1 /*M///////////////////////////////////////////////////////////////////////////////////////
3 // IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
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.
11 // For Open Source Computer Vision Library
13 // Copyright (C) 2010-2012, Institute Of Software Chinese Academy Of Science, all rights reserved.
14 // Copyright (C) 2010-2012, Advanced Micro Devices, Inc., all rights reserved.
15 // Third party copyrights are property of their respective owners.
18 // Niko Li, newlife20080214@gmail.com
19 // Wang Weiyan, wangweiyanster@gmail.com
20 // Jia Haipeng, jiahaipeng95@gmail.com
21 // Wu Xinglong, wxl370@126.com
22 // Wang Yao, bitwangyaoyao@gmail.com
23 // Sen Liu, swjtuls1987@126.com
25 // Redistribution and use in source and binary forms, with or without modification,
26 // are permitted provided that the following conditions are met:
28 // * Redistribution's of source code must retain the above copyright notice,
29 // this list of conditions and the following disclaimer.
31 // * Redistribution's in binary form must reproduce the above copyright notice,
32 // this list of conditions and the following disclaimer in the documentation
33 // and/or other materials provided with the distribution.
35 // * The name of the copyright holders may not be used to endorse or promote products
36 // derived from this software without specific prior written permission.
38 // This software is provided by the copyright holders and contributors "as is" and
39 // any express or implied warranties, including, but not limited to, the implied
40 // warranties of merchantability and fitness for a particular purpose are disclaimed.
41 // In no event shall the Intel Corporation or contributors be liable for any direct,
42 // indirect, incidental, special, exemplary, or consequential damages
43 // (including, but not limited to, procurement of substitute goods or services;
44 // loss of use, data, or profits; or business interruption) however caused
45 // and on any theory of liability, whether in contract, strict liability,
46 // or tort (including negligence or otherwise) arising in any way out of
47 // the use of this software, even if advised of the possibility of such damage.
51 #include "precomp.hpp"
52 #include "opencl_kernels.hpp"
55 using namespace cv::ocl;
57 /* these settings affect the quality of detection: change with care */
58 #define CV_ADJUST_FEATURES 1
59 #define CV_ADJUST_WEIGHTS 0
62 typedef double sqsumtype;
64 typedef struct CvHidHaarFeature
68 sumtype *p0, *p1, *p2, *p3;
71 rect[CV_HAAR_FEATURE_MAX];
76 typedef struct CvHidHaarTreeNode
78 CvHidHaarFeature feature;
86 typedef struct CvHidHaarClassifier
89 //CvHaarFeature* orig_feature;
90 CvHidHaarTreeNode *node;
96 typedef struct CvHidHaarStageClassifier
100 CvHidHaarClassifier *classifier;
103 struct CvHidHaarStageClassifier *next;
104 struct CvHidHaarStageClassifier *child;
105 struct CvHidHaarStageClassifier *parent;
107 CvHidHaarStageClassifier;
110 struct CvHidHaarClassifierCascade
114 int has_tilted_features;
116 double inv_window_area;
117 CvMat sum, sqsum, tilted;
118 CvHidHaarStageClassifier *stage_classifier;
119 sqsumtype *pq0, *pq1, *pq2, *pq3;
120 sumtype *p0, *p1, *p2, *p3;
127 int grpnumperline_totalgrp;
130 } detect_piramid_info;
132 #define _ALIGNED_ON(_ALIGNMENT) __declspec(align(_ALIGNMENT))
134 typedef _ALIGNED_ON(128) struct GpuHidHaarTreeNode
136 _ALIGNED_ON(64) int p[CV_HAAR_FEATURE_MAX][4];
137 float weight[CV_HAAR_FEATURE_MAX] ;
139 _ALIGNED_ON(16) float alpha[3] ;
140 _ALIGNED_ON(4) int left ;
141 _ALIGNED_ON(4) int right ;
146 typedef _ALIGNED_ON(32) struct GpuHidHaarClassifier
148 _ALIGNED_ON(4) int count;
149 _ALIGNED_ON(8) GpuHidHaarTreeNode *node ;
150 _ALIGNED_ON(8) float *alpha ;
152 GpuHidHaarClassifier;
155 typedef _ALIGNED_ON(64) struct GpuHidHaarStageClassifier
157 _ALIGNED_ON(4) int count ;
158 _ALIGNED_ON(4) float threshold ;
159 _ALIGNED_ON(4) int two_rects ;
160 _ALIGNED_ON(8) GpuHidHaarClassifier *classifier ;
161 _ALIGNED_ON(8) struct GpuHidHaarStageClassifier *next;
162 _ALIGNED_ON(8) struct GpuHidHaarStageClassifier *child ;
163 _ALIGNED_ON(8) struct GpuHidHaarStageClassifier *parent ;
165 GpuHidHaarStageClassifier;
168 typedef _ALIGNED_ON(64) struct GpuHidHaarClassifierCascade
170 _ALIGNED_ON(4) int count ;
171 _ALIGNED_ON(4) int is_stump_based ;
172 _ALIGNED_ON(4) int has_tilted_features ;
173 _ALIGNED_ON(4) int is_tree ;
174 _ALIGNED_ON(4) int pq0 ;
175 _ALIGNED_ON(4) int pq1 ;
176 _ALIGNED_ON(4) int pq2 ;
177 _ALIGNED_ON(4) int pq3 ;
178 _ALIGNED_ON(4) int p0 ;
179 _ALIGNED_ON(4) int p1 ;
180 _ALIGNED_ON(4) int p2 ;
181 _ALIGNED_ON(4) int p3 ;
182 _ALIGNED_ON(4) float inv_window_area ;
183 } GpuHidHaarClassifierCascade;
185 #define _ALIGNED_ON(_ALIGNMENT) __attribute__((aligned(_ALIGNMENT) ))
187 typedef struct _ALIGNED_ON(128) GpuHidHaarTreeNode
189 int p[CV_HAAR_FEATURE_MAX][4] _ALIGNED_ON(64);
190 float weight[CV_HAAR_FEATURE_MAX];// _ALIGNED_ON(16);
191 float threshold;// _ALIGNED_ON(4);
192 float alpha[3] _ALIGNED_ON(16);
193 int left _ALIGNED_ON(4);
194 int right _ALIGNED_ON(4);
198 typedef struct _ALIGNED_ON(32) GpuHidHaarClassifier
200 int count _ALIGNED_ON(4);
201 GpuHidHaarTreeNode *node _ALIGNED_ON(8);
202 float *alpha _ALIGNED_ON(8);
204 GpuHidHaarClassifier;
207 typedef struct _ALIGNED_ON(64) GpuHidHaarStageClassifier
209 int count _ALIGNED_ON(4);
210 float threshold _ALIGNED_ON(4);
211 int two_rects _ALIGNED_ON(4);
212 GpuHidHaarClassifier *classifier _ALIGNED_ON(8);
213 struct GpuHidHaarStageClassifier *next _ALIGNED_ON(8);
214 struct GpuHidHaarStageClassifier *child _ALIGNED_ON(8);
215 struct GpuHidHaarStageClassifier *parent _ALIGNED_ON(8);
217 GpuHidHaarStageClassifier;
220 typedef struct _ALIGNED_ON(64) GpuHidHaarClassifierCascade
222 int count _ALIGNED_ON(4);
223 int is_stump_based _ALIGNED_ON(4);
224 int has_tilted_features _ALIGNED_ON(4);
225 int is_tree _ALIGNED_ON(4);
226 int pq0 _ALIGNED_ON(4);
227 int pq1 _ALIGNED_ON(4);
228 int pq2 _ALIGNED_ON(4);
229 int pq3 _ALIGNED_ON(4);
230 int p0 _ALIGNED_ON(4);
231 int p1 _ALIGNED_ON(4);
232 int p2 _ALIGNED_ON(4);
233 int p3 _ALIGNED_ON(4);
234 float inv_window_area _ALIGNED_ON(4);
235 } GpuHidHaarClassifierCascade;
238 const int icv_object_win_border = 1;
239 const float icv_stage_threshold_bias = 0.0001f;
240 double globaltime = 0;
242 /* create more efficient internal representation of haar classifier cascade */
243 static GpuHidHaarClassifierCascade * gpuCreateHidHaarClassifierCascade( CvHaarClassifierCascade *cascade, int *size, int *totalclassifier)
245 GpuHidHaarClassifierCascade *out = 0;
249 int total_classifiers = 0;
253 GpuHidHaarStageClassifier *stage_classifier_ptr;
254 GpuHidHaarClassifier *haar_classifier_ptr;
255 GpuHidHaarTreeNode *haar_node_ptr;
257 CvSize orig_window_size;
258 int has_tilted_features = 0;
260 if( !CV_IS_HAAR_CLASSIFIER(cascade) )
261 CV_Error( !cascade ? CV_StsNullPtr : CV_StsBadArg, "Invalid classifier pointer" );
263 if( cascade->hid_cascade )
264 CV_Error( CV_StsError, "hid_cascade has been already created" );
266 if( !cascade->stage_classifier )
267 CV_Error( CV_StsNullPtr, "" );
269 if( cascade->count <= 0 )
270 CV_Error( CV_StsOutOfRange, "Negative number of cascade stages" );
272 orig_window_size = cascade->orig_window_size;
274 /* check input structure correctness and calculate total memory size needed for
275 internal representation of the classifier cascade */
276 for( i = 0; i < cascade->count; i++ )
278 CvHaarStageClassifier *stage_classifier = cascade->stage_classifier + i;
280 if( !stage_classifier->classifier ||
281 stage_classifier->count <= 0 )
283 sprintf( errorstr, "header of the stage classifier #%d is invalid "
284 "(has null pointers or non-positive classfier count)", i );
285 CV_Error( CV_StsError, errorstr );
288 total_classifiers += stage_classifier->count;
290 for( j = 0; j < stage_classifier->count; j++ )
292 CvHaarClassifier *classifier = stage_classifier->classifier + j;
294 total_nodes += classifier->count;
295 for( l = 0; l < classifier->count; l++ )
297 for( k = 0; k < CV_HAAR_FEATURE_MAX; k++ )
299 if( classifier->haar_feature[l].rect[k].r.width )
301 CvRect r = classifier->haar_feature[l].rect[k].r;
302 int tilted = classifier->haar_feature[l].tilted;
303 has_tilted_features |= tilted != 0;
304 if( r.width < 0 || r.height < 0 || r.y < 0 ||
305 r.x + r.width > orig_window_size.width
308 (r.x < 0 || r.y + r.height > orig_window_size.height))
310 (tilted && (r.x - r.height < 0 ||
311 r.y + r.width + r.height > orig_window_size.height)))
313 sprintf( errorstr, "rectangle #%d of the classifier #%d of "
314 "the stage classifier #%d is not inside "
315 "the reference (original) cascade window", k, j, i );
316 CV_Error( CV_StsNullPtr, errorstr );
324 // this is an upper boundary for the whole hidden cascade size
325 datasize = sizeof(GpuHidHaarClassifierCascade) +
326 sizeof(GpuHidHaarStageClassifier) * cascade->count +
327 sizeof(GpuHidHaarClassifier) * total_classifiers +
328 sizeof(GpuHidHaarTreeNode) * total_nodes;
330 *totalclassifier = total_classifiers;
332 out = (GpuHidHaarClassifierCascade *)cvAlloc( datasize );
333 memset( out, 0, sizeof(*out) );
336 out->count = cascade->count;
337 stage_classifier_ptr = (GpuHidHaarStageClassifier *)(out + 1);
338 haar_classifier_ptr = (GpuHidHaarClassifier *)(stage_classifier_ptr + cascade->count);
339 haar_node_ptr = (GpuHidHaarTreeNode *)(haar_classifier_ptr + total_classifiers);
341 out->is_stump_based = 1;
342 out->has_tilted_features = has_tilted_features;
345 /* initialize internal representation */
346 for( i = 0; i < cascade->count; i++ )
348 CvHaarStageClassifier *stage_classifier = cascade->stage_classifier + i;
349 GpuHidHaarStageClassifier *hid_stage_classifier = stage_classifier_ptr + i;
351 hid_stage_classifier->count = stage_classifier->count;
352 hid_stage_classifier->threshold = stage_classifier->threshold - icv_stage_threshold_bias;
353 hid_stage_classifier->classifier = haar_classifier_ptr;
354 hid_stage_classifier->two_rects = 1;
355 haar_classifier_ptr += stage_classifier->count;
357 for( j = 0; j < stage_classifier->count; j++ )
359 CvHaarClassifier *classifier = stage_classifier->classifier + j;
360 GpuHidHaarClassifier *hid_classifier = hid_stage_classifier->classifier + j;
361 int node_count = classifier->count;
363 float *alpha_ptr = &haar_node_ptr->alpha[0];
365 hid_classifier->count = node_count;
366 hid_classifier->node = haar_node_ptr;
367 hid_classifier->alpha = alpha_ptr;
369 for( l = 0; l < node_count; l++ )
371 GpuHidHaarTreeNode *node = hid_classifier->node + l;
372 CvHaarFeature *feature = classifier->haar_feature + l;
374 memset( node, -1, sizeof(*node) );
375 node->threshold = classifier->threshold[l];
376 node->left = classifier->left[l];
377 node->right = classifier->right[l];
379 if( fabs(feature->rect[2].weight) < DBL_EPSILON ||
380 feature->rect[2].r.width == 0 ||
381 feature->rect[2].r.height == 0 )
390 hid_stage_classifier->two_rects = 0;
392 memcpy( node->alpha, classifier->alpha, (node_count + 1)*sizeof(alpha_ptr[0]));
393 haar_node_ptr = haar_node_ptr + 1;
395 out->is_stump_based &= node_count == 1;
399 cascade->hid_cascade = (CvHidHaarClassifierCascade *)out;
400 assert( (char *)haar_node_ptr - (char *)out <= datasize );
406 #define sum_elem_ptr(sum,row,col) \
407 ((sumtype*)CV_MAT_ELEM_PTR_FAST((sum),(row),(col),sizeof(sumtype)))
409 #define sqsum_elem_ptr(sqsum,row,col) \
410 ((sqsumtype*)CV_MAT_ELEM_PTR_FAST((sqsum),(row),(col),sizeof(sqsumtype)))
412 #define calc_sum(rect,offset) \
413 ((rect).p0[offset] - (rect).p1[offset] - (rect).p2[offset] + (rect).p3[offset])
416 static void gpuSetImagesForHaarClassifierCascade( CvHaarClassifierCascade *_cascade,
420 GpuHidHaarClassifierCascade *cascade;
421 int coi0 = 0, coi1 = 0;
427 GpuHidHaarStageClassifier *stage_classifier;
429 if( !CV_IS_HAAR_CLASSIFIER(_cascade) )
430 CV_Error( !_cascade ? CV_StsNullPtr : CV_StsBadArg, "Invalid classifier pointer" );
433 CV_Error( CV_StsOutOfRange, "Scale must be positive" );
436 CV_Error( CV_BadCOI, "COI is not supported" );
438 if( !_cascade->hid_cascade )
439 gpuCreateHidHaarClassifierCascade(_cascade, &datasize, &total);
441 cascade = (GpuHidHaarClassifierCascade *) _cascade->hid_cascade;
442 stage_classifier = (GpuHidHaarStageClassifier *) (cascade + 1);
444 _cascade->scale = scale;
445 _cascade->real_window_size.width = cvRound( _cascade->orig_window_size.width * scale );
446 _cascade->real_window_size.height = cvRound( _cascade->orig_window_size.height * scale );
448 equRect.x = equRect.y = cvRound(scale);
449 equRect.width = cvRound((_cascade->orig_window_size.width - 2) * scale);
450 equRect.height = cvRound((_cascade->orig_window_size.height - 2) * scale);
451 weight_scale = 1. / (equRect.width * equRect.height);
452 cascade->inv_window_area = weight_scale;
454 cascade->pq0 = equRect.x;
455 cascade->pq1 = equRect.y;
456 cascade->pq2 = equRect.x + equRect.width;
457 cascade->pq3 = equRect.y + equRect.height;
459 cascade->p0 = equRect.x;
460 cascade->p1 = equRect.y;
461 cascade->p2 = equRect.x + equRect.width;
462 cascade->p3 = equRect.y + equRect.height;
465 /* init pointers in haar features according to real window size and
466 given image pointers */
467 for( i = 0; i < _cascade->count; i++ )
470 for( j = 0; j < stage_classifier[i].count; j++ )
472 for( l = 0; l < stage_classifier[i].classifier[j].count; l++ )
474 CvHaarFeature *feature =
475 &_cascade->stage_classifier[i].classifier[j].haar_feature[l];
476 GpuHidHaarTreeNode *hidnode = &stage_classifier[i].classifier[j].node[l];
477 double sum0 = 0, area0 = 0;
480 int base_w = -1, base_h = -1;
481 int new_base_w = 0, new_base_h = 0;
483 int flagx = 0, flagy = 0;
488 for( k = 0; k < CV_HAAR_FEATURE_MAX; k++ )
490 if(!hidnode->p[k][0])
492 r[k] = feature->rect[k].r;
493 base_w = (int)CV_IMIN( (unsigned)base_w, (unsigned)(r[k].width - 1) );
494 base_w = (int)CV_IMIN( (unsigned)base_w, (unsigned)(r[k].x - r[0].x - 1) );
495 base_h = (int)CV_IMIN( (unsigned)base_h, (unsigned)(r[k].height - 1) );
496 base_h = (int)CV_IMIN( (unsigned)base_h, (unsigned)(r[k].y - r[0].y - 1) );
504 kx = r[0].width / base_w;
507 ky = r[0].height / base_h;
512 new_base_w = cvRound( r[0].width * scale ) / kx;
513 x0 = cvRound( r[0].x * scale );
519 new_base_h = cvRound( r[0].height * scale ) / ky;
520 y0 = cvRound( r[0].y * scale );
523 for( k = 0; k < nr; k++ )
526 double correction_ratio;
530 tr.x = (r[k].x - r[0].x) * new_base_w / base_w + x0;
531 tr.width = r[k].width * new_base_w / base_w;
535 tr.x = cvRound( r[k].x * scale );
536 tr.width = cvRound( r[k].width * scale );
541 tr.y = (r[k].y - r[0].y) * new_base_h / base_h + y0;
542 tr.height = r[k].height * new_base_h / base_h;
546 tr.y = cvRound( r[k].y * scale );
547 tr.height = cvRound( r[k].height * scale );
550 #if CV_ADJUST_WEIGHTS
553 const float orig_feature_size = (float)(feature->rect[k].r.width) * feature->rect[k].r.height;
554 const float orig_norm_size = (float)(_cascade->orig_window_size.width) * (_cascade->orig_window_size.height);
555 const float feature_size = float(tr.width * tr.height);
556 //const float normSize = float(equRect.width*equRect.height);
557 float target_ratio = orig_feature_size / orig_norm_size;
558 //float isRatio = featureSize / normSize;
559 //correctionRatio = targetRatio / isRatio / normSize;
560 correction_ratio = target_ratio / feature_size;
564 correction_ratio = weight_scale * (!feature->tilted ? 1 : 0.5);
567 if( !feature->tilted )
569 hidnode->p[k][0] = tr.x;
570 hidnode->p[k][1] = tr.y;
571 hidnode->p[k][2] = tr.x + tr.width;
572 hidnode->p[k][3] = tr.y + tr.height;
576 hidnode->p[k][2] = (tr.y + tr.width) * step + tr.x + tr.width;
577 hidnode->p[k][3] = (tr.y + tr.width + tr.height) * step + tr.x + tr.width - tr.height;
578 hidnode->p[k][0] = tr.y * step + tr.x;
579 hidnode->p[k][1] = (tr.y + tr.height) * step + tr.x - tr.height;
581 hidnode->weight[k] = (float)(feature->rect[k].weight * correction_ratio);
583 area0 = tr.width * tr.height;
585 sum0 += hidnode->weight[k] * tr.width * tr.height;
587 hidnode->weight[0] = (float)(-sum0 / area0);
593 static void gpuSetHaarClassifierCascade( CvHaarClassifierCascade *_cascade)
595 GpuHidHaarClassifierCascade *cascade;
601 GpuHidHaarStageClassifier *stage_classifier;
603 if( !CV_IS_HAAR_CLASSIFIER(_cascade) )
604 CV_Error( !_cascade ? CV_StsNullPtr : CV_StsBadArg, "Invalid classifier pointer" );
606 if( !_cascade->hid_cascade )
607 gpuCreateHidHaarClassifierCascade(_cascade, &datasize, &total);
609 cascade = (GpuHidHaarClassifierCascade *) _cascade->hid_cascade;
610 stage_classifier = (GpuHidHaarStageClassifier *) cascade + 1;
612 _cascade->scale = 1.0;
613 _cascade->real_window_size.width = _cascade->orig_window_size.width ;
614 _cascade->real_window_size.height = _cascade->orig_window_size.height;
616 equRect.x = equRect.y = 1;
617 equRect.width = _cascade->orig_window_size.width - 2;
618 equRect.height = _cascade->orig_window_size.height - 2;
620 cascade->inv_window_area = weight_scale;
622 cascade->p0 = equRect.x;
623 cascade->p1 = equRect.y;
624 cascade->p2 = equRect.height;
625 cascade->p3 = equRect.width ;
626 for( i = 0; i < _cascade->count; i++ )
629 for( j = 0; j < stage_classifier[i].count; j++ )
631 for( l = 0; l < stage_classifier[i].classifier[j].count; l++ )
633 const CvHaarFeature *feature =
634 &_cascade->stage_classifier[i].classifier[j].haar_feature[l];
635 GpuHidHaarTreeNode *hidnode = &stage_classifier[i].classifier[j].node[l];
637 for( int k = 0; k < CV_HAAR_FEATURE_MAX; k++ )
639 const CvRect tr = feature->rect[k].r;
642 double correction_ratio = weight_scale * (!feature->tilted ? 1 : 0.5);
643 hidnode->p[k][0] = tr.x;
644 hidnode->p[k][1] = tr.y;
645 hidnode->p[k][2] = tr.width;
646 hidnode->p[k][3] = tr.height;
647 hidnode->weight[k] = (float)(feature->rect[k].weight * correction_ratio);
654 CvSeq *cv::ocl::OclCascadeClassifier::oclHaarDetectObjects( oclMat &gimg, CvMemStorage *storage, double scaleFactor,
655 int minNeighbors, int flags, CvSize minSize, CvSize maxSize)
657 CvHaarClassifierCascade *cascade = oldCascade;
659 const double GROUP_EPS = 0.2;
660 CvSeq *result_seq = 0;
662 cv::ConcurrentRectVector allCandidates;
663 std::vector<cv::Rect> rectList;
664 std::vector<int> rweights;
667 int totalclassifier=0;
669 GpuHidHaarClassifierCascade *gcascade;
670 GpuHidHaarStageClassifier *stage;
671 GpuHidHaarClassifier *classifier;
672 GpuHidHaarTreeNode *node;
677 bool findBiggestObject = (flags & CV_HAAR_FIND_BIGGEST_OBJECT) != 0;
679 if( maxSize.height == 0 || maxSize.width == 0 )
681 maxSize.height = gimg.rows;
682 maxSize.width = gimg.cols;
685 if( !CV_IS_HAAR_CLASSIFIER(cascade) )
686 CV_Error( !cascade ? CV_StsNullPtr : CV_StsBadArg, "Invalid classifier cascade" );
689 CV_Error( CV_StsNullPtr, "Null storage pointer" );
691 if( CV_MAT_DEPTH(gimg.type()) != CV_8U )
692 CV_Error( CV_StsUnsupportedFormat, "Only 8-bit images are supported" );
694 if( scaleFactor <= 1 )
695 CV_Error( CV_StsOutOfRange, "scale factor must be > 1" );
697 if( findBiggestObject )
698 flags &= ~CV_HAAR_SCALE_IMAGE;
700 if( !cascade->hid_cascade )
701 gpuCreateHidHaarClassifierCascade(cascade, &datasize, &totalclassifier);
703 result_seq = cvCreateSeq( 0, sizeof(CvSeq), sizeof(CvAvgComp), storage );
705 if( CV_MAT_CN(gimg.type()) > 1 )
708 cvtColor( gimg, gtemp, CV_BGR2GRAY );
712 if( findBiggestObject )
713 flags &= ~(CV_HAAR_SCALE_IMAGE | CV_HAAR_DO_CANNY_PRUNING);
715 if( gimg.cols < minSize.width || gimg.rows < minSize.height )
716 CV_Error(CV_StsError, "Image too small");
718 cl_command_queue qu = getClCommandQueue(Context::getContext());
719 if( (flags & CV_HAAR_SCALE_IMAGE) )
721 CvSize winSize0 = cascade->orig_window_size;
725 vector<CvSize> sizev;
726 vector<float> scalev;
727 for(factor = 1.f;; factor *= scaleFactor)
729 CvSize winSize = { cvRound(winSize0.width * factor), cvRound(winSize0.height * factor) };
730 sz.width = cvRound( gimg.cols / factor ) + 1;
731 sz.height = cvRound( gimg.rows / factor ) + 1;
732 CvSize sz1 = { sz.width - winSize0.width - 1, sz.height - winSize0.height - 1 };
734 if( sz1.width <= 0 || sz1.height <= 0 )
736 if( winSize.width > maxSize.width || winSize.height > maxSize.height )
738 if( winSize.width < minSize.width || winSize.height < minSize.height )
741 totalheight += sz.height;
743 scalev.push_back(factor);
746 oclMat gimg1(gimg.rows, gimg.cols, CV_8UC1);
747 oclMat gsum(totalheight + 4, gimg.cols + 1, CV_32SC1);
748 oclMat gsqsum(totalheight + 4, gimg.cols + 1, CV_32FC1);
752 cl_mem candidatebuffer;
753 cl_mem scaleinfobuffer;
755 cv::Mat imgroi, imgroisq;
756 cv::ocl::oclMat resizeroi, gimgroi, gimgroisq;
759 size_t blocksize = 8;
760 size_t localThreads[3] = { blocksize, blocksize , 1 };
761 size_t globalThreads[3] = { grp_per_CU *(gsum.clCxt->getDeviceInfo().maxComputeUnits) *localThreads[0],
764 int outputsz = 256 * globalThreads[0] / localThreads[0];
765 int loopcount = sizev.size();
766 detect_piramid_info *scaleinfo = (detect_piramid_info *)malloc(sizeof(detect_piramid_info) * loopcount);
768 for( int i = 0; i < loopcount; i++ )
772 roi = Rect(0, indexy, sz.width, sz.height);
773 roi2 = Rect(0, 0, sz.width - 1, sz.height - 1);
774 resizeroi = gimg1(roi2);
776 gimgroisq = gsqsum(roi);
777 int width = gimgroi.cols - 1 - cascade->orig_window_size.width;
778 int height = gimgroi.rows - 1 - cascade->orig_window_size.height;
779 scaleinfo[i].width_height = (width << 16) | height;
782 int grpnumperline = (width + localThreads[0] - 1) / localThreads[0];
783 int totalgrp = ((height + localThreads[1] - 1) / localThreads[1]) * grpnumperline;
785 scaleinfo[i].grpnumperline_totalgrp = (grpnumperline << 16) | totalgrp;
786 scaleinfo[i].imgoff = gimgroi.offset >> 2;
787 scaleinfo[i].factor = factor;
788 cv::ocl::resize(gimg, resizeroi, Size(sz.width - 1, sz.height - 1), 0, 0, INTER_LINEAR);
789 cv::ocl::integral(resizeroi, gimgroi, gimgroisq);
793 gcascade = (GpuHidHaarClassifierCascade *)cascade->hid_cascade;
794 stage = (GpuHidHaarStageClassifier *)(gcascade + 1);
795 classifier = (GpuHidHaarClassifier *)(stage + gcascade->count);
796 node = (GpuHidHaarTreeNode *)(classifier->node);
798 int nodenum = (datasize - sizeof(GpuHidHaarClassifierCascade) -
799 sizeof(GpuHidHaarStageClassifier) * gcascade->count - sizeof(GpuHidHaarClassifier) * totalclassifier) / sizeof(GpuHidHaarTreeNode);
801 candidate = (int *)malloc(4 * sizeof(int) * outputsz);
803 gpuSetImagesForHaarClassifierCascade( cascade, 1., gsum.step / 4 );
805 stagebuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY, sizeof(GpuHidHaarStageClassifier) * gcascade->count);
806 openCLSafeCall(clEnqueueWriteBuffer(qu, stagebuffer, 1, 0, sizeof(GpuHidHaarStageClassifier)*gcascade->count, stage, 0, NULL, NULL));
808 nodebuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY, nodenum * sizeof(GpuHidHaarTreeNode));
810 openCLSafeCall(clEnqueueWriteBuffer(qu, nodebuffer, 1, 0, nodenum * sizeof(GpuHidHaarTreeNode),
811 node, 0, NULL, NULL));
812 candidatebuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_WRITE_ONLY, 4 * sizeof(int) * outputsz);
814 scaleinfobuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY, sizeof(detect_piramid_info) * loopcount);
815 openCLSafeCall(clEnqueueWriteBuffer(qu, scaleinfobuffer, 1, 0, sizeof(detect_piramid_info)*loopcount, scaleinfo, 0, NULL, NULL));
818 int endstage = gcascade->count;
820 int pixelstep = gsum.step / 4;
822 int splitnode = stage[0].count + stage[1].count + stage[2].count;
824 p.s[0] = gcascade->p0;
825 p.s[1] = gcascade->p1;
826 p.s[2] = gcascade->p2;
827 p.s[3] = gcascade->p3;
828 pq.s[0] = gcascade->pq0;
829 pq.s[1] = gcascade->pq1;
830 pq.s[2] = gcascade->pq2;
831 pq.s[3] = gcascade->pq3;
832 float correction = gcascade->inv_window_area;
834 vector<pair<size_t, const void *> > args;
835 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&stagebuffer ));
836 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&scaleinfobuffer ));
837 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&nodebuffer ));
838 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&gsum.data ));
839 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&gsqsum.data ));
840 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&candidatebuffer ));
841 args.push_back ( make_pair(sizeof(cl_int) , (void *)&pixelstep ));
842 args.push_back ( make_pair(sizeof(cl_int) , (void *)&loopcount ));
843 args.push_back ( make_pair(sizeof(cl_int) , (void *)&startstage ));
844 args.push_back ( make_pair(sizeof(cl_int) , (void *)&splitstage ));
845 args.push_back ( make_pair(sizeof(cl_int) , (void *)&endstage ));
846 args.push_back ( make_pair(sizeof(cl_int) , (void *)&startnode ));
847 args.push_back ( make_pair(sizeof(cl_int) , (void *)&splitnode ));
848 args.push_back ( make_pair(sizeof(cl_int4) , (void *)&p ));
849 args.push_back ( make_pair(sizeof(cl_int4) , (void *)&pq ));
850 args.push_back ( make_pair(sizeof(cl_float) , (void *)&correction ));
852 const char * build_options = gcascade->is_stump_based ? "-D STUMP_BASED=1" : "-D STUMP_BASED=0";
854 openCLExecuteKernel(gsum.clCxt, &haarobjectdetect, "gpuRunHaarClassifierCascade", globalThreads, localThreads, args, -1, -1, build_options);
856 openCLReadBuffer( gsum.clCxt, candidatebuffer, candidate, 4 * sizeof(int)*outputsz );
858 for(int i = 0; i < outputsz; i++)
859 if(candidate[4 * i + 2] != 0)
860 allCandidates.push_back(Rect(candidate[4 * i], candidate[4 * i + 1],
861 candidate[4 * i + 2], candidate[4 * i + 3]));
865 openCLSafeCall(clReleaseMemObject(stagebuffer));
866 openCLSafeCall(clReleaseMemObject(scaleinfobuffer));
867 openCLSafeCall(clReleaseMemObject(nodebuffer));
868 openCLSafeCall(clReleaseMemObject(candidatebuffer));
873 CvSize winsize0 = cascade->orig_window_size;
877 cv::ocl::integral(gimg, gsum, gsqsum);
879 vector<CvSize> sizev;
880 vector<float> scalev;
881 gpuSetHaarClassifierCascade(cascade);
882 gcascade = (GpuHidHaarClassifierCascade *)cascade->hid_cascade;
883 stage = (GpuHidHaarStageClassifier *)(gcascade + 1);
884 classifier = (GpuHidHaarClassifier *)(stage + gcascade->count);
885 node = (GpuHidHaarTreeNode *)(classifier->node);
888 cl_mem candidatebuffer;
889 cl_mem scaleinfobuffer;
891 cl_mem correctionbuffer;
892 for( n_factors = 0, factor = 1;
893 cvRound(factor * winsize0.width) < gimg.cols - 10 &&
894 cvRound(factor * winsize0.height) < gimg.rows - 10;
895 n_factors++, factor *= scaleFactor )
897 CvSize winSize = { cvRound( winsize0.width * factor ),
898 cvRound( winsize0.height * factor )
900 if( winSize.width < minSize.width || winSize.height < minSize.height )
904 sizev.push_back(winSize);
905 scalev.push_back(factor);
907 int loopcount = scalev.size();
912 sizev.push_back(minSize);
913 scalev.push_back( std::min(cvRound(minSize.width / winsize0.width), cvRound(minSize.height / winsize0.height)) );
916 detect_piramid_info *scaleinfo = (detect_piramid_info *)malloc(sizeof(detect_piramid_info) * loopcount);
917 cl_int4 *p = (cl_int4 *)malloc(sizeof(cl_int4) * loopcount);
918 float *correction = (float *)malloc(sizeof(float) * loopcount);
920 size_t blocksize = 8;
921 size_t localThreads[3] = { blocksize, blocksize , 1 };
922 size_t globalThreads[3] = { grp_per_CU *gsum.clCxt->getDeviceInfo().maxComputeUnits *localThreads[0],
923 localThreads[1], 1 };
924 int outputsz = 256 * globalThreads[0] / localThreads[0];
925 int nodenum = (datasize - sizeof(GpuHidHaarClassifierCascade) -
926 sizeof(GpuHidHaarStageClassifier) * gcascade->count - sizeof(GpuHidHaarClassifier) * totalclassifier) / sizeof(GpuHidHaarTreeNode);
927 nodebuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY,
928 nodenum * sizeof(GpuHidHaarTreeNode));
929 openCLSafeCall(clEnqueueWriteBuffer(qu, nodebuffer, 1, 0,
930 nodenum * sizeof(GpuHidHaarTreeNode),
931 node, 0, NULL, NULL));
932 cl_mem newnodebuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_WRITE,
933 loopcount * nodenum * sizeof(GpuHidHaarTreeNode));
935 int endstage = gcascade->count;
936 for(int i = 0; i < loopcount; i++)
940 int ystep = cvRound(std::max(2., factor));
941 int equRect_x = (int)(factor * gcascade->p0 + 0.5);
942 int equRect_y = (int)(factor * gcascade->p1 + 0.5);
943 int equRect_w = (int)(factor * gcascade->p3 + 0.5);
944 int equRect_h = (int)(factor * gcascade->p2 + 0.5);
945 p[i].s[0] = equRect_x;
946 p[i].s[1] = equRect_y;
947 p[i].s[2] = equRect_x + equRect_w;
948 p[i].s[3] = equRect_y + equRect_h;
949 correction[i] = 1. / (equRect_w * equRect_h);
950 int width = (gsum.cols - 1 - sz.width + ystep - 1) / ystep;
951 int height = (gsum.rows - 1 - sz.height + ystep - 1) / ystep;
952 int grpnumperline = (width + localThreads[0] - 1) / localThreads[0];
953 int totalgrp = ((height + localThreads[1] - 1) / localThreads[1]) * grpnumperline;
955 scaleinfo[i].width_height = (width << 16) | height;
956 scaleinfo[i].grpnumperline_totalgrp = (grpnumperline << 16) | totalgrp;
957 scaleinfo[i].imgoff = 0;
958 scaleinfo[i].factor = factor;
959 int startnodenum = nodenum * i;
960 float factor2 = (float)factor;
962 vector<pair<size_t, const void *> > args1;
963 args1.push_back ( make_pair(sizeof(cl_mem) , (void *)&nodebuffer ));
964 args1.push_back ( make_pair(sizeof(cl_mem) , (void *)&newnodebuffer ));
965 args1.push_back ( make_pair(sizeof(cl_float) , (void *)&factor2 ));
966 args1.push_back ( make_pair(sizeof(cl_float) , (void *)&correction[i] ));
967 args1.push_back ( make_pair(sizeof(cl_int) , (void *)&startnodenum ));
969 size_t globalThreads2[3] = {nodenum, 1, 1};
970 openCLExecuteKernel(gsum.clCxt, &haarobjectdetect_scaled2, "gpuscaleclassifier", globalThreads2, NULL/*localThreads2*/, args1, -1, -1);
973 int step = gsum.step / 4;
976 stagebuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY, sizeof(GpuHidHaarStageClassifier) * gcascade->count);
977 openCLSafeCall(clEnqueueWriteBuffer(qu, stagebuffer, 1, 0, sizeof(GpuHidHaarStageClassifier)*gcascade->count, stage, 0, NULL, NULL));
978 candidatebuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_WRITE_ONLY | CL_MEM_ALLOC_HOST_PTR, 4 * sizeof(int) * outputsz);
979 scaleinfobuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY, sizeof(detect_piramid_info) * loopcount);
980 openCLSafeCall(clEnqueueWriteBuffer(qu, scaleinfobuffer, 1, 0, sizeof(detect_piramid_info)*loopcount, scaleinfo, 0, NULL, NULL));
981 pbuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY, sizeof(cl_int4) * loopcount);
982 openCLSafeCall(clEnqueueWriteBuffer(qu, pbuffer, 1, 0, sizeof(cl_int4)*loopcount, p, 0, NULL, NULL));
983 correctionbuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY, sizeof(cl_float) * loopcount);
984 openCLSafeCall(clEnqueueWriteBuffer(qu, correctionbuffer, 1, 0, sizeof(cl_float)*loopcount, correction, 0, NULL, NULL));
986 vector<pair<size_t, const void *> > args;
987 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&stagebuffer ));
988 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&scaleinfobuffer ));
989 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&newnodebuffer ));
990 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&gsum.data ));
991 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&gsqsum.data ));
992 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&candidatebuffer ));
993 args.push_back ( make_pair(sizeof(cl_int) , (void *)&gsum.rows ));
994 args.push_back ( make_pair(sizeof(cl_int) , (void *)&gsum.cols ));
995 args.push_back ( make_pair(sizeof(cl_int) , (void *)&step ));
996 args.push_back ( make_pair(sizeof(cl_int) , (void *)&loopcount ));
997 args.push_back ( make_pair(sizeof(cl_int) , (void *)&startstage ));
998 args.push_back ( make_pair(sizeof(cl_int) , (void *)&splitstage ));
999 args.push_back ( make_pair(sizeof(cl_int) , (void *)&endstage ));
1000 args.push_back ( make_pair(sizeof(cl_int) , (void *)&startnode ));
1001 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&pbuffer ));
1002 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&correctionbuffer ));
1003 args.push_back ( make_pair(sizeof(cl_int) , (void *)&nodenum ));
1004 const char * build_options = gcascade->is_stump_based ? "-D STUMP_BASED=1" : "-D STUMP_BASED=0";
1005 openCLExecuteKernel(gsum.clCxt, &haarobjectdetect_scaled2, "gpuRunHaarClassifierCascade_scaled2", globalThreads, localThreads, args, -1, -1, build_options);
1007 candidate = (int *)clEnqueueMapBuffer(qu, candidatebuffer, 1, CL_MAP_READ, 0, 4 * sizeof(int) * outputsz, 0, 0, 0, &status);
1009 for(int i = 0; i < outputsz; i++)
1011 if(candidate[4 * i + 2] != 0)
1012 allCandidates.push_back(Rect(candidate[4 * i], candidate[4 * i + 1], candidate[4 * i + 2], candidate[4 * i + 3]));
1018 clEnqueueUnmapMemObject(qu, candidatebuffer, candidate, 0, 0, 0);
1019 openCLSafeCall(clReleaseMemObject(stagebuffer));
1020 openCLSafeCall(clReleaseMemObject(scaleinfobuffer));
1021 openCLSafeCall(clReleaseMemObject(nodebuffer));
1022 openCLSafeCall(clReleaseMemObject(newnodebuffer));
1023 openCLSafeCall(clReleaseMemObject(candidatebuffer));
1024 openCLSafeCall(clReleaseMemObject(pbuffer));
1025 openCLSafeCall(clReleaseMemObject(correctionbuffer));
1028 cvFree(&cascade->hid_cascade);
1029 rectList.resize(allCandidates.size());
1030 if(!allCandidates.empty())
1031 std::copy(allCandidates.begin(), allCandidates.end(), rectList.begin());
1033 if( minNeighbors != 0 || findBiggestObject )
1034 groupRectangles(rectList, rweights, std::max(minNeighbors, 1), GROUP_EPS);
1036 rweights.resize(rectList.size(), 0);
1038 if( findBiggestObject && rectList.size() )
1040 CvAvgComp result_comp = {{0, 0, 0, 0}, 0};
1042 for( size_t i = 0; i < rectList.size(); i++ )
1044 cv::Rect r = rectList[i];
1045 if( r.area() > cv::Rect(result_comp.rect).area() )
1047 result_comp.rect = r;
1048 result_comp.neighbors = rweights[i];
1051 cvSeqPush( result_seq, &result_comp );
1055 for( size_t i = 0; i < rectList.size(); i++ )
1058 c.rect = rectList[i];
1059 c.neighbors = rweights[i];
1060 cvSeqPush( result_seq, &c );
1071 cl_mem candidatebuffer;
1072 cl_mem scaleinfobuffer;
1074 cl_mem correctionbuffer;
1075 cl_mem newnodebuffer;
1080 Rect operator()(const CvAvgComp &e) const
1086 void cv::ocl::OclCascadeClassifierBuf::detectMultiScale(oclMat &gimg, CV_OUT std::vector<cv::Rect>& faces,
1087 double scaleFactor, int minNeighbors, int flags,
1088 Size minSize, Size maxSize)
1091 int grp_per_CU = 12;
1092 size_t localThreads[3] = { blocksize, blocksize, 1 };
1093 size_t globalThreads[3] = { grp_per_CU * cv::ocl::Context::getContext()->getDeviceInfo().maxComputeUnits *localThreads[0],
1096 int outputsz = 256 * globalThreads[0] / localThreads[0];
1098 Init(gimg.rows, gimg.cols, scaleFactor, flags, outputsz, localThreads, minSize, maxSize);
1100 const double GROUP_EPS = 0.2;
1102 cv::ConcurrentRectVector allCandidates;
1103 std::vector<cv::Rect> rectList;
1104 std::vector<int> rweights;
1106 CvHaarClassifierCascade *cascade = oldCascade;
1107 GpuHidHaarClassifierCascade *gcascade;
1108 GpuHidHaarStageClassifier *stage;
1110 if( CV_MAT_DEPTH(gimg.type()) != CV_8U )
1111 CV_Error( CV_StsUnsupportedFormat, "Only 8-bit images are supported" );
1113 if( CV_MAT_CN(gimg.type()) > 1 )
1116 cvtColor( gimg, gtemp, CV_BGR2GRAY );
1121 cl_command_queue qu = getClCommandQueue(Context::getContext());
1122 if( (flags & CV_HAAR_SCALE_IMAGE) )
1128 cv::ocl::oclMat resizeroi, gimgroi, gimgroisq;
1130 for( int i = 0; i < m_loopcount; i++ )
1133 roi = Rect(0, indexy, sz.width, sz.height);
1134 roi2 = Rect(0, 0, sz.width - 1, sz.height - 1);
1135 resizeroi = gimg1(roi2);
1136 gimgroi = gsum(roi);
1137 gimgroisq = gsqsum(roi);
1139 cv::ocl::resize(gimg, resizeroi, Size(sz.width - 1, sz.height - 1), 0, 0, INTER_LINEAR);
1140 cv::ocl::integral(resizeroi, gimgroi, gimgroisq);
1141 indexy += sz.height;
1144 gcascade = (GpuHidHaarClassifierCascade *)(cascade->hid_cascade);
1145 stage = (GpuHidHaarStageClassifier *)(gcascade + 1);
1148 int endstage = gcascade->count;
1150 int pixelstep = gsum.step / 4;
1152 int splitnode = stage[0].count + stage[1].count + stage[2].count;
1154 p.s[0] = gcascade->p0;
1155 p.s[1] = gcascade->p1;
1156 p.s[2] = gcascade->p2;
1157 p.s[3] = gcascade->p3;
1158 pq.s[0] = gcascade->pq0;
1159 pq.s[1] = gcascade->pq1;
1160 pq.s[2] = gcascade->pq2;
1161 pq.s[3] = gcascade->pq3;
1162 float correction = gcascade->inv_window_area;
1164 vector<pair<size_t, const void *> > args;
1165 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->stagebuffer ));
1166 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->scaleinfobuffer ));
1167 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->nodebuffer ));
1168 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&gsum.data ));
1169 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&gsqsum.data ));
1170 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->candidatebuffer ));
1171 args.push_back ( make_pair(sizeof(cl_int) , (void *)&pixelstep ));
1172 args.push_back ( make_pair(sizeof(cl_int) , (void *)&m_loopcount ));
1173 args.push_back ( make_pair(sizeof(cl_int) , (void *)&startstage ));
1174 args.push_back ( make_pair(sizeof(cl_int) , (void *)&splitstage ));
1175 args.push_back ( make_pair(sizeof(cl_int) , (void *)&endstage ));
1176 args.push_back ( make_pair(sizeof(cl_int) , (void *)&startnode ));
1177 args.push_back ( make_pair(sizeof(cl_int) , (void *)&splitnode ));
1178 args.push_back ( make_pair(sizeof(cl_int4) , (void *)&p ));
1179 args.push_back ( make_pair(sizeof(cl_int4) , (void *)&pq ));
1180 args.push_back ( make_pair(sizeof(cl_float) , (void *)&correction ));
1182 const char * build_options = gcascade->is_stump_based ? "-D STUMP_BASED=1" : "-D STUMP_BASED=0";
1184 openCLExecuteKernel(gsum.clCxt, &haarobjectdetect, "gpuRunHaarClassifierCascade", globalThreads, localThreads, args, -1, -1, build_options);
1186 candidate = (int *)malloc(4 * sizeof(int) * outputsz);
1187 memset(candidate, 0, 4 * sizeof(int) * outputsz);
1189 openCLReadBuffer( gsum.clCxt, ((OclBuffers *)buffers)->candidatebuffer, candidate, 4 * sizeof(int)*outputsz );
1191 for(int i = 0; i < outputsz; i++)
1193 if(candidate[4 * i + 2] != 0)
1195 allCandidates.push_back(Rect(candidate[4 * i], candidate[4 * i + 1],
1196 candidate[4 * i + 2], candidate[4 * i + 3]));
1199 free((void *)candidate);
1204 cv::ocl::integral(gimg, gsum, gsqsum);
1206 gcascade = (GpuHidHaarClassifierCascade *)cascade->hid_cascade;
1208 int step = gsum.step / 4;
1213 int endstage = gcascade->count;
1215 vector<pair<size_t, const void *> > args;
1216 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->stagebuffer ));
1217 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->scaleinfobuffer ));
1218 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->newnodebuffer ));
1219 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&gsum.data ));
1220 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&gsqsum.data ));
1221 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->candidatebuffer ));
1222 args.push_back ( make_pair(sizeof(cl_int) , (void *)&gsum.rows ));
1223 args.push_back ( make_pair(sizeof(cl_int) , (void *)&gsum.cols ));
1224 args.push_back ( make_pair(sizeof(cl_int) , (void *)&step ));
1225 args.push_back ( make_pair(sizeof(cl_int) , (void *)&m_loopcount ));
1226 args.push_back ( make_pair(sizeof(cl_int) , (void *)&startstage ));
1227 args.push_back ( make_pair(sizeof(cl_int) , (void *)&splitstage ));
1228 args.push_back ( make_pair(sizeof(cl_int) , (void *)&endstage ));
1229 args.push_back ( make_pair(sizeof(cl_int) , (void *)&startnode ));
1230 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->pbuffer ));
1231 args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->correctionbuffer ));
1232 args.push_back ( make_pair(sizeof(cl_int) , (void *)&m_nodenum ));
1234 const char * build_options = gcascade->is_stump_based ? "-D STUMP_BASED=1" : "-D STUMP_BASED=0";
1235 openCLExecuteKernel(gsum.clCxt, &haarobjectdetect_scaled2, "gpuRunHaarClassifierCascade_scaled2", globalThreads, localThreads, args, -1, -1, build_options);
1237 candidate = (int *)clEnqueueMapBuffer(qu, ((OclBuffers *)buffers)->candidatebuffer, 1, CL_MAP_READ, 0, 4 * sizeof(int) * outputsz, 0, 0, 0, NULL);
1239 for(int i = 0; i < outputsz; i++)
1241 if(candidate[4 * i + 2] != 0)
1242 allCandidates.push_back(Rect(candidate[4 * i], candidate[4 * i + 1],
1243 candidate[4 * i + 2], candidate[4 * i + 3]));
1245 clEnqueueUnmapMemObject(qu, ((OclBuffers *)buffers)->candidatebuffer, candidate, 0, 0, 0);
1247 rectList.resize(allCandidates.size());
1248 if(!allCandidates.empty())
1249 std::copy(allCandidates.begin(), allCandidates.end(), rectList.begin());
1251 if( minNeighbors != 0 || findBiggestObject )
1252 groupRectangles(rectList, rweights, std::max(minNeighbors, 1), GROUP_EPS);
1254 rweights.resize(rectList.size(), 0);
1256 GenResult(faces, rectList, rweights);
1259 void cv::ocl::OclCascadeClassifierBuf::Init(const int rows, const int cols,
1260 double scaleFactor, int flags,
1261 const int outputsz, const size_t localThreads[],
1262 CvSize minSize, CvSize maxSize)
1266 return; // we only allow one time initialization
1268 CvHaarClassifierCascade *cascade = oldCascade;
1270 if( !CV_IS_HAAR_CLASSIFIER(cascade) )
1271 CV_Error( !cascade ? CV_StsNullPtr : CV_StsBadArg, "Invalid classifier cascade" );
1273 if( scaleFactor <= 1 )
1274 CV_Error( CV_StsOutOfRange, "scale factor must be > 1" );
1276 if( cols < minSize.width || rows < minSize.height )
1277 CV_Error(CV_StsError, "Image too small");
1280 int totalclassifier=0;
1282 if( !cascade->hid_cascade )
1284 gpuCreateHidHaarClassifierCascade(cascade, &datasize, &totalclassifier);
1287 if( maxSize.height == 0 || maxSize.width == 0 )
1289 maxSize.height = rows;
1290 maxSize.width = cols;
1293 findBiggestObject = (flags & CV_HAAR_FIND_BIGGEST_OBJECT) != 0;
1294 if( findBiggestObject )
1295 flags &= ~(CV_HAAR_SCALE_IMAGE | CV_HAAR_DO_CANNY_PRUNING);
1297 CreateBaseBufs(datasize, totalclassifier, flags, outputsz);
1298 CreateFactorRelatedBufs(rows, cols, flags, scaleFactor, localThreads, minSize, maxSize);
1300 m_scaleFactor = scaleFactor;
1304 m_minSize = minSize;
1305 m_maxSize = maxSize;
1308 GpuHidHaarClassifierCascade *gcascade;
1309 GpuHidHaarStageClassifier *stage;
1310 GpuHidHaarClassifier *classifier;
1311 GpuHidHaarTreeNode *node;
1312 cl_command_queue qu = getClCommandQueue(Context::getContext());
1313 if( (flags & CV_HAAR_SCALE_IMAGE) )
1315 gcascade = (GpuHidHaarClassifierCascade *)(cascade->hid_cascade);
1316 stage = (GpuHidHaarStageClassifier *)(gcascade + 1);
1317 classifier = (GpuHidHaarClassifier *)(stage + gcascade->count);
1318 node = (GpuHidHaarTreeNode *)(classifier->node);
1320 gpuSetImagesForHaarClassifierCascade( cascade, 1., gsum.step / 4 );
1322 openCLSafeCall(clEnqueueWriteBuffer(qu, ((OclBuffers *)buffers)->stagebuffer, 1, 0,
1323 sizeof(GpuHidHaarStageClassifier) * gcascade->count,
1324 stage, 0, NULL, NULL));
1326 openCLSafeCall(clEnqueueWriteBuffer(qu, ((OclBuffers *)buffers)->nodebuffer, 1, 0,
1327 m_nodenum * sizeof(GpuHidHaarTreeNode),
1328 node, 0, NULL, NULL));
1332 gpuSetHaarClassifierCascade(cascade);
1334 gcascade = (GpuHidHaarClassifierCascade *)cascade->hid_cascade;
1335 stage = (GpuHidHaarStageClassifier *)(gcascade + 1);
1336 classifier = (GpuHidHaarClassifier *)(stage + gcascade->count);
1337 node = (GpuHidHaarTreeNode *)(classifier->node);
1339 openCLSafeCall(clEnqueueWriteBuffer(qu, ((OclBuffers *)buffers)->nodebuffer, 1, 0,
1340 m_nodenum * sizeof(GpuHidHaarTreeNode),
1341 node, 0, NULL, NULL));
1343 cl_int4 *p = (cl_int4 *)malloc(sizeof(cl_int4) * m_loopcount);
1344 float *correction = (float *)malloc(sizeof(float) * m_loopcount);
1346 for(int i = 0; i < m_loopcount; i++)
1349 int equRect_x = (int)(factor * gcascade->p0 + 0.5);
1350 int equRect_y = (int)(factor * gcascade->p1 + 0.5);
1351 int equRect_w = (int)(factor * gcascade->p3 + 0.5);
1352 int equRect_h = (int)(factor * gcascade->p2 + 0.5);
1353 p[i].s[0] = equRect_x;
1354 p[i].s[1] = equRect_y;
1355 p[i].s[2] = equRect_x + equRect_w;
1356 p[i].s[3] = equRect_y + equRect_h;
1357 correction[i] = 1. / (equRect_w * equRect_h);
1358 int startnodenum = m_nodenum * i;
1359 float factor2 = (float)factor;
1361 vector<pair<size_t, const void *> > args1;
1362 args1.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->nodebuffer ));
1363 args1.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->newnodebuffer ));
1364 args1.push_back ( make_pair(sizeof(cl_float) , (void *)&factor2 ));
1365 args1.push_back ( make_pair(sizeof(cl_float) , (void *)&correction[i] ));
1366 args1.push_back ( make_pair(sizeof(cl_int) , (void *)&startnodenum ));
1368 size_t globalThreads2[3] = {m_nodenum, 1, 1};
1370 openCLExecuteKernel(Context::getContext(), &haarobjectdetect_scaled2, "gpuscaleclassifier", globalThreads2, NULL/*localThreads2*/, args1, -1, -1);
1372 openCLSafeCall(clEnqueueWriteBuffer(qu, ((OclBuffers *)buffers)->stagebuffer, 1, 0, sizeof(GpuHidHaarStageClassifier)*gcascade->count, stage, 0, NULL, NULL));
1373 openCLSafeCall(clEnqueueWriteBuffer(qu, ((OclBuffers *)buffers)->pbuffer, 1, 0, sizeof(cl_int4)*m_loopcount, p, 0, NULL, NULL));
1374 openCLSafeCall(clEnqueueWriteBuffer(qu, ((OclBuffers *)buffers)->correctionbuffer, 1, 0, sizeof(cl_float)*m_loopcount, correction, 0, NULL, NULL));
1382 void cv::ocl::OclCascadeClassifierBuf::CreateBaseBufs(const int datasize, const int totalclassifier,
1383 const int flags, const int outputsz)
1387 buffers = malloc(sizeof(OclBuffers));
1390 sizeof(GpuHidHaarStageClassifier) * ((GpuHidHaarClassifierCascade *)oldCascade->hid_cascade)->count;
1391 m_nodenum = (datasize - sizeof(GpuHidHaarClassifierCascade) - tempSize - sizeof(GpuHidHaarClassifier) * totalclassifier)
1392 / sizeof(GpuHidHaarTreeNode);
1394 ((OclBuffers *)buffers)->stagebuffer = openCLCreateBuffer(cv::ocl::Context::getContext(), CL_MEM_READ_ONLY, tempSize);
1395 ((OclBuffers *)buffers)->nodebuffer = openCLCreateBuffer(cv::ocl::Context::getContext(), CL_MEM_READ_ONLY, m_nodenum * sizeof(GpuHidHaarTreeNode));
1399 && ((m_flags & CV_HAAR_SCALE_IMAGE) ^ (flags & CV_HAAR_SCALE_IMAGE)))
1401 openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->candidatebuffer));
1404 if (flags & CV_HAAR_SCALE_IMAGE)
1406 ((OclBuffers *)buffers)->candidatebuffer = openCLCreateBuffer(cv::ocl::Context::getContext(),
1408 4 * sizeof(int) * outputsz);
1412 ((OclBuffers *)buffers)->candidatebuffer = openCLCreateBuffer(cv::ocl::Context::getContext(),
1413 CL_MEM_WRITE_ONLY | CL_MEM_ALLOC_HOST_PTR,
1414 4 * sizeof(int) * outputsz);
1418 void cv::ocl::OclCascadeClassifierBuf::CreateFactorRelatedBufs(
1419 const int rows, const int cols, const int flags,
1420 const double scaleFactor, const size_t localThreads[],
1421 CvSize minSize, CvSize maxSize)
1425 if ((m_flags & CV_HAAR_SCALE_IMAGE) && !(flags & CV_HAAR_SCALE_IMAGE))
1431 else if (!(m_flags & CV_HAAR_SCALE_IMAGE) && (flags & CV_HAAR_SCALE_IMAGE))
1433 openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->newnodebuffer));
1434 openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->correctionbuffer));
1435 openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->pbuffer));
1437 else if ((m_flags & CV_HAAR_SCALE_IMAGE) && (flags & CV_HAAR_SCALE_IMAGE))
1439 if (fabs(m_scaleFactor - scaleFactor) < 1e-6
1440 && (rows == m_rows && cols == m_cols)
1441 && (minSize.width == m_minSize.width)
1442 && (minSize.height == m_minSize.height)
1443 && (maxSize.width == m_maxSize.width)
1444 && (maxSize.height == m_maxSize.height))
1451 if (fabs(m_scaleFactor - scaleFactor) < 1e-6
1452 && (rows == m_rows && cols == m_cols)
1453 && (minSize.width == m_minSize.width)
1454 && (minSize.height == m_minSize.height)
1455 && (maxSize.width == m_maxSize.width)
1456 && (maxSize.height == m_maxSize.height))
1462 openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->newnodebuffer));
1463 openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->correctionbuffer));
1464 openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->pbuffer));
1471 int totalheight = 0;
1475 CvSize winSize0 = oldCascade->orig_window_size;
1476 detect_piramid_info *scaleinfo;
1477 cl_command_queue qu = getClCommandQueue(Context::getContext());
1478 if (flags & CV_HAAR_SCALE_IMAGE)
1480 for(factor = 1.f;; factor *= scaleFactor)
1482 CvSize winSize = { cvRound(winSize0.width * factor), cvRound(winSize0.height * factor) };
1483 sz.width = cvRound( cols / factor ) + 1;
1484 sz.height = cvRound( rows / factor ) + 1;
1485 CvSize sz1 = { sz.width - winSize0.width - 1, sz.height - winSize0.height - 1 };
1487 if( sz1.width <= 0 || sz1.height <= 0 )
1489 if( winSize.width > maxSize.width || winSize.height > maxSize.height )
1491 if( winSize.width < minSize.width || winSize.height < minSize.height )
1494 totalheight += sz.height;
1495 sizev.push_back(sz);
1496 scalev.push_back(static_cast<float>(factor));
1499 loopcount = sizev.size();
1500 gimg1.create(rows, cols, CV_8UC1);
1501 gsum.create(totalheight + 4, cols + 1, CV_32SC1);
1502 gsqsum.create(totalheight + 4, cols + 1, CV_32FC1);
1504 scaleinfo = (detect_piramid_info *)malloc(sizeof(detect_piramid_info) * loopcount);
1505 for( int i = 0; i < loopcount; i++ )
1508 roi = Rect(0, indexy, sz.width, sz.height);
1509 int width = sz.width - 1 - oldCascade->orig_window_size.width;
1510 int height = sz.height - 1 - oldCascade->orig_window_size.height;
1511 int grpnumperline = (width + localThreads[0] - 1) / localThreads[0];
1512 int totalgrp = ((height + localThreads[1] - 1) / localThreads[1]) * grpnumperline;
1514 ((detect_piramid_info *)scaleinfo)[i].width_height = (width << 16) | height;
1515 ((detect_piramid_info *)scaleinfo)[i].grpnumperline_totalgrp = (grpnumperline << 16) | totalgrp;
1516 ((detect_piramid_info *)scaleinfo)[i].imgoff = gsum(roi).offset >> 2;
1517 ((detect_piramid_info *)scaleinfo)[i].factor = scalev[i];
1519 indexy += sz.height;
1525 cvRound(factor * winSize0.width) < cols - 10 && cvRound(factor * winSize0.height) < rows - 10;
1526 factor *= scaleFactor)
1528 CvSize winSize = { cvRound( winSize0.width * factor ), cvRound( winSize0.height * factor ) };
1529 if( winSize.width < minSize.width || winSize.height < minSize.height )
1533 sizev.push_back(winSize);
1534 scalev.push_back(factor);
1537 loopcount = scalev.size();
1541 sizev.push_back(minSize);
1542 scalev.push_back( std::min(cvRound(minSize.width / winSize0.width), cvRound(minSize.height / winSize0.height)) );
1545 ((OclBuffers *)buffers)->pbuffer = openCLCreateBuffer(cv::ocl::Context::getContext(), CL_MEM_READ_ONLY,
1546 sizeof(cl_int4) * loopcount);
1547 ((OclBuffers *)buffers)->correctionbuffer = openCLCreateBuffer(cv::ocl::Context::getContext(), CL_MEM_READ_ONLY,
1548 sizeof(cl_float) * loopcount);
1549 ((OclBuffers *)buffers)->newnodebuffer = openCLCreateBuffer(cv::ocl::Context::getContext(), CL_MEM_READ_WRITE,
1550 loopcount * m_nodenum * sizeof(GpuHidHaarTreeNode));
1552 scaleinfo = (detect_piramid_info *)malloc(sizeof(detect_piramid_info) * loopcount);
1553 for( int i = 0; i < loopcount; i++ )
1557 int ystep = cvRound(std::max(2., factor));
1558 int width = (cols - 1 - sz.width + ystep - 1) / ystep;
1559 int height = (rows - 1 - sz.height + ystep - 1) / ystep;
1560 int grpnumperline = (width + localThreads[0] - 1) / localThreads[0];
1561 int totalgrp = ((height + localThreads[1] - 1) / localThreads[1]) * grpnumperline;
1563 ((detect_piramid_info *)scaleinfo)[i].width_height = (width << 16) | height;
1564 ((detect_piramid_info *)scaleinfo)[i].grpnumperline_totalgrp = (grpnumperline << 16) | totalgrp;
1565 ((detect_piramid_info *)scaleinfo)[i].imgoff = 0;
1566 ((detect_piramid_info *)scaleinfo)[i].factor = factor;
1570 if (loopcount != m_loopcount)
1574 openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->scaleinfobuffer));
1576 ((OclBuffers *)buffers)->scaleinfobuffer = openCLCreateBuffer(cv::ocl::Context::getContext(), CL_MEM_READ_ONLY, sizeof(detect_piramid_info) * loopcount);
1579 openCLSafeCall(clEnqueueWriteBuffer(qu, ((OclBuffers *)buffers)->scaleinfobuffer, 1, 0,
1580 sizeof(detect_piramid_info)*loopcount,
1581 scaleinfo, 0, NULL, NULL));
1584 m_loopcount = loopcount;
1587 void cv::ocl::OclCascadeClassifierBuf::GenResult(CV_OUT std::vector<cv::Rect>& faces,
1588 const std::vector<cv::Rect> &rectList,
1589 const std::vector<int> &rweights)
1591 MemStorage tempStorage(cvCreateMemStorage(0));
1592 CvSeq *result_seq = cvCreateSeq( 0, sizeof(CvSeq), sizeof(CvAvgComp), tempStorage );
1594 if( findBiggestObject && rectList.size() )
1596 CvAvgComp result_comp = {{0, 0, 0, 0}, 0};
1598 for( size_t i = 0; i < rectList.size(); i++ )
1600 cv::Rect r = rectList[i];
1601 if( r.area() > cv::Rect(result_comp.rect).area() )
1603 result_comp.rect = r;
1604 result_comp.neighbors = rweights[i];
1607 cvSeqPush( result_seq, &result_comp );
1611 for( size_t i = 0; i < rectList.size(); i++ )
1614 c.rect = rectList[i];
1615 c.neighbors = rweights[i];
1616 cvSeqPush( result_seq, &c );
1620 vector<CvAvgComp> vecAvgComp;
1621 Seq<CvAvgComp>(result_seq).copyTo(vecAvgComp);
1622 faces.resize(vecAvgComp.size());
1623 std::transform(vecAvgComp.begin(), vecAvgComp.end(), faces.begin(), getRect());
1626 void cv::ocl::OclCascadeClassifierBuf::release()
1630 openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->stagebuffer));
1631 openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->scaleinfobuffer));
1632 openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->nodebuffer));
1633 openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->candidatebuffer));
1635 if( (m_flags & CV_HAAR_SCALE_IMAGE) )
1637 cvFree(&oldCascade->hid_cascade);
1641 openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->newnodebuffer));
1642 openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->correctionbuffer));
1643 openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->pbuffer));
1648 initialized = false;
1653 #define _MAX_PATH 1024