void FastCornerDetector::detect
(
const image::Image<unsigned char> & ima,
std::vector<PointFeature> & regions
)
{
using FastDetectorCall =
xy* (*) (const unsigned char *, int, int, int, int, int *);
FastDetectorCall detector = nullptr;
if (size_ == 9) detector = fast9_detect_nonmax;
if (size_ == 10) detector = fast10_detect_nonmax;
if (size_ == 11) detector = fast11_detect_nonmax;
if (size_ == 12) detector = fast12_detect_nonmax;
if (!detector)
{
std::cout << "Invalid size for FAST detector: " << size_ << std::endl;
return;
}
int num_corners = 0;
xy* detections = detector(ima.data(),
ima.Width(), ima.Height(), ima.Width(),
threshold_, &num_corners);
regions.clear();
regions.reserve(num_corners);
for (int i = 0; i < num_corners; ++i)
{
regions.emplace_back(detections[i].x, detections[i].y);
}
free( detections );
}
> void LinearFilterWx1<PixelType>::applyTo(const Image::Image<PixelType> & srcImage,Image::Image<PixelType> & dstImage) const {
typedef typename PixelType::DataType PixelDataType;
typedef typename PixelType::ComputationType PixelComputationType;
BaseLinearFilterParametersType<PixelDataType,PixelComputationType> parameters(
getFilterData().getDataView(),
getXoffset(),
getYoffset(),
srcImage.getWidth(),
getTotalColor()
);
Algorithm::AlgorithmWx1<
SimpleWx1dataOperationBaseAlgorithm<
BaseLinearFilterAlgorithm<
PixelDataType,
PixelComputationType,
BaseLinearFilterParametersType<PixelDataType,PixelComputationType>,
Algorithm::BaseOperationTempType<PixelDataType,PixelComputationType>
>,
PixelDataType,
PixelComputationType,
BaseLinearFilterParametersType<PixelDataType,PixelComputationType>,
Algorithm::BaseOperationTempType<PixelDataType,PixelComputationType>
>,
PixelDataType,
BaseLinearFilterParametersType<PixelDataType,PixelComputationType>
>(
srcImage.getDataView(),
dstImage.getDataView(),
parameters
);
}
void GradientsDescriptor::fillHisto( const image::Image& _dx, const image::Image& _dy, double* _histo, const jblas::vec2& _startPoint, const jblas::vec2& _direction, double _lineAngle, int _length, double _coef )
{
jblas::vec2 currentPoint = _startPoint + 3.0 * _direction;
double lastNorm = DBL_MAX;
for(int i = 0; i < _length; ++i)
// while(true)
{
int ix = int(currentPoint(0));
int iy = int(currentPoint(1));
if( not check(ix, _dx.width() - 1) or not check(iy, _dx.height() - 1) ) return;
int dx = _dx.getPixelValue<short>( ix, iy, 0 );
int dy = _dy.getPixelValue<short>( ix, iy, 0 );
// double dx = _dx.getSubPixelValue<short>( ix, iy, 0, JfrImage_INTERP_CUBIC );
// double dy = _dy.getSubPixelValue<short>( ix, iy, 0, JfrImage_INTERP_CUBIC );
double norm = sqrt( dx * dx + dy * dy );
// if( norm > lastNorm) return ;
double angle = atan2(dy, dx);
// double diff = cos( angle - _lineAngle);
double correctedAngle = ( angle - _lineAngle);
// int idx = int( (m_count) * ( 1.0 + diff) * 0.5);
// if( idx >= m_count ) idx = m_count - 1;
// if( idx != 0 and idx != (m_count - 1 ) )
for(int j = 0; j < m_count; ++j )
{
// double lnorm = norm * ( 1.0 - fabs( fpow( -1.0 + 2.0 * j / (m_count-1) - diff ), 2.0) );
// double lnorm = norm * exp( -pow( -1.0 + 2.0 * j / (m_count-1) - diff, 2.0) );
// JFR_DEBUG( j << " " << exp( -pow2( cos( M_PI * j / (m_count-1) - 0.5 * correctedAngle ) ) ) );
double lnorm = norm * _coef * exp( -pow2( cos( M_PI * j / (m_count-1) - 0.5 * correctedAngle ) ) );
_histo[ j ] += lnorm;
}
lastNorm = norm;
currentPoint += _direction;
}
}
> void AIL_DLL_EXPORT BoxFilterWx1<PixelType>::applyTo(const Image::Image<PixelType> & srcImage,Image::Image<PixelType> & dstImage) const {
typedef typename PixelType::DataType PixelDataType;
typedef typename PixelType::ComputationType PixelComputationType;
BoxFilterWx1parametersType<PixelDataType,PixelComputationType> parameters(xOffset,filterWidth,PixelType::ComputationRange::getMinPixel(),PixelComputationType(filterWidth));
Algorithm::AlgorithmWx1<
Algorithm::BasicWx1baseAlgorithm<
BaseBoxFilterAlgorithm<
PixelDataType,
PixelComputationType,
BoxFilterWx1parametersType<PixelDataType,PixelComputationType>,
Algorithm::BaseOperationTempType<PixelDataType,PixelComputationType>
>,
PixelDataType,
PixelComputationType,
BoxFilterWx1parametersType<PixelDataType,PixelComputationType>,
Algorithm::BaseOperationTempType<PixelDataType,PixelComputationType>
>,
PixelDataType,
BoxFilterWx1parametersType<PixelDataType,PixelComputationType>
>(
srcImage.getDataView(),
dstImage.getDataView(),
parameters
);
}
void DirectSegmentsBase::coloriage( double x_1, double y_1, double x_2, double y_2, image::Image& inasegment, int _value)
{
double xStep = (x_2 - x_1);
double yStep = (y_2 - y_1);
double norm_steps = sqrt( xStep * xStep + yStep * yStep );
xStep /= norm_steps;
yStep /= norm_steps;
int max_coloriage = (int)norm_steps +1;
for( int k = 0; k < max_coloriage; ++k )
{
int x = (int)( x_1 + k * xStep );
int y = (int)( y_1 + k * yStep );
for( int j = -1; j <= 1; ++j)
{
for( int i = -1; i <= 1; ++i)
{
if( check( x + i, inasegment.width() - 1 ) and check( y + j, inasegment.height() - 1 ) )
{
inasegment.setPixelValue<int>( _value, x+i, y+j, 0 );
}
}
}
}
#if 0
double xStep = (x_2 - x_1);
double yStep = (y_2 - y_1);
double norm_steps = sqrt( xStep * xStep + yStep * yStep );
xStep /= norm_steps;
yStep /= norm_steps;
int max_coloriage = (int)norm_steps +1;
int p_x = x_1;
int p_y = y_1;
if( check( p_x, inasegment.width() - 1 ) and check( p_y, inasegment.height() - 1 ) )
{
inasegment.setPixelValue<int>( _value, p_x, p_y, 0 );
}
for( int k = -1; k <= max_coloriage; ++k )
{
int x = (int)( x_1 + k * xStep );
int y = (int)( y_1 + k * yStep );
if( check( p_x, inasegment.width() - 1 ) and check( y, inasegment.height() - 1 ) )
{
inasegment.setPixelValue<int>( _value, p_x, y, 0 );
}
if( check( x, inasegment.width() - 1 ) )
{
if( check( p_y, inasegment.height() - 1 ) )
{
inasegment.setPixelValue<int>( _value, x, p_y, 0 );
}
if( check( y, inasegment.height() - 1 ) )
{
inasegment.setPixelValue<int>( _value, x, y, 0 );
}
}
p_x = x;
p_y = y;
}
#endif
}
// suggest new feature point for tracking (count point are kept)
bool detect
(
const image::Image<unsigned char> & ima,
std::vector<features::PointFeature> & pt_to_track,
const size_t count
) const override
{
cv::Mat current_img;
cv::eigen2cv(ima.GetMat(), current_img);
std::vector<cv::KeyPoint> m_nextKeypoints;
cv::Ptr<cv::FeatureDetector> m_detector = cv::GFTTDetector::create(count);
if (m_detector == NULL)
return false;
m_detector->detect(current_img, m_nextKeypoints);
if (m_nextKeypoints.size() >= count)
{
// shuffle to avoid to sample only in one bucket
std::mt19937 gen(std::mt19937::default_seed);
std::shuffle(m_nextKeypoints.begin(), m_nextKeypoints.end(), gen);
}
const size_t kept_kp_count = std::min(m_nextKeypoints.size(), count);
m_nextKeypoints.resize(kept_kp_count);
pt_to_track.resize(kept_kp_count);
for (size_t i = 0; i < kept_kp_count; ++i)
pt_to_track[i] = features::PointFeature(m_nextKeypoints[i].pt.x, m_nextKeypoints[i].pt.y);
return kept_kp_count != 0;
// Return false if no point can be added
}
/// Try to track current point set in the provided image
/// return false when tracking failed (=> to send frame to relocalization)
bool track
(
const image::Image<unsigned char> & ima,
const std::vector<features::PointFeature> & pt_to_track,
std::vector<features::PointFeature> & pt_tracked,
std::vector<bool> & status
) override
{
cv::eigen2cv(ima.GetMat(), current_img_);
if (!pt_to_track.empty())
{
prevPts_.resize(pt_to_track.size());
nextPts_.resize(pt_to_track.size());
for (size_t i=0; i < pt_to_track.size(); ++i)
{
prevPts_[i].x = pt_to_track[i].x();
prevPts_[i].y = pt_to_track[i].y();
}
std::vector<unsigned char> status_uchar;
cv::calcOpticalFlowPyrLK(prev_img_, current_img_, prevPts_, nextPts_, status_uchar, error_);
status.assign(status_uchar.begin(), status_uchar.end());
for (size_t i=0; i < nextPts_.size(); ++i)
{
pt_tracked[i].coords() << nextPts_[i].x, nextPts_[i].y;
}
}
// swap frame for next tracking iteration
current_img_.copyTo(prev_img_);
const size_t tracked_point_count = std::accumulate(status.begin(), status.end(), 0);
return (tracked_point_count != 0);
}
/**
* Fill mask from corresponding points (each point pictured by a disk of radius _radius)
*
* \param[out] maskLeft Mask of the left image (initialized to corresponding image size).
* \param[out] maskRight Mask of the right image (initialized to corresponding image size).
*
* \return True if some pixel have been set to true.
*/
virtual bool computeMask( image::Image< unsigned char > & maskLeft, image::Image< unsigned char > & maskRight )
{
maskLeft.fill(0);
maskRight.fill(0);
for( std::vector< matching::IndMatch >::const_iterator
iter_putativeMatches = _vec_PutativeMatches.begin();
iter_putativeMatches != _vec_PutativeMatches.end();
++iter_putativeMatches )
{
const features::SIOPointFeature & L = _vec_featsL[ iter_putativeMatches->i_ ];
const features::SIOPointFeature & R = _vec_featsR[ iter_putativeMatches->j_ ];
image::FilledCircle( L.x(), L.y(), ( int )_radius, ( unsigned char ) 255, &maskLeft );
image::FilledCircle( R.x(), R.y(), ( int )_radius, ( unsigned char ) 255, &maskRight );
}
return _vec_PutativeMatches.size() > 0;
}
double Zncc::compute(image::Image const& im1, image::Image const& im2, float const* weightMatrix)
{
JFR_PRECOND(im1.depth() == im2.depth(), "The depth of both images is different");
switch(im1.depth())
{
// case CV_1U:
// if (weightMatrix == NULL)
// return computeTpl<CV_1U, bool,bool,0,1,true,false>(im1,im2);
// else
// return computeTpl<CV_1U, bool,bool,0,1,true,true>(im1,im2,weightMatrix);
case CV_8U:
if (weightMatrix == NULL)
return computeTpl<CV_8U, uint8_t,uint8_t,0,255,true,false>(im1,im2);
else
return computeTpl<CV_8U, uint8_t,uint8_t,0,255,true,true>(im1,im2,weightMatrix);
case CV_8S:
if (weightMatrix == NULL)
return computeTpl<CV_8S, int8_t,int8_t, -128,127,true,false>(im1,im2);
else
return computeTpl<CV_8S, int8_t,int8_t, -128,127,true,true>(im1,im2,weightMatrix);
case CV_16U:
if (weightMatrix == NULL)
return computeTpl<CV_16U, uint16_t,uint16_t, 0,65535,true,false>(im1,im2);
else
return computeTpl<CV_16U, uint16_t,uint16_t, 0,65535,true,true>(im1,im2,weightMatrix);
case CV_16S:
if (weightMatrix == NULL)
return computeTpl<CV_16S, int16_t,int16_t, -32768,32767,true,false>(im1,im2);
else
return computeTpl<CV_16S, int16_t,int16_t, -32768,32767,true,true>(im1,im2,weightMatrix);
case CV_32F:
if (weightMatrix == NULL) // bool and no borne because cannot use a float as a template parameter, and anyway would be useless here
return computeTpl<CV_32F, float,bool, 0,0,false,false>(im1,im2);
else
return computeTpl<CV_32F, float,bool, 0,0,false,true>(im1,im2,weightMatrix);
case CV_64F:
if (weightMatrix == NULL) // bool and no borne because cannot use a float as a template parameter, and anyway would be useless here
return computeTpl<CV_64F, double,bool, 0,0,false,false>(im1,im2);
else
return computeTpl<CV_64F, double,bool, 0,0,false,true>(im1,im2,weightMatrix);
default:
JFR_PRECOND(false, "Unknown image depth");
return FP_NAN;
}
}
/**
@brief Detect regions on the image and compute their attributes (description)
@param image Image.
@param regions The detected regions and attributes (the caller must delete the allocated data)
@param mask 8-bit gray image for keypoint filtering (optional).
Non-zero values depict the region of interest.
*/
bool Describe(const image::Image<unsigned char>& image,
std::unique_ptr<Regions> ®ions,
const image::Image<unsigned char> * mask = nullptr)
{
// Convert for opencv
cv::Mat img;
cv::eigen2cv(image.GetMat(), img);
// Convert mask image into cv::Mat
cv::Mat m_mask;
if(mask != nullptr) {
cv::eigen2cv(mask->GetMat(), m_mask);
}
// Create a SIFT detector
std::vector< cv::KeyPoint > v_keypoints;
cv::Mat m_desc;
cv::Ptr<cv::Feature2D> siftdetector = cv::xfeatures2d::SIFT::create();
// Process SIFT computation
siftdetector->detectAndCompute(img, m_mask, v_keypoints, m_desc);
Allocate(regions);
// Build alias to cached data
SIFT_Regions * regionsCasted = dynamic_cast<SIFT_Regions*>(regions.get());
// reserve some memory for faster keypoint saving
regionsCasted->Features().reserve(v_keypoints.size());
regionsCasted->Descriptors().reserve(v_keypoints.size());
// Prepare a column vector with the sum of each descriptor
cv::Mat m_siftsum;
cv::reduce(m_desc, m_siftsum, 1, cv::REDUCE_SUM);
// Copy keypoints and descriptors in the regions
int cpt = 0;
for(std::vector< cv::KeyPoint >::const_iterator i_kp = v_keypoints.begin();
i_kp != v_keypoints.end();
++i_kp, ++cpt)
{
SIOPointFeature feat((*i_kp).pt.x, (*i_kp).pt.y, (*i_kp).size, (*i_kp).angle);
regionsCasted->Features().push_back(feat);
Descriptor<unsigned char, 128> desc;
for(int j = 0; j < 128; j++)
{
desc[j] = static_cast<unsigned char>(512.0*sqrt(m_desc.at<float>(cpt, j)/m_siftsum.at<float>(cpt, 0)));
}
regionsCasted->Descriptors().push_back(desc);
}
return true;
};
void snapshot(unsigned call_no) {
if (!drawable ||
(!snapshot_prefix && !compare_prefix)) {
return;
}
Image::Image *ref = NULL;
if (compare_prefix) {
char filename[PATH_MAX];
snprintf(filename, sizeof filename, "%s%010u.png", compare_prefix, call_no);
ref = Image::readPNG(filename);
if (!ref) {
return;
}
if (retrace::verbosity >= 0) {
std::cout << "Read " << filename << "\n";
}
}
Image::Image *src = glstate::getDrawBufferImage(GL_RGBA);
if (!src) {
return;
}
if (snapshot_prefix) {
char filename[PATH_MAX];
snprintf(filename, sizeof filename, "%s%010u.png", snapshot_prefix, call_no);
if (src->writePNG(filename) && retrace::verbosity >= 0) {
std::cout << "Wrote " << filename << "\n";
}
}
if (ref) {
std::cout << "Snapshot " << call_no << " average precision of " << src->compare(*ref) << " bits\n";
delete ref;
}
delete src;
}
int main( int argc , char** argv ){
#ifdef DEBUG
MEM_ON();
TRACE_OFF();
#endif
//total execution timer
Timer totalTimer;
totalTimer.start();
std::cout<<"****************************************************************"<<std::endl;
std::cout<<"* OpenMP execution with "<<omp_get_max_threads()<<" threads *"<<std::endl;
std::cout<<"****************************************************************"<<std::endl;
std::cout<<"\n\n\n";
std::vector<std::string> imageName;
std::stringstream input(argv[4]);
double factor;
input >> factor;
int succeded = 0;
int failed = 0;
std::string operation( argv[3] );
//parallel code timer
Timer parallelTimer;
//how many images to run in parallel.Number of threads created for the program
unsigned int parallelImages = omp_get_max_threads();
//counter for total files for parallel iterations
unsigned int counter = 0;
//how many iterations to run in parallel
unsigned int parallelIterations = parallelImages;
if( GetDirFileNames ( argv[1] , imageName ) )
try {
parallelTimer.start();
for( std::vector<std::string>::iterator it = imageName.begin() ; it < imageName.end() ; it += parallelImages ){
counter += parallelImages;
parallelIterations = parallelImages;
if( counter > imageName.size() )
parallelIterations = imageName.size() - ( counter - parallelImages );
#pragma omp parallel for
for( unsigned int i = 0 ; i < parallelIterations ; i++ ) {
std::cout<<(*(it + i))<<"\n";
IMAGE::Image* oldImage = NULL;
IMAGE::Image* newImage = NULL;
std::string oldName = argv[1] + (*(it + i));
std::string newName = argv[2] + (*(it + i));
try{
oldImage= IMAGE::Image::createInstance( oldName );
newImage = IMAGE::Image::createInstance( newName );
////////////////////
try{
oldImage->open( oldName, 'r' );
newImage->open( newName , 'w' );
oldImage->readImageRaster();
newImage->raster.createRaster( oldImage->raster );
//check which operation to do
if( operation == REVERSE ){
IMAGE::PROCESS::reverseColor( newImage->raster );
}
else if( operation == BRIGHTNESS ) {
IMAGE::PROCESS::adjustBrightness( newImage->raster , atoi( argv[4] ));
}
else if( operation == CONTRAST ) {
IMAGE::PROCESS::adjustContrast( newImage->raster , atoi( argv[4] ) );
}
else if( operation == RGB2GREY ) {
IMAGE::FILTERS::convertRGB2GREY( newImage->raster , atoi( argv[4] ));
}
else if( operation == RGB2BW ) {
IMAGE::FILTERS::convertRGB2BW( newImage->raster );
}
else if( operation == RGB2SEPIA ) {
IMAGE::FILTERS::convertRGB2SEPIA( newImage->raster );
}
else if( operation == BLUR ) {
IMAGE::PROCESS::blurImage( newImage->raster , atoi(argv[4]) );
}
else if( operation == ROTATE ) {
IMAGE::PROCESS::rotateImage( newImage->raster , argv[4] ) ;
}
else if( operation == ZOOM ) {
IMAGE::PROCESS::zoomImage( newImage->raster , 1 , 1 , 400 , 300 );
}
else if( operation == SCALE ) {
IMAGE::PROCESS::scaleImage( newImage->raster , factor );
}
else {
std::cout<<"Not a valid operation \n";
exit(0);
}
newImage->writeRasterToImage();
succeded++;
}
catch( IMAGE::file_io_failed& e ) {
std::cout<<e.what()<<"\n";
failed++;
}
catch ( IMAGE::image_format_error& e ) {
std::cout<<e.what()<<"\n";
failed++;
}
catch( IMAGE::empty_image& e ){
std::cout<<e.what()<<"\n";
failed++;
}
catch ( IMAGE::empty_raster& e ) {
std::cout<<e.what()<<"\n";
failed++;
}
oldImage->close();
newImage->close();
delete oldImage;
delete newImage;
}
catch( IMAGE::not_supported_format& e ) {
std::cout<<e.what()<<"\n";
failed++;
}
}//end of omp_parallel_for
}//end of for
}
catch( IMAGE::bad_alloc& e ){
std::cout<<e.what()<<std::endl;
std::cout<<"Exiting program...\n";
exit(1);
}
catch(...){
std::cout<<"cought unexpected exception...\n";
std::cout<<"Exitig program...\n";
exit(1);
}
totalTimer.stop();
parallelTimer.stop();
//final output //////////////////////////////////////////////////////////////
std::cout<<"\n\n\n";
std::cout<<"*************************************************************"<<std::endl;
std::cout<<"* Results *"<<std::endl;
std::cout<<"* total files : "<<imageName.size()<<" *"<<std::endl;
std::cout<<"* Succeded images : "<<succeded<<" *"<<std::endl;
std::cout<<"* Failed images : "<<failed<<" *"<<std::endl;
std::cout<<"* *"<<std::endl;
std::cout<<"* Total execution time : "<<totalTimer.getReal()<<" sec *"<<std::endl;
std::cout<<"* Parallel partial execution time : "<<parallelTimer.getReal()<<" sec *"<<std::endl;
std::cout<<"* *"<<std::endl;
std::cout<<"* END OF PROGRAM *"<<std::endl;
std::cout<<"*************************************************************"<<std::endl;
///////////////////////////////////////////////////////////////////////////////////////
#ifdef DEBUG
MEM_OFF();
#endif
return 0;
}
double Zncc::compute8noborne(image::Image const& im1, image::Image const& im2)
{
JFR_PRECOND(im1.depth() == im2.depth(), "The depth of both images is different");
JFR_PRECOND(im1.depth() == CV_8U, "The depth of images must be CV_8U");
return computeTpl<CV_8U, uint8_t,uint8_t,0,255,false,false>(im1,im2);
}
/**
* Put masks to white, images are conserved
*
* \param[out] maskLeft Mask of the left image (initialized to corresponding image size).
* \param[out] maskRight Mask of the right image (initialized to corresponding image size).
*
* \return True.
*/
virtual bool computeMask(
image::Image< unsigned char > & maskLeft,
image::Image< unsigned char > & maskRight )
{
std::vector< matching::IndMatch > vec_KVLDMatches;
image::Image< unsigned char > imageL, imageR;
image::ReadImage( _sLeftImage.c_str(), &imageL );
image::ReadImage( _sRightImage.c_str(), &imageR );
image::Image< float > imgA ( imageL.GetMat().cast< float >() );
image::Image< float > imgB(imageR.GetMat().cast< float >());
std::vector< Pair > matchesFiltered, matchesPair;
for( std::vector< matching::IndMatch >::const_iterator iter_match = _vec_PutativeMatches.begin();
iter_match != _vec_PutativeMatches.end();
++iter_match )
{
matchesPair.push_back( std::make_pair( iter_match->i_, iter_match->j_ ) );
}
std::vector< double > vec_score;
//In order to illustrate the gvld(or vld)-consistant neighbors, the following two parameters has been externalized as inputs of the function KVLD.
openMVG::Mat E = openMVG::Mat::Ones( _vec_PutativeMatches.size(), _vec_PutativeMatches.size() ) * ( -1 );
// gvld-consistancy matrix, intitialized to -1, >0 consistancy value, -1=unknow, -2=false
std::vector< bool > valide( _vec_PutativeMatches.size(), true );// indices of match in the initial matches, if true at the end of KVLD, a match is kept.
size_t it_num = 0;
KvldParameters kvldparameters;//initial parameters of KVLD
//kvldparameters.K = 5;
while (
it_num < 5 &&
kvldparameters.inlierRate >
KVLD(
imgA, imgB,
_vec_featsL, _vec_featsR,
matchesPair, matchesFiltered,
vec_score, E, valide, kvldparameters ) )
{
kvldparameters.inlierRate /= 2;
std::cout<<"low inlier rate, re-select matches with new rate="<<kvldparameters.inlierRate<<std::endl;
kvldparameters.K = 2;
it_num++;
}
bool bOk = false;
if( !matchesPair.empty())
{
// Get mask
getKVLDMask(
&maskLeft, &maskRight,
_vec_featsL, _vec_featsR,
matchesPair,
valide,
E);
bOk = true;
}
else{
maskLeft.fill( 0 );
maskRight.fill( 0 );
}
return bOk;
}
/**
* @brief Extract MSER regions
* @param img Input image
* @param[out] regions Output regions
*/
void MSERExtractor::Extract( const image::Image<unsigned char> & img , std::vector<MSERRegion> & regions ) const
{
// Compute minimum and maximum region area relative to this image
const int minRegArea = img.Width() * img.Height() * m_minimum_area;
const int maxRegArea = img.Width() * img.Height() * m_maximum_area;
// List of processed pixels (maybe we can use a more efficient structure)
std::vector<std::vector<bool >> processed;
processed.resize( img.Width() );
for (int i = 0; i < img.Width(); ++i )
{
processed[ i ].resize( img.Height() );
std::fill( processed[ i ].begin() , processed[ i ].end() , false );
}
// Holds the boundary of given grayscale value (boundary[0] -> pixels in the boundary with 0 grayscale value)
std::vector<PixelStackElt> boundary[ 256 ];
// List of regions computed so far (not only valid MSER regions)
std::vector<MSERRegion *> regionStack;
// Push en empty region
regionStack.push_back( new MSERRegion );
// Start processing from top left pixel
PixelStackElt cur_pix;
cur_pix.pix_x = 0;
cur_pix.pix_y = 0;
cur_pix.pix_level = img( 0 , 0 );
cur_pix.edge_index = PIXEL_RIGHT;
processed[ cur_pix.pix_x ][ cur_pix.pix_y ] = true;
regionStack.push_back( new MSERRegion( cur_pix.pix_level , cur_pix.pix_x , cur_pix.pix_y ) );
int priority = 256;
// Start process
while (1)
{
bool restart = false;
// Process neighboring to see if there's something to search with lower grayscale level
for ( PixelNeighborsDirection curDir = cur_pix.edge_index;
curDir <= PIXEL_BOTTOM_RIGHT;
curDir = NextDirection( curDir , m_connectivity ) )
{
int nx , ny;
GetNeighbor( cur_pix.pix_x , cur_pix.pix_y , curDir , img.Width() , img.Height() , nx , ny );
// Pixel was not processed before
if (ValidPixel( nx , ny , img.Width() , img.Height() ) && ! processed[ nx ][ ny ] )
{
const int nLevel = img( ny , nx );
processed[ nx ][ ny ] = true;
// Info of the neighboring pixel
PixelStackElt n_elt;
n_elt.pix_x = nx;
n_elt.pix_y = ny;
n_elt.pix_level = nLevel;
n_elt.edge_index = PIXEL_RIGHT;
// Now look from which pixel do we have to continue
if (nLevel >= cur_pix.pix_level )
{
// Continue from the same pixel
boundary[ nLevel ].push_back( n_elt );
// Store the lowest value so far
priority = std::min( nLevel , priority );
}
else
{
// Go on with the neighboring pixel (go down)
cur_pix.edge_index = NextDirection( curDir , m_connectivity ); // Next time we have to process the next boundary pixel
boundary[ cur_pix.pix_level ].push_back( cur_pix );
// Store the lowest value so far
priority = std::min( cur_pix.pix_level , priority );
// Push the next pixel to process
cur_pix = n_elt;
restart = true;
break;
}
}
}
// Do we have to restart from a new pixel ?
if (restart )
{
// If so it's that because we found a lower grayscale value so let's start a new region
regionStack.push_back( new MSERRegion( cur_pix.pix_level , cur_pix.pix_x , cur_pix.pix_y ) );
continue;
}
// We have process all the neighboring pixels, current pixel is the lowest we have found so far
// now process the current pixel
regionStack.back()->AppendPixel( cur_pix.pix_x , cur_pix.pix_y );
// End of the process : we have no boundary region, compute MSER from graph
if (priority == 256 )
{
regionStack.back()->ComputeMSER( m_delta , minRegArea , maxRegArea , m_max_variation , m_min_diversity , regions );
break;
}
PixelStackElt next_pix = boundary[ priority ].back();
boundary[ priority ].pop_back();
// Get the next pixel level
while (boundary[ priority ].empty() && ( priority < 256 ))
{
++priority;
}
// Clear the stack
const int newLevel = next_pix.pix_level;
// Process the current stack of pixels if the next processing pixel is not at the same curent level
if (newLevel != cur_pix.pix_level )
{
// Try to merge the regions to fomr a tree
ProcessStack( newLevel , next_pix.pix_x , next_pix.pix_y , regionStack );
}
// Update next pixel for processing
cur_pix = next_pix;
}
// Clear region stack created so far
for (size_t i = 0; i < regionStack.size(); ++i )
{
delete regionStack[ i ];
}
}
double Explorer<Correl>::exploreTranslation(image::Image const& im1, image::Image const& im2_, int xmin, int xmax, int xstep, int ymin, int ymax, int ystep, double &xres, double &yres, float const* weightMatrix)
{
cv::Rect roi = im1.getROI();
// image::Image im2(im2_, cv::Rect(0,0,im2_.width(),im2_.height()));
image::Image im2(im2_);
double score;
double best_score = -1.;
int bestx = -1, besty = -1;
if (xmin < 0) xmin = 0; if (xmax >= im2.width ()) xmax = im2.width ()-1;
if (ymin < 0) ymin = 0; if (ymax >= im2.height()) ymax = im2.height()-1;
int sa_w = (xmax-xmin+1), sa_h = (ymax-ymin+1); // search area
if (sa_w < 5) xstep = 1; if (sa_h < 5) ystep = 1;
int nresults = (sa_w+2)*(sa_h+2);
double *results = new double[nresults]; // add 1 border for interpolation
for(int i = 0; i < nresults; i++) results[i] = -1e6;
// explore
for(int y = ymin; y <= ymax; y += ystep)
for(int x = xmin; x <= xmax; x += xstep)
DO_CORRELATION(im1, im2, weightMatrix, x, y, score, best_score, bestx, besty, roi);
// refine
// JFR_DEBUG("refine (" << bestx << "," << besty << " " << best_score << ")");
// TODO refine several local maxima
// TODO refine by dichotomy for large steps ?
int newbestx = bestx, newbesty = besty;
for(int y = besty-ystep+1; y <= besty+ystep-1; y++)
for(int x = bestx-xstep+1; x <= bestx+xstep-1; x++)
{
if (x == bestx && y == besty) continue;
DO_CORRELATION(im1, im2, weightMatrix, x, y, score, best_score, newbestx, newbesty, roi);
}
// ensure that all values that will be used by interpolation are computed
int newnewbestx = newbestx, newnewbesty = newbesty;
/* if (((newbestx == bestx-xstep+1 || newbestx == bestx+xstep-1) && (newbesty-ymin)%ystep) ||
((newbesty == besty-ystep+1 || newbesty == besty+ystep-1) && (newbestx-xmin)%xstep))
{
if (newbestx == bestx-xstep+1) DO_CORRELATION(im1, im2, weightMatrix, newbestx-1, newbesty, score, best_score, newnewbestx, newnewbesty, roi);
if (newbestx == bestx+xstep-1) DO_CORRELATION(im1, im2, weightMatrix, newbestx+1, newbesty, score, best_score, newnewbestx, newnewbesty, roi);
if (newbesty == besty-ystep+1) DO_CORRELATION(im1, im2, weightMatrix, newbestx, newbesty-1, score, best_score, newnewbestx, newnewbesty, roi);
if (newbesty == besty+ystep-1) DO_CORRELATION(im1, im2, weightMatrix, newbestx, newbesty+1, score, best_score, newnewbestx, newnewbesty, roi);
}*/
// JFR_DEBUG("extra interpol (" << newbestx << "," << newbesty << " " << best_score << ")");
do {
newbestx = newnewbestx, newbesty = newnewbesty;
if (newbestx>0 && RESULTS(newbesty,newbestx-1)<-1e5)
DO_CORRELATION(im1, im2, weightMatrix, newbestx-1, newbesty, score, best_score, newnewbestx, newnewbesty, roi);
if (newbestx<im2.width()-1 && RESULTS(newbesty,newbestx+1)<-1e5)
DO_CORRELATION(im1, im2, weightMatrix, newbestx+1, newbesty, score, best_score, newnewbestx, newnewbesty, roi);
if (newbesty>0 && RESULTS(newbesty-1,newbestx)<-1e5)
DO_CORRELATION(im1, im2, weightMatrix, newbestx, newbesty-1, score, best_score, newnewbestx, newnewbesty, roi);
if (newbesty<im2.height()-1 && RESULTS(newbesty+1,newbestx)<-1e5)
DO_CORRELATION(im1, im2, weightMatrix, newbestx, newbesty+1, score, best_score, newnewbestx, newnewbesty, roi);
} while (newbestx != newnewbestx || newbesty != newnewbesty);
// FIXME this could go out of bounds
// JFR_DEBUG("final : " << newnewbestx << "," << newnewbesty << " " << best_score);
bestx = newbestx;
besty = newbesty;
// TODO interpolate the score as well
// interpolate x
double a1 = RESULTS(besty,bestx-1), a2 = RESULTS(besty,bestx-0), a3 = RESULTS(besty,bestx+1);
if (a1 > -1e5 && a3 > -1e5) jmath::parabolicInterpolation(a1,a2,a3, xres); else xres = 0;
// JFR_DEBUG("interpolating " << a1 << " " << a2 << " " << a3 << " gives shift " << xres << " plus " << bestx+0.5);
xres += bestx+0.5;
// interpolate y
a1 = RESULTS(besty-1,bestx), a2 = RESULTS(besty-0,bestx), a3 = RESULTS(besty+1,bestx);
if (a1 > -1e5 && a3 > -1e5) jmath::parabolicInterpolation(a1,a2,a3, yres); else yres = 0;
// JFR_DEBUG("interpolating " << a1 << " " << a2 << " " << a3 << " gives shift " << yres << " plus " << besty+0.5);
yres += besty+0.5;
delete[] results;
return best_score;
}
double Zncc::computeTpl(image::Image const& im1_, image::Image const& im2_, float const* weightMatrix)
{
// preconds
JFR_PRECOND( im1_.depth() == depth, "Image 1 depth is different from the template parameter" );
JFR_PRECOND( im2_.depth() == depth, "Image 2 depth is different from the template parameter" );
JFR_PRECOND( im1_.channels() == im2_.channels(), "The channels number of both images are different" );
JFR_PRECOND( !useWeightMatrix || weightMatrix, "Template parameter tells to use weightMatrix but no one is given" );
// adjust ROIs to match size, assuming that it is reduced when set out of the image
// FIXME weightMatrix should be a cv::Mat in order to have a ROI too, and to adjust it
cv::Size size1; cv::Rect roi1 = im1_.getROI(size1);
cv::Size size2; cv::Rect roi2 = im2_.getROI(size2);
int dw = roi1.width - roi2.width, dh = roi1.height - roi2.height;
if (dw != 0)
{
cv::Rect &roiA = (dw<0 ? roi1 : roi2), &roiB = (dw<0 ? roi2 : roi1);
cv::Size &sizeA = (dw<0 ? size1 : size2);
if (roiA.x == 0) { roiB.x += dw; roiB.width -= dw; } else
if (roiA.x+roiA.width == sizeA.width) { roiB.width -= dw; }
}
if (dh != 0)
{
cv::Rect &roiA = (dh<0 ? roi1 : roi2), &roiB = (dh<0 ? roi2 : roi1);
cv::Size &sizeA = (dh<0 ? size1 : size2);
if (roiA.y == 0) { roiB.y += dh; roiB.height -= dh; } else
if (roiA.y+roiA.height == sizeA.height) { roiB.height -= dh; }
}
image::Image im1(im1_); im1.setROI(roi1);
image::Image im2(im2_); im2.setROI(roi2);
// some variables initialization
int height = im1.height();
int width = im1.width();
int step1 = im1.step1() - width;
int step2 = im2.step1() - width;
double mean1 = 0., mean2 = 0.;
double sigma1 = 0., sigma2 = 0., sigma12 = 0.;
double zncc_sum = 0.;
double zncc_count = 0.;
double zncc_total = 0.;
worktype const* im1ptr = reinterpret_cast<worktype const*>(im1.data());
worktype const* im2ptr = reinterpret_cast<worktype const*>(im2.data());
float const* wptr = weightMatrix;
double w;
// start the loops
for(int i = 0; i < height; ++i)
{
for(int j = 0; j < width; ++j)
{
worktype im1v = *(im1ptr++);
worktype im2v = *(im2ptr++);
if (useWeightMatrix) w = *(wptr++); else w = 1;
if (useBornes) zncc_total += w;
//std::cout << "will correl ? " << useBornes << ", " << (int)im1v << ", " << (int)im2v << std::endl;
if (!useBornes || (im1v != borneinf && im1v != bornesup && im2v != borneinf && im2v != bornesup))
{
//std::cout << "correl one pixel" << std::endl;
#if 0
double im1vw, im2vw;
if (useWeightMatrix)
{ im1vw = im1v * w; im2vw = im2v * w; } else
{ im1vw = im1v; im2vw = im2v; }
zncc_count += w;
mean1 += im1vw;
mean2 += im2vw;
sigma1 += im1v * im1vw;
sigma2 += im2v * im2vw;
zncc_sum += im1v * im2vw;
#else
zncc_count += w;
mean1 += im1v * w;
mean2 += im2v * w;
sigma1 += im1v * im1v * w;
sigma2 += im2v * im2v * w;
zncc_sum += im1v * im2v * w;
#endif
}
}
im1ptr += step1;
im2ptr += step2;
}
if (useBornes) if (zncc_count / zncc_total < 0.75)
{ /*std::cout << "zncc failed: " << zncc_count << "," << zncc_total << std::endl;*/ return -3; }
// finish
mean1 /= zncc_count;
mean2 /= zncc_count;
sigma1 = sigma1/zncc_count - mean1*mean1;
sigma2 = sigma2/zncc_count - mean2*mean2;
sigma1 = sigma1 > 0.0 ? sqrt(sigma1) : 0.0; // test for numerical rounding errors to avoid nan
sigma2 = sigma2 > 0.0 ? sqrt(sigma2) : 0.0;
sigma12 = sigma1*sigma2;
// std::cout << "normal: zncc_sum " << zncc_sum << ", count " << zncc_count << ", mean12 " << mean1*mean2 << ", sigma12 " << sigma1*sigma2 << std::endl;
zncc_sum = (sigma12 < 1e-6 ? -1 : (zncc_sum/zncc_count - mean1*mean2) / sigma12);
JFR_ASSERT(zncc_sum >= -1.01, "");
return zncc_sum;
}
/**
@brief Detect regions on the image and compute their attributes (description)
@param image Image.
@param regions The detected regions and attributes (the caller must delete the allocated data)
@param mask 8-bit gray image for keypoint filtering (optional).
Non-zero values depict the region of interest.
*/
bool Describe
(
const image::Image<unsigned char>& image,
std::unique_ptr<Regions> ®ions,
const image::Image<unsigned char> * mask = nullptr
) override
{
const int w = image.Width(), h = image.Height();
// Convert to float in range [0;1]
const image::Image<float> If(image.GetMat().cast<float>()/255.0f);
// compute sift keypoints
Allocate(regions);
// Build alias to cached data
SIFT_Regions * regionsCasted = dynamic_cast<SIFT_Regions*>(regions.get());
{
using namespace openMVG::features::sift;
const int supplementary_images = 3;
// => in order to ensure each gaussian slice is used in the process 3 extra images are required:
// +1 for dog computation
// +2 for 3d discrete extrema definition
HierarchicalGaussianScaleSpace octave_gen(
params_.num_octaves_,
params_.num_scales_,
(params_.first_octave_ == -1)
? GaussianScaleSpaceParams(1.6f/2.0f, 1.0f/2.0f, 0.5f, supplementary_images)
: GaussianScaleSpaceParams(1.6f, 1.0f, 0.5f, supplementary_images));
octave_gen.SetImage( If );
std::vector<Keypoint> keypoints;
keypoints.reserve(5000);
Octave octave;
while ( octave_gen.NextOctave( octave ) )
{
std::vector< Keypoint > keys;
// Find Keypoints
SIFT_KeypointExtractor keypointDetector(
params_.peak_threshold_ / octave_gen.NbSlice(),
params_.edge_threshold_);
keypointDetector(octave, keys);
// Find Keypoints orientation and compute their description
Sift_DescriptorExtractor descriptorExtractor;
descriptorExtractor(octave, keys);
// Concatenate the found keypoints
std::move(keys.begin(), keys.end(), std::back_inserter(keypoints));
}
for (const auto & k : keypoints)
{
// Feature masking
if (mask)
{
const image::Image<unsigned char> & maskIma = *mask;
if (maskIma(k.y, k.x) == 0)
continue;
}
Descriptor<unsigned char, 128> descriptor;
descriptor << (k.descr.cast<unsigned char>());
{
regionsCasted->Descriptors().emplace_back(descriptor);
regionsCasted->Features().emplace_back(k.x, k.y, k.sigma, k.theta);
}
}
}
return true;
};
/**
* Put masks to white, all image is considered as valid pixel selection
*
* \param[out] maskLeft Mask of the left image (initialized to corresponding image size).
* \param[out] maskRight Mask of the right image (initialized to corresponding image size).
*
* \return True.
*/
bool computeMask( image::Image< unsigned char > & maskLeft, image::Image< unsigned char > & maskRight ) override
{
maskLeft.fill( image::WHITE );
maskRight.fill( image::WHITE );
return true;
}
/**
@brief Detect regions on the image and compute their attributes (description)
@param image Image.
@param regions The detected regions and attributes (the caller must delete the allocated data)
@param mask 8-bit gray image for keypoint filtering (optional).
Non-zero values depict the region of interest.
*/
bool Describe(const image::Image<unsigned char>& image,
std::unique_ptr<Regions> ®ions,
const image::Image<unsigned char> * mask = NULL)
{
const int w = image.Width(), h = image.Height();
//Convert to float
const image::Image<float> If(image.GetMat().cast<float>());
VlSiftFilt *filt = vl_sift_new(w, h,
_params._num_octaves, _params._num_scales, _params._first_octave);
if (_params._edge_threshold >= 0)
vl_sift_set_edge_thresh(filt, _params._edge_threshold);
if (_params._peak_threshold >= 0)
vl_sift_set_peak_thresh(filt, 255*_params._peak_threshold/_params._num_scales);
Descriptor<vl_sift_pix, 128> descr;
Descriptor<unsigned char, 128> descriptor;
// Process SIFT computation
vl_sift_process_first_octave(filt, If.data());
Allocate(regions);
// Build alias to cached data
SIFT_Regions * regionsCasted = dynamic_cast<SIFT_Regions*>(regions.get());
// reserve some memory for faster keypoint saving
regionsCasted->Features().reserve(2000);
regionsCasted->Descriptors().reserve(2000);
while (true) {
vl_sift_detect(filt);
VlSiftKeypoint const *keys = vl_sift_get_keypoints(filt);
const int nkeys = vl_sift_get_nkeypoints(filt);
// Update gradient before launching parallel extraction
vl_sift_update_gradient(filt);
#ifdef OPENMVG_USE_OPENMP
#pragma omp parallel for private(descr, descriptor)
#endif
for (int i = 0; i < nkeys; ++i) {
// Feature masking
if (mask)
{
const image::Image<unsigned char> & maskIma = *mask;
if (maskIma(keys[i].y, keys[i].x) == 0)
continue;
}
double angles [4] = {0.0, 0.0, 0.0, 0.0};
int nangles = 1; // by default (1 upright feature)
if (_bOrientation)
{ // compute from 1 to 4 orientations
nangles = vl_sift_calc_keypoint_orientations(filt, angles, keys+i);
}
for (int q=0 ; q < nangles ; ++q) {
vl_sift_calc_keypoint_descriptor(filt, &descr[0], keys+i, angles[q]);
const SIOPointFeature fp(keys[i].x, keys[i].y,
keys[i].sigma, static_cast<float>(angles[q]));
siftDescToUChar(&descr[0], descriptor, _params._root_sift);
#ifdef OPENMVG_USE_OPENMP
#pragma omp critical
#endif
{
regionsCasted->Descriptors().push_back(descriptor);
regionsCasted->Features().push_back(fp);
}
}
}
if (vl_sift_process_next_octave(filt))
break; // Last octave
}
vl_sift_delete(filt);
return true;
};
