void QuickHarrisDetector::writeHarrisImagesAsPPM(image::ConvexRoi & roi) {
FILE *pFile_x = fopen("/home/jsola/im_x.ppm", "w");
FILE *pFile_y = fopen("/home/jsola/im_y.ppm", "w");
FILE *pFile_xx = fopen("/home/jsola/im_xx.ppm", "w");
FILE *pFile_xy = fopen("/home/jsola/im_xy.ppm", "w");
FILE *pFile_yy = fopen("/home/jsola/im_yy.ppm", "w");
FILE *pFile_min = fopen("/home/jsola/im_min.ppm", "w");
FILE *pFile_max = fopen("/home/jsola/im_max.ppm", "w");
HQuickData * ptr = m_quickData;
if (pFile_x != NULL) {
fprintf(pFile_x, "P2\n%d %d\n255\n", roi.w(), roi.h());
fprintf(pFile_y, "P2\n%d %d\n255\n", roi.w(), roi.h());
fprintf(pFile_xx, "P2\n%d %d\n255\n", roi.w(), roi.h());
fprintf(pFile_xy, "P2\n%d %d\n255\n", roi.w(), roi.h());
fprintf(pFile_yy, "P2\n%d %d\n255\n", roi.w(), roi.h());
fprintf(pFile_min, "P2\n%d %d\n255\n", roi.w(), roi.h());
fprintf(pFile_max, "P2\n%d %d\n255\n", roi.w(), roi.h());
for (int i = 0; i < (roi.h() * roi.w()); i++) {
fprintf(pFile_x, "%d ", 128 + (ptr->im_x) / 2);
fprintf(pFile_y, "%d ", 128 + (ptr->im_y) / 2);
fprintf(pFile_xx, "%d ", 128 + (ptr->im_conv_xx) / 2048);
fprintf(pFile_xy, "%d ", 128 + (ptr->im_conv_xy) / 2048);
fprintf(pFile_yy, "%d ", 128 + (ptr->im_conv_yy) / 2048);
fprintf(pFile_min, "%d ", (unsigned char) (ptr->im_low_curv / 1024));
fprintf(pFile_max, "%d ", (unsigned char) (ptr->im_high_curv / 2048));
ptr++;
}
fclose(pFile_x);
fclose(pFile_y);
fclose(pFile_xx);
fclose(pFile_xy);
fclose(pFile_yy);
}
}
void ActiveSearchGrid::cell2roi(const veci2 & cell, image::ConvexRoi & roi) {
veci2 ul = cellOrigin(cell);
ul(0) += separation;
ul(1) += separation;
veci2 s = cellSize;
s(0) -= 2 * separation;
s(1) -= 2 * separation;
roi.init(cv::Rect(ul(0),ul(1),s(0),s(1)));
}
void ActiveSearchGrid::setFailed(const image::ConvexRoi & roi)
{
vec2 p; p(0) = roi.x()+roi.w()/2; p(1) = roi.y()+roi.h()/2;
veci2 cell = pix2cell(p);
projectionsCount(cell(0), cell(1)) = -1;
}
bool QuickHarrisDetector::quickConvolutionWithBestPoint(
const image::ConvexRoi & roi, int pixMax[2], float & scoreMax) {
int shift_derv = 1;
int shift_derv_conv = shift_derv + shift_conv;
// bounds
int riMin = shift_derv_conv;
int riMax = roi.h() - shift_derv_conv;
int rjMin = shift_derv_conv;
int rjMax = roi.w() - shift_derv_conv;
// data structure pointers
HQuickData* int_center;
#if GAUSSIAN_MASK_APPROX
#else
HQuickData* int_upLeft;
HQuickData* int_upRight;
HQuickData* int_downLeft;
HQuickData* int_downRight;
#endif
float im_low_curv_max = 0.; // maximum smallest eigenvalue
float im_high_curv_max = 0.;
int sm, df; // sum and difference
float sr, corner_ratio;
int ri, rj; // roi coordinates
for (ri = riMin; ri < riMax; ri++) {
int_center = m_quickData + (ri * roi.w()) + rjMin;
#if GAUSSIAN_MASK_APPROX
for(int i = 0; i < nConvCoeffs; ++i)
{
int_upLeft[i] = int_center - (shift_convs[i] * (roi.w()+1));
int_upRight[i] = int_center - (shift_convs[i] * (roi.w()-1));
int_downLeft[i] = int_center + (shift_convs[i] * (roi.w()-1));
int_downRight[i] = int_center + (shift_convs[i] * (roi.w()+1));
}
#else
int_upLeft = int_center - (shift_conv * roi.w()) - shift_conv;
int_upRight = int_center - (shift_conv * roi.w()) + shift_conv;
int_downLeft = int_center + (shift_conv * roi.w()) - shift_conv;
int_downRight = int_center + (shift_conv * roi.w()) + shift_conv;
#endif
for (rj = rjMin; rj < rjMax; rj++) {
#if GAUSSIAN_MASK_APPROX
int_center->im_conv_xx = int_center->im_conv_xy = int_center->im_conv_yy = 0.;
for(int i = 0; i < nConvCoeffs; ++i)
{
int_center->im_conv_xx += convCoeffs[i]*(
int_downRight[i]->int_xx - int_upRight[i]->int_xx -
int_downLeft[i]->int_xx + int_upLeft[i]->int_xx);
int_center->im_conv_xy += convCoeffs[i]*(
int_downRight[i]->int_xy - int_upRight[i]->int_xy -
int_downLeft[i]->int_xy + int_upLeft[i]->int_xy);
int_center->im_conv_yy += convCoeffs[i]*(
int_downRight[i]->int_yy - int_upRight[i]->int_yy -
int_downLeft[i]->int_yy + int_upLeft[i]->int_yy);
}
if (nConvCoeffsCenter)
{
int_center->im_conv_xx += convCoeffs[nConvCoeffs]*int_center->im_xx;
int_center->im_conv_xy += convCoeffs[nConvCoeffs]*int_center->im_xy;
int_center->im_conv_yy += convCoeffs[nConvCoeffs]*int_center->im_yy;
}
#else
int_center->im_conv_xx = int_downRight->int_xx - int_upRight->int_xx
- int_downLeft->int_xx + int_upLeft->int_xx;
int_center->im_conv_xy = int_downRight->int_xy - int_upRight->int_xy
- int_downLeft->int_xy + int_upLeft->int_xy;
int_center->im_conv_yy = int_downRight->int_yy - int_upRight->int_yy
- int_downLeft->int_yy + int_upLeft->int_yy;
#endif
// get eigenvalues: EIG/eig = I_xx + I_yy +/- sqrt((Ixx - I_yy)^2 + 4*I_xy^2)
sm = int_center->im_conv_xx + int_center->im_conv_yy;
df = int_center->im_conv_xx - int_center->im_conv_yy;
sr = sqrt((double) ((double) df * (double) df + 4
* ((double) int_center->im_conv_xy)
* ((double) int_center->im_conv_xy)));
int_center->im_high_curv = (double) sm + sr; // Smallest eigenvalue.
int_center->im_low_curv = (double) sm - sr; // Largest eigenvalue.
// detect and write pixel corresponding to strongest corner, with score.
#ifndef JFR_NDEBUG
if (int_center->im_low_curv == 0.0) corner_ratio = 1e10; else
#endif
corner_ratio = int_center->im_high_curv / int_center->im_low_curv; // CAUTION should be high/low
if (corner_ratio < m_edge) {
if (int_center->im_low_curv > im_low_curv_max) {
im_low_curv_max = int_center->im_low_curv;
im_high_curv_max = int_center->im_high_curv;
pixMax[0] = roi.x() + rj;
pixMax[1] = roi.y() + ri;
}
}
int_center++;
#if GAUSSIAN_MASK_APPROX
for(int i = 0; i < nConvCoeffs; ++i)
{
int_upLeft[i]++;
int_upRight[i]++;
int_downLeft[i]++;
int_downRight[i]++;
}
#else
int_upLeft++;
int_upRight++;
int_downLeft++;
int_downRight++;
#endif
}
}
// normalized score: over the size of the derivative and convolution windows
scoreMax = sqrt(im_low_curv_max / normCoeff);
//std::cout << "- at " << pixMax[0] << "," << pixMax[1] << " : high " << im_high_curv_max << " low " << im_low_curv_max << " ; ratio " << im_high_curv_max/im_low_curv_max << " score " << scoreMax << std::endl;
#if FILTER_VIRTUAL_POINTS
/*
accept:
_|_ (4) |_ _/ _ (2) _ _ _ (1 or 2)
| \ \ /
reject:
|_ (3) . (0)
|
*/
int_center = m_quickData + ((pixMax[1]-roi.y()) * roi.w()) + (pixMax[0]-roi.x());
int indexes[8][2] = {{-1,-1},{-1,0},{-1,1},{0,1},{1,1},{1,0},{1,-1},{0,-1}};
int radius = m_convolutionSize/2;
for(int i = 0; i < 8; ++i) for(int j = 0; j < 2; ++j) indexes[i][j] *= radius;
double thres = im_low_curv_max * 0.75;
int npeaks = 0;
bool lastpeaked = false;
bool firstpeaked = false;
for(int i = 0; i < 8; ++i)
{
int_upLeft = int_center + indexes[i][1] * roi.w() + indexes[i][0];
if (int_upLeft->im_high_curv > thres)
{
if (i == 0) firstpeaked = true; else
if (i == 7) lastpeaked = lastpeaked || firstpeaked;
if (!lastpeaked) npeaks++; lastpeaked = true;
} else
lastpeaked = false;
std::cout << " high " << int_upLeft->im_high_curv << " " << " low " << int_upLeft->im_low_curv << std::endl;
}
bool ok = (npeaks == 1 || npeaks == 2 || npeaks == 4);
std::cout << " high " << int_center->im_high_curv << " low " << int_center->im_low_curv << " npeaks " << npeaks << " ok " << ok << std::endl;
#endif
return (scoreMax > m_threshold);
}
void QuickHarrisDetector::quickDerivatives(
const jafar::image::Image & image, image::ConvexRoi & roi) {
const uchar* pix_center;
const uchar* pix_right;
const uchar* pix_left;
const uchar* pix_down;
const uchar* pix_up;
HQuickData* int_center;
HQuickData* int_up;
HQuickData* int_upLeft;
HQuickData* int_left;
// dead zone due to derivative mask [-1 0 1]
int shift_derv = 1; // = (3-1)/2
// i: vert; j: horz image coordinates
int iMin = roi.y() + shift_derv;
int iMax = roi.y() + roi.h() - shift_derv;
// FIXME take into account possible convex part of roi x(i) and w(i) by moving this into the loop
int jMin = roi.x() + shift_derv;
int jMax = roi.x() + roi.w() - shift_derv;
int i, j; // i, j: image coordinates
int ri; // ri: vertical roi coordinate
for (i = iMin; i < iMax; i++) {
ri = i - roi.y();
pix_center = image.data() + (i * image.step()) + jMin;
pix_right = pix_center + 1;
pix_left = pix_center - 1;
pix_down = pix_center + image.step();
pix_up = pix_center - image.step();
int_center = m_quickData + (ri * roi.w()) + shift_derv;
int_up = int_center - roi.w();
int_upLeft = int_up - 1;
int_left = int_center - 1;
for (j = jMin; j < jMax; j++) {
// Build x and y derivatives, and their products: xx, xy and yy.
int_center->im_x = *pix_right - *pix_left;
int_center->im_y = *pix_down - *pix_up;
int_center->im_xx = int_center->im_x * int_center->im_x;
int_center->im_xy = int_center->im_x * int_center->im_y;
int_center->im_yy = int_center->im_y * int_center->im_y;
// build integral images of the 3 derivative products, xx, xy and yy:
if (i == iMin && j == jMin) {
int_center->int_xx = int_center->im_xx;
int_center->int_xy = int_center->im_xy;
int_center->int_yy = int_center->im_yy;
} else if (i == iMin) {
int_center->int_xx = int_center->im_xx + int_left->int_xx;
int_center->int_xy = int_center->im_xy + int_left->int_xy;
int_center->int_yy = int_center->im_yy + int_left->int_yy;
} else if (j == jMin) {
int_center->int_xx = int_center->im_xx + int_up->int_xx;
int_center->int_xy = int_center->im_xy + int_up->int_xy;
int_center->int_yy = int_center->im_yy + int_up->int_yy;
} else {
int_center->int_xx = int_center->im_xx + int_up->int_xx
+ int_left->int_xx - int_upLeft->int_xx;
int_center->int_xy = int_center->im_xy + int_up->int_xy
+ int_left->int_xy - int_upLeft->int_xy;
int_center->int_yy = int_center->im_yy + int_up->int_yy
+ int_left->int_yy - int_upLeft->int_yy;
}
// Advance all pointers
pix_center++;
pix_right++;
pix_left++;
pix_up++;
pix_down++;
int_center++;
int_up++;
int_left++;
int_upLeft++;
}
}
}
