// Copyright (C) 2010 Davis E. King (davis@dlib.net)
// License: Boost Software License See LICENSE.txt for the full license.
#ifndef DLIB_HoG_Hh_
#define DLIB_HoG_Hh_
#include "hog_abstract.h"
#include "../algs.h"
#include "../matrix.h"
#include "../array2d.h"
#include "../geometry.h"
#include <cmath>
namespace dlib
{
enum
{
hog_no_interpolation,
hog_angle_interpolation,
hog_full_interpolation,
hog_signed_gradient,
hog_unsigned_gradient
};
template <
unsigned long cell_size_,
unsigned long block_size_,
unsigned long cell_stride_,
unsigned long num_orientation_bins_,
int gradient_type_,
int interpolation_type_
>
class hog_image : noncopyable
{
COMPILE_TIME_ASSERT(cell_size_ > 1);
COMPILE_TIME_ASSERT(block_size_ > 0);
COMPILE_TIME_ASSERT(cell_stride_ > 0);
COMPILE_TIME_ASSERT(num_orientation_bins_ > 0);
COMPILE_TIME_ASSERT( gradient_type_ == hog_signed_gradient ||
gradient_type_ == hog_unsigned_gradient);
COMPILE_TIME_ASSERT( interpolation_type_ == hog_no_interpolation ||
interpolation_type_ == hog_angle_interpolation ||
interpolation_type_ == hog_full_interpolation );
public:
const static unsigned long cell_size = cell_size_;
const static unsigned long block_size = block_size_;
const static unsigned long cell_stride = cell_stride_;
const static unsigned long num_orientation_bins = num_orientation_bins_;
const static int gradient_type = gradient_type_;
const static int interpolation_type = interpolation_type_;
const static long min_size = cell_size*block_size+2;
typedef matrix<double, block_size*block_size*num_orientation_bins, 1> descriptor_type;
hog_image (
) :
num_block_rows(0),
num_block_cols(0)
{}
void clear (
)
{
num_block_rows = 0;
num_block_cols = 0;
hist_cells.clear();
}
void copy_configuration (
const hog_image&
){}
template <
typename image_type
>
inline void load (
const image_type& img
)
{
COMPILE_TIME_ASSERT( pixel_traits<typename image_traits<image_type>::pixel_type>::has_alpha == false );
load_impl(mat(img));
}
inline void unload(
) { clear(); }
inline unsigned long size (
) const { return static_cast<unsigned long>(nr()*nc()); }
inline long nr (
) const { return num_block_rows; }
inline long nc (
) const { return num_block_cols; }
long get_num_dimensions (
) const
{
return block_size*block_size*num_orientation_bins;
}
inline const descriptor_type& operator() (
long row,
long col
) const
{
// make sure requires clause is not broken
DLIB_ASSERT( 0 <= row && row < nr() &&
0 <= col && col < nc(),
"\t descriptor_type hog_image::operator()()"
<< "\n\t invalid row or col argument"
<< "\n\t row: " << row
<< "\n\t col: " << col
<< "\n\t nr(): " << nr()
<< "\n\t nc(): " << nc()
<< "\n\t this: " << this
);
row *= cell_stride;
col *= cell_stride;
++row;
++col;
int feat = 0;
for (unsigned long r = 0; r < block_size; ++r)
{
for (unsigned long c = 0; c < block_size; ++c)
{
for (unsigned long i = 0; i < num_orientation_bins; ++i)
{
des(feat++) = hist_cells[row+r][col+c].values[i];
}
}
}
des /= length(des) + 1e-8;
return des;
}
const rectangle get_block_rect (
long row,
long col
) const
{
row *= cell_stride;
col *= cell_stride;
row *= cell_size;
col *= cell_size;
// do this to account for the 1 pixel padding we use all around the image
++row;
++col;
return rectangle(col, row, col+cell_size*block_size-1, row+cell_size*block_size-1);
}
const point image_to_feat_space (
const point& p
) const
{
const long half_block = block_size/2;
if ((block_size%2) == 0)
{
return point(((p.x()-1)/(long)cell_size - half_block)/(long)cell_stride,
((p.y()-1)/(long)cell_size - half_block)/(long)cell_stride);
}
else
{
return point(((p.x()-1-(long)cell_size/2)/(long)cell_size - half_block)/(long)cell_stride,
((p.y()-1-(long)cell_size/2)/(long)cell_size - half_block)/(long)cell_stride);
}
}
const rectangle image_to_feat_space (
const rectangle& rect
) const
{
return rectangle(image_to_feat_space(rect.tl_corner()), image_to_feat_space(rect.br_corner()));
}
const point feat_to_image_space (
const point& p
) const
{
const long half_block = block_size/2;
if ((block_size%2) == 0)
{
return point((p.x()*cell_stride + half_block)*cell_size + 1,
(p.y()*cell_stride + half_block)*cell_size + 1);
}
else
{
return point((p.x()*cell_stride + half_block)*cell_size + 1 + cell_size/2,
(p.y()*cell_stride + half_block)*cell_size + 1 + cell_size/2);
}
}
const rectangle feat_to_image_space (
const rectangle& rect
) const
{
return rectangle(feat_to_image_space(rect.tl_corner()), feat_to_image_space(rect.br_corner()));
}
// these _PRIVATE_ functions are only here as a workaround for a bug in visual studio 2005.
void _PRIVATE_serialize (std::ostream& out) const
{
// serialize hist_cells
serialize(hist_cells.nc(),out);
serialize(hist_cells.nr(),out);
hist_cells.reset();
while (hist_cells.move_next())
serialize(hist_cells.element().values,out);
hist_cells.reset();
serialize(num_block_rows, out);
serialize(num_block_cols, out);
}
void _PRIVATE_deserialize (std::istream& in )
{
// deserialize item.hist_cells
long nc, nr;
deserialize(nc,in);
deserialize(nr,in);
hist_cells.set_size(nr,nc);
while (hist_cells.move_next())
deserialize(hist_cells.element().values,in);
hist_cells.reset();
deserialize(num_block_rows, in);
deserialize(num_block_cols, in);
}
private:
template <
typename image_type
>
void load_impl (
const image_type& img
)
{
// Note that we keep a border of 1 pixel all around the image so that we don't have
// to worry about running outside the image when computing the horizontal and vertical
// gradients.
// Note also that we have a border of unused cells around the hist_cells array so that we
// don't have to worry about edge effects when doing the interpolation in the main loop
// below.
// check if the window is just too small
if (img.nr() < min_size || img.nc() < min_size)
{
// If the image is smaller than our windows then there aren't any descriptors at all!
num_block_rows = 0;
num_block_cols = 0;
return;
}
// Make sure we have the right number of cell histograms and that they are
// all set to zero.
hist_cells.set_size((img.nr()-2)/cell_size+2, (img.nc()-2)/cell_size+2);
for (long r = 0; r < hist_cells.nr(); ++r)
{
for (long c = 0; c < hist_cells.nc(); ++c)
{
hist_cells[r][c].zero();
}
}
// loop over all the histogram cells and fill them out
for (long rh = 1; rh < hist_cells.nr()-1; ++rh)
{
for (long ch = 1; ch < hist_cells.nc()-1; ++ch)
{
// Fill out the current histogram cell.
// First, figure out the row and column offsets into the image for the current histogram cell.
const long roff = (rh-1)*cell_size + 1;
const long coff = (ch-1)*cell_size + 1;
for (long r = 0; r < (long)cell_size; ++r)
{
for (long c = 0; c < (long)cell_size; ++c)
{
unsigned long left;
unsigned long right;
unsigned long top;
unsigned long bottom;
assign_pixel(left, img(r+roff,c+coff-1));
assign_pixel(right, img(r+roff,c+coff+1));
assign_pixel(top, img(r+roff-1,c+coff));
assign_pixel(bottom, img(r+roff+1,c+coff));
double grad_x = (long)right-(long)left;
double grad_y = (long)top-(long)bottom;
// obtain the angle of the gradient. Make sure it is scaled between 0 and 1.
double angle = std::max(0.0, std::atan2(grad_y, grad_x)/pi + 1)/2;
if (gradient_type == hog_unsigned_gradient)
{
angle *= 2;
if (angle >= 1)
angle -= 1;
}
// now scale angle to between 0 and num_orientation_bins
angle *= num_orientation_bins;
const double strength = std::sqrt(grad_y*grad_y + grad_x*grad_x);
if (interpolation_type == hog_no_interpolation)
{
// no interpolation
hist_cells[rh][ch].values[round_to_int(angle)%num_orientation_bins] += strength;
}
else // if we should do some interpolation
{
unsigned long quantized_angle_lower = static_cast<unsigned long>(std::floor(angle));
unsigned long quantized_angle_upper = static_cast<unsigned long>(std::ceil(angle));
quantized_angle_lower %= num_orientation_bins;
quantized_angle_upper %= num_orientation_bins;
const double angle_split = (angle-std::floor(angle));
const double upper_strength = angle_split*strength;
const double lower_strength = (1-angle_split)*strength;
if (interpolation_type == hog_angle_interpolation)
{
// Stick into gradient histogram. Note that we linearly interpolate between neighboring
// histogram buckets.
hist_cells[rh][ch].values[quantized_angle_lower] += lower_strength;
hist_cells[rh][ch].values[quantized_angle_upper] += upper_strength;
}
else // here we do hog_full_interpolation
{
const double center_r = (cell_size-1)/2.0;
const double center_c = (cell_size-1)/2.0;
const double lin_neighbor_r = std::abs(center_r - r)/cell_size;
const double lin_main_r = 1-lin_neighbor_r;
const double lin_neighbor_c = std::abs(center_c - c)/cell_size;
const double lin_main_c = 1-lin_neighbor_c;
// Which neighboring cells we interpolate into depends on which
// corner of our main cell we are nearest.
if (r < center_r)
{
if (c < center_c)
{
hist_cells[rh][ch].values[quantized_angle_upper] += upper_strength * lin_main_r*lin_main_c;
hist_cells[rh][ch].values[quantized_angle_lower] += lower_strength * lin_main_r*lin_main_c;
hist_cells[rh-1][ch].values[quantized_angle_upper] += upper_strength * lin_neighbor_r*lin_main_c;
hist_cells[rh-1][ch].values[quantized_angle_lower] += lower_strength * lin_neighbor_r*lin_main_c;
hist_cells[rh][ch-1].values[quantized_angle_upper] += upper_strength * lin_neighbor_c*lin_main_r;
hist_cells[rh][ch-1].values[quantized_angle_lower] += lower_strength * lin_neighbor_c*lin_main_r;
hist_cells[rh-1][ch-1].values[quantized_angle_upper] += upper_strength * lin_neighbor_c*lin_neighbor_r;
hist_cells[rh-1][ch-1].values[quantized_angle_lower] += lower_strength * lin_neighbor_c*lin_neighbor_r;
}
else
{
hist_cells[rh][ch].values[quantized_angle_upper] += upper_strength * lin_main_r*lin_main_c;
hist_cells[rh][ch].values[quantized_angle_lower] += lower_strength * lin_main_r*lin_main_c;
hist_cells[rh-1][ch].values[quantized_angle_upper] += upper_strength * lin_neighbor_r*lin_main_c;
hist_cells[rh-1][ch].values[quantized_angle_lower] += lower_strength * lin_neighbor_r*lin_main_c;
hist_cells[rh][ch+1].values[quantized_angle_upper] += upper_strength * lin_neighbor_c*lin_main_r;
hist_cells[rh][ch+1].values[quantized_angle_lower] += lower_strength * lin_neighbor_c*lin_main_r;
hist_cells[rh-1][ch+1].values[quantized_angle_upper] += upper_strength * lin_neighbor_c*lin_neighbor_r;
hist_cells[rh-1][ch+1].values[quantized_angle_lower] += lower_strength * lin_neighbor_c*lin_neighbor_r;
}
}
else
{
if (c < center_c)
{
hist_cells[rh][ch].values[quantized_angle_upper] += upper_strength * lin_main_r*lin_main_c;
hist_cells[rh][ch].values[quantized_angle_lower] += lower_strength * lin_main_r*lin_main_c;
hist_cells[rh+1][ch].values[quantized_angle_upper] += upper_strength * lin_neighbor_r*lin_main_c;
hist_cells[rh+1][ch].values[quantized_angle_lower] += lower_strength * lin_neighbor_r*lin_main_c;
hist_cells[rh][ch-1].values[quantized_angle_upper] += upper_strength * lin_neighbor_c*lin_main_r;
hist_cells[rh][ch-1].values[quantized_angle_lower] += lower_strength * lin_neighbor_c*lin_main_r;
hist_cells[rh+1][ch-1].values[quantized_angle_upper] += upper_strength * lin_neighbor_c*lin_neighbor_r;
hist_cells[rh+1][ch-1].values[quantized_angle_lower] += lower_strength * lin_neighbor_c*lin_neighbor_r;
}
else
{
hist_cells[rh][ch].values[quantized_angle_upper] += upper_strength * lin_main_r*lin_main_c;
hist_cells[rh][ch].values[quantized_angle_lower] += lower_strength * lin_main_r*lin_main_c;
hist_cells[rh+1][ch].values[quantized_angle_upper] += upper_strength * lin_neighbor_r*lin_main_c;
hist_cells[rh+1][ch].values[quantized_angle_lower] += lower_strength * lin_neighbor_r*lin_main_c;
hist_cells[rh][ch+1].values[quantized_angle_upper] += upper_strength * lin_neighbor_c*lin_main_r;
hist_cells[rh][ch+1].values[quantized_angle_lower] += lower_strength * lin_neighbor_c*lin_main_r;
hist_cells[rh+1][ch+1].values[quantized_angle_upper] += upper_strength * lin_neighbor_c*lin_neighbor_r;
hist_cells[rh+1][ch+1].values[quantized_angle_lower] += lower_strength * lin_neighbor_c*lin_neighbor_r;
}
}
}
}
}
}
}
}
// Now figure out how many blocks we should have. Note again that the hist_cells has a border of
// unused cells (thats where that -2 comes from).
num_block_rows = (hist_cells.nr()-2 - (block_size-1) + cell_stride - 1)/cell_stride;
num_block_cols = (hist_cells.nc()-2 - (block_size-1) + cell_stride - 1)/cell_stride;
}
unsigned long round_to_int(
double val
) const
{
return static_cast<unsigned long>(std::floor(val + 0.5));
}
struct histogram
{
void zero()
{
for (unsigned long i = 0; i < num_orientation_bins; ++i)
values[i] = 0;
}
double values[num_orientation_bins];
};
array2d<histogram> hist_cells;
mutable descriptor_type des;
long num_block_rows;
long num_block_cols;
};
// ----------------------------------------------------------------------------------------
template <
unsigned long T1,
unsigned long T2,
unsigned long T3,
unsigned long T4,
int T5,
int T6
>
void serialize (
const hog_image<T1,T2,T3,T4,T5,T6>& item,
std::ostream& out
)
{
item._PRIVATE_serialize(out);
}
template <
unsigned long T1,
unsigned long T2,
unsigned long T3,
unsigned long T4,
int T5,
int T6
>
void deserialize (
hog_image<T1,T2,T3,T4,T5,T6>& item,
std::istream& in
)
{
item._PRIVATE_deserialize(in);
}
// ----------------------------------------------------------------------------------------
}
#endif // DLIB_HoG_Hh_