Classes | Public Member Functions | Static Public Member Functions

GDALSimpleSURF Class Reference

Class for searching corresponding points on images. More...

#include <gdal_simplesurf.h>

List of all members.

Classes

class  MatchedPointPairInfo
 Class stores indexes of pair of point and distance between them.

Public Member Functions

 GDALSimpleSURF (int nOctaveStart, int nOctaveEnd)
 Prepare class according to specified parameters.
std::vector< GDALFeaturePoint > * ExtractFeaturePoints (GDALIntegralImage *poImg, double dfThreshold)
 Find feature points using specified integral image.

Static Public Member Functions

static CPLErr ConvertRGBToLuminosity (GDALRasterBand *red, GDALRasterBand *green, GDALRasterBand *blue, int nXSize, int nYSize, double **padfImg, int nHeight, int nWidth)
 Convert image with RGB channels to grayscale using "luminosity" method.
static CPLErr MatchFeaturePoints (std::vector< GDALFeaturePoint * > *poMatchPairs, std::vector< GDALFeaturePoint > *poFirstCollect, std::vector< GDALFeaturePoint > *poSecondCollect, double dfThreshold)
 Find corresponding points (equal points in two collections).

Detailed Description

Class for searching corresponding points on images.

Author:
Andrew Migal migal.drew@gmail.com

Provides capability for detection feature points and finding equal points on different images. Class implements simplified version of SURF algorithm (Speeded Up Robust Features). As original, this realization is scale invariant, but sensitive to rotation. Images should have similar rotation angles (maximum difference is up to 10-15 degrees), otherwise algorithm produces incorrect and very unstable results.


Constructor & Destructor Documentation

GDALSimpleSURF::GDALSimpleSURF ( int  nOctaveStart,
int  nOctaveEnd 
)

Prepare class according to specified parameters.

Octave numbers affects to amount of detected points and their robustness. Range between bottom and top octaves also affects to required time of detection points (if range is large, algorithm should perform more operations).

Parameters:
nOctaveStart Number of bottom octave. Octave numbers starts with one
nOctaveEnd Number of top octave. Should be equal or greater than OctaveStart
Note:
Every octave finds points with specific size. For small images use small octave numbers, for high resolution - large. For 1024x1024 images it's normal to use any octave numbers from range 1-6. (for example, octave start - 1, octave end - 3, or octave start - 2, octave end - 2.) For larger images, try 1-10 range or even higher. Pay attention that number of detected point decreases quickly per octave for particular image. Algorithm finds more points in case of small octave numbers. If method detects nothing, reduce bottom bound of octave range.

NOTICE that every octave requires time to compute. Use a little range or only one octave if execution time is significant.


Member Function Documentation

CPLErr GDALSimpleSURF::ConvertRGBToLuminosity ( GDALRasterBand red,
GDALRasterBand green,
GDALRasterBand blue,
int  nXSize,
int  nYSize,
double **  padfImg,
int  nHeight,
int  nWidth 
) [static]

Convert image with RGB channels to grayscale using "luminosity" method.

Result is used in SURF-based algorithm, but may be used anywhere where grayscale images with nice contrast are required.

Parameters:
red Image's red channel
green Image's green channel
blue Image's blue channel
nXSize Width of initial image
nYSize Height of initial image
padfImg Array for resulting grayscale image
nHeight Height of resulting image
nWidth Width of resulting image
Returns:
CE_None or CE_Failure if error occurs.
std::vector< GDALFeaturePoint > * GDALSimpleSURF::ExtractFeaturePoints ( GDALIntegralImage poImg,
double  dfThreshold 
)

Find feature points using specified integral image.

Parameters:
poImg Integral image to be used
dfThreshold Threshold for feature point recognition. Detected feature point will have Hessian value greater than this provided threshold.
Note:
Typical threshold's value is 0,001. But this value can be various in each case and depends on image's nature. For example, value can be 0.002 or 0.005. Fill free to experiment with it. If threshold is high, than number of detected feature points is small, and vice versa.
CPLErr GDALSimpleSURF::MatchFeaturePoints ( std::vector< GDALFeaturePoint * > *  poMatchPairs,
std::vector< GDALFeaturePoint > *  poFirstCollect,
std::vector< GDALFeaturePoint > *  poSecondCollect,
double  dfThreshold 
) [static]

Find corresponding points (equal points in two collections).

Parameters:
poMatchPairs Resulting collection for matched points
poSecondCollect Points on the first image
poSecondCollect Points on the second image
dfThreshold Value from 0 to 1. Threshold affects to number of matched points. If threshold is lower, amount of corresponding points is larger, and vice versa
Returns:
CE_None or CE_Failure if error occurs.
Parameters:
poMatched Resulting collection for matched points
poFirstCollection Points on the first image
poSecondCollection Points on the second image
dfThreshold Value from 0 to 1. Threshold affects to number of matched points. If threshold is higher, amount of corresponding points is larger, and vice versa
Note:
Typical threshold's value is 0,1. BUT it's a very approximate guide. It can be 0,001 or even 1. This threshold provides direct adjustment of point matching. NOTICE that if threshold is lower, matches are more robust and correct, but number of matched points is smaller. Therefore if algorithm performs many false detections and produces bad results, reduce threshold. Otherwise, if algorithm finds nothing, increase threshold.
Returns:
CE_None or CE_Failure if error occurs.

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

Generated for GDAL by doxygen 1.7.1.