Skip to main content

Full text of "mit :: ai :: aim :: AIM-1158"

See other formats





A.I. Memo No. 1158 October 1989 

C.B.I.P Memo No. 43 

Computational vision: a critical review 

Shimon Edelman Daphna Weinshall 


We review the progress made in computational vision, as represented by Marr's ap- 
proach, in the last fifteen years. First, we briefly outline computational theories developed 
for low, middle and high-level vision. We then discuss in more detail solutions proposed 
to three representative problems in vision, each dealing with a different level of visual 
processing. Finally, we discuss modifications to the currently established computational 
paradigm that appear to be dictated by the recent developments in vision. 

© Massachusetts Institute of Technology (1989) 

This report describes research done at the Massachusetts Institute of Technology within 
the Artificial Intelligence Laboratory and the Center for Biological Information Processing 
in the Department of Brain and Cognitive Sciences and Whitaker College. The Center's 
research is sponsored by a grant from the Office of Naval Research (ONR), Cognitive and 
Neural Sciences Division; by the Alfred P. Sloan Foundation; and by the National Science 
Foundation. The Artificial Intelligence Laboratory's research is sponsored by the Advanced 
Research Projects Agency of the Department of Defense under Army contract DACA76-85- 
C-0010 and in part by ONR contract N00014-85-K-0124. SE and DW are supported by 
Chaim Weizmann Postdoctoral Fellowships from the Weizmann Institute of Science. 

1 Vision as an information-processing task 

Contemporary cognitive science construes mental processes as computations denned over rep- 
resentations [126]. Under this view, visual perception, in particular, is denned as the process 
whereby light patterns impinging on the retina are transformed into internal representations. 
The problem of interpreting a retinal input is illustrated in figure 1, showing a photograph 
of a natural scene and an array of the intensity values corresponding to a part of it. These 
intensities are the raw data of vision, as they arrive in the camera or at the retina. To 
appreciate the difficulty of visual recognition, one may try to identify the object (present in 
the scene) described by this array. 

Computational vision sets out to understand theoretical aspects of visual perception, to 
explicate possible strategies of solving the emerging problems, and to explore their biological 
and artificial implementations. This paradigm emphasizes the need for an interdisciplinary 
approach: machine vision can serve as a testbed for psychological and neural models of visual 
function, while biological vision can supply useful hints in choosing feasible approaches to 
problems in visual processing by providing examples of workable solutions. 

The problem of understanding visual perception in information processing terms has been 
formulated most cogently by Marr and Poggio ([92], [96], [93]), who argued that vision (and, 
indeed, any information-processing task) may be studied at three distinct levels. At the top 
level they placed the computational theory of vision, in which the problem is characterized 
as a mapping from one kind of information to another, and the abstract properties of this 
mapping are analyzed. At the middle level, there is the choice of algorithm for the mapping, 
determined, among other factors, by the nature of representation at its input and output. 
At the lowest level, there are the implementational details, such as the physical realization 
of the representation and the algorithm. 

In the original formulation, the three levels were only loosely coupled. Nevertheless, 
it appears that any commitment to a specific computational formulation of vision would 
inevitably affect the algorithm level. In particular, as we shall argue, Marr's definition of 
the computational purpose of vision have influenced the algorithm-level theories of an entire 
school of vision research. Marr held that the representation of the shape of an object is quite 
different from the representation of its use and purpose, and that vision alone can deliver 
an internal description of the shape of an object, even when the object is not recognized in 
the sense of understanding its use and purpose ([93], p.35). Together with the assumption 
that a good way to represent an object is to build a three-dimensional model of it, this 
view resulted in many current theories of visual recognition relying on the detection of 3D 
primitives [14], or requiring a library of 3D object models ([87], [156]). In biological visual 
systems, in comparison, the question of representation appears to be far from settled. As the 
retinal image is transformed with each successive processing stage, the theories concerning 
the nature of representation diverge more and more widely, reaching no consensus as to what 
is the ultimate visual representation of objects, scenes and events ([140], [130]). 

Solutions to many visual tasks appear to be inherently ambiguous. For example, assum- 
ing, as Marr seems to have done, that the purpose of vision is to reconstruct the spatial layout 
of the outside world, one is confronted with the difficult problem of ambiguity of inferring 
the third dimension from retinal projection. This ambiguity is a necessary consequence of 
the imaging process, during which depth information is lost. Marr proposed to compensate 
for this loss by constraining the solution to the reconstruction problem to conform to a priori 
assumptions, dictated by our knowledge of the physical world. The research program based 




mmmmmtainviMtmmian viinwrnixoimMwinmiMmmuxn urmiNiMininiNtvinww 
















■ ait»»na»ii>nKnnaiiiii>»niiiiiHia«iiHiii<ua»ii»anwu*K*»n 















aUMHafflsaaiNaHHHWBawannHHanaiinMaaBMai ■ aaaa 


aawaaaaMaaaawawnMaanwHaaaaawawMawwa m » nan 


muauaaaawwaaHaMawaannManaaaaaanaaaii « n » « u< 


inuaawMmiawwiiiiaiiiwiauuiiH mm »«a»i»a»i»«i»»»»ai5«i«J«« 




Figure 1: Top: a photograph of a natural scene, with two antelopes. Bottom: the array of 
the intensity values corresponding to the image of the antelope in the right part of the scene. 
These intensities are the raw data of vision, as they arrive in the camera or at the retina. 

on this approach advanced our understanding of different aspects of vision, such as perceptual 
grouping, stereopsis and motion perception. 

In areas in which physical constraints are hard to define, as in object recognition and 
scene understanding, the reconstructionist research program runs into difficulties. Recently 
implemented recognition methods ([87], [156], [145], [49], [70]) appear to circumvent the 
problem of 3D reconstruction, either by addressing other important issues in recognition, 
such as viewpoint normalization, while assuming that the 3D models of objects are built into 
the system, or by carrying out the entire recognition process in 2D. 

In this chapter, we review the progress made in computational vision, as represented 
by Marr's approach, in the last fifteen years. Section 2 outlines computational theories 
developed for low, middle and high-level vision. Section 3 contains a more detailed discussion 
of solutions proposed to three representative problems in vision, each dealing with a different 
level of visual processing. The last section suggests modifications to the currently established 
computational paradigm that appear to be dictated by the recent developments in vision. 

Since Marr's influence on vision research has been the strongest at MIT, we follow in 
our review the development of the MIT perspective on computational vision. Other recent 
reviews of computational and machine vision include ([21], [62], [41], [132] and [57]). 

2 The components of vision 

The analysis of images can be done at different levels. Low-level vision undertakes the 
reconstruction of the visual world, but only in a limited sense. Borrowing from Marr's 
terminology, its end product is a kind of 2|D sketch of the surrounding scene: a representation 
in which many attributes of the visible surfaces are made explicit and used as labels in a 
scene-based map. Among the attributes computed at this stage are the relative motion 
and depth of the visible features, as well as luminance, color and texture based descriptions 
of different surfaces. Information from the different visual cues may be further processed 
by middle-level modules that compute more abstract properties defined over the extracted 
features. Combining surface patches and building descriptions of objects is within the scope 
of high-level vision. 

The rest of this part is organized as follows: in sections 2.1, 2.2 and 2.3 we review research 
in low-level, middle and high-level vision, respectively, and in section 2.4 we review some of 
the aspects of neuronal modeling in the context of vision. 

2.1 Low-level vision 

One possible definition of the domain of low-level vision is through the notion of bottom- up, 
or data-driven, processing (as opposed to top-down, or model-driven processing). Low-level 
vision is concerned with those transformations of the input image that are common to most 
visual tasks and can be carried out as fast as the input changes, given the resources of the 

To some extent, the analysis of the different visual cues can be done independently of 
each other and in parallel. Different behavioral tasks such as navigation, recognition or high 
precision manipulation may also require the extraction of different kinds of visual information. 
It is not clear, though, how independent the analysis of the different low-level modules can 
and should be and to what extent information should actually be integrated. The question of 

module interdependence may be approached both through the development and integration 
of low-level algorithms and by the study of biological vision systems. 

Evidence from human visual psychophysics suggests there is at least some degree of in- 
dependence among the modules. For example, one can easily understand black and white 
movies that lack color and stereo, and 3D structure can be perceived in extremely impov- 
erished stimuli, such as the projections of rotating wire-frame objects [162], or random-dot 
stereograms [73]. Physiological evidence for module independence is provided by the existence 
of separate subcortical and cortical pathways carrying different kinds of visual information 
([142], [102], [174]; see also [77], p.381). This independence is probably not complete: single 
units that respond to several different cues, such as stereo disparity and direction of motion, 
are found in some visual cortical areas ([35]; cf. [27]). 

2.1.1 Stereopsis 

Since the image on the retina of an eye, or in a camera, is obtained by projecting a three- 
dimensional world onto a two-dimensional surface, the information on the third dimension, 
depth, is lost. However, a perception of depth can still be achieved by stereopsis, i.e., by 
combining information from two or more images, taken simultaneously from slightly different 
positions, e.g., by the two eyes. In human vision, binocular stereo, together with other cues 
such as motion and perspective, is an important source of depth information. 

The importance of seeing depth in manipulation tasks becomes clear if one attempts to 
thread a needle, or to insert a key into a keyhole, with one eye closed. The importance 
of stereo in recognition tasks seems to be more limited, as people can recognize objects in 
drawings and photographs that lack binocular information. Nevertheless, much attention 
has been given to the analysis of stereo vision, mainly because the goals of stereo seemed 
relatively simple and well understood, and because stereo has been demonstrated by Julesz 
to create a perception of 3D shape in the absence of any other cue ([73]; see figure 2). 

.. s . -k ■„■ ■ • ■ •«£ m , ; .. s ■ ■* V • ■ • -i . , . 

- . ■* : .■.■■•■ •*. . . -• ■ J ■■ ■ _•■ ■ •*• 

- ."• . . ' • . s ■ - ' ■ -.'• . . • . s ■ - / ■ 
• • • r ■ ■* • _■•■■ »•••■"• , ■■■■ 

,■*.. -., >v .••.;, ■;■ . *■*■■ •-, >v. •■.;, ; • ■ 

*o. ■■•'■; . ■'•■ : - : >- * j. '. •■■■•; . •'•• : -. : >- 

- /. : ■ -^ — . . t • - /. ; ■ -ds — . . f • 
s ■ ■ .■ „.» ■ • • •■_ s ■ ■ • ^ ■ • • ■■ 
+ ■ • .■ . ■ ■ • -_ ' ' • ■ . . ■ ■ -- 

■ -s ■_ • "•. . , m . 1 — m -s ■- ■ % . . ." ,^" 

■'■ ■ • r--- V .-.. . -r . .-, ■ ."■_.■# ■» . -r . 
■*"■.•■:* \^o. ■ ••■•-.'■ v 'v. 

Figure 2: A random dot stereogram. The left and the right images should be viewed by the 
left and the right eyes, respectively. When the two images are fused, a square floating in 
depth above the background appears in the center. This is an example of shape perception 
when the only cue is stereopsis: no monocular shape cues are present in either image. 

In computer vision algorithms, the extraction of depth from binocular stereo begins with 
the formation of a disparity map by matching the two images (the disparity of an object is 

defined as the difference between its positions in the two images). Thus, a disparity value 
is assigned to every location in the image. The representation of disparities in the brain is 
different: cells tuned to a relatively narrow range of disparity are found in the primary visual 
cortex along with cells that respond selectively to features whose disparities are higher or 
lower than certain values [118]. 

In a random dot stereogram, e.g. in figure 2, (and in many natural scenes) matching 
objects in the two images to compute disparity is difficult. For every object there are many 
potential target matches, all but one of which are false. Two constraints on matching were 
proposed to solve the false targets problem [95]. The first constraint, uniqueness, forces each 
object in one image to be matched to at most one object in the other image. The second 
constraint, smoothness, requires that nearby objects in one image be matched to nearby 
objects in the other image. Most of the existing stereo algorithms, as well as most of the 
theories of human stereo vision, incorporate these two constraints. 

The matching of stereo pairs of complicated scenes remains difficult even after the appli- 
cation of these constraints. For example, objects may look different in the two images, due to 
different perspective distortions. For images that have enough texture, e.g., in random dot 
stereograms, matching can be done by cross- correlation of regions in the two images ([111], 
[171]). Many attempts have been made to obtain a general matching algorithm by the use 
of more sophisticated matching techniques ([9] is a review of early algorithms; see also [101], 
[59]). Two examples are multi-resolution matching (processing the images at several levels 
of detail; [97], [48], [100]) and adjusting the fixation point to control the range of disparities 
in the images [48]. After intensive development, a few of the algorithms appear capable of 
achieving better performance on a single stereo pair than that of the human visual system, 
albeit at the cost of a high computational complexity. 

Despite the impressive performance of some existing stereo algorithms, problems with the 
basic paradigm of applying physical constraints arise already in this relatively narrow domain. 
Specifically, imposing the constraints of uniqueness and smoothness on the matching process 
may be detrimental in some situations. For example, the smoothness assumption breaks 
down for scenes that include transparent objects, such as fences. Looking through a fence, 
one may see a tree to the left of a given bar in the left eye and the same tree to the right of 
the same bar in the right eye. (It should be noted that simple modifications of the continuity 
constraint can overcome this problem [124]). A more serious problem with smoothness arises 
at those locations in the images where there are depth discontinuities. Smoothing across 
discontinuities is bound to produce errors in stereo (and other low-level vision algorithms 
that employ this constraint, such as shape from shading and motion; see sections 2.1.2, 2.1.4 
and 3.2). Among the recent studies that address this problem are [45], [18] and [144]. See 
also [122] for an overview of the use of smoothness constraints in low-level vision. 

There is also evidence that human vision may not always follow the uniqueness assump- 
tion and that people may perceive simultaneously several surfaces corresponding to multiple 
matchings between elements in the two images [167]. In some cases the relationship between 
disparities derived from matching and perceived depth is not unequivocal. For example, the 
depth perceived in a dot pattern may correspond to an average disparity, rather than to one 
of the disparities derived from a possible matching [103]. 

A major open problem in stereo is what the representation at the output of a stereo 
system should be. The disparity field, once computed by a matching algorithm, can be used 
to extract relative depth information. Objects with large disparities are usually closer to the 
viewer than objects with small disparities (when the optical axes of the cameras are almost 

parallel). If the location, the orientation and the focal length of both cameras (or eyes) are 
known, a simple geometrical transformation can be used to compute the exact depth from the 
disparities. Typically, however, the precise camera parameters are not known. Computing 
the parameters from the disparity field is possible, though computationally expensive and 
sensitive to noise [62]. It is unclear whether this computation is necessary or whether relative 
depth is sufficient for all practical purposes ([7], [168]). 

2.1.2 Motion 

The analysis of motion is similar to stereo in that it involves the analysis of a set of images 
of a given scene. The differences are that the images are taken in succession rather than 
simultaneously, more than two images may be available for analysis, and the objects may 
undergo an arbitrary transformation between the images. Specifically, in stereo it is assumed 
that the two images are taken by two fixed cameras (eyes) whose relative location is known, 
at least approximately. When images are taken at different times, any object can translate in 
any direction, rotate and deform, so that the geometrical transformation between the images 
is not sufficiently defined, unless some assumptions are made. Among the goals of motion 
analysis are the extraction of depth information for shape perception and navigation (possibly 
in a less precise way than stereo), the segmentation of an image into distinct objects, and 
the identification of moving objects. 

Psychophysical studies indicate that motion analysis alone is sufficient to extract shape 
information when the moving objects are rigid bodies (that is, the distance between any 
two points on an object remains constant with time; [162], [152]). The rigidity constraint is 
useful to make the problem computationally tractable ([152]; see [153] and [56] for a review). 
It is derived from a computational analysis of the problem and, similar to the constraints 
imposed in stereo analysis, it should be applied with care. There are many examples where 
humans perceive nonrigid motion (e.g., expansion and contraction), even though a rigid 
motion interpretation (such as a rotation in depth) is possible (see [4], [53]). In some cases, 
the motion of nonrigid objects leads to a clear perception of shape. For example, watching 
a person walking in the dark with small lights attached to the joints, one can perceive a 
human figure walking ([71], [60]). Although several recent studies address the issue of the 
interpretation of nonrigid motion, the algorithms they develop usually allow only certain 
classes of nonrigid transformations, such as bending (see, e.g., [81] and [29]). 

Similar to the analysis of stereo, the analysis of motion can be divided into two stages: the 
matching, either of features, or of regions, in the successive images, followed by a computation 
that relies on the resulting disparity information. Another approach involves the computation 
of optical flow (the instantaneous 2D velocity field) instead of the motion disparity map. 
This computation can be presumably done by local intensity change detectors (e.g. [172]). A 
principal difficulty in the first approach is matching, whereas in the optical flow formulation 
it is relatively high sensitivity to noise, associated with numerical differentiation [62]. The 
distinction between the matching and the optical flow formulations parallels the difference 
between long-range and short-range motion perception processes observed in psychophysical 
studies [20]. 

The matching problem is more difficult in motion than in stereo, for example because of 
changing illumination and because of moving shadows, whose motion differs from the true 
motion of the objects. As in stereo, a smoothness constraint, equivalent to the assumption 
that the velocities of different features change slowly over the image, appears plausible and 

is used by many algorithms ([64], [53], [107]). Typically, computer vision algorithms en- 
force smoothness of velocity either over regions of the image [64], or along its contours [53]. 
Psychophysical findings suggest that both ways of imposing smoothness may coexist in the 
human visual system. For example, the perceived motion of points on a contour is similarly 
affected by the motion of neighboring points and of more distant points that lie on the same 
contour [108]. 

Given matched images, there are many ways to recover the three-dimensional structure 
of the objects. Early studies of this problem concentrated on the issue of minimal informa- 
tion needed to recover unique structure from motion ([152], [31], [123], [104], [150], [165]). 
One major problem common to the algorithms that recover structure from motion is their 
numerical instability [153]. To decrease the resulting sensitivity to errors, motion informa- 
tion may be pooled over many features (extended space; see [62], [23]), or over many time 
frames (extended time; see [154]). Human vision appears to use both extended time and 
space analysis ([71], [55]). Some psychophysical evidence suggests that, in disagreement with 
most structure from motion algorithms, the perception of 3D structure may not depend on 
prior perception of motion [157]. 

It is still an open question to what extent the exact structure of objects is necessary to 
fulfill the ultimate goals of vision. Some shape information, such as the classification of object 
surfaces as convex, concave, planar, or hyperbolic (saddle-like) can be obtained without the 
complete recovery of the exact locations in space of all the points in the image [166]. For 
specific narrowly denned tasks, such as obstacle avoidance in navigation, an even simpler 
analysis may suffice [110]. 

2.1.3 Edge detection 

A problem common to the design of all low-level vision modules is the choice of the input 
representation. The raw input to the visual system, an array of intensity values (see figure 1), 
does not suit well tasks that require image matching, for example, because of its sensitivity 
to noise and to changes in the illumination. Intensity edges (image locations where the 
intensity changes significantly) have been proposed as a more stable initial representation 
for stereo and motion. Different definitions of significant change in intensity lead to different 
algorithms for edge detection (see [34], [6] and [54] for reviews). Finding computationally 
efficient algorithms that capture those edges that correspond to what we intuitively perceive 
as edges proved to be a difficult problem. If edge detection is meant to provide a cartoon 
of the image, i.e., the set of edges that has a physical origin, it effectively subsumes other 
difficult problems in vision, such as figure-ground separation and object recognition. 

Intensity edges may be defined as those places in the image where the rate of change 
of intensity attains a local maximum. The derivative operation included in this definition 
makes it sensitive to noise. Noise amplification can be reduced by smoothing the image before 
subjecting it to the derivative operation. A computationally efficient, biologically plausible 
edge detection operator based on this approach is a linear filter that combines smoothing 
(by a convolution with a 2D Gaussian) with differentiation (by the application of a 2D 
Laplacian, a rotationally symmetric operator). Equivalently, the image may be convolved 
with the Laplacian of a Gaussian. The edges are then found by locating the zero- crossings in 
the output of the convolution [93]. Psychophysical evidence [164] suggests that zero-crossings 
may be involved in early processing in human vision, along with additional features, such 
as intensity extrema. (For an anatomical model of zero-crossing detection in the retina see 


Since the Laplacian of Gaussian operator is spatially symmetric, it ignores the asymmetry 
of edges, which are one- dimensional curves with a preferred direction. Some algorithms 
address this prohlem by computing second directional derivatives of the input [52]. The 
computation of the second derivative in the direction of the intensity gradient has been shown 
optimal for the detection of oriented edges [149]. As in stereo, a multi-resolution approach 
(computing the edges at several levels of smoothing) proved useful in edge detection. A widely 
used well-engineered implementation of edge detection that employs both approaches is due 
to Canny ([26]; see figure 3). Labeling edges according to their physical origin (shadows, 
occluding contours etc.) is a subject of current research (e.g. [44]). Edge labeling may be 
useful in tasks such as object recognition. 

Figure 3: Left: an image of a natural scene (same as in figure 1). Right: intensity edges 
extracted from this image by the Canny algorithm [26]. 

2.1.4 Shape from shading and shadows 

Shading information seems to be quite important in human shape perception [63]. In the 
perception of shape from shading, the visual system must separate intensity changes that 
are caused by changing orientation of object surfaces from those due to changing surface 
reflectance (including color) and illumination. As an example of the perception of shape 
from shading, consider the human face in figure 4. The basic information on the 3D shape of 
the face in this image is obtained from shading analysis (e.g., the nose appears to protrude 
from the face because its flanks axe shaded darker than its tip). 

When the illumination and the viewing direction are fixed and the color is constant, the 
shading of a surface depends solely on its local orientation. When posed quantitatively, the 
problem of inferring shape from shading turns out to be one of the most difficult in low-level 
vision. The main complication is ambiguity: a shaded image can be interpreted in many 
different ways. For example, a concave surface, a convex surface and a saddle-like surface all 
appear the same from certain viewpoints. If surface color is unknown and variable, computing 
shape from shading becomes even more difficult. 

Assuming a simplified surface reflectance model, and given some a priori knowledge about 

Figure 4: A shaded image of a face. The basic information on the 3D shape of the face in 
this image is obtained from shading analysis (e.g., the nose appears to protrude from the 
face because its flanks are shaded darker than its tip). 

the objects in the image (e.g., the depth along the outlines of the objects), the exact shape 
from shading problem becomes solvable, though still computationally difficult ([170], [69], 
[62]). A further assumption of oblique illumination may substantially reduce the computa- 
tional complexity [115], although the problem of self-shadowing, neglected by most algorithms 
including [115], becomes significant. In such cases, special shape from shadows algorithms 
[138] may be employed to infer shape from the distribution of shadows in the image. One 
complicating factor, mutual illumination (a secondary illumination due to the reflectance of 
light by other objects in the scene) is neglected by all of the above algorithms. Recent work 
has shown that, despite the simplified conditions assumed by most of the shape from shading 
algorithms, mutual illumination can make the problem more ambiguous than had been pre- 
viously appreciated [43]. On the other hand, experiments with computer graphics systems 
indicate that simulating mutual illumination is important for the creation of realistic- looking 

Recent psychophysical data suggest that humans use shading information in a limited way 
([10], [147], [25]). People seem to use shading information to build qualitative descriptions 
of object surface (e.g., in terms of convex and concave regions), rather than to compute 
exact shape from shading (in light of the complexity of extracting shape from shading, these 
findings may not come as a surprise). The computation of qualitative shape from shading 
may be facilitated by using highlights [17], which are a nuisance in most exact approaches 
([16], [131]). Finally, humans appear to employ high-level heuristics, such as the assumption 
that the scene is muminated from above, to disambiguate the shape from shading problem 

2.1.5 Color 

In human vision, color contributes important information for object recognition, although its 
contribution is by no means necessary for most everyday tasks. The human visual system 
exhibits an impressive ability to infer the correct color of objects under illumination that may 
vary in direction, intensity and spectral content. Thus, while light arriving at the retina from 
a given surface patch under different illuminations may have different spectral compositions, 
the patch would normally appear to have the same color in both cases. This phenomenon 
is called color constancy. It has been suggested that color constancy can ensue if we assume 
that the color of illuminant changes slowly and smoothly, whereas surface color changes 
abruptly. Many algorithms rely on this assumption to assess the true color of the surface 
([84], [90]; [66] show that a linear color operator can be learned from examples). Under 
the assumption that all colors appear in the image with equal probabilities, the color of the 
illuminant can be estimated by averaging color over large neighborhoods of the image [83]. 
Another approach uses the color of highlights to compute the color of the illuminant [85]. 
Psychophysical evidence suggests, however, that the human visual system does not use this 
specific technique [67]. 

Humans perceive color by having three different types of receptors in the retina whose 
peak sensitivities are in the red, green and blue regions of the spectrum. Thus the dimension 
of the color space perceived by humans is only two, not considering the overall brightness. 
This means that there are many different reflectance functions that appear to humans to 
have the same color. A principal component analysis of the distribution of surface reflectance 
functions of natural objects and of ambient daylight reveals, however, that the number of 
degrees of freedom needed is actually small. More specifically, the color of many natural 
objects can be largely approximated in terms of only three base functions (see [89], [24]). 

2.1.6 Texture 

The analysis of texture deals with statistical properties of collections of features (textons 
or elementary units of texture [74]) that cover object surfaces. Statistics of texture may 
contain important perspective information and may enable the segmentation of the image 
into distinct objects. For example, in the image of a tilted uniformly textured plane the 
density of elements diminishes in the direction away from the viewer, providing a hint to 
the orientation of the plane in depth. For non-planar objects, the distribution of texture 
elements is more complex: a sphere sprayed with paint projects to a circle filled with dots, in 
which the density of the dots is low around the center and increases towards the periphery. 
Differences in the statistical distribution of texture primitives may help segment the 
image into distinct objects (see [13] for a review; [160], [28]). If the analysis is based on 
complex texture elements [12], it is first necessary to identify the elements in a hierarchical 
manner and then compute their distribution in the image. The characterization of useful 
textons (in particular, those involved in human perception of texture) has been a subject of 
extensive research [75]. One relevant question is that of the relationship between texture- 
based segmentation and the structure of the constituent textons. Another theory relies on 
a statistical analysis of the raw intensity values in the image [74]. Gradients of the size 
distribution and the density of textons can be interpreted as depth cues for receding surfaces 
([46]; cf. [141]), through Fourier analysis [5] or statistical assumptions about the scene [169]. 


2.1.7 Occluding contours 

A line drawing of a complex 3D object, containing no shading, stereo or textural cues, may 
provide information sufficient for its identification and for the perception of its shape. One 
major source of information in a line drawing is the shape of the object's outline, or the 
occluding contour. Marr [91] argued that by themselves the occluding contours are not suffi- 
cient for shape classification. He therefore concluded that our ability to perceive the shape of 
an object from its outline is achieved through the imposition of strong constraints on possible 
shapes of objects, i.e., that most shapes can be described in terms of volumetric primitives 
called generalized cones [15]. Koenderink [80] showed, however, that such restrictions are not 
necessary, and that occluding contours carry important information about shape. Specifi- 
cally, he showed that a concave segment of the object's boundary implies that the object's 
surface is locally hyperbolic (saddle-like), while a convex occluding contour implies a locally 
elliptic (convex or concave) surface. It should be noted that saddle-like regions are good 
candidates for object segmentation, and that human vision appears to use simple heuristic 
rules of this type in achieving descriptions of objects in terms of their parts [61]. 

2.1.8 Low-level vision: an interim summary 

In distinction from other theoretical approaches, computational study of vision combines an 
analysis of the computational strategies employed by biological systems with psychological 
and physiological investigations, and with building artificial vision systems. Thinking in 
terms of representations and their transformations and subjecting the resulting theories to 
empirical tests proved especially fruitful in the domain of low-level vision, or the bottom-up 
recovery of visual properties of the environment. The interaction between computational, 
biological and machine studies led to a better understanding of the difficulties involved in 
low-level vision, and to a reopening of basic questions concerning the output representation, 
the physical constraints and the modularity of low-level processing. These issues become even 
more important at the higher levels of visual processing, where at present little feedback is 
available from empirical investigations. 

2.2 Middle vision 

Two classes of visual operations appear to fall outside the scope of both the input-driven low- 
level processes and the high-level goal-oriented ones. The first class includes processes that 
operate on the "primal sketch" representation of the input, obtained in the first stage of vision 
[92]. The purpose of these processes is, in general, the completion and the enhancement of 
the primal sketch. These operations are lateral in the sense that they do not necessarily rely 
on either a more detailed input than already available, or on higher-level cues. Examples 
of phenomena that may involve such processes are boundary completion (figure 5, left), 
interpolation of depth and motion information to regions that lie in between features used 
to compute this information, spatial grouping of features (figure 5, right) and increased 
perceived saliency of some contours relative to others ([92], [27], [137]). 

Intermediate-level processes of the second type may be described computationally as mul- 
tipurpose visual routines [155], invoked at need by higher-level modules. These routines may 
enhance the intermediate representation by making explicit spatial properties and relations 
such as contiguity (of a contour) and insideness (of a feature with respect to the contour). 
This type of information is useful in many visual tasks, such as recognition and navigation, 




Figure 5: Left: Kanisza's triangles, an example of boundary completion. Illusory contours 
that form a triangle are seen to occlude three black disks. Right: an example of hierarchical 
grouping by shape similarity and physical proximity. At the highest level, the small shapes 
form a circle. 

but is too abstract and resource-intensive to warrant automatic bottom-up computation. An 
additional kind of routines may be involved in the computational substrate of visual atten- 
tion, where the basic operations are indexing (marking a location in the visual field) and the 
shifting of the processing focus to the marked location [155]. 

Grouping of edges receives increased attention in recent studies of middle vision. Lowe [87] 
proposed to start the grouping by detecting image properties that are likely to convey useful 
information about the 3D world. Among such properties are collinearity and parallelism: the 
chance that edges that appear collinear or parallel in the image do so only by accident (due 
to the particular viewpoint) is small. Another example of a middle vision operation on edges 
is the computation of the saliency of curve segments by a simple local mechanism [137]. In 
many cases, this operation can isolate perceptually important curves in a noisy edge map 
(produced, e.g., by the Canny method). Our last example, which resembles visual routines in 
that it operates "laterally" on the intermediate representations, is the integration of different 
low-level cues (e.g., [120], [3]). 

The advances in the computational theory of integrated low and intermediate level vi- 
sion are beginning to draw the attention of psychologists and physiologists. Psychophysical 
evidence indicating that the human visual system may indeed combine different cues into an 
integrated representation can be found in [27]. Some of the recent experiments that address 
the issue of intermediate level vision and support the existence of boundary tracing and in- 
dexing as visual routines are described in [72] and [128]. An example of a neurophysiologies 
study of attention in monkey is [105], which describes the modification of the receptive field 
of an extrastriate cell by attention shifts. Cells that respond to abstract visual entities such 
as illusory contours were found in the striate cortex of the monkey [159]. 

2.3 Object recognition 

An intelligent visual system is expected to allow its host to navigate through the environment 
and possibly to manipulate any objects present in it. Keeping a visual library of models of 
potentially important objects enables such a system to recognize an object and to select an 


appropriate behavior based upon its past experience. Consequently, computational theories 
of object recognition postulate that there exist in the visual system representations of fa- 
miliar objects and scenes. To recognize an object, the system compares it with each of the 
stored models, or templates [112]. An estimate of the goodness of fit between an object 
and a template can be obtained, e.g., by a correlation-like operation, in which the two are 
superimposed and the proportion of pixels that agree in their values is computed. 

Figure 6: The appearance of a 3D object can depend strongly on the viewpoint. The image 
on the right is of the same object as the image on the left, rotated in depth by 90°. The dif- 
ference between the two images illustrates the difficulties encountered by any straightforward 
template matching approach to 3D object recognition. 

Recognition of 3D objects seen from arbitrary viewpoints is difficult because an object's 
appearance may vary considerably depending on its pose relative to the observer (see figure 6). 
Thus, while straightforward template matching [2] may be useful in the recognition of 2D 
objects in a controlled environment, it will not work for 3D object recognition, unless a 
template is stored for each view that will ever have to be recognized. Although the extent to 
which people can recognize novel, radically different, views of 3D objects is not clear ([134], 
[135]), we obviously do have some ability to generalize recognition to novel views. This ability 
is termed visual object constancy (see e.g. [65]). 

Most of the schemes for object recognition proposed to date can be divided into three 
main classes, according to their approach to the problem of object constancy [156]. The 
first approach assumes that objects have certain invariant properties that are common to all 
their views and different between object classes (in practice, this twofold assumption proved 
difficult, if not impossible, to satisfy). Under this approach, objects are represented by vectors 
of property values, or, equivalently, by points in a multidimensional space. Recognition then 
becomes a problem of clustering in this space (e.g. [36]). 

A second approach to object recognition relies on the decomposition of objects into sim- 
ple generic constituents. The components are generic in the sense that all objects can be 
described in their terms. For a structural approach to succeed, the components must be 
rapidly and reliably detected in any given view of an object, otherwise the problem of com- 
ponent identification tends to become as complex as object recognition itself. An analogous 
situation results if the breakdown of objects into components is too fine-grained, e.g. when 
the components are individual edge elements or lines. In that case, component detection is 
relatively easy, but their relationships become complicated. 

Some older object recognition systems (e.g. [22]), and at least one psychological theory 
[14], describe objects in terms of elongated volumes called generalized cylinders ([15], [94]). 
The use of generalized cylinder primitives may be regarded as a compromise between the 


conflicting requirements of generality and detectability of object components. The description 
of an object under this representation scheme is said to be object-centered, in the sense that 
it involves the relations among the object's parts that do not depend on the viewer's position. 

Variants of the decomposition approach that use contour-based primitive descriptions are 
relatively successful in certain domains, such as industrial part recognition. The two major 
limitations of this approach are that in many cases the precise shape of the object matters 
more than its decomposition into parts, and that many objects have no natural structural 

A third major approach to object recognition, which may be termed normalization [114] 
or alignment [156], addresses the requirement for precise characterization of object shapes 
by employing a modified form of template matching. This approach may be illustrated using 
the simple example of the recognition of a 2D shape such as the letter A, whose size is larger 
than that of A's template stored in the system [109]. In this case, template matching can be 
still used, provided the input shape is first "normalized" by scaling it down to the standard 

The normalization approach can be extended to the domain of 3D object recognition. To 
recognize a rigid 3D object, one may first normalize its appearance by undoing the trans- 
formations (such as 3D rotation) by which it differs from a stored model. Combined with 
algorithms that compute the necessary transformations, the normalization method has been 
recently applied to the recognition of natural objects from unconstrained viewpoints ([87], 
[156]). As the main idea of normalization is the ultimate use of template matching, 3D 
recognition schemes based on normalization typically involve 3D object-centered templates 
or models. 

State of the art object recognition systems are typically based either on sophisticated 
search techniques ([47], [49]), or on variants of the normalization approach ([19], [88], [145], 
[68], [82]). These systems are mainly useful in tasks such as industrial part recognition, be- 
cause of their reliance on strict geometrical models. In addition, search-based systems tend 
to perform poorly in cluttered environments, where segmenting the object from the back- 
ground is especially difficult. A partial solution to the first shortcoming has been offered in 
the form of parametrized object models. For example, a pair of scissors can be recognized 
irrespective of the angle between the two blades, if this angle is considered as an additional 
dimension of the search space, along with the viewpoint parameters (this, of course, aggra- 
vates the complexity of the search). The problem of sensitivity to clutter and noise may 
in principle be approached through the use of distinctive labels for recognition primitives 
(edges, corners etc.). It is not clear, however, what kind of labeling would be easy enough 
to compute and at the same time informative enough to be useful in recognition. Another 
possibility here is to use middle vision techniques (section 2.2) to enhance the image before 
recognition is attempted. An additional problem, common to most recognition methods, is 
that of indexing: choosing a subset of models that is likely to include a potential match to 
the input, rather than trying to match the input to all known models. Finally, we remark 
that automatic acquisition of object models appears to be the main difficulty associated with 
those approaches that rely on 3D geometrical representation. 

The two related questions, which recognition scheme is employed by the human visual 
system and what is the nature of visual object representation, are at present unresolved [65]. 
In this respect, our understanding of recognition lags considerably behind our understanding 
of low-level tasks such as motion detection and stereopsis. We discuss possible reasons for 
this situation in section 3. 


2.4 Biophysics of computation 

The possibility to separate abstract, or computational, from concrete, or implementational, 
aspects of visual perception is a central feature of the approach to the study of vision ad- 
vocated by Marr and Poggio [96]. Experience of the last fifteen years suggests that this 
view is too idealized, and that the complexity of most problems in vision renders imprac- 
tical algorithms that ignore constraints imposed by the available hardware. Moreover, the 
choice of low-level visual modules for which a computational theory is sought depends on the 
hardware. For example, depth can be recovered either through binocular stereo or through a 
laser rangefinder. The computational problems that need to be addressed in these two cases 
are clearly different. 

One important source of differences between solution classes that are available to bio- 
logical vision and those that best suit machine vision is the relevant computational primi- 
tives. Brains process large amounts of low-precision data at a relatively slow rate and in a 
highly parallel fashion, while most digital computers are serial, fast and can carry out high- 
precision arithmetic operations (see, e.g., [66], [136]). A gradual realization of the importance 
of hardware-related constraints led to an increasing cooperation between the fields of neuro- 
science and computational vision. One research goal in this field is to find out "what are the 
biophysical mechanisms underlying information processing and how are these mechanisms 
used to perform specific tasks" ([78], p. 640). 

Although the first detailed models of a neuron were proposed in the fifties [58], compu- 
tational biophysics is still at an early stage of development. The study of neuronal function 
is now approached at different levels, from that of the biophysics and biochemistry of mem- 
branes to the level of ensembles of neurons. In particular, functional understanding is sought 
through anatomical and physiological investigation, tightly coupled with computer simula- 
tion (see [32], [79] for recent collections of papers in the field). Biochemical and electrical 
mechanisms have been invoked as explanations of phenomena such as retinal adaptation and 
spatial and temporal filtering. Models that involve networks of neurons with simple exci- 
tatory and inhibitory connections have been proposed, among other computational tasks in 
vision, for directional selectivity in the retina ([8], [148]), for the computation of a smooth 
velocity field [53], and for the winner-take-all, or maximum, operation [173]. 

3 Case studies 

From the preceding section, it appears that at the low levels of the visual processing the 
developments in machine vision closely paralleled biological and psychological findings. In 
high level tasks such as recognition, most machine vision approaches bear less resemblance to 
their putative counterparts in biological systems. We identify radical differences between the 
nature of representation at the low and high levels of the human visual system as a possible 
cause of this distinction. We support this view by an analysis of three characteristic cases. 
The first case is the measurement of visual motion, a low-level vision process whose objective, 
the computation of a projected 2D velocity field, is well-defined in representational terms and 
has support in biological studies. The second case deals with the integration of low-level visual 
cues, a middle vision process whose biological reality has not yet been demonstrated. The 
discussion of integration illustrates some of the engineering aspects of vision research. The 
third case, recognition of three-dimensional objects, is an example of a field still in search 
for a computationally efficient, biologically plausible representation paradigm, and in which 


current engineering solutions do not seem to converge on human-like performance. 

3.1 The measurement of visual motion: well-defined representations allow 
a principled solution 

A standard formulation of the computation of visual motion distinguishes between two pro- 
cessing stages [57]. In the first stage, the movement in the changing 2D image is measured. In 
the second stage, motion estimates obtained through this measurement are used in different 
ways, e.g., in navigation, or in the recovery of the 3D layout of the environment. 

The problem confronted by the visual system in the first stage of motion perception 
may be posed as the computation of a projected 2D velocity field, i.e., the assignment of a 
velocity vector to each feature in the image (alternative, more qualitative formulations are 
conceivable, but will not be discussed here). In this computation, the visual system must 
rely solely on the changes in the light intensity patterns projected on the retina. In general, 
many possible movements may give rise to the same changes in the retinal iUumination. 
The situation is further aggravated if the measurement of visual motion is carried out by a 
mechanism which examines only a limited area of the image. 

The ability of a limited-area motion detection mechanism to extract only partial infor- 
mation about the real 2D velocity field, called the aperture problem ([161], [98], [64]), may 
be illustrated by the following example. Consider an extended oriented pattern in the image, 
such as an intensity edge, moving behind a relatively small aperture, representing the limited 
area of the image analyzed by a motion detection mechanism. Because of the aperture, it 
is only possible to perceive the movement of the edge in the direction perpendicular to its 
orientation (figure 7). Some algorithms that incorporate smoothing (equivalent to using in- 
formation from extended portions of the image) can solve this problem. An example of such 
an algorithm is given below. It should be noted that differential formulations of velocity field 
computation that yield a single equation per image point (e.g. [64]) are inherently sensitive 
to the aperture problem. A complete 2D velocity may be recovered, at the expense of an 
increased sensitivity to noise, by writing down second-order differential equations [158]. 

The output of a collection of limited-aperture motion detectors must be further processed 
to recover a better approximation to the true velocity field. According to Marx's approach, 
this can be done by constraining the solution to the recovery problem to comply with prior 
assumptions that reflect the physical nature of the problem. The assumption that the velocity 
field must be smooth ([98], [64]) proved to be a good compromise between physical reality and 
computational convenience. Hildreth [53] incorporated this assumption into the measurement 
of motion by formulating the computational problem as constrained minimization. The true 
velocity V was estimated by minimizing an error functional that expressed a compromise 
between the requirement of smoothness and the compliance of the velocity component normal 
to the image contours with the measured data: 

•■/[(T>Sj]*wi'- i - i r* 

where v 1 - is the measured velocity in the direction u x perpendicular to the contour, /3 is a 
weighting factor that expresses the confidence in the measured velocity constraints, and the 
integral is taken along the image contours (in comparison, other formulations, e.g., [64], used 


x E(t=t 2 ) 

(a) \E(t=t!) 


Figure 7: An illustration of the aperture problem. Left: a bar E is moving behind a small 
aperture A, so that its ends are not visible. Only the component v± of the bar's velocity 
that is perpendicular to the bar's orientation can be measured through the aperture. Right: 
an example of a circle translating to the right. The component perpendicular to the circle's 
contour, which would be computed by aperture-limited operators, does not correspond to 
the true motion of the circle. 

area-based constraints). The first term in the above expression corresponds to an explicit 
imposition of the smoothness constraint, used to resolve the ambiguity illustrated in Figure 7. 
Computing the maximally smooth velocity field consistent with the available evidence 
was shown by Hildreth [53] to model a number of aspects of the human motion perception 
mechanism, including visual illusions in which the algorithm and human vision fail in similar 
ways. (Algorithms that use area-based formulations, e.g., [172], are also compatible with 
psychophysical data.) Hildreth's algorithm was subsequently shown to belong to a general 
framework for approaching the so called inverse problems, of which the reconstruction of 
velocity field by aperture-limited detectors is a special case [122]. Following the theoretical 
analysis of the problem, Movshon et al. [106] discovered what could be the physiological 
correlates of the two successive stages of motion measurement: cells in cortical area VI 
that are sensitive to the direction of motion in such a way that they can provide only the 
component perpendicular to image contours, and cells in area MT that appear to combine 
the outputs of simpler detectors to provide the true direction of motion in 2D. 

3.2 Integration of low-level vision cues: refining low-level representations 

Biological vision systems are more robust and flexible than any existing computer vision 
system, even in the relatively simple domain of low-level vision. One reason for this may 
be that biological vision combines information from different low-level cues to form richer 
and more robust intermediate representations. Since little is known about the nature of 
intermediate representation in human vision, attempts to integrate low-level modules tend 
to rely on engineering common sense and on mathematical tools such as probability theory. 
The task of integrating low-level modules in computer vision encounters difficulties such 


as noisy data and differences between the output representations used by the individual mod- 
ules. Most algorithms that attempt integration tend to deal only with modules that process 
related cues (e.g., stereo, vergence and focus [1]). As as example of a more comprehensive 
integration effort, we shall describe the MIT Vision Machine, a system, built around a par- 
allel supercomputer, whose purpose is the integration of stereo, motion, color, texture, and 
intensity edge data ([120], [44]). 

The two basic observations behind this particular approach to integration are (i) that 
physical edges, such as surface orientation discontinuities and object boundaries in the scene, 
cause the appearance of discontinuities in the output of at least some of the low-level visual 
modules, and (ii) that a generalized edge map would constitute a highly useful representation 
of a scene (consider, as an illustration, how informative can a mere cartoon of a scene be). 
Thus, combining information from different cues along edges, instead of over regions, may 
result in a computationally tractable and useful integrated representation. In many cases, 
the physical origin of the edges in this representation may also be deduced. For example, if 
it is known that along an edge there is no depth discontinuity (i.e., the stereo disparities are 
continuous) and no color discontinuity, it may be assumed that the edge is due to a shadow. 
A physical classification of the edges obtained through the integration of different cues may 
later be used in recognition and navigation. 

Edge-based integration should start with a reliable map of discontinuities in the individual 
low-level cues. Since the output of most low-level modules by themselves is unreliable, one 
should choose the module whose performance is the most stable and consistent and use its 
output as the basis for the integration process. A good candidate for the basic representation 
is an intensity edge map, computed by the Canny algorithm [26]. Assuming that some of 
the other edges, corresponding to discontinuities in depth, color and texture, are also found 
by this edge detector, one can improve the reliability of the depth, color and texture maps 
by forcing discontinuities in these maps to align with the intensity edges found by the more 
reliable edge detector. 

Since edge detection leads to noise amplification, unless the process includes sufficient 
smoothing, the above approach to integration faces a dilemma. If the smoothing is too 
strong, many of the discontinuities necessary for integration will be lost. Stereo disparities, 
for example, should in general be smoothed almost everywhere in the image, except at the 
boundaries between objects. The location of these is, of course, not known at the time 
the disparity map is computed, that is, before integration. Different techniques have been 
suggested to circumvent this problem, by performing a restricted smoothing ([144], [18]). 
The MIT Vision Machine uses for this purpose a statistical method, originally developed 
for image reconstruction ([45], [99]). This method employs Markov Random Fields (MRFs), 
a mathematical structure that specifies the probability for a certain parameter to attain a 
specific value at an image point in terms of the values of the parameter at the neighboring 
points. The value at a given site is allowed to affect its neighbor's value only if there is no 
discontinuity between the two. The placement of the discontinuities in this formalism is in 
itself statistical: the state of the system is allowed to fluctuate, with edges being turned on 
and off, until a stable configuration is obtained. This configuration is usually a less noisy 
version of the original image. Since the MRF technique uses an explicit discontinuity map, it 
appears well-suited for an edge-based integration system. The integration algorithm in such 
a system may include the following steps: 

1. Apply different low-level vision algorithms, in parallel. 


2. Smooth the output of each module with an MRF, constraining discontinuities to the 
locations of intensity edges, in parallel. 

3. Use the discontinuities in the MRF of each module to label edges according to their 
physical origin, i.e., occluding boundaries, corners, shadow edges, specularity edges, 
color edges. 

4. If necessary, return to step 2 and use the labeled edges to improve the quality of the 

An example of the output of this system using real images is given in figure 8. 

Figure 8: A cartoon-like representation of two objects segmented from the background using 
stereo, motion, color and texture cues. 

At present, this approach still falls far short of achieving human-like performance in scene 
segmentation and in the classification of image edges. The major open problems in the MRF- 
based approach to integration are the poor initial output of some of the low-level modules and 
the need for a manual tuning of the MRF parameters. Furthermore, it is not clear whether a 
simple and qualitative rule-based approach would not perform better than formal approach 
of combining quantitative data. 

3.3 Object recognition: looking for the right representation 

It is difficult to find in the study of high-level vision a parallel to the achievements of the 
combined mathematical, psychological and physiological approach that advanced our under- 
standing of low-level visual tasks such as motion and stereo. In low-level vision many physical 
assumptions have been found useful in dealing with the ill-posedness of inverse optics or the 
reconstruction problem. In comparison, in high-level problems such as object recognition rel- 
evant physical constraints are scarce and key computational issues, in particular the nature 
of representation, are highly controversial. 

Consider the task of recognizing objects that belong to any fixed set, such as a collection 
of typefaces. The major problem confronted by the visual system in this case is that of noise. 
For example, machines can be trained to recognize complicated and diverse 2D objects such 


as printed characters that come in a variety of sizes and typefaces [76], with an efficiency 
approaching that of a human. In another example of a restricted domain, the interpretation 
of polyhedral scenes with shadows, relying on a complete enumeration of possible types of 
polyhedral junctions, has been demonstrated by Waltz as early as 1975 [163]. In both these 
cases, the finite domain places an effective constraint on the possible solutions. 

When the shape of the objects is allowed to change continuously, but in a principled 
manner and without affecting their basically discrete and finite classification, some physical 
assumptions still apply and may be used to constrain the solution. In handwritten character 
recognition, for example, phenomenological knowledge of the physics of motor control in 
handwriting can be easily formulated in mathematical terms and leads to a simplification of 
the recognition process [39]. In comparison, when the problem is to recognize an arbitrary 
3D shape from any possible viewpoint, physical constraints can be applied to compensate for 
only one source of image variability — the imaging process. The other part of the problem, 
the variability of the intrinsic 3D shape of the objects, remains unsolved. 

In many recognition problems, the nearest conceivable thing that resembles a physical 
constraint seems too implausible or too impractical to be of any use. Consider, for example, 
face recognition, an important visual task in primates, and a challenging application of 
machine vision. A face recognition system that employs geometric models is bound to fail 
when confronted with a familiar face bearing a novel expression, even if a great number 
of templates corresponding to previously encountered expressions are stored. Applying the 
physical constraint method in this case amounts to the inclusion of the knowledge of facial and 
cranial anatomy in the recognition algorithm, an improbable requirement. It just might turn 
out that a feasible face recognition system would have to rely on stored examples of familiar 
faces bearing prototypical expressions more than on an algorithmic approach of computing 
some standard face representation using anatomical constraints. 

Even if standard representations, or models, of 3D objects are involved in recognition, it 
is not clear what method of modeling is computationally most efficient, and what method (or 
methods) is employed in human vision. A major point of controversy as to the nature of 3D 
object representation regards the question of whether it should be object-centered (symmet- 
ric with respect to the object itself), as advocated by Marr, or viewer- centered (dependent 
upon the viewer's position relative to the object). Although in principle a 3D object- centered 
representation can be readily constructed from a collection of 2^D viewer-centered ones, in 
practice building a 3D object- centered model only to use it later by comparing its 2D pro- 
jections with input images ([87], [145], [156]) seems to be redundant. Similarly, it appears 
equally unlikely that the human visual system goes through the effort of constructing com- 
putationally unwarranted 3D models, or, alternatively, that it necessarily reconstructs the 
third dimension of an input object before recognizing it (after all, we do recognize everyday 
objects in visually impoverished line drawings). 

Humans construct their library of object models through experience. It appears that 3D 
machine recognition systems that use object-centered models fail to match this crucial com- 
ponent of the human performance in recognition. The few 3D machine recognition systems 
that are designed to learn object representations from examples ([70], [151], [119]) employ 
viewer-based rather than object-based representation. Some evidence to the effect that the 
ultimate representation need not be object-centered is already available. This evidence falls 
into three classes, psychophysical, physiological and computational, discussed separately be- 


3.3.1 Hints from psychophysics 

The most straightforward way to investigate the nature of object representation in long-term 
memory is to test it in a recognition task. An object-centered representation that does not 
allow the system to infer what the object would look like from an arbitrary viewpoint is, for all 
practical purposes, equivalent to a collection of viewer-centered view-specific representations. 
Consequently, if the human visual system effectively employs object- centered representations, 
a person should be able to recognize an object, previously seen from a limited range of 
viewpoints, from a novel viewpoint. 

It should be noted that the structure from motion theorems ([152], [150], [86]) indicate 
that in principle the effect of having a full object-centered description of a 3D rigid body 
may be produced by actually storing only a few of its 2D views. The ability of the human 
visual system to infer the 3D structure of an object from a small number of its projections, 
termed the kinetic depth effect, has been known for a long time ([162], [71]). An alternative 
way to formulate the object-centered representation question is, therefore, to ask whether 
the 3D structure perceived in the kinetic depth effect is retained by the long-term memory, 
or whether this effect is a mere by-product of some other, more basic, perceptual operation. 

Until recently, the ability of the human visual system to recognize objects from novel 
viewpoints has been taken for granted. Palmer, Rosch and Chase were the first to demon- 
strate that even for familiar objects speed and accuracy of recognition varies with viewpoint 
[113]. A more direct test would involve novel objects, shown to the subjects under con- 
trolled conditions which guarantee that some views remain unseen until the test time. Such 
experiments, involving novel 3D wire-like stimuli, have been carried out by Rock and his col- 
laborators ([134], [135]). In one series of experiments [134], subjects saw novel 3D wires, each 
from a single fixed viewpoint. The subjects were subsequently shown other, similar, objects, 
one at a time, and required to decide whether these were the familiar wires, displayed at new 
attitudes, or unfamiliar wires. In these experiments, the subjects' performance approached 
chance level when the in-depth rotational distance between training and test views was about 
30°. In another experiment [135], the test called for deciding whether one of two simultane- 
ously displayed wire objects was a replica of the other object, shown at a different attitude. 
Again, subjects performed poorly, even when given the opportunity to reason explicitly about 
the relative positions of different features of the two objects. As during training the subjects 
perceived the stimuli in 3D (due to binocular stereopsis), the lack of generalization to novel 
views in this experiment could be attributed to the subjects' failure either to retain 3D infor- 
mation, or to translate it into a format that would permit later generalization, e.g., into an 
object-centered description. The likelihood of the first interpretation, namely, that subjects 
do not include 3D information in a long-term memory representation of wire-like objects, 
is diminished by the finding that the presence of binocular stereo cues significantly reduces 
recognition error rate ([38]; see also [65], p.81). 

Additional evidence in favor of the hypothesis that objects are normally represented in 
long-term memory by collections of 2^D viewpoint-specific descriptions comes from the ex- 
periments of Tarr and Pinker [143], who found that naming time for 3D objects similar to 
the ones used by Shepard and Metzler [139] increased linearly with the distance between 
the presented view and the closest learned view. This linear dependence, present in a wide 
variety of recognition tasks (cf. [113], [37]), is predicted by a theory that posits multiple- view 
representations, but is difficult to explain within the framework of a theory of recognition of 
the viewpoint normalization variety ([87], [156]). For example, Ullman's alignment scheme 


([156]; see section 2) specifies an algorithm that computes the hypothesized viewpoint of an 
object model, given the positions of a few key model features in the input image. When 
combined with a 3D object-centered representation of the model, this scheme should result 
in recognition time that is independent of the viewpoint (unless implementational constraints 
bring about such dependence, in which case the representation effectively ceases to be sym- 
metric with respect to the object, i.e. object-centered). 

3.3.2 Hints from physiology 

The notion that the primate visual system employs view-specific representations of objects 
is supported by the findings of Gross, Perrett and others ([51], [117], [50], [116]) that cells 
in the visual area IT in monkey respond preferentially to complex stimuli such as hands and 
faces. Perrett et al. [116] reported that some cells responded maximally to full views of a 
face, others to the face tilted upwards or downwards by 45°, and still others to profile and 
back views of the head. The cells' response remained strong when the preferred view of 
the face was displayed at different scales, or rotated within the image plane by as much as 
90°. If indeed these cells are at the top of the representation hierarchy in the visual system, 
then their response pattern is most easily explained in terms of viewer- centered rather than 
object-centered representations. 

3.3.3 Computational considerations 

Recent computational studies provide further support for the hypothesis that a recognition 
strategy based on multiple-view representation can in principle account for much of the 
human performance in object recognition ([40], [11], [119]). In particular, if the recognition 
problem is formulated in terms of the approximation of a mapping that associates a standard 
view of an object with any of its other views, powerful mathematical tools from function 
approximation theory are available that can construct such a mapping from examples [121]. 
In other words, a system can learn to recognize an object from any viewpoint merely by 
being exposed to a random set of a few tens of the object's views. 

4 Discussion and prognosis 

During the last decade, the combined computational and biological study of vision, pioneered 
by Marr, has resulted in progress in some areas of vision research, notably, in understanding 
low-level vision. The computational approach proved most fruitful when the problem at hand 
could be formulated as a transformation between clearly defined representations. In other 
words, principled computational solutions were first offered to those problems for which a 
notion of competence (in Chomsky's sense; see [93], p.28) was readily available. One such 
problem was binocular stereo, for which both the input representation (edge maps of the two 
images) and the output representation (a dense depth map) seemed to be obvious. 

The computational paradigm made an important contribution to the study of vision by 
encouraging the generation of concrete and testable theories of performance. In many cases, 
these theories eventually led to a revision of basic assumptions of the underlying theory 
of competence. For example, when even the sophisticated stereo algorithms appeared to 
be incomplete as models of human performance, the nature of the input and the output 
representations in stereo came under questioning. 


The computational approach has also increased the understanding of the difficulties as- 
sociated with high-level tasks, such as object recognition. Despite early enthusiasm, it now 
becomes increasingly clear that no adequate competence-level theory is yet available for these 
tasks. In particular, computational feasibility and biological (psychological and neurophysi- 
ological) plausibility of 3D object-centered models as the ultimate representations in object 
recognition is now being questioned. From the preceding review, it appears that a satis- 
factory reconstruction of the visual world is not feasible, unless its aim is a viewer-relative 
and qualitative (as opposed to absolute and quantitative) representation. Fortunately, it also 
appears that carrying out an ideal reconstruction is not a prerequisite for seeing, insofar as 
it does not seem that biological vision relies on such reconstruction. 

The roots of the present situation in computational vision may be traced to the philo- 
sophical foundations of the currently accepted computational paradigm in cognitive science. 
An important part of Marr's argument in favor of the computational approach to vision has 
been functionalist. The functionalist program in the study of mind [125] has been prompted 
by the "suspicion that there are empirical generalizations about mental states that cannot 
be formulated in the vocabulary of neurological or physical theories" ([42], p.25). Glossing 
over technical details, a common version of the functionalist approach ([126], [127]) holds 
that two intelligent agents are in the same psychological state (e.g., both believe that they 
see a cat) if they are in the same computational state. In this sense, the two agents function 
similarly, although their physical realizations may be different. Similarly, Marr and Poggio 
suggested that functional, or computational, understanding of vision may be achieved in 
separation from the understanding of its physical substrate, just as it is sufficient to study 
aerodynamics to understand flight in both birds and airplanes ([93], p.345). 

The advantages (and the feasibility) of the postulate of level separation as a research 
strategy in vision seem now less substantial than a few years ago. At the same time, func- 
tionalist theories of cognition are increasingly criticized from at least two points of view. One 
line of argument starts with the claim that the functional level is not sufficiently abstract 
to allow interesting generalizations about mental states ([127], p. 74). Others claim that any 
such generalizations are, in principle, unwarranted ([33], p.224; see also [129]), or depend on 
a better understanding of the physical substrate of cognition [30]. A speedy resolution of the 
uncertainty concerning the philosophical foundations of the functionalist program appears at 
the present time unlikely. 

Whereas a discussion of level separation in computational vision leads one to question the 
basic premises of the functionalist program, another problematic issue mentioned above, that 
of visual reconstruction as a goal of vision, can be addressed within the present paradigm. 
Persistent difficulties with solving the inverse optics problem (reconstructing the geometry 
of visible objects) prompt the search for a different competence theory of vision, one that is 
formulated in terms of more plausible representations than geometric models. Many of the 
emerging research directions can be described as an attempt to make the most of the rich in- 
termediate representation provided by the low-level modules (see the preface to [132]). These 
approaches focus on qualitative rather than quantitative representations (e.g. [146], [168]), 
on active vision ([110], [7]), on the application of high-level rules to obtain fast approximate 
solutions [131] and on massive use of parallelism, memory and learning ([120], [151], [119]). 



We are grateful to Ellen Hildreth, Norberto Grzywacz, Tomaso Poggio, Manfred Fahle and 
Anya Hurlbert for many useful comments and discussions. 


[I] A. Abbott and N. Ahuja. Surface reconstruction by dynamic integration of focus, 
camera vergence and stereo. In Proceedings of the 2nd International Conference on 
Computer Vision, pages 532-545, Tarpon Springs, FL, 1988. IEEE, Washington, DC. 

[2] Y. S. Abu-Mostafa and D. Psaltis. Optical neural computing. Scientific American, 
256:66-73, 1987. 

[3] J. Y. Aloimonos. Unification and integration of visual modules. In Proceedings Image 
Understanding Workshop, pages 507-551, San Mateo, CA, 1989. Morgan Kaufmann 
Publishers, Inc. 

[4] A. Ames. Visual perception and the rotating trapezoid window. Psychological Mono- 
graphs, 65(7), 1951. 

[5] R. Bajcsy and L. Lieberman. Texture gradient as a depth cue. Computer Graphics and 
Image Processing, 5:52-67, 1976. 

[6] D. H. Ballard and C. M. Brown. Computer Vision. Prentice-Hall, Englewood Cliffs, 
NJ, 1982. 

[7] D. H. Ballard, R. C. Nelson, and B. Yamauchi. Animate vision. Optic News, 15:9-25, 

[8] H. B. Barlow and R. W. Levick. The mechanism of directional selectivity in the rabbit's 
retina. J. Physiol, 173:477-504, 1965. 

[9] S. T. Barnard and M. A. Fischler. Computational stereo. A CM Comput. Surveys, 
143:553-572, 1982. 

[10] H. G. Barrow and J. M. Tenenbaum. Computational vision. Proc. IEEE, 69:572-595, 

[II] R. Basri and S. Ullman. Recognition by linear combinations of models, June 1989. 
forthcoming MIT AI Memo. 

[12] J. Beck. Textural segmentation. In J. Beck, editor, Organization and representation in 
perception, chapter 15. Erlbaum, Hillsdale, NJ, 1982. 

[13] J. Beck, K. Prazdny, and A. Rosenfeld. A theory of textural segmentation. In J. Beck, 
B. Hope, and A. Rosenfeld, editors, Human and Machine Vision, pages 1-38. Academic 
Press, New York, NY, 1983. 

[14] I. Biederman. Human image understanding: Recent research and a theory. Computer 
Vision, Graphics, and Image Processing, 32:29-73, 1985. 


[15] T. O. Binford. Visual perception by computer. In IEEE Conference on Systems and 
Control, Miami Beach, FL, Dec. 1971. 

[16] A. Blake and G. Brelstaff. Geometry from specularities. In Proceedings of the 2nd In- 
ternational Conference on Computer Vision, Tarpon Springs, FL, 1988. IEEE, Wash- 
ington, DC. 

[17] A. Blake and H. H. Biilthoff. Does the brain know physics? perception of shape from 
specularity, 1989. to appear. 

[18] A. Blake and A. Zisserman. Visual reconstruction. MIT Press, Cambridge, MA, 1988. 

[19] R. C. Bolles and P. Horaud. 3DP0: A three-dimensional part orientation system. 
International Journal of Robotics Research, 5:3-26, 1986. 

[20] O. J. Braddick. Low-level and high-level processes in apparent motion. Phil. Trans. R. 
Soc. London B, 290:137-151, 1980. 

[21] M. J. Brady. Computational approaches to image understanding. ACM Computing 
Surveys, 14:3-71, 1982. 

[22] R. A. Brooks. Symbolic reasoning among 3D models and 2D images. Artificial Intelli- 
gence, 17:285-348, 1981. 

[23] A. Brass and B. K. P. Horn. Passive navigation. Computer Vision, Graphics, and 
Image Processing, 21:3-20, 1983. 

[24] G. Buchsbaum and A. Gottschalk. Chromaticity coordinates of frequency-limited func- 
tions. Journal of the Optical Society of America, 1:885-887, 1984. 

[25] H. H. Bulthoff and H. A. Mallot. Interaction of different modules in depth perception. 
In Proceedings of the 1st International Conference on Computer Vision, pages 295-305, 
June 1987. 

[26] J. F. Canny. A computational approach to edge detection. IEEE Transactions on 
Pattern Analysis and Machine Intelligence, 8:679-698, 1986. 

[27] P. Cavanagh. Reconstructing the third dimension: interactions between color, tex- 
ture, motion, binocular disparity and shape. Computer Vision, Graphics, and Image 
Processing, 37:171-195, 1987. 

[28] R. Chellappa, R. Chatterjee, and R. Baghdazian. Texture synthesis and coding using 
Gaussian Markov field models. IEEE Trans. SMC, 15:298-303, 1985. 

[29] S. S. Chen and M. Penna. Shape and motion of nonrigid bodies. Computer Vision, 
Graphics, and Image Processing, 36:175-207, 1986. 

[30] P. S. Churchland. Neurophilosophy. MIT Press, Cambridge, MA, 1987. 

[31] W. F. Clocksin. Perception of surface slant and edge labels from optical flow: a com- 
putational approach. Perception, 9:253-269, 1980. 

[32] R. M. J. Cotterill, editor. Computer simulation in brain science. Cambridge Univ. 
Press, Cambridge, 1988. 


[33] D. Davidson. Essays on actions and events. Clarendon Press, Oxford, 1980. 

[34] L. S. Davis. A survey of edge detection techniques. Computer Graphics and Image 
Processing, 4:248-270, 1975. 

[35] R. Desimone, S. J. Schein, J. Moran, and L. G. Ungerleider. Contour, color and shape 
analysis beyond the striate cortex. Vision Research, 25:441-452, 1985. 

[36] R. 0. Duda and P. E. Hart. Pattern classification and scene analysis. Wiley, New 
York, 1973. 

[37] S. Edelman, H. Biilthoif, and D. Weinshall. Stimulus familiarity determines recognition 
strategy for novel 3d objects. A.I. Memo No. 1138, Artificial Intelligence Laboratory, 
Massachusetts Institute of Technology, July 1989. 

[38] S. Edelman and H. H. Biilthoff. Recognition of novel 3D objects in human vision, 1989. 

[39] S. Edelman, S. Ulhnan, and T. Flash. Reading cursive handwriting by alignment of 
letter prototypes, 1989. submitted for publication. 

[40] S. Edelman and D. Weinshall. A self- organizing multiple-view representation of 3d 
objects. A.I. Memo No. 1146, Artificial Intelligence Laboratory, Massachusetts Institute 
of Technology, August 1989. 

[41] M. A. Fischler and 0. Firschein, editors. Readings in computer vision: issues, problems, 
principles and paradigms. Morgan Kaufmann, Los Altos, CA, 1987. 

[42] J. A. Fodor. Representations. MIT Press, Cambridge, MA, 1981. 

[43] D. Forsyth and A. Zisserman. Mutual illumination. In Proceedings IEEE Conf. on 
Computer Vision and Pattern Recognition, pages 466-473, San-Diego, CA, 1989. 

[44] E. Gamble, D. Geiger, T. Poggio, and D. Weinshall. Labeling edges and the integration 
of low-level visual modules. IEEE Trans. SMC, 19(6), 1989. 

[45] S. Geman and D. Geman. Stochastic relaxation, Gibbs distributions, and the Bayesian 
restoration of images. IEEE Transactions on Pattern Analysis and Machine Intelli- 
gence, 6:721-741, 1984. 

[46] J. J. Gibson. The perception of the visual world. Houghton Mifflin, Boston, MA, 1950. 

[47] C. Goad. Fast 3D model-based vision. In A. P. Pentland, editor, From pixels to 
predicates, pages 371-391. Ablex, Norwood, NJ, 1986. 

[48] W. E. L. Grimson. From Images to Surfaces. MIT Press, Cambridge, MA, 1981. 

[49] W. E. L. Grimson and T. Lozano-Perez. Localizing overlapping parts by searching the 
interpretation tree. IEEE Transactions on Pattern Analysis and Machine Intelligence, 
9:469-482, 1987. 

[50] C. G. Gross and M. Mishkin. The neural basis of stimulus equivalence across retinal 
translation. In S. Harnad, R. W. Doty, L. Goldstein, J. Jaynes, and G. Krauthamer, 
editors, Lateralization in the nervous system. Academic Press, New York, NY, 1977. 


[51] C. G. Gross, C. E. Rocha-Miranda, and D. B. Bender. Visual properties of cells in 
inferotemporal cortex of the macaque. J. Neurophysiol., 35:96-111, 1972. 

[52] R. M. Haralick. Digital step edges from zero crossings of second directional derivatives. 
IEEE Transactions on Pattern Analysis and Machine Intelligence, 6:58-68, 1984. 

[53] E. C. Hildreth. The measurement of visual motion. MIT Press, Cambridge, MA, 1984. 

[54] E. C. Hildreth. Edge detection. In S. Shapiro, editor, Encyclopedia of artificial intelli- 
gence, pages 257-267. John Wiley, New- York, NY, 1987. 

[55] E. C. Hildreth, N. M. Grzywacz, E. H. Adelson, and V. K. Inada. The perceptual 
buildup of three-dimensional structure from motion. A.I. Memo No. 1141, Artificial 
Intelligence Laboratory, Massachusetts Institute of Technology, 1989. 

[56] E. C. Hildreth and C. Koch. The analysis of visual motion: from computational theory 
to neuronal mechanisms. Ann. Rev. Neurosci., 10:477-533, 1987. 

[57] E. C. Hildreth and S. Ullman. The computational study of vision. A.I. Memo No. 1038, 
Artificial Intelligence Laboratory, Massachusetts Institute of Technology, 1988. 

[58] A. L. Hodgkin and A. F. Huxley. A quantitative description of membrane current and 
its application to conduction and excitation in nerve. J. Physiol. Lond., 116:500-544, 

[59] W. Hoff and N. Ahuja. Extracting surfaces from stereo images: An integrated approach. 
In Proceedings of the 1st International Conference on Computer Vision, pages 284-294, 
June 1987. 

[60] D. D. Hoffman and B. E. Flinchbaugh. The interpretation of biological motion. Bio- 
logical Cybernetics, 42:195-204, 1982. 

[61] D. D. Hoffman and W. A. Richards. Parts of recognition. Cognition, 18:65-96, 1984. 

[62] B. K. P. Horn. Robot vision. MIT Press, Cambridge, Mass., 1986. 

[63] B. K. P. Horn and M. Brooks. Seeing shape from shading. MIT Press, Cambridge, 
Mass., 1989. 

[64] B. K. P. Horn and B. G. Schunck. Determining optical flow. Artificial Intelligence, 
17:185-203, 1981. 

[65] G. W. Humphreys and P. Quinlan. Normal and pathological processes in visual object 
constancy. In G. W. Humphreys and M. J. Riddoch, editors, Visual object processing: 
a cognitive neuropsychological approach, pages 43-106. Erlbaum, Hillsdale, NJ, 1987. 

[66] A. Hurlbert and T. Poggio. Synthesizing a color algorithm from examples. Science, 
239:482-485, 1988. 

[67] A. C. Hurlbert, H.-C. Lee, and H. H. Bulthoff. Cues to the color of the illuminant. 
Invest. Ophthalm. Vis. Science Suppl., 30:221, 1989. 


[68] D. P. Huttenlocher and S. Ullman. Object recognition using alignment. In Proceed- 
ings of the 1st International Conference on Computer Vision, pages 102-111, London, 
England, June 1987. IEEE, Washington, DC. 

[69] K. Ikeuchi and B. K. P. Horn. Numerical shape from shading and occluding boundaries. 
Artificial Intelligence, 15:141-184, 1981. 

[70] K. Ikeuchi and T. Kanade. Applying sensor models to automatic generation of object 
recognition programs. In Proceedings of the 2nd International Conference on Computer 
Vision, pages 228-237, Tarpon Springs, FL, 1988. 

[71] G. Johansson. Visual perception of biological motion and a model for its analysis. 
Perception and Psychophysics, 14:201-211, 1973. 

[72] P. Jolicoeur, S. Ullman, and M. Mackay. Curve tracing: a possible basic operation in 
the perception of spatial relations. Memory and Cognition, 14:129-140, 1986. 

[73] B. Julesz. Foundations of Cyclopean perception. University of Chicago Press, Chicago, 
JX, 1971. 

[74] B. Julesz. Experiments in the visual perception of texture. Scientific American, 232:34- 
43, 1975. 

[75] B. Julesz. A brief outline of the texton theory of human vision. Trends in Neurosciences, 
7:41-45, 1984. 

[76] S. Kahan, T. Pavlidis, and H. S. Baird. On the recognition of printed characters of 
any font and size. IEEE Transactions on Pattern Analysis and Machine Intelligence, 
9:274-287, 1987. 

[77] E. R. Kandel and J. H. Schwartz. Principles of neural science. Elsevier, New York, 

[78] C. Koch and T. Poggio. Biophysics of computational systems: Neurons, synapses, and 
membranes. In G. M. Edelman, W. E. Gall, and W. M. Cowan, editors, Synaptic 
Function, pages 637-697. Wiley, New York, NY, 1987. 

[79] C. Koch and I. Segev. Methods in neuronal modeling. MIT Press, Cambridge, MA, 

[80] J. J. Koenderink. What does the occluding contour tell us about solid shape? Percep- 
tion, 13:321-330, 1984. 

[81] J. J. Koenderink and A. J. van Doom. Depth and shape from differential perspective 
in the presence of bending deformations. Journal of the Optical Society of America, 
3:242-249, 1986. 

[82] Y. Lamdan and H. Wolfson. Geometric hashing: a general and efficient recognition 
scheme. In Proceedings of the 2nd International Conference on Computer Vision, pages 
238-251, Tarpon Springs, FL, 1988. IEEE, Washington, DC. 


[83] E. H. Land. An alternative technique for the computation of the designator in the 
retinex theory of color vision. Proceedings of the National Academy of Science, 83:3078- 
3080, 1986. 

[84] E. H. Land and J. J. McCann. Lightness and retinex theory. Journal of the Optical 
Society of America, 61:1-11, 1971. 

[85] H.-C. Lee. Method for computing the scene-illuminant chromaticity from specular 
highlights. Journal of the Optical Society of America, 3:1694-1699, 1986. 

[86] H. C. Longuet-Higgins. A computer algorithm for reconstructing a scene from two 
projections. Nature, 293:133-135, 1981. 

[87] D. G. Lowe. Perceptual organization and visual recognition. Kluwer Academic Pub- 
lishers, Boston, MA, 1986. 

[88] D. G. Lowe. Three-dimensional object recognition from single two-dimensional images. 
Artificial Intelligence, 31:355-395, 1987. 

[89] L. T. Maloney. Computational approaches to color vision. PhD thesis, Stanford Univ., 
Stanford, CA, 1984. 

[90] L. T. Maloney and B. Wandell. A computational model of color constancy. Journal of 
the Optical Society of America, 1:29-33, 1986. 

[91] D. Marr. Analysis of occluding contour. Proc. R. Soc. Lond. B, 197:441-475, 1976. 

[92] D. Marr. Early processing of visual information. Phil. Trans. R. Soc. Lond. B, 275:483- 
524, 1976. 

[93] D. Marr. Vision. W. H. Freeman, San Francisco, CA, 1982. 

[94] D. Marr and H. K. Nishihara. Representation and recognition of the spatial organization 
of three dimensional structure. Proceedings of the Royal Society of London B, 200:269- 
294, 1978. 

[95] D. Marr and T. Poggio. Cooperative computation of stereo disparity. Science, 194:283- 
287, 1976. 

[96] D. Marr and T. Poggio. From understanding computation to understanding neural 
circuitry. Neurosciences Res. Prog. Bull., 15:470-488, 1977. 

[97] D. Marr and T. Poggio. A computational theory of human stereo vision. Proceedings 
of the Royal Society of London B, 204:301-328, 1979. 

[98] D. Marr and S. Ullman. Directional selectivity and its use in early visual processing. 
Proceedings of the Royal Society of London B, 211:151-180, 1981. 

[99] J. Marroquin, S. Mitter, and T. Poggio. Probabilistic solution of ill-posed problems in 
computational vision. Journal of the American Statistical Association, 82:76-89, 1987. 

[100] J. E. W. Mayhew and J. P. Frisby. Psychophysical and computational studies towards 
a theory of human stereopsis. Artificial Intelligence, 17:349-386, 1981. 


[101] G. G. Medioni and R. Nevatia. Segment-based stereo matching. Computer Vision, 
Graphics, and Image Processing, 31:2-18, 1985. 

[102] M. Mishkin, L. G. Ungerleider, and K. A. Macko. Object vision and spatial vision: two 
cortical pathways. Trends in Neurosciences, 4:414-417, 1983. 

[103] G. J. Mitchison and S. P. McKee. Interpolation in stereoscopic matching. Nature, 
315:402-404, 1985. 

[104] A. Mitiche. On kineopsis and computation of structure and motion. IEEE Transactions 
on Pattern Analysis and Machine Intelligence, 8:109-112, 1986. 

[105] J. Moran and R. Desimone. Selective attention gates visual processing in the extras- 
triate cortex. Science, 229:782-784, 1985. 

[106] J. A. Movshon, E. H. Adelson, M. S. Gizzi, and W. T. Newsome. The analysis of 
moving visual patterns. In C. Chagas, R. Gattas, and C. G. Gross, editors, Pattern 
Recognition Mechanisms. Vatican Press, Rome, 1985. 

[107] H.-H. Nagel and W. Enkelmann. Towards the estimation of displacement vector fields 
by 'oriented smoothness' constraints. In Proceedings Int. Conf. on Pattern Recognition, 
pages 6-8, Montreal, Canada, July 1984. 

[108] K. Nakayama and G. H. Silverman. The aperture problem ii: spatial integration of 
velocity information along contours. Vision Research, 28:739-746, 1988. 

[109] U. Neisser. Cognitive Psychology. Appleton-Century-Crofts, New York, NY, 1967. 

[110] R. C. Nelson and J. Aloimonos. Using flow field divergence for obstacle avoidance: 
towards qualitative vision. In Proceedings of the 2nd International Conference on Com- 
puter Vision, pages 188-196, Tarpon Springs, FL, 1988. IEEE, Washington, DC. 

[Ill] H. K. Nishihara. Practical real-time imaging stereo matcher. Optical Engineering, 
23(5):536-545, 1984. 

[112] A. Paivio. The relationship between verbal and perceptual codes. In E. C. Carterette 
and M. P. Friedman, editors, Handbook of Perception, pages 375-397. Academic Press, 
New York, NY, 1978. 

[113] S. Palmer, E. Rosch, and P. Chase. Canonical perspective and the perception of objects. 
In J. Long and A. Baddeley, editors, Attention and Performance IX, pages 135-151. 
Erlbaum, Hiflsdale, NJ, 1981. 

[114] S. E. Palmer. The psychology of perceptual organization: a transformational approach. 
In J. Beck, B. Hope, and A. Rosenfeld, editors, Human and machine vision, pages 269- 
340. Academic Press, New York, 1983. 

[115] A. Pentland. Shape information from shading: a theory about human perception. In 
Proceedings of the 2nd International Conference on Computer Vision, pages 404-413, 
Tarpon Springs, FL, 1988. IEEE, Washington, DC. 

[116] D. I. Perrett, A. J. Mistlin, and A. J. Chitty. Visual neurones responsive to faces. 
Trends in Neurosciences, 10:358-364, 1989. 


[117] D. I. Perrett, E. T. Rolls, and W. Caan. Visual neurones responsive to faces in the 
monkey temporal cortex. Exp. Brain Res., 47:329-342, 1982. 

[118] G. F. Poggio and T. Poggio. The analysis of stereopsis. Ann. Rev. Neurosci., 7:379-412, 

[119] T. Poggio and S. Edelman. A network that learns to recognize 3d objects, 1989. 
submitted for publication. 

[120] T. Poggio, E. B. Gamble, and J. J. Little. Parallel integration of vision modules. 
Science, 242:436-440, 1988. 

[121] T. Poggio and F. Girosi. A theory of networks for approximation and learning. A.I. 
Memo No. 1140, Artificial Intelligence Laboratory, Massachusetts Institute of Technol- 
ogy, 1989. 

[122] T. Poggio, V. Torre, and C. Koch. Computational vision and regularization theory. 
Nature, 317:314-319, 1985. 

[123] K.Prazdny. On the information in optical flow. Computer Vision, Graphics, and Image 
Processing, 22:239-259, 1983. 

[124] K. Prazdny. Detection of binocular disparities. Biological Cybernetics, 52:93-99, 1985. 

[125] H. Putnam. Minds and machines. In S. Hook, editor, Dimensions of mind. New York 
University Press, New York, NY, 1960. 

[126] H. Putnam. Mind, language and reality. Cambridge University Press, Cambridge, 1975. 

[127] H. Putnam. Representation and reality. MIT Press, Cambridge, MA, 1988. 

[128] Z. Pylyshyn. The role of location indexes in spatial perception: a sketch of the finst 
spatial-index model. Cognition, 32:65-97, 1989. 

[129] W. V. 0. Quine. Word and object. MIT Press, Cambridge, MA, 1960. 

[130] P. Quinlan. Visual object recognition reconsidered, 1989. submitted for publication. 

[131] V. S. Ramachandran. Perception of shape from shading. Nature, 331:163-166, 1988. 

[132] W. Richards, editor. Natural computation. MIT Press, Cambridge, MA, 1988. 

[133] J. Richter and S. Ullman. A model for the temporal organization of X- and Y-type 
receptive fields in the primate retina. Biological Cybernetics, 43:127-145, 1985. 

[134] I. Rock and J. DiVita. A case of viewer- centered object perception. Cognitive Psychol- 
ogy, 19:280-293, 1987. 

[135] I. Rock, D. Wheeler, and L. Tudor. Can we imagine how objects look from other 
viewpoints? Cognitive Psychology, 21:185-210, 1989. 

[136] J. Schwartz. The new connectionism. Proc. AAAS, 117:123-141, 1988. 


[137] A. Sha'ashua and S. Ullman. Structural saliency: the detection of globally salient 
structures using a locally connected network. In Proceedings of the 2nd International 
Conference on Computer Vision, pages 321-327, Tarpon Springs, FL, 1988. IEEE, 
Washington, DC. 

[138] S. A. Shafer and T. Kanade. Using shadows in finding surface orientation. Computer 
Vision, Graphics, and Image Processing, 22:145-176, 1983. 

[139] R. N. Shepard and J. Metzler. Mental rotation of three-dimensional objects. Science, 
171:701-703, 1971. 

[140] A. Sloman. What are the purposes of vision? CSRP 066, University of Sussex, 1987. 

[141] K. Stevens. The visual interpretation of surface contours. Artificial Intelligence, 17:47- 
75, 1981. 

[142] J. Stone and B. Dreher. Parallel processing of information in the visual pathways. 
Trends in Neurosciences, 3:441-446, 1982. 

[143] M. Tarr and S. Pinker. Mental rotation and orientation- dependence in shape recogni- 
tion. Cognitive Psychology, 21, 1989. 

[144] D. Terzopoulos. Regularization of inverse visual problems involving discontinuities. 
IEEE Transactions on Pattern Analysis and Machine Intelligence, 8:413-424, 1986. 

[145] D. W. Thompson and J. L. Mundy. Three-dimensional model matching from an uncon- 
strained viewpoint. In Proceedings of IEEE Conference on Robotics and Automation, 
pages 208-220, Raleigh, NC, 1987. 

[146] W. B. Thompson and J. K. Kearney. Inexact vision. In Workshop on motion, repre- 
sentation and analysis, pages 15-22, 1986. 

[147] J. T. Todd and E. Mingolla. Perception of surface curvature and direction of illumina- 
tion from patterns of shading. J. Exp. Psychol.: HPP, 9:583-595, 1983. 

[148] V. Torre and T. Poggio. A synaptic mechanism possibly underlying directional selec- 
tivity to motion. Proc. R. Soc. Lond. B, 202:409-416, 1978. 

[149] V. Torre and T. Poggio. On edge detection. IEEE Transactions on Pattern Analysis 
and Machine Intelligence, 8:147-163, 1986. 

[150] R. Tsai and T. Huang. Uniqueness and estimation of three dimensional motion pa- 
rameters of rigid objects with curved surfaces. IEEE Transactions on Pattern Analysis 
and Machine Intelligence, 6:13-27, 1984. 

[151] L. W. Tucker, C. R. Feynman, and D. M. Fritzsche. Object recognition using the 
Connection Machine. In Proceedings IEEE Conf. on Computer Vision and Pattern 
Recognition, pages 871-878, Ann Arbor, MI, 1988. 

[152] S. Ullman. The interpretation of visual motion. MIT Press, Cambridge, MA, 1979. 

[153] S. Ullman. Computational studies in the interpretation of structure and motion: sum- 
mary and extension. In J. Beck, B. Hope, and A. Rosenfeld, editors, Human and 
Machine Vision. Academic Press, New York, 1983. 


[154] S. Ullman. Maximizing rigidity: the incremental recovery of 3D structure from rigid 
and rubbery motion. Perception, 13:255-274, 1984. 

[155] S. Ullman. Visual routines. Cognition, 18:97-159, 1984. 

[156] S. Ullman. Aligning pictorial descriptions: an approach to object recognition. Cogni- 
tion, 32:193-254, 1989. 

[157] L. Vaina and N. M. Grzywacz. Structure from motion with impaired local-speed and 
global motion-field computations, 1989. submitted. 

[158] A. Verri and T. Poggio. Against quantitative optical flow. In Proceedings of the 1st 
International Conference on Computer Vision, pages 171-180, London, England, June 
1987. IEEE, Washington, DC. 

[159] R. von der Heydt, E. Peterhans, and G. Baumgartner. Illusory contours and cortical 
neurons' responses. Science, 224:1260-1262, 1984. 

[160] H. Voorhees and T. Poggio. Computing texture boundaries from images. Nature, 
333:364-367, 1988. 

[161] H. Wallach. On perceived identity: 1. the direction of motion of straight lines. In 
H. Wallach, editor, On Perception. Quadrangle, New York, 1976. 

[162] H. Wallach and D. N. O'Connell. The kinetic depth effect. J. Exp. Psychol., 45:205-217, 

[163] D. L. Waltz. Understanding line drawings of scenes with shadows. In P. Winston, 
editor, The Psychology of Computer Vision, McGraw-Hill, New York, 1975. 

[164] R. J. Watt and M. J. Morgan. Spatial filters and the localization of luminance changes 
in human vision. Vision Research, 24:1387-1397, 1984. 

[165] A. M. Waxman and S. Ullman. Surface structure and 3D motion from image flow: a 
kinematic analysis. International Journal of Robotics Research, 4:72-94, 1985. 

[166] D. Weinshall. Direct computation of 3d shape and motion invariants. A.I. Memo No. 
1131, Artificial Intelligence Laboratory, Massachusetts Institute of Technology, May 

[167] D. Weinshall. Seeing 'ghost' solutions in stereo vision. Nature, 1989. submitted for 

[168] D. Weinshall. Qualitative depth from stereo, with applications. Computer Vision, 
Graphics, and Image Processing, in press, January 1990. 

[169] A. P. Witkin. Recovering surface shape and orientation from texture. Artificial Intel- 
ligence, 17:17-45, 1981. 

[170] R. J. Woodham. Photometric method for determining surface orientation from multiple 
images. Optical Engineering, 19:139-144, 1980. 


[171] Y. Yeshurun and E. L. Schwartz. Cepstral filtering on a columnar image architecture: 
a fast algorithm for binocular stereo segmentation. IEEE Transactions on Pattern 
Analysis and Machine Intelligence, July 1989. 

[172] A. L. Yuille and N. M. Grzywacz. A computational theory for the perception of coherent 
visual motion. Nature, 333:71-74, 1988. 

[173] A. L. Yuille and N. M. Grzywasz. A winner-take-all mechanism based on presynaptic 
inhibition feedback, 1989. in press. 

[174] S. Zeki and S. Shipp. The functional logic of cortical connections. Nature, 335:311-317, 


CS-TR Scanning Project 

Document Control Form Date : /£/£o flH 

Report # ft; rr\\\S8 

Each of the following should be identified by a checkmark: 
Originating Department: 

JS> Artificial Intellegence Laboratory (Al) 

□ Laboratory for Computer Science (LCS) 

Document Type: 

□ Technical Report (TR) ^S Technical Memo (TM) 

□ Other: 

Document Information Number of pages: is(VM'/wo ) 

Not to include OOD forms, printer instructions, etc... original pages only. 

Originals are: Intended to be printed as : 

H Single-sided or □ Single-sided or 

□ Double-sided H. Double-sided 

Print type: 

□ Typewriter fj Offset Press yrj. Laser Print 

□ InkJet Printer □ Unknown fjj Other: 

Check each if included with document: 

^ DOD Form*( fK On Funding Agent Form D Cover Page 

□ Spine D Printers Notes □ Photo negatives 

□ Other: 

Page Data: 

Blank Pages( numbed 

Photographs/Tonal Material n* woe numbe d , o^l >H 

vjtner (rate description/page number): 

Description : Page Number: 

fa-yj) (p>c(P^ ^'£Q \-3H 

(3Q Joo^^rJT 

Scanning Agent Signoff: ^ (? '' v ° 0o& '5 

Date Received: fe /«&> I ^ Date Scanned: I I II /Hs Date Returned: / / A/^L 

Scanning Agent Signature:. 

°h^AJ %.^L 

Rev 9/94 DS/LCS Document Control Forni cstrtomi.vcd 

Scanning Agent Identification Target 

Scanning of this document was supported in part by 
the Corporation for National Research Initiatives, 
using funds from the Advanced Research Projects 
Agency of the United states Government under 
Grant: MDA972-92-J1029. 

The scanning agent for this project was the 
Document Services department of the M.I.T 
Libraries. Technical support for this project was 
also provided by the M.I.T. Laboratory for 
Computer Sciences. 


Date: if n /whs. 

M.I.T. Libraries 
Document Services 

darptrgt.wpw Rev. 9/94 


SECUBITv CL ASS ir 'C»TION OF TMIS PAGE fHfcan Daia Enfrtd) 



AIM 1158 


*■ TITLE Cnd Sublllt*) 

Computational vision: 

a critical review. 

7. *UTHOR(iJ 

Shimon Edelman 
Daphna Weinshall 


Artificial Intelligence Laboratory 
545 Technology Square 
Cambridge, MA 02139 


Advanced Research Projects Agency 
1400 Wilson Blvd. 
Arlington, VA 22209 . 

I*. MONITORING AGENCY NAME ft AOORESSfM dlllmtonl from Controlling Ollleo) 

Office of Naval Research 
Information Systems 
Arlington, VA 22217 


Distribution is unlimited 










Oct. 89 



IS. SECURITY CLASS, (of IM. topotx) 



17. DISTRIBUTION STATEMENT (ol (Ha aoatract onltttd In Black 30, II dlllmronl Horn Ropotl) 




fry and Idmntlly by mlotk numbot) 

computational vision 


I. ABSTRACT (Contlnuo on tororoo old* II nocoooary and Idontltr by a/oca nvmbor) 

We review the progress made in computational vision, as represented by 
Marr's approach, in the last fifteen years. First, we briefly outline 
computational theories developed for low, middle and high vision. We 
then discuss in more detail solutions proposed to three representative 
problems in vision, each dealing with a different level of visual 
processing. Finally, we discuss modifications to the currently established 
computational paradigm that appear to be dictated by the recent developments 

DD /, 

AN 71 


S/N O'.OJ-OM- 6601 I 






t>m r mmt U Jfl ill H i I . I BM ' i» > ■ ■ > *■' 


■ ■> M l H » » i — -« f W> . Kill ' I 

T^tSu*tw;cam S fittllliPPI^lWOOO TROW ^ 

in vision. 

*s»jw3 •<?*«*<» « r**«MM i-0 #•** ■* 

m iii B i kh ] i im ii i fc i n n « ii ii 

ecu Mil'" 

r . .i j |{ l ^ l 4in« l i l M ll Vf t I I WW i| » i lJi #j > !!l lf| l *nL l ; l [j|i^ 

ji w tlipi ' . i M- ^.^-m Mji i I bw . 

?8 .i»0 

_;, ; ^; jrL M,;.. L .. 

;w*|¥*» £•**!&& * *w*t*^ i«i»i^»a*i|Wft- 

.^^. ..^ . , ...«„ . . « i . W m !> i i..,jiiii i n i i,i i »ry i l i n i » ' !lll»» ' wn""»" ■»" '» «" " » " -"»" "" " " "^'g^f U***! 

9tieup3 ^Ioa*te*f ?*? 


t*K»&& •**•$#»* *:***•••* b»s«»v%* 

#»ie»«i»*8 i**«« H iw***$© , 

~is^r^-»s:y^ :«^ »»frw.w:* H 

«r«- ■ i .ii i , » h i n.imn i ii.i m . 1 . .. ■. » lVm*."Smi ' 3*mi±» Mm±±» !At 

v«I fasSawM.,- 'i .-ft t of»l«lv tm&l$M0W&m «* »**# ••fat**! **** %r«*vsrs «V 

<i/J£juo v Hat id tw ,3.**H .**»*t «***t'H 3»*I #** «1 trfnftoiqqs » f T»«ll 

sV ..roJUJv d-^/f ^i>« »!bbia .rfol'icrt W^ol***^ mi%<3&tl3 i.anuli£3yq«c» 

'jJL ^ rff ** ^J*- """ "" " ""■■"■"'»" n "" ' ""*"' ''■i--t-^ ! — — — ~~— — ' ' " ■" ' ■—' "" <nnmn> i m i iii i .i i iuii ii !ii> i i i M i i i . i M ii ..i nmuommm--- " — " ' 



■»wrtrf'Siar*<3K^Mp* iiKTWMStES» : *iiit3|*ff»i£«tt 

i.f^NI**i««l#£S **.