fast k-means algorithm implementation using k-d tree structure More...

#include <itkKdTreeBasedKmeansEstimator.h>

Inheritance diagram for itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >:

Collaboration diagram for itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >:

Classes
class	CandidateVector
	Candidate Vector. More...
Public Types
typedef KdTreeNodeType::CentroidType	CentroidType
typedef itksys::hash_map < InstanceIdentifier, unsigned int >	ClusterLabelsType
typedef SmartPointer< const Self >	ConstPointer
typedef DistanceToCentroidMembershipFunctionType::Pointer	DistanceToCentroidMembershipFunctionPointer
typedef DistanceToCentroidMembershipFunction < MeasurementVectorType >	DistanceToCentroidMembershipFunctionType
typedef TKdTree::InstanceIdentifier	InstanceIdentifier
typedef std::vector < ParameterType >	InternalParametersType
typedef TKdTree::KdTreeNodeType	KdTreeNodeType
typedef TKdTree::MeasurementType	MeasurementType
typedef unsigned int	MeasurementVectorSizeType
typedef TKdTree::MeasurementVectorType	MeasurementVectorType
typedef MembershipFunctionType::ConstPointer	MembershipFunctionPointer
typedef MembershipFunctionBase < MeasurementVectorType >	MembershipFunctionType
typedef MembershipFunctionVectorObjectType::Pointer	MembershipFunctionVectorObjectPointer
typedef SimpleDataObjectDecorator < MembershipFunctionVectorType >	MembershipFunctionVectorObjectType
typedef std::vector < MembershipFunctionPointer >	MembershipFunctionVectorType
typedef Array< double >	ParametersType
typedef Array< double >	ParameterType
typedef SmartPointer< Self >	Pointer
typedef TKdTree::SampleType	SampleType
typedef KdTreeBasedKmeansEstimator	Self
typedef Object	Superclass
Public Member Functions
virtual ::itk::LightObject::Pointer	CreateAnother (void) const
virtual double	GetCentroidPositionChanges () const
virtual double	GetCentroidPositionChangesThreshold () const
virtual int	GetCurrentIteration () const
const TKdTree *	GetKdTree () const
virtual MeasurementVectorSizeType	GetMeasurementVectorSize () const
virtual const char *	GetNameOfClass () const
const MembershipFunctionVectorObjectType *	GetOutput () const
virtual bool	GetUseClusterLabels () const
virtual void	SetCentroidPositionChangesThreshold (double _arg)
void	SetKdTree (TKdTree *tree)
virtual void	SetUseClusterLabels (bool _arg)
void	StartOptimization ()

virtual void	SetParameters (ParametersType _arg)
virtual ParametersType	GetParameters () const

virtual void	SetMaximumIteration (int _arg)
virtual int	GetMaximumIteration () const
Static Public Member Functions
static Pointer	New ()
Protected Member Functions
void	CopyParameters (InternalParametersType &source, InternalParametersType &target)
void	CopyParameters (ParametersType &source, InternalParametersType &target)
void	CopyParameters (InternalParametersType &source, ParametersType &target)
void	FillClusterLabels (KdTreeNodeType *node, int closestIndex)
void	Filter (KdTreeNodeType *node, std::vector< int > validIndexes, MeasurementVectorType &lowerBound, MeasurementVectorType &upperBound)
int	GetClosestCandidate (ParameterType &measurements, std::vector< int > &validIndexes)
void	GetPoint (ParameterType &point, MeasurementVectorType measurements)
double	GetSumOfSquaredPositionChanges (InternalParametersType &previous, InternalParametersType &current)
bool	IsFarther (ParameterType &pointA, ParameterType &pointB, MeasurementVectorType &lowerBound, MeasurementVectorType &upperBound)
	KdTreeBasedKmeansEstimator ()
void	PrintPoint (ParameterType &point)
void	PrintSelf (std::ostream &os, Indent indent) const
virtual	~KdTreeBasedKmeansEstimator ()
Private Attributes
CandidateVector	m_CandidateVector
double	m_CentroidPositionChanges
double	m_CentroidPositionChangesThreshold
ClusterLabelsType	m_ClusterLabels
int	m_CurrentIteration
EuclideanDistanceMetric < ParameterType >::Pointer	m_DistanceMetric
bool	m_GenerateClusterLabels
TKdTree::Pointer	m_KdTree
int	m_MaximumIteration
MeasurementVectorSizeType	m_MeasurementVectorSize
MembershipFunctionVectorObjectPointer	m_MembershipFunctionsObject
ParametersType	m_Parameters
ParameterType	m_TempVertex
bool	m_UseClusterLabels

Detailed Description

template<class TKdTree>
class itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >

fast k-means algorithm implementation using k-d tree structure

It returns k mean vectors that are centroids of k-clusters using pre-generated k-d tree. k-d tree generation is done by the WeightedCentroidKdTreeGenerator. The tree construction needs to be done only once. The resulting k-d tree's non-terminal nodes that have their children nodes have vector sums of measurement vectors that belong to the nodes and the number of measurement vectors in addition to the typical node boundary information and pointers to children nodes. Instead of reassigning every measurement vector to the nearest cluster centroid and recalculating centroid, it maintain a set of cluster centroid candidates and using pruning algorithm that utilizes k-d tree, it updates the means of only relevant candidates at each iterations. It would be faster than traditional implementation of k-means algorithm. However, the k-d tree consumes a large amount of memory. The tree construction time and pruning algorithm's performance are important factors to the whole process's performance. If users want to use k-d tree for some purpose other than k-means estimation, they can use the KdTreeGenerator instead of the WeightedCentroidKdTreeGenerator. It will save the tree construction time and memory usage.

Note: There is a second implementation of k-means algorithm in ITK under the While the Kd tree based implementation is more time efficient, the GLA/LBG based algorithm is more memory efficient.

Recent API changes: The static const macro to get the length of a measurement vector, MeasurementVectorSize has been removed to allow the length of a measurement vector to be specified at run time. It is now obtained from the KdTree set as input. You may query this length using the function GetMeasurementVectorSize().

See also:: ImageKmeansModelEstimator; WeightedCentroidKdTreeGenerator, KdTree

Wiki Examples:

All Examples

Compute kmeans clusters

Definition at line 77 of file itkKdTreeBasedKmeansEstimator.h.

Member Typedef Documentation

template<class TKdTree >

typedef KdTreeNodeType::CentroidType itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::CentroidType

Definition at line 99 of file itkKdTreeBasedKmeansEstimator.h.

template<class TKdTree >

typedef itksys::hash_map< InstanceIdentifier, unsigned int > itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::ClusterLabelsType

Definition at line 162 of file itkKdTreeBasedKmeansEstimator.h.

template<class TKdTree >

typedef SmartPointer< const Self > itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::ConstPointer

Reimplemented from itk::Object.

Definition at line 85 of file itkKdTreeBasedKmeansEstimator.h.

template<class TKdTree >

typedef DistanceToCentroidMembershipFunctionType::Pointer itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::DistanceToCentroidMembershipFunctionPointer

Definition at line 116 of file itkKdTreeBasedKmeansEstimator.h.

template<class TKdTree >

typedef DistanceToCentroidMembershipFunction< MeasurementVectorType > itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::DistanceToCentroidMembershipFunctionType

Typedef requried to generate dataobject decorated output that can be plugged into SampleClassifierFilter

Definition at line 113 of file itkKdTreeBasedKmeansEstimator.h.

template<class TKdTree >

typedef TKdTree::InstanceIdentifier itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::InstanceIdentifier

Definition at line 97 of file itkKdTreeBasedKmeansEstimator.h.

template<class TKdTree >

typedef std::vector< ParameterType > itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::InternalParametersType

Definition at line 107 of file itkKdTreeBasedKmeansEstimator.h.

template<class TKdTree >

typedef TKdTree::KdTreeNodeType itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::KdTreeNodeType

Types for the KdTree data structure

Definition at line 91 of file itkKdTreeBasedKmeansEstimator.h.

template<class TKdTree >

typedef TKdTree::MeasurementType itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::MeasurementType

Definition at line 95 of file itkKdTreeBasedKmeansEstimator.h.

template<class TKdTree >

typedef unsigned int itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::MeasurementVectorSizeType

Typedef for the length of a measurement vector

Definition at line 102 of file itkKdTreeBasedKmeansEstimator.h.

template<class TKdTree >

typedef TKdTree::MeasurementVectorType itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::MeasurementVectorType

Definition at line 96 of file itkKdTreeBasedKmeansEstimator.h.

template<class TKdTree >

typedef MembershipFunctionType::ConstPointer itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::MembershipFunctionPointer

Definition at line 119 of file itkKdTreeBasedKmeansEstimator.h.

template<class TKdTree >

typedef MembershipFunctionBase< MeasurementVectorType > itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::MembershipFunctionType

Definition at line 118 of file itkKdTreeBasedKmeansEstimator.h.

template<class TKdTree >

typedef MembershipFunctionVectorObjectType::Pointer itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::MembershipFunctionVectorObjectPointer

Definition at line 124 of file itkKdTreeBasedKmeansEstimator.h.

template<class TKdTree >

typedef SimpleDataObjectDecorator< MembershipFunctionVectorType > itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::MembershipFunctionVectorObjectType

Definition at line 122 of file itkKdTreeBasedKmeansEstimator.h.

template<class TKdTree >

typedef std::vector< MembershipFunctionPointer > itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::MembershipFunctionVectorType

Definition at line 120 of file itkKdTreeBasedKmeansEstimator.h.

template<class TKdTree >

typedef Array< double > itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::ParametersType

Definition at line 108 of file itkKdTreeBasedKmeansEstimator.h.

template<class TKdTree >

typedef Array< double > itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::ParameterType

Parameters type. It defines a position in the optimization search space.

Definition at line 106 of file itkKdTreeBasedKmeansEstimator.h.

template<class TKdTree >

typedef SmartPointer< Self > itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::Pointer

Reimplemented from itk::Object.

Definition at line 84 of file itkKdTreeBasedKmeansEstimator.h.

template<class TKdTree >

typedef TKdTree::SampleType itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::SampleType

Definition at line 98 of file itkKdTreeBasedKmeansEstimator.h.

template<class TKdTree >

typedef KdTreeBasedKmeansEstimator itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::Self

Standard Self typedef.

Reimplemented from itk::Object.

Definition at line 82 of file itkKdTreeBasedKmeansEstimator.h.

template<class TKdTree >

typedef Object itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::Superclass

Reimplemented from itk::Object.

Definition at line 83 of file itkKdTreeBasedKmeansEstimator.h.

Constructor & Destructor Documentation

template<class TKdTree >

itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::KdTreeBasedKmeansEstimator ( ) [protected]

template<class TKdTree >

virtual itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::~KdTreeBasedKmeansEstimator ( ) [inline, protected, virtual]

Definition at line 168 of file itkKdTreeBasedKmeansEstimator.h.

Member Function Documentation

template<class TKdTree >

void itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::CopyParameters	(	InternalParametersType &	source,
		InternalParametersType &	target
	)		`[protected]`

copies the source parameters (k-means) to the target

template<class TKdTree >

void itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::CopyParameters	(	ParametersType &	source,
		InternalParametersType &	target
	)		`[protected]`

copies the source parameters (k-means) to the target

template<class TKdTree >

void itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::CopyParameters	(	InternalParametersType &	source,
		ParametersType &	target
	)		`[protected]`

copies the source parameters (k-means) to the target

template<class TKdTree >

virtual::itk::LightObject::Pointer itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::CreateAnother ( void ) const [virtual]

Create an object from an instance, potentially deferring to a factory. This method allows you to create an instance of an object that is exactly the same type as the referring object. This is useful in cases where an object has been cast back to a base class.

Reimplemented from itk::Object.

template<class TKdTree >

void itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::FillClusterLabels	(	KdTreeNodeType *	node,
		int	closestIndex
	)		`[protected]`

template<class TKdTree >

void itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::Filter	(	KdTreeNodeType *	node,
		std::vector< int >	validIndexes,
		MeasurementVectorType &	lowerBound,
		MeasurementVectorType &	upperBound
	)		`[protected]`

recursive pruning algorithm. the validIndexes vector contains only the indexes of the surviving candidates for the node

template<class TKdTree >

virtual double itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::GetCentroidPositionChanges ( ) const [virtual]

template<class TKdTree >

virtual double itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::GetCentroidPositionChangesThreshold ( ) const [virtual]

template<class TKdTree >

int itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::GetClosestCandidate	(	ParameterType &	measurements,
		std::vector< int > &	validIndexes
	)		`[protected]`

get the index of the closest candidate to the measurements measurement vector

template<class TKdTree >

virtual int itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::GetCurrentIteration ( ) const [virtual]

template<class TKdTree >

const TKdTree* itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::GetKdTree ( ) const

template<class TKdTree >

virtual int itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::GetMaximumIteration ( ) const [virtual]

Set/Get maximum iteration limit.

template<class TKdTree >

virtual MeasurementVectorSizeType itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::GetMeasurementVectorSize ( ) const [virtual]

Get the length of measurement vectors in the KdTree

template<class TKdTree >

virtual const char* itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::GetNameOfClass ( ) const [virtual]

Run-time type information (and related methods).

Reimplemented from itk::Object.

template<class TKdTree >

const MembershipFunctionVectorObjectType* itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::GetOutput ( ) const

Output Membership function vector containing the membership functions with the final optimized parameters

template<class TKdTree >

virtual ParametersType itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::GetParameters ( ) const [virtual]

Set the position to initialize the optimization.

template<class TKdTree >

void itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::GetPoint	(	ParameterType &	point,
		MeasurementVectorType	measurements
	)		`[protected]`

imports the measurements measurement vector data to the point

template<class TKdTree >

double itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::GetSumOfSquaredPositionChanges	(	InternalParametersType &	previous,
		InternalParametersType &	current
	)		`[protected]`

gets the sum of squared difference between the previous position and current postion of all centroid. This is the primary termination condition for this algorithm. If the return value is less than the value that was set by the SetCentroidPositionChangesThreshold method.

template<class TKdTree >

virtual bool itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::GetUseClusterLabels ( ) const [virtual]

template<class TKdTree >

bool itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::IsFarther	(	ParameterType &	pointA,
		ParameterType &	pointB,
		MeasurementVectorType &	lowerBound,
		MeasurementVectorType &	upperBound
	)		`[protected]`

returns true if the pointA is farther than pointB to the boundary

template<class TKdTree >

static Pointer itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::New ( ) [static]

Method for creation through the object factory.

Reimplemented from itk::Object.

template<class TKdTree >

void itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::PrintPoint ( ParameterType & point ) [protected]

template<class TKdTree >

void itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::PrintSelf	(	std::ostream &	os,
		Indent	indent
	)		const `[protected, virtual]`

Methods invoked by Print() to print information about the object including superclasses. Typically not called by the user (use Print() instead) but used in the hierarchical print process to combine the output of several classes.

Reimplemented from itk::Object.

template<class TKdTree >

virtual void itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::SetCentroidPositionChangesThreshold ( double _arg ) [virtual]

Set/Get the termination threshold for the squared sum of changes in centroid postions after one iteration

template<class TKdTree >

void itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::SetKdTree ( TKdTree * tree )

Set/Get the pointer to the KdTree

template<class TKdTree >

virtual void itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::SetMaximumIteration ( int _arg ) [virtual]

Set/Get maximum iteration limit.

template<class TKdTree >

virtual void itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::SetParameters ( ParametersType _arg ) [virtual]

Set the position to initialize the optimization.

template<class TKdTree >

virtual void itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::SetUseClusterLabels ( bool _arg ) [virtual]

template<class TKdTree >

void itk::Statistics::KdTreeBasedKmeansEstimator< TKdTree >::StartOptimization ( )

Start optimization Optimization will stop when it meets either of two termination conditions, the maximum iteration limit or epsilon (minimal changes in squared sum of changes in centroid positions)