Lecture 5: Statistical Methods for Classification CAP 5415: Computer Vision Fall 2006.

Lecture 5: Statistical Methods for Classification

CAP 5415: Computer VisionFall 2006

Classifiers: The Swiss Army Tool of Vision

A HUGE number of vision problems can be reduced to: Is this a _____ or not?

The next two lectures will focus on making that decision

Classifiers that we will cover Bayesian classification Logistic regression Boosting Support Vector Machines Nearest-Neighbor Classifiers

Motivating Problem Which pixels in this image are “skin pixels”?

Useful for tracking, finding people, finding images with too much skin.

How could you find skin pixels?

Step 1: Get Data

Label every pixel as skin or not skin

Getting Probabilities

Now that I have a bunch of examples, I can create probability distributions. P([r,g,b]|skin) = Probability of an [r,g,b] tuple given

that the pixel is skin P([r,g,b]|~skin) = Probability of an [r,g,b] tuple given

that the pixel is not skin

(From Jones and Rehg)

Using Bayes Rule

x – the observation y – some underlying cause (skin/not skin)

Using Bayes Rule

PriorLikelihood

Normalizing Constant

Classification

In this case P[skin|x] = 1-P[~skin|x] So the classifier reduces to

P[skin|x] > 0.5? We can change this to

P[skin|x] > c And vary c

The effect of varying c This is called a Receiver Operating Curve (or

ROC

From Jones and Rehg

Application: Finding Adult Pictures

Let's say you needed to build a web filter for a library

Could look at a few simple measurements based on the skin model

Example of Misclassified Image

Example of Correctly Classified Image

ROC Curve

Generative versus Discriminative Models

The classifier that I have just described is known as a generative model

Once you know all of the probabilities, you can generate new samples of the data

May be too much work You could also optimize a function to just

discriminate skin and not skin

Discriminative Classification using Logistic Regression

Imagine we had two measurements and we plotted each sample on a 2D chart

Discriminative Classification using Logistic Regression

Imagine we had two measurements and we plotted each sample on a 2D chart

To separate the two groups, we'll project each point onto a line

Some points will be projected to positive values and some will be projected to negative values

Discriminative Classification using Logistic Regression

This line defines a separating line Each point is classified based on where it falls

on the line

How do we get the line?

Common Option: Logistic Regression Logistic Function:

The logistic function Notice that g(x) goes from 0 to 1 We can use this to estimate the probability something being an

x or an o We need to find a function that will have large positive values

for x's And large negative values for o's

Fitting the Line

Remember, we want a line. For the diagram below, x = +1, o = -1 y = label of point (-1 or +1)

Fitting the line

The logistic function gives us an estimate of the probability of an example being either +1 or -1

We can fit the line by maximizing the conditional probability of the correct labeling of the training set

Also called features

Fitting the Line

We have multiple samples that we assume are independent, so the probability of the whole training set is

Fitting the line

It is usually easier to optimize the log conditional probability

Optimizing

Lots of options Easiest option: Gradient ascent

: The Learning Rate parameter, many ways to choose this

Choosing

My (current) personal favorite method Choose some value for Update w, Compute new probability If the new probability does not rise, divide

by 2 Otherwise multiply it by 1.1 (or something

similar) Called “Bold-Driver” heuristic

Faster Option

Computing the gradient requires summing over every training example

Could be slow for a large training set Speed-up: Stochastic Gradient Ascent Instead of computing the gradient over the

whole training set, instead choose one point at random.

Do update based on that one point

Limitations

Remember, we are only separating the two classes with a line

Separate this data with a line:

This is a fundamental problem, most things can't be separated by a line

Overcoming these limitations

Two options: Train on a more complicated function

Quadratic Cubic

Make a new set of features:

Advantages

We achieve non-linear classification by doing linear classification on non-linear transformations of the features

Only have to rewrite feature generation code Learning code stays the same

Nearest Neighbor Classifier

Is the “?” an x or an o?

?

Nearest Neighbor Classifier

Is the “?” an x or an o?

?

Nearest Neighbor Classifier

Is the “?” an x or an o?

?

Basic idea

For your new example, find the k nearest neighbors in the training set

Each neighbor casts a vote Label with the most votes wins Disadvantages:

Have to find the nearest neighbors Can be slow for a large training set Good approximate methods available (LSH - Indyk)