Merge pull request #1954 from pentschev:ocl_features2d_fast_master

modules/ocl/doc/feature_detection_and_description.rst

@@ -305,6 +305,153 @@ Converts matrices obtained via :ocv:func:`ocl::BruteForceMatcher_OCL_base::radiu
 
 If ``compactResult`` is ``true`` , the ``matches`` vector does not contain matches for fully masked-out query descriptors.
 
+
+ocl::FAST_OCL
+------------------
+.. ocv:class:: ocl::FAST_OCL
+
+Class used for corner detection using the FAST algorithm. ::
+
+        class CV_EXPORTS FAST_OCL
+        {
+        public:
+            enum
+            {
+                X_ROW = 0,
+                Y_ROW,
+                RESPONSE_ROW,
+                ROWS_COUNT
+            };
+
+            // all features have same size
+            static const int FEATURE_SIZE = 7;
+
+            explicit FAST_OCL(int threshold, bool nonmaxSupression = true, double keypointsRatio = 0.05);
+
+            //! finds the keypoints using FAST detector
+            //! supports only CV_8UC1 images
+            void operator ()(const oclMat& image, const oclMat& mask, oclMat& keypoints);
+            void operator ()(const oclMat& image, const oclMat& mask, std::vector<KeyPoint>& keypoints);
+
+            //! download keypoints from device to host memory
+            static void downloadKeypoints(const oclMat& d_keypoints, std::vector<KeyPoint>& keypoints);
+
+            //! convert keypoints to KeyPoint vector
+            static void convertKeypoints(const Mat& h_keypoints, std::vector<KeyPoint>& keypoints);
+
+            //! release temporary buffer's memory
+            void release();
+
+            bool nonmaxSupression;
+
+            int threshold;
+
+            //! max keypoints = keypointsRatio * img.size().area()
+            double keypointsRatio;
+
+            //! find keypoints and compute it's response if nonmaxSupression is true
+            //! return count of detected keypoints
+            int calcKeyPointsLocation(const oclMat& image, const oclMat& mask);
+
+            //! get final array of keypoints
+            //! performs nonmax supression if needed
+            //! return final count of keypoints
+            int getKeyPoints(oclMat& keypoints);
+
+        private:
+            // Hidden
+        };
+
+
+The class ``FAST_OCL`` implements FAST corner detection algorithm.
+
+.. seealso:: :ocv:func:`FAST`
+
+
+
+ocl::FAST_OCL::FAST_OCL
+--------------------------
+Constructor.
+
+.. ocv:function:: ocl::FAST_OCL::FAST_OCL(int threshold, bool nonmaxSupression = true, double keypointsRatio = 0.05)
+
+    :param threshold: Threshold on difference between intensity of the central pixel and pixels on a circle around this pixel.
+
+    :param nonmaxSupression: If it is true, non-maximum suppression is applied to detected corners (keypoints).
+
+    :param keypointsRatio: Inner buffer size for keypoints store is determined as (keypointsRatio * image_width * image_height).
+
+
+
+ocl::FAST_OCL::operator ()
+----------------------------
+Finds the keypoints using FAST detector.
+
+.. ocv:function:: void ocl::FAST_OCL::operator ()(const oclMat& image, const oclMat& mask, oclMat& keypoints)
+.. ocv:function:: void ocl::FAST_OCL::operator ()(const oclMat& image, const oclMat& mask, std::vector<KeyPoint>& keypoints)
+
+    :param image: Image where keypoints (corners) are detected. Only 8-bit grayscale images are supported.
+
+    :param mask: Optional input mask that marks the regions where we should detect features.
+
+    :param keypoints: The output vector of keypoints. Can be stored both in host or device memory. For device memory:
+
+            * X_ROW of keypoints will contain the horizontal coordinate of the i'th point
+            * Y_ROW of keypoints will contain the vertical coordinate of the i'th point
+            * RESPONSE_ROW will contain response of i'th point (if non-maximum suppression is applied)
+
+
+
+ocl::FAST_OCL::downloadKeypoints
+----------------------------------
+Download keypoints from device to host memory.
+
+.. ocv:function:: void ocl::FAST_OCL::downloadKeypoints(const oclMat& d_keypoints, std::vector<KeyPoint>& keypoints)
+
+
+
+ocl::FAST_OCL::convertKeypoints
+---------------------------------
+Converts keypoints from OpenCL representation to vector of ``KeyPoint``.
+
+.. ocv:function:: void ocl::FAST_OCL::convertKeypoints(const Mat& h_keypoints, std::vector<KeyPoint>& keypoints)
+
+
+
+ocl::FAST_OCL::release
+------------------------
+Releases inner buffer memory.
+
+.. ocv:function:: void ocl::FAST_OCL::release()
+
+
+
+ocl::FAST_OCL::calcKeyPointsLocation
+--------------------------------------
+Find keypoints. If ``nonmaxSupression`` is true, responses are computed and eliminates keypoints with the smaller responses from 9-neighborhood regions.
+
+.. ocv:function:: int ocl::FAST_OCL::calcKeyPointsLocation(const oclMat& image, const oclMat& mask)
+
+    :param image: Image where keypoints (corners) are detected. Only 8-bit grayscale images are supported.
+
+    :param mask: Optional input mask that marks the regions where we should detect features.
+
+The function returns the amount of detected keypoints.
+
+
+
+ocl::FAST_OCL::getKeyPoints
+-----------------------------
+Gets final array of keypoints.
+
+.. ocv:function:: int ocl::FAST_OCL::getKeyPoints(oclMat& keypoints)
+
+    :param keypoints: The output vector of keypoints.
+
+The function performs non-max suppression if needed and returns the final amount of keypoints.
+
+
+
 ocl::HOGDescriptor
 ----------------------
 

modules/ocl/include/opencv2/ocl.hpp

@@ -1482,6 +1482,65 @@ namespace cv
             harrisK = harrisK_;
         }
 
+        ////////////////////////////////// FAST Feature Detector //////////////////////////////////
+        class CV_EXPORTS FAST_OCL
+        {
+        public:
+            enum
+            {
+                X_ROW = 0,
+                Y_ROW,
+                RESPONSE_ROW,
+                ROWS_COUNT
+            };
+
+            // all features have same size
+            static const int FEATURE_SIZE = 7;
+
+            explicit FAST_OCL(int threshold, bool nonmaxSupression = true, double keypointsRatio = 0.05);
+
+            //! finds the keypoints using FAST detector
+            //! supports only CV_8UC1 images
+            void operator ()(const oclMat& image, const oclMat& mask, oclMat& keypoints);
+            void operator ()(const oclMat& image, const oclMat& mask, std::vector<KeyPoint>& keypoints);
+
+            //! download keypoints from device to host memory
+            static void downloadKeypoints(const oclMat& d_keypoints, std::vector<KeyPoint>& keypoints);
+
+            //! convert keypoints to KeyPoint vector
+            static void convertKeypoints(const Mat& h_keypoints, std::vector<KeyPoint>& keypoints);
+
+            //! release temporary buffer's memory
+            void release();
+
+            bool nonmaxSupression;
+
+            int threshold;
+
+            //! max keypoints = keypointsRatio * img.size().area()
+            double keypointsRatio;
+
+            //! find keypoints and compute it's response if nonmaxSupression is true
+            //! return count of detected keypoints
+            int calcKeyPointsLocation(const oclMat& image, const oclMat& mask);
+
+            //! get final array of keypoints
+            //! performs nonmax supression if needed
+            //! return final count of keypoints
+            int getKeyPoints(oclMat& keypoints);
+
+        private:
+            oclMat kpLoc_;
+            int count_;
+
+            oclMat score_;
+
+            oclMat d_keypoints_;
+
+            int calcKeypointsOCL(const oclMat& img, const oclMat& mask, int maxKeypoints);
+            int nonmaxSupressionOCL(oclMat& keypoints);
+        };
+
         /////////////////////////////// PyrLKOpticalFlow /////////////////////////////////////
 
         class CV_EXPORTS PyrLKOpticalFlow

modules/ocl/perf/perf_fast.cpp

@@ -0,0 +1,93 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+// Authors:
+//  * Peter Andreas Entschev, peter@entschev.com
+//
+//M*/
+
+#include "perf_precomp.hpp"
+
+using namespace perf;
+
+///////////// FAST ////////////////////////
+
+typedef std::tr1::tuple<std::string, int, bool> Image_Threshold_NonmaxSupression_t;
+typedef perf::TestBaseWithParam<Image_Threshold_NonmaxSupression_t> Image_Threshold_NonmaxSupression;
+
+PERF_TEST_P(Image_Threshold_NonmaxSupression, FAST,
+            testing::Combine(testing::Values<string>("gpu/perf/aloe.png"),
+                    testing::Values(20),
+                    testing::Bool()))
+{
+    const Image_Threshold_NonmaxSupression_t params = GetParam();
+    const std::string imgFile = std::tr1::get<0>(params);
+    const int threshold = std::tr1::get<1>(params);
+    const bool nonmaxSupression = std::tr1::get<2>(params);
+
+    const cv::Mat img = imread(getDataPath(imgFile), cv::IMREAD_GRAYSCALE);
+    ASSERT_FALSE(img.empty());
+
+    if (RUN_OCL_IMPL)
+    {
+        cv::ocl::FAST_OCL fast(threshold, nonmaxSupression, 0.5);
+
+        cv::ocl::oclMat d_img(img);
+        cv::ocl::oclMat d_keypoints;
+
+        OCL_TEST_CYCLE() fast(d_img, cv::ocl::oclMat(), d_keypoints);
+
+        std::vector<cv::KeyPoint> ocl_keypoints;
+        fast.downloadKeypoints(d_keypoints, ocl_keypoints);
+
+        sortKeyPoints(ocl_keypoints);
+
+        SANITY_CHECK_KEYPOINTS(ocl_keypoints);
+    }
+    else if (RUN_PLAIN_IMPL)
+    {
+        std::vector<cv::KeyPoint> cpu_keypoints;
+
+        TEST_CYCLE() cv::FAST(img, cpu_keypoints, threshold, nonmaxSupression);
+
+        SANITY_CHECK_KEYPOINTS(cpu_keypoints);
+    }
+    else
+        OCL_PERF_ELSE;
+}

modules/ocl/perf/perf_precomp.hpp

@@ -116,6 +116,7 @@ using namespace cv;
 #define OCL_TEST_CYCLE() for(; startTimer(), next(); cv::ocl::finish(), stopTimer())
 #define OCL_TEST_CYCLE_MULTIRUN(runsNum) for(declare.runs(runsNum); startTimer(), next(); stopTimer()) for(int r = 0; r < runsNum; cv::ocl::finish(), ++r)
 
+// TODO: Move to the ts module
 namespace cvtest {
 namespace ocl {
 inline void checkDeviceMaxMemoryAllocSize(const Size& size, int type, int factor = 1)
@@ -133,6 +134,60 @@ inline void checkDeviceMaxMemoryAllocSize(const Size& size, int type, int factor
         throw perf::TestBase::PerfSkipTestException();
     }
 }
+
+struct KeypointIdxCompare
+{
+    std::vector<cv::KeyPoint>* keypoints;
+
+    explicit KeypointIdxCompare(std::vector<cv::KeyPoint>* _keypoints) : keypoints(_keypoints) {}
+
+    bool operator ()(size_t i1, size_t i2) const
+    {
+        cv::KeyPoint kp1 = (*keypoints)[i1];
+        cv::KeyPoint kp2 = (*keypoints)[i2];
+        if (kp1.pt.x != kp2.pt.x)
+            return kp1.pt.x < kp2.pt.x;
+        if (kp1.pt.y != kp2.pt.y)
+            return kp1.pt.y < kp2.pt.y;
+        if (kp1.response != kp2.response)
+            return kp1.response < kp2.response;
+        return kp1.octave < kp2.octave;
+    }
+};
+
+inline void sortKeyPoints(std::vector<cv::KeyPoint>& keypoints, cv::InputOutputArray _descriptors = cv::noArray())
+{
+    std::vector<size_t> indexies(keypoints.size());
+    for (size_t i = 0; i < indexies.size(); ++i)
+        indexies[i] = i;
+
+    std::sort(indexies.begin(), indexies.end(), KeypointIdxCompare(&keypoints));
+
+    std::vector<cv::KeyPoint> new_keypoints;
+    cv::Mat new_descriptors;
+
+    new_keypoints.resize(keypoints.size());
+
+    cv::Mat descriptors;
+    if (_descriptors.needed())
+    {
+        descriptors = _descriptors.getMat();
+        new_descriptors.create(descriptors.size(), descriptors.type());
+    }
+
+    for (size_t i = 0; i < indexies.size(); ++i)
+    {
+        size_t new_idx = indexies[i];
+        new_keypoints[i] = keypoints[new_idx];
+        if (!new_descriptors.empty())
+            descriptors.row((int) new_idx).copyTo(new_descriptors.row((int) i));
+    }
+
+    keypoints.swap(new_keypoints);
+    if (_descriptors.needed())
+        new_descriptors.copyTo(_descriptors);
+}
+
 } // namespace cvtest::ocl
 } // namespace cvtest
 

modules/ocl/src/fast.cpp

@@ -0,0 +1,229 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+// Authors:
+//  * Peter Andreas Entschev, peter@entschev.com
+//
+//M*/
+
+#include "precomp.hpp"
+#include "opencl_kernels.hpp"
+
+using namespace cv;
+using namespace cv::ocl;
+
+cv::ocl::FAST_OCL::FAST_OCL(int _threshold, bool _nonmaxSupression, double _keypointsRatio) :
+    nonmaxSupression(_nonmaxSupression), threshold(_threshold), keypointsRatio(_keypointsRatio), count_(0)
+{
+}
+
+void cv::ocl::FAST_OCL::operator ()(const oclMat& image, const oclMat& mask, std::vector<KeyPoint>& keypoints)
+{
+    if (image.empty())
+        return;
+
+    (*this)(image, mask, d_keypoints_);
+    downloadKeypoints(d_keypoints_, keypoints);
+}
+
+void cv::ocl::FAST_OCL::downloadKeypoints(const oclMat& d_keypoints, std::vector<KeyPoint>& keypoints)
+{
+    if (d_keypoints.empty())
+        return;
+
+    Mat h_keypoints(d_keypoints);
+    convertKeypoints(h_keypoints, keypoints);
+}
+
+void cv::ocl::FAST_OCL::convertKeypoints(const Mat& h_keypoints, std::vector<KeyPoint>& keypoints)
+{
+    if (h_keypoints.empty())
+        return;
+
+    CV_Assert(h_keypoints.rows == ROWS_COUNT && h_keypoints.elemSize() == 4);
+
+    int npoints = h_keypoints.cols;
+
+    keypoints.resize(npoints);
+
+    const float* loc_x = h_keypoints.ptr<float>(X_ROW);
+    const float* loc_y = h_keypoints.ptr<float>(Y_ROW);
+    const float* response_row = h_keypoints.ptr<float>(RESPONSE_ROW);
+
+    for (int i = 0; i < npoints; ++i)
+    {
+        KeyPoint kp(loc_x[i], loc_y[i], static_cast<float>(FEATURE_SIZE), -1, response_row[i]);
+        keypoints[i] = kp;
+    }
+}
+
+void cv::ocl::FAST_OCL::operator ()(const oclMat& img, const oclMat& mask, oclMat& keypoints)
+{
+    calcKeyPointsLocation(img, mask);
+    keypoints.cols = getKeyPoints(keypoints);
+}
+
+int cv::ocl::FAST_OCL::calcKeyPointsLocation(const oclMat& img, const oclMat& mask)
+{
+    CV_Assert(img.type() == CV_8UC1);
+    CV_Assert(mask.empty() || (mask.type() == CV_8UC1 && mask.size() == img.size()));
+
+    int maxKeypoints = static_cast<int>(keypointsRatio * img.size().area());
+
+    ensureSizeIsEnough(ROWS_COUNT, maxKeypoints, CV_32SC1, kpLoc_);
+    kpLoc_.setTo(Scalar::all(0));
+
+    if (nonmaxSupression)
+    {
+        ensureSizeIsEnough(img.size(), CV_32SC1, score_);
+        score_.setTo(Scalar::all(0));
+    }
+
+    count_ = calcKeypointsOCL(img, mask, maxKeypoints);
+    count_ = std::min(count_, maxKeypoints);
+
+    return count_;
+}
+
+int cv::ocl::FAST_OCL::calcKeypointsOCL(const oclMat& img, const oclMat& mask, int maxKeypoints)
+{
+    size_t localThreads[3] = {16, 16, 1};
+    size_t globalThreads[3] = {divUp(img.cols - 6, localThreads[0]) * localThreads[0],
+                               divUp(img.rows - 6, localThreads[1]) * localThreads[1],
+                               1};
+
+    Context *clCxt = Context::getContext();
+    String kernelName = (mask.empty()) ? "calcKeypoints" : "calcKeypointsWithMask";
+    std::vector< std::pair<size_t, const void *> > args;
+
+    int counter = 0;
+    int err = CL_SUCCESS;
+    cl_mem counterCL = clCreateBuffer(*(cl_context*)clCxt->getOpenCLContextPtr(),
+                                    CL_MEM_COPY_HOST_PTR, sizeof(int),
+                                    &counter, &err);
+
+    int kpLocStep = kpLoc_.step / kpLoc_.elemSize();
+    int scoreStep = score_.step / score_.elemSize();
+    int nms = (nonmaxSupression) ? 1 : 0;
+
+    args.push_back( std::make_pair( sizeof(cl_mem), (void *)&img.data));
+    if (!mask.empty()) args.push_back( std::make_pair( sizeof(cl_mem), (void *)&mask.data));
+    args.push_back( std::make_pair( sizeof(cl_mem), (void *)&kpLoc_.data));
+    args.push_back( std::make_pair( sizeof(cl_mem), (void *)&score_.data));
+    args.push_back( std::make_pair( sizeof(cl_mem), (void *)&counterCL));
+    args.push_back( std::make_pair( sizeof(cl_int), (void *)&nms));
+    args.push_back( std::make_pair( sizeof(cl_int), (void *)&maxKeypoints));
+    args.push_back( std::make_pair( sizeof(cl_int), (void *)&threshold));
+    args.push_back( std::make_pair( sizeof(cl_int), (void *)&img.step));
+    args.push_back( std::make_pair( sizeof(cl_int), (void *)&img.rows));
+    args.push_back( std::make_pair( sizeof(cl_int), (void *)&img.cols));
+    if (!mask.empty()) args.push_back( std::make_pair( sizeof(cl_int), (void *)&mask.step));
+    args.push_back( std::make_pair( sizeof(cl_int), (void *)&kpLocStep));
+    args.push_back( std::make_pair( sizeof(cl_int), (void *)&scoreStep));
+
+    openCLExecuteKernel(clCxt, &featdetect_fast, kernelName, globalThreads, localThreads, args, -1, -1);
+
+    openCLSafeCall(clEnqueueReadBuffer(*(cl_command_queue*)clCxt->getOpenCLCommandQueuePtr(),
+                                       counterCL, CL_TRUE, 0, sizeof(int), &counter, 0, NULL, NULL));
+    openCLSafeCall(clReleaseMemObject(counterCL));
+
+    return counter;
+}
+
+int cv::ocl::FAST_OCL::nonmaxSupressionOCL(oclMat& keypoints)
+{
+    size_t localThreads[3] = {256, 1, 1};
+    size_t globalThreads[3] = {count_, 1, 1};
+
+    Context *clCxt = Context::getContext();
+    String kernelName = "nonmaxSupression";
+    std::vector< std::pair<size_t, const void *> > args;
+
+    int counter = 0;
+    int err = CL_SUCCESS;
+    cl_mem counterCL = clCreateBuffer(*(cl_context*)clCxt->getOpenCLContextPtr(),
+                                    CL_MEM_COPY_HOST_PTR, sizeof(int),
+                                    &counter, &err);
+
+    int kpLocStep = kpLoc_.step / kpLoc_.elemSize();
+    int sStep = score_.step / score_.elemSize();
+    int kStep = keypoints.step / keypoints.elemSize();
+
+    args.push_back( std::make_pair( sizeof(cl_mem), (void *)&kpLoc_.data));
+    args.push_back( std::make_pair( sizeof(cl_mem), (void *)&score_.data));
+    args.push_back( std::make_pair( sizeof(cl_mem), (void *)&keypoints.data));
+    args.push_back( std::make_pair( sizeof(cl_mem), (void *)&counterCL));
+    args.push_back( std::make_pair( sizeof(cl_int), (void *)&count_));
+    args.push_back( std::make_pair( sizeof(cl_int), (void *)&kpLocStep));
+    args.push_back( std::make_pair( sizeof(cl_int), (void *)&sStep));
+    args.push_back( std::make_pair( sizeof(cl_int), (void *)&kStep));
+
+    openCLExecuteKernel(clCxt, &featdetect_fast, kernelName, globalThreads, localThreads, args, -1, -1);
+
+    openCLSafeCall(clEnqueueReadBuffer(*(cl_command_queue*)clCxt->getOpenCLCommandQueuePtr(),
+                                       counterCL, CL_TRUE, 0, sizeof(int), &counter, 0, NULL, NULL));
+    openCLSafeCall(clReleaseMemObject(counterCL));
+
+    return counter;
+}
+
+int cv::ocl::FAST_OCL::getKeyPoints(oclMat& keypoints)
+{
+    if (count_ == 0)
+        return 0;
+
+    if (nonmaxSupression)
+    {
+        ensureSizeIsEnough(ROWS_COUNT, count_, CV_32FC1, keypoints);
+        return nonmaxSupressionOCL(keypoints);
+    }
+
+    kpLoc_.convertTo(keypoints, CV_32FC1);
+    Mat k = keypoints;
+
+    return count_;
+}
+
+void cv::ocl::FAST_OCL::release()
+{
+    kpLoc_.release();
+    score_.release();
+
+    d_keypoints_.release();
+}

modules/ocl/test/test_fast.cpp

@@ -0,0 +1,93 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+// Authors:
+//  * Peter Andreas Entschev, peter@entschev.com
+//
+//M*/
+
+#include "test_precomp.hpp"
+
+#ifdef HAVE_OPENCL
+
+////////////////////////////////////////////////////////
+// FAST
+
+namespace
+{
+    IMPLEMENT_PARAM_CLASS(FAST_Threshold, int)
+    IMPLEMENT_PARAM_CLASS(FAST_NonmaxSupression, bool)
+}
+
+PARAM_TEST_CASE(FAST, FAST_Threshold, FAST_NonmaxSupression)
+{
+    int threshold;
+    bool nonmaxSupression;
+
+    virtual void SetUp()
+    {
+        threshold = GET_PARAM(0);
+        nonmaxSupression = GET_PARAM(1);
+    }
+};
+
+OCL_TEST_P(FAST, Accuracy)
+{
+    cv::Mat image = readImage("gpu/perf/aloe.png", cv::IMREAD_GRAYSCALE);
+    ASSERT_FALSE(image.empty());
+
+    cv::ocl::FAST_OCL fast(threshold);
+    fast.nonmaxSupression = nonmaxSupression;
+
+    cv::ocl::oclMat ocl_image = cv::ocl::oclMat(image);
+
+    std::vector<cv::KeyPoint> keypoints;
+    fast(ocl_image, cv::ocl::oclMat(), keypoints);
+
+    std::vector<cv::KeyPoint> keypoints_gold;
+    cv::FAST(image, keypoints_gold, threshold, nonmaxSupression);
+
+    ASSERT_KEYPOINTS_EQ(keypoints_gold, keypoints);
+}
+
+INSTANTIATE_TEST_CASE_P(OCL_Features2D, FAST, testing::Combine(
+                        testing::Values(FAST_Threshold(25), FAST_Threshold(50)),
+                        testing::Values(FAST_NonmaxSupression(false), FAST_NonmaxSupression(true))));
+
+#endif

modules/ocl/test/utility.cpp

@@ -259,4 +259,70 @@ void showDiff(const Mat& src, const Mat& gold, const Mat& actual, double eps, bo
     }
 }
 
+namespace
+{
+    bool keyPointsEquals(const cv::KeyPoint& p1, const cv::KeyPoint& p2)
+    {
+        const double maxPtDif = 1.0;
+        const double maxSizeDif = 1.0;
+        const double maxAngleDif = 2.0;
+        const double maxResponseDif = 0.1;
+
+        double dist = cv::norm(p1.pt - p2.pt);
+
+        if (dist < maxPtDif &&
+            fabs(p1.size - p2.size) < maxSizeDif &&
+            abs(p1.angle - p2.angle) < maxAngleDif &&
+            abs(p1.response - p2.response) < maxResponseDif &&
+            p1.octave == p2.octave &&
+            p1.class_id == p2.class_id)
+        {
+            return true;
+        }
+
+        return false;
+    }
+
+    struct KeyPointLess : std::binary_function<cv::KeyPoint, cv::KeyPoint, bool>
+    {
+        bool operator()(const cv::KeyPoint& kp1, const cv::KeyPoint& kp2) const
+        {
+            return kp1.pt.y < kp2.pt.y || (kp1.pt.y == kp2.pt.y && kp1.pt.x < kp2.pt.x);
+        }
+    };
+}
+
+testing::AssertionResult assertKeyPointsEquals(const char* gold_expr, const char* actual_expr, std::vector<cv::KeyPoint>& gold, std::vector<cv::KeyPoint>& actual)
+{
+    if (gold.size() != actual.size())
+    {
+        return testing::AssertionFailure() << "KeyPoints size mistmach\n"
+                                           << "\"" << gold_expr << "\" : " << gold.size() << "\n"
+                                           << "\"" << actual_expr << "\" : " << actual.size();
+    }
+
+    std::sort(actual.begin(), actual.end(), KeyPointLess());
+    std::sort(gold.begin(), gold.end(), KeyPointLess());
+
+    for (size_t i = 0; i < gold.size(); ++i)
+    {
+        const cv::KeyPoint& p1 = gold[i];
+        const cv::KeyPoint& p2 = actual[i];
+
+        if (!keyPointsEquals(p1, p2))
+        {
+            return testing::AssertionFailure() << "KeyPoints differ at " << i << "\n"
+                                               << "\"" << gold_expr << "\" vs \"" << actual_expr << "\" : \n"
+                                               << "pt : " << testing::PrintToString(p1.pt) << " vs " << testing::PrintToString(p2.pt) << "\n"
+                                               << "size : " << p1.size << " vs " << p2.size << "\n"
+                                               << "angle : " << p1.angle << " vs " << p2.angle << "\n"
+                                               << "response : " << p1.response << " vs " << p2.response << "\n"
+                                               << "octave : " << p1.octave << " vs " << p2.octave << "\n"
+                                               << "class_id : " << p1.class_id << " vs " << p2.class_id;
+        }
+    }
+
+    return ::testing::AssertionSuccess();
+}
+
 } // namespace cvtest

modules/ocl/test/utility.hpp

@@ -54,6 +54,9 @@ extern int LOOP_TIMES;
 
 namespace cvtest {
 
+testing::AssertionResult assertKeyPointsEquals(const char* gold_expr, const char* actual_expr, std::vector<cv::KeyPoint>& gold, std::vector<cv::KeyPoint>& actual);
+#define ASSERT_KEYPOINTS_EQ(gold, actual) EXPECT_PRED_FORMAT2(assertKeyPointsEquals, gold, actual)
+
 void showDiff(const Mat& src, const Mat& gold, const Mat& actual, double eps, bool alwaysShow = false);
 
 cv::ocl::oclMat createMat_ocl(cv::RNG& rng, Size size, int type, bool useRoi);