ba_common.cpp

Go to the documentation of this file.

/**
 * Copyright (C) 2010  Hauke Strasdat
 *                     Imperial College London
 * Copyright (c) 2005-2019, Individual contributors, see AUTHORS file
 *   See: https://www.mrpt.org/Authors - All rights reserved.
 *
 * bundle_adjuster.h is part of RobotVision.
 *
 * RobotVision is free software: you can redistribute it and/or modify
 * it under the terms of the GNU Lesser General Public License as
 * published by the Free Software Foundation, either version 3 of the
 * License, or any later version.
 *
 * RobotVision is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
 * GNU Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * and the GNU Lesser General Public License along with this program.
 * If not, see <http://www.gnu.org/licenses/>.
 */

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

#include <mrpt/math/robust_kernels.h>
#include <mrpt/poses/Lie/SE.h>
#include <mrpt/vision/bundle_adjustment.h>
#include <mrpt/vision/pinhole.h>
#include "ba_internals.h"

using namespace std;
using namespace mrpt;
using namespace mrpt::vision;
using namespace mrpt::img;
using namespace mrpt::math;
using namespace mrpt::poses;

/* -------------------------------------------------------------------------------------
                            ba_initial_estimate

 Fills the frames & landmark points maps with an initial gross estimate
 from the sequence \a observations, so they can be fed to bundle adjustment
 methods.
   ------------------------------------------------------------------------------------
 */
void mrpt::vision::ba_initial_estimate(
    const TSequenceFeatureObservations& observations,
    const TCamera& camera_params, TFramePosesMap& frame_poses,
    TLandmarkLocationsMap& landmark_points)
{
    MRPT_UNUSED_PARAM(camera_params);
    MRPT_START
    // VERY CRUDE APPROACH: All camera poses at the origin, all points at
    // (0,0,1)
    // TODO: Improve with the data in "observations"...

    // Clear previous contents:
    frame_poses.clear();
    landmark_points.clear();

    // Go thru the obs list and insert poses:
    for (TSequenceFeatureObservations::const_iterator it = observations.begin();
         it != observations.end(); ++it)
    {
        const TFeatureID feat_ID = it->id_feature;
        const TCameraPoseID frame_ID = it->id_frame;

        frame_poses[frame_ID] = CPose3D(0, 0, 0, 0, 0, 0);
        landmark_points[feat_ID] = TPoint3D(0, 0, 1);
    }
    MRPT_END
}

void mrpt::vision::ba_initial_estimate(
    const TSequenceFeatureObservations& observations,
    const TCamera& camera_params, TFramePosesVec& frame_poses,
    TLandmarkLocationsVec& landmark_points)
{
    MRPT_UNUSED_PARAM(camera_params);
    MRPT_START
    // VERY CRUDE APPROACH: All camera poses at the origin, all points at
    // (0,0,1)
    // TODO: Improve with the data in "observations"...

    // Clear previous contents:
    frame_poses.clear();
    landmark_points.clear();

    if (observations.empty()) return;

    // Go thru the obs list and insert poses:

    TFeatureID max_pt_id = 0;
    TCameraPoseID max_fr_id = 0;

    for (TSequenceFeatureObservations::const_iterator it = observations.begin();
         it != observations.end(); ++it)
    {
        const TFeatureID feat_ID = it->id_feature;
        const TCameraPoseID frame_ID = it->id_frame;
        keep_max(max_fr_id, frame_ID);
        keep_max(max_pt_id, feat_ID);
    }

    // Insert N copies of the same values:
    frame_poses.assign(max_fr_id + 1, CPose3D(0, 0, 0, 0, 0, 0));
    landmark_points.assign(max_pt_id + 1, TPoint3D(0, 0, 1));

    MRPT_END
}

// This function is what to do for each feature in the reprojection loops below.
// -> residual: the raw residual, even if using robust kernel
// -> sum+= scaled squared norm of the residual (which != squared norm if using
// robust kernel)
template <bool POSES_INVERSE>
inline void reprojectionResidualsElement(
    const TCamera& camera_params, const TFeatureObservation& OBS,
    std::array<double, 2>& out_residual,
    const TFramePosesVec::value_type& frame,
    const TLandmarkLocationsVec::value_type& point, double& sum,
    const bool use_robust_kernel, const double kernel_param,
    double* out_kernel_1st_deriv)
{
    const TPixelCoordf z_pred =
        mrpt::vision::pinhole::projectPoint_no_distortion<POSES_INVERSE>(
            camera_params, frame, point);
    const TPixelCoordf& z_meas = OBS.px;

    out_residual[0] = z_meas.x - z_pred.x;
    out_residual[1] = z_meas.y - z_pred.y;

    const double sum_2 = square(out_residual[0]) + square(out_residual[1]);

    if (use_robust_kernel)
    {
        RobustKernel<rkPseudoHuber> kernel;
        kernel.param_sq = square(kernel_param);
        double kernel_1st_deriv, kernel_2nd_deriv;

        sum += kernel.eval(sum_2, kernel_1st_deriv, kernel_2nd_deriv);
        if (out_kernel_1st_deriv) *out_kernel_1st_deriv = kernel_1st_deriv;
    }
    else
    {
        sum += sum_2;
    }
}

/** Compute reprojection error vector (used from within Bundle Adjustment
 * methods, but can be used in general)
 *  See mrpt::vision::bundle_adj_full for a description of most parameters.
 *
 *  \return Overall squared reprojection error.
 */
double mrpt::vision::reprojectionResiduals(
    const TSequenceFeatureObservations& observations,
    const TCamera& camera_params, const TFramePosesMap& frame_poses,
    const TLandmarkLocationsMap& landmark_points,
    std::vector<std::array<double, 2>>& out_residuals,
    const bool frame_poses_are_inverse, const bool use_robust_kernel,
    const double kernel_param, std::vector<double>* out_kernel_1st_deriv)
{
    MRPT_START

    double sum = 0;

    const size_t N = observations.size();
    out_residuals.resize(N);
    if (out_kernel_1st_deriv) out_kernel_1st_deriv->resize(N);

    for (size_t i = 0; i < N; i++)
    {
        const TFeatureObservation& OBS = observations[i];

        const TFeatureID i_p = OBS.id_feature;
        const TCameraPoseID i_f = OBS.id_frame;

        TFramePosesMap::const_iterator itF = frame_poses.find(i_f);
        TLandmarkLocationsMap::const_iterator itP = landmark_points.find(i_p);
        ASSERTMSG_(itF != frame_poses.end(), "Frame ID is not in list!");
        ASSERTMSG_(itP != landmark_points.end(), "Landmark ID is not in list!");

        const TFramePosesMap::mapped_type& frame = itF->second;
        const TLandmarkLocationsMap::mapped_type& point = itP->second;

        double* ptr_1st_deriv =
            out_kernel_1st_deriv ? &((*out_kernel_1st_deriv)[i]) : nullptr;

        if (frame_poses_are_inverse)
            reprojectionResidualsElement<true>(
                camera_params, OBS, out_residuals[i], frame, point, sum,
                use_robust_kernel, kernel_param, ptr_1st_deriv);
        else
            reprojectionResidualsElement<false>(
                camera_params, OBS, out_residuals[i], frame, point, sum,
                use_robust_kernel, kernel_param, ptr_1st_deriv);
    }

    return sum;

    MRPT_END
}

double mrpt::vision::reprojectionResiduals(
    const TSequenceFeatureObservations& observations,
    const TCamera& camera_params, const TFramePosesVec& frame_poses,
    const TLandmarkLocationsVec& landmark_points,
    std::vector<std::array<double, 2>>& out_residuals,
    const bool frame_poses_are_inverse, const bool use_robust_kernel,
    const double kernel_param, std::vector<double>* out_kernel_1st_deriv)
{
    MRPT_START

    double sum = 0;

    const size_t N = observations.size();
    out_residuals.resize(N);
    if (out_kernel_1st_deriv) out_kernel_1st_deriv->resize(N);

    for (size_t i = 0; i < N; i++)
    {
        const TFeatureObservation& OBS = observations[i];

        const TFeatureID i_p = OBS.id_feature;
        const TCameraPoseID i_f = OBS.id_frame;

        ASSERT_BELOW_(i_p, landmark_points.size());
        ASSERT_BELOW_(i_f, frame_poses.size());
        const TFramePosesVec::value_type& frame = frame_poses[i_f];
        const TLandmarkLocationsVec::value_type& point = landmark_points[i_p];

        double* ptr_1st_deriv =
            out_kernel_1st_deriv ? &((*out_kernel_1st_deriv)[i]) : nullptr;

        if (frame_poses_are_inverse)
            reprojectionResidualsElement<true>(
                camera_params, OBS, out_residuals[i], frame, point, sum,
                use_robust_kernel, kernel_param, ptr_1st_deriv);
        else
            reprojectionResidualsElement<false>(
                camera_params, OBS, out_residuals[i], frame, point, sum,
                use_robust_kernel, kernel_param, ptr_1st_deriv);
    }

    return sum;
    MRPT_END
}

// Compute gradients & Hessian blocks
void mrpt::vision::ba_build_gradient_Hessians(
    const TSequenceFeatureObservations& observations,
    const vector<std::array<double, 2>>& residual_vec,
    const std::vector<JacData<6, 3, 2>>& jac_data_vec,
    std::vector<CMatrixFixed<double, 6, 6>>& U,
    std::vector<CVectorFixedDouble<6>>& eps_frame,
    std::vector<CMatrixFixed<double, 3, 3>>& V,
    std::vector<CVectorFixedDouble<3>>& eps_point, const size_t num_fix_frames,
    const size_t num_fix_points, const vector<double>* kernel_1st_deriv)
{
    MRPT_START

    const size_t N = observations.size();
    const bool use_robust_kernel = (kernel_1st_deriv != nullptr);

    for (size_t i = 0; i < N; i++)
    {
        const TFeatureObservation& OBS = observations[i];

        const TFeatureID i_p = OBS.id_feature;
        const TCameraPoseID i_f = OBS.id_frame;

        const Eigen::Matrix<double, 2, 1> RESID(&residual_vec[i][0]);
        const JacData<6, 3, 2>& JACOB = jac_data_vec[i];
        ASSERTDEB_(JACOB.frame_id == i_f && JACOB.point_id == i_p);

        if (i_f >= num_fix_frames)
        {
            const size_t frame_id = i_f - num_fix_frames;
            ASSERTDEB_(JACOB.J_frame_valid);
            ASSERT_BELOW_(frame_id, U.size());
            CMatrixDouble66 JtJ(UNINITIALIZED_MATRIX);
            JtJ.matProductOf_AtA(JACOB.J_frame);

            CVectorFixedDouble<6> eps_delta;
            // eps_delta = J^t * RESID
            eps_delta = JACOB.J_frame.transpose() * RESID;
            if (!use_robust_kernel)
            {
                eps_frame[frame_id] += eps_delta;
            }
            else
            {
                const double rho_1st_der = (*kernel_1st_deriv)[i];
                auto scaled_eps = eps_delta;
                scaled_eps *= rho_1st_der;
                eps_frame[frame_id] += scaled_eps;
            }
            U[frame_id] += JtJ;
        }
        if (i_p >= num_fix_points)
        {
            const size_t point_id = i_p - num_fix_points;
            ASSERTDEB_(JACOB.J_point_valid);
            ASSERT_BELOW_(point_id, V.size());
            CMatrixDouble33 JtJ(UNINITIALIZED_MATRIX);
            JtJ.matProductOf_AtA(JACOB.J_point);

            CVectorFixedDouble<3> eps_delta;
            // eps_delta = J^t * RESID
            eps_delta = JACOB.J_point.transpose() * RESID;
            if (!use_robust_kernel)
            {
                eps_point[point_id] += eps_delta;
            }
            else
            {
                const double rho_1st_der = (*kernel_1st_deriv)[i];
                auto scaled_eps = eps_delta;
                scaled_eps *= rho_1st_der;
                eps_point[point_id] += scaled_eps;
            }

            V[point_id] += JtJ;
        }
    }

    MRPT_END
}

void mrpt::vision::add_se3_deltas_to_frames(
    const TFramePosesVec& frame_poses, const CVectorDouble& delta,
    const size_t delta_first_idx, const size_t delta_num_vals,
    TFramePosesVec& new_frame_poses, const size_t num_fix_frames)
{
    MRPT_START

    new_frame_poses.resize(frame_poses.size());

    for (size_t i = 0; i < num_fix_frames; ++i)
        new_frame_poses[i] = frame_poses[i];

    size_t delta_used_vals = 0;
    const double* delta_val = &delta[delta_first_idx];

    for (size_t i = num_fix_frames; i < frame_poses.size(); i++)
    {
        const CPose3D& old_pose = frame_poses[i];
        CPose3D& new_pose = new_frame_poses[i];

        // Use the Lie Algebra methods for the increment:
        const CPose3D incrPose =
            mrpt::poses::Lie::SE<3>::exp(CVectorFixedDouble<6>(delta_val));

        // new_pose =  old_pose  [+] delta
        //         = exp(delta) (+) old_pose
        new_pose.composeFrom(incrPose, old_pose);

        // Move to the next entry in delta:
        delta_val += 6;
        delta_used_vals += 6;
    }

    ASSERT_(delta_used_vals == delta_num_vals);
    MRPT_END
}
void mrpt::vision::add_3d_deltas_to_points(
    const TLandmarkLocationsVec& landmark_points, const CVectorDouble& delta,
    const size_t delta_first_idx, const size_t delta_num_vals,
    TLandmarkLocationsVec& new_landmark_points, const size_t num_fix_points)
{
    MRPT_START

    new_landmark_points.resize(landmark_points.size());

    for (size_t i = 0; i < num_fix_points; ++i)
        new_landmark_points[i] = landmark_points[i];

    size_t delta_used_vals = 0;
    const double* delta_val = &delta[delta_first_idx];

    for (size_t i = num_fix_points; i < landmark_points.size(); i++)
    {
        const TPoint3D& old_point = landmark_points[i];
        TPoint3D& new_point = new_landmark_points[i];

        for (size_t j = 0; j < 3; j++)
            new_point[j] = old_point[j] + delta_val[j];

        // Move to the next entry in delta:
        delta_val += 3;
        delta_used_vals += 3;
    }

    ASSERT_(delta_used_vals == delta_num_vals);
    MRPT_END
}