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

refactor CUDA ORB feature detector/extractor algorithm:

Browse Source 
use new abstract interface and hidden implementation
This commit is contained in:
Vladislav Vinogradov 2015-01-13 10:40:58 +03:00
parent 554ddd2ec4
commit f960a5707d
5 changed files with 447 additions and 442 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

125
modules/cudafeatures2d/include/opencv2/cudafeatures2d.hpp
View File

@ -284,9 +284,11 @@ public:
virtual int getMaxNumPoints() const = 0;
};
/** @brief Class for extracting ORB features and descriptors from an image. :
*/
class CV_EXPORTS ORB_CUDA
//
// ORB
//
class CV_EXPORTS ORB : public cv::ORB, public Feature2DAsync
{
public:
enum
@ -300,113 +302,20 @@ public:
ROWS_COUNT
};
enum
{
DEFAULT_FAST_THRESHOLD = 20
};
/** @brief Constructor.
@param nFeatures The number of desired features.
@param scaleFactor Coefficient by which we divide the dimensions from one scale pyramid level to
the next.
@param nLevels The number of levels in the scale pyramid.
@param edgeThreshold How far from the boundary the points should be.
@param firstLevel The level at which the image is given. If 1, that means we will also look at the
image scaleFactor times bigger.
@param WTA_K
@param scoreType
@param patchSize
*/
explicit ORB_CUDA(int nFeatures = 500, float scaleFactor = 1.2f, int nLevels = 8, int edgeThreshold = 31,
int firstLevel = 0, int WTA_K = 2, int scoreType = 0, int patchSize = 31);
/** @overload */
void operator()(const GpuMat& image, const GpuMat& mask, std::vector<KeyPoint>& keypoints);
/** @overload */
void operator()(const GpuMat& image, const GpuMat& mask, GpuMat& keypoints);
/** @brief Detects keypoints and computes descriptors for them.
@param image Input 8-bit grayscale image.
@param mask Optional input mask that marks the regions where we should detect features.
@param keypoints The input/output vector of keypoints. Can be stored both in CPU and GPU memory.
For GPU memory:
- keypoints.ptr\<float\>(X_ROW)[i] contains x coordinate of the i'th feature.
- keypoints.ptr\<float\>(Y_ROW)[i] contains y coordinate of the i'th feature.
- keypoints.ptr\<float\>(RESPONSE_ROW)[i] contains the response of the i'th feature.
- keypoints.ptr\<float\>(ANGLE_ROW)[i] contains orientation of the i'th feature.
- keypoints.ptr\<float\>(OCTAVE_ROW)[i] contains the octave of the i'th feature.
- keypoints.ptr\<float\>(SIZE_ROW)[i] contains the size of the i'th feature.
@param descriptors Computed descriptors. if blurForDescriptor is true, image will be blurred
before descriptors calculation.
*/
void operator()(const GpuMat& image, const GpuMat& mask, std::vector<KeyPoint>& keypoints, GpuMat& descriptors);
/** @overload */
void operator()(const GpuMat& image, const GpuMat& mask, GpuMat& keypoints, GpuMat& descriptors);
/** @brief Download keypoints from GPU to CPU memory.
*/
static void downloadKeyPoints(const GpuMat& d_keypoints, std::vector<KeyPoint>& keypoints);
/** @brief Converts keypoints from CUDA representation to vector of KeyPoint.
*/
static void convertKeyPoints(const Mat& d_keypoints, std::vector<KeyPoint>& keypoints);
//! returns the descriptor size in bytes
inline int descriptorSize() const { return kBytes; }
inline void setFastParams(int threshold, bool nonmaxSuppression = true)
{
fastDetector_->setThreshold(threshold);
fastDetector_->setNonmaxSuppression(nonmaxSuppression);
}
/** @brief Releases inner buffer memory.
*/
void release();
static Ptr<ORB> create(int nfeatures=500,
float scaleFactor=1.2f,
int nlevels=8,
int edgeThreshold=31,
int firstLevel=0,
int WTA_K=2,
int scoreType=ORB::HARRIS_SCORE,
int patchSize=31,
int fastThreshold=20,
bool blurForDescriptor=false);
//! if true, image will be blurred before descriptors calculation
bool blurForDescriptor;
private:
enum { kBytes = 32 };
void buildScalePyramids(const GpuMat& image, const GpuMat& mask);
void computeKeyPointsPyramid();
void computeDescriptors(GpuMat& descriptors);
void mergeKeyPoints(GpuMat& keypoints);
int nFeatures_;
float scaleFactor_;
int nLevels_;
int edgeThreshold_;
int firstLevel_;
int WTA_K_;
int scoreType_;
int patchSize_;
//! The number of desired features per scale
std::vector<size_t> n_features_per_level_;
//! Points to compute BRIEF descriptors from
GpuMat pattern_;
std::vector<GpuMat> imagePyr_;
std::vector<GpuMat> maskPyr_;
GpuMat buf_;
std::vector<GpuMat> keyPointsPyr_;
std::vector<int> keyPointsCount_;
Ptr<cv::cuda::FastFeatureDetector> fastDetector_;
Ptr<cuda::Filter> blurFilter;
GpuMat d_keypoints_;
virtual void setBlurForDescriptor(bool blurForDescriptor) = 0;
virtual bool getBlurForDescriptor() const = 0;
};
//! @}

6
modules/cudafeatures2d/perf/perf_features2d.cpp
View File

@ -109,15 +109,15 @@ PERF_TEST_P(Image_NFeatures, ORB,
if (PERF_RUN_CUDA())
{
cv::cuda::ORB_CUDA d_orb(nFeatures);
cv::Ptr<cv::cuda::ORB> d_orb = cv::cuda::ORB::create(nFeatures);
const cv::cuda::GpuMat d_img(img);
cv::cuda::GpuMat d_keypoints, d_descriptors;
TEST_CYCLE() d_orb(d_img, cv::cuda::GpuMat(), d_keypoints, d_descriptors);
TEST_CYCLE() d_orb->detectAndComputeAsync(d_img, cv::noArray(), d_keypoints, d_descriptors);
std::vector<cv::KeyPoint> gpu_keypoints;
d_orb.downloadKeyPoints(d_keypoints, gpu_keypoints);
d_orb->convert(d_keypoints, gpu_keypoints);
cv::Mat gpu_descriptors(d_descriptors);

327
modules/cudafeatures2d/src/orb.cpp
View File

@ -47,18 +47,7 @@ using namespace cv::cuda;
#if !defined (HAVE_CUDA) || defined (CUDA_DISABLER)
cv::cuda::ORB_CUDA::ORB_CUDA(int, float, int, int, int, int, int, int) : fastDetector_(20) { throw_no_cuda(); }
void cv::cuda::ORB_CUDA::operator()(const GpuMat&, const GpuMat&, std::vector<KeyPoint>&) { throw_no_cuda(); }
void cv::cuda::ORB_CUDA::operator()(const GpuMat&, const GpuMat&, GpuMat&) { throw_no_cuda(); }
void cv::cuda::ORB_CUDA::operator()(const GpuMat&, const GpuMat&, std::vector<KeyPoint>&, GpuMat&) { throw_no_cuda(); }
void cv::cuda::ORB_CUDA::operator()(const GpuMat&, const GpuMat&, GpuMat&, GpuMat&) { throw_no_cuda(); }
void cv::cuda::ORB_CUDA::downloadKeyPoints(const GpuMat&, std::vector<KeyPoint>&) { throw_no_cuda(); }
void cv::cuda::ORB_CUDA::convertKeyPoints(const Mat&, std::vector<KeyPoint>&) { throw_no_cuda(); }
void cv::cuda::ORB_CUDA::release() { throw_no_cuda(); }
void cv::cuda::ORB_CUDA::buildScalePyramids(const GpuMat&, const GpuMat&) { throw_no_cuda(); }
void cv::cuda::ORB_CUDA::computeKeyPointsPyramid() { throw_no_cuda(); }
void cv::cuda::ORB_CUDA::computeDescriptors(GpuMat&) { throw_no_cuda(); }
void cv::cuda::ORB_CUDA::mergeKeyPoints(GpuMat&) { throw_no_cuda(); }
Ptr<cv::cuda::ORB> cv::cuda::ORB::create(int, float, int, int, int, int, int, int, int, bool) { throw_no_cuda(); return Ptr<cv::cuda::ORB>(); }
#else /* !defined (HAVE_CUDA) */
@ -346,7 +335,100 @@ namespace
-1,-6, 0,-11/*mean (0.127148), correlation (0.547401)*/
};
void initializeOrbPattern(const Point* pattern0, Mat& pattern, int ntuples, int tupleSize, int poolSize)
class ORB_Impl : public cv::cuda::ORB
{
public:
ORB_Impl(int nfeatures,
float scaleFactor,
int nlevels,
int edgeThreshold,
int firstLevel,
int WTA_K,
int scoreType,
int patchSize,
int fastThreshold,
bool blurForDescriptor);
virtual void detectAndCompute(InputArray _image, InputArray _mask, std::vector<KeyPoint>& keypoints, OutputArray _descriptors, bool useProvidedKeypoints);
virtual void detectAndComputeAsync(InputArray _image, InputArray _mask, OutputArray _keypoints, OutputArray _descriptors, bool useProvidedKeypoints, Stream& stream);
virtual void convert(InputArray _gpu_keypoints, std::vector<KeyPoint>& keypoints);
virtual int descriptorSize() const { return kBytes; }
virtual int descriptorType() const { return CV_8U; }
virtual int defaultNorm() const { return NORM_HAMMING; }
virtual void setMaxFeatures(int maxFeatures) { nFeatures_ = maxFeatures; }
virtual int getMaxFeatures() const { return nFeatures_; }
virtual void setScaleFactor(double scaleFactor) { scaleFactor_ = scaleFactor; }
virtual double getScaleFactor() const { return scaleFactor_; }
virtual void setNLevels(int nlevels) { nLevels_ = nlevels; }
virtual int getNLevels() const { return nLevels_; }
virtual void setEdgeThreshold(int edgeThreshold) { edgeThreshold_ = edgeThreshold; }
virtual int getEdgeThreshold() const { return edgeThreshold_; }
virtual void setFirstLevel(int firstLevel) { firstLevel_ = firstLevel; }
virtual int getFirstLevel() const { return firstLevel_; }
virtual void setWTA_K(int wta_k) { WTA_K_ = wta_k; }
virtual int getWTA_K() const { return WTA_K_; }
virtual void setScoreType(int scoreType) { scoreType_ = scoreType; }
virtual int getScoreType() const { return scoreType_; }
virtual void setPatchSize(int patchSize) { patchSize_ = patchSize; }
virtual int getPatchSize() const { return patchSize_; }
virtual void setFastThreshold(int fastThreshold) { fastThreshold_ = fastThreshold; }
virtual int getFastThreshold() const { return fastThreshold_; }
virtual void setBlurForDescriptor(bool blurForDescriptor) { blurForDescriptor_ = blurForDescriptor; }
virtual bool getBlurForDescriptor() const { return blurForDescriptor_; }
private:
int nFeatures_;
float scaleFactor_;
int nLevels_;
int edgeThreshold_;
int firstLevel_;
int WTA_K_;
int scoreType_;
int patchSize_;
int fastThreshold_;
bool blurForDescriptor_;
private:
void buildScalePyramids(InputArray _image, InputArray _mask);
void computeKeyPointsPyramid();
void computeDescriptors(OutputArray _descriptors);
void mergeKeyPoints(OutputArray _keypoints);
private:
Ptr<cv::cuda::FastFeatureDetector> fastDetector_;
//! The number of desired features per scale
std::vector<size_t> n_features_per_level_;
//! Points to compute BRIEF descriptors from
GpuMat pattern_;
std::vector<GpuMat> imagePyr_;
std::vector<GpuMat> maskPyr_;
GpuMat buf_;
std::vector<GpuMat> keyPointsPyr_;
std::vector<int> keyPointsCount_;
Ptr<cuda::Filter> blurFilter_;
GpuMat d_keypoints_;
};
static void initializeOrbPattern(const Point* pattern0, Mat& pattern, int ntuples, int tupleSize, int poolSize)
{
RNG rng(0x12345678);
@ -381,7 +463,7 @@ namespace
}
}
void makeRandomPattern(int patchSize, Point* pattern, int npoints)
static void makeRandomPattern(int patchSize, Point* pattern, int npoints)
{
// we always start with a fixed seed,
// to make patterns the same on each run
@ -393,14 +475,32 @@ namespace
pattern[i].y = rng.uniform(-patchSize / 2, patchSize / 2 + 1);
}
}
}
cv::cuda::ORB_CUDA::ORB_CUDA(int nFeatures, float scaleFactor, int nLevels, int edgeThreshold, int firstLevel, int WTA_K, int scoreType, int patchSize) :
nFeatures_(nFeatures), scaleFactor_(scaleFactor), nLevels_(nLevels), edgeThreshold_(edgeThreshold), firstLevel_(firstLevel), WTA_K_(WTA_K),
scoreType_(scoreType), patchSize_(patchSize),
fastDetector_(cuda::FastFeatureDetector::create(DEFAULT_FAST_THRESHOLD))
{
CV_Assert(patchSize_ >= 2);
ORB_Impl::ORB_Impl(int nFeatures,
float scaleFactor,
int nLevels,
int edgeThreshold,
int firstLevel,
int WTA_K,
int scoreType,
int patchSize,
int fastThreshold,
bool blurForDescriptor) :
nFeatures_(nFeatures),
scaleFactor_(scaleFactor),
nLevels_(nLevels),
edgeThreshold_(edgeThreshold),
firstLevel_(firstLevel),
WTA_K_(WTA_K),
scoreType_(scoreType),
patchSize_(patchSize),
fastThreshold_(fastThreshold),
blurForDescriptor_(blurForDescriptor)
{
CV_Assert( patchSize_ >= 2 );
CV_Assert( WTA_K_ == 2 || WTA_K_ == 3 || WTA_K_ == 4 );
fastDetector_ = cuda::FastFeatureDetector::create(fastThreshold_);
// fill the extractors and descriptors for the corresponding scales
float factor = 1.0f / scaleFactor_;
@ -420,7 +520,9 @@ cv::cuda::ORB_CUDA::ORB_CUDA(int nFeatures, float scaleFactor, int nLevels, int
int half_patch_size = patchSize_ / 2;
std::vector<int> u_max(half_patch_size + 2);
for (int v = 0; v <= half_patch_size * std::sqrt(2.f) / 2 + 1; ++v)
{
u_max[v] = cvRound(std::sqrt(static_cast<float>(half_patch_size * half_patch_size - v * v)));
}
// Make sure we are symmetric
for (int v = half_patch_size, v_0 = 0; v >= half_patch_size * std::sqrt(2.f) / 2; --v)
@ -430,7 +532,7 @@ cv::cuda::ORB_CUDA::ORB_CUDA(int nFeatures, float scaleFactor, int nLevels, int
u_max[v] = v_0;
++v_0;
}
CV_Assert(u_max.size() < 32);
CV_Assert( u_max.size() < 32 );
cv::cuda::device::orb::loadUMax(&u_max[0], static_cast<int>(u_max.size()));
// Calc pattern
@ -443,10 +545,7 @@ cv::cuda::ORB_CUDA::ORB_CUDA(int nFeatures, float scaleFactor, int nLevels, int
makeRandomPattern(patchSize_, pattern_buf, npoints);
}
CV_Assert(WTA_K_ == 2 || WTA_K_ == 3 || WTA_K_ == 4);
Mat h_pattern;
if (WTA_K_ == 2)
{
h_pattern.create(2, npoints, CV_32SC1);
@ -468,23 +567,42 @@ cv::cuda::ORB_CUDA::ORB_CUDA(int nFeatures, float scaleFactor, int nLevels, int
pattern_.upload(h_pattern);
blurFilter = cuda::createGaussianFilter(CV_8UC1, -1, Size(7, 7), 2, 2, BORDER_REFLECT_101);
blurFilter_ = cuda::createGaussianFilter(CV_8UC1, -1, Size(7, 7), 2, 2, BORDER_REFLECT_101);
}
blurForDescriptor = false;
}
void ORB_Impl::detectAndCompute(InputArray _image, InputArray _mask, std::vector<KeyPoint>& keypoints, OutputArray _descriptors, bool useProvidedKeypoints)
{
CV_Assert( useProvidedKeypoints == false );
namespace
{
inline float getScale(float scaleFactor, int firstLevel, int level)
detectAndComputeAsync(_image, _mask, d_keypoints_, _descriptors, false, Stream::Null());
convert(d_keypoints_, keypoints);
}
void ORB_Impl::detectAndComputeAsync(InputArray _image, InputArray _mask, OutputArray _keypoints, OutputArray _descriptors, bool useProvidedKeypoints, Stream& stream)
{
CV_Assert( useProvidedKeypoints == false );
buildScalePyramids(_image, _mask);
computeKeyPointsPyramid();
if (_descriptors.needed())
{
computeDescriptors(_descriptors);
}
mergeKeyPoints(_keypoints);
}
static float getScale(float scaleFactor, int firstLevel, int level)
{
return pow(scaleFactor, level - firstLevel);
}
}
void cv::cuda::ORB_CUDA::buildScalePyramids(const GpuMat& image, const GpuMat& mask)
{
CV_Assert(image.type() == CV_8UC1);
CV_Assert(mask.empty() || (mask.type() == CV_8UC1 && mask.size() == image.size()));
void ORB_Impl::buildScalePyramids(InputArray _image, InputArray _mask)
{
const GpuMat image = _image.getGpuMat();
const GpuMat mask = _mask.getGpuMat();
CV_Assert( image.type() == CV_8UC1 );
CV_Assert( mask.empty() || (mask.type() == CV_8UC1 && mask.size() == image.size()) );
imagePyr_.resize(nLevels_);
maskPyr_.resize(nLevels_);
@ -536,12 +654,10 @@ void cv::cuda::ORB_CUDA::buildScalePyramids(const GpuMat& image, const GpuMat& m
cuda::bitwise_and(maskPyr_[level], buf_, maskPyr_[level]);
}
}
}
namespace
{
//takes keypoints and culls them by the response
void cull(GpuMat& keypoints, int& count, int n_points)
// takes keypoints and culls them by the response
static void cull(GpuMat& keypoints, int& count, int n_points)
{
using namespace cv::cuda::device::orb;
@ -557,10 +673,9 @@ namespace
count = cull_gpu(keypoints.ptr<int>(cuda::FastFeatureDetector::LOCATION_ROW), keypoints.ptr<float>(cuda::FastFeatureDetector::RESPONSE_ROW), count, n_points);
}
}
}
void cv::cuda::ORB_CUDA::computeKeyPointsPyramid()
{
void ORB_Impl::computeKeyPointsPyramid()
{
using namespace cv::cuda::device::orb;
int half_patch_size = patchSize_ / 2;
@ -568,6 +683,8 @@ void cv::cuda::ORB_CUDA::computeKeyPointsPyramid()
keyPointsPyr_.resize(nLevels_);
keyPointsCount_.resize(nLevels_);
fastDetector_->setThreshold(fastThreshold_);
for (int level = 0; level < nLevels_; ++level)
{
fastDetector_->setMaxNumPoints(0.05 * imagePyr_[level].size().area());
@ -600,10 +717,10 @@ void cv::cuda::ORB_CUDA::computeKeyPointsPyramid()
// Compute orientation
IC_Angle_gpu(imagePyr_[level], keyPointsPyr_[level].ptr<short2>(0), keyPointsPyr_[level].ptr<float>(2), keyPointsCount_[level], half_patch_size, 0);
}
}
}
void cv::cuda::ORB_CUDA::computeDescriptors(GpuMat& descriptors)
{
void ORB_Impl::computeDescriptors(OutputArray _descriptors)
{
using namespace cv::cuda::device::orb;
int nAllkeypoints = 0;
@ -613,11 +730,12 @@ void cv::cuda::ORB_CUDA::computeDescriptors(GpuMat& descriptors)
if (nAllkeypoints == 0)
{
descriptors.release();
_descriptors.release();
return;
}
ensureSizeIsEnough(nAllkeypoints, descriptorSize(), CV_8UC1, descriptors);
ensureSizeIsEnough(nAllkeypoints, descriptorSize(), CV_8UC1, _descriptors);
GpuMat descriptors = _descriptors.getGpuMat();
int offset = 0;
@ -628,22 +746,22 @@ void cv::cuda::ORB_CUDA::computeDescriptors(GpuMat& descriptors)
GpuMat descRange = descriptors.rowRange(offset, offset + keyPointsCount_[level]);
if (blurForDescriptor)
if (blurForDescriptor_)
{
// preprocess the resized image
ensureSizeIsEnough(imagePyr_[level].size(), imagePyr_[level].type(), buf_);
blurFilter->apply(imagePyr_[level], buf_);
blurFilter_->apply(imagePyr_[level], buf_);
}
computeOrbDescriptor_gpu(blurForDescriptor ? buf_ : imagePyr_[level], keyPointsPyr_[level].ptr<short2>(0), keyPointsPyr_[level].ptr<float>(2),
computeOrbDescriptor_gpu(blurForDescriptor_ ? buf_ : imagePyr_[level], keyPointsPyr_[level].ptr<short2>(0), keyPointsPyr_[level].ptr<float>(2),
keyPointsCount_[level], pattern_.ptr<int>(0), pattern_.ptr<int>(1), descRange, descriptorSize(), WTA_K_, 0);
offset += keyPointsCount_[level];
}
}
}
void cv::cuda::ORB_CUDA::mergeKeyPoints(GpuMat& keypoints)
{
void ORB_Impl::mergeKeyPoints(OutputArray _keypoints)
{
using namespace cv::cuda::device::orb;
int nAllkeypoints = 0;
@ -653,11 +771,12 @@ void cv::cuda::ORB_CUDA::mergeKeyPoints(GpuMat& keypoints)
if (nAllkeypoints == 0)
{
keypoints.release();
_keypoints.release();
return;
}
ensureSizeIsEnough(ROWS_COUNT, nAllkeypoints, CV_32FC1, keypoints);
ensureSizeIsEnough(ROWS_COUNT, nAllkeypoints, CV_32FC1, _keypoints);
GpuMat& keypoints = _keypoints.getGpuMatRef();
int offset = 0;
@ -682,41 +801,41 @@ void cv::cuda::ORB_CUDA::mergeKeyPoints(GpuMat& keypoints)
offset += keyPointsCount_[level];
}
}
}
void cv::cuda::ORB_CUDA::downloadKeyPoints(const GpuMat &d_keypoints, std::vector<KeyPoint>& keypoints)
{
if (d_keypoints.empty())
void ORB_Impl::convert(InputArray _gpu_keypoints, std::vector<KeyPoint>& keypoints)
{
if (_gpu_keypoints.empty())
{
keypoints.clear();
return;
}
Mat h_keypoints(d_keypoints);
convertKeyPoints(h_keypoints, keypoints);
}
void cv::cuda::ORB_CUDA::convertKeyPoints(const Mat &d_keypoints, std::vector<KeyPoint>& keypoints)
{
if (d_keypoints.empty())
Mat h_keypoints;
if (_gpu_keypoints.kind() == _InputArray::CUDA_GPU_MAT)
{
keypoints.clear();
return;
_gpu_keypoints.getGpuMat().download(h_keypoints);
}
else
{
h_keypoints = _gpu_keypoints.getMat();
}
CV_Assert(d_keypoints.type() == CV_32FC1 && d_keypoints.rows == ROWS_COUNT);
CV_Assert( h_keypoints.rows == ROWS_COUNT );
CV_Assert( h_keypoints.type() == CV_32FC1 );
const float* x_ptr = d_keypoints.ptr<float>(X_ROW);
const float* y_ptr = d_keypoints.ptr<float>(Y_ROW);
const float* response_ptr = d_keypoints.ptr<float>(RESPONSE_ROW);
const float* angle_ptr = d_keypoints.ptr<float>(ANGLE_ROW);
const float* octave_ptr = d_keypoints.ptr<float>(OCTAVE_ROW);
const float* size_ptr = d_keypoints.ptr<float>(SIZE_ROW);
const int npoints = h_keypoints.cols;
keypoints.resize(d_keypoints.cols);
keypoints.resize(npoints);
for (int i = 0; i < d_keypoints.cols; ++i)
const float* x_ptr = h_keypoints.ptr<float>(X_ROW);
const float* y_ptr = h_keypoints.ptr<float>(Y_ROW);
const float* response_ptr = h_keypoints.ptr<float>(RESPONSE_ROW);
const float* angle_ptr = h_keypoints.ptr<float>(ANGLE_ROW);
const float* octave_ptr = h_keypoints.ptr<float>(OCTAVE_ROW);
const float* size_ptr = h_keypoints.ptr<float>(SIZE_ROW);
for (int i = 0; i < npoints; ++i)
{
KeyPoint kp;
@ -729,45 +848,21 @@ void cv::cuda::ORB_CUDA::convertKeyPoints(const Mat &d_keypoints, std::vector<Ke
keypoints[i] = kp;
}
}
}
void cv::cuda::ORB_CUDA::operator()(const GpuMat& image, const GpuMat& mask, GpuMat& keypoints)
Ptr<cv::cuda::ORB> cv::cuda::ORB::create(int nfeatures,
float scaleFactor,
int nlevels,
int edgeThreshold,
int firstLevel,
int WTA_K,
int scoreType,
int patchSize,
int fastThreshold,
bool blurForDescriptor)
{
buildScalePyramids(image, mask);
computeKeyPointsPyramid();
mergeKeyPoints(keypoints);
}
void cv::cuda::ORB_CUDA::operator()(const GpuMat& image, const GpuMat& mask, GpuMat& keypoints, GpuMat& descriptors)
{
buildScalePyramids(image, mask);
computeKeyPointsPyramid();
computeDescriptors(descriptors);
mergeKeyPoints(keypoints);
}
void cv::cuda::ORB_CUDA::operator()(const GpuMat& image, const GpuMat& mask, std::vector<KeyPoint>& keypoints)
{
(*this)(image, mask, d_keypoints_);
downloadKeyPoints(d_keypoints_, keypoints);
}
void cv::cuda::ORB_CUDA::operator()(const GpuMat& image, const GpuMat& mask, std::vector<KeyPoint>& keypoints, GpuMat& descriptors)
{
(*this)(image, mask, d_keypoints_, descriptors);
downloadKeyPoints(d_keypoints_, keypoints);
}
void cv::cuda::ORB_CUDA::release()
{
imagePyr_.clear();
maskPyr_.clear();
buf_.release();
keyPointsPyr_.clear();
d_keypoints_.release();
return makePtr<ORB_Impl>(nfeatures, scaleFactor, nlevels, edgeThreshold, firstLevel, WTA_K, scoreType, patchSize, fastThreshold, blurForDescriptor);
}
#endif /* !defined (HAVE_CUDA) */

11
modules/cudafeatures2d/test/test_features2d.cpp
View File

@ -122,7 +122,7 @@ namespace
IMPLEMENT_PARAM_CLASS(ORB_BlurForDescriptor, bool)
}
CV_ENUM(ORB_ScoreType, ORB::HARRIS_SCORE, ORB::FAST_SCORE)
CV_ENUM(ORB_ScoreType, cv::ORB::HARRIS_SCORE, cv::ORB::FAST_SCORE)
PARAM_TEST_CASE(ORB, cv::cuda::DeviceInfo, ORB_FeaturesCount, ORB_ScaleFactor, ORB_LevelsCount, ORB_EdgeThreshold, ORB_firstLevel, ORB_WTA_K, ORB_ScoreType, ORB_PatchSize, ORB_BlurForDescriptor)
{
@ -162,8 +162,9 @@ CUDA_TEST_P(ORB, Accuracy)
cv::Mat mask(image.size(), CV_8UC1, cv::Scalar::all(1));
mask(cv::Range(0, image.rows / 2), cv::Range(0, image.cols / 2)).setTo(cv::Scalar::all(0));
cv::cuda::ORB_CUDA orb(nFeatures, scaleFactor, nLevels, edgeThreshold, firstLevel, WTA_K, scoreType, patchSize);
orb.blurForDescriptor = blurForDescriptor;
cv::Ptr<cv::cuda::ORB> orb =
cv::cuda::ORB::create(nFeatures, scaleFactor, nLevels, edgeThreshold, firstLevel,
WTA_K, scoreType, patchSize, 20, blurForDescriptor);
if (!supportFeature(devInfo, cv::cuda::GLOBAL_ATOMICS))
{
@ -171,7 +172,7 @@ CUDA_TEST_P(ORB, Accuracy)
{
std::vector<cv::KeyPoint> keypoints;
cv::cuda::GpuMat descriptors;
orb(loadMat(image), loadMat(mask), keypoints, descriptors);
orb->detectAndComputeAsync(loadMat(image), loadMat(mask), keypoints, descriptors);
}
catch (const cv::Exception& e)
{
@ -182,7 +183,7 @@ CUDA_TEST_P(ORB, Accuracy)
{
std::vector<cv::KeyPoint> keypoints;
cv::cuda::GpuMat descriptors;
orb(loadMat(image), loadMat(mask), keypoints, descriptors);
orb->detectAndCompute(loadMat(image), loadMat(mask), keypoints, descriptors);
cv::Ptr<cv::ORB> orb_gold = cv::ORB::create(nFeatures, scaleFactor, nLevels, edgeThreshold, firstLevel, WTA_K, scoreType, patchSize);

6
samples/gpu/performance/tests.cpp
View File

@ -350,15 +350,15 @@ TEST(ORB)
orb->detectAndCompute(src, Mat(), keypoints, descriptors);
CPU_OFF;
cuda::ORB_CUDA d_orb;
Ptr<cuda::ORB> d_orb = cuda::ORB::create();
cuda::GpuMat d_src(src);
cuda::GpuMat d_keypoints;
cuda::GpuMat d_descriptors;
d_orb(d_src, cuda::GpuMat(), d_keypoints, d_descriptors);
d_orb->detectAndComputeAsync(d_src, cuda::GpuMat(), d_keypoints, d_descriptors);
CUDA_ON;
d_orb(d_src, cuda::GpuMat(), d_keypoints, d_descriptors);
d_orb->detectAndComputeAsync(d_src, cuda::GpuMat(), d_keypoints, d_descriptors);
CUDA_OFF;
}