drwnFilterBankResponse Class Reference

Holds the results of running an image through a bank of filters and allows for computation of features over rectangular regions. More...

Public Member Functions
	drwnFilterBankResponse ()
	default constructor
	drwnFilterBankResponse (const drwnFilterBankResponse &f)
	copy constructor
void	clear ()
	clears the filter bank responses (and releases memory)
bool	empty () const
	returns true if their are no filter responses
int	size () const
	returns the number of responses from the filter bank
int	width () const
	returns the width of each filter response (i.e., image width)
int	height () const
	returns the height of each filter response (i.e., image height)
size_t	memory () const
	returns number of bytes stored
void	addResponseImage (cv::Mat &r)
	add a filter response image to the filterbank (takes ownership)
void	addResponseImages (vector< cv::Mat > &r)
	add a number of filter response images to the filterbank (takes ownership)
void	copyResponseImage (const cv::Mat &r)
	copies a filter response image to the filterbank (called retains ownership)
void	copyResponseImages (const vector< cv::Mat > &r)
	copies a number of filter response images to the filterbank (called retains ownership)
const cv::Mat &	getResponseImage (int i) const
	return the i-th filter response image
void	deleteResponseImage (int i)
	delete the i-th filter response image (all responses above i are renumbered)
void	exponentiateResponses (double alpha=1.0)
	transform responses by exponentiation (useful, for example, when approximating max over a region, via log-sum-exp)
void	normalizeResponses ()
	transform responses by pixelwise normalization (e.g., following an exponentiation)
void	expAndNormalizeResponses (double alpha=1.0)
	transform responses by pixelwise exponentiation and normalization. More numerically stable than exponentiateResponses followed by normalizeResponses.
VectorXd	value (int x, int y) const
	value of each filter at given pixel
VectorXd	mean (int x, int y, int w, int h) const
	mean of each filter in rectangular region <x, y, x+w, y+h>
VectorXd	energy (int x, int y, int w, int h) const
	sum of squared values for each filter in rectangular region <x, y, x+w, y+h>
VectorXd	variance (int x, int y, int w, int h) const
	variance of each filter in rectangular region <x, y, x+w, y+h>
VectorXd	mean (const list< cv::Point > &pixels) const
	mean of each filter over given pixels
VectorXd	energy (const list< cv::Point > &pixels) const
	energy of each filter over given pixels
VectorXd	variance (const list< cv::Point > &pixels) const
	variance of each filter over given pixels
VectorXd	mean (const cv::Mat &mask) const
	mean of each filter over masked pixels
VectorXd	energy (const cv::Mat &mask) const
	energy of each filter over masked pixels
VectorXd	variance (const cv::Mat &mask) const
	variance of each filter over masked pixels
VectorXd	mean () const
	mean of each filter over entire image
VectorXd	energy () const
	energy of each filter over entire image
VectorXd	variance () const
	variance of each filter over entire image
cv::Mat	visualize () const
	visualize the feature responses

Protected Attributes
vector< cv::Mat >	_responses
	filter responses
vector< cv::Mat >	_sum
	sum of filter responses (integral image)
vector< cv::Mat >	_sqSum
	sum of filter responses squared

Detailed Description

Holds the results of running an image through a bank of filters and allows for computation of features over rectangular regions.

The difference between addResponseImage and copyResponseImage functions is that the former does not clone the images — the calling function must not modify the image since this will corrupt data internal to drwnFilterBankResponse. Response images should be 32-bit floating point.

The class uses integral images to allow quick computation of sums over rectangular regions. For example, the following code snippet computes the mean image intensity over a few small rectangular patches.

// load the image and convert to 32-bit greyscale
cv::Mat img = cv::imread( ... );
cv::Mat grey = drwnGreyImage(img);

// create filterbank response object and add greyscale image
drwnFilterBankResponse filterbank;
filterbank.addResponseImage(grey);

// compute average intensity over random rectangular regions
for (int i = 0; i < 100; i++) {
    const int x = (int)(drand48() * filterbank.width());
    const int y = (int)(drand48() * filterbank.height());
    const int w = (int)(drand48() * (filterbank.width() - x)) + 1;
    const int h = (int)(drand48() * (filterbank.height() - y)) + 1;

    double meanIntensity = filterbank.mean(x, y, w, h)[0];
    DRWN_LOG_VERBOSE("mean intensity over " << w << "-by-" << h << " patch at ("
      << x << ", " << y << ") is " << meanIntensity);
}

---

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

• drwnFilterBankResponse.h
• drwnFilterBankResponse.cpp