Move OpenVX implementation for FAST to HAL.

This commit is contained in:
Alexander Smorkalov 2025-02-11 17:25:48 +03:00
parent 58e557d059
commit 1de6e20463
5 changed files with 110 additions and 75 deletions

View File

@ -4,6 +4,7 @@ target_include_directories(openvx_hal PUBLIC
${OPENCV_3P_OPENVX_DIR}/include
${CMAKE_SOURCE_DIR}/modules/core/include
${CMAKE_SOURCE_DIR}/modules/imgproc/include
${CMAKE_SOURCE_DIR}/modules/features2d/include
${OPENVX_INCLUDE_DIR})
target_link_libraries(openvx_hal PUBLIC ${OPENVX_LIBRARIES})
set_target_properties(openvx_hal PROPERTIES ARCHIVE_OUTPUT_DIRECTORY ${3P_LIBRARY_OUTPUT_PATH})

View File

@ -1,5 +1,7 @@
#include "openvx_hal.hpp"
#include "opencv2/core/hal/interface.h"
#include "opencv2/imgproc/hal/interface.h"
#include "opencv2/features2d/hal/interface.h"
#define IVX_HIDE_INFO_WARNINGS
#include "ivx.hpp"
@ -1334,3 +1336,72 @@ int ovx_hal_minMaxIdxMaskStep(const uchar* src_data, size_t src_step, int width,
return CV_HAL_ERROR_OK;
}
template <> inline bool skipSmallImages<VX_KERNEL_FAST_CORNERS>(int w, int h) { return w*h < 800 * 600; }
int ovx_hal_FAST(const uchar* src_data, size_t src_step, int width, int height, uchar* keypoints_data, size_t* keypoints_count,
int threshold, bool nonmax_suppression, int /*cv::FastFeatureDetector::DetectorType*/ dtype)
{
// Nonmax suppression is done differently in OpenCV than in OpenVX
// 9/16 is the only supported mode in OpenVX
if(nonmax_suppression || dtype != CV_HAL_TYPE_9_16)
{
return CV_HAL_ERROR_NOT_IMPLEMENTED;
}
if (skipSmallImages<VX_KERNEL_FAST_CORNERS>(width, height))
{
return CV_HAL_ERROR_NOT_IMPLEMENTED;
}
try
{
ivx::Context context = getOpenVXHALContext();
ivx::Image img = ivx::Image::createFromHandle(context, VX_DF_IMAGE_U8,
ivx::Image::createAddressing(width, height, 1, (vx_int32)src_step),
const_cast<uchar*>(src_data));
ivx::Scalar vxthreshold = ivx::Scalar::create<VX_TYPE_FLOAT32>(context, threshold);
vx_size capacity = width * height;
ivx::Array corners = ivx::Array::create(context, VX_TYPE_KEYPOINT, capacity);
ivx::Scalar numCorners = ivx::Scalar::create<VX_TYPE_SIZE>(context, 0);
ivx::IVX_CHECK_STATUS(vxuFastCorners(context, img, vxthreshold, (vx_bool)nonmax_suppression, corners, numCorners));
size_t nPoints = numCorners.getValue<vx_size>();
std::vector<vx_keypoint_t> vxCorners(nPoints);
corners.copyTo(vxCorners);
cvhalKeyPoint* keypoints = (cvhalKeyPoint*)keypoints_data;
for(size_t i = 0; i < std::min(nPoints, *keypoints_count); i++)
{
//if nonmaxSuppression is false, vxCorners[i].strength is undefined
keypoints[i].x = vxCorners[i].x;
keypoints[i].y = vxCorners[i].y;
keypoints[i].size = 7;
keypoints[i].angle = -1;
keypoints[i].response = vxCorners[i].strength;
}
*keypoints_count = std::min(nPoints, *keypoints_count);
#ifdef VX_VERSION_1_1
//we should take user memory back before release
//(it's not done automatically according to standard)
img.swapHandle();
#endif
}
catch (const ivx::RuntimeError & e)
{
PRINT_HALERR_MSG(runtime);
return CV_HAL_ERROR_UNKNOWN;
}
catch (const ivx::WrapperError & e)
{
PRINT_HALERR_MSG(wrapper);
return CV_HAL_ERROR_UNKNOWN;
}
return CV_HAL_ERROR_OK;
}

View File

@ -59,7 +59,8 @@ int ovx_hal_meanStdDev(const uchar* src_data, size_t src_step, int width, int he
int ovx_hal_lut(const uchar *src_data, size_t src_step, size_t src_type, const uchar* lut_data, size_t lut_channel_size, size_t lut_channels, uchar *dst_data, size_t dst_step, int width, int height);
int ovx_hal_minMaxIdxMaskStep(const uchar* src_data, size_t src_step, int width, int height, int depth,
double* minVal, double* maxVal, int* minIdx, int* maxIdx, uchar* mask, size_t mask_step);
int ovx_hal_FAST(const uchar* src_data, size_t src_step, int width, int height, uchar* keypoints_data, size_t* keypoints_count,
int threshold, bool nonmax_suppression, int /*cv::FastFeatureDetector::DetectorType*/ dtype);
//==================================================================================================
// functions redefinition
// ...
@ -152,5 +153,6 @@ int ovx_hal_minMaxIdxMaskStep(const uchar* src_data, size_t src_step, int width,
#define cv_hal_lut ovx_hal_lut
#undef cv_hal_minMaxIdxMaskStep
#define cv_hal_minMaxIdxMaskStep ovx_hal_minMaxIdxMaskStep
#define cv_hal_FAST ovx_hal_FAST
#endif

View File

@ -49,8 +49,6 @@ The references are:
#include "opencv2/core/hal/intrin.hpp"
#include "opencv2/core/utils/buffer_area.private.hpp"
#include "opencv2/core/openvx/ovx_defs.hpp"
namespace cv
{
@ -370,72 +368,6 @@ static bool ocl_FAST( InputArray _img, std::vector<KeyPoint>& keypoints,
}
#endif
#ifdef HAVE_OPENVX
namespace ovx {
template <> inline bool skipSmallImages<VX_KERNEL_FAST_CORNERS>(int w, int h) { return w*h < 800 * 600; }
}
static bool openvx_FAST(InputArray _img, std::vector<KeyPoint>& keypoints,
int _threshold, bool nonmaxSuppression, int type)
{
using namespace ivx;
// Nonmax suppression is done differently in OpenCV than in OpenVX
// 9/16 is the only supported mode in OpenVX
if(nonmaxSuppression || type != FastFeatureDetector::TYPE_9_16)
return false;
Mat imgMat = _img.getMat();
if(imgMat.empty() || imgMat.type() != CV_8UC1)
return false;
if (ovx::skipSmallImages<VX_KERNEL_FAST_CORNERS>(imgMat.cols, imgMat.rows))
return false;
try
{
Context context = ovx::getOpenVXContext();
Image img = Image::createFromHandle(context, Image::matTypeToFormat(imgMat.type()),
Image::createAddressing(imgMat), (void*)imgMat.data);
ivx::Scalar threshold = ivx::Scalar::create<VX_TYPE_FLOAT32>(context, _threshold);
vx_size capacity = imgMat.cols * imgMat.rows;
Array corners = Array::create(context, VX_TYPE_KEYPOINT, capacity);
ivx::Scalar numCorners = ivx::Scalar::create<VX_TYPE_SIZE>(context, 0);
IVX_CHECK_STATUS(vxuFastCorners(context, img, threshold, (vx_bool)nonmaxSuppression, corners, numCorners));
size_t nPoints = numCorners.getValue<vx_size>();
keypoints.clear(); keypoints.reserve(nPoints);
std::vector<vx_keypoint_t> vxCorners;
corners.copyTo(vxCorners);
for(size_t i = 0; i < nPoints; i++)
{
vx_keypoint_t kp = vxCorners[i];
//if nonmaxSuppression is false, kp.strength is undefined
keypoints.push_back(KeyPoint((float)kp.x, (float)kp.y, 7.f, -1, kp.strength));
}
#ifdef VX_VERSION_1_1
//we should take user memory back before release
//(it's not done automatically according to standard)
img.swapHandle();
#endif
}
catch (const RuntimeError & e)
{
VX_DbgThrow(e.what());
}
catch (const WrapperError & e)
{
VX_DbgThrow(e.what());
}
return true;
}
#endif
static inline int hal_FAST(cv::Mat& src, std::vector<KeyPoint>& keypoints, int threshold, bool nonmax_suppression, FastFeatureDetector::DetectorType type)
{
if (threshold > 20)
@ -503,13 +435,12 @@ void FAST(InputArray _img, std::vector<KeyPoint>& keypoints, int threshold, bool
cv::Mat img = _img.getMat();
CALL_HAL(fast_dense, hal_FAST, img, keypoints, threshold, nonmax_suppression, type);
size_t keypoints_count;
size_t keypoints_count = 10000;
keypoints.clear();
keypoints.resize(keypoints_count);
CALL_HAL(fast, cv_hal_FAST, img.data, img.step, img.cols, img.rows,
(uchar*)(keypoints.data()), &keypoints_count, threshold, nonmax_suppression, type);
CV_OVX_RUN(true,
openvx_FAST(_img, keypoints, threshold, nonmax_suppression, type))
switch(type) {
case FastFeatureDetector::TYPE_5_8:
FAST_t<8>(_img, keypoints, threshold, nonmax_suppression);

View File

@ -118,8 +118,8 @@ void CV_FastTest::run( int )
read( fs["exp_kps2"], exp_kps2, Mat() );
fs.release();
if ( exp_kps1.size != kps1.size || 0 != cvtest::norm(exp_kps1, kps1, NORM_L2) ||
exp_kps2.size != kps2.size || 0 != cvtest::norm(exp_kps2, kps2, NORM_L2))
if ( exp_kps1.size != kps1.size || 0 != cvtest::norm(exp_kps1, kps1, NORM_L2) ||
exp_kps2.size != kps2.size || 0 != cvtest::norm(exp_kps2, kps2, NORM_L2))
{
ts->set_failed_test_info(cvtest::TS::FAIL_MISMATCH);
return;
@ -135,4 +135,34 @@ void CV_FastTest::run( int )
TEST(Features2d_FAST, regression) { CV_FastTest test; test.safe_run(); }
// #define DUMP_TEST_DATA
TEST(Features2d_FAST, noNMS)
{
Mat img = imread(string(cvtest::TS::ptr()->get_data_path()) + "inpaint/orig.png", cv::IMREAD_GRAYSCALE);
string xml = string(cvtest::TS::ptr()->get_data_path()) + "fast/result_no_nonmax.xml";
vector<KeyPoint> keypoints;
FAST(img, keypoints, 100, false, FastFeatureDetector::DetectorType::TYPE_9_16);
Mat kps(1, (int)(keypoints.size() * sizeof(KeyPoint)), CV_8U, &keypoints[0]);
Mat gt_kps;
FileStorage fs(xml, FileStorage::READ);
#ifdef DUMP_TEST_DATA
if (!fs.isOpened())
{
fs.open(xml, FileStorage::WRITE);
fs << "exp_kps" << kps;
fs.release();
fs.open(xml, FileStorage::READ);
}
#endif
ASSERT_TRUE(fs.isOpened());
fs["exp_kps"] >> gt_kps;
fs.release();
ASSERT_GT(gt_kps.total(), size_t(0));
ASSERT_EQ( 0, cvtest::norm(gt_kps, kps, NORM_L2));
}
}} // namespace