Stereopsis without Familiarity Cues

Binocular Depth Perceptionwithout Familiarity Cues

Random-dot stereo images with controlled spatial andtemporal properties clarify problems in stereopsis.

Bela Julesz

Research in stereopsis (1) is tradi-tionally devoted to quantifying the re-lationship between disparity (2) andperceived depth. Problems of thehoropter, perceptual limits of disparity,the metric of the perceived space, andso on, are all examples of this classicalproblem-posing and have been thor-oughly investigated (3). Strangelyenough, the related problem of howdisparity is derived-that is, how thecorresponding left and right retinalprojections of an object are found-has been ignored. This lack of interestis the more remarkable since thematching of the horizontally shiftedcorresponding point domains in theleft and right fields is accomplished al-most without deliberation, althoughthese point domains generally differ inbrightness and shape (owing to reflec-tions and perspective). Perhaps the in-herent limitations of the stimuli usedmay have caused researchers to shyaway from studying binocular depthperception as a pattern-matching proc-ess. Indeed, simple line drawings weretoo limited for the exploration of pat-tern matching, while real-life pictureswere unsatisfactory because of themany complex familiarity cues whichinteracted in uncontrollable ways.

Four years ago I posed two intimate-ly related questions along these lines,which constituted a new paradigm.They were: (i) Would it be possible tocreate an artificial sensory environmentdevoid of all depth and familiarity cuesexcept disparity? (ii) Could depth stillbe perceived under these conditions of

The author is a member of the technical staffin the Human Information Processing Depart-ment, Bell Telephone Laboratories, Murray Hill,N.J.

356

"familiarity deprivation" (4, 5)? Thisparadigm was never systematicallyraised before, and yet it is so familiar.It is a long-known fact, exploited inaerial reconnaissance, that objects cam-ouflaged by a complex background are

very difficult to detect monocularly butjump out if viewed stereoscopically.Nevertheless, despite the difficulty, thehidden objects can be monocularly de-tected. Even if every surface of thethree-dimensional environment were

covered with a homogeneous randomtexture, the closer surfaces would seemto have coarser granularity than theones farther away. [This retinal gra-dient of textures which is attributableto perspective yields a strong monocu-

lar depth cue (6).] Therefore the ques-tions of whether an environment canbe ideally camouflaged and of whetherobjects that are hidden when viewedmonocularly can be perceived in depthstill remained to be answered.

In order to obtain such an answera novel technique of random-dot stereoimages was devised. Such a stereo pairis shown in Fig. 1. When viewedmonocularly, both fields of Fig. 1 givea homogeneous random impressionwithout any recognizable features. Butwhen viewed stereoscopically, thisimage pair is vividly perceived in

depth, with a center square in frontof its surround. [A prism in front ofone eye greatly facilitates fusion of thestereo pairs. A satisfactory prism can

be made of gelatin, as described in(7).]The emphasis, in this brief article,

is on demonstrating this and similarrecently observed perceptual phenom-ena, with comments on certain impli-cations of these findings for stereopsis.

Figure 2 illustrates how Fig. 1 andsimilar random-dot stereo images (par-ticularly Fig. 3) are generated. It rep-resents a small stereo pair composedof a matrix of 9 X 10 picture elements.The equally probable randomly se-lected black and white picture ele-ments which are contained in corre-sponding areas in the left and rightfields are labeled in three categories.(i) Those contained in correspondingareas with zero disparity (which whenviewed stereoscopically are perceivedas the surround) are labeled 0 or 1.(ii) Those contained in correspondingareas with non-zero disparity (whichwhen viewed stereoscopically are per-ceived in front of or behind the sur-round) are labeled A or B. (iii) Thosecontained in areas which have no cor-responding areas in the other field(that is, project on only one retinaand thus have no disparity) are labeledX and Y. The 0 and 1 picture ele-ments are identical in correspondingpositions of the two fields. The posi-tions of the A and B picture elementsbelonging to corresponding areas in thetwo fields are also identical, but areshifted horizontally as if they were asolid sheet. Because of this shift someof the picture elements of the surroundare uncovered and must be assignednew brightness values (X and Y). Sincethese areas lack disparity, they can beregarded as undetermined in depth.Figure 2 contains three rectangles inthe left and right fields, composed ofA and B picture elements. Each fieldcontains an upper, middle, and lowerrectangle which can be regarded ascorresponding left and right "projec-tions" of a rectangular planar surfacelocated in depth when viewed fromdifferent angles. The projections of theupper rectangle (that is, the corre-sponding upper rectangles in the leftand right fields) are horizontally shiftedrelative to each other in the nasal di-rection by one picture element, thecorresponding lower rectangles areshifted in the temporal direction to thesame extent, while the correspondingmiddle rectangles have a one-picture-element periodicity and may be re-garded as being shifted in either direc-tion. The low density of picture ele-ments and the large disparities wouldprevent stereopsis in a pattern corre-sponding to Fig. 2. In order to achievestereopsis, the number of picture ele-ments would have to be increased con-

SCIENCE, VOL. 145

siderably. For this reason a computeris used.

It would be impractical to gen-erate without a computer ade-quately complex stimuli of severalthousand brightness elements and givenconstraints for each experiment. Figure3 is a computer-generated version ofFig. 2; it has 100 X 100 picture ele-ments, and for each of its three rec-tangles the disparity is six picture ele-ments. The disparity is always chosento be an integral multiple of the widthof the picture element; therefore, whenviewed monocularly, each of the twofields gives an impression of homoge-neous randomness, without gaps orboundaries (4, 5). The upper rectangle,when the images are fused, is seenin front of the surround (as in Fig. 1).The lower rectangle is perceived be-hind the surround, while the middlerectangle can be seen in front or be-hind at will.Of course, instead of two brightness

levels, any number of levels can beintroduced, and instead of rectangularsurfaces, any complex surface can beportrayed by this technique and giverise to stereopsis (8). Minimum areasize, dot density, disparity, perceptiontime, number of brightness levels, andother factors show strong interde-pendencies, which can be explained bystatistical analysis. But before consid-ering such interdependencies, I discusssome interesting observations whichcan be readily made when Figs. 1 and3 are viewed stereoscopically.The main result of these observa-

tions is that the paradigm mentionedabove is answered in the affirmative:Stereopsis can be obtained in the ab-sence of monocularly recognizable ob-jects or patterns. As a consequence,the many depth cues for monocularvision-cues such as the apparent sizeof familiar objects, interposition (thesuperimposing of near objects on farobjects), and linear perspective-whichin a familiar environment strongly in-fluence the final percept, do not op-erate here. In this case the complexpattern-recognition processes (whichthemselves are based on involved learn-ing and memory processes) can beovercome, and this greatly simplifiesthe study of binocular depth percep-tion. The problem is reduced to thatof finding how similar patterns arematched.

It is important to note that in theseobservations the quality of stereopsis(that is, the time required for stereop-24 JULY 1964

sis, the stability of the fused image,the amount of binocular rivalry, andso on) is excellent in spite of theabsence of all other depth cues. As amatter of fact, since every picture ele-ment has disparity (in contrast toordinary pictures, which contain largehomogeneous areas without depth in-formation) the random-dot stereoimages are usually easier to perceivein depth. For these effective stimuli,

several quantitative limits of variousparameters which were determined asborderline values for stereopsis can beextended.

It should also be mentioned that thestatistical, topological, and heuristicproperties of the random-dot stereoimages are controlled by the experi-menter; thus the observations are moreamenable to analysis.The basis of stereopsis is disparity,

Fig. 1. Basic random stereo pair. When the two fields arecenter square appears in front of the background. [See (7)useful in stereoscopic viewing.]

I Y0101 rl1 1 |101I_lIxA_B _1 1 Y B A BIA IllOi 1 B 1B 1 1 01OOOBIABIAB 110

I O AAVBAWX7 1_ 1 B B A Bi I1 O

Fig. 0xi0. iofh ed

Fig. 2. Illustration of the method by

viewed stereoscopically, thefor a description of an aid

10°1 o1 1° 1 °ll10 oA A B B Y

lol1 IOIB AIBAIXIllTO°lil1 ° ° 1 1 1 1 TO 111 1 B ABABOO0Q O A1BIAIATI 0111 0 1 0 1 11001 OY A A BAO 11 IY1 YBIBAIB -1 10100111 1

which the stereo pair of Fig. 3 was generated.

Fig. 3. Stereo pair which, when viewed stereoscopically, contains an upper rectangleperceived in front of the surround, a lower rectangle perceived behind the surround, andan ambiguous middle rectangle perceived either in front of or behind the surround.

357

as was demonstrated by Wheatstonewith his stereoscope (9). Nevertheless,there are special instances when depthcan be perceived in the absence ofdisparity. An example is the Panumphenomenon (10), where one image ofa stereo pair consists of two parallelvertical lines in a homogeneous sur-round while the other image containsa single vertical line. When one of thevertical lines in the two images isfused (when the images are viewedstereoscopically), the other line, forwhich there is no corresponding rep-resentation in the other member of thepair, is also perceived in depth; it hasa somewhat "floating" look but appearsclearly behind the fused line. Suchstimuli are particularly unsuitable forgetting better insight into this phenom-enon since they are "simple" only froma most irrelevant point of view-theyare simple to draw. In fact, line draw-ings are degenerate forms of real-lifeimages (which are composed of objectswith textured surfaces) and as a resultthe perceptual performance to be stud-ied becomes needlessly complicatedand disguised. Indeed, the areas inFigs. 2 and 3 which are without dis-parity (the areas represented in Fig. 2by X and Y and in Fig. 3 by thecorresponding dots) are a generaliza-tion of Panum's unpaired line, andtheir perception (which is quite sta-ble) can be simply described and ex-plained: Undetermined areas (areaswithout disparity) are perceived at thedepth of the most distant adjacent de-termined area (area with disparity)(4, 5). This rule is illustrated in Fig. 3,where undetermined areas at the leftand right edges of the rectangle thatis seen in front when the images areviewed stereoscopically are perceivedas continuations of the background,while, for the rectangle seen behindthe surround, the undetermined areasare perceived as belonging to the rec-tangle. (This is the reason why thelower rectangle looks wider than theupper one.)

This perceptual phenomenon is inagreement with the common experi-ence that the image of each point ofa closest surface is projected on bothretinas, whereas a surface behind ithas points which are totally or partlyhidden. Thus, an area which is partlyhidden and represented only on oneretina is perceived as a continuationof the exposed parts of the surface be-hind the superimposed one. This effectis even more apparent for random-dot

358

patterns, since the undetermined andthe determined areas which are per-ceived as being at the same depth haveidentical textures.The Panum phenomenon can be ap-

proached also in the context of binoc-ular rivalry. In this interpretation, thedetermined area exerts an additionalstabilizing effect, namely the preven-tion of binocular rivalry in theproximate undetermined areas. Aremark on the implications of thosefindings for Gestalt psychology is givenin (11).

Stereopsis under Brief Exposures

Besides disparity, there are two sec-ondary depth cues for binocular vi-sion: convergence and correlative ac-commodation (differential focusing).Both depend on muscle action. SinceDove in 1841 (12) demonstrated ster-eopsis under very brief exposures(much too short for any muscle ac-tivity), the importance of focusing andconvergence is regarded as negligible.(In addition, the fact that, in a stereoimage, areas in front and behind canbe simultaneously perceived is hard toexplain in terms of convergence.)Therefore, contrary to naive belief,stereopsis is the result of central ner-vous system processing, and the mainpurpose of convergence is the coarsealignment of corresponding retinalareas. This coarse alignment insuresthat the corresponding retinal areas arewithin the region of patent stereopsis.This does not mean that, for longerexposures, convergence motions andproprioceptive influences might not af-fect stereopsis. Dove's result, and manysimilar findings since (13), have con-clusively demonstrated that stereopsiscan occur as a result of central ner-vous system processing alone, and thisview is generally accepted by workersin this field (3). However, von Karpin-ska (14) believed that these tachisto-scopic experiments were successfulonly when the subject knew before-hand what he was expected to see.This and similar arguments are stillvoiced, and therefore the finding (15)that random stereo images (such asthat of Fig. 1) can also be perceivedin depth under conditions of tachisto-scopic presentation is not without in-terest. Since, in such experiments, thesubjects have no familiarity with thestimulus at all and nevertheless, in a1-millisecond exposure, correctly per-

ceive the middle square in front orbehind (when the two cases are pre-sented in a randomly mixed order), themost plausible objection to Dove's find-ing is removed.

Perhaps an even more importantconsequence of the finding that ran-dom stereo pairs are perceived in depthin brief exposures is that Hering's the-ory on the role of double images isdisproved. According to this theory,images not fused are seen double andare crossed or uncrossed depending onwhether they lie in front of or behindthe point of convergence. The extentto which this cue is utilized could notbe previously determined, since doubleimages were inseparable from the ap-plied stimuli. The forms in randomstereo pairs, on the other hand, arenot recognizable until the forms areperceived in depth, and thus it is im-possible to perceive double images eith-er before or after fusion (15).

It is surprising that, without second-ary depth cues, space sense can de-velop so rapidly (the effective presenta-tion time is longer than the flashes,due to the persistent afterimages, butis, nevertheless, very brief). The timerequired for stereopsis increases withlarger parallax shifts, smaller area size,and more complex (that is, nonplanar)surfaces.

Perception and Attention Time

These tachistoscopic experimentswere useful only for studying percep-tual performance in the absence of eyemotions. In order to get better insightinto the temporal aspects of percep-tion, particularly into the perceptiontime required, the afterimages have tobe "erased." A new "stereo erasing"technique was developed, in which theerasing stimulus is a random-dot stereopair.

This new technique also utilizes theambiguous depth phenomenon (ran-dom wallpaper effect) which is demon-strated in Fig. 3 (16). If the upperrectangle perceived in front of the sur-round is viewed first, then the am-biguous middle rectangle is also seenin front of the surround. On the otherhand, if the lower rectangle perceivedbehind the surround is viewed first,then the ambiguous middle rectangleis seen behind the surround. This find-ing holds for tachistoscopic exposurestoo, and thus is not the result of thesubject's maintaining the same con-

SCIENCE, VOL. 145

vergence but is, rather, the result ofhis maintaining attention for the sameperceptual organization (depth plane).

In these tachistoscopic experiments,brief presentation of an unambiguousstereo pair (a pair having a centersquare with either a temporal or anasal disparity) was followed by pre-sentation of an ambiguous stereo pair(a pair having the same disparity butin both directions). The picture ele-ments in the second stimulus differedfrom those in the first. Thus, the sec-ond stimulus erased the afterimages ofthe first stimulus, and therefore thereal presentation time for the firststimulus was known. It was found that,when presentation time was adequate,the second, or ambiguous, stimulus wasconsistently perceived at the samedepth as the first, or unambiguous,stimulus. (The unambiguous stereo pairwas presented with temporal or nasaldisparity, in mixed order.) Perceptionof the ambiguous stimulus was influ-enced by perception of the unambigu-ous stimulus even when the first stim-ulus was not consciously perceived.When the first stimulus was presentedfor a time shorter than this "percep-tion time for stereopsis," or when thesecond stimulus was delayed by aninterval longer than the "attentiontime," the second stimulus became in-dependent of the first and could beperceived as having depth opposite tothat of the first. This finding and thefact that perception and attention timeswere typically under 50 milliseconds(17) make it appear most unlikely thatconvergence motions might have beeninitiated. The subjects were unawarethat the second stimulus was ambigu-ous. These facts imply that the firststimulus serves as a "depth marker"and determines which of the possibledepth organizations should be attendedto. Such an internal attention mecha-nism was nicely demonstrated byPritchard when viewing the reversal ofa Necker-cube under conditions of ret-inal stabilization (18). [The problem ofwhether this mechanism is a parallelor a sequential process is discussedin (19).]

Binocular Similarity

In the experiments summarizedabove, pattern-matching consisted ofthe relatively simple task of findingidentical patterns in the two fields, dif-fering only in their horizontal posi-24 JULY 1964

Fig. 4. Stereo pair identical to that of Fig.panded uniformly in both dimensions by

tions. In the experiments reportednext the congruency of the correspond-ing patterns was perturbed to variousextents by several manipulations. Thereare many ways to introduce such dis-tortions. One way is to simulate real-life situations under more controlledconditions. In ordinary binocular vi-sion the two retinal projections aregenerally quite different in brightnessand shape, owing to reflections andperspective. Distortions which simu-lated such vision were introduced byblurring one of the fields, adding un-correlated random noise, expandingone field uniformly, complementingcertain points (by changing black towhite and white to black), and so on.One of the many possible perturbations(4, 5), the uniform expansion of onefield, is illustrated in Fig. 4. Differ-ences in the size of the retinal imagesfor the two eyes (aniseikonia) are neveras great as the size differences of Fig.4; nevertheless, depth is easily per-ceived in Fig. 4, which is derived fromFig. 1 by uniform expansion of oneof the fields (by 10 percent in bothdimensions). Since in these computer-

I except for the fact that one field is ex-10 percent. Stereopis is easy to obtain.

generated stimuli every point contrib-utes to stereopsis, depth can be per-ceived even under tachistoscopic con-ditions if the centers of Fig. 4 arealigned. This means that, in additionto the horizontal disparity, some verti-cal shift can be tolerated. Whenrandom-dot stereo images are used,most quantitative findings on the limitsof disparity as determined with simpleline drawings (3) seem to be very con-servative and can be extended.

In addition to expansions, rotationsof one of the computer-generated ster-eo fields by 7 degrees of arc can giverise to stereopsis during brief expo-sures, a finding which is the more re-markable since the time of exposureis too brief to permit cyclotorsionaleye movements. All these experimentsshow that the central nervous systemhas processing powers far beyond therequirements of common usage.

In the experiments described next Itried perturbations of a complexitywhich never occurs under ordinaryconditions, in order to study the limitsof pattern matching. Two correspond-ing patterns are called "binocularly

Fig. 5. Stereo pair identical to that of Fig. I except for the fact that in the leftfield the diagonal connectivity is broken; 75 percent -of the picture elements of thestereo pair are identical. Stereopsis is easy io obtain.

3S9

Fig. 6. Stereo pair generated by outlining the fields of Fig. and complementing one

of them. Stereopsis is very difficuLlt to obtain.

similar" if they can be fused and per-

ceived in depth. The quality of thepercept may be regarded as an indi-cator of the similarity of the patterns.

One surprising finding was that thenionocular similarity of two patterns

can be quite different from the binocu-lar siniilarity. This is illustrated in Fig.5, which is derived from the basicstereo pair of Fig. 1 by breaking ofthe connectivity along the diagonals inthe left field. If, along the +45-degreeand -45-degree diagonals, three adja-cent picture elements had identicalbrightness values, the middle one was

complemented (that is, was removedfrom fusion). As a result of this pro-

cedure in Fig. 5 only 25 percent ofthe picture elements became comple-niented while 75 percent were keptidentical in the two images. Althoughthe two patterns appear exceedinglydissimilar when viewed monocularly,the binocular similarity is very high,inasmuch as stereopsis is easily ob-tained. This observation has anotherimplication. It has already been provedthat monocularly recognizable objectsare not necessary for stereopsis. Nev-ertheless, one might object that, in Fig.1, similar micropatterns can be per-

ceived in the two fields, as viewedmonocularly, and that these mightserve as the basis for fusion. The factthat the patterns of Fig. 5 look so

different on both a micro and a niacro

level, when viewed monocularly, andthat the images can nevertheless beperceived in depth is strong evidencethat the pattern processing occurs afterthe binocular combination of the stereoimages has occurred. This patternprocessing reveals that 75 percent ofthe picture elements of the two fieldsare identical, a fact disguised by thedissimilarity of the fields when viewedmonocularly (5). This observation-

360

that the processing has to occur afterthe binocular conibination of theiniages-is in agreement with recent

neurophysiological findings by Hubeland Wiesel (20).

It is interesting to note that binocu-lar similarity cannot be described solelyin terms of quantitative point-by-pointidentity between a pair of patterns.

There are several ways of removing thesame percentage of picture elementsfroni fusion, and for these variousways the quality of depth perceptionmay differ greatly. Thus, binocularsimilarity depends greatly on the topol-ogy of the perturbing configurations.One crucial factor in visual perceptionis the connectivity of adjacent ele-ments. The perturbing configurationswhich destroy this connectivity to thegreatest extent in the conibined fieldproduce the greatest perceptual degra-dation (4, 5).

With such techniques many inherentpattern organizations can be studied;one interesting class, involving contourdependencies, is discussed in the nextsection.

Role of Contours in Stereopsis

One of the most comnion beliefsconcerning stereopsis is that contoursare important (3). The usual definitionof contours as boundaries betweenconfigurations that represent recogniz-able objects when viewed monocularlyhas to be modified, since the experi-ments described above illustrated thatstereopsis can be achieved in the ab-sence of such configurations. A con-

tour may be alternatively defined as a

boundary between white and blackclusters. For real-life situations the two

definitions coincide. Belief in the im-portance of contours is based on a

classical experiment by Helniholtz(21). It is a belief which has neverbeen questioned since his day. Helm-holtz used a black line drawing of asimple object in a white surround asone stereo iniage and its conmplement(negative) as the other. In spite ofsome binocular rivalry the stereo paircoLIld be fused. Because the two fieldswere evervwhere different except forthe location of the contours, it wasinferred that contours are crucial forstereopsis. On the other hand, it onefield of Fig. I is complemented, ster-eopsis is destroyed (4). Moreover, itis possible to perceive depth, withoutany binocular rivalry, for randomstereo pairs which are identical every-where except at the contours (15).

These findings seemingly contradictthe results of Helmholtz's experiment.This apparent contradiction arises fromthe spatial complexity of the stimulus.This is illustrated by Fig. 6. To gen-erate Fig. 6, the outline of the patternof Fig. 1 was generated at the boun-daries between black and white clus-ters, and one of the fields was comple-mented. These stereo fields have agreat spatial density of outline, andstereopsis is very difficult. If one re-duces this density by expanding a smallarea in each field, stereopsis is greatlyfacilitated, approximating that inHelmholtz's case.

Discussion

The techniques mentioned abovem1ake it possible to study stereopsis inits purest form. Nevertheless, the studyis limited to problems concerning thesensation of relative depth in a smallregion around the convergence point(22). (How the entire visual space isbuilt up from such regions by succes-sive convergence motions and howthese space samples are integrated ina unique percept are problems far be-vond the scope of this research.)

Under these ideal conditions severalof the observed phenomena can beexplained by relatively simple statisti-cal arguments. As an example, let us

analyze the following finding: For a

given disparity the corresponding pointdomains in the two fields (for example,the center square of Fig. 1) must pos-sess a minimum number of picture ele-ments (and thus be of a certain size)to be perceived in depth. This criticalpoint-domain size increases with in-creased disparity and decreases with

SCIENCE, VOL. 145

increased number of brightness levelsin the stimuli. These experimentalfindings can be easily explained, asfollows. Any two uncorrelated randomimages of black and white picture ele-ments have 50 percent identical ele-ments, by chance alone. With three ormore brightness levels in the stimulusthis chance identity is reduced to 33percent or less. Thus, correspondingpoint domains in the left and rightfields have to contain more correlatedpoints than the cluster formed bychance correlation. Since with a small-er number of brightness levels theprobability increases that noncorre-sponding adjacent dots will form cor-related clusters (false clusters) of con-siderable size, the critical size of cor-responding clusters has to be increased.Only then is the probability negligiblethat a false cluster will occur that issimilar in size to a critical area. If thecorresponding areas are above the crit-ical size, then they need not be identi-cal, only similar. But, to achieve ster-eopsis, this similarity has to be morethan the chance correlation (see Fig.5). The probability of finding largefalse clusters increases as the fields(which contain them) get larger. Thiscorresponds to the observation thatwith increased disparity (that is, withincreased image area to be searchedfor corresponding patterns) the size ofthe critical area has to be increasedto obtain stereopsis.

In this analysis, clusters formed byproximate points of similar (correlated)brightness played a dominant role. Thiscluster formation (or connectivity de-tection) is basic for monocular texturediscrimination too (23) and remindsone of figure and ground discrimina-tion.One can explain many of the ob-

served phenomena by regarding themas a search for connected clusters inthe combined binocular field (4). Inorder to test the validity and powerof these notions, a computer programwas written (called AUTOMAP-1)which complies a three-dimensionalcontour map from high-resolutionstereo images (24). This computersimulation is a sort of active descrip-tion, a model whose form reflectssomething of the structure of the phe-nomena represented, but which alsohas the character of a working ma-chine (25). The results were satisfac-tory (the essentials of this heuristicmodel are given in 4, 23 and 24).Some assumptions in the model were24 JULY 1964

experimentally confirmed by White(26).Random-dot stereo images are now

used in fields other than that of ster-eopsis, for studying optical illusions(27), binocular rivalry (28), and per-ceptual learning (16); some findingsmight bear on subliminal perception.Possible applications range from auto-matic map compilation (24) to clinicaluses [for example, x-ray stereofluoros-copy (29)].The paradigm itself can be general-

ized, and analogue techniques mightbe used for studying apparent motionand skin localization (30). Such a gen-eralization was recently applied in astudy of auditory memory (31). Theserefined techniques may have some im-plications for auditory localization too,where work with correlated auditorynoise was begun as early as 1948 andproduced some interesting phenomena(32).

Summary

The reported phenomena were ob-tained through the use of special tech-niques. (i) All monocular depth andfamiliarity cues were removed fromthe stimuli (through the use of random-dot stereo patterns). (ii) The statisticaland topological properties of the stim-uli were precisely known (since theywere generated according to a specificcomputer program). (iii) Convergencemotions of the eye and proprioceptivecues were eliminated (through the useof tachistoscopic illumination). (iv) Thetime of presentation was under con-trol (through erasure of the persistentafterimages). Under these conditionsstereopsis could be studied in its purestform. It was shown that depth can beperceived in the absence of monoculardepth and familiarity cues and of allbinocular depth cues except for dispar-ity. These findings have important im-plications for some existing theories ofstereopsis and open up areas for fur-ther research. Some phenomena basedon stereo erasure are reported here forthe first time. It has been demonstratedthat the perception of ambiguous depthorganizations can be influenced, evensubliminally, by a preceding unambig-uous stimulus. Perhaps the most inter-esting result is the finding that thecorrespondence of objects and patternsin the two retinal projections can beestablished without actual recognitionof the objects and patterns. This pat-

tern matching is based on some rela-tively simple processes of finding con-nected clusters formed by adjacentpoints of similar brightness, and theprocesses seem to be amenable to rigor-ous analysis.

References and Notes

1. Stereopsis is the sense of relative depth in alimited region around the point of con-vergence, attributable to disparity of cor-responding points on the two retinas. Stere-opsis is more general than fusion, since withincreased disparity (above the fusion thresh-old but within the limit of patent stereopsis)corresponding points can be seen as double(often with one image suppressed) but stillperceived in depth.

2. In this article the term disparity is used ina generic sense for similar terms such asretinal disparity, binocular disparity, geo-metrical disparity, binocular parallax (shift),and so on.

3. For an up-to-date review on the traditionalaspects of research in depth perception seeK. N. Ogle, in The Eye, H. Davson, Ed.(Academic Press, New York, 1962), vol. 4,p. 209 [abstracted in Science 135, 763 (1962)].

4. B. Julesz, Bell System Tech. J. 39, 1125(1960).

5. , in Information Theory, 4th LondonSymposium, C. Cherry, Ed. (Butterworths,London, 1961), p. 212.

6. J. J. Gibson, Am. J. Psychol. 63, 367 (1950).7. A concentrated solution of gelatin is pre-

pared by dissolving the gelatin in hot water.It is then poured into a glass or plastic con-tainer which has at least one optically trans-parent and flat surface. The container istilted by 15 to 20 degrees and is kept in thisposition until the gelatin sets. (The processcan be speeded up by refrigeration.) Theflat surface of the container and the hard-ened top surface of the gelatin form anoptical wedge which can be used as a view-ing prism.

8. B. Julesz and J. E. Miller, Bell System Tech.J. 41, 663 (1962).

9. C. Wheatstone, Phil. Trans. Roy. Soc. Lon-don 371 (1838).

10. P. L. Panum, Untersuchungen uber dasSehen mit Zwei Augen (Kiel, 1858).

11. The observed phenomena have important im-plications bearing on Gestalt psychology.According to a typical view of this school[E. Lau, Psychol. Forsch. 2, 1 (1922)]stereopsis is not a result of disparity, buteach eye works up its stimulus complex intoa Gestalt, and it is the difference betweenthese Gestalten which gives rise to an im-pression of depth. This argument is still de-bated, since some of the experiments (in-volving optical illusions) which were de-signed to illustrate Gestalt factors have givencontroversial results. The fact that stereop-sis can be obtained with the random-dotimages without any monocular cues deci-sively settles this question, since no Gestaltencan be "worked up."

It might still be argued that Gestalt fac-tors may operate after the left and rightimages are fused. In this context it is im-portant to observe what happens at thevertical boundaries of the center square ofFig. 1. The probabilities that the black-and-white picture elements will be perceived asbelonging to the rectangle or perceived asbelonging to the surround are equal. Indeed,when viewed stereoscopically, the verticalboundaries of the center square of Fig. 1look fuzzy instead of being perceived asstraight lines, as one would expect if someGestalt organization had taken place (since asquare has a "good Gestalt"). If the stimuluscontains black, white, and gray picture ele-ments of equal frequency of occurrence, thenthe probability that a dot at the boundarybelongs only to the square increases to 2/3.This statistical argument is in agreement withthe observed fact that a square looks lessfuzzy at the boundaries when it is composedof random dots having three or morebrightness levels. This comment is by nomeans to be interpreted as a critique onGestalt psychology but merely as an impli-

361

cation that stereopsis operates on a primi-tive level where Gestalt considerations donot enter.That some of the more complex Gestalt

factors-such as goodness of form, sym-metry, and closure-are not a prerequisitefor stereopsis does not mean that some morebasic factors, particularly proximity andsimilarity can be omitted. Their importancein perception was first stressed by Gestaltpsychologists, especially by Koffka andWertheimer. Similarly, throughout this articleI stress the notion of connectivity, intoclusters, of adjacent dots of similar bright-ness. But there is a difference between myapproach and that of the Gestalt psycholo-gists. They developed these powerful notionsto show that perception of form was a moreinvolved process than had been previouslybelieved, while I am using these same no-tions, restricted to random-dot patterns, toshow that certain processes in stereopsis(and in visual discrimination of texture)are simpler than had been expected andamenable to rigorous analysis.

12. H. W. Dove, Ber. Preuss. Akad. Wiss. 1841,251 (1841); Ann. Physik. 110, 494 (1860).

13. N. M. S. Langlands, Trans. Opt. Soc. Lon-don 28, 45 (1926).

14. L. von Karpinska, Z. Psychol. Physiol. Sin-nesorg. 57, 1 (1910).

15. B. Julesz, J. Opt. Soc. Am. 53, 994 (1963).16. , ibid. 54, 576 (1964).17. These findings were for a picture size of

6 degrees of arc (image size, 100 X 100picture elements; size of center squares 48 X48 picture elements); disparity, 7 minutes ofarc; stimulus brightness, 5 footlamberts (55lumens/M2) .

18. P. M. Pritchard, Quart. J. Exptl. Psychol.10, 77 (1958).

19. The ambiguous stereo pairs, when presentedby themselves for a brief period, were usuallyperceived in only one way. When a slight(10 percent) bias was introduced (to coun-teract the subject's natural bias), it provedto be excessive and reversed the perceiveddepth. If the preferred depth level were thefirst level given attention and the otherdepth levels were searched sequentially onlyafterward, then a 90-percent match at thepreferred depth level would be more thanadequate for stopping the search. That, in-stead, the 100-percent match is perceived atthe unpreferred depth level is strong evi-dence for the parallel process. Recently Dod-well and Engel came to the same conclu-sion, utilizing Efron's technique of present-ing alternate stereo images with various timedelays [see P. C. Dodwell and G. R. Engel,Nature 198, 39 (1963); R. Efron, Brit. J.Ophthalmol. 41, 709 (1963)]

20. Recent neurophysiological findings revealedbinocular neural units in the striatecortex of the cat. These units were activatedby simultaneous stimulation of correspond-ing regions in the left and right retinas. Bothsummation and antagonism between receptorfields of the two eyes were demonstrated [seeD. H. Hubel and T. N. Wiesel, J. Physiol.160, 106 (1962)].

21. H. von Helmholtz, Physiological Optics, J.P. C. Southall, Ed. (Dover, New York, rev.ed., 1962), vol. 3, p. 512 and plate IV.

22. For details on this region of patent stere-opsis and on the problem of the horopter, seeK. N. Ogle, in The Eye, H. Davson, Ed.(Academic Press, New York, 1962), vol. 4.

23. B. Julesz, "Special Issue on Sensory Informa-tion Processing," IRE PGIT [ProfessionalGroup on Information Theory] Pubi. IT-8(1962), p. 84.

24. , in Proceedings IFIP [InternationalFederation of Information Processing] Con-gress 1862, M. Popklewell, Ed. (North-Hol-land, Amsterdam, 1963), p. 439. This com-puter program takes the point-by-point dif-ferences between the left and right fieldswith all possible shifts prior to subtraction.This produces an ordered set of "analogdepth planes" in which adjacent points ofminimum value are connected into clusters.The search for these clusters of points withthe same disparity immediately gives the crosssections (contour lines) of objects in a nat-ural way.

25. M. Minsky, in Computers and Thought, E.A. Feigenbaum and J. Feldman, Eds. (Mc-Graw-Hill, New York, 1963), p. 413.

26. B. White, Am. J. Psychol. 75, 411 (1962).27. S. Papert, Nature 191, 733 (1961); J. Hoch-

berg, "Illusions and figural reversal withoutlines," address presented at the 4th annualmeeting of the Psychonomic Society, BrynMawr, Pa., 1963.

28. L. Kaufman, "Form and depth perceptionin Julesz patterns," address presented at the35th meeting of the Eastern PsychologicalAssociation, Philadephia, 1964.

29. H. M. Stauffer, The Reduction of PatientDose, R. D. Moseley, Jr., and J. H. Rust,Eds. (Thomas, Springfield, Ill., 1963), p. 193.

30. For phenomena in skin localization see G.von B6kdsy, Experiments in Hearing (Mc-Graw-Hill, New York, 1960), p. 567.

31. N. Guttman and B. Julesz, J. Acoust. Soc.Am. 35, 63 (1963); B. Julesz and N. Gutt-man, ibid., p. 1895.

32. J. C. R. Licklider, ibid. 20, 150 (1948);E. M. Cramer and W. H. Huggins, Ibid.30, 413 (1958).

Problems of Drug Development

The government, the drug industry, the universities,and the medical profession: partners or enemies?

Louis Lasagna

The past few years have beenmarked by acrimonious discussionsabout the problems of drug develop-ment. While some would like to be-lieve that Congressional hearings suchas those of the Blatnik and Kefauvercommittees stirred up previously calmwaters, it is clear that the storm windshad been gathering force for a con-siderable period of time and that theexplosive passion evident in the reac-tion to these events was not engenderedde novo. That problems exist is clear;

The author is associate professor of medicineand associate professor of pharmacology andexperimental therapeutics at Johns Hopkins Uni-versity, Baltimore, Md. A frequent contributor toprofessional and popular publications, he is alsothe author of The Doctors' Dilemmas, a study ofthe medical profession, published by Harperin 1962.

362

what is less evident is the willingnessof the interested parties to define theseproblems with clarity and to solvethem.

I should like to analyze the inter-actions of physicians, the medicalschools, government, and business firstby listing some sources of discontent,since an attack on primary causes

seems preferable to a preoccupationwith secondary manifestations. Then Ishall suggest some approaches whichmight ameliorate the present state ofaffairs, in the optimistic belief thatprogress is possible and that Heraclituswas right. ("Everything comes aboutby way of strife and necessity.")

There are almost daily complaintsabout some aspect of drug usage in

our society. The academicians are con-stantly berating industry for its moti-vations and promotional excesses.When not so engaged, they are lam-basting Congress for inadequate sup-port of clinical pharmacology or foradding to the headaches of research-ers by passing "patient consent" laws.The personnel of the Food and DrugAdministration (FDA) are rarely al-lowed to rest quietly in their foxholes:on one day they are bombed forpusillanimity, on the next for high-handedness. (If a specific issue is lack-ing, it is considered good form tobrand them as generally inept.)The drug industry, in its turn, is

bitter about the unreasonableness andextravagance of the professional at-tacks. The pharmaceutical folk are un-derstandably annoyed when their sub-stantial scientific contributions are ig-nored, or when they are asked forfunds to support research or scientificsocieties by the same academicianswho have berated them. Governmentis constantly a threat to the industry,the nature of the danger ranging frompossible patent restrictions to "arbitrar-iness" or "ignorance" on the part ofspecific FDA staffers determined toprevent a drug's being marketed or tosnatch a profitable pharmaceutical offthe market.The government, for its part, must

SCIENCE, VOL. 145