CGridMapAlignerApp.cpp

Go to the documentation of this file.

/* +------------------------------------------------------------------------+
   |                     Mobile Robot Programming Toolkit (MRPT)            |
   |                          https://www.mrpt.org/                         |
   |                                                                        |
   | Copyright (c) 2005-2020, Individual contributors, see AUTHORS file     |
   | See: https://www.mrpt.org/Authors - All rights reserved.               |
   | Released under BSD License. See: https://www.mrpt.org/License          |
   +------------------------------------------------------------------------+ */

#include "apps-precomp.h"  // Precompiled headers

#include <mrpt/apps/CGridMapAlignerApp.h>

#include <mrpt/3rdparty/tclap/CmdLine.h>
#include <mrpt/config/CConfigFile.h>
#include <mrpt/gui.h>
#include <mrpt/io/CFileGZInputStream.h>
#include <mrpt/io/CFileGZOutputStream.h>
#include <mrpt/io/CFileOutputStream.h>
#include <mrpt/io/CFileStream.h>
#include <mrpt/maps/CMultiMetricMap.h>
#include <mrpt/maps/CSimpleMap.h>
#include <mrpt/maps/CSimplePointsMap.h>
#include <mrpt/math/ops_containers.h>
#include <mrpt/math/wrap2pi.h>
#include <mrpt/obs/CSensoryFrame.h>
#include <mrpt/opengl/CGridPlaneXY.h>
#include <mrpt/opengl/CSetOfLines.h>
#include <mrpt/poses/CPoint2D.h>
#include <mrpt/poses/CPosePDFGaussian.h>
#include <mrpt/poses/CPosePDFParticles.h>
#include <mrpt/random.h>
#include <mrpt/serialization/CArchive.h>
#include <mrpt/system/datetime.h>
#include <mrpt/system/filesystem.h>
#include <mrpt/system/os.h>
#include <mrpt/system/vector_loadsave.h>

using namespace mrpt::apps;

CGridMapAlignerApp::CGridMapAlignerApp()
    : mrpt::system::COutputLogger("CGridMapAlignerApp")
{
}

void CGridMapAlignerApp::initialize(int argc, const char** argv)
{
    MRPT_START

    // Declare the supported options.
    TCLAP::CmdLine cmd(
        "grid-matching", ' ', mrpt::system::MRPT_getVersion().c_str());

    TCLAP::SwitchArg arg_match(
        "m", "match", "Operation: match two maps", cmd, false);
    TCLAP::SwitchArg arg_detect(
        "d", "detect-test", "Operation: Quality of match with one map", cmd,
        false);

    TCLAP::ValueArg<std::string> arg_filgrid1(
        "1", "map1", "Map #1 to align (*.simplemap)", true, "",
        "map1.simplemap", cmd);
    TCLAP::ValueArg<std::string> arg_filgrid2(
        "2", "map2", "Map #2 to align (*.simplemap)", false, "",
        "map2.simplemap", cmd);

    TCLAP::ValueArg<std::string> arg_out(
        "o", "out", "Output file for the results", false,
        "gridmatching_out.txt", "result_outfile", cmd);
    TCLAP::ValueArg<std::string> arg_config(
        "c", "config", "Optional config. file with more params", false, "",
        "config.ini", cmd);

    TCLAP::ValueArg<std::string> arg_aligner_method(
        "", "aligner", "The method to use for map aligning", false,
        "amModifiedRANSAC", "[amCorrelation|amRobustMatch|amModifiedRANSAC]",
        cmd);
    TCLAP::ValueArg<std::string> arg_out_dir(
        "", "out-dir", "The output directory", false, "GRID-MATCHING_RESULTS",
        "GRID-MATCHING_RESULTS", cmd);

    TCLAP::SwitchArg arg_savesog3d(
        "3", "save-sog-3d", "Save a 3D view of all the SOG modes", cmd, false);
    TCLAP::SwitchArg arg_savesogall(
        "a", "save-sog-all", "Save all the map overlaps", cmd, false);
    TCLAP::SwitchArg arg_savecorrdists(
        "t", "save-corr-dists", "Save corr & non-corr distances", cmd, false);
    TCLAP::ValueArg<std::string> arg_icpgoodness(
        "i", "save-icp-goodness", "Append all ICP goodness values here", false,
        "", "icp_goodness.txt", cmd);

    TCLAP::ValueArg<double> arg_noise_std_xy(
        "x", "noise-std-xy", "In detect-test mode,std. noise in XY", false, 0,
        "sigma", cmd);
    TCLAP::ValueArg<double> arg_noise_std_phi(
        "p", "noise-std-phi", "In detect-test mode,std. noise in PHI (deg)",
        false, 0, "sigma", cmd);
    TCLAP::ValueArg<double> arg_noise_std_laser(
        "l", "noise-std-laser", "In detect-test mode,std. noise range (m)",
        false, 0, "sigma", cmd);
    TCLAP::ValueArg<unsigned int> arg_niters(
        "N", "iters", "In detect-test mode,number of trials", false, 1,
        "rep.count", cmd);

    TCLAP::ValueArg<double> arg_Ax(
        "X", "Ax", "In detect-test mode, displacement in X (m)", false, 4, "X",
        cmd);
    TCLAP::ValueArg<double> arg_Ay(
        "Y", "Ay", "In detect-test mode, displacement in Y (m)", false, 2, "Y",
        cmd);
    TCLAP::ValueArg<double> arg_Aphi(
        "P", "Aphi", "In detect-test mode, displacement in PHI (deg)", false,
        30, "PHI", cmd);

    TCLAP::SwitchArg arg_noise_pose(
        "O", "noise-pose", "detect-test mode: enable noise in pose", cmd,
        false);
    TCLAP::SwitchArg arg_noise_laser(
        "L", "noise-laser", "detect-test mode: enable noise in laser", cmd,
        false);

    TCLAP::SwitchArg arg_verbose("v", "verbose", "Verbose output", cmd, false);
    TCLAP::SwitchArg arg_nologo(
        "g", "nologo", "skip the logo at startup", cmd, false);
    TCLAP::SwitchArg arg_nosave(
        "n", "nosave", "skip saving map images", cmd, false);
    TCLAP::SwitchArg arg_skip_icp(
        "s", "noicp", "skip ICP optimization stage", cmd, false);
    TCLAP::SwitchArg arg_most_likely(
        "", "most-likely-only",
        "Keep the most-likely Gaussian mode from the SOG", cmd, false);

    // Parse arguments:
    if (!cmd.parse(argc, argv))
        THROW_EXCEPTION("CLI arguments parsing tells we should exit.");

    fil_grid1 = arg_filgrid1.getValue();
    fil_grid2 = arg_filgrid2.getValue();
    OUTPUT_FIL = arg_out.getValue();
    CONFIG_FIL = arg_config.getValue();
    SAVE_ICP_GOODNESS_FIL = arg_icpgoodness.getValue();

    aligner_method =
        mrpt::typemeta::TEnumType<mrpt::slam::CGridMapAligner::TAlignerMethod>::
			name2value(arg_aligner_method.getValue());

    STD_NOISE_XY = arg_noise_std_xy.getValue();
    STD_NOISE_PHI = DEG2RAD(arg_noise_std_phi.getValue());
    STD_NOISE_LASER = arg_noise_std_laser.getValue();
    N_ITERS = arg_niters.getValue();

    GT_Ax = arg_Ax.getValue();
    GT_Ay = arg_Ay.getValue();
    GT_Aphi_rad = DEG2RAD(arg_Aphi.getValue());

    SAVE_SOG_3DSCENE = arg_savesog3d.getValue();
    SAVE_SOG_ALL_MAPS_OVERLAP_HYPOTHESES = arg_savesogall.getValue();
    IS_VERBOSE = arg_verbose.getValue();
    NOSAVE = arg_nosave.getValue();
    SAVE_CORR_AND_NONCORR_DISTS = arg_savecorrdists.getValue();

    if (!arg_nologo.getValue())
    {
        printf(" grid-matching - Part of the MRPT\n");
        printf(
            " MRPT C++ Library: %s - Sources timestamp: %s\n",
            mrpt::system::MRPT_getVersion().c_str(),
            mrpt::system::MRPT_getCompilationDate().c_str());
    }

    SKIP_ICP_STAGE = arg_skip_icp.getValue();
    MOST_LIKELY_SOG_MODE_ONLY = arg_most_likely.getValue();
    NOISE_IN_POSE = arg_noise_pose.getValue();
    NOISE_IN_LASER = arg_noise_laser.getValue();

    RESULTS_DIR = arg_out_dir.getValue();

    is_match = arg_match.getValue();
    is_detect_test = arg_detect.getValue();

    if (((!is_match && !is_detect_test) || (is_match && is_detect_test)) &&
        !SAVE_CORR_AND_NONCORR_DISTS)
    {
        std::cerr << std::endl
                  << "Error: One operation mode 'match' or 'detect-test' or "
                     "'save-corr-dists' must be selected."
                  << std::endl;
        TCLAP::StdOutput so;
        so.usage(cmd);
        THROW_EXCEPTION("Wrong CLI parameters.");
    }

    if (is_match)
    {
        // maps:
        if (!arg_filgrid1.isSet() || !arg_filgrid2.isSet())
        {
            std::cerr << std::endl
                      << "Error: Two maps must be passed: --map1=xxx and "
                         "--map2=xxx"
                      << std::endl;
            TCLAP::StdOutput so;
            so.usage(cmd);
            THROW_EXCEPTION("Wrong CLI parameters.");
        }
    }

    MRPT_END
}

void CGridMapAlignerApp::run()
{
    MRPT_START

    using namespace mrpt::slam;
    using namespace mrpt::maps;
    using namespace mrpt::obs;
    using namespace mrpt::opengl;
    using namespace mrpt::tfest;
    using namespace mrpt::math;
    using namespace mrpt::serialization;
    using namespace mrpt::config;
    using namespace mrpt::gui;
    using namespace mrpt::img;
    using namespace mrpt::io;
    using namespace mrpt::random;
    using namespace mrpt::poses;
    using namespace std;

    // do_grid_align: all that follows
    CImage img1, img2;

    getRandomGenerator().randomize();

    // Grid or simplemaps?
    if (is_match || (!is_detect_test && SAVE_CORR_AND_NONCORR_DISTS))
    {
        string ext1 = mrpt::system::extractFileExtension(fil_grid1, true);
        string ext2 = mrpt::system::extractFileExtension(fil_grid2, true);
        if (ext1 != ext2 || ext1 != "simplemap")
        {
            MRPT_LOG_INFO_STREAM("Map1: " << fil_grid1);
            MRPT_LOG_INFO_STREAM("Map2: " << fil_grid2);
            THROW_EXCEPTION("Both file extensions must be 'simplemap'");
        }
    }
    if (is_detect_test)
    {
        string ext1 = mrpt::system::extractFileExtension(fil_grid1, true);
        if (ext1 != "simplemap")
        {
            MRPT_LOG_INFO_STREAM("Map: " << fil_grid1);
            THROW_EXCEPTION("The file extension must be 'simplemap'");
        }
    }

    CMultiMetricMap the_map1, the_map2;

    TSetOfMetricMapInitializers map1_inits, map2_inits;
    {
        COccupancyGridMap2D::TMapDefinition def;
        def.resolution = 0.05f;
        def.insertionOpts.maxOccupancyUpdateCertainty = 0.8f;
        def.insertionOpts.maxDistanceInsertion = 30;
        map1_inits.push_back(def);
        map2_inits.push_back(def);
    }

    if (!SKIP_ICP_STAGE)
    {
        CSimplePointsMap::TMapDefinition def;
        def.insertionOpts.minDistBetweenLaserPoints = 0.10f;
        def.renderOpts.color = mrpt::img::TColorf(0, 0, 1);
        map1_inits.push_back(def);
        def.renderOpts.color = mrpt::img::TColorf(1, 0, 0);
        map2_inits.push_back(def);
    }

    mrpt::system::deleteFiles(format("%s/*.*", RESULTS_DIR.c_str()));
    mrpt::system::createDirectory(RESULTS_DIR);

    CFileOutputStream f_log(RESULTS_DIR + string("/log.txt"));
    f_log.printf(
        "%% std_noise_xy std_noise_phi std_noise_laser  covDet:mean&std "
        "stdPhi:mean&std error2D:mean&std  errorPhi:mean&phi "
        "bruteErr2D:mean&std brutErrPhi:mean&std GT_likelihood "
        "#GTcorrFound:mean&std \n");
    f_log.printf("%% -----------------------------------------\n");

    CFileStream f_out_log;
    if (!OUTPUT_FIL.empty())
        f_out_log.open(OUTPUT_FIL, fomWrite);  //, fomAppend );

    CGridMapAligner gridAlign;
    gridAlign.setMinLoggingLevel(this->getMinLoggingLevel());

    CGridMapAligner::TReturnInfo info;

    gridAlign.options.methodSelection = aligner_method;

    if (!CONFIG_FIL.empty())
    {
        // Load options for the grid-matching algorithm:
        CConfigFile cfg(CONFIG_FIL);
        gridAlign.options.loadFromConfigFile(cfg, "grid-match");

        // load alternative config. for the metric maps:
        if (cfg.sectionExists("metric_maps"))
        {
            map1_inits.loadFromConfigFile(cfg, "metric_maps");
            map2_inits = map1_inits;
        }
    }

    // Dump params:
    if (IS_VERBOSE)
    {
        gridAlign.options.dumpToConsole();
        map1_inits.dumpToConsole();
    }

    // Create the map with a points & grid-map within:
    the_map1.setListOfMaps(map1_inits);
    the_map2.setListOfMaps(map2_inits);

    COccupancyGridMap2D::Ptr grid1 = the_map1.mapByClass<COccupancyGridMap2D>();
    COccupancyGridMap2D::Ptr grid2 = the_map2.mapByClass<COccupancyGridMap2D>();

    ASSERT_(grid1);
    ASSERT_(grid2);

    // ---------------------------------------------
    //              Options: RANSAC
    // ---------------------------------------------
    gridAlign.options.ransac_SOG_sigma_m = grid1->getResolution() * 2;

    CSimpleMap map1, map2, map2noisy;

    // Load maps:
    {
        CFileGZInputStream f(fil_grid1);
        archiveFrom(f) >> map1;
    }

    // If it's detect_test only load one map:
    if (is_match || mrpt::system::fileExists(fil_grid2))
    {
        CFileGZInputStream f(fil_grid2);
        archiveFrom(f) >> map2;
    }

    // Generate the_map1 now:
    the_map1.loadFromProbabilisticPosesAndObservations(map1);

    size_t N1 = max(
        40,
        (int)round(grid1->getArea() * gridAlign.options.featsPerSquareMeter));

    COccupancyGridMapFeatureExtractor gridfeatextract;

    CLandmarksMap lm1;

    gridfeatextract.extractFeatures(
        *grid1, lm1, N1, gridAlign.options.feature_descriptor,
        gridAlign.options.feature_detector_options);

    if (!NOSAVE && (is_match || is_detect_test))  // If it's only a
    // SAVE_CORR_AND_NONCORR_DISTS
    // test, save time
    {
        grid1->saveAsBitmapFile(format("%s/map1.png", RESULTS_DIR.c_str()));
        grid1->saveAsBitmapFileWithLandmarks(
            format("%s/map1_LM.png", RESULTS_DIR.c_str()), &lm1, true);
        CImage img;
        grid1->getAsImageFiltered(img);
        img.saveToFile(format("%s/map1_filt.png", RESULTS_DIR.c_str()));
    }

    {
        CVectorFloat stats_covDet, stats_stdPhi;
        CVectorFloat stats_errorXY, stats_errorPhi;
        CVectorFloat stats_bruteErrorXY, stats_bruteErrorPhi;
        CVectorFloat stats_GT_likelihood;
        std::vector<uint32_t> overallGTcorrsFound;

        for (unsigned int iter = 0; iter < N_ITERS; iter++)
        {
            // Insert noise into the laser scans of map2:
            // ------------------------------------------------
            map2noisy = is_detect_test ? map1 : map2;

            for (unsigned int q = 0; q < map2noisy.size(); q++)
            {
                CPose3DPDF::Ptr PDF;
                CSensoryFrame::Ptr SF;

                map2noisy.get(q, PDF, SF);

                // If it's detect_test, translate the map2 by a fixed, known
                // quantity:
                if (is_detect_test)
                {
                    CPose2D gt(GT_Ax, GT_Ay, GT_Aphi_rad);
                    gt = -gt;
                    PDF->changeCoordinatesReference(CPose3D(gt));
                }

                if (NOISE_IN_LASER)
                {
                    CObservation2DRangeScan::Ptr obs =
                        SF->getObservationByClass<CObservation2DRangeScan>();
                    if (obs)
                    {
                        for (unsigned int k = 0; k < obs->getScanSize(); k++)
                        {
                            float& v = obs->getScanRange(k);
                            v += getRandomGenerator().drawGaussian1D(
                                0, STD_NOISE_LASER);
                            if (v < 0) v = 0;
                        }
                    }
                }  // end of NOISE_IN_LASER

                if (NOISE_IN_POSE)
                {
                    CPosePDFGaussian::Ptr newPDF =
                        std::make_shared<CPosePDFGaussian>();
                    newPDF->copyFrom(*PDF);

                    // Change the pose:
                    newPDF->mean.x_incr(
                        getRandomGenerator().drawGaussian1D(0, STD_NOISE_XY));
                    newPDF->mean.y_incr(
                        getRandomGenerator().drawGaussian1D(0, STD_NOISE_XY));
                    newPDF->mean.phi_incr(getRandomGenerator().drawGaussian1D(
                        0, DEG2RAD(STD_NOISE_PHI)));
                    newPDF->mean.normalizePhi();

                    // Change into the map:
                    map2noisy.set(q, newPDF, CSensoryFrame::Ptr());

                }  // end of NOISE_IN_POSE
            }

            the_map2.loadFromProbabilisticPosesAndObservations(map2noisy);

            size_t N2 = max(
                40,
                (int)round(
                    grid2->getArea() * gridAlign.options.featsPerSquareMeter));

            CLandmarksMap lm2;
            gridfeatextract.extractFeatures(
                *grid2, lm2, N2, gridAlign.options.feature_descriptor,
                gridAlign.options.feature_detector_options);

            if (!NOSAVE && (is_match || is_detect_test))  // If it's only a
            // SAVE_CORR_AND_NONCORR_DISTS
            // test, save time
            {
                // Save maps:
                grid2->saveAsBitmapFile(format(
                    "%s/map2_noise_%f.png", RESULTS_DIR.c_str(), STD_NOISE_XY));
                grid2->saveAsBitmapFileWithLandmarks(
                    format(
                        "%s/map2_LM_noise_%f.png", RESULTS_DIR.c_str(),
                        STD_NOISE_XY),
                    &lm2, true);
                CImage img;
                grid2->getAsImageFiltered(img);
                img.saveToFile(format(
                    "%s/map2_filt_noise_%f.png", RESULTS_DIR.c_str(),
                    STD_NOISE_XY));
            }

            // Only if the case of "save-corr-dists" we can do NOT align the
            // maps, since we're not interested in that
            if (is_match || is_detect_test)
            {
                // --------------------------
                //        DO ALIGN
                // --------------------------
                CPosePDF::Ptr parts = gridAlign.Align(
                    &the_map1, &the_map2, CPose2D(0, 0, 0), info);

                // Get the mean, or the best Gassian mean in the case of a SOG:
                if (IS_CLASS(*parts, CPosePDFSOG) && MOST_LIKELY_SOG_MODE_ONLY)
                {
                    CPosePDFSOG::Ptr pdf_SOG =
                        std::dynamic_pointer_cast<CPosePDFSOG>(parts);
                    pdf_SOG->getMostLikelyCovarianceAndMean(
                        estimateCOV, estimateMean);
                }
                else
                {
                    parts->getCovarianceAndMean(estimateCOV, estimateMean);
                }

                float stdPhi = sqrt(estimateCOV(2, 2));

                const auto estimateCOV22 = estimateCOV.extractMatrix<2, 2>();
                float stdXY = sqrt(estimateCOV22.det());

                float Axy = estimateMean.distance2DTo(GT_Ax, GT_Ay);
                float Aphi =
                    fabs(math::wrapToPi(estimateMean.phi() - GT_Aphi_rad));
                float AxyBrute =
                    info.noRobustEstimation.distance2DTo(GT_Ax, GT_Ay);
                float AphiBrute = fabs(math::wrapToPi(
                    info.noRobustEstimation.phi() - GT_Aphi_rad));

                MRPT_LOG_INFO_FMT(
                    "Done in %.03fms\n", 1000.0f * info.executionTime);

                MRPT_LOG_INFO_STREAM("Mean pose:\n\t " << estimateMean);
                MRPT_LOG_INFO_STREAM("Estimate covariance::\n" << estimateCOV);

                // Save particles:
                if (IS_CLASS(*parts, CPosePDFParticles))
                {
                    CPosePDFParticles::Ptr partsPdf =
                        std::dynamic_pointer_cast<CPosePDFParticles>(parts);

                    partsPdf->saveToTextFile(format(
                        "%s/particles_noise_%.03f.txt", RESULTS_DIR.c_str(),
                        STD_NOISE_XY));

                    MRPT_LOG_INFO_FMT(
                        "Goodness: %.03f%%\n", 100 * info.goodness);
                    MRPT_LOG_INFO_STREAM(
                        partsPdf->particlesCount() << " particles");
                    MRPT_LOG_INFO_STREAM("Covariance:\n\t " << estimateCOV);
                }
                else if (IS_CLASS(*parts, CPosePDFSOG))
                {
                    CPosePDFSOG::Ptr pdf_SOG =
                        std::dynamic_pointer_cast<CPosePDFSOG>(parts);
                    MRPT_LOG_INFO_FMT(
                        "SoG has %u modes.", (unsigned int)pdf_SOG->size());

                    pdf_SOG->normalizeWeights();
                    // pdf_SOG->saveToTextFile("_debug_SoG.txt");

                    stats_GT_likelihood.push_back((float)pdf_SOG->evaluatePDF(
                        CPose2D(GT_Ax, GT_Ay, GT_Aphi_rad), true));

                    if (f_out_log.fileOpenCorrectly())
                    {
                        f_out_log.printf("%% SOG_log_w   x   y   phi  \n");
                        for (size_t m = 0; m < pdf_SOG->size(); m++)
                            f_out_log.printf(
                                "    %e     %f %f %f\n", pdf_SOG->get(m).log_w,
                                pdf_SOG->get(m).mean.x(),
                                pdf_SOG->get(m).mean.y(),
                                pdf_SOG->get(m).mean.phi());
                    }

                    if (SAVE_SOG_3DSCENE)
                    {
                        COpenGLScene scene3D;
                        opengl::CSetOfObjects::Ptr thePDF3D =
                            std::make_shared<opengl::CSetOfObjects>();
                        pdf_SOG->getAs3DObject(thePDF3D);
                        opengl::CGridPlaneXY::Ptr gridXY =
                            std::make_shared<opengl::CGridPlaneXY>(
                                -10, 10, -10, 10, 0, 1);
                        scene3D.insert(gridXY);
                        scene3D.insert(thePDF3D);
                        scene3D.saveToFile("_out_SoG.3Dscene");
                    }

                    if (!SAVE_ICP_GOODNESS_FIL.empty())
                    {
                        mrpt::system::vectorToTextFile(
                            info.icp_goodness_all_sog_modes,
                            SAVE_ICP_GOODNESS_FIL, true);  // append & as column
                    }

                    // Save all the maps overlap hypotheses:
                    if (SAVE_SOG_ALL_MAPS_OVERLAP_HYPOTHESES)
                    {
                        CPosePDFSOG::iterator it;
                        size_t nNode;
                        CImage imgGrid1, imgCanvas;
                        grid1->resizeGrid(
                            min(grid1->getXMin(), -60.0f),
                            max(grid1->getXMax(), 60.0f),
                            min(-40.0f, grid1->getYMin()),
                            max(30.0f, grid1->getYMax()));
                        grid1->getAsImage(imgGrid1, true);
                        int imgGrid1LY = imgGrid1.getHeight();
                        const CPose2D nullPose(0, 0, 0);

                        const TPoint2D q1(grid1->getXMin(), grid1->getYMin());
                        const TPoint2D q2(grid1->getXMin(), grid1->getYMax());
                        const TPoint2D q3(grid1->getXMax(), grid1->getYMax());
                        const TPoint2D q4(grid1->getXMax(), grid1->getYMin());

                        const TPoint2D p1(grid2->getXMin(), grid2->getYMin());
                        const TPoint2D p2(grid2->getXMin(), grid2->getYMax());
                        const TPoint2D p3(grid2->getXMax(), grid2->getYMax());
                        const TPoint2D p4(grid2->getXMax(), grid2->getYMin());
                        for (nNode = 0, it = pdf_SOG->begin();
                             it != pdf_SOG->end(); it++, nNode++)
                        {
                            CPose2D x = it->mean;
                            const TPoint2D pp1(x + p1);
                            const TPoint2D pp2(x + p2);
                            const TPoint2D pp3(x + p3);
                            const TPoint2D pp4(x + p4);

                            // Draw the background = the_map1:
                            imgCanvas = imgGrid1;

                            // Draw the overlaped the_map2:
                            imgCanvas.line(
                                grid1->x2idx(pp1.x),
                                imgGrid1LY - 1 - grid1->y2idx(pp1.y),
                                grid1->x2idx(pp2.x),
                                imgGrid1LY - 1 - grid1->y2idx(pp2.y),
                                TColor::black());
                            imgCanvas.line(
                                grid1->x2idx(pp2.x),
                                imgGrid1LY - 1 - grid1->y2idx(pp2.y),
                                grid1->x2idx(pp3.x),
                                imgGrid1LY - 1 - grid1->y2idx(pp3.y),
                                TColor::black());
                            imgCanvas.line(
                                grid1->x2idx(pp3.x),
                                imgGrid1LY - 1 - grid1->y2idx(pp3.y),
                                grid1->x2idx(pp4.x),
                                imgGrid1LY - 1 - grid1->y2idx(pp4.y),
                                TColor::black());
                            imgCanvas.line(
                                grid1->x2idx(pp4.x),
                                imgGrid1LY - 1 - grid1->y2idx(pp4.y),
                                grid1->x2idx(pp1.x),
                                imgGrid1LY - 1 - grid1->y2idx(pp1.y),
                                TColor::black());

                            imgCanvas.saveToFile(format(
                                "%s/_OVERLAP_MAPS_SOG_MODE_%04u.png",
                                RESULTS_DIR.c_str(), (unsigned int)nNode));

                            // Save as 3D scene:
                            COpenGLScene scene;
                            CSetOfObjects::Ptr obj1 =
                                std::make_shared<CSetOfObjects>();
                            the_map1.getAs3DObject(obj1);
                            CSetOfObjects::Ptr obj2 =
                                std::make_shared<CSetOfObjects>();
                            the_map2.getAs3DObject(obj2);

                            obj2->setPose(x);

                            scene.insert(obj1);
                            scene.insert(obj2);

                            // Add also the borders of the maps:
                            CSetOfLines::Ptr lines =
                                std::make_shared<CSetOfLines>();
                            lines->setLineWidth(3);
                            lines->setColor(0, 0, 1);

                            lines->appendLine(q1.x, q1.y, 0, q2.x, q2.y, 0);
                            lines->appendLine(q2.x, q2.y, 0, q3.x, q3.y, 0);
                            lines->appendLine(q3.x, q3.y, 0, q4.x, q4.y, 0);
                            lines->appendLine(q4.x, q4.y, 0, q1.x, q1.y, 0);

                            lines->appendLine(pp1.x, pp1.y, 0, pp2.x, pp2.y, 0);
                            lines->appendLine(pp2.x, pp2.y, 0, pp3.x, pp3.y, 0);
                            lines->appendLine(pp3.x, pp3.y, 0, pp4.x, pp4.y, 0);
                            lines->appendLine(pp4.x, pp4.y, 0, pp1.x, pp1.y, 0);

                            scene.insert(lines);

                            scene.saveToFile(format(
                                "%s/_OVERLAP_MAPS_SOG_MODE_%04u.3Dscene",
                                RESULTS_DIR.c_str(), (unsigned int)nNode));
                        }

                    }  // end SAVE_SOG_ALL

#ifdef SAVE_SOG_GRID
                    // Save grid evaluation of the SOG:
                    CMatrixF gridLimits(1, 4);
                    gridLimits(0, 0) = estimateMean.x - 0.10f;
                    gridLimits(0, 1) = estimateMean.x + 0.10f,
                                  gridLimits(0, 2) = estimateMean.y - 0.10f;
                    gridLimits(0, 3) = estimateMean.y + 0.10f;
                    gridLimits.saveToTextFile(format(
                        "%s/SOG_grid_limits_noise_%f.txt", RESULTS_DIR.c_str(),
                        STD_NOISE_XY));

                    CMatrixD evalGrid;
                    pdf_SOG->evaluatePDFInArea(
                        gridLimits(0, 0), gridLimits(0, 1), gridLimits(0, 2),
                        gridLimits(0, 3), 0.002f, 0, evalGrid,
                        true  // Sum over all phis
                    );

                    evalGrid.saveToTextFile(format(
                        "%s/SOG_grid_noise_%f.txt", RESULTS_DIR.c_str(),
                        STD_NOISE_XY));
#endif
                }  // end if is SOG

                // STATS:
                stats_covDet.push_back(stdXY);
                stats_stdPhi.push_back(stdPhi);
                stats_errorXY.push_back(Axy);
                stats_errorPhi.push_back(Aphi);
                stats_bruteErrorXY.push_back(AxyBrute);
                stats_bruteErrorPhi.push_back(AphiBrute);

            }  // end if we really do align

            // Save the descriptor distances for corresponding and
            // non-corresponding features
            //  (only known if we are in the "is_detect_test" mode!)
            if (SAVE_CORR_AND_NONCORR_DISTS)
            {
                MRPT_LOG_INFO_STREAM("Generating coor & no.corr distances...");

                const bool SAVE_ALSO_COORS_DEBUG_MAPS = false;

                CLandmarksMap::Ptr lmap1 = info.landmarks_map1;
                CLandmarksMap::Ptr lmap2 = info.landmarks_map2;

                // only for the case of non "--match":
                if (!lmap1 && !lmap2)
                {
                    lmap1 = CLandmarksMap::Create();
                    lmap2 = CLandmarksMap::Create();

                    gridfeatextract.extractFeatures(
                        *grid1, *lmap1, N1,
                        gridAlign.options.feature_descriptor,
                        gridAlign.options.feature_detector_options);
                    gridfeatextract.extractFeatures(
                        *grid2, *lmap2, N2,
                        gridAlign.options.feature_descriptor,
                        gridAlign.options.feature_detector_options);
                }

                ASSERT_(lmap1 && lmap2);

                // GT transformation:
                const CPose2D GT_Ap(GT_Ax, GT_Ay, GT_Aphi_rad);
                TMatchingPairList gt_corrs;

                CFileOutputStream fout_CORR("GT_EXP_CORR.txt", true);
                CFileOutputStream fout_NCORR("GT_EXP_NCORR.txt", true);

                // Compute the distances:
                for (size_t i1 = 0; i1 < lmap1->landmarks.size(); i1++)
                {
                    // Distances in descriptor space
                    CVectorDouble D(lmap2->landmarks.size());
                    // Distances in (x,y)
                    CVectorDouble dErrs(lmap2->landmarks.size());
                    size_t i2;

                    const CLandmark* l1 = lmap1->landmarks.get(i1);

                    for (i2 = 0; i2 < lmap2->landmarks.size(); i2++)
                    {
                        CLandmark* l2 = lmap2->landmarks.get(i2);

                        CPoint2D P1 = CPoint2D(l1->pose_mean);
                        CPoint2D P2 = GT_Ap + CPoint2D(l2->pose_mean);

                        const double dErr = P1.distanceTo(P2);
                        dErrs[i2] = dErr;

                        ASSERT_(!l1->features.empty());
                        ASSERT_(!l2->features.empty());

                        D[i2] = l1->features[0].descriptorDistanceTo(
                            l2->features[0]);
                    }

                    size_t best_match = 0;
                    dErrs.minCoeff(best_match);
                    double MIN_DESCR_DIST = mrpt::math::minimum(D);
                    if (dErrs[best_match] < 0.20)
                    {
                        CLandmark* l2 = lmap2->landmarks.get(best_match);
                        gt_corrs.push_back(TMatchingPair(
                            i1, best_match, l1->pose_mean.x, l1->pose_mean.y,
                            l1->pose_mean.z, l2->pose_mean.x, l2->pose_mean.y,
                            l2->pose_mean.z));
                    }
                    else
                        best_match = (unsigned int)(-1);

                    // double m,s; // Mean & Std:
                    // mrpt::math::meanAndStd(D,m,s);
                    // D = Abs( (D-m)/s );
                    // The output files:
                    for (i2 = 0; i2 < lmap2->landmarks.size(); i2++)
                    {
                        if (i2 == best_match)
                            fout_CORR.printf(
                                "%f %f\n", D[i2], D[i2] - MIN_DESCR_DIST);
                        else
                            fout_NCORR.printf(
                                "%f %f\n", D[i2], D[i2] - MIN_DESCR_DIST);
                    }

                    // mrpt::system::pause();
                }  // end for i1

                MRPT_LOG_INFO_STREAM("GT corrs found: " << gt_corrs.size());
                overallGTcorrsFound.push_back(gt_corrs.size());
                if (SAVE_ALSO_COORS_DEBUG_MAPS)
                {
                    COccupancyGridMap2D::saveAsBitmapTwoMapsWithCorrespondences(
                        "GT_corrs.png", grid1.get(), grid2.get(), gt_corrs);
                    mrpt::system::pause();
                }
            }

        }  // end iter

        f_log.printf(
            "%f %f %f %f %f %f %f %f %f %f %f %f %f %f %f %e %f %f\n",
            STD_NOISE_XY, STD_NOISE_PHI, STD_NOISE_LASER,
            math::mean(stats_covDet), math::stddev(stats_covDet),
            math::mean(stats_stdPhi), math::stddev(stats_stdPhi),
            math::mean(stats_errorXY), math::stddev(stats_errorXY),
            math::mean(stats_errorPhi), math::stddev(stats_errorPhi),
            math::mean(stats_bruteErrorXY), math::stddev(stats_bruteErrorXY),
            math::mean(stats_bruteErrorPhi), math::stddev(stats_bruteErrorPhi),
            math::mean(stats_GT_likelihood), math::mean(overallGTcorrsFound),
            math::stddev(overallGTcorrsFound));

    }  // For each noise

    MRPT_END
}