CDifodo.cpp

Go to the documentation of this file.

/* +------------------------------------------------------------------------+
   |                     Mobile Robot Programming Toolkit (MRPT)            |
   |                          https://www.mrpt.org/                         |
   |                                                                        |
   | Copyright (c) 2005-2019, Individual contributors, see AUTHORS file     |
   | See: https://www.mrpt.org/Authors - All rights reserved.               |
   | Released under BSD License. See: https://www.mrpt.org/License          |
   +------------------------------------------------------------------------+ */

#include "vision-precomp.h"  // Precompiled headers

#include <mrpt/core/round.h>
#include <mrpt/poses/Lie/SE.h>
#include <mrpt/system/CTicTac.h>
#include <mrpt/vision/CDifodo.h>
#include <Eigen/Dense>
#include <iostream>

using namespace mrpt;
using namespace mrpt::vision;
using namespace mrpt::math;
using namespace std;
using namespace Eigen;
using mrpt::round;
using mrpt::square;

CDifodo::CDifodo()
{
    rows = 60;
    cols = 80;
    fovh = M_PIf * 58.6f / 180.0f;
    fovv = M_PIf * 45.6f / 180.0f;
    cam_mode = 1;  // (1 - 640 x 480, 2 - 320 x 240, 4 - 160 x 120)
    downsample = 1;
    ctf_levels = 1;
    m_width = 640 / (cam_mode * downsample);
    m_height = 480 / (cam_mode * downsample);
    fast_pyramid = true;

    // Resize pyramid
    const unsigned int pyr_levels =
        round(log(float(m_width / cols)) / log(2.f)) + ctf_levels;
    depth.resize(pyr_levels);
    depth_old.resize(pyr_levels);
    depth_inter.resize(pyr_levels);
    depth_warped.resize(pyr_levels);
    xx.resize(pyr_levels);
    xx_inter.resize(pyr_levels);
    xx_old.resize(pyr_levels);
    xx_warped.resize(pyr_levels);
    yy.resize(pyr_levels);
    yy_inter.resize(pyr_levels);
    yy_old.resize(pyr_levels);
    yy_warped.resize(pyr_levels);
    transformations.resize(pyr_levels);

    for (unsigned int i = 0; i < pyr_levels; i++)
    {
        unsigned int s = pow(2.f, int(i));
        cols_i = m_width / s;
        rows_i = m_height / s;
        depth[i].resize(rows_i, cols_i);
        depth_inter[i].resize(rows_i, cols_i);
        depth_old[i].resize(rows_i, cols_i);
        depth[i].fill(0.0f);
        depth_old[i].fill(0.0f);
        xx[i].resize(rows_i, cols_i);
        xx_inter[i].resize(rows_i, cols_i);
        xx_old[i].resize(rows_i, cols_i);
        xx[i].fill(0.0f);
        xx_old[i].fill(0.0f);
        yy[i].resize(rows_i, cols_i);
        yy_inter[i].resize(rows_i, cols_i);
        yy_old[i].resize(rows_i, cols_i);
        yy[i].fill(0.0f);
        yy_old[i].fill(0.0f);
        transformations[i].resize(4, 4);

        if (cols_i <= cols)
        {
            depth_warped[i].resize(rows_i, cols_i);
            xx_warped[i].resize(rows_i, cols_i);
            yy_warped[i].resize(rows_i, cols_i);
        }
    }

    depth_wf.setSize(m_height, m_width);

    fps = 30.f;  // In Hz

    previous_speed_const_weight = 0.05f;
    previous_speed_eig_weight = 0.5f;
    kai_loc_old = TTwist3D();
    num_valid_points = 0;

    // Compute gaussian mask
    VectorXf v_mask(4);
    v_mask(0) = 1.f;
    v_mask(1) = 2.f;
    v_mask(2) = 2.f;
    v_mask(3) = 1.f;
    for (unsigned int i = 0; i < 4; i++)
        for (unsigned int j = 0; j < 4; j++)
            f_mask(i, j) = v_mask(i) * v_mask(j) / 36.f;

    // Compute gaussian mask
    float v_mask2[5] = {1, 4, 6, 4, 1};
    for (unsigned int i = 0; i < 5; i++)
        for (unsigned int j = 0; j < 5; j++)
            g_mask[i][j] = v_mask2[i] * v_mask2[j] / 256.f;
}

void CDifodo::buildCoordinatesPyramid()
{
    const float max_depth_dif = 0.1f;

    // Push coordinates back
    depth_old.swap(depth);
    xx_old.swap(xx);
    yy_old.swap(yy);

    // The number of levels of the pyramid does not match the number of levels
    // used
    // in the odometry computation (because we might want to finish with lower
    // resolutions)

    unsigned int pyr_levels =
        round(log(float(m_width / cols)) / log(2.f)) + ctf_levels;

    // Generate levels
    for (unsigned int i = 0; i < pyr_levels; i++)
    {
        unsigned int s = pow(2.f, int(i));
        cols_i = m_width / s;
        rows_i = m_height / s;
        const int rows_i2 = 2 * rows_i;
        const int cols_i2 = 2 * cols_i;
        const int i_1 = i - 1;

        if (i == 0) depth[i].swap(depth_wf);

        //                              Downsampling
        //-----------------------------------------------------------------------------
        else
        {
            for (unsigned int u = 0; u < cols_i; u++)
                for (unsigned int v = 0; v < rows_i; v++)
                {
                    const int u2 = 2 * u;
                    const int v2 = 2 * v;
                    const float dcenter = depth[i_1](v2, u2);

                    // Inner pixels
                    if ((v > 0) && (v < rows_i - 1) && (u > 0) &&
                        (u < cols_i - 1))
                    {
                        if (dcenter > 0.f)
                        {
                            float sum = 0.f;
                            float weight = 0.f;

                            for (int l = -2; l < 3; l++)
                                for (int k = -2; k < 3; k++)
                                {
                                    const float abs_dif = abs(
                                        depth[i_1](v2 + k, u2 + l) - dcenter);
                                    if (abs_dif < max_depth_dif)
                                    {
                                        const float aux_w =
                                            g_mask[2 + k][2 + l] *
                                            (max_depth_dif - abs_dif);
                                        weight += aux_w;
                                        sum +=
                                            aux_w * depth[i_1](v2 + k, u2 + l);
                                    }
                                }
                            depth[i](v, u) = sum / weight;
                        }
                        else
                        {
                            float min_depth = 10.f;
                            for (int l = -2; l < 3; l++)
                                for (int k = -2; k < 3; k++)
                                {
                                    const float d = depth[i_1](v2 + k, u2 + l);
                                    if ((d > 0.f) && (d < min_depth))
                                        min_depth = d;
                                }

                            if (min_depth < 10.f)
                                depth[i](v, u) = min_depth;
                            else
                                depth[i](v, u) = 0.f;
                        }
                    }

                    // Boundary
                    else
                    {
                        if (dcenter > 0.f)
                        {
                            float sum = 0.f;
                            float weight = 0.f;

                            for (int l = -2; l < 3; l++)
                                for (int k = -2; k < 3; k++)
                                {
                                    const int indv = v2 + k, indu = u2 + l;
                                    if ((indv >= 0) && (indv < rows_i2) &&
                                        (indu >= 0) && (indu < cols_i2))
                                    {
                                        const float abs_dif = abs(
                                            depth[i_1](indv, indu) - dcenter);
                                        if (abs_dif < max_depth_dif)
                                        {
                                            const float aux_w =
                                                g_mask[2 + k][2 + l] *
                                                (max_depth_dif - abs_dif);
                                            weight += aux_w;
                                            sum +=
                                                aux_w * depth[i_1](indv, indu);
                                        }
                                    }
                                }
                            depth[i](v, u) = sum / weight;
                        }
                        else
                        {
                            float min_depth = 10.f;
                            for (int l = -2; l < 3; l++)
                                for (int k = -2; k < 3; k++)
                                {
                                    const int indv = v2 + k, indu = u2 + l;
                                    if ((indv >= 0) && (indv < rows_i2) &&
                                        (indu >= 0) && (indu < cols_i2))
                                    {
                                        const float d = depth[i_1](indv, indu);
                                        if ((d > 0.f) && (d < min_depth))
                                            min_depth = d;
                                    }
                                }

                            if (min_depth < 10.f)
                                depth[i](v, u) = min_depth;
                            else
                                depth[i](v, u) = 0.f;
                        }
                    }
                }
        }

        // Calculate coordinates "xy" of the points
        const float inv_f_i = 2.f * tan(0.5f * fovh) / float(cols_i);
        const float disp_u_i = 0.5f * (cols_i - 1);
        const float disp_v_i = 0.5f * (rows_i - 1);

        for (unsigned int u = 0; u < cols_i; u++)
            for (unsigned int v = 0; v < rows_i; v++)
                if (depth[i](v, u) > 0.f)
                {
                    xx[i](v, u) = (u - disp_u_i) * depth[i](v, u) * inv_f_i;
                    yy[i](v, u) = (v - disp_v_i) * depth[i](v, u) * inv_f_i;
                }
                else
                {
                    xx[i](v, u) = 0.f;
                    yy[i](v, u) = 0.f;
                }
    }
}

void CDifodo::buildCoordinatesPyramidFast()
{
    const float max_depth_dif = 0.1f;

    // Push coordinates back
    depth_old.swap(depth);
    xx_old.swap(xx);
    yy_old.swap(yy);

    // The number of levels of the pyramid does not match the number of levels
    // used
    // in the odometry computation (because we might want to finish with lower
    // resolutions)

    unsigned int pyr_levels =
        round(log(float(m_width / cols)) / log(2.f)) + ctf_levels;

    // Generate levels
    for (unsigned int i = 0; i < pyr_levels; i++)
    {
        unsigned int s = pow(2.f, int(i));
        cols_i = m_width / s;
        rows_i = m_height / s;
        // const int rows_i2 = 2*rows_i;
        // const int cols_i2 = 2*cols_i;
        const int i_1 = i - 1;

        if (i == 0) depth[i].swap(depth_wf);

        //                              Downsampling
        //-----------------------------------------------------------------------------
        else
        {
            for (unsigned int u = 0; u < cols_i; u++)
                for (unsigned int v = 0; v < rows_i; v++)
                {
                    const int u2 = 2 * u;
                    const int v2 = 2 * v;

                    // Inner pixels
                    if ((v > 0) && (v < rows_i - 1) && (u > 0) &&
                        (u < cols_i - 1))
                    {
                        const Matrix4f d_block =
                            depth[i_1].block<4, 4>(v2 - 1, u2 - 1);
                        float depths[4] = {d_block(5), d_block(6), d_block(9),
                                           d_block(10)};
                        float dcenter;

                        // Sort the array (try to find a good/representative
                        // value)
                        for (signed char k = 2; k >= 0; k--)
                            if (depths[k + 1] < depths[k])
                                std::swap(depths[k + 1], depths[k]);
                        for (unsigned char k = 1; k < 3; k++)
                            if (depths[k] > depths[k + 1])
                                std::swap(depths[k + 1], depths[k]);
                        if (depths[2] < depths[1])
                            dcenter = depths[1];
                        else
                            dcenter = depths[2];

                        if (dcenter > 0.f)
                        {
                            float sum = 0.f;
                            float weight = 0.f;

                            for (unsigned char k = 0; k < 16; k++)
                            {
                                const float abs_dif = abs(d_block(k) - dcenter);
                                if (abs_dif < max_depth_dif)
                                {
                                    const float aux_w =
                                        f_mask[k] * (max_depth_dif - abs_dif);
                                    weight += aux_w;
                                    sum += aux_w * d_block(k);
                                }
                            }
                            if (weight > 0) depth[i](v, u) = sum / weight;
                        }
                        else
                            depth[i](v, u) = 0.f;
                    }

                    // Boundary
                    else
                    {
                        const Matrix2f d_block = depth[i_1].block<2, 2>(v2, u2);
                        const float new_d = 0.25f * d_block.array().sum();
                        if (new_d < 0.4f)
                            depth[i](v, u) = 0.f;
                        else
                            depth[i](v, u) = new_d;
                    }
                }
        }

        // Calculate coordinates "xy" of the points
        const float inv_f_i = 2.f * tan(0.5f * fovh) / float(cols_i);
        const float disp_u_i = 0.5f * (cols_i - 1);
        const float disp_v_i = 0.5f * (rows_i - 1);

        for (unsigned int u = 0; u < cols_i; u++)
            for (unsigned int v = 0; v < rows_i; v++)
                if (depth[i](v, u) > 0.f)
                {
                    xx[i](v, u) = (u - disp_u_i) * depth[i](v, u) * inv_f_i;
                    yy[i](v, u) = (v - disp_v_i) * depth[i](v, u) * inv_f_i;
                }
                else
                {
                    xx[i](v, u) = 0.f;
                    yy[i](v, u) = 0.f;
                }
    }
}

void CDifodo::performWarping()
{
    // Camera parameters (which also depend on the level resolution)
    const float f = float(cols_i) / (2.f * tan(0.5f * fovh));
    const float disp_u_i = 0.5f * float(cols_i - 1);
    const float disp_v_i = 0.5f * float(rows_i - 1);

    // Rigid transformation estimated up to the present level
    Matrix4f acu_trans;
    acu_trans.setIdentity();
    for (unsigned int i = 1; i <= level; i++)
        acu_trans = transformations[i - 1].asEigen() * acu_trans;

    MatrixXf wacu(rows_i, cols_i);
    wacu.fill(0.f);
    depth_warped[image_level].fill(0.f);

    const auto cols_lim = float(cols_i - 1);
    const auto rows_lim = float(rows_i - 1);

    //                      Warping loop
    //---------------------------------------------------------
    for (unsigned int j = 0; j < cols_i; j++)
        for (unsigned int i = 0; i < rows_i; i++)
        {
            const float z = depth[image_level](i, j);

            if (z > 0.f)
            {
                // Transform point to the warped reference frame
                const float depth_w = acu_trans(0, 0) * z +
                                      acu_trans(0, 1) * xx[image_level](i, j) +
                                      acu_trans(0, 2) * yy[image_level](i, j) +
                                      acu_trans(0, 3);
                const float x_w = acu_trans(1, 0) * z +
                                  acu_trans(1, 1) * xx[image_level](i, j) +
                                  acu_trans(1, 2) * yy[image_level](i, j) +
                                  acu_trans(1, 3);
                const float y_w = acu_trans(2, 0) * z +
                                  acu_trans(2, 1) * xx[image_level](i, j) +
                                  acu_trans(2, 2) * yy[image_level](i, j) +
                                  acu_trans(2, 3);

                // Calculate warping
                const float uwarp = f * x_w / depth_w + disp_u_i;
                const float vwarp = f * y_w / depth_w + disp_v_i;

                // The warped pixel (which is not integer in general)
                // contributes to all the surrounding ones
                if ((uwarp >= 0.f) && (uwarp < cols_lim) && (vwarp >= 0.f) &&
                    (vwarp < rows_lim))
                {
                    const int uwarp_l = uwarp;
                    const int uwarp_r = uwarp_l + 1;
                    const int vwarp_d = vwarp;
                    const int vwarp_u = vwarp_d + 1;
                    const float delta_r = float(uwarp_r) - uwarp;
                    const float delta_l = uwarp - float(uwarp_l);
                    const float delta_u = float(vwarp_u) - vwarp;
                    const float delta_d = vwarp - float(vwarp_d);

                    // Warped pixel very close to an integer value
                    if (abs(round(uwarp) - uwarp) + abs(round(vwarp) - vwarp) <
                        0.05f)
                    {
                        depth_warped[image_level](round(vwarp), round(uwarp)) +=
                            depth_w;
                        wacu(round(vwarp), round(uwarp)) += 1.f;
                    }
                    else
                    {
                        const float w_ur = square(delta_l) + square(delta_d);
                        depth_warped[image_level](vwarp_u, uwarp_r) +=
                            w_ur * depth_w;
                        wacu(vwarp_u, uwarp_r) += w_ur;

                        const float w_ul = square(delta_r) + square(delta_d);
                        depth_warped[image_level](vwarp_u, uwarp_l) +=
                            w_ul * depth_w;
                        wacu(vwarp_u, uwarp_l) += w_ul;

                        const float w_dr = square(delta_l) + square(delta_u);
                        depth_warped[image_level](vwarp_d, uwarp_r) +=
                            w_dr * depth_w;
                        wacu(vwarp_d, uwarp_r) += w_dr;

                        const float w_dl = square(delta_r) + square(delta_u);
                        depth_warped[image_level](vwarp_d, uwarp_l) +=
                            w_dl * depth_w;
                        wacu(vwarp_d, uwarp_l) += w_dl;
                    }
                }
            }
        }

    // Scale the averaged depth and compute spatial coordinates
    const float inv_f_i = 1.f / f;
    for (unsigned int u = 0; u < cols_i; u++)
        for (unsigned int v = 0; v < rows_i; v++)
        {
            if (wacu(v, u) > 0.f)
            {
                depth_warped[image_level](v, u) /= wacu(v, u);
                xx_warped[image_level](v, u) =
                    (u - disp_u_i) * depth_warped[image_level](v, u) * inv_f_i;
                yy_warped[image_level](v, u) =
                    (v - disp_v_i) * depth_warped[image_level](v, u) * inv_f_i;
            }
            else
            {
                depth_warped[image_level](v, u) = 0.f;
                xx_warped[image_level](v, u) = 0.f;
                yy_warped[image_level](v, u) = 0.f;
            }
        }
}

void CDifodo::calculateCoord()
{
    null.resize(rows_i, cols_i);
    null.fill(false);
    num_valid_points = 0;

    for (unsigned int u = 0; u < cols_i; u++)
        for (unsigned int v = 0; v < rows_i; v++)
        {
            if ((depth_old[image_level](v, u)) == 0.f ||
                (depth_warped[image_level](v, u) == 0.f))
            {
                depth_inter[image_level](v, u) = 0.f;
                xx_inter[image_level](v, u) = 0.f;
                yy_inter[image_level](v, u) = 0.f;
                null(v, u) = true;
            }
            else
            {
                depth_inter[image_level](v, u) =
                    0.5f * (depth_old[image_level](v, u) +
                            depth_warped[image_level](v, u));
                xx_inter[image_level](v, u) =
                    0.5f *
                    (xx_old[image_level](v, u) + xx_warped[image_level](v, u));
                yy_inter[image_level](v, u) =
                    0.5f *
                    (yy_old[image_level](v, u) + yy_warped[image_level](v, u));
                null(v, u) = false;
                if ((u > 0) && (v > 0) && (u < cols_i - 1) && (v < rows_i - 1))
                    num_valid_points++;
            }
        }
}

void CDifodo::calculateDepthDerivatives()
{
    dt.resize(rows_i, cols_i);
    dt.fill(0.f);
    du.resize(rows_i, cols_i);
    du.fill(0.f);
    dv.resize(rows_i, cols_i);
    dv.fill(0.f);

    // Compute connectivity
    MatrixXf rx_ninv(rows_i, cols_i);
    MatrixXf ry_ninv(rows_i, cols_i);
    rx_ninv.fill(1.f);
    ry_ninv.fill(1.f);

    for (unsigned int u = 0; u < cols_i - 1; u++)
        for (unsigned int v = 0; v < rows_i; v++)
            if (null(v, u) == false)
            {
                rx_ninv(v, u) = sqrtf(
                    square(
                        xx_inter[image_level](v, u + 1) -
                        xx_inter[image_level](v, u)) +
                    square(
                        depth_inter[image_level](v, u + 1) -
                        depth_inter[image_level](v, u)));
            }

    for (unsigned int u = 0; u < cols_i; u++)
        for (unsigned int v = 0; v < rows_i - 1; v++)
            if (null(v, u) == false)
            {
                ry_ninv(v, u) = sqrtf(
                    square(
                        yy_inter[image_level](v + 1, u) -
                        yy_inter[image_level](v, u)) +
                    square(
                        depth_inter[image_level](v + 1, u) -
                        depth_inter[image_level](v, u)));
            }

    // Spatial derivatives
    for (unsigned int v = 0; v < rows_i; v++)
    {
        for (unsigned int u = 1; u < cols_i - 1; u++)
            if (null(v, u) == false)
                du(v, u) =
                    (rx_ninv(v, u - 1) * (depth_inter[image_level](v, u + 1) -
                                          depth_inter[image_level](v, u)) +
                     rx_ninv(v, u) * (depth_inter[image_level](v, u) -
                                      depth_inter[image_level](v, u - 1))) /
                    (rx_ninv(v, u) + rx_ninv(v, u - 1));

        du(v, 0) = du(v, 1);
        du(v, cols_i - 1) = du(v, cols_i - 2);
    }

    for (unsigned int u = 0; u < cols_i; u++)
    {
        for (unsigned int v = 1; v < rows_i - 1; v++)
            if (null(v, u) == false)
                dv(v, u) =
                    (ry_ninv(v - 1, u) * (depth_inter[image_level](v + 1, u) -
                                          depth_inter[image_level](v, u)) +
                     ry_ninv(v, u) * (depth_inter[image_level](v, u) -
                                      depth_inter[image_level](v - 1, u))) /
                    (ry_ninv(v, u) + ry_ninv(v - 1, u));

        dv(0, u) = dv(1, u);
        dv(rows_i - 1, u) = dv(rows_i - 2, u);
    }

    // Temporal derivative
    for (unsigned int u = 0; u < cols_i; u++)
        for (unsigned int v = 0; v < rows_i; v++)
            if (null(v, u) == false)
                dt(v, u) = fps * (depth_warped[image_level](v, u) -
                                  depth_old[image_level](v, u));
}

void CDifodo::computeWeights()
{
    weights.resize(rows_i, cols_i);
    weights.fill(0.f);

    // Obtain the velocity associated to the rigid transformation estimated up
    // to the present level
    CVectorFixedFloat<6> kai_level;
    kai_level.setFromMatrixLike(kai_loc_old);

    Matrix4f acu_trans;
    acu_trans.setIdentity();
    for (unsigned int i = 0; i < level; i++)
        acu_trans = transformations[i].asEigen() * acu_trans;

    // Alternative way to compute the log
    auto mat_aux = CMatrixDouble44(acu_trans.cast<double>().eval());

    auto trans_vec = poses::Lie::SE<3>::log(poses::CPose3D(mat_aux));
    trans_vec *= fps;
    const auto kai_level_acu = CVectorFixedDouble<6>(trans_vec);

    kai_level -= kai_level_acu.cast_float();

    // Parameters for the measurement error
    const float f_inv = float(cols_i) / (2.f * tan(0.5f * fovh));
    const float kz2 = 8.122e-12f;  // square(1.425e-5) / 25

    // Parameters for linearization error
    const float kduv = 20e-5f;
    const float kdt = kduv / square(fps);
    const float k2dt = 5e-6f;
    const float k2duv = 5e-6f;

    for (unsigned int u = 1; u < cols_i - 1; u++)
        for (unsigned int v = 1; v < rows_i - 1; v++)
            if (null(v, u) == false)
            {
                //                  Compute measurment error (simplified)
                //-----------------------------------------------------------------------
                const float z = depth_inter[image_level](v, u);
                const float inv_d = 1.f / z;
                // const float dycomp = du2(v,u)*f_inv_y*inv_d;
                // const float dzcomp = dv2(v,u)*f_inv_z*inv_d;
                const float z2 = z * z;
                const float z4 = z2 * z2;

                // const float var11 = kz2*z4;
                // const float var12 =
                // kz2*xx_inter[image_level](v,u)*z2*depth_inter[image_level](v,u);
                // const float var13 =
                // kz2*yy_inter[image_level](v,u)*z2*depth_inter[image_level](v,u);
                // const float var22 =
                // kz2*square(xx_inter[image_level](v,u))*z2;
                // const float var23 =
                // kz2*xx_inter[image_level](v,u)*yy_inter[image_level](v,u)*z2;
                // const float var33 =
                // kz2*square(yy_inter[image_level](v,u))*z2;
                const float var44 = kz2 * z4 * square(fps);
                const float var55 = kz2 * z4 * 0.25f;
                const float var66 = var55;

                // const float j1 =
                // -2.f*inv_d*inv_d*(xx_inter[image_level](v,u)*dycomp +
                // yy_inter[image_level](v,u)*dzcomp)*(kai_level[0] +
                // yy_inter[image_level](v,u)*kai_level[4] -
                // xx_inter[image_level](v,u)*kai_level[5])
                //              + inv_d*dycomp*(kai_level[1] -
                // yy_inter[image_level](v,u)*kai_level[3]) +
                // inv_d*dzcomp*(kai_level[2] +
                // xx_inter[image_level](v,u)*kai_level[3]);
                // const float j2 = inv_d*dycomp*(kai_level[0] +
                // yy_inter[image_level](v,u)*kai_level[4] -
                // 2.f*xx_inter[image_level](v,u)*kai_level[5]) -
                // dzcomp*kai_level[3];
                // const float j3 = inv_d*dzcomp*(kai_level[0] +
                // 2.f*yy_inter[image_level](v,u)*kai_level[4] -
                // xx_inter[image_level](v,u)*kai_level[5]) +
                // dycomp*kai_level[3];

                const float j4 = 1.f;
                const float j5 =
                    xx_inter[image_level](v, u) * inv_d * inv_d * f_inv *
                        (kai_level[0] +
                         yy_inter[image_level](v, u) * kai_level[4] -
                         xx_inter[image_level](v, u) * kai_level[5]) +
                    inv_d * f_inv *
                        (-kai_level[1] - z * kai_level[5] +
                         yy_inter[image_level](v, u) * kai_level[3]);
                const float j6 =
                    yy_inter[image_level](v, u) * inv_d * inv_d * f_inv *
                        (kai_level[0] +
                         yy_inter[image_level](v, u) * kai_level[4] -
                         xx_inter[image_level](v, u) * kai_level[5]) +
                    inv_d * f_inv *
                        (-kai_level[2] + z * kai_level[4] -
                         xx_inter[image_level](v, u) * kai_level[3]);

                // error_measurement(v,u) = j1*(j1*var11+j2*var12+j3*var13) +
                // j2*(j1*var12+j2*var22+j3*var23)
                //                      +j3*(j1*var13+j2*var23+j3*var33) +
                // j4*j4*var44
                //+
                // j5*j5*var55 + j6*j6*var66;

                const float error_m =
                    j4 * j4 * var44 + j5 * j5 * var55 + j6 * j6 * var66;

                //                  Compute linearization error
                //-----------------------------------------------------------------------
                const float ini_du = depth_old[image_level](v, u + 1) -
                                     depth_old[image_level](v, u - 1);
                const float ini_dv = depth_old[image_level](v + 1, u) -
                                     depth_old[image_level](v - 1, u);
                const float final_du = depth_warped[image_level](v, u + 1) -
                                       depth_warped[image_level](v, u - 1);
                const float final_dv = depth_warped[image_level](v + 1, u) -
                                       depth_warped[image_level](v - 1, u);

                const float dut = ini_du - final_du;
                const float dvt = ini_dv - final_dv;
                const float duu = du(v, u + 1) - du(v, u - 1);
                const float dvv = dv(v + 1, u) - dv(v - 1, u);
                const float dvu =
                    dv(v, u + 1) -
                    dv(v, u - 1);  // Completely equivalent to compute duv

                const float error_l =
                    kdt * square(dt(v, u)) +
                    kduv * (square(du(v, u)) + square(dv(v, u))) +
                    k2dt * (square(dut) + square(dvt)) +
                    k2duv * (square(duu) + square(dvv) + square(dvu));

                // Weight
                weights(v, u) = sqrt(1.f / (error_m + error_l));
            }

    // Normalize weights in the range [0,1]
    const float inv_max = 1.f / weights.maxCoeff();
    weights *= inv_max;
}

void CDifodo::solveOneLevel()
{
    MatrixXf A(num_valid_points, 6);
    MatrixXf B(num_valid_points, 1);
    unsigned int cont = 0;

    // Fill the matrix A and the vector B
    // The order of the unknowns is (vz, vx, vy, wz, wx, wy)
    // The points order will be (1,1), (1,2)...(1,cols-1), (2,1),
    // (2,2)...(row-1,cols-1).

    const float f_inv = float(cols_i) / (2.f * tan(0.5f * fovh));

    for (unsigned int u = 1; u < cols_i - 1; u++)
        for (unsigned int v = 1; v < rows_i - 1; v++)
            if (null(v, u) == false)
            {
                // Precomputed expressions
                const float d = depth_inter[image_level](v, u);
                const float inv_d = 1.f / d;
                const float x = xx_inter[image_level](v, u);
                const float y = yy_inter[image_level](v, u);
                const float dycomp = du(v, u) * f_inv * inv_d;
                const float dzcomp = dv(v, u) * f_inv * inv_d;
                const float tw = weights(v, u);

                // Fill the matrix A
                A(cont, 0) =
                    tw * (1.f + dycomp * x * inv_d + dzcomp * y * inv_d);
                A(cont, 1) = tw * (-dycomp);
                A(cont, 2) = tw * (-dzcomp);
                A(cont, 3) = tw * (dycomp * y - dzcomp * x);
                A(cont, 4) = tw * (y + dycomp * inv_d * y * x +
                                   dzcomp * (y * y * inv_d + d));
                A(cont, 5) = tw * (-x - dycomp * (x * x * inv_d + d) -
                                   dzcomp * inv_d * y * x);
                B(cont, 0) = tw * (-dt(v, u));

                cont++;
            }

    // Solve the linear system of equations using weighted least squares
    const MatrixXf AtA = A.transpose() * A;
    const MatrixXf AtB = A.transpose() * B;
    VectorXf Var = AtA.ldlt().solve(AtB);

    // Covariance matrix calculation
    MatrixXf res = -B;
    for (unsigned int k = 0; k < 6; k++) res += Var(k) * A.col(k);

    est_cov =
        (1.f / float(num_valid_points - 6)) * AtA.inverse() * res.squaredNorm();

    // Update last velocity in local coordinates
    // (vx, vy, vz, wx, wy, wz)
    kai_loc_level.fromVector(Var);
}

void CDifodo::odometryCalculation()
{
    // Clock to measure the runtime
    mrpt::system::CTicTac clock;
    clock.Tic();

    // Build the gaussian pyramid
    if (fast_pyramid)
        buildCoordinatesPyramidFast();
    else
        buildCoordinatesPyramid();

    // Coarse-to-fines scheme
    for (unsigned int i = 0; i < ctf_levels; i++)
    {
        // Previous computations
        transformations[i].setIdentity();

        level = i;
        unsigned int s = pow(2.f, int(ctf_levels - (i + 1)));
        cols_i = cols / s;
        rows_i = rows / s;
        image_level =
            ctf_levels - i + round(log(float(m_width / cols)) / log(2.f)) - 1;

        // 1. Perform warping
        if (i == 0)
        {
            depth_warped[image_level] = depth[image_level];
            xx_warped[image_level] = xx[image_level];
            yy_warped[image_level] = yy[image_level];
        }
        else
            performWarping();

        // 2. Calculate inter coords and find null measurements
        calculateCoord();

        // 3. Compute derivatives
        calculateDepthDerivatives();

        // 4. Compute weights
        computeWeights();

        // 5. Solve odometry
        if (num_valid_points > 6) solveOneLevel();

        // 6. Filter solution
        filterLevelSolution();
    }

    // Update poses
    poseUpdate();

    // Save runtime
    execution_time = 1000.f * clock.Tac();
}

void CDifodo::filterLevelSolution()
{
    //      Calculate Eigenvalues and Eigenvectors
    //----------------------------------------------------------
    CMatrixFloat66 Bii;
    std::vector<float> eigenVals;

    if (est_cov.eig_symmetric(Bii, eigenVals))
    {
        std::cerr
            << "\n Eigensolver couldn't find a solution. Pose is not updated\n";
        return;
    }

    // First, we have to describe both the new linear and angular velocities in
    // the "eigenvector" basis
    //-------------------------------------------------------------------------------------------------
    auto kai_b = CVectorFixedFloat<6>(Bii.asEigen().colPivHouseholderQr().solve(
        kai_loc_level.asVector<Matrix<float, 6, 1>>()));

    // Second, we have to describe both the old linear and angular velocities in
    // the "eigenvector" basis
    //-------------------------------------------------------------------------------------------------
    auto kai_loc_sub = kai_loc_old.asVector<Eigen::Matrix<float, 6, 1>>();

    // Important: we have to substract the previous levels' solutions from the
    // old velocity.
    Matrix4f acu_trans;
    acu_trans.setIdentity();
    for (unsigned int i = 0; i < level; i++)
        acu_trans = transformations[i].asEigen() * acu_trans;

    auto mat_aux = CMatrixDouble44(acu_trans.cast<double>());
    auto acu_trans_vec = poses::Lie::SE<3>::log(poses::CPose3D(mat_aux));
    acu_trans_vec *= fps;
    const auto kai_level_acu = CVectorFixedDouble<6>(acu_trans_vec);

    kai_loc_sub -= kai_level_acu.asEigen().cast<float>();

    // Matrix<float, 4, 4> log_trans = fps*acu_trans.log();
    // kai_loc_sub(0) -= log_trans(0,3); kai_loc_sub(1) -= log_trans(1,3);
    // kai_loc_sub(2) -= log_trans(2,3);
    // kai_loc_sub(3) += log_trans(1,2); kai_loc_sub(4) -= log_trans(0,2);
    // kai_loc_sub(5) += log_trans(0,1);

    // Transform that local representation to the "eigenvector" basis
    const Matrix<float, 6, 1> kai_b_old =
        Bii.asEigen().colPivHouseholderQr().solve(kai_loc_sub);

    //                                  Filter velocity
    //--------------------------------------------------------------------------------
    const float cf = previous_speed_eig_weight * expf(-int(level)),
                df = previous_speed_const_weight * expf(-int(level));
    Matrix<float, 6, 1> kai_b_fil;
    for (unsigned int i = 0; i < 6; i++)
        kai_b_fil(i) =
            (kai_b.asEigen()(i) + (cf * eigenVals[i] + df) * kai_b_old(i)) /
            (1.f + cf * eigenVals[i] + df);

    // Transform filtered velocity to the local reference frame
    Matrix<float, 6, 1> kai_loc_fil =
        Bii.asEigen().inverse().colPivHouseholderQr().solve(kai_b_fil);

    // Compute the rigid transformation
    auto aux_vel =
        mrpt::math::CVectorFixedDouble<6>(kai_loc_fil.cast<double>() / fps);
    const poses::CPose3D aux2 = mrpt::poses::Lie::SE<3>::exp(aux_vel);

    CMatrixDouble44 trans;
    aux2.getHomogeneousMatrix(trans);
    transformations[level] = trans.cast_float();
}

void CDifodo::poseUpdate()
{
    // First, compute the overall transformation
    //---------------------------------------------------
    Matrix4f acu_trans;
    acu_trans.setIdentity();
    for (unsigned int i = 1; i <= ctf_levels; i++)
        acu_trans = transformations[i - 1].asEigen() * acu_trans;

    // Compute the new estimates in the local and absolutes reference frames
    //---------------------------------------------------------------------
    auto mat_aux = CMatrixDouble44(acu_trans.cast<double>().eval());
    auto acu_trans_vec = poses::Lie::SE<3>::log(poses::CPose3D(mat_aux));
    acu_trans_vec *= fps;
    const CVectorFixedDouble<6> kai_level_acu(acu_trans_vec);
    kai_loc.fromVector(kai_level_acu.cast_float());

    //---------------------------------------------------------------------------------------------
    // Directly from Eigen:
    //- Eigen 3.1.0 needed for Matrix::log()
    //- The line "#include <unsupported/Eigen/MatrixFunctions>" should be
    // uncommented (CDifodo.h)
    //
    // Matrix<float, 4, 4> log_trans = fps*acu_trans.log();
    // kai_loc(0) = log_trans(0,3); kai_loc(1) = log_trans(1,3); kai_loc(2) =
    // log_trans(2,3);
    // kai_loc(3) = -log_trans(1,2); kai_loc(4) = log_trans(0,2); kai_loc(5) =
    // -log_trans(0,1);
    //---------------------------------------------------------------------------------------------

    CMatrixDouble33 inv_trans;

    cam_pose.getRotationMatrix(inv_trans);
    const auto v_abs =
        (inv_trans.asEigen() *
         (kai_loc.asVector<Eigen::Matrix<double, 6, 1>>().topRows(3)))
            .eval();
    const auto w_abs =
        (inv_trans.asEigen() *
         (kai_loc.asVector<Eigen::Matrix<double, 6, 1>>().bottomRows(3)))
            .eval();
    kai_abs.vx = v_abs.x();
    kai_abs.vy = v_abs.y();
    kai_abs.vz = v_abs.z();

    kai_abs.wx = w_abs.x();
    kai_abs.wy = w_abs.y();
    kai_abs.wz = w_abs.z();

    //                      Update poses
    //-------------------------------------------------------
    cam_oldpose = cam_pose;
    const auto pose_aux = poses::CPose3D(CMatrixDouble44(acu_trans));
    cam_pose = cam_pose + pose_aux;

    // Compute the velocity estimate in the new ref frame (to be used by the
    // filter in the next iteration)
    //---------------------------------------------------------------------------------------------------
    cam_pose.getRotationMatrix(inv_trans);
    const auto old_vtrans =
        (inv_trans.asEigen().inverse() *
         (kai_abs.asVector<Eigen::Matrix<double, 6, 1>>().topRows(3)))
            .eval();
    const auto old_w =
        (inv_trans.asEigen().inverse() *
         (kai_abs.asVector<Eigen::Matrix<double, 6, 1>>().bottomRows(3)))
            .eval();

    kai_loc_old.vx = old_vtrans.x();
    kai_loc_old.vy = old_vtrans.y();
    kai_loc_old.vz = old_vtrans.z();

    kai_loc_old.wx = old_w.x();
    kai_loc_old.wy = old_w.y();
    kai_loc_old.wz = old_w.z();
}

void CDifodo::setFOV(float new_fovh, float new_fovv)
{
    fovh = M_PI * new_fovh / 180.0;
    fovv = M_PI * new_fovv / 180.0;
}

void CDifodo::getPointsCoord(CMatrixFloat& x, CMatrixFloat& y, CMatrixFloat& z)
{
    x.resize(rows, cols);
    y.resize(rows, cols);
    z.resize(rows, cols);

    z = depth_inter[0];
    x = xx_inter[0];
    y = yy_inter[0];
}

void CDifodo::getDepthDerivatives(
    CMatrixFloat& cur_du, CMatrixFloat& cur_dv, CMatrixFloat& cur_dt)
{
    cur_du.resize(rows, cols);
    cur_dv.resize(rows, cols);
    cur_dt.resize(rows, cols);

    cur_du = du;
    cur_dv = dv;
    cur_dt = dt;
}

void CDifodo::getWeights(CMatrixFloat& w)
{
    w.resize(rows, cols);
    w = weights;
}