generic-library/opencv
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

OpenCV with the refactored features2d compiles! contrib is broken for now; the tests are not tried yet

Browse Source
This commit is contained in:
Vadim Pisarevsky
2014-10-15 22:49:17 +04:00
parent 2e915026a0
commit 09df1a286b
27 changed files with 1939 additions and 2316 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
modules/features2d/src/kaze/AKAZEConfig.h
View File

@@ -22,12 +22,12 @@ struct AKAZEOptions {
, soffset(1.6f)
, derivative_factor(1.5f)
, sderivatives(1.0)
, diffusivity(cv::DIFF_PM_G2)
, diffusivity(KAZE::DIFF_PM_G2)
, dthreshold(0.001f)
, min_dthreshold(0.00001f)
, descriptor(cv::DESCRIPTOR_MLDB)
, descriptor(AKAZE::DESCRIPTOR_MLDB)
, descriptor_size(0)
, descriptor_channels(3)
, descriptor_pattern_size(10)

203
modules/features2d/src/kaze/AKAZEFeatures.cpp
View File

@@ -61,14 +61,14 @@ void AKAZEFeatures::Allocate_Memory_Evolution(void) {
for (int j = 0; j < options_.nsublevels; j++) {
TEvolution step;
step.Lx = cv::Mat::zeros(level_height, level_width, CV_32F);
step.Ly = cv::Mat::zeros(level_height, level_width, CV_32F);
step.Lxx = cv::Mat::zeros(level_height, level_width, CV_32F);
step.Lxy = cv::Mat::zeros(level_height, level_width, CV_32F);
step.Lyy = cv::Mat::zeros(level_height, level_width, CV_32F);
step.Lt = cv::Mat::zeros(level_height, level_width, CV_32F);
step.Ldet = cv::Mat::zeros(level_height, level_width, CV_32F);
step.Lsmooth = cv::Mat::zeros(level_height, level_width, CV_32F);
step.Lx = Mat::zeros(level_height, level_width, CV_32F);
step.Ly = Mat::zeros(level_height, level_width, CV_32F);
step.Lxx = Mat::zeros(level_height, level_width, CV_32F);
step.Lxy = Mat::zeros(level_height, level_width, CV_32F);
step.Lyy = Mat::zeros(level_height, level_width, CV_32F);
step.Lt = Mat::zeros(level_height, level_width, CV_32F);
step.Ldet = Mat::zeros(level_height, level_width, CV_32F);
step.Lsmooth = Mat::zeros(level_height, level_width, CV_32F);
step.esigma = options_.soffset*pow(2.f, (float)(j) / (float)(options_.nsublevels) + i);
step.sigma_size = fRound(step.esigma);
step.etime = 0.5f*(step.esigma*step.esigma);
@@ -97,7 +97,7 @@ void AKAZEFeatures::Allocate_Memory_Evolution(void) {
* @param img Input image for which the nonlinear scale space needs to be created
* @return 0 if the nonlinear scale space was created successfully, -1 otherwise
*/
int AKAZEFeatures::Create_Nonlinear_Scale_Space(const cv::Mat& img)
int AKAZEFeatures::Create_Nonlinear_Scale_Space(const Mat& img)
{
CV_Assert(evolution_.size() > 0);
@@ -107,8 +107,8 @@ int AKAZEFeatures::Create_Nonlinear_Scale_Space(const cv::Mat& img)
evolution_[0].Lt.copyTo(evolution_[0].Lsmooth);
// Allocate memory for the flow and step images
cv::Mat Lflow = cv::Mat::zeros(evolution_[0].Lt.rows, evolution_[0].Lt.cols, CV_32F);
cv::Mat Lstep = cv::Mat::zeros(evolution_[0].Lt.rows, evolution_[0].Lt.cols, CV_32F);
Mat Lflow = Mat::zeros(evolution_[0].Lt.rows, evolution_[0].Lt.cols, CV_32F);
Mat Lstep = Mat::zeros(evolution_[0].Lt.rows, evolution_[0].Lt.cols, CV_32F);
// First compute the kcontrast factor
options_.kcontrast = compute_k_percentile(img, options_.kcontrast_percentile, 1.0f, options_.kcontrast_nbins, 0, 0);
@@ -121,8 +121,8 @@ int AKAZEFeatures::Create_Nonlinear_Scale_Space(const cv::Mat& img)
options_.kcontrast = options_.kcontrast*0.75f;
// Allocate memory for the resized flow and step images
Lflow = cv::Mat::zeros(evolution_[i].Lt.rows, evolution_[i].Lt.cols, CV_32F);
Lstep = cv::Mat::zeros(evolution_[i].Lt.rows, evolution_[i].Lt.cols, CV_32F);
Lflow = Mat::zeros(evolution_[i].Lt.rows, evolution_[i].Lt.cols, CV_32F);
Lstep = Mat::zeros(evolution_[i].Lt.rows, evolution_[i].Lt.cols, CV_32F);
}
else {
evolution_[i - 1].Lt.copyTo(evolution_[i].Lt);
@@ -136,16 +136,16 @@ int AKAZEFeatures::Create_Nonlinear_Scale_Space(const cv::Mat& img)
// Compute the conductivity equation
switch (options_.diffusivity) {
case cv::DIFF_PM_G1:
case KAZE::DIFF_PM_G1:
pm_g1(evolution_[i].Lx, evolution_[i].Ly, Lflow, options_.kcontrast);
break;
case cv::DIFF_PM_G2:
case KAZE::DIFF_PM_G2:
pm_g2(evolution_[i].Lx, evolution_[i].Ly, Lflow, options_.kcontrast);
break;
case cv::DIFF_WEICKERT:
case KAZE::DIFF_WEICKERT:
weickert_diffusivity(evolution_[i].Lx, evolution_[i].Ly, Lflow, options_.kcontrast);
break;
case cv::DIFF_CHARBONNIER:
case KAZE::DIFF_CHARBONNIER:
charbonnier_diffusivity(evolution_[i].Lx, evolution_[i].Ly, Lflow, options_.kcontrast);
break;
default:
@@ -155,7 +155,7 @@ int AKAZEFeatures::Create_Nonlinear_Scale_Space(const cv::Mat& img)
// Perform FED n inner steps
for (int j = 0; j < nsteps_[i - 1]; j++) {
cv::details::kaze::nld_step_scalar(evolution_[i].Lt, Lflow, Lstep, tsteps_[i - 1][j]);
nld_step_scalar(evolution_[i].Lt, Lflow, Lstep, tsteps_[i - 1][j]);
}
}
@@ -167,7 +167,7 @@ int AKAZEFeatures::Create_Nonlinear_Scale_Space(const cv::Mat& img)
* @brief This method selects interesting keypoints through the nonlinear scale space
* @param kpts Vector of detected keypoints
*/
void AKAZEFeatures::Feature_Detection(std::vector<cv::KeyPoint>& kpts)
void AKAZEFeatures::Feature_Detection(std::vector<KeyPoint>& kpts)
{
kpts.clear();
Compute_Determinant_Hessian_Response();
@@ -176,7 +176,7 @@ void AKAZEFeatures::Feature_Detection(std::vector<cv::KeyPoint>& kpts)
}
/* ************************************************************************* */
class MultiscaleDerivativesAKAZEInvoker : public cv::ParallelLoopBody
class MultiscaleDerivativesAKAZEInvoker : public ParallelLoopBody
{
public:
explicit MultiscaleDerivativesAKAZEInvoker(std::vector<TEvolution>& ev, const AKAZEOptions& opt)
@@ -185,7 +185,7 @@ public:
{
}
void operator()(const cv::Range& range) const
void operator()(const Range& range) const
{
std::vector<TEvolution>& evolution = *evolution_;
@@ -219,7 +219,7 @@ private:
*/
void AKAZEFeatures::Compute_Multiscale_Derivatives(void)
{
cv::parallel_for_(cv::Range(0, (int)evolution_.size()),
parallel_for_(Range(0, (int)evolution_.size()),
MultiscaleDerivativesAKAZEInvoker(evolution_, options_));
}
@@ -253,7 +253,7 @@ void AKAZEFeatures::Compute_Determinant_Hessian_Response(void) {
* @brief This method finds extrema in the nonlinear scale space
* @param kpts Vector of detected keypoints
*/
void AKAZEFeatures::Find_Scale_Space_Extrema(std::vector<cv::KeyPoint>& kpts)
void AKAZEFeatures::Find_Scale_Space_Extrema(std::vector<KeyPoint>& kpts)
{
float value = 0.0;
@@ -261,8 +261,8 @@ void AKAZEFeatures::Find_Scale_Space_Extrema(std::vector<cv::KeyPoint>& kpts)
int npoints = 0, id_repeated = 0;
int sigma_size_ = 0, left_x = 0, right_x = 0, up_y = 0, down_y = 0;
bool is_extremum = false, is_repeated = false, is_out = false;
cv::KeyPoint point;
vector<cv::KeyPoint> kpts_aux;
KeyPoint point;
vector<KeyPoint> kpts_aux;
// Set maximum size
if (options_.descriptor == AKAZE::DESCRIPTOR_MLDB_UPRIGHT || options_.descriptor == AKAZE::DESCRIPTOR_MLDB) {
@@ -365,7 +365,7 @@ void AKAZEFeatures::Find_Scale_Space_Extrema(std::vector<cv::KeyPoint>& kpts)
for (size_t i = 0; i < kpts_aux.size(); i++) {
is_repeated = false;
const cv::KeyPoint& pt = kpts_aux[i];
const KeyPoint& pt = kpts_aux[i];
for (size_t j = i + 1; j < kpts_aux.size(); j++) {
// Compare response with the upper scale
@@ -392,7 +392,7 @@ void AKAZEFeatures::Find_Scale_Space_Extrema(std::vector<cv::KeyPoint>& kpts)
* @brief This method performs subpixel refinement of the detected keypoints
* @param kpts Vector of detected keypoints
*/
void AKAZEFeatures::Do_Subpixel_Refinement(std::vector<cv::KeyPoint>& kpts)
void AKAZEFeatures::Do_Subpixel_Refinement(std::vector<KeyPoint>& kpts)
{
float Dx = 0.0, Dy = 0.0, ratio = 0.0;
float Dxx = 0.0, Dyy = 0.0, Dxy = 0.0;
@@ -433,7 +433,7 @@ void AKAZEFeatures::Do_Subpixel_Refinement(std::vector<cv::KeyPoint>& kpts)
b(0) = -Dx;
b(1) = -Dy;
cv::solve(A, b, dst, DECOMP_LU);
solve(A, b, dst, DECOMP_LU);
if (fabs(dst(0)) <= 1.0f && fabs(dst(1)) <= 1.0f) {
kpts[i].pt.x = x + dst(0);
@@ -456,10 +456,10 @@ void AKAZEFeatures::Do_Subpixel_Refinement(std::vector<cv::KeyPoint>& kpts)
/* ************************************************************************* */
class SURF_Descriptor_Upright_64_Invoker : public cv::ParallelLoopBody
class SURF_Descriptor_Upright_64_Invoker : public ParallelLoopBody
{
public:
SURF_Descriptor_Upright_64_Invoker(std::vector<cv::KeyPoint>& kpts, cv::Mat& desc, std::vector<TEvolution>& evolution)
SURF_Descriptor_Upright_64_Invoker(std::vector<KeyPoint>& kpts, Mat& desc, std::vector<TEvolution>& evolution)
: keypoints_(&kpts)
, descriptors_(&desc)
, evolution_(&evolution)
@@ -474,18 +474,18 @@ public:
}
}
void Get_SURF_Descriptor_Upright_64(const cv::KeyPoint& kpt, float* desc) const;
void Get_SURF_Descriptor_Upright_64(const KeyPoint& kpt, float* desc) const;
private:
std::vector<cv::KeyPoint>* keypoints_;
cv::Mat* descriptors_;
std::vector<KeyPoint>* keypoints_;
Mat* descriptors_;
std::vector<TEvolution>* evolution_;
};
class SURF_Descriptor_64_Invoker : public cv::ParallelLoopBody
class SURF_Descriptor_64_Invoker : public ParallelLoopBody
{
public:
SURF_Descriptor_64_Invoker(std::vector<cv::KeyPoint>& kpts, cv::Mat& desc, std::vector<TEvolution>& evolution)
SURF_Descriptor_64_Invoker(std::vector<KeyPoint>& kpts, Mat& desc, std::vector<TEvolution>& evolution)
: keypoints_(&kpts)
, descriptors_(&desc)
, evolution_(&evolution)
@@ -501,18 +501,18 @@ public:
}
}
void Get_SURF_Descriptor_64(const cv::KeyPoint& kpt, float* desc) const;
void Get_SURF_Descriptor_64(const KeyPoint& kpt, float* desc) const;
private:
std::vector<cv::KeyPoint>* keypoints_;
cv::Mat* descriptors_;
std::vector<KeyPoint>* keypoints_;
Mat* descriptors_;
std::vector<TEvolution>* evolution_;
};
class MSURF_Upright_Descriptor_64_Invoker : public cv::ParallelLoopBody
class MSURF_Upright_Descriptor_64_Invoker : public ParallelLoopBody
{
public:
MSURF_Upright_Descriptor_64_Invoker(std::vector<cv::KeyPoint>& kpts, cv::Mat& desc, std::vector<TEvolution>& evolution)
MSURF_Upright_Descriptor_64_Invoker(std::vector<KeyPoint>& kpts, Mat& desc, std::vector<TEvolution>& evolution)
: keypoints_(&kpts)
, descriptors_(&desc)
, evolution_(&evolution)
@@ -527,18 +527,18 @@ public:
}
}
void Get_MSURF_Upright_Descriptor_64(const cv::KeyPoint& kpt, float* desc) const;
void Get_MSURF_Upright_Descriptor_64(const KeyPoint& kpt, float* desc) const;
private:
std::vector<cv::KeyPoint>* keypoints_;
cv::Mat* descriptors_;
std::vector<KeyPoint>* keypoints_;
Mat* descriptors_;
std::vector<TEvolution>* evolution_;
};
class MSURF_Descriptor_64_Invoker : public cv::ParallelLoopBody
class MSURF_Descriptor_64_Invoker : public ParallelLoopBody
{
public:
MSURF_Descriptor_64_Invoker(std::vector<cv::KeyPoint>& kpts, cv::Mat& desc, std::vector<TEvolution>& evolution)
MSURF_Descriptor_64_Invoker(std::vector<KeyPoint>& kpts, Mat& desc, std::vector<TEvolution>& evolution)
: keypoints_(&kpts)
, descriptors_(&desc)
, evolution_(&evolution)
@@ -554,18 +554,18 @@ public:
}
}
void Get_MSURF_Descriptor_64(const cv::KeyPoint& kpt, float* desc) const;
void Get_MSURF_Descriptor_64(const KeyPoint& kpt, float* desc) const;
private:
std::vector<cv::KeyPoint>* keypoints_;
cv::Mat* descriptors_;
std::vector<KeyPoint>* keypoints_;
Mat* descriptors_;
std::vector<TEvolution>* evolution_;
};
class Upright_MLDB_Full_Descriptor_Invoker : public cv::ParallelLoopBody
class Upright_MLDB_Full_Descriptor_Invoker : public ParallelLoopBody
{
public:
Upright_MLDB_Full_Descriptor_Invoker(std::vector<cv::KeyPoint>& kpts, cv::Mat& desc, std::vector<TEvolution>& evolution, AKAZEOptions& options)
Upright_MLDB_Full_Descriptor_Invoker(std::vector<KeyPoint>& kpts, Mat& desc, std::vector<TEvolution>& evolution, AKAZEOptions& options)
: keypoints_(&kpts)
, descriptors_(&desc)
, evolution_(&evolution)
@@ -581,24 +581,24 @@ public:
}
}
void Get_Upright_MLDB_Full_Descriptor(const cv::KeyPoint& kpt, unsigned char* desc) const;
void Get_Upright_MLDB_Full_Descriptor(const KeyPoint& kpt, unsigned char* desc) const;
private:
std::vector<cv::KeyPoint>* keypoints_;
cv::Mat* descriptors_;
std::vector<KeyPoint>* keypoints_;
Mat* descriptors_;
std::vector<TEvolution>* evolution_;
AKAZEOptions* options_;
};
class Upright_MLDB_Descriptor_Subset_Invoker : public cv::ParallelLoopBody
class Upright_MLDB_Descriptor_Subset_Invoker : public ParallelLoopBody
{
public:
Upright_MLDB_Descriptor_Subset_Invoker(std::vector<cv::KeyPoint>& kpts,
cv::Mat& desc,
Upright_MLDB_Descriptor_Subset_Invoker(std::vector<KeyPoint>& kpts,
Mat& desc,
std::vector<TEvolution>& evolution,
AKAZEOptions& options,
cv::Mat descriptorSamples,
cv::Mat descriptorBits)
Mat descriptorSamples,
Mat descriptorBits)
: keypoints_(&kpts)
, descriptors_(&desc)
, evolution_(&evolution)
@@ -616,22 +616,22 @@ public:
}
}
void Get_Upright_MLDB_Descriptor_Subset(const cv::KeyPoint& kpt, unsigned char* desc) const;
void Get_Upright_MLDB_Descriptor_Subset(const KeyPoint& kpt, unsigned char* desc) const;
private:
std::vector<cv::KeyPoint>* keypoints_;
cv::Mat* descriptors_;
std::vector<KeyPoint>* keypoints_;
Mat* descriptors_;
std::vector<TEvolution>* evolution_;
AKAZEOptions* options_;
cv::Mat descriptorSamples_; // List of positions in the grids to sample LDB bits from.
cv::Mat descriptorBits_;
Mat descriptorSamples_; // List of positions in the grids to sample LDB bits from.
Mat descriptorBits_;
};
class MLDB_Full_Descriptor_Invoker : public cv::ParallelLoopBody
class MLDB_Full_Descriptor_Invoker : public ParallelLoopBody
{
public:
MLDB_Full_Descriptor_Invoker(std::vector<cv::KeyPoint>& kpts, cv::Mat& desc, std::vector<TEvolution>& evolution, AKAZEOptions& options)
MLDB_Full_Descriptor_Invoker(std::vector<KeyPoint>& kpts, Mat& desc, std::vector<TEvolution>& evolution, AKAZEOptions& options)
: keypoints_(&kpts)
, descriptors_(&desc)
, evolution_(&evolution)
@@ -648,28 +648,28 @@ public:
}
}
void Get_MLDB_Full_Descriptor(const cv::KeyPoint& kpt, unsigned char* desc) const;
void Get_MLDB_Full_Descriptor(const KeyPoint& kpt, unsigned char* desc) const;
void MLDB_Fill_Values(float* values, int sample_step, int level,
float xf, float yf, float co, float si, float scale) const;
void MLDB_Binary_Comparisons(float* values, unsigned char* desc,
int count, int& dpos) const;
private:
std::vector<cv::KeyPoint>* keypoints_;
cv::Mat* descriptors_;
std::vector<KeyPoint>* keypoints_;
Mat* descriptors_;
std::vector<TEvolution>* evolution_;
AKAZEOptions* options_;
};
class MLDB_Descriptor_Subset_Invoker : public cv::ParallelLoopBody
class MLDB_Descriptor_Subset_Invoker : public ParallelLoopBody
{
public:
MLDB_Descriptor_Subset_Invoker(std::vector<cv::KeyPoint>& kpts,
cv::Mat& desc,
MLDB_Descriptor_Subset_Invoker(std::vector<KeyPoint>& kpts,
Mat& desc,
std::vector<TEvolution>& evolution,
AKAZEOptions& options,
cv::Mat descriptorSamples,
cv::Mat descriptorBits)
Mat descriptorSamples,
Mat descriptorBits)
: keypoints_(&kpts)
, descriptors_(&desc)
, evolution_(&evolution)
@@ -688,16 +688,16 @@ public:
}
}
void Get_MLDB_Descriptor_Subset(const cv::KeyPoint& kpt, unsigned char* desc) const;
void Get_MLDB_Descriptor_Subset(const KeyPoint& kpt, unsigned char* desc) const;
private:
std::vector<cv::KeyPoint>* keypoints_;
cv::Mat* descriptors_;
std::vector<KeyPoint>* keypoints_;
Mat* descriptors_;
std::vector<TEvolution>* evolution_;
AKAZEOptions* options_;
cv::Mat descriptorSamples_; // List of positions in the grids to sample LDB bits from.
cv::Mat descriptorBits_;
Mat descriptorSamples_; // List of positions in the grids to sample LDB bits from.
Mat descriptorBits_;
};
/**
@@ -705,7 +705,7 @@ private:
* @param kpts Vector of detected keypoints
* @param desc Matrix to store the descriptors
*/
void AKAZEFeatures::Compute_Descriptors(std::vector<cv::KeyPoint>& kpts, cv::Mat& desc)
void AKAZEFeatures::Compute_Descriptors(std::vector<KeyPoint>& kpts, Mat& desc)
{
for(size_t i = 0; i < kpts.size(); i++)
{
@@ -714,17 +714,17 @@ void AKAZEFeatures::Compute_Descriptors(std::vector<cv::KeyPoint>& kpts, cv::Mat
// Allocate memory for the matrix with the descriptors
if (options_.descriptor < AKAZE::DESCRIPTOR_MLDB_UPRIGHT) {
desc = cv::Mat::zeros((int)kpts.size(), 64, CV_32FC1);
desc = Mat::zeros((int)kpts.size(), 64, CV_32FC1);
}
else {
// We use the full length binary descriptor -> 486 bits
if (options_.descriptor_size == 0) {
int t = (6 + 36 + 120)*options_.descriptor_channels;
desc = cv::Mat::zeros((int)kpts.size(), (int)ceil(t / 8.), CV_8UC1);
desc = Mat::zeros((int)kpts.size(), (int)ceil(t / 8.), CV_8UC1);
}
else {
// We use the random bit selection length binary descriptor
desc = cv::Mat::zeros((int)kpts.size(), (int)ceil(options_.descriptor_size / 8.), CV_8UC1);
desc = Mat::zeros((int)kpts.size(), (int)ceil(options_.descriptor_size / 8.), CV_8UC1);
}
}
@@ -732,28 +732,28 @@ void AKAZEFeatures::Compute_Descriptors(std::vector<cv::KeyPoint>& kpts, cv::Mat
{
case AKAZE::DESCRIPTOR_KAZE_UPRIGHT: // Upright descriptors, not invariant to rotation
{
cv::parallel_for_(cv::Range(0, (int)kpts.size()), MSURF_Upright_Descriptor_64_Invoker(kpts, desc, evolution_));
parallel_for_(Range(0, (int)kpts.size()), MSURF_Upright_Descriptor_64_Invoker(kpts, desc, evolution_));
}
break;
case AKAZE::DESCRIPTOR_KAZE:
{
cv::parallel_for_(cv::Range(0, (int)kpts.size()), MSURF_Descriptor_64_Invoker(kpts, desc, evolution_));
parallel_for_(Range(0, (int)kpts.size()), MSURF_Descriptor_64_Invoker(kpts, desc, evolution_));
}
break;
case AKAZE::DESCRIPTOR_MLDB_UPRIGHT: // Upright descriptors, not invariant to rotation
{
if (options_.descriptor_size == 0)
cv::parallel_for_(cv::Range(0, (int)kpts.size()), Upright_MLDB_Full_Descriptor_Invoker(kpts, desc, evolution_, options_));
parallel_for_(Range(0, (int)kpts.size()), Upright_MLDB_Full_Descriptor_Invoker(kpts, desc, evolution_, options_));
else
cv::parallel_for_(cv::Range(0, (int)kpts.size()), Upright_MLDB_Descriptor_Subset_Invoker(kpts, desc, evolution_, options_, descriptorSamples_, descriptorBits_));
parallel_for_(Range(0, (int)kpts.size()), Upright_MLDB_Descriptor_Subset_Invoker(kpts, desc, evolution_, options_, descriptorSamples_, descriptorBits_));
}
break;
case AKAZE::DESCRIPTOR_MLDB:
{
if (options_.descriptor_size == 0)
cv::parallel_for_(cv::Range(0, (int)kpts.size()), MLDB_Full_Descriptor_Invoker(kpts, desc, evolution_, options_));
parallel_for_(Range(0, (int)kpts.size()), MLDB_Full_Descriptor_Invoker(kpts, desc, evolution_, options_));
else
cv::parallel_for_(cv::Range(0, (int)kpts.size()), MLDB_Descriptor_Subset_Invoker(kpts, desc, evolution_, options_, descriptorSamples_, descriptorBits_));
parallel_for_(Range(0, (int)kpts.size()), MLDB_Descriptor_Subset_Invoker(kpts, desc, evolution_, options_, descriptorSamples_, descriptorBits_));
}
break;
}
@@ -766,7 +766,7 @@ void AKAZEFeatures::Compute_Descriptors(std::vector<cv::KeyPoint>& kpts, cv::Mat
* @note The orientation is computed using a similar approach as described in the
* original SURF method. See Bay et al., Speeded Up Robust Features, ECCV 2006
*/
void AKAZEFeatures::Compute_Main_Orientation(cv::KeyPoint& kpt, const std::vector<TEvolution>& evolution_)
void AKAZEFeatures::Compute_Main_Orientation(KeyPoint& kpt, const std::vector<TEvolution>& evolution_)
{
/* ************************************************************************* */
/// Lookup table for 2d gaussian (sigma = 2.5) where (0,0) is top left and (6,6) is bottom right
@@ -854,7 +854,7 @@ void AKAZEFeatures::Compute_Main_Orientation(cv::KeyPoint& kpt, const std::vecto
* from Agrawal et al., CenSurE: Center Surround Extremas for Realtime Feature Detection and Matching,
* ECCV 2008
*/
void MSURF_Upright_Descriptor_64_Invoker::Get_MSURF_Upright_Descriptor_64(const cv::KeyPoint& kpt, float *desc) const {
void MSURF_Upright_Descriptor_64_Invoker::Get_MSURF_Upright_Descriptor_64(const KeyPoint& kpt, float *desc) const {
float dx = 0.0, dy = 0.0, mdx = 0.0, mdy = 0.0, gauss_s1 = 0.0, gauss_s2 = 0.0;
float rx = 0.0, ry = 0.0, len = 0.0, xf = 0.0, yf = 0.0, ys = 0.0, xs = 0.0;
@@ -977,7 +977,7 @@ void MSURF_Upright_Descriptor_64_Invoker::Get_MSURF_Upright_Descriptor_64(const
* from Agrawal et al., CenSurE: Center Surround Extremas for Realtime Feature Detection and Matching,
* ECCV 2008
*/
void MSURF_Descriptor_64_Invoker::Get_MSURF_Descriptor_64(const cv::KeyPoint& kpt, float *desc) const {
void MSURF_Descriptor_64_Invoker::Get_MSURF_Descriptor_64(const KeyPoint& kpt, float *desc) const {
float dx = 0.0, dy = 0.0, mdx = 0.0, mdy = 0.0, gauss_s1 = 0.0, gauss_s2 = 0.0;
float rx = 0.0, ry = 0.0, rrx = 0.0, rry = 0.0, len = 0.0, xf = 0.0, yf = 0.0, ys = 0.0, xs = 0.0;
@@ -1101,7 +1101,7 @@ void MSURF_Descriptor_64_Invoker::Get_MSURF_Descriptor_64(const cv::KeyPoint& kp
* @param kpt Input keypoint
* @param desc Descriptor vector
*/
void Upright_MLDB_Full_Descriptor_Invoker::Get_Upright_MLDB_Full_Descriptor(const cv::KeyPoint& kpt, unsigned char *desc) const {
void Upright_MLDB_Full_Descriptor_Invoker::Get_Upright_MLDB_Full_Descriptor(const KeyPoint& kpt, unsigned char *desc) const {
float di = 0.0, dx = 0.0, dy = 0.0;
float ri = 0.0, rx = 0.0, ry = 0.0, xf = 0.0, yf = 0.0;
@@ -1114,9 +1114,9 @@ void Upright_MLDB_Full_Descriptor_Invoker::Get_Upright_MLDB_Full_Descriptor(cons
const std::vector<TEvolution>& evolution = *evolution_;
// Matrices for the M-LDB descriptor
cv::Mat values_1 = cv::Mat::zeros(4, options.descriptor_channels, CV_32FC1);
cv::Mat values_2 = cv::Mat::zeros(9, options.descriptor_channels, CV_32FC1);
cv::Mat values_3 = cv::Mat::zeros(16, options.descriptor_channels, CV_32FC1);
Mat values_1 = Mat::zeros(4, options.descriptor_channels, CV_32FC1);
Mat values_2 = Mat::zeros(9, options.descriptor_channels, CV_32FC1);
Mat values_3 = Mat::zeros(16, options.descriptor_channels, CV_32FC1);
// Get the information from the keypoint
ratio = (float)(1 << kpt.octave);
@@ -1395,7 +1395,7 @@ void MLDB_Full_Descriptor_Invoker::MLDB_Binary_Comparisons(float* values, unsign
* @param kpt Input keypoint
* @param desc Descriptor vector
*/
void MLDB_Full_Descriptor_Invoker::Get_MLDB_Full_Descriptor(const cv::KeyPoint& kpt, unsigned char *desc) const {
void MLDB_Full_Descriptor_Invoker::Get_MLDB_Full_Descriptor(const KeyPoint& kpt, unsigned char *desc) const {
const int max_channels = 3;
CV_Assert(options_->descriptor_channels <= max_channels);
@@ -1428,7 +1428,7 @@ void MLDB_Full_Descriptor_Invoker::Get_MLDB_Full_Descriptor(const cv::KeyPoint&
* @param kpt Input keypoint
* @param desc Descriptor vector
*/
void MLDB_Descriptor_Subset_Invoker::Get_MLDB_Descriptor_Subset(const cv::KeyPoint& kpt, unsigned char *desc) const {
void MLDB_Descriptor_Subset_Invoker::Get_MLDB_Descriptor_Subset(const KeyPoint& kpt, unsigned char *desc) const {
float di = 0.f, dx = 0.f, dy = 0.f;
float rx = 0.f, ry = 0.f;
@@ -1449,7 +1449,7 @@ void MLDB_Descriptor_Subset_Invoker::Get_MLDB_Descriptor_Subset(const cv::KeyPoi
float si = sin(angle);
// Allocate memory for the matrix of values
cv::Mat values = cv::Mat_<float>::zeros((4 + 9 + 16)*options.descriptor_channels, 1);
Mat values = Mat_<float>::zeros((4 + 9 + 16)*options.descriptor_channels, 1);
// Sample everything, but only do the comparisons
vector<int> steps(3);
@@ -1522,7 +1522,7 @@ void MLDB_Descriptor_Subset_Invoker::Get_MLDB_Descriptor_Subset(const cv::KeyPoi
* @param kpt Input keypoint
* @param desc Descriptor vector
*/
void Upright_MLDB_Descriptor_Subset_Invoker::Get_Upright_MLDB_Descriptor_Subset(const cv::KeyPoint& kpt, unsigned char *desc) const {
void Upright_MLDB_Descriptor_Subset_Invoker::Get_Upright_MLDB_Descriptor_Subset(const KeyPoint& kpt, unsigned char *desc) const {
float di = 0.0f, dx = 0.0f, dy = 0.0f;
float rx = 0.0f, ry = 0.0f;
@@ -1540,7 +1540,7 @@ void Upright_MLDB_Descriptor_Subset_Invoker::Get_Upright_MLDB_Descriptor_Subset(
float xf = kpt.pt.x / ratio;
// Allocate memory for the matrix of values
Mat values = cv::Mat_<float>::zeros((4 + 9 + 16)*options.descriptor_channels, 1);
Mat values = Mat_<float>::zeros((4 + 9 + 16)*options.descriptor_channels, 1);
vector<int> steps(3);
steps.at(0) = options.descriptor_pattern_size;
@@ -1614,7 +1614,7 @@ void Upright_MLDB_Descriptor_Subset_Invoker::Get_Upright_MLDB_Descriptor_Subset(
* @note The function keeps the 18 bits (3-channels by 6 comparisons) of the
* coarser grid, since it provides the most robust estimations
*/
void generateDescriptorSubsample(cv::Mat& sampleList, cv::Mat& comparisons, int nbits,
void generateDescriptorSubsample(Mat& sampleList, Mat& comparisons, int nbits,
int pattern_size, int nchannels) {
int ssz = 0;
@@ -1718,4 +1718,3 @@ void generateDescriptorSubsample(cv::Mat& sampleList, cv::Mat& comparisons, int
}
}
}

8
modules/features2d/src/kaze/AKAZEFeatures.h
View File

@@ -11,10 +11,12 @@
/* ************************************************************************* */
// Includes
#include "../precomp.hpp"
#include "AKAZEConfig.h"
#include "TEvolution.h"
namespace cv
{
/* ************************************************************************* */
// AKAZE Class Declaration
class AKAZEFeatures {
@@ -22,7 +24,7 @@ class AKAZEFeatures {
private:
AKAZEOptions options_; ///< Configuration options for AKAZE
std::vector<TEvolution> evolution_; ///< Vector of nonlinear diffusion evolution
std::vector<TEvolution> evolution_; ///< Vector of nonlinear diffusion evolution
/// FED parameters
int ncycles_; ///< Number of cycles
@@ -59,4 +61,6 @@ public:
void generateDescriptorSubsample(cv::Mat& sampleList, cv::Mat& comparisons,
int nbits, int pattern_size, int nchannels);
}
#endif

216
modules/features2d/src/kaze/KAZEFeatures.cpp
View File

@@ -20,14 +20,15 @@
* @date Jan 21, 2012
* @author Pablo F. Alcantarilla
*/
#include "../precomp.hpp"
#include "KAZEFeatures.h"
#include "utils.h"
namespace cv
{
// Namespaces
using namespace std;
using namespace cv;
using namespace cv::details::kaze;
/* ************************************************************************* */
/**
@@ -52,19 +53,20 @@ KAZEFeatures::KAZEFeatures(KAZEOptions& options)
void KAZEFeatures::Allocate_Memory_Evolution(void) {
// Allocate the dimension of the matrices for the evolution
for (int i = 0; i <= options_.omax - 1; i++) {
for (int j = 0; j <= options_.nsublevels - 1; j++) {
for (int i = 0; i <= options_.omax - 1; i++)
{
for (int j = 0; j <= options_.nsublevels - 1; j++)
{
TEvolution aux;
aux.Lx = cv::Mat::zeros(options_.img_height, options_.img_width, CV_32F);
aux.Ly = cv::Mat::zeros(options_.img_height, options_.img_width, CV_32F);
aux.Lxx = cv::Mat::zeros(options_.img_height, options_.img_width, CV_32F);
aux.Lxy = cv::Mat::zeros(options_.img_height, options_.img_width, CV_32F);
aux.Lyy = cv::Mat::zeros(options_.img_height, options_.img_width, CV_32F);
aux.Lt = cv::Mat::zeros(options_.img_height, options_.img_width, CV_32F);
aux.Lsmooth = cv::Mat::zeros(options_.img_height, options_.img_width, CV_32F);
aux.Ldet = cv::Mat::zeros(options_.img_height, options_.img_width, CV_32F);
aux.esigma = options_.soffset*pow((float)2.0f, (float)(j) / (float)(options_.nsublevels)+i);
aux.Lx = Mat::zeros(options_.img_height, options_.img_width, CV_32F);
aux.Ly = Mat::zeros(options_.img_height, options_.img_width, CV_32F);
aux.Lxx = Mat::zeros(options_.img_height, options_.img_width, CV_32F);
aux.Lxy = Mat::zeros(options_.img_height, options_.img_width, CV_32F);
aux.Lyy = Mat::zeros(options_.img_height, options_.img_width, CV_32F);
aux.Lt = Mat::zeros(options_.img_height, options_.img_width, CV_32F);
aux.Lsmooth = Mat::zeros(options_.img_height, options_.img_width, CV_32F);
aux.Ldet = Mat::zeros(options_.img_height, options_.img_width, CV_32F);
aux.esigma = options_.soffset*pow((float)2.0f, (float)(j) / (float)(options_.nsublevels)+i);
aux.etime = 0.5f*(aux.esigma*aux.esigma);
aux.sigma_size = fRound(aux.esigma);
aux.octave = i;
@@ -74,7 +76,8 @@ void KAZEFeatures::Allocate_Memory_Evolution(void) {
}
// Allocate memory for the FED number of cycles and time steps
for (size_t i = 1; i < evolution_.size(); i++) {
for (size_t i = 1; i < evolution_.size(); i++)
{
int naux = 0;
vector<float> tau;
float ttime = 0.0;
@@ -92,47 +95,43 @@ void KAZEFeatures::Allocate_Memory_Evolution(void) {
* @param img Input image for which the nonlinear scale space needs to be created
* @return 0 if the nonlinear scale space was created successfully. -1 otherwise
*/
int KAZEFeatures::Create_Nonlinear_Scale_Space(const cv::Mat &img)
int KAZEFeatures::Create_Nonlinear_Scale_Space(const Mat &img)
{
CV_Assert(evolution_.size() > 0);
// Copy the original image to the first level of the evolution
img.copyTo(evolution_[0].Lt);
gaussian_2D_convolution(evolution_[0].Lt, evolution_[0].Lt, 0, 0, options_.soffset);
gaussian_2D_convolution(evolution_[0].Lt, evolution_[0].Lsmooth, 0, 0, options_.sderivatives);
gaussian_2D_convolution(evolution_[0].Lt, evolution_[0].Lt, 0, 0, options_.soffset);
gaussian_2D_convolution(evolution_[0].Lt, evolution_[0].Lsmooth, 0, 0, options_.sderivatives);
// Firstly compute the kcontrast factor
Compute_KContrast(evolution_[0].Lt, options_.kcontrast_percentille);
// Allocate memory for the flow and step images
cv::Mat Lflow = cv::Mat::zeros(evolution_[0].Lt.rows, evolution_[0].Lt.cols, CV_32F);
cv::Mat Lstep = cv::Mat::zeros(evolution_[0].Lt.rows, evolution_[0].Lt.cols, CV_32F);
Mat Lflow = Mat::zeros(evolution_[0].Lt.rows, evolution_[0].Lt.cols, CV_32F);
Mat Lstep = Mat::zeros(evolution_[0].Lt.rows, evolution_[0].Lt.cols, CV_32F);
// Now generate the rest of evolution levels
for (size_t i = 1; i < evolution_.size(); i++) {
for (size_t i = 1; i < evolution_.size(); i++)
{
evolution_[i - 1].Lt.copyTo(evolution_[i].Lt);
gaussian_2D_convolution(evolution_[i - 1].Lt, evolution_[i].Lsmooth, 0, 0, options_.sderivatives);
gaussian_2D_convolution(evolution_[i - 1].Lt, evolution_[i].Lsmooth, 0, 0, options_.sderivatives);
// Compute the Gaussian derivatives Lx and Ly
Scharr(evolution_[i].Lsmooth, evolution_[i].Lx, CV_32F, 1, 0, 1, 0, BORDER_DEFAULT);
Scharr(evolution_[i].Lsmooth, evolution_[i].Ly, CV_32F, 0, 1, 1, 0, BORDER_DEFAULT);
// Compute the conductivity equation
if (options_.diffusivity == cv::DIFF_PM_G1) {
pm_g1(evolution_[i].Lx, evolution_[i].Ly, Lflow, options_.kcontrast);
}
else if (options_.diffusivity == cv::DIFF_PM_G2) {
pm_g2(evolution_[i].Lx, evolution_[i].Ly, Lflow, options_.kcontrast);
}
else if (options_.diffusivity == cv::DIFF_WEICKERT) {
weickert_diffusivity(evolution_[i].Lx, evolution_[i].Ly, Lflow, options_.kcontrast);
}
if (options_.diffusivity == KAZE::DIFF_PM_G1)
pm_g1(evolution_[i].Lx, evolution_[i].Ly, Lflow, options_.kcontrast);
else if (options_.diffusivity == KAZE::DIFF_PM_G2)
pm_g2(evolution_[i].Lx, evolution_[i].Ly, Lflow, options_.kcontrast);
else if (options_.diffusivity == KAZE::DIFF_WEICKERT)
weickert_diffusivity(evolution_[i].Lx, evolution_[i].Ly, Lflow, options_.kcontrast);
// Perform FED n inner steps
for (int j = 0; j < nsteps_[i - 1]; j++) {
for (int j = 0; j < nsteps_[i - 1]; j++)
nld_step_scalar(evolution_[i].Lt, Lflow, Lstep, tsteps_[i - 1][j]);
}
}
return 0;
@@ -144,7 +143,7 @@ int KAZEFeatures::Create_Nonlinear_Scale_Space(const cv::Mat &img)
* @param img Input image
* @param kpercentile Percentile of the gradient histogram
*/
void KAZEFeatures::Compute_KContrast(const cv::Mat &img, const float &kpercentile)
void KAZEFeatures::Compute_KContrast(const Mat &img, const float &kpercentile)
{
options_.kcontrast = compute_k_percentile(img, kpercentile, options_.sderivatives, options_.kcontrast_bins, 0, 0);
}
@@ -181,7 +180,7 @@ void KAZEFeatures::Compute_Detector_Response(void)
* @brief This method selects interesting keypoints through the nonlinear scale space
* @param kpts Vector of keypoints
*/
void KAZEFeatures::Feature_Detection(std::vector<cv::KeyPoint>& kpts)
void KAZEFeatures::Feature_Detection(std::vector<KeyPoint>& kpts)
{
kpts.clear();
Compute_Detector_Response();
@@ -190,14 +189,14 @@ void KAZEFeatures::Feature_Detection(std::vector<cv::KeyPoint>& kpts)
}
/* ************************************************************************* */
class MultiscaleDerivativesKAZEInvoker : public cv::ParallelLoopBody
class MultiscaleDerivativesKAZEInvoker : public ParallelLoopBody
{
public:
explicit MultiscaleDerivativesKAZEInvoker(std::vector<TEvolution>& ev) : evolution_(&ev)
{
}
void operator()(const cv::Range& range) const
void operator()(const Range& range) const
{
std::vector<TEvolution>& evolution = *evolution_;
for (int i = range.start; i < range.end; i++)
@@ -226,74 +225,79 @@ private:
*/
void KAZEFeatures::Compute_Multiscale_Derivatives(void)
{
cv::parallel_for_(cv::Range(0, (int)evolution_.size()),
parallel_for_(Range(0, (int)evolution_.size()),
MultiscaleDerivativesKAZEInvoker(evolution_));
}
/* ************************************************************************* */
class FindExtremumKAZEInvoker : public cv::ParallelLoopBody
class FindExtremumKAZEInvoker : public ParallelLoopBody
{
public:
explicit FindExtremumKAZEInvoker(std::vector<TEvolution>& ev, std::vector<std::vector<cv::KeyPoint> >& kpts_par,
explicit FindExtremumKAZEInvoker(std::vector<TEvolution>& ev, std::vector<std::vector<KeyPoint> >& kpts_par,
const KAZEOptions& options) : evolution_(&ev), kpts_par_(&kpts_par), options_(options)
{
}
void operator()(const cv::Range& range) const
void operator()(const Range& range) const
{
std::vector<TEvolution>& evolution = *evolution_;
std::vector<std::vector<cv::KeyPoint> >& kpts_par = *kpts_par_;
std::vector<std::vector<KeyPoint> >& kpts_par = *kpts_par_;
for (int i = range.start; i < range.end; i++)
{
float value = 0.0;
bool is_extremum = false;
for (int ix = 1; ix < options_.img_height - 1; ix++) {
for (int jx = 1; jx < options_.img_width - 1; jx++) {
for (int ix = 1; ix < options_.img_height - 1; ix++)
{
for (int jx = 1; jx < options_.img_width - 1; jx++)
{
is_extremum = false;
value = *(evolution[i].Ldet.ptr<float>(ix)+jx);
is_extremum = false;
value = *(evolution[i].Ldet.ptr<float>(ix)+jx);
// Filter the points with the detector threshold
if (value > options_.dthreshold) {
if (value >= *(evolution[i].Ldet.ptr<float>(ix)+jx - 1)) {
// First check on the same scale
if (check_maximum_neighbourhood(evolution[i].Ldet, 1, value, ix, jx, 1)) {
// Now check on the lower scale
if (check_maximum_neighbourhood(evolution[i - 1].Ldet, 1, value, ix, jx, 0)) {
// Now check on the upper scale
if (check_maximum_neighbourhood(evolution[i + 1].Ldet, 1, value, ix, jx, 0)) {
is_extremum = true;
}
}
}
}
}
// Add the point of interest!!
if (is_extremum == true) {
cv::KeyPoint point;
point.pt.x = (float)jx;
point.pt.y = (float)ix;
point.response = fabs(value);
point.size = evolution[i].esigma;
point.octave = (int)evolution[i].octave;
point.class_id = i;
// We use the angle field for the sublevel value
// Then, we will replace this angle field with the main orientation
point.angle = static_cast<float>(evolution[i].sublevel);
kpts_par[i - 1].push_back(point);
// Filter the points with the detector threshold
if (value > options_.dthreshold)
{
if (value >= *(evolution[i].Ldet.ptr<float>(ix)+jx - 1))
{
// First check on the same scale
if (check_maximum_neighbourhood(evolution[i].Ldet, 1, value, ix, jx, 1))
{
// Now check on the lower scale
if (check_maximum_neighbourhood(evolution[i - 1].Ldet, 1, value, ix, jx, 0))
{
// Now check on the upper scale
if (check_maximum_neighbourhood(evolution[i + 1].Ldet, 1, value, ix, jx, 0))
is_extremum = true;
}
}
}
}
// Add the point of interest!!
if (is_extremum)
{
KeyPoint point;
point.pt.x = (float)jx;
point.pt.y = (float)ix;
point.response = fabs(value);
point.size = evolution[i].esigma;
point.octave = (int)evolution[i].octave;
point.class_id = i;
// We use the angle field for the sublevel value
// Then, we will replace this angle field with the main orientation
point.angle = static_cast<float>(evolution[i].sublevel);
kpts_par[i - 1].push_back(point);
}
}
}
}
}
private:
std::vector<TEvolution>* evolution_;
std::vector<std::vector<cv::KeyPoint> >* kpts_par_;
std::vector<std::vector<KeyPoint> >* kpts_par_;
KAZEOptions options_;
};
@@ -304,7 +308,7 @@ private:
* @param kpts Vector of keypoints
* @note We compute features for each of the nonlinear scale space level in a different processing thread
*/
void KAZEFeatures::Determinant_Hessian(std::vector<cv::KeyPoint>& kpts)
void KAZEFeatures::Determinant_Hessian(std::vector<KeyPoint>& kpts)
{
int level = 0;
float dist = 0.0, smax = 3.0;
@@ -325,12 +329,14 @@ void KAZEFeatures::Determinant_Hessian(std::vector<cv::KeyPoint>& kpts)
kpts_par_.push_back(aux);
}
cv::parallel_for_(cv::Range(1, (int)evolution_.size()-1),
FindExtremumKAZEInvoker(evolution_, kpts_par_, options_));
parallel_for_(Range(1, (int)evolution_.size()-1),
FindExtremumKAZEInvoker(evolution_, kpts_par_, options_));
// Now fill the vector of keypoints!!!
for (int i = 0; i < (int)kpts_par_.size(); i++) {
for (int j = 0; j < (int)kpts_par_[i].size(); j++) {
for (int i = 0; i < (int)kpts_par_.size(); i++)
{
for (int j = 0; j < (int)kpts_par_[i].size(); j++)
{
level = i + 1;
is_extremum = true;
is_repeated = false;
@@ -388,7 +394,7 @@ void KAZEFeatures::Determinant_Hessian(std::vector<cv::KeyPoint>& kpts)
* @brief This method performs subpixel refinement of the detected keypoints
* @param kpts Vector of detected keypoints
*/
void KAZEFeatures::Do_Subpixel_Refinement(std::vector<cv::KeyPoint> &kpts) {
void KAZEFeatures::Do_Subpixel_Refinement(std::vector<KeyPoint> &kpts) {
int step = 1;
int x = 0, y = 0;
@@ -482,10 +488,10 @@ void KAZEFeatures::Do_Subpixel_Refinement(std::vector<cv::KeyPoint> &kpts) {
}
/* ************************************************************************* */
class KAZE_Descriptor_Invoker : public cv::ParallelLoopBody
class KAZE_Descriptor_Invoker : public ParallelLoopBody
{
public:
KAZE_Descriptor_Invoker(std::vector<cv::KeyPoint> &kpts, cv::Mat &desc, std::vector<TEvolution>& evolution, const KAZEOptions& options)
KAZE_Descriptor_Invoker(std::vector<KeyPoint> &kpts, Mat &desc, std::vector<TEvolution>& evolution, const KAZEOptions& options)
: kpts_(&kpts)
, desc_(&desc)
, evolution_(&evolution)
@@ -497,10 +503,10 @@ public:
{
}
void operator() (const cv::Range& range) const
void operator() (const Range& range) const
{
std::vector<cv::KeyPoint> &kpts = *kpts_;
cv::Mat &desc = *desc_;
std::vector<KeyPoint> &kpts = *kpts_;
Mat &desc = *desc_;
std::vector<TEvolution> &evolution = *evolution_;
for (int i = range.start; i < range.end; i++)
@@ -526,13 +532,13 @@ public:
}
}
private:
void Get_KAZE_Upright_Descriptor_64(const cv::KeyPoint& kpt, float* desc) const;
void Get_KAZE_Descriptor_64(const cv::KeyPoint& kpt, float* desc) const;
void Get_KAZE_Upright_Descriptor_128(const cv::KeyPoint& kpt, float* desc) const;
void Get_KAZE_Descriptor_128(const cv::KeyPoint& kpt, float *desc) const;
void Get_KAZE_Upright_Descriptor_64(const KeyPoint& kpt, float* desc) const;
void Get_KAZE_Descriptor_64(const KeyPoint& kpt, float* desc) const;
void Get_KAZE_Upright_Descriptor_128(const KeyPoint& kpt, float* desc) const;
void Get_KAZE_Descriptor_128(const KeyPoint& kpt, float *desc) const;
std::vector<cv::KeyPoint> * kpts_;
cv::Mat * desc_;
std::vector<KeyPoint> * kpts_;
Mat * desc_;
std::vector<TEvolution> * evolution_;
KAZEOptions options_;
};
@@ -543,7 +549,7 @@ private:
* @param kpts Vector of keypoints
* @param desc Matrix with the feature descriptors
*/
void KAZEFeatures::Feature_Description(std::vector<cv::KeyPoint> &kpts, cv::Mat &desc)
void KAZEFeatures::Feature_Description(std::vector<KeyPoint> &kpts, Mat &desc)
{
for(size_t i = 0; i < kpts.size(); i++)
{
@@ -558,7 +564,7 @@ void KAZEFeatures::Feature_Description(std::vector<cv::KeyPoint> &kpts, cv::Mat
desc = Mat::zeros((int)kpts.size(), 64, CV_32FC1);
}
cv::parallel_for_(cv::Range(0, (int)kpts.size()), KAZE_Descriptor_Invoker(kpts, desc, evolution_, options_));
parallel_for_(Range(0, (int)kpts.size()), KAZE_Descriptor_Invoker(kpts, desc, evolution_, options_));
}
/* ************************************************************************* */
@@ -568,7 +574,7 @@ void KAZEFeatures::Feature_Description(std::vector<cv::KeyPoint> &kpts, cv::Mat
* @note The orientation is computed using a similar approach as described in the
* original SURF method. See Bay et al., Speeded Up Robust Features, ECCV 2006
*/
void KAZEFeatures::Compute_Main_Orientation(cv::KeyPoint &kpt, const std::vector<TEvolution>& evolution_, const KAZEOptions& options)
void KAZEFeatures::Compute_Main_Orientation(KeyPoint &kpt, const std::vector<TEvolution>& evolution_, const KAZEOptions& options)
{
int ix = 0, iy = 0, idx = 0, s = 0, level = 0;
float xf = 0.0, yf = 0.0, gweight = 0.0;
@@ -647,7 +653,7 @@ void KAZEFeatures::Compute_Main_Orientation(cv::KeyPoint &kpt, const std::vector
* from Agrawal et al., CenSurE: Center Surround Extremas for Realtime Feature Detection and Matching,
* ECCV 2008
*/
void KAZE_Descriptor_Invoker::Get_KAZE_Upright_Descriptor_64(const cv::KeyPoint &kpt, float *desc) const
void KAZE_Descriptor_Invoker::Get_KAZE_Upright_Descriptor_64(const KeyPoint &kpt, float *desc) const
{
float dx = 0.0, dy = 0.0, mdx = 0.0, mdy = 0.0, gauss_s1 = 0.0, gauss_s2 = 0.0;
float rx = 0.0, ry = 0.0, len = 0.0, xf = 0.0, yf = 0.0, ys = 0.0, xs = 0.0;
@@ -775,7 +781,7 @@ void KAZE_Descriptor_Invoker::Get_KAZE_Upright_Descriptor_64(const cv::KeyPoint
* from Agrawal et al., CenSurE: Center Surround Extremas for Realtime Feature Detection and Matching,
* ECCV 2008
*/
void KAZE_Descriptor_Invoker::Get_KAZE_Descriptor_64(const cv::KeyPoint &kpt, float *desc) const
void KAZE_Descriptor_Invoker::Get_KAZE_Descriptor_64(const KeyPoint &kpt, float *desc) const
{
float dx = 0.0, dy = 0.0, mdx = 0.0, mdy = 0.0, gauss_s1 = 0.0, gauss_s2 = 0.0;
float rx = 0.0, ry = 0.0, rrx = 0.0, rry = 0.0, len = 0.0, xf = 0.0, yf = 0.0, ys = 0.0, xs = 0.0;
@@ -904,7 +910,7 @@ void KAZE_Descriptor_Invoker::Get_KAZE_Descriptor_64(const cv::KeyPoint &kpt, fl
* from Agrawal et al., CenSurE: Center Surround Extremas for Realtime Feature Detection and Matching,
* ECCV 2008
*/
void KAZE_Descriptor_Invoker::Get_KAZE_Upright_Descriptor_128(const cv::KeyPoint &kpt, float *desc) const
void KAZE_Descriptor_Invoker::Get_KAZE_Upright_Descriptor_128(const KeyPoint &kpt, float *desc) const
{
float gauss_s1 = 0.0, gauss_s2 = 0.0;
float rx = 0.0, ry = 0.0, len = 0.0, xf = 0.0, yf = 0.0, ys = 0.0, xs = 0.0;
@@ -1056,7 +1062,7 @@ void KAZE_Descriptor_Invoker::Get_KAZE_Upright_Descriptor_128(const cv::KeyPoint
* from Agrawal et al., CenSurE: Center Surround Extremas for Realtime Feature Detection and Matching,
* ECCV 2008
*/
void KAZE_Descriptor_Invoker::Get_KAZE_Descriptor_128(const cv::KeyPoint &kpt, float *desc) const
void KAZE_Descriptor_Invoker::Get_KAZE_Descriptor_128(const KeyPoint &kpt, float *desc) const
{
float gauss_s1 = 0.0, gauss_s2 = 0.0;
float rx = 0.0, ry = 0.0, rrx = 0.0, rry = 0.0, len = 0.0, xf = 0.0, yf = 0.0, ys = 0.0, xs = 0.0;
@@ -1202,3 +1208,5 @@ void KAZE_Descriptor_Invoker::Get_KAZE_Descriptor_128(const cv::KeyPoint &kpt, f
desc[i] /= len;
}
}
}

4
modules/features2d/src/kaze/KAZEFeatures.h
View File

@@ -22,8 +22,8 @@ namespace cv
/* ************************************************************************* */
// KAZE Class Declaration
class KAZEFeatures {
class KAZEFeatures
{
private:
/// Parameters of the Nonlinear diffusion class

19
modules/features2d/src/kaze/TEvolution.h
View File

@@ -8,10 +8,13 @@
#ifndef __OPENCV_FEATURES_2D_TEVOLUTION_H__
#define __OPENCV_FEATURES_2D_TEVOLUTION_H__
namespace cv
{
/* ************************************************************************* */
/// KAZE/A-KAZE nonlinear diffusion filtering evolution
struct TEvolution {
struct TEvolution
{
TEvolution() {
etime = 0.0f;
esigma = 0.0f;
@@ -20,11 +23,11 @@ struct TEvolution {
sigma_size = 0;
}
cv::Mat Lx, Ly; ///< First order spatial derivatives
cv::Mat Lxx, Lxy, Lyy; ///< Second order spatial derivatives
cv::Mat Lt; ///< Evolution image
cv::Mat Lsmooth; ///< Smoothed image
cv::Mat Ldet; ///< Detector response
Mat Lx, Ly; ///< First order spatial derivatives
Mat Lxx, Lxy, Lyy; ///< Second order spatial derivatives
Mat Lt; ///< Evolution image
Mat Lsmooth; ///< Smoothed image
Mat Ldet; ///< Detector response
float etime; ///< Evolution time
float esigma; ///< Evolution sigma. For linear diffusion t = sigma^2 / 2
int octave; ///< Image octave
@@ -32,4 +35,6 @@ struct TEvolution {
int sigma_size; ///< Integer esigma. For computing the feature detector responses
};
}
#endif

976
modules/features2d/src/kaze/nldiffusion_functions.cpp
View File

@@ -22,502 +22,500 @@
* @author Pablo F. Alcantarilla
*/
#include "../precomp.hpp"
#include "nldiffusion_functions.h"
#include <iostream>
// Namespaces
using namespace std;
using namespace cv;
/* ************************************************************************* */
namespace cv {
namespace details {
namespace kaze {
/* ************************************************************************* */
/**
* @brief This function smoothes an image with a Gaussian kernel
* @param src Input image
* @param dst Output image
* @param ksize_x Kernel size in X-direction (horizontal)
* @param ksize_y Kernel size in Y-direction (vertical)
* @param sigma Kernel standard deviation
*/
void gaussian_2D_convolution(const cv::Mat& src, cv::Mat& dst, int ksize_x, int ksize_y, float sigma) {
int ksize_x_ = 0, ksize_y_ = 0;
// Compute an appropriate kernel size according to the specified sigma
if (sigma > ksize_x || sigma > ksize_y || ksize_x == 0 || ksize_y == 0) {
ksize_x_ = (int)ceil(2.0f*(1.0f + (sigma - 0.8f) / (0.3f)));
ksize_y_ = ksize_x_;
}
// The kernel size must be and odd number
if ((ksize_x_ % 2) == 0) {
ksize_x_ += 1;
}
if ((ksize_y_ % 2) == 0) {
ksize_y_ += 1;
}
// Perform the Gaussian Smoothing with border replication
GaussianBlur(src, dst, Size(ksize_x_, ksize_y_), sigma, sigma, BORDER_REPLICATE);
}
/* ************************************************************************* */
/**
* @brief This function computes image derivatives with Scharr kernel
* @param src Input image
* @param dst Output image
* @param xorder Derivative order in X-direction (horizontal)
* @param yorder Derivative order in Y-direction (vertical)
* @note Scharr operator approximates better rotation invariance than
* other stencils such as Sobel. See Weickert and Scharr,
* A Scheme for Coherence-Enhancing Diffusion Filtering with Optimized Rotation Invariance,
* Journal of Visual Communication and Image Representation 2002
*/
void image_derivatives_scharr(const cv::Mat& src, cv::Mat& dst, int xorder, int yorder) {
Scharr(src, dst, CV_32F, xorder, yorder, 1.0, 0, BORDER_DEFAULT);
}
/* ************************************************************************* */
/**
* @brief This function computes the Perona and Malik conductivity coefficient g1
* g1 = exp(-|dL|^2/k^2)
* @param Lx First order image derivative in X-direction (horizontal)
* @param Ly First order image derivative in Y-direction (vertical)
* @param dst Output image
* @param k Contrast factor parameter
*/
void pm_g1(const cv::Mat& Lx, const cv::Mat& Ly, cv::Mat& dst, float k) {
Size sz = Lx.size();
float inv_k = 1.0f / (k*k);
for (int y = 0; y < sz.height; y++) {
const float* Lx_row = Lx.ptr<float>(y);
const float* Ly_row = Ly.ptr<float>(y);
float* dst_row = dst.ptr<float>(y);
for (int x = 0; x < sz.width; x++) {
dst_row[x] = (-inv_k*(Lx_row[x]*Lx_row[x] + Ly_row[x]*Ly_row[x]));
}
}
exp(dst, dst);
}
/* ************************************************************************* */
/**
* @brief This function computes the Perona and Malik conductivity coefficient g2
* g2 = 1 / (1 + dL^2 / k^2)
* @param Lx First order image derivative in X-direction (horizontal)
* @param Ly First order image derivative in Y-direction (vertical)
* @param dst Output image
* @param k Contrast factor parameter
*/
void pm_g2(const cv::Mat &Lx, const cv::Mat& Ly, cv::Mat& dst, float k) {
Size sz = Lx.size();
dst.create(sz, Lx.type());
float k2inv = 1.0f / (k * k);
for(int y = 0; y < sz.height; y++) {
const float *Lx_row = Lx.ptr<float>(y);
const float *Ly_row = Ly.ptr<float>(y);
float* dst_row = dst.ptr<float>(y);
for(int x = 0; x < sz.width; x++) {
dst_row[x] = 1.0f / (1.0f + ((Lx_row[x] * Lx_row[x] + Ly_row[x] * Ly_row[x]) * k2inv));
}
}
}
/* ************************************************************************* */
/**
* @brief This function computes Weickert conductivity coefficient gw
* @param Lx First order image derivative in X-direction (horizontal)
* @param Ly First order image derivative in Y-direction (vertical)
* @param dst Output image
* @param k Contrast factor parameter
* @note For more information check the following paper: J. Weickert
* Applications of nonlinear diffusion in image processing and computer vision,
* Proceedings of Algorithmy 2000
*/
void weickert_diffusivity(const cv::Mat& Lx, const cv::Mat& Ly, cv::Mat& dst, float k) {
Size sz = Lx.size();
float inv_k = 1.0f / (k*k);
for (int y = 0; y < sz.height; y++) {
const float* Lx_row = Lx.ptr<float>(y);
const float* Ly_row = Ly.ptr<float>(y);
float* dst_row = dst.ptr<float>(y);
for (int x = 0; x < sz.width; x++) {
float dL = inv_k*(Lx_row[x]*Lx_row[x] + Ly_row[x]*Ly_row[x]);
dst_row[x] = -3.315f/(dL*dL*dL*dL);
}
}
exp(dst, dst);
dst = 1.0 - dst;
}
/* ************************************************************************* */
/**
* @brief This function computes Charbonnier conductivity coefficient gc
* gc = 1 / sqrt(1 + dL^2 / k^2)
* @param Lx First order image derivative in X-direction (horizontal)
* @param Ly First order image derivative in Y-direction (vertical)
* @param dst Output image
* @param k Contrast factor parameter
* @note For more information check the following paper: J. Weickert
* Applications of nonlinear diffusion in image processing and computer vision,
* Proceedings of Algorithmy 2000
*/
void charbonnier_diffusivity(const cv::Mat& Lx, const cv::Mat& Ly, cv::Mat& dst, float k) {
Size sz = Lx.size();
float inv_k = 1.0f / (k*k);
for (int y = 0; y < sz.height; y++) {
const float* Lx_row = Lx.ptr<float>(y);
const float* Ly_row = Ly.ptr<float>(y);
float* dst_row = dst.ptr<float>(y);
for (int x = 0; x < sz.width; x++) {
float den = sqrt(1.0f+inv_k*(Lx_row[x]*Lx_row[x] + Ly_row[x]*Ly_row[x]));
dst_row[x] = 1.0f / den;
}
}
}
/* ************************************************************************* */
/**
* @brief This function computes a good empirical value for the k contrast factor
* given an input image, the percentile (0-1), the gradient scale and the number of
* bins in the histogram
* @param img Input image
* @param perc Percentile of the image gradient histogram (0-1)
* @param gscale Scale for computing the image gradient histogram
* @param nbins Number of histogram bins
* @param ksize_x Kernel size in X-direction (horizontal) for the Gaussian smoothing kernel
* @param ksize_y Kernel size in Y-direction (vertical) for the Gaussian smoothing kernel
* @return k contrast factor
*/
float compute_k_percentile(const cv::Mat& img, float perc, float gscale, int nbins, int ksize_x, int ksize_y) {
int nbin = 0, nelements = 0, nthreshold = 0, k = 0;
float kperc = 0.0, modg = 0.0;
float npoints = 0.0;
float hmax = 0.0;
// Create the array for the histogram
std::vector<int> hist(nbins, 0);
// Create the matrices
Mat gaussian = Mat::zeros(img.rows, img.cols, CV_32F);
Mat Lx = Mat::zeros(img.rows, img.cols, CV_32F);
Mat Ly = Mat::zeros(img.rows, img.cols, CV_32F);
// Perform the Gaussian convolution
gaussian_2D_convolution(img, gaussian, ksize_x, ksize_y, gscale);
// Compute the Gaussian derivatives Lx and Ly
Scharr(gaussian, Lx, CV_32F, 1, 0, 1, 0, cv::BORDER_DEFAULT);
Scharr(gaussian, Ly, CV_32F, 0, 1, 1, 0, cv::BORDER_DEFAULT);
// Skip the borders for computing the histogram
for (int i = 1; i < gaussian.rows - 1; i++) {
const float *lx = Lx.ptr<float>(i);
const float *ly = Ly.ptr<float>(i);
for (int j = 1; j < gaussian.cols - 1; j++) {
modg = lx[j]*lx[j] + ly[j]*ly[j];
// Get the maximum
if (modg > hmax) {
hmax = modg;
}
}
}
hmax = sqrt(hmax);
// Skip the borders for computing the histogram
for (int i = 1; i < gaussian.rows - 1; i++) {
const float *lx = Lx.ptr<float>(i);
const float *ly = Ly.ptr<float>(i);
for (int j = 1; j < gaussian.cols - 1; j++) {
modg = lx[j]*lx[j] + ly[j]*ly[j];
// Find the correspondent bin
if (modg != 0.0) {
nbin = (int)floor(nbins*(sqrt(modg) / hmax));
if (nbin == nbins) {
nbin--;
}
hist[nbin]++;
npoints++;
}
}
}
// Now find the perc of the histogram percentile
nthreshold = (int)(npoints*perc);
for (k = 0; nelements < nthreshold && k < nbins; k++) {
nelements = nelements + hist[k];
}
if (nelements < nthreshold) {
kperc = 0.03f;
}
else {
kperc = hmax*((float)(k) / (float)nbins);
}
return kperc;
}
/* ************************************************************************* */
/**
* @brief This function computes Scharr image derivatives
* @param src Input image
* @param dst Output image
* @param xorder Derivative order in X-direction (horizontal)
* @param yorder Derivative order in Y-direction (vertical)
* @param scale Scale factor for the derivative size
*/
void compute_scharr_derivatives(const cv::Mat& src, cv::Mat& dst, int xorder, int yorder, int scale) {
Mat kx, ky;
compute_derivative_kernels(kx, ky, xorder, yorder, scale);
sepFilter2D(src, dst, CV_32F, kx, ky);
}
/* ************************************************************************* */
/**
* @brief Compute derivative kernels for sizes different than 3
* @param _kx Horizontal kernel ues
* @param _ky Vertical kernel values
* @param dx Derivative order in X-direction (horizontal)
* @param dy Derivative order in Y-direction (vertical)
* @param scale_ Scale factor or derivative size
*/
void compute_derivative_kernels(cv::OutputArray _kx, cv::OutputArray _ky, int dx, int dy, int scale) {
int ksize = 3 + 2 * (scale - 1);
// The standard Scharr kernel
if (scale == 1) {
getDerivKernels(_kx, _ky, dx, dy, 0, true, CV_32F);
return;
}
_kx.create(ksize, 1, CV_32F, -1, true);
_ky.create(ksize, 1, CV_32F, -1, true);
Mat kx = _kx.getMat();
Mat ky = _ky.getMat();
float w = 10.0f / 3.0f;
float norm = 1.0f / (2.0f*scale*(w + 2.0f));
for (int k = 0; k < 2; k++) {
Mat* kernel = k == 0 ? &kx : &ky;
int order = k == 0 ? dx : dy;
std::vector<float> kerI(ksize, 0.0f);
if (order == 0) {
kerI[0] = norm, kerI[ksize / 2] = w*norm, kerI[ksize - 1] = norm;
}
else if (order == 1) {
kerI[0] = -1, kerI[ksize / 2] = 0, kerI[ksize - 1] = 1;
}
Mat temp(kernel->rows, kernel->cols, CV_32F, &kerI[0]);
temp.copyTo(*kernel);
}
}
class Nld_Step_Scalar_Invoker : public cv::ParallelLoopBody
{
public:
Nld_Step_Scalar_Invoker(cv::Mat& Ld, const cv::Mat& c, cv::Mat& Lstep, float _stepsize)
: _Ld(&Ld)
, _c(&c)
, _Lstep(&Lstep)
, stepsize(_stepsize)
{
}
virtual ~Nld_Step_Scalar_Invoker()
{
}
void operator()(const cv::Range& range) const
{
cv::Mat& Ld = *_Ld;
const cv::Mat& c = *_c;
cv::Mat& Lstep = *_Lstep;
for (int i = range.start; i < range.end; i++)
{
const float *c_prev = c.ptr<float>(i - 1);
const float *c_curr = c.ptr<float>(i);
const float *c_next = c.ptr<float>(i + 1);
const float *ld_prev = Ld.ptr<float>(i - 1);
const float *ld_curr = Ld.ptr<float>(i);
const float *ld_next = Ld.ptr<float>(i + 1);
float *dst = Lstep.ptr<float>(i);
for (int j = 1; j < Lstep.cols - 1; j++)
{
float xpos = (c_curr[j] + c_curr[j+1])*(ld_curr[j+1] - ld_curr[j]);
float xneg = (c_curr[j-1] + c_curr[j]) *(ld_curr[j] - ld_curr[j-1]);
float ypos = (c_curr[j] + c_next[j]) *(ld_next[j] - ld_curr[j]);
float yneg = (c_prev[j] + c_curr[j]) *(ld_curr[j] - ld_prev[j]);
dst[j] = 0.5f*stepsize*(xpos - xneg + ypos - yneg);
}
}
}
private:
cv::Mat * _Ld;
const cv::Mat * _c;
cv::Mat * _Lstep;
float stepsize;
};
/* ************************************************************************* */
/**
* @brief This function performs a scalar non-linear diffusion step
* @param Ld2 Output image in the evolution
* @param c Conductivity image
* @param Lstep Previous image in the evolution
* @param stepsize The step size in time units
* @note Forward Euler Scheme 3x3 stencil
* The function c is a scalar value that depends on the gradient norm
* dL_by_ds = d(c dL_by_dx)_by_dx + d(c dL_by_dy)_by_dy
*/
void nld_step_scalar(cv::Mat& Ld, const cv::Mat& c, cv::Mat& Lstep, float stepsize) {
cv::parallel_for_(cv::Range(1, Lstep.rows - 1), Nld_Step_Scalar_Invoker(Ld, c, Lstep, stepsize), (double)Ld.total()/(1 << 16));
float xneg, xpos, yneg, ypos;
float* dst = Lstep.ptr<float>(0);
const float* cprv = NULL;
const float* ccur = c.ptr<float>(0);
const float* cnxt = c.ptr<float>(1);
const float* ldprv = NULL;
const float* ldcur = Ld.ptr<float>(0);
const float* ldnxt = Ld.ptr<float>(1);
for (int j = 1; j < Lstep.cols - 1; j++) {
xpos = (ccur[j] + ccur[j+1]) * (ldcur[j+1] - ldcur[j]);
xneg = (ccur[j-1] + ccur[j]) * (ldcur[j] - ldcur[j-1]);
ypos = (ccur[j] + cnxt[j]) * (ldnxt[j] - ldcur[j]);
dst[j] = 0.5f*stepsize*(xpos - xneg + ypos);
}
dst = Lstep.ptr<float>(Lstep.rows - 1);
ccur = c.ptr<float>(Lstep.rows - 1);
cprv = c.ptr<float>(Lstep.rows - 2);
ldcur = Ld.ptr<float>(Lstep.rows - 1);
ldprv = Ld.ptr<float>(Lstep.rows - 2);
for (int j = 1; j < Lstep.cols - 1; j++) {
xpos = (ccur[j] + ccur[j+1]) * (ldcur[j+1] - ldcur[j]);
xneg = (ccur[j-1] + ccur[j]) * (ldcur[j] - ldcur[j-1]);
yneg = (cprv[j] + ccur[j]) * (ldcur[j] - ldprv[j]);
dst[j] = 0.5f*stepsize*(xpos - xneg - yneg);
}
ccur = c.ptr<float>(1);
ldcur = Ld.ptr<float>(1);
cprv = c.ptr<float>(0);
ldprv = Ld.ptr<float>(0);
int r0 = Lstep.cols - 1;
int r1 = Lstep.cols - 2;
for (int i = 1; i < Lstep.rows - 1; i++) {
cnxt = c.ptr<float>(i + 1);
ldnxt = Ld.ptr<float>(i + 1);
dst = Lstep.ptr<float>(i);
xpos = (ccur[0] + ccur[1]) * (ldcur[1] - ldcur[0]);
ypos = (ccur[0] + cnxt[0]) * (ldnxt[0] - ldcur[0]);
yneg = (cprv[0] + ccur[0]) * (ldcur[0] - ldprv[0]);
dst[0] = 0.5f*stepsize*(xpos + ypos - yneg);
xneg = (ccur[r1] + ccur[r0]) * (ldcur[r0] - ldcur[r1]);
ypos = (ccur[r0] + cnxt[r0]) * (ldnxt[r0] - ldcur[r0]);
yneg = (cprv[r0] + ccur[r0]) * (ldcur[r0] - ldprv[r0]);
dst[r0] = 0.5f*stepsize*(-xneg + ypos - yneg);
cprv = ccur;
ccur = cnxt;
ldprv = ldcur;
ldcur = ldnxt;
}
Ld += Lstep;
}
/* ************************************************************************* */
/**
* @brief This function downsamples the input image using OpenCV resize
* @param img Input image to be downsampled
* @param dst Output image with half of the resolution of the input image
*/
void halfsample_image(const cv::Mat& src, cv::Mat& dst) {
// Make sure the destination image is of the right size
CV_Assert(src.cols / 2 == dst.cols);
CV_Assert(src.rows / 2 == dst.rows);
resize(src, dst, dst.size(), 0, 0, cv::INTER_AREA);
}
/* ************************************************************************* */
/**
* @brief This function checks if a given pixel is a maximum in a local neighbourhood
* @param img Input image where we will perform the maximum search
* @param dsize Half size of the neighbourhood
* @param value Response value at (x,y) position
* @param row Image row coordinate
* @param col Image column coordinate
* @param same_img Flag to indicate if the image value at (x,y) is in the input image
* @return 1->is maximum, 0->otherwise
*/
bool check_maximum_neighbourhood(const cv::Mat& img, int dsize, float value, int row, int col, bool same_img) {
bool response = true;
for (int i = row - dsize; i <= row + dsize; i++) {
for (int j = col - dsize; j <= col + dsize; j++) {
if (i >= 0 && i < img.rows && j >= 0 && j < img.cols) {
if (same_img == true) {
if (i != row || j != col) {
if ((*(img.ptr<float>(i)+j)) > value) {
response = false;
return response;
}
}
}
else {
if ((*(img.ptr<float>(i)+j)) > value) {
response = false;
return response;
}
}
}
}
}
return response;
}
namespace cv
{
using namespace std;
/* ************************************************************************* */
/**
* @brief This function smoothes an image with a Gaussian kernel
* @param src Input image
* @param dst Output image
* @param ksize_x Kernel size in X-direction (horizontal)
* @param ksize_y Kernel size in Y-direction (vertical)
* @param sigma Kernel standard deviation
*/
void gaussian_2D_convolution(const cv::Mat& src, cv::Mat& dst, int ksize_x, int ksize_y, float sigma) {
int ksize_x_ = 0, ksize_y_ = 0;
// Compute an appropriate kernel size according to the specified sigma
if (sigma > ksize_x || sigma > ksize_y || ksize_x == 0 || ksize_y == 0) {
ksize_x_ = (int)ceil(2.0f*(1.0f + (sigma - 0.8f) / (0.3f)));
ksize_y_ = ksize_x_;
}
// The kernel size must be and odd number
if ((ksize_x_ % 2) == 0) {
ksize_x_ += 1;
}
if ((ksize_y_ % 2) == 0) {
ksize_y_ += 1;
}
// Perform the Gaussian Smoothing with border replication
GaussianBlur(src, dst, Size(ksize_x_, ksize_y_), sigma, sigma, BORDER_REPLICATE);
}
/* ************************************************************************* */
/**
* @brief This function computes image derivatives with Scharr kernel
* @param src Input image
* @param dst Output image
* @param xorder Derivative order in X-direction (horizontal)
* @param yorder Derivative order in Y-direction (vertical)
* @note Scharr operator approximates better rotation invariance than
* other stencils such as Sobel. See Weickert and Scharr,
* A Scheme for Coherence-Enhancing Diffusion Filtering with Optimized Rotation Invariance,
* Journal of Visual Communication and Image Representation 2002
*/
void image_derivatives_scharr(const cv::Mat& src, cv::Mat& dst, int xorder, int yorder) {
Scharr(src, dst, CV_32F, xorder, yorder, 1.0, 0, BORDER_DEFAULT);
}
/* ************************************************************************* */
/**
* @brief This function computes the Perona and Malik conductivity coefficient g1
* g1 = exp(-|dL|^2/k^2)
* @param Lx First order image derivative in X-direction (horizontal)
* @param Ly First order image derivative in Y-direction (vertical)
* @param dst Output image
* @param k Contrast factor parameter
*/
void pm_g1(const cv::Mat& Lx, const cv::Mat& Ly, cv::Mat& dst, float k) {
Size sz = Lx.size();
float inv_k = 1.0f / (k*k);
for (int y = 0; y < sz.height; y++) {
const float* Lx_row = Lx.ptr<float>(y);
const float* Ly_row = Ly.ptr<float>(y);
float* dst_row = dst.ptr<float>(y);
for (int x = 0; x < sz.width; x++) {
dst_row[x] = (-inv_k*(Lx_row[x]*Lx_row[x] + Ly_row[x]*Ly_row[x]));
}
}
exp(dst, dst);
}
/* ************************************************************************* */
/**
* @brief This function computes the Perona and Malik conductivity coefficient g2
* g2 = 1 / (1 + dL^2 / k^2)
* @param Lx First order image derivative in X-direction (horizontal)
* @param Ly First order image derivative in Y-direction (vertical)
* @param dst Output image
* @param k Contrast factor parameter
*/
void pm_g2(const cv::Mat &Lx, const cv::Mat& Ly, cv::Mat& dst, float k) {
Size sz = Lx.size();
dst.create(sz, Lx.type());
float k2inv = 1.0f / (k * k);
for(int y = 0; y < sz.height; y++) {
const float *Lx_row = Lx.ptr<float>(y);
const float *Ly_row = Ly.ptr<float>(y);
float* dst_row = dst.ptr<float>(y);
for(int x = 0; x < sz.width; x++) {
dst_row[x] = 1.0f / (1.0f + ((Lx_row[x] * Lx_row[x] + Ly_row[x] * Ly_row[x]) * k2inv));
}
}
}
/* ************************************************************************* */
/**
* @brief This function computes Weickert conductivity coefficient gw
* @param Lx First order image derivative in X-direction (horizontal)
* @param Ly First order image derivative in Y-direction (vertical)
* @param dst Output image
* @param k Contrast factor parameter
* @note For more information check the following paper: J. Weickert
* Applications of nonlinear diffusion in image processing and computer vision,
* Proceedings of Algorithmy 2000
*/
void weickert_diffusivity(const cv::Mat& Lx, const cv::Mat& Ly, cv::Mat& dst, float k) {
Size sz = Lx.size();
float inv_k = 1.0f / (k*k);
for (int y = 0; y < sz.height; y++) {
const float* Lx_row = Lx.ptr<float>(y);
const float* Ly_row = Ly.ptr<float>(y);
float* dst_row = dst.ptr<float>(y);
for (int x = 0; x < sz.width; x++) {
float dL = inv_k*(Lx_row[x]*Lx_row[x] + Ly_row[x]*Ly_row[x]);
dst_row[x] = -3.315f/(dL*dL*dL*dL);
}
}
exp(dst, dst);
dst = 1.0 - dst;
}
/* ************************************************************************* */
/**
* @brief This function computes Charbonnier conductivity coefficient gc
* gc = 1 / sqrt(1 + dL^2 / k^2)
* @param Lx First order image derivative in X-direction (horizontal)
* @param Ly First order image derivative in Y-direction (vertical)
* @param dst Output image
* @param k Contrast factor parameter
* @note For more information check the following paper: J. Weickert
* Applications of nonlinear diffusion in image processing and computer vision,
* Proceedings of Algorithmy 2000
*/
void charbonnier_diffusivity(const cv::Mat& Lx, const cv::Mat& Ly, cv::Mat& dst, float k) {
Size sz = Lx.size();
float inv_k = 1.0f / (k*k);
for (int y = 0; y < sz.height; y++) {
const float* Lx_row = Lx.ptr<float>(y);
const float* Ly_row = Ly.ptr<float>(y);
float* dst_row = dst.ptr<float>(y);
for (int x = 0; x < sz.width; x++) {
float den = sqrt(1.0f+inv_k*(Lx_row[x]*Lx_row[x] + Ly_row[x]*Ly_row[x]));
dst_row[x] = 1.0f / den;
}
}
}
/* ************************************************************************* */
/**
* @brief This function computes a good empirical value for the k contrast factor
* given an input image, the percentile (0-1), the gradient scale and the number of
* bins in the histogram
* @param img Input image
* @param perc Percentile of the image gradient histogram (0-1)
* @param gscale Scale for computing the image gradient histogram
* @param nbins Number of histogram bins
* @param ksize_x Kernel size in X-direction (horizontal) for the Gaussian smoothing kernel
* @param ksize_y Kernel size in Y-direction (vertical) for the Gaussian smoothing kernel
* @return k contrast factor
*/
float compute_k_percentile(const cv::Mat& img, float perc, float gscale, int nbins, int ksize_x, int ksize_y) {
int nbin = 0, nelements = 0, nthreshold = 0, k = 0;
float kperc = 0.0, modg = 0.0;
float npoints = 0.0;
float hmax = 0.0;
// Create the array for the histogram
std::vector<int> hist(nbins, 0);
// Create the matrices
Mat gaussian = Mat::zeros(img.rows, img.cols, CV_32F);
Mat Lx = Mat::zeros(img.rows, img.cols, CV_32F);
Mat Ly = Mat::zeros(img.rows, img.cols, CV_32F);
// Perform the Gaussian convolution
gaussian_2D_convolution(img, gaussian, ksize_x, ksize_y, gscale);
// Compute the Gaussian derivatives Lx and Ly
Scharr(gaussian, Lx, CV_32F, 1, 0, 1, 0, cv::BORDER_DEFAULT);
Scharr(gaussian, Ly, CV_32F, 0, 1, 1, 0, cv::BORDER_DEFAULT);
// Skip the borders for computing the histogram
for (int i = 1; i < gaussian.rows - 1; i++) {
const float *lx = Lx.ptr<float>(i);
const float *ly = Ly.ptr<float>(i);
for (int j = 1; j < gaussian.cols - 1; j++) {
modg = lx[j]*lx[j] + ly[j]*ly[j];
// Get the maximum
if (modg > hmax) {
hmax = modg;
}
}
}
hmax = sqrt(hmax);
// Skip the borders for computing the histogram
for (int i = 1; i < gaussian.rows - 1; i++) {
const float *lx = Lx.ptr<float>(i);
const float *ly = Ly.ptr<float>(i);
for (int j = 1; j < gaussian.cols - 1; j++) {
modg = lx[j]*lx[j] + ly[j]*ly[j];
// Find the correspondent bin
if (modg != 0.0) {
nbin = (int)floor(nbins*(sqrt(modg) / hmax));
if (nbin == nbins) {
nbin--;
}
hist[nbin]++;
npoints++;
}
}
}
// Now find the perc of the histogram percentile
nthreshold = (int)(npoints*perc);
for (k = 0; nelements < nthreshold && k < nbins; k++) {
nelements = nelements + hist[k];
}
if (nelements < nthreshold) {
kperc = 0.03f;
}
else {
kperc = hmax*((float)(k) / (float)nbins);
}
return kperc;
}
/* ************************************************************************* */
/**
* @brief This function computes Scharr image derivatives
* @param src Input image
* @param dst Output image
* @param xorder Derivative order in X-direction (horizontal)
* @param yorder Derivative order in Y-direction (vertical)
* @param scale Scale factor for the derivative size
*/
void compute_scharr_derivatives(const cv::Mat& src, cv::Mat& dst, int xorder, int yorder, int scale) {
Mat kx, ky;
compute_derivative_kernels(kx, ky, xorder, yorder, scale);
sepFilter2D(src, dst, CV_32F, kx, ky);
}
/* ************************************************************************* */
/**
* @brief Compute derivative kernels for sizes different than 3
* @param _kx Horizontal kernel ues
* @param _ky Vertical kernel values
* @param dx Derivative order in X-direction (horizontal)
* @param dy Derivative order in Y-direction (vertical)
* @param scale_ Scale factor or derivative size
*/
void compute_derivative_kernels(cv::OutputArray _kx, cv::OutputArray _ky, int dx, int dy, int scale) {
int ksize = 3 + 2 * (scale - 1);
// The standard Scharr kernel
if (scale == 1) {
getDerivKernels(_kx, _ky, dx, dy, 0, true, CV_32F);
return;
}
_kx.create(ksize, 1, CV_32F, -1, true);
_ky.create(ksize, 1, CV_32F, -1, true);
Mat kx = _kx.getMat();
Mat ky = _ky.getMat();
float w = 10.0f / 3.0f;
float norm = 1.0f / (2.0f*scale*(w + 2.0f));
for (int k = 0; k < 2; k++) {
Mat* kernel = k == 0 ? &kx : &ky;
int order = k == 0 ? dx : dy;
std::vector<float> kerI(ksize, 0.0f);
if (order == 0) {
kerI[0] = norm, kerI[ksize / 2] = w*norm, kerI[ksize - 1] = norm;
}
else if (order == 1) {
kerI[0] = -1, kerI[ksize / 2] = 0, kerI[ksize - 1] = 1;
}
Mat temp(kernel->rows, kernel->cols, CV_32F, &kerI[0]);
temp.copyTo(*kernel);
}
}
class Nld_Step_Scalar_Invoker : public cv::ParallelLoopBody
{
public:
Nld_Step_Scalar_Invoker(cv::Mat& Ld, const cv::Mat& c, cv::Mat& Lstep, float _stepsize)
: _Ld(&Ld)
, _c(&c)
, _Lstep(&Lstep)
, stepsize(_stepsize)
{
}
virtual ~Nld_Step_Scalar_Invoker()
{
}
void operator()(const cv::Range& range) const
{
cv::Mat& Ld = *_Ld;
const cv::Mat& c = *_c;
cv::Mat& Lstep = *_Lstep;
for (int i = range.start; i < range.end; i++)
{
const float *c_prev = c.ptr<float>(i - 1);
const float *c_curr = c.ptr<float>(i);
const float *c_next = c.ptr<float>(i + 1);
const float *ld_prev = Ld.ptr<float>(i - 1);
const float *ld_curr = Ld.ptr<float>(i);
const float *ld_next = Ld.ptr<float>(i + 1);
float *dst = Lstep.ptr<float>(i);
for (int j = 1; j < Lstep.cols - 1; j++)
{
float xpos = (c_curr[j] + c_curr[j+1])*(ld_curr[j+1] - ld_curr[j]);
float xneg = (c_curr[j-1] + c_curr[j]) *(ld_curr[j] - ld_curr[j-1]);
float ypos = (c_curr[j] + c_next[j]) *(ld_next[j] - ld_curr[j]);
float yneg = (c_prev[j] + c_curr[j]) *(ld_curr[j] - ld_prev[j]);
dst[j] = 0.5f*stepsize*(xpos - xneg + ypos - yneg);
}
}
}
private:
cv::Mat * _Ld;
const cv::Mat * _c;
cv::Mat * _Lstep;
float stepsize;
};
/* ************************************************************************* */
/**
* @brief This function performs a scalar non-linear diffusion step
* @param Ld2 Output image in the evolution
* @param c Conductivity image
* @param Lstep Previous image in the evolution
* @param stepsize The step size in time units
* @note Forward Euler Scheme 3x3 stencil
* The function c is a scalar value that depends on the gradient norm
* dL_by_ds = d(c dL_by_dx)_by_dx + d(c dL_by_dy)_by_dy
*/
void nld_step_scalar(cv::Mat& Ld, const cv::Mat& c, cv::Mat& Lstep, float stepsize) {
cv::parallel_for_(cv::Range(1, Lstep.rows - 1), Nld_Step_Scalar_Invoker(Ld, c, Lstep, stepsize), (double)Ld.total()/(1 << 16));
float xneg, xpos, yneg, ypos;
float* dst = Lstep.ptr<float>(0);
const float* cprv = NULL;
const float* ccur = c.ptr<float>(0);
const float* cnxt = c.ptr<float>(1);
const float* ldprv = NULL;
const float* ldcur = Ld.ptr<float>(0);
const float* ldnxt = Ld.ptr<float>(1);
for (int j = 1; j < Lstep.cols - 1; j++) {
xpos = (ccur[j] + ccur[j+1]) * (ldcur[j+1] - ldcur[j]);
xneg = (ccur[j-1] + ccur[j]) * (ldcur[j] - ldcur[j-1]);
ypos = (ccur[j] + cnxt[j]) * (ldnxt[j] - ldcur[j]);
dst[j] = 0.5f*stepsize*(xpos - xneg + ypos);
}
dst = Lstep.ptr<float>(Lstep.rows - 1);
ccur = c.ptr<float>(Lstep.rows - 1);
cprv = c.ptr<float>(Lstep.rows - 2);
ldcur = Ld.ptr<float>(Lstep.rows - 1);
ldprv = Ld.ptr<float>(Lstep.rows - 2);
for (int j = 1; j < Lstep.cols - 1; j++) {
xpos = (ccur[j] + ccur[j+1]) * (ldcur[j+1] - ldcur[j]);
xneg = (ccur[j-1] + ccur[j]) * (ldcur[j] - ldcur[j-1]);
yneg = (cprv[j] + ccur[j]) * (ldcur[j] - ldprv[j]);
dst[j] = 0.5f*stepsize*(xpos - xneg - yneg);
}
ccur = c.ptr<float>(1);
ldcur = Ld.ptr<float>(1);
cprv = c.ptr<float>(0);
ldprv = Ld.ptr<float>(0);
int r0 = Lstep.cols - 1;
int r1 = Lstep.cols - 2;
for (int i = 1; i < Lstep.rows - 1; i++) {
cnxt = c.ptr<float>(i + 1);
ldnxt = Ld.ptr<float>(i + 1);
dst = Lstep.ptr<float>(i);
xpos = (ccur[0] + ccur[1]) * (ldcur[1] - ldcur[0]);
ypos = (ccur[0] + cnxt[0]) * (ldnxt[0] - ldcur[0]);
yneg = (cprv[0] + ccur[0]) * (ldcur[0] - ldprv[0]);
dst[0] = 0.5f*stepsize*(xpos + ypos - yneg);
xneg = (ccur[r1] + ccur[r0]) * (ldcur[r0] - ldcur[r1]);
ypos = (ccur[r0] + cnxt[r0]) * (ldnxt[r0] - ldcur[r0]);
yneg = (cprv[r0] + ccur[r0]) * (ldcur[r0] - ldprv[r0]);
dst[r0] = 0.5f*stepsize*(-xneg + ypos - yneg);
cprv = ccur;
ccur = cnxt;
ldprv = ldcur;
ldcur = ldnxt;
}
Ld += Lstep;
}
/* ************************************************************************* */
/**
* @brief This function downsamples the input image using OpenCV resize
* @param img Input image to be downsampled
* @param dst Output image with half of the resolution of the input image
*/
void halfsample_image(const cv::Mat& src, cv::Mat& dst) {
// Make sure the destination image is of the right size
CV_Assert(src.cols / 2 == dst.cols);
CV_Assert(src.rows / 2 == dst.rows);
resize(src, dst, dst.size(), 0, 0, cv::INTER_AREA);
}
/* ************************************************************************* */
/**
* @brief This function checks if a given pixel is a maximum in a local neighbourhood
* @param img Input image where we will perform the maximum search
* @param dsize Half size of the neighbourhood
* @param value Response value at (x,y) position
* @param row Image row coordinate
* @param col Image column coordinate
* @param same_img Flag to indicate if the image value at (x,y) is in the input image
* @return 1->is maximum, 0->otherwise
*/
bool check_maximum_neighbourhood(const cv::Mat& img, int dsize, float value, int row, int col, bool same_img) {
bool response = true;
for (int i = row - dsize; i <= row + dsize; i++) {
for (int j = col - dsize; j <= col + dsize; j++) {
if (i >= 0 && i < img.rows && j >= 0 && j < img.cols) {
if (same_img == true) {
if (i != row || j != col) {
if ((*(img.ptr<float>(i)+j)) > value) {
response = false;
return response;
}
}
}
else {
if ((*(img.ptr<float>(i)+j)) > value) {
response = false;
return response;
}
}
}
}
}
return response;
}
}

48
modules/features2d/src/kaze/nldiffusion_functions.h
View File

@@ -11,43 +11,37 @@
#ifndef __OPENCV_FEATURES_2D_NLDIFFUSION_FUNCTIONS_H__
#define __OPENCV_FEATURES_2D_NLDIFFUSION_FUNCTIONS_H__
/* ************************************************************************* */
// Includes
#include "../precomp.hpp"
/* ************************************************************************* */
// Declaration of functions
namespace cv {
namespace details {
namespace kaze {
namespace cv
{
// Gaussian 2D convolution
void gaussian_2D_convolution(const cv::Mat& src, cv::Mat& dst, int ksize_x, int ksize_y, float sigma);
// Gaussian 2D convolution
void gaussian_2D_convolution(const cv::Mat& src, cv::Mat& dst, int ksize_x, int ksize_y, float sigma);
// Diffusivity functions
void pm_g1(const cv::Mat& Lx, const cv::Mat& Ly, cv::Mat& dst, float k);
void pm_g2(const cv::Mat& Lx, const cv::Mat& Ly, cv::Mat& dst, float k);
void weickert_diffusivity(const cv::Mat& Lx, const cv::Mat& Ly, cv::Mat& dst, float k);
void charbonnier_diffusivity(const cv::Mat& Lx, const cv::Mat& Ly, cv::Mat& dst, float k);
// Diffusivity functions
void pm_g1(const cv::Mat& Lx, const cv::Mat& Ly, cv::Mat& dst, float k);
void pm_g2(const cv::Mat& Lx, const cv::Mat& Ly, cv::Mat& dst, float k);
void weickert_diffusivity(const cv::Mat& Lx, const cv::Mat& Ly, cv::Mat& dst, float k);
void charbonnier_diffusivity(const cv::Mat& Lx, const cv::Mat& Ly, cv::Mat& dst, float k);
float compute_k_percentile(const cv::Mat& img, float perc, float gscale, int nbins, int ksize_x, int ksize_y);
float compute_k_percentile(const cv::Mat& img, float perc, float gscale, int nbins, int ksize_x, int ksize_y);
// Image derivatives
void compute_scharr_derivatives(const cv::Mat& src, cv::Mat& dst, int xorder, int yorder, int scale);
void compute_derivative_kernels(cv::OutputArray _kx, cv::OutputArray _ky, int dx, int dy, int scale);
void image_derivatives_scharr(const cv::Mat& src, cv::Mat& dst, int xorder, int yorder);
// Image derivatives
void compute_scharr_derivatives(const cv::Mat& src, cv::Mat& dst, int xorder, int yorder, int scale);
void compute_derivative_kernels(cv::OutputArray _kx, cv::OutputArray _ky, int dx, int dy, int scale);
void image_derivatives_scharr(const cv::Mat& src, cv::Mat& dst, int xorder, int yorder);
// Nonlinear diffusion filtering scalar step
void nld_step_scalar(cv::Mat& Ld, const cv::Mat& c, cv::Mat& Lstep, float stepsize);
// Nonlinear diffusion filtering scalar step
void nld_step_scalar(cv::Mat& Ld, const cv::Mat& c, cv::Mat& Lstep, float stepsize);
// For non-maxima suppresion
bool check_maximum_neighbourhood(const cv::Mat& img, int dsize, float value, int row, int col, bool same_img);
// For non-maxima suppresion
bool check_maximum_neighbourhood(const cv::Mat& img, int dsize, float value, int row, int col, bool same_img);
// Image downsampling
void halfsample_image(const cv::Mat& src, cv::Mat& dst);
// Image downsampling
void halfsample_image(const cv::Mat& src, cv::Mat& dst);
}
}
}
#endif