stitching/src/exposure_compensate.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*/

#include "precomp.hpp"
#ifdef HAVE_EIGEN
#include <Eigen/Core>
#include <Eigen/Dense>
#endif

namespace cv {
namespace detail {

Ptr<ExposureCompensator> ExposureCompensator::createDefault(int type)
{
    Ptr<ExposureCompensator> e;
    if (type == NO)
        e = makePtr<NoExposureCompensator>();
    else if (type == GAIN)
        e = makePtr<GainCompensator>();
    else if (type == GAIN_BLOCKS)
        e = makePtr<BlocksGainCompensator>();
    else if (type == CHANNELS)
        e = makePtr<ChannelsCompensator>();
    else if (type == CHANNELS_BLOCKS)
        e = makePtr<BlocksChannelsCompensator>();

    if (e.get() != nullptr)
        return e;

    CV_Error(Error::StsBadArg, "unsupported exposure compensation method");
}


void ExposureCompensator::feed(const std::vector<Point> &corners, const std::vector<UMat> &images,
                               const std::vector<UMat> &masks)
{
    std::vector<std::pair<UMat,uchar> > level_masks;
    for (size_t i = 0; i < masks.size(); ++i)
        level_masks.push_back(std::make_pair(masks[i], (uchar)255));
    feed(corners, images, level_masks);
}


void GainCompensator::feed(const std::vector<Point> &corners, const std::vector<UMat> &images,
                           const std::vector<std::pair<UMat,uchar> > &masks)
{
    LOGLN("Exposure compensation...");
#if ENABLE_LOG
    int64 t = getTickCount();
#endif

    const int num_images = static_cast<int>(images.size());
    Mat accumulated_gains;
    prepareSimilarityMask(corners, images);

    for (int n = 0; n < nr_feeds_; ++n)
    {
        if (n > 0)
        {
            // Apply previous iteration gains
            for (int i = 0; i < num_images; ++i)
                apply(i, corners[i], images[i], masks[i].first);
        }

        singleFeed(corners, images, masks);

        if (n == 0)
            accumulated_gains = gains_.clone();
        else
            multiply(accumulated_gains, gains_, accumulated_gains);
    }
    gains_ = accumulated_gains;

    LOGLN("Exposure compensation, time: " << ((getTickCount() - t) / getTickFrequency()) << " sec");
}

void GainCompensator::singleFeed(const std::vector<Point> &corners, const std::vector<UMat> &images,
                                 const std::vector<std::pair<UMat,uchar> > &masks)
{
    CV_Assert(corners.size() == images.size() && images.size() == masks.size());

    if (images.size() == 0)
        return;

    const int num_channels = images[0].channels();
    CV_Assert(std::all_of(images.begin(), images.end(),
        [num_channels](const UMat& image) { return image.channels() == num_channels; }));
    CV_Assert(num_channels == 1 || num_channels == 3);

    const int num_images = static_cast<int>(images.size());
    Mat_<int> N(num_images, num_images); N.setTo(0);
    Mat_<double> I(num_images, num_images); I.setTo(0);
    Mat_<bool> skip(num_images, 1); skip.setTo(true);

    Mat subimg1, subimg2;
    Mat_<uchar> submask1, submask2, intersect;

    std::vector<UMat>::iterator similarity_it = similarities_.begin();

    for (int i = 0; i < num_images; ++i)
    {
        for (int j = i; j < num_images; ++j)
        {
            Rect roi;
            if (overlapRoi(corners[i], corners[j], images[i].size(), images[j].size(), roi))
            {
                subimg1 = images[i](Rect(roi.tl() - corners[i], roi.br() - corners[i])).getMat(ACCESS_READ);
                subimg2 = images[j](Rect(roi.tl() - corners[j], roi.br() - corners[j])).getMat(ACCESS_READ);

                submask1 = masks[i].first(Rect(roi.tl() - corners[i], roi.br() - corners[i])).getMat(ACCESS_READ);
                submask2 = masks[j].first(Rect(roi.tl() - corners[j], roi.br() - corners[j])).getMat(ACCESS_READ);
                intersect = (submask1 == masks[i].second) & (submask2 == masks[j].second);

                if (!similarities_.empty())
                {
                    CV_Assert(similarity_it != similarities_.end());
                    UMat similarity = *similarity_it++;
                    // in-place operation has an issue. don't remove the swap
                    // detail https://github.com/opencv/opencv/issues/19184
                    Mat_<uchar> intersect_updated;
                    bitwise_and(intersect, similarity, intersect_updated);
                    std::swap(intersect, intersect_updated);
                }

                int intersect_count = countNonZero(intersect);
                N(i, j) = N(j, i) = std::max(1, intersect_count);

                // Don't compute Isums if subimages do not intersect anyway
                if (intersect_count == 0)
                    continue;

                // Don't skip images that intersect with at least one other image
                if (i != j)
                {
                    skip(i, 0) = false;
                    skip(j, 0) = false;
                }

                double Isum1 = 0, Isum2 = 0;
                for (int y = 0; y < roi.height; ++y)
                {
                    if (num_channels == 3)
                    {
                        const Vec<uchar, 3>* r1 = subimg1.ptr<Vec<uchar, 3> >(y);
                        const Vec<uchar, 3>* r2 = subimg2.ptr<Vec<uchar, 3> >(y);
                        for (int x = 0; x < roi.width; ++x)
                        {
                            if (intersect(y, x))
                            {
                                Isum1 += norm(r1[x]);
                                Isum2 += norm(r2[x]);
                            }
                        }
                    }
                    else // if (num_channels == 1)
                    {
                        const uchar* r1 = subimg1.ptr<uchar>(y);
                        const uchar* r2 = subimg2.ptr<uchar>(y);
                        for (int x = 0; x < roi.width; ++x)
                        {
                            if (intersect(y, x))
                            {
                                Isum1 += r1[x];
                                Isum2 += r2[x];
                            }
                        }
                    }
                }
                I(i, j) = Isum1 / N(i, j);
                I(j, i) = Isum2 / N(i, j);
            }
        }
    }
    if (getUpdateGain() || gains_.rows != num_images)
    {
        double alpha = 0.01;
        double beta = 100;
        int num_eq = num_images - countNonZero(skip);
        gains_.create(num_images, 1);
        gains_.setTo(1);

        // No image process, gains are all set to one, stop here
        if (num_eq == 0)
            return;

        Mat_<double> A(num_eq, num_eq); A.setTo(0);
        Mat_<double> b(num_eq, 1); b.setTo(0);
        for (int i = 0, ki = 0; i < num_images; ++i)
        {
            if (skip(i, 0))
                continue;

            for (int j = 0, kj = 0; j < num_images; ++j)
            {
                if (skip(j, 0))
                    continue;

                b(ki, 0) += beta * N(i, j);
                A(ki, ki) += beta * N(i, j);
                if (j != i)
                {
                    A(ki, ki) += 2 * alpha * I(i, j) * I(i, j) * N(i, j);
                    A(ki, kj) -= 2 * alpha * I(i, j) * I(j, i) * N(i, j);
                }
                ++kj;
            }
            ++ki;
        }

        Mat_<double> l_gains;

#ifdef HAVE_EIGEN
        Eigen::MatrixXf eigen_A, eigen_b, eigen_x;
        cv2eigen(A, eigen_A);
        cv2eigen(b, eigen_b);

        Eigen::LLT<Eigen::MatrixXf> solver(eigen_A);
#if ENABLE_LOG
        if (solver.info() != Eigen::ComputationInfo::Success)
            LOGLN("Failed to solve exposure compensation system");
#endif
        eigen_x = solver.solve(eigen_b);

        Mat_<float> l_gains_float;
        eigen2cv(eigen_x, l_gains_float);
        l_gains_float.convertTo(l_gains, CV_64FC1);
#else
        solve(A, b, l_gains);
#endif
        CV_CheckTypeEQ(l_gains.type(), CV_64FC1, "");

        for (int i = 0, j = 0; i < num_images; ++i)
        {
            // Only assign non-skipped gains. Other gains are already set to 1
            if (!skip(i, 0))
                gains_.at<double>(i, 0) = l_gains(j++, 0);
        }
    }
}


void GainCompensator::apply(int index, Point /*corner*/, InputOutputArray image, InputArray /*mask*/)
{
    CV_INSTRUMENT_REGION();

    multiply(image, gains_(index, 0), image);
}


std::vector<double> GainCompensator::gains() const
{
    std::vector<double> gains_vec(gains_.rows);
    for (int i = 0; i < gains_.rows; ++i)
        gains_vec[i] = gains_(i, 0);
    return gains_vec;
}

void GainCompensator::getMatGains(std::vector<Mat>& umv)
{
    umv.clear();
    for (int i = 0; i < gains_.rows; ++i)
        umv.push_back(Mat(1,1,CV_64FC1,Scalar(gains_(i, 0))));
}
void GainCompensator::setMatGains(std::vector<Mat>& umv)
{
    gains_=Mat_<double>(static_cast<int>(umv.size()),1);
    for (int i = 0; i < static_cast<int>(umv.size()); i++)
    {
        int type = umv[i].type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
        CV_CheckType(type, depth == CV_64F && cn == 1, "Only double images are supported for gain");
        CV_Assert(umv[i].rows == 1 && umv[i].cols == 1);
        gains_(i, 0) = umv[i].at<double>(0, 0);
    }
}

void GainCompensator::prepareSimilarityMask(
    const std::vector<Point> &corners, const std::vector<UMat> &images)
{
    if (similarity_threshold_ >= 1)
    {
        LOGLN("  skipping similarity mask: disabled");
        return;
    }
    if (!similarities_.empty())
    {
        LOGLN("  skipping similarity mask: already set");
        return;
    }

    LOGLN("  calculating similarity mask");
    const int num_images = static_cast<int>(images.size());
    for (int i = 0; i < num_images; ++i)
    {
        for (int j = i; j < num_images; ++j)
        {
            Rect roi;
            if (overlapRoi(corners[i], corners[j], images[i].size(), images[j].size(), roi))
            {
                UMat subimg1 = images[i](Rect(roi.tl() - corners[i], roi.br() - corners[i]));
                UMat subimg2 = images[j](Rect(roi.tl() - corners[j], roi.br() - corners[j]));
                UMat similarity = buildSimilarityMask(subimg1, subimg2);
                similarities_.push_back(similarity);
            }
        }
    }
}

UMat GainCompensator::buildSimilarityMask(InputArray src_array1, InputArray src_array2)
{
    CV_Assert(src_array1.rows() == src_array2.rows() && src_array1.cols() == src_array2.cols());
    CV_Assert(src_array1.type() == src_array2.type());
    CV_Assert(src_array1.type() == CV_8UC3 || src_array1.type() == CV_8UC1);

    Mat src1 = src_array1.getMat();
    Mat src2 = src_array2.getMat();

    UMat umat_similarity(src1.rows, src1.cols, CV_8UC1);
    Mat similarity = umat_similarity.getMat(ACCESS_WRITE);

    if (src1.channels() == 3)
    {
        for (int y = 0; y < similarity.rows; ++y)
        {
            for (int x = 0; x < similarity.cols; ++x)
            {
                Vec<float, 3> vec_diff =
                    Vec<float, 3>(*src1.ptr<Vec<uchar, 3>>(y, x))
                    - Vec<float, 3>(*src2.ptr<Vec<uchar, 3>>(y, x));
                double diff = norm(vec_diff * (1.f / 255.f));

                *similarity.ptr<uchar>(y, x) = diff <= similarity_threshold_ ? 255 : 0;
            }
        }
    }
    else // if (src1.channels() == 1)
    {
        for (int y = 0; y < similarity.rows; ++y)
        {
            for (int x = 0; x < similarity.cols; ++x)
            {
                float diff = std::abs(static_cast<int>(*src1.ptr<uchar>(y, x))
                    - static_cast<int>(*src2.ptr<uchar>(y, x))) / 255.f;

                *similarity.ptr<uchar>(y, x) = diff <= similarity_threshold_ ? 255 : 0;
            }
        }
    }
    similarity.release();

    Mat kernel = getStructuringElement(MORPH_RECT, Size(3,3));
    UMat umat_erode;
    erode(umat_similarity, umat_erode, kernel);
    dilate(umat_erode, umat_similarity, kernel);

    return umat_similarity;
}

void ChannelsCompensator::feed(const std::vector<Point> &corners, const std::vector<UMat> &images,
                               const std::vector<std::pair<UMat,uchar> > &masks)
{
    std::array<std::vector<UMat>, 3> images_channels;

    // Split channels of each input image
    for (const UMat& image: images)
    {
        std::vector<UMat> image_channels;
        image_channels.resize(3);
        split(image, image_channels);

        for (int i = 0; i < int(images_channels.size()); ++i)
            images_channels[i].emplace_back(std::move(image_channels[i]));
    }

    // For each channel, feed the channel of each image in a GainCompensator
    gains_.clear();
    gains_.resize(images.size());

    GainCompensator compensator(getNrFeeds());
    compensator.setSimilarityThreshold(getSimilarityThreshold());
    compensator.prepareSimilarityMask(corners, images);

    for (int c = 0; c < 3; ++c)
    {
        const std::vector<UMat>& channels = images_channels[c];

        compensator.feed(corners, channels, masks);

        std::vector<double> gains = compensator.gains();
        for (int i = 0; i < int(gains.size()); ++i)
            gains_.at(i)[c] = gains[i];
    }
}

void ChannelsCompensator::apply(int index, Point /*corner*/, InputOutputArray image, InputArray /*mask*/)
{
    CV_INSTRUMENT_REGION();

    multiply(image, gains_.at(index), image);
}

void ChannelsCompensator::getMatGains(std::vector<Mat>& umv)
{
    umv.clear();
    for (int i = 0; i < static_cast<int>(gains_.size()); ++i)
    {
        Mat m;
        Mat(gains_[i]).copyTo(m);
        umv.push_back(m);
    }
}

void ChannelsCompensator::setMatGains(std::vector<Mat>& umv)
{
    for (int i = 0; i < static_cast<int>(umv.size()); i++)
    {
        Scalar s;
        umv[i].copyTo(s);
        gains_.push_back(s);
    }
}


template<class Compensator>
void BlocksCompensator::feed(const std::vector<Point> &corners, const std::vector<UMat> &images,
                             const std::vector<std::pair<UMat,uchar> > &masks)
{
    CV_Assert(corners.size() == images.size() && images.size() == masks.size());

    const int num_images = static_cast<int>(images.size());

    std::vector<Size> bl_per_imgs(num_images);
    std::vector<Point> block_corners;
    std::vector<UMat> block_images;
    std::vector<std::pair<UMat,uchar> > block_masks;

    // Construct blocks for gain compensator
    for (int img_idx = 0; img_idx < num_images; ++img_idx)
    {
        Size bl_per_img((images[img_idx].cols + bl_width_ - 1) / bl_width_,
                        (images[img_idx].rows + bl_height_ - 1) / bl_height_);
        int bl_width = (images[img_idx].cols + bl_per_img.width - 1) / bl_per_img.width;
        int bl_height = (images[img_idx].rows + bl_per_img.height - 1) / bl_per_img.height;
        bl_per_imgs[img_idx] = bl_per_img;
        for (int by = 0; by < bl_per_img.height; ++by)
        {
            for (int bx = 0; bx < bl_per_img.width; ++bx)
            {
                Point bl_tl(bx * bl_width, by * bl_height);
                Point bl_br(std::min(bl_tl.x + bl_width, images[img_idx].cols),
                            std::min(bl_tl.y + bl_height, images[img_idx].rows));

                block_corners.push_back(corners[img_idx] + bl_tl);
                block_images.push_back(images[img_idx](Rect(bl_tl, bl_br)));
                block_masks.push_back(std::make_pair(masks[img_idx].first(Rect(bl_tl, bl_br)),
                                                masks[img_idx].second));
            }
        }
    }

    if (getUpdateGain() || int(gain_maps_.size()) != num_images)
    {
        Compensator compensator;
        compensator.setNrFeeds(getNrFeeds());
        compensator.setSimilarityThreshold(getSimilarityThreshold());
        compensator.feed(block_corners, block_images, block_masks);

        gain_maps_.clear();
        gain_maps_.resize(num_images);

        Mat_<float> ker(1, 3);
        ker(0, 0) = 0.25; ker(0, 1) = 0.5; ker(0, 2) = 0.25;

        int bl_idx = 0;
        for (int img_idx = 0; img_idx < num_images; ++img_idx)
        {
            Size bl_per_img = bl_per_imgs[img_idx];
            UMat gain_map = getGainMap(compensator, bl_idx, bl_per_img);
            bl_idx += bl_per_img.width*bl_per_img.height;

            for (int i=0; i<nr_gain_filtering_iterations_; ++i)
            {
                UMat tmp;
                sepFilter2D(gain_map, tmp, CV_32F, ker, ker);
                swap(gain_map, tmp);
            }

            gain_maps_[img_idx] = gain_map;
        }
    }
}

UMat BlocksCompensator::getGainMap(const GainCompensator& compensator, int bl_idx, Size bl_per_img)
{
    std::vector<double> gains = compensator.gains();

    UMat u_gain_map(bl_per_img, CV_32F);
    Mat_<float> gain_map = u_gain_map.getMat(ACCESS_WRITE);

    for (int by = 0; by < bl_per_img.height; ++by)
        for (int bx = 0; bx < bl_per_img.width; ++bx, ++bl_idx)
            gain_map(by, bx) = static_cast<float>(gains[bl_idx]);

    return u_gain_map;
}

UMat BlocksCompensator::getGainMap(const ChannelsCompensator& compensator, int bl_idx, Size bl_per_img)
{
    std::vector<Scalar> gains = compensator.gains();

    UMat u_gain_map(bl_per_img, CV_32FC3);
    Mat_<Vec3f> gain_map = u_gain_map.getMat(ACCESS_WRITE);

    for (int by = 0; by < bl_per_img.height; ++by)
        for (int bx = 0; bx < bl_per_img.width; ++bx, ++bl_idx)
            for (int c = 0; c < 3; ++c)
                gain_map(by, bx)[c] = static_cast<float>(gains[bl_idx][c]);

    return u_gain_map;
}

void BlocksCompensator::apply(int index, Point /*corner*/, InputOutputArray _image, InputArray /*mask*/)
{
    CV_INSTRUMENT_REGION();

    CV_Assert(_image.type() == CV_8UC3);

    UMat u_gain_map;
    if (gain_maps_.at(index).size() == _image.size())
        u_gain_map = gain_maps_.at(index);
    else
        resize(gain_maps_.at(index), u_gain_map, _image.size(), 0, 0, INTER_LINEAR);

    if (u_gain_map.channels() != 3)
    {
        std::vector<UMat> gains_channels;
        gains_channels.push_back(u_gain_map);
        gains_channels.push_back(u_gain_map);
        gains_channels.push_back(u_gain_map);
        merge(gains_channels, u_gain_map);
    }

    multiply(_image, u_gain_map, _image, 1, _image.type());
}

void BlocksCompensator::getMatGains(std::vector<Mat>& umv)
{
    umv.clear();
    for (int i = 0; i < static_cast<int>(gain_maps_.size()); ++i)
    {
        Mat m;
        gain_maps_[i].copyTo(m);
        umv.push_back(m);
    }
}

void BlocksCompensator::setMatGains(std::vector<Mat>& umv)
{
    for (int i = 0; i < static_cast<int>(umv.size()); i++)
    {
        UMat m;
        umv[i].copyTo(m);
        gain_maps_.push_back(m);
    }
}

void BlocksGainCompensator::feed(const std::vector<Point> &corners, const std::vector<UMat> &images,
                                 const std::vector<std::pair<UMat,uchar> > &masks)
{
    BlocksCompensator::feed<GainCompensator>(corners, images, masks);
}

void BlocksChannelsCompensator::feed(const std::vector<Point> &corners, const std::vector<UMat> &images,
                                     const std::vector<std::pair<UMat,uchar> > &masks)
{
    BlocksCompensator::feed<ChannelsCompensator>(corners, images, masks);
}


} // namespace detail
} // namespace cv