void run_teste_integral_2(){
std::string crop_image = Config::PROJECT_PATH + "/dataset/24x24/Instancia_Sun_300_400/training_images/non_faces/group_0/crop/1403571922.05/crop_399.pgm";
std::string full_image = Config::PROJECT_PATH + "/dataset/24x24/Instancia_Sun_nocrop/training_images/non_faces/SUN2012/16155.pgm";
IntegralImage ii1(crop_image);
teste_integral_2(ii1);
ulong*** data;
int crop_start_index = 10;
int ncrops = 1;
int maxCrops = 10;
int crop_width = 24;
int crop_height = 24;
int shift_step = 24;
Point img_size;
int last_crop;
int random_hop = 23;
ulong** img_data;
getImageCrops(&data,&img_data,&img_size,full_image.c_str(),&crop_start_index,ncrops,maxCrops,crop_width,crop_height,shift_step,random_hop,checkData,NULL);
Point p;
p.x = 24;
p.y = 24;
IntegralImage ii2(data[0],p,true);
teste_integral_2(ii2);
}
// Calculates the array of values used as the integral image, lest values need calculation.
void JCHaarFinder::generateIIArray() {
// if true, the gnerated II array will be printed to a file after calculation
bool takeSnapshot = false;
int width = curImage->getWidth(),
height = curImage->getHeight();
if (_iiArray == NULL) {
free(_iiArray);
}
_iiArray = (int*) malloc(width * height * sizeof(int));
_ii2Array = (int*) malloc(width * height * sizeof(int));
for (int i = 0; i<width*height; i++) {
_iiArray[i] = 0;
_ii2Array[i] = 0;
}
for (int row=0; row < height; row++) {
int rowsum = 0;
int rowsum2 = 0;
/* The strategy here is to keep a sum of the current row that is calculated.
* The values of the fields in the rows are continuously added to the rowsum
* variable, and the columns are taken care of by looking earlier in the array
* as we're building it (we look in the previous row at the same column). This
* eventually yields the integral image.
*
* Also, we keep an integral image over all the squares of the pixel values. This
* is used for the standard deviations in the normalization stage when the feature
* sums are compared to their thresholds.
*/
for(int col=0; col < width; col++) {
rowsum += i(col, row);
rowsum2 += i(col, row)*i(col, row);
if (row > 0) {
_iiArray[row*width + col] = rowsum + ii(col, row-1);
_ii2Array[row*width + col] = rowsum2 + ii2(col, row-1);
} else {
_iiArray[row*width + col] = rowsum;
_ii2Array[row*width + col] = rowsum2;
}
}
}
// Debug functionality that writes out the integral images in CSV structures.
if (takeSnapshot) {
ofFile iiFile;
ofFile ii2File;
iiFile.open(ofToDataPath("ii.csv"), ofFile::WriteOnly, false);
ii2File.open(ofToDataPath("ii2.csv"), ofFile::WriteOnly, false);
for (int row= 0; row<height; row++) {
for (int col = 0; col<width; col++) {
iiFile << ofToString(ii(col, row)) + ";";
ii2File << ofToString(ii2(col, row)) + ";";
}
iiFile << endl;
ii2File << endl;
}
iiFile.close();
}
}
// Returns a vector of rectangles that pass the requirements for faces.
vector<ofxCvBlob> JCHaarFinder::getRectsFromImage(ofImage* inputImage) {
curImage = inputImage;
float scale = 1;
float scaleMultiplier = 1.25;
blobs.clear();
generateIIArray();
/*
* Strategy: Loop through the image such that a window moves over the
* picture, checking stages. Once the window has moved through the image
* scale it a bit up, and re-run the algorithm.
*
* The values to be for'ed over are nested as follows:
*
* the scale of the cascade window
* window's position on y axis
* window's position on x axis
* stages of the cascade
* features of the stage
* rectangles of the feature
*
* Since we don't support larger trees than the ones which only have a
* root node, we can ignore traversing trees and just consider a list
* of rectangles for any one feature.
*/
// This implementation keeps enlarging the window until it is bigger than the picture.
for (scale = 1; scale*casc.height <= curImage->getHeight() || scale*casc.width <= curImage->getWidth(); scale *= scaleMultiplier) {
int window_area = scale*casc.height * scale*casc.width;
// Keep moving the window down as long as its offset and its height are within the image. Likewise to the left
// As an experiment, we're shifting the window in chunks equal to about a tenth of the current window height and width, respectively.
for (int offsetY = 0; (offsetY+scale*casc.height) < curImage->getHeight(); offsetY += (int) ((scale*casc.height)/10) ) {
for (int offsetX = 0; (offsetX + scale*casc.width) < curImage->getWidth(); offsetX += (int) ((scale*casc.width)/10)) {
bool passed = true;
// *** LOOPING OVER STAGES
for (int curStageIdx = 0; curStageIdx < casc.stages.size(); curStageIdx++) {
if (passed == false) break;
stage* s = &casc.stages.at(curStageIdx);
float stage_sum = 0.0;
// *** LOOPING OVER FEATURES
for (int featureIdx = 0; featureIdx < s->features.size(); featureIdx++) {
feature* f = & (s->features.at(featureIdx));
float feature_sum = 0.0;
// These coordinates are for the sliding window
int px1 = offsetX,
py1 = offsetY,
px2 = offsetX + scale*casc.width,
py2 = offsetY + scale*casc.height;
float mean = ((float) (ii(px2, py2) + ii(px1, py1) - ii(px1, py2) - ii(px2, py1))) / window_area;
float stddev = sqrt((ii2(px2, py2) + ii2(px1, py1) - ii2(px1, py2) - ii2(px2, py1)) / window_area - mean*mean);
// *** LOOPING OVER RECTANGLES
for (int rectangleIdx = 0; rectangleIdx < f->rectangles.size(); rectangleIdx++) {
featureRect* r = & ( f->rectangles.at(rectangleIdx));
// These coordinates are for the rectangles inside the sliding window. The coordinates
// are absolute; i.e. they share the same origin as the coordinates of the window.
int x1 = offsetX + (r->rectangle.x * scale),
y1 = offsetY + (r->rectangle.y * scale),
x2 = x1 + r->rectangle.width * scale,
y2 = y1 + r->rectangle.height * scale;
int thisRect = (ii(x2, y2) + ii(x1, y1) - ii(x1, y2) - ii(x2, y1)) * r->weight;
feature_sum += thisRect;
}
// Determine in which direction the cascade should "fall". If the feature sum is less than
// its threshold, fall left, otherwise right.
// http://stackoverflow.com/questions/978742/what-do-the-left-and-right-values-mean-in-the-haar-cascade-xml-files
if (feature_sum/window_area < f->threshold*stddev) {
stage_sum += f->leftVal;
} else {
stage_sum += f->rightVal;
}
}
// The stage is passed if its sum is above its threshold.
passed = (stage_sum > s->threshold);
}
// passed will be true iff all stages passed; if so, we detected a face.
if (passed) {
ofLog(OF_LOG_NOTICE, "detected face (maybe)!");
blobs.push_back(makeBlob(offsetX, offsetY, scale*casc.width, scale*casc.height));
}
}
}
}
return blobs;
}
