CSSE463: Image Recognition Day 17 Today: Bayesian classifiers - - PowerPoint PPT Presentation

CSSE463: Image Recognition Day 17

 Today: Bayesian classifiers  Tomorrow: Lightning talks and exam questions  Weds night: Lab 4 due  Thursday: Exam

 Bring your calculator.

 Questions?

Exam Thursday



Closed book, notes, computer



BUT you may bring handwritten notes (1-side of paper)



You may also want a calculator.



Similar to the written assignment in format:



Some questions from daily quizzes



Some extensions of quizzes



Some applications of image-processing algorithms



Some questions asking about process you followed in lab



Also a few Matlab-specific questions



Write 2-3 lines of code to…

Bayesian classifiers

 Use training data

 Assume that you know

probabilities of each feature.

 If 2 classes:

 Classes

and

 Say, circles vs. non-circles  A single feature, x  Both classes equally likely  Both types of errors equally

bad

 Where should we set the

threshold between classes? Here?

 Where in graph are 2 types of

errors?

x p(x) P(x|

1)

Non-circles P(x|

2)

Circles Detected as circles Q1-4

What if we have prior information?

 Bayesian probabilities say that if we only

expect 10% of the objects to be circles, that should affect our classification

Q5-8

Bayesian classifier in general

 Bayes rule:

 Verify with example

 For classifiers:

 x = feature(s) 

i = class

 P( |x) = posterior probability  P( ) = prior  P(x) = unconditional probability  Find best class by maximum a

posteriori (MAP) priniciple. Find class i that maximizes P(

i|x).  Denominator doesn’t affect

calculations

 Example:

 indoor/outdoor classification

) ( ) ( ) | ( ) | ( b p a p a b p b a p ) ( ) ( ) | ( ) | ( x p p x p x p

i i i Learned from examples (histogram) Learned from training set (or leave out if unknown) Fixed

Indoor vs. outdoor classification

 I can use low-level image info (color,

texture, etc)

 But there’s another source of really helpful

info!

Camera Metadata Distributions

p(FF|I) p(FF|O) 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 On Off

p(FF|I) p(FF|O)

1 2 3 4 5 7 9 17 p(SD|I) 0.05 0.1 0.15 0.2 0.25 0.3 0.35 p(SD|I) p(SD|O) 0.01 0.017 0.022 0.03 0.05 0.07 0.1 0.12 p(ET|I) p(ET|O) 0.2 0.4 0.6 p(ET|I) p(ET|O)

Exposure Time Flash Subject Distance

• 6
• 0.5

1 2.5 4 5.5 7 8.5 10 11.5 p(BV|I) 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 p(BV|I) p(BV|O)

Subject Distance Scene Brightness

Why we need Bayes Rule

Problem: We know conditional probabilities like P(flash was on | indoor) We want to find conditional probabilities like P(indoor | flash was on, exp time = 0.017, sd=8 ft, SVM output) Let = class of image, and x = all the evidence. More generally, we know P( x | ) from the training set (why?) But we want P( | x)

) ( ) ( ) | ( ) | ( x p p x p x p

i i i

Using Bayes Rule

P( |x) = P(x| )P( )/P(x)

The denominator is constant for an image, so

P( |x) = P(x| )P( )

Q9

Using Bayes Rule

P( |x) = P(x| )P( )/P(x)

The denominator is constant for an image, so

P( |x) = P(x| )P( ) We have two types of features, from image metadata (M) and from low-level features, like color (L) Conditional independence means P(x| ) = P(M| )P(L| ) P( |X) = P(M| ) P(L| ) P( )

From histograms From SVM Priors (initial bias)

Bayesian network

 Efficient way to encode conditional

probability distributions and calculate marginals

 Use for classification by having the

classification node at the root

 Examples

 Indoor-outdoor classification  Automatic image orientation detection

Indoor vs. outdoor classification

SVM KL Divergence Color Features SVM Texture Features EXIF header

Each edge in the graph has an associated matrix of conditional probabilities

Effects of Image Capture Context

Recall for a class C is fraction of C classified correctly

Orientation detection

 See IEEE TPAMI paper

 Hardcopy or posted

 Also uses single-feature Bayesian classifier

(answer to #1-4)

 Keys:

 4-class problem (North, South, East, West)  Priors really helped here!

 You should be able to understand the two

papers (both posted)