I would like to implement a HOG descriptor in C++. I found an implementation of this code in Open Pattern Recognition Project - Histograms of Oriented Gradients (HOG) Feature Extraction.
An extract of the code follows:
//extract HOG features from a group of key points
void HOGExtractor::extract(ublas::matrix<float> &dscr_list, const
ublas::matrix<int> &bbox_list, const IntegralHistogram &inthist)
{
unsigned int length = _param.xgrid * _param.ygrid * inthist.dirnum();
if (dscr_list.size1() != bbox_list.size1() || dscr_list.size2() !=
length)
throw std::runtime_error( boost::str(
boost::format("Allocate proper memory for dscr_list [%1% %2%] != [%3% %4%]") % dscr_list.size1() % dscr_list.size2() % dscr_list.size2() % length) );
ublas::matrix<float>::iterator1 idscr_list;
ublas::matrix<int>::const_iterator1 ibbox_list;
ublas::matrix<int>::const_iterator1 ibbox_list_end(bbox_list.end1());
ublas::vector<int> bbox_grids( _param.xgrid * _param.ygrid * 4 );
for ( ibbox_list = bbox_list.begin1(), idscr_list = dscr_list.begin1(); ibbox_list != ibbox_list_end; ++ibbox_list, ++idscr_list)
{
ublas::matrix<int>::const_iterator2 ibbox = ibbox_list.begin();
float x0 = std::max( *ibbox++, 0 );
float y0 = std::max( *ibbox++, 0 );
float x1 = std::min( *ibbox++, inthist.width() - 1 );
float y1 = std::min( *ibbox++, inthist.height() - 1 );
if ( x1 - x0 < _param.xgrid || y1 - y0 < _param.ygrid )
throw std::runtime_error( boost::str(boost::format("Grid number should be less than bbox width/height [%1% %2% %3% %4%]/[%5% %6%]") % x0 % y0 % x1 % y1 % _param.xgrid % _param.ygrid) );
float xstep = (x1 - x0 + 1) / _param.xgrid;
float ystep = (y1 - y0 + 1) / _param.ygrid;
for (int i = 0; i < _param.xgrid; ++i)
for (int j = 0; j < _param.ygrid; ++j)
{
bbox_grids( (i * _param.ygrid + j)*4 + 0) = static_cast<int>(x0 + xstep*i); //x0
bbox_grids( (i * _param.ygrid + j)*4 + 1) = static_cast<int>(y0 + ystep*j); //y0
bbox_grids( (i * _param.ygrid + j)*4 + 2) = static_cast<int>(x0 + xstep*(i + 1) - 1); //x1
bbox_grids( (i * _param.ygrid + j)*4 + 3) = static_cast<int>(y0 + ystep*(j + 1) - 1); //y1
}
inthist.get_hist< ublas::matrix<float>::iterator2, ublas::vector<int>::const_iterator>( idscr_list.begin(), bbox_grids.begin(), bbox_grids.end(), _param.normalize);
}
}
The "key points" it is referred to here are (if I'm not mistaken) the coordinates of the box containing an object we want to extract the features.
However, I am wondering how I can get these coordinates in a image? (Given an image and an object I want to extract the features, how can I know the values of x0, x1, y0, y1?)
