potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-30 15:53:03 +04:00
Code Packages Releases Wiki Activity

fixed octave computation in SURF_GPU

Browse Source 
used random images in gpu filter tests
This commit is contained in:
Vladislav Vinogradov
2012-03-26 18:07:03 +00:00
parent 903c05db1a
commit 089a835c0a
7 changed files with 200 additions and 385 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/gpu/src/cuda/surf.cu
+7 -6
View File
@@ -117,7 +117,7 @@ namespace cv { namespace gpu { namespace device
template <int N> __device__ float icvCalcHaarPatternSum(const float src[][5], int oldSize, int newSize, int y, int x)
{
#if __CUDA_ARCH__ >= 200
typedef double real_t;
typedef double real_t;
#else
typedef float real_t;
#endif
@@ -248,7 +248,7 @@ namespace cv { namespace gpu { namespace device
template <typename Mask>
__global__ void icvFindMaximaInLayer(const PtrStepf det, const PtrStepf trace, int4* maxPosBuffer, unsigned int* maxCounter)
{
#if defined (__CUDA_ARCH__) && __CUDA_ARCH__ >= 110
#if __CUDA_ARCH__ >= 110
extern __shared__ float N9[];
@@ -368,10 +368,10 @@ namespace cv { namespace gpu { namespace device
// INTERPOLATION
__global__ void icvInterpolateKeypoint(const PtrStepf det, const int4* maxPosBuffer,
float* featureX, float* featureY, int* featureLaplacian, float* featureSize, float* featureHessian,
float* featureX, float* featureY, int* featureLaplacian, int* featureOctave, float* featureSize, float* featureHessian,
unsigned int* featureCounter)
{
#if defined (__CUDA_ARCH__) && __CUDA_ARCH__ >= 110
#if __CUDA_ARCH__ >= 110
const int4 maxPos = maxPosBuffer[blockIdx.x];
@@ -459,6 +459,7 @@ namespace cv { namespace gpu { namespace device
featureX[ind] = px;
featureY[ind] = py;
featureLaplacian[ind] = maxPos.w;
featureOctave[ind] = c_octave;
featureSize[ind] = psize;
featureHessian[ind] = N9[1][1][1];
}
@@ -471,7 +472,7 @@ namespace cv { namespace gpu { namespace device
}
void icvInterpolateKeypoint_gpu(const PtrStepf& det, const int4* maxPosBuffer, unsigned int maxCounter,
float* featureX, float* featureY, int* featureLaplacian, float* featureSize, float* featureHessian,
float* featureX, float* featureY, int* featureLaplacian, int* featureOctave, float* featureSize, float* featureHessian,
unsigned int* featureCounter)
{
dim3 threads;
@@ -482,7 +483,7 @@ namespace cv { namespace gpu { namespace device
dim3 grid;
grid.x = maxCounter;
icvInterpolateKeypoint<<<grid, threads>>>(det, maxPosBuffer, featureX, featureY, featureLaplacian, featureSize, featureHessian, featureCounter);
icvInterpolateKeypoint<<<grid, threads>>>(det, maxPosBuffer, featureX, featureY, featureLaplacian, featureOctave, featureSize, featureHessian, featureCounter);
cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() );
modules/gpu/src/filtering.cpp
+3 -1
View File
@@ -948,7 +948,9 @@ namespace
{
DeviceInfo devInfo;
int cc = devInfo.majorVersion() * 10 + devInfo.minorVersion();
CV_Assert(cc >= 20 || ksize <= 16);
if (ksize > 16 && cc < 20)
CV_Error(CV_StsNotImplemented, "column linear filter doesn't implemented for kernel size > 16 for device with compute capabilities less than 2.0");
func(src, dst, kernel.ptr<float>(), ksize, anchor, brd_type, cc, StreamAccessor::getStream(s));
}
modules/gpu/src/surf.cpp
+30 -50
View File
@@ -80,7 +80,7 @@ namespace cv { namespace gpu { namespace device
int img_rows, int img_cols, int octave, bool use_mask, int nLayers);
void icvInterpolateKeypoint_gpu(const PtrStepf& det, const int4* maxPosBuffer, unsigned int maxCounter,
float* featureX, float* featureY, int* featureLaplacian, float* featureSize, float* featureHessian,
float* featureX, float* featureY, int* featureLaplacian, int* featureOctave, float* featureSize, float* featureHessian,
unsigned int* featureCounter);
void icvCalcOrientation_gpu(const float* featureX, const float* featureY, const float* featureSize, float* featureDir, int nFeatures);
@@ -161,7 +161,7 @@ namespace
ensureSizeIsEnough(img_rows * (surf_.nOctaveLayers + 2), img_cols, CV_32FC1, surf_.trace);
ensureSizeIsEnough(1, maxCandidates, CV_32SC4, surf_.maxPosBuffer);
ensureSizeIsEnough(SURF_GPU::SF_FEATURE_STRIDE, maxFeatures, CV_32FC1, keypoints);
ensureSizeIsEnough(SURF_GPU::ROWS_COUNT, maxFeatures, CV_32FC1, keypoints);
keypoints.setTo(Scalar::all(0));
for (int octave = 0; octave < surf_.nOctaves; ++octave)
@@ -183,9 +183,10 @@ namespace
if (maxCounter > 0)
{
icvInterpolateKeypoint_gpu(surf_.det, surf_.maxPosBuffer.ptr<int4>(), maxCounter,
keypoints.ptr<float>(SURF_GPU::SF_X), keypoints.ptr<float>(SURF_GPU::SF_Y),
keypoints.ptr<int>(SURF_GPU::SF_LAPLACIAN), keypoints.ptr<float>(SURF_GPU::SF_SIZE),
keypoints.ptr<float>(SURF_GPU::SF_HESSIAN), counters.ptr<unsigned int>());
keypoints.ptr<float>(SURF_GPU::X_ROW), keypoints.ptr<float>(SURF_GPU::Y_ROW),
keypoints.ptr<int>(SURF_GPU::LAPLACIAN_ROW), keypoints.ptr<int>(SURF_GPU::OCTAVE_ROW),
keypoints.ptr<float>(SURF_GPU::SIZE_ROW), keypoints.ptr<float>(SURF_GPU::HESSIAN_ROW),
counters.ptr<unsigned int>());
}
}
unsigned int featureCounter;
@@ -195,7 +196,7 @@ namespace
keypoints.cols = featureCounter;
if (surf_.upright)
keypoints.row(SURF_GPU::SF_DIR).setTo(Scalar::all(90.0));
keypoints.row(SURF_GPU::ANGLE_ROW).setTo(Scalar::all(90.0));
else
findOrientation(keypoints);
}
@@ -205,8 +206,8 @@ namespace
const int nFeatures = keypoints.cols;
if (nFeatures > 0)
{
icvCalcOrientation_gpu(keypoints.ptr<float>(SURF_GPU::SF_X), keypoints.ptr<float>(SURF_GPU::SF_Y),
keypoints.ptr<float>(SURF_GPU::SF_SIZE), keypoints.ptr<float>(SURF_GPU::SF_DIR), nFeatures);
icvCalcOrientation_gpu(keypoints.ptr<float>(SURF_GPU::X_ROW), keypoints.ptr<float>(SURF_GPU::Y_ROW),
keypoints.ptr<float>(SURF_GPU::SIZE_ROW), keypoints.ptr<float>(SURF_GPU::ANGLE_ROW), nFeatures);
}
}
@@ -216,8 +217,8 @@ namespace
if (nFeatures > 0)
{
ensureSizeIsEnough(nFeatures, descriptorSize, CV_32F, descriptors);
compute_descriptors_gpu(descriptors, keypoints.ptr<float>(SURF_GPU::SF_X), keypoints.ptr<float>(SURF_GPU::SF_Y),
keypoints.ptr<float>(SURF_GPU::SF_SIZE), keypoints.ptr<float>(SURF_GPU::SF_DIR), nFeatures);
compute_descriptors_gpu(descriptors, keypoints.ptr<float>(SURF_GPU::X_ROW), keypoints.ptr<float>(SURF_GPU::Y_ROW),
keypoints.ptr<float>(SURF_GPU::SIZE_ROW), keypoints.ptr<float>(SURF_GPU::ANGLE_ROW), nFeatures);
}
}
@@ -266,20 +267,22 @@ void cv::gpu::SURF_GPU::uploadKeypoints(const vector<KeyPoint>& keypoints, GpuMa
keypointsGPU.release();
else
{
Mat keypointsCPU(SURF_GPU::SF_FEATURE_STRIDE, static_cast<int>(keypoints.size()), CV_32FC1);
Mat keypointsCPU(SURF_GPU::ROWS_COUNT, static_cast<int>(keypoints.size()), CV_32FC1);
float* kp_x = keypointsCPU.ptr<float>(SURF_GPU::SF_X);
float* kp_y = keypointsCPU.ptr<float>(SURF_GPU::SF_Y);
int* kp_laplacian = keypointsCPU.ptr<int>(SURF_GPU::SF_LAPLACIAN);
float* kp_size = keypointsCPU.ptr<float>(SURF_GPU::SF_SIZE);
float* kp_dir = keypointsCPU.ptr<float>(SURF_GPU::SF_DIR);
float* kp_hessian = keypointsCPU.ptr<float>(SURF_GPU::SF_HESSIAN);
float* kp_x = keypointsCPU.ptr<float>(SURF_GPU::X_ROW);
float* kp_y = keypointsCPU.ptr<float>(SURF_GPU::Y_ROW);
int* kp_laplacian = keypointsCPU.ptr<int>(SURF_GPU::LAPLACIAN_ROW);
int* kp_octave = keypointsCPU.ptr<int>(SURF_GPU::OCTAVE_ROW);
float* kp_size = keypointsCPU.ptr<float>(SURF_GPU::SIZE_ROW);
float* kp_dir = keypointsCPU.ptr<float>(SURF_GPU::ANGLE_ROW);
float* kp_hessian = keypointsCPU.ptr<float>(SURF_GPU::HESSIAN_ROW);
for (size_t i = 0, size = keypoints.size(); i < size; ++i)
{
const KeyPoint& kp = keypoints[i];
kp_x[i] = kp.pt.x;
kp_y[i] = kp.pt.y;
kp_octave[i] = kp.octave;
kp_size[i] = kp.size;
kp_dir[i] = kp.angle;
kp_hessian[i] = kp.response;
@@ -290,30 +293,6 @@ void cv::gpu::SURF_GPU::uploadKeypoints(const vector<KeyPoint>& keypoints, GpuMa
}
}
namespace
{
int getPointOctave(float size, const SURF_GPU& params)
{
int best_octave = 0;
float min_diff = numeric_limits<float>::max();
for (int octave = 1; octave < params.nOctaves; ++octave)
{
for (int layer = 0; layer < params.nOctaveLayers; ++layer)
{
float diff = std::abs(size - (float)calcSize(octave, layer));
if (min_diff > diff)
{
min_diff = diff;
best_octave = octave;
if (min_diff == 0)
return best_octave;
}
}
}
return best_octave;
}
}
void cv::gpu::SURF_GPU::downloadKeypoints(const GpuMat& keypointsGPU, vector<KeyPoint>& keypoints)
{
const int nFeatures = keypointsGPU.cols;
@@ -322,18 +301,19 @@ void cv::gpu::SURF_GPU::downloadKeypoints(const GpuMat& keypointsGPU, vector<Key
keypoints.clear();
else
{
CV_Assert(keypointsGPU.type() == CV_32FC1 && keypointsGPU.rows == SF_FEATURE_STRIDE);
CV_Assert(keypointsGPU.type() == CV_32FC1 && keypointsGPU.rows == ROWS_COUNT);
Mat keypointsCPU(keypointsGPU);
keypoints.resize(nFeatures);
float* kp_x = keypointsCPU.ptr<float>(SF_X);
float* kp_y = keypointsCPU.ptr<float>(SF_Y);
int* kp_laplacian = keypointsCPU.ptr<int>(SF_LAPLACIAN);
float* kp_size = keypointsCPU.ptr<float>(SF_SIZE);
float* kp_dir = keypointsCPU.ptr<float>(SF_DIR);
float* kp_hessian = keypointsCPU.ptr<float>(SF_HESSIAN);
float* kp_x = keypointsCPU.ptr<float>(SURF_GPU::X_ROW);
float* kp_y = keypointsCPU.ptr<float>(SURF_GPU::Y_ROW);
int* kp_laplacian = keypointsCPU.ptr<int>(SURF_GPU::LAPLACIAN_ROW);
int* kp_octave = keypointsCPU.ptr<int>(SURF_GPU::OCTAVE_ROW);
float* kp_size = keypointsCPU.ptr<float>(SURF_GPU::SIZE_ROW);
float* kp_dir = keypointsCPU.ptr<float>(SURF_GPU::ANGLE_ROW);
float* kp_hessian = keypointsCPU.ptr<float>(SURF_GPU::HESSIAN_ROW);
for (int i = 0; i < nFeatures; ++i)
{
@@ -341,10 +321,10 @@ void cv::gpu::SURF_GPU::downloadKeypoints(const GpuMat& keypointsGPU, vector<Key
kp.pt.x = kp_x[i];
kp.pt.y = kp_y[i];
kp.class_id = kp_laplacian[i];
kp.octave = kp_octave[i];
kp.size = kp_size[i];
kp.angle = kp_dir[i];
kp.response = kp_hessian[i];
kp.octave = getPointOctave(kp.size, *this);
}
}
}