This commit is contained in:
Konstantin Matskevich 2014-01-16 14:10:17 +04:00
parent 4a4151ec97
commit b5f717b6b3
5 changed files with 443 additions and 10 deletions

View File

@ -0,0 +1,76 @@
/*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) 2010-2012, Multicoreware, Inc., all rights reserved.
// Copyright (C) 2010-2012, Advanced Micro Devices, Inc., 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.
//
//M*/
#include "perf_precomp.hpp"
#include "opencv2/ts/ocl_perf.hpp"
#ifdef HAVE_OPENCL
namespace cvtest {
namespace ocl {
typedef std::tr1::tuple<int, int> StereoBMFixture_t;
typedef TestBaseWithParam<StereoBMFixture_t> StereoBMFixture;
OCL_PERF_TEST_P(StereoBMFixture, StereoBM, ::testing::Combine(OCL_PERF_ENUM(32, 64, 128), OCL_PERF_ENUM(11,21) ) )
{
const int n_disp = get<0>(GetParam()), winSize = get<1>(GetParam());
UMat left, right, disp;
imread(getDataPath("gpu/stereobm/aloe-L.png"), IMREAD_GRAYSCALE).copyTo(left);
imread(getDataPath("gpu/stereobm/aloe-R.png"), IMREAD_GRAYSCALE).copyTo(right);
ASSERT_FALSE(left.empty());
ASSERT_FALSE(right.empty());
declare.in(left, right);
Ptr<StereoBM> bm = createStereoBM( n_disp, winSize );
bm->setPreFilterType(bm->PREFILTER_NORMALIZED_RESPONSE);
OCL_TEST_CYCLE() bm->compute(left, right, disp);
SANITY_CHECK(disp, 1e-2, ERROR_RELATIVE);
}
}//ocl
}//cvtest
#endif

View File

@ -0,0 +1,159 @@
/*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) 2010-2012, Institute Of Software Chinese Academy Of Science, all rights reserved.
// Copyright (C) 2010-2012, Advanced Micro Devices, Inc., 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.
//
//M*/
//////////////////////////////////////////////////////////////////////////////////////////////////
////////////////////////////////////////// stereoBM //////////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////
#ifdef SIZE
__kernel void stereoBM(__global const uchar * left, __global const uchar * right, __global uchar * dispptr,
int disp_step, int disp_offset, int rows, int cols, int mindisp, int ndisp,
int preFilterCap, int winsize, int textureTreshold, int uniquenessRatio)
{
int x = get_global_id(0);
int y = get_global_id(1);
int wsz2 = winsize/2;
short FILTERED = (mindisp - 1)<<4;
if(x < cols && y < rows )
{
int dispIdx = mad24(y, disp_step, disp_offset + x*(int)sizeof(short) );
__global short * disp = (__global short*)(dispptr + dispIdx);
disp[0] = FILTERED;
if( (x > mindisp+ndisp+wsz2-2) && (y > wsz2-1) && (x < cols-wsz2-mindisp) && (y < rows - wsz2))
{
int cost[SIZE];
int textsum = 0;
for(int d = mindisp; d < ndisp+mindisp; d++)
{
cost[d-mindisp] = 0;
for(int i = -wsz2; i < wsz2+1; i++)
for(int j = -wsz2; j < wsz2+1; j++)
{
textsum += abs( left[min( y+i, rows-1 ) * cols + min( x+j, cols-1 )] - preFilterCap );
cost[d-mindisp] += abs( left[min( y+i, rows-1 ) * cols + min( x+j, cols-1 )]
- right[min( y+i, rows-1 ) * cols + min( x+j-d, cols-1 )] );
}
}
int best_disp = mindisp, best_cost = cost[0];
for(int d = mindisp; d < ndisp+mindisp; d++)
{
best_cost = (cost[d-mindisp] < best_cost) ? cost[d-mindisp] : best_cost;
best_disp = (best_cost == cost[d-mindisp]) ? d : best_disp;
}
int thresh = best_cost + (best_cost * uniquenessRatio/100);
for(int d = mindisp; (d < ndisp + mindisp) && (uniquenessRatio > 0); d++)
{
best_disp = ( (cost[d-mindisp] <= thresh) && (d < best_disp-1 || d > best_disp + 1) ) ? FILTERED : best_disp;
}
disp[0] = textsum < textureTreshold ? (FILTERED) : (best_disp == FILTERED) ? (short)(best_disp) : (short)(best_disp);
if( best_disp != FILTERED )
{
int y1 = (best_disp > mindisp) ? cost[best_disp-mindisp-1] : cost[best_disp-mindisp+1],
y2 = cost[best_disp-mindisp],
y3 = (best_disp < mindisp+ndisp-1) ? cost[best_disp-mindisp+1] : cost[best_disp-mindisp-1];
float a = (y3 - ((best_disp+1)*(y2-y1) + best_disp*y1 - (best_disp-1)*y2)/(best_disp - (best_disp-1)) )/
((best_disp+1)*((best_disp+1) - (best_disp-1) - best_disp) + (best_disp-1)*best_disp);
float b = (y2 - y1)/(best_disp - (best_disp-1)) - a*((best_disp-1)+best_disp);
disp[0] = (y1 == y2 || y2 == y3) ? (short)(best_disp*16) : (short)(-b/(2*a)*16);
}
}
}
}
#endif
//////////////////////////////////////////////////////////////////////////////////////////////////
/////////////////////////////////////// Norm Prefiler ////////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////
__kernel void prefilter_norm(__global unsigned char *input, __global unsigned char *output,
int rows, int cols, int prefilterCap, int winsize, int scale_g, int scale_s)
{
int x = get_global_id(0);
int y = get_global_id(1);
int wsz2 = winsize/2;
if(x < cols && y < rows)
{
int cov1 = input[ max(y-1, 0) * cols + x] * 1 +
input[y * cols + max(x-1,0)] * 1 + input[ y * cols + x] * 4 + input[y * cols + min(x+1, cols-1)] * 1 +
input[min(y+1, rows-1) * cols + x] * 1;
int cov2 = 0;
for(int i = -wsz2; i < wsz2+1; i++)
for(int j = -wsz2; j < wsz2+1; j++)
cov2 += input[min( max( (y+i),0 ),rows-1 ) * cols + min( max( (x+j),0 ),cols-1 )];
int res = (cov1*scale_g - cov2*scale_s)>>10;
res = min(min(max(-prefilterCap, res), prefilterCap) + prefilterCap, 255);
output[y * cols + x] = res & 0xFF;
}
}
//////////////////////////////////////////////////////////////////////////////////////////////////
////////////////////////////////////// Sobel Prefiler ////////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////
__kernel void prefilter_xsobel(__global unsigned char *input, __global unsigned char *output,
int rows, int cols, int prefilterCap)
{
int x = get_global_id(0);
int y = get_global_id(1);
output[y * cols + x] = min(prefilterCap, 255) & 0xFF;
if(x < cols && y < rows-1 && x > 0)
{
int cov = input[((y > 0) ? y-1 : y+1) * cols + (x-1)] * (-1) + input[((y > 0) ? y-1 : y+1) * cols + ((x<cols-1) ? x+1 : x-1)] * (1) +
input[ (y) * cols + (x-1)] * (-2) + input[ (y) * cols + ((x<cols-1) ? x+1 : x-1)] * (2) +
input[ (y+1) * cols + (x-1)] * (-1) + input[ (y+1) * cols + ((x<cols-1) ? x+1 : x-1)] * (1);
cov = min(min(max(-prefilterCap, cov), prefilterCap) + prefilterCap, 255);
output[y * cols + x] = cov & 0xFF;
}
}

View File

@ -49,6 +49,8 @@
#include "opencv2/core/private.hpp"
#include "opencv2/core/ocl.hpp"
#ifdef HAVE_TEGRA_OPTIMIZATION
#include "opencv2/calib3d/calib3d_tegra.hpp"
#else

View File

@ -48,6 +48,7 @@
#include "precomp.hpp"
#include <stdio.h>
#include <limits>
#include "opencl_kernels.hpp"
namespace cv
{
@ -85,6 +86,26 @@ struct StereoBMParams
int dispType;
};
static bool ocl_prefilter_norm(InputArray _input, OutputArray _output, int winsize, int prefilterCap)
{
ocl::Kernel k("prefilter_norm", ocl::calib3d::stereobm_oclsrc);
if(k.empty())
return false;
int scale_g = winsize*winsize/8, scale_s = (1024 + scale_g)/(scale_g*2);
scale_g *= scale_s;
UMat input = _input.getUMat(), output;
_output.create(input.size(), input.type());
output = _output.getUMat();
size_t globalThreads[3] = { input.cols, input.rows, 1 };
k.args(ocl::KernelArg::PtrReadOnly(input), ocl::KernelArg::PtrWriteOnly(output), input.rows, input.cols,
prefilterCap, winsize, scale_g, scale_s);
return k.run(2, globalThreads, NULL, false);
}
static void prefilterNorm( const Mat& src, Mat& dst, int winsize, int ftzero, uchar* buf )
{
@ -149,6 +170,24 @@ static void prefilterNorm( const Mat& src, Mat& dst, int winsize, int ftzero, uc
}
}
static bool ocl_prefilter_xsobel(InputArray _input, OutputArray _output, int prefilterCap)
{
ocl::Kernel k("prefilter_xsobel", ocl::calib3d::stereobm_oclsrc);
if(k.empty())
return false;
UMat input = _input.getUMat(), output;
_output.create(input.size(), input.type());
output = _output.getUMat();
size_t blockSize = 1;
size_t globalThreads[3] = { input.cols, input.rows, 1 };
size_t localThreads[3] = { blockSize, blockSize, 1 };
k.args(ocl::KernelArg::PtrReadOnly(input), ocl::KernelArg::PtrWriteOnly(output), input.rows, input.cols, prefilterCap);
return k.run(2, globalThreads, localThreads, false);
}
static void
prefilterXSobel( const Mat& src, Mat& dst, int ftzero )
@ -534,7 +573,6 @@ findStereoCorrespondenceBM( const Mat& left, const Mat& right,
hsad = hsad0 - dy0*ndisp; cbuf = cbuf0 + (x + wsz2 + 1)*cstep - dy0*ndisp;
lptr = lptr0 + std::min(std::max(x, -lofs), width-lofs-1) - dy0*sstep;
rptr = rptr0 + std::min(std::max(x, -rofs), width-rofs-1) - dy0*sstep;
for( y = -dy0; y < height + dy1; y++, hsad += ndisp, cbuf += ndisp, lptr += sstep, rptr += sstep )
{
int lval = lptr[0];
@ -651,6 +689,25 @@ findStereoCorrespondenceBM( const Mat& left, const Mat& right,
}
}
static bool ocl_prefiltering(InputArray left0, InputArray right0, OutputArray left, OutputArray right, StereoBMParams* state)
{
if( state->preFilterType == StereoBM::PREFILTER_NORMALIZED_RESPONSE )
{
if(!ocl_prefilter_norm( left0, left, state->preFilterSize, state->preFilterCap))
return false;
if(!ocl_prefilter_norm( right0, right, state->preFilterSize, state->preFilterCap))
return false;
}
else
{
if(!ocl_prefilter_xsobel( left0, left, state->preFilterCap ))
return false;
if(!ocl_prefilter_xsobel( right0, right, state->preFilterCap))
return false;
}
return true;
}
struct PrefilterInvoker : public ParallelLoopBody
{
PrefilterInvoker(const Mat& left0, const Mat& right0, Mat& left, Mat& right,
@ -679,6 +736,32 @@ struct PrefilterInvoker : public ParallelLoopBody
StereoBMParams* state;
};
static bool ocl_stereo( InputArray _left, InputArray _right,
OutputArray _disp, StereoBMParams* state)
{
ocl::Kernel k("stereoBM", ocl::calib3d::stereobm_oclsrc, cv::format("-D SIZE=%d", state->numDisparities ) );
if(k.empty())
return false;
UMat left = _left.getUMat(), right = _right.getUMat();
_disp.create(_left.size(), CV_16S);
UMat disp = _disp.getUMat();
size_t globalThreads[3] = { left.cols, left.rows, 1 };
int idx = 0;
idx = k.set(idx, ocl::KernelArg::PtrReadOnly(left));
idx = k.set(idx, ocl::KernelArg::PtrReadOnly(right));
idx = k.set(idx, ocl::KernelArg::WriteOnly(disp));
idx = k.set(idx, state->minDisparity);
idx = k.set(idx, state->numDisparities);
idx = k.set(idx, state->preFilterCap);
idx = k.set(idx, state->SADWindowSize);
idx = k.set(idx, state->textureThreshold);
idx = k.set(idx, state->uniquenessRatio);
return k.run(2, globalThreads, NULL, false);
}
struct FindStereoCorrespInvoker : public ParallelLoopBody
{
@ -776,21 +859,18 @@ public:
void compute( InputArray leftarr, InputArray rightarr, OutputArray disparr )
{
Mat left0 = leftarr.getMat(), right0 = rightarr.getMat();
int dtype = disparr.fixedType() ? disparr.type() : params.dispType;
Size leftsize = leftarr.size();
if (left0.size() != right0.size())
if (leftarr.size() != rightarr.size())
CV_Error( Error::StsUnmatchedSizes, "All the images must have the same size" );
if (left0.type() != CV_8UC1 || right0.type() != CV_8UC1)
if (leftarr.type() != CV_8UC1 || rightarr.type() != CV_8UC1)
CV_Error( Error::StsUnsupportedFormat, "Both input images must have CV_8UC1" );
if (dtype != CV_16SC1 && dtype != CV_32FC1)
CV_Error( Error::StsUnsupportedFormat, "Disparity image must have CV_16SC1 or CV_32FC1 format" );
disparr.create(left0.size(), dtype);
Mat disp0 = disparr.getMat();
if( params.preFilterType != PREFILTER_NORMALIZED_RESPONSE &&
params.preFilterType != PREFILTER_XSOBEL )
CV_Error( Error::StsOutOfRange, "preFilterType must be = CV_STEREO_BM_NORMALIZED_RESPONSE" );
@ -802,7 +882,7 @@ public:
CV_Error( Error::StsOutOfRange, "preFilterCap must be within 1..63" );
if( params.SADWindowSize < 5 || params.SADWindowSize > 255 || params.SADWindowSize % 2 == 0 ||
params.SADWindowSize >= std::min(left0.cols, left0.rows) )
params.SADWindowSize >= std::min(leftsize.width, leftsize.height) )
CV_Error( Error::StsOutOfRange, "SADWindowSize must be odd, be within 5..255 and be not larger than image width or height" );
if( params.numDisparities <= 0 || params.numDisparities % 16 != 0 )
@ -814,6 +894,26 @@ public:
if( params.uniquenessRatio < 0 )
CV_Error( Error::StsOutOfRange, "uniqueness ratio must be non-negative" );
int FILTERED = (params.minDisparity - 1) << DISPARITY_SHIFT;
if(ocl::useOpenCL() && disparr.isUMat())
{
UMat left, right;
CV_Assert(ocl_prefiltering(leftarr, rightarr, left, right, &params));
CV_Assert(ocl_stereo(left, right, disparr, &params));
if( params.speckleRange >= 0 && params.speckleWindowSize > 0 )
filterSpeckles(disparr.getMat(), FILTERED, params.speckleWindowSize, params.speckleRange, slidingSumBuf);
if (dtype == CV_32F)
disparr.getUMat().convertTo(disparr, CV_32FC1, 1./(1 << DISPARITY_SHIFT), 0);
return;
}
Mat left0 = leftarr.getMat(), right0 = rightarr.getMat();
disparr.create(left0.size(), dtype);
Mat disp0 = disparr.getMat();
preFilteredImg0.create( left0.size(), CV_8U );
preFilteredImg1.create( left0.size(), CV_8U );
cost.create( left0.size(), CV_16S );
@ -828,7 +928,6 @@ public:
int lofs = std::max(ndisp - 1 + mindisp, 0);
int rofs = -std::min(ndisp - 1 + mindisp, 0);
int width1 = width - rofs - ndisp + 1;
int FILTERED = (params.minDisparity - 1) << DISPARITY_SHIFT;
if( lofs >= width || rofs >= width || width1 < 1 )
{
@ -855,7 +954,7 @@ public:
bufSize2 = width*height*(sizeof(Point_<short>) + sizeof(int) + sizeof(uchar));
#if CV_SSE2
bool useShorts = params.preFilterCap <= 31 && params.SADWindowSize <= 21 && checkHardwareSupport(CV_CPU_SSE2);
bool useShorts = false;//params.preFilterCap <= 31 && params.SADWindowSize <= 21 && checkHardwareSupport(CV_CPU_SSE2);
#else
const bool useShorts = false;
#endif
@ -870,6 +969,7 @@ public:
slidingSumBuf.create( 1, bufSize, CV_8U );
uchar *_buf = slidingSumBuf.data;
parallel_for_(Range(0, 2), PrefilterInvoker(left0, right0, left, right, _buf, _buf + bufSize1, &params), 1);
Rect validDisparityRect(0, 0, width, height), R1 = params.roi1, R2 = params.roi2;

View File

@ -0,0 +1,96 @@
///////////////////////////////////////////////////////////////////////////////////////
//
// 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) 2010-2012, Institute Of Software Chinese Academy Of Science, all rights reserved.
// Copyright (C) 2010-2012, Advanced Micro Devices, Inc., 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.
//
//M*/
#include "test_precomp.hpp"
#include "cvconfig.h"
#include "opencv2/ts/ocl_test.hpp"
#ifdef HAVE_OPENCL
namespace cvtest {
namespace ocl {
PARAM_TEST_CASE(StereoBMFixture, int, int)
{
int n_disp;
int winSize;
Mat left, right, disp;
UMat uleft, uright, udisp;
virtual void SetUp()
{
n_disp = GET_PARAM(0);
winSize = GET_PARAM(1);
left = readImage("gpu/stereobm/aloe-L.png", IMREAD_GRAYSCALE);
right = readImage("gpu/stereobm/aloe-R.png", IMREAD_GRAYSCALE);
ASSERT_FALSE(left.empty());
ASSERT_FALSE(right.empty());
left.copyTo(uleft);
right.copyTo(uright);
}
void Near(double eps = 0.0)
{
EXPECT_MAT_NEAR_RELATIVE(disp, udisp, eps);
}
};
OCL_TEST_P(StereoBMFixture, StereoBM)
{
Ptr<StereoBM> bm = createStereoBM( n_disp, winSize);
bm->setPreFilterType(bm->PREFILTER_XSOBEL);
OCL_OFF(bm->compute(left, right, disp));
OCL_ON(bm->compute(uleft, uright, udisp));
Near(1e-2);
}
OCL_INSTANTIATE_TEST_CASE_P(StereoMatcher, StereoBMFixture, testing::Combine(testing::Values(128),
testing::Values(15)));
}//ocl
}//cvtest
#endif //HAVE_OPENCL