Rapid object detection using boosted cascade of simple features

Rapid Object Detection using a Boosted
Cascade of Simple Features
Authors: Paul Viola and Michael Jones
Presented by: Hirantha Pradeep Weerarathna
Supervisor: Dr. Anuja Dharmaratne

University of Colombo School of Computing

Authors

Paul Viola

– Got PhD in 1995 from MIT
– Current Designations:
• Manger, Architect, Developer, and Researcher at Microsoft
• Professor at MIT
– Research Interests are Machine Learning, Web Search and Image
Processing

Authors

Michael Jones

– Got PhD from Massachusetts Institute of Technology in 1997
– Currently senior member in MERL Research
– Research Interests are Computer Vision and Machine Learning

Road Map

− Introduction
− Key Contributions
− Results and Comparisons
− Conclusion

 A prominent paper in the domain of Object Detection.
 More than 6000 citations to the date in related articles.
 First solution provided real time Object Detection.
 Still most widely used solution for the Object Detection.

Object detection

Detecting instances of semantic objects of a certain class (such as
humans, buildings, or cars) in digital images and videos.
~wikipedia.org

Applications:
 Robotic Applications
 Surveillance Systems
 Sports Applications
 Security Systems
 And many more..

Face Detection
Determining the locations and sizes of human faces in arbitrary
images.
~wikipedia.org

Image Representation

 Represent and classify images based on simple features rather than
pixels directly. Reasons are:
1. Feature can encode ad-hoc domain knowledge that is difficult to learn using
finite quantity of training data.
2. Feature based systems operates much faster than pixel based systems.

Features
 Three kinds of simple features are used.
1. Two-Rectangles features
2. Three-Rectangles features
3. Four-Rectangles features

 Feature value Calculation

∑pixel values in white area - ∑pixel values in gray area

Key Contributions

 Three main contributions
1. Introduction of Integral Image
2. Learning algorithm based on AdaBoost
3. Combine Classifiers in Cascade

Integral Image
 This concept was first introduced with this solution framework.
 Integral Image is computed from an image using few operations on
pixels.

y y
10 20 10 20 10 30 40 60
20 10 10 10 30 60 80 110
x 30 10 10 20 x 60 100 130 180
10 20 30 20 70 130 190 260

Original Image Integral Image

Integral Image
 Using Integral Image, pixel sum of a rectangle are can be calculated
using 4 array references.
 It leads to a rapid evaluation of rectangle features
 Feature evaluation in constant time

A 1 B 2

C D
3 4

∑Pixel sum of area D= ii(4) + ii(1) – ii(2) –ii(3)

Learning Algorithm based on
AdaBoost
 AdaBoost is used for feature selection and classifier training
 Capable of selecting a small set of good features from a large
number of feature set
 AdaBoost use a set of weak learners to form a strong one
 It guarantees that training error of the strong classifier reach zero
exponentially in number of rounds

Learning Algorithm based on
AdaBoost
 A weak learner select a single rectangle feature which best seperates
positive and negative examples
 Weak learner determines the optimal threshold function, such that
misclassification is minimized

Rapid object detection using boosted cascade of simple features

Combine Classifiers in Cascade
 Building cascade of classifiers,
 Increase detection performance
 Rapidly reduce computation power
 Simpler classifiers apply early and reject majority of sub windows,
then apply complex classifiers to achieve low false positive
 Subsequent classifiers are trained using examples, which pass
through all the previous stages

 Cascade Training process involves two trade-offs
1. Classifier with more features will achieve higher DR and lower FPR
2. Classifier with more features need more computations

 Can define a optimization framework in which
1. Number of classifier stages
2. Number of features in each stage
3. Threshold of each stage

 Minimum number of features are selected such that, expected DR
and FPR are achieved

 Simple framework is used to produce effective cascade which is
highly efficient
1. User selects maximum acceptable FPR and minimum acceptable DR per
each stage
2. User selects target overall FPR and DR
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

Results
 Testing has done on MIT+CMU test set, which consists with 507
faces in 130 images
 Using a cascade of 38 layers
 Cascade has trained using 4916 facial images and 9544 non-facial
images
 Testing has been done with scaling factor of 1.25 and windows
shifting scale of 1.0 on images
 On a conventional Pentium III machine with 700Mhz processor.
 They have achieved Detection Speed of 15 frames/sec

Results
ROC Curve for Face Detector

Results
Detection Rate Comparison of Cotemporary Solution

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

Rapid object detection using boosted cascade of simple features

More Related Content

What's hot (20)

Viewers also liked (9)

Similar to Rapid object detection using boosted cascade of simple features (20)

Recently uploaded (20)

Rapid object detection using boosted cascade of simple features

Editor's Notes