samples(gpu): cleanup samples for legacy API

This commit is contained in:
Alexander Alekhin 2018-10-31 23:41:49 +00:00
parent 850053f9ca
commit fc21b15d6e
4 changed files with 5 additions and 1072 deletions

View File

@ -21,7 +21,6 @@ set(OPENCV_CUDA_SAMPLES_REQUIRED_DEPS
opencv_cudaoptflow
opencv_cudabgsegm
opencv_cudastereo
opencv_cudalegacy
opencv_cudaobjdetect)
ocv_check_dependencies(${OPENCV_CUDA_SAMPLES_REQUIRED_DEPS})

View File

@ -4,7 +4,6 @@
#include "opencv2/core.hpp"
#include "opencv2/core/utility.hpp"
#include "opencv2/cudabgsegm.hpp"
#include "opencv2/cudalegacy.hpp"
#include "opencv2/video.hpp"
#include "opencv2/highgui.hpp"
@ -16,8 +15,6 @@ enum Method
{
MOG,
MOG2,
GMG,
FGD_STAT
};
int main(int argc, const char** argv)
@ -25,7 +22,7 @@ int main(int argc, const char** argv)
cv::CommandLineParser cmd(argc, argv,
"{ c camera | | use camera }"
"{ f file | ../data/vtest.avi | input video file }"
"{ m method | mog | method (mog, mog2, gmg, fgd) }"
"{ m method | mog | method (mog, mog2) }"
"{ h help | | print help message }");
if (cmd.has("help") || !cmd.check())
@ -40,9 +37,7 @@ int main(int argc, const char** argv)
string method = cmd.get<string>("method");
if (method != "mog"
&& method != "mog2"
&& method != "gmg"
&& method != "fgd")
&& method != "mog2")
{
cerr << "Incorrect method" << endl;
return -1;
@ -50,8 +45,8 @@ int main(int argc, const char** argv)
Method m = method == "mog" ? MOG :
method == "mog2" ? MOG2 :
method == "fgd" ? FGD_STAT :
GMG;
(Method)-1;
CV_Assert(m != (Method)-1);
VideoCapture cap;
@ -73,8 +68,6 @@ int main(int argc, const char** argv)
Ptr<BackgroundSubtractor> mog = cuda::createBackgroundSubtractorMOG();
Ptr<BackgroundSubtractor> mog2 = cuda::createBackgroundSubtractorMOG2();
Ptr<BackgroundSubtractor> gmg = cuda::createBackgroundSubtractorGMG(40);
Ptr<BackgroundSubtractor> fgd = cuda::createBackgroundSubtractorFGD();
GpuMat d_fgmask;
GpuMat d_fgimg;
@ -93,23 +86,12 @@ int main(int argc, const char** argv)
case MOG2:
mog2->apply(d_frame, d_fgmask);
break;
case GMG:
gmg->apply(d_frame, d_fgmask);
break;
case FGD_STAT:
fgd->apply(d_frame, d_fgmask);
break;
}
namedWindow("image", WINDOW_NORMAL);
namedWindow("foreground mask", WINDOW_NORMAL);
namedWindow("foreground image", WINDOW_NORMAL);
if (m != GMG)
{
namedWindow("mean background image", WINDOW_NORMAL);
}
namedWindow("mean background image", WINDOW_NORMAL);
for(;;)
{
@ -132,15 +114,6 @@ int main(int argc, const char** argv)
mog2->apply(d_frame, d_fgmask);
mog2->getBackgroundImage(d_bgimg);
break;
case GMG:
gmg->apply(d_frame, d_fgmask);
break;
case FGD_STAT:
fgd->apply(d_frame, d_fgmask);
fgd->getBackgroundImage(d_bgimg);
break;
}
double fps = cv::getTickFrequency() / (cv::getTickCount() - start);

View File

@ -1,388 +0,0 @@
#if defined _MSC_VER && _MSC_VER >= 1400
#pragma warning( disable : 4201 4408 4127 4100)
#endif
#include <iostream>
#include <iomanip>
#include <cstdio>
#include "opencv2/core/cuda.hpp"
#include "opencv2/cudalegacy.hpp"
#include "opencv2/highgui.hpp"
#include "opencv2/imgproc.hpp"
#include "opencv2/objdetect.hpp"
#include "opencv2/objdetect/objdetect_c.h"
using namespace std;
using namespace cv;
#if !defined(HAVE_CUDA) || defined(__arm__)
int main( int, const char** )
{
#if !defined(HAVE_CUDA)
std::cout << "CUDA support is required (CMake key 'WITH_CUDA' must be true)." << std::endl;
#endif
#if defined(__arm__)
std::cout << "Unsupported for ARM CUDA library." << std::endl;
#endif
return 0;
}
#else
const Size2i preferredVideoFrameSize(640, 480);
const cv::String wndTitle = "NVIDIA Computer Vision :: Haar Classifiers Cascade";
static void matPrint(Mat &img, int lineOffsY, Scalar fontColor, const string &ss)
{
int fontFace = FONT_HERSHEY_DUPLEX;
double fontScale = 0.8;
int fontThickness = 2;
Size fontSize = cv::getTextSize("T[]", fontFace, fontScale, fontThickness, 0);
Point org;
org.x = 1;
org.y = 3 * fontSize.height * (lineOffsY + 1) / 2;
putText(img, ss, org, fontFace, fontScale, Scalar(0,0,0), 5*fontThickness/2, 16);
putText(img, ss, org, fontFace, fontScale, fontColor, fontThickness, 16);
}
static void displayState(Mat &canvas, bool bHelp, bool bGpu, bool bLargestFace, bool bFilter, double fps)
{
Scalar fontColorRed(0,0,255);
Scalar fontColorNV(0,185,118);
ostringstream ss;
ss << "FPS = " << setprecision(1) << fixed << fps;
matPrint(canvas, 0, fontColorRed, ss.str());
ss.str("");
ss << "[" << canvas.cols << "x" << canvas.rows << "], " <<
(bGpu ? "GPU, " : "CPU, ") <<
(bLargestFace ? "OneFace, " : "MultiFace, ") <<
(bFilter ? "Filter:ON" : "Filter:OFF");
matPrint(canvas, 1, fontColorRed, ss.str());
if (bHelp)
{
matPrint(canvas, 2, fontColorNV, "Space - switch GPU / CPU");
matPrint(canvas, 3, fontColorNV, "M - switch OneFace / MultiFace");
matPrint(canvas, 4, fontColorNV, "F - toggle rectangles Filter");
matPrint(canvas, 5, fontColorNV, "H - toggle hotkeys help");
}
else
{
matPrint(canvas, 2, fontColorNV, "H - toggle hotkeys help");
}
}
static NCVStatus process(Mat *srcdst,
Ncv32u width, Ncv32u height,
NcvBool bFilterRects, NcvBool bLargestFace,
HaarClassifierCascadeDescriptor &haar,
NCVVector<HaarStage64> &d_haarStages, NCVVector<HaarClassifierNode128> &d_haarNodes,
NCVVector<HaarFeature64> &d_haarFeatures, NCVVector<HaarStage64> &h_haarStages,
INCVMemAllocator &gpuAllocator,
INCVMemAllocator &cpuAllocator,
cudaDeviceProp &devProp)
{
ncvAssertReturn(!((srcdst == NULL) ^ gpuAllocator.isCounting()), NCV_NULL_PTR);
NCVStatus ncvStat;
NCV_SET_SKIP_COND(gpuAllocator.isCounting());
NCVMatrixAlloc<Ncv8u> d_src(gpuAllocator, width, height);
ncvAssertReturn(d_src.isMemAllocated(), NCV_ALLOCATOR_BAD_ALLOC);
NCVMatrixAlloc<Ncv8u> h_src(cpuAllocator, width, height);
ncvAssertReturn(h_src.isMemAllocated(), NCV_ALLOCATOR_BAD_ALLOC);
NCVVectorAlloc<NcvRect32u> d_rects(gpuAllocator, 100);
ncvAssertReturn(d_rects.isMemAllocated(), NCV_ALLOCATOR_BAD_ALLOC);
NCV_SKIP_COND_BEGIN
for (Ncv32u i=0; i<(Ncv32u)srcdst->rows; i++)
{
memcpy(h_src.ptr() + i * h_src.stride(), srcdst->ptr(i), srcdst->cols);
}
ncvStat = h_src.copySolid(d_src, 0);
ncvAssertReturnNcvStat(ncvStat);
ncvAssertCUDAReturn(cudaStreamSynchronize(0), NCV_CUDA_ERROR);
NCV_SKIP_COND_END
NcvSize32u roi;
roi.width = d_src.width();
roi.height = d_src.height();
Ncv32u numDetections;
ncvStat = ncvDetectObjectsMultiScale_device(
d_src, roi, d_rects, numDetections, haar, h_haarStages,
d_haarStages, d_haarNodes, d_haarFeatures,
haar.ClassifierSize,
(bFilterRects || bLargestFace) ? 4 : 0,
1.2f, 1,
(bLargestFace ? NCVPipeObjDet_FindLargestObject : 0)
| NCVPipeObjDet_VisualizeInPlace,
gpuAllocator, cpuAllocator, devProp, 0);
ncvAssertReturnNcvStat(ncvStat);
ncvAssertCUDAReturn(cudaStreamSynchronize(0), NCV_CUDA_ERROR);
NCV_SKIP_COND_BEGIN
ncvStat = d_src.copySolid(h_src, 0);
ncvAssertReturnNcvStat(ncvStat);
ncvAssertCUDAReturn(cudaStreamSynchronize(0), NCV_CUDA_ERROR);
for (Ncv32u i=0; i<(Ncv32u)srcdst->rows; i++)
{
memcpy(srcdst->ptr(i), h_src.ptr() + i * h_src.stride(), srcdst->cols);
}
NCV_SKIP_COND_END
return NCV_SUCCESS;
}
int main(int argc, const char** argv)
{
cout << "OpenCV / NVIDIA Computer Vision" << endl;
cout << "Face Detection in video and live feed" << endl;
cout << "Syntax: exename <cascade_file> <image_or_video_or_cameraid>" << endl;
cout << "=========================================" << endl;
ncvAssertPrintReturn(cv::cuda::getCudaEnabledDeviceCount() != 0, "No GPU found or the library is compiled without CUDA support", -1);
ncvAssertPrintReturn(argc == 3, "Invalid number of arguments", -1);
cv::cuda::printShortCudaDeviceInfo(cv::cuda::getDevice());
string cascadeName = argv[1];
string inputName = argv[2];
NCVStatus ncvStat;
NcvBool bQuit = false;
VideoCapture capture;
Size2i frameSize;
//open content source
Mat image = imread(inputName);
Mat frame;
if (!image.empty())
{
frameSize.width = image.cols;
frameSize.height = image.rows;
}
else
{
if (!capture.open(inputName))
{
int camid = -1;
istringstream ss(inputName);
int x = 0;
ss >> x;
ncvAssertPrintReturn(capture.open(camid) != 0, "Can't open source", -1);
}
capture >> frame;
ncvAssertPrintReturn(!frame.empty(), "Empty video source", -1);
frameSize.width = frame.cols;
frameSize.height = frame.rows;
}
NcvBool bUseGPU = true;
NcvBool bLargestObject = false;
NcvBool bFilterRects = true;
NcvBool bHelpScreen = false;
CascadeClassifier classifierOpenCV;
ncvAssertPrintReturn(classifierOpenCV.load(cascadeName) != 0, "Error (in OpenCV) opening classifier", -1);
int devId;
ncvAssertCUDAReturn(cudaGetDevice(&devId), -1);
cudaDeviceProp devProp;
ncvAssertCUDAReturn(cudaGetDeviceProperties(&devProp, devId), -1);
cout << "Using GPU: " << devId << "(" << devProp.name <<
"), arch=" << devProp.major << "." << devProp.minor << endl;
//==============================================================================
//
// Load the classifier from file (assuming its size is about 1 mb)
// using a simple allocator
//
//==============================================================================
NCVMemNativeAllocator gpuCascadeAllocator(NCVMemoryTypeDevice, static_cast<Ncv32u>(devProp.textureAlignment));
ncvAssertPrintReturn(gpuCascadeAllocator.isInitialized(), "Error creating cascade GPU allocator", -1);
NCVMemNativeAllocator cpuCascadeAllocator(NCVMemoryTypeHostPinned, static_cast<Ncv32u>(devProp.textureAlignment));
ncvAssertPrintReturn(cpuCascadeAllocator.isInitialized(), "Error creating cascade CPU allocator", -1);
Ncv32u haarNumStages, haarNumNodes, haarNumFeatures;
ncvStat = ncvHaarGetClassifierSize(cascadeName, haarNumStages, haarNumNodes, haarNumFeatures);
ncvAssertPrintReturn(ncvStat == NCV_SUCCESS, "Error reading classifier size (check the file)", -1);
NCVVectorAlloc<HaarStage64> h_haarStages(cpuCascadeAllocator, haarNumStages);
ncvAssertPrintReturn(h_haarStages.isMemAllocated(), "Error in cascade CPU allocator", -1);
NCVVectorAlloc<HaarClassifierNode128> h_haarNodes(cpuCascadeAllocator, haarNumNodes);
ncvAssertPrintReturn(h_haarNodes.isMemAllocated(), "Error in cascade CPU allocator", -1);
NCVVectorAlloc<HaarFeature64> h_haarFeatures(cpuCascadeAllocator, haarNumFeatures);
ncvAssertPrintReturn(h_haarFeatures.isMemAllocated(), "Error in cascade CPU allocator", -1);
HaarClassifierCascadeDescriptor haar;
ncvStat = ncvHaarLoadFromFile_host(cascadeName, haar, h_haarStages, h_haarNodes, h_haarFeatures);
ncvAssertPrintReturn(ncvStat == NCV_SUCCESS, "Error loading classifier", -1);
NCVVectorAlloc<HaarStage64> d_haarStages(gpuCascadeAllocator, haarNumStages);
ncvAssertPrintReturn(d_haarStages.isMemAllocated(), "Error in cascade GPU allocator", -1);
NCVVectorAlloc<HaarClassifierNode128> d_haarNodes(gpuCascadeAllocator, haarNumNodes);
ncvAssertPrintReturn(d_haarNodes.isMemAllocated(), "Error in cascade GPU allocator", -1);
NCVVectorAlloc<HaarFeature64> d_haarFeatures(gpuCascadeAllocator, haarNumFeatures);
ncvAssertPrintReturn(d_haarFeatures.isMemAllocated(), "Error in cascade GPU allocator", -1);
ncvStat = h_haarStages.copySolid(d_haarStages, 0);
ncvAssertPrintReturn(ncvStat == NCV_SUCCESS, "Error copying cascade to GPU", -1);
ncvStat = h_haarNodes.copySolid(d_haarNodes, 0);
ncvAssertPrintReturn(ncvStat == NCV_SUCCESS, "Error copying cascade to GPU", -1);
ncvStat = h_haarFeatures.copySolid(d_haarFeatures, 0);
ncvAssertPrintReturn(ncvStat == NCV_SUCCESS, "Error copying cascade to GPU", -1);
//==============================================================================
//
// Calculate memory requirements and create real allocators
//
//==============================================================================
NCVMemStackAllocator gpuCounter(static_cast<Ncv32u>(devProp.textureAlignment));
ncvAssertPrintReturn(gpuCounter.isInitialized(), "Error creating GPU memory counter", -1);
NCVMemStackAllocator cpuCounter(static_cast<Ncv32u>(devProp.textureAlignment));
ncvAssertPrintReturn(cpuCounter.isInitialized(), "Error creating CPU memory counter", -1);
ncvStat = process(NULL, frameSize.width, frameSize.height,
false, false, haar,
d_haarStages, d_haarNodes,
d_haarFeatures, h_haarStages,
gpuCounter, cpuCounter, devProp);
ncvAssertPrintReturn(ncvStat == NCV_SUCCESS, "Error in memory counting pass", -1);
NCVMemStackAllocator gpuAllocator(NCVMemoryTypeDevice, gpuCounter.maxSize(), static_cast<Ncv32u>(devProp.textureAlignment));
ncvAssertPrintReturn(gpuAllocator.isInitialized(), "Error creating GPU memory allocator", -1);
NCVMemStackAllocator cpuAllocator(NCVMemoryTypeHostPinned, cpuCounter.maxSize(), static_cast<Ncv32u>(devProp.textureAlignment));
ncvAssertPrintReturn(cpuAllocator.isInitialized(), "Error creating CPU memory allocator", -1);
printf("Initialized for frame size [%dx%d]\n", frameSize.width, frameSize.height);
//==============================================================================
//
// Main processing loop
//
//==============================================================================
namedWindow(wndTitle, 1);
Mat frameDisp;
do
{
Mat gray;
cvtColor((image.empty() ? frame : image), gray, cv::COLOR_BGR2GRAY);
//
// process
//
NcvSize32u minSize = haar.ClassifierSize;
if (bLargestObject)
{
Ncv32u ratioX = preferredVideoFrameSize.width / minSize.width;
Ncv32u ratioY = preferredVideoFrameSize.height / minSize.height;
Ncv32u ratioSmallest = min(ratioX, ratioY);
ratioSmallest = max((Ncv32u)(ratioSmallest / 2.5f), (Ncv32u)1);
minSize.width *= ratioSmallest;
minSize.height *= ratioSmallest;
}
Ncv32f avgTime;
NcvTimer timer = ncvStartTimer();
if (bUseGPU)
{
ncvStat = process(&gray, frameSize.width, frameSize.height,
bFilterRects, bLargestObject, haar,
d_haarStages, d_haarNodes,
d_haarFeatures, h_haarStages,
gpuAllocator, cpuAllocator, devProp);
ncvAssertPrintReturn(ncvStat == NCV_SUCCESS, "Error in memory counting pass", -1);
}
else
{
vector<Rect> rectsOpenCV;
classifierOpenCV.detectMultiScale(
gray,
rectsOpenCV,
1.2f,
bFilterRects ? 4 : 0,
(bLargestObject ? CV_HAAR_FIND_BIGGEST_OBJECT : 0)
| CV_HAAR_SCALE_IMAGE,
Size(minSize.width, minSize.height));
for (size_t rt = 0; rt < rectsOpenCV.size(); ++rt)
rectangle(gray, rectsOpenCV[rt], Scalar(255));
}
avgTime = (Ncv32f)ncvEndQueryTimerMs(timer);
cvtColor(gray, frameDisp, cv::COLOR_GRAY2BGR);
displayState(frameDisp, bHelpScreen, bUseGPU, bLargestObject, bFilterRects, 1000.0f / avgTime);
imshow(wndTitle, frameDisp);
//handle input
switch (cv::waitKey(3))
{
case ' ':
bUseGPU = !bUseGPU;
break;
case 'm':
case 'M':
bLargestObject = !bLargestObject;
break;
case 'f':
case 'F':
bFilterRects = !bFilterRects;
break;
case 'h':
case 'H':
bHelpScreen = !bHelpScreen;
break;
case 27:
bQuit = true;
break;
}
// For camera and video file, capture the next image
if (capture.isOpened())
{
capture >> frame;
if (frame.empty())
{
break;
}
}
} while (!bQuit);
cv::destroyWindow(wndTitle);
return 0;
}
#endif //!defined(HAVE_CUDA)

View File

@ -1,651 +0,0 @@
#if defined _MSC_VER && _MSC_VER >= 1400
#pragma warning( disable : 4201 4408 4127 4100)
#endif
#include <iostream>
#include <iomanip>
#include <memory>
#include <exception>
#include <ctime>
#include <ctype.h>
#include <iostream>
#include <iomanip>
#include "opencv2/core/cuda.hpp"
#include "opencv2/cudalegacy.hpp"
#include "opencv2/highgui.hpp"
#include "opencv2/core/core_c.h" // FIXIT legacy API
#include "opencv2/highgui/highgui_c.h" // FIXIT legacy API
#if !defined(HAVE_CUDA)
int main( int, const char** )
{
std::cout << "Please compile the library with CUDA support" << std::endl;
return -1;
}
#else
//using std::shared_ptr;
using cv::Ptr;
#define PARAM_LEFT "--left"
#define PARAM_RIGHT "--right"
#define PARAM_SCALE "--scale"
#define PARAM_ALPHA "--alpha"
#define PARAM_GAMMA "--gamma"
#define PARAM_INNER "--inner"
#define PARAM_OUTER "--outer"
#define PARAM_SOLVER "--solver"
#define PARAM_TIME_STEP "--time-step"
#define PARAM_HELP "--help"
Ptr<INCVMemAllocator> g_pGPUMemAllocator;
Ptr<INCVMemAllocator> g_pHostMemAllocator;
class RgbToMonochrome
{
public:
float operator ()(unsigned char b, unsigned char g, unsigned char r)
{
float _r = static_cast<float>(r)/255.0f;
float _g = static_cast<float>(g)/255.0f;
float _b = static_cast<float>(b)/255.0f;
return (_r + _g + _b)/3.0f;
}
};
class RgbToR
{
public:
float operator ()(unsigned char /*b*/, unsigned char /*g*/, unsigned char r)
{
return static_cast<float>(r)/255.0f;
}
};
class RgbToG
{
public:
float operator ()(unsigned char /*b*/, unsigned char g, unsigned char /*r*/)
{
return static_cast<float>(g)/255.0f;
}
};
class RgbToB
{
public:
float operator ()(unsigned char b, unsigned char /*g*/, unsigned char /*r*/)
{
return static_cast<float>(b)/255.0f;
}
};
template<class T>
NCVStatus CopyData(IplImage *image, Ptr<NCVMatrixAlloc<Ncv32f> >& dst)
{
dst = Ptr<NCVMatrixAlloc<Ncv32f> > (new NCVMatrixAlloc<Ncv32f> (*g_pHostMemAllocator, image->width, image->height));
ncvAssertReturn (dst->isMemAllocated (), NCV_ALLOCATOR_BAD_ALLOC);
unsigned char *row = reinterpret_cast<unsigned char*> (image->imageData);
T convert;
for (int i = 0; i < image->height; ++i)
{
for (int j = 0; j < image->width; ++j)
{
if (image->nChannels < 3)
{
dst->ptr ()[j + i*dst->stride ()] = static_cast<float> (*(row + j*image->nChannels))/255.0f;
}
else
{
unsigned char *color = row + j * image->nChannels;
dst->ptr ()[j +i*dst->stride ()] = convert (color[0], color[1], color[2]);
}
}
row += image->widthStep;
}
return NCV_SUCCESS;
}
template<class T>
NCVStatus CopyData(const IplImage *image, const NCVMatrixAlloc<Ncv32f> &dst)
{
unsigned char *row = reinterpret_cast<unsigned char*> (image->imageData);
T convert;
for (int i = 0; i < image->height; ++i)
{
for (int j = 0; j < image->width; ++j)
{
if (image->nChannels < 3)
{
dst.ptr ()[j + i*dst.stride ()] = static_cast<float>(*(row + j*image->nChannels))/255.0f;
}
else
{
unsigned char *color = row + j * image->nChannels;
dst.ptr ()[j +i*dst.stride()] = convert (color[0], color[1], color[2]);
}
}
row += image->widthStep;
}
return NCV_SUCCESS;
}
static NCVStatus LoadImages (const char *frame0Name,
const char *frame1Name,
int &width,
int &height,
Ptr<NCVMatrixAlloc<Ncv32f> > &src,
Ptr<NCVMatrixAlloc<Ncv32f> > &dst,
IplImage *&firstFrame,
IplImage *&lastFrame)
{
IplImage *image;
image = cvLoadImage (frame0Name);
if (image == 0)
{
std::cout << "Could not open '" << frame0Name << "'\n";
return NCV_FILE_ERROR;
}
firstFrame = image;
// copy data to src
ncvAssertReturnNcvStat (CopyData<RgbToMonochrome> (image, src));
IplImage *image2;
image2 = cvLoadImage (frame1Name);
if (image2 == 0)
{
std::cout << "Could not open '" << frame1Name << "'\n";
return NCV_FILE_ERROR;
}
lastFrame = image2;
ncvAssertReturnNcvStat (CopyData<RgbToMonochrome> (image2, dst));
width = image->width;
height = image->height;
return NCV_SUCCESS;
}
template<typename T>
inline T Clamp (T x, T a, T b)
{
return ((x) > (a) ? ((x) < (b) ? (x) : (b)) : (a));
}
template<typename T>
inline T MapValue (T x, T a, T b, T c, T d)
{
x = Clamp (x, a, b);
return c + (d - c) * (x - a) / (b - a);
}
static NCVStatus ShowFlow (NCVMatrixAlloc<Ncv32f> &u, NCVMatrixAlloc<Ncv32f> &v, const char *name)
{
IplImage *flowField;
NCVMatrixAlloc<Ncv32f> host_u(*g_pHostMemAllocator, u.width(), u.height());
ncvAssertReturn(host_u.isMemAllocated(), NCV_ALLOCATOR_BAD_ALLOC);
NCVMatrixAlloc<Ncv32f> host_v (*g_pHostMemAllocator, u.width (), u.height ());
ncvAssertReturn (host_v.isMemAllocated (), NCV_ALLOCATOR_BAD_ALLOC);
ncvAssertReturnNcvStat (u.copySolid (host_u, 0));
ncvAssertReturnNcvStat (v.copySolid (host_v, 0));
float *ptr_u = host_u.ptr ();
float *ptr_v = host_v.ptr ();
float maxDisplacement = 1.0f;
for (Ncv32u i = 0; i < u.height (); ++i)
{
for (Ncv32u j = 0; j < u.width (); ++j)
{
float d = std::max ( fabsf(*ptr_u), fabsf(*ptr_v) );
if (d > maxDisplacement) maxDisplacement = d;
++ptr_u;
++ptr_v;
}
ptr_u += u.stride () - u.width ();
ptr_v += v.stride () - v.width ();
}
CvSize image_size = cvSize (u.width (), u.height ());
flowField = cvCreateImage (image_size, IPL_DEPTH_8U, 4);
if (flowField == 0) return NCV_NULL_PTR;
unsigned char *row = reinterpret_cast<unsigned char *> (flowField->imageData);
ptr_u = host_u.ptr();
ptr_v = host_v.ptr();
for (int i = 0; i < flowField->height; ++i)
{
for (int j = 0; j < flowField->width; ++j)
{
(row + j * flowField->nChannels)[0] = 0;
(row + j * flowField->nChannels)[1] = static_cast<unsigned char> (MapValue (-(*ptr_v), -maxDisplacement, maxDisplacement, 0.0f, 255.0f));
(row + j * flowField->nChannels)[2] = static_cast<unsigned char> (MapValue (*ptr_u , -maxDisplacement, maxDisplacement, 0.0f, 255.0f));
(row + j * flowField->nChannels)[3] = 255;
++ptr_u;
++ptr_v;
}
row += flowField->widthStep;
ptr_u += u.stride () - u.width ();
ptr_v += v.stride () - v.width ();
}
cvShowImage (name, flowField);
return NCV_SUCCESS;
}
static IplImage *CreateImage (NCVMatrixAlloc<Ncv32f> &h_r, NCVMatrixAlloc<Ncv32f> &h_g, NCVMatrixAlloc<Ncv32f> &h_b)
{
CvSize imageSize = cvSize (h_r.width (), h_r.height ());
IplImage *image = cvCreateImage (imageSize, IPL_DEPTH_8U, 4);
if (image == 0) return 0;
unsigned char *row = reinterpret_cast<unsigned char*> (image->imageData);
for (int i = 0; i < image->height; ++i)
{
for (int j = 0; j < image->width; ++j)
{
int offset = j * image->nChannels;
int pos = i * h_r.stride () + j;
row[offset + 0] = static_cast<unsigned char> (h_b.ptr ()[pos] * 255.0f);
row[offset + 1] = static_cast<unsigned char> (h_g.ptr ()[pos] * 255.0f);
row[offset + 2] = static_cast<unsigned char> (h_r.ptr ()[pos] * 255.0f);
row[offset + 3] = 255;
}
row += image->widthStep;
}
return image;
}
static void PrintHelp ()
{
std::cout << "Usage help:\n";
std::cout << std::setiosflags(std::ios::left);
std::cout << "\t" << std::setw(15) << PARAM_ALPHA << " - set alpha\n";
std::cout << "\t" << std::setw(15) << PARAM_GAMMA << " - set gamma\n";
std::cout << "\t" << std::setw(15) << PARAM_INNER << " - set number of inner iterations\n";
std::cout << "\t" << std::setw(15) << PARAM_LEFT << " - specify left image\n";
std::cout << "\t" << std::setw(15) << PARAM_RIGHT << " - specify right image\n";
std::cout << "\t" << std::setw(15) << PARAM_OUTER << " - set number of outer iterations\n";
std::cout << "\t" << std::setw(15) << PARAM_SCALE << " - set pyramid scale factor\n";
std::cout << "\t" << std::setw(15) << PARAM_SOLVER << " - set number of basic solver iterations\n";
std::cout << "\t" << std::setw(15) << PARAM_TIME_STEP << " - set frame interpolation time step\n";
std::cout << "\t" << std::setw(15) << PARAM_HELP << " - display this help message\n";
}
static int ProcessCommandLine(int argc, char **argv,
Ncv32f &timeStep,
char *&frame0Name,
char *&frame1Name,
NCVBroxOpticalFlowDescriptor &desc)
{
timeStep = 0.25f;
for (int iarg = 1; iarg < argc; ++iarg)
{
if (strcmp(argv[iarg], PARAM_LEFT) == 0)
{
if (iarg + 1 < argc)
{
frame0Name = argv[++iarg];
}
else
return -1;
}
if (strcmp(argv[iarg], PARAM_RIGHT) == 0)
{
if (iarg + 1 < argc)
{
frame1Name = argv[++iarg];
}
else
return -1;
}
else if(strcmp(argv[iarg], PARAM_SCALE) == 0)
{
if (iarg + 1 < argc)
desc.scale_factor = static_cast<Ncv32f>(atof(argv[++iarg]));
else
return -1;
}
else if(strcmp(argv[iarg], PARAM_ALPHA) == 0)
{
if (iarg + 1 < argc)
desc.alpha = static_cast<Ncv32f>(atof(argv[++iarg]));
else
return -1;
}
else if(strcmp(argv[iarg], PARAM_GAMMA) == 0)
{
if (iarg + 1 < argc)
desc.gamma = static_cast<Ncv32f>(atof(argv[++iarg]));
else
return -1;
}
else if(strcmp(argv[iarg], PARAM_INNER) == 0)
{
if (iarg + 1 < argc)
desc.number_of_inner_iterations = static_cast<Ncv32u>(atoi(argv[++iarg]));
else
return -1;
}
else if(strcmp(argv[iarg], PARAM_OUTER) == 0)
{
if (iarg + 1 < argc)
desc.number_of_outer_iterations = static_cast<Ncv32u>(atoi(argv[++iarg]));
else
return -1;
}
else if(strcmp(argv[iarg], PARAM_SOLVER) == 0)
{
if (iarg + 1 < argc)
desc.number_of_solver_iterations = static_cast<Ncv32u>(atoi(argv[++iarg]));
else
return -1;
}
else if(strcmp(argv[iarg], PARAM_TIME_STEP) == 0)
{
if (iarg + 1 < argc)
timeStep = static_cast<Ncv32f>(atof(argv[++iarg]));
else
return -1;
}
else if(strcmp(argv[iarg], PARAM_HELP) == 0)
{
PrintHelp ();
return 0;
}
}
return 0;
}
int main(int argc, char **argv)
{
char *frame0Name = 0, *frame1Name = 0;
Ncv32f timeStep = 0.01f;
NCVBroxOpticalFlowDescriptor desc;
desc.alpha = 0.197f;
desc.gamma = 50.0f;
desc.number_of_inner_iterations = 10;
desc.number_of_outer_iterations = 77;
desc.number_of_solver_iterations = 10;
desc.scale_factor = 0.8f;
int result = ProcessCommandLine (argc, argv, timeStep, frame0Name, frame1Name, desc);
if (argc == 1 || result)
{
PrintHelp();
return result;
}
cv::cuda::printShortCudaDeviceInfo(cv::cuda::getDevice());
std::cout << "OpenCV / NVIDIA Computer Vision\n";
std::cout << "Optical Flow Demo: Frame Interpolation\n";
std::cout << "=========================================\n";
std::cout << "Press:\n ESC to quit\n 'a' to move to the previous frame\n 's' to move to the next frame\n";
int devId;
ncvAssertCUDAReturn(cudaGetDevice(&devId), -1);
cudaDeviceProp devProp;
ncvAssertCUDAReturn(cudaGetDeviceProperties(&devProp, devId), -1);
std::cout << "Using GPU: " << devId << "(" << devProp.name <<
"), arch=" << devProp.major << "." << devProp.minor << std::endl;
g_pGPUMemAllocator = Ptr<INCVMemAllocator> (new NCVMemNativeAllocator (NCVMemoryTypeDevice, static_cast<Ncv32u>(devProp.textureAlignment)));
ncvAssertPrintReturn (g_pGPUMemAllocator->isInitialized (), "Device memory allocator isn't initialized", -1);
g_pHostMemAllocator = Ptr<INCVMemAllocator> (new NCVMemNativeAllocator (NCVMemoryTypeHostPageable, static_cast<Ncv32u>(devProp.textureAlignment)));
ncvAssertPrintReturn (g_pHostMemAllocator->isInitialized (), "Host memory allocator isn't initialized", -1);
int width, height;
Ptr<NCVMatrixAlloc<Ncv32f> > src_host;
Ptr<NCVMatrixAlloc<Ncv32f> > dst_host;
IplImage *firstFrame, *lastFrame;
if (frame0Name != 0 && frame1Name != 0)
{
ncvAssertReturnNcvStat (LoadImages (frame0Name, frame1Name, width, height, src_host, dst_host, firstFrame, lastFrame));
}
else
{
ncvAssertReturnNcvStat (LoadImages ("frame10.bmp", "frame11.bmp", width, height, src_host, dst_host, firstFrame, lastFrame));
}
Ptr<NCVMatrixAlloc<Ncv32f> > src (new NCVMatrixAlloc<Ncv32f> (*g_pGPUMemAllocator, src_host->width (), src_host->height ()));
ncvAssertReturn(src->isMemAllocated(), -1);
Ptr<NCVMatrixAlloc<Ncv32f> > dst (new NCVMatrixAlloc<Ncv32f> (*g_pGPUMemAllocator, src_host->width (), src_host->height ()));
ncvAssertReturn (dst->isMemAllocated (), -1);
ncvAssertReturnNcvStat (src_host->copySolid ( *src, 0 ));
ncvAssertReturnNcvStat (dst_host->copySolid ( *dst, 0 ));
#if defined SAFE_MAT_DECL
#undef SAFE_MAT_DECL
#endif
#define SAFE_MAT_DECL(name, allocator, sx, sy) \
NCVMatrixAlloc<Ncv32f> name(*allocator, sx, sy);\
ncvAssertReturn(name.isMemAllocated(), -1);
SAFE_MAT_DECL (u, g_pGPUMemAllocator, width, height);
SAFE_MAT_DECL (v, g_pGPUMemAllocator, width, height);
SAFE_MAT_DECL (uBck, g_pGPUMemAllocator, width, height);
SAFE_MAT_DECL (vBck, g_pGPUMemAllocator, width, height);
SAFE_MAT_DECL (h_r, g_pHostMemAllocator, width, height);
SAFE_MAT_DECL (h_g, g_pHostMemAllocator, width, height);
SAFE_MAT_DECL (h_b, g_pHostMemAllocator, width, height);
std::cout << "Estimating optical flow\nForward...\n";
if (NCV_SUCCESS != NCVBroxOpticalFlow (desc, *g_pGPUMemAllocator, *src, *dst, u, v, 0))
{
std::cout << "Failed\n";
return -1;
}
std::cout << "Backward...\n";
if (NCV_SUCCESS != NCVBroxOpticalFlow (desc, *g_pGPUMemAllocator, *dst, *src, uBck, vBck, 0))
{
std::cout << "Failed\n";
return -1;
}
// matrix for temporary data
SAFE_MAT_DECL (d_temp, g_pGPUMemAllocator, width, height);
// first frame color components (GPU memory)
SAFE_MAT_DECL (d_r, g_pGPUMemAllocator, width, height);
SAFE_MAT_DECL (d_g, g_pGPUMemAllocator, width, height);
SAFE_MAT_DECL (d_b, g_pGPUMemAllocator, width, height);
// second frame color components (GPU memory)
SAFE_MAT_DECL (d_rt, g_pGPUMemAllocator, width, height);
SAFE_MAT_DECL (d_gt, g_pGPUMemAllocator, width, height);
SAFE_MAT_DECL (d_bt, g_pGPUMemAllocator, width, height);
// intermediate frame color components (GPU memory)
SAFE_MAT_DECL (d_rNew, g_pGPUMemAllocator, width, height);
SAFE_MAT_DECL (d_gNew, g_pGPUMemAllocator, width, height);
SAFE_MAT_DECL (d_bNew, g_pGPUMemAllocator, width, height);
// interpolated forward flow
SAFE_MAT_DECL (ui, g_pGPUMemAllocator, width, height);
SAFE_MAT_DECL (vi, g_pGPUMemAllocator, width, height);
// interpolated backward flow
SAFE_MAT_DECL (ubi, g_pGPUMemAllocator, width, height);
SAFE_MAT_DECL (vbi, g_pGPUMemAllocator, width, height);
// occlusion masks
SAFE_MAT_DECL (occ0, g_pGPUMemAllocator, width, height);
SAFE_MAT_DECL (occ1, g_pGPUMemAllocator, width, height);
// prepare color components on host and copy them to device memory
ncvAssertReturnNcvStat (CopyData<RgbToR> (firstFrame, h_r));
ncvAssertReturnNcvStat (CopyData<RgbToG> (firstFrame, h_g));
ncvAssertReturnNcvStat (CopyData<RgbToB> (firstFrame, h_b));
ncvAssertReturnNcvStat (h_r.copySolid ( d_r, 0 ));
ncvAssertReturnNcvStat (h_g.copySolid ( d_g, 0 ));
ncvAssertReturnNcvStat (h_b.copySolid ( d_b, 0 ));
ncvAssertReturnNcvStat (CopyData<RgbToR> (lastFrame, h_r));
ncvAssertReturnNcvStat (CopyData<RgbToG> (lastFrame, h_g));
ncvAssertReturnNcvStat (CopyData<RgbToB> (lastFrame, h_b));
ncvAssertReturnNcvStat (h_r.copySolid ( d_rt, 0 ));
ncvAssertReturnNcvStat (h_g.copySolid ( d_gt, 0 ));
ncvAssertReturnNcvStat (h_b.copySolid ( d_bt, 0 ));
std::cout << "Interpolating...\n";
std::cout.precision (4);
std::vector<IplImage*> frames;
frames.push_back (firstFrame);
// compute interpolated frames
for (Ncv32f timePos = timeStep; timePos < 1.0f; timePos += timeStep)
{
ncvAssertCUDAReturn (cudaMemset (ui.ptr (), 0, ui.pitch () * ui.height ()), NCV_CUDA_ERROR);
ncvAssertCUDAReturn (cudaMemset (vi.ptr (), 0, vi.pitch () * vi.height ()), NCV_CUDA_ERROR);
ncvAssertCUDAReturn (cudaMemset (ubi.ptr (), 0, ubi.pitch () * ubi.height ()), NCV_CUDA_ERROR);
ncvAssertCUDAReturn (cudaMemset (vbi.ptr (), 0, vbi.pitch () * vbi.height ()), NCV_CUDA_ERROR);
ncvAssertCUDAReturn (cudaMemset (occ0.ptr (), 0, occ0.pitch () * occ0.height ()), NCV_CUDA_ERROR);
ncvAssertCUDAReturn (cudaMemset (occ1.ptr (), 0, occ1.pitch () * occ1.height ()), NCV_CUDA_ERROR);
NppStInterpolationState state;
// interpolation state should be filled once except pSrcFrame0, pSrcFrame1, and pNewFrame
// we will only need to reset buffers content to 0 since interpolator doesn't do this itself
state.size = NcvSize32u (width, height);
state.nStep = d_r.pitch ();
state.pSrcFrame0 = d_r.ptr ();
state.pSrcFrame1 = d_rt.ptr ();
state.pFU = u.ptr ();
state.pFV = v.ptr ();
state.pBU = uBck.ptr ();
state.pBV = vBck.ptr ();
state.pos = timePos;
state.pNewFrame = d_rNew.ptr ();
state.ppBuffers[0] = occ0.ptr ();
state.ppBuffers[1] = occ1.ptr ();
state.ppBuffers[2] = ui.ptr ();
state.ppBuffers[3] = vi.ptr ();
state.ppBuffers[4] = ubi.ptr ();
state.ppBuffers[5] = vbi.ptr ();
// interpolate red channel
nppiStInterpolateFrames (&state);
// reset buffers
ncvAssertCUDAReturn (cudaMemset (ui.ptr (), 0, ui.pitch () * ui.height ()), NCV_CUDA_ERROR);
ncvAssertCUDAReturn (cudaMemset (vi.ptr (), 0, vi.pitch () * vi.height ()), NCV_CUDA_ERROR);
ncvAssertCUDAReturn (cudaMemset (ubi.ptr (), 0, ubi.pitch () * ubi.height ()), NCV_CUDA_ERROR);
ncvAssertCUDAReturn (cudaMemset (vbi.ptr (), 0, vbi.pitch () * vbi.height ()), NCV_CUDA_ERROR);
ncvAssertCUDAReturn (cudaMemset (occ0.ptr (), 0, occ0.pitch () * occ0.height ()), NCV_CUDA_ERROR);
ncvAssertCUDAReturn (cudaMemset (occ1.ptr (), 0, occ1.pitch () * occ1.height ()), NCV_CUDA_ERROR);
// interpolate green channel
state.pSrcFrame0 = d_g.ptr ();
state.pSrcFrame1 = d_gt.ptr ();
state.pNewFrame = d_gNew.ptr ();
nppiStInterpolateFrames (&state);
// reset buffers
ncvAssertCUDAReturn (cudaMemset (ui.ptr (), 0, ui.pitch () * ui.height ()), NCV_CUDA_ERROR);
ncvAssertCUDAReturn (cudaMemset (vi.ptr (), 0, vi.pitch () * vi.height ()), NCV_CUDA_ERROR);
ncvAssertCUDAReturn (cudaMemset (ubi.ptr (), 0, ubi.pitch () * ubi.height ()), NCV_CUDA_ERROR);
ncvAssertCUDAReturn (cudaMemset (vbi.ptr (), 0, vbi.pitch () * vbi.height ()), NCV_CUDA_ERROR);
ncvAssertCUDAReturn (cudaMemset (occ0.ptr (), 0, occ0.pitch () * occ0.height ()), NCV_CUDA_ERROR);
ncvAssertCUDAReturn (cudaMemset (occ1.ptr (), 0, occ1.pitch () * occ1.height ()), NCV_CUDA_ERROR);
// interpolate blue channel
state.pSrcFrame0 = d_b.ptr ();
state.pSrcFrame1 = d_bt.ptr ();
state.pNewFrame = d_bNew.ptr ();
nppiStInterpolateFrames (&state);
// copy to host memory
ncvAssertReturnNcvStat (d_rNew.copySolid (h_r, 0));
ncvAssertReturnNcvStat (d_gNew.copySolid (h_g, 0));
ncvAssertReturnNcvStat (d_bNew.copySolid (h_b, 0));
// convert to IplImage
IplImage *newFrame = CreateImage (h_r, h_g, h_b);
if (newFrame == 0)
{
std::cout << "Could not create new frame in host memory\n";
break;
}
frames.push_back (newFrame);
std::cout << timePos * 100.0f << "%\r";
}
std::cout << std::setw (5) << "100%\n";
frames.push_back (lastFrame);
Ncv32u currentFrame;
currentFrame = 0;
ShowFlow (u, v, "Forward flow");
ShowFlow (uBck, vBck, "Backward flow");
cvShowImage ("Interpolated frame", frames[currentFrame]);
bool qPressed = false;
while ( !qPressed )
{
int key = toupper (cvWaitKey (10));
switch (key)
{
case 27:
qPressed = true;
break;
case 'A':
if (currentFrame > 0) --currentFrame;
cvShowImage ("Interpolated frame", frames[currentFrame]);
break;
case 'S':
if (currentFrame < frames.size()-1) ++currentFrame;
cvShowImage ("Interpolated frame", frames[currentFrame]);
break;
}
}
cvDestroyAllWindows ();
std::vector<IplImage*>::iterator iter;
for (iter = frames.begin (); iter != frames.end (); ++iter)
{
cvReleaseImage (&(*iter));
}
return 0;
}
#endif