Part 3: classifier based methodsAntonio Torralba

Overview of section• A short story of discriminative methods

• Object detection with classifiers

• Boosting– Gentle boosting– Weak detectors– Object model– Object detection

• Multiclass object detection

• Context based object recognition

Classifier based methodsObject detection and recognition is formulated as a classification problem.

Bag of image patches

… and a decision is taken at each window about if it contains a target object or not.

Decision boundary

Computer screen

Background

In some feature space

Where are the screens?

The image is partitioned into a set of overlapping windows

(The lousy painter)

Discriminative vs. generative

0 10 20 30 40 50 60 700

0.05

0.1

x = data

• Generative model

0 10 20 30 40 50 60 700

0.5

1

x = data

• Discriminative model

0 10 20 30 40 50 60 70 80

-1

1

x = data

• Classification function

(The artist)

• The representation and matching of pictorial structures Fischler, Elschlager (1973). • Face recognition using eigenfaces M. Turk and A. Pentland (1991). • Human Face Detection in Visual Scenes - Rowley, Baluja, Kanade (1995) • Graded Learning for Object Detection - Fleuret, Geman (1999) • Robust Real-time Object Detection - Viola, Jones (2001)• Feature Reduction and Hierarchy of Classifiers for Fast Object Detection in Video Images - Heisele, Serre, Mukherjee, Poggio (2001)•….

• The representation and matching of pictorial structures Fischler, Elschlager (1973). • Face recognition using eigenfaces M. Turk and A. Pentland (1991). • Human Face Detection in Visual Scenes - Rowley, Baluja, Kanade (1995) • Graded Learning for Object Detection - Fleuret, Geman (1999) • Robust Real-time Object Detection - Viola, Jones (2001)• Feature Reduction and Hierarchy of Classifiers for Fast Object Detection in Video Images - Heisele, Serre, Mukherjee, Poggio (2001)•….

Face detection

• The representation and matching of pictorial structures Fischler, Elschlager (1973). • Face recognition using eigenfaces M. Turk and A. Pentland (1991). • Human Face Detection in Visual Scenes - Rowley, Baluja, Kanade (1995) • Graded Learning for Object Detection - Fleuret, Geman (1999) • Robust Real-time Object Detection - Viola, Jones (2001)• Feature Reduction and Hierarchy of Classifiers for Fast Object Detection in Video Images - Heisele, Serre, Mukherjee, Poggio (2001)•….

Face detection

• Formulation: binary classification

Formulation

+1-1

x1 x2 x3 xN

…

… xN+1 xN+2 xN+M

-1 -1 ? ? ?

…

Training data: each image patch is labeledas containing the object or background

Test data

Features x =

Labels y =

Where belongs to some family of functions

• Classification function

• Minimize misclassification error(Not that simple: we need some guarantees that there will be generalization)

Discriminative methods

106 examples

Nearest neighbor

Shakhnarovich, Viola, Darrell 2003Berg, Berg, Malik 2005…

Neural networks

LeCun, Bottou, Bengio, Haffner 1998Rowley, Baluja, Kanade 1998…

Support Vector Machines and Kernels Conditional Random Fields

McCallum, Freitag, Pereira 2000Kumar, Hebert 2003…

Guyon, VapnikHeisele, Serre, Poggio, 2001…

A simple object detector with Boosting

Download

• Toolbox for manipulating dataset

• Code and dataset

Matlab code

• Gentle boosting

• Object detector using a part based model

Dataset with cars and computer monitors

http://people.csail.mit.edu/torralba/iccv2005/

• A simple algorithm for learning robust classifiers– Freund & Shapire, 1995– Friedman, Hastie, Tibshhirani, 1998

• Provides efficient algorithm for sparse visual feature selection– Tieu & Viola, 2000– Viola & Jones, 2003

• Easy to implement, not requires external optimization tools.

Why boosting?

Boosting

Boosting fits the additive model

by minimizing the exponential loss

Training samples

The exponential loss is a differentiable upper bound to the misclassification error.

Boosting

Sequential procedure. At each step we add

For more details: Friedman, Hastie, Tibshirani. “Additive Logistic Regression: a Statistical View of Boosting” (1998)

to minimize the residual loss

inputDesired outputParametersweak classifier

Weak classifiers • The input is a set of weighted training

samples (x,y,w)

• Regression stumps: simple but commonly used in object detection.

Four parameters:

b=Ew(y [x> ])

a=Ew(y [x< ])x

fm(x)

Flavors of boosting

• AdaBoost (Freund and Shapire, 1995)

• Real AdaBoost (Friedman et al, 1998)

• LogitBoost (Friedman et al, 1998)

• Gentle AdaBoost (Friedman et al, 1998)

• BrownBoosting (Freund, 2000)

• FloatBoost (Li et al, 2002)

• …

From images to features:A myriad of weak detectors

We will now define a family of visual features that can be used as weak classifiers (“weak detectors”)

Takes image as input and the output is binary response.The output is a weak detector.

A myriad of weak detectors

• Yuille, Snow, Nitzbert, 1998• Amit, Geman 1998• Papageorgiou, Poggio, 2000• Heisele, Serre, Poggio, 2001• Agarwal, Awan, Roth, 2004• Schneiderman, Kanade 2004 • Carmichael, Hebert 2004• …

Weak detectors

Textures of textures Tieu and Viola, CVPR 2000

Every combination of three filters generates a different feature

This gives thousands of features. Boosting selects a sparse subset, so computations on test time are very efficient. Boosting also avoids overfitting to some extend.

Haar wavelets

Haar filters and integral imageViola and Jones, ICCV 2001

The average intensity in the block is computed with four sums independently of the block size.

Haar waveletsPapageorgiou & Poggio (2000)

Polynomial SVM

Edges and chamfer distance

Gavrila, Philomin, ICCV 1999

Edge fragments

Weak detector = k edge fragments and threshold. Chamfer distance uses 8 orientation planes

Opelt, Pinz, Zisserman, ECCV 2006

Histograms of oriented gradients

• Dalal & Trigs, 2006

• Shape context

Belongie, Malik, Puzicha, NIPS 2000• SIFT, D. Lowe, ICCV 1999

Weak detectors

Part based: similar to part-based generative models. We create weak detectors by using parts and voting for the object center location

Car model Screen model

These features are used for the detector on the course web site.

Weak detectors

First we collect a set of part templates from a set of training objects.

Vidal-Naquet, Ullman, Nature Neuroscience 2003

…

Weak detectors

We now define a family of “weak detectors” as:

= =

Better than chance

*

Weak detectors

We can do a better job using filtered images

Still a weak detectorbut better than before

* * ===

TrainingFirst we evaluate all the N features on all the training images.

Then, we sample the feature outputs on the object center and at random locations in the background:

Representation and object model

…

4 10

Selected features for the screen detector

1 2 3

…

100

Lousy painter

Representation and object modelSelected features for the car detector

1 2 3 4 10 100

… …

Detection• Invariance: search strategy

• Part based

fi, Pigi

Here, invariance in translation and scale is achieved by the search strategy: the classifier is evaluated at all locations (by translating the image) and at all scales (by scaling the image in small steps).

The search cost can be reduced using a cascade.

Example: screen detectionFeature output

Example: screen detectionFeature output

Thresholded output

Weak ‘detector’Produces many false alarms.

Example: screen detectionFeature output

Thresholded output

Strong classifier at iteration 1

Example: screen detectionFeature output

Thresholded output

Strongclassifier

Second weak ‘detector’Produces a different set of false alarms.

Example: screen detection

+

Feature output

Thresholded output

Strongclassifier

Strong classifier at iteration 2

Example: screen detection

+

…

Feature output

Thresholded output

Strongclassifier

Strong classifier at iteration 10

Example: screen detection

+

…

Feature output

Thresholded output

Strongclassifier

Adding features

Finalclassification

Strong classifier at iteration 200

We want the complexity of the 3 features classifier with the performance of the 100 features classifier:

Cascade of classifiersFleuret and Geman 2001, Viola and Jones 2001

Recall

Precision

0% 100%

100%

3 features

30 features

100 features

Select a threshold with high recall for each stage.

We increase precision using the cascade

Some goals for object recognition

• Able to detect and recognize many object classes

• Computationally efficient

• Able to deal with data starving situations:– Some training samples might be harder to

collect than others– We want on-line learning to be fast

Multiclass object detection

Multiclass object detection

Number of classes

Amount of labeled data

Number of classes

Complexity

Shared features• Is learning the object class 1000 easier

than learning the first?

• Can we transfer knowledge from one object to another?

• Are the shared properties interesting by themselves?

…

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

Primary 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

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

through shared densities on transforms. In IEEE Computer Vision and Pattern Recognition.

Transformations are sharedand can be learnt from other tasks.

Models of object recognitionI. Biederman, “Recognition-by-components: A theory of human image understanding,” Psychological Review, 1987.

M. Riesenhuber and T. Poggio, “Hierarchical models of object recognition in cortex,” Nature Neuroscience 1999.

T. Serre, L. Wolf and T. Poggio. “Object recognition with features inspired by visual cortex”. CVPR 2005

Sharing in constellation models

Pictorial StructuresFischler & Elschlager, IEEE Trans. Comp. 1973

Constellation ModelBurl, Liung,Perona, 1996; Weber, Welling, Perona, 2000

Fergus, Perona, & Zisserman, CVPR 2003

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

Model-Guided SegmentationMori, Ren, Efros, & Malik, CVPR 2004

E-Step

Random initialization

Variational EMVariational EM

prior knowledge of p()

new estimate of p(|train)

M-Step

new ’s

(Attias, Hinton, Beal, etc.)Slide from Fei Fei Li

Fei-Fei, Fergus, & Perona, ICCV 2003

Reusable Parts

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

Krempp, Geman, & Amit “Sequential Learning of Reusable Parts for Object Detection”. TR 2002

Num

ber

of f

eatu

res

Number of classes

Examples of reused parts

Sharing patches• Bart and Ullman, 2004

For a new class, use only features similar to features that where good for other classes:

Proposed Dog features

Multiclass boosting

• Adaboost.MH (Shapire & Singer, 2000)

• Error correcting output codes (Dietterich & Bakiri, 1995; …)

• Lk-TreeBoost (Friedman, 2001)

• ...

Shared features

Screen detector

Car detector

Face detector

• Independent binary classifiers:

Torralba, Murphy, Freeman. CVPR 2004. PAMI 2007

Screen detector

Car detector

Face detector

• Binary classifiers that share features:

50 training samples/class29 object classes2000 entries in the dictionary

Results averaged on 20 runsError bars = 80% interval

Krempp, Geman, & Amit, 2002

Torralba, Murphy, Freeman. CVPR 2004

Shared features

Class-specific features

Generalization as a function of object similarities

12 viewpoints12 unrelated object classes

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

Are

a un

der

RO

C

Are

a un

der

RO

C K = 2.1 K = 4.8

Torralba, Murphy, Freeman. CVPR 2004. PAMI 2007

Opelt, Pinz, Zisserman, CVPR 2006

Efficiency Generalization

Some references on multiclass

• Caruana 1997• Schapire, Singer, 2000• Thrun, Pratt 1997• Krempp, Geman, Amit, 2002• E.L.Miller, Matsakis, Viola, 2000• Mahamud, Hebert, Lafferty, 2001• Fink 2004• LeCun, Huang, Bottou, 2004• Holub, Welling, Perona, 2005• …

Context based methodsAntonio Torralba

Why is this hard?

2

1

What are the hidden objects?

What are the hidden objects?

Chance ~ 1/30000

Context-based object recognition

• Cognitive psychology– Palmer 1975 – Biederman 1981– …

• Computer vision– Noton and Stark (1971)– Hanson and Riseman (1978)– Barrow & Tenenbaum (1978) – Ohta, kanade, Skai (1978)– Haralick (1983)– Strat and Fischler (1991)– Bobick and Pinhanez (1995)– Campbell et al (1997)

Global and local representations

building

car

sidewalk

Urban street scene

Global and local representations

Image index: Summary statistics, configuration of textures

Urban street scene

features

histogram

building

car

sidewalk

Urban street scene

Global scene representations

Spatial structure is important in order to provide context for object localization

Sivic, Russell, Freeman, Zisserman, ICCV 2005Fei-Fei and Perona, CVPR 2005Bosch, Zisserman, Munoz, ECCV 2006

Bag of words Spatially organized textures

Non localized textons

S. Lazebnik, et al, CVPR 2006Walker, Malik. Vision Research 2004

…

M. Gorkani, R. Picard, ICPR 1994A. Oliva, A. Torralba, IJCV 2001

…

Contextual object relationshipsCarbonetto, de Freitas & Barnard (2004) Kumar, Hebert (2005)

Torralba Murphy Freeman (2004)

Fink & Perona (2003)E. Sudderth et al (2005)

Context

• Murphy, Torralba & Freeman (NIPS 03)Use global context to predict presence and location of objects

Op1,c1

vp1,c1

OpN,c1

vpN,c1. . .

Op1,c2

vp1,c2

OpN,c2

vpN,c2. . .

Class 1 Class 2

E1 E2

S

c2maxVc1

maxV

X1X2vg

Keyboards

3d Scene Context

Image World

[Hoiem, Efros, Hebert ICCV 2005]

3d Scene Context

Image Support Vertical Sky

V-Left V-Center V-Right V-Porous V-Solid

[Hoiem, Efros, Hebert ICCV 2005]

ObjectSurface?

Support?

Object-Object Relationships

Enforce spatial consistency between labels using MRF

Carbonetto, de Freitas & Barnard (04)

Object-Object Relationships

Use latent variables to induce long distance correlations between labels in a Conditional Random Field (CRF)

He, Zemel & Carreira-Perpinan (04)

• Fink & Perona (NIPS 03)Use output of boosting from other objects at previous

iterations as input into boosting for this iteration

Object-Object Relationships

CRFsObject-Object Relationships

[Torralba Murphy Freeman 2004] [Kumar Hebert 2005]

Hierarchical Sharing and Context

• Scenes share objects

• Objects share parts

• Parts share features

E. Sudderth, A. Torralba, W. T. Freeman, and A. Wilsky.

Some references on context

• Strat & Fischler (PAMI 91) • Torralba & Sinha (ICCV 01), • Torralba (IJCV 03) • Fink & Perona (NIPS 03)• Murphy, Torralba & Freeman (NIPS 03)• Kumar and M. Hebert (NIPS 04)• Carbonetto, Freitas & Barnard (ECCV 04)• He, Zemel & Carreira-Perpinan (CVPR 04)• Sudderth, Torralba, Freeman, Wilsky (ICCV 05)• Hoiem, Efros, Hebert (ICCV 05)• …

With a mixture of generative and discriminative approaches

A car out of context …

Banksy

Integrated models for scene and object recognition

Summary

Many techniques are used for training discriminative models that I have not mention here:

• Conditional random fields

• Kernels for object recognition

• Learning object similarities

• …