NICE_SemSeg/progs/testSemanticSegmentation3D.cpp

// Beispielhafter Aufruf: BUILD_x86_64/progs/testSemanticSegmentation -config <CONFIGFILE>

/**
* @file testSemanticSegmentation.cpp
* @brief test semantic segmentation routines for 3d images and 2d images
* @author Erik Rodner, Björn Fröhlich, Sven Sickert
* @date 03/20/2008
*/

#ifdef NICE_USELIB_OPENMP
#include <omp.h>
#endif

#include "core/basics/Config.h"
#include "core/basics/StringTools.h"
#include "vislearning/baselib/ICETools.h"

#include "core/image/MultiChannelImage3DT.h"
#include "semseg/semseg/SemSegContextTree3D.h"

#include "core/basics/ResourceStatistics.h"
#include "core/image/Morph.h"

#include <fstream>
#include <vector>

#undef DEBUG

using namespace OBJREC;

using namespace NICE;

using namespace std;

void segmentToOverlay ( const NICE::Image *orig, const NICE::ColorImage & segment,
                        NICE::ColorImage & result )
{
  int xsize = orig->width();
  int ysize = orig->height();
  
  result.resize( xsize, ysize );
  vector< NICE::MatrixT<double> > channelMat;
  
  double alpha = .5;
  
  for (int c = 0; c < 3; c++)
  {
    NICE::MatrixT<double> chan ( xsize, ysize );
    channelMat.push_back( chan );
  }

  for (int y = 0; y < ysize; y++)
  {
    for (int x = 0; x < xsize; x++)
    {
      uchar val = orig->getPixelQuick(x,y);
      for (int c = 0; c < 3; c++)
        channelMat[c](x,y) = (double)val + alpha*(double)segment.getPixel( x, y, c );
    }
  }
  
  for (int c = 0; c < 3; c++)
  {
    channelMat[c] /= channelMat[c].Max();
    channelMat[c] *= 255;
  }
  
  for (int y = 0; y < ysize; y++)
  {
    for (int x = 0; x < xsize; x++)
    {
      for (int c = 0; c < 3; c++)
      {
        int val = channelMat[c](x,y);
        result.setPixel( x, y, c, (uchar)val);
      }
    }
  }
}

void updateMatrix ( const NICE::Image & img,
                    const NICE::Image & gt,
                    NICE::Matrix & M,
                    const set<int> & forbidden_classes )
{
  double subsamplex = gt.width() / ( double ) img.width();
  double subsampley = gt.height() / ( double ) img.height();

  for ( int y = 0 ; y < gt.height() ; y++ )
    for ( int x = 0 ; x < gt.width() ; x++ )
    {
      int xx = ( int ) ( x / subsamplex );
      int yy = ( int ) ( y / subsampley );

      if ( xx < 0 ) xx = 0;

      if ( yy < 0 ) yy = 0;

      if ( xx > img.width() - 1 ) xx = img.width() - 1;

      if ( yy > img.height() - 1 ) yy = img.height() - 1;

      int cimg = img.getPixel ( xx, yy );

      int gimg = gt.getPixel ( x, y );

      if ( forbidden_classes.find ( gimg ) == forbidden_classes.end() )
      {
        M ( gimg, cimg ) ++;
      }
    }
}

void startClassification (SemanticSegmentation *semseg,
                          std::vector< NICE::Matrix > & M_vec,
                          const Config & conf,
                          const LabeledSet* testFiles,
                          const ClassNames & classNames,
                          const set<int> & forbidden_classes,
                          const string & resultdir,
                          const bool doCrossVal)
{
  bool show_results = conf.gB ( "debug", "show_results", false );
  bool write_results = conf.gB ( "debug", "write_results", false );
  if (doCrossVal)
    write_results = false;

  bool run_3Dseg = conf.gB( "SSContextTree", "run_3dseg", false);
  bool postProcessing = conf.gB( "main", "post_process", false);
  string output_type = conf.gS ( "debug", "output_type", "ppm" );
  string output_postfix = conf.gS ( "debug", "output_postfix", "" );

  vector< int > zsizeVec;
  semseg->getDepthVector ( testFiles, zsizeVec, run_3Dseg );
  int depthCount = 0, idx = 0;
  vector< string > filelist;
  NICE::MultiChannelImageT<double> segresult;
  NICE::MultiChannelImageT<double> gt;

  for (LabeledSet::const_iterator it = testFiles->begin(); it != testFiles->end(); it++)
  {
    for (std::vector<ImageInfo *>::const_iterator jt = it->second.begin();
         jt != it->second.end(); jt++)
    {
      ImageInfo & info = *(*jt);
      std::string file = info.img();
      filelist.push_back ( file );
      depthCount++;

      NICE::Image lm;
      NICE::Image lm_gt;
      if ( info.hasLocalizationInfo() )
      {
        const LocalizationResult *l_gt = info.localization();

        lm.resize ( l_gt->xsize, l_gt->ysize );
        lm.set ( 0 );

        lm_gt.resize ( l_gt->xsize, l_gt->ysize );
        lm_gt.set ( 0 );

        l_gt->calcLabeledImage ( lm, classNames.getBackgroundClass() );
#ifdef DEBUG
        cout << "testSemanticSegmentation3D: Generating Labeled NICE::Image (Ground-Truth)" << endl;
#endif
        l_gt->calcLabeledImage ( lm_gt, classNames.getBackgroundClass() );
      }
      segresult.addChannel ( lm );
      gt.addChannel ( lm_gt );

      int depthBoundary = 0;
      if ( run_3Dseg )
      {
        depthBoundary = zsizeVec[idx];
      }

      if ( depthCount < depthBoundary ) continue;

      NICE::MultiChannelImage3DT<double> probabilities;

      semseg->classify ( filelist, segresult, probabilities );

      // save to file
      for ( int z = 0; z < segresult.channels(); z++ )
      {
        std::string fname = StringTools::baseName ( filelist[z], false );

        if ( show_results || write_results )
        {
          NICE::ColorImage orig ( filelist[z] );
          NICE::ColorImage rgb;
          NICE::ColorImage rgb_gt;
          NICE::ColorImage ov_rgb;
          NICE::ColorImage ov_rgb_gt;

          for ( int y = 0 ; y < orig.height(); y++ )
          {
            for ( int x = 0 ; x < orig.width(); x++ )
            {
              lm.setPixel ( x, y, segresult.get ( x, y, ( uint ) z ) );
              if ( run_3Dseg )
                lm_gt.setPixel ( x, y, gt.get ( x, y, ( uint ) z ) );
            }
          }

          // confusion matrix
          NICE::Matrix M ( classNames.getMaxClassno() + 1, classNames.getMaxClassno() + 1 );
          M.set ( 0 );
          updateMatrix ( lm, lm_gt, M, forbidden_classes );
          M_vec.push_back ( M );

          classNames.labelToRGB ( lm, rgb );
          classNames.labelToRGB ( lm_gt, rgb_gt );

          if (postProcessing)
          {
            // median filter
            for (int r = 0; r < 3; r++)
            {
              NICE::Image postIm(rgb.width(), rgb.height());
              NICE::median(*(rgb.getChannel(r)), &postIm, 1);
              for (int y = 0; y < rgb.height(); y++)
                for (int x = 0; x < rgb.width(); x++)
                  rgb.setPixel(x,y,r, postIm.getPixelQuick(x,y));
            }
          }

          if ( write_results )
          {
            segmentToOverlay ( orig.getChannel(1), rgb, ov_rgb );
            segmentToOverlay ( orig.getChannel(1), rgb_gt, ov_rgb_gt );

            std::stringstream out;
            if ( output_postfix.size() > 0 )
              out << resultdir << "/" << fname << output_postfix;
            else
              out << resultdir << "/" << fname;
#ifdef DEBUG
            cout << "Writing to file " << out.str() << "_*." << output_type << endl;
#endif
            orig.write ( out.str() + "_orig." + output_type );
            rgb.write ( out.str() + "_result." + output_type );
            rgb_gt.write ( out.str() + "_groundtruth." + output_type );
            ov_rgb.write ( out.str() + "_overlay_res." + output_type );
            ov_rgb_gt.write ( out.str() + "_overlay_gt." + output_type );

            // write Probability maps
            NICE::ColorImage prob_map( probabilities.width(), probabilities.height() );
            prob_map.set(0,0,0);
            int iNumChannels = probabilities.channels();
            for ( int idxProbMap = 0; idxProbMap < iNumChannels; idxProbMap++)
            {
                for ( int y = 0 ; y < probabilities.height(); y++ )
                {
                    for ( int x = 0 ; x < probabilities.width(); x++ )
                    {
                        double probVal = probabilities.get( x, y, z, idxProbMap ) * 255.0;
                        int tmp = round(probVal);
                        for ( int c = 0 ; c < 3 ; c++ )
                            prob_map.setPixel( x, y, c, tmp );
                    }
                }
                std::stringstream ssFileProbMap;
                //ssFileProbMap << out.str() << "_probs." << "c" << idxProbMap << "." << output_type;
                ssFileProbMap << out.str() << "_probs." << "c-" << classNames.code( idxProbMap ) << "." << output_type;
                //classNames
                prob_map.write ( ssFileProbMap.str() );
            }
          }
        }
      }

      // prepare for new 3d image
      filelist.clear();
      segresult.reInit(0,0,0);
      gt.reInit(0,0,0);
      depthCount = 0;
      idx++;
    }
  }

  segresult.freeData();
}

/**
 test semantic segmentation routines
*/
int main ( int argc, char **argv )
{
  std::set_terminate ( __gnu_cxx::__verbose_terminate_handler );

  Config conf ( argc, argv );

  ResourceStatistics rs;

  /*---------------CONFIGURATION---------------*/
  bool doCrossVal = conf.gB ( "debug", "do_crossval", false );
  string resultdir = conf.gS ( "debug", "resultdir", "." );
  /*-------------------------------------------*/

#ifdef DEBUG
  cerr << "Writing Results to " << resultdir << endl;
#endif

  std::vector< NICE::Matrix > M_vec;

  MultiDataset md ( &conf );

  const ClassNames & classNames = md.getClassNames ( "train" );
  set<int> forbidden_classes;
  classNames.getSelection ( conf.gS ( "analysis", "forbidden_classes", "" ),
                            forbidden_classes );

  // initialize semantic segmentation method
  SemanticSegmentation *semseg = NULL;
  
  // TRAINING AND TESTING
  if (!doCrossVal)
  {
    semseg = new SemSegContextTree3D ( &conf, &classNames );

    // STANDARD EVALUATION
    cout << "\nTRAINING" << endl;
    cout << "########\n" << endl;
    semseg->train( &md );

    cout << "\nCLASSIFICATION" << endl;
    cout << "##############\n" << endl;
    const LabeledSet *testFiles = md["test"];
    startClassification (semseg, M_vec, conf, testFiles, classNames,
                         forbidden_classes, resultdir, doCrossVal );

    delete semseg;
  }
  else
  {
    // CROSS-VALIDATION
    for (int cval = 1; cval <= 10; cval++)
    {
      semseg = new SemSegContextTree3D ( &conf, &classNames );

      stringstream ss;
      ss << cval;
      string cvaltrain = "train_cv" + ss.str();
      string cvaltest = "test_cv" + ss.str();

      cout << "\nTRAINING " << cval << endl;
      cout << "###########\n" << endl;
      const LabeledSet *trainFiles = md[cvaltrain];
      semseg->train( trainFiles );

      cout << "\nCLASSIFICATION " << cval << endl;
      cout << "#################\n" << endl;
      const LabeledSet *testFiles = md[cvaltest];
      startClassification (semseg, M_vec, conf, testFiles, classNames,
                           forbidden_classes, resultdir, doCrossVal );

      delete semseg;
    }
  }

  cout << "\nSTATISTICS" << endl;
  cout << "##########\n" << endl;

  long maxMemory;
  double userCPUTime, sysCPUTime;
  rs.getStatistics ( maxMemory, userCPUTime, sysCPUTime );
  cout << "Memory (max):    " << maxMemory << " KB" << endl;
  cout << "CPU Time (user): " << userCPUTime << " seconds" << endl;
  cout << "CPU Time (sys):  " << sysCPUTime << " seconds" << endl;

  cout << "\nPERFORMANCE" << endl;
  cout << "###########\n" << endl;

  double overall = 0.0;
  double sumall = 0.0;

  NICE::Matrix M ( classNames.getMaxClassno() + 1, classNames.getMaxClassno() + 1 );
  M.set ( 0 );
  for ( int s = 0; s < ( int ) M_vec.size(); s++ )
  {
    NICE::Matrix M_tmp = M_vec[s];
    for ( int r = 0; r < ( int ) M_tmp.rows(); r++ )
    {
      for ( int c = 0; c < ( int ) M_tmp.cols(); c++ )
      {
        if ( r == c )
          overall += M_tmp ( r, c );

        sumall += M_tmp ( r, c );
        M ( r, c ) += M_tmp ( r, c );
      }
    }
  }
  overall /= sumall;

  cout << "Confusion Matrix:" << endl;
  for (int r = 0; r < (int) M.rows(); r++)
  {
    for (int c = 0; c < (int) M.cols(); c++)
    {
      cout << M(r,c) << "  ";
    }
    cout << endl;
  }

  // metrics for binary classification
  double precision, recall, f1score = -1.0;
  if (classNames.getMaxClassno()+1 == 2)
  {
    precision = (double)M(1,1) / (double)(M(1,1)+M(0,1));
    recall = (double)M(1,1) / (double)(M(1,1)+M(1,0));
    f1score = 2.0*(precision*recall)/(precision+recall);
  }

  // normalizing M using rows
  for ( int r = 0 ; r < ( int ) M.rows() ; r++ )
  {
    double sum = 0.0;

    for ( int c = 0 ; c < ( int ) M.cols() ; c++ )
      sum += M ( r, c );

    if ( fabs ( sum ) > 1e-4 )
      for ( int c = 0 ; c < ( int ) M.cols() ; c++ )
        M ( r, c ) /= sum;
  }

  double avg_perf = 0.0;
  int classes_trained = 0;

  for ( int r = 0 ; r < ( int ) M.rows() ; r++ )
  {
    if ( ( classNames.existsClassno ( r ) ) && ( forbidden_classes.find ( r ) == forbidden_classes.end() ) )
    {
      avg_perf += M ( r, r );
      double lsum = 0.0;
      for ( int r2 = 0; r2 < ( int ) M.rows(); r2++ )
      {
        lsum += M ( r,r2 );
      }
      if ( lsum != 0.0 )
      {
        classes_trained++;
      }
    }
  }

  // print results of evaluation
  cout << "Overall Recogntion Rate: " << overall << endl;
  cout << "Average Recogntion Rate: " << avg_perf / ( classes_trained ) << endl;
  cout << "Lower Bound: " << 1.0  / classes_trained << endl;
  cout << "Precision: " << precision << endl;
  cout << "Recall: " << recall << endl;
  cout << "F1Score: " << f1score << endl;

  cout <<"\nClasses:" << endl;
  for ( int r = 0 ; r < ( int ) M.rows() ; r++ )
  {
    if ( ( classNames.existsClassno ( r ) ) && ( forbidden_classes.find ( r ) == forbidden_classes.end() ) )
    {
      std::string classname = classNames.text ( r );
      cout << classname.c_str() << ": " << M ( r, r ) << endl;
    }
  }

  return 0;
}