• Andrej Muhič's avatar
    Merge pull request #1552 from amuhic:master · ced5aa76
    Andrej Muhič authored
    Implementation of CSR-DCF tracker (#1552)
    
    * Initial commit for CSR-DCF tracker implementation
    
    * Fixes for automatic build
    
    * General code fixes
    
    * Removed unused parameters. Added CSRT to automatic tests.
    
    * Fixed VS build warnings. Fixed a bug with gray sequences.
    
    * Fixed VS build errors for samples file.
    ced5aa76
trackerCSRTSegmentation.cpp 17.7 KB
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450
// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html.

#include "precomp.hpp"

#include "trackerCSRTSegmentation.hpp"

#include <fstream>
#include <iostream>
#include <vector>
#include <iostream>

//-------------------- HISTOGRAM CLASS --------------------
namespace cv
{

Histogram::Histogram(int numDimensions, int numBinsPerDimension)
{
    m_numBinsPerDim = numBinsPerDimension;
    m_numDim = numDimensions;
    p_size = cvFloor(std::pow(m_numBinsPerDim, m_numDim));
    p_bins.resize(p_size, 0);
    p_dimIdCoef.resize(m_numDim, 1);
    for (int i = 0; i < m_numDim-1; ++i)
        p_dimIdCoef[i] = static_cast<int>(std::pow(numBinsPerDimension, m_numDim - 1 - i));

}

void Histogram::extractForegroundHistogram(std::vector<cv::Mat> & imgChannels,
        cv::Mat weights, bool useMatWeights, int x1, int y1, int x2, int y2)
{
    //just for code clarity
    cv::Mat & img = imgChannels[0];

    if (!useMatWeights){
        //weights are epanechnikov distr. with peek at the center of the image;
        double cx = x1 + (x2-x1)/2.;
        double cy = y1 + (y2-y1)/2.;
        double kernelSize_width = 1.0/(0.5*static_cast<double>(x2-x1)*1.4142+1);  //sqrt(2)
        double kernelSize_height = 1.0/(0.5*static_cast<double>(y2-y1)*1.4142+1);

        cv::Mat kernelWeight(img.rows, img.cols, CV_64FC1);
        for (int y = y1; y < y2+1; ++y){
            double * weightPtr = kernelWeight.ptr<double>(y);
            double tmp_y = std::pow((cy-y)*kernelSize_height, 2);
            for (int x = x1; x < x2+1; ++x){
                weightPtr[x] = kernelProfile_Epanechnikov(std::pow((cx-x)*kernelSize_width,2) + tmp_y);
            }
        }
        weights = kernelWeight;
    }
    //extract pixel values and compute histogram
    double rangePerBinInverse = static_cast<double>(m_numBinsPerDim)/256.0;  // 1 / (imgRange/numBinsPerDim)
    double sum = 0;
    for (int y = y1; y < y2+1; ++y){
        std::vector<const uchar *> dataPtr(m_numDim);
        for (int dim = 0; dim < m_numDim; ++dim)
            dataPtr[dim] = imgChannels[dim].ptr<uchar>(y);
        const double * weightPtr = weights.ptr<double>(y);

        for (int x = x1; x < x2+1; ++x){
            int id = 0;
            for (int dim = 0; dim < m_numDim; ++dim){
                id += p_dimIdCoef[dim]*cvFloor(rangePerBinInverse*dataPtr[dim][x]);
            }
            p_bins[id] += weightPtr[x];
            sum += weightPtr[x];
        }
    }
    //normalize
    sum = 1./sum;
    for(int i = 0; i < p_size; ++i)
        p_bins[i] *= sum;
}

void Histogram::extractBackGroundHistogram(
        std::vector<cv::Mat> & imgChannels,
        int x1, int y1, int x2, int y2,
        int outer_x1, int outer_y1, int outer_x2, int outer_y2)
{
    //extract pixel values and compute histogram
    double rangePerBinInverse = static_cast<double>(m_numBinsPerDim)/256.0;  // 1 / (imgRange/numBinsPerDim)
    double sum = 0;
    for (int y = outer_y1; y < outer_y2; ++y){

        std::vector<const uchar *> dataPtr(m_numDim);
        for (int dim = 0; dim < m_numDim; ++dim)
            dataPtr[dim] = imgChannels[dim].ptr<uchar>(y);

        for (int x = outer_x1; x < outer_x2; ++x){
            if (x >= x1 && x <= x2 && y >= y1 && y <= y2)
                continue;

            int id = 0;
            for (int dim = 0; dim < m_numDim; ++dim){
                id += p_dimIdCoef[dim]*cvFloor(rangePerBinInverse*dataPtr[dim][x]);
            }
            p_bins[id] += 1.0;
            sum += 1.0;
        }
    }
    //normalize
    sum = 1./sum;
    for(int i = 0; i < p_size; ++i)
        p_bins[i] *= sum;
}

cv::Mat Histogram::backProject(std::vector<cv::Mat> & imgChannels)
{
    //just for code clarity
    cv::Mat & img = imgChannels[0];

    cv::Mat backProject(img.rows, img.cols, CV_64FC1);
    double rangePerBinInverse = static_cast<double>(m_numBinsPerDim)/256.0;  // 1 / (imgRange/numBinsPerDim)

    for (int y = 0; y < img.rows; ++y){
        double * backProjectPtr = backProject.ptr<double>(y);
        std::vector<const uchar *> dataPtr(m_numDim);
        for (int dim = 0; dim < m_numDim; ++dim)
            dataPtr[dim] = imgChannels[dim].ptr<uchar>(y);

        for (int x = 0; x < img.cols; ++x){
            int id = 0;
            for (int dim = 0; dim < m_numDim; ++dim){
                id += p_dimIdCoef[dim]*cvFloor(rangePerBinInverse*dataPtr[dim][x]);
            }
            backProjectPtr[x] = p_bins[id];
        }
    }
    return backProject;
}

// add new methods
std::vector<double> Histogram::getHistogramVector() {
    return p_bins;
}

void Histogram::setHistogramVector(double *vector) {
    for (size_t i=0; i<p_bins.size(); i++) {
        p_bins[i] = vector[i];
    }
}

//-------------------- SEGMENT CLASS --------------------
std::pair<cv::Mat, cv::Mat> Segment::computePosteriors(
        std::vector<cv::Mat> &imgChannels,
        int x1, int y1, int x2, int y2,
        cv::Mat weights, cv::Mat fgPrior, cv::Mat bgPrior,
        const Histogram &fgHistPrior, int numBinsPerChannel)
{
    //preprocess and normalize all data
    CV_Assert(imgChannels.size() > 0);

    //fit target to the image
    x1 = std::min(std::max(x1, 0), imgChannels[0].cols-1);
    y1 = std::min(std::max(y1, 0), imgChannels[0].rows-1);
    x2 = std::max(std::min(x2, imgChannels[0].cols-1), 0);
    y2 = std::max(std::min(y2, imgChannels[0].rows-1), 0);

    //enlarge bbox by 1/3 of its size for background area
    int offsetX = (x2-x1)/3;
    int offsetY = (y2-y1)/3;
    int outer_y1 = std::max(0, (int)(y1-offsetY));
    int outer_y2 = std::min(imgChannels[0].rows, (int)(y2+offsetY+1));
    int outer_x1 = std::max(0, (int)(x1-offsetX));
    int outer_x2 = std::min(imgChannels[0].cols, (int)(x2+offsetX+1));

    //extract histogram from original data -> more pixels better representation of distr. by histograms
    Histogram hist_target =
        (fgHistPrior.m_numBinsPerDim == numBinsPerChannel && (size_t)fgHistPrior.m_numDim == imgChannels.size())
        ? fgHistPrior : Histogram(static_cast<int>(imgChannels.size()), numBinsPerChannel);
    Histogram hist_background(static_cast<int>(imgChannels.size()), numBinsPerChannel);
    if (weights.cols == 0)
        hist_target.extractForegroundHistogram(imgChannels, cv::Mat(), false, x1, y1, x2, y2);
    else
        hist_target.extractForegroundHistogram(imgChannels, weights, true, x1, y1, x2, y2);
    hist_background.extractBackGroundHistogram(imgChannels, x1, y1, x2, y2,
            outer_x1, outer_y1, outer_x2, outer_y2);

    //compute resize factor so that the max area is 1000 (=avg. size ~ 32x32)
    double factor = sqrt(1000.0/((x2-x1)*(y2-y1)));
    if (factor > 1)
        factor = 1.0;
    cv::Size newSize(cvFloor((x2-x1)*factor), cvFloor((y2-y1)*factor));

    //rescale input data
    cv::Rect roiRect_inner = cv::Rect(x1, y1, x2-x1, y2-y1);
    std::vector<cv::Mat> imgChannelsROI_inner(imgChannels.size());
    for (size_t i = 0; i < imgChannels.size(); ++i)
        cv::resize(imgChannels[i](roiRect_inner), imgChannelsROI_inner[i], newSize);

    //initialize priors if there is no external source and rescale
    cv::Mat fgPriorScaled;
    if (fgPrior.cols == 0)
        fgPriorScaled = 0.5*cv::Mat::ones(newSize, CV_64FC1);
    else
        cv::resize(fgPrior(roiRect_inner), fgPriorScaled, newSize);
    cv::Mat bgPriorScaled;
    if (bgPrior.cols == 0)
        bgPriorScaled = 0.5*cv::Mat::ones(newSize, CV_64FC1);
    else
        cv::resize(bgPrior(roiRect_inner), bgPriorScaled, newSize);

    //backproject pixels likelihood
    cv::Mat foregroundLikelihood = hist_target.backProject(imgChannelsROI_inner).mul(fgPriorScaled);
    cv::Mat backgroundLikelihood = hist_background.backProject(imgChannelsROI_inner).mul(bgPriorScaled);

    double p_b = std::sqrt((std::pow(outer_x2-outer_x1, 2) + std::pow(outer_y2-outer_y1, 2)) /
            (std::pow(x2-x1, 2) + std::pow(y2-y1, 2))) ;
    double p_o = 1./(p_b + 1);

    //convert likelihoods to posterior prob. (Bayes rule)
    cv::Mat prob_o(newSize, foregroundLikelihood.type());
    prob_o = p_o*foregroundLikelihood / (p_o*foregroundLikelihood + p_b*backgroundLikelihood);
    cv::Mat prob_b = 1.0 - prob_o;

    std::pair<cv::Mat, cv::Mat> sizedProbs = getRegularizedSegmentation(prob_o, prob_b, fgPriorScaled, bgPriorScaled);

    //resize probs to original size
    std::pair<cv::Mat, cv::Mat> probs;
    cv::resize(sizedProbs.first, probs.first, cv::Size(roiRect_inner.width, roiRect_inner.height));
    cv::resize(sizedProbs.second, probs.second, cv::Size(roiRect_inner.width, roiRect_inner.height));

    return probs;
}

std::pair<cv::Mat, cv::Mat> Segment::computePosteriors2(
    std::vector<cv::Mat> &imgChannels, int x1, int y1, int x2, int y2, double p_b,
    cv::Mat fgPrior, cv::Mat bgPrior, Histogram hist_target, Histogram hist_background)
{
    //preprocess and normalize all data
    CV_Assert(imgChannels.size() > 0);

    //fit target to the image
    x1 = std::min(std::max(x1, 0), imgChannels[0].cols-1);
    y1 = std::min(std::max(y1, 0), imgChannels[0].rows-1);
    x2 = std::max(std::min(x2, imgChannels[0].cols-1), 0);
    y2 = std::max(std::min(y2, imgChannels[0].rows-1), 0);

    // calculate width and height of the region
    int w = x2 - x1 + 1;
    int h = y2 - y1 + 1;
    w = std::min(std::max(w, 1), imgChannels[0].cols);
    h = std::min(std::max(h, 1), imgChannels[0].rows);

    //double p_o = 1./(p_b + 1);
    double p_o = 1. - p_b;

    //compute resize factor so that the max area is 1000 (=avg. size ~ 32x32)
    double factor = sqrt(1000.0/(w*h));
    if (factor > 1)
        factor = 1.0;
    cv::Size newSize(cvFloor(w*factor), cvFloor(h*factor));

    //rescale input data
    cv::Rect roiRect_inner = cv::Rect(x1, y1, w, h);
    std::vector<cv::Mat> imgChannelsROI_inner(imgChannels.size());
    for (size_t i = 0; i < imgChannels.size(); ++i)
        cv::resize(imgChannels[i](roiRect_inner), imgChannelsROI_inner[i], newSize);

    //initialize priors if there is no external source and rescale
    cv::Mat fgPriorScaled;
    if (fgPrior.cols == 0)
        fgPriorScaled = 0.5*cv::Mat::ones(newSize, CV_64FC1);
    else
        cv::resize(fgPrior(roiRect_inner), fgPriorScaled, newSize);
    cv::Mat bgPriorScaled;
    if (bgPrior.cols == 0)
        bgPriorScaled = 0.5*cv::Mat::ones(newSize, CV_64FC1);
    else
        cv::resize(bgPrior(roiRect_inner), bgPriorScaled, newSize);

    //backproject pixels likelihood
    cv::Mat foregroundLikelihood = hist_target.backProject(imgChannelsROI_inner).mul(fgPriorScaled);
    cv::Mat backgroundLikelihood = hist_background.backProject(imgChannelsROI_inner).mul(bgPriorScaled);

    //convert likelihoods to posterior prob. (Bayes rule)
    cv::Mat prob_o(newSize, foregroundLikelihood.type());
    prob_o = p_o*foregroundLikelihood / (p_o*foregroundLikelihood + p_b*backgroundLikelihood);
    cv::Mat prob_b = 1.0 - prob_o;

    std::pair<cv::Mat, cv::Mat> sizedProbs = getRegularizedSegmentation(prob_o, prob_b,
            fgPriorScaled, bgPriorScaled);
    //std::pair<cv::Mat, cv::Mat> sizedProbs = std::pair<cv::Mat, cv::Mat>(prob_o, prob_b);

    //resize probs to original size
    std::pair<cv::Mat, cv::Mat> probs;
    cv::resize(sizedProbs.first, probs.first, cv::Size(roiRect_inner.width, roiRect_inner.height));
    cv::resize(sizedProbs.second, probs.second, cv::Size(roiRect_inner.width, roiRect_inner.height));

    return probs;
}

std::pair<cv::Mat, cv::Mat> Segment::computePosteriors2(std::vector<cv::Mat> &imgChannels,
        cv::Mat fgPrior, cv::Mat bgPrior, Histogram hist_target, Histogram hist_background)
{
    //preprocess and normalize all data
    CV_Assert(imgChannels.size() > 0);

    //fit target to the image
    int x1 = 0;
    int y1 = 0;
    int x2 = imgChannels[0].cols-1;
    int y2 = imgChannels[0].rows-1;

    //compute resize factor so that we control the max area ~32^2
    double factor = sqrt(1000./((x2-x1)*(y2-y1)));
    //double factor = 1;
    if (factor > 1)
        factor = 1.0;
    cv::Size newSize(cvFloor((x2-x1)*factor), cvFloor((y2-y1)*factor));

    //rescale input data
    cv::Rect roiRect_inner = cv::Rect(x1, y1, x2-x1+1, y2-y1+1);
    std::vector<cv::Mat> imgChannelsROI_inner(imgChannels.size());
    for (size_t i = 0; i < imgChannels.size(); ++i)
        cv::resize(imgChannels[i](roiRect_inner), imgChannelsROI_inner[i], newSize);

    //initialize priors if there is no external source and rescale
    cv::Mat fgPriorScaled;
    if (fgPrior.cols == 0)
        fgPriorScaled = 0.5*cv::Mat::ones(newSize, CV_64FC1);
    else
        cv::resize(fgPrior(roiRect_inner), fgPriorScaled, newSize);

    cv::Mat bgPriorScaled;
    if (bgPrior.cols == 0)
        bgPriorScaled = 0.5*cv::Mat::ones(newSize, CV_64FC1);
    else
        cv::resize(bgPrior(roiRect_inner), bgPriorScaled, newSize);

    //backproject pixels likelihood
    cv::Mat foregroundLikelihood = hist_target.backProject(imgChannelsROI_inner).mul(fgPriorScaled);
    cv::Mat backgroundLikelihood = hist_background.backProject(imgChannelsROI_inner).mul(bgPriorScaled);

    //prior for posterior, relative to the number of pixels in bg and fg
    double p_b = 5./3.;
    double p_o = 1./(p_b + 1);

    //convert likelihoods to posterior prob. (Bayes rule)
    cv::Mat prob_o(newSize, foregroundLikelihood.type());
    prob_o = p_o*foregroundLikelihood / (p_o*foregroundLikelihood + p_b*backgroundLikelihood);
    cv::Mat prob_b = 1.0 - prob_o;

    std::pair<cv::Mat, cv::Mat> sizedProbs = getRegularizedSegmentation(prob_o, prob_b, fgPriorScaled, bgPriorScaled);

    //resize probs to original size
    std::pair<cv::Mat, cv::Mat> probs;
    cv::resize(sizedProbs.first, probs.first, cv::Size(roiRect_inner.width, roiRect_inner.height));
    cv::resize(sizedProbs.second, probs.second, cv::Size(roiRect_inner.width, roiRect_inner.height));

    return probs;
}

std::pair<cv::Mat, cv::Mat> Segment::getRegularizedSegmentation(
        cv::Mat &prob_o, cv::Mat &prob_b, cv::Mat & prior_o, cv::Mat & prior_b)
{
    int hsize = cvFloor(std::max(1.0, (double)cvFloor(static_cast<double>(prob_b.cols)*3./50. + 0.5)));
    int lambdaSize = hsize*2+1;

    //compute gaussian kernel
    cv::Mat lambda(lambdaSize, lambdaSize, CV_64FC1);
    double std2 = std::pow(hsize/3.0, 2);
    double sumLambda = 0.0;
    for (int y = -hsize; y < hsize + 1; ++y){
        double * lambdaPtr = lambda.ptr<double>(y+hsize);
        double tmp_y = y*y;
        for (int x = -hsize; x < hsize +1; ++x){
            double tmp_gauss = gaussian(x*x, tmp_y, std2);
            lambdaPtr[x+hsize] = tmp_gauss;
            sumLambda += tmp_gauss;
        }
    }
    sumLambda -= lambda.at<double>(hsize, hsize);
    //set center of kernel to 0
    lambda.at<double>(hsize, hsize) = 0.0;
    sumLambda = 1.0/sumLambda;
    //normalize kernel to sum to 1
    lambda = lambda*sumLambda;

    //create lambda2 kernel
    cv::Mat lambda2 = lambda.clone();
    lambda2.at<double>(hsize, hsize) = 1.0;

    double terminateThr = 1e-1;
    double logLike = std::numeric_limits<double>::max();
    int maxIter = 50;

    //return values
    cv::Mat Qsum_o(prior_o.rows, prior_o.cols, prior_o.type());
    cv::Mat Qsum_b(prior_o.rows, prior_o.cols, prior_o.type());

    //algorithm temporal
    cv::Mat Si_o(prior_o.rows, prior_o.cols, prior_o.type());
    cv::Mat Si_b(prior_o.rows, prior_o.cols, prior_o.type());
    cv::Mat Ssum_o(prior_o.rows, prior_o.cols, prior_o.type());
    cv::Mat Ssum_b(prior_o.rows, prior_o.cols, prior_o.type());
    cv::Mat Qi_o(prior_o.rows, prior_o.cols, prior_o.type());
    cv::Mat Qi_b(prior_o.rows, prior_o.cols, prior_o.type());
    cv::Mat logQo(prior_o.rows, prior_o.cols, prior_o.type());
    cv::Mat logQb(prior_o.rows, prior_o.cols, prior_o.type());

    int i;
    for (i = 0; i < maxIter; ++i){
        //follows the equations from Kristan et al. ACCV2014 paper
        //"A graphical model for rapid obstacle image-map estimation from unmanned surface vehicles"
        cv::Mat P_Io = prior_o.mul(prob_o) + std::numeric_limits<double>::epsilon();
        cv::Mat P_Ib = prior_b.mul(prob_b) + std::numeric_limits<double>::epsilon();

        cv::filter2D(prior_o, Si_o, -1, lambda, cv::Point(-1, -1), 0, cv::BORDER_REFLECT);
        cv::filter2D(prior_b, Si_b, -1, lambda, cv::Point(-1, -1), 0, cv::BORDER_REFLECT);
        Si_o = Si_o.mul(prior_o);
        Si_b = Si_b.mul(prior_b);
        cv::Mat normSi = 1.0/(Si_o + Si_b);
        Si_o = Si_o.mul(normSi);
        Si_b = Si_b.mul(normSi);
        cv::filter2D(Si_o, Ssum_o, -1, lambda2, cv::Point(-1, -1), 0, cv::BORDER_REFLECT);
        cv::filter2D(Si_b, Ssum_b, -1, lambda2, cv::Point(-1, -1), 0, cv::BORDER_REFLECT);

        cv::filter2D(P_Io, Qi_o, -1, lambda, cv::Point(-1, -1), 0, cv::BORDER_REFLECT);
        cv::filter2D(P_Ib, Qi_b, -1, lambda, cv::Point(-1, -1), 0, cv::BORDER_REFLECT);
        Qi_o = Qi_o.mul(P_Io);
        Qi_b = Qi_b.mul(P_Ib);
        cv::Mat normQi = 1.0/(Qi_o + Qi_b);
        Qi_o = Qi_o.mul(normQi);
        Qi_b = Qi_b.mul(normQi);
        cv::filter2D(Qi_o, Qsum_o, -1, lambda2, cv::Point(-1, -1), 0, cv::BORDER_REFLECT);
        cv::filter2D(Qi_b, Qsum_b, -1, lambda2, cv::Point(-1, -1), 0, cv::BORDER_REFLECT);

        prior_o = (Qsum_o + Ssum_o)*0.25;
        prior_b = (Qsum_b + Ssum_b)*0.25;
        cv::Mat normPI = 1.0/(prior_o + prior_b);
        prior_o = prior_o.mul(normPI);
        prior_b = prior_b.mul(normPI);

        //converge ?
        cv::log(Qsum_o, logQo);
        cv::log(Qsum_b, logQb);
        cv::Scalar mean = cv::sum(logQo+logQb);
        double logLikeNew = -mean.val[0]/(2*Qsum_o.rows*Qsum_o.cols);
        if (std::abs(logLike - logLikeNew) < terminateThr)
            break;
        logLike = logLikeNew;
    }
    return std::pair<cv::Mat, cv::Mat>(Qsum_o, Qsum_b);
}

} //cv namespace
//---------------------------------------------------------------------------------------------------------------------