org.openimaj.image.objectdetection.haar

Class StageTreeClassifier

java.lang.Object

org.openimaj.image.objectdetection.haar.StageTreeClassifier

@Reference(type=Inproceedings,
           author={"Viola, P.","Jones, M."},
           title="Rapid object detection using a boosted cascade of simple features",
           year="2001",
           booktitle="Computer Vision and Pattern Recognition, 2001. CVPR 2001. Proceedings of the 2001 IEEE Computer Society Conference on",
           pages={" I","511 "," I","518 vol.1"},
           number="",
           volume="1",
           customData={"keywords"," AdaBoost; background regions; boosted simple feature cascade; classifiers; face detection; image processing; image representation; integral image; machine learning; object specific focus-of-attention mechanism; rapid object detection; real-time applications; statistical guarantees; visual object detection; feature extraction; image classification; image representation; learning (artificial intelligence); object detection;","doi","10.1109/CVPR.2001.990517","ISSN","1063-6919 "})
public class StageTreeClassifier
extends Object

A tree of classifier stages. In the case that the tree is degenerate and all Stages have null Stage.failureStage()s, then the tree is known as a cascade.

The general idea is that for a given window in the image being tested (defined by an x,y position and scale), the stage tree is evaluated. If when evaluating the tree a leaf node is hit (i.e. a Stage that passes successfully, but has a null Stage.successStage()) then the tree is said to have passed, and indicates a potential object detection within the window. If a Stage fails to pass and has a null Stage.failureStage() then the tree is said to have failed, indicating the object in question was not found.

In order to achieve good performance, this implementation pre-computes and caches variables related to a given detection scale. This means that it is NOT safe to use a detector based on this stage implementation in a multi-threaded environment such that multiple images are being tested at a given time. It is however safe to use this implementation with a detector that multi-threads its detection across the x and y window positions for a fixed scale:

  StageTreeClassifier cascade = ...
 
        for each scale {
                cascade.setScale(scale);
 
                //the x and y search could be threaded...
                for each y {
                        for each x {
                                cascade.matches(sat, x, y); {
                        }
                }
 }
 

Author:

Jonathon Hare (jsh2@ecs.soton.ac.uk)

Constructor Summary

Constructors

Constructor and Description
StageTreeClassifier(int width, int height, String name, boolean hasTiltedFeatures, Stage root) Construct the StageTreeClassifier with the given parameters.

Method Summary

Modifier and Type	Method and Description
int	classify(SummedSqTiltAreaTable sat, int x, int y) Apply the classifier to the given image at the given position.
int	getHeight() Get the classifier height
String	getName() Get the classifier name
Stage	getRoot() Get the root Stage of the classifier
int	getWidth() Get the classifier width
boolean	hasTiltedFeatures() Does the classifier use tilted haar-like features?
void	setScale(float scale) Set the current detection scale.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Detail

StageTreeClassifier

public StageTreeClassifier(int width,
                           int height,
                           String name,
                           boolean hasTiltedFeatures,
                           Stage root)

Construct the StageTreeClassifier with the given parameters.

Parameters:

width - the width of the classifier

height - the height of the classifier

name - the name of the classifier

hasTiltedFeatures - are there tilted haar-like features in the classifiers?

root - the root of the tree of stages

Method Detail

setScale

public void setScale(float scale)

Set the current detection scale. This must be called before calling classify(SummedSqTiltAreaTable, int, int).

Internally, this goes through all the stages and their individual classifiers and pre-caches information related to the current scale to avoid lots of expensive recomputation of values that don't change for a given scale.

Parameters:

scale - the current scale

classify

public int classify(SummedSqTiltAreaTable sat,
                    int x,
                    int y)

Apply the classifier to the given image at the given position. Internally, this will apply each stage to the image. If all stages complete successfully a detection is indicated.

This method returns the number of stages passed if all stages pass; if a stage fails, then (-1 * number of successful stages) is returned. For example a value of 20 indicates the successful detection from a total of 20 stages, whilst -10 indicates an unsuccessful detection due to a failure on the 11th stage.

Parameters:

sat - the summed area table(s) for the image in question. If there are tilted features, this must include the tilted SAT.

x - the x-ordinate of the top-left of the current window

y - the y-ordinate of the top-left of the current window

Returns:

> 0 if a detection was made; <=0 if no detection was made. The magnitude indicates the number of stages that passed.

getWidth

public int getWidth()

Get the classifier width

Returns:

the width

getHeight

public int getHeight()

Get the classifier height

Returns:

the height

getName

public String getName()

Get the classifier name

Returns:

the name

hasTiltedFeatures

public boolean hasTiltedFeatures()

Does the classifier use tilted haar-like features?

Returns:

true if tilted features are used; false otherwise.

getRoot

public Stage getRoot()

Get the root Stage of the classifier

Returns:

the root