MIT6870_ORSU_lecture3: Levels of categorization and multiclass object recognition

transcript

8/3/2019 MIT6870_ORSU_lecture3: Levels of categorization and multiclass object recognition

http://slidepdf.com/reader/full/mit6870orsulecture3-levels-of-categorization-and-multiclass-object-recognition 1/97

Lecture 3Levels of categorization

and multiclass object recognition

6.870 Object Recognition and Scene Understandinghttp://people.csail.mit.edu/torralba/courses/6.870/6.870.recognition.htm

Class business

Start thinking about the class project

Choose a partner (if you want to do it alone, thatis fine too).

Brainstorm about ideas soon! It is also fine if it ispart of your own research.

Consult with me before starting

Best presentation award (for creditonly), after popular vote: a copy of thebook ³Vision Science´ (or another bookif the winner already has this one)

Class business

I have to travel

I am very flexible, but, how flexible are

Wednesday

Sharing parts for intraclass transfer learning

Presenter: Sharat Chikkerur

Evaluator: Hueihan Jhuang

Some goals for this lecture

What are categories?

How many categories are there?

Some theories about the structure of object categories

Multiclass object recognition and transfer

learning in computer vision

An example of categorical perception

Continuous perception: graded response

5 0 1 00 1 50 20 0 25 0

Many perceptual phenomena are a mixture of the two: categorical at an everyday

level of magnification, but continuous at a more microscopic level. It can also

depend on cultural aspects, expertise, task, attention, «

5 0 10 0 15 0 2 0 0 2 5 0

Categorical perception: ³sharp´ boundaries

Another example

Continuous perception: graded response

20-24 25-29 30-34 35-39 40-44 45-49 50-54

Categorical perception: ³sharp´ boundaries

happinessfear

Emotions have categorical boundaries

Identification Task

e n t i f i c a t i o n

Happiness

Why do we care about categories?

Perception of function: We can perceive the 3D

shape, texture, material properties, withoutknowing about objects. But, the concept of

category encapsulates also information about

what can we do with those objects.

³We therefore include the perception of function as a proper ±indeed, crucial- subject

for vision science´, from Vision Science, chapter 9, Palmer .

Why do we care about categories?

When we recognize an object we can makepredictions about its behavior in the future,

beyond of what is immediately perceived.

The perception of function

Direct perception (affordances): GibsonFlat surface

Horizontal

Knee-high

Sittable

One caveat of this comparison: deciding that

something is a chair might require access to more

features than the ones needed to decide that we can

sit on something« (it is a different level of

categorization)

Chair Chair

Chair?

Flat surface

Horizontal

Knee-high

Sittable

uponChair

Mediated perception (Categorization)

Direct perception

Some aspects of an object function can be

perceived directly

Functional form: Some forms clearly

indicate to a function (³sittable-upon´,

container, cutting device, «)

Sittable-upon Sittable-upon

Sittable-upon

It does not seem easy

to sit-upon this«

Direct perception

Some aspects of an object function can be

perceived directly

Observer relativity: Function is observer

dependent

From http://lastchancerescueflint.org

Limitations of Direct Perception

The functions are the same at some level of description: we can put things

inside in both and somebody will come later to empty them. However, weare not expected to put inside the same kinds of things«

Objects of similar structure might have very different functions

Not all functions seem to be available from direct visual information only.

Limitations of Direct Perception

Propulsion system

Strong protective surface

Something that looks like a door

Sure, I can travel to space onthis object

Visual appearance might be a very weak cue to function

Indirect perception of function by

categorization

Well« this requires object recognition (for

more details, see entire course)

So, what do we use direct or indirect?

³It seems exceedingly unlikely (though

logically possible) that we categorize

everything in our visual fields´, Palmer.

Hypothesis: we categorize the objects that

are relevant for a specific task that we

have at hand, but we only extract

affordances from the others.

How many categories?

Slide by Aude Oliva

³Muchas´

How many object categories are there?

Biederman 1987

How many categories?

Probably this question is not even specific

enough to have an answer

Which level of categorization

is the right one?Car is an object composed of:

a few doors, four wheels (not all visible at all times), a roof,

front lights, windshield

If you are thinking in buying a car, you might want to be a bit more specific about

your categorization.

psychological

cognition

content

belief

entity

object

artifact

structure

substance

Livingthing

organism

person

leader scientist

phenomenon

instrumentality

location

region

animal

chordate

vertebrate

chemist

stream

Categorical hierarchies

From Wordnet

Categories can be organized in hierarchies (tree structures are commonly used)

This is a mapping of the Wordnet tree into the 2D plane

Organizing ³things´ into categories

(1) Feature-based (2) prototype

Definition:disassembling aconcept into a set of

featural components

Each feature is anessential element of the category: ³ for athing to be an X, itmust have thatfeature. Otherwise it isnot an X´

Categories are formedon the basis of characteristics features,

which describe thetypical model of thecategory

Characteristics featuresare commonly present inexemplars of concepts,but they are not alwayspresent.

Prototypical Theory

According to theprototype view, an objectwill be classified as aninstance of a category if itis sufficiently similar to

the prototype.

Similarity ~ the number of features shared betweenan object and the

prototype (however,some features should beweighted more heavily asbeing more central to theprototype than are other

features).

Prototype or Sum of Exemplars ?

Prototype Model Exemplars Model

Category judgments are made

by comparing a new exemplar

to the prototype.

Category judgments are made

by comparing a new exemplar

to all the old exemplars of a category

or to the exemplar that is the most

appropriate

Levels of Categorization

The idea of prototypes and typicality led to the study of levels of categorization.

Rosch et al: ³albeit concepts exist at many different

levels of a hierarchy, one level is fundamental: basic

level.´

Basic level: the best compromise between grouping

together similar objects, and distinguish among

objects from the same category.

Willingham (247)

B ASIC-LEVELCATEGORIES

SUPERORDIN ATE LEVELCATEGORIES

SUBORDIN ATE LEVELCATEGORIES

Rosch¶s Levels of CategorizationDefinition of Basic Level:

Similar shape: Basic level categories are the highest-level category for which their members have similar shapes.

Similar motor interactions: « for which people interact with its members

using similar motor sequences.

Common attributes: « there are a significant number of attributes in

common between pairs of members.

Sub Basic Superordinate

similarity

Similarity declines only slightly

going from subordinate tobasic level, and then drops

dramatically.

Rosch et al (1976) found that when people

are shown pictures of objects, they identify

objects at a basic level more quickly than

they identified objects at higher or lower

levels.

Objects appear to be recognized first at

their basic level, and only afterwards theyare classified in terms of higher or lowers

level categories

Typicality effects

Typicality: how good or common an

item is a member of a given category.

The typical exemplar is like arepresentation of the average (or central tendency)

But, the representation of a categoryvary with experience, so does the³typical´ exemplar

Entry-level categories(Jolicoeur, Gluck, Kosslyn 1984)

Typical member of a basic-level categoryare categorized at the expected level

Atypical members tend to be classified at

a subordinate level.

A bird An ostrich

We do not need to recognize the exact

category

A new class can borrow information from

similar categories

So, where is computer vision?

Well«

The first problem« the data

Datasets

Language 106 samples

Datasets

Character recognition

(MNIST)

104 samples

Datasets

Character recognition

(MNIST)

104 samples

Visual object recognition 10

samples

Early datasetsLena: a dataset in one picture

The Columbia Object Image Library (COIL-100) dataset consists of color images of 100objects where the images of the objects were taken at pose intervals of 5 degrees, i.e., 72

poses per object.

http://www.cs.columbia.edu/CA

S. A. Nene, S. K. Nayar and H. Murase,

Technical Report CUCS-006-96, February 1996.

Collecting datasets

(towards 106-7 examples)

ESP game (CMU)Luis Von Ahn and Laura Dabbish 2004

LabelMe (MIT)Russell, Torralba, Freeman, 2005

StreetScenes (C

L-MIT)Bileschi, Poggio, 2006

WhatWhere (Caltech)Perona et al, 2007

P ASCAL challenge

2006, 2007

Lotus Hill InstituteSong-Chun Zhu et al 2007

Cortina

Labeling with games

L. von Ahn, L. Dabbish, 2004; L. von Ahn, R. Liu and M. Blum, 2006

The PASCAL Visual Object

Classes Challenge 2007

M. Everingham, Luc van Gool , C. Williams, J. Winn, A. Zisserman 2007

The twenty object classes that have been selected are:

Person: person

Animal: bird, cat, cow, dog, horse, sheep

Vehicle: aeroplane, bicycle, boat, bus, car, motorbike, train

Indoor: bottle, chair, dining table, potted plant, sofa, tv/monitor

Caltech 101 & 256

Griffin, Holub, Perona, 2007

Fei-Fei, Fergus, Perona, 2004

How to evaluate datasets?

How many labeled examples? How many classes? Segments or bounding

boxes? How many instances per image? How small are the targets? Variability

across instances of the same classes (viewpoint, style, illumination). How

different are the images?

How representative of the visual world is? What happens if you nail it?

Labelme.csail.mit.edu B. Russell, A. Torralba, K. Murphy, W.T. Freeman. IJCV 2008

Tool went online July 1st, 2005

290,000 object annotations

Polygon quality

« things do not always look good«

Polygon quality

Testing

Most common labels:

adksdsa

woiieiie

Online HooligansDo not try this at home

Tags quality

M tl b t lb

Matlab toolbox

LMquer y (database, 'object.name', 'car,building,road,tree')

How many categories are in LabelMe?

How many levels there are?

10% of the objects

account for 90% of

the data

~Zipf¶s law

Caltech 101

Tiny images

LabelMe

We need transfer learning

Slid i +

Slide spain+perona

We have better low and mid-level vision

Better learning algorithms

Lot¶s of computational power

And lot¶s of data

We are running out of excuses

M lti l bj t d t ti

Multiclass object detectionthe not so early days

M lti l bj t d t ti

Multiclass object detectionthe not so early days

Schneiderman-Kanade multiclass object detection

Using a set of independent binary classifiers was a common strategy: Viola-Jones extension for dealing with rotations

- two cascades for each view

(a) One detector for each class

There is nothing wrong with this approach if you have access to

lots of training data and you do not care about efficiency.

Some symptoms of one vs all

Some symptoms of one-vs-all

multiclass approaches

Some of these parts cannot be used for anything else than this object.

What is the best representation to detect a traffic sign?

Very regular object: template matching will do the job

Parts derived from

training a binaryclassifier.

detection ratewith 0 false alarms

multiclass approachesPart-based object representation (looking for meaningful parts):

A. Agarwal and D. Roth

These studies try to recover parts that are meaningful. But is this the

right thing to do? The derived parts may be too specific, and they are

not likely to be useful in a general system.

M. Weber, M. Welling and P. Perona

multiclass approachesComputational cost grows linearly with Nclasses * Nviews * Nstyles «

Shared features

Is learning the object class 1000 easier than learning the first?

Can we transfer knowledge from one

object to another?

re the shared properties interesting bythemselves?

M ltit k l i

Multitask learning

R. Caruana. Multitask Learning. ML 1997

³MTL improves generalization by leveraging the domain-specific information

contained in the training signals of related tasks. It does this by training tasks in

parallel while using a shared representation´.

Sejnowski & Rosenberg 1986; Hinton 1986; Le Cun et al. 1989; Suddarth &

Kergosien 1990; Pratt et al. 1991; Sharkey & Sharkey 1992; «

M ltit k l i

Multitask learning

horizontal location of doorknob

single or double door

horizontal location of doorway center

width of doorway

horizontal location of left door jamb

horizontal location of right door jamb

width of left door jamb

width of right door jamb

horizontal location of left edge of door

horizontal location of right edge of door

Primar y task: detect door knobs

Tasks used:

R. Caruana. Multitask Learning. ML 1997

Sharing invariancesS. Thrun. Is Learning the n-th Thing Any Easier Than Learning The First?

NIPS 1996

Knowledge is transferred between tasks via a learned model of the

invariances of the domain: object recognition is invariant to rotation,

translation, scaling, lighting, « These invariances are common to all

object recognition tasks.

Toy world

Without sharing

With sharing

Convolutional Neural Network

Translation invariance is already built into the network

The output neurons share all the intermediate levels

Le Cun et al, 98

Sharing transformations

Miller, E., Matsakis, N., and Viola, P. (2000). Learning from one example

through shared densities on transforms. In IEEE Computer Vision andPattern Recognition.

Transformations are sharedand can be learnt from other tasks.

M d l f bj t iti

Models of object recognitionI. Biederman, ³Recognition-by-components: A theory of human image

understanding,´ Psychological Review , 1987.

M. Riesenhuber and T. Poggio, ³Hierarchical models of object recognition in

cortex,´ Nature Neuroscience 1999.

T. Serre, L. Wolf and T. Poggio. ³Object recognition with features inspired

by visual cortex´. CVPR 2005

Sharing in constellation models

Sharing in constellation models(next Wednesday)

Pictorial StructuresF ischler & Elschlager, IEEE Trans. Comp. 1973

ConstellationModel

F ergus, Perona, & Zisserman, CVPR 2003

SVM DetectorsHeisele, Poggio, et. al., NIPS 2001

Model-Guided Segmentation

Mori, Ren, Efros, & Malik, CVPR 2004

R bl P t

Reusable Parts

Goal: Look for a vocabulary of edges that reduces the number of

features.

Krempp, Geman, & Amit ³Sequential Learning of Reusable Parts for Object

Detection´. TR 2002

N u m b e r o f f e a

t u r e s

Number of classes

Examples of reused parts

Additive models and boosting

Additive models and boosting(more details on Wednesday)

Torralba, Murphy, Freeman. CVPR 2004. P AMI 2007

Screen detector

Car detector

Face detector

Binary classifiers that share features:

Screen detector

Car detector

Face detector

Independent binary classifiers:

S ifi f t

Specific feature

Non-shared feature: this feature

is too specific to faces.

pedestrian

Traffic light

Background class

Sh d f t

Shared feature

shared feature

50 training samples/class

29 object classes

2000 entries in the dictionary

Results averaged on 20 runs

Error bars = 80% interval

Shared features

Class-specific features

enera za on as a unc on o o ec

similarities

12 viewpoints12 unrelated object classes

Number of training samples per class Number of training samples per class

A r e a u n d e r R O C

A r e a u n d e r R O C K = 2.1 K = 4.8

Sharing patches

Bart and Ullman, 2004For a new class, use only features similar to features that where good for other

classes:

Proposed Dog

features

Transfer Learning for Image Classification with Sparse

Prototype Representations

md d d

,22,21,2

,12,11,1

Coefficients for

for feature 2

Coefficients for

classifier 2

A. Quattoni, M. Collins, T. Darrell, CVPR 2008

§ §§! !

D y xk

W C y x f l D

k 1 1),(

|)(|max)),((||

Hi hi l T i M d l

Hierarchical Topic Models

Topic models typically use a³bag of words´ approx.: ± Learning topics allows transfer

of information within a corpus of related documents

± Mixing proportions capture thedistinctive features of particular documents

Latent Dirichlet Allocation (LDA)Blei, Ng, & Jordan, JMLR 2003

Pr(topic | doc)

Pr(word | topic)

Hi hi l T i M d l

Pr(topic | doc)

Pr(word | topic)

Pr(x=word | z=topic) Pr(z=topic | doc)7topic

Pr(x=word | doc) =

Pr(topic | doc)

Pr(word | topic)

³bag of features´ models:

Object Recognition ( Sivic et. al., ICCV 2005)

Scene Recognition ( F ei-F ei et. al., CVPR 2005)

Learning Shared Parts

Objects are often locally similar in appearance

Discover parts shared across categories

± How many total parts should we share?

± How many parts should each category use?

HDP Object Model

We learn the

number of parts.

Each object

uses a different

number of parts.

The model

assumes aknown number

of object

categories.

Parts are distributions

over appearances andlocations

HDP Object Model

There is no context, so the model is happy in creating

impossible part combinations.

Sharing Parts: 16 Categories

Caltech 101 Dataset (Li & Perona)

Horses (Borenstein & Ullman)

at & dog faces (Vidal-Naquet & Ullman)

Bikes from Graz-02 (Opelt & Pinz)

Google«

Visualization of Shared Parts

Pr(appearance | part)

Pr(position | part)

Pr(appearance | part)

Pr(position | part)

Detection Results

6 Training Images per Categor y Detection vs. Training Set Size

Recognition Task

Recognition Results

6 Training Images per Categor y Recognition vs. Training Set Size

Recognition performance decreases. By sharing features, the classes look more similar.

Some more references

Baxter 1996

Caruana 1997

Schapire, Singer, 2000

Thrun, Pratt 1997

Krempp, Geman, Amit, 2002

E.L.Miller, Matsakis, Viola, 2000

Mahamud, Hebert, Lafferty, 2001

Fink et al. 2003, 2004

LeCun, Huang, Bottou, 2004

Holub, Welling, Perona, 2005

Wednesday

Sharing parts for intraclass transfer learning

Presenter: SharatC

hikkerur

Evaluator: Hueihan Jhuang

MIT6870_ORSU_lecture3: Levels of categorization and multiclass object recognition

Documents