video/src/tvl1flow.cpp

/*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) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage 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*/

/*
//
// This implementation is based on Javier Sánchez Pérez <jsanchez@dis.ulpgc.es> implementation.
// Original BSD license:
//
// Copyright (c) 2011, Javier Sánchez Pérez, Enric Meinhardt Llopis
// All rights reserved.
//
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are met:
//
// * Redistributions of source code must retain the above copyright notice, this
//   list of conditions and the following disclaimer.
//
// * Redistributions 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.
//
// 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 COPYRIGHT HOLDER 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.
//
*/

#include "precomp.hpp"
#include "opencl_kernels_video.hpp"

#include <limits>
#include <iomanip>
#include <iostream>
#include "opencv2/core/opencl/ocl_defs.hpp"


using namespace cv;

namespace {

class OpticalFlowDual_TVL1 : public DualTVL1OpticalFlow
{
public:
    OpticalFlowDual_TVL1();

    void calc(InputArray I0, InputArray I1, InputOutputArray flow);
    void collectGarbage();

    CV_IMPL_PROPERTY(double, Tau, tau)
    CV_IMPL_PROPERTY(double, Lambda, lambda)
    CV_IMPL_PROPERTY(double, Theta, theta)
    CV_IMPL_PROPERTY(double, Gamma, gamma)
    CV_IMPL_PROPERTY(int, ScalesNumber, nscales)
    CV_IMPL_PROPERTY(int, WarpingsNumber, warps)
    CV_IMPL_PROPERTY(double, Epsilon, epsilon)
    CV_IMPL_PROPERTY(int, InnerIterations, innerIterations)
    CV_IMPL_PROPERTY(int, OuterIterations, outerIterations)
    CV_IMPL_PROPERTY(bool, UseInitialFlow, useInitialFlow)
    CV_IMPL_PROPERTY(double, ScaleStep, scaleStep)
    CV_IMPL_PROPERTY(int, MedianFiltering, medianFiltering)

protected:
    double tau;
    double lambda;
    double theta;
    double gamma;
    int nscales;
    int warps;
    double epsilon;
    int innerIterations;
    int outerIterations;
    bool useInitialFlow;
    double scaleStep;
    int medianFiltering;

private:
   void procOneScale(const Mat_<float>& I0, const Mat_<float>& I1, Mat_<float>& u1, Mat_<float>& u2, Mat_<float>& u3);

    bool procOneScale_ocl(const UMat& I0, const UMat& I1, UMat& u1, UMat& u2);

    bool calc_ocl(InputArray I0, InputArray I1, InputOutputArray flow);
    struct dataMat
    {
        std::vector<Mat_<float> > I0s;
        std::vector<Mat_<float> > I1s;
        std::vector<Mat_<float> > u1s;
        std::vector<Mat_<float> > u2s;
        std::vector<Mat_<float> > u3s;

        Mat_<float> I1x_buf;
        Mat_<float> I1y_buf;

        Mat_<float> flowMap1_buf;
        Mat_<float> flowMap2_buf;

        Mat_<float> I1w_buf;
        Mat_<float> I1wx_buf;
        Mat_<float> I1wy_buf;

        Mat_<float> grad_buf;
        Mat_<float> rho_c_buf;

        Mat_<float> v1_buf;
        Mat_<float> v2_buf;
        Mat_<float> v3_buf;

        Mat_<float> p11_buf;
        Mat_<float> p12_buf;
        Mat_<float> p21_buf;
        Mat_<float> p22_buf;
        Mat_<float> p31_buf;
        Mat_<float> p32_buf;

        Mat_<float> div_p1_buf;
        Mat_<float> div_p2_buf;
        Mat_<float> div_p3_buf;

        Mat_<float> u1x_buf;
        Mat_<float> u1y_buf;
        Mat_<float> u2x_buf;
        Mat_<float> u2y_buf;
        Mat_<float> u3x_buf;
        Mat_<float> u3y_buf;
    } dm;
    struct dataUMat
    {
        std::vector<UMat> I0s;
        std::vector<UMat> I1s;
        std::vector<UMat> u1s;
        std::vector<UMat> u2s;

        UMat I1x_buf;
        UMat I1y_buf;

        UMat I1w_buf;
        UMat I1wx_buf;
        UMat I1wy_buf;

        UMat grad_buf;
        UMat rho_c_buf;

        UMat p11_buf;
        UMat p12_buf;
        UMat p21_buf;
        UMat p22_buf;

        UMat diff_buf;
        UMat norm_buf;
    } dum;
};

namespace cv_ocl_tvl1flow
{
    bool centeredGradient(const UMat &src, UMat &dx, UMat &dy);

    bool warpBackward(const UMat &I0, const UMat &I1, UMat &I1x, UMat &I1y,
        UMat &u1, UMat &u2, UMat &I1w, UMat &I1wx, UMat &I1wy,
        UMat &grad, UMat &rho);

    bool estimateU(UMat &I1wx, UMat &I1wy, UMat &grad,
        UMat &rho_c, UMat &p11, UMat &p12,
        UMat &p21, UMat &p22, UMat &u1,
        UMat &u2, UMat &error, float l_t, float theta, char calc_error);

    bool estimateDualVariables(UMat &u1, UMat &u2,
        UMat &p11, UMat &p12, UMat &p21, UMat &p22, float taut);
}

bool cv_ocl_tvl1flow::centeredGradient(const UMat &src, UMat &dx, UMat &dy)
{
    size_t globalsize[2] = { src.cols, src.rows };

    ocl::Kernel kernel;
    if (!kernel.create("centeredGradientKernel", cv::ocl::video::optical_flow_tvl1_oclsrc, ""))
        return false;

    int idxArg = 0;
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(src));//src mat
    idxArg = kernel.set(idxArg, (int)(src.cols));//src mat col
    idxArg = kernel.set(idxArg, (int)(src.rows));//src mat rows
    idxArg = kernel.set(idxArg, (int)(src.step / src.elemSize()));//src mat step
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrWriteOnly(dx));//res mat dx
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrWriteOnly(dy));//res mat dy
    idxArg = kernel.set(idxArg, (int)(dx.step/dx.elemSize()));//res mat step
    return kernel.run(2, globalsize, NULL, false);
}

bool cv_ocl_tvl1flow::warpBackward(const UMat &I0, const UMat &I1, UMat &I1x, UMat &I1y,
    UMat &u1, UMat &u2, UMat &I1w, UMat &I1wx, UMat &I1wy,
    UMat &grad, UMat &rho)
{
    size_t globalsize[2] = { I0.cols, I0.rows };

    ocl::Kernel kernel;
    if (!kernel.create("warpBackwardKernel", cv::ocl::video::optical_flow_tvl1_oclsrc, ""))
        return false;

    int idxArg = 0;
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(I0));//I0 mat
    int I0_step = (int)(I0.step / I0.elemSize());
    idxArg = kernel.set(idxArg, I0_step);//I0_step
    idxArg = kernel.set(idxArg, (int)(I0.cols));//I0_col
    idxArg = kernel.set(idxArg, (int)(I0.rows));//I0_row
    ocl::Image2D imageI1(I1);
    ocl::Image2D imageI1x(I1x);
    ocl::Image2D imageI1y(I1y);
    idxArg = kernel.set(idxArg, imageI1);//image2d_t tex_I1
    idxArg = kernel.set(idxArg, imageI1x);//image2d_t tex_I1x
    idxArg = kernel.set(idxArg, imageI1y);//image2d_t tex_I1y
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(u1));//const float* u1
    idxArg = kernel.set(idxArg, (int)(u1.step / u1.elemSize()));//int u1_step
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(u2));//const float* u2
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrWriteOnly(I1w));///float* I1w
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrWriteOnly(I1wx));//float* I1wx
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrWriteOnly(I1wy));//float* I1wy
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrWriteOnly(grad));//float* grad
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrWriteOnly(rho));//float* rho
    idxArg = kernel.set(idxArg, (int)(I1w.step / I1w.elemSize()));//I1w_step
    idxArg = kernel.set(idxArg, (int)(u2.step / u2.elemSize()));//u2_step
    int u1_offset_x = (int)((u1.offset) % (u1.step));
    u1_offset_x = (int)(u1_offset_x / u1.elemSize());
    idxArg = kernel.set(idxArg, (int)u1_offset_x );//u1_offset_x
    idxArg = kernel.set(idxArg, (int)(u1.offset/u1.step));//u1_offset_y
    int u2_offset_x = (int)((u2.offset) % (u2.step));
    u2_offset_x = (int) (u2_offset_x / u2.elemSize());
    idxArg = kernel.set(idxArg, (int)u2_offset_x);//u2_offset_x
    idxArg = kernel.set(idxArg, (int)(u2.offset / u2.step));//u2_offset_y
    return kernel.run(2, globalsize, NULL, false);
}

bool cv_ocl_tvl1flow::estimateU(UMat &I1wx, UMat &I1wy, UMat &grad,
    UMat &rho_c, UMat &p11, UMat &p12,
    UMat &p21, UMat &p22, UMat &u1,
    UMat &u2, UMat &error, float l_t, float theta, char calc_error)
{
    size_t globalsize[2] = { I1wx.cols, I1wx.rows };

    ocl::Kernel kernel;
    if (!kernel.create("estimateUKernel", cv::ocl::video::optical_flow_tvl1_oclsrc, ""))
        return false;

    int idxArg = 0;
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(I1wx)); //const float* I1wx
    idxArg = kernel.set(idxArg, (int)(I1wx.cols)); //int I1wx_col
    idxArg = kernel.set(idxArg, (int)(I1wx.rows)); //int I1wx_row
    idxArg = kernel.set(idxArg, (int)(I1wx.step/I1wx.elemSize())); //int I1wx_step
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(I1wy)); //const float* I1wy
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(grad)); //const float* grad
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(rho_c)); //const float* rho_c
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(p11)); //const float* p11
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(p12)); //const float* p12
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(p21)); //const float* p21
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(p22)); //const float* p22
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadWrite(u1)); //float* u1
    idxArg = kernel.set(idxArg, (int)(u1.step / u1.elemSize())); //int u1_step
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadWrite(u2)); //float* u2
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrWriteOnly(error)); //float* error
    idxArg = kernel.set(idxArg, (float)l_t); //float l_t
    idxArg = kernel.set(idxArg, (float)theta); //float theta
    idxArg = kernel.set(idxArg, (int)(u2.step / u2.elemSize()));//int u2_step
    int u1_offset_x = (int)(u1.offset % u1.step);
    u1_offset_x = (int) (u1_offset_x  / u1.elemSize());
    idxArg = kernel.set(idxArg, (int)u1_offset_x); //int u1_offset_x
    idxArg = kernel.set(idxArg, (int)(u1.offset/u1.step)); //int u1_offset_y
    int u2_offset_x = (int)(u2.offset % u2.step);
    u2_offset_x = (int)(u2_offset_x / u2.elemSize());
    idxArg = kernel.set(idxArg, (int)u2_offset_x ); //int u2_offset_x
    idxArg = kernel.set(idxArg, (int)(u2.offset / u2.step)); //int u2_offset_y
    idxArg = kernel.set(idxArg, (char)calc_error);    //char calc_error

    return kernel.run(2, globalsize, NULL, false);
}

bool cv_ocl_tvl1flow::estimateDualVariables(UMat &u1, UMat &u2,
    UMat &p11, UMat &p12, UMat &p21, UMat &p22, float taut)
{
    size_t globalsize[2] = { u1.cols, u1.rows };

    ocl::Kernel kernel;
    if (!kernel.create("estimateDualVariablesKernel", cv::ocl::video::optical_flow_tvl1_oclsrc, ""))
        return false;

    int idxArg = 0;
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(u1));// const float* u1
    idxArg = kernel.set(idxArg, (int)(u1.cols)); //int u1_col
    idxArg = kernel.set(idxArg, (int)(u1.rows)); //int u1_row
    idxArg = kernel.set(idxArg, (int)(u1.step/u1.elemSize())); //int u1_step
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(u2)); // const float* u2
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadWrite(p11)); // float* p11
    idxArg = kernel.set(idxArg, (int)(p11.step/p11.elemSize())); //int p11_step
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadWrite(p12)); // float* p12
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadWrite(p21)); // float* p21
    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadWrite(p22)); // float* p22
    idxArg = kernel.set(idxArg, (float)(taut));    //float taut
    idxArg = kernel.set(idxArg, (int)(u2.step/u2.elemSize())); //int u2_step
    int u1_offset_x = (int)(u1.offset % u1.step);
    u1_offset_x = (int)(u1_offset_x / u1.elemSize());
    idxArg = kernel.set(idxArg, u1_offset_x); //int u1_offset_x
    idxArg = kernel.set(idxArg, (int)(u1.offset / u1.step)); //int u1_offset_y
    int u2_offset_x = (int)(u2.offset % u2.step);
    u2_offset_x = (int)(u2_offset_x / u2.elemSize());
    idxArg = kernel.set(idxArg, u2_offset_x); //int u2_offset_x
    idxArg = kernel.set(idxArg, (int)(u2.offset / u2.step)); //int u2_offset_y

    return kernel.run(2, globalsize, NULL, false);

}

OpticalFlowDual_TVL1::OpticalFlowDual_TVL1()
{
    tau            = 0.25;
    lambda         = 0.15;
    theta          = 0.3;
    nscales        = 5;
    warps          = 5;
    epsilon        = 0.01;
    gamma          = 0.;
    innerIterations = 30;
    outerIterations = 10;
    useInitialFlow = false;
    medianFiltering = 5;
    scaleStep      = 0.8;
}

void OpticalFlowDual_TVL1::calc(InputArray _I0, InputArray _I1, InputOutputArray _flow)
{
    CV_OCL_RUN(_flow.isUMat() &&
               ocl::Image2D::isFormatSupported(CV_32F, 1, false),
               calc_ocl(_I0, _I1, _flow))

    Mat I0 = _I0.getMat();
    Mat I1 = _I1.getMat();

    CV_Assert( I0.type() == CV_8UC1 || I0.type() == CV_32FC1 );
    CV_Assert( I0.size() == I1.size() );
    CV_Assert( I0.type() == I1.type() );
    CV_Assert( !useInitialFlow || (_flow.size() == I0.size() && _flow.type() == CV_32FC2) );
    CV_Assert( nscales > 0 );
    bool use_gamma = gamma != 0;
    // allocate memory for the pyramid structure
    dm.I0s.resize(nscales);
    dm.I1s.resize(nscales);
    dm.u1s.resize(nscales);
    dm.u2s.resize(nscales);
    dm.u3s.resize(nscales);

    I0.convertTo(dm.I0s[0], dm.I0s[0].depth(), I0.depth() == CV_8U ? 1.0 : 255.0);
    I1.convertTo(dm.I1s[0], dm.I1s[0].depth(), I1.depth() == CV_8U ? 1.0 : 255.0);

    dm.u1s[0].create(I0.size());
    dm.u2s[0].create(I0.size());
    if (use_gamma) dm.u3s[0].create(I0.size());

    if (useInitialFlow)
    {
        Mat_<float> mv[] = { dm.u1s[0], dm.u2s[0] };
        split(_flow.getMat(), mv);
    }

    dm.I1x_buf.create(I0.size());
    dm.I1y_buf.create(I0.size());

    dm.flowMap1_buf.create(I0.size());
    dm.flowMap2_buf.create(I0.size());

    dm.I1w_buf.create(I0.size());
    dm.I1wx_buf.create(I0.size());
    dm.I1wy_buf.create(I0.size());

    dm.grad_buf.create(I0.size());
    dm.rho_c_buf.create(I0.size());

    dm.v1_buf.create(I0.size());
    dm.v2_buf.create(I0.size());
    dm.v3_buf.create(I0.size());

    dm.p11_buf.create(I0.size());
    dm.p12_buf.create(I0.size());
    dm.p21_buf.create(I0.size());
    dm.p22_buf.create(I0.size());
    dm.p31_buf.create(I0.size());
    dm.p32_buf.create(I0.size());

    dm.div_p1_buf.create(I0.size());
    dm.div_p2_buf.create(I0.size());
    dm.div_p3_buf.create(I0.size());

    dm.u1x_buf.create(I0.size());
    dm.u1y_buf.create(I0.size());
    dm.u2x_buf.create(I0.size());
    dm.u2y_buf.create(I0.size());
    dm.u3x_buf.create(I0.size());
    dm.u3y_buf.create(I0.size());

    // create the scales
    for (int s = 1; s < nscales; ++s)
    {
        resize(dm.I0s[s - 1], dm.I0s[s], Size(), scaleStep, scaleStep);
        resize(dm.I1s[s - 1], dm.I1s[s], Size(), scaleStep, scaleStep);

        if (dm.I0s[s].cols < 16 || dm.I0s[s].rows < 16)
        {
            nscales = s;
            break;
        }

        if (useInitialFlow)
        {
            resize(dm.u1s[s - 1], dm.u1s[s], Size(), scaleStep, scaleStep);
            resize(dm.u2s[s - 1], dm.u2s[s], Size(), scaleStep, scaleStep);

            multiply(dm.u1s[s], Scalar::all(scaleStep), dm.u1s[s]);
            multiply(dm.u2s[s], Scalar::all(scaleStep), dm.u2s[s]);
        }
        else
        {
            dm.u1s[s].create(dm.I0s[s].size());
            dm.u2s[s].create(dm.I0s[s].size());
        }
        if (use_gamma) dm.u3s[s].create(dm.I0s[s].size());
    }
    if (!useInitialFlow)
    {
        dm.u1s[nscales - 1].setTo(Scalar::all(0));
        dm.u2s[nscales - 1].setTo(Scalar::all(0));
    }
    if (use_gamma) dm.u3s[nscales - 1].setTo(Scalar::all(0));
    // pyramidal structure for computing the optical flow
    for (int s = nscales - 1; s >= 0; --s)
    {
        // compute the optical flow at the current scale
        procOneScale(dm.I0s[s], dm.I1s[s], dm.u1s[s], dm.u2s[s], dm.u3s[s]);

        // if this was the last scale, finish now
        if (s == 0)
            break;

        // otherwise, upsample the optical flow

        // zoom the optical flow for the next finer scale
        resize(dm.u1s[s], dm.u1s[s - 1], dm.I0s[s - 1].size());
        resize(dm.u2s[s], dm.u2s[s - 1], dm.I0s[s - 1].size());
        if (use_gamma) resize(dm.u3s[s], dm.u3s[s - 1], dm.I0s[s - 1].size());

        // scale the optical flow with the appropriate zoom factor (don't scale u3!)
        multiply(dm.u1s[s - 1], Scalar::all(1 / scaleStep), dm.u1s[s - 1]);
        multiply(dm.u2s[s - 1], Scalar::all(1 / scaleStep), dm.u2s[s - 1]);
    }

    Mat uxy[] = { dm.u1s[0], dm.u2s[0] };
    merge(uxy, 2, _flow);
}

bool OpticalFlowDual_TVL1::calc_ocl(InputArray _I0, InputArray _I1, InputOutputArray _flow)
{
    UMat I0 = _I0.getUMat();
    UMat I1 = _I1.getUMat();

    CV_Assert(I0.type() == CV_8UC1 || I0.type() == CV_32FC1);
    CV_Assert(I0.size() == I1.size());
    CV_Assert(I0.type() == I1.type());
    CV_Assert(!useInitialFlow || (_flow.size() == I0.size() && _flow.type() == CV_32FC2));
    CV_Assert(nscales > 0);

    // allocate memory for the pyramid structure
    dum.I0s.resize(nscales);
    dum.I1s.resize(nscales);
    dum.u1s.resize(nscales);
    dum.u2s.resize(nscales);
    //I0s_step == I1s_step
    double alpha = I0.depth() == CV_8U ? 1.0 : 255.0;

    I0.convertTo(dum.I0s[0], CV_32F, alpha);
    I1.convertTo(dum.I1s[0], CV_32F, I1.depth() == CV_8U ? 1.0 : 255.0);

    dum.u1s[0].create(I0.size(), CV_32FC1);
    dum.u2s[0].create(I0.size(), CV_32FC1);

    if (useInitialFlow)
    {
        std::vector<UMat> umv;
        umv.push_back(dum.u1s[0]);
        umv.push_back(dum.u2s[0]);
        cv::split(_flow,umv);
    }

    dum.I1x_buf.create(I0.size(), CV_32FC1);
    dum.I1y_buf.create(I0.size(), CV_32FC1);

    dum.I1w_buf.create(I0.size(), CV_32FC1);
    dum.I1wx_buf.create(I0.size(), CV_32FC1);
    dum.I1wy_buf.create(I0.size(), CV_32FC1);

    dum.grad_buf.create(I0.size(), CV_32FC1);
    dum.rho_c_buf.create(I0.size(), CV_32FC1);

    dum.p11_buf.create(I0.size(), CV_32FC1);
    dum.p12_buf.create(I0.size(), CV_32FC1);
    dum.p21_buf.create(I0.size(), CV_32FC1);
    dum.p22_buf.create(I0.size(), CV_32FC1);

    dum.diff_buf.create(I0.size(), CV_32FC1);

    // create the scales
    for (int s = 1; s < nscales; ++s)
    {
        resize(dum.I0s[s - 1], dum.I0s[s], Size(), scaleStep, scaleStep);
        resize(dum.I1s[s - 1], dum.I1s[s], Size(), scaleStep, scaleStep);

        if (dum.I0s[s].cols < 16 || dum.I0s[s].rows < 16)
        {
            nscales = s;
            break;
        }

        if (useInitialFlow)
        {
            resize(dum.u1s[s - 1], dum.u1s[s], Size(), scaleStep, scaleStep);
            resize(dum.u2s[s - 1], dum.u2s[s], Size(), scaleStep, scaleStep);

            //scale by scale factor
            multiply(dum.u1s[s], Scalar::all(scaleStep), dum.u1s[s]);
            multiply(dum.u2s[s], Scalar::all(scaleStep), dum.u2s[s]);
        }
    }

    // pyramidal structure for computing the optical flow
    for (int s = nscales - 1; s >= 0; --s)
    {
        // compute the optical flow at the current scale
        if (!OpticalFlowDual_TVL1::procOneScale_ocl(dum.I0s[s], dum.I1s[s], dum.u1s[s], dum.u2s[s]))
            return false;

        // if this was the last scale, finish now
        if (s == 0)
            break;

        // zoom the optical flow for the next finer scale
        resize(dum.u1s[s], dum.u1s[s - 1], dum.I0s[s - 1].size());
        resize(dum.u2s[s], dum.u2s[s - 1], dum.I0s[s - 1].size());

        // scale the optical flow with the appropriate zoom factor
        multiply(dum.u1s[s - 1], Scalar::all(1 / scaleStep), dum.u1s[s - 1]);
        multiply(dum.u2s[s - 1], Scalar::all(1 / scaleStep), dum.u2s[s - 1]);
    }

    std::vector<UMat> uxy;
    uxy.push_back(dum.u1s[0]);
    uxy.push_back(dum.u2s[0]);
    merge(uxy, _flow);
    return true;
}

////////////////////////////////////////////////////////////
// buildFlowMap

struct BuildFlowMapBody : ParallelLoopBody
{
    void operator() (const Range& range) const;

    Mat_<float> u1;
    Mat_<float> u2;
    mutable Mat_<float> map1;
    mutable Mat_<float> map2;
};

void BuildFlowMapBody::operator() (const Range& range) const
{
    for (int y = range.start; y < range.end; ++y)
    {
        const float* u1Row = u1[y];
        const float* u2Row = u2[y];

        float* map1Row = map1[y];
        float* map2Row = map2[y];

        for (int x = 0; x < u1.cols; ++x)
        {
            map1Row[x] = x + u1Row[x];
            map2Row[x] = y + u2Row[x];
        }
    }
}

void buildFlowMap(const Mat_<float>& u1, const Mat_<float>& u2, Mat_<float>& map1, Mat_<float>& map2)
{
    CV_DbgAssert( u2.size() == u1.size() );
    CV_DbgAssert( map1.size() == u1.size() );
    CV_DbgAssert( map2.size() == u1.size() );

    BuildFlowMapBody body;

    body.u1 = u1;
    body.u2 = u2;
    body.map1 = map1;
    body.map2 = map2;

    parallel_for_(Range(0, u1.rows), body);
}

////////////////////////////////////////////////////////////
// centeredGradient

struct CenteredGradientBody : ParallelLoopBody
{
    void operator() (const Range& range) const;

    Mat_<float> src;
    mutable Mat_<float> dx;
    mutable Mat_<float> dy;
};

void CenteredGradientBody::operator() (const Range& range) const
{
    const int last_col = src.cols - 1;

    for (int y = range.start; y < range.end; ++y)
    {
        const float* srcPrevRow = src[y - 1];
        const float* srcCurRow = src[y];
        const float* srcNextRow = src[y + 1];

        float* dxRow = dx[y];
        float* dyRow = dy[y];

        for (int x = 1; x < last_col; ++x)
        {
            dxRow[x] = 0.5f * (srcCurRow[x + 1] - srcCurRow[x - 1]);
            dyRow[x] = 0.5f * (srcNextRow[x] - srcPrevRow[x]);
        }
    }
}

void centeredGradient(const Mat_<float>& src, Mat_<float>& dx, Mat_<float>& dy)
{
    CV_DbgAssert( src.rows > 2 && src.cols > 2 );
    CV_DbgAssert( dx.size() == src.size() );
    CV_DbgAssert( dy.size() == src.size() );

    const int last_row = src.rows - 1;
    const int last_col = src.cols - 1;

    // compute the gradient on the center body of the image
    {
        CenteredGradientBody body;

        body.src = src;
        body.dx = dx;
        body.dy = dy;

        parallel_for_(Range(1, last_row), body);
    }

    // compute the gradient on the first and last rows
    for (int x = 1; x < last_col; ++x)
    {
        dx(0, x) = 0.5f * (src(0, x + 1) - src(0, x - 1));
        dy(0, x) = 0.5f * (src(1, x) - src(0, x));

        dx(last_row, x) = 0.5f * (src(last_row, x + 1) - src(last_row, x - 1));
        dy(last_row, x) = 0.5f * (src(last_row, x) - src(last_row - 1, x));
    }

    // compute the gradient on the first and last columns
    for (int y = 1; y < last_row; ++y)
    {
        dx(y, 0) = 0.5f * (src(y, 1) - src(y, 0));
        dy(y, 0) = 0.5f * (src(y + 1, 0) - src(y - 1, 0));

        dx(y, last_col) = 0.5f * (src(y, last_col) - src(y, last_col - 1));
        dy(y, last_col) = 0.5f * (src(y + 1, last_col) - src(y - 1, last_col));
    }

    // compute the gradient at the four corners
    dx(0, 0) = 0.5f * (src(0, 1) - src(0, 0));
    dy(0, 0) = 0.5f * (src(1, 0) - src(0, 0));

    dx(0, last_col) = 0.5f * (src(0, last_col) - src(0, last_col - 1));
    dy(0, last_col) = 0.5f * (src(1, last_col) - src(0, last_col));

    dx(last_row, 0) = 0.5f * (src(last_row, 1) - src(last_row, 0));
    dy(last_row, 0) = 0.5f * (src(last_row, 0) - src(last_row - 1, 0));

    dx(last_row, last_col) = 0.5f * (src(last_row, last_col) - src(last_row, last_col - 1));
    dy(last_row, last_col) = 0.5f * (src(last_row, last_col) - src(last_row - 1, last_col));
}

////////////////////////////////////////////////////////////
// forwardGradient

struct ForwardGradientBody : ParallelLoopBody
{
    void operator() (const Range& range) const;

    Mat_<float> src;
    mutable Mat_<float> dx;
    mutable Mat_<float> dy;
};

void ForwardGradientBody::operator() (const Range& range) const
{
    const int last_col = src.cols - 1;

    for (int y = range.start; y < range.end; ++y)
    {
        const float* srcCurRow = src[y];
        const float* srcNextRow = src[y + 1];

        float* dxRow = dx[y];
        float* dyRow = dy[y];

        for (int x = 0; x < last_col; ++x)
        {
            dxRow[x] = srcCurRow[x + 1] - srcCurRow[x];
            dyRow[x] = srcNextRow[x] - srcCurRow[x];
        }
    }
}

void forwardGradient(const Mat_<float>& src, Mat_<float>& dx, Mat_<float>& dy)
{
    CV_DbgAssert( src.rows > 2 && src.cols > 2 );
    CV_DbgAssert( dx.size() == src.size() );
    CV_DbgAssert( dy.size() == src.size() );

    const int last_row = src.rows - 1;
    const int last_col = src.cols - 1;

    // compute the gradient on the central body of the image
    {
        ForwardGradientBody body;

        body.src = src;
        body.dx = dx;
        body.dy = dy;

        parallel_for_(Range(0, last_row), body);
    }

    // compute the gradient on the last row
    for (int x = 0; x < last_col; ++x)
    {
        dx(last_row, x) = src(last_row, x + 1) - src(last_row, x);
        dy(last_row, x) = 0.0f;
    }

    // compute the gradient on the last column
    for (int y = 0; y < last_row; ++y)
    {
        dx(y, last_col) = 0.0f;
        dy(y, last_col) = src(y + 1, last_col) - src(y, last_col);
    }

    dx(last_row, last_col) = 0.0f;
    dy(last_row, last_col) = 0.0f;
}

////////////////////////////////////////////////////////////
// divergence

struct DivergenceBody : ParallelLoopBody
{
    void operator() (const Range& range) const;

    Mat_<float> v1;
    Mat_<float> v2;
    mutable Mat_<float> div;
};

void DivergenceBody::operator() (const Range& range) const
{
    for (int y = range.start; y < range.end; ++y)
    {
        const float* v1Row = v1[y];
        const float* v2PrevRow = v2[y - 1];
        const float* v2CurRow = v2[y];

        float* divRow = div[y];

        for(int x = 1; x < v1.cols; ++x)
        {
            const float v1x = v1Row[x] - v1Row[x - 1];
            const float v2y = v2CurRow[x] - v2PrevRow[x];

            divRow[x] = v1x + v2y;
        }
    }
}

void divergence(const Mat_<float>& v1, const Mat_<float>& v2, Mat_<float>& div)
{
    CV_DbgAssert( v1.rows > 2 && v1.cols > 2 );
    CV_DbgAssert( v2.size() == v1.size() );
    CV_DbgAssert( div.size() == v1.size() );

    {
        DivergenceBody body;

        body.v1 = v1;
        body.v2 = v2;
        body.div = div;

        parallel_for_(Range(1, v1.rows), body);
    }

    // compute the divergence on the first row
    for(int x = 1; x < v1.cols; ++x)
        div(0, x) = v1(0, x) - v1(0, x - 1) + v2(0, x);

    // compute the divergence on the first column
    for (int y = 1; y < v1.rows; ++y)
        div(y, 0) = v1(y, 0) + v2(y, 0) - v2(y - 1, 0);

    div(0, 0) = v1(0, 0) + v2(0, 0);
}

////////////////////////////////////////////////////////////
// calcGradRho

struct CalcGradRhoBody : ParallelLoopBody
{
    void operator() (const Range& range) const;

    Mat_<float> I0;
    Mat_<float> I1w;
    Mat_<float> I1wx;
    Mat_<float> I1wy;
    Mat_<float> u1;
    Mat_<float> u2;
    mutable Mat_<float> grad;
    mutable Mat_<float> rho_c;
};

void CalcGradRhoBody::operator() (const Range& range) const
{
    for (int y = range.start; y < range.end; ++y)
    {
        const float* I0Row = I0[y];
        const float* I1wRow = I1w[y];
        const float* I1wxRow = I1wx[y];
        const float* I1wyRow = I1wy[y];
        const float* u1Row = u1[y];
        const float* u2Row = u2[y];

        float* gradRow = grad[y];
        float* rhoRow = rho_c[y];

        for (int x = 0; x < I0.cols; ++x)
        {
            const float Ix2 = I1wxRow[x] * I1wxRow[x];
            const float Iy2 = I1wyRow[x] * I1wyRow[x];

            // store the |Grad(I1)|^2
            gradRow[x] = Ix2 + Iy2;

            // compute the constant part of the rho function
            rhoRow[x] = (I1wRow[x] - I1wxRow[x] * u1Row[x] - I1wyRow[x] * u2Row[x] - I0Row[x]);
        }
    }
}

void calcGradRho(const Mat_<float>& I0, const Mat_<float>& I1w, const Mat_<float>& I1wx, const Mat_<float>& I1wy, const Mat_<float>& u1, const Mat_<float>& u2,
    Mat_<float>& grad, Mat_<float>& rho_c)
{
    CV_DbgAssert( I1w.size() == I0.size() );
    CV_DbgAssert( I1wx.size() == I0.size() );
    CV_DbgAssert( I1wy.size() == I0.size() );
    CV_DbgAssert( u1.size() == I0.size() );
    CV_DbgAssert( u2.size() == I0.size() );
    CV_DbgAssert( grad.size() == I0.size() );
    CV_DbgAssert( rho_c.size() == I0.size() );

    CalcGradRhoBody body;

    body.I0 = I0;
    body.I1w = I1w;
    body.I1wx = I1wx;
    body.I1wy = I1wy;
    body.u1 = u1;
    body.u2 = u2;
    body.grad = grad;
    body.rho_c = rho_c;

    parallel_for_(Range(0, I0.rows), body);
}

////////////////////////////////////////////////////////////
// estimateV

struct EstimateVBody : ParallelLoopBody
{
    void operator() (const Range& range) const;

    Mat_<float> I1wx;
    Mat_<float> I1wy;
    Mat_<float> u1;
    Mat_<float> u2;
    Mat_<float> u3;
    Mat_<float> grad;
    Mat_<float> rho_c;
    mutable Mat_<float> v1;
    mutable Mat_<float> v2;
    mutable Mat_<float> v3;
    float l_t;
    float gamma;
};

void EstimateVBody::operator() (const Range& range) const
{
    bool use_gamma = gamma != 0;
    for (int y = range.start; y < range.end; ++y)
    {
        const float* I1wxRow = I1wx[y];
        const float* I1wyRow = I1wy[y];
        const float* u1Row = u1[y];
        const float* u2Row = u2[y];
        const float* u3Row = use_gamma?u3[y]:NULL;
        const float* gradRow = grad[y];
        const float* rhoRow = rho_c[y];

        float* v1Row = v1[y];
        float* v2Row = v2[y];
        float* v3Row = use_gamma ? v3[y]:NULL;

        for (int x = 0; x < I1wx.cols; ++x)
        {
            const float rho = use_gamma ? rhoRow[x] + (I1wxRow[x] * u1Row[x] + I1wyRow[x] * u2Row[x]) + gamma * u3Row[x] :
                                          rhoRow[x] + (I1wxRow[x] * u1Row[x] + I1wyRow[x] * u2Row[x]);
            float d1 = 0.0f;
            float d2 = 0.0f;
            float d3 = 0.0f;
            if (rho < -l_t * gradRow[x])
            {
                d1 = l_t * I1wxRow[x];
                d2 = l_t * I1wyRow[x];
                if (use_gamma) d3 = l_t * gamma;
            }
            else if (rho > l_t * gradRow[x])
            {
                d1 = -l_t * I1wxRow[x];
                d2 = -l_t * I1wyRow[x];
                if (use_gamma) d3 = -l_t * gamma;
            }
            else if (gradRow[x] > std::numeric_limits<float>::epsilon())
            {
                float fi = -rho / gradRow[x];
                d1 = fi * I1wxRow[x];
                d2 = fi * I1wyRow[x];
                if (use_gamma) d3 = fi * gamma;
            }

            v1Row[x] = u1Row[x] + d1;
            v2Row[x] = u2Row[x] + d2;
            if (use_gamma) v3Row[x] = u3Row[x] + d3;
        }
    }
}

void estimateV(const Mat_<float>& I1wx, const Mat_<float>& I1wy, const Mat_<float>& u1, const Mat_<float>& u2, const Mat_<float>& u3, const Mat_<float>& grad, const Mat_<float>& rho_c,
   Mat_<float>& v1, Mat_<float>& v2, Mat_<float>& v3, float l_t, float gamma)
{
    CV_DbgAssert( I1wy.size() == I1wx.size() );
    CV_DbgAssert( u1.size() == I1wx.size() );
    CV_DbgAssert( u2.size() == I1wx.size() );
    CV_DbgAssert( grad.size() == I1wx.size() );
    CV_DbgAssert( rho_c.size() == I1wx.size() );
    CV_DbgAssert( v1.size() == I1wx.size() );
    CV_DbgAssert( v2.size() == I1wx.size() );

    EstimateVBody body;
    bool use_gamma = gamma != 0;
    body.I1wx = I1wx;
    body.I1wy = I1wy;
    body.u1 = u1;
    body.u2 = u2;
    if (use_gamma) body.u3 = u3;
    body.grad = grad;
    body.rho_c = rho_c;
    body.v1 = v1;
    body.v2 = v2;
    if (use_gamma) body.v3 = v3;
    body.l_t = l_t;
    body.gamma = gamma;
    parallel_for_(Range(0, I1wx.rows), body);
}

////////////////////////////////////////////////////////////
// estimateU

float estimateU(const Mat_<float>& v1, const Mat_<float>& v2, const Mat_<float>& v3,
            const Mat_<float>& div_p1, const Mat_<float>& div_p2, const Mat_<float>& div_p3,
            Mat_<float>& u1, Mat_<float>& u2, Mat_<float>& u3,
            float theta, float gamma)
{
    CV_DbgAssert( v2.size() == v1.size() );
    CV_DbgAssert( div_p1.size() == v1.size() );
    CV_DbgAssert( div_p2.size() == v1.size() );
    CV_DbgAssert( u1.size() == v1.size() );
    CV_DbgAssert( u2.size() == v1.size() );

    float error = 0.0f;
    bool use_gamma = gamma != 0;
    for (int y = 0; y < v1.rows; ++y)
    {
        const float* v1Row = v1[y];
        const float* v2Row = v2[y];
        const float* v3Row = use_gamma?v3[y]:NULL;
        const float* divP1Row = div_p1[y];
        const float* divP2Row = div_p2[y];
        const float* divP3Row = use_gamma?div_p3[y]:NULL;

        float* u1Row = u1[y];
        float* u2Row = u2[y];
        float* u3Row = use_gamma?u3[y]:NULL;


        for (int x = 0; x < v1.cols; ++x)
        {
            const float u1k = u1Row[x];
            const float u2k = u2Row[x];
            const float u3k = use_gamma?u3Row[x]:0;

            u1Row[x] = v1Row[x] + theta * divP1Row[x];
            u2Row[x] = v2Row[x] + theta * divP2Row[x];
            if (use_gamma) u3Row[x] = v3Row[x] + theta * divP3Row[x];
            error += use_gamma?(u1Row[x] - u1k) * (u1Row[x] - u1k) + (u2Row[x] - u2k) * (u2Row[x] - u2k) + (u3Row[x] - u3k) * (u3Row[x] - u3k):
                               (u1Row[x] - u1k) * (u1Row[x] - u1k) + (u2Row[x] - u2k) * (u2Row[x] - u2k);
        }
    }

    return error;
}

////////////////////////////////////////////////////////////
// estimateDualVariables

struct EstimateDualVariablesBody : ParallelLoopBody
{
    void operator() (const Range& range) const;

    Mat_<float> u1x;
    Mat_<float> u1y;
    Mat_<float> u2x;
    Mat_<float> u2y;
    Mat_<float> u3x;
    Mat_<float> u3y;
    mutable Mat_<float> p11;
    mutable Mat_<float> p12;
    mutable Mat_<float> p21;
    mutable Mat_<float> p22;
    mutable Mat_<float> p31;
    mutable Mat_<float> p32;
    float taut;
    bool use_gamma;
};

void EstimateDualVariablesBody::operator() (const Range& range) const
{
    for (int y = range.start; y < range.end; ++y)
    {
        const float* u1xRow = u1x[y];
        const float* u1yRow = u1y[y];
        const float* u2xRow = u2x[y];
        const float* u2yRow = u2y[y];
        const float* u3xRow = u3x[y];
        const float* u3yRow = u3y[y];

        float* p11Row = p11[y];
        float* p12Row = p12[y];
        float* p21Row = p21[y];
        float* p22Row = p22[y];
        float* p31Row = p31[y];
        float* p32Row = p32[y];

        for (int x = 0; x < u1x.cols; ++x)
        {
            const float g1 = static_cast<float>(hypot(u1xRow[x], u1yRow[x]));
            const float g2 = static_cast<float>(hypot(u2xRow[x], u2yRow[x]));
            const float g3 = static_cast<float>(hypot(u3xRow[x], u3yRow[x]));

            const float ng1  = 1.0f + taut * g1;
            const float ng2 =  1.0f + taut * g2;
            const float ng3 = 1.0f + taut * g3;

            p11Row[x] = (p11Row[x] + taut * u1xRow[x]) / ng1;
            p12Row[x] = (p12Row[x] + taut * u1yRow[x]) / ng1;
            p21Row[x] = (p21Row[x] + taut * u2xRow[x]) / ng2;
            p22Row[x] = (p22Row[x] + taut * u2yRow[x]) / ng2;
            if (use_gamma) p31Row[x] = (p31Row[x] + taut * u3xRow[x]) / ng3;
            if (use_gamma) p32Row[x] = (p32Row[x] + taut * u3yRow[x]) / ng3;
        }
    }
}

void estimateDualVariables(const Mat_<float>& u1x, const Mat_<float>& u1y,
                     const Mat_<float>& u2x, const Mat_<float>& u2y,
                     const Mat_<float>& u3x, const Mat_<float>& u3y,
                           Mat_<float>& p11, Mat_<float>& p12,
                     Mat_<float>& p21, Mat_<float>& p22,
                     Mat_<float>& p31, Mat_<float>& p32,
                     float taut, bool use_gamma)
{
    CV_DbgAssert( u1y.size() == u1x.size() );
    CV_DbgAssert( u2x.size() == u1x.size() );
    CV_DbgAssert( u3x.size() == u1x.size() );
    CV_DbgAssert( u2y.size() == u1x.size() );
    CV_DbgAssert( u3y.size() == u1x.size() );
    CV_DbgAssert( p11.size() == u1x.size() );
    CV_DbgAssert( p12.size() == u1x.size() );
    CV_DbgAssert( p21.size() == u1x.size() );
    CV_DbgAssert( p22.size() == u1x.size() );
    CV_DbgAssert( p31.size() == u1x.size() );
    CV_DbgAssert( p32.size() == u1x.size() );

    EstimateDualVariablesBody body;

    body.u1x = u1x;
    body.u1y = u1y;
    body.u2x = u2x;
    body.u2y = u2y;
    body.u3x = u3x;
    body.u3y = u3y;
    body.p11 = p11;
    body.p12 = p12;
    body.p21 = p21;
    body.p22 = p22;
    body.p31 = p31;
    body.p32 = p32;
    body.taut = taut;
    body.use_gamma = use_gamma;

    parallel_for_(Range(0, u1x.rows), body);
}

bool OpticalFlowDual_TVL1::procOneScale_ocl(const UMat& I0, const UMat& I1, UMat& u1, UMat& u2)
{
    using namespace cv_ocl_tvl1flow;

    const double scaledEpsilon = epsilon * epsilon * I0.size().area();

    CV_DbgAssert(I1.size() == I0.size());
    CV_DbgAssert(I1.type() == I0.type());
    CV_DbgAssert(u1.empty() || u1.size() == I0.size());
    CV_DbgAssert(u2.size() == u1.size());

    if (u1.empty())
    {
        u1.create(I0.size(), CV_32FC1);
        u1.setTo(Scalar::all(0));

        u2.create(I0.size(), CV_32FC1);
        u2.setTo(Scalar::all(0));
    }

    UMat I1x = dum.I1x_buf(Rect(0, 0, I0.cols, I0.rows));
    UMat I1y = dum.I1y_buf(Rect(0, 0, I0.cols, I0.rows));

    if (!centeredGradient(I1, I1x, I1y))
        return false;

    UMat I1w = dum.I1w_buf(Rect(0, 0, I0.cols, I0.rows));
    UMat I1wx = dum.I1wx_buf(Rect(0, 0, I0.cols, I0.rows));
    UMat I1wy = dum.I1wy_buf(Rect(0, 0, I0.cols, I0.rows));

    UMat grad = dum.grad_buf(Rect(0, 0, I0.cols, I0.rows));
    UMat rho_c = dum.rho_c_buf(Rect(0, 0, I0.cols, I0.rows));

    UMat p11 = dum.p11_buf(Rect(0, 0, I0.cols, I0.rows));
    UMat p12 = dum.p12_buf(Rect(0, 0, I0.cols, I0.rows));
    UMat p21 = dum.p21_buf(Rect(0, 0, I0.cols, I0.rows));
    UMat p22 = dum.p22_buf(Rect(0, 0, I0.cols, I0.rows));
    p11.setTo(Scalar::all(0));
    p12.setTo(Scalar::all(0));
    p21.setTo(Scalar::all(0));
    p22.setTo(Scalar::all(0));

    UMat diff = dum.diff_buf(Rect(0, 0, I0.cols, I0.rows));

    const float l_t = static_cast<float>(lambda * theta);
    const float taut = static_cast<float>(tau / theta);
    int n;

    for (int warpings = 0; warpings < warps; ++warpings)
    {
        if (!warpBackward(I0, I1, I1x, I1y, u1, u2, I1w, I1wx, I1wy, grad, rho_c))
            return false;

        double error = std::numeric_limits<double>::max();
        double prev_error = 0;

        for (int n_outer = 0; error > scaledEpsilon && n_outer < outerIterations; ++n_outer)
        {
            if (medianFiltering > 1) {
                cv::medianBlur(u1, u1, medianFiltering);
                cv::medianBlur(u2, u2, medianFiltering);
            }
            for (int n_inner = 0; error > scaledEpsilon && n_inner < innerIterations; ++n_inner)
            {
                // some tweaks to make sum operation less frequently
                n = n_inner + n_outer*innerIterations;
                char calc_error = (n & 0x1) && (prev_error < scaledEpsilon);
                if (!estimateU(I1wx, I1wy, grad, rho_c, p11, p12, p21, p22,
                    u1, u2, diff, l_t, static_cast<float>(theta), calc_error))
                    return false;
                if (calc_error)
                {
                    error = cv::sum(diff)[0];
                    prev_error = error;
                }
                else
                {
                    error = std::numeric_limits<double>::max();
                    prev_error -= scaledEpsilon;
                }
                if (!estimateDualVariables(u1, u2, p11, p12, p21, p22, taut))
                    return false;
            }
        }
    }
    return true;
}

void OpticalFlowDual_TVL1::procOneScale(const Mat_<float>& I0, const Mat_<float>& I1, Mat_<float>& u1, Mat_<float>& u2, Mat_<float>& u3)
{
    const float scaledEpsilon = static_cast<float>(epsilon * epsilon * I0.size().area());

    CV_DbgAssert( I1.size() == I0.size() );
    CV_DbgAssert( I1.type() == I0.type() );
    CV_DbgAssert( u1.size() == I0.size() );
    CV_DbgAssert( u2.size() == u1.size() );

    Mat_<float> I1x = dm.I1x_buf(Rect(0, 0, I0.cols, I0.rows));
    Mat_<float> I1y = dm.I1y_buf(Rect(0, 0, I0.cols, I0.rows));
    centeredGradient(I1, I1x, I1y);

    Mat_<float> flowMap1 = dm.flowMap1_buf(Rect(0, 0, I0.cols, I0.rows));
    Mat_<float> flowMap2 = dm.flowMap2_buf(Rect(0, 0, I0.cols, I0.rows));

    Mat_<float> I1w = dm.I1w_buf(Rect(0, 0, I0.cols, I0.rows));
    Mat_<float> I1wx = dm.I1wx_buf(Rect(0, 0, I0.cols, I0.rows));
    Mat_<float> I1wy = dm.I1wy_buf(Rect(0, 0, I0.cols, I0.rows));

    Mat_<float> grad = dm.grad_buf(Rect(0, 0, I0.cols, I0.rows));
    Mat_<float> rho_c = dm.rho_c_buf(Rect(0, 0, I0.cols, I0.rows));

    Mat_<float> v1 = dm.v1_buf(Rect(0, 0, I0.cols, I0.rows));
    Mat_<float> v2 = dm.v2_buf(Rect(0, 0, I0.cols, I0.rows));
    Mat_<float> v3 = dm.v3_buf(Rect(0, 0, I0.cols, I0.rows));

    Mat_<float> p11 = dm.p11_buf(Rect(0, 0, I0.cols, I0.rows));
    Mat_<float> p12 = dm.p12_buf(Rect(0, 0, I0.cols, I0.rows));
    Mat_<float> p21 = dm.p21_buf(Rect(0, 0, I0.cols, I0.rows));
    Mat_<float> p22 = dm.p22_buf(Rect(0, 0, I0.cols, I0.rows));
    Mat_<float> p31 = dm.p31_buf(Rect(0, 0, I0.cols, I0.rows));
    Mat_<float> p32 = dm.p32_buf(Rect(0, 0, I0.cols, I0.rows));
    p11.setTo(Scalar::all(0));
    p12.setTo(Scalar::all(0));
    p21.setTo(Scalar::all(0));
    p22.setTo(Scalar::all(0));
    bool use_gamma = gamma != 0.;
    if (use_gamma) p31.setTo(Scalar::all(0));
    if (use_gamma) p32.setTo(Scalar::all(0));

    Mat_<float> div_p1 = dm.div_p1_buf(Rect(0, 0, I0.cols, I0.rows));
    Mat_<float> div_p2 = dm.div_p2_buf(Rect(0, 0, I0.cols, I0.rows));
    Mat_<float> div_p3 = dm.div_p3_buf(Rect(0, 0, I0.cols, I0.rows));

    Mat_<float> u1x = dm.u1x_buf(Rect(0, 0, I0.cols, I0.rows));
    Mat_<float> u1y = dm.u1y_buf(Rect(0, 0, I0.cols, I0.rows));
    Mat_<float> u2x = dm.u2x_buf(Rect(0, 0, I0.cols, I0.rows));
    Mat_<float> u2y = dm.u2y_buf(Rect(0, 0, I0.cols, I0.rows));
    Mat_<float> u3x = dm.u3x_buf(Rect(0, 0, I0.cols, I0.rows));
    Mat_<float> u3y = dm.u3y_buf(Rect(0, 0, I0.cols, I0.rows));

    const float l_t = static_cast<float>(lambda * theta);
    const float taut = static_cast<float>(tau / theta);

    for (int warpings = 0; warpings < warps; ++warpings)
    {
        // compute the warping of the target image and its derivatives
        buildFlowMap(u1, u2, flowMap1, flowMap2);
        remap(I1, I1w, flowMap1, flowMap2, INTER_CUBIC);
        remap(I1x, I1wx, flowMap1, flowMap2, INTER_CUBIC);
        remap(I1y, I1wy, flowMap1, flowMap2, INTER_CUBIC);
        //calculate I1(x+u0) and its gradient
        calcGradRho(I0, I1w, I1wx, I1wy, u1, u2, grad, rho_c);

        float error = std::numeric_limits<float>::max();
        for (int n_outer = 0; error > scaledEpsilon && n_outer < outerIterations; ++n_outer)
        {
            if (medianFiltering > 1) {
                cv::medianBlur(u1, u1, medianFiltering);
                cv::medianBlur(u2, u2, medianFiltering);
            }
            for (int n_inner = 0; error > scaledEpsilon && n_inner < innerIterations; ++n_inner)
            {
                // estimate the values of the variable (v1, v2) (thresholding operator TH)
                estimateV(I1wx, I1wy, u1, u2, u3, grad, rho_c, v1, v2, v3, l_t, static_cast<float>(gamma));

                // compute the divergence of the dual variable (p1, p2, p3)
                divergence(p11, p12, div_p1);
                divergence(p21, p22, div_p2);
                if (use_gamma) divergence(p31, p32, div_p3);

                // estimate the values of the optical flow (u1, u2)
                error = estimateU(v1, v2, v3, div_p1, div_p2, div_p3, u1, u2, u3, static_cast<float>(theta), static_cast<float>(gamma));

                // compute the gradient of the optical flow (Du1, Du2)
                forwardGradient(u1, u1x, u1y);
                forwardGradient(u2, u2x, u2y);
                if (use_gamma) forwardGradient(u3, u3x, u3y);

                // estimate the values of the dual variable (p1, p2, p3)
                estimateDualVariables(u1x, u1y, u2x, u2y, u3x, u3y, p11, p12, p21, p22, p31, p32, taut, use_gamma);
            }
        }
    }
}

void OpticalFlowDual_TVL1::collectGarbage()
{
    //dataMat structure dm
    dm.I0s.clear();
    dm.I1s.clear();
    dm.u1s.clear();
    dm.u2s.clear();

    dm.I1x_buf.release();
    dm.I1y_buf.release();

    dm.flowMap1_buf.release();
    dm.flowMap2_buf.release();

    dm.I1w_buf.release();
    dm.I1wx_buf.release();
    dm.I1wy_buf.release();

    dm.grad_buf.release();
    dm.rho_c_buf.release();

    dm.v1_buf.release();
    dm.v2_buf.release();

    dm.p11_buf.release();
    dm.p12_buf.release();
    dm.p21_buf.release();
    dm.p22_buf.release();

    dm.div_p1_buf.release();
    dm.div_p2_buf.release();

    dm.u1x_buf.release();
    dm.u1y_buf.release();
    dm.u2x_buf.release();
    dm.u2y_buf.release();

    //dataUMat structure dum
    dum.I0s.clear();
    dum.I1s.clear();
    dum.u1s.clear();
    dum.u2s.clear();

    dum.I1x_buf.release();
    dum.I1y_buf.release();

    dum.I1w_buf.release();
    dum.I1wx_buf.release();
    dum.I1wy_buf.release();

    dum.grad_buf.release();
    dum.rho_c_buf.release();

    dum.p11_buf.release();
    dum.p12_buf.release();
    dum.p21_buf.release();
    dum.p22_buf.release();

    dum.diff_buf.release();
    dum.norm_buf.release();
}

} // namespace

Ptr<DualTVL1OpticalFlow> cv::createOptFlow_DualTVL1()
{
    return makePtr<OpticalFlowDual_TVL1>();
}