// Copyright (C) 2011 Davis E. King (davis@dlib.net)
// License: Boost Software License See LICENSE.txt for the full license.
#ifndef DLIB_STRUCTURAL_SVM_ObJECT_DETECTION_PROBLEM_Hh_
#define DLIB_STRUCTURAL_SVM_ObJECT_DETECTION_PROBLEM_Hh_
#include "structural_svm_object_detection_problem_abstract.h"
#include "../matrix.h"
#include "structural_svm_problem_threaded.h"
#include <sstream>
#include "../string.h"
#include "../array.h"
#include "../image_processing/full_object_detection.h"
#include "../image_processing/box_overlap_testing.h"
namespace dlib
{
// ----------------------------------------------------------------------------------------
template <
typename image_scanner_type,
typename image_array_type
>
class structural_svm_object_detection_problem : public structural_svm_problem_threaded<matrix<double,0,1> >,
noncopyable
{
public:
structural_svm_object_detection_problem(
const image_scanner_type& scanner,
const test_box_overlap& overlap_tester,
const bool auto_overlap_tester,
const image_array_type& images_,
const std::vector<std::vector<full_object_detection> >& truth_object_detections_,
const std::vector<std::vector<rectangle> >& ignore_,
const test_box_overlap& ignore_overlap_tester_,
unsigned long num_threads = 2
) :
structural_svm_problem_threaded<matrix<double,0,1> >(num_threads),
boxes_overlap(overlap_tester),
images(images_),
truth_object_detections(truth_object_detections_),
ignore(ignore_),
ignore_overlap_tester(ignore_overlap_tester_),
match_eps(0.5),
loss_per_false_alarm(1),
loss_per_missed_target(1)
{
#ifdef ENABLE_ASSERTS
// make sure requires clause is not broken
DLIB_ASSERT(is_learning_problem(images_, truth_object_detections_) &&
ignore_.size() == images_.size() &&
scanner.get_num_detection_templates() > 0,
"\t structural_svm_object_detection_problem::structural_svm_object_detection_problem()"
<< "\n\t Invalid inputs were given to this function "
<< "\n\t scanner.get_num_detection_templates(): " << scanner.get_num_detection_templates()
<< "\n\t is_learning_problem(images_,truth_object_detections_): " << is_learning_problem(images_,truth_object_detections_)
<< "\n\t ignore.size(): " << ignore.size()
<< "\n\t images.size(): " << images.size()
<< "\n\t this: " << this
);
for (unsigned long i = 0; i < truth_object_detections.size(); ++i)
{
for (unsigned long j = 0; j < truth_object_detections[i].size(); ++j)
{
DLIB_ASSERT(truth_object_detections[i][j].num_parts() == scanner.get_num_movable_components_per_detection_template(),
"\t trained_function_type structural_object_detection_trainer::train()"
<< "\n\t invalid inputs were given to this function"
<< "\n\t truth_object_detections["<<i<<"]["<<j<<"].num_parts(): " <<
truth_object_detections[i][j].num_parts()
<< "\n\t scanner.get_num_movable_components_per_detection_template(): " <<
scanner.get_num_movable_components_per_detection_template()
<< "\n\t all_parts_in_rect(truth_object_detections["<<i<<"]["<<j<<"]): " << all_parts_in_rect(truth_object_detections[i][j])
);
}
}
#endif
// The purpose of the max_num_dets member variable is to give us a reasonable
// upper limit on the number of detections we can expect from a single image.
// This is used in the separation_oracle to put a hard limit on the number of
// detections we will consider. We do this purely for computational reasons
// since otherwise we can end up wasting large amounts of time on certain
// pathological cases during optimization which ultimately do not influence the
// result. Therefore, we force the separation oracle to only consider the
// max_num_dets strongest detections.
max_num_dets = 0;
for (unsigned long i = 0; i < truth_object_detections.size(); ++i)
{
if (truth_object_detections[i].size() > max_num_dets)
max_num_dets = truth_object_detections[i].size();
}
max_num_dets = max_num_dets*3 + 10;
initialize_scanners(scanner, num_threads);
if (auto_overlap_tester)
{
auto_configure_overlap_tester();
}
}
test_box_overlap get_overlap_tester (
) const
{
return boxes_overlap;
}
void set_match_eps (
double eps
)
{
// make sure requires clause is not broken
DLIB_ASSERT(0 < eps && eps < 1,
"\t void structural_svm_object_detection_problem::set_match_eps(eps)"
<< "\n\t Invalid inputs were given to this function "
<< "\n\t eps: " << eps
<< "\n\t this: " << this
);
match_eps = eps;
}
double get_match_eps (
) const
{
return match_eps;
}
double get_loss_per_missed_target (
) const
{
return loss_per_missed_target;
}
void set_loss_per_missed_target (
double loss
)
{
// make sure requires clause is not broken
DLIB_ASSERT(loss > 0,
"\t void structural_svm_object_detection_problem::set_loss_per_missed_target(loss)"
<< "\n\t Invalid inputs were given to this function "
<< "\n\t loss: " << loss
<< "\n\t this: " << this
);
loss_per_missed_target = loss;
}
double get_loss_per_false_alarm (
) const
{
return loss_per_false_alarm;
}
void set_loss_per_false_alarm (
double loss
)
{
// make sure requires clause is not broken
DLIB_ASSERT(loss > 0,
"\t void structural_svm_object_detection_problem::set_loss_per_false_alarm(loss)"
<< "\n\t Invalid inputs were given to this function "
<< "\n\t loss: " << loss
<< "\n\t this: " << this
);
loss_per_false_alarm = loss;
}
private:
void auto_configure_overlap_tester(
)
{
std::vector<std::vector<rectangle> > mapped_rects(truth_object_detections.size());
for (unsigned long i = 0; i < truth_object_detections.size(); ++i)
{
mapped_rects[i].resize(truth_object_detections[i].size());
for (unsigned long j = 0; j < truth_object_detections[i].size(); ++j)
{
mapped_rects[i][j] = scanners[i].get_best_matching_rect(truth_object_detections[i][j].get_rect());
}
}
boxes_overlap = find_tight_overlap_tester(mapped_rects);
}
virtual long get_num_dimensions (
) const
{
return scanners[0].get_num_dimensions() +
1;// for threshold
}
virtual long get_num_samples (
) const
{
return images.size();
}
virtual void get_truth_joint_feature_vector (
long idx,
feature_vector_type& psi
) const
{
const image_scanner_type& scanner = scanners[idx];
psi.set_size(get_num_dimensions());
std::vector<rectangle> mapped_rects;
psi = 0;
for (unsigned long i = 0; i < truth_object_detections[idx].size(); ++i)
{
mapped_rects.push_back(scanner.get_best_matching_rect(truth_object_detections[idx][i].get_rect()));
scanner.get_feature_vector(truth_object_detections[idx][i], psi);
}
psi(scanner.get_num_dimensions()) = -1.0*truth_object_detections[idx].size();
// check if any of the boxes overlap. If they do then it is impossible for
// us to learn to correctly classify this sample
for (unsigned long i = 0; i < mapped_rects.size(); ++i)
{
for (unsigned long j = i+1; j < mapped_rects.size(); ++j)
{
if (boxes_overlap(mapped_rects[i], mapped_rects[j]))
{
const double area_overlap = mapped_rects[i].intersect(mapped_rects[j]).area();
const double match_amount = area_overlap/(double)( mapped_rects[i]+mapped_rects[j]).area();
const double overlap_amount = area_overlap/std::min(mapped_rects[i].area(),mapped_rects[j].area());
using namespace std;
ostringstream sout;
sout << "An impossible set of object labels was detected. This is happening because ";
sout << "the truth labels for an image contain rectangles which overlap according to the ";
sout << "test_box_overlap object supplied for non-max suppression. To resolve this, you ";
sout << "either need to relax the test_box_overlap object so it doesn't mark these rectangles as ";
sout << "overlapping or adjust the truth rectangles in your training dataset. ";
// make sure the above string fits nicely into a command prompt window.
string temp = sout.str();
sout.str(""); sout << wrap_string(temp,0,0) << endl << endl;
sout << "image index: "<< idx << endl;
sout << "The offending rectangles are:\n";
sout << "rect1: "<< mapped_rects[i] << endl;
sout << "rect2: "<< mapped_rects[j] << endl;
sout << "match amount: " << match_amount << endl;
sout << "overlap amount: " << overlap_amount << endl;
throw dlib::impossible_labeling_error(sout.str());
}
}
}
// make sure the mapped rectangles are within match_eps of the
// truth rectangles.
for (unsigned long i = 0; i < mapped_rects.size(); ++i)
{
const double area = (truth_object_detections[idx][i].get_rect().intersect(mapped_rects[i])).area();
const double total_area = (truth_object_detections[idx][i].get_rect() + mapped_rects[i]).area();
if (area/total_area <= match_eps)
{
using namespace std;
ostringstream sout;
sout << "An impossible set of object labels was detected. This is happening because ";
sout << "none of the object locations checked by the supplied image scanner is a close ";
sout << "enough match to one of the truth boxes in your training dataset. To resolve this ";
sout << "you need to either lower the match_eps, adjust the settings of the image scanner ";
sout << "so that it is capable of hitting this truth box, or adjust the offending truth rectangle so it ";
sout << "can be matched by the current image scanner. Also, if you ";
sout << "are using the scan_fhog_pyramid object then you could try using a finer image pyramid. ";
sout << "Additionally, the scan_fhog_pyramid scans a fixed aspect ratio box across the image when it ";
sout << "searches for objects. So if you are getting this error and you are using the scan_fhog_pyramid, ";
sout << "it's very likely the problem is that your training dataset contains truth rectangles of widely ";
sout << "varying aspect ratios. The solution is to make sure your training boxes all have about the same aspect ratio. ";
// make sure the above string fits nicely into a command prompt window.
string temp = sout.str();
sout.str(""); sout << wrap_string(temp,0,0) << endl << endl;
sout << "image index "<< idx << endl;
sout << "match_eps: "<< match_eps << endl;
sout << "best possible match: "<< area/total_area << endl;
sout << "truth rect: "<< truth_object_detections[idx][i].get_rect() << endl;
sout << "truth rect width/height: "<< truth_object_detections[idx][i].get_rect().width()/(double)truth_object_detections[idx][i].get_rect().height() << endl;
sout << "truth rect area: "<< truth_object_detections[idx][i].get_rect().area() << endl;
sout << "nearest detection template rect: "<< mapped_rects[i] << endl;
sout << "nearest detection template rect width/height: "<< mapped_rects[i].width()/(double)mapped_rects[i].height() << endl;
sout << "nearest detection template rect area: "<< mapped_rects[i].area() << endl;
throw dlib::impossible_labeling_error(sout.str());
}
}
}
virtual void separation_oracle (
const long idx,
const matrix_type& current_solution,
scalar_type& loss,
feature_vector_type& psi
) const
{
const image_scanner_type& scanner = scanners[idx];
std::vector<std::pair<double, rectangle> > dets;
const double thresh = current_solution(scanner.get_num_dimensions());
scanner.detect(current_solution, dets, thresh-loss_per_false_alarm);
// The loss will measure the number of incorrect detections. A detection is
// incorrect if it doesn't hit a truth rectangle or if it is a duplicate detection
// on a truth rectangle.
loss = truth_object_detections[idx].size()*loss_per_missed_target;
// Measure the loss augmented score for the detections which hit a truth rect.
std::vector<double> truth_score_hits(truth_object_detections[idx].size(), 0);
// keep track of which truth boxes we have hit so far.
std::vector<bool> hit_truth_table(truth_object_detections[idx].size(), false);
std::vector<rectangle> final_dets;
// The point of this loop is to fill out the truth_score_hits array.
for (unsigned long i = 0; i < dets.size() && final_dets.size() < max_num_dets; ++i)
{
if (overlaps_any_box(boxes_overlap, final_dets, dets[i].second))
continue;
const std::pair<double,unsigned int> truth = find_best_match(truth_object_detections[idx], dets[i].second);
final_dets.push_back(dets[i].second);
const double truth_match = truth.first;
// if hit truth rect
if (truth_match > match_eps)
{
// if this is the first time we have seen a detect which hit truth_object_detections[idx][truth.second]
const double score = dets[i].first - thresh;
if (hit_truth_table[truth.second] == false)
{
hit_truth_table[truth.second] = true;
truth_score_hits[truth.second] += score;
}
else
{
truth_score_hits[truth.second] += score + loss_per_false_alarm;
}
}
}
hit_truth_table.assign(hit_truth_table.size(), false);
final_dets.clear();
#ifdef ENABLE_ASSERTS
double total_score = 0;
#endif
// Now figure out which detections jointly maximize the loss and detection score sum. We
// need to take into account the fact that allowing a true detection in the output, while
// initially reducing the loss, may allow us to increase the loss later with many duplicate
// detections.
for (unsigned long i = 0; i < dets.size() && final_dets.size() < max_num_dets; ++i)
{
if (overlaps_any_box(boxes_overlap, final_dets, dets[i].second))
continue;
const std::pair<double,unsigned int> truth = find_best_match(truth_object_detections[idx], dets[i].second);
const double truth_match = truth.first;
if (truth_match > match_eps)
{
if (truth_score_hits[truth.second] > loss_per_missed_target)
{
if (!hit_truth_table[truth.second])
{
hit_truth_table[truth.second] = true;
final_dets.push_back(dets[i].second);
#ifdef ENABLE_ASSERTS
total_score += dets[i].first;
#endif
loss -= loss_per_missed_target;
}
else
{
final_dets.push_back(dets[i].second);
#ifdef ENABLE_ASSERTS
total_score += dets[i].first;
#endif
loss += loss_per_false_alarm;
}
}
}
else if (!overlaps_ignore_box(idx,dets[i].second))
{
// didn't hit anything
final_dets.push_back(dets[i].second);
#ifdef ENABLE_ASSERTS
total_score += dets[i].first;
#endif
loss += loss_per_false_alarm;
}
}
psi.set_size(get_num_dimensions());
psi = 0;
for (unsigned long i = 0; i < final_dets.size(); ++i)
scanner.get_feature_vector(scanner.get_full_object_detection(final_dets[i], current_solution), psi);
#ifdef ENABLE_ASSERTS
const double psi_score = dot(psi, current_solution);
DLIB_CASSERT(std::abs(psi_score-total_score) <= 1e-4 * std::max(1.0,std::max(std::abs(psi_score),std::abs(total_score))),
"\t The get_feature_vector() and detect() methods of image_scanner_type are not in sync."
<< "\n\t The relative error is too large to be attributed to rounding error."
<< "\n\t error: " << std::abs(psi_score-total_score)
<< "\n\t psi_score: " << psi_score
<< "\n\t total_score: " << total_score
);
#endif
psi(scanner.get_num_dimensions()) = -1.0*final_dets.size();
}
bool overlaps_ignore_box (
const long idx,
const dlib::rectangle& rect
) const
{
for (unsigned long i = 0; i < ignore[idx].size(); ++i)
{
if (ignore_overlap_tester(ignore[idx][i], rect))
return true;
}
return false;
}
std::pair<double,unsigned int> find_best_match(
const std::vector<full_object_detection>& boxes,
const rectangle rect
) const
/*!
ensures
- determines which rectangle in boxes matches rect the most and
returns the amount of this match. Specifically, the match is
a number O with the following properties:
- 0 <= O <= 1
- Let R be the maximum matching rectangle in boxes, then
O == (R.intersect(rect)).area() / (R + rect).area()
- O == 0 if there is no match with any rectangle.
!*/
{
double match = 0;
unsigned int best_idx = 0;
for (unsigned long i = 0; i < boxes.size(); ++i)
{
const unsigned long area = rect.intersect(boxes[i].get_rect()).area();
if (area != 0)
{
const double new_match = area / static_cast<double>((rect + boxes[i].get_rect()).area());
if (new_match > match)
{
match = new_match;
best_idx = i;
}
}
}
return std::make_pair(match,best_idx);
}
struct init_scanners_helper
{
init_scanners_helper (
array<image_scanner_type>& scanners_,
const image_array_type& images_
) :
scanners(scanners_),
images(images_)
{}
array<image_scanner_type>& scanners;
const image_array_type& images;
void operator() (long i ) const
{
scanners[i].load(images[i]);
}
};
void initialize_scanners (
const image_scanner_type& scanner,
unsigned long num_threads
)
{
scanners.set_max_size(images.size());
scanners.set_size(images.size());
for (unsigned long i = 0; i < scanners.size(); ++i)
scanners[i].copy_configuration(scanner);
// now load the images into all the scanners
parallel_for(num_threads, 0, scanners.size(), init_scanners_helper(scanners, images));
}
test_box_overlap boxes_overlap;
mutable array<image_scanner_type> scanners;
const image_array_type& images;
const std::vector<std::vector<full_object_detection> >& truth_object_detections;
const std::vector<std::vector<rectangle> >& ignore;
const test_box_overlap ignore_overlap_tester;
unsigned long max_num_dets;
double match_eps;
double loss_per_false_alarm;
double loss_per_missed_target;
};
// ----------------------------------------------------------------------------------------
}
#endif // DLIB_STRUCTURAL_SVM_ObJECT_DETECTION_PROBLEM_Hh_