double overlapJaccard(cv::Rect& r1,cv::Rect& r2){
double overlap=(r1 & r2).area();
double overlapJaccard=overlap/(r1.area()+r2.area()-overlap);
return overlapJaccard;
}
//[0; 1] (0.5 when different_area == common_area)
inline float rect_similarity(const cv::Rect &r1, const cv::Rect &r2)
{
float common = (r1 & r2).area();
float different = (r1.area() + r2.area() - 2.0f * common);
if (different > FLT_EPSILON)
return std::min(0.5f * common / different, 1.0f);
else
return 1.0f;
}
inline void CalcVarianceAndSD(cv::Rect &block, cv::Mat &sum, cv::Mat &sqsum, double &mean, double &stdvar)
{
double brs = sum.at<int>(block.y+block.height,block.x+block.width); // D
double bls = sum.at<int>(block.y+block.height,block.x); // C
double trs = sum.at<int>(block.y,block.x+block.width); // B
double tls = sum.at<int>(block.y,block.x); // A
double brsq = sqsum.at<double>(block.y+block.height,block.x+block.width); // D
double blsq = sqsum.at<double>(block.y+block.height,block.x); // C
double trsq = sqsum.at<double>(block.y,block.x+block.width); // B
double tlsq = sqsum.at<double>(block.y,block.x); // A
mean = (brs + tls-trs-bls)/((double)block.area() + 1); // D + A - B - C
double sqmean = (brsq+tlsq-trsq-blsq)/((double)block.area() + 1); // D + A - B - C
stdvar = sqrt(sqmean - mean * mean);
return;
}
void CallBackFunc(int evnt, int x, int y, int flags, void* userdata) {
if (evnt == cv::EVENT_LBUTTONDOWN) {
mouseButtonDown = true;
targetSelected = false;
boundingRect = cv::Rect(0,0,0,0);
point1 = cv::Point(x,y);
cv::destroyWindow(targetName);
cv::destroyWindow(ColorTracker.getColorSquareWindowName());
targetImage.release();
}
if (evnt == cv::EVENT_MOUSEMOVE) {
if (x < 0) x = 0;
else if (x > image.cols) x = image.cols;
if (y < 0) y = 0;
else if (y > image.rows) y = image.rows;
point2 = cv::Point(x,y);
if (mouseButtonDown) {
boundingRect = cv::Rect(point1,point2);
}
cv::imshow(imageName,image);
}
if (evnt == cv::EVENT_LBUTTONUP) {
mouseButtonDown = false;
if (boundingRect.area() != 0) {
targetImage = image(calibratedRect(boundingRect));
cv::imshow(targetName, targetImage);
}
else {
boundingRect = cv::Rect(point1-cv::Point(5,5),point1+cv::Point(5,5));
targetImage = image(calibratedRect(boundingRect));
cv::imshow(targetName, targetImage);
}
targetSelected = true;
}
}
bool MotionDetector::isInMotion(cv::Rect boundingBox, float interSection)
{
if (cv::sum(m_motionMap(boundingBox))[0] > interSection * boundingBox.area())
return true;
else
return false;
}
bool EndToEndTest::rectMatches(cv::Rect actualPlate, PlateRegion candidate)
{
// Determine if this region matches our plate in the image
// Do this simply by verifying that the center point of the plate is within the region
// And that the plate region is not x% larger or smaller
const float MAX_SIZE_PERCENT_LARGER = 0.65;
//int plateCenterX = actualPlate.x + (int) (((float) actualPlate.width) / 2.0);
//int plateCenterY = actualPlate.y + (int) (((float) actualPlate.height) / 2.0);
//Point centerPoint(plateCenterX, plateCenterY);
vector<Point> requiredPoints;
requiredPoints.push_back(Point( actualPlate.x + (int) (((float) actualPlate.width) * 0.2),
actualPlate.y + (int) (((float) actualPlate.height) * 0.15)
));
requiredPoints.push_back(Point( actualPlate.x + (int) (((float) actualPlate.width) * 0.8),
actualPlate.y + (int) (((float) actualPlate.height) * 0.15)
));
requiredPoints.push_back(Point( actualPlate.x + (int) (((float) actualPlate.width) * 0.2),
actualPlate.y + (int) (((float) actualPlate.height) * 0.85)
));
requiredPoints.push_back(Point( actualPlate.x + (int) (((float) actualPlate.width) * 0.8),
actualPlate.y + (int) (((float) actualPlate.height) * 0.85)
));
float sizeDiff = 1.0 - ((float) actualPlate.area()) / ((float) candidate.rect.area());
//cout << "Candidate: " << candidate.rect.x << "," << candidate.rect.y << " " << candidate.rect.width << "-" << candidate.rect.height << endl;
//cout << "Actual: " << actualPlate.x << "," << actualPlate.y << " " << actualPlate.width << "-" << actualPlate.height << endl;
//cout << "size diff: " << sizeDiff << endl;
bool hasAllPoints = true;
for (int z = 0; z < requiredPoints.size(); z++)
{
if (candidate.rect.contains(requiredPoints[z]) == false)
hasAllPoints = false;
break;
}
if ( hasAllPoints &&
(sizeDiff < MAX_SIZE_PERCENT_LARGER) )
{
return true;
}
else
{
for (int i = 0; i < candidate.children.size(); i++)
{
if (rectMatches(actualPlate, candidate.children[i]))
return true;
}
}
return false;
}
bool MotionDetector::containsBoundingBox(cv::Rect outer, cv::Rect inner)
{
float intersectionParam = 0.45;
cv::Rect intersect = outer & inner;
if (intersect .area() >= intersectionParam * inner.area())
return true;
else
return false;
}
bool fusionRects(cv::Rect prevRectK, cv::Rect rectK)
{
if(prevRectK.area() == 0)
return false;
if(rectK.width > 1.8*prevRectK.width)
return true;
else
return false;
}
region_descriptor create_rectangle(cv::Rect rect)
{
region_descriptor result;
for(int i = 0; i < rect.height; ++i)
result.lineIntervals.push_back(region_interval(i+rect.y, rect.x, rect.x + rect.width));
result.bounding_box = rect;
result.m_size = rect.area();
return result;
}
void ImageSegmentation::thresholding(const cv::Mat &grey, cv::Mat &bin, const double threshold, const int method, const cv::Rect roi)
{
if(roi.area() == 0)
{
cv::threshold(grey, bin, threshold, 255, method);
}
else
{
cv::threshold(grey(roi), bin, threshold, 255, method);
}
}
void Detector::set_state(cv::Rect state, float confidence) {
CHECK(0.0f <= confidence && confidence <= 1.0f);
if (!IsRectInsideFrame(state, frame_)) {
WARNING("given state " << state << " is outside the frame");
} else if (state.area() == 0) {
WARNING("given state is empty");
}
state_ = state;
confidence_ = confidence;
}
bool getBox()
{
// Crops the image based on user input and creates a template for the tracker with it.
printf("Reading image!!\n");
ImageOf<PixelRgb> *imgIn = imInPort.read(); // read an image
cv::Mat img((IplImage*) imgIn->getIplImage());
printf("Click first top left and then bottom right from the object !!\n");
bool boxOK = false;
//Bottle &out = coordsOutPort.prepare();
cv::Point tl, br;
while (!boxOK){
printf("Click on top left!\n");
Bottle *point1 = coordsInPort.read(true);
tl.x = point1->get(0).asDouble();
tl.y = point1->get(1).asDouble();
printf("Point read at %d, %d!!\n", tl.x, tl.y);
printf("Click on bottom right!\n");
Bottle *point2 = coordsInPort.read(true);
br.x = point2->get(0).asDouble();
br.y = point2->get(1).asDouble();
printf("Point read at %d, %d!!\n", br.x, br.y);
BBox = cv::Rect(tl,br);
if (BBox.area() > 20) {
printf("valid coordinates, cropping image!\n");
boxOK = true;}
else {printf("Coordinates not valid, click again!\n");}
}
printf("Prep out mat !!\n");
ImageOf<PixelRgb> &templateOut = tempOutPort.prepare();
templateOut.resize(BBox.width, BBox.height);
cv::Mat tempOut((IplImage*)templateOut.getIplImage(),false);
img(BBox).copyTo(tempOut);
//cv::GaussianBlur(img(BBox), imOut, cv::Size(1,1), 1, 1);
double t0 = Time::now();
while(Time::now()-t0 < 1) { //send the template for one second
printf("Writing Template!\n");
tempOutPort.write();
Time::delay(0.1);
}
tracking = true;
return true;
}
void DrawCircles(Mat& frame, std::vector<Rect>& hands, cv::Rect& maxRect, int& posX, int& posY)
{
// Draw circles on the detected hands
for (int i = 0; i < hands.size(); i++)
{
if (hands[i].area() > maxRect.area())
maxRect = hands[i];
}
Point center(maxRect.x + maxRect.width*0.5, maxRect.y + maxRect.height*0.5);
ellipse(frame, center, Size(maxRect.width*0.5, maxRect.height*0.5), 0, 0, 360, Scalar(255, 0, 255), 4, 8, 0);
circle(frame, center, 5, Scalar(144, 144, 255), 3);
posX = maxRect.x + maxRect.width*0.5;
posY = maxRect.y + maxRect.height*0.5;
}
FrameDescriptor::FrameDescriptor(int numberOfFeatures, cv::Rect roi_rect) : numberOfFeatures(numberOfFeatures), roi_rect(roi_rect),frame_number(0) {
//Allocate KeyPoint Vector
featurePoints.reserve(numberOfFeatures);
//if we have an roi to generate, set the flag, otherwise, set process frame to point to refFrame
has_been_normalized = false;
if(roi_rect.area() > 0) {
roi_set = true;
roi_offset = cv::Point2f(roi_rect.x,roi_rect.y);
process_frame = new cv::Mat();
} else {
roi_set = false;
roi_offset = cv::Point2f(0,0);
process_frame = &refFrame;
}
}
void FaceProcessor::enrollImagePart(const cv::Mat &rgbImage, double &resRed, double &resGreen, double &resBlue, double &resT, cv::Rect roirect)
{
if(roirect == cv::Rect()) {
roirect = cv::Rect(0,0,rgbImage.cols,rgbImage.rows);
} else {
roirect = roirect & cv::Rect(0,0,rgbImage.cols,rgbImage.rows);
}
unsigned long red = 0;
unsigned long green = 0;
unsigned long blue = 0;
unsigned long area = 0;
if(roirect.area() > 0) {
cv::Mat region = cv::Mat(rgbImage,roirect);
unsigned char *ptr;
unsigned char tR = 0, tG = 0, tB = 0;
#pragma omp parallel for private(ptr,tB,tG,tR) reduction(+:area,green)
for(int j = 0; j < roirect.height; j++) {
ptr = region.ptr(j);
for(int i = 0; i < roirect.width; i++) {
tB = ptr[3*i];
tG = ptr[3*i+1];
tR = ptr[3*i+2];
if( /*__skinColor(tR, tG, tB)*/ true) {
area++;
red += tR;
green += tG;
blue += tB;
}
}
}
}
resT = ((double)cv::getTickCount() - (double)m_markTime)*1000.0 / cv::getTickFrequency();
m_markTime = cv::getTickCount();
if(area > 16) {
resRed = ((double)red) / area;
resGreen = ((double)green) / area;
resBlue = ((double)blue) / area;
} else {
resRed = 0.0;
resGreen = 0.0;
resBlue = 0.0;
}
}
bool rectHasLargerArea(cv::Rect a, cv::Rect b) { return a.area() < b.area(); };
static PixelType mean_from_integral(const cv::Mat& mat, cv::Rect region)
{
return sum_from_integral<PixelType>(mat, region) / region.area();
}
cv::Mat_<byte> DepthSegmenter::calcMostCommonDepthMask( const cv::Rect subRect, cv::Mat_<float> mask) const
{
cv::Mat_<float> subMat = _rngMat;
if ( subRect.area() > 0)
subMat = _rngMat(subRect);
if ( mask.empty())
mask = cv::Mat_<float>::ones(subMat.size());
assert( mask.rows == subMat.rows && mask.cols == subMat.cols);
// Find the most common depth
int *bins = (int*)calloc( _depthLevels, sizeof(int)); // Zero'd
float *means = (float*)calloc( _depthLevels, sizeof(float)); // Zero'd within bin depth means
std::vector< std::vector<float> > dvals( _depthLevels);
const double rngDelta = _maxRng - _minRng;
int topIdx = 0; // Remember top index (most hits)
int maxCnt = 0;
const int rows = subMat.rows;
const int cols = subMat.cols;
for ( int i = 0; i < rows; ++i)
{
const float *pxRow = subMat.ptr<float>(i);
const float *maskRow = mask.ptr<float>(i);
for ( int j = 0; j < cols; ++j)
{
if ( maskRow[j]) // Ignore zero values
{
const float depth = pxRow[j];
if ( depth <= 0)
continue;
const int b = binVal( bins, means, depth, rngDelta);
if ( b < 0) // Out of range so continue
continue;
dvals[b].push_back(depth); // Store the depth value itself for later std-dev calc
if ( bins[b] > maxCnt)
{
topIdx = b;
maxCnt = bins[b];
} // end if
} // end if
} // end for - cols
} // end for - rows
/*
for ( int i = 0; i < _depthLevels; ++i)
cerr << "Bin " << i << ": " << bins[i] << endl;
cerr << "Top bin = " << topIdx << endl;
*/
free(bins);
const float meanDepth = means[topIdx]; // Mean of most common depth values
//free(means);
// Calculate the std-dev for the most common depth interval
const std::vector<float>& vds = dvals[topIdx];
const int sz = (int)vds.size();
double sumSqDiffs = 0;
for ( int i = 0; i < sz; ++i)
sumSqDiffs += pow(vds[i]-meanDepth,2);
const double stddev = sqrt(sumSqDiffs/sz);
_lastStdDev = float(stddev);
// Identify the pixels closest to meanDepth within c*stddev
const float withinRng = _inlierFactor * _lastStdDev;
float minDepth = FLT_MAX;
float maxDepth = 0;
float totDepth = 0;
int pxCount = 0;
cv::Mat_<byte> outMat( rows, cols);
for ( int i = 0; i < rows; ++i)
{
byte* outRow = outMat.ptr<byte>(i);
const float* inRow = subMat.ptr<float>(i);
for ( int j = 0; j < cols; ++j)
{
const float depth = inRow[j];
if ( depth > 0 && (fabs(depth - meanDepth) < withinRng))
{
outRow[j] = 0xff;
totDepth += depth;
pxCount++;
if ( depth < minDepth)
minDepth = depth;
if ( depth > maxDepth)
maxDepth = depth;
} // end if
else
outRow[j] = 0;
} // end for
} // end for
_lastMinDepth = minDepth;
_lastMaxDepth = maxDepth;
_lastAvgDepth = totDepth/pxCount;
//cv::fastFree(dvals);
return outMat;
} // end calcMostCommonDepthMask
void displacement(const cv::Rect& initial, const cv::Rect& last, cv::Point& centerDiff, double& areaRatio) {
cv::Point centerInitial = center(initial);
cv::Point centerLast = center(last);
centerDiff = centerInitial - centerLast;
areaRatio = ((double)(initial.area())) / last.area();
}
double GridFitter::evaluateCandidate(PipelineGrid& grid, cv::Mat const& roi, cv::Mat const& binarizedROI, cv::Mat const& sobelXRoi, const cv::Mat& sobelYRoi, const settings_cache_t &settings)
{
double error = 0;
const cv::Rect boundingBox = grid.getBoundingBox();
// return max error if either width or height is zero
if (!boundingBox.area()) return std::numeric_limits<double>::max();
// also return max error if grid bounding box is not within roi
if (boundingBox.x < 0 || boundingBox.y < 0) return std::numeric_limits<double>::max();
if (boundingBox.x + boundingBox.width >= roi.rows ||
boundingBox.y + boundingBox.height >= roi.cols) return std::numeric_limits<double>::max();
enum ROI {
BINARY = 0,
GRAYSCALE
};
static const ROI roiKind = ROI::GRAYSCALE;
static const double numErrorMeasurements = 6.;
const cv::Mat& selectedRoi = roiKind == ROI::BINARY ? binarizedROI : roi;
{
error_counter_t<expected_white_error_fun_t> errorFun(selectedRoi);
errorFun = grid.processInnerWhiteRingCoordinates(std::move(errorFun));
error += errorFun.getNormalizedError() * settings.err_func_alpha_inner;
}
{
error_counter_t<expected_black_error_fun_t> errorFun(selectedRoi);
errorFun = grid.processInnerBlackRingCoordinates(std::move(errorFun));
error += errorFun.getNormalizedError() * settings.err_func_alpha_inner;
}
for (size_t cellIdx = 0; cellIdx < Grid::NUM_MIDDLE_CELLS; ++cellIdx) {
variance_online_calculator_t errorFun(selectedRoi);
errorFun = grid.processGridCellCoordinates(cellIdx, std::move(errorFun));
error += (errorFun.getNormalizedVariance() / Grid::NUM_MIDDLE_CELLS) * settings.err_func_alpha_variance;
}
{
error_counter_t<expected_white_error_fun_t> errorFun(selectedRoi);
errorFun = grid.processOuterRingCoordinates(std::move(errorFun));
error += errorFun.getNormalizedError() * settings.err_func_alpha_outer;
}
{
sobel_error_counter_t errorFun(sobelXRoi, sobelYRoi, settings.sobel_threshold);
errorFun = grid.processOuterRingEdgeCoordinates(std::move(errorFun));
error += errorFun.getNormalizedError() * settings.err_func_alpha_outer_edge;
}
{
sobel_error_counter_t errorFun(sobelXRoi, sobelYRoi, settings.sobel_threshold);
errorFun = grid.processInnerLineCoordinates(std::move(errorFun));
error += errorFun.getNormalizedError() * settings.err_func_alpha_inner_edge;
}
error /= numErrorMeasurements;
return error;
}
bool testThreshold( const cv::Mat_<double>& ii, const cv::Rect& rct, double thresh)
{
return ( RFeatures::getIntegralImageSum<double>( ii, rct) / rct.area()) >= thresh;
} // end testThreshold
Mat PicOpt::Optimize9Patch::ResizeImageRect(const Mat &img,
const cv::Rect &rc,
bool is_hrz,
Vec2i &new_patch)
{
if (rc.area() <= 0)
{
return img;
}
auto size = img.size();
if (is_hrz)
{
cv::Rect left(Point(), size);
cv::Rect center = left;
cv::Rect right = left;
int width = (std::min)(rc.width, center_width_);
left.width = rc.x;
center.x = rc.x + (rc.width - width) / 2;
center.width = width;
right.x = rc.x + rc.width; // note that the line is [x, y)
right.width = size.width - right.x;
Mat out(size.height, left.width + center.width + right.width, img.type());
cv::Rect out_center = center;
out_center.x = left.width;
cv::Rect out_right = right;
out_right.x = left.width + center.width;
CopyImageRect(img, left, out, left);
CopyImageRect(img, center, out, out_center);
CopyImageRect(img, right, out, out_right);
new_patch[0] = left.width;
new_patch[1] = right.width;
return out;
}
else
{
cv::Rect top(Point(), size);
cv::Rect center = top;
cv::Rect bottom = top;
int height = (std::min)(rc.height, center_width_);
top.height = rc.y;
center.y = rc.y + (rc.height - height) / 2;
center.height = height;
bottom.y = rc.y + rc.height;
bottom.height = size.height - bottom.y;
Mat out(top.height + center.height + bottom.height, size.width, img.type());
cv::Rect out_center = center;
out_center.y = top.height;
cv::Rect out_bottom = bottom;
out_bottom.y = top.height + center.height;
CopyImageRect(img, top, out, top);
CopyImageRect(img, center, out, out_center);
CopyImageRect(img, bottom, out, out_bottom);
new_patch[0] = top.height;
new_patch[1] = bottom.height;
return out;
}
}
void TrackAssociation::SuperviseTrainingForSubTracks()
{
std::list<ObjSubTrack*> valid_track;
for (size_t i = 0; i < subTrackVector.size(); ++i)
{
ObjSubTrack *subTrack = subTrackVector[i];
if (subTrack->appModel.valid)
{
valid_track.push_back(subTrack);
}
}
for (size_t i = 0; i<subTrackVector.size(); ++i)
{
ObjSubTrack *subTrack = subTrackVector[i];
if (subTrack->appModel.valid)
{
const DetectionModel &model1 = subTrack->detectionSubTrack.back();
const cv::Rect r1 = model1.ConvertToRect();
Rect r1_rect;
r1_rect = r1;
Rect negative_region_rect = subTrack->appModel.tracker.getTrackingROI(r1_rect,7);
cv::Rect negative_region = negative_region_rect.ConvertToRect();
bool notTrain = false;
std::list<ObjSubTrack*> negative_list; //other objTracks should be discriminated
if (subTrack->appModel.tracker.tracking_lost)
{
notTrain = true;
// for (std::list<ObjSubTrack *>::iterator iter = valid_track.begin(); iter!=valid_track.end(); ++iter)
// {//push all other than itself
// if ((*iter)!=subTrack)
// {
// negative_list.push_back(*iter);
// }
// }
}
//collect neg tracks
for (std::list<ObjSubTrack *>::iterator iter = valid_track.begin(); iter!=valid_track.end(); ++iter)
{
if ((*iter) != subTrack)
{
const DetectionModel &model2 = (*iter)->detectionSubTrack.back();
const cv::Rect r2 = model2.ConvertToRect();
cv::Rect intersect = r1 & r2;
float overlap = intersect.area()/float(r1.area());
if (overlap > 0.1f)
{
notTrain = true;
}
if ((*iter)->appModel.tracker.tracking_lost)
{
negative_list.push_back(*iter);
}else
{
if ((r2 & negative_region).area()>0)
{
negative_list.push_back(*iter);
}
}
}
}
//positive samples
if (!notTrain)
{
if (subTrack->detectionSubTrack.back().external_detection)
{
subTrack->SupervisedTraining(frame_pool);
}else
{
subTrack->SemiSupervisedTraining(frame_pool);
}
}
subTrack->AddToNegList(negative_list);//Add to negative set (construct Nearest Neighbor Classifier)
}
}
}
bool vpSortLargestFace(cv::Rect rect1, cv::Rect rect2)
{
return (rect1.area() > rect2.area());
}
bool MotionDetector::evaluateStaticObjRect(cv::Rect& tmp_rect)
{
int tmp_area = tmp_rect.area();
bool isChild = false;
bool isParent = false;
bool result = false;
if (tmp_area > m_MIN_BLOBSIZE) // reject rects that are to small
{
if (!isInMotion(tmp_rect, 0.95)) // reject rects that are in motion
{
if (!(m_staticObjectCandidates.size() > m_MAX_STATIC_OBJ)) // reject rects after max count
{
AmbiguousCandidate tmp;
tmp.boundingBox = tmp_rect;
tmp.counter = 0;
tmp.accMovement = 0;
tmp.state = DETECTOR_UNCLASSIFIED_OBJECT;
std::set<int> deleteIdx;
int iteration = 0;
for (int k = 0; k < m_staticObjectCandidates.size(); ++k)
{
if (containsBoundingBox(m_staticObjectCandidates[k].boundingBox, tmp_rect)) // is within
{
isChild = true;
result = false;
return result;
}
if (containsBoundingBox(tmp_rect, m_staticObjectCandidates[k].boundingBox)) // contains existing bounding box
{
if (isParent)
{
deleteIdx.insert(k);
}
else
{
isParent = true;
result = true;
m_staticObjectCandidates[k].boundingBox = tmp_rect;
}
}
}
for (auto it = deleteIdx.begin(); it != deleteIdx.end(); ++it)
{
m_staticObjectCandidates.erase(m_staticObjectCandidates.begin() + *it - iteration);
++iteration;
}
if (!isParent && !isChild)
{
m_staticObjectCandidates.push_back(tmp);
result = true;
}
}
}
}
return result;
}
