Skip to main content

Full text of "mit :: ai :: aim :: AIM-1218"

See other formats


MASSACHUSETTS INSTITUTE OF TECHNOLOGY 
ARTIFICIAL INTELLIGENCE LABORATORY 



A.I. Memo No. 1218 August, 1991 

Perceptual Organization, Figure-Ground, Attention And Saliency: 

Figure Has A Fuzzy Boundary, Its Outside/Near/Top/Incoming Regions Are More Salient (Or 
Not; Or What?), And Holes Are Independent Of The Whole 



J. Brian Subirana-Vilanova and Whitman Richards 



Abstract: Figure and ground are often viewed as binary complements to one another, with 
a well denned boundary between them. A simple experiment shows otherwise: if the contour of a 
simple convex shape is perturbed to create a distinctive texture, it is typically the outside of the 
contour that provides the basis for similarity judgement, not the inside. The introduction of the 
appropriate task, however, can make the inside part of the contour become more salient. A similar 
result occurs for concave shapes, such as a C, where notions of "inside" and "outside" are not well 
denned. Here, as well as with "holes", any proposal that directly relates figure to fixed aspects of 
objects fails. This leads us to propose an operational definition of "figure". 

Measures that assess similarity between shapes using a distance metric, cannot explain 
the above results. This leads us to suggest that there is a task-dependent bias in visual perception 
according to which the saliency of the two sides of a contour (inside and outside) is not the same. 
We suggest novel related biases such as "near is more salient than far", "top is more salient than 
bottom" and "expansion is more salient than contraction". We also discuss implications to visual 
perception; our findings seem to indicate that a frame is set in the image prior to recognition, and 
agree with a model in which recognition proceeds by the successive processing of convex chunks of 
image structures defined by this frame. 



Acknowledgements: This report describes research done at the Artificial Intelligence Lab- 
oratory of the Massachusetts Institute of Technology. Support for the Laboratory's Artificial In- 
telligence Research is provided in part by grant Sl-801534-2 from Hughes Aircraft Company and 
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. WR was also supported 
under NSF-IRI8900267. 



1 Introduction 

The natural world is usually conceived as being composed of different objects such 
as chairs, dogs or trees. This conception carries with it a notion that objects occupy a 
region of space, and have an "inside". By default, things outside this region of space 
are considered "outside" the object. Thus, the lungs of a dog are inside the dog, but 
the chair occupies a different region and is outside the object dog. When an object 
is projected into an image, these simple notions lead to what appears to be a clear 
disjunction between what is considered figure, and what is ground. Customarily, the 
figure is seen as the inside of the imaged shape as denned by its bounding contours 
(i.e. its silhouette). The region outside this boundary is ground. This implies that 
the points of an image are either figure or ground. Such a view is reinforced by 
reversible figures, such as Rubin's vase- face or Escher's patterns of birds and fish. 
Here, we show that such a simple disjunctive notion of figure and ground is incorrect, 
and that in general, the assignment of "figure" to the region of an image is ill-posed 
and consists of a fuzzy boundary. 

If figure has an ill-defined boundary, then are there some regions of the image that 
are receiving "more attention" than others? We contend that the answer is yes and 
that this is due not only to processing constraints but also to some computational 
needs of the perceiver. In particular, a fuzzy boundary leaves room to address regions 
of the image that are of immediate concern, such as the handle of a mug (its outside) 
that we are trying to grasp, or the inside surface of a hole that we are trying to 
penetrate. 

The ambiguity in defining a precise region of the image as figure arises in part 
because many objects in the world do not have clearly defined boundaries. Although 
objects occupy a region of space, the inside and outside regions of this space are 
uncertain. For example, what is the inside of a fir tree? Does it include the region 
between the branches where birds might nest, or the air space between the needles? 
If we attempt to be quite literal, then perhaps only the solid parts define the tree's 
exterior. But clearly such a definition is not consistent with our conceptual view of 
the fir tree which includes roughly everything within its convex hull. Just like the 
simple donut, we really have at least two and perhaps more conceptualizations of 
inside and outside. For the donut, the hole is inside it, in one sense, whereas the 
dough is inside it in another. But the region occupied by the donut for the most 
part includes both. Similarly for the fir tree, or for the air space of the mouth of a 
dog when it barks. Which of these two quite distinct inclusions of inside should be 
associated with the notion of figure? 

In this paper we will present some suggestions that attempt to answer these is- 
sues. Since they are coupled to some fundamental problems of visual perception such 
as perceptual organization, attention, reference frames and recognition, it will be 
necessary to address these, too. The suggestions are based on some simple observa- 






Figure 1: Fir tree at several scales of resolution. What is inside the tree? 



tions. The essence of them can be easily grasped by the reader by glancing at the 
figures of the paper. The text alternates the presentation of such observations with 
the discussion of the suggested implications. 

We begin, in the next section, by presenting some demonstrations that clarify 
how figural assignments are given to image regions. Along the way, we use these 
demonstrations to suggest a new, operational definition of "figure". In the following 
two sections we suggest that outside is more salient than inside; or not; or what? 
In section 5 we review the notion of "hole" and in sections 6 and 7 that of figure. 
In sections 8 and 9 we discuss the implications of our findings to visual perception. 
In section 10, we suggest that typically "near is more salient than far" and point to 
other similar biases. We end in section 11 with a summary of what is new about this 
paper. 

2 Fuzzy Boundaries 

Typically, figure-ground assignments are disjunctive, as in the Escher-drawings. 
However, when the image of a fractal-like object is considered, the exact boundary of 
the image shape is unclear, and depends upon the scale used to analyze the image. 
For the finest scale, perhaps the finest details are explicit, such as the needles of a 
spruce or the small holes through which a visual ray can pass unobstructed. But at 
the coarsest scale, most fractal objects including trees will appear as a smooth, solid 
convex shape. Any definition of figure must address this scale issue. Consider then, 
the following definition of figure: 

Definition 1 (Figure): "Figure" is that collection of image structures which cur- 
rently are supporting the analysis of a scene. 

By this definition, we mean to imply that the perceiver is trying to build or recover 
the description of an object (or scene) in the world, and his information-processing 



(L C x O 

aj b; c ; 

Figure 2: The notion of "what is figure" does not require that the figure be a region enclosed by 
a visible contour. In (a) the x is seen to lie within the C, and is associated with the figure, whereas 
in (b) the x lies outside the figure. In (c) the answer is unclear. 



capability is focused on certain regions in the scene that are directly relevant to this 
task. The precise regions of the scene that are being analyzed, and their level of detail 
will be set by the demands of the goal in mind. Such a definition implies that the 
regions of the scene assigned as figure may not have a well-defined, visible contour; 
and leads to the following claim: 

Claim 1 The region of the image currently assigned as "figure" may not have a 
well-defined boundary earmarked by a visible image contour. 

In support of this claim, and hence our initial definition of "Figure", consider 
the C of Figure 2. Although the image contour by itself is well-defined, the region 
enclosed by the C is not. We would like to argue that the region "enclosed" by the 
"C" is a legitimate figural assertion. For example, if one asks the question does the 
"X" lie inside the C, our immediate answer is yes for case (a), and no for case (b). 
To make this judgement, the visual system must evaluate the size of the interior 
region of the C. Thus, by our definition, the concept "inside of C" must lead to an 
assignment of certain pixels of the display as figure. Without an explicit contour in 
the image, however, where should one draw the boundary between the figure, and 
its complement, the ground? For example, should we choose to close the figure with 
a straight line between the two endpoints? Another possibility would be to find a 
spline that completes the curve in such a way that the tangent at the two endpoints 
of the C is continuous for the complete figure. (Oddly, most observers choose neither, 
but rather something more closely approaching a "blurred" version of the C, as if 
they were using a large Gaussian mask on a colored closed C.) We contend that 
such "fuzzy" figural boundaries occur not only within regions that are incompletely 
specified, such as that within the incomplete closing of the C, but also within regions 
that appear more properly defined by explicit image contours. 

To further clarify our definition of "figure", note that it is not prescribed by the 
retinal image, but rather by the collection of image structures in view. Any pixel- 
based definition of figure tied exclusively to the retinal image is inadequate, for it 
will not allow figural assertions to be made by a sequence of fixations of the object. 












Figure 3: Top row: Use the middle pattern as reference. Most see the left pattern as more similar 
to the reference. This could be because it has a smaller number of modified corners (with respect to 
the center) than the right one, and therefore, a pictorial match is better. Second row: In this case, 
the left and right stars look equally similar to the center one. This seems natural if we consider 
that both have a similar number of corners smoothed. Third row: Most see the left pattern as 
more similar despite the fact that both, left and right, have the same number of smoothed corners 
with respect to the center star. Therefore, in order to explain these observations, one cannot base 
an argument on just the number of smoothed corners. The position of the smoothed corners need 
be taken into account, i.e. preferences are not based on just pictorial matches. Rather, here the 
convexities on the outside of the patterns seem to drive our similarity judgement. 



Rather, a structure-based definition of figure presumes that the observer is building 
a description of an object or event, perhaps by recovering object properties. The 
support required to build these object properties is what we define as figure. This 
support corresponds closely to Ullman's incremental representations [Ullman 1984] 
upon which visual routines may act, and consequently the operations involved in 
figural assertions should include such procedures as indexing the sub-regions, marking 
these regions, and the setting of a coordinate frame. We continue with some simple 
observations that bear on these problems. 




Figure 4: Top: Reversible figure. Second Row: The contour (shown again in the center) that 
defined the previous reversible figure is modified in two similar ways (left and right contours). Third 
and fourth row: When such three contours are closed a preference exists, and this preference depends 
for most on the side used to close the contour. Use the center shape as reference in both rows. As 
in the example of the previous Figure most favor the outer portions of the shape to judge similarity. 
A distance metric, based solely on a pictorial match and that does not take into account the relative 
location of the different points of the shape, cannot account for these observations. 



3 Outside is More Salient than Inside 



When binary figure-ground assignments are made for an image shape with a 
well-defined, simple, closed contour, such as an "0", the assignment is equivalent 
to partitioning the image into two regions, one lying inside the contour, the other 
outside. For such a simple shape as the "0" , the immediate intuition is that it is the 
inside of the contour which is given the figural assignment, and this does not include 
any of the outside of the contour (see [Hoffman and Richards 1984] for example, where 
shape descriptors depend on such a distinction). By our definition, however, "figure" 
might also include at the very least a small band or ribbon outside the contour, 
simply because contour analysis demands such. As a step toward testing this notion, 
namely that a ribbon along the outer boundary of the shape should also be included 



when figural assignments are made, we perturb the contour to create simple textures 
such a those illustrated in Figure 3 (bottom row). 

In this figure, let the middle star-pattern be your reference. Given this reference 
pattern, which of the two adjacent patterns is the most similar? Virtually every- 
body we test immediately pick the pattern on the left (N > 20) 1 . Now look more 
closely at these two adjacent patterns. In the left pattern, the intrusions have been 
smoothed, whereas in the right pattern the protrusions are smooth. Clearly the sim- 
ilarity judgement is based upon the similarity of the protrusions, which are viewed 
as sharp convex angles. The inner discrepancy is almost neglected. 

The same conclusion is reached even when the contour has a more part-based 
flavor, rather than being a contour texture, as in Figure 4. Here, a rectangle has 
been modified to have only two protrusions 2 . Again, subjects will base their similarity 
judgments on the shape of the convex portion of the protrusion, rather than the inner 
concavity. 

This result is not surprising if shape recognition is to make any use of the fact that 
most objects in nature can be decomposed into parts. The use of such a property 
should indeed place more emphasis upon the outer portions of the object silhouette, 
because it is here that the character of a part is generally determined, not by the 
nature of its attachment. Almost all attachments lead to concavities, such as when 
a stick is thrust into a marshmallow. Trying to classify a three-dimensional object 
by its attachments is usually misguided, not only because many different parts can 
have similar attachments, but also because the precise form of the attachments is 
not reliably visible in the image. Hence the indexing of parts (or textures) for shape 
recognition can proceed more effectively by concentrating on the outer extremities. 

Another possible justification for such observations is that, in tasks such as grasp- 
ing or collision avoidance, the outer part is also more important and deserves more 
attention because it is the one that we are likely to encounter first 3 . 

The outer region of a shape is thus more salient than its inner region. This implies 
that the region of scene pixels assigned to figure places more weight on the outer, 
convex portions of the contour than on its interior concave elements (or to interior 
homogeneous regions), and leads to the following claim: 

Claim 2 The human visual system assigns a non-binary figure-ground function to 
scene pixels with greater weight given to regions near the outside of shapes, which 
become more salient. 



1 The results are virtually independent of the viewing conditions. However, if the stars sustain 
an angle larger than 10 degrees, the preferences may reverse. 

2 This is one sample from a collection of similar shapes that was used in an experiment too 
informal to report, but which clearly corroborated the observations presented in this paper. 

3 For example, in Figure 3 (bottom), the center and left stars would "hurt" when grasped, whereas 
the right star would not because it has a "smooth" outside. 






Figure 5: Left: Alligator. Center: Alligator with frame curve superimposed. The frame curve has 
been computed using a standard smoothing algorithm. Right: The outside of the different alligator's 
parts has been shaded by a simple coloring operation. The frame curve and the silhouette of the 
shape have been used as bounding contours for the coloring operation. Note that there is one shaded 
region per part. 



Note that this claim simply refers to "regions near the outside", not to whether 
the region is convex or concave. In Figure 4, the outer portion of the protrusion 
contains a small concavity, which presumably is the basis for the figural comparison. 

Exactly what region of the contour is involved in this judgement is unclear, and 
may depend upon the property being assessed. All we wish to claim at this point 
is that whatever this property, its principal region of support is the outer portion 
of the contour. The process of specifying just which image elements constitute this 
outer contour is still not clear, nor is the measure (nor weight) to be applied to these 
elements. One possibility is an insideness measure. Such a measure could be easily 
computed as a function of the distance of the image elements to the "smoothed" 
version of the contour (a circle in Figure 3 and something close to a rectangle in 
Figure 4). In this context, the smoothed contour corresponds to the notion of frame 
curves as used in [Subirana-Vilanova 1991]. 

This leads us to the following definition of frame curve which has to be read 
bearing in mind claim 1: 

Definition 2 (Frame Curve): A frame curve is a virtual curve in the image which 
lies in "the center" of the figure's boundary. 

In general, the frame curve can be computed by smoothing the silhouette of the 
shape. This is not always a well-defined process because the silhouette may be ill- 
defined or fragmented, and because there is no known way of determining a unique 
scale at which to apply the smoothing. Figure 5 (center) shows the frame curve for 
an alligator computed using such scheme. On the right, the regions of the shape 
that are "outside" the frame curve have been colored; note that these regions do not 
intersect, and correspond closely to the outer portions of the different parts of the 



shape. As mentioned above, these outer portions are both more stable and more 
likely to be of immediate interest. 

Note that Claim 2 supports Claim 1 because the figure-ground function mentioned 
in Claim 2 is to be taken to represent a fuzzy boundary for figure. The notion of 
frame curve should not be seen as a discrete boundary for figure (perhaps only at a 
first approximation). Indeed, we contend that a discrete boundary is not a realistic 
concept. 

4 Inside is More Salient than Outside 

Consider once more the three-star patterns of Figure 3. Imagine now that each of 
these patterns is expanded to occupy 20 degrees of visual angle (roughly your hand 
at 30 centimeters distance). In this case the inner protrusions may become more 
prominent and now the left pattern may be more similar to the middle reference 
pattern. (A similar effect can be obtained if one imagines trying to look through 
these patterns, as if in preparation for reaching an object through a hole or window.) 
Is this reversal of saliency simply due to a change in image size, or does the notion 
of a "hole" carry with it a special weighting function for figural assignments? 

For example, perhaps by viewing the central region of any of the patterns of 
Figure 3 as a "hole", the specification of what is outside the contour has been reversed. 
Claim 2 would then continue to hold. However, now we require that pixel assignments 
to "figure" be gated by a higher level cognitive operator which decides whether an 
image region should be regarded as an object "hole" or not. 

5 When a Hole Is Not a Hole 

Consider next the star patterns in Figure 6, which consist of two superimposed 
convex shapes, one inside the other. Again with the middle pattern as reference, most 
will pick as most similar the adjacent pattern to the right. This is surprising, because 
these patterns are generally regarded as textured donuts, with the inner-most region 
a hole. But if this is the case and our previous claim is to hold, then the left pattern 
should have been most similar. The favored choice is thus as if the inner star pattern 
were viewed as one object occluding another. Indeed, if we now force ourselves to take 
this view, ignoring the outer pattern, then the right patterns are again more similar 
as in Figure 3. So in either case, regardless of whether we view the combination as 
a donut with a hole, or as one shape occluding part of another, we still use the same 
portion of the inner contour to make our similarity judgement. The hole of the donut 
thus does not act like a hole. The only exception is when we explicitly try to put our 
hand through this donut hole. Then the inner-most protrusions become more salient 
as previously described for "holes". 

These results lead to the following claim: 

8 




Figure 6: Star patterns with "holes" treat the inside ring of the shape as if this ring was an 
occluding shape, i.e. as if it was independent from the surrounding contours, even if one perceives 
a donut like shape. 

Claim 2 (revisited): Once the attentional frame is chosen, then (conceptually) a 
sign is given to radial vectors converging or diverging from the center of this frame 
(i.e. the focal point). If the vector is directed outward (as if an object representation 
is accessed), then the outer portion of the encountered contours are salient. If the 
vector is directed inward to the focal point (as if a passageway is explored), then the 
inner portion of the contour becomes salient? 

So far we have been ignoring one important step in the processing of visual infor- 
mation: attention. It has long been known that humans concentrate the processing 
of images in certain regions or structures of the visual array. Attention has several 
forms: one of them, perhaps the most obvious, is gaze. We cannot explore a station- 
ary scene by swinging our eyes past it in continuous movements. Instead, the eyes 
jump with a saccadic movement, come to rest momentarily and then jump to a new 
locus of interest (see [Yarbus 1967]). 

4 There is an interesting exception to the rule: if the si»e of the contours is very big (in retinal 
terms) then the inside is always more salient (as if we where only interested in the inside of large 
objects). 



In the star patterns that we discussed in Section 2 (see Figure 3) the attention 
was focused primarily on the stars as whole objects. That is, there is a center of 
the figure that appears as the "natural" place to begin to direct our attention. The 
default location of this center, which is to become the center of a local coordinate 
frame, seems to be, roughly, the center of gravity of the figure [Richards and Kaufman 
1969], [Kaufman and Richards 1969], [Palmer 1983]. Attention is then allowed to be 
directed to locations within this frame. Consider next the shapes shown in the top 
of Figure 7. Each ribbon-like shape has one clear center on which we first focus our 
attention. So now let us bend each of the ribbons to create a new frame center which 
lies near the inner left edge of each figure (Figure 7, lower). Whereas before most 
regarded the left pattern as more similar to the middle reference, now the situation 
is starting to become confused. When the ribbons are finally closed to create the 
donuts of figure 6, the favored similarity judgement is for the right pattern. The 
primary effect of bending and closing the ribbon seems to be a shift in the relation 
between the attentional frame and the contours. Following the center of gravity 
rule, this center eventually will move outside the original body of the ribbon. This 
suggests that the judgments of texture similarity are dependent on the location of 
the attentional coordinate frame. 

Typically, as we move our gaze around the scene, the center of the coordinate 
frame will shift with respect to the imaged contours, thus altering the pixel assign- 
ments. A shift in the focus of attention without image movements can create an effect 
similar to altered gaze. More details regarding these proposed computations will be 
given in the next sections. What is important at the moment is that the saliency of 
figural assignments will depend upon the position of the contour with respect to the 
location of the center of the attentional coordinate frame. The reader can test this 
effect himself by forcing his attention to lie either within or outside the boundary of 
the ribbons. Depending upon the position chosen, the similarity judgments change 
consistently with Claim 2. 

As we move our attention around the figure, the focus of attention will shift, but 
the frame need not. But if the frame moves, then so consequently will the assignment 
of scene pixels to (potentially) active image pixels. The visual system first picks 
a (virtual) focal point in the scene, typically bounded by contours, and based on 
this focal point, defines the extent of the region (containing the focal point) to be 
included as "figure". If all events in the selected region are treated as one object or a 
collection of superimposed objects, then the radially distant (convex) portions of the 
contours drive the similarity judgments and are weighted more heavily in the figural 
computations. On the other hand, if the choice is made to regard the focal point as a 
visual ray along which something must pass through (such as a judgement regarding 
the size of a hole), then the contours that lie radially the closest are given greater 
weight (i.e. those that were previously concave). This led us to the revised version of 
claim 2, namely that the attentional coordinate frame has associated with it either 

10 








Figure 7: Top: Using the middle shape as a reference, most see the left shape as more similar. 
Bottom: If this same shape is bent, the situation becomes confused. 

an inward or outward pointing vector that dictates which portion of a contour will be 
salient (i.e. outer versus inner). We have argued that the orientation of this vector 
is task- dependent. 

Here we must introduce a contrary note. We do not propose that the attentional 
frame is imposed only upon a 3D structure seen as an object. Such a view would 
require that object recognition (or a 2 1/2 sketch) had already taken place. Rather, 
our claim is that the attentional coordinate frame is imposed upon a (frontal plane) 
silhouette or region prior to recognition, and is used to support the object recognition 
process, such as by indexing the image elements to a model. Hence, because a 
commitment is made to a coordinate frame and the sign of its object-associated 
vectors (inward or outward), proper object recognition could be blocked if either the 
location of the frame or the sign of its radial vectors were chosen improperly. When 
3D structure is obtained and analogous 3D frames are imposed, similar claims are 
possible but they are not the subject of this section. 



11 



6 Against Ground 

Our latest claim introduces a complementary, mutually exclusive state to the 
attentional coordinate frame within which figural processing is integrated. When the 
attentions! vector is pointing outward, as in "object mode", does this then imply 
that the contour regions associated with the inward state of this vector should be 
assigned to a separate state called "ground"? We argue against such an assignment, 
especially the view that objects in the foreground yield figure and that the rest of 
the display is ground. 

Consider the following experiment of [Rock and Sigman 1973] in which they 
showed a dot moving up and down behind a slit or opening, as if a sinusoidal curve 
was being translated behind it. The experiments were performed with slits of differ- 
ent shapes, so that in some cases the slit was perceived as an occluded surface and in 
others as an occluding one. They found that the perception of the curve is achieved 
only if the slit is perceived as an occluded region and not when it is perceived as 
an occluding region. Using their terms, the "correct" perception is achieved only if 
the slit is part of ground but not when it is part of the figure. Instead, we suggest 
that what is figure has not changed but rather its attributes did because the slit was 
viewed as a passageway between objects, and not as an object with a hole. This is 
an entirely different concept of "ground" . 

In support of our view, another experiment by [Rock and Gilchrist 1975] shows 
that figure need not correspond to the occluding surface. In this second experiment, 
they showed a horizontal line moving up and down with one end remaining in contact 
with one side of an outline figure of a face. Consequently, the line in the display 
changes in length. When the line is on the side of the face most observers see it 
changing size, adapting to the outline, while when it is on the other side of the 
contour, it is seen with constant length but occluded by the face. This has been 
described as a situation in which no figure-ground reversal occurs. However, in our 
terms the figure has changed because the attended region changes. In the first case 
figure corresponds to the occluding surface, and in the second to the occluded one. 
Thus, figure need not correspond to the occluding surface, even when the surfaces 
that are occluded are known. Again, this conclusion is consistent with our definition 
of figure and our claims. 

7 What's Figure? 

Our least controversial claim is that the image region taken as figure depends upon 
one's attention and goal. Reversible illusory patterns, such as the Escher drawings 
or Rubin's face-vase support this claim. The more controversial claim is that the 
image region taken as figure does not have a boundary that can be defined solely in 
terms of an image contour, even if we include virtual contours such as those cognitive 



12 



edges formed by the Kanizsa figures. The reason is two fold: First, the focal position 
of our attentional coordinate frame with respect to the contours determines that 
part of the contour used in figural similarity judgments, implying that the region 
assigned to figure has changed, or at the very least has been given altered weights. 
Second, whether the focal position is viewed as part of a passageway or alternatively 
simply as a hole in an object affects the figural boundary. In each case, the region 
is understood to lie within an object, but the chosen task affects the details of the 
region being processed. This effect is also seen clearly in textured C-shaped patterns, 
and becomes acute when one is asked to judge whether X lies inside the C, or if Y will 
fit into the C, etc. The virtual boundary assigned to close the C when making such 
judgments of necessity will also depend in part upon the size of the second object, 
Y. To simply assert that figure is that region lying inside an image contour misses 
the point of what the visual information processor is up to. 

Again, a simple experiment from the Rock laboratory demonstrates that "figure" 
is not simply a region of the display, but rather a collection of scene elements. Some 
of the elements within this "attended" region may be ignored, and thus, not be part 
of the structures at which higher level visual operations are currently being applied. 
[Rock and Gutman 1981] showed two overlapping novel outline figures, one red and 
one green, for a brief period, e.g. one second. Subjects were instructed to rate figures 
of a given color on the basis of how much they liked them (this attracts attention to 
one of the figures). They later presented subjects with a new set of outline figures 
and asked subjects whether they had seen these figures in the previous phase of the 
experiment, regardless of their color. They found that subjects were very good at 
remembering the attended shapes but failed on the unattended ones. This experiment 
agrees with the model presented in this paper, in which only the attended set of 
structures is being processed. Thus, figure is, clearly, not a region of pixels contained 
in the attended figure because the attended figure was partly contained in such a 
region and did not yield any high-level perception. 

In order to show that what they were studying was failure of perception and 
not merely selective memory for what was being attended or not attended, [Rock 
and Gutman 1981] did another experiment. They presented a series, just like in the 
previous case but with two familiar figures in different pairs of the series, one in 
the attended color and one in the unattended color. They found that the attended 
familiar figure was readily recognized but that the unattended familiar figure was 
not. It is natural that if the unattended figure is perceived and recognized it would 
stand out. Failure of recognition therefore supports the belief that the fundamental 
deficit is of perception. The extent of such deficit is unclear; it may be that the level 
of processing reached for the unattended figures is not complete but goes beyond that 
of figures not contained in the attended region. 

Therefore, an operational definition of "what is figure?" seems more fruitful in 
trying to understand how images are interpreted. Our definition is in this spirit, and 

13 



leads to a slightly different view of the initial steps involved in the processing of visual 
images than those now in vogue in computational vision. This is the subject of the 
next two sections. 

8 Perceptual Organization, Object Types and Figural Assignments: 
Which Comes First and the Role of Convexity 

There have been many proposals on what are the different steps involved in visual 
perception and it is not the main goal of this paper to make yet another such proposal. 
Nevertheless, our findings have some relevant implications to what should be the 
nature of these steps which we will now discuss. 

We suggest that figure and objects are not strongly coupled. Figure is simply 
the image-based structures which support some high-level processing, regardless of 
whether the region is assumed to be an object in the foreground, an occluded object, 
a hole, a passageway or none of the above. This does not mean that the current 
assumption of which type of region is figure does not play any role in the processing. 
Rather, we have shown several examples where the assumptions or role of the region 
is transformed onto an attribute of figure that governs both which portions of the 
contours are included in the processing and the type of processing to be done in it. 
Curiously, this suggests that a cognitive judgement proceeds and selects that portion 
of an image or contour to be processed for the task at hand. 

But how can a cognitive judgement anticipate where attention will be directed 
without some preliminary image processing that notes the current contours and 
edges? We are thus required to postulate an earlier, more reflexive mechanism that 
directs the eye, and hence the principal focus of attention, to various regions of the 
image. Computational studies suggest that the location of such focus may involve a 
bottom- up process such as the one described in [Subirana-Vilanova 1990]. Subirana- 
Vilanova's scheme computes points upon which further processing is directed using 
either image contours or image intensities directly. Regions corresponding to each 
potential point can also be obtained using bottom-up computations 5 . There is other 
computational evidence that bottom-up grouping and perceptual organization pro- 
cesses can correctly identify candidate interesting structures (see [Marroquin 1976], 
[Witkin and Tenenbaum 1983], [Mahoney 1985], [Harlick and Shapiro 1985], [Lowe 
1984, 1987], [Sha'ashua and Ullman 1988], [Jacobs 1989], [Grimson 1990], [Subirana- 



5 The result of these computations may affect strongly the choice of reference frames. For example, 
if the inner stars in Figure 6 are rotated so as to align with the outer stars (creating convexities in 
the space between the two), our attention seems more likely to shift to the region in-between the 
two stars and in this case the similarities will change in agreement with claim 2. 

Another way of increasing the preference for the "in-between" reference frame in Figures 6 and 
7 is by coloring the donut black and leaving the surrounding white (because in human perception 
there is a bias towards dark objects). 



14 



Vilanova 1990]). 

Psychological results in line with the Gestalt tradition [Wertheimer 1923], [Koffka 
1935], [Kohler 1940] argue for bottom-up processes too. However, they also provide 
evidence that top-down processing is involved. Other experiments argue in this di- 
rection, such as the one performed by [Kundel and Nodine 1983] in which a poor copy 
of a shape is difficult, if not impossible to segment correctly unless one is given some 
high level help such as "this image contains an object of this type". With the hint, 
perceptual organization and recognition proceed effortlessly. Other examples of top- 
down processing include [Newhall 54], [Rock 1983], [Cavanagh 1991], CM. Mooney 
and P.B. Porter's binary faces, and R.C. Jones' spotted dog. The role of top-down 
processing may be really simple, such as controlling or tweaking the behavior of an 
otherwise purely bottom-up process; or perhaps it involves selecting an appropriate 
model or structure among several ones computed bottom-up; or perhaps just index- 
ing. In either case the role of top-down processing cannot be ignored. Indeed, here 
we claim that the setting up of the attentional coordinate frame is an important early 
step in image interpretation. 

Our observations suggest that perceptual organization results in regions that are 
closed or convex (at a coarse scale) as discussed (see also section 5). This corrob- 
orates computational studies on perceptual organization which also point in that 
direction, demonstrating the effectiveness and the viability of perceptual organiza- 
tion schemes which limit themselves to finding convex or "enclosed" regions (or at 
least favor them) [Jacobs 1989], [Huttenlocher and Wayner 1990], [Subirana-Vilanova 
1990], [Clemens 1991], [Subirana-Vilanova and Sung]. It is still unclear if this is a 
general limitation of the visual system, a compound effect with inside and outside, or 
rather specific to shape perception. There are, however, several areas that may bring 
some more light onto the question. One of them is the study of the gamma effect: 
When a visual object is abruptly presented on a homogeneous background, its sudden 
appearance is accompanied by an expansion of the object. Similarly, a contraction 
movement is perceived if the object suddenly disappears from the visual field. Such 
movements were observed a long time ago and were named "gamma" movements by 
[Kenkel 1913], (see [Kanizsa 1979] for an introduction). For non-elongated shapes, 
the direction of movement of the figure is generally centrifugal (from the center out- 
ward for expansion and from the periphery toward the center for contraction). For 
elongated shapes, the movement occurs mainly along the perceptual privileged axes. 
It is unclear whether the movements are involved in the selection of figure or if, on 
the contrary are subsequent to it. In any case they might be related to a coloring 
process (perhaps responsible for the expansion movements) involved in figure selec- 
tion that would determine a non-discrete boundary upon which saliency judgments 
are established (see also [Mumford, Kosslyn, Hillger and Herrnstein 1987]). If this 
is true, studying the effect on non-convex shapes (such as those on Figure 7) may 
provide cues to what sort of computation is used when the figures are not convex, 

15 



and to the nature of the inside/outside asymmetry. 

Another area that may be interesting to study is motion capture which was ob- 
served informally by [Ramachandran and Anstis 83]: When an empty shape is moved 
in a dynamic image of random dots it "captures" the points that are inside it. This 
means that the points inside the shape are perceived as moving in the same direc- 
tion of the shape even though they are, in fact, stationary (randomly appearing for 
a short interval). This can be informally verified by the reader by drawing a circle 
on a transparency and sliding it through the screen of a connected TV with noise: 
The points inside the circle will be perceived as moving along with the circle. The 
results hold even if the circle has some gaps and it has been shown that they also 
hold when the shapes are defined by subjective contours [Ramachandran 86]. There 
is no clear study of what happens for non-convex shapes such as a C. What portions 
are captured? Informal experiments done in our laboratory seem to confirm that 
the boundary of the captured region is somewhat fuzzy for unclosed shapes like a 
C which supports the notion of a fuzzy boundary. In addition, the shape for the 
captured region seems to have convexity restrictions similar to the ones suggested 
for the inside-outside relations. It is unclear if both mechanisms are related but the 
similarity is intriguing. This seems a very promising direction for future research. 

Further evidence for the bias towards convex structures is provided by an aston- 
ishing result obtained recently by [Cumming, Hurlbert, Johnson and Parker 1991]: 
when a textured cycle of a sine wave in depth (the upper half convex, the lower half 
concave) is seen rotating both halfs may appear convex 6 , despite the fact that this 
challenges rigidity 7 (in fact, a narrow band between the two ribbons is seen as moving 
non-rigidly!). 

It is also of interest to study how people perceive ambiguous patterns or tilings 
[Tuijl 1980], [Shimaya and Yoroizawa 1990] that can be organized in several different 
ways. It has been shown that in some cases the preference for convex structures 
can overcome the preference for symmetric structures that are convex [Kanizsa and 
Gerbino 1976]. The interaction between convex and concave regions is still unclear, 
especially if the tilings are not complete. 

Studies with pigeons 8 [Herrnstein, Vaughan, Mumford and Kosslyn 1989] indicate 
that they can deal with inside-outside relations so long as the objects are convex but 
not when they are concave. It is unclear if some sort of "inside-outside" is used at 
all by the pigeons. More detailed studies could reveal the computation involved, and 



6 The surface can be described by the equation Z = sin(y) where Z is the depth from the fixation 
plane. The rotation is along the F-axis by +/ — 10 degrees at 1 Hz. 

This observation will be relevant later because it supports the notion that a frame is set in the 
image before structure from motion is recovered (see claim 3 and related discussion). 

8 The pigeon visual system, despite its reduced dimensions and simplicity, is capable of some 
remarkable recognition tasks that do not involve explicit inside/outside relations. See [Herrnstein 
and Loveland 1964], [Cerella 1982], [Herrnstein 1984] for an introduction. 



16 



perhaps whether they use a local feature strategy or a global one. This, in turn, may 
provide some insights into the limitations of our visual system. 

9 What is the Shape of Reference Frames? 

As described in the previous sections, our proposal implies that the establishment 
of a frame of reference is required prior to recognition. In other words, without the 
frame, which is used to set the saliency of the different image regions, recognition 
cannot proceed. We have pinned down three aspects of it: its location, its size and its 
inside and outside. Previous research on frames has focused on the orientation of such 
a frame (relevant results include, to name but a few [Attneave 1967], [Shepard and 
Metzler 1971], [Rock 1973], [Cooper 1976], [Wiser 1980], [Schwartz 1981], [Shepard 
and Cooper 1982], [Jolicoeur and Landau 1984], [Jolicoeur 1985], [Palmer 1985], 
[Corballis and Cullen 86], [Maki 1986], [Jolicoeur, Snow and Murray 1987], [Parsons 
and Shimojo 1987], [Robertson, Palmer and Gomez 1987], [Shepard and Metzler 
1988], [Corballis 1988], [Palmer, Simone and Kube 1988], [Georgopoulos, Lurito, 
Petrides, Schwartz and Massey 1989], [Tarr and Pinker 1989]), on the influence of the 
environment ([Mach 1914], [Attneave 1968], [Palmer 1980], [Palmer and Bucher 1981], 
[Humphreys 1983], [Palmer 1989]), on its location ([Richards and Kaufman 1969], 
[Kaufman and Richards 1969], [Cavanagh 1978], [Palmer 1983], [Cavanagh 1985], 
[Nazir and O'Reagan 1990]), and on its size ([Sekuler and Nash 1972], [Cavanagh 
1978], [Jolicoeur and Besner 1987], [Jolicoeur 1987], [Larsen and Bundsen 1987]). 
Exciting results have been obtained in this directions but it is not the purpose of this 
paper to review them. 

The shape of the frame, instead, has received very little attention. [Subirana- 
Vilanova 1990] proposed that in some cases, a curved frame might be useful (see also 
[Palmer 1989]). In particular, he proposed to recognize elongated curved objects, 
such as the ones shown in Figure 7, by unbending them using their main curved axis 
as a frame to match the unbended versions. If human vision used such a scheme, 
one would expect no differences in the perception of the shapes shown on the top of 
Figure 7 from those on the bottom of the same figure. As we have discussed, our 
findings suggest otherwise, which argues against such a mechanism in human vision. 

10 Related Effects: What Do You Want to be More Salient? 

The shapes used so far in our examples have been defined by image contours. 
The results, however, do not seem to depend on how such contours are established 
and similar results seem to hold when the shapes are defined by motion or other 
discontinuities. Thus, the results seem to reflect the true nature of shape perception. 
In this section we will suggest that similar biases in saliency occur in other dimensions 
of visual perception. What all of them have in common is that they require the 

17 



Figure 8: Top is more salient than bottom: Using the middle pattern as reference, most see the 
right contour as more similar. 



establishment of an attentional frame of reference at a very early stage, and that the 
nature of the frame depends on the task at hand. In particular, we will suggest that: 
top is more salient than bottom, near is more salient than far and outward motion 
is more salient than inward motion. 

Top is more salient than bottom; or not. 

Consider the contours in Figure 8, the center contour appears more similar to the 
one on the right than to the one on the left. We suggest that this is because the top of 
the contours is, in general, more salient than its bottom. We can provide functional 
justification similar to that given in the inside-outside case: the top is more salient 
because, by default, the visual system is more interested in it, as if it were the part of 
a surface that we contact first. Just like with our inside-outside notion, the outcome 
can be reversed by changing the task (consider they are the roof of a small room that 
you are about to enter). Thus, there is an asymmetry on the saliency of the two sides 
of such contour (top and bottom) similar to the inside/outside one discussed in the 
previous sections. 

Near is more salient than far; or not. 

When looking for a fruit tree of a certain species it is likely that, in addition we 
are interested in finding the one that is closer to us. Similarly, if we are trying to 
grasp something that is surrounded by other objects, the regions that are closer to 
our hand are likely to be of more interest than the rest of the scene. We suggest that 
when three-dimensional information is available, the visual system emphasizes the 
closer regions of the scene. Evidence is shown in Figure 9 in which we show a stereo 
pair with surfaces similar to the silhouette of the star of Figure 3. 

At a first glance, most see two of the three surfaces of Figure 3 as being more 
similar. The preference, as in the previous case, can be reversed if we change the 
task: imagine, for example, that you are flying above such surfaces and are looking 
for a place to land. Your attention will change to the far portions of the surfaces and 
with it your preferred similarities. Therefore, attention and the task at hand play 
an important role in determining how we perceive the three-dimensional world. Note 
also, that, as in the previous examples, a matching measure based on the distance 
between two surfaces cannot account for our observations. For in this case, such 



18 






Figure 9: This random dot stereo diagram illustrates that close structures are more salient in 
visual perception than those that are further away. 



distance to the center surface is the same for both bounding surfaces. 

Expansion is more salient than contraction; or not. 

Is there a certain type of motion that should be of most interest to the human 
visual system? Presumably, motion coming directly toward the observer is more rel- 
evant than motion away from it. Or, similarly, expanding motion should be more 
salient than contracting motion. Evidence in support of this suggestion is provided 
by a simple experiment illustrated in Figure 10 9 . Like in the previous cases, two 
seemingly symmetric percepts are not perceived equally by the visual system. This 
distinction, again, seems to bear on some simple task- related objectives of the ob- 
server. 

So, what's more salient? How does perception work? 

Inside/outside, near/far, expansion/contraction and top/bottom are generally not 
correlated. If saliency were determined independently for each of these relations, then 
conflicts could arise in some cases. For example, the inside of an object may be near 
or far, in the top or in the bottom of the image. Will, in this case, the outside regions 
on the bottom be more salient than those that are inside and on the top? 

This is an important issue that will not be addressed in this paper. A more 
detailed understanding of how attention and perceptual organization interact with 
the early vision modules is required. In any case, it would be interesting to find a 
modular division showing how these processes may interact. Unfortunately, this is 
a no-win situation. Either the modules are too few to be interesting or the division 

9 In a pool of 7 MIT graduate students, all but one reported that their attention was directed 
first at the expanding pattern. 



19 



Figure 10: When the dots in this two similar figures are moving towards the center on one of them, 
and towards the outside on the other, attention focuses first on the expanding flow. This provides 
evidence that motion coming directly toward the observer is more salient than motion away from 
it. 



is easily proven to be wrong. Nevertheless, it may be useful to give a proposal as 
precise as possible to illustrate what has been said so far. Figure 11 is it. 

Like in [Witkin and Tenenbaum 1983], our proposal is that grouping is done very 
early (before any 2 1/2 D sketch-like processing), but we point out the importance 
of selecting a coordinate frame which, among other things, is involved in top-down 
processing and can be used to index into a class of models. Indexing can be based 
on the coarse description of the shape that the frame can produce, or on the image 
features associated with the frame. As shown in Figure 11, this frame may later be 
enhanced by 3D information coming from the different early vision modules. Like 
in [Jepson and Richards 91], we suggest that one of the most important roles of the 
frame is to select and articulate the processing on the "relevant" structures of the 
image (see also footnote 7). This leads us to the last claim of this paper: 

Claim 3 An attentional "coordinate" frame is imposed in the image prior to con- 
structing an object description for recognition. 

11 What's New 

The fact that figure and ground reversals are attention related has been known 
for some time [Rubin 1921] 10 . However, there appears to be no precise statement of 
the relation between "figure" and notions of "inside" and "object", nor has it been 



10 [Rubin 1921] showed subjects simple contours where there was a two way ambiguity in what 
should be figure (similar to the reversible figure in the top of Figure 4). He found that if one region 
was found as figure when shown the image for the first time then, if on subsequent presentations 
the opposite region was found as figure, recognition would not occur. 



20 



Feature indexing 




Figure 11: A modular description of visual perception that illustrates some of the concepts dis- 
cussed in this paper. The different frames and image features depicted may share data structures; 
this accounts for some implicit feed-back in the diagram. (The Figure emphasizes the order in 
which events take place, not the detailed nature of the data structures involved.) It is suggested 
that perception begins by some simple image features which are used to compute a frame that is 
used to interpret these image features. The frame is an active structure which can be modified by 
visual routines. In this diagram, shape appears as the main source for recognition but indexing 
plays also an important role. Indexing can be based on features and on a rough description of the 
selected frame. 



21 



noted previously that contour saliency depends on inside/outside, near/far, expan- 
sion/contraction and top/bottom relations, and changes when the task is changed, 
such as viewing a region as something to pass thru, rather than as a shape to be 
recognized. 

These new observations support an operational definition of figure which is based 
on attention and which rules ground out of the picture. We have suggested that 
occlusion be treated as an attribute of figure. An alternative way of saying the same, 
is to call figure (as defined in this paper) the processing focus and to associate a 
figure/ ground attribute to it. Regardless of the syntaxis, any proposal that relates 
figure to object fails: A hole can become figure while the inside of a shape can become 
"ground". The idea that figure has a non-discrete boundary has not been suggested 
previously either. This leads to the concept of frame curve which can be used for 
shape segmentation in conjunction with Inside/Outside relationships. 

Our findings also demonstrate that the task at hand controls top-down processing. 
Existing evidence for top-down processing shows its role in increasing the speed and 
performance of recognition (by providing hints, such as restricting the set of models 
to be considered). However, a qualitative role of top-down processing (such as deter- 
mining whether we are looking for an object or a hole), not dependent on the image, 
like the one presented here has not been suggested previously. 

We have shown that "matching to model" will not correspond with human percep- 
tion unless inside/outside, top/bottom, expansion/contraction and near/far relations 
are factored early in the recognition strategy. We have also discussed several ways 
in which the role of convexity can be studied in human vision, such as inside/ outside 
relations, gamma movements and motion capture. Our observations provide new 
light into the nature of the attention and perceptual organization processes involved 
in visual perception. In particular, they indicate that a frame is set in the image 
prior to recognition and agree with a model in which recognition proceeds by the 
successive processing of convex chunks of image structures denned by this frame. 

Acknowledgments 

Part of this work was done while JBSV was at the NTT Human Interface Labora- 
tories in Yokosuka (Japan) during the winter of 1990; thanks to Kosugi-san, Kaneda- 
san and Sato-san for interesting discussions. Thanks also to Shimon Ullman, Ellen 
Hildreth, David Beymer, Tomaso Poggio and Eric Grimson for interesting discussions 
and suggestions for improving the presentation of this paper. 

References 

[1] F. Attneave. Triangles as ambiguous figures. American Journal of Psychology, 
81:447-453, 1968. 

22 



[2] F. Attneave and R.K. Olson. Discriminability of stimuli varying in physical and 
retinal orientation. Journal of Experimental Psychology: Human Perception and 
Performance, 74:149-157, 1967. 

[3] E.G. Boring. A history of experimental psychology. Appleton Century Crofts, 
Inc., 1964. Second Edition. Originally published 1929. 

[4] P. Cavanagh. Size and position invariance in the visual system. Perception, 
7:167-177, 1978. 

[5] P. Cavanagh. Local log polar frequency analysis in the striate cortex as a basis 
for size and orientation invariance. In D. Rose and V. Dobson, editors, Models 
of the visual cortex, pages 85-95. Wiley, 1985. 

[6] P. Cavanagh. What's up in top-down processing? In Andrei Gorea, editor, 
Proceedings of the Xlllth european conference on visual perception. 1990. 

[7] J. Cerella. Mechanisms of concept formation in the pigeon. In D.J. Ingle, M.A. 
Goodale, and R.J.W. Mansfield, editors, Analysis of visual behavior, pages 241- 
260. The MIT Press, Cambridge and London, 1982. 

[8] D. Clemens. Region-based feature interpretation for recognizing 3D models in 
2D images. PhD thesis, Massachusetts Institute of Technology, Cambridge, MA, 
1991. 

[9] L.A. Cooper. Demonstration of a mental analog to an external rotation. Perce- 
pion and Psychophysics, 1:20-43, 1976. 

[10] M.C. Corbalis. Recognition of disoriented shapes. Psychological Review, 95:115- 
123, 1988. 

[11] M.C. Corbalis and S Cullen. Decisions about the axes of disoriented shapes. 
Mem. Cognition, 14:27-38, 1986. 

[12] B.G. Cumming, A.C. Hurlbert, E.B. Johnson, and A.J. Parker. Effects of texture 
and shading on the KDE. In The Association for Research in Vision and Oph- 
thalmology. Annual Meeting Abstract Issue. Vol 32, NO. 4, page 1277, Bethesda, 
Maryland 20814-3928, 1991. 

[13] A. P. Georgopoulos, J.T. Lurito, M. Petrides, A.B. Schwartz, and J.T. Massey. 
Mental rotation of the neuronal population vector. Science, 243:234-236, 1989. 

[14] R.L. Gregory. The intelligent eye. McGraw-Hill Book Company, New York, St. 
Louis and San Francisco, 1970. 



23 



[15] W.E.L. Grimson. Object Recognition By Computer. The Role Of Geometric 
Constraints. The MIT Press, Cambridge and London, 1990. 

[16] R.M. Haralick and L.G. Shapiro. Image segmentation techniques. Computer 
Vision, Graphics, and Image Processing, 29:100-132, 1985. 

[17] R.J. Herrnstein. Objects, categories, and discriminative stimulus. In H.L. Roit- 
blat, T.G. Bever, and H.S. Terrace, editors, Animal Cognition: Proceedings of 
the Frank Guggenheim Conference. Lawrence Erlbaum Associates, Hillsdale N.J, 
1984. 

[18] R.J. Herrnstein, W. Vaughan Jr., D.B. Mumford, and S.M. Kosslyn. Teaching 
pigeons an abstract relational rule: Insideness. Percepion and Psychophysics, 
46(l):56-64, 1989. 

[19] R.J. Herrnstein and D. Loveland. Complex visual concept in the pigeon. Science, 
46:549-551, 1964. 

[20] D.D. Hoffman and W.A. Richards. Parts of recognition. In S. Pinker, editor, 
Visual Cognition, pages 2-96. The MIT Press, Cambridge, MA, 1984. 

[21] G.W. Humphreys. Reference frames and shape perception. Cognitive Psychology, 
15:151-196, 1983. 

[22] D.P. Huttenlocher and P. Wayner. Finding convex edge groupings in an image. 
TR 90-1116, Department of Computer Sciences, Cornell University, Ithaca, New 
York, 1990. 

[23] D.W. Jacobs. Grouping for recognition. A.I. Technical Report No. 1117, Artifi- 
cial Intelligence Laboratory, Massachusetts Institute of Technology, Cambridge, 
MA, 1989. 

[24] R.C. James. The dalmatian dog, photographed by R.C. James has been repro- 
duced in a number of different publications. A binary version can be found in 
[Marr 82] and a non-discretized one in [Gregory 1970]. 

[25] A. Jepson and W. Richards. Integrating vision modules. IEEE Systems and 
Cybernetics, 1991. Special Issue on Assimilation. Editor: A. Jain. To appear. 

[26] P. Jolicoeur. The time to name disoriented natural objects. Mem. Cognition, 
13(4):289-303, 1985. 

[27] P. Jolicoeur. A size-congruency effect in memory for visual shape. Mem. Cogni- 
tion, 15(6):531-543, 1987. 



24 



[28] P. Jolicoeur and D. Besner. Additivity and interaction between size ratio and 
response category in the comparison of size-discrepant shapes. Journal of Exper- 
imental Psychology: Human Perception and Performance, 13(3):478-487, 1987. 

[29] P. Jolicoeur and M.J. Landau. Effects of orientation on the identification of 
simple visual patterns. Canadian Journal of Psychology, 38(l):80-93, 1984. 

[30] P. Jolicoeur, D. Snow, and J. Murray. The time to identify disoriented letters: 
Effects of practice and font. Canadian Journal of Psychology, 41(3):303-316, 
1987. 

[31] G. Kanizsa. Organization in Vision. Praeger, 1979. 

[32] G. Kanizsa and W. Gerbino. Convexity and symmetry in figure-ground organi- 
zation. In M. Hele, editor, Vision and Artifact. Springer, New York, 1976. 

[33] L. Kaufman and W. Richards. Spontaneous fixation tendencies for visual forms. 
Percepion and Psychophysics, 5(2):85-88, 1969. 

[34] F. Kenkel. Untersuchungen uber den zusammenhang zwischen erscheinungs- 
grosse und erscheinungsbewegung bei einigen sogennanten optischen tauschun- 
gen. Zeitschrift fur Psychologie, 67:358-449, 1913. 

[35] A. Koffka. The principles of Gestalt psychology. Harcourt, Brace, New York, 
1940. 

[36] W. Kohler. Dynamics in psychology. Liveright, New York, 1940. 

[37] H.L. Kundel and C.F. Nodine. A visual concept shapes image perception. Ra- 
diology, 146(2):363-368, 1983. 

[38] A. Larsen and C. Bundesen. Size scaling in visual pattern recognition. Journal of 
Experimental Psychology: Human Perception and Performance, 4(l):l-20, 1978. 

[39] D.G. Lowe. Perceptual Organization and Visual Recognition. PhD thesis, Stand- 
ford University, 1984. 

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

[41] E. Mach. The analysis of sensations. Chicago: Open Court, 1914. 

[42] J.V. Mahoney. Image chunking: defining spatial building blocks for scene anal- 
ysis. A.I. Technical Report No. 980, Artificial Intelligence Laboratory, Mas- 
sachusetts Institute of Technology, Cambridge, MA, 1987. 



25 



[43] R. Maki. Naming and location the tops of rotated pictures. Canadian Journal 
of Psychology, 40(l):368-387, 1986. 

[44] D. Marr. Analysis of occluding contour. Proceedings of the Royal Society of 
London B, 197:441-475, 1977. 

[45] D. Marr. Vision: A Computational Investigation into the Human Representation 
and Processing of Visual Information. W.H. Freeman and Company, New York, 
1982. 

[46] J.L. Marroquin. Human perception of structure. Master's thesis, Massachusetts 
Institute of Technology, 1976. 

[47] D. Mumford, S.M. Kosslyn, L.A. Hillger, and R.J. Herrnstein. Discriminating 
figure from ground: The role of edge detection and region growing. Proceedings 
of the National Academy of Science, 87:7354-7358, 1984. 

[48] T.A. Nazir and J.K. O'Reagan. Some results on translation invariance in the 
human visual system. Spatial Vision, 5(2):81-100, 1990. 

[49] S.M. Newhall. Hidden cow puzzle picture. American Journal of Psychology, 
65(110), 1954. 

[50] S.E. Palmer. Structural aspects of visual similarity. Mem. Cognition, 6(2):91-97, 
1978. 

[51] S.E. Palmer. What makes triangles point: Local and global effects in configura- 
tions of ambiguous triangles. Cognitive Psychology, 12:285-305, 1980. 

[52] S.E. Palmer. On goodness, gestalt, groups, and garner. Paper presented at 
Annual Meeting of Psychonomic Society. San Diego, California. Unpublished 
manuscript, 1983. 

[53] S.E. Palmer. The role of symmetry in shape perception. Acta Psychologica, 
59:67-90, 1985. 

[54] S.E. Palmer. Reference frames in the perception of shape and orientation. In 
B.E. Shepp and S. Ballesteros, editors, Object perception: Structure and process, 
pages 121-163. Hillsdale, NJ: Lawrence Erlbaum Associates, 1989. 

[55] S.E. Palmer and N.M. Bucher. Configural effects in perceived pointing of am- 
biguous triangles. Journal of Experimental Psychology: Human Perception and 
Performance, 7:88-114, 1981. 

[56] S.E. Palmer, E. Simone, and Paul Kube. Reference frame effects on shape 
perception in two versus three dimensions. Perception, 17:147-163, 1988. 

26 



[57] L.M. Parsons and S. Shimojo. Percieved spatial organitzation of cutaneous pat- 
terns on surfaces of the human body in various positions. Journal of Experimen- 
tal Psychology: Human Perception and Performance, 13(3):488-504, 1987. 

[58] P.B. Porter. Another picture puzzle. American Journal of Psychology, 67:550- 
551, 1954. 

[59] V.S. Ramachandran. Capture of stereopsis and apparent motion by illusory 
contours. Percepion and Psychophysics, 39(5):361-373, 1986. 

[60] V.S. Ramachandran and S.M. Anstis. Displacement thresholds for coherent 
apparent motion in random dot-patterns. Vision Research, 23(12):1719-1724, 
1983. 

[61] W. Richards and L. Kaufman, "center-of-gravity" tendencies for fixations and 
flow patterns. Percepion and Psychophysics, 5(2):81-84, 1969. 

[62] L.C. Roberson, S.E. Palmer, and L.M. Gomez. Reference frames in mental rota- 
tion. Journal of Experimental Psychology: Human Perception and Performance, 
13(3):368-379, 1987. 

[63] I. Rock. Orientation and Form. Academic Press, New York, 1973. 

[64] I. Rock. The logic of perception. The MIT Press, Cambridge, MA, 1983. 

[65] I. Rock and A. Gilchrist. The conditions for the perception of the covering and 
uncovering of a line. American Journal of Psychology, 88:571-582, 1975. 

[66] I. Rock and D. Gutman. The effect of inattention on form perception. Journal 
of Experimental Psychology: Human Perception and Performance, 7:275-285, 
1983. 

[67] I. Rock and E. Sigman. Intelligence factors in the perception of form through a 
moving slit. Perception, 2:357-369, 1973. 

[68] E. Rubin. Visuell Wahrgenommene Figuren. Glydendalske, 1921. See [Boring 
1964] Pg. 605, or [Rock 83] Pg. 306. 

[69] R. Sekuler and D. Nash. Speed of size scaling in human vision. Psychonomic 
Science, 27:93-94, 1972. 

[70] A. Sha'ashua and S. Ullman. Structural saliency: The detection of globally 
salient structures using a locally connected network. In Proceedings of the In- 
ternational Conference on Computer Vision, pages 321-327, 1988. 



27 



[71] R.N. Shepard and L.A. Cooper. Mental Images and their Transformations. The 
MIT Press, Cambridge, MA, 1982. 

[72] S. Shepard and D. Metzler. Mental rotation of three dimensional objects. Sci- 
ence, 171:701-703, 1971. 

[73] S. Shepard and D. Metzler. Mental rotation: Effects of dimensionality of objects 
and type of task. Journal of Experimental Psychology: Human Perception and 
Performance, 14(1):3-11, 1988. 

[74] A. Shimaya and I. Yoroizawa. Automatic creation of reasonable interpretations 
for complex line figures. In Proceedings Int. Conf. on Pattern Recognition, pages 
480-484, Atlantic City, New Jersey, 1990. 

[75] S.P. Shwartz. The perception of disoriented complex objects. In Proceedings of 
the 3rd Conference on Cognitive Sciences, pages 181-183, Berkeley, 1981. 

[76] J.B. Subirana-Vilanova. The skeleton sketch: Finding salient frames of reference. 
In Proceedings Image Understanding Workshop, 1990, pages 399-414. Morgan 
and Kaufman, 1990. Also in the proceedings of ICCV'90. 

[77] J.B. Subirana-Vilanova. On contour texture. In Proceedings IEEE Conf. on 
Computer Vision and Pattern Recognition, pages 753-754, Ann Arbor, MI, 1991. 

[78] J.B. Subirana-Vilanova and W. Richards. Figure-ground in visual perception. 
In The Association for Research in Vision and Ophthalmology. Annual Meeting 
Abstract Issue. Vol 32, NO. 4, page 697, Bethesda, Maryland 20814-3928, 1991. 

[79] J.B. Subirana-Vilanova and K.K. Sung. In preparation. 

[80] M. Tarr and S. Pinker. Mental rotation and orientation- dependence in shape 
recognition. Cognitive Psychology, 21:233-282, 1989. 

[81] A. Treisman and G. Gelade. A feature integration theory of attention. Cognitive 
Psychology, 12:97-136, 1980. 

[82] H. Tuijil. Perceptual interpretation of complex line patterns. Journal of Exper- 
imental Psychology: Human Perception and Performance, 6(2), 1983. 

[83] S. Ullman. Visual routines. Cognition, 18, 1984. 

[84] D.L. Waltz. Understanding line drawings of scenes with shadows. In P. Winston, 
editor, The Psychology of Computer Vision. McGraw-Hill, New York, 1972. 



28 



[85] M. Wertheimer. Principles of perceptual organization. In B. Beardslee and 
M. Wertheimer, editors, Readings in perception. Van Nostrand, Princeton, 1958. 
Originally published in 1923. 

[86] M.A. Wiser. The role of intrinsic axes in the mental representation of shapes. 
PhD thesis, Massachusetts Institute of Technology, Cambridge, MA, 1980. See 
also the Proceedings of 3rd Conference on Cognitive Sciences, Berkeley, pp. 184- 
186, 1981. 

[87] A.P. Witkin and J.M. Tenenbaum. On the role of structure in vision. In J. Beck, 
B. Hope, and A. Rosenfeld, editors, Human and Machine Vision. Academic 
Press, New York, 1983. 

[88] A.L. Yarbus. Eye movements and Vision. Plenum, New York, 1967. 



29 



REPORT DOCUMENTATION PAGE 



form Approved 
OMB No 07044188 



Public reporting burden for this collection of information is estimated to averse 1 hour per response, including the time for reviewing instructions, searching existing data sources, 
gathering and maintaining the data needed, and completing and reviewing the collection of information. Send comments regarding this burden estimate or any other aspect of this 
collection of information, including suggestions for reducing this burden, to Washington Headquarters Services. Directorate tor Information Operations and Reports. 1 215 Jefferson 
Davis Highway. Suite 1 204. Arlington. V A 22202-4302. and to the Office of Management and Budget. Paperwork Reduction Protect (0704-0 188), Washington. DC 20S03 



1. AGENCY USE ONLY (Leave blink) 



REPORT DATE 

August 1991 



3. REPORT TYPE AND DATES COVERED 

memorandum 



4. TITLE AND SUBTITLE 

Perceptual Organization, Figure-Ground, Attention and 
Saliency 



6. AUTHOR(S) 

J. Brian Subirana 



5. FUNDING NUMBERS 

Sl-801534-2 
DACA76-85-C-0010 
N00014-85-K-0124 
NSF-IRI8900267 



7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) 

Artificial Intelligence Laboratory 
545 Technology Square 
Cambridge, Massachusetts 02139 



8. PERFORMING ORGANIZATION 
REPORT NUMBER 



AP '+?•&%? 



9. SPONSORING /MONITORING AGENCY NAME(S) AND ADDRESS(ES) 

Office of Naval Research 
Information Systems 
Arlington, Virginia 22217 



10. SPONSORING /MONITORING 
AGENCY REPORT NUMBER 



AIM 1218 



11. SUPPLEMENTARY NOTES 

None 



12a. DISTRIBUTION /AVAILABILITY STATEMENT 

Distribution of this document is unlimited 



12b. DISTRIBUTION CODE 



13. ABSTRACT (Maximum 200 words) 

Abstract: Figure and ground are often viewed as binary complements to one another, with 
a well denned boundary between them. A simple experiment shows otherwise: if the contour of a 
simple convex shape is perturbed to create a distinctive texture, it is typically the outside of the 
contour that provides the basis for similarity judgement, not the inside. The introduction of the 
appropriate task, however, can make the inside part of the contour become more salient. A similar 
result occurs for concave shapes, such as a C, where notions of "inside" and "outside" are not well 
denned. Here, as well as with "holes", any proposal that directly relates figure to fixed aspects of 
objects fails. This leads us to propose an operational definition of "figure". 

Measures that assess similarity between shapes using a distance metric, cannot explain 
the above results. This leads us to suggest that there is a task-dependent bias in visual perception 
according to which the saliency of the two sides of a contour (inside and outside) is not the same. 



(continued on back) 



14. SUBJECT TERMS ( key wor ds) 

vision 



15. NUMBER OF PAGES 

29 



16. PRICE CODE 



17. SECURITY CLASSIFICATION 
OF REPORT 

UNCLASSIFIED 



18. SECURITY CLASSIFICATION 
OF THIS PAGE 

UNCLASSIFIED 



19. SECURITY CLASSIFICATION 
OF ABSTRACT 

UNCLASSIFIED 



20. LIMITATION OF ABSTRACT 



UNCLASSIFIED 



IMSN 7540-01-280-5500 



Standard Form 298 (Rev. 2-89) 

Prescribed by ANSI Std Z39-18 
298-102 



Block 13 continued: 

We suggest novel related biases such as "near is mote salient than far", "top is more salient than 
bottom" and "expansion is more salient than contraction". We also discuss implications to visual 
perception; our findings seem to indicate that a frame is set in the image prior to recognition, and 
agree with a model in which recognition proceeds by the successive processing of convex chunks of 
image structures defined by this frame. 



CS-TR Scanning Project 

Document Control Form Date : fi / 3 /^M 

Report # Ai fh- I3J'* 

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

Bf Artificial Intellegence Laboratory (AI) 
D Laboratory for Computer Science (LCS) 

Document Type: 

□ Technical Report (TR) ^K. Technical Memo (TM) 

□ Other: 

Document Information Number of pages: 3o 

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

Originals are: Intended to be printed as : 

£f Single-sided or □ Single-sided or 

□ Double-sided J2( Double-sided 

Print type: 

fj Typewriter fj Offset Press )^ Laser Print 

□ InkJet Printer □ Unknown □ Other: 

Check each if included with document: 

J5^ DOD Form,).(Pc5 □ Funding Agent Form D Cover Page 

D Spine □ Printers Notes □ Photo negatives 

□ Other: 

Page Data: 

Blank Pageso^paoenumbeo: 



Photographs/Tonal Material (*(»«• nU mi»o: 

Utner (note description/page number). 

Description : Page Number: 



Scanning Agent Signoff: 

Date Received: /) lodl'H Date Scanned: /I /OV/lfr Date Returned: H lt& /?ft 



Scanning Agent Signature:. 



0^1,l/k/.r4> 



Rev 9/94 DS/LCS Document Control Form cstrform vsd