A Robust Real Time Face

Detection
 Outline
 AdaBoost – Learning Algorithm
 Face Detection in real life
 Using AdaBoost for Face Detection
 Improvements
 Demonstration
 AdaBoost

A short Introduction to Boosting (Freund & Schapire, 1999)

Logistic Regression, AdaBoost and Bregman Distances
(Collins, Schapire, Singer, 2002)
 Boosting
 The Horse-Racing Gambler Problem
– Rules of thumb for a set of races
– How should we choose the set of races in order
to get the best rules of thumb?
– How should the rules be combined into a single
highly accurate prediction rule?
 Boosting !
 AdaBoost - the idea
 AdaBoost agglomerates many weak
classifiers into one strong classifier.

IQ  Initialize sample weights

 For each cycle:
– Find a classifier that performs
well on the weighted sample
– Increase weights of
misclassified examples
 Return a weighted list of
classifiers
Shoe
size
 AdaBoost - algorithm
Given ( x1 , y1 ),.., ( xm , ym ) where xi  X , yi  Y  {1,1}
Initialize D1 (i )  1 / m
For y  1..T
Select the best weak classifier using distributi on Dt
Get weak hypothesis ht : X  {1,1} with error  t  Pri ~ Dt [ht ( xi )  yi ]
1 1 t
Choose  t  ln( )
2 t


e t if hi ( xi )  yi
D (i ) 
Update Dt 1 (i )  t   where Z t is a normalizat ion factor
Zt  et if hi ( xi )  y

Output the final hypothesis :

T
H ( x)  sign (  t ht ( x))
t 1
 AdaBoost – training error
 Freund and Schapire (1997) proved that:
T
   t2 1
err ' ( H )  e t 1
, where γt    t
2

 AdaBoost ADApts to the error rates of the

individual weak hypotheses.
 AdaBoost – generalization error
 Freund and Schapire (1997) showed that:

Td
err ( H )  Pr'[ H ( x)  y ]  O( ),
m
where :

Pr '[H(x)  y] - the empirical probability on the training sample

d - VC dimension

T - number of rounds

m - training set size

 AdaBoost – generalization error
 The analysis implies that boosting will overfit
if run for too many rounds
 However, it was observed empirically that
AdaBoost does not overfit, even when run
thousands of rounds.
 Moreover, it was observed that the
generalization error continues to drive down
long after training error reached zero
 AdaBoost – generalization error
 An alternative analysis was presented by
Schapire et al. (1998), that suits the empirical
findings
d
Pr[err ( H )  Pr'[margin ( x, y )   ]  O ( ) ]  1- 
m 2

where :

margin( x , y ) 
y  t ht ( x )
 t
 AdaBoost – different point of view
 We try to solve the problem of approximating the
y’s using a linear combination of weak hypotheses
 In other words, we are interested in the problem of
finding a vector of parameters α such that
n
f ( xi )    j h j ( xi )
j 1

is a ‘good approximation’ of yi
 For classification problems we try to match the
sign of f(xi) to yi
 AdaBoost – different point of view
 Sometimes it is advantageous to minimize some
other (non-negative) loss function instead of the
number of classification errors
 For AdaBoost the loss function is

 exp( y
i 1
i f ( xi ))

 This point of view was used by Collins, Schapire

and Singer (2002) to demonstrate that AdaBoost
converges to optimality
Face Detection
(not face recognition)
 Face Detection in Monkeys
There are cells that
‘detect faces’
 Face Detection in Human
There are ‘processes of
face detection’
 Faces Are Special
We analyze faces in a
‘different way’
 Faces Are Special
We analyze faces in a
‘different way’
 Faces Are Special
We analyze faces in a
‘different way’
 Face Recognition in Human
We analyze faces ‘in a
specific location’
Robust Real-Time Face
Detection
Viola and Jones, 2003
 Features

Picture analysis, Integral Image

 Features
 The system classifies images based on the value
of simple features

Two-rectangle

Value =
Three-rectangle
∑ (pixels in white area) - ∑
(pixels in black area)

Four-rectangle
Contrast Features
Source

Result
 Features
 Notice that each feature is related to a
special location in the sub-window
 Why features and not pixels?
– Encode domain knowledge
– Feature based system operates faster
– Inspiration from human V1
 Features
 Later we will see that there are other
features that can be used to implement an
efficient face detector
 The original system of Viola and Jones used
only rectangle features
 Computing Features
 Given a detection resolution of 24x24, and
size of ~200x200, the set of rectangle
features is ~160,000 !
 We need to find a way to rapidly compute
the features
 Integral Image
 Intermediate
representation of the
image

ii ( x, y )   i ( x' , y ' )
x '  x , y ' y

 Computed in one pass

over the original image
s ( x, y )  s ( x, y  1)  i ( x, y )
ii ( x, y )  ii ( x  1, y )  s ( x, y )
s ( x,1)  0
ii ( 1, y )  0
 Integral Image
x
(0,0)
s(x,y) = s(x,y-1) + i(x,y)
ii(x,y) = ii(x-1,y) + s(x,y)

y (x,y)

Using the integral image representation

one can compute the value of any
rectangular sum in constant time.
For example the integral sum inside
rectangle D we can compute as:
ii(4) + ii(1) – ii(2) – ii(3)
 Integral Image

(x,y)

Integral
Image
-1 +1
+2 -2
-1 +1
(x,y)
 Building a Detector

Cascading, training a cascade

 Main Ideas
 The Features will be used as weak
classifiers
 We will concatenate several detectors
serially into a cascade
 We will boost (using a version of AdaBoost)
a number of features to get ‘good enough’
detectors
 Main Ideas
 The Features will be used as weak
classifiers
 We will concatenate several detectors
serially into a cascade
 We will boost (using a version of AdaBoost)
a number of features to get ‘good enough’
detectors
 Weak Classifiers
Weak Classifier : A feature which best separates
the examples
Given a sub-window (x), a feature (f), a threshold
(Θ), and a polarity (p) indicating the direction of
the inequality:

h( x, f , p, )  1  pf ( x)  p
 Weak Classifiers
 A weak classifier is a combination of a
feature and a threshold
 We have K features
 We have N thresholds where N is the
number of examples
 Thus there are KN weak classifiers
 Weak Classifier Selection
 For each feature sort the examples based on
feature value
 For each element evaluate the total sum of
positive/negative example weights (T+/T-) and
the sum of positive/negative weights below the
current example (S+/S-)
 The error for a threshold which splits the range
between the current and previous example in the
sorted list is
e  min(S   (T   S  ), S   (T   S  ))
 An example
e B A S- S+ T- T+ W f y x

2/5 3/5 2/5 0 0 2/5 3/5 1/5 2 -1 X1

1/5 4/5 1/5 1/5 0 2/5 3/5 1/5 3 -1 X2

0 5/5 0 2/5 0 2/5 3/5 1/5 5 1 X3

1/5 4/5 1/5 2/5 1/5 2/5 3/5 1/5 7 1 X4

2/5 3/5 2/5 2/5 2/5 2/5 3/5 1/5 8 1 X5

 Main Ideas
 The Features will be used as weak
classifiers
 We will concatenate several detectors
serially into a cascade
 We will boost (using a version of AdaBoost)
a number of features to get ‘good enough’
detectors
 Main Ideas
 The Features will be used as weak
classifiers
 We will concatenate several detectors
serially into a cascade
 We will boost (using a version of AdaBoost)
a number of features to get ‘good enough’
detectors
 Cascading
 We start with simple classifiers which reject
many of the negative sub-windows while
detecting almost all positive sub-windows
 Positive results from the first classifier
triggers the evaluation of a second (more
complex) classifier, and so on
 A negative outcome at any point leads to the
immediate rejection of the sub-window
Cascading
 Main Ideas
 The Features will be used as weak
classifiers
 We will concatenate several detectors
serially into a cascade
 We will boost (using a version of AdaBoost)
a number of features to get ‘good enough’
detectors
 Main Ideas
 The Features will be used as weak
classifiers
 We will concatenate several detectors
serially into a cascade
 We will boost (using a version of
AdaBoost) a number of features to get
‘good enough’ detectors
 Training a cascade
 User selects values for:
– Maximum acceptable false positive rate per
layer
– Minimum acceptable detection rate per layer
– Target overall false positive rate
 User gives a set of positive and negative
examples
 Training a cascade (cont.)
 While the overall false positive rate is not met:
– While the false positive rate of current layer is less than
the maximum per layer:
 Train a classifier with n features using AdaBoost on set of
positive and negative examples
 Decrease threshold for current classifier detection rate of the
layer is more than the minimum
 Evaluate current cascade classifier on validation set
– Evaluate current cascade detector on a set of non faces
images and put any false detections into the negative
training set
Results
 Training Data Set
 4916 hand labeled faces
 Aligned to base resolution
(24x24)
 Non faces for first layer
were collected from 9500
non faces images
 Non faces for subsequent
layers were obtained by
scanning the partial
cascade across non faces
and collecting false
positives (max 6000 for
each layer)
 Structure of the Detector
 38 layer cascade
 6060 features

Layer number 1 2 3 to 4 5 to 38
Number of feautures 2 10 50 -
Detection rate 100% 100% - -
Rejection rate 50% 80% - -
 Speed of final Detector
 On a 700Mhz Pentium III processor, the
face detector can process a 384 by 288
pixel image in about .067 seconds
 Improvements

Learning Object Detection from a Small

Number of Examples: the Importance of
Good Features (Levy & Weiss, 2004)
 Improvements
 Performance depends crucially on the
features that are used to represent the
objects (Levy & Weiss, 2004)
 Good Features imply:
– Good results from small training databases
– Better generalization abilities
– Shorter (faster) classifiers
 Edge Orientation Histogram

 Invariant to global illumination changes

 Captures geometric properties of faces
 Domain knowledge represented:
– Inner part of the face includes more horizontal edges then vertical
– The ration between vertical and horizontal edges is bounded
– The area of the eyes includes mainly horizontal edges
– The chin has more or less the same number of oblique edges on
both sides
 Edge Orientation Histogram
 The EOH can be calculated using some kind
of Integral Image:
– We find the gradients at the point (x,y) using
Sobel masks
– We calculate the orientation of the edge (x,y)
– We divide the edges into K bins
– The result is stored in K matrices
– We use the same idea of Integral Image for the
matrices
 EOH Features
 The ratio between two
orientations
 The dominance of a given
orientation
 Symmetry Features
 Results
 Already with only 250 positive examples we
can see above 90% detection rate
 Faster classifier
 Better performance in profile faces
 Demo
Implementing Viola & Jones system
Frank Fritze, 2004