train_HOG.cpp 14.4 KB
Newer Older
wester committed
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
#include <opencv2/opencv.hpp>

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

#include <time.h>

using namespace cv;
using namespace cv::ml;
using namespace std;

void get_svm_detector(const Ptr<SVM>& svm, vector< float > & hog_detector );
void convert_to_ml(const std::vector< cv::Mat > & train_samples, cv::Mat& trainData );
void load_images( const string & prefix, const string & filename, vector< Mat > & img_lst );
void sample_neg( const vector< Mat > & full_neg_lst, vector< Mat > & neg_lst, const Size & size );
Mat get_hogdescriptor_visu(const Mat& color_origImg, vector<float>& descriptorValues, const Size & size );
void compute_hog( const vector< Mat > & img_lst, vector< Mat > & gradient_lst, const Size & size );
void train_svm( const vector< Mat > & gradient_lst, const vector< int > & labels );
void draw_locations( Mat & img, const vector< Rect > & locations, const Scalar & color );
void test_it( const Size & size );

void get_svm_detector(const Ptr<SVM>& svm, vector< float > & hog_detector )
{
    // get the support vectors
    Mat sv = svm->getSupportVectors();
    const int sv_total = sv.rows;
    // get the decision function
    Mat alpha, svidx;
    double rho = svm->getDecisionFunction(0, alpha, svidx);

    CV_Assert( alpha.total() == 1 && svidx.total() == 1 && sv_total == 1 );
    CV_Assert( (alpha.type() == CV_64F && alpha.at<double>(0) == 1.) ||
               (alpha.type() == CV_32F && alpha.at<float>(0) == 1.f) );
    CV_Assert( sv.type() == CV_32F );
    hog_detector.clear();

    hog_detector.resize(sv.cols + 1);
    memcpy(&hog_detector[0], sv.ptr(), sv.cols*sizeof(hog_detector[0]));
    hog_detector[sv.cols] = (float)-rho;
}


/*
* Convert training/testing set to be used by OpenCV Machine Learning algorithms.
* TrainData is a matrix of size (#samples x max(#cols,#rows) per samples), in 32FC1.
* Transposition of samples are made if needed.
*/
void convert_to_ml(const std::vector< cv::Mat > & train_samples, cv::Mat& trainData )
{
    //--Convert data
    const int rows = (int)train_samples.size();
    const int cols = (int)std::max( train_samples[0].cols, train_samples[0].rows );
    cv::Mat tmp(1, cols, CV_32FC1); //< used for transposition if needed
    trainData = cv::Mat(rows, cols, CV_32FC1 );
    vector< Mat >::const_iterator itr = train_samples.begin();
    vector< Mat >::const_iterator end = train_samples.end();
    for( int i = 0 ; itr != end ; ++itr, ++i )
    {
        CV_Assert( itr->cols == 1 ||
            itr->rows == 1 );
        if( itr->cols == 1 )
        {
            transpose( *(itr), tmp );
            tmp.copyTo( trainData.row( i ) );
        }
        else if( itr->rows == 1 )
        {
            itr->copyTo( trainData.row( i ) );
        }
    }
}

void load_images( const string & prefix, const string & filename, vector< Mat > & img_lst )
{
    string line;
    ifstream file;

    file.open( (prefix+filename).c_str() );
    if( !file.is_open() )
    {
        cerr << "Unable to open the list of images from " << filename << " filename." << endl;
        exit( -1 );
    }

    bool end_of_parsing = false;
    while( !end_of_parsing )
    {
        getline( file, line );
        if( line.empty() ) // no more file to read
        {
            end_of_parsing = true;
            break;
        }
        Mat img = imread( (prefix+line).c_str() ); // load the image
        if( img.empty() ) // invalid image, just skip it.
            continue;
#ifdef _DEBUG
        imshow( "image", img );
        waitKey( 10 );
#endif
        img_lst.push_back( img.clone() );
    }
}

void sample_neg( const vector< Mat > & full_neg_lst, vector< Mat > & neg_lst, const Size & size )
{
    Rect box;
    box.width = size.width;
    box.height = size.height;

    const int size_x = box.width;
    const int size_y = box.height;

    srand( (unsigned int)time( NULL ) );

    vector< Mat >::const_iterator img = full_neg_lst.begin();
    vector< Mat >::const_iterator end = full_neg_lst.end();
    for( ; img != end ; ++img )
    {
        box.x = rand() % (img->cols - size_x);
        box.y = rand() % (img->rows - size_y);
        Mat roi = (*img)(box);
        neg_lst.push_back( roi.clone() );
#ifdef _DEBUG
        imshow( "img", roi.clone() );
        waitKey( 10 );
#endif
    }
}

// From http://www.juergenwiki.de/work/wiki/doku.php?id=public:hog_descriptor_computation_and_visualization
Mat get_hogdescriptor_visu(const Mat& color_origImg, vector<float>& descriptorValues, const Size & size )
{
    const int DIMX = size.width;
    const int DIMY = size.height;
    float zoomFac = 3;
    Mat visu;
    resize(color_origImg, visu, Size( (int)(color_origImg.cols*zoomFac), (int)(color_origImg.rows*zoomFac) ) );

    int cellSize        = 8;
    int gradientBinSize = 9;
    float radRangeForOneBin = (float)(CV_PI/(float)gradientBinSize); // dividing 180 into 9 bins, how large (in rad) is one bin?

    // prepare data structure: 9 orientation / gradient strenghts for each cell
    int cells_in_x_dir = DIMX / cellSize;
    int cells_in_y_dir = DIMY / cellSize;
    float*** gradientStrengths = new float**[cells_in_y_dir];
    int** cellUpdateCounter   = new int*[cells_in_y_dir];
    for (int y=0; y<cells_in_y_dir; y++)
    {
        gradientStrengths[y] = new float*[cells_in_x_dir];
        cellUpdateCounter[y] = new int[cells_in_x_dir];
        for (int x=0; x<cells_in_x_dir; x++)
        {
            gradientStrengths[y][x] = new float[gradientBinSize];
            cellUpdateCounter[y][x] = 0;

            for (int bin=0; bin<gradientBinSize; bin++)
                gradientStrengths[y][x][bin] = 0.0;
        }
    }

    // nr of blocks = nr of cells - 1
    // since there is a new block on each cell (overlapping blocks!) but the last one
    int blocks_in_x_dir = cells_in_x_dir - 1;
    int blocks_in_y_dir = cells_in_y_dir - 1;

    // compute gradient strengths per cell
    int descriptorDataIdx = 0;
    int cellx = 0;
    int celly = 0;

    for (int blockx=0; blockx<blocks_in_x_dir; blockx++)
    {
        for (int blocky=0; blocky<blocks_in_y_dir; blocky++)
        {
            // 4 cells per block ...
            for (int cellNr=0; cellNr<4; cellNr++)
            {
                // compute corresponding cell nr
                cellx = blockx;
                celly = blocky;
                if (cellNr==1) celly++;
                if (cellNr==2) cellx++;
                if (cellNr==3)
                {
                    cellx++;
                    celly++;
                }

                for (int bin=0; bin<gradientBinSize; bin++)
                {
                    float gradientStrength = descriptorValues[ descriptorDataIdx ];
                    descriptorDataIdx++;

                    gradientStrengths[celly][cellx][bin] += gradientStrength;

                } // for (all bins)


                // note: overlapping blocks lead to multiple updates of this sum!
                // we therefore keep track how often a cell was updated,
                // to compute average gradient strengths
                cellUpdateCounter[celly][cellx]++;

            } // for (all cells)


        } // for (all block x pos)
    } // for (all block y pos)


    // compute average gradient strengths
    for (celly=0; celly<cells_in_y_dir; celly++)
    {
        for (cellx=0; cellx<cells_in_x_dir; cellx++)
        {

            float NrUpdatesForThisCell = (float)cellUpdateCounter[celly][cellx];

            // compute average gradient strenghts for each gradient bin direction
            for (int bin=0; bin<gradientBinSize; bin++)
            {
                gradientStrengths[celly][cellx][bin] /= NrUpdatesForThisCell;
            }
        }
    }

    // draw cells
    for (celly=0; celly<cells_in_y_dir; celly++)
    {
        for (cellx=0; cellx<cells_in_x_dir; cellx++)
        {
            int drawX = cellx * cellSize;
            int drawY = celly * cellSize;

            int mx = drawX + cellSize/2;
            int my = drawY + cellSize/2;

            rectangle(visu, Point((int)(drawX*zoomFac), (int)(drawY*zoomFac)), Point((int)((drawX+cellSize)*zoomFac), (int)((drawY+cellSize)*zoomFac)), Scalar(100,100,100), 1);

            // draw in each cell all 9 gradient strengths
            for (int bin=0; bin<gradientBinSize; bin++)
            {
                float currentGradStrength = gradientStrengths[celly][cellx][bin];

                // no line to draw?
                if (currentGradStrength==0)
                    continue;

                float currRad = bin * radRangeForOneBin + radRangeForOneBin/2;

                float dirVecX = cos( currRad );
                float dirVecY = sin( currRad );
                float maxVecLen = (float)(cellSize/2.f);
                float scale = 2.5; // just a visualization scale, to see the lines better

                // compute line coordinates
                float x1 = mx - dirVecX * currentGradStrength * maxVecLen * scale;
                float y1 = my - dirVecY * currentGradStrength * maxVecLen * scale;
                float x2 = mx + dirVecX * currentGradStrength * maxVecLen * scale;
                float y2 = my + dirVecY * currentGradStrength * maxVecLen * scale;

                // draw gradient visualization
                line(visu, Point((int)(x1*zoomFac),(int)(y1*zoomFac)), Point((int)(x2*zoomFac),(int)(y2*zoomFac)), Scalar(0,255,0), 1);

            } // for (all bins)

        } // for (cellx)
    } // for (celly)


    // don't forget to free memory allocated by helper data structures!
    for (int y=0; y<cells_in_y_dir; y++)
    {
        for (int x=0; x<cells_in_x_dir; x++)
        {
            delete[] gradientStrengths[y][x];
        }
        delete[] gradientStrengths[y];
        delete[] cellUpdateCounter[y];
    }
    delete[] gradientStrengths;
    delete[] cellUpdateCounter;

    return visu;

} // get_hogdescriptor_visu

void compute_hog( const vector< Mat > & img_lst, vector< Mat > & gradient_lst, const Size & size )
{
    HOGDescriptor hog;
    hog.winSize = size;
    Mat gray;
    vector< Point > location;
    vector< float > descriptors;

    vector< Mat >::const_iterator img = img_lst.begin();
    vector< Mat >::const_iterator end = img_lst.end();
    for( ; img != end ; ++img )
    {
        cvtColor( *img, gray, COLOR_BGR2GRAY );
        hog.compute( gray, descriptors, Size( 8, 8 ), Size( 0, 0 ), location );
        gradient_lst.push_back( Mat( descriptors ).clone() );
#ifdef _DEBUG
        imshow( "gradient", get_hogdescriptor_visu( img->clone(), descriptors, size ) );
        waitKey( 10 );
#endif
    }
}

void train_svm( const vector< Mat > & gradient_lst, const vector< int > & labels )
{

    Mat train_data;
    convert_to_ml( gradient_lst, train_data );

    clog << "Start training...";
    Ptr<SVM> svm = SVM::create();
    /* Default values to train SVM */
    svm->setCoef0(0.0);
    svm->setDegree(3);
    svm->setTermCriteria(TermCriteria( CV_TERMCRIT_ITER+CV_TERMCRIT_EPS, 1000, 1e-3 ));
    svm->setGamma(0);
    svm->setKernel(SVM::LINEAR);
    svm->setNu(0.5);
    svm->setP(0.1); // for EPSILON_SVR, epsilon in loss function?
    svm->setC(0.01); // From paper, soft classifier
    svm->setType(SVM::EPS_SVR); // C_SVC; // EPSILON_SVR; // may be also NU_SVR; // do regression task
    svm->train(train_data, ROW_SAMPLE, Mat(labels));
    clog << "...[done]" << endl;

    svm->save( "my_people_detector.yml" );
}

void draw_locations( Mat & img, const vector< Rect > & locations, const Scalar & color )
{
    if( !locations.empty() )
    {
        vector< Rect >::const_iterator loc = locations.begin();
        vector< Rect >::const_iterator end = locations.end();
        for( ; loc != end ; ++loc )
        {
            rectangle( img, *loc, color, 2 );
        }
    }
}

void test_it( const Size & size )
{
    char key = 27;
    Scalar reference( 0, 255, 0 );
    Scalar trained( 0, 0, 255 );
    Mat img, draw;
    Ptr<SVM> svm;
    HOGDescriptor hog;
    HOGDescriptor my_hog;
    my_hog.winSize = size;
    VideoCapture video;
    vector< Rect > locations;

    // Load the trained SVM.
    svm = StatModel::load<SVM>( "my_people_detector.yml" );
    // Set the trained svm to my_hog
    vector< float > hog_detector;
    get_svm_detector( svm, hog_detector );
    my_hog.setSVMDetector( hog_detector );
    // Set the people detector.
    hog.setSVMDetector( hog.getDefaultPeopleDetector() );
    // Open the camera.
    video.open(0);
    if( !video.isOpened() )
    {
        cerr << "Unable to open the device 0" << endl;
        exit( -1 );
    }

    bool end_of_process = false;
    while( !end_of_process )
    {
        video >> img;
        if( img.empty() )
            break;

        draw = img.clone();

        locations.clear();
        hog.detectMultiScale( img, locations );
        draw_locations( draw, locations, reference );

        locations.clear();
        my_hog.detectMultiScale( img, locations );
        draw_locations( draw, locations, trained );

        imshow( "Video", draw );
        key = (char)waitKey( 10 );
        if( 27 == key )
            end_of_process = true;
    }
}

int main( int argc, char** argv )
{
    cv::CommandLineParser parser(argc, argv, "{help h|| show help message}"
            "{pd||pos_dir}{p||pos.lst}{nd||neg_dir}{n||neg.lst}");
    if (parser.has("help"))
    {
        parser.printMessage();
        exit(0);
    }
    vector< Mat > pos_lst;
    vector< Mat > full_neg_lst;
    vector< Mat > neg_lst;
    vector< Mat > gradient_lst;
    vector< int > labels;
    string pos_dir = parser.get<string>("pd");
    string pos = parser.get<string>("p");
    string neg_dir = parser.get<string>("nd");
    string neg = parser.get<string>("n");
    if( pos_dir.empty() || pos.empty() || neg_dir.empty() || neg.empty() )
    {
        cout << "Wrong number of parameters." << endl
            << "Usage: " << argv[0] << " --pd=pos_dir -p=pos.lst --nd=neg_dir -n=neg.lst" << endl
            << "example: " << argv[0] << " --pd=/INRIA_dataset/ -p=Train/pos.lst --nd=/INRIA_dataset/ -n=Train/neg.lst" << endl;
        exit( -1 );
    }
    load_images( pos_dir, pos, pos_lst );
    labels.assign( pos_lst.size(), +1 );
    const unsigned int old = (unsigned int)labels.size();
    load_images( neg_dir, neg, full_neg_lst );
    sample_neg( full_neg_lst, neg_lst, Size( 96,160 ) );
    labels.insert( labels.end(), neg_lst.size(), -1 );
    CV_Assert( old < labels.size() );

    compute_hog( pos_lst, gradient_lst, Size( 96, 160 ) );
    compute_hog( neg_lst, gradient_lst, Size( 96, 160 ) );

    train_svm( gradient_lst, labels );

    test_it( Size( 96, 160 ) ); // change with your parameters

    return 0;
}