Parsing visual scenes via dynamic cues(Parsing visual scenes via dynamic cues,Invited Iecture at the 24th Annual Meeting)

NII-Electronic Library Service TheJLiPaneselournalofIZsNchonomicScience

2006, VoL25,No.1,22-29

Lecture

Parsing

visualscenesvia

dynamic

cues

YuriOsTRovsKy,

Ethan

MEyERs,

and

Pawan

Massachusetts, institute

of

7lechnology*SINHA*

Objects

encountered in dailyvisual experience often consist of reglons that differin color,

luminance and shape, The

human

visual system is adept at binding together

these

various regions

to_perceive

the

whole object, while simultaneously separating them from those _that

_belong

_toother

objects or the

background,

How

this

region

integration

isachieved and

how

the visua] system

develops

these capabilities is not well understood.

We

recently

had

a unique opportunity to

investigate

this _question when we met an individual

(SK)

in

_India,who had an unusual visual

history,

At

the

time of our meeting,

SK

was

28

yearsold and had untreated congenital bilateral aphakia, resulting

in

highly _{compromised visual abilities, After providing treatrnenttoSK, we} studied the early stages of hisvisual skills. Specifically,we examined

SK's

_performance on simple image _parsmg and recognitjon tasks.We found that _prominent Figuralcues of _grouping, such as

good-continuation,

junction

structure and symmetry, were

largely

ineffective

for

image

_parsing.

By

contrast, motion _{cues were of profound significance and played two critical roles inSK's object}

processingabilities. First,they enabled intra-objectintegration,and segregation from

background.

Second.

they facilitatedthe

develepment

of object representations which _permitted recognition

in

static images.

Together

with data from earlier

infant

studies. these results suggest that motion

information

_plays

a fundamental role inorganizing early visual experience.

Key words: aphakia, object processing motion

information,

object segregation from

background

Individuals who _gain sight latein life_provide a

unique window

into

$everal aspects of visual

devel-opment,

Such

cases, however, are extrernely rare;

fewer than 20 have been studied

in

any detailover

the

course of

the

past

1OOO

years

(von

Senden,

1932;

Valvo, 1971; Gregory & Wallace, 1963;

Fine

et al.

2003).

Our

werk with SK _provides an opportunity to

add tothisimportant

but

sparse

body

oi work. We

focused our efforts on theissueof visual image

pars-ing.

Real-world

images

typically comprise many

re-gions of differentcolors and

luminances

_(Figure

_1).

Our visual systems are adept at integrating subsets

of theseregions intomeaningful entities. How _this

is

achieved

is

afundamental _question,which has been researched extensively

in

the domains of

experimen-tal and computational neuroscience

(Wertheimer,

1938;

Marr,

1982; Hummel & Biederman,

1992;

Ull-* _{Department of Brain and}

_Cognitive

_Sciences,

Massachusetts

Institute

of Technology,

Cam-bridge,

MA

02139

Correspondence

to

<psinha@mit,edu>

Copyright

2006.

man, 1996; Hupe et al.

1998;

Needham,

2001; Tu et

a].,20e3).

Much

ofthework

has

focused

on the use of

heuristicssuch asalignment of contours, and

similar-ityof texturestatistics

(Grossberg

&

Mingolla.

1985;

Mumford

& Shah, 1985;

Field

et al. 1993;Kovacs

&

Julesz,

1993;

Leung

&

Malik, 1998; August et aL,

1999).

In

circumscribed domains, these heuristics

can account rather well for

human

_performance

(Koffka,

1935;

Kanizsa,

1979;

Elder

&

Zucker,

1998L

but they tend to prove inadequate for analyzing

real-world

imagery

under lessconstrained settings

(Shi

&

Malik,

1997; Borenstein

&

Ullman,

2002).

Furthermore, itisunclear whether theseheuristics

serve toorganize

inforrnation

during

the

_{early stages} of visual experience. Determining thenaturc of cues active at thistime isimportant for elucidating

the

principles of visual

Iearning

and bootstrapping. With thismotivation, we undertook studies with

SK,

an individualwho afforded us a rare opportunity to

examine visual skills intheirearly stages.

SK i$a 29 _year old male,

born

in

Bihar,India. By

..-thetimc

SK

was

4

months old, members of

his

farnily

The

Japanese

Psychonomic

Society.

Allrights reserved,

NII-Electronic Library Service

Y.OsTRovsKy,

E,

MEyERs,

and P.SINHAi Parsing visual scenesvia dynamic cues

23

o

'

(a)

(b)

Figure 1. Natural

images,

such as the one shown in

(a)

are collections of many regions o

luminances,

as

indicated

in

(b),

The human visual systern has to accomplish the

subsets of theseregions

into

coherent objects, as

in

(c).

(c)f

different

hues and

task of

integrating

noticed that

he

had

an inabUity to

fixate.

Due

to

financial

and logisticalconstraints, medical

interven-tionwas not sought until SK was an aclolescent. At theage of

12

years,

SK

was examined

by

an

ophthal-mologist, who recornrnended surgery tocorrect

his

sight.

However,

the operation was cancelled due to

SK's

father

becQrning

$ick, which completely

de-pletedhis

family's

finances.SK was admitted to the

State

School for the

Blind

in Darbangha,

Bihar,

where he studied for 12yearsand

learned

Braille.

In

2000,

he

moved toa hostelforthe

blind

inNew Delhi

and enrolled

in

a correspondence course, which

earned him a masters

degree

inpolitical science in

August,

2005.

Itwas

during

a visit tothishostel

thatthe authors met SK

in

January

2004.

Examinations

by

three independent

ophthalmolo-gistsinNew

Delhi

yielded

identical

asses$ments. SK

has

secondary congenital

bilateral

aphakia

(Pratt

&

Richards,

1968;

Johnson

&

Cheng,

1997),

with the

lensesalmost completely absorbed

in

the anterior

and _posterior charnbers of the right and

left

eye

respectively, The optical pathways inthe eyes are clear.

SK's

acuity was assessed to

be

20/900.

SK

had

never

been

able to afford a _pairof eye-glasses that could compensate

for

his

aphakia. During our next visit to India,in

July,

2004, we had SK examined again

by

optometrists and ophthalmologists inNew

Delhi and _purchased a _pair of eyeglasses for him.

Post-correctionacuity was deterrninedte

be

20f120.

The residual acuity

impairrnent

islikelydue to neu-ralamblyopia

(Kiorpes

&

McKee,

1999).

Beginning

two weeks after the refractive

correc-tion,we performed a series of experiments to assess

SK's visual abilities.

Tests

of low-levelvisual

func-tion revealed that

SK

had

normal color and motion

'

perception,

We

then

proceeded to

investigate

his

image _parsing and object recognition skills.

Our

visual _parsing studies comprised seven expen-ments, which assessed

SK's

responses with images of simple shapes. Histaskwas tosay

how

many objects there were

in

each ofthe

images,

point towhere they were, and whenever _possible,name them,

Figure

2(a)

shows

the

specific tasks and representative stimuli

corresponding tothese experiments.

Further

cletails

are _provided

in

the methods section, The graph

in

Figure

2(b)shows

SK's

perfor,mance on these tasks

relative to thatof four control subjects, matched on socio-economic and educational

background.

For the contrels, viewing

distance

was increased to

incQrpo-rate theinformation lossresulting from SK'sreduced

acuity,

SK'sresponses on thisexperimental

battery

iteda consistent _pattern, He had no

difliculty

in

enumerating

individual

geometric shapes when pre-sented by themselves, or even in the _presence of other shapes, so long as they were non-overlapping

NII-Electronic Library Service

24

(a)

=

leoo2gs8 sofiEe8

(b)

The

_Japanese

_Journal

of Psychonomic

Science

Vol.

25,No, 1

i

O

Hovvmany objects?Howmanv oblects? ,{IX'･Mlge.,ii.ili.i-,l"....I

Howmany

objects?

,,ll,.,1il･11i･1/･ll/I...tttt/ulllltt:-.ttt

Hcrwmany objects?

,,lll/////i!'i'/'i'-g.･l,1/l,-'a'

Whichobject

isinfront? N/l":"..Nfn--N.Ik11krl.:f-l Tracethe longcurve

tt/ttttttt

Hewmanv

objeets?

ewContro[s

SK

di-fu

m

(c)

'

Figure 2,

_(a)

The

experimental battery we used to assess

SK's

image

_parsing skil]s.

(b}

SK's _performance,

relative to control subjects' on simple

image

segrnentation and shape analysis tasks.

(c)

SK's

tracing

(in

red) _{of a pattern} drawn

by

one of the authors, indicatinga

fragmented

percept of the input.

(experiment

`A').

However, with overlapped Figures,

presented either as

line-drawings

or as

fi11ed

trans-parent surfaces

(experiments

`B'

and `C'),

SK's

re-sponses were very differentfrom controls'.

He

per-ceived all closed loops,and regions of uniform lumi-nance as distinctobjects.

To

ensure that

he

under-stood the task,we

had

toldhim at the start of the

experiment that Figures may be overlapped

(a

no-tionhe was familiarwith

from

his

prior

haptic

ex-perience)and that

he

had to indicate the number

of "objects'

rather than `regionsl

Besides SK's reports with overlapped Figures,

therewere some other notable response _patterns as

well.

With

opaque overlapping shapes

(experiment

'D'},

SK

was able to correctly indicate _the number,

but

was at chance at

determining

their

depth-ordering

(experiment

'E').

_Extended contours made up of a series of separated linescgments, embedded

in

a

field

of randomly oriented ones

(experiment

LF')

NII-Electronic Library Service

Y.

OsTRovsKy,

E.MEyERs, and P.

'

SINHA: Parsmgvisual scenes via dynamic cues 25

along the contour was minimal. Inimages of

three-dimensional shapes, such as cubes er _pyramids

(cx-periment

LG'),

with surfaces of differentluminance,

SK

reported _perceiving rnultiple objects, one corre-sponding toeach facet.He was unable tointegrate

the

facets

into

thepercept of a single

3-D

object.

Insummary, SK's performance with thisstimulu$

set indicated a _profound

inability

to use cues of

contour continuation,

junction

structure and

Figural

symmetry to analyze the images presented.

SK's

tendency to_pcrceive theaforementioned stimuli

in

a

fragmented

manner

is

also refiected

in

his

tracingsof

simple

Figures.

Figure 2{c)shows histracings

(in

red)

over a Figure

drawn

by

one of theauthQrs.

Next we investigatedthe

functional

significance of

SK]s

atypical

image

parsing skills.

Given

his

pre-nounced tendency toover-fragment

images,

we

hy-pothesized that

SK's

ability to veridically segrnent

and recognize real-world

images

would be compro-mised,

To

testthis hypothesis,we assessed

his

nam-ing _performance on a set of 50 images of objects

common

in

the

Indian

setting.

SK

was able to

recog-nize only 20% of all images shown tohim. We asked

SK

topoint toobjects inseveral of these

images

and

also to indicate theirextent, even

if

he

could not

recognize the objects,

A

few

results are $hown in

Figure

3{a),

It

isevident that

SK

is

_greatly

over-segmenting thescenes and is_partitioningthem

into

meaningless regions, which are changeable and

unin-formative

regarding object identity.A robust object

representation isdifliculttoconstruct

based

on such

fragments.

SK's

parsing can belargelyaccounted

for

by a simple algorithm of

luminance

and hue-based

segmentation, The results obtained using such an

algorithm on the same images that

SK

saw, are

in-cluded in Figure

3(b).

Further detailsof the algo-rithm are _provided

in

the

Methods

section,

So

far,

we have described SK's performance

exclu-sively with static

imagery.

In order tomake our

experiments more representative of everyday visual experience, which typically

involves

dynamic

inputs, we created a new set of stimuli that

incorporated

motion cues

(Figure

4(a)).SK's task here was the

same as before. to

indicate

the number of objects

shown The individualshapes underwent small

inde-pendent

translations, For overlapping

figures,

the

extent of movement was such as to mamtam an

overlap at all times.

The inclusien

,of

motion brought about a

dramatic

change

in

SK's

responses, As data

in

Figure

4(b)

indicate,SK was able toproduce correct responses

for

a majority of the dynamic stimulL Motion also allowed

SK

toperceive shapes embedded

in

noise, a

task that

he

was entirely unable todo inthe static condition,

Motion

thus appears to be instrumental

forenabling SK tolinktogether _partsof an object, and segregate them from thebackground,

Beyond

facilitating

intra-object

integration and

object-background segregation, our results suggest that motion _plays another

important

role

in

SK's

object

_perception

skills. Itappears tohelp construct representations thatcan

be

used to recognize objects

instatic images,

We

had

mentioned above

that

in

an

image

set of 50 common objects,

SK

was able to

recognize only

20%

of

the

images shown.

It

is

inter-esting to consider whether there isan underlying

principlethat could explain the observed

partition-ing into`recognizable'

and `unrecognizable' sets,

In

examining

SK's

responses, an interestingpattern be-came evident.

As

shown

in

Figure

4(c),

the

partition-ing

corresponding to SK's responses

is

very similar tothe one obtained under the criterion of motility.

objects that move versus those that do not.

The

congruence isstatistically significant

to<O.OO05;

x2

test),

Before

considering the

implications

of

this

re-sult, we need toaddress theconcern that

it

might

be

driven

by some

low-level

image artifacts. This

would imply that the partitioning relates simply to

the

_particular

choice of images rather than the

eb-jects

shown therein.

To

test

this

possibility,we

created a separate set of

images,

which comprised a new exemplar foreach of the

50

categories included

in

the original set, The photographs showed the objects against natural backgrounds and under

nor-mal lighting.We assessed

SK's

object naming

per-forrnance

on thisnew set.

SK's

responses with

these

images

were consistent with those corresponding to

the original set

O<O.OO05;

x2

test),

suggesting that thenaming _performance islikely

being

driven

by the

im-NII-Electronic Library Service

26

The

_Japanese

_Journal

ofPsychonomic Science Vol.25,

No,

1

eeiwarv'i

'i,i/ilieewwlli/i･ll,/',,,,,2E,i-ll,ltt'k-ls:/;,#,?i･/ltmp.,,.,:i･:1/;/,

(a)

(b)

Figure 3,

(a)

SK's segmentation of real-world images. Outlined

in

green are

the

regions

that

SK

saw as

being distinctobjects. He was unablc torecognize any of these images.

_{b)

Segmentation results from a

simple algorithm

based

on

hue

and luminance-based _grouping

(see

methods section

fer

algorithm details), age artifacts.

A

plausible explanation of the congruence

be-tween the partitions derived

from

SK's

responses and

thatbased on the criterion of motility, isthatmotion of objects heips

bind

_theirconstituent regions

intQ

cohesivc representations, which can then

be

used to

recognize mstances m new mputs, that may wcll be statlc,

Taken together,theseresults point towards a

cru-cialrole of motion information

jn

the early stages of

object

learning.

Motion appears to help both in

segregating objects and also

in

binding their con-stituents

into

representatiens forrecognition.

Given

thisearly effectiveness of motion cues,

it

seems

plau-sible that motion might bootstrap the

learning

of

Figural cues oi grouping

(such

as contour

continua-tion and symmetry), which can then beused on _their

own to_{perform grouping} instatic

images

_(Cavanagh,

1993),

Consistent

with thispossibility, the

develop-mental _progression of motion _processing in_the

pri-mate

brain

begins

with a significant

level

of

func-tionalityright at birth

(Kiorpes

&

_{Mevshon, 2003,}

2004), Behavioral studies with human infantshave

shown that sensitivity to motion based cucs

for

grouping and other visual analyses arises wetl

be-tore

sensitivity tostatic Figural cues

{Arterberry

&

Yonas, 2000;

_Johnson,

2003),

Even inadult

'

ers, motion cues greatlyiacihtatenovel ebject

]earn-ing

(Brady

&

Kersten. 2003).

Furthermore,

damage te putative motion sensitive areas in the human

visual cortex has

been

observed

to

lead

_to

pro-found

impairments

in object integrationskills

(Da-masio. 1985).

These results from

SK

are subject tothe cautions

NII-Electronic Library Service

Y.

OsTRovsKy,

E.MEyERs, and

P.

SINHA:

Parsing

vi$ualscenes via

dynamic

cues 27

(a)

:"i?\

'

N.

-.-)}x

Namethe object

s

I

,j-lil

"

flssis.

-

g.

Name the object fi5loo96oge-U2 soraE6tvtu

(b)

e

ee

Controls

ma

SK

(c)Figure

4.

_(a)

Stimulus sets for assessing role

SK's

performance relative tocontrols'.

(c)

Le

panel:

SK's

recognit{on resuits.

i"herent

in

interpreting

any case-study, Our

adop-tion of a case-study

format

is

necessitated by the rarity of

SK's

condition. We deriveconfidence inthe

generalityof theseresults based on theircongruence with results from _previouslyreported case-studies of sight-recovery, and the infant

literature.

While the

earlier papers on sight recovery

in

adulthood

(Gre-gory

&

Wallace,

1963;

Fine et al. 2003>

did

not

specifically

focus

on the

individuals'

region integra-tionskills, they reported difuculties

in

natural image recognition

(`iWe

formed the impression that

he

saw

[the

natural scenes as]

little

more than patches of colour," Gregory

&

Wallace,

1963, _page 24)and

in

simple image _parsing

("[MM]

described

two

overlap-ping transparent squares as three surfaces with the

central square

in

front."

Fine et al,,2003,page

915).

Furthermore,

in

these past cases, as inthe present

one, motion sensitivity was evident soon after

treat-of dynamic

information

in

ft_panel:

Partition

obtained

SK's

object segregation skills,

(b)

via

the

criterion of `motiljty. _Right

ment. Inthe infantliterature,

it

has

been

shown that

babies'

perception of static visual scenes

is

_quite

fragmented. unti] approximately six months of age.

During

･this

_period,theirability tolinkspatially sepa-rated parts of a _partiallyoccluded object

is

driven

strongly

by

common motion

(Kellman

&

Spelke,

1983;

Johnson

et al. 2002),

Our results cornplement these elegant studies

in

significant ways, First,they show that region inte-gration via

Figural

cues isnot merely a maturational

process,unfolding with age, but rather a visually

driven

developmental one, Second, they explicate

the nature of _processing

in

the absence of occlusion,

when _purely

Figural

cues

(such

asspatially

contigu-ous collinear contours) can conceivably allow

for

grouping. And, third,they relate _grouping to

real-world object recognition

by

showing that motion

NII-Electronic Library Service

28

The

Japanese

Journal

of

Psychonomic

Science

Vol.25,No.1

regions thatcan serve as representations

for

recogni-tion innew, _possiblystatic,inputs, Taken together,

theseresults suggest thatdynamic

information

pro-vides a

key

organizing infiuenceon early visual proc-essmg.

Methods:

Image segmentation algorithm

The image segmentation algorithm used tocreate

the images

in

Figure

3b works as follows:

1)

The

image

istransformed

in

to

CIE

LAB

color

space.

2) Each pixel

in

the

image

is

added

to

a `region

table'thatspecifies theLAB color ofthe region

(which

isinitially

just

the pixel'scolor value),

and

the

region size

(which

is

initially

1).

The

pixelsof an image are added _{to the table}

in

randorn order.

3) A predefinedthreshold

T

is

set by theuser-this

istheone

free

_parameter

of thealgorithm,

For

the

images

shown inthis_paper,

T

was set to

be

10.

4) The algorithrn then

loops

through each entry

in

the region table,merging each region

R

with itsspatially adjacent regions

Ai

if

_the

distancebetween

R

and

Ai

in

LAB space isless

than thethresholdT. The LAB color entry in

theregion tableisupdated toa weighted

age oi the merged regions

based

on their

tivesizes.

The

_pixelcount entry isupdated to

reflect the totalnumber of _pixels

in

_the new

merged region,

5) ThealgorithmstopswhenthedistanceinLAB

space

between

all adjacent regions

is

_greater

than the threshold T,

Experimental methods

All experimental stimu]i were presented on a 15"

CRT

monitor with a display resolution of

1280

×

1024

pixels and 85Hz refresh rate. Each of the conditions comprised 10

distinct

trials,SK's viewing

distance

was not constrained during the testjng,

but

averaged 40cms. The individualobjects spanned, on average, 8 degrees of visual angle.

In

dynamic

dis-plays, objects' speed averaged

6

degrees/second.

All

presentationsstayed up until a response.

SK

volun-teeredhisparticipationand was not _paid,other than

being

compensated

for

_{transportation}costs. He was

free

totake asmany rest breaks as hewished

during

the course of thetesting,

Acknowledgments

The authors wish tothank SK, the

doctors

at

New

Delhi's

Shroff

Eye

Hospital and

Drs.

Richard

Held,

Scott

Johnson,

Elizabeth

Spelke

and Michael

Ober-dorfer. This work was supported

by

the Alfred

P,

Sloan Foundation, the

John

Merck

Scholars

Fund

and the National Eye Institute

_(NIH).

This

study was conducted as _part of Project Prakash-a recently

launched charitable and scientific endeavor whose

'

t

t

goal

is

to

locate

congenitally

blind

children

in

India,

'

and treatthose whose blindness

is

curable,

References

Arterberry, M.E.

&

Yonas

A. 2000. Perception of

three-dimensional shape specified by optic

flow

by

8-week-old infants.Ptercoption

&

RsychQPh3,sics,

62

(3),

550-556,

August,

J.

Siddiqi,K,,& Zucker,

S.

W. 1999.Contour

Fragment

Grouping and

Shared,

Simple Occluders,

Computer

Vision

and

image

Uitderstanding,

76{2):

146-162.

Borenstein,

E.

&

Ullman, S.2002,

Class-specific,

down

segmentation. A. Heyden etaL

(Eds.),

LNCS

2351,

Springer,

Verlag,

pp,

109-122.

Brady,

M.J.

&

Kersten,

D, 2003. Bootstrapped

learning of novel objects.

Ibuvaal

of

Vision,

3(6),

413-422,

http:1/journalofvision.orgl316/2L doi:

10.1167/3.6.2.

Cavanagh,

P.1993.

The

perception of form and

tion,Current

QPinion

in

IVlaurobiolagy,

3,

177-182.

Damasio,

A.R.

1985, Disorders of complex visual

processing: agnosia, achromatopsia, Balint's

drome.

and related

difficulties

of orientation and

construction. In:M,

M,

Mesulam

(Ed,),

_PrinciPtes

of

BehaviorutIVburotogy,Davis,Philadelphia,

Elder,

J.

and Zucker, S.W. 1998.Evidence

for

ary-specific _grouping in

human

vision, Vision

Re-search, 38(1},143-152.

Field,

D,,Hayes,A.

&

Hess,

R.

1993.

Contour

tionby the

human

visual system: evidence

for

a

local"Association

Field"

Vision

Res.

33(2),

193.Fine,

I.

Wade,

A.R,,

Brewer,

A,

A.

May,

M,G.

Good-man, D.F. Boynton,

G.

M.

WandelL

B.A. &

Ma-cLeod, D.I. 2003. Long-term deprivation affects visua] _perception and cortex. Ndture

IVeuroscience,

NII-Electronic Library Service

Y.

OSTROVSKy,

E,MEyERs, and P.SINHA: Parsing vi$ual scenesvia dynamic cues 29

6(9),

915-916,

Gregory,

R.

L.& Wallace,

J.

G,

1963. Recovery from

ear!y

blindness:

A case study.

Quarterly

fournat

of

PsychoZczgy.

Monograph

no.

2.

Grossberg,

S,

&

Mingolla,

E.1985.Neural Dynamics

of Perceptual

Grouping:

Textures, boundaries and

emergent segmentations, Percoption and

Psycho-Physics,

38(2),141-171.

HumrneL

J,

E,& Biederman,

L

1992,

Dynamic binding

in

a neural network forshape recognition.

Rsycho-togical

Review,

99,480-517.

Hupe,

J.

M,,

James,

A.C,,Payne,

B,

R.

Lomber,

S.G.,

Girard, P.

&

Bullier,

J,

1998. Cortical

feedback

improves discriminationbetween Figure and

ground by Vl,

V2

and

V3

neurons. Ndture,

394,

784-787,

Johnson,

B.L. & Cheng, K.P. 1997.

Congenital

akia: a clinicopathologic report of threecases.

nat

of

PediatricCiphthalmotagical

Strzibismus,

34(1), 35-39.

Johnson,

S,

P,,

Bremner,

J,

G.,

SIater,A.,Mason,

U.,

&

Foster,

K.2002.

Young

infants'_perceptionof unity

and forrninocclusion

displays.

foumaal

of

EmpenL

mental

Child

RsycholQgy,

81,

358-374,

Johnson,

S.

P.

2003,

Development of fragmented vs.

holisticobject _perception.In

G.

Schwarzer

& H.

Leder

{Ecls.),

The

devetoPment

of,lace

Processing

(pp.

3-17).

Cambridge,

MA:

Hogrefe

&

Huber,

Kanizsa.

G, 1979.

0rganization

in Vision:Essays on

Gestalt

Plarception,New

York:

Praeger.

Kellman,

P.

J.

&

Spelke,

E.

S,

1983.

Perception

of

tlyoccluded objects in infancy.

CQg?iitive

qgy,15,

483rc524.

Kiorpes,L,

&

McKee,

S,

P.1999.

Neural

mechanisms

underlying amblyopia.

Current

Qpinion

in

ology 9,480-486,

Kiorpes,L.& Movshon,

J,

A.

2003.

Neural limitations

on visual development

in

primates.

In

Chalupa,

L.

M.

&

Werner,

J.

S.

{Eds,),

The

Visuae

Nigurosctences,

(pp,

159-173}. Cambridge, Massachusetts: MIT

Press.

Kiorpes, L,

&

Movshon,

J,

A, 2004. Development of

sensitivity tovisual motion inmacaque monkeys,

VisualIVeuroscience,

21(6),

851-859.

Koffka,

K,

(1935),

f'le'inciples

of

Gestalt IlsycholQgl,,

New York: Harcourt, Brace and

World.

Kovacs,

I.

&

Julesz,

B,1993,

A

CIosed

Curve

is

Much

Morc

than an

Incomplete

One: Effectof

Closure

in

Figure-Ground

Segrnentation.

Proc.IVdt,Acad,

Sci.

USA, 90,7495-7497.

Leung,

T.

&

Malik,

J,

1998.

Contour

continuity

in

region based

image

segmentation. In FijthEuro.

Coof

Computer Vision,Freiburg,

Germany.

Marr,

D.

1982,

Vision:

A

comPutationat

investigation

into

the

human

rePresentation and

Processing

of

vis-ual

injbrmation,

IVlew}'brk:W. H.

Freeman

and

Company.

Mumford,

D.

&

Shah,

J,

1985.Boundary detection

by

minimizing functionals,

In

IEEE Conj1on ComPuter

Vision

and Patter?z

Recognition,

San

Francisco,

Needham,

A

20el. Object recognition and object

segregation in4.5-month-o]d

infants.

10urnal

of

perimental

Chitd

Rsychology,

78,

3-24.

Pratt,

J.

C.

&

Richards,

R.

D.

1968,Bilateralsecondary

congenital aphakia,

Arch,

QPhthalmology,

80{4), 420-422.

Shi,

J.

and Malik,

J.

(1997),

Norrnalized

cuts and

age segmentation, Conjlarence

in

ComPuter

Vision

and

Ptittem

Recagnition

(pp

731-737},

San

Juan,

Tu, Z.Chen,

X,,

Yuille,A.L,,& Zhu,

S.

C.

2003.

Image

Parsing:

Unifying

Segmentation,

Detection, and

Recognition.

I}roceedings

of

the IE[IEintl.Coof on

Computer Vision.

Ullman,

S.1996.

High-level

visien,

Cambridge,

MA:

MIT Press.

Valvo,

A,

!971. Sight Restoration coflerLong-Term

Blindness:

The

lh,oblemsand Behavior

Patterns

of

Visual

Rehabilitation.

American Foundation forthe

Blind,

Von Senden, M. 1932.

SPace

and

Sight:

The RircePtion

of

Space

and

Shmpe

in

the

Congenitally

BtindBqfore

and Afler

QPeratien.

Reprint,

Glencoe,

IL: Free

Press,

196e,

Wertheimer,

M.

1938. Laws of organization in

ceptual

forms

(partial

translation).In

A

Sourcebook

of

GestattRsychology,W. Ellis

_(Ed.),

(pp.

71-88),