26
AdaBoost: an Introduction Carlo Ciliberto SINA – Genova – 28/11/2011

AdaBoost: an Introduction - liralab.it · AdaBoost Training: For : • Find the current best weak learner: WeakTraining learners h h h h h h h h h h h h h h h h h S S S S S S S S

Embed Size (px)

Citation preview

AdaBoost: an IntroductionCarlo Ciliberto

SINA – Genova – 28/11/2011

AdaBoost

AdaBoostInputs:

• Training set

S

S

S SS

S

S

SSS

S

S

SS

S

S

Training set

AdaBoostInputs:

• Training set

• Weak Learners

Weak learners

h h h hhhh

h h

hhhhh

h

hh

S

S

S SS

S

S

SSS

S

S

SS

S

S

Training set

AdaBoostInputs:

• Training set

• Weak Learners

Initialization:• Uniform initial distribution

Weak learners

h h h hhhh

h h

hhhhh

h

hh

S

S

S SS

S

S

SSS

S

S

SS

S

S

Weight distributionTraining set

AdaBoostTraining:

For : 

Weak learners

h h h hhhh

h h

hhhhh

h

hh

S

S

S SS

S

S

SSS

S

S

SS

S

S

Training set

AdaBoostTraining:

For : 

• Find the current best weak learner: 

Weak learners

h h h hhhh

h h

hhhhh

h

hh

S

S

S SS

S

S

SSS

S

S

SS

S

S

Training set

Evaluate performances 

AdaBoostTraining:

For : 

• Find the current best weak learner: 

Weak learners

h h h hhhh

h h

hhhhh

h

hh

S

S

S SS

S

S

SSS

S

S

SS

S

S

Training set

ht

Best weak learner 

AdaBoostTraining:

For : 

• Find the current best weak learner: 

Weak learners

h h h hhhh

h h

hhhhh

h

hh

S

S

S SS

S

S

SSS

S

S

SS

S

S

Training set

ht

Best weak learner 

• Set ,

AdaBoostTraining:

For : 

• Find the current best weak learner: 

Weak learners

h h h hhhh

h h

hhhhh

h

hh

S

S

S SS

S

S

SSS

S

S

SS

S

S

Training set

ht

weight 

• Set ,

AdaBoostTraining:

For : 

• Find the current best weak learner: 

• Set ,

Weak learners

h h h hhhh

h h

hhhhh

h

hh

S

S

S SS

S

S

SSS

S

S

SS

S

S

Training set

• Update the weights 

ht

UpdateDistribution

AdaBoostTraining:

For : 

• Find the current best weak learner: 

• Set ,

Weak learners

h h h hhhh

h h

hhhhh

h

hh

S

S

S SS

S

S

SSS

S

S

SS

S

S

Training set

• Update the weights 

ht

UpdateDistribution

AdaBoostTest:

ht

AdaBoostTest:

hthth1 hT

AdaBoostTest:

ht

S

For any novel example  : 

hth1 hT

AdaBoostTest:

hth1 hT

S

For any novel example  : 

• The strong classifier returns

AdaBoostTest:

hth1 hT

S

For any novel example  : 

• The strong classifier returns

Face Detection

Viola and Jones 2001

Face Detection

Viola and Jones 2001

• Training set

• Weak Learners

Integral Image

Viola and Jones 2001

Integral Image

Viola and Jones 2001

Integral Image

Viola and Jones 2001

Integral Image

Viola and Jones 2001

Integral Image

Viola and Jones 2001

Integral Image

Viola and Jones 2001

Integral Image

Viola and Jones 2001