parsing.cc

   1
   2 ///////////////////////////////////////////////////////////////////////////
   3 // This program is free software: you can redistribute it and/or modify  //
   4 // it under the terms of the version 3 of the GNU General Public License //
   5 // as published by the Free Software Foundation.                         //
   6 //                                                                       //
   7 // This program is distributed in the hope that it will be useful, but   //
   8 // WITHOUT ANY WARRANTY; without even the implied warranty of            //
   9 // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU      //
  10 // General Public License for more details.                              //
  11 //                                                                       //
  12 // You should have received a copy of the GNU General Public License     //
  13 // along with this program. If not, see <http://www.gnu.org/licenses/>.  //
  14 //                                                                       //
  15 // Written by Francois Fleuret                                           //
  16 // (C) Idiap Research Institute                                          //
  17 //                                                                       //
  18 // Contact <francois.fleuret@idiap.ch> for comments & bug reports        //
  19 ///////////////////////////////////////////////////////////////////////////
  20
  21 #include "parsing.h"
  22 #include "fusion_sort.h"
  23
  24 Parsing::Parsing(LabelledImagePool *image_pool,
  25                  PoseCellHierarchy *hierarchy,
  26                  scalar_t proportion_negative_cells,
  27                  int image_index) {
  28
  29   _image_pool = image_pool;
  30   _image_index = image_index;
  31
  32   PoseCellSet cell_set;
  33   LabelledImage *image;
  34
  35   image = _image_pool->grab_image(_image_index);
  36
  37   hierarchy->add_root_cells(image, &cell_set);
  38
  39   int *kept = new int[cell_set.nb_cells()];
  40
  41   _nb_cells = 0;
  42
  43   for(int c = 0; c < cell_set.nb_cells(); c++) {
  44     int l = image->pose_cell_label(cell_set.get_cell(c));
  45     kept[c] = (l > 0) || (l < 0 && drand48() < proportion_negative_cells);
  46     if(kept[c]) _nb_cells++;
  47   }
  48
  49   _cells = new PoseCell[_nb_cells];
  50   _responses = new scalar_t[_nb_cells];
  51   _labels = new int[_nb_cells];
  52   _nb_positives = 0;
  53   _nb_negatives = 0;
  54
  55   int d = 0;
  56   for(int c = 0; c < cell_set.nb_cells(); c++) {
  57     if(kept[c]) {
  58       _cells[d] = *(cell_set.get_cell(c));
  59       _labels[d] = image->pose_cell_label(&_cells[d]);
  60       _responses[d] = 0;
  61       if(_labels[d] < 0) {
  62         _nb_negatives++;
  63       } else if(_labels[d] > 0) {
  64         _nb_positives++;
  65       }
  66       d++;
  67     }
  68   }
  69
  70   delete[] kept;
  71
  72   _image_pool->release_image(_image_index);
  73 }
  74
  75 Parsing::~Parsing() {
  76   delete[] _cells;
  77   delete[] _responses;
  78   delete[] _labels;
  79 }
  80
  81 void Parsing::down_one_level(PoseCellHierarchy *hierarchy,
  82                              int level, int *sample_nb_occurences, scalar_t *sample_responses) {
  83   PoseCellSet cell_set;
  84   LabelledImage *image;
  85
  86   int new_nb_cells = 0;
  87   for(int c = 0; c < _nb_cells; c++) {
  88     new_nb_cells += sample_nb_occurences[c];
  89   }
  90
  91   PoseCell *new_cells = new PoseCell[new_nb_cells];
  92   scalar_t *new_responses = new scalar_t[new_nb_cells];
  93   int *new_labels = new int[new_nb_cells];
  94
  95   image = _image_pool->grab_image(_image_index);
  96   int b = 0;
  97
  98   for(int c = 0; c < _nb_cells; c++) {
  99
 100     if(sample_nb_occurences[c] > 0) {
 101
 102       cell_set.erase_content();
 103       hierarchy->add_subcells(level, _cells + c, &cell_set);
 104
 105       if(_labels[c] > 0) {
 106         ASSERT(sample_nb_occurences[c] == 1);
 107         int e = -1;
 108         for(int d = 0; d < cell_set.nb_cells(); d++) {
 109           if(image->pose_cell_label(cell_set.get_cell(d)) > 0) {
 110             ASSERT(e < 0);
 111             e = d;
 112           }
 113         }
 114         ASSERT(e >= 0);
 115         ASSERT(b < new_nb_cells);
 116         new_cells[b] = *(cell_set.get_cell(e));
 117         new_responses[b] = sample_responses[c];
 118         new_labels[b] = 1;
 119         b++;
 120       }
 121
 122       else if(_labels[c] < 0) {
 123         for(int d = 0; d < sample_nb_occurences[c]; d++) {
 124           ASSERT(b < new_nb_cells);
 125           new_cells[b] = *(cell_set.get_cell(int(drand48() * cell_set.nb_cells())));
 126           new_responses[b] = sample_responses[c];
 127           new_labels[b] = -1;
 128           b++;
 129         }
 130       }
 131
 132       else {
 133         cerr << "INCONSISTENCY" << endl;
 134         abort();
 135       }
 136     }
 137   }
 138
 139   ASSERT(b == new_nb_cells);
 140
 141   _image_pool->release_image(_image_index);
 142
 143   delete[] _cells;
 144   delete[] _labels;
 145   delete[] _responses;
 146   _nb_cells = new_nb_cells;
 147   _cells = new_cells;
 148   _labels = new_labels;
 149   _responses = new_responses;
 150 }
 151
 152 void Parsing::update_cell_responses(PiFeatureFamily *pi_feature_family,
 153                                     Classifier *classifier) {
 154   LabelledImage *image;
 155
 156   image = _image_pool->grab_image(_image_index);
 157   image->compute_rich_structure();
 158
 159   SampleSet *samples = new SampleSet(pi_feature_family->nb_features(), 1);
 160
 161   for(int c = 0; c < _nb_cells; c++) {
 162     samples->set_sample(0, pi_feature_family, image, &_cells[c], 0);
 163     _responses[c] += classifier->response(samples, 0);
 164     ASSERT(!isnan(_responses[c]));
 165   }
 166
 167   _image_pool->release_image(_image_index);
 168   delete samples;
 169 }
 170
 171 void Parsing::collect_samples(SampleSet *samples,
 172                               PiFeatureFamily *pi_feature_family,
 173                               int s,
 174                               int *to_collect) {
 175   LabelledImage *image;
 176
 177   image = _image_pool->grab_image(_image_index);
 178   image->compute_rich_structure();
 179
 180   for(int c = 0; c < _nb_cells; c++) {
 181     if(to_collect[c]) {
 182       samples->set_sample(s, pi_feature_family, image, &_cells[c], _labels[c]);
 183       s++;
 184     }
 185   }
 186
 187   _image_pool->release_image(_image_index);
 188 }