www.pudn.com > OPENCV_SIFT_VC6.rar > ScaleSpace.cpp
// ScaleSpace.cpp: implementation of the CScaleSpace class.
//
//////////////////////////////////////////////////////////////////////
#include "stdafx.h"
#include "TestSIFT.h"
#include "ScaleSpace.h"
#ifdef _DEBUG
#undef THIS_FILE
static char THIS_FILE[]=__FILE__;
#define new DEBUG_NEW
#endif
//////////////////////////////////////////////////////////////////////
// Construction/Destruction
//////////////////////////////////////////////////////////////////////
// Variable s is from explanation in Lowe paper
CScaleSpace::CScaleSpace(int s,double basePixScale)
{
m_scaleLevel = s;
m_dogCount = m_scaleLevel+2;
m_gaussCount = m_scaleLevel+3;
m_gaussImgs = NULL;
m_dogImgs = NULL;
m_magnitudes = NULL;
m_directions = NULL;
m_basePixScale = basePixScale;
}
CScaleSpace::~CScaleSpace()
{
if (m_dogImgs != NULL)
{
for (int i = 0; i < m_dogCount; i++)
cvReleaseImage (&m_dogImgs[i]);
delete m_dogImgs;
}
if (m_gaussImgs != NULL)
{
for (int i = 0; i < m_gaussCount; i++)
cvReleaseImage (&m_gaussImgs[i]);
delete m_gaussImgs;
}
}
void CScaleSpace::buildDoG(IplImage *img,double startsigma,double firstScale)
{
m_gaussImgs = new IplImage*[m_gaussCount];
for (int i = 0; i < m_gaussCount; i++)
m_gaussImgs[i] = cvCreateImage (cvGetSize(img),IPL_DEPTH_32F,1);
if (img->depth == IPL_DEPTH_32F)
cvCopy(img,m_gaussImgs[0]);
else
{
for (int y = 0; y < cvGetSize(img).height; y++)
for (int x = 0; x < cvGetSize(img).width; x++)
cvSet2D (m_gaussImgs[0],y,x,cvGet2D(img,y,x));
}
m_dogImgs = new IplImage*[m_dogCount];
for (i = 0; i < m_dogCount; i++)
m_dogImgs[i] = cvCreateImage (cvGetSize(img),IPL_DEPTH_32F,1);
/* This explaination is from Sebastian Nowozin's C# code
*
* Gaussian G(sigma), with relation
* G(sigma_1) * G(sigma_2) = G(sqrt(sigma_1^2 + * sigma_2^2))
*
* Then, we have:
*
* G(k^{p+1}) = G(k^p) * G(sigma),
* and our goal is to compute every iterations sigma value so this
* equation iteratively produces the next level. Hence:
*
* sigma = \sqrt{\left(k^{p+1}\right)^2 - \left(k^p\right)^2}
* = \sqrt{k^{2p+2} - k^{2p}}
* = \sqrt{k^2p (k^2 - 1)}
* = k^p \sqrt{k^2 - 1}
*
* In the code below, 'w' is the running k^p, where p increases by one
* each iteration. kTerm is the constant \sqrt{k^2 - 1} term.
*/
// SToK (scales) = 2^(1/scales) == k
double k = exp (1.0/m_scaleLevel*log(2));
double kTerm = sqrt (k*k - 1.0);
double w = startsigma;
for (i = 1 ; i < m_gaussCount ; i++) {
cvSmooth( m_gaussImgs[i-1], m_gaussImgs[i],CV_GAUSSIAN,0,0,w);
cvAbsDiff (m_gaussImgs[i-1],m_gaussImgs[i],m_dogImgs[i-1]);
w *= k;
}
}
IplImage* CScaleSpace::getDoG(int i)
{
return m_dogImgs[i];
}
int CScaleSpace::getDoGCount()
{
return m_dogCount;
}
void CScaleSpace::findPeak(double r,double dogThresh,std::vector* vec)
{
// dogThresh is in range [0,1]
dogThresh *= 255;
// {x,y,scale offset}
int dir[27][3] =
{{-1,-1,0},{0,-1,0},{1,-1,0},{-1,0,0},{1,0,0},{-1,1,0},{0,1,0},{1,1,0},
{-1,-1,-1},{0,-1,-1},{1,-1,-1},{-1,0,-1},{0,0,-1},{1,0,-1},{-1,1,-1},{0,1,-1},{1,1,-1},
{-1,-1,1},{0,-1,1},{1,-1,1},{-1,0,1},{0,0,1},{1,0,1},{-1,1,1},{0,1,1},{1,1,1}};
for (int i = 1; i < m_dogCount-1; i++)
{
for (int y = 1; y < cvGetSize (m_dogImgs[i]).height-1; y++)
for (int x = 1; x < cvGetSize (m_dogImgs[i]).width-1; x++)
{
double val = cvGet2D (m_dogImgs[i],y,x).val[0];
bool maxima = true;
for (int k = 0; k < 27; k++)
{
if (val < dogThresh || val < cvGet2D (m_dogImgs[i+dir[k][2]],y+dir[k][1],x+dir[k][0]).val[0])
{
maxima = false;
break;
}
}
if (maxima)
{
// Eliminate edge response
double dxx = cvGet2D(m_dogImgs[i],y,x+1).val[0] +
cvGet2D(m_dogImgs[i],y,x-1).val[0] -
2 * cvGet2D(m_dogImgs[i],y,x).val[0];
double dyy = cvGet2D(m_dogImgs[i],y+1,x).val[0] +
cvGet2D(m_dogImgs[i],y-1,x).val[0] -
2 * cvGet2D(m_dogImgs[i],y,x).val[0];
double dxy = 0.25 * (cvGet2D(m_dogImgs[i],y+1,x+1).val[0] -
cvGet2D(m_dogImgs[i],y-1,x+1).val[0] -
cvGet2D(m_dogImgs[i],y+1,x-1).val[0] +
cvGet2D(m_dogImgs[i],y-1,x-1).val[0]);
double tr = dxx+dyy; // trace
double det = dxx*dyy - dxy*dxy; //det
double rsq = r+1;
rsq *= rsq;
if (tr*tr/(det+1e-20) > (rsq)/r)
maxima = false;
}
if (maxima)
{
CScalePoint p (x,y,i);
vec->push_back(p);
}
//else
// cvSet2D (dogImg[i],y,x,cvScalar(0));
}
}
}
void CScaleSpace::filterAndLocalizePeaks(std::vector& peaks,std::vector* filtered,
double dValueLoThresh, double scaleAdjustThresh, int relocationMaximum)
{
dValueLoThresh *= 255;
CvMat *processedMap = cvCreateMat (cvGetSize(m_dogImgs[0]).height,cvGetSize(m_dogImgs[0]).width,CV_8UC1);
cvSetZero (processedMap);
for (int i = 0; i < peaks.size(); i++)
{
// Kuas : already reject edge like in findPeak
//if (IsTooEdgelike (spaces[peak.Level], peak.X, peak.Y, edgeRatio))
// continue;
// When the localization hits some problem, i.e. while moving the
// point a border is reached, then skip this point.
if (LocalizeIsWeak (peaks[i], relocationMaximum, processedMap))
continue;
if (abs (peaks[i].m_fineS) > scaleAdjustThresh)
continue;
// Additional local pixel information is now available, threshhold
// the D(^x)
// Console.WriteLine ("{0} {1} {2} # == DVALUE", peak.Y, peak.X, peak.Local.DValue);
if (abs (peaks[i].m_DValue) <= dValueLoThresh)
continue;
// its edgy enough, add it
filtered->push_back (peaks[i]);
}
cvReleaseMat (&processedMap);
//return (filtered);
}
bool CScaleSpace::LocalizeIsWeak(CScalePoint& peak, int steps, CvMat* processed)
{
bool needToAdjust = true;
int adjusted = steps;
while (needToAdjust)
{
int x = peak.m_x;
int y = peak.m_y;
// Points we cannot say anything about, as they lie on the border
// of the scale space
// Kuas : seem to be not happen because in FindPeaks not get these peaks
//if (point.Level <= 0 || point.Level >= (spaces.Length - 1))
// return (true);
IplImage* dog = m_dogImgs[peak.m_level];
if (x <= 0 || x >= (cvGetSize(dog).width-1))
return (true);
if (y <= 0 || y >= (cvGetSize(dog).height-1))
return (true);
double dp;
CvMat *adj = cvCreateMat (3,1,CV_32F);
GetAdjustment (peak, peak.m_level, x, y, &dp, adj);
// Get adjustments and check if we require further adjustments due
// to pixel level moves. If so, turn the adjustments into real
// changes and continue the loop. Do not adjust the plane, as we
// are usually quite low on planes in thie space and could not do
// further adjustments from the top/bottom planes.
double adjS = cvGet2D (adj,0,0).val[0];//adj[0, 0];
double adjY = cvGet2D (adj,1,0).val[0];//adj[1, 0];
double adjX = cvGet2D (adj,2,0).val[0];//adj[2, 0];
cvReleaseMat (&adj);
if (abs (adjX) > 0.5 || abs (adjY) > 0.5)
{
// Already adjusted the last time, give up
if (adjusted == 0) {
//Console.WriteLine ("too many adjustments, returning");
return (true);
}
adjusted -= 1;
// Check that just one pixel step is needed, otherwise discard
// the point
double distSq = adjX * adjX + adjY * adjY;
if (distSq > 2.0)
return (true);
peak.m_x = (int) (peak.m_x + adjX + 0.5);
peak.m_y = (int) (peak.m_y + adjY + 0.5);
//point.Level = (int) (point.Level + adjS + 0.5);
//TRACE ("moved point by (%f,%f: %f) to (%d,%d: %d)",
// adjX, adjY, adjS, peak.m_x, peak.m_y, peak.m_level);
continue;
}
// Check if we already have a keypoint within this octave for this
// pixel position in order to avoid dupes. (Maybe we can move this
// check earlier after any adjustment, so we catch dupes earlier).
// If its not in there, mark it for later searches.
//
// FIXME: check why there does not seem to be a dupe at all
if (cvGet2D (processed,peak.m_y,peak.m_x).val[0] != 0)
return (true);
cvSet2D (processed,peak.m_y,peak.m_x,cvScalar(1));
// Save final sub-pixel adjustments.
peak.m_fineS = adjS;
peak.m_fineX = adjX;
peak.m_fineY = adjY;
peak.m_DValue = cvGet2D (dog,peak.m_y,peak.m_x).val[0] + 0.5 * dp;
needToAdjust = false;
}
return false;
}
void CScaleSpace::GetAdjustment(CScalePoint& point, int level, int x, int y, double* dp,CvMat* adj)
{
*dp = 0.0;
//if (point.Level <= 0 || point.Level >= (spaces.Length - 1))
// throw (new ArgumentException ("point.Level is not within [bottom-1;top-1] range"));
IplImage* below = m_dogImgs[level - 1];
IplImage* current = m_dogImgs[level];
IplImage* above = m_dogImgs[level + 1];
CvMat* H = cvCreateMat( 3, 3, CV_32F);
//H[0, 0] = below[x, y] - 2 * current[x, y] + above[x, y];
cvSet2D (H,0,0,cvScalar (cvGet2D (below,y,x).val[0] - 2 * cvGet2D(current,y,x).val[0] + cvGet2D(above,y,x).val[0]) );
cvSet2D (H,0,1,cvScalar(0.25 * (cvGet2D(above,y+1,x).val[0] - cvGet2D(above,y-1,x).val[0] - (cvGet2D(below,y+1,x).val[0] - cvGet2D(below,y-1,x).val[0]))));
cvSet2D (H,1,0,cvGet2D (H,0,1));
cvSet2D (H,0,2,cvScalar(0.25 * (cvGet2D(above,y,x+1).val[0] - cvGet2D(above,y,x-1).val[0] - (cvGet2D(below,y,x+1).val[0] - cvGet2D(below,y,x-1).val[0]))));
cvSet2D (H,2,0,cvGet2D(H,0,2));
cvSet2D (H,1,1,cvScalar(cvGet2D(current,y-1,x).val[0] - 2 * cvGet2D(current,y,x).val[0] + cvGet2D(current,y+1,x).val[0]));
cvSet2D (H,1,2,cvScalar(0.25 * (cvGet2D(current,y+1,x+1).val[0] - cvGet2D(current,y+1,x-1).val[0] -(cvGet2D(current,y-1,x+1).val[0] - cvGet2D(current,y-1,x-1).val[0]))));
cvSet2D (H,2,1,cvGet2D(H,1,2));
cvSet2D (H,2,2,cvScalar(cvGet2D(current,y,x-1).val[0] - 2 * cvGet2D(current,y,x).val[0] + cvGet2D(current,y,x+1).val[0]));
CvMat* d = cvCreateMat(3, 1, CV_32F);
cvSet2D (d,0,0,cvScalar(0.5 * (cvGet2D(above,y,x).val[0] - cvGet2D(below,y,x).val[0])));
cvSet2D (d,1,0,cvScalar(0.5 * (cvGet2D(current,y+1,x).val[0] - cvGet2D(current,y-1,x).val[0])));
cvSet2D (d,2,0,cvScalar(0.5 * (cvGet2D(current,y,x+1).val[0] - cvGet2D(current,y,x-1).val[0])));
CvMat* b = cvCloneMat(d);
//b.Negate ();
cvSet2D (b,0,0,cvScalar(-cvGet2D(b,0,0).val[0]));
cvSet2D (b,1,0,cvScalar(-cvGet2D(b,1,0).val[0]));
cvSet2D (b,2,0,cvScalar(-cvGet2D(b,2,0).val[0]));
// Solve: A x = b
//H.SolveLinear (b);
// adj ::= x
cvSolve (H,b,adj);
*dp = cvDotProduct(adj,d);
cvReleaseMat (&H);
cvReleaseMat (&d);
cvReleaseMat (&b);
}
#define sqr(x) (x*x)
void CScaleSpace::GenMagnitudeAndDirectionMaps()
{
// We leave the first entry to null, and ommit the last. This way, the
// magnitudes and directions maps have the same index as the
// imgScaled maps they belong to.
m_magnitudes = new IplImage*[m_dogCount - 1];
m_directions = new IplImage*[m_dogCount - 1];
// Build the maps, omitting the border pixels, as we cannot build
// gradient information there.
for (int s = 1 ; s < (m_dogCount - 1); s++)
{
m_magnitudes[s] = cvCreateImage (cvGetSize(m_dogImgs[0]),IPL_DEPTH_32F,1);
m_directions[s] = cvCreateImage (cvGetSize(m_dogImgs[0]),IPL_DEPTH_32F,1);
for (int y = 1 ; y < cvGetSize(m_gaussImgs[s]).height - 1 ; y++)
{
for (int x = 1 ; x < cvGetSize(m_gaussImgs[s]).width - 1 ; x++)
{
// gradient magnitude m
double dx = cvGet2D(m_gaussImgs[s],y,x+1).val[0] -
cvGet2D(m_gaussImgs[s],y,x-1).val[0];
double dy = cvGet2D(m_gaussImgs[s],y+1,x).val[0] -
cvGet2D(m_gaussImgs[s],y-1,x).val[0];
cvSet2D (m_magnitudes[s],y,x,cvScalar( sqrt (dx*dx + dy*dy) ) );
// gradient direction theta
cvSet2D (m_directions[s],y,x,cvScalar(atan2(dy,dx) ));
}
}
}
}
void CScaleSpace::ClearMagnitudeAndDirectionMaps()
{
for (int i = 1; i < m_dogCount-1; i++)
{
cvReleaseImage (&m_magnitudes[i]);
cvReleaseImage (&m_directions[i]);
}
delete m_magnitudes;
delete m_directions;
}
void CScaleSpace::GenerateKeypoints (std::vector& localizedPeaks, std::vector* keypoints,int scaleCount, double octaveSigma)
{
for (int i = 0; i < localizedPeaks.size(); i++)
{
std::vector thisKeyPoints;
// Generate zero or more keypoints from the scale point locations.
// TODO: make the values configurable
// Songkran : make 36,0.8 be constant
GenerateKeypointSingle (m_basePixScale, localizedPeaks[i], 36, 0.8, scaleCount, octaveSigma, &thisKeyPoints);
// Generate the feature descriptor.
CreateDescriptors (&thisKeyPoints,m_magnitudes[localizedPeaks[i].m_level], m_directions[localizedPeaks[i].m_level], 2.0, 4, 8, 0.2);
// Only copy over those keypoints that have been successfully
// assigned a descriptor (feature vector).
// Songkran : all key must be already assign feature vector i think
for (int j = 0; j < thisKeyPoints.size(); j++)
{
//if (kp.HasFV == false)
// throw (new Exception ("should not happen"));
// Transform the this level image relative coordinate system
// to the original image coordinates by multiplying with the
// current img scale (which starts with either 0.5 or 1.0 and
// then always doubles: 2.0, 4.0, ..)
// Note that the kp coordinates are not used for processing by
// the detection methods and this has to be the last step.
// Also transform the relative-to-image scale to an
// absolute-to-original-image scale.
// Songkran :
// m_kpScale = octaveSigma * exp (((point.m_level + point.m_fineS) / scaleCount) * log(2));
CKeyPoint* kp = thisKeyPoints[j];
if (kp->m_featureVec == NULL)
TRACE ("Songkran : Something wrong with feature point descriptor\n");
kp->m_x *= kp->m_imgScale;
kp->m_y *= kp->m_imgScale;
kp->m_kpScale *= kp->m_imgScale;
keypoints->push_back (kp);
}
}
}
// Average the content of the direction bins.
void CScaleSpace::AverageWeakBins (double* bins, int binCount)
{
// TODO: make some tests what number of passes is the best. (its clear
// one is not enough, as we may have something like
// ( 0.4, 0.4, 0.3, 0.4, 0.4 ))
for (int sn = 0 ; sn < 4 ; ++sn)
{
double firstE = bins[0];
double last = bins[binCount - 1];
for (int sw = 0 ; sw < binCount ; ++sw)
{
double cur = bins[sw];
double next = (sw == (binCount - 1)) ? firstE : bins[(sw + 1) % binCount];
bins[sw] = (last + cur + next) / 3.0;
last = cur;
}
}
}
// Fit a parabol to the three points (-1.0 ; left), (0.0 ; middle) and
// (1.0 ; right).
//
// Formulas:
// f(x) = a (x - c)^2 + b
//
// c is the peak offset (where f'(x) is zero), b is the peak value.
//
// In case there is an error false is returned, otherwise a correction
// value between [-1 ; 1] is returned in 'degreeCorrection', where -1
// means the peak is located completely at the left vector, and -0.5 just
// in the middle between left and middle and > 0 to the right side. In
// 'peakValue' the maximum estimated peak value is stored.
bool CScaleSpace::InterpolateOrientation (double left, double middle,
double right, double* degreeCorrection, double* peakValue)
{
// Kuas:
// if we set f(x) = 0 then
// a ( x^2 -2cx + c^2) + b= 0
// x1+x2 = 2c/a ,so we get
// a = 2c/(x1+x2)
// and so .. i don't under stand this code
double a = ((left + right) - 2.0 * middle) / 2.0;
// Not a parabola
if (a == 0.0)
return (false);
*peakValue = 0; //Double.NaN;
*degreeCorrection = 0; //Double.NaN;
double c = (((left - middle) / a) - 1.0) / 2.0;
double b = middle - c * c * a;
if (c < -0.5 || c > 0.5)
{
TRACE ("KUAS : ERROR IN InterpolateOrientation\n");
}
*degreeCorrection = c;
*peakValue = b;
return (true);
}
// Create the descriptor vector for a list of keypoints.
//
// keypoints: The list of keypoints to be processed. Everything but the
// descriptor must be filled in already.
// magnitude/direction: The precomputed gradient magnitude and direction
// maps.
// considerScaleFactor: The downscale factor, which describes the amount
// of pixels in the circular region relative to the keypoint scale.
// Low values means few pixels considered, large values extend the
// range. (Use values between 1.0 and 6.0)
// descDim: The dimension size of the output descriptor. There will be
// descDim * descDim * directionCount elements in the feature vector.
// directionCount: The dimensionality of the low level gradient vectors.
// fvGradHicap: The feature vector gradient length hi-cap threshhold.
// (Should be: 0.2)
//
// Some parts modelled after Alexandre Jenny's Matlab implementation.
//
// Songkran : I think this is not implements according to Lowe's paper
// but the idia behind is quite the same
// in this implementation he use variable size of window (change by scale of DoG img)
void CScaleSpace::CreateDescriptors (std::vector* keypoints,
IplImage* magnitude, IplImage* direction,
double considerScaleFactor, int descDim, int directionCount,
double fvGradHicap)
// 2.0, 4, 8,
// 0.2);
{
if (keypoints->size() <= 0)
return;
considerScaleFactor *= (*keypoints)[0]->m_kpScale;
double dDim05 = ((double) descDim) / 2.0;
// Now calculate the radius: We consider pixels in a square with
// dimension 'descDim' plus 0.5 in each direction. As the feature
// vector elements at the diagonal borders are most distant from the
// center pixel we have scale up with sqrt(2).
// Songkran : I think in Lowe paper he suggest to use radius = (descDim*4)/2 * sqrt(2)
int radius = (int) (((descDim + 1.0) / 2) *
sqrt (2.0) * considerScaleFactor + 0.5);
// Precompute the sigma for the "center-most, border-less" gaussian
// weighting.
// (We are operating to dDim05, CV book tells us G(x), x > 3 \sigma
// negligible, but this range seems much shorter!?)
//
// In Lowe03, page 15 it says "A Gaussian weighting function with
// \sigma equal to one half the width of the descriptor window is
// used", so we just use his advice.
// Songkran : should the line below recorrect to 2.0 * (r/2) * (r/2)
double sigma2Sq = 2.0 * dDim05 * dDim05;
for (int i = 0; i < keypoints->size(); i++)
{
CKeyPoint* kp = (*keypoints)[i];
// The angle to rotate with: negate the orientation.
double angle = -kp->m_orientation;
kp->CreateVector (descDim, descDim, directionCount);
//Console.WriteLine (" FV allocated");
for (int y = -radius ; y < radius ; ++y)
{
for (int x = -radius ; x < radius ; ++x)
{
// Rotate and scale
double yR = sin (angle) * x + cos (angle) * y;
double xR = cos (angle) * x - sin (angle) * y;
yR /= considerScaleFactor;
xR /= considerScaleFactor;
// Now consider all (xR, yR) that are anchored within
// (- descDim/2 - 0.5 ; -descDim/2 - 0.5) to
// (descDim/2 + 0.5 ; descDim/2 + 0.5),
// as only those can influence the FV.
if (yR >= (dDim05 + 0.5) || xR >= (dDim05 + 0.5) ||
xR <= -(dDim05 + 0.5) || yR <= -(dDim05 + 0.5))
continue;
int currentX = (int) (x + kp->m_x + 0.5);
int currentY = (int) (y + kp->m_y + 0.5);
if (currentX < 1 || currentX >= (cvGetSize(magnitude).width - 1) ||
currentY < 1 || currentY >= (cvGetSize(magnitude).height - 1))
continue;
// Weight the magnitude relative to the center of the
// whole FV. We do not need a normalizing factor now, as
// we normalize the whole FV later anyway (see below).
// xR, yR are each in -(dDim05 + 0.5) to (dDim05 + 0.5)
// range
double magW = exp (-(xR * xR + yR * yR) / sigma2Sq) *
cvGet2D (magnitude,currentY,currentX).val[0];
// Anchor to (-1.0, -1.0)-(dDim + 1.0, dDim + 1.0), where
// the FV points are located at (x, y)
// Songkran : i think this should be yR += dDim05;
// Songkran : and range is from (-0.5,-0.5) to (dDim+0.5,dDim+0.5)
//yR += dDim05 - 0.5;
//xR += dDim05 - 0.5;
yR += dDim05;
xR += dDim05;
// Build linear interpolation weights:
// A B
// C D
//
// The keypoint is located between A, B, C and D.
// Songkran : this is trilinear interpolation (3 variable x,y,dir)
int xIdx[2];
int yIdx[2];
int dirIdx[2];
double xWeight[2];
double yWeight[2];
double dirWeight[2];
bool flagx[2],flagy[2],flagdir[2];
memset (flagx,false,2*sizeof(bool));
memset (flagy,false,2*sizeof(bool));
memset (flagdir,false,2*sizeof(bool));
// Songkran : have to check xR < descDim also ??
if (xR >= 0 && xR < descDim)
{
xIdx[0] = (int) xR;
xWeight[0] = (1.0 - (xR - xIdx[0]));
flagx[0] = true;
}
if (yR >= 0 && yR < descDim)
{
yIdx[0] = (int) yR;
yWeight[0] = (1.0 - (yR - yIdx[0]));
flagy[0] = true;
}
if (xR+1 < descDim)
{
xIdx[1] = (int) (xR + 1.0);
// xWeight[0] + xWeight[1] = 1
// 1 - xR + xIdx[0] + xR - xIdx[0] - 1 + 1
xWeight[1] = xR - xIdx[1] + 1.0;
flagx[1] = true;
}
if (yR+1 < descDim)
{
// yWeight[0] + yWeight[1] = 1
// 1 - yR + yIdy[0] + yR - yIdy[0] - 1 + 1
yIdx[1] = (int) (yR + 1.0);
yWeight[1] = yR - yIdx[1] + 1.0;
flagy[1] = true;
}
// Rotate the gradient direction by the keypoint
// orientation, then normalize to [-pi ; pi] range.
// Songkran : should correct to dir += 2*Math.PI ?
double dir = cvGet2D(direction,currentY,currentX).val[0] - kp->m_orientation;
if (dir <= -PI)
//dir += Math.PI;
dir += 2*PI;
if (dir > PI)
//dir -= Math.PI;
dir -= 2*PI;
double idxDir = (dir * directionCount) / (2.0 * PI);
if (idxDir < 0.0)
idxDir += directionCount;
dirIdx[0] = (int) idxDir;
dirIdx[1] = (dirIdx[0] + 1) % directionCount;
dirWeight[0] = 1.0 - (idxDir - dirIdx[0]);
dirWeight[1] = idxDir - dirIdx[0];
for (int iy = 0 ; iy < 2 ; ++iy)
for (int ix = 0 ; ix < 2 ; ++ix)
for (int id = 0 ; id < 2 ; ++id)
if (flagx[ix] && flagy[iy])
kp->FVSet (xIdx[ix], yIdx[iy], dirIdx[id],
kp->FVGet (xIdx[ix], yIdx[iy], dirIdx[id]) +
xWeight[ix] * yWeight[iy] * dirWeight[id] * magW);
}
}
// Normalize and hicap the feature vector, as recommended on page
// 16 in Lowe03.
kp->CapAndNormalizeFV (fvGradHicap);
}
}
// Assign each feature point one or more standardized orientations.
// (section 5 in Lowe's paper)
//
// We use an orientation histogram with 36 bins, 10 degrees each. For
// this, every pixel (x,y) lieing in a circle of 'squareDim' diameter
// within in a 'squareDim' sized field within the image L ('gaussImg') is
// examined and two measures calculated:
//
// m = \sqrt{ (L_{x+1,y} - L_{x-1,y})^2 + (L_{x,y+1} - L_{x,y-1})^2 }
// theta = tan^{-1} ( \frac{ L_{x,y+1} - L_{x,y-1} }
// { L_{x+1,y} - L_{x-1,y} } )
//
// Where m is the gradient magnitude around the pixel and theta is the
// gradient orientation. The 'imgScale' value is the octave scale,
// starting with 1.0 at the finest-detail octave, and doubling every
// octave. The gradient orientations are discreetized to 'binCount'
// directions (should be: 36). For every peak orientation that lies within
// 'peakRelThresh' of the maximum peak value, a keypoint location is
// added (should be: 0.8).
//
// Note that 'space' is the gaussian smoothed original image, not the
// difference-of-gaussian one used for peak-search.
// Kuas : imgScale is basePixScale (start scale (relative dimension with input img) of that octave)
void CScaleSpace::GenerateKeypointSingle (double imgScale, CScalePoint point,
int binCount, double peakRelThresh, int scaleCount,
double octaveSigma, std::vector*vec)
{
// The relative estimated keypoint scale. The actual absolute keypoint
// scale to the original image is yielded by the product of imgScale.
// But as we operate in the current octave, the size relative to the
// anchoring images is missing the imgScale factor.
//double kpScale = octaveSigma * exp (((point.m_level + point.m_fineS) / scaleCount) * log(2));
double kpScale = octaveSigma * (exp (((point.m_level) / scaleCount) * log(2))+ point.m_fineS);
// Lowe03, "A gaussian-weighted circular window with a \sigma three
// times that of the scale of the keypoint".
//
// With \sigma = 3.0 * kpScale, the square dimension we have to
// consider is (3 * \sigma) (until the weight becomes very small).
// Kuas : Lowe04 say \sigma = 1.5 * kpScale
double sigma = 1.5 * kpScale;
int radius = (int) (3.0 * sigma / 2.0 + 0.5);
int radiusSq = radius * radius;
IplImage* magnitude = m_magnitudes[point.m_level];
IplImage* direction = m_directions[point.m_level];
// As the point may lie near the border, build the rectangle
// coordinates we can still reach, minus the border pixels, for which
// we do not have gradient information available.
int xMin = point.m_x - radius > 1 ? point.m_x - radius : 1;
int xMax = point.m_x + radius < cvGetSize(magnitude).width-1 ? point.m_x + radius : cvGetSize(magnitude).width-1;
int yMin = point.m_y - radius > 1 ? point.m_y - radius : 1;
int yMax = point.m_y + radius < cvGetSize(magnitude).height-1 ? point.m_y + radius : cvGetSize(magnitude).height-1;
// Precompute 1D gaussian divisor (2 \sigma^2) in:
// G(r) = e^{-\frac{r^2}{2 \sigma^2}}
double gaussianSigmaFactor = 2.0 * sigma * sigma;
double *bins = new double[binCount];
memset (bins,0,sizeof(double)*binCount);
// Build the direction histogram
for (int y = yMin ; y < yMax ; ++y) {
for (int x = xMin ; x < xMax ; ++x) {
// Only consider pixels in the circle, else we might skew the
// orientation histogram by considering more pixels into the
// corner directions
int relX = x - point.m_x;
int relY = y - point.m_y;
if (sqr(relX) + sqr(relY) > radiusSq)
continue;
// The gaussian weight factor.
double gaussianWeight = exp (- ((relX * relX + relY * relY) / gaussianSigmaFactor));
// find the closest bin and add the direction
//int binIdx = FindClosestRotationBin (binCount, direction[x, y]);
double angle = cvGet2D (direction,y,x).val[0];
angle += PI;
angle /= 2.0 * PI;
// calculate the aligned bin
angle *= binCount;
int binIdx = (int) angle;
if (binIdx == binCount)
binIdx = 0;
bins[binIdx] += cvGet2D(magnitude,y,x).val[0] * gaussianWeight;
}
}
// As there may be succeeding histogram entries like this:
// ( ..., 0.4, 0.3, 0.4, ... ) where the real peak is located at the
// middle of this three entries, we can improve the distinctiveness of
// the bins by applying an averaging pass.
//
// TODO: is this really the best method? (we also loose a bit of
// information. Maybe there is a one-step method that conserves more)
AverageWeakBins (bins, binCount);
// find the maximum peak in gradient orientation
double maxGrad = 0.0;
int maxBin = 0;
for (int b = 0 ; b < binCount ; ++b) {
if (bins[b] > maxGrad) {
maxGrad = bins[b];
maxBin = b;
}
}
// First determine the real interpolated peak high at the maximum bin
// position, which is guaranteed to be an absolute peak.
//
// XXX: should we use the estimated peak value as reference for the
// 0.8 check or the original bin-value?
double maxPeakValue, maxDegreeCorrection;
InterpolateOrientation (bins[maxBin == 0 ? (binCount - 1) : (maxBin - 1)],
bins[maxBin], bins[(maxBin + 1) % binCount],
&maxDegreeCorrection, &maxPeakValue);
// Now that we know the maximum peak value, we can find other keypoint
// orientations, which have to fulfill two criterias:
//
// 1. They must be a local peak themselves. Else we might add a very
// similar keypoint orientation twice (imagine for example the
// values: 0.4 1.0 0.8, if 1.0 is maximum peak, 0.8 is still added
// with the default threshhold, but the maximum peak orientation
// was already added).
// 2. They must have at least peakRelThresh times the maximum peak
// value.
bool* binIsKeypoint = new bool[binCount];
for (b = 0 ; b < binCount ; ++b)
{
binIsKeypoint[b] = false;
// The maximum peak of course is
if (b == maxBin)
{
binIsKeypoint[b] = true;
continue;
}
// Local peaks are, too, in case they fulfill the threshhold
if (bins[b] < (peakRelThresh * maxPeakValue))
continue;
int leftI = (b == 0) ? (binCount - 1) : (b - 1);
int rightI = (b + 1) % binCount;
if (bins[b] <= bins[leftI] || bins[b] <= bins[rightI])
continue; // not local peak
binIsKeypoint[b] = true;
}
// All the valid keypoint bins are now marked in binIsKeypoint, now
// build them.
// find other possible locations
double oneBinRad = (2.0 * PI) / binCount;
for (b = 0 ; b < binCount ; ++b)
{
if (binIsKeypoint[b] == false)
continue;
int bLeft = (b == 0) ? (binCount - 1) : (b - 1);
int bRight = (b + 1) % binCount;
// Get an interpolated peak direction and value guess.
double peakValue;
double degreeCorrection;
if (InterpolateOrientation (bins[bLeft], bins[b], bins[bRight],
°reeCorrection, &peakValue) == false)
{
//throw (new InvalidOperationException ("BUG: Parabola fitting broken"));
TRACE ("BUG: Parabola fitting broken\n");
}
// [-1.0 ; 1.0] -> [0 ; binrange], and add the fixed absolute bin
// position.
// We subtract PI because bin 0 refers to 0, binCount-1 bin refers
// to a bin just below 2PI, so -> [-PI ; PI]. Note that at this
// point we determine the canonical descriptor anchor angle. It
// does not matter where we set it relative to the peak degree,
// but it has to be constant. Also, if the output of this
// implementation is to be matched with other implementations it
// must be the same constant angle (here: -PI).
double degree = (b + degreeCorrection) * oneBinRad - PI;
if (degree < -PI)
degree += 2.0 * PI;
else if (degree > PI)
degree -= 2.0 * PI;
CKeyPoint* kp = new CKeyPoint (m_gaussImgs[point.m_level],
point.m_x + point.m_fineX,
point.m_y + point.m_fineY,
imgScale, kpScale, degree);
vec->push_back (kp);
}
delete binIsKeypoint;
delete bins;
}