git.maemo.org Git - opencv/blob - apps/haartraining/doc/haartraining.htm

   1 <html>
   2
   3 <head>
   4 <meta http-equiv=Content-Type content="text/html; charset=windows-1251">
   5 <meta name=Generator content="Microsoft Word 11 (filtered)">
   6 <title>Object Detection Using Haar-like Features with Cascade of Boosted
   7 Classifiers</title>
   8 <style>
   9 <!--
  10  /* Style Definitions */
  11  p.MsoNormal, li.MsoNormal, div.MsoNormal
  12         {margin:0in;
  13         margin-bottom:.0001pt;
  14         text-align:justify;
  15         font-size:12.0pt;
  16         font-family:"Times New Roman";}
  17 h1
  18         {margin-top:12.0pt;
  19         margin-right:0in;
  20         margin-bottom:3.0pt;
  21         margin-left:0in;
  22         text-align:justify;
  23         page-break-after:avoid;
  24         font-size:16.0pt;
  25         font-family:Arial;}
  26 h2
  27         {margin-top:12.0pt;
  28         margin-right:0in;
  29         margin-bottom:3.0pt;
  30         margin-left:0in;
  31         text-align:justify;
  32         page-break-after:avoid;
  33         font-size:14.0pt;
  34         font-family:Arial;
  35         font-style:italic;}
  36 h3
  37         {margin-top:12.0pt;
  38         margin-right:0in;
  39         margin-bottom:3.0pt;
  40         margin-left:0in;
  41         text-align:justify;
  42         page-break-after:avoid;
  43         font-size:13.0pt;
  44         font-family:Arial;}
  45 span.Typewch
  46         {font-family:"Courier New";
  47         font-weight:bold;}
  48 @page Section1
  49         {size:595.3pt 841.9pt;
  50         margin:56.7pt 88.0pt 63.2pt 85.05pt;}
  51 div.Section1
  52         {page:Section1;}
  53  /* List Definitions */
  54  ol
  55         {margin-bottom:0in;}
  56 ul
  57         {margin-bottom:0in;}
  58 -->
  59 </style>
  60
  61 </head>
  62
  63 <body lang=RU>
  64
  65 <div class=Section1>
  66
  67 <h1><span lang=EN-US>Rapid Object Detection With A Cascade of Boosted
  68 Classifiers Based on Haar-like Features</span></h1>
  69
  70 <h2><span lang=EN-US>Introduction</span></h2>
  71
  72 <p class=MsoNormal><span lang=EN-US>This document describes how to train and
  73 use a cascade of boosted classifiers for rapid object detection. A large set of
  74 over-complete haar-like features provide the basis for the simple individual
  75 classifiers. Examples of object detection tasks are face, eye and nose
  76 detection, as well as logo detection. </span></p>
  77
  78 <p class=MsoNormal><span lang=EN-US>&nbsp;</span></p>
  79
  80 <p class=MsoNormal><span lang=EN-US>The sample detection task in this document
  81 is logo detection, since logo detection does not require the collection of
  82 large set of registered and carefully marked object samples. Instead we assume
  83 that from one prototype image, a very large set of derived object examples can
  84 be derived (</span><span class=Typewch><span lang=EN-US>createsamples</span></span><span
  85 lang=EN-US> utility, see below).</span></p>
  86
  87 <p class=MsoNormal><span lang=EN-US>&nbsp;</span></p>
  88
  89 <p class=MsoNormal><span lang=EN-US>A detailed description of the training/evaluation
  90 algorithm can be found in [1] and [2].</span></p>
  91
  92 <h2><span lang=EN-US>Samples Creation</span></h2>
  93
  94 <p class=MsoNormal><span lang=EN-US>For training a training samples must be
  95 collected. There are two sample types: negative samples and positive samples.
  96 Negative samples correspond to non-object images. Positive samples correspond
  97 to object images.</span></p>
  98
  99 <h3><span lang=EN-US>Negative Samples</span></h3>
 100
 101 <p class=MsoNormal><span lang=EN-US>Negative samples are taken from arbitrary
 102 images. These images must not contain object representations. Negative samples
 103 are passed through background description file. It is a text file in which each
 104 text line contains the filename (relative to the directory of the description
 105 file) of negative sample image. This file must be created manually. Note that
 106 the negative samples and sample images are also called background samples or
 107 background samples images, and are used interchangeably in this document</span></p>
 108
 109 <p class=MsoNormal><span lang=EN-US>&nbsp;</span></p>
 110
 111 <p class=MsoNormal><span lang=EN-US>Example of negative description file:</span></p>
 112
 113 <p class=MsoNormal><span lang=EN-US>&nbsp;</span></p>
 114
 115 <p class=MsoNormal><span lang=EN-US>Directory structure:</span></p>
 116
 117 <p class=MsoNormal><span class=Typewch><span lang=EN-US>/img</span></span></p>
 118
 119 <p class=MsoNormal><span class=Typewch><span lang=EN-US>  img1.jpg</span></span></p>
 120
 121 <p class=MsoNormal><span class=Typewch><span lang=EN-US>  img2.jpg</span></span></p>
 122
 123 <p class=MsoNormal><span class=Typewch><span lang=EN-US>bg.txt</span></span></p>
 124
 125 <p class=MsoNormal><span class=Typewch><span lang=EN-US>&nbsp;</span></span></p>
 126
 127 <p class=MsoNormal><span class=Typewch><span style='font-family:"Times New Roman";
 128 font-weight:normal'>File </span></span><span class=Typewch><span lang=EN-US>bg.txt:</span></span></p>
 129
 130 <p class=MsoNormal><span class=Typewch><span lang=EN-US>img/img1.jpg</span></span></p>
 131
 132 <p class=MsoNormal><span class=Typewch><span lang=EN-US>img/img2.jpg</span></span></p>
 133
 134 <h3><span lang=EN-US>Positive Samples</span></h3>
 135
 136 <p class=MsoNormal><span lang=EN-US>Positive samples are created by </span><span
 137 class=Typewch><span lang=EN-US>createsamples</span></span><span lang=EN-US>
 138 utility. They may be created from single object image or from collection of
 139 previously marked up images.<br>
 140 <br>
 141 </span></p>
 142
 143 <p class=MsoNormal><span lang=EN-US>The single object image may for instance
 144 contain a company logo. Then are large set of positive samples are created from
 145 the given object image by randomly rotating, changing the logo color as well as
 146 placing the logo on arbitrary background.</span></p>
 147
 148 <p class=MsoNormal><span lang=EN-US>The amount and range of randomness can be
 149 controlled by command line arguments. </span></p>
 150
 151 <p class=MsoNormal><span lang=EN-US>Command line arguments:</span></p>
 152
 153 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 154 class=Typewch><span lang=EN-US>- vec &lt;vec_file_name&gt;</span></span><span
 155 lang=EN-US> </span></p>
 156
 157 <p class=MsoNormal style='margin-left:17.1pt'><span lang=EN-US>name of the
 158 output file containing the positive samples for training</span></p>
 159
 160 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 161 class=Typewch><span lang=EN-US>- img &lt;image_file_name&gt;</span></span><span
 162 lang=EN-US> </span></p>
 163
 164 <p class=MsoNormal style='margin-left:17.1pt'><span lang=EN-US>source object
 165 image (e.g., a company logo)</span></p>
 166
 167 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 168 class=Typewch><span lang=EN-US>- bg &lt;background_file_name&gt;</span></span><span
 169 lang=EN-US> </span></p>
 170
 171 <p class=MsoNormal style='margin-left:17.1pt'><span lang=EN-US>background
 172 description file; contains a list of images into which randomly distorted
 173 versions of the object are pasted for positive sample generation</span></p>
 174
 175 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 176 class=Typewch><span lang=EN-US>- num &lt;number_of_samples&gt;</span></span><span
 177 lang=EN-US> </span></p>
 178
 179 <p class=MsoNormal style='margin-left:17.1pt'><span lang=EN-US>number of
 180 positive samples to generate </span></p>
 181
 182 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 183 class=Typewch><span lang=EN-US>- bgcolor &lt;background_color&gt;</span></span></p>
 184
 185 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 186 lang=EN-US>      background color (currently grayscale images are assumed); the
 187 background color denotes the transparent color. Since there might be
 188 compression artifacts, the amount of color tolerance can be specified by </span><span
 189 class=Typewch><span lang=EN-US>\96bgthresh</span></span><span class=Typewch><span
 190 lang=EN-US style='font-family:Arial;font-weight:normal'>. </span></span><span
 191 lang=EN-US>All pixels between </span><span class=Typewch><span lang=EN-US>bgcolor-bgthresh</span></span><span
 192 lang=EN-US> and </span><span class=Typewch><span lang=EN-US>bgcolor+bgthresh</span></span><span
 193 lang=EN-US> are regarded as transparent.</span></p>
 194
 195 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 196 class=Typewch><span lang=EN-US>- bgthresh &lt;background_color_threshold&gt;</span></span></p>
 197
 198 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 199 class=Typewch><span lang=EN-US>- inv</span></span></p>
 200
 201 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 202 lang=EN-US>      if specified, the colors will be inverted</span></p>
 203
 204 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 205 class=Typewch><span lang=EN-US>- randinv</span></span><span lang=EN-US> </span></p>
 206
 207 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 208 lang=EN-US>      if specified, the colors will be inverted randomly</span></p>
 209
 210 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 211 class=Typewch><span lang=EN-US>- maxidev &lt;max_intensity_deviation&gt;</span></span></p>
 212
 213 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 214 class=Typewch><span lang=EN-US>  </span></span><span lang=EN-US>maximal
 215 intensity deviation of foreground samples pixels</span></p>
 216
 217 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 218 class=Typewch><span lang=EN-US>- maxxangle &lt;max_x_rotation_angle&gt;,</span></span></p>
 219
 220 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 221 class=Typewch><span lang=EN-US>- maxyangle &lt;max_y_rotation_angle&gt;,</span></span></p>
 222
 223 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 224 class=Typewch><span lang=EN-US>- maxzangle &lt;max_z_rotation_angle&gt;</span></span></p>
 225
 226 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 227 lang=EN-US>      maximum rotation angles in radians</span></p>
 228
 229 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 230 class=Typewch><span lang=EN-US>-show</span></span></p>
 231
 232 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 233 lang=EN-US>      if specified, each sample will be shown. Pressing \91Esc\92 will
 234 continue creation process without samples showing. Useful debugging option.</span></p>
 235
 236 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 237 class=Typewch><span lang=EN-US>- w &lt;sample_width&gt;</span></span></p>
 238
 239 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 240 class=Typewch><span lang=EN-US>  </span></span><span class=Typewch><span
 241 lang=EN-US style='font-family:"Times New Roman";font-weight:normal'>width (in
 242 pixels) of the output samples</span></span></p>
 243
 244 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 245 class=Typewch><span lang=EN-US>- h &lt;sample_height&gt;</span></span></p>
 246
 247 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 248 class=Typewch><span lang=EN-US>  </span></span><span class=Typewch><span
 249 lang=EN-US style='font-family:"Times New Roman";font-weight:normal'>height (in
 250 pixels) of the output samples</span></span></p>
 251
 252 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 253 class=Typewch><span lang=EN-US>&nbsp;</span></span></p>
 254
 255 <p class=MsoNormal><span lang=EN-US>For following procedure is used to create a
 256 sample object instance:</span></p>
 257
 258 <p class=MsoNormal><span lang=EN-US>The source image is rotated random around
 259 all three axes. The chosen angle is limited my</span><span class=Typewch><span
 260 lang=EN-US> -max?angle</span></span><span lang=EN-US>. Next pixels of
 261 intensities in the range of </span><span class=Typewch><span lang=EN-US>[bg_color-bg_color_threshold;
 262 bg_color+bg_color_threshold]</span></span><span lang=EN-US> are regarded as
 263 transparent. White noise is added to the intensities of the foreground. If </span><span
 264 class=Typewch><span lang=EN-US>\96inv</span></span><span lang=EN-US> key is
 265 specified then foreground pixel intensities are inverted. If </span><span
 266 class=Typewch><span lang=EN-US>\96randinv</span></span><span lang=EN-US> key is
 267 specified then it is randomly selected whether for this sample inversion will
 268 be applied. Finally, the obtained image is placed onto arbitrary background
 269 from the background description file, resized to the pixel size specified by </span><span
 270 class=Typewch><span lang=EN-US>\96w</span></span><span lang=EN-US> and </span><span
 271 class=Typewch><span lang=EN-US>\96h</span></span><span lang=EN-US> and stored
 272 into the file specified by the </span><span class=Typewch><span lang=EN-US>\96vec</span></span><span
 273 lang=EN-US> command line parameter.</span></p>
 274
 275 <p class=MsoNormal><span lang=EN-US>&nbsp;</span></p>
 276
 277 <p class=MsoNormal><span lang=EN-US>Positive samples also may be obtained from
 278 a collection of previously marked up images. This collection is described by
 279 text file similar to background description file. Each line of this file
 280 corresponds to collection image. The first element of the line is image file
 281 name. It is followed by number of object instances. The following numbers are
 282 the coordinates of bounding rectangles (x, y, width, height).</span></p>
 283
 284 <p class=MsoNormal><span lang=EN-US>&nbsp;</span></p>
 285
 286 <p class=MsoNormal><span lang=EN-US>Example of description file:</span></p>
 287
 288 <p class=MsoNormal><span lang=EN-US>&nbsp;</span></p>
 289
 290 <p class=MsoNormal><span lang=EN-US>Directory structure:</span></p>
 291
 292 <p class=MsoNormal><span class=Typewch><span lang=EN-US>/img</span></span></p>
 293
 294 <p class=MsoNormal><span class=Typewch><span lang=EN-US>  img1.jpg</span></span></p>
 295
 296 <p class=MsoNormal><span class=Typewch><span lang=EN-US>  img2.jpg</span></span></p>
 297
 298 <p class=MsoNormal><span class=Typewch><span lang=EN-US>info.dat</span></span></p>
 299
 300 <p class=MsoNormal><span class=Typewch><span lang=EN-US>&nbsp;</span></span></p>
 301
 302 <p class=MsoNormal><span class=Typewch><span lang=EN-US style='font-family:
 303 "Times New Roman";font-weight:normal'>File </span></span><span class=Typewch><span
 304 lang=EN-US>info.dat:</span></span></p>
 305
 306 <p class=MsoNormal><span class=Typewch><span lang=EN-US>img/img1.jpg  1  140
 307 100 45 45</span></span></p>
 308
 309 <p class=MsoNormal><span class=Typewch><span lang=EN-US>img/img2.jpg  2  100
 310 200 50 50   50 30 25 25</span></span></p>
 311
 312 <p class=MsoNormal><span lang=EN-US>&nbsp;</span></p>
 313
 314 <p class=MsoNormal><span lang=EN-US>Image </span><span class=Typewch><span
 315 lang=EN-US>img1.jpg</span></span><span lang=EN-US> contains single object
 316 instance with bounding rectangle (140, 100, 45, 45). Image </span><span
 317 class=Typewch><span lang=EN-US>img2.jpg</span></span><span lang=EN-US> contains
 318 two object instances.</span></p>
 319
 320 <p class=MsoNormal><span lang=EN-US>&nbsp;</span></p>
 321
 322 <p class=MsoNormal><span lang=EN-US>In order to create positive samples from
 323 such collection </span><span class=Typewch><span lang=EN-US>\96info</span></span><span
 324 lang=EN-US> argument should be specified instead of </span><span class=Typewch><span
 325 lang=EN-US>\96img</span></span><span class=Typewch><span style='font-family:"Times New Roman";
 326 font-weight:normal'>:</span></span></p>
 327
 328 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 329 class=Typewch><span lang=EN-US>- info &lt;collection_file_name&gt;</span></span><span
 330 lang=EN-US> </span></p>
 331
 332 <p class=MsoNormal style='margin-left:17.1pt'><span lang=EN-US>description file
 333 of marked up images collection</span></p>
 334
 335 <p class=MsoNormal><span lang=EN-US>&nbsp;</span></p>
 336
 337 <p class=MsoNormal><span lang=EN-US>The scheme of sample creation in this case
 338 is as follows. The object instances are taken from images. Then they are
 339 resized to samples size and stored in output file. No distortion is applied, so
 340 the only affecting arguments are </span><span class=Typewch><span lang=EN-US>\96w</span></span><span
 341 lang=EN-US>, </span><span class=Typewch><span lang=EN-US>-h</span></span><span
 342 lang=EN-US>, </span><span class=Typewch><span lang=EN-US>-show</span></span><span
 343 lang=EN-US> and </span><span class=Typewch><span lang=EN-US>\96num</span></span><span
 344 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 345 normal'>.</span></span></p>
 346
 347 <p class=MsoNormal><span lang=EN-US>&nbsp;</span></p>
 348
 349 <p class=MsoNormal><span class=Typewch><span lang=EN-US>createsamples</span></span><span
 350 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 351 normal'> utility may be used for examining samples stored in positive samples
 352 file. In order to do this only </span></span><span class=Typewch><span
 353 lang=EN-US>\96vec</span></span><span class=Typewch><span lang=EN-US
 354 style='font-family:"Times New Roman";font-weight:normal'>, </span></span><span
 355 class=Typewch><span lang=EN-US>\96w</span></span><span class=Typewch><span
 356 lang=EN-US style='font-family:"Times New Roman";font-weight:normal'> and </span></span><span
 357 class=Typewch><span lang=EN-US>\96h</span></span><span class=Typewch><span
 358 lang=EN-US style='font-family:"Times New Roman";font-weight:normal'> parameters
 359 should be specified.</span></span></p>
 360
 361 <p class=MsoNormal><span lang=EN-US>&nbsp;</span></p>
 362
 363 <p class=MsoNormal><span lang=EN-US>Note that for training, it does not matter
 364 how positive samples files are generated. So the </span><span class=Typewch><span
 365 lang=EN-US>createsamples</span></span><span lang=EN-US> utility is only one way
 366 to collect/create a vector file of positive samples.</span></p>
 367
 368 <h2><span lang=EN-US>Training</span></h2>
 369
 370 <p class=MsoNormal><span lang=EN-US>The next step after samples creation is
 371 training of classifier. It is performed by the </span><span class=Typewch><span
 372 lang=EN-US>haartraining</span></span><span lang=EN-US> utility.</span></p>
 373
 374 <p class=MsoNormal><span lang=EN-US>&nbsp;</span></p>
 375
 376 <p class=MsoNormal><span lang=EN-US>Command line arguments:</span><span
 377 class=Typewch><span lang=EN-US> </span></span></p>
 378
 379 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 380 class=Typewch><span lang=EN-US>- data &lt;dir_name&gt;</span></span><span
 381 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 382 normal'> </span></span></p>
 383
 384 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 385 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 386 normal'>      directory name in which the trained classifier is stored</span></span></p>
 387
 388 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 389 class=Typewch><span lang=EN-US>- vec &lt;vec_file_name&gt;</span></span><span
 390 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 391 normal'> </span></span></p>
 392
 393 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 394 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 395 normal'>      file name of positive sample file (created by </span></span><span
 396 class=Typewch><span lang=EN-US>trainingsamples</span></span><span
 397 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 398 normal'> utility or by any other means)</span></span></p>
 399
 400 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 401 class=Typewch><span lang=EN-US>- bg &lt;background_file_name&gt;</span></span></p>
 402
 403 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 404 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 405 normal'>      background description file</span></span></p>
 406
 407 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 408 class=Typewch><span lang=EN-US>- npos &lt;number_of_positive_samples&gt;,</span></span></p>
 409
 410 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 411 class=Typewch><span lang=EN-US>- nneg &lt;number_of_negative_samples&gt;</span></span><span
 412 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 413 normal'> </span></span></p>
 414
 415 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 416 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 417 normal'>      number of positive/negative samples used in training of each
 418 classifier stage. Reasonable values are npos = 7000 and nneg = 3000.</span></span></p>
 419
 420 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 421 class=Typewch><span lang=EN-US>- nstages &lt;number_of_stages&gt;</span></span></p>
 422
 423 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 424 class=Typewch><span lang=EN-US>  </span></span><span class=Typewch><span
 425 lang=EN-US style='font-family:"Times New Roman";font-weight:normal'>number of
 426 stages to be trained</span></span></p>
 427
 428 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 429 class=Typewch><span lang=EN-US>- nsplits &lt;number_of_splits&gt;</span></span><span
 430 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 431 normal'> </span></span></p>
 432
 433 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 434 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 435 normal'>      determines the weak classifier used in stage classifiers. If </span></span><span
 436 class=Typewch><span lang=EN-US style='font-family:"Times New Roman"'>1</span></span><span
 437 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 438 normal'>, then a simple stump classifier is used, if </span></span><span
 439 class=Typewch><span lang=EN-US style='font-family:"Times New Roman"'>2</span></span><span
 440 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 441 normal'> and more, then CART classifier with </span></span><span class=Typewch><span
 442 lang=EN-US>number_of_splits</span></span><span class=Typewch><span lang=EN-US
 443 style='font-family:"Times New Roman";font-weight:normal'> internal (split)
 444 nodes is used</span></span></p>
 445
 446 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 447 class=Typewch><span lang=EN-US>- mem &lt;memory_in_MB&gt;</span></span><span
 448 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 449 normal'> </span></span></p>
 450
 451 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 452 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 453 normal'>      Available memory in MB for precalculation. The more memory you
 454 have the faster the training process</span></span></p>
 455
 456 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 457 class=Typewch><span lang=EN-US>- sym (default),</span></span></p>
 458
 459 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 460 class=Typewch><span lang=EN-US>- nonsym</span></span><span class=Typewch><span
 461 lang=EN-US style='font-family:"Times New Roman";font-weight:normal'> </span></span></p>
 462
 463 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 464 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 465 normal'>      specifies whether the object class under training has vertical
 466 symmetry or not. Vertical symmetry speeds up training process. For instance,
 467 frontal faces show off vertical symmetry</span></span></p>
 468
 469 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 470 class=Typewch><span lang=EN-US>- minhitrate &lt;min_hit_rate&gt;</span></span></p>
 471
 472 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 473 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 474 normal'>      minimal desired hit rate for each stage classifier. Overall hit
 475 rate may be estimated as </span></span><span class=Typewch><span lang=EN-US>(min_hit_rate^number_of_stages)</span></span></p>
 476
 477 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 478 class=Typewch><span lang=EN-US>- maxfalsealarm &lt;max_false_alarm_rate&gt;</span></span><span
 479 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 480 normal'> </span></span></p>
 481
 482 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 483 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 484 normal'>      maximal desired false alarm rate for each stage classifier. </span></span><span
 485 class=Typewch><span style='font-family:"Times New Roman";font-weight:normal'>Overall
 486 false alarm rate may be estimated as</span></span><span class=Typewch><span
 487 lang=EN-US> (max_false_alarm_rate^number_of_stages)</span></span></p>
 488
 489 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 490 class=Typewch><span lang=EN-US>- weighttrimming &lt;weight_trimming&gt;</span></span></p>
 491
 492 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 493 class=Typewch><span lang=EN-US>  </span></span><span class=Typewch><span
 494 lang=EN-US style='font-family:"Times New Roman";font-weight:normal'>Specifies
 495 wheter and how much weight trimming should be used. A decent choice is 0.90.</span></span></p>
 496
 497 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 498 class=Typewch><span lang=EN-US>- eqw</span></span></p>
 499
 500 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 501 class=Typewch><span lang=EN-US>- mode &lt;BASIC (default) | CORE | ALL&gt;</span></span><span
 502 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 503 normal'> </span></span></p>
 504
 505 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 506 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 507 normal'>      selects the type of haar features set used in training. BASIC use
 508 only upright features, while ALL uses the full set of upright and 45 degree
 509 rotated feature set. See [1] for more details.</span></span></p>
 510
 511 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 512 class=Typewch><span lang=EN-US>- w &lt;sample_width&gt;,</span></span></p>
 513
 514 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 515 class=Typewch><span lang=EN-US>- h &lt;sample_height&gt;</span></span><span
 516 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 517 normal'> </span></span></p>
 518
 519 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 520 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 521 normal'>      Size of training samples (in pixels). Must have exactly the same
 522 values as used during training samples creation (utility </span></span><span
 523 class=Typewch><span lang=EN-US>trainingsamples</span></span><span
 524 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 525 normal'>)</span></span></p>
 526
 527 <p class=MsoNormal><span class=Typewch><span lang=EN-US style='font-family:
 528 "Times New Roman";font-weight:normal'>&nbsp;</span></span></p>
 529
 530 <p class=MsoNormal><span class=Typewch><span lang=EN-US style='font-family:
 531 "Times New Roman";font-weight:normal'>Note: in order to use multiprocessor
 532 advantage a compiler that supports OpenMP 1.0 standard should be used.</span></span></p>
 533
 534 <h2><span lang=EN-US>Application</span></h2>
 535
 536 <p class=MsoNormal><span lang=EN-US>OpenCV cvHaarDetectObjects() function (in
 537 particular haarFaceDetect demo) is used for detection.</span></p>
 538
 539 <h3><span lang=EN-US>Test Samples</span></h3>
 540
 541 <p class=MsoNormal><span lang=EN-US>In order to evaluate the performance of
 542 trained classifier a collection of marked up images is needed. When such
 543 collection is not available test samples may be created from single object
 544 image by </span><span class=Typewch><span lang=EN-US>createsamples</span></span><span
 545 lang=EN-US> utility. The scheme of test samples creation in this case is
 546 similar to training samples creation since each test sample is a background
 547 image into which a randomly distorted and randomly scaled instance of the
 548 object picture is pasted at a random position. </span></p>
 549
 550 <p class=MsoNormal><span lang=EN-US>&nbsp;</span></p>
 551
 552 <p class=MsoNormal><span lang=EN-US>If both </span><span class=Typewch><span
 553 lang=EN-US>\96img</span></span><span lang=EN-US> and </span><span class=Typewch><span
 554 lang=EN-US>\96info</span></span><span lang=EN-US> arguments are specified then
 555 test samples will be created by </span><span class=Typewch><span lang=EN-US>createsamples</span></span><span
 556 lang=EN-US> utility. The sample image is arbitrary distorted as it was
 557 described below, then it is placed at random location to background image and
 558 stored. The corresponding description line is added to the file specified by </span><span
 559 class=Typewch><span lang=EN-US>\96info</span></span><span lang=EN-US> argument.</span></p>
 560
 561 <p class=MsoNormal><span lang=EN-US>&nbsp;</span></p>
 562
 563 <p class=MsoNormal><span lang=EN-US>The </span><span class=Typewch><span
 564 lang=EN-US>\96w</span></span><span lang=EN-US> and </span><span class=Typewch><span
 565 lang=EN-US>\96h</span></span><span lang=EN-US> keys determine the minimal size of
 566 placed object picture.</span></p>
 567
 568 <p class=MsoNormal><span lang=EN-US>&nbsp;</span></p>
 569
 570 <p class=MsoNormal><span lang=EN-US>The test image file name format is as
 571 follows:</span></p>
 572
 573 <p class=MsoNormal><span class=Typewch><span lang=EN-US>imageOrderNumber_x_y_width_height.jpg</span></span><span
 574 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 575 normal'>, where </span></span><span class=Typewch><span lang=EN-US>x</span></span><span
 576 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 577 normal'>, </span></span><span class=Typewch><span lang=EN-US>y</span></span><span
 578 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 579 normal'>, </span></span><span class=Typewch><span lang=EN-US>width</span></span><span
 580 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 581 normal'> and </span></span><span class=Typewch><span lang=EN-US>height</span></span><span
 582 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 583 normal'> are the coordinates of placed object bounding rectangle.</span></span></p>
 584
 585 <p class=MsoNormal><span class=Typewch><span lang=EN-US style='font-family:
 586 "Times New Roman";font-weight:normal'>Note that you should use a background
 587 images set different from the background image set used during training.</span></span></p>
 588
 589 <h3><span class=Typewch><span lang=EN-US style='font-family:"Times New Roman"'>Performance
 590 Evaluation</span></span></h3>
 591
 592 <p class=MsoNormal><span lang=EN-US>In order to evaluate the performance of the
 593 classifier </span><span class=Typewch><span lang=EN-US>performance</span></span><span
 594 lang=EN-US> utility may be used. It takes a collection of marked up images,
 595 applies the classifier and outputs the performance, i.e. number of found
 596 objects, number of missed objects, number of false alarms and other
 597 information.</span></p>
 598
 599 <p class=MsoNormal><span lang=EN-US>&nbsp;</span></p>
 600
 601 <p class=MsoNormal><span lang=EN-US>Command line arguments:</span><span
 602 class=Typewch><span lang=EN-US> </span></span></p>
 603
 604 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 605 class=Typewch><span lang=EN-US>- data &lt;dir_name&gt;</span></span><span
 606 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 607 normal'> </span></span></p>
 608
 609 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 610 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 611 normal'>      directory name in which the trained classifier is stored</span></span></p>
 612
 613 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 614 class=Typewch><span lang=EN-US>- info &lt;collection_file_name&gt;</span></span><span
 615 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 616 normal'> </span></span></p>
 617
 618 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 619 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 620 normal'>      file with test samples description</span></span></p>
 621
 622 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 623 class=Typewch><span lang=EN-US>- maxSizeDiff &lt;max_size_difference&gt;</span></span><span
 624 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 625 normal'>,</span></span></p>
 626
 627 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 628 class=Typewch><span lang=EN-US>- maxPosDiff &lt;max_position_difference&gt;</span></span><span
 629 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 630 normal'> </span></span></p>
 631
 632 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 633 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 634 normal'>      determine the criterion of reference and detected rectangles
 635 coincidence. Default values are 1.5 and 0.3 respectively.</span></span></p>
 636
 637 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 638 class=Typewch><span lang=EN-US>- sf &lt;scale_factor&gt;</span></span><span
 639 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 640 normal'>,</span></span></p>
 641
 642 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 643 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 644 normal'>      detection parameter. Default value is 1.2.</span></span></p>
 645
 646 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 647 class=Typewch><span lang=EN-US>- w &lt;sample_width&gt;,</span></span></p>
 648
 649 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 650 class=Typewch><span lang=EN-US>- h &lt;sample_height&gt;</span></span><span
 651 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 652 normal'> </span></span></p>
 653
 654 <p class=MsoNormal style='margin-left:17.1pt;text-indent:-17.1pt'><span
 655 class=Typewch><span lang=EN-US style='font-family:"Times New Roman";font-weight:
 656 normal'>      Size of training samples (in pixels). Must have exactly the same
 657 values as used during training (utility </span></span><span class=Typewch><span
 658 lang=EN-US>haartraining</span></span><span class=Typewch><span lang=EN-US
 659 style='font-family:"Times New Roman";font-weight:normal'>)</span></span></p>
 660
 661 <h2><span lang=EN-US>References</span></h2>
 662
 663 <p class=MsoNormal><span lang=EN-US>[1] Rainer Lienhart and Jochen Maydt. An
 664 Extended Set of Haar-like Features for Rapid Object Detection. Submitted to
 665 ICIP2002.</span></p>
 666
 667 <p class=MsoNormal><span lang=EN-US>[2] Alexander Kuranov, Rainer Lienhart, and
 668 Vadim Pisarevsky. An Empirical Analysis of Boosting Algorithms for Rapid
 669 Objects With an Extended Set of Haar-like Features. Intel Technical Report
 670 MRL-TR-July02-01, 2002.</span></p>
 671
 672 </div>
 673
 674 </body>
 675
 676 </html>