The Viola/Jones Face Detector

1
The Viola/Jones Face Detector

• Prepared with figures taken from
• Robust real-time object detection
• CRL 2001/01, February 2001

2
Three Measures Toward Speeded Up Detection

• Integral image a fast way to compute simple
  features
• In Adaboost the weak learner is nothing but a
  feature selector. The advantage is that if there
  are N weak learners there are merely N features
  to compute.
• Cascaded combination of classifiers. Most of true
  negatives are rejected very fast at the at the
  first few stages. Can keep high detection rate
  and low false positive rate.

3
Image Features
Rectangle filters Similar to Haar wavelets
Base resolution is 24-by-24 11 scales, scaling
factor of 1.25 45396 features
4
Rectangular Features for Face Detection
Forehead, eye features can be captured
5
Fast Feature Computation Integral Image

• Integral image value at a pixel (x, y) is the sum
  of the pixel values of the original image above
  and to the left of (x, y), inclusive.
• Integral image can be computed by one pass
  through the image

6
Computing Sum within a Rectangle by Integral Image

• The sum of the pixels within rectangle D can be
  computed with four array references.
• The value of the integral image at location 1 is
  the sum of the pixels in rectangle A.
• The value at location 2 is A B, at location 3
  is A C, and at location 4 is A B C D.
• The sum within D can be computed as 4 1 - (2
  3).

7
Constrained Classifier Feature Selection

• Restrict the weak learner to a single feature
• A weak classifier hj(x) consists of a feature fj,
  a threshold ?j, and a parity pj indicating the
  direction of inequality sign
• x is a 24-by-24 sub-window of an image

8
Boosting Algorithm
9
Learning Result
Must do better
10
Cascading Classifiers
The initial classifier eliminates a large number
of negative examples with very little processing.
Subsequent layers eliminate additional
negatives but require additional computation.
After several stages of processing the number
of sub-windows have been reduced radically.
11
How Cascading Can Meet Performance?
For K stages of cascading with each stage having
fi as the false positive rate, the overall false
positive rate for the cascade is
Similarly, the overall detection rate is
To keep F very low and D very high, for each
stage the goal is to have very high detection
rate (close to 100), but moderate false
positive rate (say, 30)
12
Cascaded Classifier
50
20
2
IMAGE SUB-WINDOW
5 Features
20 Features
FACE
1 Feature
F
F
F
NON-FACE
NON-FACE
NON-FACE

• A 1 feature classifier achieves 100 detection
  rate and about 50 false positive rate.
• A 5 feature classifier achieves 100 detection
  rate and 40 false positive rate (20 cumulative)
• using data from previous stage.
• A 20 feature classifier achieve 100 detection
  rate with 10 false positive rate (2 cumulative)

13
Building A Cascaded Detector
14
Classifier is Learned from Labeled Data

• Training Data
• 4916 hand labeled faces
• All frontal
• 10000 non faces
• Faces are normalized
• Scale, translation
• Many variations
• Across individuals
• Illumination
• Pose (rotation both in plane and out)

15
Boosted Face Detection

• For each round of boosting
• Evaluate each rectangle filter on each example
• Sort examples by filter values
• Select best threshold for each filter (min Z)
• Select best filter/threshold ( Feature)
• Reweight examples
• Weeks to learn train
• 15 frames per second to detect faces from unknown
  images.

16
Performance
17
Output of Face Detector on Test Images