11/3 Notes

Face Detection

APplications

• Image retrival
• Law enforcement
• Biometrics
• Survallence

Given an image, Determine

• if faces are present
• the number of faces present
• the location and extent of each face in image
• the pose of each face - yaw, pitch, roll
• the identity of each face

There is a part of the brain dedicated to detecting faces

Methods

Appearance Based Methods - use classifiers that operate directly on the normalized pixel intensity of the imgage without extracting any facial Features

Rule based methods - employ knowledge of the components of the fase (rules human knows, not machine)

Feature based methods - uses grouping of edges, skin shape and color, template matching to detect faces

Texture based methods - uses textural feature to represent and detect facial patterns

Only talking about Appearance based Methods

Challenges

Detecting faces as they move, or if faces enter frame - tracking

Determining pose of the face

Viola Jones Technique

An image representation based on itegral image that allows for very fasst feature extraction

A simple and efficiaent calssifier based on adaboost

See slide?

Features used:

• 2 rectangle feature
• 3 rectangle feature
• 4 rectangle feature

f(x) = sum(W_i_j) - sum(B_i_j)

W = white regions

B = Black/grey regions

See slides?

Adaboost

The weak learning algorithm is designed to select the single rectangle feature which best separates the positive and negative examples

A weak classifier h(x,f,p,theta) consists of a feature f, a threshold theta and a polarity p, indicating the direction of the inequality:

h(x,f,p,theta) = 1 if pf(x) < p*theta, 0 otherwise

SEE SLIDES FOR STEPS OF ADABOOST