Correction of

Geometric Perceptual

in Pictures

Thesis by

Denis N Zorin

In Partial Fulllment of the Requirements

for the Degree of

Master of Science

California Institute of Technology

Pasadena California

i

Abstract

We suggest an approach for correcting several typ es of p erceived geometric distortions in computer

generated and photographic images The approach is based on a mathematical formalization of

desirable prop erties of pictures

Weprovide a review of p erception of pictures and identify p erceptually imp ortant geometric

prop erties of images

From a small set of simple assumptions we obtain p erceptually preferable pro jections of three

dimensional space into the plane and show that these pro jections can b e decomp osed into a p er

sp ective or parallel pro jection followed by a planar transformation The decomp osition is easily

implemented and provides a convenient framework for further analysis of the image mapping

Weprove that two p erceptually imp ortant prop erties are incompatible and cannot b e satised

simultaneously It is imp ossible to construct a pro jection such that the images of all lines are straight

and the images of all spheres are exact circles Perceptually preferable tradeos b etween these two

typ es of distortions can dep end on the content of the picture We construct parametric families of

pro jections with parameters representing the relative imp ortance of the p erceptual characteristics

By adjusting the settings of the parameters we can minimize the overall of the picture

It turns out that a simple family of transformations pro duces results that are suciently close to

optimal We implement the prop osed transformations and apply them to computergenerated and

photographic p ersp ective pro jection images Our transformations can considerably reduce distortion in wideangle motion pictures and computergenerated animations

ii

Contents

Abstract i

Intro duction

Motivation

Contributions

Persp ective Pro jection and Viewing Position Nomenclature

Summary

Perception of Pictures

Assumptions and Denitions

Perceptual Nomenclature

Dening Pictures

Assumptions ab out Perception

Robustness of Pictures

Perception of Ob jects in Pictures

Rectangular Ob jects

Spheres Cylinders Humans and Animals

Foreshortening

Other Prop erties of Visual Perception

Visual Field

Perception of Straightness and Curvature

Perception of Verticality

Linear Persp ective

Contents iii

Rectangular Ob jects in Linear Persp ective

Spheres Cylinders Humans

Straightness VerticalityTexture DensityMovement

Other Traditional Persp ective Systems

Summary

Formalization of Perceptual Requirements

Technical Requirements

Structural Requirements

Denitions and Preliminary Lemmas

Holes in Images

Lo ops Folds and Twists

Lo ops and Twists

Decomp osition

Nonstructural Requirements

Denitions

Zero Curvature Condition

Direct View Condition

Error Functionals

Optimization Problem

Construction of Pro jections

Previous Work

Perceptuallybased systems

Artistic systems

Technical Pro jections

Axially Symmetric Pro jection

Reformulation of the Problem

Limiting Cases

Lower Bound of the Optimal Values of F

Second Bound

Contents iv

Interp olations

Extentions

Implementation and Applications

Implementation

Deep and Shallow Scenes When Do We Need a WideAngle Pro jection

Applications of Correcting Transformations

Other Applications

Summary

Conclusions and Future Work

Conclusions

Future Work

App endices

A Review of the Exp erimental Studies of Perception of Pictures

A Robustness and Perception of the Picture Surface and Frame

A Rigidity in Motion Pictures

A Size and Distance in Pictures

A Orientation and Layout

A Rectangular Corners

A Foreshortening

A Visual eld

A Straightness and Curvature

A Verticality

ter of pro jection from a pro jection of a rectangle B Reconstruction of the cen

Bibliography

Chapter

Intro duction

Motivation

The pinhole photograph from Pirenne Figure demonstrates the main problem that is

considered in this thesis The sphere in the picture app ears to b e deformed and only the fact that

ellipsoidal architectural decorations are very uncommon in ancient buildings makes us assume that

the ob ject in the picture is indeed a sphere If we strip the picture of additional cues we are likely

to p erceive the ob ject in the picture as an ellipsoidlike shap e rather than a sphere

The shap e of twodimensional images in a photograph is determined by the pro jection of three

dimensional space into the plane pro duced by the camera For a and for most

standard this pro jection is the perspective projection The distortion in Figure is one of

several typ es of distortion that are quite common for wideangle p ersp ective images

There are several questions that we can ask ab out these distortions Why do es p ersp ective

pro jection result in distorted images Whydowe p erceive p ersp ective pro jection images as realistic

in most cases Howdowe obtain a picture similar to Figure which lo oks less distorted The

p ossibility of decreasing destortion is demonstrated in Figure

These questions are imp ortant for image synthesis b ecause most existing rendering systems

Contributions

Figure Left Wideangle pinhole photograph taken on the ro of of the Church of St Ignatzio in Rome classical

example of p ersp ective distortions from Pirenne

Figure Right Corrected version of the picture with transformation describ ed in Section applied

mo del some asp ects of photographic pro cess practically all of them use p ersp ective pro jection as

the viewing transformation

Answering the questions that wehave p osed requires analysis of some features of human p er

ception of pictures for example the p erception of shap e Such analysis can b e used to evaluate

commonly used pro jections and create new ones

The main goals of the work describ ed in this thesis were

to collect and systematize available data on the p erception of pictures of threedimensional

scenes

todevelop a framework for evaluation and construction of viewing transformation based on p er

ceptual considerations

to construct sample families of transformations in this framework

Contributions

We formalize a numb er of p erceptual requirements as mathematical constraints imp osed on

the viewing transformation

Weprove that a viewing transformation can pro duce p erceptually acceptable images for arbi

Viewing transformation is the computer graphics term for the pro jection of threedimensional space into the plane

Persp ective Projection and Viewing Position Nomenclature

principal ray projection(picture) projection(picture) plane center plane of projection picture picture α α center of center of the picture the picture

projection angle

direction of projection

direction of projection

Figure Left Central pro jection

Figure Right Parallel pro jection

trary scene geometry only if it can b e represented as a comp osition of a linear pro jection and

planar transformation for each disjoint part of the scene

Weintro duce quantitative measures for twotyp es of p erceptual distortion that can b e used

for evaluation of viewing transformations

We demonstrate that two imp ortanttyp es of distortion cannot b e eliminated simultaneously

We construct a useful family of axially symmetric viewing transformations whichallow the

reduction p erceptual distortion

We describ e a nonsymmetric extension of this family which allows further reduction of geo

metric distortion

Persp ective Projection and Viewing Position Nomenclature

We shall use a numb er of terms describing various characteristics of p ersp ective pro jections and

viewing p ositions There is a lot of variation in usage of some of these terms we try to use a

consistent nomenclature whenever p ossible even when describing results of authors who originally

used a dierentoneTerminology more sp ecic to perception can b e found in Section and

denitions related to more general typ es of pro jection are given in Section and Section

central pro jection Let C be a pointinR letp b e a plane such that C pFor anypoint

l to b e the line through C and xThenacentral projection with the x R x C dene

x

center C to the plane p is the mapping R n C p which maps anypoint x R to l p x

Persp ective Projection and Viewing Position Nomenclature

parallel pro jection Let p a plane v avector which is not parallel to the plane pFor any p oint

x R dene l to b e the line through x parallel to v Then a paral lel projection in the

x

direction v to the plane p is the mapping R p wqhich maps any p oint x R to l p

x

perspective pro jection Either central or paral lel projection of a part of the threedimensional

space into a plane the projection or picture plane

orthogonal pro jection A p ersp ective pro jection is orthogonal if the direction of pro jection is

p erep endicular to the plane of pro jection this denition applies b oth to the parallel and

central pro jections

principal ray In central pro jection the principal ray is the ray through the center of pro jection

hintersects the pro jection plane p erp endicularly whic

direction of pro jection For the parallel pro jection the direction of projection is a part of the

standard denition For the central pro jection we dene the direction of projection to b e the

direction from the center of pro jection to the center of the picture This denition is precise

for ob jects with central symmetrysuch as a rectangle For a picture of arbitrary shap e the

center can b e dened to b e the geometric center the center of mass

When wesay direction of pro jection we mean that the angle between the pro jection

plane and the direction of pro jection Figure Figure is

pro jection angle For a central pro jection with the center C wecho ose a characteristic line segment

AB in the picture passing through the center of the picture and endp oints AB on the

AC B For rectangular pictures b oundary of the picture Dene the projection angle to b e

we can use the horizontal or vertical line segment going through the center of the picture The

projection angle is not dened uniquely for a given picture even if it is rectangular If the ratio

of the minimal linear dimension of the picture to the maximal is not to o small the projection

angle do esnt vary much with the choice of AB

For the parallel pro jection we dene the pro jection angle to b e zero

pro jection distance For the central pro jection the distance from the center of pro jection to the

center of the picture is the projection distance

Persp ective Projection and Viewing Position Nomenclature

projection(picture) viewing point plane picture α center of the picture

visual angle

viewing direction

Figure

viewing p oint For mono cular viewing the viewing point is dened as the center of the of the

eye For bino cular viewing the viewing point is the middle of the line segmentbetween the

eyes

viewing angle The viewing angle is equal to the projection angle ofacentral pro jection from the

viewing p oint to the picture plane

viewing direction The viewing direction is the direction from the viewing p oint to the center of

the picture

viewing distance The distance from the viewing p oint to the center of the picture

visual eld visual eld is the solid angle with the ap ex at at the viewing p ointsuchthelightfrom

all directions within this angle reaches the eyes

eld of view If an ap erture is used for viewing a large part of the visual eld is o ccupied by the

ap erture T he remaining part is called the eld of view

Summary

Summary

We start with an analysis of the p erception of geometry in pictures Chapter There is no general

theory that explains all the asp ects of p erception of pictures Instead of using any particular theory

we describ e some minimal requirements of p erceptual realism that arise from everyday exp erience

and some exp erimental work in psychology This part of the thesis together with App endix A

attempts to provide a concise but sucient description of facts that are known ab out the p erception

of pictures The most imp ortant topics that are discussed in Chapter are the robustness of pictures

the p erception of common geometric shap es such as rectangular corners and spheres the p erception

of curvature and verticality and the p erception of rigidity in motion pictures

Perceptual requirements provide us with a set of constraints that we imp ose on the pro jection

Violation of some of these constraints results in unacceptable artifacts in the images suchasloops

in the images of lines or twisted images of planes Section Other constraints are more exible

and some deviation from the exact satisfaction of these constraints can b e tolerated An example

of such a constraint is the curvature of the image of a straight line it can b e nonzero for relatively

short line segments

The rst group of p erceptual constraints structural constraints allows us to limit the set of

viewing transformations that we consider Section Using additional practical considerations

we prove that any pro jection satisfying these constraints is a comp osition of a p ersp ective pro jection

and a twodimensional transformation

Some constraints in the second group cannot b e satised simultaneously the curvature of lines

and the distortion of shap e cannot b e eliminated completely at the same time Section We

intro duce error functions for each constraint whichcharacterize the quality of a pro jection with

resp ect to this constraintatapoint We can use the average or the maximum error to obtain an

overall characteristic of a pro jection

Using global error functionals for the direct view constraint and the zero curvature constraint

we formulate a parametric optimization problem which allows one to nd pro jections with minimal

global errors for a given tradeo b etween the twotyp es of the error Section

We consider a simple parametric family of pro jections Section and show that this family

approximates exact solutions of the optimization problem in the axiallysymmetric case well enough

for practical purp oses The resulting family of transformations has a very simple form

Summary

If weallow dierent tradeos b etween twotyp es of constraints in dierent parts of the picture

we can pro duce nonsymmetric pro jections which can b e b etter adjusted to the sp ecic scene con

guration The resulting extended family of transformations is still quite easy to use but is much

more exible

Our families of pro jections havetwo principal comp onents that can b e chosen indep endently

the p ersp ective pro jection and the twodimensional transformation We describ e how the choice of

appropriate pro jection is inuenced by the geometry of the scene Section

The fact that the rst part of our decomp osition of pro jections is the standard p ersp ective

pro jection makes it p ossible to implementthetwodimensional transformation as a p ostpro cessing

step of rendering It is also p ossible to apply this transformation to photographs if the

is known or if it can b e computed from the information in the picture Section App endix B

In the wideangle motion pictures varying distortion of shap e results in an impression of nonrigidity

whichmay b e undesirable our transformations allow to reduce this eect

Chapter

Perception of Pictures

One of the goals of image synthesis is to achievetherealism in pictures Realism is hard to dene

although most p eople haveanintuitive notion of realism Wewillcallthisintuitive notion perceptual

realismFor computer graphics applications a precise denition is required The most common

solution is to dene realism as equivalent to photorealism Photographs are considered to b e a

standard of truth in images and the mo deling of an idealized photographic pro cess b ecomes the

main task of the image synthesis This solution has a numb er of advantages but inevitably suers

from a fundamental aw the image synthesis algorithms are constructed and evaluated as mo dels

of a photographic pro cess while the output of these algorithms pictures is evaluated primarily by

a p erceptual pro cess If we use this approach the p erceptual quality of computergenerated images

is inherently limited by that of photographs

It is well known that an arbitrary photograph do es not necessarily lo ok go o d Although photog

raphers use some typ es of p erceptual distortions o ccurring in photographs to achievevarious artistic

eects in many cases distortions are undesirable and must b e avoided Avoiding these distortions by

purely photographic means choice of the viewing p oint or of the eld of view results in restrictions

on the scenes that can b e realistically repro duced in photographs Computer generation of images

is more exible and instead of faithfully imitating the photographic pro cess wecantrytouse

Assumptions and Denitions

p erceptual principles directly instead of mo deling a camera we can redene our working concept

of realism formalizing some facts ab out picture p erception Moreover photographic images can b e

pro cessed digitally and we can eliminate distortion in already exiting images

These goals require a more precise denition of the class of pictures that we are going to consider

after all any pattern on a at surface can b e a picture Section describ es the class of pictures

in whichwe are interested

There are three main sources of information ab out picture p erception available to us psy

chophysical research art history and theoryandoureveryday exp erience Quite a few facts ab out

p erception app ear to b e so obvious that nob o dy ever states them explicitly or b others to test them

exp erimentallyWe will see that some of these obvious facts are quite imp ortant Section

There is a considerable dierence b etween our approach to the problems of p erception and the

approach that is typically used in psychophysical research Our main goal is not to prove or disprove

some particular theory of p erception but to try to collect facts ab out p erception that can b e used

for making b etter pictures Wewishtoavoid relying on any particular theory as much as p ossible

We cannot avoid making some theoretical assumptions altogether we describ e our assumptions in

Section

Most exp erimental studies consider linear p ersp ective images but their results apply to anypro

jection pro ducing similar images Whenever p ossible we will emphasize twodimensional character

istics of images rather than the characteristics of the central pro jection used for their construction

Then we can extend our conclusions to any mappings from threedimensional space into the plane

Linear p ersp ectivenevertheless is extremely imp ortant it do esnt pro duce any distortion in most

cases and our main practical application is correction of distortions in the cases in whichitdoesIn

Section we discuss linear p ersp ective in greater detail

Assumptions and Denitions

Perceptual Nomenclature

One of the problems with using data from the psychological literature is the lack of consistent

and clear terminologyWe attempted to dene and use consistently a limited numb er of terms for

frequentlyused concepts and quantities Some of our denitions are nonstandard such as parallel

Assumptions and Denitions

and p erp endicular foreshortening Our motivation in these cases was to create an unambiguous and

concise way to describ e a particular concept

The word quantity here refers to numerical characteristics of threedimensional ob jects and scenes

represented in a picture suchasslant size distance or orientation The numerical value of a quantity

can b e

depicted A depicted quantity size slant distance etc is the actual value for the ob ject or scene

represented in the picture

apparent An apparent quantity is the value as p erceived or deduced from the picture Apparent

quantity cannot b e measureed directly it can b e estimated from the judged quantity or from

some other resp onse such as ballthrowing in Smith

judged A judged quantity is the value as rep orted by a sub ject in a go o d exp eriment it is deter

mined by the apparent measure but can b e biased bythemechanism of rep orting or by the

sp ecied task This bias however should b e the same for identical apparentquantities

The totality of apparentquantities can b e thoughtofastheperceptual spaceTheperceptual space

tity need not p ossess a consistent geometry in mathematical sense the value of an apparentquan

can dep end on the task attention fo cus and many other factors In most cases the perceptual space

of a picture is not connected to the real space around the observer A p ossible reason for this is

the scalability of the pictures in most cases familiar ob jects in the pictures are smaller or larger

than their real size and they cannot b e directly placed in the real space There are several notable

exceptions the p erceptual space of trompeloeil pictures Milman is merged with the real

space the orientation of ob jects in the p erceptual space can b e judged with resp ect to the viewer

Most names of quantities eg size slant distance are unambiguous The denition of orienta

tion requires some comments

orientation We shall distinguish b etween twotyp e of orientation relative to the viewer and relative

to the other ob jects in the picture We can estimate the angle b etween the viewing direction

and some direction sp ecied by an elongated ob ject in a picture We shall call this angle

orientation with resp ect to the viewer It is also p ossible to estimate the angle b etween two

directions in a picture We shall call this angle relative orientation

Assumptions and Denitions

Information ab out the geometry of threedimensional ob jects is partly contained in the two

dimensional geometry of images Most terms that we use to sp ecify the twodimensional geometry of

images are just general geometric terms eg angles lines curves areas etc A sp ecic characteristic

of images which is inherently twodimensional but cannot b e describ ed without referring to the three

dimensional ob ject is foreshortening For p ersp ective pro jections we shall distinguish twotyp es of

foreshortening paral lel and perpendicular

foreshortening If the size of the image of an ob ject dep ends on its p osition in the scene we shall

call this eect foreshortening

parallel foreshortening The decrease in the size of the image of an ob ject when it is placed at

increasing distances from the picture plane It is also called perspective convergence

p erp endicular foreshortening Supp ose it is p ossible to destinguish a feature of the ob ject p er

p endicular to the pro jection plane and another feature which is parallel to the pro jection plane

such as faces in a cub e or the lateral and frontal side of the head The dierence b etween the

ratio of the sizes of these features and the ratio of the sizes of their images is the perpendicular

foreshortening

It is p ossible to extend these denitions to the case of nonp ersp ective pro jections if wecan

somehow dene the distance to the pro jection plane which need not coincide with the physical

distance

It is imp ortanttomake a clear distinction b etween several related p erceptual phenomena

distortion Some images of familiar ob jects can b e identied as distorted if p erceptual information

in the picture is sucienttoidentify the ob ject with a high degree of condence yet some

tradict exp erience part of this information results in conclusions ab out the ob ject that con

Returning to the example in Figure one is reasonably condent that the depicted ob ject is

a sphere b ecause a sphere is more likely to b e a part of architectural decoration than a tilted

ellipsoid At the same time the shap e of the image suggests that the ob ject is not spherical

which results in a contradiction

misp erception Some images of ob jects can convey incorrect information ab out the ob jects without

causing distortion If an ob ject is part of a class of ob jects whichvary in size some images

Assumptions and Denitions

result in an apparent size which is close to the actual size or shap e while others will result

in a dierent apparent size We will call the later case misperception For example the right

monitor in Figure app ears to have an asp ect ratio signicantly greater than the actual

one no more than and than that of the left monitor the monitors are known to b e

identical

illusion Some pictures under sp ecial conditions can b e mistaken for the real scene that they repre

sent We shall call this condition il lusion

Figure Photo from the article Navigating Close to Shore byDave Do oling IEEE Sp ectrum Dec

c

IEEE photo byIntergraph Corp viewing angle The two monitors have the same size But the left one

app ears to b e wider misp erception

Dening Pictures

This section denes more precisely what wemeanby a picture It is also imp ortant to describ e

assumptions ab out viewing conditions the p erception of a picture dep ends on the p osition of the

viewer with resp ect to the picture presence or absence of ap ertures frames etc

In this thesis our main ob ject of study will b e at or nearly at pictures on pap er a pro jection

screen or any other at surface observed bino cularly without any restrictions on the p osition of

the head and without any sp ecial devices Pictures of this typ e include b o ok illustrations photos

p osters screen pro jections of slides motion pictures and pictures on computer displays Excluded

are stereograms of all typ es pictures that are designed for observation through a xed small ap erture

pictures in headmounted displays and anamorphic pictures

Assumptions and Denitions

Further we are interested in pictures that are representations of threedimensional ob jects There

are many dierenttyp es of representation from purely symb olic likeaverbal description or a kanji

character to a highly realistic photographic image which when viewed from a correct p osition

mono cularly can b e confused for a real ob ject Linear p ersp ective is used to some extent in most

pictures of the latter typ e

Some authors argued that essentially all forms of pictorial representation are based on convention

Arnheim Go o dman For example it has sometimes b een claimed that linear p ersp ective

is a matter of convention and p ersp ective images can b e understo o d only within the context of

a particular culture It app ears that such radical approach cannot b e accepted Crosscultural

Deregowski developmental Hagen a Hagen and Jones studies contain indications

that no culturedep endent learning is required for adequate p erception of p ersp ective images The

existence of various phenomena in p erception of pictures that we will discuss b elow also demonstrate

that there is a fundamental dierence b etween the p erception of most pictures and the p erception

of purely conventional representations such as text

Figure Examples of highly conventional pictures a road signa map a carto on

The presence or absence of a conventional comp onent in a picture is dicult to determine In

some cases Figure a it is easy to classify a drawing as almost purely conventional We will b e

most interested in drawings that dont use a symb olic metho d of representation or use it in a very

limited way This includes a wide range of pictures Figure

Assumptions ab out Perception

We will try to make as few assumptions ab out p erception as p ossible Wehave however to assume

that pictorial representation of information is mostly nonsymb olic Our assumptions can b e most

clearly formulated in terms of retinal images and internal representations

Assumptions and Denitions

a c

b d

Figure Examples of pictures that use symb ols in a limited way or dont use them line drawings a painting

Piano lesson by A Renoir a photograph Beckman Institute at Caltech

Our ability to recognize ob jects and textures that wehave seen b efore is crucial b oth to p erception

of the threedimensional world and pictures Wehave to assume

existenceofsomeinternalrepresentation of objects and a process that compares information

extractedfrom the retinal image with this representation

All the visual information ab out our environmentiscontained in the images formed on the retina

of the eye

Therefore

our internal representation of threedimensional objects should bebased on the information con

tained in these images

We dont know what part of the information contained in retinal images is utilized in this internal

representation

As retinal images are twodimensional central pro jections there is onetoone corresp ondence

between a retinal image of an ob ject and a pro jection of the ob ject onto a plane a linear p ersp ective

Assumptions and Denitions

picture of an ob ject Therefore we can consider geometry of p ersp ective pro jections instead of

geometry of retinal images

We can distinguish at least twotyp es of geometric information in retinal images structural and

nonstructural

Structural features have more qualitative nature and are the easiest to detect Examples of

structural information are the numb er of holes in an image of an ob ject dimension of the image or

its parts the numb er of edges meeting at a vertex Figure Mathematically this information

corresp onds to the top ological prop erties of the image We assume that

structural information about projections of an object is present in the internal representation of

the object If an image of an object has structural features that dont match those of any perspective

projection it is perceivedascontradictory distortedoritisnotrecognizedatall

a

b

c

d

e

Figure Examples of structural dierences in images a the numb er of holes b numb er of edges meeting at a

vertex c dimension d numb er of selfintersections lo cal structure e number of intersections of edges lo cal structure

Weshouldpoint out that our assumptions ab out the imp ortance of structural features are mostly

motivated by common sense and intuitive ideas ab out p erception due to the lack of exp erimental

Robustness of Pictures

data It is also imp ortant to note that the precision of the visual system is nite Finke and Kurtzman

and the ab ove statement is an idealization for example if a hole in an image is to o small it

is not detected by the visual system

Nonstructural features typically can vary without causing any signicantchange in p erception

Examples of nonstructural features are the degree of convergence foreshortening gradients of tex

ture angles and curvature of edges Figure The imp ortant dierence from structural features

is that variation in nonstructural features can b e registered by the visual system without pro ducing

considerable change in p erception

a

b

c

d

Figure Examples of nonstructural dierences in images a texture gradients b parallel foreshortening c p er

p endicular foreshortening d curvature

Robustness of Pictures

The term robustness of pictures was intro duced byKub ovy The p erception of a picture

typically do esnt dep end on the viewing p oint wecanwalk past a painting a b o ok moveour

head away closer or to the side of a computer display and the ob jects in the picture are unlikely

Robustness of Pictures

to change their shap e or p osition This is one of the most imp ortant prop erties of pictures were

it not true a viewing apparatus or correct choice of viewing p osition would b e required for correct

p erception Robustness is not total some p erceptual variables are less robust then others

Anumb er of particular p erceptual variables were conrmed to b e indep endent of the viewing

p oint

Previous studies Rosinski and Fab er Rosinski et al showed that for normal viewing

conditions p erception of slant do esnt dep end on viewing p osition This study was extended by

Halloran who showed that for extremely oblique viewing directions with the angle b etween the

viewing direction and the picture surface less then robustness of the p erception of slant breaks

down

There is some evidence Hagen bSmith bRosinski and Fab er Exp eriment that

relative size judgements are robust as well only two viewing directions and were used in

the rst study and only the viewing distance was varied in the last two

Robustness is very imp ortant for moving pictures for example inverse p ersp ective constructions

such as describ ed in Deregowski andLa Gournerie applied to the images of rotating b o dies

might result in nonrigid ob jects if the pro jection p oint used for reconstruction is dierent from the

actual pro jection p oint However this do esnt happ en in most cases As it was shown byCutting

for images of rotating nearrectangular solids there is a general tendency to ignore small

deformations esp ecially if they preserve angles b etween edges and that rigidity is preserved at least

up to a viewing direction

Relative orientations and layout are quite robust Halloran Exp eriment Goldstein

Exp eriment Goldstein Exp eriment although relative distances b etween ob jects in the

direction p erp endicular to the picture plane dep end somewhat on the viewing distance see discussion

b elow

Not all p erceptual variables are robust and the range of viewing p oints within which robustness

is preserved might dep end on the depicted ob ject

Wehave already mentioned that relative distance b etween ob jects in the direction p erp endic

ular to the picture plane is not very robust Exp eriments show that the viewing p osition aects

judgements of relative distance Smith bSmith a although the eect is less than predicted

by geometric reconstruction It should b e noted however that reduced viewing conditions were

Robustness of Pictures

used in these studies mono cular viewing through a p eephole and the absence of robustness can b e

attributed to the lack of information ab out the surface of the picture see b elow

The orientation of the ob jects with resp ect to the observer is not robust at all in some cases

It is the most apparent nonrobust p erceptual variable it was noted bymany authors that ob jects

p ointing p erp endicular to the plane of the picture such as gun barrels or ngers Figure app ear

to b e following the observer as he moves past the picture Portraits often app ear to b e following

the observer with their eyes It turns out Goldstein Goldstein Halloran that this

eect is more pronounced for orientations close to the p erp endicular to the picture and decreases

with the angle of orientation This results in the following p erceptual paradox while spatial layout

is p erceived as more or less invariant orientations of dierent ob jects change in dierentways for

example the rut in the road in Figure A do esnt rotate much while the direction of the road

rotates considerablyThus the dierence b etween apparent orientations changes while the relative

orientation of the rut and the road do esnt change This is an example of a more general phenomenon

the structure of p erceived space need not b e consistent dierent p erceptual mechanisms might

pro duce contradictory information without causing any p erceptual problems

Robustness of p erception may dep end on the contents of the picture For example parallel pro

jections of rotating rigid parallelepip eds are p erceived as rigid in a wider range of viewing directions

than central pro jections Cutting Exp eriments and

Robustness of pictures is related to the dual nature of pictures weperceive b oth the picture

surface and the threedimensional scene represented in this surface Surface texture atness and

visible frame are imp ortant factors in p erception of pictures Numerous exp eriments provide evidence

that the absence of these factors results in a decrease in the robustness

Perception of slant b ecomes completely unrobust when all the information ab out the picture

surface is removed Rosinski and Fab er Rosinski et al In unpublished exp eriments by

W Purdy Lumsden only the picture frame was absent and the information ab out the picture

surface was not completely removed Still for the simple pictures used in the exp eriments slanted

strip ed surfaces picture frame removal was sucient for almost total suppression of robustness

Perception of Objects in Pictures

Figure A World War I p oster from Taylor The nger app ears to b e following the viewer when he

walks past the picture Similar p osters were created in the US Uncle Sam wants you and in Russia Did YOU

volunteer during times when the government felt it necessary to p oint ngers at each citizen

Perception of Objects in Pictures

The distortions that wementioned in Chapter are asso ciated with images of particular ob jects

Two groups of ob jects which frequently app ear to b e distorted in photographs Kub ovy The

rst group included ob jects with rectangular threedimensional corners ie corners formed by

three edges with all angles b etween them equal to The second group included spherical and

cylindrical ob jects and humans

Rectangular Objects

Perception of pictures of rectangular corners is well describ ed by simple rules Perkins Perkins

Shepard

These basic facts ab out p erception of rectangular corners are usually called Perkins laws We

will call three line segments meeting at a p ointathreestar Perkins We will call an image of

Perception of Objects in Pictures

a rectangular corner twofaced if only two out of three b oundary surfaces are visible and threefaced if all three are visible Figure

Three-faced image Two-faced image

Figure Twotyp es of images of rectangular corners

Perkins rst law A threestar is acceptable as a twofaced image of a rectangular

corner if and only if it contains two angles less than or equal to whose sum is greater

then

Perkins second law A threestar is acceptable as a threefaced image of a rectangular

corner if and only if all three angles are greater than

Small deviations from Perkins laws can o ccur but in general there is a go o d agreementbetween

the exp eriment and theory Perkins Shepard

It is imp ortanttonotethatPerkins laws haveavery simple geometric interpretation acceptable

pro jections coincide with orthogonal p ersp ective pro jections There is no dierence b etween central

and parallel pro jections in this case b ecause only angles b etween lines are imp ortant For the

central orthogonal pro jection wehave to assume that the principal ray go es through the center of

the threestar

From Perkins data Figure A we can make a rough estimate of the deviations from Perkins

laws that dont result in considerable distortion the threestars deviating from Perkins laws by less

then still can b e p erceived as rectangular corners Of course this is a rough estimate We will

discuss Perkins laws in greater detail in Section

Perception of Objects in Pictures

Spheres Cylinders Humans and Animals

Perceptual requirements on the images of spheres are remarkably restrictive only disks are gen

erally accepted as go o d images of spheres Pirenne Kub ovy No data on detection of

noncircularity were available to us It app ears to b e safe to assume that asp ect ratio is de

tectable Figure and accept it as a rough upp er b oundary Spheres are not that common in real

environments although quite p opular in computer graphics images They are also convenient test ob jects b ecause the distortion of shap e in the image of a sphere is easy to detect and describ e

1.00 1.01 1.02

1.03 1.04 1.05

1.06 1.07 1.08

1.10 1.11 1.12

Figure Ellipses with dierent asp ect ratios the asp ect ratio is shown in the center of each ellipse

There are two distinct problems asso ciated with images of cylinders such as columns the one that

is most often mentioned arises when a row of cylinders is depicted It is not a problem asso ciated

with the image of any particular cylinder but rather a problem of unacceptable relative sizes of

images We will discuss it in the Section

The second problem can b e considered a separate case of a more general class of problems that

is asso ciated with axiallysymmetric ob jects

The image of the foundation of the column or the upp er part of a cup or a b owl is generally

acceptable if it is an ellipse with ma jor axis oriented p erp endicular to the axis of symmetry cf

Section Figure

Our p erception of humans is likelytobevery sp ecialized Humans come in many shap es colors

and sizes so the result of p erception of an inadequate image is more often a misp erception a

false conclusion ab out the p erson in the picture rather than direct p erception of distortion as with

spheres If we stretch or shrink the image vertically or horizontally within a wide range these

changes are likely to pro duce acceptable although sometimes misleading pictures For example the

Perception of Objects in Pictures

men in Figure are all identical but those close to the edges of the picture app ear to b e quite

dierent from those in the middle The part of the b o dy which is most aected by deformation is

the head due to less variation in the shap e of the head b etween individuals compared to other b o dy

parts

Figure Deformationsofhuman gures in a wideangle p ersp ective picture horizontally

There are typ es of distortion that are more likely to pro duce deformed not just misleading

images An imp ortant feature of acceptable frontal pictures of humans is their axial symmetry

When this symmetry is broken the image b ecomes unacceptable

Similar conclusions can b e reached ab out images of animals although larger deformations are

tolerated

Foreshortening

We consider twotyp es of foreshortening parallel and p erp endicular Our terms here are not quite

standard we b elieve it is useful to destinguish b etweeen these twotyp es of foreshortening

For rectangular solids the preferred amount of p erp endicular foreshortening dep ends on the

asp ect ratio Nicholls and Kennedy b it is almost constant for cub es the preferred amountwas

close to for all viewing conditions It should b e noted that the use of line drawings in this

study could considerably aect the results p erception of the whole ob ject could b e imp ortantin

this study the absence of any distinctive features in the drawing sp ecifying absolute size markings

Perception of Objects in Pictures

Figure Left pro jections of cub es from Pirenne

Figure Right pro jections of cylinders from Pirenne

on a matchbox or windows on a building could aect the preferred amount of foreshortening in an

unpredictable way

The amount of p erp endicular foreshortening deviating signicantly from the preferred one results

in distortion if it is known that the depicted ob jects are supp osed to b e cub es and in misp erception

if no such information is available Figure

Perp endicular foreshortening is easy to dene only for ob jects with distinct edges parallel and

p erp endicular to the picture plane it can b e also dened for familiar ob jects with distinctive features

parallel or p erp endicular to the picture plane Some distortions of the human gures can b e describ ed

in terms of the p erp endicular foreshortening ratio

Parallel foreshortening across ob jects can vary in a wide range without causing p erceptual prob

lems In p ersp ective images the ratio of parallel foreshortening is directly related to the viewing

angle the greater the viewing angle the greater the ratio for an ob ject of a given size Nicholls and

Kennedy a studied the eects of the change in the viewing angle on the p erception of cub es it

was found that a mo derate degree of parallel foreshortening approximately was consistently

preferred indep endent of the viewing conditions

Other Prop erties of Visual Perception

Other Prop erties of Visual Perception

Visual Field

The total size of the visual eld is quite large it extends more than horizontally and

vertically for eacheyeCarterette and Friedman However the density of the receptors is very

nonuniform and maximal resolution is achieved only in the fovea of the eyewhich has an angular

size of only ab out Lateral vision is very limited in resolution its main function is presumably

motion detection

As the resolution decreases less and less detail and precision is available to the rest of the visual

pro cessing system Finke and Kurtzman found that the size of the eld where gratings apart

can b e resolved is approximately For example the angular size of the mo on is Outside

eld of view we cannot detect any details of the lunar surface and cannot even tell if the mo on

is round or not

This also means that only lowresolution information in the retinal pro jections of the ob jects

that are far enough o the viewing direction can b e incorp orated into the internal representation

The same study Finke and Kurtzman provides some evidence that indeed the resolution of the

internal imagery corresp onds to the resolution of the actual visual eld

Perception of Straightness and Curvature

Intuitively it is clear that the images of straight edges and lines should b e straight It is imp ortant

to know the accuracy of the p erception of straightness Exp erimental studies that we knowabout

considered the p erception of straightness only for very short lines with angular size close to Itwas

found that for small p erturbations p erceptual threshold is determined by the solid angle subtended

y the maximal bump with resp ect to the leastsquares straight line Figure Watt et al b

The threshold value was determined to b e close to sq arc min

In Ogilvie and Daicar a similar measure was used although a chord rather than a least

squares line was considered and there was only one bump Resulting thresholds had the same order

of magnitude sq arc min These values are remarkably small Watt et al p oints out

that the receptor density is not sucient for the detection of p erturbations of that size and some

typ e of higherlevel pro cessing should b e involved to account for these thresholds hyp eracuity It

LinearPersp ective

app ears that this threshold is likely to increase with the angular size of the line However it is b ound

to stay quite lowandany considerable deviation from straightness will cause p erceptual distortion

Least-squares straight line Maximal bump

Curved line

Figure Watts measure of p erceptual curvature

Perception of Verticality

The p erception of verticality is inuenced bytwo main factors vestibular p erception of the force

of gravity and visual p erception Ob jects that we assume to b e vertical walls trees etc or

horizontal the surface of the ground determine the visual vertical A numberofexperiments Ong

and Kessinger DiLorenzo and Ro ck show that when there is a conict b etween vestibular

and visual information there is no clear preference for the ob jectivevertical determined by the

vestibular system The ro dandframe eect DiLorenzo and Ro ck indicates that a visible

frame aects p erception of actual vertical when a vertical ro d is viewed inside a tilted frame it

app ears to b e tilted in the direction opp osite to the direction of the frame

In pictures lines that are not parallel to the edges of a rectangular frame are p erceived as non

vertical If these lines representvertical or horizontal edges of ob jects buildings or furniture the

picture creates a feeling of instabilitywhich is sometimes undesirable The ro dandframe eect

explains to some extent the origin of this feeling

LinearPersp ective

Linear p ersp ectivewas intro duced as a systematic to ol during the Renaissance To some extent

linear p ersp ectivewas known to the Romans fresco es of Pomp eii and Chinese see discussion of

other p ersp ective systems in Section but during the Renaissance it was intro duced as a rigorous

system whichwas considered to b e the foundation of painting and drawing till the second half

LinearPersp ective

of the nineteenth century The rst artist to use p ersp ective during the Renaissance was Filipp o

Bruneleschi but the rst written description and analysis was done by Leon Alb erti in Alb erti

and Leonardo da Vinci in da Vinci He was also the rst to observe the limitations of the

metho d As it is discussed in Kub ovy Leonardo was not convinced that p ersp ective pictures

are robust hence his overly stringent requirements to the application of p ersp ective Leonardos

rule

do not trouble yourself ab out representing anything unless you takeyour viewp oint

at a distance of at least twenty times the maximum width and height of the thing that

you represent and this will satisfy every b eholder who places himself in frontofthework

at any angle whatso ever

This means that angular size of any ob ject in a picture should b e less than clearly an

overkill Leonardo oers the following example of p ersp ective distortion Figure in the image

ofarow of columns depicted from a close pro jection p oint the width of the images of columns closer

to the edges of a picture would increase which is quite counterintuitive

Figure Leonardos example of a p ersp ective distortion

Statements similar to Leonardos rule are quite common For example Glaeser doesnot

recommend using viewing angles ab ove Olmer recommends viewing angles no more

than horizontally and vertically

Thus the limitation of linear p ersp ectivewere recognized simultaneously with its discoveryAs

it is p ointed out byKub ovy it is not correct to identify artistic linear p ersp ective with central

pro jection of a halfspace onto a plane only a limited range of viewing angles is used and as we

will see the images of many ob jects are often painted with deviations from linear p ersp ective

LinearPersp ective

In photographs and computer generated images linear p ersp ective is practically identical to

central pro jection exact linear p ersp ective wideangle images are not uncommon and due to the

physics of the pro cess or nature of the algorithms resulting images follow the rules of the central

pro jection exactlyIntuitive comp ensation used by the artists is not available in this case Two

metho ds are helpful in the analysis of exact linear p ersp ective rst we can checkhowwell linear

p ersp ective images satisfy various p erceptually desirable requirements describ ed in the previous

sections second we can examine delib erate deviations from linear p ersp ective that are common in

painting

Structural features Linear p ersp ective images have no undesirable structural features see

Section

Robustness The studies of robustness were done using p ersp ective images Summarizing their

results wecansay that under normal viewing conditions p ersp ective images are robust when the

viewing angle is suciently small or the viewing direction is not to o oblique In general robustness

is a concern only for moving images

Rectangular Objects in LinearPersp ective

In p ersp ective images Perkins laws are always violated to some extent with the only exception of

orthogonal parallel pro jection The extent of this violation dep ends for the parallel pro jection on

the direction of pro jection and for the central pro jection on the direction from the center to the

ap ex of the rectangular corner We will call b oth directions directions of pro jection

Wehave selected two parameters to characterize the violation of Perkins laws fraction of

rectangular corners with pro jections violating Perkins laws for a given direction of pro jection and

maximal deviation of the oending angle from

There are innitely many p ossible p ositions of rectangular corners in space and each of them

o angles and determining the orientation of one of the can b e represented by three angles tw

edges and the angle of rotation of the remaining pair of edges around the rst one

Dene a triple to b e go o d if the pro jection of the corner in the given direction satises

Perkins laws

Then we can dene the fraction of go o d pro jections to b e

LinearPersp ective

A B A B C

C

Figure Threesided and twosided images of a rectangular corner

R

Z

d d d

all go o d

R

d d d

d d d

all good

all

We dene the oending angle in the following wayLetA B C b e the angles b etween the edges

of the threestar always less then Let A B C Then there are two p ossible violations of

Perkins laws either all three angles are less than or only one of them is less than C In

the rst case A B C and A should b e greater than while B and C can b e less than

In this case we dene A the largest angle to b e oending angle In the second case there are two

p ossibilities A B C or A B C In the rst case B should b e less than second

Perkins law applies and we dene it to b e the oending angle In the second case all angles should

b e greater than and we dene C to b e the oending angle

It can b e easily shown that our denition is equivalent to the following simpler one

the oending angle is the angle b etween edges of the threestar which is the closest to

Figure shows the fraction of go o d corners as a function of the angle b etween the pro jection

direction and the picture plane MonteCarlo evaluation It also shows the fractions of almost

go o d corners those with deviations less than degrees

Figure shows the maximal oending angle as a function of the angle b etween the pro jection

direction and the picture plane

We can see that all factors are negligibly small for small angles b etween the pro jection direction

and the pro jection plane From Perkins we can see that deviations of approximately degrees

are still tolerated in almost of the cases Therefore in most cases pictures with pro jection

angles that are small enough not to create distortions of more than can b e considered acceptable

LinearPersp ective

1 90

0.9 80

0.8 70

0.7 60

0.6 50

0.5 40 0.4

fraction of rectangular corners 30 0.3 less then 1 deg violation of Perkins’ laws less then 3 deg violation of Perkins’ laws less then 5 deg violation of Perkins’ laws 20

0.2 Maximal deviation from Perkins’ laws, degrees

0.1 10

0 0 0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90

Projection direction, degrees Projection direction, degrees

Figure Left Fraction of go o d rectangular corners vs pro jection angle

Figure Right Maximal oending angle vs pro jection angle

Hagen and Jones and Nicholls and Kennedy a examined preferred parallel foreshortening

for parallelepip eds and other typ es of prisms Again the general trend of their results is in the

direction of lesser pro jection angles although Nicholls and Kennedy a argues that a mo derate

degree of parallel foreshortening is preferred to parallel pro jection while Hagen and Jones

claims that parallel pro jection has the highest rating For reasons explained in our comments to

Hagen and Jones in App endix A webelieve that the former pap er is more reliable

The p erp endicular foreshortening in p ersp ective images of rectangles might also create problems

Figure Comparing to the results of Nicholls and Kennedy b we can see that the ratio

of foreshortening can b e quite far from the p erceptually acceptable range It should b e noted that

cub es are not that common and much wider range of foreshortening ratios can b e tolerated for

arbitrary parallelepip eds unless this is something familiarlike a computer terminal or a TV

Spheres Cylinders Humans

All nonorthogonal pro jections of spheres are ellipses The asp ect ratio of these ellipses is a useful

measure of deviation of the image from p erceptually acceptable Figure shows the asp ect ratio

of the image of a small sphere as a function of the pro jection direction Assuming upp er b ound

for the acceptable asp ect ratio we can see that the pro jection direction should not b e less than

pro jection angle

Images of spheres are not that common in painting When they are present they are depicted as

circles regardless of their p osition The most famous example of this typ e is The Scho ol of Athens

by Rafael Figure Pirenne La Gournerie

LinearPersp ective

2

1.9

1.8

1.7

1.6

1.5 aspect ratio 1.4

1.3

1.2

1.1

1 30 40 50 60 70 80 90

projection direction, degrees

Figure Asp ect ratio of the image of a small sphere vs pro jection direction angle

In the b eginning of this chapter we p ointed out the problem with rows of cylinders that is sp ecic

to the linear p ersp ective Another more general problem is the tilt of the axes of the image of a

crosssection Figure ma jor axis of the horizontal crosssection is always oriented towards the

center of the image The most apparent distortions o ccur in the images close to the lines through

the center of the picture tilted at

Images of humans in exact linear p ersp ectivemay lo ok grossly distorted b ecause of extreme

changes of the asp ect ratios and violations of symmetry Figure a No such asymmetry is

observed in the art human b o dies are typically drawn as if the pro jection p ointwas lo cated directly

in frontofthem

Another problem whichiswellknown to photographers is the distortion of the features of a

face resulting from high degrees of p erp endicular foreshortening

Straightness VerticalityTexture Density Movement

Linear p ersp ective images of straight lines are straight Linear p ersp ective is the only pro jection

that has this prop ertyKlein

Anumb er of problems are asso ciated with the p erception of verticality in linear p ersp ective

images Unless the picture plane is p erp endicular to the ground vertical lines converge or diverge

In some cases Figure this divergence creates desirable artistic eect in other cases Figure

it is undesirable Even small tilts of the picture plane create considerable distortion in wideangle

images partly due to the violation of the Perkins laws and partly to the divergence of the vertical line and the conict with the frame of the picture

LinearPersp ective

a

b

Figure The Scho ol of Athens by Rafael a General view b Detail compare the image of the sphere as

painted by Rafael with reconstruction of the central pro jection of the sphere p osition of the center of pro jection were

determined from the images of the columns and walls

Another problem is related to the change in the texture density texture elements of identical

size have more extended pro jections when they are close to the edges of the image This eect can

b e observed in Figure the tree app ears to b e stretched to the left partly b ecause of the texture

density gradient

For moving linear p ersp ective images computer animations cinema some of the problems as

so ciated with oblique and wideangle pro jections are amplied

For example as it is describ ed in Cutting rotating rectangular solids are more often

p erceived as nonrigid for oblique pro jections and wideangle p ersp ective Cutting attributes this

fact to the greater p ercentage of frames that dont satisfy Perkins laws for a given viewing p oint

Because of the changes in p erp endicular foreshortening many ob jects app ear to stretch or shrink

when they move across the eld of view

The height of remote ob jects suchasmountains or clouds ab ove the horizon changes when

Other Traditional Persp ective Systems

Figure Two views of New York CityfromFeininger it is interesting to note that the author oers the

second image as an example of rather nonortho dox picture far more interesting than the conventional view

Figure A picture of a building taken with a tilted camera from Feininger we get the impression of

lo oking up as if wewere standing in the street in front of the building

they move from the p eriphery of the picture to the center These changes are quite apparentin

wideangle pictures and in many cases are p erceptually undesirable Figure although correct

geometrically

Other Traditional Persp ective Systems

From the Renaissance until the second half of nineteenth century Western painting and drawing

was based on linear p ersp ective As wehavementioned b efore some deviations from its laws were

common and analysis of p erception provides some explanations for them Artistic p ersp ectiveis

mostly a to ol for depiction of relative p ositions and sizes of ob jects in space its role in depiction of

separate ob jects is much less imp ortant

Other Traditional Persp ective Systems

Figure Beckman Institute courtyard at Caltech photo by the author

In the Middle Ages Europ ean art was quite advanced but didnt makeuseofany pro jection

system in a consistent manner The same can b e said ab out Byzantine and ancient Russian art

Comp ositions created by the artists were often quite complicated and some facts ab out their or

ganization can b e useful to understand Other cultures pro duced evolved forms of pictorial art

indep endently

Perhaps Chinese and Japanese art is the most advanced system dierentfromWestern art Their

approach to the depiction of ob jects might dier in some details but similarities are considerable

the use of color and shading is quite dierent but geometrically there is little or no dierence Most

imp ortantly the approach to the depiction of space adopted in Oriental art is quite consistentand

somewhat dierent from the concept of linear p ersp ective

As it was p ointed out bymany authors for example Hagen Japanese and Chinese artists

typically used oblique parallel pro jection in their drawings As in Western art pro jection was used

mostly for depiction of relative p ositions of ob jects rather than for separate ob jects

Parallel pro jection is just a sp ecial case of p ersp ective pro jection with the pro jection p oint

Other Traditional Persp ective Systems

































Figure Tra jectories of p oints in a wideangle motion picture when the camera turns The pro jection angle is

in vertical and horizontal directions

removed to innity The main problem of this pro jection is that there is no way to combine images

of very large ob jects mountains and small ob jects p eople in the same picture This problem was

often solved in the following way parallel pro jection was used for each part of the picture but some

amount of p ersp ective deminution with distance was intro duced for ob jects lo cated further awayAs

a smo oth transition b etween these separate areas was imp ossible they were separated by symb olic

clouds

In a sense inside the continuous areas Oriental practice is more consistentthanWestern As

wehave p ointed outthe shap e of linear p ersp ective images dep ends up on its p osition within the

image and this fact is typically ignored bythepainters For parallel pro jection this is no longer true

therefore a rigorous construction similar to an aerial photograph is p ossible For oblique pro jections

however it means not the absence of the distortion of shap e but a constant distortion If the angle

between the direction of pro jection and the pro jection plane is greater than all typ es of

Other Traditional Persp ective Systems

Figure Chang Tsetuan c Going Up River at Chingming Festival Time Detail of a handscroll

Palace Museum Peking from Sullivan

distortion are quite small and can b e ignored Of course this is mostly theoretical observation and

no precise geometric constructions were ever utilized

Oriental artists didnt use linear p ersp ective not b ecause they were comp etely unfamiliar with

it A painting by ChangTsetuan dating as early as XI I century Figure clearly exhibits

p ersp ective deminution Sullivan quotes a Sung dynasty critic Shen Kua who criticized a

tenthcentury landscapist Li Cheng for his use of linear p ersp ective Why lo ok at a building said

Shen Kua from only one p oint of view Li Chengs angles and corners of buildings and his eaves

seen from b elow are all very well but only continually shifting p ersp ective enables us to grasp the

whole

Another interesting asp ect of Oriental art is o ccasional use of divergent p ersp ective for rectan

gular ob jects The same trend can b e even more consistently observed in the Byzantine and ancient

Other Traditional Persp ective Systems

Figure Eucharist XV century State Russian Museum from Raushenbakh

Russian art Figure It cannot b e attributed to the lack of skill other asp ects of images demon

strate very high skill and artistic talent In Byzantine art in particular it b ecame to large extent

a matter of convention But no religious or symb olic basis is rmly established for this particular

convention and instances of such images in Oriental and Persian art demonstrate that it mighthave

some p erceptual basis

Deregowski and Parker Deregowski et al argue that laterally displaced ob jects ie

those lo cated in the p eriphery of the visual eld are p erceived in this wayAnumberofother

explanations were attempted Raushenbakh

This phenomenon is mostly of historical interest b ecause it is likely that most contemp orary

p eople wouldnt nd these images to b e go o d representations of rectangular ob jects It demonstrates

Summary

that p otentially the conventional element in representation is very signicant it could b e that some

of these images were p erceived as quite acceptable bycontemp oraries

The organization of space in Byzantine and ancient Russian paintings is of greater interest in

some sense the general principle of the representation of space is similar to that of the Japanese

art The scene is sub divided into several parts without connecting elements Each part of the

scene is depicted separately and then they are overlayed The dierence from the Oriental art is in

considerable o cclusion and much less consistent size relationships

Summarizing our observations on dierent artistic cultures that have attempted to depict space

rather than separate ob jects we can observethattheyhave at least one p oint in common all

straight lines are depicted as straight lines This a ma jor restriction on the images of rectangular

ob jects Indeed in all systems parallel or central pro jection is used Divergent p ersp ectivecan

b e considered as central pro jection applied backwards A common trend can b e observed in the

depiction of space the whole comp osition is divided into several parts and each of this parts is

drawn indep endently The pro jection used for each part is close to parallel There are quite a few

exceptions to the ab ovebutifany system is ever used consistently it follows this pattern

It should b e noted that other cultures develop ed even more dierent systems of pictorial repre

sentation For example Hagen describ es the system that is used in the art of Northwest Coast

Indians These images app ear even less realistic than Byzantine art to a mo dern Western observer

ant to our discussion and are not particularly relev

Summary

The overview of the p erception of pictures that wehave provided in this chapter will serve as a basis

for the construction of pro jections allowing us to achieve in some cases p erceptually b etter results

than regular linear p ersp ective

In this summary we rep eat the main p oints of our discussion of p erception which are relevantto

the task of constructing pro jections of the threedimensional world into the picture plane

Images of ob jects should b e top ologically similar to a linear p ersp ective image of the ob ject

To b e p erceptually acceptable images of ob jects should satisfy certain criteria suchasPerkins

Summary

laws for rectangular corners have asp ect ratio close to for the images of spheres b ounds on

the p erp endicular foreshortening ratio for parallelepip eds cylinders and humans

Linear p ersp ectiveworks b est for smaller pro jection angles Estimates of the critical angle

vary but for pro jection angles less than we are guaranteed that there will b e practically no

distortion Within these b ounds most of the p erceptual criteria mentioned ab ove are satised

This fact is conrmed by the artistic practice in dierent cultures

For wider pro jection angles distortions may o ccur They b ecome quite ob jectionable when the

pro jection angle exceeds Some familiar geometric shap es rectangular corners spheres

cylinders human gures are esp ecially sensitivetosuch distortions Distortions are amplied

in moving pictures

Summary

Figure Funaki Screens Rakuchurakugaizabyobu right panel Tokyo National Museum from

Mason

Chapter

Formalization of Perceptual

Requirements

In this chapter we showhow it is p ossible to formalize p erceptual requirements describ ed in Chap

ter use them together with additional technical restrictions to restrict the class of pro jections that

we consider and derive p erceptual metrics that can help to evaluate pro jections from p erceptual

p oint of view

Technical Requirements

To narrowdown the area of the search for p erceptually acceptable pro jections we are going to sp ecify

several additional design considerations They dont haveany p erceptual basis and some of them

are quite restrictive however they make the task of constructing pro jections manageable Resulting

pro jections can b e applied to a wide class of images and the choice of correct tranformation can b e made simple

Structural Requirements

Wewant a parametric family of viewing transformations so that an appropriate one can b e

chosen for eachimage

The numb er of parameters should b e small and they must have a clear intuitive meaning

Construction of the family of pro jections should not dep end on the scene Matching the scene

to a particular pro jection in the family should b e achieved bycho osing an appropriate set of

parameters

Pictures should b e scalable any part of a picture can b e made a separate picture Therefore

no area in a picture can b e considered small enough to b e ignored

Structural Requirements

Structural distortions as describ ed in Section are in general more apparent than nonstructural

distortions It is not dicult to nd exceptions to this rule but given the additional condition of

scalability Section almost any structural distortion would b e highly ob jectionable

This idea allows us to use structural and nonstructural requirements in dierentways we p ostu

late that some structural requirements should b e satised exactly considerably reducing the variety

of pro jections that we consider Then nonstructural distortions can b e minimized within the class

of pro jections satisfying structural requirements

Even given the scalability assumption some structural requirements app ear to b e more imp ortant

than others if a requirement is asso ciated with unusual and complicated ob ject geometry it is less

imp ortant than those asso ciated with simpler and more common ob jects

In this section wecho ose several formalizations of certain structural requirements and apply them

to images of al l p ossible ob jects in al l p ossible p ositions wewant our family of pro jections to b e

indep endent of the ob jects that wewant to depict and wewant the resulting images to satisfy a

given set of structural requirements exactly

ts result in the same More sp ecicallywe will show that three sets of structural requiremen

structure for the pro jections

the bers of the projection should besubsetsofthebers of a perspective projection see Sec tion for the denition of a b er

Structural Requirements

It do esnt mean that any p erceptually acceptable pro jection necessarily has this structure It

rather means that if a pro jection do esnt have this structure there are ob jects that will have unac

ceptable images if they are in certain p ositions in space

For each set of structural requirements wewillprove the statementaboveby assuming the

contrary ie presence of curved or noncoplanar b ers and construct an ob ject that will havean

image violating some structural requirement

As the ob jects that we consider in Section are quite simple and common ob ject with planar

faces and straight edges it seems reasonable to assume that most mappings with curved b ers are

likely to pro duce considerable structural distortions for at least some scenes and to concentrate in

the rest of our study on the pro jections with the b er structure of a p ersp ective pro jection which

guarantees absence of structural distortions

Denitions and Preliminary Lemmas

We will use x y for the p oints in the domain of a pro jection a volume in D space and

for the p oints in the range a p oint in the picture plane

A denotes the b oundary of the set A

A denotes the closure of the set A

Denition By a line segment we mean any connected subset of a straight line

We will use the following notation for line segments

a b denotes the closed line segmentbetween the p oints a and b

a b denotes the op en line segmentbetween the p oints a and b

a b denotes the straightlinecontaining the p oints a and b

n m

Denition We wil l cal l a mapping P R R smooth if al l the components of this mapping

areatleast twicecontinuously dierentiable

Denition The set of al l points of the domain of a mapping that map to a xedpoint is cal led

the b er of the mapping at the point

If a b er is a line segment and is not a p oint we will call the straight line containing it the b er

in the domain of a mapping there is b er going through this p oint We will lineFor each p oint x

denote it F x

Structural Requirements

Denition A pro jection is a smooth mapping P V R R where V is an open

pathconnected domain in R with the rank of the Jacobian of P equal to in al l points of V

Denition We wil l cal l a continuous smooth curve I R where I is a line segment

an p erturbation of a continuous smooth curve I R if thereisahomeomorphism

I dieomorphism g I I such that the distance jg x xj is less than for any x

Denition A homotopy of a continuous curve I A to another curve I A in

n

A R is a continuous mapping h I A such that ht t and ht t for

al l t I

Denition Two smooth planar curves intersect transversally at a point x if their tangents

at this point arenotparal lel A smooth curve and a smooth surface intersect transversally if the

tangent vector of the curve is not in the tangent plane of the surface

A crossing is an intersection of twocurves that cannot b e eliminated by an arbitrarily small

p erturbation of the curves More precisely

Denition Two planar curves in an open set A R have a crossing inside A if they have

an intersection point in A and thereis such that for any perturbation of any of the intersecting

curves where the perturbed curves stil l intersect inside A

If twocurves intersect transversally in A they have a crossing in A Some nontransversal

intersections result in crossing some dont Figure

B A

abc d



Figure Examples of intersections a Transversal curves and have a crossing in any op en set containing the



intersection b Nontransversal curves and still have a crossing in any op en set containing the intersection c

 

nontransversal and have a crossing in A and dont have a crossing in B d nontransversal and dont have

a crossing in any set

Two closed curves are homotopic in a set if one can b e continuously deformed to the other

without leaving the set

The following denition is a mathematical expression for having a dierentnumberofholes

We will use without pro of the fact that a smo oth closed curve without selfintersections separates the

Structural Requirements

plane into two pathdisconnected sets one b ounded inside and the other unb ounded outside

The curve can b e retracted to a p ointby a homotopy in the inside area and cannot b e retracted

to a p oint in the outside area

Denition We wil l dene a hole in the image P O of an object O to bea pathconnected

set of points H that dont belong to P O if in P O there is a closedsmooth curve without self

intersections such that H lies inside the curve Two holes are dierent if thereisasmooth closed

curve without selfintersections such that one hole lies inside the curve and the other outside

Denition The convex hull of a set is the union of al l line segments connecting two points

from this set CHA denotes the convex hul l of the set A

Nowwe proveseveral lemmas that we will use later

are divided into two subsets S and Lemma If the points of a continuous curve R

S so that S S S S then thereisacommon limit point of S and S

Pro of

Supp ose for eachpointofS and S there is a neighb orho o d which do esnt contain anypoints

of the other set Let x S x S Then there is a continuous curvefromx to x Consider the

inverse images I S and I S It follows from our assumption that for eachpoint

in I there is a neighborhood not containing p oints in I Let the union of all these neighborhoods

be N and the union of neighb orho o ds constructed in the same way for S be N ThenN and N

are op en sets whichcover the whole interval Their intersection is an op en set that do esnt contain

any p oints of I and I therefore it is emptyI I Thus wehave a decomp osition of the

interval into two nonempty op en sets which is imp ossible as the interval is connected

Lemma If P V R R is a continuously dierentiable mapping with the rankofthe

Jacobian everywhereequal to then

anyber P of the mapping is a dimensional submanifold of R

the unit tangent vector to this manifold at any point x isacontinuously dierentiable function

of x in a neighborhoodofx for some choice of orientations

the curvature of the ber is a continuous function of x

Structural Requirements

The rst part of the statement immediately follows from Theorem in Warner Lets

prove the second and the third part

Consider a lo cal onetoone parametrization of P in a neighb orho o d of a p oint x

x x x x xt The tangentvector at a p oint x isx t P xt Dierentiat

ing this expression with resp ect to tweget

dP x

x t

dx

dP x

As the Jacobian has rank solution of this system exists and is unique up to a constant

dx

Therefore the unit tangentvector is dened uniquely by the system and the formula

x t

jx tj

dP x

Comp onents of are continous functions of the partial derivatives of the comp onents of

dx

As P is assumed to b e continuously dierentiable the derivatives are continuous functions of x

Thereforex t is a continuous function of x

As the curvature is simply the length of t it can b e proven to b e continuous by dierentiating

the expression for It is imp ortant to note thatx t do esnt dep end explicitly on t

Lemma If al l bers of a projection arelinesegments and any two bers arecoplanar the

fol lowing is true if any two ber lines intersect but do not coincide al l ber lines intersect at the

same point Otherwise al l ber lines areparal lel

Pro of Let f and f be twointersecting but not coinciding b er lines Let x V be a point

outside the plane p dened by f and f and f be the ber line of xAsf should b e coplanar to

f it can b e either parallel to it or intersect it But if it is parallel to f it is parallel to the plane

p and therefore do esnt intersect f But it isnt parallel to f either therefore f and f are not

coplanar Thus f should intersect f and f But it can intersect the plane p only in one p oint so

it should b e the common p ointoff and f

Consider a p oint x in the plane p and the b er line f corresp onding to it The same reasoning

applies to f and the pair f f

Structural Requirements

Supp ose f and f do not intersect Supp ose some b er f intersects f Ifitisnotinpthenit

cannot b e parallel to f and it do esnt intersect it which is imp ossible If it is in p then a b er line

f which is not in the plane should intersect it in twopoints f f and f f which is imp ossible

Holes in Images

Given our denition of holes description of example a on Figure obtains exact meaning

Condition H An image of an ob ject should not haveanumb er of holes dierent from

the numb er of holes for any p ersp ective pro jection of the ob ject

Lemma Given Condition H al l bers of the projection should be line segments

Idea of the Pro of Supp ose there is a b er which is not a linear segment We construct a

tub e around this b er in suchaway that no straight line can get through the tub e Therefore no

p ersp ective pro jection of the tub e can have holes in it The b er go es through the tub e without

intersecting the walls therefore there is at least one hole in the image of the tub e under P which

contradicts the Condition H

A2 x2 N

f y

T

A1

x1

Figure Construction of Lemma Pro of

Structural Requirements

Supp ose a b er f is not a line segment Cho ose a p oint x on the b er where the curvature is not

zero There is a neighb orho o d of the p oint x where the curvature is not zero b ecause the curvature

is continuous Lemma

As the b er f is a onedimensional submanifold of R with induced top ologywe can nd an

op en neigb orhho o d N of x in f which is dieomorphic to an op en interval and N N f where

f f

NisanopensetinR Wecancho ose N and N so that N is a ball

f

Consider the endp oints x and x of N IfN is chosen in the interior of f they should b e

f f

outside N As they are limit p oints of N they should b e on N These are the only intersections

f

of f with N

ver of f with op en balls centered at the p oints of f Let f b e the closure of N Consider a co

c c c f

Cho ose the balls so that for the p oints of N their closures are inside N Asf is b ounded and

f c

closed it is compact Therefore there is a nite sub cover Cover and there is suchthatanyball

f

c

of the radius less than centered at a p ointof f is inside Cover Let T Cover N

c f f

c c

T has the following prop erties T f N f T fx x gIfapoint y is inside T there is

f

apoint y of f such that jy y j AsN is a ball the convex hull CHT will b e contained in

f f

T will b e contained in N Therefore CHT f N N and CH

f

Consider the line segmentx x As the curvature of f in N is not zero ther is a p oint y on

this line segmentwhichisnotonf there are neigb orho o ds M of f and M of y whichdont

c f c y

intersect Cho ose small enough for T to b e inside M and for the ball of radius to b e inside M

f y

Dene the ob ject O to b e CH T Consider T N By construction the only part of T that

can intersect N should b elong to the b oundary of the balls B x and B x centered at x and

x Let A B x N A B x N For suciently small A and A are disjoint ie

T N is a disjoint union of A and A AsA and A are intersections of a ball and a sphere A

and A in N are circles Dene the ob ject O as O n T n A A A A Wethrow

T with N away T and the interior of the intersection of

Consider a p ersp ective pro jection Q with center outside O Letl b e a straight line going through

the center of pro jection a b er line of the p ersp ective pro jection and let b e the corresp onding

p oint in the picture plane Wewanttoshow that there are no holes in the image of O AsO is

convex its p ersp ective pro jection image is convextooanditiseasytoshow that it do esnt have

holes As QO QO no p oint that do esnt b elong to QO can b e a p oint of a hole in QO

Structural Requirements

Wehave to consider only the p oints of QO n QO If QO n QO the line l intersects O

but do esnt intersect O ie intersects O inside T

No p oint outside T can b e connected byacontinuous curvetoapointinT without crossing T

All the p oints of T except those that are in N are in O If the straight line l intersects O but

do esnt intersect T it should go through the p oints of A or A

Supp ose l intersects only one of these sets Consider a plane through l whichintersects N only

are on one side of this plane the only p oints of O that are on inside A Then all the p oints of O

the other side b elong to T Therefore for anycurveinO there is a homotopyin R that contracts

it to a p oint without intersecting l The image of O has the same prop erty with resp ect to the

image of l whichisapoint Therefore there is no closed curveinQO thathas inside and

cannot b elong to a hole in QO

Supp ose l intersects b oth sets Any line going through the balls of radius centered at x and

x go es through a ball of radius centered at y But this ball do esnt intersect T therefore there

isapointonl whichisoutside T and inside N Then the line should intersect T n N whichis

part of O Thus is in the image of O

ever has a hole Fib er f do esnt intersect We conclude that QO has no holes in it P O how

O by construction It intersects the set A at the p oint x Consider the image of A A A

is a circle therefore its image can b e considered as a smo oth closed curve Again if is suciently

small we can assume that the directions of the b ers going through the p oints of A are within a

small solid angle and are transversal to the sphere N The image of A A is onetoone and

the curve S P A do esnt have selfintersections

Clearly x P A is in the interior of this curve and therefore there is a hole in the image of

O

Threfore P O and QO have dierentnumb er of holes for anychoice of Q and Condition H is

violated

Lemma Given Condition H al l bers should becoplanar

Idea of the Pro of The idea is similar to the pro of of the previous lemma If there are two

noncoplanar b ers we can construct an ob ject with tub es through the ob ject along the b ers in

Structural Requirements

suchaway that one cannot see through b oth tub es at the same time There will b e at least two

holes in the image of the ob ject under P and no more than one under P

a 2 F(x 2) C ε

a’

b’

1 C ε b

F(x 1)

Figure Construction from Lemma

Pro of

Supp ose wehavetwo noncoplanar b ers f and f

As wehave proven all the b ers are line segments Cho ose a curve connecting two p oints on

the b ers f and f We can separate all the p oints on the path into two sets S thesetofall

p oints for which the b er F x is coplanar with f and S the rest of the path Using Lemma

we can nd a p oint x suchthatinany neighborhood of this point there are p oints from S and S

Let B x b e a ball contained in V Cho osing x tobeapointfromS and x to b e a p oint from

that are not coplanar and b oth intersect B x We can pick S we obtain b ers F x and F x

p oints a a F x andb b F x inside B x

Consider two cylinders C and C of radius and with axes a a andb b Let l and l be

resp ectivelyWe can cho ose and C line segments connecting p oints on the opp osite faces of C

in suchaway that for anychoice of l l they are not coplanar and C and C are inside B x

Appropriate choice of a a band b can also ensure that the faces of the cylinders are inside the

b oundary of the convex hull CHC C

Structural Requirements

Consider an ob ject O dened as CHC C n IntC IntC Any p ersp ective pro jection

of this ob ject has at most one hole a straightlinewhichintersects the convex hull but do esnt

intersect O should enter at one of the cylinders faces and exit at the opp osite face of the same

cylinder If there are two disconnected regions in the image in which b ers of the p ersp ective

pro jection b ehave in this waywehavetwo coplanar straight lines b ers of a p ersp ective pro jection

are coplanar that intersect CHC C only inside C and C resp ectively But by construction

this is imp ossible It can b e shown that each cylinder can pro duce no more than one hole in a

p ersp ective pro jection

On the other hand as a a andb b b elong to the b ers and the b oundaries of the faces of

P aandP b the image P O has at least two holes cylinders map to curves enclosing the p oints

By Lemma all b er lines are either parallel or intersect at one p oint Therefore b ers of P

are subsets of b ers of a p ersp ective pro jection

The main problem with the argumentabove is that the ob jects that we construct are rather

unusual and the resulting structural distortion is not very signicant it is not obvious when weare

supp osed to b e able to see through a curved pip e and when we are not In the following subsections

we will to consider simpler ob jects

Lo ops Folds and Twists

In this section we consider the following conditions which formalize Figure de but are more

restricitve We describ e this set of conditions mostly for historical reasons this is the set describ ed

in Zorin and Barr The conditions used in the next section formalize essentially the same

requirements in a b etter way

The main advantage of this formalization is that we dont have to deal with smo othness prop erties

and the pro ofs b ecome simpler

Condition L The mapping of any line segment l to its image P l is onetoone every

where or it is a p oint

This condition prevents lo ops in the images of lines It is more restrictive not only it do esnt

allow lo ops but also folds that is situations when the image is not onetoone but is still simply connected

Structural Requirements

Condition T The mapping of a subset of a plane m to the image P m is onetoone

everywhere or nowhere

This condition prevents twists in the images of planes

Lemma Given condition L al l bers of the mapping P are line segments

x 3

x2 -1 ξ) P ( y

x

1

Figure Construction from Lemma

Pro of

Consider a b er P Supp ose it is not a line segment Let x and x be two p oints of the

b er Then P x x

Let S P x x S P n x x It is p ossible to show that there is a p oint

x P such thatinany neighb orho o d of this p oint there are p oints of S and S Let B xbe

a ball contained in V the domain of the mapping We can cho ose x x and x P tobe

three noncollinear p oints in B x

Consider the image of the triangle x x x As the whole triangle cannot map into this

would imply that dimP contradicting Condition L there is a p oint y in the triangle such

that P y Consider the segmentx y y where y x x If P y then the

image of x y is not onetoone but contains at least twopoints If P y then the same

Thus the assumption that a b er can contain noncollinear p oints contradicts is true for x x

Condition L

Note that this lemma do esnt rely on the Condition T

Lemma Any two bers arecoplanar

Structural Requirements

x 1 y 1 x2 y 2 -1 P ( ξ ) 1 -1 ξ

P ( 2)

Figure Construction from Lemma

Pro of

Supp ose two b ers are not coplanar Cho ose two p oints on these b ers Consider a continuous

path connecting these twopoints that is contained in V Such a path exists b ecause V is assumed

to b e pathconnected As in the previous lemma we can separate all the p oints on the path into

two sets S the set of all p oints x for which the b er P P x is coplanar with P P x and

S the rest of the path In the same manner we can nd a p oint x suchthatinany neighb orho o d

of this p oint there are p oints from S and S Again let B x b e a ball contained in V Let P

x and P betwo noncoplanar b ers corresp onding to p oints in B x Consider two p oints

and x on P and y and y on P The image of x x is not onetoone it follows

from Condition T that the image of x x y is not onetoone anywhere Let be a pointin

the image As the intersection of the b er P with the plane of the triangle consists of more

than one p oint and the b er is a line segment the whole b er b elongs to the plane of the triangle

But the same thing will b e true for y y x Cho ose P x y As two triangles share

P x x y and the b er P shouldbeinthe the edge x y P y y x and

planes of b oth triangles ie in x y As it follows from continuity should contain x y This

is imp ossible b ecause x and y map to dierent p oints Therefore there can b e no two b ers that

are not coplanar

As b efore from the two lemmas that wehaveproven it follows that the b ers ofP are subsets

of b ers of a p ersep ective pro jection

Structural Requirements

p ’ 2 y2 x2 y2 l y2 y2’

M 2 c P P(c) ’ y1 P(a) x1 y1 l1 P(M 1) -1 P ( ξ ) a ψ P(M 2) M 1 ψ yy 11’

ξ ψ’

ψ



Figure Constructions from Lemma the image of the triangle y ay is twisted



Lo ops and Twists

The following statement is the precise expression of the idea that the image of a linesegment should

b e a curve with no lo ops although not necessarily onetoone

Condition L The image of a line segment should b e a simply connnected curve

Twists in the image of a plane can b e describ ed more precisely in the following way supp ose

an ob ject has a at face with several straight edges intersecting only at the endp oints In this case

no p ersp ective pro jection pro duces images of edges crossing each other at some p oint in the middle

unless they coincide This leads us to the following formalization

Condition T The image of a pair of coplanar line segments should have no crossing

points that are not images of an intersection p ointofthesegments

Again we prove the same twotyp es of lemmas as in previous sections

Lemma Given Conditions L and T every ber is a line segment

Pro of

Supp ose a b er P has nonzero curvature in a neighb orho o d of a p oint

Cho ose twopoints x x on the b er P inside this neighb orhho o d Consider two parallel

going through the p oints x and x resp ectively l l V Cho ose these line segments l and l

intervals so that x and x are their interior p oints and the plane through l and l is transversal to

all the b ers going through l and l

Structural Requirements

The images of the line segments l and l intersect at the p oint Suppose is an intersection

p oint such that there is no p ossibly onesided neighborhood of where the curves P l andP l

coincide

There should not b e a crossing Condition T so the intersection at shouldnt b e transversal

Consider a plane p through the line segment l which is crossed by P transversallyThe

b ers of the mapping P going through the line segment l map it into a smo oth curve in the plane

p which is a onedimensional submanifold Condition L Therefore the tangentvector to the

image of l is unique at all p oints As the plane p is transversal to the b er P wecan

in suchaway that for each b er going through cho ose a onedimensional neighborhood of x in l

apoint in this neighb orho o d there is only one intersection with the plane p this is p ossible due to

continuity of the tangentvector Lemma Restriction P j of the mapping P R R to the

p

plane p is dierentiable and the p oint x is nonsingular b ecause of the transversality assumption

Therefore in a neighborhood of x it can b e inverted and the inverse mapping will have the same

smo othness inverse function theorem Consider T P j It maps the neighb orhho o d N of to a

p

neighboorhood N of x the curve P l to a smo oth curve in the plane p As is an intersection

l andP l p oint such that there is no p ossibly onesided neighborhood of where the curves P

coincide the same is true for x T l andl in p Then there is a p oint x in N on T P l where

the tangentvector is not parallel to l Consider l which lies in the plane p and go es through the

p oint x The curve T P l intersects l at that p oint and is transversal to it Therefore the same

is true for the images P l and P l contradicting Condition T We conclude that anyintersection

p ointof P l andP l m ust have a neighb orho o d p ossibly onesided where the curves coincide

Picktwopoints and in this neighborhood Let y P l y P l

y P l y P l

Consider the planar quadrangle y y y y The images of the edges y y and y y coincide

by construction

ery p ointofl there is a planar neighb orho o d in the plane p of the quadrangle y y y y For ev

q

where the mapping P j is onetoone b ecause at the p oints of l b ers are transversal to the plane

q

by construction Due to the compactness of l there is an such that there are no suchpoints that

are closer than toy y l Pickapoint a on y y which is closer to l than Denote

p oints on y y andy y which are at the distance from y and y as b and b Denote the

Structural Requirements

quadrangle y y b b as M In the same waywe can construct the quadrangle M near l We

can cho ose M M and a in suchaway that P M P M and P y a P M

Consider the line segments y y and y a

The image of y y is a curve connecting P y andP y P y P y y crosses P b b

at a unique internal p oint P c b ecause the mapping of M to the picture plane is onetoone and

y y intersects b b at one interior p oint c P c y lies in the interior of P M and connects

two p oints on the b oundary P y a lies in the interior of P M and connects two p oints on the

b oundary

WecaneasilyshowthatM n y a is separated byy ainto twosets S and S with the

y there is a path connecting twopoints of M and lying inside M that do esnt following prop ert

intersect y a if and only if the p oints b elong to the same set Dene these sets to b e the sets of

p oints on the dierent sides of the line y a in the plane of the quadrangle Consider twopoints

s and s from dierent sets and a connecting path t Cho ose a unit normal vector n to the line

y a Let d be the vector from y to t Consider the scalar pro duct d n If it is negativefor

s it should b e p ositivefors Therefore there is a p ointonthecurve where it is zero continuity

and so the curveintersects y a

This prop erty transfers to P M and to P M due to the inclusion P M P M The

p oint P c b elongs to P b b As P b b is part of the b oundary and the only b oundary p oints

of P y a are P y and P a they dont intersect as P j is onetoone Therefore P b b

M

lies entirely in one of the sets P S orP S It also lies in the same set as P bbecause P b and

P b can b e connected by a path that do esnt cross P y a The p oint P y lies in the other set

y as y and b are in dierent sets and P P y We conclude that P y a should intersect

P c y at an interior p oint This intersection is a crossing takeinterior p oints s and s of P S

and P S lying on the curve P y a There are neighb orho o ds of these p oints that are contained

in P S andP S resp ectivelyThus if the p erturbation do esnt take s and s of the curveoutof

these neighb orho o ds and do esnt take the part of the curvebetween s and s out of P M then the

intersection is not eliminated This is the case for suciently small p erturbations b ecause the part

Therefore of the curvebetween s and s lies entirely inside P M The same is true for P c y

P y a and P c y violate Condition T and wegetacontradiction Therefore all b ers should

have zero curvature ieall b ers should b e line segments

Structural Requirements

Lemma Given Conditions L and T allbers arecoplanar

Pro of

Assume the that the opp osite is true In the same way as in Lemma we construct two

continuous parts of noncoplanar b ers f and f inside a ball which lies completely in V Cho ose

two p oints x and x on b ers f and f resp ectively

Cho ose a p oint x on x x sothatifwe draw a plane p through the b er going through x and

x x the b ers f and f wont lie in the plane

The tangentvector to the b er is a dierentiable function of the p ointin R Therefore the

angle b etween a xed direction and the b er F xt where xt is a parametrization of x x is

a dieren tiable function We can cho ose the direction so that it is not a constant Then there is a

line segment inside x x where the derivative of the angle with resp ect to t has a constant sign

Then on this line segment all b ers have dierent angles with the xed direction

d

A

n’ C n

B

Figure Unit sphere construction from Lemma

In addition we can cho ose x and x close enough to ensure that no b ers along x x lie in

the plane p which is p erp endicular to p and go es through x x The following construction on

the unit sphere where p oints corresp ond to the directions of the b ers helps to explain whywecan

makesuchchoice of x x p and p

The directions of b ers ll a closed area A on the unit sphere This area do esnt include the

p oint corresp onding to the direction of x x x and x b elong to dierent b ers Let d be a unit

vector corresp onding to the direction of x x with an arbitrary choice of orientation

Structural Requirements

p’ f2 pp’ x2 f 1 y2 y y1 T T(y 1) 2 x

x2

x1 p x1 x2 x1 y

1 T(y 2)

Figure Construction from Lemma the image of the quadrangle x x y y is twisted

 

Possible directions of the normals n and n of the planes p and p form a circle C in the plane

p erp endicular to d with arbitrary choice of orientation

Directions p erp endicular to d form a band B a union of the large circles of the sphere This

band do esnt include the circle of p ossible directions of p p b ecause A do esnt include dAsa

result the intersection of B and C can b e made arbitrarily small This intersection consists of two

connected parts coresp onding to sets of opp osite directions By making them suciently small we

can guarantee that the directions p erp endicular to those in C B are not in B Thenwe can cho ose

n in C B and n in C p erp endicular to n This will satisfy the conditions stated ab ove

Cho ose a halfspace with resp ect to the plane p and pro ject the parts of the b er contained in this

halfspace into pWe can always cho ose p in suchaway that these pro jections are on dierentsides

x the p oint corresp onding to n in the circle C should b e b etween the p oints of intersection of x

of large circles corresp onding to the p erp endiculars to the directions of f and f with C

The plane p is transversal to the b ers in some neighborhood N of x x Therefore the

mapping of this neighb orho o d to the picture plane will b e onetoone and smo oth On the plane p

wecancho ose a neighborhood N of x x which maps into the image of the neighb orho o d N

Cho ose the p oint y N on the pro jection of the b er f into the plane p and y N on the

pro jection of the b er f so that y and y are on one side of x x Cho ose the pairs of p oints

in suchaway that the intervals y x and y x dont intersect

Lets show that there is a neighborhood of x x such that all b ers that lie in the plane p and

intersect this neighb orho o d also intersect x x Supp ose for any there is a b er f that lies in

pintersects x x at a p oint outside x x and go es through a p ointwhich is closer than to

x x By construction there is a b er f intersecting but not containing x x that lies in the

plane pBycho osing small we can make the lines containing f and f intersect arbitrarily close to

Structural Requirements

x x ie inside V This is imp ossible b ers cannot intersect Therefore we can cho ose y and

y in suchaway that the only b ers in p intersecting the quadrangle y x x y intersect x x

The mapping T P j P maps this quadrangle to the plane p Consider the b er going

N

through y Ify is suciently close to x the angle b etween the b er going through y and the

plane p is close to the angle b etween f and p

The plane p sub divides the space into two halfspaces H and H LetH b e the halfspace that

p lies in the same halfplane with resp ect contains part of the b er f whose pro jection into the plane

to x x asy and y By our choice of p the pro jection of f H lies in the other halfplane

Therefore the pro jection into p of the part of F y which lies in H will b e close to the pro jection

of f Bycho osing y close enough to x x we can guarantee that it lies entirely in the same

halfplane of p as the pro jection of f and do esnt intersect x x As p p F y p T y

pro jects to a p ointonx x We conclude that T y isnotinH and is therefore in H Inthe

same wa ywe can show that T y isinH Thus T y and T y are in dierent halfplanes of p

with resp ect to x x T y y intersects x x atthepoint xastheber F x lies in p and

do esnt b elong to x x If there is a path inside the image which go es from a p oint of the image

on one side of the line x x toapoint on the other side there should b e a p oint z x x on

the path b ecause no p ointinx x y y is mapp ed to a p ointofx x outside x x Therefore

x y y is sub divided into two disconnected parts byx x the set x

Let D b e an op en neighb orho o d of T x x y y with x x n x x excluded It has the

same prop erty with resp ect to x x and T y y iscontained in the interior of D AsF x p

and F x p there are neighb orho o ds of the endp oints of x x inx x that dont contain

any p oints of of b ers that lie entirely in p Therefore we can picktwopoints x and x on x x

that dont coincide with x or x and such that for all x x x n x x F x p

Dene D D n x x n x x D has the same prop ertyasD Consider p erturbations of

x x and T y y Suciently small p erturbation do esnt movex x out of a band around

ve x and x into D andtheendpoints of T y y out of their neigh x x It also do esnt mo

b orho o ds that have emptyintersection with the band T y y Then p erturb ed x x retains the

dividing prop erty for D and the endp oints of T y y are still in dierent sets Therefore the

intersection is preserved under small p erturbations and is a crossing This contradicts Condition T

We conclude that the b ers should b e coplanar

Structural Requirements

This is sucient Lemma to prove that the b ers of the pro jection P coincide with the

b ers of a p ersp ective pro jection

Decomp osition

Wehaveproved in previous sections that various sets of structural requirements result in the following

prop erty for pro jections

Theorem Given Condition H from Section or Conditions L and T from Section

or Conditions L and T from Section a projection P satisfying these conditions should have

bers that are subsets of the bers of a perspective projection

From this theorem it immediately follows that we can decomp ose the pro jection P in twoways

describ ed in the following

Corollary Assume that the conditions of Theorem hold Let betheperspective projection

coresponding to P

is a central projection with center C Assume in addition that the region V lies entirely in

ewithrespect to a plane going through the center of ThenP can bedecomposed one halfspac

in two ways

asacomposition of a central projection into a plane fol lowedbya smooth onetoone

plane

transformation T of the plane or

plane

asacentral projection into a sphere with center C fol lowed by onetoone smooth

spher e

mapping T of the sphere into the picture plane

spher e

is a paral lel projection Then P can bedecomposedasacomposition of an orthogonal central

projection into a plane fol lowed by a smooth onetoone transformation T of the

plane plane

plane

It is imp ortanttopoint out that the theorems were proven for a pathconnected domain V

Therefore if all the ob jects in a picture that are sub ject to the structural conditions describ ed ab ove

can b e separated into sets with each set lying in a separate pathconnected domain for example

foreground middleground and background a dierent p ersp ective pro jection can b e chosen in the

Nonstructural Requirements

decomp osition for each set It is interesting to see actual realization of similar principles in some

Western and Japanese paintings Section Section

Nonstructural Requirements

As wehavementioned in the b eginning of this chapter we will use nonstructural requirements to

cho ose pro jections in the class dened by the structural requirements Due to the more quantitative

nature of these requirements the appropriate way to formalize them is dierent Rather than

dening which images satisfy the requirement and which images dont wedene error functions

that give us a measure of a nonstructural distortion Formally sp eaking these are functions on the

set of all p ossible images Functions of this typ e are dicult to deal with Therefore it will b e more

convenient to dene error functions on the p oint of the picture plane in suchaway that the error

function at each p ointgives an estimate of the maximal distortion of a small image at this p oint

In this section we consider only pro jections on a pathconnected op en domain V satisfying

Theorem Therefore for each pro jection we can dene the center of pro jection unless all b ers

are parallel and decomp ositions as describ ed in Section We will assume that the b er structure

of the pro jection is xed ie pro jections and in the decomp ositons are xed

spher e plane

Denitions

Lets establish some notation for the pro jections that satisfy the conditions of the theorem

plane

W R p

V R plane

H

H

P

H

T

spher e plane

H

H

Hj

T

spher e

W S

spher e R

We will consider pro jections P V R R from an op en connected domain V into

the picture plane which are comp ositions of the pro jection from the center O into the

plane

intermediate plane p and a mapping T W R R W V Wecan

plane plane plane plane

cho ose the plane p so that that the distance from O to the plane is L

P can b e also represented as a comp osition of central pro jection from O into the sphere of

and some mapping radius L with center at O intermediate sphere V R S spher e

Nonstructural Requirements

T W S R W V We will assume that the image of V in the sphere

spher e spher e spher e

b elongs to a hemisphere

If the b er structure of the pro jection P is that of a parallel pro jection we will use only the rst decomp osition

picture plane π

η ψ ρ intermediate sphere (η,ξ) ξ O

ζ (θ,ζ) L θ p x y φ r (x,y) intermediate plane

(x,y,z)

Figure Co ordinate systems

Lets intro duce rectangular co ordinates xy and p olar co ordinates r on the plane prect

angular co ordinates and p olar co ordinates in the plane On the sphere we will cho ose

angular co ordinate system and lo cal co ordinates in the neighborhoodofaxedpoint

u L v L sin Figure

The corresp ondence b etween T and T is given by the stereographic pro jection S R

plane

r tan

Zero Curvature Condition

The requirement from Section is p erhaps the easiest to formalize

Zerocurvature condition Images of line segments should have zero curvature at each

point

It would b e desirable to use a measure of deviation from this condition similar to Watts measure

Section Watts measure however is dened for a particular viewing p osition while our

Nonstructural Requirements

measure has to b e indep endent of the viewing conditions A natural candidate for such measure is

the curvature of the image of a straight line at a p oint

As the curvature has the dimension of inverse length we need a scaling factor to make the measure

indep endent of the size of the picture As a scaling factor wecancho ose some characteristic length

The most appropriate choice app ears to b e either a typical size of a line segment that can b e found

in a picture or the size of the picture The later choice gives a conservative measure assuming the

worst case there are images of straight lines crossing the whole eld of view

Lets qualitatively compare this measure with Watts measure Watts measure dep ends on the

viewing conditions and therefore wehavetomake assumptions ab out these conditions Due to the

spatial limits of acute vision describ ed in Section we can assume that the picture o ccupies no

more than of the eld of view Most pictures are viewed from a distance when their angular

size is not less then Given these estimates wehave the viewing distance l equal roughly to

to times the size of the picture Let l k R where R is the size of the picture The area

between a chord of length d and a segment of a curve with radius of curvature r is to the rst order

of approximation d r Assuming d to b e of the order of magnitude of the size of the picture

d k R where k and close to one by the order of magnitude The Watts functional can b e

written as C k k R r where C is less than unity b ecause we usedachord instead of a least

squares lines Therefore it is roughly prop ortional to the curvature times the size of the picture

with the co ecient of the order

If we consider the decomp osition P T we can observethat satises the

plane plane plane

zerocurvature condition Therefore wehave to consider only the mapping T We will denote

plane

comp onents of T x y whichisapoint in the picture plane by x y x y

plane

The curvature dep ends not only on the p oint but also on the direction of the line whose image

we are considering As an error function for the zerocurvature condition at a p oint x we will use

an estimate of the maximum curvature of the image of a line going through x

Lets estimate the maximum curvature of the image of a line dened by x v t in the intermediate

pro jection plane in the p oint T x

plane

d

where d is the change in the direction of the tangentvector of the curve By denition

ds

along the segment ds

d

Let w b e the tangent of the curve t w Then dt

Nonstructural Requirements

dw dw dw

jw j w

dt dt dt

jj

jw j jw j jw j

In our case the curve is dened by the equation tT x v t To estimate the curvature

plane

dw

j j

jw j

dt

and the minimum of for a xed p oint x from ab ove we will estimate the maximum of

jv j jv j

d

Using F x v trF v for a scalar function F we can write

dt

dw dv dw t

rw v rw v

dt dt dt

d d

jrw jrw j jv j jr j jr j jv j

dt dt

jrr v j jrr v j jv j jv rr j jv rr j jv j

j jv j je rr j je rr j e rr j e rr

x y x y

Expanding je rr j j j j j and other terms in similar waywe get

x xx xy

q

dw t

j j j j j j j j j j j j jv j

xx yy xy xx yy xy

dt

Now lets estimate jw j from b elow

jw j r v r v

v v v v

x y x y x y

y y y x x x

Av Bv v Cv

y x

x y

We can nd suchthat

w w

u v

p p

cos sin

w w w w

u v u v

Thus

jw j A cos B cos sin C sin

Nonstructural Requirements

Minimizing this expression wrt weget

p

A C B jv j jw j A C

Finallywe get the estimate for the curvature

q

j j j j j j j j j j j j

xx yy xy xx yy xy

p

jj

A C A C B

To simplify calculations we will use the square of the curvature

j j j j j j j j j j j j

xx yy xy xx yy xy

p

K T xy

plane

A C A C B

If wesetK T xy we see that all of the second derivatives of and should b e equal

plane

to zero and so T should b e a linear transformation This coincides with the main theorem of

plane

ane geometry whichsays that the only transformations of the plane which map lines into lines

are linear transformations

Direct View Condition

Perceptual distortions of the shap e of rectangular corners spheres b o dies cylinders etc Sec

tion dont o ccur when we use narrowangle p ersp ective pro jection ie when the principal ray

go es through the ob ject and the angular subtense of the ob ject when viewed from the pro jection

p oint is small Therefore the following condition can b e considered a generalization of a number of

nonstructural p erceptual requirements

Directview condition It is desirable for the pro jection to b e close in a neighborhood

of each p oint to the p ersp ective pro jection into a plane p erp endicular to the b er line

going through this p oint

First we consider the case when the b er structure of the mapping coincides with the b er

structure of a central pro jection

We can observe that the pro jection into the sphere is lo cally close to the pro jection into

spher e

the tangent plane of the sphere which is p erp endicular to the b er lines of P

Nonstructural Requirements

Therefore if we use the decomp osition P T wehave to construct the mapping T

spher e spher e

which is lo cally as close to a similarity mapping as p ossible Formally it means that the dierential

of the mapping T which maps the tangent plane of the sphere at each p oint x to the plane

spher e

T R R coinciding with the picture plane at the p oint f x should b e close to a similarity

f x

mapping The dierential DT can b e represented by the Jacobian matrix J of the mapping

f x

T at the p oint x

spher e

In co ordinate form this can b e written as

J I

fx

for some

For lo cal co ordinates u v on the sphere the matrix J can b e written as

fx

u v

C B

J

A

fx

v u

Then for to b e satised it is necessary and sucientthat

jJ j

fx u v v u

Note that these equations formally coincide with CauchyRiemann conditions If we consider the

sphere and the plane to b e complex manifolds then the equations ab ove are necessary and sucient

conditions for T to b e a conformal mapping

spher e

For global spherical co ordinates wehave

u v

sin

u v

sin

Wewillsay that mappings T and P satisfy direct view condition if they satisfy the condi

spher e tions

Nonstructural Requirements

We also need of measure of deviation from the direct view condition A nondegenerate linear

transformation J is a similarity transformation if and only if jJwjjw j do esnt dep end on w If this

ratio dep ends on w thenwe dene the direct view error function to b e

jJwj jJwj

min max D T

spher e

jw j jw j

D T can b e used as the measure of nondirectness of the transformation at the

spher e

p oint

This measure has a simple geometric meaning Supp ose wehave a small sphere at some p oint

in space The image of the sphere on the intermediate sphere is a small disk Then this disk is

mapp ed by T into the picture plane Ideally this image should b e a circle and as it was noted

spher e

in Section the asp ect ratio of the image is a reasonable measure of distortion in this case

The functional that wehave dened is the deviation from of the asp ect ratio of the image of a

small sphere This measure can b e also related to the maximal oending angle in the Perkins laws

Section

jJwj jJwj

and max We will use the following notation Lets nd the explicit expression for min

jw j jw j

E

u u

F

u v u v

G

v v

Then jJwj Ew Fw Gw

v

u v

This expression is identical to the expression in Section so we immediately get

p

jJwj

E G E G F min

jw j

p

jJwj

max E G F E G

jw j

Finally for the direct pro jection error function D weget

Nonstructural Requirements

p

E G E G F

p

D T

spher e

E G E G F

Using the corresp ondence b etween the intermediate sphere and the plane we can write D as a

function of T x and y

plane

The only transformations that satisfy the direct view condition exactly are the conformal map

pings from the sphere to the plane

As it can b e easily shown if T is linear corresp onding T is not conformal

plane spher e

Nowwe consider the case of pro jections P with parallel b ers In this case the pro jection to

the intermediate plane has the direct view prop erty All conformal transformations of the plane

will preserve the direct view prop erty Then the conditions b ecome standard CauchyRiemann

equations

The same measure of deviation from the direct view condition can b e used but in this case

wetake the lo cal co ordinates u v to b e the same as

In this case any comp osition of orthogonal and similarity transformation of the plane will b e con

formal and linear and will satisfy the direct view condition But any pro jection T where

plane plane

T is a comp osition of orthogonal and similarity transformation is equivalent to an orthogonal

plane

parallel pro jection up to a scaling factor Therefore we arrive at an imp ortant conclusion

The direct view condition can b e satised simultaneously with the zerocurvature con

dition within the class of pro jections describ ed by Theorem only if the pro jection

coincides with an orthogonal parallel pro jection to the plane up to a similarity transfor

mation of the plane

ErrorFunctionals

In view of the statement in the previous subsection we can see that only orthogonal parallel pro jec

ultaneously It is quite obvious tion satises b oth the directview and zerocurvature conditions sim

that parallel pro jection is not a satisfactory choice in many cases In Section we examine some

reasons for that Therefore we are also interested in pro jections that have the b er structure of a

central pro jection

Since for central pro jections the directview and zerocurvature conditions cannot b e satised

Nonstructural Requirements

exactlywe should nd a way to dene the optimal tradeo b etween these twotyp es of distortions

In order to do this it is convenient to use global error functionals which indicate the quality of the

pro jection with resp ect to a given condition

In order to estimate the total error for a pro jection wehavetocho ose a measure and integrate

the error function over the domain of the T we will assume that the direct pro jection error

plane

function is converted to the intermediate pro jection plane co ordinates D D T xy The

plane

most conservativeway to dene the curvature and direct pro jection functionals is to use L measure

which results for continuous D and K in

K T sup K T xy D T sup D T xy

plane xy plane plane xy plane

p

Other p ossibilities are for example L measures

 

ZZ ZZ

p p

p p

K T xy dxdy D T D T xy dxdy K T

plane plane plane plane

xy xy

Optimization Problem

Initially we dont makeany assumptions ab out the form of the functional that wecho ose except

that it should b e some functional norm in the space of smo oth functions Assuming T to b e

plane

smo oth K T if and only if K T xy almost everywhere The same is true for

plane plane

D Then the only functions satisfying K T are linear functions and the only functions

plane

satisfying D T are those derived from conformal mappings of the sphere onto the plane

plane

Therefore we cannot make b oth functionals equal to zero for anychoice of the norm So wewant

foragiven value of one functional to nd a function T such that the other functional has the

plane

smallest p ossible value

minimize K T sub ject to D T

plane plane

or

minimize D T sub ject to K T

plane plane

Nonstructural Requirements

Wewant to nd a family of functions T orT solving the constrained optimization

plane plane

problems stated ab ove In each case we get a function or

We can reduce the constrained problem stated ab ove to an unconstrained one in the following

way

Consider a linear combination of the functionals F T K T D T If

plane plane plane

T minimizes F T for a xed then K T isminimal for the xed value of D T

plane plane plane plane

D T As wechange thevalue of D T will increase up to the maximal value whichis

plane plane

attained when Therefore instead of minimizing K T for xed values of D T we

plane plane

can minimize the functional F T for varying from to The same is true for minimizing

plane

D T for xed values of K T The problem can b e restated as follows

plane

minimize F D K for

Wealsohave to sp ecify the b oundary conditions in order to make the problem welldened This

can b e done by xing the frame of the picture that is the values of T on the b oundary of W

plane

To ensure that the mapping is onetoone and the inverse is smo oth the Jacobian of the mapping

should not b e zero anywhere

It would also b e desirable not to increase one distortion anywhere in the image b eyond the

level that is p ossible when the other distortion is It is a somewhat restrictive requirementand

p otentially can b e relaxed if it is p ossible to obtain considerable reductions in maximal distortion

at the exp ense of a small increase of of small values of distortion It is unlikely for the distortion

of curvature if the curvature is in some part of the image small increases in the curvature are

likely to b e observable b ecause of the threshold nature of the p erception of straightness Watt et al

It is p otentially p ossible to gain some improvementby relaxing the same requirement for the

direct view condition small distortions of shap e are undetectable for most ob jects

Mathematically the restrictions ab ovecanbeformulated as

K x y K x y D x y D x y

conf lin

where K x y is the value of the error function at the p ointx y for a conformal mapping

conf

D x y and D x y is the value of the error function for the identity mapping the only lin

Nonstructural Requirements

one which has K x y K x y dep ends on the choice of the conformal mapping this

conf

choice can b e arbitrary b ecause all conformal mappings result in D x y and any is sucient

if condition holds

In Section we will explore a particular case of this optimization problem for T with

plane

central symmetry

It is also interesting to consider weighted error functions that is w x y K T xy and

K plane

w x y D T xy where w x y and w x y are nonnegativeweights

D plane K D

The weights of the error functions can b e interpreted as the relative imp ortance of a particular

typ e of p erceptual error close to the given p oint in the image It can b e usersp ecied or calculated

using information ab out the ob ject space

In this case the total error can be dened as the norm kw x y K T xy

K plane

w x y D T xyk and the optimal transformation

D plane

K D dep endencies As wehaveshown the original optimization problem can b e reduced to

minimization of one functional F This functional do esnt have a go o d p erceptual interpretation A

good waytocharacterize the optimal family of transformation is bytheK D pairs corresp onding

to each memb er of the family

If we solve the optimization problem we get a oneparametric family parametrized by

which denes a parametric curveK D This curveallows to see to what extent distortions

can b e eliminated simultaneously

It is necessary to emphasize that the optimization problem formulated in this section is not equiv

alent to the task of minimizing distortions It is a sp ecic formalization of this task with relatively

low dep endence on the scene contents Pro jections can b e improved p erceptually byintro ducing

more dep endence on scene contents The trado is b etween the diculty of the construction of an

ween appropriate pro jection and the quality of the resulting image We tried to cho ose a path in b et

in which the choice of pro jections is limited but some improvement can still b e achieved

Chapter

Construction of Pro jections

Previous Work

Anumb er of reasons lead to the constructions of picture pro jections that were dierent from the

linear p ersp ective There were three main motivations p erceptual Raushenbakh Reggini

Stark artistic Ernst Inakage Flo con and Taton and technical

Goncharenko et al considerations

Perceptuallybased systems

Regginis p ersp ective Regginis system of p ersp ectiveReggini is based on Thouless for

mula Thouless a Thouless b Thouless rewritten as

i

t p

r d

where t is the distance from the pro jection p oint to the ob ject d is the distance from the

pro jection p oint to the plane and i is a constantbetween and The pro jection of a p ointis

computed in the following way let r b e the distance from the p oint to the principal ray the ray

Perceptuallybased systems

through the pro jection p oint p erp endicular to the pro jection plane Let t b e the distance from

the p oint to the pro jection plane A p oint X in threedimensional space can b e sp ecied by three

numbers rt in the cylindrical co ordinate system with the origin coinciding with the pro jection

p oint and the axis coinciding with the principal ray

In this system the rays sets of p oints mapping to a single p oint in the pro jection plane are

curved In the previous chapter wehave discussed the problems that are asso ciated with this

curvature

The main goal of Regginis system is to provide for changing parallel foreshortening as a result

close ob jects have smaller degree of parallel foreshortening The eective viewp oint determined by

the intersection p oint of the tangent lines to the rays at a given distance from the picture plane

moves away from the pro jection plane as the distance to the ob ject increases

Raushenbakhs system Raushenbakhs system Raushenbakh is based on an idea similar to

Regginis Instead of using Thouless formula he derives similar curvilinear systems under more gen

eral assumptions ab out the relationship of the apparent size on the pro jected size More sp ecically

Raushenbakh assumes that the apparent size is related to the pro jected size by the formula

s sF D

where s is the apparentsize s is the size of pro jection onto a xed plane D is the distance

from the plane to the ob ject He suggest a technique for measuring F D

The result of his derivations is a family of pro jections dierent systems corresp ond to dierent

tradeos b etween vertical and horizontal deviation of the depicted size from the apparentsize

Both systems that were describ ed ab ove are based on the assumption that the apparentsize

should b e depicted We should note however that the size constancy works also for pictures

although less than for the real scenes Still there is a dierence in the p erception of size in pictures

and in the real world and some intermediate values b etween apparent and pro jected size might create

the most appropriate impression These systems are also useful for understanding some deviations

from linear p ersp ective in paintings and naive art some of the representations of ob jects that are

consistently present in the pictures could result from unconscious attempts to depict the apparent size

Perceptuallybased systems

We b elieve that main problem of these systems is that the distortion asso ciated with relativesize

can b e found mostly in wideangle images and a numb er of other distortions such as the distortion

of shap e in these image tend to b e more prominent The systems describ ed ab ove might result

in reduction of these distortions but nothing in their construction guarantees that The imp ortant

problem of the curvature of images of straight lines is not addressed either

We dont describ e two more systems Flo cons and Starks b ecause we didnt have the original

references available and we could not nd satisfactory accounts of their work

Artistic systems

In a sense any pro jection system falls into this category b ecause any system can b e used for artistic

purp oses We describ e several systems that were created sp ecically for artistic purp oses

Eschers p ersp ective M C Escher is famous for geometric construction in his paintings He

used unuasual p ersp ective constructions in manyworks and one of his exp eriments Figure is

particularly relevant to our discussion From Eschers notes that he had made when he was working

on this woodcut weknow that he was trying to create a smo oth transition b etween dierent views

of a building with the viewing direction changing from horizontal to almost vertical Resulting

construction is similar to the panoramic pro jection describ ed b elow and allows to avoid distortion

of rectangular corners in all parts of this complicated construction

Systems describ ed byMInakage Anumb er of pro jections were describ ed in Inakage

Inakage denes curvilinear perspective as a linear pro jection of the threedimensional space in to

a curved surface This denition implies that the surface is pro jected into the plane to get the nal

image Therefore this denition is equivalent to the decomp osition given by Theorem

Dwarping extends the curvilinear p ersp ective allowing to change the metrics of space pro duc

ing curved rays Any continuous system can b e considered an instance of curvilinear p ersp ective

combined with Dwarping

Sp ecic case of inverse perspective is describ ed by the formulas

z z

y y x x

d d

Perceptuallybased systems

Figure M C Escher High and Lowlithograph

Perceptuallybased systems

where x y z isapoint in space x y is its pro jection and d is the distance from the pro jection

p oint to the pro jection plane

Another example of p ersp ective constructions are the photographic collages byDavid Ho ckney

Ho ckney

As wehave mentioned in Chapter deviations from linear p ersp ective in the art are probably

more common than rigorous p ersp ective constructions Rejection of linear p ersp ectivewas one of

the theoretical foundations of cubism The systems that wehave describ ed ab ove are just several

examples of attempts to use alternative metho ds of pro jection

Technical Projections

Extreme distortions of linear p ersp ective for angles close to and imp ossibility of using it for

pro jection angles greater than prompted several practical solutions primarily in lens design

For extremely wideangles greater than degrees linear pro jection is never used b ecause of

the distortions and diculties in creating appropriate rectilinear lenses Fisheye lenses are based

on pro jections that are dierent from linear p ersp ective All of these pro jections have straightrays

unless the light propagates in a medium with variable refraction co ecient

Several pro jections used in lens design are describ ed in Goncharenko et al The b er

structure of all these pro jections is identical to the central pro jection all of the b ers are straight

lines intersecting at the center of the lens which can b e called the pro jection p oint Moreover these

pro jections have axial symmetry Let p b e the axis of symmetrywhich corresp onds to the raygoing

ts Let b e the angle b etween the direction from through the center of the lens and its fo cal p oin

the pro jection p oint to a p oint in space X We will use a p olar co ordinate system r on the

pro jection plane with the origin at the intersection p oint of the axis and the plane Then the image

of the p oint X is determined by a function r F

Three pro jections describ ed in Goncharenko et al are shown in the table b elowf is the fo cal length of the lens

Axially Symmetric Projection

name formula examples of lenses

orthographic r f sin FisheyeNikkor OP

Kowa Fisheye Olympus ZuikoFisheye

equidistant r f

NikkorFisheye

Asahi Sup er TakumarFisheye Minolta MC Fish

equalized r f sin

eye

Another example of a pro jection that is widely used in photography is the panoramic pro jection

If the lm in a panoramic camera is placed on a cylindrical surface the resulting pro jection resem

bles equidistant pro jection in one direction and regular linear p ersp ective in the other Panoramic

pro jection was advo cated byFeininger as the only correct p ersp ective on the grounds that

parallel lines should always b e seen as convergenttoapointatinnity

The pro jections that wehave describ ed in this section were mostly ad hoc solutions to the

problems of wideangle images No consistent eort was made to minimize geometric distortions

Axially Symmetric Projection

In general the optimization problems formulated in Section are very dicult In this section

we will consider the case of axially symmetric pro jections The transformation T for these

plane

pro jections can b e written in p olar co ordinates as

r

In this case the problem can b e reduced to a onedimensional problem Unfortunatelyeven in

this form direct solution of the optimization problem is quite dicult We shall make some estimates

of the optimal values of the functional and show that some simple families of pro jections are close

enough to optimal We shall consider some applications of one of these families in Chapter

Reformulation of the Problem

Using Equations for the transformation T we can simplify the expressions for error functions

plane

After substitutions we obtain

Axially Symmetric Projection

r r r r

R K r R

min

r r r

where R is the scaling constant

The expression ab ove is for the transformation T R R The direct pro jection

plane

functional was derived for T S R The formulas relating T and T are

spher e plane spher e

r tan

The direct view functional has the form

d

max

d sin

d

min

d sin

Using r tan weget

max r

r

D r

min r

r

We can notice that in b oth cases the dep endence on the angular co ordinate completely disap

p eared so now the problem is onedimensional

Weshouldalsonotethatif do esnt dep end on the only consistentway to dene the b oundary

conditions is to sp ecify the value of on a circle ie for r R where R is the size of the picture

we will assume that we use it as the scaling factor in the curvature error function

As the transformation T is required to b e onetoone and continuous everywhere

plane

and r everywhere In order for the error functions to b e nite at In general it

r for all r b ecause otherwise we get an innite compression factor at is desirable to have

the p oints where

Moreover if wewant the transformation to b e smo oth at zero

The optimization problem for the supnorm now takes the form

r D r minimize F max K

r

sub ject to RR r

Axially Symmetric Projection

Limiting Cases

Consider the cases and In the rst case the optimal solution is r r In the second

case we obtain a dierential equation

p

r r

which can b e easily solved The solution satisfying RR is

p

r R

p

r

r R

The expression for the comp osition of the stereographic pro jection and which can b e obtained

by substituting r tan into the expression for is more intuitive

tan C tan

where C is a normalization co ecient and results in the following expression for

R tan arctan r

r

tan arctan R

This mapping has a simple geometric interpretation Figure rst the plane is pro jected to

the intermediate unit sphere and then repro jected back from the p ole

1

r ρ( r )

Figure Construction of DP r

Axially Symmetric Projection

This mapping has an imp ortant role in the rest of the thesis We will call it direct view trans

formation and denote it DV rR

In Figure we compare this transformation with transformations implemented in the sheye lenses

θ 1

θ _θ sinθ 2sin_ θ 2tan tanθ 2 2

abcde

Figure Pro jections orthographic a equalized b equidistantc DP r d linear e

Lower Bound of the Optimal Values of F

In this section we will nd a lower b ound of the optimal values of the functional

First we observe that

min F min max F r

R

min F R R RR min F R R RR

min F R x x R

x x



Therefore we can estimate the minimal value of the functional from b elowby minimizing the

function

x x

R

x R x max R F x

minx R x

If x should b e zero at the p oint where the minimum is achieved We can consider the

Axially Symmetric Projection

function F of only one variable x Assume we shall consider the case case later

Then we can divide the functional by and use

p p

Consider three ranges of x x R R x and x

p

x R In this case the function can b e rewritten as

x

R

F x x R

x R x

The derivative of the function is

x x

R

F x x R

x R x

In this range of x the derivative is clearly negative Therefore if the minimum is attained in

p

R this range it must b e attained at the rightendpoint

x In this case the function takes the form

x R x F x x R

R

The derivative of the function is

Axially Symmetric Projection

F x x R x x R

R

The derivative is p ositive in this range therefore the function is increasing and the minimum

can b e attained only at the left endp oint x

p

R x In this case the function and the derivative are

x

R

F x x R R x

x

x x

R

x R F x x R

x

At the right endp oint of the interval the derivativeisalways nonnegative at the left endp oint

it might b e p ositive or negative

The second derivative of the function in this range is

Axially Symmetric Projection

x x x x

R R R

F x x R R

x x x

It can b e easily seen that each term is p ositive inside the interval Therefore the rst derivative

increases inside the interval and has no more than one ro ot If it is p ositive at the left endp oint

there are no ro ots otherwise there is one ro ot

Using the continuity of the function we conclude that the minimum is attained at a p oint in the

p

R x foranyvalue of interval

If it is easy to see that in this case the minmal value of F is and is attained at

p

x R

If the value of D at R is known we can nd r

The K D curve for the b ound is shown in Figure for R eld of view

Second Bound

The b ound that wehavederived in the previous section is not very tight which it can b e easily

seen for In this case the minimum of the functional is and the optimal solution is unique

Section The values of K in this case are much higher then the b ound found in the previous

section In this section we shall make some additional assumptions ab out and nd a b etter b ound

for the optimal pairs K D in a narrower class of functions We b elieve that this b ound is still

far from the optimal it is however the b est we could obtain Although the optimization problem

that we consider can b e easily reduced to a problem of optimal control in the form of Mayer with

inequality constraints Pontriagin the necessary conditions that we could nd dont yield a

go o d estimate of the optimal values of the functional F or of the optimal solution in particular

b ecause the second derivative of the Hamiltonian with resp ect to the control vanishes identically

The elementry estimate that we derive in this section supp orts the idea that the interp olations

derived in the next section are close to optimal but of course do esnt prove it

Axially Symmetric Projection

We make the following assumptions ab out

D K are nondecreasing functions of r

p

r D r is less than for r r for all r ie D r

p

r r

The function is concave

The rst assumption is motivated by the desire to avoid larger distortions closer to the center

of the picture ie to preserve the same typ e of distribution of distortions as for the limiting cases

r r and r DP r

The second assumption follows from inequalities

The third assumption means that there is no overcompression in the radial direction Consider

an image of a small sphere For p ersp ective pro jection it is an ellipse with its ma jor axis oriented

along the direction to the center of the picture For DP r the image is a circle Wewould exp ect

intermediate pro jections to pro duce images of intermediate shap es ie ellipses with ma jor axis still

oriented along the direction to the center but with smaller asp ect ratio

The last assumption is motivated by the fact that in a regular p ersp ective image the amountof

the shap e distortion increases towards the edges and the compression should increase

K r K where



r r

K

Due to Assumption K is p ositive Note that As Assumption

j j

Let u Then

u u

K K u K

r r

Wehave assumed that K increases Supp ose u decreases at some p oint than K should

dK



increase Therefore at that p oint Dierentiating we get the following expression for the

dr

derivative

Axially Symmetric Projection

dK

u u u u

dr r r r r

It is nonnegativeif

u

u

r

u

u r

r

Supp ose that u increases on r Then

Z

u ur

u maxu u du

r r r



r



If wecho ose r to b e the maximalinterval on which u do esnt decrease it will decrease on

an interval r r As u r increases on r the following estimate can b e made

r



ur

ur u ur

r



u u r max

r



r r

ur

r



p

It remains to estimate ur r First we estimate Denote d D Then

p

r r d

p

r and integrating from r to we get Dividing by r

Z Z

dr dr

p p

ln d

r r

r r

The integral in is equal to the function DP r dened in Section Finallyweget

d

DP r DP r

From and we obtain

p p

r r r r

u

d

DP d DP These b ounds can b e directly used in b ecause the rst b ound is an increasing function of

Axially Symmetric Projection

ur and the second b ound is a decreasing function of ur Due to the Assumption the value

p

of uisknown as D r is nondecreasing D d therefore d

We can further improvethelower b ound for D close to in the following way the estimate

still holds for maximum of u on r Supp ose the estimate is maximal for some r Then the

value of K will b e greater than the lower estimate for K on the interval r plus the maximum

of the estimate of u

The plot of the resulting b ound is shown in Figure

K 2

1.5

1

0.5

0 0 0.2 0.4 0.6 0.8 1

D

Figure Bounds from Section upp er curves and Section lower curve

We can further narrowdown the range of the functions that we are considering and assume that

u increases In this case the estimate can b e further improved as the value of u will b e greater

than the maximum of Figure

Interp olations

We couldnt nd a direct solution of the optimization problem but given the estimates of

the previous sections we can evaluate interp olations b etween the solutions for the cases and

describ ed in Section

We suggest two p ossible interp olations

linear interpolation describ ed bytheformula

Axially Symmetric Projection

p

R r

p

r r R

r R

geometric interpolation describ ed by

arctan r

tan

r R

arctan R

tan

The second case corresp onds to the following geometric construction DP r is the pro jection of

the plane to the unit sphere from the center of the sphere followed by repro jection from the p ole If

wemove the p ointfromwhichwe repro ject from the sphere to the plane from the p ole to the center

we get a smo oth interp olation from r to DP r describ ed by

If we compare the K D curves for these interp olations with the estimates derived in the previous

sections b oth families of functions satisfy the assumptions for all estimates wecanseethatnot

much improvement can b e gained by nding b etter approximations to the solution of the optimization problem Figure

K

2

1.5

1

0.5

0 0 0.2 0.4 0.6 0.8 1

D

Figure The KD curves from top to b ottom are linear interp olation geometric interp olation estimate from

Section estimate from Section lower curve

It is imp ortant to rememb er that p erception of curvature and of some shap es such as spheres and

rectangular corners has a threshold nature Therefore once a p erceptual requirement is violated

by a suciently large margin p erceptual saturation is likely to o ccur and even large dierences

Axially Symmetric Projection

do not makemuch dierence Our rst estimate was made without any assumptions ab out and

it still indicates that for lowvalues of D K should b e quite large

Nevertheless the dierence b etween the interp olations that we prop ose and the estimates is still

considerable and it is at least of theoretical interest to obtain b etter estimates of the optimal K D

pairs and of the optimal solution

Extentions

Both oneparameter families that wehave describ ed can b e easily extended to larger families

parametrized by the function By allowing dep ending on the angle we obtain greater exi

bility in the choice of transformation appropriate for a particular picture

The equations for the pro jections take the form

p

r R

p

r rR

r R

arctan r

tan

r R

arctan R

tan

In this case is similar to the usersp ecied weights describ ed in Section If we assume

that changes slowly the expressions for direct view and curvature error functions can b e applied

and for each direction the problem can b e solved indep endently Therefore optimal values for

each direction are close to those obtained for the axiallysymmetric case

For fast changes in however this do es not applyandtwodimensional optimization is necces sary to compare the family to the optimal solution as dened in Section

Chapter

Implementation and Applications

Implementation

Wehave implemented the transformations describ ed in Section and Section The im

plementation of our pro jections is quite straightforward The pro jection of Theorem

plane

practically coincides with the standard p ersp ectiveparallel pro jection There is however an im

p ortant implementation detail that is absent in some rendering systems As wementioned b efore

our center of pro jection need not coincide with the p osition of the camera or the eye It is chosen

according to p erceptual requirements For instance it can happ en that the most appropriate center

of pro jection for an oce scene is outside the ro om Figure c In these cases it is necessary

to haveamechanism for making parts of the mo del invisible These parts of the mo del should

participate in lighting calculations but should b e ignored by the viewing transformation This can

b e done using clipping planes

The D part of the viewing transformation T can b e implemented as a separate p ostpro

plane

cessing stage In this way the transformation can b e applied to any p ersp ective image computer

generated or photographic The only additional information required is the p osition of the center of

pro jection relative to the image The basic structure of the implementation is very simple

Implementation

for all output pixels i j do

p

r i j

setpixel i j

interpolated color r ir r jr

end

For the geometric interp olation the inverse function can b e calculated directly using

the formula

arctan R r

r tan arctan tan

R

Equation can b e used in the cases of constant and variable

For linear interp olation the inverse function canbecomputednumerically with any

of the standard ro otnding metho ds such as those found in Numerical Recipes Press et al

As wehave mentioned in Section for mo derately wide angles the dierence b etween any

transformation from the family that we are considering and the identity transformation is quite small

That means that the initial approximation of identity for anynumerical ro otnding pro cedure is

good and even the dichotomymethodworks quite well There is no particular advantage in using

the linear interp olation at least in our implementation but it was the rst one that wehave

implemented and some of our images were generated using this transformation

color x y function computes the color for any p ointx y with real co ordinates The interp olated

in the original image byinterp olating the colors of the integer pixels This interp olation can b e done

using convolution with any reasonable lter Using a wider lter slows down the computation but

virtually guarantees the absence of aliasing

Implementations of and are practically identical to the implementation describ ed

ab ove There is a signicant dierence in the case of the values of cannot b e precomputed

b ecause they dep end on Using is preferable b ecause the inverse function can b e computed

explicitly

We also need a mechanism that allows user to sp ecify desired values of for various directions

in the picture One way to implement this is to allow user to sp ecify the values of atanumber of

p oints compute the direction from the center of the picture to each p oint and use a closed spline

curvetointerp olate the values of for all directions see for example Hill

Deep and Shallow Scenes When Do We Need a WideAngle Projection

The algorithm is linear in n and should b e ab out as fast as any other image transformation

change of the size ltering or other geometric transformations Our nonoptimized program in

which no attempt was made to eliminate extra function calls in inner lo ops etc to ok ab out sec

on an HP to transform a by image with a lter of width

Deep and Shallow Scenes When Do We Need a WideAngle

Projection

As wehave discussed in detail in Chapter narrowangle p ersp ective pro jections are likely to pro

duce p erceptually acceptable images while wideangle pro jections are likely to pro duce distortions

and are practically unusable for extremely wide angles Figure a As wehaveshown in Chap

ter and Chapter it is in general imp ossible to nd a pro jection whichwould pro duce undistorted

images for any scene given a xed angle of pro jection

Before we try to use compromises suggested in Chapter wehavetoanswer the following

question when is it p ossible for a given scene to use a narrowangle pro jection instead of a wide

angle one

A general direction from the scene to the pro jection p oint is determined byourchoice of the

side of the ob jects that wewant to see in the picture we assume that our preferences are consistent

with the b er structure of the p ersp ective pro jection ie we dont want to see around the corners

or oppp osite sides of one ob ject at the same time We are only free to cho ose the distance b etween

the pro jection p oint and the scene

Supp ose wewantanobjectofsize s is lo cated at the distance l from the center of pro jection

to have approximately the same size as an ob ject of size S lo cated at the distance l L Ss

s the distance l As the ratio of the sizes of their images for a p ersp ective pro jection is Sll L

determining the pro jection angle can b e estimated from Sll Ls The pro jection angle

can b e found from the expression tan Sl L The resulting estimate fo r l is S sLOf

course S need not b e a size of an ob ject it can b e a distance b etween ob jects and in general any

characteristic length in the direction p erp endicular to the principal rayWe can consider S to b e the

width of the scene in the background and s to b e the width of the scene in the foreground L

is just the distance b etween the closest and most remote ob jects in the scene depth of the scene

Deep and Shallow Scenes When Do We Need a WideAngle Projection

Therefore the pro jection angle has to b e large if the dierence b etween the foreground width of the

scene and the background width of the scene is large compared to the depth of the scene

We shall call a scene shal low if the ratio S sL is small if this ratio is large then we

shall call the scene deep

Deep scenes require wideangle pro jections while narrowangle pro jections can b e used for shal

low scenes

Almost any indo or scene is inherently shallow b ecause the ob jects in such scenes typically have

comparable sizes Therefore in most cases for closed spaces of relatively small size we can get away

with using the p ersp ective pro jection It is imp ortanttonotethatthechoice of the pro jection p oint

for a scene do esnt have to reect an actual physical p osition of the observer in the scene For

example any reallife photo showing a whole wall of an oce would require a very wide pro jection

angle However in computer graphics nothing prevents us from cho osing a pro jection p ointinan

unreachable p osition in or outside the ro om Our choice can b e based purely on the p erceptual

quality of the picture Minimization of shap e distortion requires narrow angles of pro jection At

the same time extremely small angles of pro jection result in a lack of parallel foreshortening and

excessive p erp endicular foreshortening for some ob jects Therefore a mo derate pro jection angle is

most preferable In Figure we can see pictures of the same part of the oce from three pro jection

p oints corresp onding to the pro jection angles and Most p eople prefer the pro jection in

the middle In fact the number was determined by showing to p eople a set of pictures with the

pro jection angle varying from to and asking them to c ho ose the nicest picture The variance

of the results was small nob o dy chose any image outside the limits

The eect of the change of the pro jection angle on the image of a deep scene is radically dierent

For a deep scene the general layout and relative sizes of the ob jects are unchanged only for a small

range of angles Consider Figure The image has the desired comp osition but distortions

are clearly visible the men close to the edges of the picture app ear to b e bulkier than the men in

the center of the picture although their shap e was identical The shap e of the head app ears to b e

distorted in some cases and the men close to the edge of the picture app ear to b e asymmetrical

The two images are radically dierent in the rst case when all the ob jects are still present

in the picture the men in the picture almost disapp ear In the second case the pyramids in the

background are absent and relative sizes of the images of the men are changed

Deep and Shallow Scenes When Do We Need a WideAngle Projection

a

b

c

Figure Example of a shallow scene a pro jection angle b pro jection angle c pro jection angle

Position of the camera is shown on the right for each picture Note that for image the camera is outside the ro om

Applications of Correcting Transformations

For deep scenes the pro jection angle is practically determined by the scene and comp osition of

the picture and cannot b e varied In this case transformations of the typ e describ ed in Section

can b e applied to decrease the distortions in the p ersp ective image The results of applying the

transformation to the picture is shown in Figure b Note the reduction in the distortion

of shap e of the men As a tradeo the barrel distortion is intro duced but due to the absence of

straight lines its only manifestation is a slant of some gures

Striking distortions o ccur in wideangle motion pictures esp ecially when the camera rotates

around a xed vertical axis pan They are primarily asso ciated with two factors

whenanobjectmoves to from the edge to the center of the picture the shap e of its image

changes from distorted to nondistorted This change is noticeable and pro duces an impression of

nonrigidity

the vertical co ordinate of ob jects that are lo cated at a large angle from the horizon changes

considerably cf Figure This eect app ears unnatural

These eects are reduced if transformations like or are applied to the frames

Summarizing the discussion ab ove we suggest the following recip e for the choice of an appropriate

pro jection for an image

Determine if the scene is shallowordeep

If the scene is shallow cho ose a p erceptually optimal pro jection angle which is likely to b e in

the range

If the scene is deep determine the minimal pro jection angle that is acceptable for the scene

Then cho ose a transformation from or which pro duces the b est results

Applications of Correcting Transformations

In the previous section we describ ed the conditions which require a wideangle pro jection under the

assumption that we are restricted in our choice of pro jection only by the comp osition of the scene

In the case of photography there are several additional restrictions there should b e no obstacles

between the camera and the ob jects that must app ear in the image and the size of the camera

cannot b e made arbitrarily large For example it is imp ossible to make a picture of a ro om from

Applications of Correcting Transformations

a

b

c

d

Figure Example of a deep scene a pro jection angle b pro jection angle correcting transformationwith

applied c pro jection angle keeping the gures of men approximately the right size d pro jection

angle keeping the pyramids in the eld of view

Applications of Correcting Transformations

a pro jection p oint b ehind a wall As a result the pro jection angle can b e chosen only in a limited

range and some pictures must use a wideangle pro jection

In many cases the b est camera p ositions are physically inaccessible or there is no time to make

agoodchoice of the p osition

All these reasons explain why wideangle images are inevitable in many cases even for shallow

scenes Transformations can b e used to correct for distortions of shap e in such

images

The p osition of the center of pro jection is usually known for computergenerated images but is

more dicult to obtain for photographs For photographs it can b e calculated if weknow the size

of the lm and the fo cal distance of the lens used in the camera Alternatively it can b e computed

directly from the image if there is a rectangular ob ject of known asp ect ratio present in the picture

App endix B

If our task is correcting a distorted photographic image it is imp ortant to understand what typ e

of distortion is present

For example distortions related to verticality Section can b e corrected by a linear trans

formation as it was known to photographers for a long time Feininger Figure

ab

Figure Photos from Feininger a The original photo with convergentverticals b Same photo printed

with slanted easel the slantischosen so that the verticals b ecome parallel

Distortion of shap e can b e corrected using our transformations In some cases these trans

formations also decrease distortions of relative size due to the excessive p erp endicular or parallel foreshortening

Other Applications

Lets consider in greater detail an example of application of transformations and

The head of the author in Figure a is considerably distorted and his right shoulder app ears

to b e longer than the left He also app ears to havemuch more muscle than in real life The shap e of

small ob jects on the table is distorted The rectangular corner indicated in the left part of Figure a

also lo oks distorted cf Section

After application of the transformation for the author lo oks much more like himself

and the rectangular corner lo oks b etter The straight edges in the left part of the picture are curved

If we use the transformation with gradually increasing from on the left to on the

right we can eliminate the curvature of edges while leaving the corrections in the left part o the

picture intact Unfortunately the rectangular corner do esnt lo ok right again we cannot haveboth

zerocurvature and correct shap e at the same place

Again wegive a short recip e for correction of wideangle photos

If vertical ob jects in the picture app ear to b e falling apply a linear transformation which

makes convergentverticals parallel

If distortions of shap e are present apply or Transformation is sucient

when all long straight lines in the picture pass close to the center of the picture otherwise

cho ose close to in for the directions where there are straight lines and no ob jects

of simple or familiar shap e and close to in the directions where there are such ob jects and

no long straight edges

Other Applications

There are two more applications of our transformations that might b e useful

Zo oming If it is necessary to extract a part of a wideangle image the ob jects in the subimage

are likely to lo ok quite distorted In the context of the whole image this distortion app ears to b e more

acceptable but when a part of the picture is clipp ed the distortions b ecome more ob jectionable

If the clipp ed part is small no long straight edges will app ear in the picture and transformation

with close to can b e applied b efore clipping The results are demonstrated in Figure d and e

Other Applications

d

a

e b

c

Figure An example of application of to a wideangle photo a Original picture note the

distortion of the shap e of the head and of the marked rectangular corner in the right part of the picture b Transfor

mation with applied the straight edges in the right part of the picture are curved c Transformation

applied changes from to d Zo om of the head in the original photo e Zo om of the head in the transformed

photo b

Summary

Extremely wide angle images As wehave seen Figure a linear p ersp ective is unsuitable

for extremely wide angles Rather than using traditional sheye views whichovercomp ensate

for the distortions of the p ersp ective pro jection we can use one of our transformations Visible

distortions are practically inevitable but for our transformations they are likely to b e less than for

the p ersp ective pro jection and sheye views

Summary

Here is a summary of p ossible applications of the transformations that we prop ose

Creation of wideangle computergenerated pictures with minimal distortions

Reduction of distortions in photographic images

Creation of wideangle motion pictures with reduced distortion

A b etter alternative to sheye views

Zo oming of parts of a wideangle picture

Summary

a b

c d

Figure Extremely wide angle images a Linear p ersp ective b Angular sheye equidistant pro jection

c Hemispherical sheye orthographic pro jection d Our direct pro jection Images b and d are quite similar

but in c we can observe excessive compression of the ob jects in the margins the tables app ear to b e narrower and

the ball do esnt lo ok spherical

Chapter

Conclusions and Future Work

Conclusions

Lets reiterate the main p oints of the thesis b elow

Wehaveprovided a review of p erception of geometry in pictures and of the usage of p ersp ective

systems in the art Chapter

Wehave suggested a classication of p erceptually imp ortant geometric prop erties of images

structural and nonstructural whichisconvenient for the study of general pro jections of the

space into the plane from a p erceptual p oint of view Section

Wehave shown that a numb er of simple structural requirements such as the absence of lo ops

in the images of straight lines the absence of twists in the images of planes when applied to

all p ossible ob jects that can o ccur in a picture lead to the conclusion that the b ers of the

pro jection should b e subsets of the b ers of a p ersp ective pro jection Section

Wehaveintro duced twotyp es of error functions which can b e used as quantitative measures

of distortions of curvature and shap e Section

Future Work

Wehave shown that the error functions cannot b e simultaneously identically zero for any

pro jection except orthogonal parallel For each error function however there are pro jections

for which the error function vanishes identically Section

Wehave describ ed error functionals and formulated optimization problems that can b e used

to construct pro jections with minimized distortions Section

Wehave found estimates of the optimal values of error functionals for the axially symmetric

case and demonstrated that under certain assumptions simple interp olations havevalues of

error functionals close to the optimal Section

Wehave describ ed situations when distortions are likely o ccur and suggested metho ds to

decrease them Chapter

Future Work

There are a numb er of directions in whichthiswork can b e completed and extended

The most immediate goals are

to nd b etter estimates of the optimal values of functionals and approximations to the solution

of the optimization problem in Section

to consider other cases of simple symmetries for example pro jections similar to cylindrical

A promising direction of work whichwe hop e to pursue is suggested by the fact that conformal

transformations preserve the direct view error function That means that we can try to decrease cur

vature or achieveimprovement in some other p erceptual factor without causing additional distortion

of shap e if we use conformal transformations of the picture plane

The decomp osition theorem from Chapter applies to the case of connected regions of space It

is interesting to consider the ways of splitting the space into several disconnected parts and using a

dierent pro jection for each of these parts

ts of the scene It is also p ossible to consider We tried to avoid dep endence on the conten

transformations that dep end strongly on our knowledge of the geometry of the ob jects in the scene

In this case error functions havetobeevaluated only for the actual images of ob jects rather than

for all p ossible images and b etter results can b e achieved at the exp ense of much greater complexity

and lesser generality of the algorithms

App endix A

Review of the Exp erimental Studies of

Perception of Pictures

This is a brief exp osition of exp eriments describ ed byvarious authors It is not in anyway complete

Our main purp ose here is to give the reader a general idea ab out stimuli viewing conditions tasks

and results of the exp eriments

We tried to describ e all the exp eriments using consistent terminology from Section Thus

on numerous o ccasions we had to change the original terms to their equivalents

In most cases wesay nothing ab out theories of p erception that the exp eriments were supp osed

to test On the contrarywe tried to separate the exp erimental data and the theory and to get rid

of the bias intro duced by the theory into the interpretation of the results

We sub divided the pap ers into several sections according to what we b elievewas the main

exp erimental sub ject of the pap er

A Robustness and Perception of the Picture Surface and Frame

Figure A Stimulus from Hagen b

A Robustness and Perception of the Picture Surface and Frame

Hagen b Hagen M A Inuence of picture surface and station p oint on the ability

to comp ensate for oblique view in pictorial p erception Developmental Psychology

Exp eriment

Sub jects Children years old years old and adults in each group

Stimuli Photographs of triangles and squares against a textured background horizontal plane

Figure A Three relative sizes of the images in the pictures bigger was and of

the smaller wotyp es of stimuli were used color prints and slides

Viewing conditions Peephole mono cular viewing two viewing directions and viewing

distance equal to the pro jection distance

Task The sub jects were instructed to p oint to the bigger ob ject

Results Correct station p oint increased the p ercent of correct choices for adults and yearold

children for slides and decreased for prints For yearold children the p ercent of correct answers

was very small indicating insucient understanding of p ersp ective pictures

Lumsden Lumsden E A Problems of magnication and minication An ex

A Robustness and Perception of the Picture Surface and Frame

planation of the distortions of distance slant shap e and velo city In Hagen

pages

Rosinski and Fab er Rosinski R R and Fab er J Comp ensation for viewing p oint

in the p erception of pictured space In Hagen pages

E A Lumsden R R Rosinski and J Fab er describ e some unpublished exp eriments by W C Purdy

Sub jects Unknown

Stimuli Texture gradients pro jected onto a screen Visible p ortion of the pro jection was circular

and lled with texture oblique central pro jection of a piece of transparent lm with a grid of lines

approximately inch apart Two p ositions of the light source and the lm sourcetolm distance

remained constant were chosen so that they resulted in magnication and of the tilted lm

Four pro jection directions of the lm were used and The pro jection

directions were chosen so that each image coincides with the image one step b elow if pro jected from

the viewing p oint

Viewing conditions Fixed viewing p osition on the other side of the screen bino cular viewing un

certain viewing angle limited to by an ap erture

Task Sub jects were asked to adjust a rectangular plate to a p osition parallel with the slantofthe

textured surface in the picture

Results For magnication judged slant corresp onds to the depicted slant one step b elow that

is the slantisperceived as if the pro jection was made from the viewing p oint For magnication

viewing p oint coincides with the pro jection p oint judged slantwas in corresp ondence with depicted

slant with the same constant error

alence of magnication to magnication one step b elow is not surprising as Note The equiv

the distance from the viewing p oint to the screen remained constant and the resulting stimulus was

practically identical

Rosinski and Fab er Rosinski R R and Fab er J Comp ensation for viewing p oint

in the p erception of pictured space In Hagen pages

Exp eriment

Sub jects Unknown

A Robustness and Perception of the Picture Surface and Frame

Stimuli Photos and line drawings of gray square parallelepip eds cm in height and varying in

width Three faces of each parallelepip ed were visible Pro jection p ointwas cm away from the

picture plane Orthogonal central pro jection was used

Viewing conditions The viewing p ointwas lo cated cm cm cm away from the picture on

the principal ray This corresp onds to te viewing angles of approximately and and

magnications

Task Sub jects were asked to estimate the size more sp ecically width of the ob jects in the pictures

Estimate were mo dulusfree in abstract units

Note it is entirely unclear from the text what were the sp ecic instructions to the sub jects and

what abstract units were to b e used

Results There was no signicantchange in apparent relative size of the ob jects when the

viewing p ointwas changed

Exp eriment

Sub jects Unknown

Stimuli Computergenerated images of slanted lattices displayed on a CRT Pro jection p oint for the

pictures was cm away from the screen Central orthogonal pro jection

was used Depicted orientation varied from to in increments

y this Viewing conditions Viewing p ointwas on the principal ray to the picture cm awa

corresp onds to magnications in the range Viewing was bino cular

Task Observers were asked to make direct judgement of the apparent slant in degrees

Results Judged dep ended on magnicationminication this dep endence was more signicant

for slants closer to or and less signicant for the medium range For magnied pictures

dep endence was more apparent than for minied

Note The task of judging the slant in degrees is quite complicated it app ears that the data were

distorted by the absence of exp erience of matching numerical measure with a particular slant The

main p oint however is that the apparent slant dep ends to some extent on magnication Purdys

exp eriment app ears to b e more reliable

Exp eriment

Sub jects Unknown

A Robustness and Perception of the Picture Surface and Frame

Stimuli Computergenerated images of slanted lattices in this exp eriment pro jection p ointwas

always at cm from the screen The range was in increments same as in Exp eri

ment

Viewing conditions Viewing p ointwas lo cated on the principal ray cm

away from the picture which corresp onded to magnication in the range Viewing was

bino cular

Task Same as in Exp eriment

Results Practically no dep endence of the judged slant on the magnication was observed

Rosinski et al Rosinski R R Mulholland T Degelman D and Farb er J Pic

ture p erception An analysis of visual comp ensation Perception Psychophysics

Same exp eriments are describ ed in Rosinski and Fab er in less detail

Exp eriment

Sub jects college students

Stimuli Photographs of a slanted strip ed rectangle slantvaries in the range orthogonal

central pro jection Surface was rotated around a vertical axis exact lo cation of the axis was not

sp ecied Kub ovy in his account of the exp eriment sp ecies the vertical axis going through the

center of the rectangle Photographs were cut into trap ezoidal shap es coinciding with the outlines

of the images of the rectangle

Viewing conditions Viewing was mono cular through an ap erture and the photograph was placed

in a viewing b ox where no frame was visible Viewing direction was either or the distance

from the picture plane was equal to the pro jection distance

Task Sub jects were asked to set the inclination of a vertically pivoted palm b oard to match the

apparent slant

Results

Judged slant diered considerably for and viewing direction

Dene Projected slant as equal to the slant of the rectangle that would pro duce the same image

if pro jected from the viewing p oint When the viewing p oint coincides with the pro jection p oint

A Robustness and Perception of the Picture Surface and Frame

pro jected slant is equal to the depicted slant It was found that the dep endence of the judged slant

on the pro jected slant for viewing direction was very similar to the dep endence of the judged

slant on the depicted slantfor viewing direction The authors conclude that under restricted

viewing conditions apparentslant is determined by the pro jected slant

Exp eriment

Sub jects Same

Stimuli Same

Viewing conditions Viewing was bino cular the photos were in a frame without a viewing b ox

no ap erture was present Viewing direction was or viewing distance was equal to the

pro jection distance

Task Same

Results There were practically no dierence b etween judged slants As pro jected slant in this

exp eriment diered considerably the authors conclude that the apparent slant is determined by the

depicted slant for unrestricted viewing conditions

Hagen et al b Hagen M A Jones R K and Reed E S On a neglected

variable in theories of pictorial p erception Truncation of the visual eld Perception

Psychophysics

Exp eriment

Sub jects

Stimuli a Iso celes triangles whose height and base were in The triangles were placed

vertically on the table Distance was measured from a round marker placed on the table in away

from the sub ject The distance b etween the markers and the triangles was and

in b Slides of the scene cm

Viewing conditions Viewing was mono cular Three viewing conditions were used for the real scene

xed p osition with untruncated eld of view p eephole mm rectangular truncation cm

slot cm away Slides were viewed from the pro jection p oint conditions for the slides coincided with the rectangular truncation

A Rigidity in Motion Pictures

Task Sub jects were asked to repro duce with a tap e the size of the triangles and the distance from

the marker to the triangles

Results

The distance was consistently underestimated for the truncated eld of view there was a consid

erable compression of the scene ie the relative distances decreased The eect was similar for the

three truncated conditions slides rectangular truncation p eephole

Intercept extrap olated apparent zero distance was the greatest for slides and signicantly

nonzero for all conditions except the untruncated viewing

Size estimates for the truncated eld of view were lower than the estimates for the untruncateed

eld of view for the real scene for slides they were higher at smaller distances but lower at longer

distances

A Rigidity in Motion Pictures

Sub jects

Stimuli Orthogonal and oblique direction of pro jection parallel and central pro jections pro

jection angle of nearly rectangular rotating rigid and nonrigid solids on a computer display

The images were wireframe with hidden lines removed Twotyp es of deformations were used for

nonrigid solids ane angles b etween edges are unchanged and nonane Figure A Defor

mations were p erio dical with the p erio d equal to the p erio d of rotation Extent of deformation was

of the smallest dimension of the solid For the oblique pro jection the asp ect ratio

of the ob jects was changed to make them app ear less narrow

Viewing conditions Unrestricted viewing conditions the viewing angle not sp ecied but constant

Duration of each trial was sec p erio ds of rotation

Task Sub jects were asked to rate the rigidity on the scale of to indicating high condence

that the ob ject was rigid

Results

Parallel pro jection makes it more dicult to detect nonrigidity

in viewing direction do esnt makeany dierence The change of

Ane deformations are more dicult to p erceive

A Rigidity in Motion Pictures

Figure A Left Nonrigid stimuli from Cutting

Figure A Right Stimuli from Exp erimentCutting

Small deformations are not recognized very well

Exp eriment

Sub jects Same

Stimuli Stimuli of the same typ e Central pro jections were used Pro jection angles were and

pro jection directions andavariable direction in the range withthe

p erio d equal to one half of the p erio d of rotation Figure A

Viewing conditions Two viewing angles were used and

Task Same

Results

For parallel pro jection wehave less nonrigidity p erceived for all slants of the screen

For p ersp ective pro jection there is considerable nonrigidity for slant

Simulated and physical distance pro jection p oint p osition and viewing p oint p osition did not

aect judgements

Among nonrigid stimuli nonane transformations were much easier to p erceive

Relation b etween wider angle p ersp ective and viewing direction radii pro jection just

lo oked bad from viewing direction

Exp eriment

Sub jects

Stimuli Rigidnonrigid stimuli pro jection direction was comp ensated by the viewing

A Size and Distance in Pictures

direction of the same magnitude radii pro jection distances corresp onding viewing angles

are

Viewing conditions Viewing distances were radii

Task Same

Results Rigiditywas preserved no comp ensation for slant o ccurred

A Size and Distance in Pictures

Smith b Smith O W Judgements of size and distance in photographs The

American Journal of Psychology

Exp eriment

Sub jects p er viewing condition

Stimuli Photographs of a plowed eld extending from the foreground to distant hills stakes

in each photographs were lo cated on a circular arc yd away from the pro jection p oint the height

of the stakes varied in the range in in in increments An ft gap in the middle of the arc

p ermitted sub jects to see the th stake at a distance stakes were identical in all pictures th

stakewas or in high and it was or yd All p ossible combinations

were used Two more photographs were obtained by airbrushing out all the detail except for the

image of one stake in high at yd from the pro jection p oint On one of them shadows were

painted All photos were by in

Viewing conditions Mono cular viewing from a xed p osition through an ap erture Two viewing

boxes were used one resulted in magnication and the other in magnication the viewing

distance was equal to and of the pro jection distance The b oxes were black inside backlit

slides were used Data for similar pictures from two dierent studies were used restricted viewing

from the pro jection p oint bino cular viewing and real scene viewing

Task Sub jects were asked to judge the size of the distant stakebymatching it with a near stake

estimate the distance to the distant stake and stakes in the foreground in yards and the heightof

the distant stake in feet Imp overished photographs were judged either b efore or b efore and after the regular photographs

A Size and Distance in Pictures

Results

Matches of height and judgements of height in feet didnt vary muchwiththechange of viewing

conditions Height judgements were overestimates

The judgements of distance increased for magnication and decreased for magnication

Standard deviation increased with the distance and was quite high so only qualitative conclusions

are valid

Qualitative trend for the imp overished photographs is the same distance judgements dep end on

the viewing p oint but distances were considerably underestimated if the imp overished photographs

were viewed rst Size judgements in imp overished photographs were not signicantly dierent

Smith Smith Patricia C S O W Ball throwing resp onses to photographically

p ortrayed targets Journal of Experimental Psychology

Exp eriment

Sub jects for Exp eriment for Exp eriment

Stimuli In a ro om by ft there were marked target lo cations at ve distances

mfromthe viewer Figure A The viewer was used with ap ertures which resulted

in approximately elds of view Below the viewer a curtain could b e

op ened and through this op ening the sub jects could toss a ball at a target Next to the viewer

another curtain could b e op ened and through this op ening the sub jects could observe the targets

with unrestricted head motion

Figure A Diagram of the ro om of the exp eriments Smith

A Size and Distance in Pictures

Viewing conditions viewing conditions were used mono cular viewing through the viewer with

one of the p ossible ap ertures bino cular and mono cular unrestricted viewing

Task Sub jects were asked to throw the ball at one of the targets for dierent viewing conditions re

stricted conditions rst in Exp eriment unrestricted conditions rst in Exp eriment Immediately

after the toss the viewer or the curtain was closed

Results Restricted eld of view resulted in larger distance to the p oint of impact of the ball There

was no signicant dep endence on the size of the ap erture and b etween mono cular and bino cular

unrestricted viewing The order of the unrestricted and restricted viewing conditions didnt aect

the qualitative relationship b etween the results for restricted and unrestricted viewing conditions

Exp eriment

Sub jects

Stimuli Same except that the sub jects observed photographs of the scene in the viewing apparatus

instead of the scene itself For unrestricted viewing conditions the real scene was used Photographs

were made from the correct viewing p oint no attempt was made to match the colors or intensities

photographs were blackandwhite with yellow target markers paintedonthem

Task Same Sub jects were not told that photos are used after the exp erimenttheywere asked if

they sawanything unusual in the scene visible through the viewer

Results

The data were adjusted in the following way the dierence b etween means in unrestricted viewing

conditions identical for Exp eriments and was added to the means for restricted conditions

to correct for the dierence b etween groups of sub jects Adjusted results were very close for pho

tographs and real scene restricted viewing

There was practically no dierence b etween bino cular and mono cular unrestricted viewing

No sub ject rep orted that he had seen photos in the viewer or noted anyunusual qualityofwhat

he had seen

Exp eriment We omit this exp eriment b ecause we dont nd its results relevant to our purp ose

Smith a Smith O W Comparison of apparent depth in a photograph viewed

from two distances Perceptual and Motor Skil ls

A Orientation and Layout

Exp eriment

Sub jects

Stimuli in backlit photograph of a corridor ft long

Viewing conditions Mono cular viewing through a p eephole from two viewing distances and

of the pro jection distance from the viewing p oints on the principal ray to the picture Viewing

angle was and magnication and A viewing b oxwas used to restrict sub jects

eld of view to the image area

Task Sub jects were asked to estimate the distance to the front edge of the corridor and the length

of the corridor in paces

Results The means for the greater viewing distance were more than two times greater than the

means for the smaller viewing distance

A Orientation and Layout

Halloran Halloran T O Picture p erception is arraysp ecic Viewing angle

versus apparent orientation Perception Psychophysics

Exp eriment

Sub jects

Stimuli Pictures similar to those used in Rosinski et al Blackandwhite strip ed rectangles

depicted at slant around a vertical axis the axis was lo cated

at the right edge of the rectangle

Viewing conditions Unrestricted viewing from the viewing distance equal to the pro jection distance

and the viewing direction of

Task Sub jects were asked to repro duce the slant with a p ointer pivoted around a vertical axis

Results For large slants and large displacements viewing direction of or exceeding those

of Rosinski et al there was a considerable dep endence of the estimated slant on the viewing p osition

A Orientation and Layout

Figure A Left Winslow Homer The Fog Warning from Halloran

Figure A Right Scene reconstruction from Halloran

Exp eriment

Sub jects Same

Stimuli Pictures from Exp erimentwhich depicted and slant transformed in the manner

that corresp onds to the viewing p ositions of Exp erimenttwoframeswere present the trans

formation of the frame of the original and pictures and a rectangular frame of the nal

picture

Viewing conditions Viewing p ointwas xed viewing direction was

Task Same

Results No dierence from Exp eriment except for a small nearly constant digression towards pic

ture plane Compare to the apparent distortions of the pictures of slanted pictures as demonstrated

in Pirenne

Exp eriment

Sub jects

Stimuli Multicolored repro duction of Winslow Homers painting The Fog Warning bycm

Figure A

Viewing conditions Unrestricted viewing from p ositions the following are pairs viewing distance

viewing direction m m m m m m

m

ere instructed a to orient the p ointer in the direction that the dory in the Task The sub jects w

picture was headed with resp ect to the picture plane b to orientthepointer in the direction of the mother ship on the horizon from the mans p osition

A Orientation and Layout

Results

There is a consistent shift in b oth judged orientations with the change in the viewing p oint As

the viewing direction angle increases the judged angle of orientation of the b oat and of the direction

to the ship increases

The rotation of the judged orientation of the b oat approximately corresp onds to the mean rotation

of the ob jects reconstructed using sets of cues Figure A

Set the xed height of the man and the length of the b oat

Set the rectangle from the stern width to the corresp onding width in the front of the b oat

Set the xed length and width of the b oat except and viewing directions

Set oars are of equal length

The b est t is given byaveraging set which results in no rotation with any of the sets

The change of dierence of the angle of orientation of the b oat and the direction to the ship is

small over interval the judged angle of orientation of the b oat

Exp eriment

Sub jects

Stimuli Mono chrome repro duction of the picture from Exp eriment

Viewing conditions Unrestricted viewing from viewing p oints from Exp eriment with the viewing

direction the viewing distance was scaled as the size of the picture was dierent

Task

a Sub jects were asked to make separates judgementof

the planview orientation of the dory that is indicated by its long axis disregarding all other cues

the orientation with resp ect to the picture plane of a straight line connecting the blades of the

oars

the angles b etween the line connecting the blades of the oars and the b oat axis

To accomplish this a pair of p ointers pivoted on a common axis were used to representtwosides

of the angle to b e estimated

b Sub jects were asked to identify the correct b oat among planview outline drawings with

lengthtowidth ratios of and Results

A Orientation and Layout

Mean length axis orientation was approx For a similar viewing p osition in Exp eriment

the mean orientation was indicating some inuence of other cues

The oarcrossing angle implied by separate judgements of the oar direction and the b oat axis

direction was while the mean oarcrossing angle judged directlywas

There were strong individual dierences

The results of this exp eriment suggest that cues can inuence judgments in dierentways

dep ending on the task and that the ability to separate cues diers among sub jects

Exp eriment

Sub jects

Stimuli stimuli similar to Goldstein Figure A except they were p ersp ective line drawings

each depicted a cm ball with one or two cm square dots on it One dot was lo cated on the

intersection of the central p erp endicular to the picture plane and the sphere the image of this dot

was in the center of the image of the ball the other one was on the big circle of the ball which lies

in the plane p erp endicular to the picture plane rotated from the p osition of the rst dot

Viewing conditions Unrestricted viewing the viewing distance equals the pro jection distance

cm viewing direction was

Task For the twodot picture sub jects were asked to indicate the angle b etween dots with reference

to the center of the ball for the singledot pictures sub jects were asked the dots direction from the

picture plane Pointer arrangement from Exp erimentwas used

or singledot pictures the average judgements were for the center dot and for Results F

the side dot resulting in the relative angle of For the twodot pictures the mean judgement

was

Exp eriment

Sub jects

Stimuli Same as in Exp eriment

Viewing conditions Same as in Exp eriment and also from the viewing p oint on the principal ray

cm away from the picture

A Orientation and Layout

Figure A Left Stimuli from Exp eriment Goldstein

Figure A Right The Village of Bequigny by Theo dore Rousseau Goldstein

Task Sub jects were asked to indicate spatial orientation of the dory using a p ointer with two degrees

of freedom

Results

Mean judged ab out angles were for viewing direction and for viewing

direction Compared to the Exp eriment the rotation b etween p ositions was twice as much

Mean judged pitch angle b etween the axis of the b oat and the horizontal plane was and

observers indicated no or little increase others indicated substantial increase

Goldstein Goldstein E B Rotation of ob jects in pictures viewed at an an

gle Evidence for dierent prop erties of twotyp es of pictorial space Journal of

Experimental Psychology Human Perception Performance

Exp eriment

Sub jects

Stimuli Line drawings of ro ds Figure A with dierent swings around a vertical axis Depicted

swings were The ro d was mm across

Viewing conditions Viewing was mono cular p osition of the righteyewas xed The pictures

ere rotated so that the viewing direction was and The viewing w

distance was cm corresp onding to the viewing angle of approximately for viewing direction

Task The task was to set the direction of a p ointer parallel to the direction of the horizontal ro d in the picture

A Orientation and Layout

Results All the ro ds app ear to b e rotating as the viewing direction changes The amountof

rotation increases with depicted angle from for the ro d up to for the ro d

Exp eriment

Sub jects Same

Stimuli Blackandwhite photograph of the painting The Village of Bequigny by Theo dore

Rousseau Figure A

Viewing conditions Same as in Exp eriment The viewing angle was for the viewing direction

Task Observers were asked to judge the orientation of the road A the rut in the road B the

house C and the line dened bytwo trees D

Results Road rotates the house rotates the trees rotate the rut rotates Mostly

conrms the results of the Exp eriment although the house rotates more then the direction b etween

trees while app earing to b e more p erp endicular to the road with viewing direction

Exp eriment

Sub jects Same

Stimuli Similar to those in the Exp eriment but the length of the horizontal ro ds was only mm

only disks are left

Viewing conditions Same as in Exp eriment

Task Same as in Exp eriment

Results The amount of rotation signicantly increased for small angles up to for ro d

but remained unchanged for larger angles The dierential rotation decreased

Exp eriment

Sub jects Same

Stimuli Same as in Exp eriment

Viewing conditions Same as in Exp eriment

Task Observers were given a sheet of pap er with two parallel lines on it indicating the direction of

the road The task was to indicate the p osition of the rut of the house and of the trees relativeto the road

A Orientation and Layout

Figure A Left Stimuli from Exp erimentGoldstein

Figure A Right Results for Exp eriments andGoldstein

Results There was little dep endence on the viewing direction There was some change in the

relative distances but almost no change in the orientation

Goldstein Spatial layout orientation relative to the observer and p erceived

pro jection in pictures viewed at an angle Journal of Experimental Psychology

Human Perception Performance

Exp eriment

Sub jects

Stimuli Blackandwhite photographs of three vertical dowels with horizontal strip es on a homege

neous white surface against a homogeneous white background Figure A The size of the picture

was bycm

Viewing conditions Mono cular viewing from a xed p osition with viewing distance equal to the

pro jection distance and viewing direction Viewing direction

was changed by rotating the picture

Task Observers judged the spatial layout of the dowels by arranging three discs on a piece of pap er

A Orientation and Layout

Figure A Left Drawings of the face stimuli from Exp erimentGoldstein

Figure A Right A line drawing of the sphere from Exp erimentGoldstein

Results Averaged results are shown in Figure A for viewing directions Size BC

was scaled to the same size for all observers Judged layout changed only slightly with the viewing

direction

Exp eriment

Sub jects Same

Stimuli Same

Viewing conditions Same

Task Observers were asked to set a p ointer parallel to the directions BA BC CA

Results Total changes of the directions were BA CABCThelayouts

reconstructed from the directions are shown in Figure A These layout dier greatly from each

other

Exp eriment

Sub jects

Stimuli Seven photographs of a frontally oriented human face with dierent gaze directions resulting

from lo oking at seven dierent targets p ositioned at intervals in the range of directions from

to Figure A

Viewing conditions The viewing was mono cular p osition of the righteyewas xed The pictures

as were rotated so that the viewing direction w

Task The task was to set the direction of a p ointer parallel to the direction of the gaze in the picture

Results Total rotation of the judged direction of the gaze was for the depicted gaze of

A Orientation and Layout

and only for the gaze and ab out for gaze Total amount of rotation increased

towards gaze

Exp eriment

Sub jects

Stimuli Similar to the photographs from Exp eriment in addition to the frontal p osition of the

face three additional head orientations were used and

Viewing conditions Same

Task Same

Results

The results were qualitatively similar to the results of Exp eriment The maximum of total

rotation is shifted to smaller angles for and head orientations

Judged gaze direction is relatively indep endent from the head orientation

Exp eriment

Sub jects

Stimuli Line drawings of twovertical cylinders similar to those used in Exp eriments and Black

onwhite pictures were used for viewing in the light and negatives were used for viewing in the

dark

Viewing conditions Similar to Exp eriments and Two mo des of viewing were used in the light

regular line drawings and in the dark backlit negatives

Task Sub jects were asked to orientapointer in the direction parallel to the direction from cylinder

A to cylinder B

Results Viewing the picture in the dark when the picture frame and surface are invisible increases

total amount of rotation of the judged direction for oblique orientation

Exp eriment notnumb ered in the pap er

Sub jects

Stimuli Photograph of a sphere with a marked p oint either at horizontal direction from the center

of the sphere or at counted from the direction parallel to the picture plane Figure A

Viewing conditions Exact conditions unknown Viewing direction was

A RectangularCorners

Task Sub jects were asked to judge the directions from the center of the sphere to the dots presum

ably using a p ointer

Results The average judgements of the orientations of the p oints A and B were and

which results in calculated angular separation of

A RectangularCorners

Perkins Perkins D N Visual discrimination b etween rectangular and nonrect

angular parallelepip eds Perception Psychophysics

Exp eriment

Sub jects

Stimuli drawings of b oxes some of which could have b een and some of which could not havebeen

orthogonal parallel pro jections of rectangular b oxes according to the rst Perkins law Section

There were dierent drawings the rest were pro duced by rotating the original drawings by

The original drawings all had angle C Figure A and angle A varied from

to Three edges radiating from the central vertex were cm size of the pictures was in

Viewing conditions Unrestricted viewing from a p osition feet away from the pictures Three

metho ds of presentation were used straight when pictures were presen ted one by one

samples when two pictures with A equal to nonrectangular case and rectangular

case were p ermanently displayed next to the judged picture

pairs when pictures were sorted into two piles and the rectangularity status rectangular

nonrectangular b orderline was dierent for corresp onding cards in two piles

Task In the rst two cases the task was to sayifthebox in the picture app eared to b e rectangular

In the last case the task was to saywhichbox app ears to b e more rectangular

Results For each pro cedure the p ercentage of correct judgements for rectangular pictures was close

to for nonrectangular it varied from for the straight pro cedure to for pairs Any

boxwas shown in orientations Distribution of the numb er of rectangular judgements maximum

is shown in Figure A

A RectangularCorners

B

A C

Figure A Left One of the stimuli for Perkins

Figure A Right Results of Perkins

Perkins Perkins D N Comp ensating for distortion in viewing pictures

obliquely Perception Psychophysics

Exp eriment

Sub jects

Stimuli Same as in Perkins

Viewing conditions First sub ject judged rectangularity of the pictures from viewing direction

the viewing distance ft cards only Next two viewing directions were used and

A susp ended wire frame ensured correct viewing direction otherwise viewing conditions were

unrestricted The choice of the viewing directions was motivated by the following observation

classication of the repro jection of the drawings into an oblique plane according to the Perkins laws

change for a large numb er of pictures around In the neighb orho o d of and classication

of these pro jections were stable cf Section

Task Sub jects were asked to judge each picture as rectangular or nonrectangular b ox

Results All b oxes for each of the two viewing directions could b e classied into one of four cate

gories using four p ossible combinations of two criteria whether the picture satises the rst Perkins

law and whether the pro jection of the picture onto the plane p erp endicular to the viewing direction

satises this law The b orderline case was excluded

A RectangularCorners

Figure A Left Stimuli from Shepard

Figure A Right Threedimensional ob jects from Shepard

Judgements were made mostly according to the classication of the picture not of the pro jection In

the cases of a conict the p ercent of judgments coinciding with the picture classication was lower

but still high ab ove in all cases except orthogonally nonrectangular pro jectively rectangular

There was very little pro jective inuence at

Shepard Shepard R N Psychophysical complimentarityInPerceptual Or

ganization Kub ovy M and Pomerantz J R editors pages Lawrence

Erlbaum Asso ciates Publishers Hillsdale New Jersey

Exp eriment

Sub jects Unknown

Stimuli line drawings consisting of three radial lines meeting at dierent angles at one p oint

Figure A

Viewing conditions Circular ap erture in frontofthedrawing otherwise unrestricted viewing

Task Sub jects were asked if each pattern could b e a pro jection of a a corner of a cub e all three

angles of a tetrahedron all three angle or of a planar MercedesBenz sign all three

A RectangularCorners

angles Figure A Sub ject could rotate eachdrawing in its plane around the center with

a remote control

Results Dashed lines in Figure A indicate the b oundaries of the areas of orthogonal parallel

pro jections of corresp onding spatial corners For the corners there is no dierence b etween orthogonal

parallel and central pro jections assuming that the principal ray of the central pro jection go es through

the p oint where the lines meet The size of the circles is prop ortional to the fraction of p ositive

answers

Figure A Results of Shepard

For the cubic corner there is close agreement of threshold of p ositiveanswers with orthogonal

pro jection b oundaries This conrms Perkins rst law

For the MersedesBenz gure the agreementwas less than for the cubic corner and for corner

even less However most of the change was in the direction of decreasing the ranges of acceptable congurations

A Foreshortening

A Foreshortening

Nicholls and Kennedy a Nicholls A L and KennedyJM Angular subtense ef

fects on p erception of p olar and parallel pro jections of cub es Perception Psy

chophysics

Exp eriment

Sub jects

Stimuli Line drawings of a cub e orthogonal parallel pro jection and central pro jection with and

pro jection angles

Viewing conditions Pro jections were viewed mono cularly viewing angles were

Task The sub jects were asked to compare the pictures with a standard the picture with the

pro jection angle viewed at a distance corresp onding to the viewing angle

Results

Mo derate parallel foreshortening was always the b est parallel pro jection was second b est

and strong parallel foreshortening was the worst

Parallel pro jection is p erceived as distorted only at close distances only for it was worse than

strongly convergent

Exp eriment

Sub jects

Stimuli Same as but the image with mo derate degree of parallel foreshortening pro jection

angle p ositioned at the distance corresp onding to the viewing angle was used as standard

Viewing conditions Same

Results Order of preference was preserved except for the closest distance when the pictures

with the strongest parallel foreshortening was preferred over the parallel pro jection no parallel

foreshortening but they were still judged to b e slightly worse than the pictures with mo derate

parallel foreshortening

Exp eriment

Sub jects

A Foreshortening

Figure A Left Stimuli from Nicholls and Kennedy b Exp eriment

Figure A Right Stimuli from Nicholls and Kennedy b Exp eriment

Stimuli Same

Viewing conditions The viewing angles were and Standard image

pro jection mo derate degree of parallel foreshorteningset at a p osition corresp onding to the viewing

angle

Results Mo derate parallel foreshortening was still rated the b est at all distances but now parallel

pro jection had lower ratings than strongly divergent

Nicholls and Kennedy b Nicholls A L and KennedyJM Foreshortening and

the p erception of parallel pro jections Perception Psychophysics

Exp eriment

Sub jects

Stimuli Line drawings of a cub e in oblique parallel pro jection One face was always depicted as a

square oblique lines representing the lateral edges of the cub e met the horizontal at

the degree of p erp endicular foreshortening the ratio of the length of the edges of the fronttothe

length of the oblique lines was and Figure A

Viewing conditions Unrestricted viewing

Task Sub jects were asked to cho ose b est drawings among the set with a xed angle

A Foreshortening

Results ratio was consistently preferred

Exp eriment

Sub jects

Stimuli Three wo o den blo cks with square crosssections with lengthtowidth ratios of and

line drawings with receding lines equal to of the frontedges

Figure A

Viewing conditions Unrestricted viewing

Task The sub jects examined one out of three wooden blocks After that they were asked to cho ose

the b est drawing

Results

For blo ck picture was preferred for blo ck was preferred and for blo ck

was preferred In rst two cases the preferred p erp endicular foreshortening was in the

third case Ab out of the sub jects made the dominating choice

Exp eriment

Sub jects

Stimuli Same blo cks as b efore but the pictures were either frontal orthogonal parallel pro jections

or frontal orthogonal central pro jections Figure A

Viewing conditions Unrestricted viewing

Task Same

Results For parallel pro jection and blo cks the preferred p erp endicular foreshortening

was for blo ck For central pro jections it was and resp ectively Ab out of

the sub jects made the dominating choice

Hagen and Elliot Hagen M A and Elliot H B An investigation of the re

lationship b etween viewing condition and preference for true and mo died linear

p ersp ective with adults Journal of Experimental Psychology Human Perception

e Performanc

A Foreshortening

Figure A Stimuli from Nicholls and Kennedy b Exp eriment

Figure A Stimulus from Hagen and Elliot

Exp eriment

Sub jects college students in Exp eriment college students in Exp eriment

Stimuli sets of images of p olyhedra Figure A generated by the metho d describ ed in Reggini

see also Section convergence degree from to Angular size from to

Viewing conditions a mono cular p eephole from xed distance from and viewing directions

b free viewing D solids were shown prior to the exp eriment

Task a assign rank in the range to b Sort the pictures from the most to the least natural

lo oking

Results Parallel pro jection is consistently rated the b est from b oth viewing p oints For the free

A Foreshortening

viewing the result is same the only exception are the pictures of cub es is the most frequently

chosen

Exp eriment

Sub jects college students

Stimuli Same

Viewing conditions Viewing p oint coincided with the pro jection p oint for the degree of convergence

regular p ersp ective Mono cular viewing p eephole

Task Sub jects were asked to rank the pictures from most to least natural lo oking or from most to

least accurate drawing

Results Parallel pro jection has the highest rating No dep endence on the form of instructions was

observed

Exp eriment

Sub jects college students

Stimuli Sets DF from Exp eriments

Viewing conditions Viewing p oint coincided with the pro jection p oint for the degree of convergence

mono cular viewing p eephole Stimuli were presented three at a time each picture was oriented

p erp endicular to the viewing direction

Task Sub jects were asked to cho ose the most and the least natural drawing from each triplet

Results Parallel pro jection was chosen as the most realistic more frequently than any other degree

of convergence

Note This study was critisized by other authors The images in sets C E F diered not only in the

degree of convergence but also in visibility of faces of the p olyhedra The use of Regginis p ersp ective

seems to b e unnecessary b ecause the degree p ersp ectiveconvergence parallel foreshortening could

be varied simply bychanging the viewing p oint

Hagen M A Elliott H B and Jones R K A distinctive Hagen et al a

characteristic of pictorial p erception The zo om eect Perception

A Visual eld

Exp eriment

Sub jects college students

Stimuli Picture sets DF from Hagen and Elliot typ es of stimuli a black and white line

drawings normal light b Dayglo uorescent pap er under normal light c Dayglo stimuli under

ultraviolet light

Viewing conditions Mono cular viewing p eephole the viewing p oint coincides with the pro jection

p oint for the degree of convergence Three pictures from the same set were presented simultane

ously The viewing direction was D ob jects were shown to the sub jects b efore the exp eriment

Task Sub jects were instructed to cho ose the most realistic and natural lo oking drawing out of three

Results

Under the normal light parallel pro jection was preferred for all typ es of stimuli

For Dayglo stimuli under ultraviolet light there was no particular preference

Exp eriment

Sub jects

Stimuli Dayglo stimuli from Exp eriment under ultraviolet light

Viewing conditions Mono cular viewing with head motion allowed

Task Same as in Exp eriment

Results Results are similar to those obtained under normal light parallel pro jection is preferred

Exp eriment

Sub jects

Stimuli Same

Viewing conditions Bino cular viewing with xed p osition of the head

ask Same T

Results Slightchange in the direction of normal light preferences but not nearly as radical

A Visual eld

Finke and Kurtzman Finke R A and Kurtzman H S Mapping the visual eld

A Visual eld

in mental imagery Journal of Experimental Psychology General

Exp eriment

Sub jects

Stimuli A circle cm in diameter with eight radial lines A disk cm in diameter was placed in

the middle each half of the disk contained gratings or cyclep erdegree cp d Gratings had

the same numb er of cp d and were p erp endicular to each other Figure A The disk in the middle

was oriented so that the gratings were parallel or p erp endicular to one of the drawn diameters of

the circle

Figure A Stimuli from Finke and Kurtzman

Viewing conditions The viewing p ointwas xed cm away from the disk resulting in the angular

size of the small disk approximately and the size of the circle approximately Viewing was

bino cular

Task

a Sub jects were asked to move their eyes along each diameter until they could not see that the

disk in the middle consists of two distinct parts A small red dot on the end of a ro d was used as a

xation p oint

b The sub jects were trained to imagine the pattern in the middle the actual pattern was removed

and the same exp erimentwas p erformed

Note The details of instructions are imp ortant for part b see the pap er for details Results

A Straightness and Curvature

The average eld size b oth for real and imagined stimuli decreased with increase in the spatial

frequency of the gratings for cp d for cp d for cp d

The measured visual eld was longer horizontally than vertically asp ect ratio for

cp d resp ectively

Imagery elds were in close corresp ondence with the p erceptual elds

Exp eriment

Sub jects

This exp erimentwas designed to check the inuence of knowledge or exp ectations on the results of

part b of the previous exp eriment Sub jects were asked to guess the outcome of the exp eriment

that was describ ed to them actual stimuli were demonstrated

Results Resulting eld sizes for and cp d gratings were considerably lower than in Exp eriment

the eld size estimate for cp d was closer to the cp d estimate contrary to the results of the

Exp erimentAverage predicted visual elds were circular no eccentricity sub jects however

did predict considerable eccentricity in the correct direction

Note Exp erimentwas mostly irrelevant for our purp oses and is omitted

A Straightness and Curvature

Watt et al Watt R J Ward R M and Casco C The detection of deviation

from straightness in lines Vision Research

Exp eriment

Sub jects authors

Stimuli Various p erturbations of a straight line segment presented on the screen of an oscilloscop e

The segments were arc min in length the size of the p erturbations along the line was in the

e Slike Eachtyp e range arc min There were typ es of p erturbations vernierlike bumplik

of p erturbation had various typ es of smo othing none discontinuous constant bumps Gaussian

linear Figure A Figure A Figure A

The maximal amplitude of the p erturbation in the direction p erp endicular to the bump varied The

A Straightness and Curvature

Figure A Left VernierlikestimulifromWatt et al

Figure A Right Bumplikestimuli from Watt et al

Figure A Slike stimuli from Watt et al

resolution of the displaywas arc min The direction sign of the p erturbation varied randomly

The orientation of the line varied randomly in a range around the vertical axis

Viewing conditions Unrestricted viewing from the distance m

Task The sub jects indicated the sign of the presented stimulus if they could not decide they

guessed

Results On the basis of the data threshold amplitudes for various stimulus typ es and size were

computed

The threshold amplitude was never b elowarcsec

For vernier p erturbations the threshold decreased with size of the p erturbations for the other two

typ es it decreased for smaller sizes of the bump but increased for larger sizes

The b est t to the data was provided bythehyp othesis that the threshold is determined by

the area of the maximal bump with resp ect to the leastsquarero ot straight line through the data

threshold value arc min if the amplitude is greater than arc min

Exp eriment a

Sub jects Unknown

Stimuli Pairs of lines on the screen of an oscilloscop e One of the lines was straight the dots of the

A Straightness and Curvature

other line were p erturb ed according to a Gaussian distribution with varying means The length of

the lines varied

Viewing conditions Same

Task Sub jects were asked to cho ose the p erturb ed line

Results According to a computer simulation the largest bump rule would pro duce threshold vs

length dep endence with exp onent Probability summation bumps are detected with some

probability rule gives exp onents b etween and Complete summation all bumps are

used corresp onds to the exp onent The exp onents found in the exp erimentwere signicantly

less than and greater than weak probability summation

Exp eriment b

Sub jects Unknown

Stimuli Bumplike p erturbation with no smo othing from the Exp erimentwas used Line length

was arc min A wide range of p erturbation lengths was considered

Viewing conditions Same

Task Same

Results It app ears that for larger p erturbation sizes more than haveofthelinethebumpson

the ends of the line are detected rather than the large bump in the middle This results in increase

of the threshold for larger p erturbation sizes cf Exp eriment

Exp eriment a

Sub jects Unknown

Stimuli Lines with two Gausslike bumps on one side with length arc min displayed on the

screen of a oscilloscop e A random Gaussian p erturbation with variable standard deviation and zero

mean was added

Viewing conditions Same

Task Same as in Exp eriment

Results Amplitude threshold was plotted against the standard deviation of the added noise As

suming the existence of internal noise in determination of p osition of the p oints it was estimated to

b e approximately arc sec

A Straightness and Curvature

Exp eriment b

Sub jects Unknown

Stimuli Lines with three typ es of p erturbations the rst second and third derivatives of a Gaussian

with variable spatial extent

Viewing conditions Same

Task Same as in Exp eriment

Results Threshold decreases up to certain size of the p erturbation than slightly increases This

critical size value is relatively indep endentofthetyp e of the p erturbation approximately arc min

This might b e an indication that there is a limit for spatial integration

Ogilvie and Daicar Ogilvie J and Daicar E The p erception of curvature Cana

dian Journal of Psychology

Exp eriment

Sub jects

Stimuli Black curved lines mm wide on a in photographic plate The radii of curvature

were and m Twochord length were used and m Dene

normalized curvature as the chord length divided by the radius of curvature Normalized curvature

approximates the displacement of the end p oint of the line with resp ect to the tangent at the starting

p oint expressed as a fraction of the chord length For mm chord length it was equal to

Viewing conditions The viewing p ointwas m away from the display resulting in angular subtens

of the lines of and arc min A cm ap erture was used to restrict the eld of view to

Viewing was mono cular

torientations vertical horizontal oblique Task Lines were presented to the sub jects in dieren

Sub jects were asked to indicate if the line is curved up or down left or right Each line was

presented times lines were presented in random order

Results Calculated average thresholds were b est correlated with the segment angular area which

was calculated using an approximate formula chord length sagitta where sagitta is the

A Verticality

distance from the middle of the arc to the chord Threshold was in the range log sq sec

sq min

A Verticality

DiLorenzo and Ro ck DiLorenzo J R and Ro ck I The ro dandframe eect as a

function of the righting of the frame Journal of Experimental Psychology Human

Perception Performance

Exp eriment

Sub jects

Stimuli A luminous vertical ro d inside a luminous frame The tilt of the frame could b e adjusted

Viewing conditions Unrestricted viewing from a xed p osition viewing angle in a dark ro om

only the frame and the ro d were visible

Task The sub jects were shown the frame tilted at After min the sub jects were asked to

adjust the frame to the vertical p osition After that the ro d tilted at and the frame tilted at

in the same direction were shown The sub jects were asked to adjust the ro d to the vertical p osition

the frame remained tilted Same exp erimentwas done with the head of the observer tilted at

Results The average righting eect the tilt of the single frame after adjustment was the

average ro dandframe RFE eect was and for counterclo ckwise and clo ckwise directions

When the head was tilted the eects increased

Exp eriment

Sub jects

Stimuli Similar to the Exp eriment with an internal frame added angular subtense

Viewing conditions Same

Task The outer frame was presented in an upright p osition for min Than adjusted if the observer

p erceived it as tilted After that it was returned to the ob jectivevertical Then the inner frame was

added at a tilt of and Exp eriment was p erformed with the inner frame as the frame Same

exp erimentwas p erformed with tilted head

A Verticality

Results No righting eect or RFE were observed in b oth cases

Exp eriment

Sub jects

Stimuli Same as Exp eriment

Viewing conditions Same

Task Similar to Exp eriment but with outer frame tilted and inner frame vertical

Results Approximately righting eect for b oth frames and RFE were observed When

the head was tilted the righting eect increased for b oth frames RFE eect increased mostly

insignicantly

App endix B

Reconstruction of the center of

pro jection from a pro jection of a

rectangle

Supp ose we know that a picture contains a p ersp ective pro jection of a rectangle which is a nonde

generate quadrangle Let l b e the line going through the center of pro jection O p erp endicular to

the pro jection plane p Figure B

Lets assume in addition that we know the p osition of the intersection p oint O of the line l and

the pro jection plane p If the picture was made with a regular camera then O is the center of the

picture

Intro duce a co ordinate system with the origin at p oint O z axis directed along the normal to the

plane and x and y axes parallel to the pro jection plane

Let x x x x b e the known pro jections of the vertices of a rectangle with unknown vertices v

v v v Let d OO d Let e b e the unit vector in the direction of d Then for all j

x d j

App endix B Reconstruction of the center of projection from a projection of a rectangle

vj

x y j

O O’ z d l

x p

Figure B

We will use parentheses for the scalar pro duct of a pair of vectors and for the triple pro duct of

a triple of vectors We will use brackets for the vector pro duct

The fact that x is the pro jection of v can b e expressed as

j j

x v d

j j

v B

j

d

The sides of the rectangle should b e equal

v v v v B

The lengths of the diagonals should b e equal

jv v j jv v j

which is equivalentto

v v v v B

Subtracting v v v v which follows from B from B and simplifying we get

v v v v B

Substituting the expression for v from B into B and B results in j

App endix B Reconstruction of the center of projection from a projection of a rectangle

v ex v ex v ex v ex B

d

x x v ev e x x v ev e B

d d

Multiplying the vector equation B by x x and x we get the following system of equations in

variables v e v ev ewe leavev e as a parameter

x x v e x x x x x x

C C B C B B

C C B C B B

C C B C B B

v e B

x x v e x x x x x x

C C B C B B

A A A

x x v e x x x x x x

The matrix of the lefthand side of this system is the Grams matrix of the vectors x x x By

assumption the pro jection of the rectangle is not a line segment Therefore the rank of this matrix

must b e and the system has a unique solution

We can also note that we can express the determinant of the matrix in the system B as

D x x x where x x x is the triple pro duct x x x This is a consequence of

the following general identity

ae af ag

ab cefg B

be bf bg

ce cf cg

The solutions of the system are

x x x x x x

v e

v e

x x x x x x

D

x x x x x x

x x x x x x

v e

v e

x x x x x x

D

x x x x x x

App endix B Reconstruction of the center of projection from a projection of a rectangle

x x x x x x

v e

v e

x x x x x x

D

x x x x x x

Using identity B the solutions can b e simplied

x x x x x x x x x

v e v e v e

x x x

x x x

x x x x x x

v e v e v e v e

x x x x x x

Substituting these expressions for v e v e v e into B and dividing by

v e x x x we get the following equation

x x x x x x x x x x x x x x x x B

The zcomp onentofthevectors x x and x is equal to d Manipulating the determinants we get

x x x x x x

i j k i j k

d x x x

x x x i j k x x x

i j i j

k k

d d d

x x x x x x

k j j i k i

C B

d

A

x x x x x x

k j k i j i

x x x x x x x x x x

i j k j j k i j j k

B C

d d

A

x x x x x x x x x x

j i j j j j i

k k k

We denote this determinant D Note that

ij k

jD j jx x x x j B

ij k i j j k

aretwodimensional vectors in the pro jection plane x x x

i

i i

Further x x x x d x x x x d Equation B takes the form

App endix B Reconstruction of the center of projection from a projection of a rectangle

D D x x d D D x x d

which is equivalentto

D D x x D D x x d D D D D

This equation has a unique solution if

D D D D B

Consider the geometric interpretation of the quantities x x x It is the volume of the paral

i j k

x x taken with the p ositivesignifthevectors form a righthand triple lelepip ed spanned by x

j k i

and with the negative sign if the vectors form a lefthand triple

Assume counterclo ckwise ordering of the vertices of the quadrangle x x x x if welookdown from

O Figure B This assumption can b e made without loss of generality b ecause if wechange the

ordering the signs of all the volumes that we consider will change simultaneously which do esnt aect the form of B

O

3 p 4 2

1

Figure B

From Figure B we see that

x x x V ol umeO x x x VolumeO

App endix B Reconstruction of the center of projection from a projection of a rectangle

x x x V ol umeO x x x VolumeO

where Oijk denotes the pyramid with ap ex O and base ij k The heights of all pyramids are equal

to dAsD x x x d jD j S where S is the area of the triangle ij k Thiscanbe

ij k i j k ij k ij k ij k

also seen from B

We can use the information ab out the signs of x x x to write B in the following form

i j k

d S S S S S S x x S S x x

Condition B is

S S S S B

4

c d C 3

1 a b

2

Figure B

The diagonals of the quadrangle divide it into four parts Denote their areas by a b c d

Figure B Then condition B means that

c bd a d ca b

After rearrangementwe get

c ab d

which means that nonadjacent pieces must have unequal areas This is a necessary condition for d

to b e determined from B

Denote y x x Then the formula for d is

ij i j

App endix B Reconstruction of the center of projection from a projection of a rectangle

x x jy y jjy y jx x jy y jjy y j

B d

jy y jjy y jjy y jjy y j

The distancefrom the center of projection to the projection plane is given by B whenever the

righthand side is dened and positive For the righthand side of B to be deneditisneccessary

and sucient that the area of the triangle C is not equal to the area of the triangle C and the

area of the triangle C is not equal to the area of the triangle C FigureB

Bibliography

Alb erti Alb erti L B On paintingGreenwo o d Press Westp ort Connecticut

Arnheim Arnheim R Art and visual perception a psychology of the creative eyeUniversity

of California Press Berkeley

Carterette and Friedman Carterette E C and Friedman M P editors Handbook of Percep

tion Seeingvolume Academic Press New York

Cutting Cutting J E Rigidity in cinema seen from the frontrow side aisle Journal of

Experimental Psychology Human Perception and Performance

da Vinci daVinciLNotebooks I IIDover New York

Deregowski Deregowski J B Geometric restitution of p ersp ective Bartels metho d Per

ception

Deregowski Deregowski J B Intercultural search for the origins of p ersp ective In Con

temporary issues in crosscultural psychology Bleichro dt N and Drenth P J D editors pages

Swets Zeitlinger Amsterdam Netherlands

Deregowski and Parker Deregowski J B and Parker D M Convergent and divergentper

sp ective Perception

Deregowski et al Deregowski J B Parker D M and Massironi M The p erception of

spatial structure with oblique viewing An explanation for byzantine p ersp ective Perception

Bibliography

DiLorenzo and Ro ck DiLorenzo J R and Ro ck I The ro dandframe eect as a function of

the righting of the frame Journal of Experimental Psychology Human Perception Perfor

mance

Ernst Ernst b The Magic Mirror of M C Escher Random House New York

Feininger Feininger A Feininger on photography Crown Publishers Inc New York revised

edition

Finke and Kurtzman Finke R A and Kurtzman H S Mapping the visual eld in mental

imagery Journal of Experimental Psychology General

Flo con and Taton Flo con A and Taton R Laperspective Que saisje Le p oint des connais

sances actuelle Presses universitaires de France Paris

Glaeser Glaeser G Fast Algorithms for D Graphics SpringerVerlag New York

Goldstein Goldstein E B Rotation of ob jects in pictures viewed at an angle Evidence for

dierent prop erties of twotyp es of pictorial space Journal of Experimental Psychology Human

Perception Performance

Goldstein Goldstein E B Spatial layout orientation relative to the observer and p erceiv ed

pro jection in pictures viewed at an angle Journal of Experimental Psychology Human Percep

tion Performance

Goncharenko et al Goncharenko E N Osip ova L P and Tomilin M G Lenses for wide

angle and panoramic photography Soviet Journal of Optical Technology

Go o dman Go o dman N Languages of artHackett Indianap olis IN nd edition edition

constructivist approach to picture p erception

Hagen a Hagen M A Development of ability to p erceive and pro duce pictorial depth cue of

overlapping Perceptual and Motor Skil ls part

Hagen b Hagen M A Inuence of picture surface and station p oint on the abilitytocom

p ensate for oblique view in pictorial p erception Developmental Psychology

Bibliography

Hagen Hagen M A editor Academic Press series in cognition and p erception Academic

Press New York

Hagen Hagen M A Varieties of realism geometries of representational art Cambridge

University Press Cambridge New York

Hagen and Elliot Hagen M A and Elliot H B An investigation of the relationship b etween

viewing condition and preference for true and mo died linear p ersp ective with adults Journal

of Experimental Psychology Human Perception and Performance

Hagen et al a Hagen M A Elliott H B and Jones R K A distinctivecharacteristic of

pictorial p erception The zo om eect Perception

Hagen and Jones Hagen M A and Jones R K Dierential patterns of preference for

mo died linear p ersp ectiveinchildren and adults Journal of Experimental Child Psychology

Hagen et al b Hagen M A Jones R K and Reed E S On a neglected variable in theories

of pictorial p erception Truncation of the visual eld Perception and Psychophysics

Halloran Halloran T O Picture p erception is arraysp ecic Viewing angle versus apparent

orientation Perception and Psychophysics

Hill Hill F S Computer Graphics Macmillan Publishing Co Collier Macmillan Publishers

New York London

D David Hockney a retrospective organized by MauriceTuchman and Ho ckneyHockney

Stephanie Barron Los Angeles County Museum of Art Abrams Los Angeles New York

Inakage Inakage M Nonlinear p ersp ective pro jections In Modeling in Computer Graphics

IFIP WG Working Conference Tokyo Japan pages Tokyo SpringerVerlag

Klein Klein F Elementary mathematics from an advanced standpointgeometry Macmillan

company New York

Kub ovy Kub ovyM The psychology of perspective and RenaissanceartCambridge University

Press Cambridge Cambridgeshire New York

Bibliography

La Gournerie La Gournerie J d Traitedeperspective linearecontenant les traces pour les

tableaux plans et courbes les basreliefs et les decorations theatrales avecunetheorie des eets

de perspective Dalmont et Duno d MalletBachelier Paris volume and atlas of plates

Lumsden Lumsden E A Problems of magnication and minication An explanation of the

distortions of distance slant shap e and velo city In Hagen pages

Mason MasonP History of Japanese Art Harry N Abrams Inc New York

Milman Milman M Trompeloeil paintedarchitecture SkiraRizzoli New York

Nicholls and Kennedy a Nicholls A L and Kennedy J M Angular subtense eects on p er

ception of p olar and parallel pro jections of cub es Perception Psychophysics

Nicholls and Kennedy b Nicholls A L and KennedyJMForeshortening and the p erception

ception Psychophysics of parallel pro jections Per

Ogilvie and Daicar Ogilvie J and Daicar E The p erception of curvature Canadian Journal

of Psychology

Olmer Olmer P Perspective artistique Plon Paris

Ong and Kessinger Ong J and Kessinger D J Perception of verticality with a ro d and frame

apparatus American Journal of Optometry Archives of American Academy of Optometry

Perkins Perkins D N Cubic corners Quarterly Progress Rep ort MIT Research Lab

oratory of Electronics

Perkins Perkins D N Visual discrimination b etween rectangular and nonrectangular paral

lelopip eds Perception Psychophysics

Perkins Perkins D N Comp ensating for distortion in viewing pictures obliquely Perception

and Psychophysics

Pirenne Pirenne M H Optics painting and photographyCambridge University Press New

York

Bibliography

Pontriagin Pontriagin L S The mathematical theory of optimal processesInterscience Pub

lishers New York

Press et al Press W H FlanneryBP Teukolsky S A and Vetterling W T Numerical

Recipes in C The Art of Scientic ComputingCambridge University Press Cambridge UK

Raushenbakh Raushenbakh B V Prostranstvennye postroenya v zhivopisiNauka Moskva

in Russian

Raushenbakh Raushenbakh B V Sistemy perspektivy v izobrazitelnom iskusstve obshchaia

teoriia perspektivyNauka Moskva in Russian

Reggini Reggini H C Persp ective using curved pro jection rays and its computer applica

tion Leonardo

Rosinski and Fab er Rosinski R R and Fab er J Comp ensation for viewing p ointinthe

p erception of pictured space In Hagen pages

Rosinski et al Rosinski R R Mulholland T Degelman D and Farb er J Picture p ercep

tion An analysis of visual comp ensation Perception Psychophysics

Shepard Shepard R N Psychophysical complimentarity In Perceptual Organization

Kub ovy M and Pomerantz J R editors pages Lawrence Erlbaum Asso ciates Pub

lishers Hillsdale New Jersey

Smith a Smith O W Comparison of apparent depth in a photograph viewed from two

distances Perceptual and Motor Skil ls

Smith b Smith O W Judgements of size and distance in photographs The American

Journal of Psychology

Smith Smith Patricia C S O W Ball throwing resp onses to photographically p ortrayed

targets Journal of Experimental Psychology

Stark Stark F Das Netzhautbild Verfahren zur Herstel lung des wahren Sehbildes nach dem

Grundprinzip des menslichen Sehens angewandt auf die zeihnerische Konstruktion der perspective

by the author Neuss am Rhein

Bibliography

Sullivan Sullivan M The Meeting of Eastern and Western ArtUniversity of California Press

Berkeley and Los Angeles

Taylor Taylor A J P The First World War an il lustratedhistory Hamish Hamilton

London

Thouless a Thouless R H Phenomenal regression to the real ob ject i British Journal of

Psychology

Thouless b Thouless R H Phenomenal regression to the real ob ject ii British Journal of

Psychology

Thouless Thouless R H Individual dierences in phenomenal regression to the real ob ject

British Journal of Psychology

Warner Warner F W Foundations of dierentiable manifolds and Lie groupsvolume of

Graduate Texts in Mathematics Springer New York reprint edition

Watt et al Watt R J Ward R M and Casco C The detection of deviation from straight

ness in lines Vision Research

Zorin and Barr Zorin D and Barr A H Corection of geometric p erceptual distortions in

pictures In Computer Graphics SIGGRAPH Proceedings