fastMatch的c++实现(需要opencv支持)

#include "MatchTemp.h" #include <ctime> struct Min_Val { int index; float val; }; Min_Val index_min_val(vector<float> fv) { int index = -1; float minVal = FLT_MAX; for (size_t i=0;i<fv.size();++i) { if (minVal>fv[i]) { minVal = fv[i]; index = i; } } Min_Val mv; mv.index = index; mv.val = minVal; return mv; } inline Mat repmat(const Mat inM,int rows) { Mat outM = Mat(0,inM.cols,inM.type()); for ( int i=0;i<rows;++i) { outM.push_back(inM); } return outM; } CMatchTemp::CMatchTemp(void) { m_epsilon = 0.15f; m_delta = 0.25f; m_minScale = 0.5f; m_maxScale = 2.0f; } CMatchTemp::~CMatchTemp(void) { } //使行列为奇数,数值范围[0,1] Mat CMatchTemp::makeOdd(const Mat& img) { int rows = img.rows; int cols = img.cols; if (rows%2==0) { rows -= 1; } if (cols%2==0) { cols -= 1; } Mat roi = img(Rect(0,0,cols,rows)); Mat out ; roi.convertTo(out,CV_32FC1,1.0f/255); return out; } //生成约束边界和步长 void CMatchTemp::GenerateGrid(float sourceW,float sourceH,float _delta, float minTx,float maxTx, float minTy, float maxTy, float minR,float maxR,float minS,float maxS) { m_bounds.tx[0] = minTx; m_bounds.tx[1] = maxTx ; m_bounds.ty[0] = minTy; m_bounds.ty[1] = maxTy; m_bounds.r[0] = minR; m_bounds.r[1] = maxR; m_bounds.s[0] = minS; m_bounds.s[1] = maxS; // [1/c,c] m_steps.tx = _delta*sourceW/sqrt(2.0f); m_steps.ty = _delta*sourceH/sqrt(2.0f); m_steps.r = _delta*sqrt(2.0f); m_steps.s = _delta/sqrt(2.0f); } Mat CMatchTemp::match(Mat img,Mat temp) { float blur_sigma = 2.0; GaussianBlur(img,img,Size(0,0),blur_sigma,blur_sigma); GaussianBlur(temp,temp,Size(0,0),blur_sigma,blur_sigma); src_img = makeOdd(img); temp_img = makeOdd(temp); h1 = temp_img.rows; w1 = temp_img.cols; h2 = src_img.rows; w2 = src_img.cols; r1x = 0.5f*(w1-1); r1y = 0.5f*(h1-1); r2x = 0.5f*(w2-1); r2y = 0.5f*(h2-1); //搜索范围 float minTx = -(r2x-r1x*m_minScale); float maxTx = r2x-r1x*m_minScale; float minTy = -(r2y-r1y*m_minScale); float maxTy = r2y-r1y*m_minScale; float minR =(float) -CV_PI; float maxR = (float) CV_PI; //生成限制 GenerateGrid(h1,w1,m_delta,minTx,maxTx,minTy,maxTy,minR,maxR,m_minScale,m_maxScale); //寻找最佳变换 bestTransMat = FindBestTransformation(); return bestTransMat; } Mat CMatchTemp::FindBestTransformation() { static Mat src_cpy = src_img.clone(); Mat bestTransMat; RNG rng(time(NULL)); //生成 Theta(1/eps^2) 个随机样本点 long numPoints = cvRound(8/(m_epsilon*m_epsilon)); vector<float> xs(numPoints); vector<float> ys(numPoints); for (int i=0;i<numPoints;++i) { xs[i] = (float) rng.uniform(5,w1-5); ys[i] = (float) rng.uniform(5,h1-5); } //生成网络 Mat configs; int gridSize = CreateListOfConfigs(configs); if ( gridSize > 71000000) cout<<"more than 35 million configs!"<<endl; Mat goodConfigsStats = Mat(0,6,CV_32FC1); //主循环 float deltaFact = 1.511f; int level = 0; vector<float> bestDists; float newDelta = m_delta; while(1) { // 1.选出变换后的坐标落在图像区域中的变换坐标 Mat matrixConfigs_mex; vector<bool> insiders; GetAffineFromConfig(configs,matrixConfigs_mex,insiders); cout<<"gridSize: "<<configs.rows<<" "; GetInsideMat(configs,matrixConfigs_mex,insiders); // 2.评估这些变换 double tic = getTickCount(); vector<float> distances = EvaluateConfigs_mex(matrixConfigs_mex, xs, ys); cout<<"EvaluateConfigs_mex 耗时："<<(getTickCount()-tic)/getTickFrequency()<<"s"<<endl; Min_Val mini_ = index_min_val(distances); int ind = mini_.index; float bestDist = mini_.val; Mat bestConfig = configs.row(ind); CreateAffineTransformation(bestConfig,bestTransMat); // cout<<"best tran:"<<bestTransMat<<endl; bestDists.push_back(bestDist); Mat temp = drawMatch(src_cpy,temp_img,bestTransMat,255 - level * 60); imshow("re",temp); waitKey(1000); //if (exist('goodConfigs','var')) // goodConfigsStats = size(goodConfigs,1); //end cout<<" level---"<<level<<",best dist:"<<bestDist<<endl; //退出机制 if ( (bestDist < 0.009) || ((level > 0) && (bestDist < 0.019)) || (level >= 10) /* || ((level > 3) && (bestDist > mean(bestDists(level-3:level-1))*0.97)) */ /*|| ((level > 8) && (goodConfigsStats(end) > 20 ) )*/ ) { break; } // 3 . 更新变换网络 //3.1 挑选 Mat goodConfigs ; bool ishighPer = GetGoodConfigsByDistance(configs,bestDist,newDelta,distances,goodConfigs); //3.2 扩展 if (ishighPer && level==0) { float fact = 0.99f; //fprintf('##### RESTARTING!!! changing from delta: %.3f, to delta: %.3f\n', newDelta, newDelta*fact); newDelta = newDelta*fact; //level = 0; m_steps.tx *= fact; m_steps.ty *= fact; m_steps.r *= fact; m_steps.s *=fact; CreateListOfConfigs(configs); } else { //float prevDelta = newDelta; newDelta = newDelta/deltaFact; Mat expandedConfigs = ExpandConfigsRandom(goodConfigs,m_steps,level,80,deltaFact); configs.release(); configs = Mat(0,6,CV_32FC1); configs.push_back(goodConfigs); configs.push_back(expandedConfigs); } // 4 . 更新 level++; for (int i=0;i<numPoints;++i) { xs[i] = (float) rng.uniform(5,w1-5); ys[i] = (float) rng.uniform(5,h1-5); } } // 返回最佳变换 return bestTransMat; } long CMatchTemp::CreateListOfConfigs(Mat& configs) { configs.release(); vector<float> tx_steps = get_vec_by_setp(m_bounds.tx[0] , m_steps.tx , m_bounds.tx[1]); vector<float> ty_steps = get_vec_by_setp(m_bounds.ty[0] , m_steps.ty , m_bounds.ty[1]); vector<float> r_steps = get_vec_by_setp(m_bounds.r[0] , m_steps.r , m_bounds.r[1]); vector<float> s_steps = get_vec_by_setp(m_bounds.s[0] , m_steps.s , m_bounds.s[1]); if (abs(m_steps.s) ==0) { s_steps.clear(); s_steps.push_back(m_bounds.s[0]); } if (m_bounds.tx[1]-tx_steps.back() > 0.5 * m_steps.tx) { tx_steps .push_back( tx_steps.back() + m_steps.tx ) ; } if (m_bounds.ty[1]-ty_steps.back() > 0.5 * m_steps.ty) { ty_steps .push_back ( ty_steps.back() + m_steps.ty); } if (m_bounds.r[1]-r_steps.back() > m_steps.r) // 这个具有周期性 { r_steps .push_back ( r_steps.back() + m_steps.r); } if (m_bounds.s[1]-s_steps.back() > 0.5 * m_steps.s) { s_steps.push_back( s_steps.back() + m_steps.s); } size_t ntx_steps = tx_steps.size(); size_t nty_steps = (ty_steps).size(); size_t ns_steps = (s_steps).size(); size_t nr_steps = (r_steps).size(); size_t NR2_steps = 0; for (size_t i=0;i<r_steps.size();++i) { if (r_steps[i] < -CV_PI/2 + m_steps.r/2 ) { NR2_steps++; } } long gridSize = ntx_steps*nty_steps*(ns_steps*ns_steps)*(nr_steps*NR2_steps); configs = Mat(0,6,CV_32FC1); int gridInd = 0; float tx,ty,r2,sx,sy,r1; for (size_t tx_ind = 0 ;tx_ind< ntx_steps;tx_ind++) { tx = tx_steps[tx_ind]; for(size_t ty_ind = 0 ;ty_ind< nty_steps ;ty_ind++) { ty = ty_steps[ty_ind]; for(size_t r1_ind = 0;r1_ind< nr_steps;++r1_ind) { r1 = r_steps[r1_ind]; for(size_t r2_ind = 0; r2_ind<NR2_steps ;++r2_ind) { r2 = r_steps[r2_ind]; for(size_t sx_ind = 0;sx_ind< ns_steps;++sx_ind) { sx = s_steps[sx_ind]; for(size_t sy_ind = 0; sy_ind<ns_steps;++sy_ind) { sy = s_steps[sy_ind]; gridInd = gridInd + 1; float f6[] = {tx,ty,r2,sx,sy,r1}; Mat temp = Mat(1,6,CV_32FC1,f6); configs.push_back(temp); } } } } } } return gridSize; } inline vector<float> CMatchTemp::get_vec_by_setp(float start,float setp,float end) { vector<float> out; while (start<=end) { out.push_back(start); start += setp; } return out; } void CMatchTemp::GetAffineFromConfig(const Mat& conf,Mat& out1