Paper_Reading 《Rapid Object Detection using a Boosted Cascade of Simple Features》

Paper Link

Author:Paul Viola,Michael Jones

Key Contributions:

1.Integral Image

2.Learning algorithm based on AdaBoost

3.Combine Classifiers in Cascade

Methods:

• Represent and classify images based on simple features rather than pixels directly(operates faster and easily encode ad-hoc domain knowledge)
• Three kinds of simple features are used in the paper
  1. Two-Rectangles features (A and B)
  2. Three-Rectangles features (C)
  3. Four-Rectangles features (D)
     

The feature value calculation:
$$\sum{pixel\ values\ in\ white}-\sum{pixel\ values\ in\ gray}$$

• Contributions 1:Integral Image
  
  

• Contributions 2:Learning Algorithm based on AdaBoost

  The advantages of AdaBoost:
  1.Used for feature selection and classifier training
  2.Selecting small set of good features from large feature set
  3.Used a set of weak learners to form a strong one
  4.Guarantees the training error of strong classifier very low

  The Pesudo of Adaboost:
  

• Contributions 3:Combine Classifiers in Cascade

  • Building cascade of classifiers(Increase Performance & Reduce computation)
  • Simpler classifiers apply early to reject majority of sub windows and apply complex classifiers to achieve low false positive
  • Subsequent classifiers are trained using examples,which pass through all the previous stages
    

• How to use

  1. User selects maximun FPR(False Prediction Rate) and minimun acceptable DR(Detection Rate) per each stage
  2. User selects target
  3. Each stage is trained by adding features until the target DR and FPRs are met
  4. Stages are added until the overall target for DR and FPR are met

• Further optimization

  1. Number of classifier stages
  2. Number of features in each stage
  3. Threshold of each stage
  4. Minimun number of features that achieved accuracy

• Conclusion

  1. Solution achieves the goal of real time object detection
  2. Conjunction of simple rectangle features and integral image gives a efficient feature representation
  3. AdaBoost is used for the feature selection and classifier training
  4. Cascade of classifiers allows to quickly discard background regions and concentrate more on ojbect-like regions

• Approximately Performance
  Accuracy:Front-Face around 68.8%,Profile around 33%
  Time:every picture with 67 millisecond