computer Vision in a Modern Art Installation

brasscoffeeΤεχνίτη Νοημοσύνη και Ρομποτική

17 Νοε 2013 (πριν από 3 χρόνια και 8 μήνες)

69 εμφανίσεις


“15 SECONDS OF FAME”

Use of Computer Vision in
a Modern Art Installation

Franc Solina

Computer Vision Laboratory

Faculty of Computer and Information

Science

University of Ljubljana
, Slovenia

Motivation for this work





collaboration with the Academy of Fine Arts in
Ljubljana since 1995



new media, computer
-
based art installations
(internet, virtual galleries, video, mobile robots,
remote operation)



work of scientist and conceptual artist Ken
Goldberg, UC Berkeley (TELEGARDEN)




COMPUTER VISION + ART INSTALLATION = ?

Video cameras in art installations




wooden mirror (Daniel Rozin)



touch me (Alba d’Urbano)



liquid views (Monika Fleischman)







TECHNICAL LIMITATIONS:
precise positioning of the subject


In the future everybody will be

famous for 15 minutes
.”


Andy Warhol

Marilyn Monroe (Andy Warhol, 1964)

Image mediated culture



people like to look at themselves (mirrors, photos,


paintings, video)



vanity, self
-
discovery, self
-
assertion



a face in mass culture
-
> FAME



media attention
-

a mirror of the indivudual’s


self
-
perception




WARHOL: celebrity photo
-
> portrait



warhol
-
like portrait
-
> instant celebrity

Faces in computer vision



images of people



find people, identify them, determine their activity



video surveillance



face recognition <
-

FACE DETECTION

15 seconds of fame

Hardware

Digital camera

LCD monitor

computer

USB

Software

input photo

transformation

color filters

pop
-
art portrait

illumination compensation

learning

15 second loop


find faces +

randomly select one

Roadmap


color
-
based face detection


illumination compensation


pop
-
art color transformations


display
and ordering
of portraits over the
Internet


conclusions

Our original face detection

Simplified face detection 1

Simplified face detection 2


ADVANTAGES:
faster,
detected also faces from profile



DISADVANTAGES:
faces of dark complexion not detected,
other body parts can be detected

Eliminating the influence of
non
-
standard illumination


different from daylight illumination


color constancy/compensation methods



eestimate the present illumination


reconstruct the image under standard
illumination


run face detection algorithm



Color compensation methods


close to standard illumination


low time complexity


Grey World


Average surface color in the image is achromatic


Illumination estimation: average color


Mean gray value


Modified Grey World


Illumination estimation: each color is counted only once


White
-
Patch Retinex


On each image white surface is present


Illumination estimation: maximal color

Color compensation methods

NO

GW

MGW

RET

NO


original

GW


Gray World

MGW


Modified GW

RET


White
-
Patch Retinex

Color constancy methods


far from standard illumination


Color by Correlation


(1) LEARNING: Take images of the Macbeth color
checker under present illum. and under standard illum.


Use correlation to compute the transform. Parameters


(2) APPLY TRANSFORMATION

Color comp. + correll. method

NO

GW

MGW

RET

COR

NO


original

GW


Gray World

MGW


Modified GW

RET


White
-
Patch Retinex

COR


Color by Correlation


Face detection results #1

0
20
40
60
80
100
standard
incandescent
flashlight
neon
Without preprocessing
GW
MGW
RET
Face detection after GW

GW

Face detection results #2

0
20
40
60
80
100
white
yellow
green
blue
red
Without preprocessing
Color by Correlation
Face detection after COR

COR

Warhol’s celebrity portraits


segment the face from the background


delineate the contours


highlight some facial features (mouth, eyes, hair)


overlay with color screens



above transformations
-
> shape grammar


BUT: requires automatic segmentation into
constituent face parts

pop
-
art color filters

+
color
-
balance

+
random coloring

+
posterize

+
hue
-
saturation

+
color
-
balance

+
posterize

+
hue
-
saturation

=
17 universal filters

Display of portraits


4 smaller portraits


same filter


different configurations


1 big portrait


each with a different filter


horizontal flip


each time a different person


no detection
-
> last detected face with a
different pop
-
art filter


15 second counter

E
-
m
ail ordering of portraits

Ordering
system


Beside the portrait is displayed
an unique ID number


Sending e
-
mail to

15sec@lrv.fri.uni
-
lj.si


Sending the requested picture

Creating of the web page

The
g
allery of “famous” people

from the project web page:

black.fri.uni
-
lj.si/15sec

Audience interactions


people quickly realize that portraits of people
present at the moment are displayed


if several people are present, becoming famous
is elusive


subtle staging to get one’s most favourable
image on the screen


subdued competition for “media” attention


narcissistic and voyeristic use of the “electronic
mirror”

Exhibitions in art galleries


Forum Stadtpark, Graz, Austria, 19
-
26 Sep. 2003


Finzgar Gallery, Ljubljana, 14
-
26 Nov. 2002


8th International Festival of Computer Arts,
Maribor, 28 May
-
1 June 2002


Conclusions


well accepted by the audience


no visible interface


a group of people can interact at once


exact positioning of observers not necessary



at least one face should be found in the input image


-
> high percentage of true positive face detections


-
> percentage of true negative face detections can be low


a huge database for testing face detection is generated



The goal was not to mimic Andy Warhol’s portraits per se
but to play upon the celebrification process and the
discourse taking place in front of the installation.

From the first public showing