High level vision

High level vision

• Models of object recognition• Top down influences• Navigation/Movement

Last Time. . ..

We spent a lot of time focusing on lines; how you get them, why you would want them, and so on.

We need to move from lines to objects. How do you recognize an object from an organization of lines?How does perception connect to memory?

Models of object recognition

• Template• Feature• “New wave” of feature models (3D

features)

Template model

Template--problems

Problems: Size

Orientation

Need too many templates

Feature model--pandemonium

Feature modelsGood: visual input does seem to be decomposed into features

Good: Physiological evidence about simple features from Hubel & Wiesel

Problems: orientation missing features

natural objects

Natural objects: What are the features of a dog?

• Nose • Ear • Front Leg• Tail• Back Leg

Principle from Gestalt Psych

A

B

C

D

Good continuation

Good continuation can be used to find the parts of objects

“new wave” of feature models

These models use three-dimensional features.

Biederman’s geon model

You usually only need to see the edges of a geonGeons have properties that are invariant to rotations

Geons

Simple objects

Experiment

Which can be better identifiedat a very briefexposure?

Problems with Geons

• Do geons really represent all shapes?• How are relationships among geons coded?

An alternative: Local viewpoints

Note that you can identify objects from many different orientations. Templates & Feature

models couldn’t account for this—geons can.BUT how good are you at doing this really?

AlternativeSome researchers have suggested that object representations are NOT viewpoint independent. Rather, we store views of objects the way we see them.

HUH? Isn’t that the template theory?

The difference is that you do some “fixing” (size, rotation) of the image to fit the template

Tarr’s local view experimentsO° 45° 90° O° 45° 90°

Tarr Results

Problems with local view

• Chicken & egg: how do you know how to rotate the image before you can identify it?

• What is stored is clearly not literal pictures…but what is it?

• How is what you see and what is in memory matched?

Chicken & egg problem. . .

Bottom up processing refers to beginning with relatively raw, unprocessed sensory information, and building towards more conceptual representations. Top-down processing refers to conceptual knowledge influencing the processing or interpretation of lower-level perceptual processes

NOTE--we’ve been acting as though all processing were

bottom up.

Example: ambiguous figures

Example

Example:

It appears that in top down processing you use conceptual information to generate hypotheses about what the stimulus might be, then test these hypotheses

More formal work

Watch for the object appearing

The Parsing Paradox

If perceptual organization is a matter of mapping sensations onto structural schema, which happens first: interpreting the whole or interpreting the parts? How can someone recognize a face until he has first recognized the eyes, nose, mouth and ears? Then again, how can you recognize the parts until you know that they are part of a face?

--Stephen Palmer

Question: do you process the top-down information atthe same time as the bottom-up info?

You’ll see a circle--try to identify the objectthat appears in the circle.

The result: people are better at identifying the objectwhen the scene makes sense, compared to when it’s jumbled

How is this possible?

Word superiority effectIt is easier to recognize a letter in a

word than in isolation.

Identify the letter that will appear in the circle

TAKE

WOLP

Word superiority effect

Faster and more reliable in identifying a letter when it’s part of a

word than a non-word.

Isn’t there the same chicken-egg problem? Don’t you need to know the letters to identify the word? So then how is the word helping to identify the letters (which you already know?)

Model of word identification

Navigation vs. Object identification

There is increasing evidence that spatial information that helps us get around is independent of the information that helps us identify objects.

Mishkin & Ungerlieder

Mishkin & Ungerlieger

Mishkin & Ungerliedger

Object

Spatial