drwnPixelSegModel Class Reference

Implements a pixel-level CRF model for multi-class image segmentation (pixel labeling). More...

Inheritance diagram for drwnPixelSegModel:

Public Member Functions
	drwnPixelSegModel ()
	default constructor
	drwnPixelSegModel (const drwnPixelSegModel &model)
	copy constructor
const char *	type () const
	returns object type as a string (e.g., Foo::type() { return "Foo"; })
virtual void	clear ()
	clear learned model parameters
virtual bool	save (drwnXMLNode &xml) const
	write object to XML node (see also write)
virtual bool	load (drwnXMLNode &xml)
	read object from XML node (see also read)
int	numLabels () const
	return the number of labels that this model has been trained to recognize
void	setFeatureGenerator (const drwnSegImagePixelFeatures &featureGenerator)
	set the object used to generate pixel features from the image instance
const drwnSegImagePixelFeatures &	getFeatureGenerator () const
	return the object used to generate pixel features from the image instance
void	learnTrainingClassWeights (const vector< string > &baseNames)
	learn the training class weights by counting class occurrance
void	learnBoostedPixelModels (const vector< string > &baseNames, int subSample=0)
	learn boosted classifiers for predicting classes from pixel features
void	learnPixelUnaryModel (const vector< string > &baseNames, int subSample=0)
	learn unary potentials for calibrating pixel predictions
void	learnPixelContrastWeight (const vector< string > &baseNames)
	learn the weight of the pairwise constrast term
void	learnPixelContrastWeight (double weight)
	set the weight of the pairwise contrast term
void	learnLongRangePairwiseWeight (const vector< string > &baseNames, double threshold, unsigned radius=4)
	learn the weight of the long range pairwise term
void	learnLongRangePairwiseWeight (double weight, double threshold, unsigned radius=4)
	set the weight of the long range pairwise term
void	learnRobustPottsWeight (const vector< string > &baseNames)
	learn the weight of the robust potts consistency term
void	learnRobustPottsWeight (double weight)
	set the weight of the robust potts consistency term
void	learnPixelContrastAndRobustPottsWeights (const vector< string > &baseNames)
	learn the weights of the pairwise contrast and robust potts consistency terms jointly
void	learnPixelContrastAndLongRangeWeights (const vector< string > &baseNames, double threshold, unsigned radius=4)
	learn the contrast and long-range weights jointly
double	getPairwiseContrastWeight () const
	get the weight of the pairwise contrast term
double	getLongRangePairwiseWeight () const
	get the weight of the long range pairwise term
double	getRobustPottsWeight () const
	get the weight of the robust potts term
void	cacheUnaryPotentials (drwnSegImageInstance *instance) const
	cache the unary potentials inside the instance object
void	cacheLongRangeEdges (drwnSegImageInstance *instance) const
	cache the long-range edges inside the instance object
double	inferPixelLabels (drwnSegImageInstance *instance) const
	infer the pixel labels using learned model parameters and store the predicted labels inside the instance object
virtual double	inferPixelLabels (const vector< string > &baseNames, vector< MatrixXi > &predictedLabels) const
	infer pixel labels for a set of images and return predicted labels in predictedLabels
virtual double	energy (drwnSegImageInstance *instance) const
	return the energy of a given labeling according to the model parameters
Public Member Functions inherited from drwnWriteable
bool	write (const char *filename) const
	write object to file (calls save)
bool	read (const char *filename)
	read object from file (calls load)
void	dump () const
	print object's current state to standard output (for debugging)

Protected Member Functions
void	computeBoostedResponses (const vector< double > &x, vector< double > &y) const
	compute boosted pixel features y as a function of raw features x
void	cacheBoostedPixelResponses (drwnSegImageInstance &instance) const
	cache all boosted pixel features in instance More...
void	buildSampledTrainingSet (const vector< string > &baseNames, const char *labelExt, int nLabels, vector< vector< double > > &featureVectors, vector< int > &featureLabels, int subSample, bool bRawFeatures) const
	dataset creation for training
void	crossValidateWeights (const vector< string > &baseNames, const vector< double > &pairwiseContrastValues, const vector< double > &robustPottsValues, const vector< double > &longRangeValues)
	cross-validate pairwise contrast and robust potts weights

Protected Attributes
drwnSegImagePixelFeatures *	_featureGenerator
	pixel feature generator
vector< double >	_classTrainingWeights
	class training weights (normalized to sum to one)
vector< drwnBoostedClassifier * >	_pixelClassModels
	boosted pixel classifiers for generating pixel features (if boosted pixel classifiers are not learned then the multi-class logistic is run on the raw features)
drwnFeatureWhitener	_pixelFeatureWhitener
	feature whitener for unary model
drwnTMultiClassLogistic< drwnBiasJointFeatureMap >	_pixelUnaryModel
	pixelwise unary model
double	_pixelContrastWeight
	weight for pairwise constrast-dependent smoothness term
double	_longRangeEdgeWeight
	weight for (long range) auxiliary edges
unsigned	_longRangeMatchRadius
	radius for long range edge match
double	_longRangeEdgeThreshold
	percentage of long range edges to include
double	_robustPottsWeight
	weight for robust potts consistency term

Detailed Description

Implements a pixel-level CRF model for multi-class image segmentation (pixel labeling).

The model supports a contrast-sensitive pairwise smoothness prior, long range pairwise terms determined from patch matching, and a higher-order robust generalized potts consistency prior. The pairwise smoothness prior is defined over an 8-connected neighbourhood for each pixel. The robust potts prior is implemented on the superpixels defined in each drwnSegImageInstance.

See also

drwnSegImageInstance

Multi-class Image Segmentation

Member Function Documentation

cacheBoostedPixelResponses()

void drwnPixelSegModel::cacheBoostedPixelResponses ( drwnSegImageInstance &  instance ) const

protected

cache all boosted pixel features in instance

Todo:

refactor using persistent storage class

---

The documentation for this class was generated from the following files:

• drwnPixelSegModel.h
• drwnPixelSegModel.cpp