63
Theories of Vision: a swift overview Learning-based Methods in Visio A. Efros, CMU, Spring 201 Most slides from Steve Palmer

Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

Embed Size (px)

Citation preview

Page 1: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

Theories of Vision: a swift overview

Learning-based Methods in VisionA. Efros, CMU, Spring 2012

Most slides from Steve Palmer

Page 2: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

What is a Theory?What is a Theory?

Theory An integrated set of related hypotheses about underlying mechanisms or principles that organize, explain, and predict facts about the phenomena of a given domain.

© Stephen E. Palmer, 2002

Page 3: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

What is a Theory?What is a Theory?

Three important functions of a theory:Integrate “old” factsPredict “new” factsLead to “understanding”

© Stephen E. Palmer, 2002

Three criteria for evaluating a theory:Logical consistencyEmpirical adequacyParsimony (“Occam’s razor”)

Page 4: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

Categorization of Theories of Vision

“Why do things looks as they do?”-- Kurt Koffka (1935)

Nativism vs. Empiricism• “Because we were born (evolved) to see them that way” vs.• “Because we have learned to see them that way”

Atomisn vs. Holism• “because of the way each pixel appears” vs.• “because of the way the entire scene appears”

Organism vs. Environment• “Because we are the way we are” vs.• “Because the world is the way it is”

Page 5: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

StructuralismStructuralism

© Stephen E. Palmer, 2002

THEORY NATIVISM vs.EMPIRICISM

ATOMISM vs. HOLISM

ORGANISM vs.ENVIRONMENT

PRINCIPAL ANALOGY METHOD

EcologicalOptics

Structuralism

Gestaltism

Constructivism

Page 6: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

StructuralismStructuralism

© Stephen E. Palmer, 2002

Structuralism: Perception results from the association of basic sensory atoms in memory via repeated, prior joint occurrences.

Derived from philosophy of British Empiricists (e.g., Locke, Berkeley, Hume, and Mills).

Proposed by Wilhelm Wundt, the father of modern Psychology.

Page 7: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

StructuralismStructuralism

© Stephen E. Palmer, 2002

Sensory Atoms

Retinal mosaic

Greennessat (x3,y3)

Yellownessat (x2,y2)

Rednessat (x1,y1)

Page 8: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

StructuralismStructuralism

© Stephen E. Palmer, 2002

Perceptual Complexes

Retinal mosaic

Page 9: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

StructuralismStructuralism

© Stephen E. Palmer, 2002

Perceptual Complexes

Retinal mosaic

Red appleat (x0,y0)

Page 10: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

StructuralismStructuralism

© Stephen E. Palmer, 2002

Perceptual Complexes

Retinal mosaic

Red appleat (x0,y0) Association

Page 11: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

StructuralismStructuralism

© Stephen E. Palmer, 2002

Chemical Analogy

Perceptions are made of basic sensory experiences just as molecules are made of basic atoms.

Page 12: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

StructuralismStructuralism

© Stephen E. Palmer, 2002

THEORY NATIVISM vs.EMPIRICISM

ATOMISM vs. HOLISM

ORGANISM vs.ENVIRONMENT

PRINCIPAL ANALOGY METHOD

EcologicalOptics

Structuralism

Gestaltism

Constructivism

Empiricism Atomism Organism ChemistryTrained

Introspection

Page 13: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

GestaltismGestaltism

© Stephen E. Palmer, 2002

Gestaltism: Perception results from the interactionbetween the intrinsic structure of the stimulusand the intrinsic structure of the brain.

MaxWertheimer

WolfgangKöhler

KurtKoffka

Page 14: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

GestaltismGestaltism

© Stephen E. Palmer, 2002

The Gestalt movement in perceptual theory was primarily a reaction against Structuralism:

Successful in arguing against Structuralism,but less successful in promoting its own theoretical agenda.

Rejected atomismRejected empiricismRejected associationism

Page 15: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

GestaltismGestaltism

Holism: The whole is different from the sum of its parts.

Principles of Gestalt Theory

Prägnanz: The percept will be as “good” as the prevailing conditions allow, I.e. simplest explanation

Nativism:Not a total rejection of learning, but rejection of Its primacy.

Page 16: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

GestaltismGestaltism

© Stephen E. Palmer, 2002

Prägnanz: the percept will be as “good” as the prevailing conditions allow

What is this?

?square &

circle?

?square &pacman?

?squigit &pacman?

etc….

Page 17: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

GestaltismGestaltism

© Stephen E. Palmer, 2002

Holism: The whole is different from the sum of its parts.

Emergent properties:Features of a configurationthat are not features ofits components, e.g.:

• length• orientation• curvature• closure• connectedness

Page 18: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

© Stephen E. Palmer, 2002

The problem of grouping: “I stand at the window and see a house, trees, sky.”Theoretically, I might say that there were 327 brightnesses and nuances of color. Do I have “327”? No. I have sky,house, and trees. (Wertheimer, 1923)

The Problem of OrganizationThe Problem of Organization

Page 19: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

© Stephen E. Palmer, 2002

Wertheimer’s “laws” of grouping

Rows

Perceptual GroupingPerceptual Grouping

Columns

Page 20: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

© Stephen E. Palmer, 2002

14.20

Proximity

Rows Columns

Perceptual GroupingPerceptual Grouping

Page 21: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

© Stephen E. Palmer, 2002

14.21

Color Similarity

Rows Columns

Perceptual GroupingPerceptual Grouping

Page 22: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

© Stephen E. Palmer, 2002

14.22

Size Similarity

Rows Columns

Perceptual GroupingPerceptual Grouping

Page 23: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

© Stephen E. Palmer, 2002

14.23

Orientation Similarity

Rows Columns

Perceptual GroupingPerceptual Grouping

Page 24: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

© Stephen E. Palmer, 2002

14.24

Similarity of texture

Rows

Perceptual GroupingPerceptual Grouping

Columns

Page 25: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

© Stephen E. Palmer, 2002

14.26

Common Fate

Columns

Perceptual GroupingPerceptual Grouping

Page 26: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

© Stephen E. Palmer, 2002

14.27

Good Continuation

ColumnsRows

Perceptual GroupingPerceptual Grouping

Page 27: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

© Stephen E. Palmer, 2002

14.28

Closure

ColumnsRows

Perceptual GroupingPerceptual Grouping

Page 28: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

© Stephen E. Palmer, 2002

14.29

Common Region

Rows

Perceptual GroupingPerceptual Grouping

Columns

Page 29: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

© Stephen E. Palmer, 2002

14.30

Element Connectedness

Rows

Perceptual GroupingPerceptual Grouping

Columns

Page 30: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

© Stephen E. Palmer, 2002

14.32

Past Experience

Perceptual GroupingPerceptual Grouping

Page 31: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

GestaltismGestaltism

© Stephen E. Palmer, 2002

Field Theoretic Analogy

Field ofa singlepositivecharge

Field ofa pair ofpositivecharges

Field of apositive &a negative

charge

Repulsion Attraction

Page 32: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

GestaltismGestaltism

© Stephen E. Palmer, 2002

THEORY NATIVISM vs.EMPIRICISM

ATOMISM vs. HOLISM

ORGANISM vs.ENVIRONMENT

PRINCIPAL ANALOGY METHOD

EcologicalOptics

Structuralism

Gestaltism

Constructivism

Nativism Holism Organism EM FieldsNaive

Introspection

Trained Introspection

Empiricism Atomism Organism Chemistry

Page 33: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

Ecological OpticsEcological Optics

© Stephen E. Palmer, 2002

Ecological Optics: Perception is the direct apprehension of the visible environment by extracting invariants in the dynamic ambient optic array and the affordances of objects.

James J. Gibson: Ask notwhat’s inside your head, butwhat your head’s inside of!(Ecological optics is a theory of stimulus structure.)

Page 34: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

Ecological OpticsEcological Optics

© Stephen E. Palmer, 2002

Direct Perception: Perception is not mediatedby processes like unconscious thinking

Perception of the 3-D world is NOT ambiguous for an actively exploring organism.

Page 35: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

Ecological OpticsEcological Optics

© Stephen E. Palmer, 2002

Texture Gradients are gradual changes in the size & shape of texture elements in an image when a plane recedes in depth.

Tilefloor

Wheatfield

Geometric surface

Page 36: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

Ecological OpticsEcological Optics

© Stephen E. Palmer, 2002

Motion Gradients are gradual changes in the speed and direction of elements in optic flow within the dynamic ambient optic array.

Bird flyingPlane landing

Page 37: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

Ecological OpticsEcological Optics

© Stephen E. Palmer, 2002

THEORY NATIVISM vs.EMPIRICISM

ATOMISM vs. HOLISM

ORGANISM vs.ENVIRONMENT

PRINCIPAL ANALOGY METHOD

EcologicalOptics

Structuralism

Gestaltism

Constructivism

Trained Introspection

Empiricism Atomism Organism Chemistry

PhysicalField Theory

NaiveIntrospection

Nativism Holism Organism

Nativism Holism EnvironmentMechanicalResonance

EcologicalAnalysis

Page 38: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

ConstructivismConstructivism

© Stephen E. Palmer, 2002

Constructivism: Perception is the result ofunconscious inferences about the scene mostlikely to have caused the retinal image or event.

Hermann von Helmholtzoriginated the idea ofunconscious inference andthe likelihood principle.

Page 39: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

ConstructivismConstructivism

© Stephen E. Palmer, 2002

Unconscious Inference: the process of recovering environmental information by logically combiningretinal information with heuristic assumptions.

Tilted room illusion:If you assume that thewalls and floor of the room are vertical andhorizontal, then youmust be tilted —and youfeel that way!

Page 40: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

ConstructivismConstructivism

© Stephen E. Palmer, 2002

Tilted room illusion

A. Actual Situation

B. Perceived Situation

Page 41: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

ConstructivismConstructivism

© Stephen E. Palmer, 2002

Prägnanz vs. LikelihoodWhat governs what we see: goodness or probability?

What is this?

? ? ?square &

circle?square &pacman?

squigit &pacman?

etc….

Page 42: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

ConstructivismConstructivism

© Stephen E. Palmer, 2002

THEORY NATIVISM vs.EMPIRICISM

ATOMISM vs. HOLISM

ORGANISM vs.ENVIRONMENT

PRINCIPAL ANALOGY METHOD

EcologicalOptics

Structuralism

Gestaltism

Constructivism

Trained Introspection

Empiricism Atomism Organism Chemistry

PhysicalField Theory

NaiveIntrospection

Nativism Holism Organism

(both) (both) (both)Logical

InferenceLikelihoodAnalysis

Mechanical Resonance

EcologicalAnalysis

Nativism Holism Environment

Page 43: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

© Stephen E. Palmer, 2002

So which theory is correct?

Probably none of them!

Or maybe all of them, to some degree!

It’s good to keep them in mind, whendesigning your algorithms!

“You can’t play 20-questionswith nature and win!”

(Allen Newell, 1973)

Page 44: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

Four Stages of Visual PerceptionFour Stages of Visual Perception

© Stephen E. Palmer, 2002

Image- BasedProcessing

Surface- BasedProcessing

Object-Based

Processing

Category- BasedProcessing

Light

Vision

Audition

STM

LTM

Motor

Sound

LightMove-ment

Odor (etc.)

Ceramiccup on a table

David Marr, 1982

Page 45: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

Four Stages of Visual PerceptionFour Stages of Visual Perception

© Stephen E. Palmer, 2002

The Retinal Image

An Image (blowup) Receptor Output

Page 46: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

Four Stages of Visual PerceptionFour Stages of Visual Perception

© Stephen E. Palmer, 2002

Image-basedRepresentation

Primal Sketch(Marr)

An Image

(Line Drawing)

RetinalImage

Image-based

processes

EdgesLinesBlobsetc.

Page 47: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

Four Stages of Visual PerceptionFour Stages of Visual Perception

© Stephen E. Palmer, 2002

Surface-basedRepresentation

Primal Sketch 2.5-D Sketch

Image-basedRepresentation

Surface-based

processes

StereoShadingMotion

etc.

Page 48: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

Koenderink’s trick

Page 49: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

Four Stages of Visual PerceptionFour Stages of Visual Perception

© Stephen E. Palmer, 2002

Object-basedRepresentation

Object-based

processes

GroupingParsing

Completionetc.

Surface-basedRepresentation

2.5-D Sketch Volumetric Sketch

Page 50: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

Geons(Biederman '87)

Page 51: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

Four Stages of Visual PerceptionFour Stages of Visual Perception

© Stephen E. Palmer, 2002

Category-basedRepresentation

Category-based

processes

Pattern-Recognition

Spatial-description

Object-basedRepresentation

Volumetric Sketch Basic-level Category

Category: cup

Color: light-gray

Size: 6”

Location: table

Page 52: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

We likely throw away a lot

Page 53: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

line drawings are universal

Page 54: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

However, things are not so simple…

● Problems with feed-forward model of processing…

Page 55: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

two-tone images

Page 56: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer
Page 57: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer
Page 58: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

hair (not shadow!)

inferred external contours

“attached shadow” contour

“cast shadow” contour

Page 59: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer
Page 60: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

Finding 3D structure in two-tone images requires distinguishing cast shadows, attached shadows, and areas of low reflectivity

The images do not contain this information a priori (at low level)

Cavanagh's argument

Page 61: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

Marr's model (circa 1980) Cavanagh’s Model (circa 1990s)

Feedforward vs. feedback models

stimulusstimulus

2D shape

memory

3D shape

2½D sketch

Object

3D model

feedback

basic recognition with 2D primitives

reconstruction of shape from image features

object recognition by matching 3D models

primal sketch

Page 62: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

A Classical View of Vision

Grouping /Segmentation

Figure/GroundOrganization

Object and Scene Recognition

pixels, features, edges, etc.Low-level

Mid-level

High-level

Page 63: Theories of Vision: a swift overview Learning-based Methods in Vision A. Efros, CMU, Spring 2012 Most slides from Steve Palmer

A Contemporary View of Vision

Figure/GroundOrganization

Grouping /Segmentation

Object and Scene Recognition

pixels, features, edges, etc.Low-level

Mid-level

High-level

But where we draw this line?