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, (C) IDIAP                                //
  16 // Contact <francois.fleuret@idiap.ch> for comments & bug reports        //
  17 ///////////////////////////////////////////////////////////////////////////
  18
  19 #include "parsing.h"
  20 #include "fusion_sort.h"
  21
  22 Parsing::Parsing(LabelledImagePool *image_pool,
  23                  PoseCellHierarchy *hierarchy,
  24                  scalar_t proportion_negative_cells,
  25                  int image_index) {
  26
  27   _image_pool = image_pool;
  28   _image_index = image_index;
  29
  30   PoseCellSet cell_set;
  31   LabelledImage *image;
  32
  33   image = _image_pool->grab_image(_image_index);
  34
  35   hierarchy->add_root_cells(image, &cell_set);
  36
  37   int *kept = new int[cell_set.nb_cells()];
  38
  39   _nb_cells = 0;
  40
  41   for(int c = 0; c < cell_set.nb_cells(); c++) {
  42     int l = image->pose_cell_label(cell_set.get_cell(c));
  43     kept[c] = (l > 0) || (l < 0 && drand48() < proportion_negative_cells);
  44     if(kept[c]) _nb_cells++;
  45   }
  46
  47   _cells = new PoseCell[_nb_cells];
  48   _responses = new scalar_t[_nb_cells];
  49   _labels = new int[_nb_cells];
  50   _nb_positives = 0;
  51   _nb_negatives = 0;
  52
  53   int d = 0;
  54   for(int c = 0; c < cell_set.nb_cells(); c++) {
  55     if(kept[c]) {
  56       _cells[d] = *(cell_set.get_cell(c));
  57       _labels[d] = image->pose_cell_label(&_cells[d]);
  58       _responses[d] = 0;
  59       if(_labels[d] < 0) {
  60         _nb_negatives++;
  61       } else if(_labels[d] > 0) {
  62         _nb_positives++;
  63       }
  64       d++;
  65     }
  66   }
  67
  68   delete[] kept;
  69
  70   _image_pool->release_image(_image_index);
  71 }
  72
  73 Parsing::~Parsing() {
  74   delete[] _cells;
  75   delete[] _responses;
  76   delete[] _labels;
  77 }
  78
  79 void Parsing::down_one_level(PoseCellHierarchy *hierarchy,
  80                              int level, int *sample_nb_occurences, scalar_t *sample_responses) {
  81   PoseCellSet cell_set;
  82   LabelledImage *image;
  83
  84   int new_nb_cells = 0;
  85   for(int c = 0; c < _nb_cells; c++) {
  86     new_nb_cells += sample_nb_occurences[c];
  87   }
  88
  89   PoseCell *new_cells = new PoseCell[new_nb_cells];
  90   scalar_t *new_responses = new scalar_t[new_nb_cells];
  91   int *new_labels = new int[new_nb_cells];
  92
  93   image = _image_pool->grab_image(_image_index);
  94   int b = 0;
  95
  96   for(int c = 0; c < _nb_cells; c++) {
  97
  98     if(sample_nb_occurences[c] > 0) {
  99
 100       cell_set.erase_content();
 101       hierarchy->add_subcells(level, _cells + c, &cell_set);
 102
 103       if(_labels[c] > 0) {
 104         ASSERT(sample_nb_occurences[c] == 1);
 105         int e = -1;
 106         for(int d = 0; d < cell_set.nb_cells(); d++) {
 107           if(image->pose_cell_label(cell_set.get_cell(d)) > 0) {
 108             ASSERT(e < 0);
 109             e = d;
 110           }
 111         }
 112         ASSERT(e >= 0);
 113         ASSERT(b < new_nb_cells);
 114         new_cells[b] = *(cell_set.get_cell(e));
 115         new_responses[b] = sample_responses[c];
 116         new_labels[b] = 1;
 117         b++;
 118       }
 119
 120       else if(_labels[c] < 0) {
 121         for(int d = 0; d < sample_nb_occurences[c]; d++) {
 122           ASSERT(b < new_nb_cells);
 123           new_cells[b] = *(cell_set.get_cell(int(drand48() * cell_set.nb_cells())));
 124           new_responses[b] = sample_responses[c];
 125           new_labels[b] = -1;
 126           b++;
 127         }
 128       }
 129
 130       else {
 131         cerr << "INCONSISTENCY" << endl;
 132         abort();
 133       }
 134     }
 135   }
 136
 137   ASSERT(b == new_nb_cells);
 138
 139   _image_pool->release_image(_image_index);
 140
 141   delete[] _cells;
 142   delete[] _labels;
 143   delete[] _responses;
 144   _nb_cells = new_nb_cells;
 145   _cells = new_cells;
 146   _labels = new_labels;
 147   _responses = new_responses;
 148 }
 149
 150 void Parsing::update_cell_responses(PiFeatureFamily *pi_feature_family,
 151                                     Classifier *classifier) {
 152   LabelledImage *image;
 153
 154   image = _image_pool->grab_image(_image_index);
 155   image->compute_rich_structure();
 156
 157   SampleSet *samples = new SampleSet(pi_feature_family->nb_features(), 1);
 158
 159   for(int c = 0; c < _nb_cells; c++) {
 160     samples->set_sample(0, pi_feature_family, image, &_cells[c], 0);
 161     _responses[c] += classifier->response(samples, 0);
 162     ASSERT(!isnan(_responses[c]));
 163   }
 164
 165   _image_pool->release_image(_image_index);
 166   delete samples;
 167 }
 168
 169 void Parsing::collect_samples(SampleSet *samples,
 170                               PiFeatureFamily *pi_feature_family,
 171                               int s,
 172                               int *to_collect) {
 173   LabelledImage *image;
 174
 175   image = _image_pool->grab_image(_image_index);
 176   image->compute_rich_structure();
 177
 178   for(int c = 0; c < _nb_cells; c++) {
 179     if(to_collect[c]) {
 180       samples->set_sample(s, pi_feature_family, image, &_cells[c], _labels[c]);
 181       s++;
 182     }
 183   }
 184
 185   _image_pool->release_image(_image_index);
 186 }