Machine Perception Primitive: An implementation of the "Fast Salience Using Natural-statistics" algorithm from Butko, et al., 2008. FastSUN is an efficient implementation of Zhang et al.'s SUN algorithm, which is documented in Zhang, et al., 2008 (see Related Publications). More...

#include <FastSalience.h>

Collaboration diagram for FastSalience:

[legend]

List of all members.

Public Member Functions
	FastSalience (int numtemporal, int numspatial, float firsttau=1.0, int firstrad=0)
	Detailed Constructor.
	FastSalience ()
	Default Constructor: construct a FastSalience feature detector with default parameters.
virtual	~FastSalience ()
	Simple Destructor.
void	updateSalience (const cv::Mat &image)
	Compute the salience map for a new input frame.
void	getSalMap (cv::Mat &dest) const
	The current salience map -- negative log likelihood values.
void	getSalImage (cv::Mat &dest) const
	The current salience iamge, normalized to the range 0-1 for display.
void	setGGDistributionPower (double value)
	Set the Generalized-Gaussian feature distribution to be something other than Laplacian.
void	setUseDoEFeatures (int flag)
	Toggle use of Difference of Exponential (motion) features. On by default.
void	setUseDoBFeatures (int flag)
	Toggle use of Difference of Box (spatial) features. On by default.
void	setUseColorInformation (int flag)
	Toggle use of Color Contrast (red-green and blue-yellow) information. On by default.
void	setUseGGDistributionParams (int flag)
	Toggle use of custom Generalized Gaussian Distribution parameters. On by default.
void	setEstimateGGDistributionParams (int flag)
	Toggle estimation of custom Generalized Gaussian Distribution parameters. On by default.
std::vector< cv::KeyPoint >	getKeyPoints (int radius=2) const
	Find key-point interest detectors using non-maximal suppression on the salience map.

Detailed Description

Machine Perception Primitive: An implementation of the "Fast Salience Using Natural-statistics" algorithm from Butko, et al., 2008. FastSUN is an efficient implementation of Zhang et al.'s SUN algorithm, which is documented in Zhang, et al., 2008 (see Related Publications).

Author:: Nicholas Butko

Date:: 2010 version 0.4

Constructor & Destructor Documentation

FastSalience::FastSalience	(	int	numtemporal,
		int	numspatial,
		float	firsttau = `1.0`,
		int	firstrad = `0`
	)

Detailed Constructor.

Parameters:

salwidth	The width of the input images supplied to the algorithm. Note: No image scaling is done by the FastSalience algorithm, so if you want a salience map of a downscaled image, you must scale the image down to `salwidth` before calling updateSalience.
salheight	The height of the input images supplied to the algorithm. Note: No image scaling is done by the FastSalience algorithm, so if you want a salience map of a downscaled image, you must scale the image down to `salheight` before calling updateSalience.
numtemporal	Number of timescales of Difference of Expontential filters to track.
numspatial	Number of sizes of Difference of Box filters to use.
firsttau	Exponential Falloff parameter for the first Difference of Exponentials scale. Lower numbers give slower falloff. Must be greater than 0.
firstrad	Radius of smallest Difference of Boxes filter center. The diameter of the box is 2*rad+1, so the smallest allowed first radius is 0.

FastSalience::~FastSalience ( ) [virtual]

Simple Destructor.

Deallocates all memory associated with the FastSalience object.

Member Function Documentation

vector< KeyPoint > FastSalience::getKeyPoints ( int radius = 2 ) const

Find key-point interest detectors using non-maximal suppression on the salience map.

Parameters:

radius Non-maximal suppression radius.

void FastSalience::setEstimateGGDistributionParams ( int flag )

Toggle estimation of custom Generalized Gaussian Distribution parameters. On by default.

Initially we assume that image features are drawn from a Laplace distribution of zero mean and unit variance. However, these parameters can be estimated and adjusted online to better reflect the statistics of the current environment. If off, the current estimate of the mean and variance are used. If on, the estimate updates with each new frame (at small computational cost). One approach would be to estimate the parameters for a few frames, and then when the estimates no longer change much, to turn off estimation and just use the "learned" parameters.

Parameters:

flag	If 0, turn off estimation of histogram information. Otherwise, calculate it.

void FastSalience::setGGDistributionPower ( double value )

Set the Generalized-Gaussian feature distribution to be something other than Laplacian.

Image features often exhibit marginal histograms that are of a generalized Gaussian form. The power of these distributions is often between .5 and .7, but performing this power computation is numerically slow. By using a Laplacian distribution (power of one) by defuault, the salience estimate is worse, but much faster to compute. Setting "power" to some lower value may give better but slower resultes.

Parameters:

value The power of a generalized-gaussian feature distribution.

void FastSalience::setUseColorInformation ( int flag )

Toggle use of Color Contrast (red-green and blue-yellow) information. On by default.

Parameters:

flag	If 0, turn off Color Contrast information. Otherwise, calculate it.

void FastSalience::setUseDoBFeatures ( int flag )

Toggle use of Difference of Box (spatial) features. On by default.

Parameters:

flag	If 0, turn off DoB features. Otherwise, calculate them.

void FastSalience::setUseDoEFeatures ( int flag )

Toggle use of Difference of Exponential (motion) features. On by default.

Parameters:

flag	If 0, turn off DoE features. Otherwise, calculate them.

void FastSalience::setUseGGDistributionParams ( int flag )

Toggle use of custom Generalized Gaussian Distribution parameters. On by default.

If this is off, we assume that image features are drawn from a Laplace distribution of zero mean and unit variance. However, these parameters can be estimated and adjusted online to better reflect the statistics of the current environment. If on, these online estimates will be used (at small computational cost).

Parameters:

flag	If 0, turn off the use of histogram estimate information. Otherwise, use it.

void FastSalience::updateSalience ( const cv::Mat & image )

Compute the salience map for a new input frame.

Parameters:

colorframe A BGR Color image with width=salwidth and height=salheight. The input image type should be CV_8U3, CV_8U1, CV_32FC1, or CV_32FC3. The salience map that is computed is accessible via getSalImageDouble and getSalImageFloat

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

/Users/nick/projects/NickThesis/Code/OpenCV/src/FastSalience.h
/Users/nick/projects/NickThesis/Code/OpenCV/src/FastSalience.cpp

FastSalience Class Reference [Machine Perception Primitives]

Public Member Functions

Detailed Description

Constructor & Destructor Documentation

Member Function Documentation

FastSalience Class Reference
[Machine Perception Primitives]