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

added sipport of BORDER_REPLICATE mode into gpu::corner* functions

Browse Source
This commit is contained in:
Alexey Spizhevoy 2010-12-06 07:47:26 +00:00
parent bb9bf53eb1
commit 6ef4d9b1dd
6 changed files with 162 additions and 26 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
modules/gpu/src/cuda/border_interpolate.hpp
View File

@ -44,7 +44,7 @@ namespace cv { namespace gpu {
struct BrdReflect101
{
BrdReflect101(int len) : last(len - 1) {}
BrdReflect101(int len): last(len - 1) {}
__device__ int idx_low(int i) const
{
@ -73,7 +73,7 @@ namespace cv { namespace gpu {
template <typename T>
struct BrdRowReflect101: BrdReflect101
{
BrdRowReflect101(int len) : BrdReflect101(len) {}
BrdRowReflect101(int len): BrdReflect101(len) {}
__device__ float at_low(int i, const T* data) const
{
@ -90,7 +90,7 @@ namespace cv { namespace gpu {
template <typename T>
struct BrdColReflect101: BrdReflect101
{
BrdColReflect101(int len, int step) : BrdReflect101(len), step(step) {}
BrdColReflect101(int len, int step): BrdReflect101(len), step(step) {}
__device__ float at_low(int i, const T* data) const
{
@ -108,7 +108,7 @@ namespace cv { namespace gpu {
struct BrdReplicate
{
BrdReplicate(int len) : last(len - 1) {}
BrdReplicate(int len): last(len - 1) {}
__device__ int idx_low(int i) const
{
@ -122,7 +122,7 @@ namespace cv { namespace gpu {
__device__ int idx(int i) const
{
return min(max(i, last), 0);
return max(min(i, last), 0);
}
bool is_range_safe(int mini, int maxi) const
@ -137,7 +137,7 @@ namespace cv { namespace gpu {
template <typename T>
struct BrdRowReplicate: BrdReplicate
{
BrdRowReplicate(int len) : BrdReplicate(len) {}
BrdRowReplicate(int len): BrdReplicate(len) {}
__device__ float at_low(int i, const T* data) const
{
@ -154,7 +154,7 @@ namespace cv { namespace gpu {
template <typename T>
struct BrdColReplicate: BrdReplicate
{
BrdColReplicate(int len, int step) : BrdReplicate(len), step(step) {}
BrdColReplicate(int len, int step): BrdReplicate(len), step(step) {}
__device__ float at_low(int i, const T* data) const
{

4
modules/gpu/src/cuda/cuda_shared.hpp
View File

@ -99,9 +99,9 @@ namespace cv
// border interpolation modes)
enum
{
BORDER_REFLECT101 = 0
BORDER_REFLECT101 = 0,
BORDER_REPLICATE
};
}
}

87
modules/gpu/src/cuda/imgproc.cu
View File

@ -42,6 +42,7 @@
#include "cuda_shared.hpp"
#include "border_interpolate.hpp"
#include <stdio.h>
using namespace cv::gpu;
@ -498,6 +499,39 @@ namespace cv { namespace gpu { namespace imgproc
texture<float, 2> harrisDxTex;
texture<float, 2> harrisDyTex;
__global__ void cornerHarris_kernel(const int cols, const int rows, const int block_size, const float k,
PtrStep dst)
{
const unsigned int x = blockIdx.x * blockDim.x + threadIdx.x;
const unsigned int y = blockIdx.y * blockDim.y + threadIdx.y;
if (x < cols && y < rows)
{
float a = 0.f;
float b = 0.f;
float c = 0.f;
const int ibegin = y - (block_size / 2);
const int jbegin = x - (block_size / 2);
const int iend = ibegin + block_size;
const int jend = jbegin + block_size;
for (int i = ibegin; i < iend; ++i)
{
for (int j = jbegin; j < jend; ++j)
{
float dx = tex2D(harrisDxTex, j, i);
float dy = tex2D(harrisDyTex, j, i);
a += dx * dx;
b += dx * dy;
c += dy * dy;
}
}
((float*)dst.ptr(y))[x] = a * c - b * b - k * (a + c) * (a + c);
}
}
template <typename B>
__global__ void cornerHarris_kernel(const int cols, const int rows, const int block_size, const float k,
PtrStep dst, B border_row, B border_col)
@ -555,6 +589,13 @@ namespace cv { namespace gpu { namespace imgproc
cornerHarris_kernel<<<grid, threads>>>(
cols, rows, block_size, k, dst, BrdReflect101(cols), BrdReflect101(rows));
break;
case BORDER_REPLICATE:
harrisDxTex.addressMode[0] = cudaAddressModeClamp;
harrisDxTex.addressMode[1] = cudaAddressModeClamp;
harrisDyTex.addressMode[0] = cudaAddressModeClamp;
harrisDyTex.addressMode[1] = cudaAddressModeClamp;
cornerHarris_kernel<<<grid, threads>>>(cols, rows, block_size, k, dst);
break;
}
cudaSafeCall(cudaThreadSynchronize());
@ -567,6 +608,42 @@ namespace cv { namespace gpu { namespace imgproc
texture<float, 2> minEigenValDxTex;
texture<float, 2> minEigenValDyTex;
__global__ void cornerMinEigenVal_kernel(const int cols, const int rows, const int block_size,
PtrStep dst)
{
const unsigned int x = blockIdx.x * blockDim.x + threadIdx.x;
const unsigned int y = blockIdx.y * blockDim.y + threadIdx.y;
if (x < cols && y < rows)
{
float a = 0.f;
float b = 0.f;
float c = 0.f;
const int ibegin = y - (block_size / 2);
const int jbegin = x - (block_size / 2);
const int iend = ibegin + block_size;
const int jend = jbegin + block_size;
for (int i = ibegin; i < iend; ++i)
{
for (int j = jbegin; j < jend; ++j)
{
float dx = tex2D(minEigenValDxTex, j, i);
float dy = tex2D(minEigenValDyTex, j, i);
a += dx * dx;
b += dx * dy;
c += dy * dy;
}
}
a *= 0.5f;
c *= 0.5f;
((float*)dst.ptr(y))[x] = (a + c) - sqrtf((a - c) * (a - c) + b * b);
}
}
template <typename B>
__global__ void cornerMinEigenVal_kernel(const int cols, const int rows, const int block_size,
PtrStep dst, B border_row, B border_col)
@ -624,8 +701,14 @@ namespace cv { namespace gpu { namespace imgproc
{
case BORDER_REFLECT101:
cornerMinEigenVal_kernel<<<grid, threads>>>(
cols, rows, block_size, dst,
BrdReflect101(cols), BrdReflect101(rows));
cols, rows, block_size, dst, BrdReflect101(cols), BrdReflect101(rows));
break;
case BORDER_REPLICATE:
minEigenValDxTex.addressMode[0] = cudaAddressModeClamp;
minEigenValDxTex.addressMode[1] = cudaAddressModeClamp;
minEigenValDyTex.addressMode[0] = cudaAddressModeClamp;
minEigenValDyTex.addressMode[1] = cudaAddressModeClamp;
cornerMinEigenVal_kernel<<<grid, threads>>>(cols, rows, block_size, dst);
break;
}

8
modules/gpu/src/cuda/linear_filters_beta.cu
View File

@ -150,7 +150,8 @@ void rowFilterCaller(const DevMem2D_<T> src, PtrStepf dst, int anchor,
static const Caller callers[] =
{
rowFilterCaller<T, BrdRowReflect101<T> >
rowFilterCaller<T, BrdRowReflect101<T> >,
rowFilterCaller<T, BrdRowReplicate<T> >
};
callers[brd_interp](src, dst, anchor, kernel, ksize);
@ -251,7 +252,8 @@ void colFilterCaller(const DevMem2D_<T> src, PtrStepf dst, int anchor,
static const Caller callers[] =
{
colFilterCaller<T, BrdColReflect101<T> >
colFilterCaller<T, BrdColReflect101<T> >,
colFilterCaller<T, BrdColReplicate<T> >
};
callers[brd_interp](src, dst, anchor, kernel, ksize);
@ -261,4 +263,4 @@ void colFilterCaller(const DevMem2D_<T> src, PtrStepf dst, int anchor,
template void colFilterCaller<unsigned char>(const DevMem2D_<unsigned char>, PtrStepf, int, const float*, int, int);
template void colFilterCaller<float>(const DevMem2D_<float>, PtrStepf, int, const float*, int, int);
}}}
}}}

12
modules/gpu/src/imgproc_gpu.cpp
View File

@ -944,10 +944,13 @@ void cv::gpu::cornerHarris(const GpuMat& src, GpuMat& dst, int blockSize, int ks
switch (borderType)
{
case cv::BORDER_REFLECT101:
gpuBorderType = cv::gpu::BORDER_REFLECT101;
gpuBorderType = cv::gpu::BORDER_REFLECT101;
break;
case cv::BORDER_REPLICATE:
gpuBorderType = cv::gpu::BORDER_REPLICATE;
break;
default:
CV_Error(CV_StsBadArg, "cornerHarris: unsupported border type");
CV_Error(CV_StsBadArg, "cornerHarris: unsupported border extrapolation mode");
}
GpuMat Dx, Dy;
@ -964,8 +967,11 @@ void cv::gpu::cornerMinEigenVal(const GpuMat& src, GpuMat& dst, int blockSize, i
case cv::BORDER_REFLECT101:
gpuBorderType = cv::gpu::BORDER_REFLECT101;
break;
case cv::BORDER_REPLICATE:
gpuBorderType = cv::gpu::BORDER_REPLICATE;
break;
default:
CV_Error(CV_StsBadArg, "cornerMinEigenVal: unsupported border type");
CV_Error(CV_StsBadArg, "cornerMinEigenVal: unsupported border extrapolation mode");
}
GpuMat Dx, Dy;

63
tests/gpu/src/imgproc_gpu.cpp
View File

@ -640,15 +640,37 @@ struct CV_GpuCornerHarrisTest: CvTest
rng.fill(src, RNG::UNIFORM, cv::Scalar(0), cv::Scalar(256));
double k = 0.1;
int borderType = BORDER_REFLECT101;
cv::Mat dst_gold;
cv::cornerHarris(src, dst_gold, blockSize, apertureSize, k, borderType);
cv::gpu::GpuMat dst;
cv::Mat dsth;
int borderType;
borderType = BORDER_REFLECT101;
cv::cornerHarris(src, dst_gold, blockSize, apertureSize, k, borderType);
cv::gpu::cornerHarris(cv::gpu::GpuMat(src), dst, blockSize, apertureSize, k, borderType);
cv::Mat dsth = dst;
dsth = dst;
for (int i = 0; i < dst.rows; ++i)
{
for (int j = 0; j < dst.cols; ++j)
{
float a = dst_gold.at<float>(i, j);
float b = dsth.at<float>(i, j);
if (fabs(a - b) > 1e-3f)
{
ts->printf(CvTS::CONSOLE, "%d %d %f %f %d\n", i, j, a, b, apertureSize);
ts->set_failed_test_info(CvTS::FAIL_INVALID_OUTPUT);
return false;
};
}
}
borderType = BORDER_REPLICATE;
cv::cornerHarris(src, dst_gold, blockSize, apertureSize, k, borderType);
cv::gpu::cornerHarris(cv::gpu::GpuMat(src), dst, blockSize, apertureSize, k, borderType);
dsth = dst;
for (int i = 0; i < dst.rows; ++i)
{
for (int j = 0; j < dst.cols; ++j)
@ -703,15 +725,17 @@ struct CV_GpuCornerMinEigenValTest: CvTest
else if (depth == CV_8U)
rng.fill(src, RNG::UNIFORM, cv::Scalar(0), cv::Scalar(256));
int borderType = BORDER_REFLECT101;
cv::Mat dst_gold;
cv::cornerMinEigenVal(src, dst_gold, blockSize, apertureSize, borderType);
cv::gpu::GpuMat dst;
cv::Mat dsth;
int borderType;
borderType = BORDER_REFLECT101;
cv::cornerMinEigenVal(src, dst_gold, blockSize, apertureSize, borderType);
cv::gpu::cornerMinEigenVal(cv::gpu::GpuMat(src), dst, blockSize, apertureSize, borderType);
cv::Mat dsth = dst;
dsth = dst;
for (int i = 0; i < dst.rows; ++i)
{
for (int j = 0; j < dst.cols; ++j)
@ -726,6 +750,27 @@ struct CV_GpuCornerMinEigenValTest: CvTest
};
}
}
borderType = BORDER_REPLICATE;
cv::cornerMinEigenVal(src, dst_gold, blockSize, apertureSize, borderType);
cv::gpu::cornerMinEigenVal(cv::gpu::GpuMat(src), dst, blockSize, apertureSize, borderType);
dsth = dst;
for (int i = 0; i < dst.rows; ++i)
{
for (int j = 0; j < dst.cols; ++j)
{
float a = dst_gold.at<float>(i, j);
float b = dsth.at<float>(i, j);
if (fabs(a - b) > 1e-2f)
{
ts->printf(CvTS::CONSOLE, "%d %d %f %f %d %d\n", i, j, a, b, apertureSize, blockSize);
ts->set_failed_test_info(CvTS::FAIL_INVALID_OUTPUT);
return false;
};
}
}
return true;
}
};