Lecture #23

Preview:

DESCRIPTION

Lecture #23. Object recognition 4/23/13. MBEX final survey. If you took survey at the beginning of the semester and then complete it now, you will receive 5 pts towards homework Please take it by April 30 th Thanks for helping biology education research. The end is in sight. - PowerPoint PPT Presentation

Citation preview

Lecture #23

Object recognition4/23/13

MBEX final survey

• If you took survey at the beginning of the semester and then complete it now, you will receive 5 pts towards homework

• Please take it by April 30th

• Thanks for helping biology education research

The end is in sight

4/25 One of main wiki pages done5/2 Question for exam5/9 Wiki due last class: Intro page and

3 main pages done5/16 Final Thursday 1:30-3:30

Wiki references

• With text, refer to your references with either (author, year) or [1]

• Make sure to write out the reference to include author, year, title, journalThis helps the reader see what they are Links are great but they are in addition

Writing

• Don’t obfuscate your writing with jargon that is recondite and abstruse

• Write in your own voice and not that of a medical encyclopedia

• Make it so your parents could understand it, while still relying on the primary literature!!

Writing

• Most important thing that you do as a scientist / business person / medical professional

• Do it oftenStrive to improve it

Once you’ve written something

• Getting your first version done is just the first step

• Read it again

• Edit, edit, edit

• Simplify, simplify, simplify

How do people write?

What do you do when you get stuck?

What to do when you get stuck

• Deal with subunitsDoes each paragraph hang togetherTopic sentence

• Outline paper based on topic sentencesTests organization

• Have many people read itDifferent learning styles and perspectives

I. Object perception• What enables us to

recognize two objects as being the same thing?

Low level vision

• Retina detects dotsCenter - surround wiring

• Visual cortex detects lines, edges and blobs of color

Gestalt

• Visual perception is more than just detection of dots of color and lines

• Whole is greater than sum of partsGestalt = “whole”

• There are rules by which visual scenes are interpreted as combinations of “perceptual groups”

Middle vision

• To build an object you need to:Find edgesGroup similar areas

“White”Decide what goes together to make a whole

• Organize elements into groups by grouping rules

Walls + windows + door + columns + roof = White House

Higher level vision

• Determine what an object is

• Match middle level views with memory of what object is

• Independent of Viewing angleWhether seen before or not

Grouping rules

• How do you think you organize what you see into individual elements or groups?

Finding edges• Edge detection can be

difficult and produce partial lines

Rule #1 : Good continuation• This figure

is most likely the result of

these lines and not these

Rule #2 : Occlusion• If an edge stops, it is

likely being occluded• So assume it must

actually extend

Kanizsa figure• You see the object

even though it is only hinted at from objects present

Rule # 3 - Texture segmentation

• Group parts of image that have similarly sized texture

• This is based on two principles:SimilarityProximity

Similarity and proximity

• Image bits, that are similar and close to each other, are grouped together

• Can be based onColorShapeSizeOrientation

Similarity and proximity

• Can fool visual system if overlapped characteristicsSame forms and colors occur in both groups

Rule #4: Parallelism and symmetry

• Lines that are parallel are seen as a group

• Lines that are mirror images are seen as group

Additional rules• Group by proximity

Additional rules• Group by proximity

• Group by common regions

• Group by connectedness

Sinauer web site

Camouflage

• An attempt to thwart the visual system’s ability to discriminate an object from its background – disrupt similarity/ proximity grouping

Camouflage

• Color and texture matching - flounder

Camouflage to the extreme

• Flounder can even match a checkerboard!

Zebra uses stripes to break up outline

May work even better for lion under dim light

Stripes make it hard to tell where one zebra stops and the next starts

Insects use shape, texture and color

Katydids

Walking stick

Video by Roger Hanlon

BBC: Vision and photography

Perceiving the alphabet

• Each letter is composed of shapes - lines and curves

Perceptual committees• Low level vision provides

shape and orientation info• Feature demons are

specialists which detect certain aspects

• Cognitive demons combine information to recognize each letter

• Decision demon makes decision out of pandemonium

Oliver Selfridge 1959

Similarities with brain

• Areas of visual cortex that detect features

• Feed info to other areas where cognition happens

• Detection happens in parallel

http://www.sinauer.com/wolfe/chap4/pandemoniumF.htm

Similarities with brain• Probably are neurons

which interconnect and make these comparisons

• Have to train neurons on the alphabet characters for your language

It is possible to wear out the demons

• After images result if you have two demons which oppose each other

• If wear one out, the other wins

It is possible to wear out the demons

• After images result if you have two demons which oppose each other

• If one gets tired, the other wins

Tilt after effect

Tilt after effect

Tilt after effect

Additional grouping rules:Figure and ground

• Like grouping, there are rules which help the brain decide which object is in front of which

• Ideas?

Rules

• Surroundedness - if one object surrounds other, the surrounded object is figure

• Size - smaller object is figure• Symmetry - symmetric object is figure• Parallelism - regions which are parallel are

part of a figure

Which are figures?

Which are figures and which are holes?

Now which are figures vs holes?

Ambiguous figures

• Made by alternate interpretations which equally likely

• Necker cube

Necker cube

• Either of these views is equally likely

• Brain may flip back and forth between them

Stairs

• Another ambiguous figure which can flip

What do you see?

• There are two equally likely interpretation

• Brain switches between them

Do you see a young or an old woman?

Interpret image as the most probable solution

This explanation never happens

How middle vision turns parts into an object

• Bring together that which should be brought together (similar, proximate..)

• Split asunder that which should be splitEdges and figure/ground

• Use what you know (physics)• Avoid accidents • Seek consensus and avoid ambiguity

Object recognition• Perhaps object gets

compared to a template?A

Object recognition• Perhaps object gets

compared to a template?

• But need a lot of templates

A A A aa A aA A a

Recognition by components• Biederman - All

objects composed of cylindrical shapes = geons (geometric ions)

• Like a shape alphabet• Objects contain

particular shapes in particular orientations

Object recognition• Capital A is two

slanted lines joined by one horizontal line part way up

• Small a is circle joined to a line

A A A aa A aA A a

Enables viewpoint invariance• Since made up of set

of components, doesn’t matter how you view the object because you will recognize the components

Not everything is viewpoint invariant

• In general you can recognize text

• However, it is a lot harder to understandLetter recognition depends on orientation

Even if it is turned upside down

Faces are also a bit of a special case Jard to discriminate upside down

Wolfe ch 4.31

Faces are also hard to discriminate upside down.. But easy right side up

Wolfe ch 4.31

Object recognition

Wolfe ch 4.32

Some of middle processing occurs in visual cortex - grouping - texture segmentation Other tasks handled by extrastriate cortex

Higher processing Where - parietal lobe

Location of objects What - temporal lobe

Object recognition

Inferotemporal lobe

• Visual cortex - responds to simple shapes in particular orientation in small field of view (3-4 deg)

• Inferotemporal (IT) Responds to complex shapes (hands, faces)Responds to over half field of view

Hierarchical processing

• Simple to complex• Visual cortex to inferotemporal cortex

Add more and more levels features to identify objects

• Propose there are complex cells which respond to particular objects - grandmother cellLearn their receptive fieldEveryone’s grandmother is different!

IT cortex cells in macaque

These IT cells respond well to faces and so would be “grandmother” cells

Certain aspects of faces are more important than others

Strokes which cause lesions in the IT lead to agnosia Inability to recognize objects

Bill Choisser’s web site

Prosopagnosia - face blind

Visual perception is all about context

• The eye receives shapes with lightness and color

• The brain can interpret identical objects in very different ways

Lightness illusion

Lightness illusion

These backgrounds do not make circles seem as different