lti::guyMedioniSaliency Class Reference

The Guy-Medioni saliency is a procedure for perceptual grouping of low level features like points and edges. More...

#include <ltiGuyMedioniSaliency.h>

Inheritance diagram for lti::guyMedioniSaliency:

[legend]

Collaboration diagram for lti::guyMedioniSaliency:

[legend]

List of all members.

Classes
class	parameters
	the parameters for the class guyMedioniSaliency More...
Public Member Functions
	guyMedioniSaliency ()
	guyMedioniSaliency (const parameters &par)
	guyMedioniSaliency (const guyMedioniSaliency &other)
virtual	~guyMedioniSaliency ()
virtual const char *	getTypeName () const
bool	apply (const channel &edgeRelevance, const channel &orientation, channel &saliencyLambdaX, channel &saliencyLambdaY)
bool	apply (const channel &edgeRelevance, const channel &orientation, channel &saliency)
bool	apply (const channel &edgeRelevance, const channel &orientation, channel &saliency, channel &angle, float &maxSaliency)
const matrix< float > &	getExtensionFieldX () const
const matrix< float > &	getExtensionFieldY () const
const matrix< float > &	getExtensionFieldM () const
guyMedioniSaliency &	copy (const guyMedioniSaliency &other)
guyMedioniSaliency &	operator= (const guyMedioniSaliency &other)
virtual functor *	clone () const
const parameters &	getParameters () const
virtual bool	updateParameters ()
Protected Member Functions
void	accumulate (const channel &edgeRelevance, const channel &orientation, const float threshold)
void	getSaliency (channel &salLambdaMax, channel &salLambdaMin)
void	getSaliency (channel &sal)
void	getSaliency (channel &sal, channel &angle, float &maxSaliency)
Protected Attributes
arctanLUT	atan2
Saliency Accumulators
Six accumulators are necessary to compute the saliency
matrix< float >	sumx2
matrix< float >	sumy2
matrix< float >	sumxy
Extension Field

matrix< float >	extensionFieldX
matrix< float >	extensionFieldY
matrix< float >	extensionFieldM
bilinearInterpolator< float >	ifieldX
bilinearInterpolator< float >	ifieldY
point	extFieldCenter
int	extFieldRadius
imatrix	extFieldCircle
float	fieldThreshold
bool	generateExtensionField ()
void	vote (const point p, const float angle, const float weight)

Detailed Description

The Guy-Medioni saliency is a procedure for perceptual grouping of low level features like points and edges.

For this task it computes first a extension field, an based on it a saliency map, where only perceptual relevant features survive, or illusory ones appear.

As input it expects the output of an orientation map, with a relevance channel and an orientation channel. The relevance can be replaced by an edge map. With a threshold-parameter you can decide how much relevant a pixel must be in order to be considered in the saliency map.

The original theory can be found in:

G. Guy and G. Medioni. "Inferring global perceptual contours from local features" International Journal of Computer Vision, Vol.20, No.1/2, 1996, pp. 113-133

http://citeseer.nj.nec.com/guy96inferring.html

Example:

 lti::image img;
 lti::channel chnl,rel,ori;
 lti::channel sal,sala;
 lti::channel8 rel8;
 float maxSal;

 // load an image (suppose you have one in the directory ../img/ called
 //                image1.bmp)
 lti::loadImage loader;
 loader.load("../img/image1.bmp",img);
 chnl.castFrom(img); // convert the color image into a gray valued one

 // get the edges and orientation channels
 lti::cannyEdges canny;
 canny.apply(chnl,rel8,ori);
 rel.castFrom(rel8);

 // get the Guy-Medioni saliency (sal) and the orientation (sala)
 // and the maximal value found in the sal channel (maxSal).
 lti::guyMedioniSaliency gmSaliency;
 gmSaliency.apply(rel,ori,sal,sala,maxSal);

The saliency channel is somehow blurred. You can use the lti::nonMaximaSuppression to get again an edge channel from the saliency:

 lti::channel8 newEdges;
 lti::nonMaximaSuppression suppressor;
 suppressor.apply(sal,sala,newEdges,maxSal);

Constructor & Destructor Documentation

lti::guyMedioniSaliency::guyMedioniSaliency ( )

default constructor

lti::guyMedioniSaliency::guyMedioniSaliency ( const parameters & par )

Construct a functor using the given parameters.

lti::guyMedioniSaliency::guyMedioniSaliency ( const guyMedioniSaliency & other )

copy constructor

Parameters:

other

the object to be copied

virtual lti::guyMedioniSaliency::~guyMedioniSaliency ( ) [virtual]

destructor

Member Function Documentation

void lti::guyMedioniSaliency::accumulate	(	const channel &	edgeRelevance,
		const channel &	orientation,
		const float	threshold
	)			`[protected]`

compute the first step of the saliency, accumulating the fields of each edge-element.

bool lti::guyMedioniSaliency::apply	(	const channel &	edgeRelevance,
		const channel &	orientation,
		channel &	saliency,
		channel &	angle,
		float &	maxSaliency
	)

Generates the enhanced saliency measure.

This method is more efficient than to compute both lambda channels with the alternative apply() method and to subtract them.

Parameters:

	edgeRelevance	Relevance for each pixel. It can be an edge image.
	orientation	Orientation of each pixel. Please note that the orientation channel expected by this functor contains the angle perpendicular to the edgels as returned by the gradient or orientationMap functors.
	saliency	resulting saliency channel.
	angle	resulting direction for each pixel
	maxSaliency	maximum saliency value.

Returns:: true if apply successful or false otherwise.

Warning:: The orientation channel expected differs from the one returned by orientation map functors. It must represent the real direction of the edgels instead of the direction perpendicular to them.

bool lti::guyMedioniSaliency::apply	(	const channel &	edgeRelevance,
		const channel &	orientation,
		channel &	saliency
	)

Generates the enhanced saliency measure.

This method is more efficient than to compute both lambda channels with the alternative apply() method and to subtract them.

Parameters:

	edgeRelevance	Relevance for each pixel. It can be an edge image.
	orientation	Orientation of each pixel. Please note that the orientation channel expected by this functor contains the angle perpendicular to the edgels as returned by the gradient or orientationMap functors.
	saliency	resulting saliency channel.

Returns:: true if apply successful or false otherwise.

Warning:: The orientation channel expected differs from the one returned by orientation map functors. It must represent the real direction of the edgels instead of the direction perpendicular to them.

bool lti::guyMedioniSaliency::apply	(	const channel &	edgeRelevance,
		const channel &	orientation,
		channel &	saliencyLambdaX,
		channel &	saliencyLambdaY
	)

Generates the raw saliency and junction saliency maps using Guy-Medioni's approach.

To obtain the final saliency it is more efficient to use the alternative apply() method than to subtract both lambda channels.

Parameters:

	edgeRelevance	Relevance for each pixel. It can be an edge image.
	orientation	Orientation of each pixel. Please note that the orientation channel expected by this functor contains the angle perpendicular to the edgels as returned by the gradient or orientationMap functors.
	saliencyLambdaX	resulting major eigenvalue: raw saliency.
	saliencyLambdaY	resulting minor eigenvalues: junction saliency.