C++ UMat::type方法代码示例

static bool ocl_fastNlMeansDenoisingColored( InputArray _src, OutputArray _dst,
                                      float h, float hForColorComponents,
                                      int templateWindowSize, int searchWindowSize)
{
    UMat src = _src.getUMat();
    _dst.create(src.size(), src.type());
    UMat dst = _dst.getUMat();

    UMat src_lab;
    cvtColor(src, src_lab, COLOR_LBGR2Lab);

    UMat l(src.size(), CV_8U);
    UMat ab(src.size(), CV_8UC2);
    std::vector<UMat> l_ab(2), l_ab_denoised(2);
    l_ab[0] = l;
    l_ab[1] = ab;
    l_ab_denoised[0].create(src.size(), CV_8U);
    l_ab_denoised[1].create(src.size(), CV_8UC2);

    int from_to[] = { 0,0, 1,1, 2,2 };
    mixChannels(std::vector<UMat>(1, src_lab), l_ab, from_to, 3);

    fastNlMeansDenoising(l_ab[0], l_ab_denoised[0], h, templateWindowSize, searchWindowSize);
    fastNlMeansDenoising(l_ab[1], l_ab_denoised[1], hForColorComponents, templateWindowSize, searchWindowSize);

    UMat dst_lab(src.size(), CV_8UC3);
    mixChannels(l_ab_denoised, std::vector<UMat>(1, dst_lab), from_to, 3);

    cvtColor(dst_lab, dst, COLOR_Lab2LBGR, src.channels());
    return true;
}

void SURF_OCL::downloadKeypoints(const UMat &keypointsGPU, std::vector<KeyPoint> &keypoints)
{
    const int nFeatures = keypointsGPU.cols;

    if (nFeatures == 0)
        keypoints.clear();
    else
    {
        CV_Assert(keypointsGPU.type() == CV_32FC1 && keypointsGPU.rows == ROWS_COUNT);

        Mat keypointsCPU = keypointsGPU.getMat(ACCESS_READ);
        keypoints.resize(nFeatures);

        float *kp_x = keypointsCPU.ptr<float>(SURF_OCL::X_ROW);
        float *kp_y = keypointsCPU.ptr<float>(SURF_OCL::Y_ROW);
        int *kp_laplacian = keypointsCPU.ptr<int>(SURF_OCL::LAPLACIAN_ROW);
        int *kp_octave = keypointsCPU.ptr<int>(SURF_OCL::OCTAVE_ROW);
        float *kp_size = keypointsCPU.ptr<float>(SURF_OCL::SIZE_ROW);
        float *kp_dir = keypointsCPU.ptr<float>(SURF_OCL::ANGLE_ROW);
        float *kp_hessian = keypointsCPU.ptr<float>(SURF_OCL::HESSIAN_ROW);

        for (int i = 0; i < nFeatures; ++i)
        {
            KeyPoint &kp = keypoints[i];
            kp.pt.x = kp_x[i];
            kp.pt.y = kp_y[i];
            kp.class_id = kp_laplacian[i];
            kp.octave = kp_octave[i];
            kp.size = kp_size[i];
            kp.angle = kp_dir[i];
            kp.response = kp_hessian[i];
        }
    }
}

UMat cv::superres::convertToType(const UMat& src, int type, UMat& buf0, UMat& buf1)
{
    CV_INSTRUMENT_REGION();

    if (src.type() == type)
        return src;

    const int depth = CV_MAT_DEPTH(type);
    const int cn = CV_MAT_CN(type);

    if (src.depth() == depth)
    {
        convertToCn(src, buf0, cn);
        return buf0;
    }

    if (src.channels() == cn)
    {
        convertToDepth(src, buf1, depth);
        return buf1;
    }

    convertToCn(src, buf0, cn);
    convertToDepth(buf0, buf1, depth);
    return buf1;
}

UMat& UMat::setTo(InputArray _value, InputArray _mask)
{
    bool haveMask = !_mask.empty();
#ifdef HAVE_OPENCL
    int tp = type(), cn = CV_MAT_CN(tp), d = CV_MAT_DEPTH(tp);

    if( dims <= 2 && cn <= 4 && CV_MAT_DEPTH(tp) < CV_64F && ocl::useOpenCL() )
    {
        Mat value = _value.getMat();
        CV_Assert( checkScalar(value, type(), _value.kind(), _InputArray::UMAT) );
        int kercn = haveMask || cn == 3 ? cn : std::max(cn, ocl::predictOptimalVectorWidth(*this)),
                kertp = CV_MAKE_TYPE(d, kercn);

        double buf[16] = { 0, 0, 0, 0, 0, 0, 0, 0,
                           0, 0, 0, 0, 0, 0, 0, 0 };
        convertAndUnrollScalar(value, tp, (uchar *)buf, kercn / cn);

        int scalarcn = kercn == 3 ? 4 : kercn, rowsPerWI = ocl::Device::getDefault().isIntel() ? 4 : 1;
        String opts = format("-D dstT=%s -D rowsPerWI=%d -D dstST=%s -D dstT1=%s -D cn=%d",
                             ocl::memopTypeToStr(kertp), rowsPerWI,
                             ocl::memopTypeToStr(CV_MAKETYPE(d, scalarcn)),
                             ocl::memopTypeToStr(d), kercn);

        ocl::Kernel setK(haveMask ? "setMask" : "set", ocl::core::copyset_oclsrc, opts);
        if( !setK.empty() )
        {
            ocl::KernelArg scalararg(0, 0, 0, 0, buf, CV_ELEM_SIZE(d) * scalarcn);
            UMat mask;

            if( haveMask )
            {
                mask = _mask.getUMat();
                CV_Assert( mask.size() == size() && mask.type() == CV_8UC1 );
                ocl::KernelArg maskarg = ocl::KernelArg::ReadOnlyNoSize(mask),
                        dstarg = ocl::KernelArg::ReadWrite(*this);
                setK.args(maskarg, dstarg, scalararg);
            }
            else
            {
                ocl::KernelArg dstarg = ocl::KernelArg::WriteOnly(*this, cn, kercn);
                setK.args(dstarg, scalararg);
            }

            size_t globalsize[] = { cols * cn / kercn, (rows + rowsPerWI - 1) / rowsPerWI };
            if( setK.run(2, globalsize, NULL, false) )
            {
                CV_IMPL_ADD(CV_IMPL_OCL);
                return *this;
            }
        }
    }
#endif
    Mat m = getMat(haveMask ? ACCESS_RW : ACCESS_WRITE);
    m.setTo(_value, _mask);
    return *this;
}

// draw both pure-C++ and ocl square results onto a single image
static UMat drawSquaresBoth( const UMat& image,
                            const vector<vector<Point> >& sqs)
{
    UMat imgToShow(Size(image.cols, image.rows), image.type());
    image.copyTo(imgToShow);

    drawSquares(imgToShow, sqs);

    return imgToShow;
}

UMat UMat::diag(const UMat& d)
{
    CV_Assert( d.cols == 1 || d.rows == 1 );
    int len = d.rows + d.cols - 1;
    UMat m(len, len, d.type(), Scalar(0));
    UMat md = m.diag();
    if( d.cols == 1 )
        d.copyTo(md);
    else
        transpose(d, md);
    return m;
}

static bool ocl_pyrUp( InputArray _src, OutputArray _dst, const Size& _dsz, int borderType)
{
    int type = _src.type(), depth = CV_MAT_DEPTH(type), channels = CV_MAT_CN(type);

    if (channels > 4 || borderType != BORDER_DEFAULT)
        return false;

    bool doubleSupport = ocl::Device::getDefault().doubleFPConfig() > 0;
    if (depth == CV_64F && !doubleSupport)
        return false;

    Size ssize = _src.size();
    if ((_dsz.area() != 0) && (_dsz != Size(ssize.width * 2, ssize.height * 2)))
        return false;

    UMat src = _src.getUMat();
    Size dsize = Size(ssize.width * 2, ssize.height * 2);
    _dst.create( dsize, src.type() );
    UMat dst = _dst.getUMat();

    int float_depth = depth == CV_64F ? CV_64F : CV_32F;
    const int local_size = 16;
    char cvt[2][50];
    String buildOptions = format(
            "-D T=%s -D FT=%s -D convertToT=%s -D convertToFT=%s%s "
            "-D T1=%s -D cn=%d -D LOCAL_SIZE=%d",
            ocl::typeToStr(type), ocl::typeToStr(CV_MAKETYPE(float_depth, channels)),
            ocl::convertTypeStr(float_depth, depth, channels, cvt[0]),
            ocl::convertTypeStr(depth, float_depth, channels, cvt[1]),
            doubleSupport ? " -D DOUBLE_SUPPORT" : "",
            ocl::typeToStr(depth), channels, local_size
    );
    size_t globalThreads[2] = { dst.cols, dst.rows };
    size_t localThreads[2] = { local_size, local_size };
    ocl::Kernel k;
    if (ocl::Device::getDefault().isIntel() && channels == 1)
    {
        k.create("pyrUp_unrolled", ocl::imgproc::pyr_up_oclsrc, buildOptions);
        globalThreads[0] = dst.cols/2; globalThreads[1] = dst.rows/2;
    }
    else
        k.create("pyrUp", ocl::imgproc::pyr_up_oclsrc, buildOptions);

    if (k.empty())
        return false;

    k.args(ocl::KernelArg::ReadOnly(src), ocl::KernelArg::WriteOnly(dst));
    return k.run(2, globalThreads, localThreads, false);
}

static bool ocl_pyrDown( InputArray _src, OutputArray _dst, const Size& _dsz, int borderType)
{
    int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);

    bool doubleSupport = ocl::Device::getDefault().doubleFPConfig() > 0;
    if (cn > 4 || (depth == CV_64F && !doubleSupport))
        return false;

    Size ssize = _src.size();
    Size dsize = _dsz.area() == 0 ? Size((ssize.width + 1) / 2, (ssize.height + 1) / 2) : _dsz;
    if (dsize.height < 2 || dsize.width < 2)
        return false;

    CV_Assert( ssize.width > 0 && ssize.height > 0 &&
            std::abs(dsize.width*2 - ssize.width) <= 2 &&
            std::abs(dsize.height*2 - ssize.height) <= 2 );

    UMat src = _src.getUMat();
    _dst.create( dsize, src.type() );
    UMat dst = _dst.getUMat();

    int float_depth = depth == CV_64F ? CV_64F : CV_32F;
    const int local_size = 256;
    int kercn = 1;
    if (depth == CV_8U && float_depth == CV_32F && cn == 1 && ocl::Device::getDefault().isIntel())
        kercn = 4;
    const char * const borderMap[] = { "BORDER_CONSTANT", "BORDER_REPLICATE", "BORDER_REFLECT", "BORDER_WRAP",
                                       "BORDER_REFLECT_101" };
    char cvt[2][50];
    String buildOptions = format(
            "-D T=%s -D FT=%s -D convertToT=%s -D convertToFT=%s%s "
            "-D T1=%s -D cn=%d -D kercn=%d -D fdepth=%d -D %s -D LOCAL_SIZE=%d",
            ocl::typeToStr(type), ocl::typeToStr(CV_MAKETYPE(float_depth, cn)),
            ocl::convertTypeStr(float_depth, depth, cn, cvt[0]),
            ocl::convertTypeStr(depth, float_depth, cn, cvt[1]),
            doubleSupport ? " -D DOUBLE_SUPPORT" : "", ocl::typeToStr(depth),
            cn, kercn, float_depth, borderMap[borderType], local_size
    );
    ocl::Kernel k("pyrDown", ocl::imgproc::pyr_down_oclsrc, buildOptions);
    if (k.empty())
        return false;

    k.args(ocl::KernelArg::ReadOnly(src), ocl::KernelArg::WriteOnly(dst));

    size_t localThreads[2]  = { local_size/kercn, 1 };
    size_t globalThreads[2] = { (src.cols + (kercn-1))/kercn, (dst.rows + 1) / 2 };
    return k.run(2, globalThreads, localThreads, false);
}

UMat& UMat::setTo(InputArray _value, InputArray _mask)
{
    bool haveMask = !_mask.empty();
    int tp = type(), cn = CV_MAT_CN(tp);
    if( dims <= 2 && cn <= 4 && cn != 3 && ocl::useOpenCL() )
    {
        Mat value = _value.getMat();
        CV_Assert( checkScalar(value, type(), _value.kind(), _InputArray::UMAT) );
        double buf[4];
        convertAndUnrollScalar(value, tp, (uchar*)buf, 1);

        char opts[1024];
        sprintf(opts, "-D dstT=%s", ocl::memopTypeToStr(tp));

        ocl::Kernel setK(haveMask ? "setMask" : "set", ocl::core::copyset_oclsrc, opts);
        if( !setK.empty() )
        {
            ocl::KernelArg scalararg(0, 0, 0, buf, CV_ELEM_SIZE(tp));
            UMat mask;

            if( haveMask )
            {
                mask = _mask.getUMat();
                CV_Assert( mask.size() == size() && mask.type() == CV_8U );
                ocl::KernelArg maskarg = ocl::KernelArg::ReadOnlyNoSize(mask);
                ocl::KernelArg dstarg = ocl::KernelArg::ReadWrite(*this);
                setK.args(maskarg, dstarg, scalararg);
            }
            else
            {
                ocl::KernelArg dstarg = ocl::KernelArg::WriteOnly(*this);
                setK.args(dstarg, scalararg);
            }

            size_t globalsize[] = { cols, rows };
            if( setK.run(2, globalsize, 0, false) )
                return *this;
        }
    }
    Mat m = getMat(haveMask ? ACCESS_RW : ACCESS_WRITE);
    m.setTo(_value, _mask);
    return *this;
}

static bool ocl_pyrDown( InputArray _src, OutputArray _dst, const Size& _dsz, int borderType)
{
    int type = _src.type(), depth = CV_MAT_DEPTH(type), channels = CV_MAT_CN(type);

    if (channels > 4 || borderType != BORDER_DEFAULT)
        return false;

    bool doubleSupport = ocl::Device::getDefault().doubleFPConfig() > 0;
    if ((depth == CV_64F) && !(doubleSupport))
        return false;

    Size ssize = _src.size();
    Size dsize = _dsz.area() == 0 ? Size((ssize.width + 1) / 2, (ssize.height + 1) / 2) : _dsz;
    CV_Assert( ssize.width > 0 && ssize.height > 0 &&
            std::abs(dsize.width*2 - ssize.width) <= 2 &&
            std::abs(dsize.height*2 - ssize.height) <= 2 );

    UMat src = _src.getUMat();
    _dst.create( dsize, src.type() );
    UMat dst = _dst.getUMat();

    int float_depth = depth == CV_64F ? CV_64F : CV_32F;
    char cvt[2][50];
    String buildOptions = format(
            "-D T=%s -D FT=%s -D convertToT=%s -D convertToFT=%s%s "
            "-D T1=%s -D cn=%d",
            ocl::typeToStr(type), ocl::typeToStr(CV_MAKETYPE(float_depth, channels)),
            ocl::convertTypeStr(float_depth, depth, channels, cvt[0]),
            ocl::convertTypeStr(depth, float_depth, channels, cvt[1]),
            doubleSupport ? " -D DOUBLE_SUPPORT" : "",
            ocl::typeToStr(depth), channels
    );
    ocl::Kernel k("pyrDown", ocl::imgproc::pyr_down_oclsrc, buildOptions);
    if (k.empty())
        return false;

    k.args(ocl::KernelArg::ReadOnly(src), ocl::KernelArg::WriteOnly(dst));

    size_t localThreads[2]  = { 256, 1 };
    size_t globalThreads[2] = { src.cols, dst.rows };
    return k.run(2, globalThreads, localThreads, false);
}

    static bool ocl_calcBtvRegularization(InputArray _src, OutputArray _dst, int btvKernelSize, const UMat & ubtvWeights)
    {
        int cn = _src.channels();
        ocl::Kernel k("calcBtvRegularization", ocl::superres::superres_btvl1_oclsrc,
                      format("-D cn=%d", cn));
        if (k.empty())
            return false;

        UMat src = _src.getUMat();
        _dst.create(src.size(), src.type());
        _dst.setTo(Scalar::all(0));
        UMat dst = _dst.getUMat();

        const int ksize = (btvKernelSize - 1) / 2;

        k.args(ocl::KernelArg::ReadOnlyNoSize(src), ocl::KernelArg::WriteOnly(dst),
              ksize, ocl::KernelArg::PtrReadOnly(ubtvWeights));

        size_t globalsize[2] = { (size_t)src.cols, (size_t)src.rows };
        return k.run(2, globalsize, NULL, false);
    }

static bool ocl_pyrUp( InputArray _src, OutputArray _dst, const Size& _dsz, int borderType)
{
    int type = _src.type(), depth = CV_MAT_DEPTH(type), channels = CV_MAT_CN(type);

    if ((channels != 1 && channels != 2 && channels != 4) || borderType != BORDER_DEFAULT)
        return false;

    bool doubleSupport = ocl::Device::getDefault().doubleFPConfig() > 0;
    if (depth == CV_64F && !doubleSupport)
        return false;

    Size ssize = _src.size();
    if ((_dsz.area() != 0) && (_dsz != Size(ssize.width * 2, ssize.height * 2)))
        return false;

    UMat src = _src.getUMat();
    Size dsize = Size(ssize.width * 2, ssize.height * 2);
    _dst.create( dsize, src.type() );
    UMat dst = _dst.getUMat();

    int float_depth = depth == CV_64F ? CV_64F : CV_32F;
    char cvt[2][50];
    ocl::Kernel k("pyrUp", ocl::imgproc::pyr_up_oclsrc,
                 format("-D T=%s -D FT=%s -D convertToT=%s -D convertToFT=%s%s",
                 ocl::typeToStr(type), ocl::typeToStr(CV_MAKETYPE(float_depth, channels)),
                 ocl::convertTypeStr(float_depth, depth, channels, cvt[0]),
                 ocl::convertTypeStr(depth, float_depth, channels, cvt[1]),
                 doubleSupport ? " -D DOUBLE_SUPPORT" : ""));
    if (k.empty())
        return false;

    k.args(ocl::KernelArg::ReadOnly(src), ocl::KernelArg::WriteOnly(dst));
    size_t globalThreads[2] = {dst.cols, dst.rows};
    size_t localThreads[2]  = {16, 16};

    return k.run(2, globalThreads, localThreads, false);
}

void MultiBandBlender::feed(InputArray _img, InputArray mask, Point tl)
{
#if ENABLE_LOG
    int64 t = getTickCount();
#endif

    UMat img = _img.getUMat();
    CV_Assert(img.type() == CV_16SC3 || img.type() == CV_8UC3);
    CV_Assert(mask.type() == CV_8U);

    // Keep source image in memory with small border
    int gap = 3 * (1 << num_bands_);
    Point tl_new(std::max(dst_roi_.x, tl.x - gap),
                 std::max(dst_roi_.y, tl.y - gap));
    Point br_new(std::min(dst_roi_.br().x, tl.x + img.cols + gap),
                 std::min(dst_roi_.br().y, tl.y + img.rows + gap));

    // Ensure coordinates of top-left, bottom-right corners are divided by (1 << num_bands_).
    // After that scale between layers is exactly 2.
    //
    // We do it to avoid interpolation problems when keeping sub-images only. There is no such problem when
    // image is bordered to have size equal to the final image size, but this is too memory hungry approach.
    tl_new.x = dst_roi_.x + (((tl_new.x - dst_roi_.x) >> num_bands_) << num_bands_);
    tl_new.y = dst_roi_.y + (((tl_new.y - dst_roi_.y) >> num_bands_) << num_bands_);
    int width = br_new.x - tl_new.x;
    int height = br_new.y - tl_new.y;
    width += ((1 << num_bands_) - width % (1 << num_bands_)) % (1 << num_bands_);
    height += ((1 << num_bands_) - height % (1 << num_bands_)) % (1 << num_bands_);
    br_new.x = tl_new.x + width;
    br_new.y = tl_new.y + height;
    int dy = std::max(br_new.y - dst_roi_.br().y, 0);
    int dx = std::max(br_new.x - dst_roi_.br().x, 0);
    tl_new.x -= dx; br_new.x -= dx;
    tl_new.y -= dy; br_new.y -= dy;

    int top = tl.y - tl_new.y;
    int left = tl.x - tl_new.x;
    int bottom = br_new.y - tl.y - img.rows;
    int right = br_new.x - tl.x - img.cols;

    // Create the source image Laplacian pyramid
    UMat img_with_border;
    copyMakeBorder(_img, img_with_border, top, bottom, left, right,
                   BORDER_REFLECT);
    LOGLN("  Add border to the source image, time: " << ((getTickCount() - t) / getTickFrequency()) << " sec");
#if ENABLE_LOG
    t = getTickCount();
#endif

    std::vector<UMat> src_pyr_laplace;
    if (can_use_gpu_ && img_with_border.depth() == CV_16S)
        createLaplacePyrGpu(img_with_border, num_bands_, src_pyr_laplace);
    else
        createLaplacePyr(img_with_border, num_bands_, src_pyr_laplace);

    LOGLN("  Create the source image Laplacian pyramid, time: " << ((getTickCount() - t) / getTickFrequency()) << " sec");
#if ENABLE_LOG
    t = getTickCount();
#endif

    // Create the weight map Gaussian pyramid
    UMat weight_map;
    std::vector<UMat> weight_pyr_gauss(num_bands_ + 1);

    if(weight_type_ == CV_32F)
    {
        mask.getUMat().convertTo(weight_map, CV_32F, 1./255.);
    }
    else // weight_type_ == CV_16S
    {
        mask.getUMat().convertTo(weight_map, CV_16S);
        UMat add_mask;
        compare(mask, 0, add_mask, CMP_NE);
        add(weight_map, Scalar::all(1), weight_map, add_mask);
    }

    copyMakeBorder(weight_map, weight_pyr_gauss[0], top, bottom, left, right, BORDER_CONSTANT);

    for (int i = 0; i < num_bands_; ++i)
        pyrDown(weight_pyr_gauss[i], weight_pyr_gauss[i + 1]);

    LOGLN("  Create the weight map Gaussian pyramid, time: " << ((getTickCount() - t) / getTickFrequency()) << " sec");
#if ENABLE_LOG
    t = getTickCount();
#endif

    int y_tl = tl_new.y - dst_roi_.y;
    int y_br = br_new.y - dst_roi_.y;
    int x_tl = tl_new.x - dst_roi_.x;
    int x_br = br_new.x - dst_roi_.x;

    // Add weighted layer of the source image to the final Laplacian pyramid layer
    for (int i = 0; i <= num_bands_; ++i)
    {
        Rect rc(x_tl, y_tl, x_br - x_tl, y_br - y_tl);
#ifdef HAVE_OPENCL
        if ( !cv::ocl::useOpenCL() ||
             !ocl_MultiBandBlender_feed(src_pyr_laplace[i], weight_pyr_gauss[i],
                    dst_pyr_laplace_[i](rc), dst_band_weights_[i](rc)) )
#endif
//.........这里部分代码省略.........

    bool BTVL1_Base::ocl_process(InputArrayOfArrays _src, OutputArray _dst, InputArrayOfArrays _forwardMotions,
                                 InputArrayOfArrays _backwardMotions, int baseIdx)
    {
        std::vector<UMat> & src = *(std::vector<UMat> *)_src.getObj(),
                & forwardMotions = *(std::vector<UMat> *)_forwardMotions.getObj(),
                & backwardMotions = *(std::vector<UMat> *)_backwardMotions.getObj();

        // update blur filter and btv weights
        if (!filter_ || blurKernelSize_ != curBlurKernelSize_ || blurSigma_ != curBlurSigma_ || src[0].type() != curSrcType_)
        {
            filter_ = createGaussianFilter(src[0].type(), Size(blurKernelSize_, blurKernelSize_), blurSigma_);
            curBlurKernelSize_ = blurKernelSize_;
            curBlurSigma_ = blurSigma_;
            curSrcType_ = src[0].type();
        }

        if (btvWeights_.empty() || btvKernelSize_ != curBtvKernelSize_ || alpha_ != curAlpha_)
        {
            calcBtvWeights(btvKernelSize_, alpha_, btvWeights_);
            Mat(btvWeights_, true).copyTo(ubtvWeights_);

            curBtvKernelSize_ = btvKernelSize_;
            curAlpha_ = alpha_;
        }

        // calc high res motions
        calcRelativeMotions(forwardMotions, backwardMotions, ulowResForwardMotions_, ulowResBackwardMotions_, baseIdx, src[0].size());

        upscaleMotions(ulowResForwardMotions_, uhighResForwardMotions_, scale_);
        upscaleMotions(ulowResBackwardMotions_, uhighResBackwardMotions_, scale_);

        uforwardMaps_.resize(uhighResForwardMotions_.size());
        ubackwardMaps_.resize(uhighResForwardMotions_.size());
        for (size_t i = 0; i < uhighResForwardMotions_.size(); ++i)
            buildMotionMaps(uhighResForwardMotions_[i], uhighResBackwardMotions_[i], uforwardMaps_[i], ubackwardMaps_[i]);

        // initial estimation
        const Size lowResSize = src[0].size();
        const Size highResSize(lowResSize.width * scale_, lowResSize.height * scale_);

        resize(src[baseIdx], uhighRes_, highResSize, 0, 0, INTER_LINEAR); // TODO

        // iterations
        udiffTerm_.create(highResSize, uhighRes_.type());
        ua_.create(highResSize, uhighRes_.type());
        ub_.create(highResSize, uhighRes_.type());
        uc_.create(lowResSize, uhighRes_.type());

        for (int i = 0; i < iterations_; ++i)
        {
            udiffTerm_.setTo(Scalar::all(0));

            for (size_t k = 0; k < src.size(); ++k)
            {
                // a = M * Ih
                remap(uhighRes_, ua_, ubackwardMaps_[k], noArray(), INTER_NEAREST);
                // b = HM * Ih
                GaussianBlur(ua_, ub_, Size(blurKernelSize_, blurKernelSize_), blurSigma_);
                // c = DHM * Ih
                resize(ub_, uc_, lowResSize, 0, 0, INTER_NEAREST);

                diffSign(src[k], uc_, uc_);

                // a = Dt * diff
                upscale(uc_, ua_, scale_);

                // b = HtDt * diff
                GaussianBlur(ua_, ub_, Size(blurKernelSize_, blurKernelSize_), blurSigma_);
                // a = MtHtDt * diff
                remap(ub_, ua_, uforwardMaps_[k], noArray(), INTER_NEAREST);

                add(udiffTerm_, ua_, udiffTerm_);
            }

            if (lambda_ > 0)
            {
                calcBtvRegularization(uhighRes_, uregTerm_, btvKernelSize_, btvWeights_, ubtvWeights_);
                addWeighted(udiffTerm_, 1.0, uregTerm_, -lambda_, 0.0, udiffTerm_);
            }

            addWeighted(uhighRes_, 1.0, udiffTerm_, tau_, 0.0, uhighRes_);
        }

        Rect inner(btvKernelSize_, btvKernelSize_, uhighRes_.cols - 2 * btvKernelSize_, uhighRes_.rows - 2 * btvKernelSize_);
        uhighRes_(inner).copyTo(_dst);

        return true;
    }