1/*M///////////////////////////////////////////////////////////////////////////////////////
2//
3//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
4//
5//  By downloading, copying, installing or using the software you agree to this license.
6//  If you do not agree to this license, do not download, install,
7//  copy or use the software.
8//
9//
10//                        Intel License Agreement
11//                For Open Source Computer Vision Library
12//
13// Copyright (C) 2000, Intel Corporation, all rights reserved.
14// Third party copyrights are property of their respective owners.
15//
16// Redistribution and use in source and binary forms, with or without modification,
17// are permitted provided that the following conditions are met:
18//
19//   * Redistribution's of source code must retain the above copyright notice,
20//     this list of conditions and the following disclaimer.
21//
22//   * Redistribution's in binary form must reproduce the above copyright notice,
23//     this list of conditions and the following disclaimer in the documentation
24//     and/or other materials provided with the distribution.
25//
26//   * The name of Intel Corporation may not be used to endorse or promote products
27//     derived from this software without specific prior written permission.
28//
29// This software is provided by the copyright holders and contributors "as is" and
30// any express or implied warranties, including, but not limited to, the implied
31// warranties of merchantability and fitness for a particular purpose are disclaimed.
32// In no event shall the Intel Corporation or contributors be liable for any direct,
33// indirect, incidental, special, exemplary, or consequential damages
34// (including, but not limited to, procurement of substitute goods or services;
35// loss of use, data, or profits; or business interruption) however caused
36// and on any theory of liability, whether in contract, strict liability,
37// or tort (including negligence or otherwise) arising in any way out of
38// the use of this software, even if advised of the possibility of such damage.
39//
40//M*/
41
42/* Haar features calculation */
43
44#include "precomp.hpp"
45#include "opencv2/imgproc/imgproc_c.h"
46#include "opencv2/objdetect/objdetect_c.h"
47#include <stdio.h>
48
49#if CV_SSE2
50#   if 1 /*!CV_SSE4_1 && !CV_SSE4_2*/
51#       define _mm_blendv_pd(a, b, m) _mm_xor_pd(a, _mm_and_pd(_mm_xor_pd(b, a), m))
52#       define _mm_blendv_ps(a, b, m) _mm_xor_ps(a, _mm_and_ps(_mm_xor_ps(b, a), m))
53#   endif
54#endif
55
56#if 0 /*CV_AVX*/
57#  define CV_HAAR_USE_AVX 1
58#  if defined _MSC_VER
59#    pragma warning( disable : 4752 )
60#  endif
61#else
62#  if CV_SSE2
63#    define CV_HAAR_USE_SSE 1
64#  endif
65#endif
66
67/* these settings affect the quality of detection: change with care */
68#define CV_ADJUST_FEATURES 1
69#define CV_ADJUST_WEIGHTS  0
70
71typedef int sumtype;
72typedef double sqsumtype;
73
74typedef struct CvHidHaarFeature
75{
76    struct
77    {
78        sumtype *p0, *p1, *p2, *p3;
79        float weight;
80    }
81    rect[CV_HAAR_FEATURE_MAX];
82} CvHidHaarFeature;
83
84
85typedef struct CvHidHaarTreeNode
86{
87    CvHidHaarFeature feature;
88    float threshold;
89    int left;
90    int right;
91} CvHidHaarTreeNode;
92
93
94typedef struct CvHidHaarClassifier
95{
96    int count;
97    //CvHaarFeature* orig_feature;
98    CvHidHaarTreeNode* node;
99    float* alpha;
100} CvHidHaarClassifier;
101
102
103typedef struct CvHidHaarStageClassifier
104{
105    int  count;
106    float threshold;
107    CvHidHaarClassifier* classifier;
108    int two_rects;
109
110    struct CvHidHaarStageClassifier* next;
111    struct CvHidHaarStageClassifier* child;
112    struct CvHidHaarStageClassifier* parent;
113} CvHidHaarStageClassifier;
114
115
116typedef struct CvHidHaarClassifierCascade
117{
118    int  count;
119    int  isStumpBased;
120    int  has_tilted_features;
121    int  is_tree;
122    double inv_window_area;
123    CvMat sum, sqsum, tilted;
124    CvHidHaarStageClassifier* stage_classifier;
125    sqsumtype *pq0, *pq1, *pq2, *pq3;
126    sumtype *p0, *p1, *p2, *p3;
127
128    void** ipp_stages;
129} CvHidHaarClassifierCascade;
130
131
132const int icv_object_win_border = 1;
133const float icv_stage_threshold_bias = 0.0001f;
134
135static CvHaarClassifierCascade*
136icvCreateHaarClassifierCascade( int stage_count )
137{
138    CvHaarClassifierCascade* cascade = 0;
139
140    int block_size = sizeof(*cascade) + stage_count*sizeof(*cascade->stage_classifier);
141
142    if( stage_count <= 0 )
143        CV_Error( CV_StsOutOfRange, "Number of stages should be positive" );
144
145    cascade = (CvHaarClassifierCascade*)cvAlloc( block_size );
146    memset( cascade, 0, block_size );
147
148    cascade->stage_classifier = (CvHaarStageClassifier*)(cascade + 1);
149    cascade->flags = CV_HAAR_MAGIC_VAL;
150    cascade->count = stage_count;
151
152    return cascade;
153}
154
155static void
156icvReleaseHidHaarClassifierCascade( CvHidHaarClassifierCascade** _cascade )
157{
158    if( _cascade && *_cascade )
159    {
160#ifdef HAVE_IPP
161        CvHidHaarClassifierCascade* cascade = *_cascade;
162        if( CV_IPP_CHECK_COND && cascade->ipp_stages )
163        {
164            int i;
165            for( i = 0; i < cascade->count; i++ )
166            {
167                if( cascade->ipp_stages[i] )
168                    ippiHaarClassifierFree_32f( (IppiHaarClassifier_32f*)cascade->ipp_stages[i] );
169            }
170        }
171        cvFree( &cascade->ipp_stages );
172#endif
173        cvFree( _cascade );
174    }
175}
176
177/* create more efficient internal representation of haar classifier cascade */
178static CvHidHaarClassifierCascade*
179icvCreateHidHaarClassifierCascade( CvHaarClassifierCascade* cascade )
180{
181    CvRect* ipp_features = 0;
182    float *ipp_weights = 0, *ipp_thresholds = 0, *ipp_val1 = 0, *ipp_val2 = 0;
183    int* ipp_counts = 0;
184
185    CvHidHaarClassifierCascade* out = 0;
186
187    int i, j, k, l;
188    int datasize;
189    int total_classifiers = 0;
190    int total_nodes = 0;
191    char errorstr[1000];
192    CvHidHaarClassifier* haar_classifier_ptr;
193    CvHidHaarTreeNode* haar_node_ptr;
194    CvSize orig_window_size;
195    int has_tilted_features = 0;
196    int max_count = 0;
197
198    if( !CV_IS_HAAR_CLASSIFIER(cascade) )
199        CV_Error( !cascade ? CV_StsNullPtr : CV_StsBadArg, "Invalid classifier pointer" );
200
201    if( cascade->hid_cascade )
202        CV_Error( CV_StsError, "hid_cascade has been already created" );
203
204    if( !cascade->stage_classifier )
205        CV_Error( CV_StsNullPtr, "" );
206
207    if( cascade->count <= 0 )
208        CV_Error( CV_StsOutOfRange, "Negative number of cascade stages" );
209
210    orig_window_size = cascade->orig_window_size;
211
212    /* check input structure correctness and calculate total memory size needed for
213       internal representation of the classifier cascade */
214    for( i = 0; i < cascade->count; i++ )
215    {
216        CvHaarStageClassifier* stage_classifier = cascade->stage_classifier + i;
217
218        if( !stage_classifier->classifier ||
219            stage_classifier->count <= 0 )
220        {
221            sprintf( errorstr, "header of the stage classifier #%d is invalid "
222                     "(has null pointers or non-positive classfier count)", i );
223            CV_Error( CV_StsError, errorstr );
224        }
225
226        max_count = MAX( max_count, stage_classifier->count );
227        total_classifiers += stage_classifier->count;
228
229        for( j = 0; j < stage_classifier->count; j++ )
230        {
231            CvHaarClassifier* classifier = stage_classifier->classifier + j;
232
233            total_nodes += classifier->count;
234            for( l = 0; l < classifier->count; l++ )
235            {
236                for( k = 0; k < CV_HAAR_FEATURE_MAX; k++ )
237                {
238                    if( classifier->haar_feature[l].rect[k].r.width )
239                    {
240                        CvRect r = classifier->haar_feature[l].rect[k].r;
241                        int tilted = classifier->haar_feature[l].tilted;
242                        has_tilted_features |= tilted != 0;
243                        if( r.width < 0 || r.height < 0 || r.y < 0 ||
244                            r.x + r.width > orig_window_size.width
245                            ||
246                            (!tilted &&
247                            (r.x < 0 || r.y + r.height > orig_window_size.height))
248                            ||
249                            (tilted && (r.x - r.height < 0 ||
250                            r.y + r.width + r.height > orig_window_size.height)))
251                        {
252                            sprintf( errorstr, "rectangle #%d of the classifier #%d of "
253                                     "the stage classifier #%d is not inside "
254                                     "the reference (original) cascade window", k, j, i );
255                            CV_Error( CV_StsNullPtr, errorstr );
256                        }
257                    }
258                }
259            }
260        }
261    }
262
263    // this is an upper boundary for the whole hidden cascade size
264    datasize = sizeof(CvHidHaarClassifierCascade) +
265               sizeof(CvHidHaarStageClassifier)*cascade->count +
266               sizeof(CvHidHaarClassifier) * total_classifiers +
267               sizeof(CvHidHaarTreeNode) * total_nodes +
268               sizeof(void*)*(total_nodes + total_classifiers);
269
270    out = (CvHidHaarClassifierCascade*)cvAlloc( datasize );
271    memset( out, 0, sizeof(*out) );
272
273    /* init header */
274    out->count = cascade->count;
275    out->stage_classifier = (CvHidHaarStageClassifier*)(out + 1);
276    haar_classifier_ptr = (CvHidHaarClassifier*)(out->stage_classifier + cascade->count);
277    haar_node_ptr = (CvHidHaarTreeNode*)(haar_classifier_ptr + total_classifiers);
278
279    out->isStumpBased = 1;
280    out->has_tilted_features = has_tilted_features;
281    out->is_tree = 0;
282
283    /* initialize internal representation */
284    for( i = 0; i < cascade->count; i++ )
285    {
286        CvHaarStageClassifier* stage_classifier = cascade->stage_classifier + i;
287        CvHidHaarStageClassifier* hid_stage_classifier = out->stage_classifier + i;
288
289        hid_stage_classifier->count = stage_classifier->count;
290        hid_stage_classifier->threshold = stage_classifier->threshold - icv_stage_threshold_bias;
291        hid_stage_classifier->classifier = haar_classifier_ptr;
292        hid_stage_classifier->two_rects = 1;
293        haar_classifier_ptr += stage_classifier->count;
294
295        hid_stage_classifier->parent = (stage_classifier->parent == -1)
296            ? NULL : out->stage_classifier + stage_classifier->parent;
297        hid_stage_classifier->next = (stage_classifier->next == -1)
298            ? NULL : out->stage_classifier + stage_classifier->next;
299        hid_stage_classifier->child = (stage_classifier->child == -1)
300            ? NULL : out->stage_classifier + stage_classifier->child;
301
302        out->is_tree |= hid_stage_classifier->next != NULL;
303
304        for( j = 0; j < stage_classifier->count; j++ )
305        {
306            CvHaarClassifier* classifier = stage_classifier->classifier + j;
307            CvHidHaarClassifier* hid_classifier = hid_stage_classifier->classifier + j;
308            int node_count = classifier->count;
309            float* alpha_ptr = (float*)(haar_node_ptr + node_count);
310
311            hid_classifier->count = node_count;
312            hid_classifier->node = haar_node_ptr;
313            hid_classifier->alpha = alpha_ptr;
314
315            for( l = 0; l < node_count; l++ )
316            {
317                CvHidHaarTreeNode* node = hid_classifier->node + l;
318                CvHaarFeature* feature = classifier->haar_feature + l;
319                memset( node, -1, sizeof(*node) );
320                node->threshold = classifier->threshold[l];
321                node->left = classifier->left[l];
322                node->right = classifier->right[l];
323
324                if( fabs(feature->rect[2].weight) < DBL_EPSILON ||
325                    feature->rect[2].r.width == 0 ||
326                    feature->rect[2].r.height == 0 )
327                    memset( &(node->feature.rect[2]), 0, sizeof(node->feature.rect[2]) );
328                else
329                    hid_stage_classifier->two_rects = 0;
330            }
331
332            memcpy( alpha_ptr, classifier->alpha, (node_count+1)*sizeof(alpha_ptr[0]));
333            haar_node_ptr =
334                (CvHidHaarTreeNode*)cvAlignPtr(alpha_ptr+node_count+1, sizeof(void*));
335
336            out->isStumpBased &= node_count == 1;
337        }
338    }
339/*
340#ifdef HAVE_IPP
341    int can_use_ipp = CV_IPP_CHECK_COND && (!out->has_tilted_features && !out->is_tree && out->isStumpBased);
342
343    if( can_use_ipp )
344    {
345        int ipp_datasize = cascade->count*sizeof(out->ipp_stages[0]);
346        float ipp_weight_scale=(float)(1./((orig_window_size.width-icv_object_win_border*2)*
347            (orig_window_size.height-icv_object_win_border*2)));
348
349        out->ipp_stages = (void**)cvAlloc( ipp_datasize );
350        memset( out->ipp_stages, 0, ipp_datasize );
351
352        ipp_features = (CvRect*)cvAlloc( max_count*3*sizeof(ipp_features[0]) );
353        ipp_weights = (float*)cvAlloc( max_count*3*sizeof(ipp_weights[0]) );
354        ipp_thresholds = (float*)cvAlloc( max_count*sizeof(ipp_thresholds[0]) );
355        ipp_val1 = (float*)cvAlloc( max_count*sizeof(ipp_val1[0]) );
356        ipp_val2 = (float*)cvAlloc( max_count*sizeof(ipp_val2[0]) );
357        ipp_counts = (int*)cvAlloc( max_count*sizeof(ipp_counts[0]) );
358
359        for( i = 0; i < cascade->count; i++ )
360        {
361            CvHaarStageClassifier* stage_classifier = cascade->stage_classifier + i;
362            for( j = 0, k = 0; j < stage_classifier->count; j++ )
363            {
364                CvHaarClassifier* classifier = stage_classifier->classifier + j;
365                int rect_count = 2 + (classifier->haar_feature->rect[2].r.width != 0);
366
367                ipp_thresholds[j] = classifier->threshold[0];
368                ipp_val1[j] = classifier->alpha[0];
369                ipp_val2[j] = classifier->alpha[1];
370                ipp_counts[j] = rect_count;
371
372                for( l = 0; l < rect_count; l++, k++ )
373                {
374                    ipp_features[k] = classifier->haar_feature->rect[l].r;
375                    //ipp_features[k].y = orig_window_size.height - ipp_features[k].y - ipp_features[k].height;
376                    ipp_weights[k] = classifier->haar_feature->rect[l].weight*ipp_weight_scale;
377                }
378            }
379
380            if( ippiHaarClassifierInitAlloc_32f( (IppiHaarClassifier_32f**)&out->ipp_stages[i],
381                (const IppiRect*)ipp_features, ipp_weights, ipp_thresholds,
382                ipp_val1, ipp_val2, ipp_counts, stage_classifier->count ) < 0 )
383                break;
384        }
385
386        if( i < cascade->count )
387        {
388            for( j = 0; j < i; j++ )
389                if( out->ipp_stages[i] )
390                    ippiHaarClassifierFree_32f( (IppiHaarClassifier_32f*)out->ipp_stages[i] );
391            cvFree( &out->ipp_stages );
392        }
393    }
394#endif
395*/
396    cascade->hid_cascade = out;
397    assert( (char*)haar_node_ptr - (char*)out <= datasize );
398
399    cvFree( &ipp_features );
400    cvFree( &ipp_weights );
401    cvFree( &ipp_thresholds );
402    cvFree( &ipp_val1 );
403    cvFree( &ipp_val2 );
404    cvFree( &ipp_counts );
405
406    return out;
407}
408
409
410#define sum_elem_ptr(sum,row,col)  \
411    ((sumtype*)CV_MAT_ELEM_PTR_FAST((sum),(row),(col),sizeof(sumtype)))
412
413#define sqsum_elem_ptr(sqsum,row,col)  \
414    ((sqsumtype*)CV_MAT_ELEM_PTR_FAST((sqsum),(row),(col),sizeof(sqsumtype)))
415
416#define calc_sum(rect,offset) \
417    ((rect).p0[offset] - (rect).p1[offset] - (rect).p2[offset] + (rect).p3[offset])
418
419#define calc_sumf(rect,offset) \
420    static_cast<float>((rect).p0[offset] - (rect).p1[offset] - (rect).p2[offset] + (rect).p3[offset])
421
422
423CV_IMPL void
424cvSetImagesForHaarClassifierCascade( CvHaarClassifierCascade* _cascade,
425                                     const CvArr* _sum,
426                                     const CvArr* _sqsum,
427                                     const CvArr* _tilted_sum,
428                                     double scale )
429{
430    CvMat sum_stub, *sum = (CvMat*)_sum;
431    CvMat sqsum_stub, *sqsum = (CvMat*)_sqsum;
432    CvMat tilted_stub, *tilted = (CvMat*)_tilted_sum;
433    CvHidHaarClassifierCascade* cascade;
434    int coi0 = 0, coi1 = 0;
435    int i;
436    CvRect equRect;
437    double weight_scale;
438
439    if( !CV_IS_HAAR_CLASSIFIER(_cascade) )
440        CV_Error( !_cascade ? CV_StsNullPtr : CV_StsBadArg, "Invalid classifier pointer" );
441
442    if( scale <= 0 )
443        CV_Error( CV_StsOutOfRange, "Scale must be positive" );
444
445    sum = cvGetMat( sum, &sum_stub, &coi0 );
446    sqsum = cvGetMat( sqsum, &sqsum_stub, &coi1 );
447
448    if( coi0 || coi1 )
449        CV_Error( CV_BadCOI, "COI is not supported" );
450
451    if( !CV_ARE_SIZES_EQ( sum, sqsum ))
452        CV_Error( CV_StsUnmatchedSizes, "All integral images must have the same size" );
453
454    if( CV_MAT_TYPE(sqsum->type) != CV_64FC1 ||
455        CV_MAT_TYPE(sum->type) != CV_32SC1 )
456        CV_Error( CV_StsUnsupportedFormat,
457        "Only (32s, 64f, 32s) combination of (sum,sqsum,tilted_sum) formats is allowed" );
458
459    if( !_cascade->hid_cascade )
460        icvCreateHidHaarClassifierCascade(_cascade);
461
462    cascade = _cascade->hid_cascade;
463
464    if( cascade->has_tilted_features )
465    {
466        tilted = cvGetMat( tilted, &tilted_stub, &coi1 );
467
468        if( CV_MAT_TYPE(tilted->type) != CV_32SC1 )
469            CV_Error( CV_StsUnsupportedFormat,
470            "Only (32s, 64f, 32s) combination of (sum,sqsum,tilted_sum) formats is allowed" );
471
472        if( sum->step != tilted->step )
473            CV_Error( CV_StsUnmatchedSizes,
474            "Sum and tilted_sum must have the same stride (step, widthStep)" );
475
476        if( !CV_ARE_SIZES_EQ( sum, tilted ))
477            CV_Error( CV_StsUnmatchedSizes, "All integral images must have the same size" );
478        cascade->tilted = *tilted;
479    }
480
481    _cascade->scale = scale;
482    _cascade->real_window_size.width = cvRound( _cascade->orig_window_size.width * scale );
483    _cascade->real_window_size.height = cvRound( _cascade->orig_window_size.height * scale );
484
485    cascade->sum = *sum;
486    cascade->sqsum = *sqsum;
487
488    equRect.x = equRect.y = cvRound(scale);
489    equRect.width = cvRound((_cascade->orig_window_size.width-2)*scale);
490    equRect.height = cvRound((_cascade->orig_window_size.height-2)*scale);
491    weight_scale = 1./(equRect.width*equRect.height);
492    cascade->inv_window_area = weight_scale;
493
494    cascade->p0 = sum_elem_ptr(*sum, equRect.y, equRect.x);
495    cascade->p1 = sum_elem_ptr(*sum, equRect.y, equRect.x + equRect.width );
496    cascade->p2 = sum_elem_ptr(*sum, equRect.y + equRect.height, equRect.x );
497    cascade->p3 = sum_elem_ptr(*sum, equRect.y + equRect.height,
498                                     equRect.x + equRect.width );
499
500    cascade->pq0 = sqsum_elem_ptr(*sqsum, equRect.y, equRect.x);
501    cascade->pq1 = sqsum_elem_ptr(*sqsum, equRect.y, equRect.x + equRect.width );
502    cascade->pq2 = sqsum_elem_ptr(*sqsum, equRect.y + equRect.height, equRect.x );
503    cascade->pq3 = sqsum_elem_ptr(*sqsum, equRect.y + equRect.height,
504                                          equRect.x + equRect.width );
505
506    /* init pointers in haar features according to real window size and
507       given image pointers */
508    for( i = 0; i < _cascade->count; i++ )
509    {
510        int j, k, l;
511        for( j = 0; j < cascade->stage_classifier[i].count; j++ )
512        {
513            for( l = 0; l < cascade->stage_classifier[i].classifier[j].count; l++ )
514            {
515                CvHaarFeature* feature =
516                    &_cascade->stage_classifier[i].classifier[j].haar_feature[l];
517                /* CvHidHaarClassifier* classifier =
518                    cascade->stage_classifier[i].classifier + j; */
519                CvHidHaarFeature* hidfeature =
520                    &cascade->stage_classifier[i].classifier[j].node[l].feature;
521                double sum0 = 0, area0 = 0;
522                CvRect r[3];
523
524                int base_w = -1, base_h = -1;
525                int new_base_w = 0, new_base_h = 0;
526                int kx, ky;
527                int flagx = 0, flagy = 0;
528                int x0 = 0, y0 = 0;
529                int nr;
530
531                /* align blocks */
532                for( k = 0; k < CV_HAAR_FEATURE_MAX; k++ )
533                {
534                    if( !hidfeature->rect[k].p0 )
535                        break;
536                    r[k] = feature->rect[k].r;
537                    base_w = (int)CV_IMIN( (unsigned)base_w, (unsigned)(r[k].width-1) );
538                    base_w = (int)CV_IMIN( (unsigned)base_w, (unsigned)(r[k].x - r[0].x-1) );
539                    base_h = (int)CV_IMIN( (unsigned)base_h, (unsigned)(r[k].height-1) );
540                    base_h = (int)CV_IMIN( (unsigned)base_h, (unsigned)(r[k].y - r[0].y-1) );
541                }
542
543                nr = k;
544
545                base_w += 1;
546                base_h += 1;
547                kx = r[0].width / base_w;
548                ky = r[0].height / base_h;
549
550                if( kx <= 0 )
551                {
552                    flagx = 1;
553                    new_base_w = cvRound( r[0].width * scale ) / kx;
554                    x0 = cvRound( r[0].x * scale );
555                }
556
557                if( ky <= 0 )
558                {
559                    flagy = 1;
560                    new_base_h = cvRound( r[0].height * scale ) / ky;
561                    y0 = cvRound( r[0].y * scale );
562                }
563
564                for( k = 0; k < nr; k++ )
565                {
566                    CvRect tr;
567                    double correction_ratio;
568
569                    if( flagx )
570                    {
571                        tr.x = (r[k].x - r[0].x) * new_base_w / base_w + x0;
572                        tr.width = r[k].width * new_base_w / base_w;
573                    }
574                    else
575                    {
576                        tr.x = cvRound( r[k].x * scale );
577                        tr.width = cvRound( r[k].width * scale );
578                    }
579
580                    if( flagy )
581                    {
582                        tr.y = (r[k].y - r[0].y) * new_base_h / base_h + y0;
583                        tr.height = r[k].height * new_base_h / base_h;
584                    }
585                    else
586                    {
587                        tr.y = cvRound( r[k].y * scale );
588                        tr.height = cvRound( r[k].height * scale );
589                    }
590
591#if CV_ADJUST_WEIGHTS
592                    {
593                    // RAINER START
594                    const float orig_feature_size =  (float)(feature->rect[k].r.width)*feature->rect[k].r.height;
595                    const float orig_norm_size = (float)(_cascade->orig_window_size.width)*(_cascade->orig_window_size.height);
596                    const float feature_size = float(tr.width*tr.height);
597                    //const float normSize    = float(equRect.width*equRect.height);
598                    float target_ratio = orig_feature_size / orig_norm_size;
599                    //float isRatio = featureSize / normSize;
600                    //correctionRatio = targetRatio / isRatio / normSize;
601                    correction_ratio = target_ratio / feature_size;
602                    // RAINER END
603                    }
604#else
605                    correction_ratio = weight_scale * (!feature->tilted ? 1 : 0.5);
606#endif
607
608                    if( !feature->tilted )
609                    {
610                        hidfeature->rect[k].p0 = sum_elem_ptr(*sum, tr.y, tr.x);
611                        hidfeature->rect[k].p1 = sum_elem_ptr(*sum, tr.y, tr.x + tr.width);
612                        hidfeature->rect[k].p2 = sum_elem_ptr(*sum, tr.y + tr.height, tr.x);
613                        hidfeature->rect[k].p3 = sum_elem_ptr(*sum, tr.y + tr.height, tr.x + tr.width);
614                    }
615                    else
616                    {
617                        hidfeature->rect[k].p2 = sum_elem_ptr(*tilted, tr.y + tr.width, tr.x + tr.width);
618                        hidfeature->rect[k].p3 = sum_elem_ptr(*tilted, tr.y + tr.width + tr.height,
619                                                              tr.x + tr.width - tr.height);
620                        hidfeature->rect[k].p0 = sum_elem_ptr(*tilted, tr.y, tr.x);
621                        hidfeature->rect[k].p1 = sum_elem_ptr(*tilted, tr.y + tr.height, tr.x - tr.height);
622                    }
623
624                    hidfeature->rect[k].weight = (float)(feature->rect[k].weight * correction_ratio);
625
626                    if( k == 0 )
627                        area0 = tr.width * tr.height;
628                    else
629                        sum0 += hidfeature->rect[k].weight * tr.width * tr.height;
630                }
631
632                hidfeature->rect[0].weight = (float)(-sum0/area0);
633            } /* l */
634        } /* j */
635    }
636}
637
638
639// AVX version icvEvalHidHaarClassifier.  Process 8 CvHidHaarClassifiers per call. Check AVX support before invocation!!
640#ifdef CV_HAAR_USE_AVX
641CV_INLINE
642double icvEvalHidHaarClassifierAVX( CvHidHaarClassifier* classifier,
643                                    double variance_norm_factor, size_t p_offset )
644{
645    int  CV_DECL_ALIGNED(32) idxV[8] = {0,0,0,0,0,0,0,0};
646    uchar flags[8] = {0,0,0,0,0,0,0,0};
647    CvHidHaarTreeNode* nodes[8];
648    double res = 0;
649    uchar exitConditionFlag = 0;
650    for(;;)
651    {
652        float CV_DECL_ALIGNED(32) tmp[8] = {0,0,0,0,0,0,0,0};
653        nodes[0] = (classifier+0)->node + idxV[0];
654        nodes[1] = (classifier+1)->node + idxV[1];
655        nodes[2] = (classifier+2)->node + idxV[2];
656        nodes[3] = (classifier+3)->node + idxV[3];
657        nodes[4] = (classifier+4)->node + idxV[4];
658        nodes[5] = (classifier+5)->node + idxV[5];
659        nodes[6] = (classifier+6)->node + idxV[6];
660        nodes[7] = (classifier+7)->node + idxV[7];
661
662        __m256 t = _mm256_set1_ps(static_cast<float>(variance_norm_factor));
663
664        t = _mm256_mul_ps(t, _mm256_set_ps(nodes[7]->threshold,
665                                           nodes[6]->threshold,
666                                           nodes[5]->threshold,
667                                           nodes[4]->threshold,
668                                           nodes[3]->threshold,
669                                           nodes[2]->threshold,
670                                           nodes[1]->threshold,
671                                           nodes[0]->threshold));
672
673        __m256 offset = _mm256_set_ps(calc_sumf(nodes[7]->feature.rect[0], p_offset),
674                                      calc_sumf(nodes[6]->feature.rect[0], p_offset),
675                                      calc_sumf(nodes[5]->feature.rect[0], p_offset),
676                                      calc_sumf(nodes[4]->feature.rect[0], p_offset),
677                                      calc_sumf(nodes[3]->feature.rect[0], p_offset),
678                                      calc_sumf(nodes[2]->feature.rect[0], p_offset),
679                                      calc_sumf(nodes[1]->feature.rect[0], p_offset),
680                                      calc_sumf(nodes[0]->feature.rect[0], p_offset));
681
682        __m256 weight = _mm256_set_ps(nodes[7]->feature.rect[0].weight,
683                                      nodes[6]->feature.rect[0].weight,
684                                      nodes[5]->feature.rect[0].weight,
685                                      nodes[4]->feature.rect[0].weight,
686                                      nodes[3]->feature.rect[0].weight,
687                                      nodes[2]->feature.rect[0].weight,
688                                      nodes[1]->feature.rect[0].weight,
689                                      nodes[0]->feature.rect[0].weight);
690
691        __m256 sum = _mm256_mul_ps(offset, weight);
692
693        offset = _mm256_set_ps(calc_sumf(nodes[7]->feature.rect[1], p_offset),
694                               calc_sumf(nodes[6]->feature.rect[1], p_offset),
695                               calc_sumf(nodes[5]->feature.rect[1], p_offset),
696                               calc_sumf(nodes[4]->feature.rect[1], p_offset),
697                               calc_sumf(nodes[3]->feature.rect[1], p_offset),
698                               calc_sumf(nodes[2]->feature.rect[1], p_offset),
699                               calc_sumf(nodes[1]->feature.rect[1], p_offset),
700                               calc_sumf(nodes[0]->feature.rect[1], p_offset));
701
702        weight = _mm256_set_ps(nodes[7]->feature.rect[1].weight,
703                               nodes[6]->feature.rect[1].weight,
704                               nodes[5]->feature.rect[1].weight,
705                               nodes[4]->feature.rect[1].weight,
706                               nodes[3]->feature.rect[1].weight,
707                               nodes[2]->feature.rect[1].weight,
708                               nodes[1]->feature.rect[1].weight,
709                               nodes[0]->feature.rect[1].weight);
710
711        sum = _mm256_add_ps(sum, _mm256_mul_ps(offset, weight));
712
713        if( nodes[0]->feature.rect[2].p0 )
714            tmp[0] = calc_sumf(nodes[0]->feature.rect[2], p_offset) * nodes[0]->feature.rect[2].weight;
715        if( nodes[1]->feature.rect[2].p0 )
716            tmp[1] = calc_sumf(nodes[1]->feature.rect[2], p_offset) * nodes[1]->feature.rect[2].weight;
717        if( nodes[2]->feature.rect[2].p0 )
718            tmp[2] = calc_sumf(nodes[2]->feature.rect[2], p_offset) * nodes[2]->feature.rect[2].weight;
719        if( nodes[3]->feature.rect[2].p0 )
720            tmp[3] = calc_sumf(nodes[3]->feature.rect[2], p_offset) * nodes[3]->feature.rect[2].weight;
721        if( nodes[4]->feature.rect[2].p0 )
722            tmp[4] = calc_sumf(nodes[4]->feature.rect[2], p_offset) * nodes[4]->feature.rect[2].weight;
723        if( nodes[5]->feature.rect[2].p0 )
724            tmp[5] = calc_sumf(nodes[5]->feature.rect[2], p_offset) * nodes[5]->feature.rect[2].weight;
725        if( nodes[6]->feature.rect[2].p0 )
726            tmp[6] = calc_sumf(nodes[6]->feature.rect[2], p_offset) * nodes[6]->feature.rect[2].weight;
727        if( nodes[7]->feature.rect[2].p0 )
728            tmp[7] = calc_sumf(nodes[7]->feature.rect[2], p_offset) * nodes[7]->feature.rect[2].weight;
729
730        sum = _mm256_add_ps(sum,_mm256_load_ps(tmp));
731
732        __m256 left  = _mm256_set_ps(static_cast<float>(nodes[7]->left), static_cast<float>(nodes[6]->left),
733                                     static_cast<float>(nodes[5]->left), static_cast<float>(nodes[4]->left),
734                                     static_cast<float>(nodes[3]->left), static_cast<float>(nodes[2]->left),
735                                     static_cast<float>(nodes[1]->left), static_cast<float>(nodes[0]->left));
736        __m256 right = _mm256_set_ps(static_cast<float>(nodes[7]->right),static_cast<float>(nodes[6]->right),
737                                     static_cast<float>(nodes[5]->right),static_cast<float>(nodes[4]->right),
738                                     static_cast<float>(nodes[3]->right),static_cast<float>(nodes[2]->right),
739                                     static_cast<float>(nodes[1]->right),static_cast<float>(nodes[0]->right));
740
741        _mm256_store_si256((__m256i*)idxV, _mm256_cvttps_epi32(_mm256_blendv_ps(right, left, _mm256_cmp_ps(sum, t, _CMP_LT_OQ))));
742
743        for(int i = 0; i < 8; i++)
744        {
745            if(idxV[i]<=0)
746            {
747                if(!flags[i])
748                {
749                    exitConditionFlag++;
750                    flags[i] = 1;
751                    res += (classifier+i)->alpha[-idxV[i]];
752                }
753                idxV[i]=0;
754            }
755        }
756        if(exitConditionFlag == 8)
757            return res;
758    }
759}
760#endif //CV_HAAR_USE_AVX
761
762CV_INLINE
763double icvEvalHidHaarClassifier( CvHidHaarClassifier* classifier,
764                                 double variance_norm_factor,
765                                 size_t p_offset )
766{
767    int idx = 0;
768    /*#if CV_HAAR_USE_SSE && !CV_HAAR_USE_AVX
769        if(cv::checkHardwareSupport(CV_CPU_SSE2))//based on old SSE variant. Works slow
770        {
771            double CV_DECL_ALIGNED(16) temp[2];
772            __m128d zero = _mm_setzero_pd();
773            do
774            {
775                CvHidHaarTreeNode* node = classifier->node + idx;
776                __m128d t = _mm_set1_pd((node->threshold)*variance_norm_factor);
777                __m128d left = _mm_set1_pd(node->left);
778                __m128d right = _mm_set1_pd(node->right);
779
780                double _sum = calc_sum(node->feature.rect[0],p_offset) * node->feature.rect[0].weight;
781                _sum += calc_sum(node->feature.rect[1],p_offset) * node->feature.rect[1].weight;
782                if( node->feature.rect[2].p0 )
783                    _sum += calc_sum(node->feature.rect[2],p_offset) * node->feature.rect[2].weight;
784
785                __m128d sum = _mm_set1_pd(_sum);
786                t = _mm_cmplt_sd(sum, t);
787                sum = _mm_blendv_pd(right, left, t);
788
789                _mm_store_pd(temp, sum);
790                idx = (int)temp[0];
791            }
792            while(idx > 0 );
793
794        }
795        else
796    #endif*/
797    {
798        do
799        {
800            CvHidHaarTreeNode* node = classifier->node + idx;
801            double t = node->threshold * variance_norm_factor;
802
803            double sum = calc_sum(node->feature.rect[0],p_offset) * node->feature.rect[0].weight;
804            sum += calc_sum(node->feature.rect[1],p_offset) * node->feature.rect[1].weight;
805
806            if( node->feature.rect[2].p0 )
807                sum += calc_sum(node->feature.rect[2],p_offset) * node->feature.rect[2].weight;
808
809            idx = sum < t ? node->left : node->right;
810        }
811        while( idx > 0 );
812    }
813    return classifier->alpha[-idx];
814}
815
816
817
818static int
819cvRunHaarClassifierCascadeSum( const CvHaarClassifierCascade* _cascade,
820                               CvPoint pt, double& stage_sum, int start_stage )
821{
822#ifdef CV_HAAR_USE_AVX
823    bool haveAVX = false;
824    if(cv::checkHardwareSupport(CV_CPU_AVX))
825    if(__xgetbv()&0x6)// Check if the OS will save the YMM registers
826       haveAVX = true;
827#else
828#  ifdef CV_HAAR_USE_SSE
829    bool haveSSE2 = cv::checkHardwareSupport(CV_CPU_SSE2);
830#  endif
831#endif
832
833    int p_offset, pq_offset;
834    int i, j;
835    double mean, variance_norm_factor;
836    CvHidHaarClassifierCascade* cascade;
837
838    if( !CV_IS_HAAR_CLASSIFIER(_cascade) )
839        CV_Error( !_cascade ? CV_StsNullPtr : CV_StsBadArg, "Invalid cascade pointer" );
840
841    cascade = _cascade->hid_cascade;
842    if( !cascade )
843        CV_Error( CV_StsNullPtr, "Hidden cascade has not been created.\n"
844            "Use cvSetImagesForHaarClassifierCascade" );
845
846    if( pt.x < 0 || pt.y < 0 ||
847        pt.x + _cascade->real_window_size.width >= cascade->sum.width ||
848        pt.y + _cascade->real_window_size.height >= cascade->sum.height )
849        return -1;
850
851    p_offset = pt.y * (cascade->sum.step/sizeof(sumtype)) + pt.x;
852    pq_offset = pt.y * (cascade->sqsum.step/sizeof(sqsumtype)) + pt.x;
853    mean = calc_sum(*cascade,p_offset)*cascade->inv_window_area;
854    variance_norm_factor = cascade->pq0[pq_offset] - cascade->pq1[pq_offset] -
855                           cascade->pq2[pq_offset] + cascade->pq3[pq_offset];
856    variance_norm_factor = variance_norm_factor*cascade->inv_window_area - mean*mean;
857    if( variance_norm_factor >= 0. )
858        variance_norm_factor = std::sqrt(variance_norm_factor);
859    else
860        variance_norm_factor = 1.;
861
862    if( cascade->is_tree )
863    {
864        CvHidHaarStageClassifier* ptr = cascade->stage_classifier;
865        assert( start_stage == 0 );
866
867        while( ptr )
868        {
869            stage_sum = 0.0;
870            j = 0;
871
872#ifdef CV_HAAR_USE_AVX
873            if(haveAVX)
874            {
875                for( ; j <= ptr->count - 8; j += 8 )
876                {
877                    stage_sum += icvEvalHidHaarClassifierAVX(
878                        ptr->classifier + j,
879                        variance_norm_factor, p_offset );
880                }
881            }
882#endif
883            for( ; j < ptr->count; j++ )
884            {
885                stage_sum += icvEvalHidHaarClassifier( ptr->classifier + j, variance_norm_factor, p_offset );
886            }
887
888            if( stage_sum >= ptr->threshold )
889            {
890                ptr = ptr->child;
891            }
892            else
893            {
894                while( ptr && ptr->next == NULL ) ptr = ptr->parent;
895                if( ptr == NULL )
896                    return 0;
897                ptr = ptr->next;
898            }
899        }
900    }
901    else if( cascade->isStumpBased )
902    {
903#ifdef CV_HAAR_USE_AVX
904        if(haveAVX)
905        {
906            CvHidHaarClassifier* classifiers[8];
907            CvHidHaarTreeNode* nodes[8];
908            for( i = start_stage; i < cascade->count; i++ )
909            {
910                stage_sum = 0.0;
911                j = 0;
912                float CV_DECL_ALIGNED(32) buf[8];
913                if( cascade->stage_classifier[i].two_rects )
914                {
915                    for( ; j <= cascade->stage_classifier[i].count - 8; j += 8 )
916                    {
917                        classifiers[0] = cascade->stage_classifier[i].classifier + j;
918                        nodes[0] = classifiers[0]->node;
919                        classifiers[1] = cascade->stage_classifier[i].classifier + j + 1;
920                        nodes[1] = classifiers[1]->node;
921                        classifiers[2] = cascade->stage_classifier[i].classifier + j + 2;
922                        nodes[2] = classifiers[2]->node;
923                        classifiers[3] = cascade->stage_classifier[i].classifier + j + 3;
924                        nodes[3] = classifiers[3]->node;
925                        classifiers[4] = cascade->stage_classifier[i].classifier + j + 4;
926                        nodes[4] = classifiers[4]->node;
927                        classifiers[5] = cascade->stage_classifier[i].classifier + j + 5;
928                        nodes[5] = classifiers[5]->node;
929                        classifiers[6] = cascade->stage_classifier[i].classifier + j + 6;
930                        nodes[6] = classifiers[6]->node;
931                        classifiers[7] = cascade->stage_classifier[i].classifier + j + 7;
932                        nodes[7] = classifiers[7]->node;
933
934                        __m256 t = _mm256_set1_ps(static_cast<float>(variance_norm_factor));
935                        t = _mm256_mul_ps(t, _mm256_set_ps(nodes[7]->threshold,
936                                                           nodes[6]->threshold,
937                                                           nodes[5]->threshold,
938                                                           nodes[4]->threshold,
939                                                           nodes[3]->threshold,
940                                                           nodes[2]->threshold,
941                                                           nodes[1]->threshold,
942                                                           nodes[0]->threshold));
943
944                        __m256 offset = _mm256_set_ps(calc_sumf(nodes[7]->feature.rect[0], p_offset),
945                                                      calc_sumf(nodes[6]->feature.rect[0], p_offset),
946                                                      calc_sumf(nodes[5]->feature.rect[0], p_offset),
947                                                      calc_sumf(nodes[4]->feature.rect[0], p_offset),
948                                                      calc_sumf(nodes[3]->feature.rect[0], p_offset),
949                                                      calc_sumf(nodes[2]->feature.rect[0], p_offset),
950                                                      calc_sumf(nodes[1]->feature.rect[0], p_offset),
951                                                      calc_sumf(nodes[0]->feature.rect[0], p_offset));
952
953                        __m256 weight = _mm256_set_ps(nodes[7]->feature.rect[0].weight,
954                                                      nodes[6]->feature.rect[0].weight,
955                                                      nodes[5]->feature.rect[0].weight,
956                                                      nodes[4]->feature.rect[0].weight,
957                                                      nodes[3]->feature.rect[0].weight,
958                                                      nodes[2]->feature.rect[0].weight,
959                                                      nodes[1]->feature.rect[0].weight,
960                                                      nodes[0]->feature.rect[0].weight);
961
962                        __m256 sum = _mm256_mul_ps(offset, weight);
963
964                        offset = _mm256_set_ps(calc_sumf(nodes[7]->feature.rect[1], p_offset),
965                                               calc_sumf(nodes[6]->feature.rect[1], p_offset),
966                                               calc_sumf(nodes[5]->feature.rect[1], p_offset),
967                                               calc_sumf(nodes[4]->feature.rect[1], p_offset),
968                                               calc_sumf(nodes[3]->feature.rect[1], p_offset),
969                                               calc_sumf(nodes[2]->feature.rect[1], p_offset),
970                                               calc_sumf(nodes[1]->feature.rect[1], p_offset),
971                                               calc_sumf(nodes[0]->feature.rect[1], p_offset));
972
973                        weight = _mm256_set_ps(nodes[7]->feature.rect[1].weight,
974                                               nodes[6]->feature.rect[1].weight,
975                                               nodes[5]->feature.rect[1].weight,
976                                               nodes[4]->feature.rect[1].weight,
977                                               nodes[3]->feature.rect[1].weight,
978                                               nodes[2]->feature.rect[1].weight,
979                                               nodes[1]->feature.rect[1].weight,
980                                               nodes[0]->feature.rect[1].weight);
981
982                        sum = _mm256_add_ps(sum, _mm256_mul_ps(offset,weight));
983
984                        __m256 alpha0 = _mm256_set_ps(classifiers[7]->alpha[0],
985                                                      classifiers[6]->alpha[0],
986                                                      classifiers[5]->alpha[0],
987                                                      classifiers[4]->alpha[0],
988                                                      classifiers[3]->alpha[0],
989                                                      classifiers[2]->alpha[0],
990                                                      classifiers[1]->alpha[0],
991                                                      classifiers[0]->alpha[0]);
992                        __m256 alpha1 = _mm256_set_ps(classifiers[7]->alpha[1],
993                                                      classifiers[6]->alpha[1],
994                                                      classifiers[5]->alpha[1],
995                                                      classifiers[4]->alpha[1],
996                                                      classifiers[3]->alpha[1],
997                                                      classifiers[2]->alpha[1],
998                                                      classifiers[1]->alpha[1],
999                                                      classifiers[0]->alpha[1]);
1000
1001                        _mm256_store_ps(buf, _mm256_blendv_ps(alpha0, alpha1, _mm256_cmp_ps(t, sum, _CMP_LE_OQ)));
1002                        stage_sum += (buf[0]+buf[1]+buf[2]+buf[3]+buf[4]+buf[5]+buf[6]+buf[7]);
1003                    }
1004
1005                    for( ; j < cascade->stage_classifier[i].count; j++ )
1006                    {
1007                        CvHidHaarClassifier* classifier = cascade->stage_classifier[i].classifier + j;
1008                        CvHidHaarTreeNode* node = classifier->node;
1009
1010                        double t = node->threshold*variance_norm_factor;
1011                        double sum = calc_sum(node->feature.rect[0],p_offset) * node->feature.rect[0].weight;
1012                        sum += calc_sum(node->feature.rect[1],p_offset) * node->feature.rect[1].weight;
1013                        stage_sum += classifier->alpha[sum >= t];
1014                    }
1015                }
1016                else
1017                {
1018                    for( ; j <= (cascade->stage_classifier[i].count)-8; j+=8 )
1019                    {
1020                        float  CV_DECL_ALIGNED(32) tmp[8] = {0,0,0,0,0,0,0,0};
1021
1022                        classifiers[0] = cascade->stage_classifier[i].classifier + j;
1023                        nodes[0] = classifiers[0]->node;
1024                        classifiers[1] = cascade->stage_classifier[i].classifier + j + 1;
1025                        nodes[1] = classifiers[1]->node;
1026                        classifiers[2] = cascade->stage_classifier[i].classifier + j + 2;
1027                        nodes[2] = classifiers[2]->node;
1028                        classifiers[3] = cascade->stage_classifier[i].classifier + j + 3;
1029                        nodes[3] = classifiers[3]->node;
1030                        classifiers[4] = cascade->stage_classifier[i].classifier + j + 4;
1031                        nodes[4] = classifiers[4]->node;
1032                        classifiers[5] = cascade->stage_classifier[i].classifier + j + 5;
1033                        nodes[5] = classifiers[5]->node;
1034                        classifiers[6] = cascade->stage_classifier[i].classifier + j + 6;
1035                        nodes[6] = classifiers[6]->node;
1036                        classifiers[7] = cascade->stage_classifier[i].classifier + j + 7;
1037                        nodes[7] = classifiers[7]->node;
1038
1039                        __m256 t = _mm256_set1_ps(static_cast<float>(variance_norm_factor));
1040
1041                        t = _mm256_mul_ps(t, _mm256_set_ps(nodes[7]->threshold,
1042                                                           nodes[6]->threshold,
1043                                                           nodes[5]->threshold,
1044                                                           nodes[4]->threshold,
1045                                                           nodes[3]->threshold,
1046                                                           nodes[2]->threshold,
1047                                                           nodes[1]->threshold,
1048                                                           nodes[0]->threshold));
1049
1050                        __m256 offset = _mm256_set_ps(calc_sumf(nodes[7]->feature.rect[0], p_offset),
1051                                                      calc_sumf(nodes[6]->feature.rect[0], p_offset),
1052                                                      calc_sumf(nodes[5]->feature.rect[0], p_offset),
1053                                                      calc_sumf(nodes[4]->feature.rect[0], p_offset),
1054                                                      calc_sumf(nodes[3]->feature.rect[0], p_offset),
1055                                                      calc_sumf(nodes[2]->feature.rect[0], p_offset),
1056                                                      calc_sumf(nodes[1]->feature.rect[0], p_offset),
1057                                                      calc_sumf(nodes[0]->feature.rect[0], p_offset));
1058
1059                        __m256 weight = _mm256_set_ps(nodes[7]->feature.rect[0].weight,
1060                                                      nodes[6]->feature.rect[0].weight,
1061                                                      nodes[5]->feature.rect[0].weight,
1062                                                      nodes[4]->feature.rect[0].weight,
1063                                                      nodes[3]->feature.rect[0].weight,
1064                                                      nodes[2]->feature.rect[0].weight,
1065                                                      nodes[1]->feature.rect[0].weight,
1066                                                      nodes[0]->feature.rect[0].weight);
1067
1068                        __m256 sum = _mm256_mul_ps(offset, weight);
1069
1070                        offset = _mm256_set_ps(calc_sumf(nodes[7]->feature.rect[1], p_offset),
1071                                               calc_sumf(nodes[6]->feature.rect[1], p_offset),
1072                                               calc_sumf(nodes[5]->feature.rect[1], p_offset),
1073                                               calc_sumf(nodes[4]->feature.rect[1], p_offset),
1074                                               calc_sumf(nodes[3]->feature.rect[1], p_offset),
1075                                               calc_sumf(nodes[2]->feature.rect[1], p_offset),
1076                                               calc_sumf(nodes[1]->feature.rect[1], p_offset),
1077                                               calc_sumf(nodes[0]->feature.rect[1], p_offset));
1078
1079                        weight = _mm256_set_ps(nodes[7]->feature.rect[1].weight,
1080                                               nodes[6]->feature.rect[1].weight,
1081                                               nodes[5]->feature.rect[1].weight,
1082                                               nodes[4]->feature.rect[1].weight,
1083                                               nodes[3]->feature.rect[1].weight,
1084                                               nodes[2]->feature.rect[1].weight,
1085                                               nodes[1]->feature.rect[1].weight,
1086                                               nodes[0]->feature.rect[1].weight);
1087
1088                        sum = _mm256_add_ps(sum, _mm256_mul_ps(offset, weight));
1089
1090                        if( nodes[0]->feature.rect[2].p0 )
1091                            tmp[0] = calc_sumf(nodes[0]->feature.rect[2],p_offset) * nodes[0]->feature.rect[2].weight;
1092                        if( nodes[1]->feature.rect[2].p0 )
1093                            tmp[1] = calc_sumf(nodes[1]->feature.rect[2],p_offset) * nodes[1]->feature.rect[2].weight;
1094                        if( nodes[2]->feature.rect[2].p0 )
1095                            tmp[2] = calc_sumf(nodes[2]->feature.rect[2],p_offset) * nodes[2]->feature.rect[2].weight;
1096                        if( nodes[3]->feature.rect[2].p0 )
1097                            tmp[3] = calc_sumf(nodes[3]->feature.rect[2],p_offset) * nodes[3]->feature.rect[2].weight;
1098                        if( nodes[4]->feature.rect[2].p0 )
1099                            tmp[4] = calc_sumf(nodes[4]->feature.rect[2],p_offset) * nodes[4]->feature.rect[2].weight;
1100                        if( nodes[5]->feature.rect[2].p0 )
1101                            tmp[5] = calc_sumf(nodes[5]->feature.rect[2],p_offset) * nodes[5]->feature.rect[2].weight;
1102                        if( nodes[6]->feature.rect[2].p0 )
1103                            tmp[6] = calc_sumf(nodes[6]->feature.rect[2],p_offset) * nodes[6]->feature.rect[2].weight;
1104                        if( nodes[7]->feature.rect[2].p0 )
1105                            tmp[7] = calc_sumf(nodes[7]->feature.rect[2],p_offset) * nodes[7]->feature.rect[2].weight;
1106
1107                        sum = _mm256_add_ps(sum, _mm256_load_ps(tmp));
1108
1109                        __m256 alpha0 = _mm256_set_ps(classifiers[7]->alpha[0],
1110                                                      classifiers[6]->alpha[0],
1111                                                      classifiers[5]->alpha[0],
1112                                                      classifiers[4]->alpha[0],
1113                                                      classifiers[3]->alpha[0],
1114                                                      classifiers[2]->alpha[0],
1115                                                      classifiers[1]->alpha[0],
1116                                                      classifiers[0]->alpha[0]);
1117                        __m256 alpha1 = _mm256_set_ps(classifiers[7]->alpha[1],
1118                                                      classifiers[6]->alpha[1],
1119                                                      classifiers[5]->alpha[1],
1120                                                      classifiers[4]->alpha[1],
1121                                                      classifiers[3]->alpha[1],
1122                                                      classifiers[2]->alpha[1],
1123                                                      classifiers[1]->alpha[1],
1124                                                      classifiers[0]->alpha[1]);
1125
1126                        __m256 outBuf = _mm256_blendv_ps(alpha0, alpha1, _mm256_cmp_ps(t, sum, _CMP_LE_OQ ));
1127                        outBuf = _mm256_hadd_ps(outBuf, outBuf);
1128                        outBuf = _mm256_hadd_ps(outBuf, outBuf);
1129                        _mm256_store_ps(buf, outBuf);
1130                        stage_sum += (buf[0] + buf[4]);
1131                    }
1132
1133                    for( ; j < cascade->stage_classifier[i].count; j++ )
1134                    {
1135                        CvHidHaarClassifier* classifier = cascade->stage_classifier[i].classifier + j;
1136                        CvHidHaarTreeNode* node = classifier->node;
1137
1138                        double t = node->threshold*variance_norm_factor;
1139                        double sum = calc_sum(node->feature.rect[0],p_offset) * node->feature.rect[0].weight;
1140                        sum += calc_sum(node->feature.rect[1],p_offset) * node->feature.rect[1].weight;
1141                        if( node->feature.rect[2].p0 )
1142                            sum += calc_sum(node->feature.rect[2],p_offset) * node->feature.rect[2].weight;
1143                        stage_sum += classifier->alpha[sum >= t];
1144                    }
1145                }
1146                if( stage_sum < cascade->stage_classifier[i].threshold )
1147                    return -i;
1148            }
1149        }
1150        else
1151#elif defined CV_HAAR_USE_SSE //old SSE optimization
1152        if(haveSSE2)
1153        {
1154            for( i = start_stage; i < cascade->count; i++ )
1155            {
1156                __m128d vstage_sum = _mm_setzero_pd();
1157                if( cascade->stage_classifier[i].two_rects )
1158                {
1159                    for( j = 0; j < cascade->stage_classifier[i].count; j++ )
1160                    {
1161                        CvHidHaarClassifier* classifier = cascade->stage_classifier[i].classifier + j;
1162                        CvHidHaarTreeNode* node = classifier->node;
1163
1164                        // ayasin - NHM perf optim. Avoid use of costly flaky jcc
1165                        __m128d t = _mm_set_sd(node->threshold*variance_norm_factor);
1166                        __m128d a = _mm_set_sd(classifier->alpha[0]);
1167                        __m128d b = _mm_set_sd(classifier->alpha[1]);
1168                        __m128d sum = _mm_set_sd(calc_sum(node->feature.rect[0],p_offset) * node->feature.rect[0].weight +
1169                                                 calc_sum(node->feature.rect[1],p_offset) * node->feature.rect[1].weight);
1170                        t = _mm_cmpgt_sd(t, sum);
1171                        vstage_sum = _mm_add_sd(vstage_sum, _mm_blendv_pd(b, a, t));
1172                    }
1173                }
1174                else
1175                {
1176                    for( j = 0; j < cascade->stage_classifier[i].count; j++ )
1177                    {
1178                        CvHidHaarClassifier* classifier = cascade->stage_classifier[i].classifier + j;
1179                        CvHidHaarTreeNode* node = classifier->node;
1180                        // ayasin - NHM perf optim. Avoid use of costly flaky jcc
1181                        __m128d t = _mm_set_sd(node->threshold*variance_norm_factor);
1182                        __m128d a = _mm_set_sd(classifier->alpha[0]);
1183                        __m128d b = _mm_set_sd(classifier->alpha[1]);
1184                        double _sum = calc_sum(node->feature.rect[0],p_offset) * node->feature.rect[0].weight;
1185                        _sum += calc_sum(node->feature.rect[1],p_offset) * node->feature.rect[1].weight;
1186                        if( node->feature.rect[2].p0 )
1187                            _sum += calc_sum(node->feature.rect[2],p_offset) * node->feature.rect[2].weight;
1188                        __m128d sum = _mm_set_sd(_sum);
1189
1190                        t = _mm_cmpgt_sd(t, sum);
1191                        vstage_sum = _mm_add_sd(vstage_sum, _mm_blendv_pd(b, a, t));
1192                    }
1193                }
1194                __m128d i_threshold = _mm_set1_pd(cascade->stage_classifier[i].threshold);
1195                if( _mm_comilt_sd(vstage_sum, i_threshold) )
1196                    return -i;
1197            }
1198        }
1199        else
1200#endif // AVX or SSE
1201        {
1202            for( i = start_stage; i < cascade->count; i++ )
1203            {
1204                stage_sum = 0.0;
1205                if( cascade->stage_classifier[i].two_rects )
1206                {
1207                    for( j = 0; j < cascade->stage_classifier[i].count; j++ )
1208                    {
1209                        CvHidHaarClassifier* classifier = cascade->stage_classifier[i].classifier + j;
1210                        CvHidHaarTreeNode* node = classifier->node;
1211                        double t = node->threshold*variance_norm_factor;
1212                        double sum = calc_sum(node->feature.rect[0],p_offset) * node->feature.rect[0].weight;
1213                        sum += calc_sum(node->feature.rect[1],p_offset) * node->feature.rect[1].weight;
1214                        stage_sum += classifier->alpha[sum >= t];
1215                    }
1216                }
1217                else
1218                {
1219                    for( j = 0; j < cascade->stage_classifier[i].count; j++ )
1220                    {
1221                        CvHidHaarClassifier* classifier = cascade->stage_classifier[i].classifier + j;
1222                        CvHidHaarTreeNode* node = classifier->node;
1223                        double t = node->threshold*variance_norm_factor;
1224                        double sum = calc_sum(node->feature.rect[0],p_offset) * node->feature.rect[0].weight;
1225                        sum += calc_sum(node->feature.rect[1],p_offset) * node->feature.rect[1].weight;
1226                        if( node->feature.rect[2].p0 )
1227                            sum += calc_sum(node->feature.rect[2],p_offset) * node->feature.rect[2].weight;
1228                        stage_sum += classifier->alpha[sum >= t];
1229                    }
1230                }
1231                if( stage_sum < cascade->stage_classifier[i].threshold )
1232                    return -i;
1233            }
1234        }
1235    }
1236    else
1237    {
1238        for( i = start_stage; i < cascade->count; i++ )
1239        {
1240            stage_sum = 0.0;
1241            int k = 0;
1242
1243#ifdef CV_HAAR_USE_AVX
1244            if(haveAVX)
1245            {
1246                for( ; k < cascade->stage_classifier[i].count - 8; k += 8 )
1247                {
1248                    stage_sum += icvEvalHidHaarClassifierAVX(
1249                        cascade->stage_classifier[i].classifier + k,
1250                        variance_norm_factor, p_offset );
1251                }
1252            }
1253#endif
1254            for(; k < cascade->stage_classifier[i].count; k++ )
1255            {
1256
1257                stage_sum += icvEvalHidHaarClassifier(
1258                    cascade->stage_classifier[i].classifier + k,
1259                    variance_norm_factor, p_offset );
1260            }
1261
1262            if( stage_sum < cascade->stage_classifier[i].threshold )
1263                return -i;
1264        }
1265    }
1266    return 1;
1267}
1268
1269
1270CV_IMPL int
1271cvRunHaarClassifierCascade( const CvHaarClassifierCascade* _cascade,
1272                            CvPoint pt, int start_stage )
1273{
1274    double stage_sum;
1275    return cvRunHaarClassifierCascadeSum(_cascade, pt, stage_sum, start_stage);
1276}
1277
1278namespace cv
1279{
1280
1281class HaarDetectObjects_ScaleImage_Invoker : public ParallelLoopBody
1282{
1283public:
1284    HaarDetectObjects_ScaleImage_Invoker( const CvHaarClassifierCascade* _cascade,
1285                                          int _stripSize, double _factor,
1286                                          const Mat& _sum1, const Mat& _sqsum1, Mat* _norm1,
1287                                          Mat* _mask1, Rect _equRect, std::vector<Rect>& _vec,
1288                                          std::vector<int>& _levels, std::vector<double>& _weights,
1289                                          bool _outputLevels, Mutex *_mtx )
1290    {
1291        cascade = _cascade;
1292        stripSize = _stripSize;
1293        factor = _factor;
1294        sum1 = _sum1;
1295        sqsum1 = _sqsum1;
1296        norm1 = _norm1;
1297        mask1 = _mask1;
1298        equRect = _equRect;
1299        vec = &_vec;
1300        rejectLevels = _outputLevels ? &_levels : 0;
1301        levelWeights = _outputLevels ? &_weights : 0;
1302        mtx = _mtx;
1303    }
1304
1305    void operator()( const Range& range ) const
1306    {
1307        Size winSize0 = cascade->orig_window_size;
1308        Size winSize(cvRound(winSize0.width*factor), cvRound(winSize0.height*factor));
1309        int y1 = range.start*stripSize, y2 = std::min(range.end*stripSize, sum1.rows - 1 - winSize0.height);
1310
1311        if (y2 <= y1 || sum1.cols <= 1 + winSize0.width)
1312            return;
1313
1314        Size ssz(sum1.cols - 1 - winSize0.width, y2 - y1);
1315        int x, y, ystep = factor > 2 ? 1 : 2;
1316
1317#ifdef HAVE_IPP
1318        if(CV_IPP_CHECK_COND && cascade->hid_cascade->ipp_stages )
1319        {
1320            IppiRect iequRect = {equRect.x, equRect.y, equRect.width, equRect.height};
1321            ippiRectStdDev_32f_C1R(sum1.ptr<float>(y1), (int)sum1.step,
1322                                   sqsum1.ptr<double>(y1), (int)sqsum1.step,
1323                                   norm1->ptr<float>(y1), (int)norm1->step,
1324                                   ippiSize(ssz.width, ssz.height), iequRect );
1325
1326            int positive = (ssz.width/ystep)*((ssz.height + ystep-1)/ystep);
1327
1328            if( ystep == 1 )
1329                (*mask1) = Scalar::all(1);
1330            else
1331                for( y = y1; y < y2; y++ )
1332                {
1333                    uchar* mask1row = mask1->ptr(y);
1334                    memset( mask1row, 0, ssz.width );
1335
1336                    if( y % ystep == 0 )
1337                        for( x = 0; x < ssz.width; x += ystep )
1338                            mask1row[x] = (uchar)1;
1339                }
1340
1341            for( int j = 0; j < cascade->count; j++ )
1342            {
1343                if( ippiApplyHaarClassifier_32f_C1R(
1344                            sum1.ptr<float>(y1), (int)sum1.step,
1345                            norm1->ptr<float>(y1), (int)norm1->step,
1346                            mask1->ptr<uchar>(y1), (int)mask1->step,
1347                            ippiSize(ssz.width, ssz.height), &positive,
1348                            cascade->hid_cascade->stage_classifier[j].threshold,
1349                            (IppiHaarClassifier_32f*)cascade->hid_cascade->ipp_stages[j]) < 0 )
1350                    positive = 0;
1351                if( positive <= 0 )
1352                    break;
1353            }
1354            CV_IMPL_ADD(CV_IMPL_IPP|CV_IMPL_MT);
1355
1356            if( positive > 0 )
1357                for( y = y1; y < y2; y += ystep )
1358                {
1359                    uchar* mask1row = mask1->ptr(y);
1360                    for( x = 0; x < ssz.width; x += ystep )
1361                        if( mask1row[x] != 0 )
1362                        {
1363                            mtx->lock();
1364                            vec->push_back(Rect(cvRound(x*factor), cvRound(y*factor),
1365                                                winSize.width, winSize.height));
1366                            mtx->unlock();
1367                            if( --positive == 0 )
1368                                break;
1369                        }
1370                    if( positive == 0 )
1371                        break;
1372                }
1373        }
1374        else
1375#endif // IPP
1376            for( y = y1; y < y2; y += ystep )
1377                for( x = 0; x < ssz.width; x += ystep )
1378                {
1379                    double gypWeight;
1380                    int result = cvRunHaarClassifierCascadeSum( cascade, cvPoint(x,y), gypWeight, 0 );
1381                    if( rejectLevels )
1382                    {
1383                        if( result == 1 )
1384                            result = -1*cascade->count;
1385                        if( cascade->count + result < 4 )
1386                        {
1387                            mtx->lock();
1388                            vec->push_back(Rect(cvRound(x*factor), cvRound(y*factor),
1389                                           winSize.width, winSize.height));
1390                            rejectLevels->push_back(-result);
1391                            levelWeights->push_back(gypWeight);
1392                            mtx->unlock();
1393                        }
1394                    }
1395                    else
1396                    {
1397                        if( result > 0 )
1398                        {
1399                            mtx->lock();
1400                            vec->push_back(Rect(cvRound(x*factor), cvRound(y*factor),
1401                                           winSize.width, winSize.height));
1402                            mtx->unlock();
1403                        }
1404                    }
1405                }
1406    }
1407
1408    const CvHaarClassifierCascade* cascade;
1409    int stripSize;
1410    double factor;
1411    Mat sum1, sqsum1, *norm1, *mask1;
1412    Rect equRect;
1413    std::vector<Rect>* vec;
1414    std::vector<int>* rejectLevels;
1415    std::vector<double>* levelWeights;
1416    Mutex* mtx;
1417};
1418
1419
1420class HaarDetectObjects_ScaleCascade_Invoker : public ParallelLoopBody
1421{
1422public:
1423    HaarDetectObjects_ScaleCascade_Invoker( const CvHaarClassifierCascade* _cascade,
1424                                            Size _winsize, const Range& _xrange, double _ystep,
1425                                            size_t _sumstep, const int** _p, const int** _pq,
1426                                            std::vector<Rect>& _vec, Mutex* _mtx )
1427    {
1428        cascade = _cascade;
1429        winsize = _winsize;
1430        xrange = _xrange;
1431        ystep = _ystep;
1432        sumstep = _sumstep;
1433        p = _p; pq = _pq;
1434        vec = &_vec;
1435        mtx = _mtx;
1436    }
1437
1438    void operator()( const Range& range ) const
1439    {
1440        int iy, startY = range.start, endY = range.end;
1441        const int *p0 = p[0], *p1 = p[1], *p2 = p[2], *p3 = p[3];
1442        const int *pq0 = pq[0], *pq1 = pq[1], *pq2 = pq[2], *pq3 = pq[3];
1443        bool doCannyPruning = p0 != 0;
1444        int sstep = (int)(sumstep/sizeof(p0[0]));
1445
1446        for( iy = startY; iy < endY; iy++ )
1447        {
1448            int ix, y = cvRound(iy*ystep), ixstep = 1;
1449            for( ix = xrange.start; ix < xrange.end; ix += ixstep )
1450            {
1451                int x = cvRound(ix*ystep); // it should really be ystep, not ixstep
1452
1453                if( doCannyPruning )
1454                {
1455                    int offset = y*sstep + x;
1456                    int s = p0[offset] - p1[offset] - p2[offset] + p3[offset];
1457                    int sq = pq0[offset] - pq1[offset] - pq2[offset] + pq3[offset];
1458                    if( s < 100 || sq < 20 )
1459                    {
1460                        ixstep = 2;
1461                        continue;
1462                    }
1463                }
1464
1465                int result = cvRunHaarClassifierCascade( cascade, cvPoint(x, y), 0 );
1466                if( result > 0 )
1467                {
1468                    mtx->lock();
1469                    vec->push_back(Rect(x, y, winsize.width, winsize.height));
1470                    mtx->unlock();
1471                }
1472                ixstep = result != 0 ? 1 : 2;
1473            }
1474        }
1475    }
1476
1477    const CvHaarClassifierCascade* cascade;
1478    double ystep;
1479    size_t sumstep;
1480    Size winsize;
1481    Range xrange;
1482    const int** p;
1483    const int** pq;
1484    std::vector<Rect>* vec;
1485    Mutex* mtx;
1486};
1487
1488
1489}
1490
1491
1492CvSeq*
1493cvHaarDetectObjectsForROC( const CvArr* _img,
1494                     CvHaarClassifierCascade* cascade, CvMemStorage* storage,
1495                     std::vector<int>& rejectLevels, std::vector<double>& levelWeights,
1496                     double scaleFactor, int minNeighbors, int flags,
1497                     CvSize minSize, CvSize maxSize, bool outputRejectLevels )
1498{
1499    const double GROUP_EPS = 0.2;
1500    CvMat stub, *img = (CvMat*)_img;
1501    cv::Ptr<CvMat> temp, sum, tilted, sqsum, normImg, sumcanny, imgSmall;
1502    CvSeq* result_seq = 0;
1503    cv::Ptr<CvMemStorage> temp_storage;
1504
1505    std::vector<cv::Rect> allCandidates;
1506    std::vector<cv::Rect> rectList;
1507    std::vector<int> rweights;
1508    double factor;
1509    int coi;
1510    bool doCannyPruning = (flags & CV_HAAR_DO_CANNY_PRUNING) != 0;
1511    bool findBiggestObject = (flags & CV_HAAR_FIND_BIGGEST_OBJECT) != 0;
1512    bool roughSearch = (flags & CV_HAAR_DO_ROUGH_SEARCH) != 0;
1513    cv::Mutex mtx;
1514
1515    if( !CV_IS_HAAR_CLASSIFIER(cascade) )
1516        CV_Error( !cascade ? CV_StsNullPtr : CV_StsBadArg, "Invalid classifier cascade" );
1517
1518    if( !storage )
1519        CV_Error( CV_StsNullPtr, "Null storage pointer" );
1520
1521    img = cvGetMat( img, &stub, &coi );
1522    if( coi )
1523        CV_Error( CV_BadCOI, "COI is not supported" );
1524
1525    if( CV_MAT_DEPTH(img->type) != CV_8U )
1526        CV_Error( CV_StsUnsupportedFormat, "Only 8-bit images are supported" );
1527
1528    if( scaleFactor <= 1 )
1529        CV_Error( CV_StsOutOfRange, "scale factor must be > 1" );
1530
1531    if( findBiggestObject )
1532        flags &= ~CV_HAAR_SCALE_IMAGE;
1533
1534    if( maxSize.height == 0 || maxSize.width == 0 )
1535    {
1536        maxSize.height = img->rows;
1537        maxSize.width = img->cols;
1538    }
1539
1540    temp.reset(cvCreateMat( img->rows, img->cols, CV_8UC1 ));
1541    sum.reset(cvCreateMat( img->rows + 1, img->cols + 1, CV_32SC1 ));
1542    sqsum.reset(cvCreateMat( img->rows + 1, img->cols + 1, CV_64FC1 ));
1543
1544    if( !cascade->hid_cascade )
1545        icvCreateHidHaarClassifierCascade(cascade);
1546
1547    if( cascade->hid_cascade->has_tilted_features )
1548        tilted.reset(cvCreateMat( img->rows + 1, img->cols + 1, CV_32SC1 ));
1549
1550    result_seq = cvCreateSeq( 0, sizeof(CvSeq), sizeof(CvAvgComp), storage );
1551
1552    if( CV_MAT_CN(img->type) > 1 )
1553    {
1554        cvCvtColor( img, temp, CV_BGR2GRAY );
1555        img = temp;
1556    }
1557
1558    if( findBiggestObject )
1559        flags &= ~(CV_HAAR_SCALE_IMAGE|CV_HAAR_DO_CANNY_PRUNING);
1560
1561    if( flags & CV_HAAR_SCALE_IMAGE )
1562    {
1563        CvSize winSize0 = cascade->orig_window_size;
1564#ifdef HAVE_IPP
1565        int use_ipp = CV_IPP_CHECK_COND && (cascade->hid_cascade->ipp_stages != 0);
1566
1567        if( use_ipp )
1568            normImg.reset(cvCreateMat( img->rows, img->cols, CV_32FC1));
1569#endif
1570        imgSmall.reset(cvCreateMat( img->rows + 1, img->cols + 1, CV_8UC1 ));
1571
1572        for( factor = 1; ; factor *= scaleFactor )
1573        {
1574            CvSize winSize(cvRound(winSize0.width*factor),
1575                                cvRound(winSize0.height*factor));
1576            CvSize sz(cvRound( img->cols/factor ), cvRound( img->rows/factor ));
1577            CvSize sz1(sz.width - winSize0.width + 1, sz.height - winSize0.height + 1);
1578
1579            CvRect equRect(icv_object_win_border, icv_object_win_border,
1580                winSize0.width - icv_object_win_border*2,
1581                winSize0.height - icv_object_win_border*2);
1582
1583            CvMat img1, sum1, sqsum1, norm1, tilted1, mask1;
1584            CvMat* _tilted = 0;
1585
1586            if( sz1.width <= 0 || sz1.height <= 0 )
1587                break;
1588            if( winSize.width > maxSize.width || winSize.height > maxSize.height )
1589                break;
1590            if( winSize.width < minSize.width || winSize.height < minSize.height )
1591                continue;
1592
1593            img1 = cvMat( sz.height, sz.width, CV_8UC1, imgSmall->data.ptr );
1594            sum1 = cvMat( sz.height+1, sz.width+1, CV_32SC1, sum->data.ptr );
1595            sqsum1 = cvMat( sz.height+1, sz.width+1, CV_64FC1, sqsum->data.ptr );
1596            if( tilted )
1597            {
1598                tilted1 = cvMat( sz.height+1, sz.width+1, CV_32SC1, tilted->data.ptr );
1599                _tilted = &tilted1;
1600            }
1601            norm1 = cvMat( sz1.height, sz1.width, CV_32FC1, normImg ? normImg->data.ptr : 0 );
1602            mask1 = cvMat( sz1.height, sz1.width, CV_8UC1, temp->data.ptr );
1603
1604            cvResize( img, &img1, CV_INTER_LINEAR );
1605            cvIntegral( &img1, &sum1, &sqsum1, _tilted );
1606
1607            int ystep = factor > 2 ? 1 : 2;
1608            const int LOCS_PER_THREAD = 1000;
1609            int stripCount = ((sz1.width/ystep)*(sz1.height + ystep-1)/ystep + LOCS_PER_THREAD/2)/LOCS_PER_THREAD;
1610            stripCount = std::min(std::max(stripCount, 1), 100);
1611
1612#ifdef HAVE_IPP
1613            if( use_ipp )
1614            {
1615                cv::Mat fsum(sum1.rows, sum1.cols, CV_32F, sum1.data.ptr, sum1.step);
1616                cv::cvarrToMat(&sum1).convertTo(fsum, CV_32F, 1, -(1<<24));
1617            }
1618            else
1619#endif
1620                cvSetImagesForHaarClassifierCascade( cascade, &sum1, &sqsum1, _tilted, 1. );
1621
1622            cv::Mat _norm1 = cv::cvarrToMat(&norm1), _mask1 = cv::cvarrToMat(&mask1);
1623            cv::parallel_for_(cv::Range(0, stripCount),
1624                         cv::HaarDetectObjects_ScaleImage_Invoker(cascade,
1625                                (((sz1.height + stripCount - 1)/stripCount + ystep-1)/ystep)*ystep,
1626                                factor, cv::cvarrToMat(&sum1), cv::cvarrToMat(&sqsum1), &_norm1, &_mask1,
1627                                cv::Rect(equRect), allCandidates, rejectLevels, levelWeights, outputRejectLevels, &mtx));
1628        }
1629    }
1630    else
1631    {
1632        int n_factors = 0;
1633        cv::Rect scanROI;
1634
1635        cvIntegral( img, sum, sqsum, tilted );
1636
1637        if( doCannyPruning )
1638        {
1639            sumcanny.reset(cvCreateMat( img->rows + 1, img->cols + 1, CV_32SC1 ));
1640            cvCanny( img, temp, 0, 50, 3 );
1641            cvIntegral( temp, sumcanny );
1642        }
1643
1644        for( n_factors = 0, factor = 1;
1645             factor*cascade->orig_window_size.width < img->cols - 10 &&
1646             factor*cascade->orig_window_size.height < img->rows - 10;
1647             n_factors++, factor *= scaleFactor )
1648            ;
1649
1650        if( findBiggestObject )
1651        {
1652            scaleFactor = 1./scaleFactor;
1653            factor *= scaleFactor;
1654        }
1655        else
1656            factor = 1;
1657
1658        for( ; n_factors-- > 0; factor *= scaleFactor )
1659        {
1660            const double ystep = std::max( 2., factor );
1661            CvSize winSize(cvRound( cascade->orig_window_size.width * factor ),
1662                                cvRound( cascade->orig_window_size.height * factor ));
1663            CvRect equRect;
1664            int *p[4] = {0,0,0,0};
1665            int *pq[4] = {0,0,0,0};
1666            int startX = 0, startY = 0;
1667            int endX = cvRound((img->cols - winSize.width) / ystep);
1668            int endY = cvRound((img->rows - winSize.height) / ystep);
1669
1670            if( winSize.width < minSize.width || winSize.height < minSize.height )
1671            {
1672                if( findBiggestObject )
1673                    break;
1674                continue;
1675            }
1676
1677            if ( winSize.width > maxSize.width || winSize.height > maxSize.height )
1678            {
1679                if( !findBiggestObject )
1680                    break;
1681                continue;
1682            }
1683
1684            cvSetImagesForHaarClassifierCascade( cascade, sum, sqsum, tilted, factor );
1685            cvZero( temp );
1686
1687            if( doCannyPruning )
1688            {
1689                equRect.x = cvRound(winSize.width*0.15);
1690                equRect.y = cvRound(winSize.height*0.15);
1691                equRect.width = cvRound(winSize.width*0.7);
1692                equRect.height = cvRound(winSize.height*0.7);
1693
1694                p[0] = (int*)(sumcanny->data.ptr + equRect.y*sumcanny->step) + equRect.x;
1695                p[1] = (int*)(sumcanny->data.ptr + equRect.y*sumcanny->step)
1696                            + equRect.x + equRect.width;
1697                p[2] = (int*)(sumcanny->data.ptr + (equRect.y + equRect.height)*sumcanny->step) + equRect.x;
1698                p[3] = (int*)(sumcanny->data.ptr + (equRect.y + equRect.height)*sumcanny->step)
1699                            + equRect.x + equRect.width;
1700
1701                pq[0] = (int*)(sum->data.ptr + equRect.y*sum->step) + equRect.x;
1702                pq[1] = (int*)(sum->data.ptr + equRect.y*sum->step)
1703                            + equRect.x + equRect.width;
1704                pq[2] = (int*)(sum->data.ptr + (equRect.y + equRect.height)*sum->step) + equRect.x;
1705                pq[3] = (int*)(sum->data.ptr + (equRect.y + equRect.height)*sum->step)
1706                            + equRect.x + equRect.width;
1707            }
1708
1709            if( scanROI.area() > 0 )
1710            {
1711                //adjust start_height and stop_height
1712                startY = cvRound(scanROI.y / ystep);
1713                endY = cvRound((scanROI.y + scanROI.height - winSize.height) / ystep);
1714
1715                startX = cvRound(scanROI.x / ystep);
1716                endX = cvRound((scanROI.x + scanROI.width - winSize.width) / ystep);
1717            }
1718
1719            cv::parallel_for_(cv::Range(startY, endY),
1720                cv::HaarDetectObjects_ScaleCascade_Invoker(cascade, winSize, cv::Range(startX, endX),
1721                                                           ystep, sum->step, (const int**)p,
1722                                                           (const int**)pq, allCandidates, &mtx ));
1723
1724            if( findBiggestObject && !allCandidates.empty() && scanROI.area() == 0 )
1725            {
1726                rectList.resize(allCandidates.size());
1727                std::copy(allCandidates.begin(), allCandidates.end(), rectList.begin());
1728
1729                groupRectangles(rectList, std::max(minNeighbors, 1), GROUP_EPS);
1730
1731                if( !rectList.empty() )
1732                {
1733                    size_t i, sz = rectList.size();
1734                    cv::Rect maxRect;
1735
1736                    for( i = 0; i < sz; i++ )
1737                    {
1738                        if( rectList[i].area() > maxRect.area() )
1739                            maxRect = rectList[i];
1740                    }
1741
1742                    allCandidates.push_back(maxRect);
1743
1744                    scanROI = maxRect;
1745                    int dx = cvRound(maxRect.width*GROUP_EPS);
1746                    int dy = cvRound(maxRect.height*GROUP_EPS);
1747                    scanROI.x = std::max(scanROI.x - dx, 0);
1748                    scanROI.y = std::max(scanROI.y - dy, 0);
1749                    scanROI.width = std::min(scanROI.width + dx*2, img->cols-1-scanROI.x);
1750                    scanROI.height = std::min(scanROI.height + dy*2, img->rows-1-scanROI.y);
1751
1752                    double minScale = roughSearch ? 0.6 : 0.4;
1753                    minSize.width = cvRound(maxRect.width*minScale);
1754                    minSize.height = cvRound(maxRect.height*minScale);
1755                }
1756            }
1757        }
1758    }
1759
1760    rectList.resize(allCandidates.size());
1761    if(!allCandidates.empty())
1762        std::copy(allCandidates.begin(), allCandidates.end(), rectList.begin());
1763
1764    if( minNeighbors != 0 || findBiggestObject )
1765    {
1766        if( outputRejectLevels )
1767        {
1768            groupRectangles(rectList, rejectLevels, levelWeights, minNeighbors, GROUP_EPS );
1769        }
1770        else
1771        {
1772            groupRectangles(rectList, rweights, std::max(minNeighbors, 1), GROUP_EPS);
1773        }
1774    }
1775    else
1776        rweights.resize(rectList.size(),0);
1777
1778    if( findBiggestObject && rectList.size() )
1779    {
1780        CvAvgComp result_comp = {CvRect(),0};
1781
1782        for( size_t i = 0; i < rectList.size(); i++ )
1783        {
1784            cv::Rect r = rectList[i];
1785            if( r.area() > cv::Rect(result_comp.rect).area() )
1786            {
1787                result_comp.rect = r;
1788                result_comp.neighbors = rweights[i];
1789            }
1790        }
1791        cvSeqPush( result_seq, &result_comp );
1792    }
1793    else
1794    {
1795        for( size_t i = 0; i < rectList.size(); i++ )
1796        {
1797            CvAvgComp c;
1798            c.rect = rectList[i];
1799            c.neighbors = !rweights.empty() ? rweights[i] : 0;
1800            cvSeqPush( result_seq, &c );
1801        }
1802    }
1803
1804    return result_seq;
1805}
1806
1807CV_IMPL CvSeq*
1808cvHaarDetectObjects( const CvArr* _img,
1809                     CvHaarClassifierCascade* cascade, CvMemStorage* storage,
1810                     double scaleFactor,
1811                     int minNeighbors, int flags, CvSize minSize, CvSize maxSize )
1812{
1813    std::vector<int> fakeLevels;
1814    std::vector<double> fakeWeights;
1815    return cvHaarDetectObjectsForROC( _img, cascade, storage, fakeLevels, fakeWeights,
1816                                scaleFactor, minNeighbors, flags, minSize, maxSize, false );
1817
1818}
1819
1820
1821static CvHaarClassifierCascade*
1822icvLoadCascadeCART( const char** input_cascade, int n, CvSize orig_window_size )
1823{
1824    int i;
1825    CvHaarClassifierCascade* cascade = icvCreateHaarClassifierCascade(n);
1826    cascade->orig_window_size = orig_window_size;
1827
1828    for( i = 0; i < n; i++ )
1829    {
1830        int j, count, l;
1831        float threshold = 0;
1832        const char* stage = input_cascade[i];
1833        int dl = 0;
1834
1835        /* tree links */
1836        int parent = -1;
1837        int next = -1;
1838
1839        sscanf( stage, "%d%n", &count, &dl );
1840        stage += dl;
1841
1842        assert( count > 0 );
1843        cascade->stage_classifier[i].count = count;
1844        cascade->stage_classifier[i].classifier =
1845            (CvHaarClassifier*)cvAlloc( count*sizeof(cascade->stage_classifier[i].classifier[0]));
1846
1847        for( j = 0; j < count; j++ )
1848        {
1849            CvHaarClassifier* classifier = cascade->stage_classifier[i].classifier + j;
1850            int k, rects = 0;
1851            char str[100];
1852
1853            sscanf( stage, "%d%n", &classifier->count, &dl );
1854            stage += dl;
1855
1856            classifier->haar_feature = (CvHaarFeature*) cvAlloc(
1857                classifier->count * ( sizeof( *classifier->haar_feature ) +
1858                                      sizeof( *classifier->threshold ) +
1859                                      sizeof( *classifier->left ) +
1860                                      sizeof( *classifier->right ) ) +
1861                (classifier->count + 1) * sizeof( *classifier->alpha ) );
1862            classifier->threshold = (float*) (classifier->haar_feature+classifier->count);
1863            classifier->left = (int*) (classifier->threshold + classifier->count);
1864            classifier->right = (int*) (classifier->left + classifier->count);
1865            classifier->alpha = (float*) (classifier->right + classifier->count);
1866
1867            for( l = 0; l < classifier->count; l++ )
1868            {
1869                sscanf( stage, "%d%n", &rects, &dl );
1870                stage += dl;
1871
1872                assert( rects >= 2 && rects <= CV_HAAR_FEATURE_MAX );
1873
1874                for( k = 0; k < rects; k++ )
1875                {
1876                    CvRect r;
1877                    int band = 0;
1878                    sscanf( stage, "%d%d%d%d%d%f%n",
1879                            &r.x, &r.y, &r.width, &r.height, &band,
1880                            &(classifier->haar_feature[l].rect[k].weight), &dl );
1881                    stage += dl;
1882                    classifier->haar_feature[l].rect[k].r = r;
1883                }
1884                sscanf( stage, "%s%n", str, &dl );
1885                stage += dl;
1886
1887                classifier->haar_feature[l].tilted = strncmp( str, "tilted", 6 ) == 0;
1888
1889                for( k = rects; k < CV_HAAR_FEATURE_MAX; k++ )
1890                {
1891                    memset( classifier->haar_feature[l].rect + k, 0,
1892                            sizeof(classifier->haar_feature[l].rect[k]) );
1893                }
1894
1895                sscanf( stage, "%f%d%d%n", &(classifier->threshold[l]),
1896                                       &(classifier->left[l]),
1897                                       &(classifier->right[l]), &dl );
1898                stage += dl;
1899            }
1900            for( l = 0; l <= classifier->count; l++ )
1901            {
1902                sscanf( stage, "%f%n", &(classifier->alpha[l]), &dl );
1903                stage += dl;
1904            }
1905        }
1906
1907        sscanf( stage, "%f%n", &threshold, &dl );
1908        stage += dl;
1909
1910        cascade->stage_classifier[i].threshold = threshold;
1911
1912        /* load tree links */
1913        if( sscanf( stage, "%d%d%n", &parent, &next, &dl ) != 2 )
1914        {
1915            parent = i - 1;
1916            next = -1;
1917        }
1918        stage += dl;
1919
1920        cascade->stage_classifier[i].parent = parent;
1921        cascade->stage_classifier[i].next = next;
1922        cascade->stage_classifier[i].child = -1;
1923
1924        if( parent != -1 && cascade->stage_classifier[parent].child == -1 )
1925        {
1926            cascade->stage_classifier[parent].child = i;
1927        }
1928    }
1929
1930    return cascade;
1931}
1932
1933#ifndef _MAX_PATH
1934#define _MAX_PATH 1024
1935#endif
1936
1937CV_IMPL CvHaarClassifierCascade*
1938cvLoadHaarClassifierCascade( const char* directory, CvSize orig_window_size )
1939{
1940    if( !directory )
1941        CV_Error( CV_StsNullPtr, "Null path is passed" );
1942
1943    char name[_MAX_PATH];
1944
1945    int n = (int)strlen(directory)-1;
1946    const char* slash = directory[n] == '\\' || directory[n] == '/' ? "" : "/";
1947    int size = 0;
1948
1949    /* try to read the classifier from directory */
1950    for( n = 0; ; n++ )
1951    {
1952        sprintf( name, "%s%s%d/AdaBoostCARTHaarClassifier.txt", directory, slash, n );
1953        FILE* f = fopen( name, "rb" );
1954        if( !f )
1955            break;
1956        fseek( f, 0, SEEK_END );
1957        size += ftell( f ) + 1;
1958        fclose(f);
1959    }
1960
1961    if( n == 0 && slash[0] )
1962        return (CvHaarClassifierCascade*)cvLoad( directory );
1963
1964    if( n == 0 )
1965        CV_Error( CV_StsBadArg, "Invalid path" );
1966
1967    size += (n+1)*sizeof(char*);
1968    const char** input_cascade = (const char**)cvAlloc( size );
1969
1970    if( !input_cascade )
1971      CV_Error( CV_StsNoMem, "Could not allocate memory for input_cascade" );
1972
1973    char* ptr = (char*)(input_cascade + n + 1);
1974
1975    for( int i = 0; i < n; i++ )
1976    {
1977        sprintf( name, "%s/%d/AdaBoostCARTHaarClassifier.txt", directory, i );
1978        FILE* f = fopen( name, "rb" );
1979        if( !f )
1980            CV_Error( CV_StsError, "" );
1981        fseek( f, 0, SEEK_END );
1982        size = (int)ftell( f );
1983        fseek( f, 0, SEEK_SET );
1984        size_t elements_read = fread( ptr, 1, size, f );
1985        CV_Assert(elements_read == (size_t)(size));
1986        fclose(f);
1987        input_cascade[i] = ptr;
1988        ptr += size;
1989        *ptr++ = '\0';
1990    }
1991
1992    input_cascade[n] = 0;
1993
1994    CvHaarClassifierCascade* cascade = icvLoadCascadeCART( input_cascade, n, orig_window_size );
1995
1996    if( input_cascade )
1997        cvFree( &input_cascade );
1998
1999    return cascade;
2000}
2001
2002
2003CV_IMPL void
2004cvReleaseHaarClassifierCascade( CvHaarClassifierCascade** _cascade )
2005{
2006    if( _cascade && *_cascade )
2007    {
2008        int i, j;
2009        CvHaarClassifierCascade* cascade = *_cascade;
2010
2011        for( i = 0; i < cascade->count; i++ )
2012        {
2013            for( j = 0; j < cascade->stage_classifier[i].count; j++ )
2014                cvFree( &cascade->stage_classifier[i].classifier[j].haar_feature );
2015            cvFree( &cascade->stage_classifier[i].classifier );
2016        }
2017        icvReleaseHidHaarClassifierCascade( &cascade->hid_cascade );
2018        cvFree( _cascade );
2019    }
2020}
2021
2022
2023/****************************************************************************************\
2024*                                  Persistence functions                                 *
2025\****************************************************************************************/
2026
2027/* field names */
2028
2029#define ICV_HAAR_SIZE_NAME            "size"
2030#define ICV_HAAR_STAGES_NAME          "stages"
2031#define ICV_HAAR_TREES_NAME           "trees"
2032#define ICV_HAAR_FEATURE_NAME         "feature"
2033#define ICV_HAAR_RECTS_NAME           "rects"
2034#define ICV_HAAR_TILTED_NAME          "tilted"
2035#define ICV_HAAR_THRESHOLD_NAME       "threshold"
2036#define ICV_HAAR_LEFT_NODE_NAME       "left_node"
2037#define ICV_HAAR_LEFT_VAL_NAME        "left_val"
2038#define ICV_HAAR_RIGHT_NODE_NAME      "right_node"
2039#define ICV_HAAR_RIGHT_VAL_NAME       "right_val"
2040#define ICV_HAAR_STAGE_THRESHOLD_NAME "stage_threshold"
2041#define ICV_HAAR_PARENT_NAME          "parent"
2042#define ICV_HAAR_NEXT_NAME            "next"
2043
2044static int
2045icvIsHaarClassifier( const void* struct_ptr )
2046{
2047    return CV_IS_HAAR_CLASSIFIER( struct_ptr );
2048}
2049
2050static void*
2051icvReadHaarClassifier( CvFileStorage* fs, CvFileNode* node )
2052{
2053    CvHaarClassifierCascade* cascade = NULL;
2054
2055    char buf[256];
2056    CvFileNode* seq_fn = NULL; /* sequence */
2057    CvFileNode* fn = NULL;
2058    CvFileNode* stages_fn = NULL;
2059    CvSeqReader stages_reader;
2060    int n;
2061    int i, j, k, l;
2062    int parent, next;
2063
2064    stages_fn = cvGetFileNodeByName( fs, node, ICV_HAAR_STAGES_NAME );
2065    if( !stages_fn || !CV_NODE_IS_SEQ( stages_fn->tag) )
2066        CV_Error( CV_StsError, "Invalid stages node" );
2067
2068    n = stages_fn->data.seq->total;
2069    cascade = icvCreateHaarClassifierCascade(n);
2070
2071    /* read size */
2072    seq_fn = cvGetFileNodeByName( fs, node, ICV_HAAR_SIZE_NAME );
2073    if( !seq_fn || !CV_NODE_IS_SEQ( seq_fn->tag ) || seq_fn->data.seq->total != 2 )
2074        CV_Error( CV_StsError, "size node is not a valid sequence." );
2075    fn = (CvFileNode*) cvGetSeqElem( seq_fn->data.seq, 0 );
2076    if( !CV_NODE_IS_INT( fn->tag ) || fn->data.i <= 0 )
2077        CV_Error( CV_StsError, "Invalid size node: width must be positive integer" );
2078    cascade->orig_window_size.width = fn->data.i;
2079    fn = (CvFileNode*) cvGetSeqElem( seq_fn->data.seq, 1 );
2080    if( !CV_NODE_IS_INT( fn->tag ) || fn->data.i <= 0 )
2081        CV_Error( CV_StsError, "Invalid size node: height must be positive integer" );
2082    cascade->orig_window_size.height = fn->data.i;
2083
2084    cvStartReadSeq( stages_fn->data.seq, &stages_reader );
2085    for( i = 0; i < n; ++i )
2086    {
2087        CvFileNode* stage_fn;
2088        CvFileNode* trees_fn;
2089        CvSeqReader trees_reader;
2090
2091        stage_fn = (CvFileNode*) stages_reader.ptr;
2092        if( !CV_NODE_IS_MAP( stage_fn->tag ) )
2093        {
2094            sprintf( buf, "Invalid stage %d", i );
2095            CV_Error( CV_StsError, buf );
2096        }
2097
2098        trees_fn = cvGetFileNodeByName( fs, stage_fn, ICV_HAAR_TREES_NAME );
2099        if( !trees_fn || !CV_NODE_IS_SEQ( trees_fn->tag )
2100            || trees_fn->data.seq->total <= 0 )
2101        {
2102            sprintf( buf, "Trees node is not a valid sequence. (stage %d)", i );
2103            CV_Error( CV_StsError, buf );
2104        }
2105
2106        cascade->stage_classifier[i].classifier =
2107            (CvHaarClassifier*) cvAlloc( trees_fn->data.seq->total
2108                * sizeof( cascade->stage_classifier[i].classifier[0] ) );
2109        for( j = 0; j < trees_fn->data.seq->total; ++j )
2110        {
2111            cascade->stage_classifier[i].classifier[j].haar_feature = NULL;
2112        }
2113        cascade->stage_classifier[i].count = trees_fn->data.seq->total;
2114
2115        cvStartReadSeq( trees_fn->data.seq, &trees_reader );
2116        for( j = 0; j < trees_fn->data.seq->total; ++j )
2117        {
2118            CvFileNode* tree_fn;
2119            CvSeqReader tree_reader;
2120            CvHaarClassifier* classifier;
2121            int last_idx;
2122
2123            classifier = &cascade->stage_classifier[i].classifier[j];
2124            tree_fn = (CvFileNode*) trees_reader.ptr;
2125            if( !CV_NODE_IS_SEQ( tree_fn->tag ) || tree_fn->data.seq->total <= 0 )
2126            {
2127                sprintf( buf, "Tree node is not a valid sequence."
2128                         " (stage %d, tree %d)", i, j );
2129                CV_Error( CV_StsError, buf );
2130            }
2131
2132            classifier->count = tree_fn->data.seq->total;
2133            classifier->haar_feature = (CvHaarFeature*) cvAlloc(
2134                classifier->count * ( sizeof( *classifier->haar_feature ) +
2135                                      sizeof( *classifier->threshold ) +
2136                                      sizeof( *classifier->left ) +
2137                                      sizeof( *classifier->right ) ) +
2138                (classifier->count + 1) * sizeof( *classifier->alpha ) );
2139            classifier->threshold = (float*) (classifier->haar_feature+classifier->count);
2140            classifier->left = (int*) (classifier->threshold + classifier->count);
2141            classifier->right = (int*) (classifier->left + classifier->count);
2142            classifier->alpha = (float*) (classifier->right + classifier->count);
2143
2144            cvStartReadSeq( tree_fn->data.seq, &tree_reader );
2145            for( k = 0, last_idx = 0; k < tree_fn->data.seq->total; ++k )
2146            {
2147                CvFileNode* node_fn;
2148                CvFileNode* feature_fn;
2149                CvFileNode* rects_fn;
2150                CvSeqReader rects_reader;
2151
2152                node_fn = (CvFileNode*) tree_reader.ptr;
2153                if( !CV_NODE_IS_MAP( node_fn->tag ) )
2154                {
2155                    sprintf( buf, "Tree node %d is not a valid map. (stage %d, tree %d)",
2156                             k, i, j );
2157                    CV_Error( CV_StsError, buf );
2158                }
2159                feature_fn = cvGetFileNodeByName( fs, node_fn, ICV_HAAR_FEATURE_NAME );
2160                if( !feature_fn || !CV_NODE_IS_MAP( feature_fn->tag ) )
2161                {
2162                    sprintf( buf, "Feature node is not a valid map. "
2163                             "(stage %d, tree %d, node %d)", i, j, k );
2164                    CV_Error( CV_StsError, buf );
2165                }
2166                rects_fn = cvGetFileNodeByName( fs, feature_fn, ICV_HAAR_RECTS_NAME );
2167                if( !rects_fn || !CV_NODE_IS_SEQ( rects_fn->tag )
2168                    || rects_fn->data.seq->total < 1
2169                    || rects_fn->data.seq->total > CV_HAAR_FEATURE_MAX )
2170                {
2171                    sprintf( buf, "Rects node is not a valid sequence. "
2172                             "(stage %d, tree %d, node %d)", i, j, k );
2173                    CV_Error( CV_StsError, buf );
2174                }
2175                cvStartReadSeq( rects_fn->data.seq, &rects_reader );
2176                for( l = 0; l < rects_fn->data.seq->total; ++l )
2177                {
2178                    CvFileNode* rect_fn;
2179                    CvRect r;
2180
2181                    rect_fn = (CvFileNode*) rects_reader.ptr;
2182                    if( !CV_NODE_IS_SEQ( rect_fn->tag ) || rect_fn->data.seq->total != 5 )
2183                    {
2184                        sprintf( buf, "Rect %d is not a valid sequence. "
2185                                 "(stage %d, tree %d, node %d)", l, i, j, k );
2186                        CV_Error( CV_StsError, buf );
2187                    }
2188
2189                    fn = CV_SEQ_ELEM( rect_fn->data.seq, CvFileNode, 0 );
2190                    if( !CV_NODE_IS_INT( fn->tag ) || fn->data.i < 0 )
2191                    {
2192                        sprintf( buf, "x coordinate must be non-negative integer. "
2193                                 "(stage %d, tree %d, node %d, rect %d)", i, j, k, l );
2194                        CV_Error( CV_StsError, buf );
2195                    }
2196                    r.x = fn->data.i;
2197                    fn = CV_SEQ_ELEM( rect_fn->data.seq, CvFileNode, 1 );
2198                    if( !CV_NODE_IS_INT( fn->tag ) || fn->data.i < 0 )
2199                    {
2200                        sprintf( buf, "y coordinate must be non-negative integer. "
2201                                 "(stage %d, tree %d, node %d, rect %d)", i, j, k, l );
2202                        CV_Error( CV_StsError, buf );
2203                    }
2204                    r.y = fn->data.i;
2205                    fn = CV_SEQ_ELEM( rect_fn->data.seq, CvFileNode, 2 );
2206                    if( !CV_NODE_IS_INT( fn->tag ) || fn->data.i <= 0
2207                        || r.x + fn->data.i > cascade->orig_window_size.width )
2208                    {
2209                        sprintf( buf, "width must be positive integer and "
2210                                 "(x + width) must not exceed window width. "
2211                                 "(stage %d, tree %d, node %d, rect %d)", i, j, k, l );
2212                        CV_Error( CV_StsError, buf );
2213                    }
2214                    r.width = fn->data.i;
2215                    fn = CV_SEQ_ELEM( rect_fn->data.seq, CvFileNode, 3 );
2216                    if( !CV_NODE_IS_INT( fn->tag ) || fn->data.i <= 0
2217                        || r.y + fn->data.i > cascade->orig_window_size.height )
2218                    {
2219                        sprintf( buf, "height must be positive integer and "
2220                                 "(y + height) must not exceed window height. "
2221                                 "(stage %d, tree %d, node %d, rect %d)", i, j, k, l );
2222                        CV_Error( CV_StsError, buf );
2223                    }
2224                    r.height = fn->data.i;
2225                    fn = CV_SEQ_ELEM( rect_fn->data.seq, CvFileNode, 4 );
2226                    if( !CV_NODE_IS_REAL( fn->tag ) )
2227                    {
2228                        sprintf( buf, "weight must be real number. "
2229                                 "(stage %d, tree %d, node %d, rect %d)", i, j, k, l );
2230                        CV_Error( CV_StsError, buf );
2231                    }
2232
2233                    classifier->haar_feature[k].rect[l].weight = (float) fn->data.f;
2234                    classifier->haar_feature[k].rect[l].r = r;
2235
2236                    CV_NEXT_SEQ_ELEM( sizeof( *rect_fn ), rects_reader );
2237                } /* for each rect */
2238                for( l = rects_fn->data.seq->total; l < CV_HAAR_FEATURE_MAX; ++l )
2239                {
2240                    classifier->haar_feature[k].rect[l].weight = 0;
2241                    classifier->haar_feature[k].rect[l].r = cvRect( 0, 0, 0, 0 );
2242                }
2243
2244                fn = cvGetFileNodeByName( fs, feature_fn, ICV_HAAR_TILTED_NAME);
2245                if( !fn || !CV_NODE_IS_INT( fn->tag ) )
2246                {
2247                    sprintf( buf, "tilted must be 0 or 1. "
2248                             "(stage %d, tree %d, node %d)", i, j, k );
2249                    CV_Error( CV_StsError, buf );
2250                }
2251                classifier->haar_feature[k].tilted = ( fn->data.i != 0 );
2252                fn = cvGetFileNodeByName( fs, node_fn, ICV_HAAR_THRESHOLD_NAME);
2253                if( !fn || !CV_NODE_IS_REAL( fn->tag ) )
2254                {
2255                    sprintf( buf, "threshold must be real number. "
2256                             "(stage %d, tree %d, node %d)", i, j, k );
2257                    CV_Error( CV_StsError, buf );
2258                }
2259                classifier->threshold[k] = (float) fn->data.f;
2260                fn = cvGetFileNodeByName( fs, node_fn, ICV_HAAR_LEFT_NODE_NAME);
2261                if( fn )
2262                {
2263                    if( !CV_NODE_IS_INT( fn->tag ) || fn->data.i <= k
2264                        || fn->data.i >= tree_fn->data.seq->total )
2265                    {
2266                        sprintf( buf, "left node must be valid node number. "
2267                                 "(stage %d, tree %d, node %d)", i, j, k );
2268                        CV_Error( CV_StsError, buf );
2269                    }
2270                    /* left node */
2271                    classifier->left[k] = fn->data.i;
2272                }
2273                else
2274                {
2275                    fn = cvGetFileNodeByName( fs, node_fn, ICV_HAAR_LEFT_VAL_NAME );
2276                    if( !fn )
2277                    {
2278                        sprintf( buf, "left node or left value must be specified. "
2279                                 "(stage %d, tree %d, node %d)", i, j, k );
2280                        CV_Error( CV_StsError, buf );
2281                    }
2282                    if( !CV_NODE_IS_REAL( fn->tag ) )
2283                    {
2284                        sprintf( buf, "left value must be real number. "
2285                                 "(stage %d, tree %d, node %d)", i, j, k );
2286                        CV_Error( CV_StsError, buf );
2287                    }
2288                    /* left value */
2289                    if( last_idx >= classifier->count + 1 )
2290                    {
2291                        sprintf( buf, "Tree structure is broken: too many values. "
2292                                 "(stage %d, tree %d, node %d)", i, j, k );
2293                        CV_Error( CV_StsError, buf );
2294                    }
2295                    classifier->left[k] = -last_idx;
2296                    classifier->alpha[last_idx++] = (float) fn->data.f;
2297                }
2298                fn = cvGetFileNodeByName( fs, node_fn,ICV_HAAR_RIGHT_NODE_NAME);
2299                if( fn )
2300                {
2301                    if( !CV_NODE_IS_INT( fn->tag ) || fn->data.i <= k
2302                        || fn->data.i >= tree_fn->data.seq->total )
2303                    {
2304                        sprintf( buf, "right node must be valid node number. "
2305                                 "(stage %d, tree %d, node %d)", i, j, k );
2306                        CV_Error( CV_StsError, buf );
2307                    }
2308                    /* right node */
2309                    classifier->right[k] = fn->data.i;
2310                }
2311                else
2312                {
2313                    fn = cvGetFileNodeByName( fs, node_fn, ICV_HAAR_RIGHT_VAL_NAME );
2314                    if( !fn )
2315                    {
2316                        sprintf( buf, "right node or right value must be specified. "
2317                                 "(stage %d, tree %d, node %d)", i, j, k );
2318                        CV_Error( CV_StsError, buf );
2319                    }
2320                    if( !CV_NODE_IS_REAL( fn->tag ) )
2321                    {
2322                        sprintf( buf, "right value must be real number. "
2323                                 "(stage %d, tree %d, node %d)", i, j, k );
2324                        CV_Error( CV_StsError, buf );
2325                    }
2326                    /* right value */
2327                    if( last_idx >= classifier->count + 1 )
2328                    {
2329                        sprintf( buf, "Tree structure is broken: too many values. "
2330                                 "(stage %d, tree %d, node %d)", i, j, k );
2331                        CV_Error( CV_StsError, buf );
2332                    }
2333                    classifier->right[k] = -last_idx;
2334                    classifier->alpha[last_idx++] = (float) fn->data.f;
2335                }
2336
2337                CV_NEXT_SEQ_ELEM( sizeof( *node_fn ), tree_reader );
2338            } /* for each node */
2339            if( last_idx != classifier->count + 1 )
2340            {
2341                sprintf( buf, "Tree structure is broken: too few values. "
2342                         "(stage %d, tree %d)", i, j );
2343                CV_Error( CV_StsError, buf );
2344            }
2345
2346            CV_NEXT_SEQ_ELEM( sizeof( *tree_fn ), trees_reader );
2347        } /* for each tree */
2348
2349        fn = cvGetFileNodeByName( fs, stage_fn, ICV_HAAR_STAGE_THRESHOLD_NAME);
2350        if( !fn || !CV_NODE_IS_REAL( fn->tag ) )
2351        {
2352            sprintf( buf, "stage threshold must be real number. (stage %d)", i );
2353            CV_Error( CV_StsError, buf );
2354        }
2355        cascade->stage_classifier[i].threshold = (float) fn->data.f;
2356
2357        parent = i - 1;
2358        next = -1;
2359
2360        fn = cvGetFileNodeByName( fs, stage_fn, ICV_HAAR_PARENT_NAME );
2361        if( !fn || !CV_NODE_IS_INT( fn->tag )
2362            || fn->data.i < -1 || fn->data.i >= cascade->count )
2363        {
2364            sprintf( buf, "parent must be integer number. (stage %d)", i );
2365            CV_Error( CV_StsError, buf );
2366        }
2367        parent = fn->data.i;
2368        fn = cvGetFileNodeByName( fs, stage_fn, ICV_HAAR_NEXT_NAME );
2369        if( !fn || !CV_NODE_IS_INT( fn->tag )
2370            || fn->data.i < -1 || fn->data.i >= cascade->count )
2371        {
2372            sprintf( buf, "next must be integer number. (stage %d)", i );
2373            CV_Error( CV_StsError, buf );
2374        }
2375        next = fn->data.i;
2376
2377        cascade->stage_classifier[i].parent = parent;
2378        cascade->stage_classifier[i].next = next;
2379        cascade->stage_classifier[i].child = -1;
2380
2381        if( parent != -1 && cascade->stage_classifier[parent].child == -1 )
2382        {
2383            cascade->stage_classifier[parent].child = i;
2384        }
2385
2386        CV_NEXT_SEQ_ELEM( sizeof( *stage_fn ), stages_reader );
2387    } /* for each stage */
2388
2389    return cascade;
2390}
2391
2392static void
2393icvWriteHaarClassifier( CvFileStorage* fs, const char* name, const void* struct_ptr,
2394                        CvAttrList attributes )
2395{
2396    int i, j, k, l;
2397    char buf[256];
2398    const CvHaarClassifierCascade* cascade = (const CvHaarClassifierCascade*) struct_ptr;
2399
2400    /* TODO: parameters check */
2401
2402    cvStartWriteStruct( fs, name, CV_NODE_MAP, CV_TYPE_NAME_HAAR, attributes );
2403
2404    cvStartWriteStruct( fs, ICV_HAAR_SIZE_NAME, CV_NODE_SEQ | CV_NODE_FLOW );
2405    cvWriteInt( fs, NULL, cascade->orig_window_size.width );
2406    cvWriteInt( fs, NULL, cascade->orig_window_size.height );
2407    cvEndWriteStruct( fs ); /* size */
2408
2409    cvStartWriteStruct( fs, ICV_HAAR_STAGES_NAME, CV_NODE_SEQ );
2410    for( i = 0; i < cascade->count; ++i )
2411    {
2412        cvStartWriteStruct( fs, NULL, CV_NODE_MAP );
2413        sprintf( buf, "stage %d", i );
2414        cvWriteComment( fs, buf, 1 );
2415
2416        cvStartWriteStruct( fs, ICV_HAAR_TREES_NAME, CV_NODE_SEQ );
2417
2418        for( j = 0; j < cascade->stage_classifier[i].count; ++j )
2419        {
2420            CvHaarClassifier* tree = &cascade->stage_classifier[i].classifier[j];
2421
2422            cvStartWriteStruct( fs, NULL, CV_NODE_SEQ );
2423            sprintf( buf, "tree %d", j );
2424            cvWriteComment( fs, buf, 1 );
2425
2426            for( k = 0; k < tree->count; ++k )
2427            {
2428                CvHaarFeature* feature = &tree->haar_feature[k];
2429
2430                cvStartWriteStruct( fs, NULL, CV_NODE_MAP );
2431                if( k )
2432                {
2433                    sprintf( buf, "node %d", k );
2434                }
2435                else
2436                {
2437                    sprintf( buf, "root node" );
2438                }
2439                cvWriteComment( fs, buf, 1 );
2440
2441                cvStartWriteStruct( fs, ICV_HAAR_FEATURE_NAME, CV_NODE_MAP );
2442
2443                cvStartWriteStruct( fs, ICV_HAAR_RECTS_NAME, CV_NODE_SEQ );
2444                for( l = 0; l < CV_HAAR_FEATURE_MAX && feature->rect[l].r.width != 0; ++l )
2445                {
2446                    cvStartWriteStruct( fs, NULL, CV_NODE_SEQ | CV_NODE_FLOW );
2447                    cvWriteInt(  fs, NULL, feature->rect[l].r.x );
2448                    cvWriteInt(  fs, NULL, feature->rect[l].r.y );
2449                    cvWriteInt(  fs, NULL, feature->rect[l].r.width );
2450                    cvWriteInt(  fs, NULL, feature->rect[l].r.height );
2451                    cvWriteReal( fs, NULL, feature->rect[l].weight );
2452                    cvEndWriteStruct( fs ); /* rect */
2453                }
2454                cvEndWriteStruct( fs ); /* rects */
2455                cvWriteInt( fs, ICV_HAAR_TILTED_NAME, feature->tilted );
2456                cvEndWriteStruct( fs ); /* feature */
2457
2458                cvWriteReal( fs, ICV_HAAR_THRESHOLD_NAME, tree->threshold[k]);
2459
2460                if( tree->left[k] > 0 )
2461                {
2462                    cvWriteInt( fs, ICV_HAAR_LEFT_NODE_NAME, tree->left[k] );
2463                }
2464                else
2465                {
2466                    cvWriteReal( fs, ICV_HAAR_LEFT_VAL_NAME,
2467                        tree->alpha[-tree->left[k]] );
2468                }
2469
2470                if( tree->right[k] > 0 )
2471                {
2472                    cvWriteInt( fs, ICV_HAAR_RIGHT_NODE_NAME, tree->right[k] );
2473                }
2474                else
2475                {
2476                    cvWriteReal( fs, ICV_HAAR_RIGHT_VAL_NAME,
2477                        tree->alpha[-tree->right[k]] );
2478                }
2479
2480                cvEndWriteStruct( fs ); /* split */
2481            }
2482
2483            cvEndWriteStruct( fs ); /* tree */
2484        }
2485
2486        cvEndWriteStruct( fs ); /* trees */
2487
2488        cvWriteReal( fs, ICV_HAAR_STAGE_THRESHOLD_NAME, cascade->stage_classifier[i].threshold);
2489        cvWriteInt( fs, ICV_HAAR_PARENT_NAME, cascade->stage_classifier[i].parent );
2490        cvWriteInt( fs, ICV_HAAR_NEXT_NAME, cascade->stage_classifier[i].next );
2491
2492        cvEndWriteStruct( fs ); /* stage */
2493    } /* for each stage */
2494
2495    cvEndWriteStruct( fs ); /* stages */
2496    cvEndWriteStruct( fs ); /* root */
2497}
2498
2499static void*
2500icvCloneHaarClassifier( const void* struct_ptr )
2501{
2502    CvHaarClassifierCascade* cascade = NULL;
2503
2504    int i, j, k, n;
2505    const CvHaarClassifierCascade* cascade_src =
2506        (const CvHaarClassifierCascade*) struct_ptr;
2507
2508    n = cascade_src->count;
2509    cascade = icvCreateHaarClassifierCascade(n);
2510    cascade->orig_window_size = cascade_src->orig_window_size;
2511
2512    for( i = 0; i < n; ++i )
2513    {
2514        cascade->stage_classifier[i].parent = cascade_src->stage_classifier[i].parent;
2515        cascade->stage_classifier[i].next = cascade_src->stage_classifier[i].next;
2516        cascade->stage_classifier[i].child = cascade_src->stage_classifier[i].child;
2517        cascade->stage_classifier[i].threshold = cascade_src->stage_classifier[i].threshold;
2518
2519        cascade->stage_classifier[i].count = 0;
2520        cascade->stage_classifier[i].classifier =
2521            (CvHaarClassifier*) cvAlloc( cascade_src->stage_classifier[i].count
2522                * sizeof( cascade->stage_classifier[i].classifier[0] ) );
2523
2524        cascade->stage_classifier[i].count = cascade_src->stage_classifier[i].count;
2525
2526        for( j = 0; j < cascade->stage_classifier[i].count; ++j )
2527            cascade->stage_classifier[i].classifier[j].haar_feature = NULL;
2528
2529        for( j = 0; j < cascade->stage_classifier[i].count; ++j )
2530        {
2531            const CvHaarClassifier* classifier_src =
2532                &cascade_src->stage_classifier[i].classifier[j];
2533            CvHaarClassifier* classifier =
2534                &cascade->stage_classifier[i].classifier[j];
2535
2536            classifier->count = classifier_src->count;
2537            classifier->haar_feature = (CvHaarFeature*) cvAlloc(
2538                classifier->count * ( sizeof( *classifier->haar_feature ) +
2539                                      sizeof( *classifier->threshold ) +
2540                                      sizeof( *classifier->left ) +
2541                                      sizeof( *classifier->right ) ) +
2542                (classifier->count + 1) * sizeof( *classifier->alpha ) );
2543            classifier->threshold = (float*) (classifier->haar_feature+classifier->count);
2544            classifier->left = (int*) (classifier->threshold + classifier->count);
2545            classifier->right = (int*) (classifier->left + classifier->count);
2546            classifier->alpha = (float*) (classifier->right + classifier->count);
2547            for( k = 0; k < classifier->count; ++k )
2548            {
2549                classifier->haar_feature[k] = classifier_src->haar_feature[k];
2550                classifier->threshold[k] = classifier_src->threshold[k];
2551                classifier->left[k] = classifier_src->left[k];
2552                classifier->right[k] = classifier_src->right[k];
2553                classifier->alpha[k] = classifier_src->alpha[k];
2554            }
2555            classifier->alpha[classifier->count] =
2556                classifier_src->alpha[classifier->count];
2557        }
2558    }
2559
2560    return cascade;
2561}
2562
2563
2564CvType haar_type( CV_TYPE_NAME_HAAR, icvIsHaarClassifier,
2565                  (CvReleaseFunc)cvReleaseHaarClassifierCascade,
2566                  icvReadHaarClassifier, icvWriteHaarClassifier,
2567                  icvCloneHaarClassifier );
2568
2569/* End of file. */
2570