void CmSaliencyRC::GetCmplx(CMat& mag32F, CMat& ang32F, Mat& cmplx32FC2)
{
CV_Assert(mag32F.type() == CV_32FC1 && ang32F.type() == CV_32FC1 && mag32F.size() == ang32F.size());
cmplx32FC2.create(mag32F.size(), CV_32FC2);
for (int y = 0; y < mag32F.rows; y++)
{
float* cmpD = cmplx32FC2.ptr<float>(y);
const float* dataA = ang32F.ptr<float>(y);
const float* dataM = mag32F.ptr<float>(y);
for (int x = 0; x < mag32F.cols; x++, cmpD += 2)
{
cmpD[0] = dataM[x] * cos(dataA[x]);
cmpD[1] = dataM[x] * sin(dataA[x]);
}
}
}
void CmSaliencyRC::SmoothSaliency(CMat &colorNum1i, Mat &sal1f, float delta, const vector<vector<CostfIdx>> &similar)
{
if (sal1f.cols < 2)
return;
CV_Assert(sal1f.rows == 1 && sal1f.type() == CV_32FC1);
CV_Assert(colorNum1i.size() == sal1f.size() && colorNum1i.type() == CV_32SC1);
int binN = sal1f.cols;
Mat newSal1d= Mat::zeros(1, binN, CV_64FC1);
float *sal = (float*)(sal1f.data);
double *newSal = (double*)(newSal1d.data);
int *pW = (int*)(colorNum1i.data);
// Distance based smooth
int n = max(cvRound(binN * delta), 2);
vecD dist(n, 0), val(n), w(n);
for (int i = 0; i < binN; i++){
const vector<CostfIdx> &similari = similar[i];
double totalDist = 0, totoalWeight = 0;
for (int j = 0; j < n; j++){
int ithIdx =similari[j].second;
dist[j] = similari[j].first;
val[j] = sal[ithIdx];
w[j] = pW[ithIdx];
totalDist += dist[j];
totoalWeight += w[j];
}
double valCrnt = 0;
for (int j = 0; j < n; j++)
valCrnt += val[j] * (totalDist - dist[j]) * w[j];
newSal[i] = valCrnt / (totalDist * totoalWeight);
}
normalize(newSal1d, sal1f, 0, 1, NORM_MINMAX, CV_32FC1);
}
// Training SVM with feature vector X and label Y.
// Each row of X is a feature vector, with corresponding label in Y.
// Return a CV_32F weight Mat
Mat Objectness::trainSVM(CMat &X1f, const vecI &Y, int sT, double C, double bias, double eps)
{
// Set SVM parameters
parameter param; {
param.solver_type = sT; // L2R_L2LOSS_SVC_DUAL;
param.C = C;
param.eps = eps; // see setting below
param.p = 0.1;
param.nr_weight = 0;
param.weight_label = NULL;
param.weight = NULL;
set_print_string_function(print_null);
CV_Assert(X1f.rows == Y.size() && X1f.type() == CV_32F);
}
// Initialize a problem
feature_node *x_space = NULL;
problem prob;{
prob.l = X1f.rows;
prob.bias = bias;
prob.y = Malloc(double, prob.l);
prob.x = Malloc(feature_node*, prob.l);
const int DIM_FEA = X1f.cols;
prob.n = DIM_FEA + (bias >= 0 ? 1 : 0);
x_space = Malloc(feature_node, (prob.n + 1) * prob.l);
int j = 0;
for (int i = 0; i < prob.l; i++){
prob.y[i] = Y[i];
prob.x[i] = &x_space[j];
const float* xData = X1f.ptr<float>(i);
for (int k = 0; k < DIM_FEA; k++){
x_space[j].index = k + 1;
x_space[j++].value = xData[k];
}
if (bias >= 0){
x_space[j].index = prob.n;
x_space[j++].value = bias;
}
x_space[j++].index = -1;
}
CV_Assert(j == (prob.n + 1) * prob.l);
}
// Training SVM for current problem
const char* error_msg = check_parameter(&prob, ¶m);
if(error_msg){
fprintf(stderr,"ERROR: %s\n",error_msg);
exit(1);
}
model *svmModel = train(&prob, ¶m);
Mat wMat(1, prob.n, CV_64F, svmModel->w);
wMat.convertTo(wMat, CV_32F);
free_and_destroy_model(&svmModel);
destroy_param(¶m);
free(prob.y);
free(prob.x);
free(x_space);
return wMat;
}
void BenchMarkLatex::printMatTransport(CMat &mat1di, CStr texFile, const char* rowDes[], bool *descendOrder)
{
FILE* f = fopen(_S(texFile), "w");
if (f == NULL){
printf("Can't open file %s\n", _S(texFile));
return;
}
CV_Assert(mat1di.cols == _numMethod);
const int MAX_COL = min(20, _numMethod);
const int NUM_BLOCK = (_numMethod + MAX_COL - 1) / MAX_COL;
string strAlign = "|l||";
for (int i = 0; i < MAX_COL; i++)
strAlign += "c|";
const char* rankCommand[3] = {"\\first", "\\second", "\\third"};
Mat mat1d = mat1di.t();
mat1d.convertTo(mat1d, CV_64F);
Mat rnk1i = getRankIdx(mat1d, true).t();
Mat rnkInv = getRankIdx(mat1d, false).t();
for (int i = 0; i < mat1di.rows; i++)
if (!descendOrder[i])
rnkInv.row(i).copyTo(rnk1i.row(i)); // Revert ordering for row i
for (int b = 0; b < NUM_BLOCK; b++){
fprintf(f, "\\begin{tabular}{%s} \\hline\n\tMethod ", _S(strAlign));
for (int i = 0; i < MAX_COL; i++){
int colIdx = i + b * MAX_COL;
fprintf(f, "& %4s", colIdx < _numMethod ? _S(_methodNames[colIdx]) : "");
}
fprintf(f, "\\\\\\hline\n");
for (int i = 0; i < mat1di.rows; i++){
fprintf(f, "\t%-5s ", rowDes[i]);
for (int j = 0; j < MAX_COL; j++){
int colIdx = j + b * MAX_COL;
if (colIdx >= _numMethod){
fprintf(f, "& ");
continue;
}
int idx = rnk1i.at<int>(i, colIdx);
if (mat1di.type() == CV_32S){
if (idx < 3)
fprintf(f, "& %s{%d} ", rankCommand[idx], mat1di.at<int>(i, colIdx));
else
fprintf(f, "& %d ", mat1di.at<int>(i, colIdx));
}
else{// CV_64F
if (idx < 3)
fprintf(f, "& %s{%5.3f} ", rankCommand[idx], mat1di.at<double>(i, colIdx));
else
fprintf(f, "& %5.3f ", mat1di.at<double>(i, colIdx));
}
}
fprintf(f, "\\\\\n");
}
fprintf(f, "\\hline\n\\end{tabular}\n");
}
fclose(f);
}
void CmSaliencyRC::SmoothByHist(CMat &img3f, Mat &sal1f, float delta)
{
//imshow("Before", sal1f); imshow("Src", img3f);
// Quantize colors
CV_Assert(img3f.size() == sal1f.size() && img3f.type() == CV_32FC3 && sal1f.type() == CV_32FC1);
Mat idx1i, binColor3f, colorNums1i;
int binN = Quantize(img3f, idx1i, binColor3f, colorNums1i);
//CmShow::HistBins(binColor3f, colorNums1i, "Frequency");
// Get initial color saliency
Mat _colorSal = Mat::zeros(1, binN, CV_64FC1);
int rows = img3f.rows, cols = img3f.cols;{
double* colorSal = (double*)_colorSal.data;
if (img3f.isContinuous() && sal1f.isContinuous())
cols *= img3f.rows, rows = 1;
for (int y = 0; y < rows; y++){
const int* idx = idx1i.ptr<int>(y);
const float* initialS = sal1f.ptr<float>(y);
for (int x = 0; x < cols; x++)
colorSal[idx[x]] += initialS[x];
}
const int *colorNum = (int*)(colorNums1i.data);
for (int i = 0; i < binN; i++)
colorSal[i] /= colorNum[i];
normalize(_colorSal, _colorSal, 0, 1, NORM_MINMAX, CV_32F);
}
// Find similar colors & Smooth saliency value for color bins
vector<vector<CostfIdx>> similar(binN); // Similar color: how similar and their index
Vec3f* color = (Vec3f*)(binColor3f.data);
cvtColor(binColor3f, binColor3f, CV_BGR2Lab);
for (int i = 0; i < binN; i++){
vector<CostfIdx> &similari = similar[i];
similari.push_back(make_pair(0.f, i));
for (int j = 0; j < binN; j++)
if (i != j)
similari.push_back(make_pair(vecDist<float, 3>(color[i], color[j]), j));
sort(similari.begin(), similari.end());
}
cvtColor(binColor3f, binColor3f, CV_Lab2BGR);
//CmShow::HistBins(binColor3f, _colorSal, "BeforeSmooth", true);
SmoothSaliency(colorNums1i, _colorSal, delta, similar);
//CmShow::HistBins(binColor3f, _colorSal, "AfterSmooth", true);
// Reassign pixel saliency values
float* colorSal = (float*)(_colorSal.data);
for (int y = 0; y < rows; y++){
const int* idx = idx1i.ptr<int>(y);
float* resSal = sal1f.ptr<float>(y);
for (int x = 0; x < cols; x++)
resSal[x] = colorSal[idx[x]];
}
//imshow("After", sal1f);
//waitKey(0);
}
void CmShow::SaveShow(CMat& img, CStr& title)
{
if (title.size() == 0)
return;
int mDepth = CV_MAT_DEPTH(img.type());
double scale = (mDepth == CV_32F || mDepth == CV_64F ? 255 : 1);
if (title.size() > 4 && title[title.size() - 4] == '.')
imwrite(title, img*scale);
else if (title.size())
imshow(title, img);
}
int SegmentImage(CMat &_src3f, Mat &pImgInd, float sigma, float c, int min_size)
{
CV_Assert(_src3f.type() == CV_32FC3);
int width(_src3f.cols), height(_src3f.rows);
pImgInd.create(height, width, CV_32S);
image<RGB_f> *im = new image<RGB_f>(width, height, _src3f.data);
image<int> *regIdx = new image<int>(width, height, pImgInd.data);
int regNum = SegmentImage(im, regIdx, sigma, c, min_size);
im->data = NULL; regIdx->data = NULL;
delete im; delete regIdx;
return regNum;
}
Mat CmSaliencyRC::GetFT(CMat &img3f)
{
CV_Assert(img3f.data != NULL && img3f.type() == CV_32FC3);
Mat sal(img3f.size(), CV_32F), tImg;
GaussianBlur(img3f, tImg, Size(3, 3), 0);
cvtColor(tImg, tImg, CV_BGR2Lab);
Scalar colorM = mean(tImg);
for (int r = 0; r < tImg.rows; r++) {
float *s = sal.ptr<float>(r);
float *lab = tImg.ptr<float>(r);
for (int c = 0; c < tImg.cols; c++, lab += 3)
s[c] = (float)(sqr(colorM[0] - lab[0]) + sqr(colorM[1] - lab[1]) + sqr(colorM[2] - lab[2]));
}
normalize(sal, sal, 0, 1, NORM_MINMAX);
return sal;
}
void CmSaliencyRC::AbsAngle(CMat& cmplx32FC2, Mat& mag32FC1, Mat& ang32FC1)
{
CV_Assert(cmplx32FC2.type() == CV_32FC2);
mag32FC1.create(cmplx32FC2.size(), CV_32FC1);
ang32FC1.create(cmplx32FC2.size(), CV_32FC1);
for (int y = 0; y < cmplx32FC2.rows; y++)
{
const float* cmpD = cmplx32FC2.ptr<float>(y);
float* dataA = ang32FC1.ptr<float>(y);
float* dataM = mag32FC1.ptr<float>(y);
for (int x = 0; x < cmplx32FC2.cols; x++, cmpD += 2)
{
dataA[x] = atan2(cmpD[1], cmpD[0]);
dataM[x] = sqrt(cmpD[0] * cmpD[0] + cmpD[1] * cmpD[1]);
}
}
}
// 3, 30, 20, 10
GrabCutMF::GrabCutMF(CMat &img3f, CMat &img3u, CStr &nameNE, float w1, float w2, float w3, float alpha, float beta, float gama, float mu)
:_fGMM(5), _bGMM(5), _w(img3f.cols), _h(img3f.rows)
,_crf(_w, _h, 2), _nameNE(nameNE)
{
CV_Assert(img3f.data != NULL && img3f.type() == CV_32FC3);
_imgBGR3f = img3f;
_img3u = img3u;
_trimap1i.create(_h, _w, CV_32S);
_segVal1f.create(_h, _w, CV_32F);
_unary2f.create(_h, _w, CV_32FC2);
_res1u.create(_h, _w, CV_8U);
if (w1 != 0)
_crf.addPairwiseBilateral(alpha, alpha, beta, beta, beta, img3u.data, w1);
if (w2 != 0)
_crf.addPairwiseGaussian(gama, gama, w2);
if (w3 != 0)
_crf.addPairwiseColorGaussian(mu, mu, mu, img3u.data, w3);
}
// img3f is the input BGR image
void CmColorQua::S_Quantize(CMat& _img3f, Mat &idx1i, int method)
{
CV_Assert(method >= 0 && method < S_Q_NUM && _img3f.data != NULL && _img3f.type() == CV_32FC3);
S_QUANTIZE_FUNC SQ_Function = sqFuns[method];
Mat img;
switch (method) {
case S_BGR: img = _img3f; break;
case S_HSV: cvtColor(_img3f, img, CV_BGR2HSV); break;
case S_LAB: cvtColor(_img3f, img, CV_BGR2Lab); break;
}
idx1i.create(img.size(), CV_32S);
for (int r = 0; r < img.rows; r++) {
const Vec3f * imgD = img.ptr<Vec3f>(r);
int *idx = idx1i.ptr<int>(r);
for (int c = 0; c < img.cols; c++)
idx[c] = (*SQ_Function)(imgD[c]);
}
}
// src3f are BGR, color3f are 1xBinDim matrix represent color fore each histogram bin
int CmColorQua::S_BinInf(CMat& idx1i, Mat &color3f, vecI &colorNum, int method, CMat &src3f)
{
int totalBinNum = 0;
CV_Assert(idx1i.data != NULL && idx1i.type() == CV_32S && method >= 0 && method < S_Q_NUM);
// Find colors for each bin
color3f = Mat::zeros(1, binNum[method], CV_32FC3);
Vec3f* color = (Vec3f*)(color3f.data);
vector<Vec3d> colorD(color3f.cols, 0);
colorNum.resize(color3f.cols, 0);
if (src3f.size() != Size() && src3f.data != NULL) {
for (int r = 0; r < idx1i.rows; r++) {
const int *idx = idx1i.ptr<int>(r);
const Vec3f *src = src3f.ptr<Vec3f>(r);
for (int c = 0; c < idx1i.cols; c++) {
colorD[idx[c]] += src[c];
colorNum[idx[c]] ++;
}
}
}
S_RECOVER_FUNC SR_Function = srFuns[method];
for (int i = 0; i < color3f.cols; i++) {
if (colorNum[i] == 0)
(*SR_Function)(i, color[i]);
else
totalBinNum += colorNum[i];
}
if (method == 1)
cvtColor(color3f, color3f, CV_HSV2BGR);
else if (method == 2)
cvtColor(color3f, color3f, CV_Lab2BGR);
for (int i = 0; i < color3f.cols; i++)
if (colorNum[i] > 0)
color[i] = Vec3f((float)(colorD[i][0]/colorNum[i]), (float)(colorD[i][1]/colorNum[i]), (float)(colorD[i][2]/colorNum[i]));
return totalBinNum;
}
/*
* Segment an image
*
* Returns a color image representing the segmentation.
*
* Input:
* im: image to segment.
* sigma: to smooth the image.
* c: constant for threshold function.
* min_size: minimum component size (enforced by post-processing stage).
* nu_ccs: number of connected components in the segmentation.
* Output:
* colors: colors assigned to each components
* pImgInd: index of each components, [0, colors.size() -1]
*/
int SegmentImage(CMat &_src3f, Mat &pImgInd, double sigma, double c, int min_size)
{
CV_Assert(_src3f.type() == CV_32FC3);
int width(_src3f.cols), height(_src3f.rows);
Mat smImg3f;
GaussianBlur(_src3f, smImg3f, Size(), sigma, 0, BORDER_REPLICATE);
// build graph
edge *edges = new edge[width*height*4];
int num = 0;
{
for (int y = 0; y < height; y++) {
for (int x = 0; x < width; x++) {
if (x < width-1) {
edges[num].a = y * width + x;
edges[num].b = y * width + (x+1);
edges[num].w = diff(smImg3f, x, y, x+1, y);
num++;
}
if (y < height-1) {
edges[num].a = y * width + x;
edges[num].b = (y+1) * width + x;
edges[num].w = diff(smImg3f, x, y, x, y+1);
num++;
}
if ((x < width-1) && (y < height-1)) {
edges[num].a = y * width + x;
edges[num].b = (y+1) * width + (x+1);
edges[num].w = diff(smImg3f, x, y, x+1, y+1);
num++;
}
if ((x < width-1) && (y > 0)) {
edges[num].a = y * width + x;
edges[num].b = (y-1) * width + (x+1);
edges[num].w = diff(smImg3f, x, y, x+1, y-1);
num++;
}
}
}
}
// segment
universe *u = segment_graph(width*height, num, edges, (float)c);
// post process small components
for (int i = 0; i < num; i++) {
int a = u->find(edges[i].a);
int b = u->find(edges[i].b);
if ((a != b) && ((u->size(a) < min_size) || (u->size(b) < min_size)))
u->join(a, b);
}
delete [] edges;
// pick random colors for each component
map<int, int> marker;
pImgInd.create(smImg3f.size(), CV_32S);
int idxNum = 0;
for (int y = 0; y < height; y++) {
int *imgIdx = pImgInd.ptr<int>(y);
for (int x = 0; x < width; x++) {
int comp = u->find(y * width + x);
if (marker.find(comp) == marker.end())
marker[comp] = idxNum++;
int idx = marker[comp];
imgIdx[x] = idx;
}
}
delete u;
return idxNum;
}
