Understanding the HOG feature layout - c++

Question

I am doing a project which involve multi-class object detection. My aim is to detect the following objects.
1. Truck
2. Car
3. Person
As I have three different objects, it means I will have three different window sizes. But HOG features for blocks will remain the same. I hacked into OpenCV hog.cpp and made two new functions to calculate the HOG descriptors for blocks only. Here is my code.
void cv::gpu::HOGDescriptor::getDescriptorsBlock(const GpuMat& img, Size win_stride, GpuMat& descriptors, FileStorage fs3, string fileName, double scale, int width, int height, size_t lev)
{
CV_Assert(win_stride.width % block_stride.width == 0 && win_stride.height % block_stride.height == 0);
size_t block_hist_size = getBlockHistogramSize();
computeBlockHistograms(img);
Size blocks_per_img = numPartsWithin(img.size(), block_size, block_stride);
// Size blocks_per_win = numPartsWithin(win_size, block_size, block_stride);
// Size wins_per_img = numPartsWithin(img.size(), win_size, win_stride);
// copy block_hists from GPU to CPU/
float dest_ptr[block_hist_size * blocks_per_img.area()];
cudaMemcpy( &dest_ptr[0], block_hists.ptr<float>(), block_hist_size *blocks_per_img.area()*sizeof(CV_32F), cudaMemcpyDeviceToHost);
std::cout<<"( "<<width<< " ," << height<< ")"<< std::endl;
std::cout <<lev<< std::endl;
// write to yml file
int level = lev;
fs3<<"Scale"<<scale;
fs3 <<"Level"<<level;
fs3<<"Width"<<width<<"Height"<<height;
fs3 << "features" << "[";
for (unsigned int i = 0; i < (block_hist_size * blocks_per_img.area()) ; i++ )
{
fs3 << dest_ptr[i];
}
fs3 << "]";
}
Similarly to get block descriptors for multi-scale
void cv::gpu::HOGDescriptor::getDescriptorsMultiScale(const GpuMat& img,
Size win_stride, double scale0, unsigned int count)
{
CV_Assert(img.type() == CV_8UC1 || img.type() == CV_8UC4);
vector<double> level_scale;
double scale = 1.;
int levels = 0;
for (levels = 0; levels < nlevels; levels++)
{
level_scale.push_back(scale);
if (cvRound(img.cols/scale) < win_size.width ||
cvRound(img.rows/scale) < win_size.height || scale0 <= 1)
break;
scale *= scale0;
}
levels = std::max(levels, 1);
level_scale.resize(levels);
image_scales.resize(levels);
// open yml file with image ID
FileStorage fs3;
char fileName[20];
GpuMat descriptors;
sprintf (fileName, "%04d", count);
fs3.open(fileName, FileStorage::WRITE);
for (size_t i = 0; i < level_scale.size(); i++)
{
scale = level_scale[i];
Size sz(cvRound(img.cols / scale), cvRound(img.rows / scale));
GpuMat smaller_img;
if (sz == img.size())
smaller_img = img;
else
{
image_scales[i].create(sz, img.type());
switch (img.type())
{
case CV_8UC1: hog::resize_8UC1(img, image_scales[i]); break;
case CV_8UC4: hog::resize_8UC4(img, image_scales[i]); break;
}
smaller_img = image_scales[i];
}
std::cout<<"scale "<<level_scale[i]<<std::endl;
// calculate descriptors for blocks
getDescriptorsBlock( smaller_img, win_stride, descriptors, fs3, fileName, scale, smaller_img.cols, smaller_img.rows, i);
// detect(smaller_img, locations, hit_threshold, win_stride, padding);
}
// close yml file
fs3.release();
}
My question would be to understand the layout structure of HOG descriptors for block only. Can some one share his thoughts

Answer 1

Generally, using an image pyramid is often applied to become scale invariant. If you want to get more sophisticated, have a look at this paper "Object Detection with Discriminatively Trained
Part Based Models" [1]. They were really successful in using HoG on different scales. Of course the original HoG paper might be of use to understand the structure of the feature itself [2], if that is more what you are after.
[1] http://vision.ics.uci.edu/papers/FelzenszwalbGMR_PAMI_2009/FelzenszwalbGMR_PAMI_2009.pdf
[2] http://lear.inrialpes.fr/people/triggs/pubs/Dalal-cvpr05.pdf