You are on page 1of 33

Computational Perception

15-485/785

Perceptual Constancy 1
Perceptual Constancy
How do we perceive stimuli as being the
same?
• Sensory patterns can be radically
different.
• Shading is one example.
• What are others?

example from Dan Kersten

CP08: Perceptual Constancy 1 2 Michael S. Lewicki ◇ Carnegie Mellon


A schematic of visual computation

object and scene representation

perceptual constancy

visual structure
representation

Sensory Coding

Perceptual constancy is inherently an inferential process that depends on both


lower-level and higher-level information.

CP08: Perceptual Constancy 1 3 Michael S. Lewicki ◇ Carnegie Mellon


Lightness constancy

• Is the perception really constant?


• Is the retinal image accessible?
• Drawing well is hard, but recognizing
poor drawings is easy

image from Ted Adelson

CP08: Perceptual Constancy 1 4 Michael S. Lewicki ◇ Carnegie Mellon


Perceptual constancy demos

Purves lab demos:


www.purveslab.net/seeforyourself/

CP08: Perceptual Constancy 1 5 Michael S. Lewicki ◇ Carnegie Mellon


Brightness contrast illusions

images from www.purveslab.net


CP08: Perceptual Constancy 1 6 Michael S. Lewicki ◇ Carnegie Mellon
Brightness contrast with color

images from www.purveslab.net


CP08: Perceptual Constancy 1 7 Michael S. Lewicki ◇ Carnegie Mellon
Effect of texture on contrast perception: Chubb Illusion

images from www.purveslab.net


CP08: Perceptual Constancy 1 8 Michael S. Lewicki ◇ Carnegie Mellon
More subtle types of invariance

Attached Shadows Braje, Legge, and Kersten, 1999:


• Test object recognition with or
without cast shadows
• Cast shadows have no effect on
object recognition response time or
accuracy.
• Same is true for color.

Cast Shadows

CP08: Perceptual Constancy 1 9 Michael S. Lewicki ◇ Carnegie Mellon


Size constancy

• People are very good at estimating


actual size, esp. when multiple cues
are available
• If cues are removed, they choose
closer to retinal size.
• With depth cues, it is very difficult to
judge retinal size.

from Rock, 1984

from Palmer, 1999

CP08: Perceptual Constancy 1 10 Michael S. Lewicki ◇ Carnegie Mellon


How big is the cylinder?

Estimate the height of


the third vs the first. from Palmer, 1999

CP08: Perceptual Constancy 1 11 Michael S. Lewicki ◇ Carnegie Mellon


How big is the cylinder?

The third is 1/3 the


height of the first. from Palmer, 1999

CP08: Perceptual Constancy 1 12 Michael S. Lewicki ◇ Carnegie Mellon


How big is the cylinder?

The third is 1/3 the


height of the first. from Palmer, 1999

CP08: Perceptual Constancy 1 13 Michael S. Lewicki ◇ Carnegie Mellon


Direct estimation of size: texture gradients

from Goldstein, 2001 from Rock, 1984

Objects covering same number texture gradient units have same size.
CP08: Perceptual Constancy 1 14 Michael S. Lewicki ◇ Carnegie Mellon
Relative estimation of size

from Palmer, 1999

Relative size is invariant with distance.


CP08: Perceptual Constancy 1 15 Michael S. Lewicki ◇ Carnegie Mellon
Shape constancy

• Shapes are also perceived in their 3D


form over many viewing positions.
• We have little access to the retinal
shape.
• Common shapes are perceived
accurately largely independent of
projective transformation.

from Palmer, 1999

CP08: Perceptual Constancy 1 16 Michael S. Lewicki ◇ Carnegie Mellon


Even complex shapes are perceived accurately

from Palmer, 1999

CP08: Perceptual Constancy 1 17 Michael S. Lewicki ◇ Carnegie Mellon


Perceptual shape constancy is limited

• For unfamiliar shapes, subjects to not


generalize to novel views.
• Subjects’ memory was much worse
when tested on a novel view
compared to the same view.
• Subjects also show poor shape
constancy on unfamiliar solid 3D
shapes.

from Palmer, 1999

CP08: Perceptual Constancy 1 18 Michael S. Lewicki ◇ Carnegie Mellon


Shape constancy and familiar objects

Subjects do show shape constancy for familiar objects.

from Palmer, 1999

CP08: Perceptual Constancy 1 19 Michael S. Lewicki ◇ Carnegie Mellon


Measuring the effect of view on recognition

View is related to
ratings of how well an
image depicts object.

from Palmer, 1999

CP08: Perceptual Constancy 1 20 Michael S. Lewicki ◇ Carnegie Mellon


Canonical views of common objects

Subjects are quickest to


name object from
canonical view.

CP08: Perceptual Constancy 1 21 Michael S. Lewicki ◇ Carnegie Mellon


Reference frames

Wiser (1981)
• With poor intrinsic axis (A), subjects
recognition memory was much
better when figure’s orientation was
the same.
• With good intrinsic axis (B), subjects
showed now difference when
orientation was changed.

from Palmer, 1999

CP08: Perceptual Constancy 1 22 Michael S. Lewicki ◇ Carnegie Mellon


Reference frames are not completely general

Same or Different?

Now?

from Hancock, 2000

CP08: Perceptual Constancy 1 23 Michael S. Lewicki ◇ Carnegie Mellon


Context affects perception

How do these faces differ?

Now can you tell?


from Hancock, 2000
CP08: Perceptual Constancy 1 24 Michael S. Lewicki ◇ Carnegie Mellon
Generalized
a: Learn a basis fortemplates
a specific class of images, e.g. faces (Turk and Pentland,
91).
Idea: Learn a basis for a specific class of images, e.g. faces (Turk and Pentland, 1991).
• Normalize
Normalize object
object set by set by position,
position, size,size,
etc.etc.Usually
Usually by
byhand.
hand.
Model •theModel
objects using using
the objects linearlinear
superposition:
superposition:

!
I(x, y) = aiφi(x, y)
i

• Derive optimal basis functions for object set


Derive optimal {φ1, ..., φM } for object set.
under a• Gaussian
Under a Gaussian
model model,
this isthis is PCA,
PCA, which
which is isfeasible
feasible for
for large
largepatterns.
patterns.

CP08: Perceptual Constancy 1 25 Michael S. Lewicki ◇ Carnegie Mellon


Eigenfaces

from Hancock, 2000

CP08: Perceptual Constancy 1 26 Michael S. Lewicki ◇ Carnegie Mellon


Eigen face distortions

from Hancock, 2000

CP08: Perceptual Constancy 1 27 Michael S. Lewicki ◇ Carnegie Mellon


Face image principal components

The face image components move


back and forth along their axes. from Hancock, 2000

CP08: Perceptual Constancy 1 28 Michael S. Lewicki ◇ Carnegie Mellon


Face shape principal components

The face shape components move


back and forth along their axes. from Hancock, 2000

CP08: Perceptual Constancy 1 29 Michael S. Lewicki ◇ Carnegie Mellon


Making a caricature: step 1

Identify positions of facial features. from Bruce and Young, 1998

CP08: Perceptual Constancy 1 30 Michael S. Lewicki ◇ Carnegie Mellon


Making a caricature: step 2

Distort positions from average. from Bruce and Young, 1998

CP08: Perceptual Constancy 1 31 Michael S. Lewicki ◇ Carnegie Mellon


Making a caricature: step 3

Morph original face according to caricature distortion.


from Bruce and Young, 1998

CP08: Perceptual Constancy 1 32 Michael S. Lewicki ◇ Carnegie Mellon


Computer-generated caricatures of famous actors

Caricatures were more recognizable (i.e. faster RT) than originals.


from Bruce and Young, 1998

CP08: Perceptual Constancy 1 33 Michael S. Lewicki ◇ Carnegie Mellon

You might also like