Fast and Robust Classification using Asymmetric AdaBoost and a Detector Cascade

1
Fast and Robust Classification using Asymmetric
AdaBoost and a Detector Cascade

• Paul viola and Michael Jones

2
Outline

• Motivation and objective
• Definition of feature vectors
• Asymmetric Adaboosting
• Experiments and Results
• Conclusions

3
Motivation

• Given a set of images find regions in these
  images which contain instances of faces.

4
Objectives and Applications

• Objective
• Fast classification
• High detection rate
• Low false positive
• Application domains
• Face detection
• Database retrieval

5
Case Study Face Detection

• Difficulty in using raw grayscale pixel values
• To capture (learn) ad-hoc domain knowledge
  classifiers for images.
• Computation overhead.
• Derive another features by applying simple
  filters to the pixels.

6
Definition of Simple Features

• Using a 24x24 pixel base detection window,
  with all the possible combination of horizontal
  and vertical location and scale of these feature
  types, the full set of features sums up to
  54,864.

7
Example

• The motivation behind using rectangular features,
  as opposed to more expressive steerable filters
  is due to their extreme computational efficiency.

8
Integral image

• Definition The integral image at location (x,y),
  is the sum of the pixel values above and to the
  left of (x,y), inclusive.
• Function Using the following two recurrences,
  where i(x,y) is the pixel value of original image
  at the given location and s(x,y) is the
  cumulative column sum, we can calculate the
  integral image representation of the image in a
  single pass.

9
Rapid Evaluation of Features
10
Features
11
Algorithm Overview

• Main components
• Cascading
• Asymmetric adaboost
• Simple, boosted classifiers can reject many of
  negative subwindows while detecting all positive
  instances.

12
(No Transcript)
13
Limitation of Adaboost

• AdaBoost minimizes a quantity related to
  classification error it does not minimize the
  number of false negatives.
• Unfortunately feature selection proceeds as if
  classification error were the only goal, and the
  features selected are not optimal for the task of
  rejecting negative examples

14
Asymmetric Adaboost

• If we hope to minimize false negatives then the
  weight on positive examples could be increased so
  that the minimum error criteria will also have
  very few false negatives.

15
Asymmetric Adaboost

• Minimization of this bound can be achieved using
  AdaBoost by pre-weighting each example by
  exp(0.5yi log k )
• Using the steps followed in Adaboost we obtain

16
Training
5000 faces 5000 false positive from layer 1
Layer1
Weak classifier 1
Weak classifier 2
Weak classifier 3
50,000
17
Cascading
18
Experiment

• Naive asymmetric v.s. asymmetric boosting

19
Experiment

• asymmetric boosting v.s. normal boosting

20
Results
21
Results
18 Layers
22 Layers
22
Result
26 Layers
23
Questions
24

• Thank You