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

Merge pull request #3635 from jet47:cuda-optflow-refactoring

Browse Source
This commit is contained in:
Vadim Pisarevsky
2015-01-22 09:45:19 +00:00
parent 9c81338cb9 710617034b
commit 3f1fb281be
27 changed files with 1577 additions and 1858 deletions
Download Patch File Download Diff File Expand all files Collapse all files

270
samples/gpu/brox_optical_flow.cpp
View File

@@ -1,270 +0,0 @@
#include <iostream>
#include <iomanip>
#include <string>
#include <ctype.h>
#include "opencv2/core.hpp"
#include "opencv2/core/utility.hpp"
#include "opencv2/highgui.hpp"
#include "opencv2/imgproc.hpp"
#include "opencv2/cudaoptflow.hpp"
#include "opencv2/cudaarithm.hpp"
using namespace std;
using namespace cv;
using namespace cv::cuda;
void getFlowField(const Mat& u, const Mat& v, Mat& flowField);
int main(int argc, const char* argv[])
{
try
{
const char* keys =
"{ h help | | print help message }"
"{ l left | | specify left image }"
"{ r right | | specify right image }"
"{ s scale | 0.8 | set pyramid scale factor }"
"{ a alpha | 0.197 | set alpha }"
"{ g gamma | 50.0 | set gamma }"
"{ i inner | 10 | set number of inner iterations }"
"{ o outer | 77 | set number of outer iterations }"
"{ si solver | 10 | set number of basic solver iterations }"
"{ t time_step | 0.1 | set frame interpolation time step }";
CommandLineParser cmd(argc, argv, keys);
if (cmd.has("help") || !cmd.check())
{
cmd.printMessage();
cmd.printErrors();
return 0;
}
string frame0Name = cmd.get<string>("left");
string frame1Name = cmd.get<string>("right");
float scale = cmd.get<float>("scale");
float alpha = cmd.get<float>("alpha");
float gamma = cmd.get<float>("gamma");
int inner_iterations = cmd.get<int>("inner");
int outer_iterations = cmd.get<int>("outer");
int solver_iterations = cmd.get<int>("solver");
float timeStep = cmd.get<float>("time_step");
if (frame0Name.empty() || frame1Name.empty())
{
cerr << "Missing input file names" << endl;
return -1;
}
Mat frame0Color = imread(frame0Name);
Mat frame1Color = imread(frame1Name);
if (frame0Color.empty() || frame1Color.empty())
{
cout << "Can't load input images" << endl;
return -1;
}
cv::cuda::printShortCudaDeviceInfo(cv::cuda::getDevice());
cout << "OpenCV / NVIDIA Computer Vision" << endl;
cout << "Optical Flow Demo: Frame Interpolation" << endl;
cout << "=========================================" << endl;
namedWindow("Forward flow");
namedWindow("Backward flow");
namedWindow("Interpolated frame");
cout << "Press:" << endl;
cout << "\tESC to quit" << endl;
cout << "\t'a' to move to the previous frame" << endl;
cout << "\t's' to move to the next frame\n" << endl;
frame0Color.convertTo(frame0Color, CV_32F, 1.0 / 255.0);
frame1Color.convertTo(frame1Color, CV_32F, 1.0 / 255.0);
Mat frame0Gray, frame1Gray;
cv::cvtColor(frame0Color, frame0Gray, COLOR_BGR2GRAY);
cv::cvtColor(frame1Color, frame1Gray, COLOR_BGR2GRAY);
GpuMat d_frame0(frame0Gray);
GpuMat d_frame1(frame1Gray);
cout << "Estimating optical flow" << endl;
BroxOpticalFlow d_flow(alpha, gamma, scale, inner_iterations, outer_iterations, solver_iterations);
cout << "\tForward..." << endl;
GpuMat d_fu, d_fv;
d_flow(d_frame0, d_frame1, d_fu, d_fv);
Mat flowFieldForward;
getFlowField(Mat(d_fu), Mat(d_fv), flowFieldForward);
cout << "\tBackward..." << endl;
GpuMat d_bu, d_bv;
d_flow(d_frame1, d_frame0, d_bu, d_bv);
Mat flowFieldBackward;
getFlowField(Mat(d_bu), Mat(d_bv), flowFieldBackward);
cout << "Interpolating..." << endl;
// first frame color components
GpuMat d_b, d_g, d_r;
// second frame color components
GpuMat d_bt, d_gt, d_rt;
// prepare color components on host and copy them to device memory
Mat channels[3];
cv::split(frame0Color, channels);
d_b.upload(channels[0]);
d_g.upload(channels[1]);
d_r.upload(channels[2]);
cv::split(frame1Color, channels);
d_bt.upload(channels[0]);
d_gt.upload(channels[1]);
d_rt.upload(channels[2]);
// temporary buffer
GpuMat d_buf;
// intermediate frame color components (GPU memory)
GpuMat d_rNew, d_gNew, d_bNew;
GpuMat d_newFrame;
vector<Mat> frames;
frames.reserve(static_cast<int>(1.0f / timeStep) + 2);
frames.push_back(frame0Color);
// compute interpolated frames
for (float timePos = timeStep; timePos < 1.0f; timePos += timeStep)
{
// interpolate blue channel
interpolateFrames(d_b, d_bt, d_fu, d_fv, d_bu, d_bv, timePos, d_bNew, d_buf);
// interpolate green channel
interpolateFrames(d_g, d_gt, d_fu, d_fv, d_bu, d_bv, timePos, d_gNew, d_buf);
// interpolate red channel
interpolateFrames(d_r, d_rt, d_fu, d_fv, d_bu, d_bv, timePos, d_rNew, d_buf);
GpuMat channels3[] = {d_bNew, d_gNew, d_rNew};
cuda::merge(channels3, 3, d_newFrame);
frames.push_back(Mat(d_newFrame));
cout << setprecision(4) << timePos * 100.0f << "%\r";
}
frames.push_back(frame1Color);
cout << setw(5) << "100%" << endl;
cout << "Done" << endl;
imshow("Forward flow", flowFieldForward);
imshow("Backward flow", flowFieldBackward);
int currentFrame = 0;
imshow("Interpolated frame", frames[currentFrame]);
for(;;)
{
int key = toupper(waitKey(10) & 0xff);
switch (key)
{
case 27:
return 0;
case 'A':
if (currentFrame > 0)
--currentFrame;
imshow("Interpolated frame", frames[currentFrame]);
break;
case 'S':
if (currentFrame < static_cast<int>(frames.size()) - 1)
++currentFrame;
imshow("Interpolated frame", frames[currentFrame]);
break;
}
}
}
catch (const exception& ex)
{
cerr << ex.what() << endl;
return -1;
}
catch (...)
{
cerr << "Unknow error" << endl;
return -1;
}
}
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);
}
void getFlowField(const Mat& u, const Mat& v, Mat& flowField)
{
float maxDisplacement = 1.0f;
for (int i = 0; i < u.rows; ++i)
{
const float* ptr_u = u.ptr<float>(i);
const float* ptr_v = v.ptr<float>(i);
for (int j = 0; j < u.cols; ++j)
{
float d = max(fabsf(ptr_u[j]), fabsf(ptr_v[j]));
if (d > maxDisplacement)
maxDisplacement = d;
}
}
flowField.create(u.size(), CV_8UC4);
for (int i = 0; i < flowField.rows; ++i)
{
const float* ptr_u = u.ptr<float>(i);
const float* ptr_v = v.ptr<float>(i);
Vec4b* row = flowField.ptr<Vec4b>(i);
for (int j = 0; j < flowField.cols; ++j)
{
row[j][0] = 0;
row[j][1] = static_cast<unsigned char> (mapValue (-ptr_v[j], -maxDisplacement, maxDisplacement, 0.0f, 255.0f));
row[j][2] = static_cast<unsigned char> (mapValue ( ptr_u[j], -maxDisplacement, maxDisplacement, 0.0f, 255.0f));
row[j][3] = 255;
}
}
}

19
samples/gpu/farneback_optical_flow.cpp
View File

@@ -7,6 +7,7 @@
#include "opencv2/highgui.hpp"
#include "opencv2/video.hpp"
#include "opencv2/cudaoptflow.hpp"
#include "opencv2/cudaarithm.hpp"
using namespace std;
using namespace cv;
@@ -70,8 +71,8 @@ int main(int argc, char **argv)
if (frameL.empty() || frameR.empty()) return -1;
GpuMat d_frameL(frameL), d_frameR(frameR);
GpuMat d_flowx, d_flowy;
FarnebackOpticalFlow d_calc;
GpuMat d_flow;
Ptr<cuda::FarnebackOpticalFlow> d_calc = cuda::FarnebackOpticalFlow::create();
Mat flowxy, flowx, flowy, image;
bool running = true, gpuMode = true;
@@ -86,17 +87,21 @@ int main(int argc, char **argv)
if (gpuMode)
{
tc0 = getTickCount();
d_calc(d_frameL, d_frameR, d_flowx, d_flowy);
d_calc->calc(d_frameL, d_frameR, d_flow);
tc1 = getTickCount();
d_flowx.download(flowx);
d_flowy.download(flowy);
GpuMat planes[2];
cuda::split(d_flow, planes);
planes[0].download(flowx);
planes[1].download(flowy);
}
else
{
tc0 = getTickCount();
calcOpticalFlowFarneback(
frameL, frameR, flowxy, d_calc.pyrScale, d_calc.numLevels, d_calc.winSize,
d_calc.numIters, d_calc.polyN, d_calc.polySigma, d_calc.flags);
frameL, frameR, flowxy, d_calc->getPyrScale(), d_calc->getNumLevels(), d_calc->getWinSize(),
d_calc->getNumIters(), d_calc->getPolyN(), d_calc->getPolySigma(), d_calc->getFlags());
tc1 = getTickCount();
Mat planes[] = {flowx, flowy};

61
samples/gpu/optical_flow.cpp
View File

@@ -5,6 +5,7 @@
#include <opencv2/core/utility.hpp>
#include "opencv2/highgui.hpp"
#include "opencv2/cudaoptflow.hpp"
#include "opencv2/cudaarithm.hpp"
using namespace std;
using namespace cv;
@@ -122,10 +123,13 @@ static void drawOpticalFlow(const Mat_<float>& flowx, const Mat_<float>& flowy,
}
}
static void showFlow(const char* name, const GpuMat& d_flowx, const GpuMat& d_flowy)
static void showFlow(const char* name, const GpuMat& d_flow)
{
Mat flowx(d_flowx);
Mat flowy(d_flowy);
GpuMat planes[2];
cuda::split(d_flow, planes);
Mat flowx(planes[0]);
Mat flowy(planes[1]);
Mat out;
drawOpticalFlow(flowx, flowy, out, 10);
@@ -171,14 +175,12 @@ int main(int argc, const char* argv[])
GpuMat d_frame0(frame0);
GpuMat d_frame1(frame1);
GpuMat d_flowx(frame0.size(), CV_32FC1);
GpuMat d_flowy(frame0.size(), CV_32FC1);
GpuMat d_flow(frame0.size(), CV_32FC2);
BroxOpticalFlow brox(0.197f, 50.0f, 0.8f, 10, 77, 10);
PyrLKOpticalFlow lk; lk.winSize = Size(7, 7);
FarnebackOpticalFlow farn;
OpticalFlowDual_TVL1_CUDA tvl1;
FastOpticalFlowBM fastBM;
Ptr<cuda::BroxOpticalFlow> brox = cuda::BroxOpticalFlow::create(0.197f, 50.0f, 0.8f, 10, 77, 10);
Ptr<cuda::DensePyrLKOpticalFlow> lk = cuda::DensePyrLKOpticalFlow::create(Size(7, 7));
Ptr<cuda::FarnebackOpticalFlow> farn = cuda::FarnebackOpticalFlow::create();
Ptr<cuda::OpticalFlowDual_TVL1> tvl1 = cuda::OpticalFlowDual_TVL1::create();
{
GpuMat d_frame0f;
@@ -189,68 +191,45 @@ int main(int argc, const char* argv[])
const int64 start = getTickCount();
brox(d_frame0f, d_frame1f, d_flowx, d_flowy);
brox->calc(d_frame0f, d_frame1f, d_flow);
const double timeSec = (getTickCount() - start) / getTickFrequency();
cout << "Brox : " << timeSec << " sec" << endl;
showFlow("Brox", d_flowx, d_flowy);
showFlow("Brox", d_flow);
}
{
const int64 start = getTickCount();
lk.dense(d_frame0, d_frame1, d_flowx, d_flowy);
lk->calc(d_frame0, d_frame1, d_flow);
const double timeSec = (getTickCount() - start) / getTickFrequency();
cout << "LK : " << timeSec << " sec" << endl;
showFlow("LK", d_flowx, d_flowy);
showFlow("LK", d_flow);
}
{
const int64 start = getTickCount();
farn(d_frame0, d_frame1, d_flowx, d_flowy);
farn->calc(d_frame0, d_frame1, d_flow);
const double timeSec = (getTickCount() - start) / getTickFrequency();
cout << "Farn : " << timeSec << " sec" << endl;
showFlow("Farn", d_flowx, d_flowy);
showFlow("Farn", d_flow);
}
{
const int64 start = getTickCount();
tvl1(d_frame0, d_frame1, d_flowx, d_flowy);
tvl1->calc(d_frame0, d_frame1, d_flow);
const double timeSec = (getTickCount() - start) / getTickFrequency();
cout << "TVL1 : " << timeSec << " sec" << endl;
showFlow("TVL1", d_flowx, d_flowy);
}
{
const int64 start = getTickCount();
GpuMat buf;
calcOpticalFlowBM(d_frame0, d_frame1, Size(7, 7), Size(1, 1), Size(21, 21), false, d_flowx, d_flowy, buf);
const double timeSec = (getTickCount() - start) / getTickFrequency();
cout << "BM : " << timeSec << " sec" << endl;
showFlow("BM", d_flowx, d_flowy);
}
{
const int64 start = getTickCount();
fastBM(d_frame0, d_frame1, d_flowx, d_flowy);
const double timeSec = (getTickCount() - start) / getTickFrequency();
cout << "Fast BM : " << timeSec << " sec" << endl;
showFlow("Fast BM", d_flowx, d_flowy);
showFlow("TVL1", d_flow);
}
imshow("Frame 0", frame0);

81
samples/gpu/performance/tests.cpp
View File

@@ -1187,87 +1187,6 @@ TEST(GoodFeaturesToTrack)
CUDA_OFF;
}
TEST(PyrLKOpticalFlow)
{
Mat frame0 = imread(abspath("../data/rubberwhale1.png"));
if (frame0.empty()) throw runtime_error("can't open ../data/rubberwhale1.png");
Mat frame1 = imread(abspath("../data/rubberwhale2.png"));
if (frame1.empty()) throw runtime_error("can't open ../data/rubberwhale2.png");
Mat gray_frame;
cvtColor(frame0, gray_frame, COLOR_BGR2GRAY);
for (int points = 1000; points <= 8000; points *= 2)
{
SUBTEST << points;
vector<Point2f> pts;
goodFeaturesToTrack(gray_frame, pts, points, 0.01, 0.0);
vector<Point2f> nextPts;
vector<unsigned char> status;
vector<float> err;
calcOpticalFlowPyrLK(frame0, frame1, pts, nextPts, status, err);
CPU_ON;
calcOpticalFlowPyrLK(frame0, frame1, pts, nextPts, status, err);
CPU_OFF;
cuda::PyrLKOpticalFlow d_pyrLK;
cuda::GpuMat d_frame0(frame0);
cuda::GpuMat d_frame1(frame1);
cuda::GpuMat d_pts;
Mat pts_mat(1, (int)pts.size(), CV_32FC2, (void*)&pts[0]);
d_pts.upload(pts_mat);
cuda::GpuMat d_nextPts;
cuda::GpuMat d_status;
cuda::GpuMat d_err;
d_pyrLK.sparse(d_frame0, d_frame1, d_pts, d_nextPts, d_status, &d_err);
CUDA_ON;
d_pyrLK.sparse(d_frame0, d_frame1, d_pts, d_nextPts, d_status, &d_err);
CUDA_OFF;
}
}
TEST(FarnebackOpticalFlow)
{
const string datasets[] = {"../data/rubberwhale", "../data/basketball"};
for (size_t i = 0; i < sizeof(datasets)/sizeof(*datasets); ++i) {
for (int fastPyramids = 0; fastPyramids < 2; ++fastPyramids) {
for (int useGaussianBlur = 0; useGaussianBlur < 2; ++useGaussianBlur) {
SUBTEST << "dataset=" << datasets[i] << ", fastPyramids=" << fastPyramids << ", useGaussianBlur=" << useGaussianBlur;
Mat frame0 = imread(abspath(datasets[i] + "1.png"), IMREAD_GRAYSCALE);
Mat frame1 = imread(abspath(datasets[i] + "2.png"), IMREAD_GRAYSCALE);
if (frame0.empty()) throw runtime_error("can't open " + datasets[i] + "1.png");
if (frame1.empty()) throw runtime_error("can't open " + datasets[i] + "2.png");
cuda::FarnebackOpticalFlow calc;
calc.fastPyramids = fastPyramids != 0;
calc.flags |= useGaussianBlur ? OPTFLOW_FARNEBACK_GAUSSIAN : 0;
cuda::GpuMat d_frame0(frame0), d_frame1(frame1), d_flowx, d_flowy;
CUDA_ON;
calc(d_frame0, d_frame1, d_flowx, d_flowy);
CUDA_OFF;
Mat flow;
CPU_ON;
calcOpticalFlowFarneback(frame0, frame1, flow, calc.pyrScale, calc.numLevels, calc.winSize, calc.numIters, calc.polyN, calc.polySigma, calc.flags);
CPU_OFF;
}}}
}
#ifdef HAVE_OPENCV_BGSEGM
TEST(MOG)

62
samples/gpu/pyrlk_optical_flow.cpp
View File

@@ -77,44 +77,6 @@ template <typename T> inline T mapValue(T x, T a, T b, T c, T d)
return c + (d - c) * (x - a) / (b - a);
}
static void getFlowField(const Mat& u, const Mat& v, Mat& flowField)
{
float maxDisplacement = 1.0f;
for (int i = 0; i < u.rows; ++i)
{
const float* ptr_u = u.ptr<float>(i);
const float* ptr_v = v.ptr<float>(i);
for (int j = 0; j < u.cols; ++j)
{
float d = max(fabsf(ptr_u[j]), fabsf(ptr_v[j]));
if (d > maxDisplacement)
maxDisplacement = d;
}
}
flowField.create(u.size(), CV_8UC4);
for (int i = 0; i < flowField.rows; ++i)
{
const float* ptr_u = u.ptr<float>(i);
const float* ptr_v = v.ptr<float>(i);
Vec4b* row = flowField.ptr<Vec4b>(i);
for (int j = 0; j < flowField.cols; ++j)
{
row[j][0] = 0;
row[j][1] = static_cast<unsigned char> (mapValue (-ptr_v[j], -maxDisplacement, maxDisplacement, 0.0f, 255.0f));
row[j][2] = static_cast<unsigned char> (mapValue ( ptr_u[j], -maxDisplacement, maxDisplacement, 0.0f, 255.0f));
row[j][3] = 255;
}
}
}
int main(int argc, const char* argv[])
{
const char* keys =
@@ -186,12 +148,8 @@ int main(int argc, const char* argv[])
// Sparse
PyrLKOpticalFlow d_pyrLK;
d_pyrLK.winSize.width = winSize;
d_pyrLK.winSize.height = winSize;
d_pyrLK.maxLevel = maxLevel;
d_pyrLK.iters = iters;
Ptr<cuda::SparsePyrLKOpticalFlow> d_pyrLK = cuda::SparsePyrLKOpticalFlow::create(
Size(winSize, winSize), maxLevel, iters);
GpuMat d_frame0(frame0);
GpuMat d_frame1(frame1);
@@ -199,7 +157,7 @@ int main(int argc, const char* argv[])
GpuMat d_nextPts;
GpuMat d_status;
d_pyrLK.sparse(useGray ? d_frame0Gray : d_frame0, useGray ? d_frame1Gray : d_frame1, d_prevPts, d_nextPts, d_status);
d_pyrLK->calc(useGray ? d_frame0Gray : d_frame0, useGray ? d_frame1Gray : d_frame1, d_prevPts, d_nextPts, d_status);
// Draw arrows
@@ -216,20 +174,6 @@ int main(int argc, const char* argv[])
imshow("PyrLK [Sparse]", frame0);
// Dense
GpuMat d_u;
GpuMat d_v;
d_pyrLK.dense(d_frame0Gray, d_frame1Gray, d_u, d_v);
// Draw flow field
Mat flowField;
getFlowField(Mat(d_u), Mat(d_v), flowField);
imshow("PyrLK [Dense] Flow Field", flowField);
waitKey();
return 0;