Contrast-Sensitive Potts training class. More...

#include <TrainEdgePottsCS.h>

Inheritance diagram for DirectGraphicalModels::CTrainEdgePottsCS:

[legend]

Collaboration diagram for DirectGraphicalModels::CTrainEdgePottsCS:

[legend]

Public Member Functions
	CTrainEdgePottsCS (byte nStates, word nFeatures, ePotPenalApproach penApproach=eP_APP_PEN_EXP)
	Constructor. More...

virtual	~CTrainEdgePottsCS (void)

Public Member Functions inherited from DirectGraphicalModels::CTrainEdgePotts
	CTrainEdgePotts (byte nStates, word nFeatures)
	Constructor. More...

virtual	~CTrainEdgePotts (void)

virtual void	reset (void)
	Resets class variables. More...

virtual void	addFeatureVecs (const Mat &featureVector1, byte gt1, const Mat &featureVector2, byte gt2)
	Adds a pair of feature vectors. More...

Public Member Functions inherited from DirectGraphicalModels::CTrainEdge
	CTrainEdge (byte nStates, word nFeatures)
	Constructor. More...

virtual	~CTrainEdge (void)=default

virtual void	train (bool doClean=false)
	Random model training. More...

Mat	getEdgePotentials (const Mat &featureVector1, const Mat &featureVector2, const vec_float_t &vParams, float weight=1.0f) const
	Returns the edge potential, based on the feature vectors. More...

Public Member Functions inherited from DirectGraphicalModels::ITrain
	ITrain (byte nStates, word nFeatures)
	Constructor. More...

virtual	~ITrain (void)=default

word	getNumFeatures (void) const
	Returns number of features. More...

Public Member Functions inherited from DirectGraphicalModels::CBaseRandomModel
	CBaseRandomModel (byte nStates)
	Constructor. More...

virtual	~CBaseRandomModel (void)

virtual void	save (const std::string &path, const std::string &name=std::string(), short idx=-1) const
	Saves the training data. More...

virtual void	load (const std::string &path, const std::string &name=std::string(), short idx=-1)
	Loads the training data. More...

byte	getNumStates (void) const
	Returns number of states (classes) More...

Protected Member Functions
virtual Mat	calculateEdgePotentials (const Mat &featureVector1, const Mat &featureVector2, const vec_float_t &vParams) const
	Returns the contrast-sensitive edge potentials. More...

Protected Member Functions inherited from DirectGraphicalModels::CTrainEdgePotts
virtual void	saveFile (FILE *pFile) const
	Saves the random model into the file. More...

virtual void	loadFile (FILE *pFile)
	Loads the random model from the file. More...

Protected Member Functions inherited from DirectGraphicalModels::CBaseRandomModel
std::string	generateFileName (const std::string &path, const std::string &name, short idx) const
	Generates name of the data file for storing random model parameters. More...

Private Attributes
ePotPenalApproach	m_penApproach

Additional Inherited Members
Static Public Member Functions inherited from DirectGraphicalModels::CTrainEdge
static std::shared_ptr< CTrainEdge >	create (byte edgeRandomModel, byte nStates, word nFeatures)
	Factory method returning edge trainer object. More...

static Mat	getDefaultEdgePotentials (float val, byte nStates)
	Returns the data-independent edge potentials. More...

static Mat	getDefaultEdgePotentials (const vec_float_t &values)
	Returns the data-independent edge potentials. More...

Protected Attributes inherited from DirectGraphicalModels::CBaseRandomModel
byte	m_nStates
	The number of states (classes) More...

Detailed Description

Contrast-Sensitive Potts training class.

This class improves the CTrainEdgePotts class by considering the contrast in test-data between edge's nodes. In other respects this class is still training-data-independent, and thus may be applied without edge training procedure, i.e. functions reset(), save(), load(), addFeatureVecs() are unnecessary here.

This class may be used in case when the training-data for edges is not available.

Author: Sergey G. Kosov, serge.nosp@m.y.ko.nosp@m.sov@p.nosp@m.roje.nosp@m.ct-10.nosp@m..de

Definition at line 35 of file TrainEdgePottsCS.h.

Constructor & Destructor Documentation

◆ CTrainEdgePottsCS()

DirectGraphicalModels::CTrainEdgePottsCS::CTrainEdgePottsCS	(	byte	nStates,
		word	nFeatures,
		ePotPenalApproach	penApproach = `eP_APP_PEN_EXP`
	)

inline

Constructor.

Parameters

nStates	Number of states (classes)
nFeatures	Number of features
penApproach	Flag specifying the penalization approach for the edge potential matrix (Ref. ePotPenalApproach)

Definition at line 44 of file TrainEdgePottsCS.h.

◆ ~CTrainEdgePottsCS()

virtual DirectGraphicalModels::CTrainEdgePottsCS::~CTrainEdgePottsCS ( void )

inlinevirtual

Definition at line 49 of file TrainEdgePottsCS.h.

Member Function Documentation

◆ calculateEdgePotentials()

Mat DirectGraphicalModels::CTrainEdgePottsCS::calculateEdgePotentials	(	const Mat &	featureVector1,
		const Mat &	featureVector2,
		const vec_float_t &	vParams
	)		const

protectedvirtual

Returns the contrast-sensitive edge potentials.

This function returns matrix with diagonal elements equal to parameter \(\vec{\theta}\) provided through argument params and penalized with a penalization function \(\mathcal{P}\), which depends on the difference between arguments featureVector1 and featureVector2; all the other elements are 1's:

\[edgePot[nStates][nStates] = \begin{bmatrix} \theta_1\mathcal{P} & 1 & \cdots & 1 \\ 1 & \theta_2\mathcal{P} & \cdots & 1 \\ \vdots & \vdots & \ddots & \vdots \\ 1 & 1 & \cdots & \theta_{nStates}\mathcal{P} \end{bmatrix}, \]

where \(\mathcal{P}\equiv\mathcal{P}(d;\,\lambda)\) is the penalization function, where \(d = ||\textbf{f}_1 - \textbf{f}_2||_2\) is the Euclidean distance between feature vectors,
and \(\lambda\) is a parameter provided through argument params. For more details on penalization fuction, please refere to ePotPenalApproach.

Parameters

featureVector1	Multi-dimensinal point \(\textbf{f}_1\): Mat(size: nFeatures x 1; type: CV_{XX}C1), corresponding to the first node of the edge
featureVector2	Multi-dimensinal point \(\textbf{f}_2\): Mat(size: nFeatures x 1; type: CV_{XX}C1), corresponding to the second node of the edge
vParams	Array of control parameters \(\{\vec{\theta},\lambda\} \). \(\vec{\theta}\) may consist either from one parameter (in this case all the diagonal elemets will be the same), or from nStates parameters, specifying smoothness strength for each state (class) individually; \(\lambda\) consists from one parameter.

Returns: The edge potential matrix: Mat(size: nStates x nStates; type: CV_32FC1)

If featureVector1 or featureVector2 is empty, the function returns the test-data-independent Potts potential: CTrainEdgePotts::calculateEdgePotentials()

Reimplemented from DirectGraphicalModels::CTrainEdgePotts.

Reimplemented in DirectGraphicalModels::CTrainEdgePrior.

Definition at line 52 of file TrainEdgePottsCS.cpp.

Here is the call graph for this function: