LCS 11: Cognitive Science 1. Gestalt principles 2. Recognition by - - PowerPoint PPT Presentation

Pomona College

LCS 11: Cognitive Science

Object recognition

Jesse A. Harris April 28, 2013

Jesse A. Harris: LCS 11: Cognitive Science, Object recognition 1

Agenda

֠ Object perception

• 1. Gestalt principles
• 2. Recognition by components theory
• 3. Hoffman’s transversal intersection

֠ Presentation signup sheet ֠ Reading for Wednesday: Hoffman, 1998: ch 5 ֠ Radiolab podcast (optional, but awesome)

Jesse A. Harris: LCS 11: Cognitive Science, Object recognition 2

What is this?

I challenge you to tell me:

• 1. What is its

purpose?

• 2. Where does it come

from?

• 3. What is the name

for this type of

• bject?

http://marksvegplot.blogspot.com/2010/12/identify-mystery-object.html

Jesse A. Harris: LCS 11: Cognitive Science, Object recognition 3

What is this?

I challenge you to tell me:

• 1. What is its

purpose?

• 2. Where does it come

from?

• 3. What is the name

for this type of

• bject?

http://marksvegplot.blogspot.com/2010/12/identify-mystery-object.html

Jesse A. Harris: LCS 11: Cognitive Science, Object recognition 4

Guesses

Hazel

Salt? Snuff? Spittoon? Holder for some rare spice from the Spice Islands? Finger bowl? Hard to tell how big it is.

Sherie Buck

I think it is old, not something you can go and buy now? Spice crusher of some sort?

Sherie Buck

Does it hold insects?

Kim Swins

Is it a Bento Box? They are japanese lunch boxes.

Jesse A. Harris: LCS 11: Cognitive Science, Object recognition 5

What is this?

Mystery object revealed:

• 1. It is used for

DRINKING millet beer.

• 2. It comes from

NEPAL.

• 3. It is called a

TONGBA.

http://marksvegplot.blogspot.com/2010/12/identify-mystery-object.html

Jesse A. Harris: LCS 11: Cognitive Science, Object recognition 6

What is this?

What really basic questions were NOT asked?

• 1. Where is the
• bject?
• 2. What are its parts?

http://marksvegplot.blogspot.com/2010/12/identify-mystery-object.html

Jesse A. Harris: LCS 11: Cognitive Science, Object recognition 7

Object recognition

What makes object recognition complex?

The problem of perceptual segregation:

• 1. Massive overlap in environment – where does one object

stop and another one begin?

• 2. Objects of the same type vary in characteristics – on

what basis do we categorize them as the same?

• 3. Objects perceived as the same, despite differences in

viewing – how are we able to account for viewpoint?

Jesse A. Harris: LCS 11: Cognitive Science, Object recognition 8

Massive overlap in environment – where does one object stop and another one begin? Objects of the same type vary in characteristics – on what basis do we categorize them as the same? Objects perceived as the same, despite differences in viewing – how are we able to account for viewpoint?

Gestalt psychology

Fundamental principle of Prägnanz

We tend to perceive and experience the world as regular,

• rderly, symmetric, and simple.

Of several geometrically possible organizations [the]

• ne [that] will remain will actually occur which

possess the best, simplest, and most stable shape. (Koffka, 1935)

Jesse A. Harris: LCS 11: Cognitive Science, Object recognition 12

Gestalt psychology

Fundamental principle of Prägnanz

We tend to perceive and experience the world as regular,

• rderly, symmetric, and simple.

Gestalt laws

Reconstruct and reify objects according to several principles, including: Proximity Group nearby things as one. Similarity Group similar elements together. Continuation Avoid disrupting straight or smoothly curving lines. Closure Fill in missing parts if need be.

Jesse A. Harris: LCS 11: Cognitive Science, Object recognition 13

Proximity Similarity Similarity

Continuation Continuation Closure Closure

Closure Figure/ground Models of object perception

◮ Early selectivity ◮ Three stages of representation

to derive shape information

◮ Transform primitive features of

the objective stimulus to a figure in 3D space

◮ Object-centered representation

transformed to viewer-centric representation

Published 1982

Jesse A. Harris: LCS 11: Cognitive Science, Object recognition 23

Deriving shape information

• 1. Primal sketch

Makes available basic features such as blobs, edges, groups, stereopsis, and boundaries

• 2. 2.5 sketch

Inferences about orientation, depth, basic relationships between features and objects inferred

• 3. 3D model representation

Components of object represented in “volumetric” and shape primitives.

Published 1982

Jesse A. Harris: LCS 11: Cognitive Science, Object recognition 24

3D representation

Jesse A. Harris: LCS 11: Cognitive Science, Object recognition 25

Biedermann’s 1987 component geons

36 fundamental geon types were originally proposed.

Biedermann’s 1987 Recognition-by-components

Edge extraction

Early stage sensitive to surface characteristics, like luminosity

Deletion of non-accidental properties

• 1. Extract invariants, like curvature, parallelism, symmetry.
• 2. Construct geons

Parsing concavity

Attend to concave joints, as in Hoffman

Components + match

Determine components and match to object representations.

Jesse A. Harris: LCS 11: Cognitive Science, Object recognition 27

Viewpoint invariance

Major prediction of theory is that we consult geons in a viewpoint invariant representation.

Familiar objects (Biedermann & Gerhardstein 1993)

◮ Orientation didn’t affect priming effect of familiar

• bjects.

Unfamiliar objects (Tarr and colleagues)

◮ Novel objects show viewpoint dependence, but

performance improves over time.

◮ Performance dependent on object type: speed increases

with expertise, but accuracy still viewpoint sensitive

Jesse A. Harris: LCS 11: Cognitive Science, Object recognition 28

Greebles

http://www.tarrlab.org

Transversal intersection

Concave crease

Sharp edges pointing into object (vs. convex creases, which point out) Can tell where arbitrary shapes interpenetrate by concave crease at surface.

Rule of concave creases.

Divide shapes into parts along concave creases.

Jesse A. Harris: LCS 11: Cognitive Science, Object recognition 30

Transversal intersection

Minima rule.

Divide shapes into parts at negative minima, along lines of curvature of the principal curvature (where curvature grows least).

Gradient relationships

The more evidence you have for transversal intersections, the easier they are to spot. Evidence includes salience of cusp boundary in accordance with sharpness of cusp angle (Rules 17, 18)

Jesse A. Harris: LCS 11: Cognitive Science, Object recognition 31

Boundaries

Which side of a curve corresponds to the object (figure)? Check both sides of curve and choose the figure leading to a recognizable object.

Salient boundaries

Choose figure and ground so that the figure has the more salient boundaries.

Salient parts

Choose figure and ground so that the figure has the more salient parts.

Jesse A. Harris: LCS 11: Cognitive Science, Object recognition 32

Sign up for presentation times

Jesse A. Harris: LCS 11: Cognitive Science, Object recognition 33