1
Object and Face
Recognition
2
What is visual agnosia?
3
The impairment of visual object recognition in people who
possess sufficiently preserved visual fields, acuity and other
elementary forms of visual ability to enable object
recognition, and in whom the object recognition impairment
cannot be attributed to... loss of knowledge about objects...
[The] impairment is one of visual recognition rather than
naming, and is therefore manifest on naming and non-verbal
tasks alike.
Martha Farah
Two types of visual agnosia
Apperceptive agnosia
Object recognition is impaired because of deficits
in perceptual processing.
Associative agnosia
Perceptual processes remain intact but object
recognition is impaired because of difficulties in
accessing relevant knowledge about objects from
memory.
4
Perception and Action
8
We typically consider object recognition as involving
perceptual and memory processes - you see an
object and then identify it. This is the “sitting in the
chair” approach.
But object recognition often goes beyond visual
recognition alone - we often interact with objects.
Physical interactions with objects often facilitates
object recognition.
The What and Where Pathways
10
Ungerleider and Mishkin (1982) performed lesions
in either the temporal and parietal lobes of
monkeys.
The monkeys were asked to perform an object
discrimination or a landmark discrimination task.
Leslie
Ungerleider
Monkeys with parietal lesion could not perform
the landmark discrimination task.
Monkeys with temporal lesion could not perform
the object discrimination task.
Mortimer
Mishkin
A Double Dissociation
Object Discrimination (What)
Landmark Discrimination (Where)
Pick the correct shape (triangle
or rectangle) for a reward.
Pick the shape closer to the cylinder
for a reward.
11
12
13
Explaining Visual Agnosia
Hierarchical model of object recognition
14
Riddoch and Humphreys proposed a model to account for
the various deficits in object recognition shown by brain
damaged patients.
• Edge grouping by collinearity
• Feature binding into shapes
• View normalization
• Structural description
• Semantic system
Glyn
Humphreys
Jane
Riddoch
15
Edge Grouping by Collinearity
An early stage of processing where edges of objects
are derived.
Collinearity means having a common line.
...........
:::
16
Evidence for Edge Grouping
17
Patient DF had severely impaired object recognition
ability. She recognized only a few real objects and
could not recognize any objects in line drawings (D.
Milner, 1991).
She also had trouble with recognizing line
orientation, which is important for detecting edges.
18
Principle of Good Continuation
Separate points/sections are perceived as if they
form a straight line or smooth curve when they are
connected.
Principle of Similarity
Similar things appear to be grouped together.
Feature Binding
19
Object features that have been extracted during the
edge grouping stage and the feature detection stage
are combined to form shapes.
This stage is akin to recognizing geons and/or geon
assemblies in the RBC theory.
Evidence for Feature Binding
Integrative agnosia: Patients with this condition have
trouble combining or integrating features of an object.
Patient HJA cannot find an inverted T among upright Ts,
presumably because he found it hard to group the
distractors together.
“I have come to cope with recognizing many common
objects, if they are standing alone. When objects are
placed together, though, I have more difficulties. To
recognize one sausage on its own is far from picking
one out from a dish of cold foods in a salad.”
20
Evidence for Feature Binding
23
Integrative agnosia: Patients with this condition have
trouble combining or integrating features of an object.
Patient HJA cannot find an inverted T among upright Ts,
presumably because he found it hard to group the
distractors together.
“I have come to cope with recognizing many common
objects, if they are standing alone. When objects are
placed together, though, I have more difficulties. To
recognize one sausage on its own is far from picking
one out from a dish of cold foods in a salad.”
More Evidence for Feature Binding
24
HJA did OK with
configurations like
this.
But not with
these.
View normalization
25
View normalization allows a viewpoint-invariant
representation to be derived. This is a controversial
idea, because evidence generally suggests that
successful object recognition does not require
viewpoint-invariant representations.
Evidence for normalization
Warrington and Taylor asked patients to recognize
objects presented at either a normal or an unusual view.
These patients performed particularly poorly when
objects were shown at an unusual angle.
The same occurred when Warrington and
Taylor showed two pictures simultaneously,
with each depicting an object at a different
angle, and had patients judge whether they
were the same object.
Elizabeth
Warrington
26
Structural Description
29
During this stage, individuals gain access to stored
knowledge about the structure (i.e., visual
appearance) of objects.
Evidence for Structural Description
30
Object-decision task -- Ss see pictures or drawings of
real and pseudo-objects, and they must decide which
are real.
Some patients perform poorly on this task even though
they perform at normal levels on tasks designed to
assess earlier stages of object recognition (e.g.,
matching objects from different viewing angles).
Some patients perform extremely poorly when they have
to name objects presented visually, but can perform
normally when naming objects presented verbally.
Semantic System
31
At this stage, people gain access to stored
knowledge of semantic (nonvisual) information
relevant to the object.
Evidence for Semantic System
Patients with an impaired semantic system can show
category-specific deficits in object recognition.
Category-Specific Deficit
Impaired for living things but intact for non-living
things.
Impaired for man-made things (though much less
common).
33
Category Specific Deficits
34
1.00
Recognition Accuracy
Living
Non-living
0.75
0.50
Living things can be
animals, vegetables...
0.25
Non-living things can
be tools, stationary,
utensils, appliances...
0.00
PS
JJ
Why do more patients have trouble with living things?
35
Living things are more visually similar to each other than
non-living things.
37
Although some evidence supports the living vs. non-living
contrast, another possibility is visual/perceptual properties vs.
functional properties.
Living things are defined by visual features.
Non-living things are defined by functions.
Neuroimaging experiments failed to provide clear support for
the living vs. nonliving view.
More fine-grained deficits have also been found. Some
patients have shown major deficits in naming animals but not
plants and vice versa.
38
Face Recognition
39
Are Faces Special?
Yes!
One evidence comes from individuals with
prosopagnosia.
Patients with prosopagnosia lose their ability to
recognize faces, though they can recognize other
objects.
fMRI evidence suggests that the fusiform face area
specializes in face recognition.
40
41
Patient HJA
“…when shopping he (John) would wait for me outside the shop. I
suggested he might recognise me by my height, spectacles, shape
of coat, shoes, and handbag but he never spots me and I always
have to call him. Once he tried to be very clever and began helping
to unload a trolley at the check out, much to the surprise of the lady
who must have been mystified by his apology.”
“I’m sorry. I thought you were my wife.”
Humphreys & Riddoch (1987)
To see but not to see: A case study of visual agnosia.
Featural vs. Configural Processing
Farah proposed two pattern-recognition systems.
Featural Processing: Recognition of simple parts
and the assembly of those parts into larger wholes.
Configural Processing: Recognition of larger
configurations. Responsible for analyzing the
spatial relations among features in a face.
Farah claims that the second system is damaged in
people with prosopagnosia.
Martha Farah
43
44
Another distinguishing principle of face recognition
is its strong dependence on orientation.
In one experiment, Yin (1969) showed subjects
pictures of houses or faces during a study phase.
During the test phase, subjects were shown upright
or inverted houses and faces and must make
recognition judgments on them.
The Face Inversion Effect
62
Recognition Probability
1.00
Houses
Faces
0.90
0.80
0.70
0.60
0.50
Upright
Inverted
63
More Evidence for the
importance of orientation in
recognizing faces...
The Thatcher Illusion (Thompson, 1980)
64
What does this tell us?
67
Our perception of faces are specifically built to detect
featural details (and to decode spatial relation among
the features) from an upright position. When a face is
inverted, that ability is lost. Our perception of faces
upside down and right side up is very different!
Contrast Reversal
68
Contrast reversed faces are also much more difficult
to recognized than contrast reversed objects.
Are Faces Special?
82
No!
A prosopagnosic farmer could not recognize his individual
cows.
A prosopagnosic bird watcher lost his ability to discriminate
warblers.
Another one lost his ability to tell cars apart. She had to read
the license plate of each car in a parking lot to find her own car.
Dog show judges show the inversion effect for dogs as they do
for humans.
Experience-Dependent Plasticity
Faces are typically recognized at the individual level but
other objects are not (within- vs between-class
discrimination).
We have more expertise with faces.
If we develop expertise at identifying objects in a particular
category, recognition of these objects can invoke the same
brain structure that is allegedly used only for face
recognition — the fusiform face area.
83
© 1999 Nature America Inc. • http://neurosci.nature.com
a r t icles
Gauthier and Tarr (1999)
ortant implications for intere fusiform ‘face area’ in visuon. First, our experiments
rsion effect can be obtained
cific areas, and that a similar
be produced for novel objects
ng. A previous study24 found
inversion effect in the right
s ‘face area’ for grayscale faces
h two-tone faces. The authors
ce area’ may be involved simbut not in the identification
tection is very difficult for
ces, but note that recognition
ompromised). In contrast, our
he face selective region of the
d in recognition at the indiraining at this level led to an
is account converges with
ociating the inversion effect Fig. 4. Activation maps for three novices and three experts in the passive-viewing tasks
ecognition at the individual with faces and greebles. A baseline of passive viewing of objects is used in both conditions,
ore consistent with the syn- and only the voxels showing more activation for faces or greebles than objects are shown.
osia, in which face recogni- Images are thresholded at an arbitrary value of t = 0.75. Note that we do not attribute any
sions in the ventral temporal statistical meaning to individual subjects’ t-values. The statistical significance of the effects is
patients typically have no dif- determined by their representation in the group sample. White squares, middle fusiform
but cannot recognize them at gyri ROI; arrows, lateral occipital gyrus foci for one expert (bottom right).
,27.
s from two independent tasks
n in face-specific areas can increase with
was only perceived as a face after presentation of a grayscale verjects. In a recent review article28, it was sugsion of the same stimuli, leading to increased activation in the
ding would negate any role for a face-specific
fusiform gyrus. In the other study32, a similar area was activat‘special’ face processing system. Specificaled as a result of a previous study episode, when subjects made
n upright greebles were compared to invertobject decisions about structurally coherent (but not structuralwas obtained in the right hemisphere ‘face
ly incoherent) objects. Note that these learning effects are
er in the right middle fusiform gyrus. Using
obtained following a single brief presentation of an image and
eebles versus objects, the effect was stronger
are specific to particular exemplars of a category (that is,
m gyri than anterior fusiform gyri ROIs with
increased activation occurring only for the actual pictures seen
n hemispheres. Left hemisphere middle and
but not for similar ones that are unfamiliar). In contrast, the
ri ‘face areas’ have been reported1,2,5–9, but
expertise effect that we describe requires hours of intensive training and is not specific to particular exemplars because the greehoroughly studied than the right middle
bles presented during the fMRI sessions were unfamiliar and
(The left areas typically show effects of a
different from those used during the training procedure. The
an the right areas, and the anterior fusiform
expertise effect may therefore be the signature of a different
n fMRI studies using a surface coil.)
mechanism, sufficiently long-term and general to produce catmost strongly to show an expertise effect in
egory preferences in the visual cortex.
orm gyrus. Comparison of mean Talaraich
Expertise is not the only factor that is thought to contribute to
middle fusiform gyri ‘face area’ with activathe specialization of the middle fusiform gyri for face processselected studies (Table 1) shows that our
ing. Faces are recognized most often at a very specific (or subori activation is most similar to previously
dinate) level (for example, Bob versus Jim), whereas objects are
’6, as well as an area activated by selective
typically recognized in a less specific manner, another important
a PET study. The fusiform gyrus has been
difference between face and object recognition. Even when
ing of pictures29 and words (left hemisphere
objects are selected from a single category33 or when subjects are
r, the latter study showed activation in the
gyrus for a semantic minus a non-semantic
required to discriminate between visually similar objects6, more
should note that it is very unlikely that our
activation is found for faces than objects in this area. Although
e to experts naming the greebles more than
this suggests that categorization level is not the only factor that
were scanned with unfamiliar greebles in
determines specialization of the middle fusiform gyri, it does not
greebles used during a single session were all
preclude some role for this factor. In two studies34,35, we found
ly, making family names non-diagnostic.
evidence that recognizing non-face familiar objects at a more
r simple learning effects in the fusiform
specific level (for example, pelican rather than bird) leads to activation in the face-selective part of the middle fusiform gyri. When
that falls about halfway between our middle
we compared the activation for passive viewing of faces minus
terior fusiform gyri ROIs (no coordinates
objects to that of specific non-face object recognition, the actiIn one study31, a two-tone image of a face
olume 2 no 6 • june 1999
571
Isabel Gauthier
Michael Tarr
85