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.
10 // Intel License Agreement
11 // For Open Source Computer Vision Library
13 // Copyright (C) 2000, Intel Corporation, all rights reserved.
14 // Third party copyrights are property of their respective owners.
16 // Redistribution and use in source and binary forms, with or without modification,
17 // are permitted provided that the following conditions are met:
19 // * Redistribution's of source code must retain the above copyright notice,
20 // this list of conditions and the following disclaimer.
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.
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.
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.
48 // class for grouping object candidates, detected by Cascade Classifier, HOG etc.
49 // instance of the class is to be passed to cv::partition (see cxoperations.hpp)
50 class CV_EXPORTS SimilarRects
53 SimilarRects(double _eps) : eps(_eps) {}
54 inline bool operator()(const Rect& r1, const Rect& r2) const
56 double delta = eps*(std::min(r1.width, r2.width) + std::min(r1.height, r2.height))*0.5;
57 return std::abs(r1.x - r2.x) <= delta &&
58 std::abs(r1.y - r2.y) <= delta &&
59 std::abs(r1.x + r1.width - r2.x - r2.width) <= delta &&
60 std::abs(r1.y + r1.height - r2.y - r2.height) <= delta;
65 void groupRectangles(vector<Rect>& rectList, int groupThreshold, double eps)
67 if( groupThreshold <= 0 || rectList.empty() )
71 int nclasses = partition(rectList, labels, SimilarRects(eps));
72 vector<Rect> rrects(nclasses);
73 vector<int> rweights(nclasses, 0);
74 int i, nlabels = (int)labels.size();
75 for( i = 0; i < nlabels; i++ )
78 rrects[cls].x += rectList[i].x;
79 rrects[cls].y += rectList[i].y;
80 rrects[cls].width += rectList[i].width;
81 rrects[cls].height += rectList[i].height;
85 for( i = 0; i < nclasses; i++ )
88 if( rweights[i] <= groupThreshold )
90 float s = 1.f/rweights[i];
91 rectList.push_back(Rect(saturate_cast<int>(r.x*s),
92 saturate_cast<int>(r.y*s),
93 saturate_cast<int>(r.width*s),
94 saturate_cast<int>(r.height*s)));
98 #define CC_CASCADE_PARAMS "cascadeParams"
99 #define CC_STAGE_TYPE "stageType"
100 #define CC_FEATURE_TYPE "featureType"
101 #define CC_HEIGHT "height"
102 #define CC_WIDTH "width"
104 #define CC_STAGE_NUM "stageNum"
105 #define CC_STAGES "stages"
106 #define CC_STAGE_PARAMS "stageParams"
108 #define CC_BOOST "BOOST"
109 #define CC_MAX_DEPTH "maxDepth"
110 #define CC_WEAK_COUNT "maxWeakCount"
111 #define CC_STAGE_THRESHOLD "stageThreshold"
112 #define CC_WEAK_CLASSIFIERS "weakClassifiers"
113 #define CC_INTERNAL_NODES "internalNodes"
114 #define CC_LEAF_VALUES "leafValues"
116 #define CC_FEATURES "features"
117 #define CC_FEATURE_PARAMS "featureParams"
118 #define CC_MAX_CAT_COUNT "maxCatCount"
120 #define CC_HAAR "HAAR"
121 #define CC_RECTS "rects"
122 #define CC_TILTED "tilted"
125 #define CC_RECT "rect"
127 #define CV_SUM_PTRS( p0, p1, p2, p3, sum, rect, step ) \
129 (p0) = sum + (rect).x + (step) * (rect).y, \
131 (p1) = sum + (rect).x + (rect).width + (step) * (rect).y, \
133 (p2) = sum + (rect).x + (step) * ((rect).y + (rect).height), \
134 /* (x + w, y + h) */ \
135 (p3) = sum + (rect).x + (rect).width + (step) * ((rect).y + (rect).height)
137 #define CV_TILTED_PTRS( p0, p1, p2, p3, tilted, rect, step ) \
139 (p0) = tilted + (rect).x + (step) * (rect).y, \
140 /* (x - h, y + h) */ \
141 (p1) = tilted + (rect).x - (rect).height + (step) * ((rect).y + (rect).height), \
142 /* (x + w, y + w) */ \
143 (p2) = tilted + (rect).x + (rect).width + (step) * ((rect).y + (rect).width), \
144 /* (x + w - h, y + w + h) */ \
145 (p3) = tilted + (rect).x + (rect).width - (rect).height \
146 + (step) * ((rect).y + (rect).width + (rect).height)
148 #define CALC_SUM_(p0, p1, p2, p3, offset) \
149 ((p0)[offset] - (p1)[offset] - (p2)[offset] + (p3)[offset])
151 #define CALC_SUM(rect,offset) CALC_SUM_((rect)[0], (rect)[1], (rect)[2], (rect)[3], offset)
153 FeatureEvaluator::~FeatureEvaluator() {}
154 bool FeatureEvaluator::read(const FileNode&) {return true;}
155 Ptr<FeatureEvaluator> FeatureEvaluator::clone() const { return Ptr<FeatureEvaluator>(); }
156 int FeatureEvaluator::getFeatureType() const {return -1;}
157 bool FeatureEvaluator::setImage(const Mat&, Size) {return true;}
158 bool FeatureEvaluator::setWindow(Point) { return true; }
159 double FeatureEvaluator::calcOrd(int) const { return 0.; }
160 int FeatureEvaluator::calcCat(int) const { return 0; }
162 //---------------------------------------------- HaarEvaluator ---------------------------------------
163 class HaarEvaluator : public FeatureEvaluator
170 float calc( int offset ) const;
171 void updatePtrs( const Mat& sum );
172 bool read( const FileNode& node );
176 enum { RECT_NUM = 3 };
184 const int* p[RECT_NUM][4];
188 virtual ~HaarEvaluator();
190 virtual bool read( const FileNode& node );
191 virtual Ptr<FeatureEvaluator> clone() const;
192 virtual int getFeatureType() const { return FeatureEvaluator::HAAR; }
194 virtual bool setImage(const Mat&, Size origWinSize);
195 virtual bool setWindow(Point pt);
197 double operator()(int featureIdx) const
198 { return featuresPtr[featureIdx].calc(offset) * varianceNormFactor; }
199 virtual double calcOrd(int featureIdx) const
200 { return (*this)(featureIdx); }
203 Ptr<vector<Feature> > features;
204 Feature* featuresPtr; // optimization
205 bool hasTiltedFeatures;
207 Mat sum0, sqsum0, tilted0;
208 Mat sum, sqsum, tilted;
215 double varianceNormFactor;
218 inline HaarEvaluator::Feature :: Feature()
221 rect[0].r = rect[1].r = rect[2].r = Rect();
222 rect[0].weight = rect[1].weight = rect[2].weight = 0;
223 p[0][0] = p[0][1] = p[0][2] = p[0][3] =
224 p[1][0] = p[1][1] = p[1][2] = p[1][3] =
225 p[2][0] = p[2][1] = p[2][2] = p[2][3] = 0;
228 inline float HaarEvaluator::Feature :: calc( int offset ) const
230 float ret = rect[0].weight * CALC_SUM(p[0], offset) + rect[1].weight * CALC_SUM(p[1], offset);
232 if( rect[2].weight != 0.0f )
233 ret += rect[2].weight * CALC_SUM(p[2], offset);
238 inline void HaarEvaluator::Feature :: updatePtrs( const Mat& sum )
240 const int* ptr = (const int*)sum.data;
241 size_t step = sum.step/sizeof(ptr[0]);
244 CV_TILTED_PTRS( p[0][0], p[0][1], p[0][2], p[0][3], ptr, rect[0].r, step );
245 CV_TILTED_PTRS( p[1][0], p[1][1], p[1][2], p[1][3], ptr, rect[1].r, step );
247 CV_TILTED_PTRS( p[2][0], p[2][1], p[2][2], p[2][3], ptr, rect[2].r, step );
251 CV_SUM_PTRS( p[0][0], p[0][1], p[0][2], p[0][3], ptr, rect[0].r, step );
252 CV_SUM_PTRS( p[1][0], p[1][1], p[1][2], p[1][3], ptr, rect[1].r, step );
254 CV_SUM_PTRS( p[2][0], p[2][1], p[2][2], p[2][3], ptr, rect[2].r, step );
258 bool HaarEvaluator::Feature :: read( const FileNode& node )
260 FileNode rnode = node[CC_RECTS];
261 FileNodeIterator it = rnode.begin(), it_end = rnode.end();
264 for( ri = 0; ri < RECT_NUM; ri++ )
267 rect[ri].weight = 0.f;
270 for(ri = 0; it != it_end; ++it, ri++)
272 FileNodeIterator it2 = (*it).begin();
273 it2 >> rect[ri].r.x >> rect[ri].r.y >>
274 rect[ri].r.width >> rect[ri].r.height >> rect[ri].weight;
277 tilted = (int)node[CC_TILTED] != 0;
281 HaarEvaluator::HaarEvaluator()
283 features = new vector<Feature>();
285 HaarEvaluator::~HaarEvaluator()
289 bool HaarEvaluator::read(const FileNode& node)
291 features->resize(node.size());
292 featuresPtr = &(*features)[0];
293 FileNodeIterator it = node.begin(), it_end = node.end();
294 hasTiltedFeatures = false;
296 for(int i = 0; it != it_end; ++it, i++)
298 if(!featuresPtr[i].read(*it))
300 if( featuresPtr[i].tilted )
301 hasTiltedFeatures = true;
306 Ptr<FeatureEvaluator> HaarEvaluator::clone() const
308 HaarEvaluator* ret = new HaarEvaluator;
309 ret->origWinSize = origWinSize;
310 ret->features = features;
311 ret->featuresPtr = &(*ret->features)[0];
312 ret->hasTiltedFeatures = hasTiltedFeatures;
313 ret->sum0 = sum0, ret->sqsum0 = sqsum0, ret->tilted0 = tilted0;
314 ret->sum = sum, ret->sqsum = sqsum, ret->tilted = tilted;
315 ret->normrect = normrect;
316 memcpy( ret->p, p, 4*sizeof(p[0]) );
317 memcpy( ret->pq, pq, 4*sizeof(pq[0]) );
318 ret->offset = offset;
319 ret->varianceNormFactor = varianceNormFactor;
323 bool HaarEvaluator::setImage( const Mat &image, Size _origWinSize )
325 int rn = image.rows+1, cn = image.cols+1;
326 origWinSize = _origWinSize;
327 normrect = Rect(1, 1, origWinSize.width-2, origWinSize.height-2);
329 if (image.cols < origWinSize.width || image.rows < origWinSize.height)
332 if( sum0.rows < rn || sum0.cols < cn )
334 sum0.create(rn, cn, CV_32S);
335 sqsum0.create(rn, cn, CV_64F);
336 if (hasTiltedFeatures)
337 tilted0.create( rn, cn, CV_32S);
339 sum = Mat(rn, cn, CV_32S, sum0.data);
340 sqsum = Mat(rn, cn, CV_32S, sqsum0.data);
342 if( hasTiltedFeatures )
344 tilted = Mat(rn, cn, CV_32S, tilted0.data);
345 integral(image, sum, sqsum, tilted);
348 integral(image, sum, sqsum);
349 const int* sdata = (const int*)sum.data;
350 const double* sqdata = (const double*)sqsum.data;
351 size_t sumStep = sum.step/sizeof(sdata[0]);
352 size_t sqsumStep = sqsum.step/sizeof(sqdata[0]);
354 CV_SUM_PTRS( p[0], p[1], p[2], p[3], sdata, normrect, sumStep );
355 CV_SUM_PTRS( pq[0], pq[1], pq[2], pq[3], sqdata, normrect, sqsumStep );
357 size_t fi, nfeatures = features->size();
359 for( fi = 0; fi < nfeatures; fi++ )
360 featuresPtr[fi].updatePtrs( !featuresPtr[fi].tilted ? sum : tilted );
364 bool HaarEvaluator::setWindow( Point pt )
366 if( pt.x < 0 || pt.y < 0 ||
367 pt.x + origWinSize.width >= sum.cols-2 ||
368 pt.y + origWinSize.height >= sum.rows-2 )
371 size_t pOffset = pt.y * (sum.step/sizeof(int)) + pt.x;
372 size_t pqOffset = pt.y * (sqsum.step/sizeof(double)) + pt.x;
373 int valsum = CALC_SUM(p, pOffset);
374 double valsqsum = CALC_SUM(pq, pqOffset);
376 double nf = (double)normrect.area() * valsqsum - (double)valsum * valsum;
381 varianceNormFactor = 1./nf;
382 offset = (int)pOffset;
386 //---------------------------------------------- LBPEvaluator -------------------------------------
388 class LBPEvaluator : public FeatureEvaluator
394 Feature( int x, int y, int _block_w, int _block_h ) :
395 rect(x, y, _block_w, _block_h) {}
397 int calc( int offset ) const;
398 void updatePtrs( const Mat& sum );
399 bool read(const FileNode& node );
401 Rect rect; // weight and height for block
402 const int* p[16]; // fast
406 virtual ~LBPEvaluator();
408 virtual bool read( const FileNode& node );
409 virtual Ptr<FeatureEvaluator> clone() const;
410 virtual int getFeatureType() const { return FeatureEvaluator::LBP; }
412 virtual bool setImage(const Mat& image, Size _origWinSize);
413 virtual bool setWindow(Point pt);
415 int operator()(int featureIdx) const
416 { return featuresPtr[featureIdx].calc(offset); }
417 virtual int calcCat(int featureIdx) const
418 { return (*this)(featureIdx); }
421 Ptr<vector<Feature> > features;
422 Feature* featuresPtr; // optimization
430 inline LBPEvaluator::Feature :: Feature()
433 for( int i = 0; i < 16; i++ )
437 inline int LBPEvaluator::Feature :: calc( int offset ) const
439 int cval = CALC_SUM_( p[5], p[6], p[9], p[10], offset );
441 return (CALC_SUM_( p[0], p[1], p[4], p[5], offset ) >= cval ? 128 : 0) | // 0
442 (CALC_SUM_( p[1], p[2], p[5], p[6], offset ) >= cval ? 64 : 0) | // 1
443 (CALC_SUM_( p[2], p[3], p[6], p[7], offset ) >= cval ? 32 : 0) | // 2
444 (CALC_SUM_( p[6], p[7], p[10], p[11], offset ) >= cval ? 16 : 0) | // 5
445 (CALC_SUM_( p[10], p[11], p[14], p[15], offset ) >= cval ? 8 : 0)| // 8
446 (CALC_SUM_( p[9], p[10], p[13], p[14], offset ) >= cval ? 4 : 0)| // 7
447 (CALC_SUM_( p[8], p[9], p[12], p[13], offset ) >= cval ? 2 : 0)| // 6
448 (CALC_SUM_( p[4], p[5], p[8], p[9], offset ) >= cval ? 1 : 0);
451 inline void LBPEvaluator::Feature :: updatePtrs( const Mat& sum )
453 const int* ptr = (const int*)sum.data;
454 size_t step = sum.step/sizeof(ptr[0]);
456 CV_SUM_PTRS( p[0], p[1], p[4], p[5], ptr, tr, step );
457 tr.x += 2*rect.width;
458 CV_SUM_PTRS( p[2], p[3], p[6], p[7], ptr, tr, step );
459 tr.y += 2*rect.height;
460 CV_SUM_PTRS( p[10], p[11], p[14], p[15], ptr, tr, step );
461 tr.x -= 2*rect.width;
462 CV_SUM_PTRS( p[8], p[9], p[12], p[13], ptr, tr, step );
465 bool LBPEvaluator::Feature :: read(const FileNode& node )
467 FileNode rnode = node[CC_RECT];
468 FileNodeIterator it = rnode.begin();
469 it >> rect.x >> rect.y >> rect.width >> rect.height;
473 LBPEvaluator::LBPEvaluator()
475 features = new vector<Feature>();
477 LBPEvaluator::~LBPEvaluator()
481 bool LBPEvaluator::read( const FileNode& node )
483 features->resize(node.size());
484 featuresPtr = &(*features)[0];
485 FileNodeIterator it = node.begin(), it_end = node.end();
486 for(int i = 0; it != it_end; ++it, i++)
488 if(!featuresPtr[i].read(*it))
494 Ptr<FeatureEvaluator> LBPEvaluator::clone() const
496 LBPEvaluator* ret = new LBPEvaluator;
497 ret->origWinSize = origWinSize;
498 ret->features = features;
499 ret->featuresPtr = &(*ret->features)[0];
500 ret->sum0 = sum0, ret->sum = sum;
501 ret->normrect = normrect;
502 ret->offset = offset;
506 bool LBPEvaluator::setImage( const Mat& image, Size _origWinSize )
508 int rn = image.rows+1, cn = image.cols+1;
509 origWinSize = _origWinSize;
511 if( image.cols < origWinSize.width || image.rows < origWinSize.height )
514 if( sum0.rows < rn || sum0.cols < cn )
515 sum0.create(rn, cn, CV_32S);
516 sum = Mat(rn, cn, CV_32S, sum0.data);
517 integral(image, sum);
519 size_t fi, nfeatures = features->size();
521 for( fi = 0; fi < nfeatures; fi++ )
522 featuresPtr[fi].updatePtrs( sum );
526 bool LBPEvaluator::setWindow( Point pt )
528 if( pt.x < 0 || pt.y < 0 ||
529 pt.x + origWinSize.width >= sum.cols-2 ||
530 pt.y + origWinSize.height >= sum.rows-2 )
532 offset = pt.y * ((int)sum.step/sizeof(int)) + pt.x;
536 Ptr<FeatureEvaluator> FeatureEvaluator::create(int featureType)
538 return featureType == HAAR ? Ptr<FeatureEvaluator>(new HaarEvaluator) :
539 featureType == LBP ? Ptr<FeatureEvaluator>(new LBPEvaluator) : Ptr<FeatureEvaluator>();
542 //---------------------------------------- Classifier Cascade --------------------------------------------
544 CascadeClassifier::CascadeClassifier()
548 CascadeClassifier::CascadeClassifier(const string& filename)
551 CascadeClassifier::~CascadeClassifier()
555 bool CascadeClassifier::empty() const
557 return oldCascade.empty() && stages.empty();
560 bool CascadeClassifier::load(const string& filename)
562 oldCascade.release();
564 FileStorage fs(filename, FileStorage::READ);
568 if( read(fs.getFirstTopLevelNode()) )
573 oldCascade = Ptr<CvHaarClassifierCascade>((CvHaarClassifierCascade*)cvLoad(filename.c_str(), 0, 0, 0));
574 return !oldCascade.empty();
577 template<class FEval>
578 inline int predictOrdered( CascadeClassifier& cascade, Ptr<FeatureEvaluator> &_feval )
580 int si, nstages = (int)cascade.stages.size();
581 int nodeOfs = 0, leafOfs = 0;
582 FEval& feval = (FEval&)*_feval;
583 float* cascadeLeaves = &cascade.leaves[0];
584 CascadeClassifier::DTreeNode* cascadeNodes = &cascade.nodes[0];
585 CascadeClassifier::DTree* cascadeWeaks = &cascade.classifiers[0];
586 CascadeClassifier::Stage* cascadeStages = &cascade.stages[0];
588 for( si = 0; si < nstages; si++ )
590 CascadeClassifier::Stage& stage = cascadeStages[si];
591 int wi, ntrees = stage.ntrees;
594 for( wi = 0; wi < ntrees; wi++ )
596 CascadeClassifier::DTree& weak = cascadeWeaks[stage.first + wi];
597 int idx = 0, root = nodeOfs;
601 CascadeClassifier::DTreeNode& node = cascadeNodes[root + idx];
602 double val = feval(node.featureIdx);
603 idx = val < node.threshold ? node.left : node.right;
606 sum += cascadeLeaves[leafOfs - idx];
607 nodeOfs += weak.nodeCount;
608 leafOfs += weak.nodeCount + 1;
610 if( sum < stage.threshold )
616 template<class FEval>
617 inline int predictCategorical( CascadeClassifier& cascade, Ptr<FeatureEvaluator> &_feval )
619 int si, nstages = (int)cascade.stages.size();
620 int nodeOfs = 0, leafOfs = 0;
621 FEval& feval = (FEval&)*_feval;
622 size_t subsetSize = (cascade.ncategories + 31)/32;
623 int* cascadeSubsets = &cascade.subsets[0];
624 float* cascadeLeaves = &cascade.leaves[0];
625 CascadeClassifier::DTreeNode* cascadeNodes = &cascade.nodes[0];
626 CascadeClassifier::DTree* cascadeWeaks = &cascade.classifiers[0];
627 CascadeClassifier::Stage* cascadeStages = &cascade.stages[0];
629 for( si = 0; si < nstages; si++ )
631 CascadeClassifier::Stage& stage = cascadeStages[si];
632 int wi, ntrees = stage.ntrees;
635 for( wi = 0; wi < ntrees; wi++ )
637 CascadeClassifier::DTree& weak = cascadeWeaks[stage.first + wi];
638 int idx = 0, root = nodeOfs;
641 CascadeClassifier::DTreeNode& node = cascadeNodes[root + idx];
642 int c = feval(node.featureIdx);
643 const int* subset = &cascadeSubsets[(root + idx)*subsetSize];
644 idx = (subset[c>>5] & (1 << (c & 31))) ? node.left : node.right;
647 sum += cascadeLeaves[leafOfs - idx];
648 nodeOfs += weak.nodeCount;
649 leafOfs += weak.nodeCount + 1;
651 if( sum < stage.threshold )
657 template<class FEval>
658 inline int predictOrderedStump( CascadeClassifier& cascade, Ptr<FeatureEvaluator> &_feval )
660 int si, nstages = (int)cascade.stages.size();
661 int nodeOfs = 0, leafOfs = 0;
662 FEval& feval = (FEval&)*_feval;
663 float* cascadeLeaves = &cascade.leaves[0];
664 CascadeClassifier::DTreeNode* cascadeNodes = &cascade.nodes[0];
665 CascadeClassifier::Stage* cascadeStages = &cascade.stages[0];
666 for( si = 0; si < nstages; si++ )
668 CascadeClassifier::Stage& stage = cascadeStages[si];
669 int wi, ntrees = stage.ntrees;
671 for( wi = 0; wi < ntrees; wi++, nodeOfs++, leafOfs+= 2 )
673 CascadeClassifier::DTreeNode& node = cascadeNodes[nodeOfs];
674 double val = feval(node.featureIdx);
675 sum += cascadeLeaves[ val < node.threshold ? leafOfs : leafOfs+1 ];
677 if( sum < stage.threshold )
683 template<class FEval>
684 inline int predictCategoricalStump( CascadeClassifier& cascade, Ptr<FeatureEvaluator> &_feval )
686 int si, nstages = (int)cascade.stages.size();
687 int nodeOfs = 0, leafOfs = 0;
688 FEval& feval = (FEval&)*_feval;
689 size_t subsetSize = (cascade.ncategories + 31)/32;
690 int* cascadeSubsets = &cascade.subsets[0];
691 float* cascadeLeaves = &cascade.leaves[0];
692 CascadeClassifier::DTreeNode* cascadeNodes = &cascade.nodes[0];
693 CascadeClassifier::Stage* cascadeStages = &cascade.stages[0];
695 for( si = 0; si < nstages; si++ )
697 CascadeClassifier::Stage& stage = cascadeStages[si];
698 int wi, ntrees = stage.ntrees;
701 for( wi = 0; wi < ntrees; wi++ )
703 CascadeClassifier::DTreeNode& node = cascadeNodes[nodeOfs];
704 int c = feval(node.featureIdx);
705 const int* subset = &cascadeSubsets[nodeOfs*subsetSize];
706 sum += cascadeLeaves[ subset[c>>5] & (1 << (c & 31)) ? leafOfs : leafOfs+1];
710 if( sum < stage.threshold )
716 int CascadeClassifier::runAt( Ptr<FeatureEvaluator> &_feval, Point pt )
718 CV_Assert( oldCascade.empty() );
719 /*if( !oldCascade.empty() )
720 return cvRunHaarClassifierCascade(oldCascade, pt, 0);*/
722 assert(featureType == FeatureEvaluator::HAAR ||
723 featureType == FeatureEvaluator::LBP);
724 return !_feval->setWindow(pt) ? -1 :
725 is_stump_based ? ( featureType == FeatureEvaluator::HAAR ?
726 predictOrderedStump<HaarEvaluator>( *this, _feval ) :
727 predictCategoricalStump<LBPEvaluator>( *this, _feval ) ) :
728 ( featureType == FeatureEvaluator::HAAR ?
729 predictOrdered<HaarEvaluator>( *this, _feval ) :
730 predictCategorical<LBPEvaluator>( *this, _feval ) );
734 bool CascadeClassifier::setImage( Ptr<FeatureEvaluator> &_feval, const Mat& image )
736 /*if( !oldCascade.empty() )
738 Mat sum(image.rows+1, image.cols+1, CV_32S);
739 Mat tilted(image.rows+1, image.cols+1, CV_32S);
740 Mat sqsum(image.rows+1, image.cols+1, CV_64F);
741 integral(image, sum, sqsum, tilted);
742 CvMat _sum = sum, _sqsum = sqsum, _tilted = tilted;
743 cvSetImagesForHaarClassifierCascade( oldCascade, &_sum, &_sqsum, &_tilted, 1. );
746 return empty() ? false : _feval->setImage(image, origWinSize );
750 struct getRect { Rect operator ()(const CvAvgComp& e) const { return e.rect; } };
752 void CascadeClassifier::detectMultiScale( const Mat& image, vector<Rect>& objects,
753 double scaleFactor, int minNeighbors,
754 int flags, Size minSize )
756 CV_Assert( scaleFactor > 1 && image.depth() == CV_8U );
761 if( !oldCascade.empty() )
763 MemStorage storage(cvCreateMemStorage(0));
764 CvMat _image = image;
765 CvSeq* _objects = cvHaarDetectObjects( &_image, oldCascade, storage, scaleFactor,
766 minNeighbors, flags, minSize );
767 vector<CvAvgComp> vecAvgComp;
768 Seq<CvAvgComp>(_objects).copyTo(vecAvgComp);
769 objects.resize(vecAvgComp.size());
770 std::transform(vecAvgComp.begin(), vecAvgComp.end(), objects.begin(), getRect());
776 Mat img = image, imgbuf(image.rows+1, image.cols+1, CV_8U);
778 if( img.channels() > 1 )
781 cvtColor(img, temp, CV_BGR2GRAY);
785 int maxNumThreads = 1;
787 maxNumThreads = cv::getNumThreads();
789 vector<vector<Rect> > rects( maxNumThreads );
790 vector<Rect>* rectsPtr = &rects[0];
791 vector<Ptr<FeatureEvaluator> > fevals( maxNumThreads );
793 Ptr<FeatureEvaluator>* fevalsPtr = &fevals[0];
795 for( double factor = 1; ; factor *= scaleFactor )
797 int stripCount, stripSize;
798 Size winSize( cvRound(origWinSize.width*factor), cvRound(origWinSize.height*factor) );
799 Size sz( cvRound( img.cols/factor ), cvRound( img.rows/factor ) );
800 Size sz1( sz.width - origWinSize.width, sz.height - origWinSize.height );
802 if( sz1.width <= 0 || sz1.height <= 0 )
804 if( winSize.width < minSize.width || winSize.height < minSize.height )
807 int yStep = factor > 2. ? 1 : 2;
808 if( maxNumThreads > 1 )
810 stripCount = max(min(sz1.height/yStep, maxNumThreads*3), 1);
811 stripSize = (sz1.height + stripCount - 1)/stripCount;
812 stripSize = (stripSize/yStep)*yStep;
817 stripSize = sz1.height;
820 Mat img1( sz, CV_8U, imgbuf.data );
821 resize( img, img1, sz, 0, 0, CV_INTER_LINEAR );
822 if( !feval->setImage( img1, origWinSize ) )
824 for( int i = 1; i < maxNumThreads; i++ )
825 fevalsPtr[i] = feval->clone();
828 #pragma omp parallel for num_threads(maxNumThreads) schedule(dynamic)
830 for( int i = 0; i < stripCount; i++ )
832 int threadIdx = cv::getThreadNum();
833 int y1 = i*stripSize, y2 = (i+1)*stripSize;
834 if( i == stripCount - 1 || y2 > sz1.height )
836 Size ssz(sz1.width, y2 - y1);
838 for( int y = y1; y < y2; y += yStep )
839 for( int x = 0; x < ssz.width; x += yStep )
841 int r = runAt(fevalsPtr[threadIdx], Point(x,y));
843 rectsPtr[threadIdx].push_back(Rect(cvRound(x*factor), cvRound(y*factor),
844 winSize.width, winSize.height));
850 for( vector< vector<Rect> >::const_iterator it = rects.begin(); it != rects.end(); it++ )
851 objects.insert( objects.end(), it->begin(), it->end() );
852 groupRectangles( objects, minNeighbors, 0.2 );
856 bool CascadeClassifier::read(const FileNode& root)
859 string stageTypeStr = (string)root[CC_STAGE_TYPE];
860 if( stageTypeStr == CC_BOOST )
865 string featureTypeStr = (string)root[CC_FEATURE_TYPE];
866 if( featureTypeStr == CC_HAAR )
867 featureType = FeatureEvaluator::HAAR;
868 else if( featureTypeStr == CC_LBP )
869 featureType = FeatureEvaluator::LBP;
873 origWinSize.width = (int)root[CC_WIDTH];
874 origWinSize.height = (int)root[CC_HEIGHT];
875 CV_Assert( origWinSize.height > 0 && origWinSize.width > 0 );
877 is_stump_based = (int)(root[CC_STAGE_PARAMS][CC_MAX_DEPTH]) == 1 ? true : false;
879 // load feature params
880 FileNode fn = root[CC_FEATURE_PARAMS];
884 ncategories = fn[CC_MAX_CAT_COUNT];
885 int subsetSize = (ncategories + 31)/32,
886 nodeStep = 3 + ( ncategories>0 ? subsetSize : 1 );
889 fn = root[CC_STAGES];
893 stages.reserve(fn.size());
897 FileNodeIterator it = fn.begin(), it_end = fn.end();
899 for( int si = 0; it != it_end; si++, ++it )
903 stage.threshold = fns[CC_STAGE_THRESHOLD];
904 fns = fns[CC_WEAK_CLASSIFIERS];
907 stage.ntrees = (int)fns.size();
908 stage.first = (int)classifiers.size();
909 stages.push_back(stage);
910 classifiers.reserve(stages[si].first + stages[si].ntrees);
912 FileNodeIterator it1 = fns.begin(), it1_end = fns.end();
913 for( ; it1 != it1_end; ++it1 ) // weak trees
916 FileNode internalNodes = fnw[CC_INTERNAL_NODES];
917 FileNode leafValues = fnw[CC_LEAF_VALUES];
918 if( internalNodes.empty() || leafValues.empty() )
921 tree.nodeCount = (int)internalNodes.size()/nodeStep;
922 classifiers.push_back(tree);
924 nodes.reserve(nodes.size() + tree.nodeCount);
925 leaves.reserve(leaves.size() + leafValues.size());
927 subsets.reserve(subsets.size() + tree.nodeCount*subsetSize);
929 FileNodeIterator it2 = internalNodes.begin(), it2_end = internalNodes.end();
931 for( ; it2 != it2_end; ) // nodes
934 node.left = (int)*it2; ++it2;
935 node.right = (int)*it2; ++it2;
936 node.featureIdx = (int)*it2; ++it2;
939 for( int j = 0; j < subsetSize; j++, ++it2 )
940 subsets.push_back((int)*it2);
941 node.threshold = 0.f;
945 node.threshold = (float)*it2; ++it2;
947 nodes.push_back(node);
950 it2 = leafValues.begin(), it2_end = leafValues.end();
952 for( ; it2 != it2_end; ++it2 ) // leaves
953 leaves.push_back((float)*it2);
958 feval = FeatureEvaluator::create(featureType);
959 fn = root[CC_FEATURES];
963 return feval->read(fn);