VISion Res. Vol. 31, No. II, pp. 1923-1930, 1991Printed in Great Britam. All rights reserved

0042-6989/91 $3.00 +0.00Copyright © 1991 Pergamon Press pIc

THE ROLE OF COLOUR AS A MONOCULAR DEPTH CUE

TOM TROSCIANKO,l RACHEL MONTAGNON,2 JACQUES LE CLERC,3 EMMANUELLE MALBERT3

and PIERRE-LoUIS CHANTEAU3

IPerceptual Systems Research Centre, Department of Psychology, Umversity of Bristol, 8 WoodlandRoad, Bristol BS8 ITN, 2Department of Anatomy, University of Bristol, University Walk, Bristol BS8

lTD, England and JESIEE 2, Boulevard Blaise Pascal, B.P.99, 93162 Noisy-Ie-Grand, France

(ReceIVed 27 July 1990; in revised form 18 February 1991)

Abstract-Does colour information playa role tn the perception of depth? Its input to stereopsis is weak,and it has been suggeted that depth from monocular cues, such as texture gradients, is also abolishedat isoluminance (colour contrast with no luminance contrast). We first investigated whether depth fromtexture gradients disappears at isoluminance. The percept remained unaltered. Further experimentsrevealed that certain colour gradients (at isoluminance) markedly affected the perceived depth. A gradientin saturation (e.g. red-to-grey) was particularly effective, whereas a red-green hue gradient had no effecton perceived slant. We concluded that colour information can be used by the visual system to encodedepth, especially in situatIOns where the visual environment is rich in cues which could be used to signaldepth in this way.

Colour Monocular depth perception Texture Texture gradients Isoluminance

INTRODUCTION

The human visual system expends much effortgenerating a stable three-dimensional perceptfrom two-dimensional retinal images. In stere­opsis it uses the angular disparities between setsof points in each monocular image and musttherefore solve the "correspondence problem".Although colour might provide a useful con­straint here, experimental evidence shows itdoes not lead to a strong perception of depth inhuman vision (Lu & Fender, 1972), althoughthere is some evidence that colour does have aweak input to stereopsis (De Weert & Sadza,1983; Grinberg & Williams, 1985). However,the visual system can estimate depth frommany monocular depth cues, such as perspec­tive, texture-gradient, and occlusion for staticimages; and motion parallax for moving ones.Livingstone and Hubel (1987) claimed that themonocular cues also fail to elicit depth at isolu­minance, and that motion perception is im­paired under these conditions (see Cavanagh,Boeglin & Favreau, 1984; Troscianko & Fahle,1988-but this issue will not be consideredfurther in this paper). However, some studiescast doubt on Livingstone and Hubel's claimthat no depth can be seen under monocularviewing conditions. Cavanagh (1987) showed

that line drawings showing occlusion and/orperspective can be seen in the correct manner atisoluminance. On the other hand, shading infor­mation and subjective contours did not givestable depth percepts. Furthermore, Zimmermanand Cavanagh (1990) found that slant percep­tion was unaffected at isoluminance in slantedplanes where the border of the frame was visibleand a (random) texture was embedded in theplane. This study investigated the generation ofdepth from surface detail alone, without theadded information from occluding contours ofany other source.

There were two separate questions asked inthis study. First, we wanted to see whetherquantitative data from naive subjects wouldsupport Livingstone and Hubel's argument thatno depth is seen in monocular displays having,say, a texture gradient or perspective. This isequivalent to asking whether colour can act asa carrier for the spatial information required toelicit such depth perception. Secondly, ifit turnedout that colour can act as an appropriate carrierfor such spatial information, we wanted to askwhether colour can, in itself, contribute tomonocular depth perception. Thus, the secondquestion was whether the visual system canuse colour gradient information to estimatedepth, for a fixed spatial pattern. The visual

1923

1924 TOM TRosclANKo el at.

environment contains colour-gradient cues torelative distance: for example, atmospheric scat­tering renders distant colours less saturated thanclosely-viewed ones. This cue is effective overdistances of many kilometres where other visualmechanisms (such as stereopsis and motionparallax) would not be expected to provideinformation about depth.

METHODS

Stimuli were generated using a Pluto IIgraphics system driven by an IBM-PC compat­ible computer. This graphics system gives 8-bitlevel resolution to each of the R, G, and B gunson a Digivision monitor, model CD14 3112 H3.The frame rate was 50 Hz. Using a chinrest,subjects viewed the display monocularly, usingtheir right eye, at a distance of 1 m, through atube lined with black velvet. This eliminated therest of the visual field (including the edges of theTV monitor) and hence cues to the true fronto­parallel orientation of the display. It should bepointed out that any residual cues (raster lines,specks ofdust, and any specular reflections fromthe screen) would code zero slant. Thus, anyperceived slant found in our experiments wouldnot be expected to arise from these artefacts. Inpilot experiments (with a luminance-modulateddisplay) perceived depth reduced with viewingtime. An electro-mechanical shutter with a 2 secon, 2 sec off cycle rendered the depth perceptstable over time and was therefore used in allexperiments.

The isoluminant point between the red squaresand the white background was found by flickerphotometry at a rate of 25 Hz for each subject.The procedure (described in detail in Troscianko& Low, 1985) gives an approximate indicationof the iso1uminant point; it is then necessary tobracket this in small steps to be sure of includingeach subject's actual isoluminance point for thatpattern. Note that each subject produced aslightly different absolute match point (typicalbetween-subject variation was about 2% incontrast). However, the subject's individualisoluminance point was then used as a referencefor that subject and is represented, for example,by the zero value of the abscissa of Fig. 2.Estimates of perceived depth were made asfollows: with their left arms, subjects slanted ahinged plane to match the perceived slant of thescreen and this inclination was measured by thecomputer through a 8-bit analogue-to-digitalconverter. The range of slants available between

the end-stops on the hinged plane was about- 30 to +90 deg (a negative slant means thatthe top of the plane was closer than the bottom;a slant of zero means the plane was fronto­parallel; a positive slant means that the top wasfurther than the bottom). Thus, there was noparticular apparatus cue to the position of thevertical, although of course gravity would stillprovide such a cue. The intention was to avoidceiling effects around zero slant. Stimuli werepresented in random order, 60 times each. Eachstimulus was presented for as many 4 sec presen­tation cycles as was necessary for the subjectto match the perceived slant; the subject presseda pushbutton to indicate that the match wassatisfactory, whereupon the computer stored theslant value and generated the next stimulus.Four subjects were used: two authors and twonaive subjects.

There were two main experiments, and a con­trol experiment. Experiment 1 asked whetherslant is perceived in an isoluminant texture-gradi­ent display. The control experiment was aimed atestablishing whether any slant found in Exper­iment 1 could arise due to a failure of achievingisoluminance (i.e. could be due to residual lumi­nance information in the image). Experiment 2asked whether colour gradients, can, in them­selves, give a perception of slant, either whenthere is also a texture gradient present (in whichcase the colour gradient can be added either tooppose or enhance the texture gradient) or whenthere is no texture gradient in the image. Thedetails of the colours and gradients used in thestimuli are given in the captions to the figuresshowing the stimuli used (Figs 1 and 3).

RESULTS

Figure 1 shows a typical texture-gradientstimulus used in the first experiment. Theindependent variable was the luminance of the(neutral) background. The luminance of the(red) squares was kept constant. Half the stimulihad no texture gradient, and were entirely com­posed of red squares similar in size to those inthe middle row of Fig. 1.

Figure 2 gives results for four subjects viewingthese texture-gradient stimuli. Slant estimatesdo not vary with luminance contrast, even atisoluminance. (Stimuli with no texture gradientgave slants of about zero.) Note that there areindividual differences between subjects but thefunction for each subject is essentially flat.These results suggest that colour information

'C>

IV U>

Fig

.I.

Ate

xtur

egr

adie

nto

fsq

uare

sde

crea

sing

insi

zew

ith

incr

easi

nghe

ight

onth

esc

reen

.S

ubje

cts

view

edth

edi

spla

ym

onoc

ular

lyw

itho

utse

eing

the

edge

so

fth

esc

reen

.W

ith

thei

rle

ftar

m,

they

adju

sted

asl

anti

ngpl

ane

tom

atch

the

perc

eive

dsl

ant

of

the

disp

lay.

The

lum

inan

ceo

fth

ere

dsq

uare

sw

as4

cd/m

2 •T

hech

rom

atic

ity

of

the

squa

res

was

x=

0.55

,y

=0.

32,

and

of

the

neut

ral

back

grou

ndx

=0.

26,

y=

0.24

(Min

olta

chro

ma

met

er).

(A)-·-------------'----...---....-mm....--n.n.r.n•••n•••••_UtS)

-If

-,(D) i?tllll, ,

Fig. 3. Stimuli used in second experiment. From top left, (A) Texture gradient, regular pattern; (B) Notexture gradient, irregular pattern; (C) Texture gradient, irregular pattern; (D) No texture gradient, regularpattern. The experiment compared perceived slant between a uniform red pattern and one having the samespatial structure but a gradient in saturation (e.g. red-grey), hue (e.g. red-green), or luminance (e.g.red-black). Chromaticities of the colour gradients were as follows: red-grey from x = 0.60, y = 0.35(bottom row) to x = 0.26, y = 0.25 (top row); red-green from x = 0.61, y = 0.35 (bottom row) to x = 0.30,y = 0.49 (top row). All displays except the luminance-gradient stimuli were isoluminant with a luminance

value of 4 cd/m2•

1926

The role of colour as a monocular depth cue 1927

Mean differences30

1m! r-black

25 • r-grey

j 20 Dr-green.,uc 15~

~,~:0 10c0

~ 5

0

-5A B C 0

Texture condition

Fig. 4. Results of second experiment. The ordinate showsthe difference in perceived slant in the "gradient" conditioncompared to the uniform red condition. Thus, a positivevalue shows that the gradient increases the amount ofperceived depth. Three kinds of gradient were investigated:(I) red-black (luminance gradient); (2) red-grey (saturationgradient) and (3) red-green (hue gradient). Further experi­ments showed that a reversed gradient reduces the perceived

depth. Standard deviations on each mean are shown.

dow" containing what looked like scratchmarksof random length. Thus, the colour informationwas effective in producing perceptual segregationof the two planes, of which one was defined bycolour and the other by luminance.

The second experiment investigated the effectof colour gradients on the perception of slant.Figure 3 shows the types of stimulus used in thisexperiment. There were four spatial configur­ations, labelled A, B, C and D. A consisted ofa texture gradient and a regular pattern (withperspective cues). B had no texture gradient,and positional noise was added to make thedisplay irregular. C was an irregular patternwith a texture gradient (but no perspective cuesdue to the irregularity). D was a regular patternwithout a texture gradient.

Figure 4 shows the main results obtained inExperiment 2. The data are presented as a differ­ence in matched slant between a pattern with theappropriate colour gradient, and the same spatialpattern with a uniform colour whose chroma­ticity was that of the middle row of the colour­gradient pattern. All displays in Experiment 2were presented at isoluminance, with the obviousexception of the luminance-gradient condition(where the mean luminance was equated to thatof the isoluminant stimuli). The data of Fig. 4are presented such that a positive value indicatedthat the gradient in question added to the per­ceived slant (i.e. the top seemed proportionatelymore distant than the bottom) whereas a nega­tive value indicated that the gradient reduced theamount of slant seen. Differences, rather than

25-15 -5 5 15

Luminance controst ("!o)

oL-_~__L-_-L__-'---_-...J

-25

Lum contrast and perceived tilt50

--.-+-+-+-t+___+-:_x~x-x---

40 x--x-

j ~30

::::;:;

i 20.<=~ •___a-a-a-·-.___•0:::;:

• TT X RM10

o JH + NM

Fig. 2. Mean results for four subjects showing the relation­ship between perceived slant and luminance contrast. Theordinate labelled "0" is each subject's isoluminant point asdetermined by flicker photometry. There is no reduction in

perceived slant at this point.

can code depth arising from texture gradients.However, it has been argued that all isolurninantdisplay have some residual luminance inform­ation (Livingstone & Hubel, 1987; Troscianko& Fahle, 1988), implying that the depth in thesedisplays might have arisen from low-contrastluminance edges around each square, perhapsin some part of the visual field only, since theabove authors have also argued that it is difficultsimultaneously to achieve isoluminance in allparts of a large display. In the contol experiment,we measured the amount of perceived slant atisoluminance and then the amount of luminancecontrast needed to produce the same amount ofslant in a monochrome display. Both (naive)subjects used in this experiment needed a lumin­ance contrast of around 30% to give the sameamount of depth as in the isoluminant display.Errors of isoluminance are typically around4% contrast (Troscianko, 1987), i.e. an orderof magnitude less than the amount of contrastrequired for this effect. It is therefore unlikelythat such errors can account for the depth seenin these displays. Note that additional luminancemasking-noise (high contrast white stripes ofrandom length and position, oriented verticallyand horizontally) was added to the display in thecontrol experiment in order to raise the lumin­ance threshold and effectively mask any weaknon-isoluminant edge effects of the sides of thesquares-but the same noise was also added tothe isoluminant display, and hardly reduced theperceived slant, again implying that the depthwas being coded by colour, and not residualluminance. Indeed, in the isoluminant case withadded masking noise, subjects reported seeinga slanted plane through a (frontoparallel) "win-

1928 TOM TROSCIANKO et al.

absolute measures, are shown here because of thebetween-subject variation in absolute matchedvalues which has been seen in Fig. 2. We foundthat the matched slant differences gave a stablemeasure across subjects. The results for thesaturation gradient and the luminance gradientare clear-cut. Adding each of these gradients suchthat the top of the display is less saturated ordarker than the bottom enhances the perceptionof depth. This effect is independent of whetherthe pattern already has a texture gradient or not.Note, however, that a red-green hue gradientdoes not produce a change in perceived slant(unlike gradients of saturation and luminance).Informal experiments not reported in detail heresuggest, moreover, that a textured field of somekind is necesary to elicit a perception of depthfrom a colour gradient. In other words, a uni­form field changing from red at the bottom togrey at the top does not look slanted. Finally,we found that adding a reversed saturationgradient (i.e. red at the top, grey at the bottom)to a texture-gradient display could cancel theperceived slant, making the display look fronto­parallel. This nulling possibility enhances thevalidity of the slant-matching psychophysicalmethod.

DISCUSSION

The results of Experiment 1 suggest that slantarising from texture gradients is clearly perceivedat isoluminance; indeed, there is no reductionin perceived slant between the isoluminantcondition and conditions in which there was aluminance contrast present in the image. Thesefindings are markedly different from thosereported by Livingstone and Hubel (1987). Wefeel that the perceived slant at isoluminance isunlikely to have arisen as a result of errors ofachieving isoluminance, as shown by the resultsof the control experiment. The difference inresults may possibly be attributed to the fact thatwe eliminated context cues such as the rectangu­lar surround of the video monitor, and that weused a shutter to prevent total adaptation tothe slant after prolonged viewing. Finally, weobtained quantitative data using a slant-match­ing technique rather than relying on binaryverbal reports (slant vs no-slant).

If the results of Experiment 1 suggest thatcolour is a useful vehicle for encoding slant fromtexture gradients, the results in Experiment 2suggest that colour can, in itself, code monoculardepth. A saturation gradient is particularly effec-

tive at achieving this. The amount of depth thusgenerated seems to depend only on the colourgradient and not on whether the colour gradientis superimposed on a texture gradient or a flattexture. If the colour gradient is introduced inopposition to the texture gradient, cancellation ofdepth can result if the parameters are accuratelybalanced. In general, the difficulties of achievingsuch a balance prevented us from using this asan experimental technique here, but such a seriesof experiments is certainly possible in principle.Finally, we established (albeit informally) thata texture of some kind is necessary for colourto elicit depth-a uniform colour gradient(e.g. a rainbow) is not sufficient. This suggeststhat, while colour can code depth, its contribu­tion is contingent upon the presence of texturecues. Such a contingency implies strong linksbetween texture and colour processing in humanvision.

Why should colour information be used indepth perception? In landscape, more distantcolours appear much less saturated. This iscaused by atmospheric scattering. The presentstudy has found that saturation and luminancegradients, which do occur in natural scenes,produce a perception of slant. On the otherhand, a red-green hue gradient, which does notoccur in nature, does not change perceivedslant. Strong monocular depth effects have beenreported under water (Ross, 1967) where diversperceive distant objects as being larger than theywould otherwise be for a given retinal subtense.The "constancy scaling" hypothesis (Emmert'sLaw) predicts the direction of the illusion. Theexplanation is given in terms of a reduction inbrightness contrast, but our findings suggest thata change in colour could be important as well.It therefore appears that variations (gradients)in colour can in themselves code depth in theabsence of stereo and convergence cues. Al­though our viewing distance was 1m, our displayeliminated stereo, convergence and context cues.Thus, the subjects saw a display whose distancewas inherently ambiguous. Several subjectsspontaneously reported that it was like lookingat terrain several miles away. Ifcolour is a depthcue, it would be reasonable for the cue to act overlarge distances since it is under such conditionsthat stereopsis and convergence cues are notavailable to the system.

Acknowledgements-We thank the Medical Research Councilfor support, Mark Georgeson for comments, and SusanBlackmore for help in preparing the manuscript.

The role of colour as a monocular depth cue 1929

REFERENCES

Cavanagh, P. (1987). Reconstructing the third dimension:Interactions between color, texture, motion, binoculardisparity, and shape. Computer Vision, Graphics, and ImageProcessing, 37, 171-195.

Cavanagh, P., Boeglin, J. & Favreau, O. E. (1985). Perceptionof motion in equiluminous kinematograms. Perception, 14,151-162.

De Weert, C. M. M. & Sadza, K. J. (1983). Newdata concerning the contribution of colour differencesto stereopsis. In Mollon, J. D. & Sharpe, L. T.(Eds), Colour Vision (pp. 553-562). London: AcademicPress.

Grinberg, D. L. & Williams, D. R. (1985). Stereopsis withchromatic signals from the blue-sensitive mechanism.Vision Research, 25, 531-537.

Livingstone, M. S. & Hubel, D. H. (1987). Psychophysicalevidence for separate channels for the perception of form,

colour, movement, and depth. Journal ofNeuroscience, 7,3416-3468.

Lu, C. & Fender, D. H. (1972). The interaction of colourand luminance in stereoscopic vision. InvestigativeOphthalmology, 2, 482-489.

Ross, H. E. (1967). Water, fog, and the size-distance invari­ance hypothesis. British Journal ofPsychology, 58,301-313.

Troscianko, T. (1987). Perception of random-dot symmetryand apparent movement at and near isoluminance. VisionResearch, 27, 547-554.

Troscianko, T. & Fahle, M. (1988). Who do isoluminantstimuli appear slower? Journal of the Optical Society ofAmerica, A5, 871--880.

Troscianko, T. & Low, I. (1985). A technique for presentingisoluminant stimuli using a microcomputer. Spatial Vision,1, 197-202.

Zimmerman, G. L. & Cavanagh, P. (1990). Slant perceptionat equiluminance. Investigative Ophthalmology and VisualScience Suppl., 31, 411.