edu.berkeley.guir.lib.gesture
Class Classifier

java.lang.Object
  |
  +--edu.berkeley.guir.lib.gesture.Classifier

Direct Known Subclasses:

RubineClassifier

public class Classifier

extends Object

Gesture classifier/recogizer, using Dean Rubine's algorithm, with guidance from Rob Miller's C++ port, which is part of the Amulet package. Note: for performance reasons, the Classifier does not retrain automatically when the underlying GestureSet changes. It will retrain automatically the next time it needs to (e.g., for classify()). Known bugs: training may take some time, but train() does not check to see if its thread gets interrupted.

Inner Class Summary

class	Classifier.FeatureDirection
protected class	Classifier.MyCollectionListener
protected class	Classifier.MyPropChangeListener
class	Classifier.Result Used to return the results of a classification.

Field Summary

protected double[][]	ccm
protected double[][]	ccmInv
protected GestureCategory	dotCategory
protected List	enabledCategories
protected static double	EPSILON
protected Class[]	featureClasses
protected BitSet	featuresUsed
protected GestureSet	gestureSet
protected double[][]	meanFeatureValues
static String	MISRECOGNIZED_PROP Set on gestures that are misrecognized, with a Result value
protected CollectionListener	myCollectionListener
protected PropertyChangeListener	myPropChangeListener
protected PropertyChangeSupport	propChangeSupport
protected boolean	trained
static String	TRAINED_PROP a Boolean property indicating whether the classifier is trained or not
protected double[][]	weights

Constructor Summary

Classifier()

Classifier(GestureSet gs)
Does not automatically train.

Method Summary

void	addPropertyChangeListener(PropertyChangeListener listener)
void	addPropertyChangeListener(String propertyName, PropertyChangeListener listener)
double	categoryDistance(GestureCategory categoryA, GestureCategory categoryB) Return the square of the distance between the two categories (which must in the GestureSet).
double	categoryDistance(int categoryA, int categoryB) Return the square of the distance between the two categories.
protected void	checkForInterrupt()
Classifier.Result	classify(Gesture gesture) Determine which GestureCategory the Gesture is most likely a member of, and return an appropriate Result.
Classifier.Result	classifyWithoutTraining(Gesture gesture) Only call this if you know the GestureSet the classifier is using hasn't changed since the last time classify() or train() was called.
protected double[][]	computeCovarianceMatrix(GestureCategory gestureCategory, double[] meanFeatureVector) Compute and return covariance matrix and fill in the mean feature value array (if provided) (Note: only upper triangular part computed)
double	distanceToCategory(Gesture gesture, GestureCategory category)
double	distanceToCategory(Gesture gesture, int catIndex)
void	dump(PrintStream out)
void	dumpMFV(PrintStream out) Return a matrix where A[Ci][Cj] is a Vector of category Ci examples (i.e.
void	dumpRelativeVariance(PrintStream out)
int	findPrincipleComponent(double[] featureVals) Return the index of the largest component of featureVals, using the current training weights.
Classifier.FeatureDirection	findPrincipleFeature(int catA, int catB) Find the feature class on which catA and catB differ most (based on current weights).
protected boolean	fixClassifier(double[][] ccm, double[][] ccmInv)
static Class[]	getDefaultFeatureClasses()
double[]	getDistancesByFeature(int catA, int catB) Use the current weights to get normalized distances along each feature between two categories.
protected Feature	getFeature(GestureCategory gc, int gIndex, Class featureClass)
protected Feature	getFeature(GestureCategory gc, int gIndex, int fIndex)
Class[]	getFeatureClasses()
BitSet	getFeaturesUsed() Features actually used in current classification (null if untrained)
GestureSet	getGestureSet()
double[]	getNormalizedDistancesByFeature(Gesture gesture, GestureCategory gc) Use the weights for the category to get normalized distance of the gesture from the category.
double[]	getNormalizedDistancesByFeature(Gesture gesture, int categoryIndex) Use the weights for the category to get normalized distance of the gesture from the category.
double[]	getNormalizedDistancesByFeature(Gesture gesture, String categoryName) Use the weights for the category to get normalized distance of the gesture from the category.
List	getTrainingCategories() Returns an unmodifiable List of the categories that were used to train the classifier.
double[][]	getWeights()
boolean	isTrained()
double	MahalanobisDistance(double[] v, double[] u) Return the square of the distance between the two vectors.
static double	MahalanobisDistance(double[] v, double[] u, double[][] sigma) Return the square of the distance between the two vectors, using the specified dispersion matrix.
void	removePropertyChangeListener(PropertyChangeListener listener)
void	removePropertyChangeListener(String propertyName, PropertyChangeListener listener)
void	setFeatureClasses(Class[] featureClasses)
void	setGestureSet(GestureSet gs) Does not automatically train.
protected void	setTrained(boolean isTrained)
boolean	testRecognition(Gesture gesture) return true iff the gesture was correctly recognized
List	testRecognition(GestureContainer container)
List	testRecognition(Iterator iter) iter should contain Gestures or GestureContainers.
List	testRecognition(Iterator iter, boolean onlyEnabled) iter should contain Gestures or GestureContainers.
void	train() (Re)trains the current Classifier.

Methods inherited from class java.lang.Object

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

Field Detail

TRAINED_PROP

public static final String TRAINED_PROP

a Boolean property indicating whether the classifier is trained or not

MISRECOGNIZED_PROP

public static final String MISRECOGNIZED_PROP

Set on gestures that are misrecognized, with a Result value

featureClasses

protected Class[] featureClasses

EPSILON

protected static final double EPSILON

gestureSet

protected GestureSet gestureSet

weights

protected double[][] weights

meanFeatureValues

protected double[][] meanFeatureValues

ccm

protected double[][] ccm

ccmInv

protected double[][] ccmInv

dotCategory

protected GestureCategory dotCategory

trained

protected boolean trained

featuresUsed

protected BitSet featuresUsed

enabledCategories

protected List enabledCategories

myCollectionListener

protected CollectionListener myCollectionListener

myPropChangeListener

protected PropertyChangeListener myPropChangeListener

propChangeSupport

protected transient PropertyChangeSupport propChangeSupport

Constructor Detail

Classifier

public Classifier()

Classifier

public Classifier(GestureSet gs)

Does not automatically train.

Method Detail

setFeatureClasses

public void setFeatureClasses(Class[] featureClasses)

getFeatureClasses

public Class[] getFeatureClasses()

getDefaultFeatureClasses

public static Class[] getDefaultFeatureClasses()

getWeights

public double[][] getWeights()

setGestureSet

public void setGestureSet(GestureSet gs)

Does not automatically train.

getGestureSet

public GestureSet getGestureSet()

getFeaturesUsed

public BitSet getFeaturesUsed()

Features actually used in current classification (null if untrained)

getFeature

protected Feature getFeature(GestureCategory gc,
                             int gIndex,
                             int fIndex)

getFeature

protected Feature getFeature(GestureCategory gc,
                             int gIndex,
                             Class featureClass)

computeCovarianceMatrix

protected double[][] computeCovarianceMatrix(GestureCategory gestureCategory,
                                             double[] meanFeatureVector)

Compute and return covariance matrix and fill in the mean feature value array (if provided) (Note: only upper triangular part computed)

setTrained

protected void setTrained(boolean isTrained)

isTrained

public boolean isTrained()

checkForInterrupt

protected void checkForInterrupt()
                          throws InterruptedException

train

public void train()
           throws TrainingException,
                  InterruptedException

(Re)trains the current Classifier.

fixClassifier

protected boolean fixClassifier(double[][] ccm,
                                double[][] ccmInv)

classify

public Classifier.Result classify(Gesture gesture)
                           throws TrainingException,
                                  InterruptedException

Determine which GestureCategory the Gesture is most likely a member of, and return an appropriate Result. Resturns null if no GestureSet has been assigned, or if the GestureSet has no members.

classifyWithoutTraining

public Classifier.Result classifyWithoutTraining(Gesture gesture)

Only call this if you know the GestureSet the classifier is using hasn't changed since the last time classify() or train() was called. See classify() for return value information.

testRecognition

public List testRecognition(GestureContainer container)
                     throws InterruptedException,
                            TrainingException

testRecognition

public List testRecognition(Iterator iter)
                     throws InterruptedException,
                            TrainingException

iter should contain Gestures or GestureContainers. returns the misrecognized Gestures. If all are correctly recognized, returns an empty List.

testRecognition

public List testRecognition(Iterator iter,
                            boolean onlyEnabled)
                     throws InterruptedException,
                            TrainingException

iter should contain Gestures or GestureContainers. returns the misrecognized Gestures. If all are correctly recognized, returns an empty List. If onlyEnabled is true, skip GestureObjects that aren't enabled.

testRecognition

public boolean testRecognition(Gesture gesture)
                        throws InterruptedException,
                               TrainingException

return true iff the gesture was correctly recognized

categoryDistance

public double categoryDistance(GestureCategory categoryA,
                               GestureCategory categoryB)

Return the square of the distance between the two categories (which must in the GestureSet). If the classifier is not trained, it throws an IllegalStateException.

categoryDistance

public double categoryDistance(int categoryA,
                               int categoryB)

Return the square of the distance between the two categories. If the classifier is not trained, it throws an IllegalStateException.

distanceToCategory

public double distanceToCategory(Gesture gesture,
                                 GestureCategory category)

distanceToCategory

public double distanceToCategory(Gesture gesture,
                                 int catIndex)

getNormalizedDistancesByFeature

public double[] getNormalizedDistancesByFeature(Gesture gesture,
                                                String categoryName)

Use the weights for the category to get normalized distance of the gesture from the category. This distance is not comparable with the mahalanobis distance (I don't think).

getNormalizedDistancesByFeature

public double[] getNormalizedDistancesByFeature(Gesture gesture,
                                                GestureCategory gc)

Use the weights for the category to get normalized distance of the gesture from the category. This distance is not comparable with the mahalanobis distance (I don't think).

getNormalizedDistancesByFeature

public double[] getNormalizedDistancesByFeature(Gesture gesture,
                                                int categoryIndex)

Use the weights for the category to get normalized distance of the gesture from the category. This distance is not comparable with the mahalanobis distance (I don't think). categoryIndex is the index into the enabled classes that the classifier has.

getDistancesByFeature

public double[] getDistancesByFeature(int catA,
                                      int catB)

Use the current weights to get normalized distances along each feature between two categories. This distance is not comparable with the mahalanobis distance (I don't think). catA and catB are indices into the enabled classes that the classifier has.

getTrainingCategories

public List getTrainingCategories()

Returns an unmodifiable List of the categories that were used to train the classifier.

MahalanobisDistance

public double MahalanobisDistance(double[] v,
                                  double[] u)

Return the square of the distance between the two vectors.

MahalanobisDistance

public static double MahalanobisDistance(double[] v,
                                         double[] u,
                                         double[][] sigma)

Return the square of the distance between the two vectors, using the specified dispersion matrix.

findPrincipleComponent

public int findPrincipleComponent(double[] featureVals)

Return the index of the largest component of featureVals, using the current training weights. (If all features were weighted equally, we'd be in normal Euclidean space and this would simply be the largest feature value.)

findPrincipleFeature

public Classifier.FeatureDirection findPrincipleFeature(int catA,
                                                        int catB)

Find the feature class on which catA and catB differ most (based on current weights). catA and catB are the indices of categories in the current enabled categories list.

dumpMFV

public void dumpMFV(PrintStream out)

Return a matrix where A[Ci][Cj] is a Vector of category Ci examples (i.e. Gestures) that are classified as Cj.

dump

public void dump(PrintStream out)

dumpRelativeVariance

public void dumpRelativeVariance(PrintStream out)

addPropertyChangeListener

public void addPropertyChangeListener(PropertyChangeListener listener)

addPropertyChangeListener

public void addPropertyChangeListener(String propertyName,
                                      PropertyChangeListener listener)

removePropertyChangeListener

public void removePropertyChangeListener(PropertyChangeListener listener)

removePropertyChangeListener

public void removePropertyChangeListener(String propertyName,
                                         PropertyChangeListener listener)