ba_full.cpp

Go to the documentation of this file.

/* +------------------------------------------------------------------------+
   |                     Mobile Robot Programming Toolkit (MRPT)            |
   |                          http://www.mrpt.org/                          |
   |                                                                        |
   | Copyright (c) 2005-2018, Individual contributors, see AUTHORS file     |
   | See: http://www.mrpt.org/Authors - All rights reserved.                |
   | Released under BSD License. See details in http://www.mrpt.org/License |
   +------------------------------------------------------------------------+ */

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

#include <mrpt/vision/bundle_adjustment.h>
#include <mrpt/system/CTimeLogger.h>
#include <mrpt/math/CSparseMatrix.h>
#include <mrpt/math/ops_containers.h>
#include <mrpt/core/aligned_std_map.h>

#include <memory>  // unique_ptr

#include "ba_internals.h"

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

// Define -> estimate the inverse of poses; Undefined -> estimate the camera
// poses (read the doc of mrpt::vision::bundle_adj_full)
#define USE_INVERSE_POSES

#ifdef USE_INVERSE_POSES
#define INV_POSES_BOOL true
#else
#define INV_POSES_BOOL false
#endif

/* ----------------------------------------------------------
                    bundle_adj_full

    See bundle_adjustment.h for docs.

  This implementation is almost entirely an adapted version from
  RobotVision (at OpenSLAM.org) (C) by Hauke Strasdat (Imperial
  College London), licensed under GNU LGPL.
  See the related paper:  H. Strasdat, J.M.M. Montiel,
  A.J. Davison: "Scale Drift-Aware Large Scale Monocular SLAM",
  RSS2010, http://www.roboticsproceedings.org/rss06/p10.html

   ---------------------------------------------------------- */
double mrpt::vision::bundle_adj_full(
    const TSequenceFeatureObservations& observations,
    const TCamera& camera_params, TFramePosesVec& frame_poses,
    TLandmarkLocationsVec& landmark_points,
    const mrpt::system::TParametersDouble& extra_params,
    const TBundleAdjustmentFeedbackFunctor user_feedback)
{
    MRPT_START

    // Generic BA problem dimension numbers:
    static const unsigned int FrameDof = 6;  // Poses: x y z yaw pitch roll
    static const unsigned int PointDof = 3;  // Landmarks: x y z
    static const unsigned int ObsDim = 2;  // Obs: x y (pixels)

    // Typedefs for this specific BA problem:
    using MyJacData = JacData<FrameDof, PointDof, ObsDim>;
    using MyJacDataVec = mrpt::aligned_std_vector<MyJacData>;

    using Array_O = std::array<double, ObsDim>;
    using Array_F = CArrayDouble<FrameDof>;
    using Array_P = CArrayDouble<PointDof>;
    using Matrix_FxF = CMatrixFixedNumeric<double, FrameDof, FrameDof>;
    using Matrix_PxP = CMatrixFixedNumeric<double, PointDof, PointDof>;
    using Matrix_FxP = CMatrixFixedNumeric<double, FrameDof, PointDof>;

    // Extra params:
    const bool use_robust_kernel =
        0 != extra_params.getWithDefaultVal("robust_kernel", 1);
    const bool verbose = 0 != extra_params.getWithDefaultVal("verbose", 0);
    const double initial_mu = extra_params.getWithDefaultVal("mu", -1);
    const size_t max_iters =
        extra_params.getWithDefaultVal("max_iterations", 50);
    const size_t num_fix_frames =
        extra_params.getWithDefaultVal("num_fix_frames", 1);
    const size_t num_fix_points =
        extra_params.getWithDefaultVal("num_fix_points", 0);
    const double kernel_param =
        extra_params.getWithDefaultVal("kernel_param", 3.0);

    const bool enable_profiler =
        0 != extra_params.getWithDefaultVal("profiler", 0);

    mrpt::system::CTimeLogger profiler(enable_profiler);

    profiler.enter("bundle_adj_full (complete run)");

    // Input data sizes:
    const size_t num_points = landmark_points.size();
    const size_t num_frames = frame_poses.size();
    const size_t num_obs = observations.size();

    ASSERT_ABOVE_(num_frames, 0);
    ASSERT_ABOVE_(num_points, 0);
    ASSERT_(num_fix_frames >= 1);
    ASSERT_ABOVEEQ_(num_frames, num_fix_frames);
    ASSERT_ABOVEEQ_(num_points, num_fix_points);

#ifdef USE_INVERSE_POSES
    // *Warning*: This implementation assumes inverse camera poses: inverse them
    // at the entrance and at exit:
    profiler.enter("invert_poses");
    for (size_t i = 0; i < num_frames; i++) frame_poses[i].inverse();
    profiler.leave("invert_poses");
#endif

    MyJacDataVec jac_data_vec(num_obs);
    vector<Array_O> residual_vec(num_obs);
    vector<double> kernel_1st_deriv(num_obs);

    // prepare structure of sparse Jacobians:
    for (size_t i = 0; i < num_obs; i++)
    {
        jac_data_vec[i].frame_id = observations[i].id_frame;
        jac_data_vec[i].point_id = observations[i].id_feature;
    }

    // Compute sparse Jacobians:
    profiler.enter("compute_Jacobians");
    ba_compute_Jacobians<INV_POSES_BOOL>(
        frame_poses, landmark_points, camera_params, jac_data_vec,
        num_fix_frames, num_fix_points);
    profiler.leave("compute_Jacobians");

    profiler.enter("reprojectionResiduals");
    double res = mrpt::vision::reprojectionResiduals(
        observations, camera_params, frame_poses, landmark_points, residual_vec,
        INV_POSES_BOOL,  // are poses inverse?
        use_robust_kernel, kernel_param,
        use_robust_kernel ? &kernel_1st_deriv : nullptr);
    profiler.leave("reprojectionResiduals");

    MRPT_CHECK_NORMAL_NUMBER(res);

    VERBOSE_COUT << "res: " << res << endl;

    // Auxiliary vars:
    Array_F arrF_zeros;
    arrF_zeros.assign(0);
    Array_P arrP_zeros;
    arrP_zeros.assign(0);

    const size_t num_free_frames = num_frames - num_fix_frames;
    const size_t num_free_points = num_points - num_fix_points;
    const size_t len_free_frames = FrameDof * num_free_frames;
    const size_t len_free_points = PointDof * num_free_points;

    mrpt::aligned_std_vector<Matrix_FxF> H_f(num_free_frames);
    mrpt::aligned_std_vector<Array_F> eps_frame(num_free_frames, arrF_zeros);
    mrpt::aligned_std_vector<Matrix_PxP> H_p(num_free_points);
    mrpt::aligned_std_vector<Array_P> eps_point(num_free_points, arrP_zeros);

    profiler.enter("build_gradient_Hessians");
    ba_build_gradient_Hessians(
        observations, residual_vec, jac_data_vec, H_f, eps_frame, H_p,
        eps_point, num_fix_frames, num_fix_points,
        use_robust_kernel ? &kernel_1st_deriv : nullptr);
    profiler.leave("build_gradient_Hessians");

    double nu = 2;
    double eps = 1e-16;  // 0.000000000000001;
    bool stop = false;
    double mu = initial_mu;

    // Automatic guess of "mu":
    if (mu < 0)
    {
        double norm_max_A = 0;
        for (size_t j = 0; j < num_free_frames; ++j)
            for (size_t dim = 0; dim < FrameDof; dim++)
                keep_max(norm_max_A, H_f[j](dim, dim));

        for (size_t i = 0; i < num_free_points; ++i)
            for (size_t dim = 0; dim < PointDof; dim++)
                keep_max(norm_max_A, H_p[i](dim, dim));
        double tau = 1e-3;
        mu = tau * norm_max_A;
    }

    Matrix_FxF I_muFrame(UNINITIALIZED_MATRIX);
    Matrix_PxP I_muPoint(UNINITIALIZED_MATRIX);

    // Cholesky object, as a pointer to reuse it between iterations:
    using SparseCholDecompPtr = std::unique_ptr<CSparseMatrix::CholeskyDecomp>;

    SparseCholDecompPtr ptrCh;

    for (size_t iter = 0; iter < max_iters; iter++)
    {
        VERBOSE_COUT << "iteration: " << iter << endl;

        // provide feedback to the user:
        if (user_feedback)
            user_feedback(
                iter, res, max_iters, observations, frame_poses,
                landmark_points);

        bool has_improved = false;
        do
        {
            profiler.enter("COMPLETE_ITER");

            VERBOSE_COUT << "mu: " << mu << endl;

            I_muFrame.unit(FrameDof, mu);
            I_muPoint.unit(PointDof, mu);

            mrpt::aligned_std_vector<Matrix_FxF> U_star(
                num_free_frames, I_muFrame);
            mrpt::aligned_std_vector<Matrix_PxP> V_inv(num_free_points);

            for (size_t i = 0; i < U_star.size(); ++i) U_star[i] += H_f[i];

            for (size_t i = 0; i < H_p.size(); ++i)
                (H_p[i] + I_muPoint).inv_fast(V_inv[i]);

            using WMap = mrpt::aligned_std_map<pair<TCameraPoseID, TLandmarkID>,
                                               Matrix_FxP>;
            WMap W, Y;

            // For quick look-up of entries in W affecting a given point ID:
            vector<vector<WMap::iterator>> W_entries(
                num_points);  // Index is "TLandmarkID"

            MyJacDataVec::const_iterator jac_iter = jac_data_vec.begin();
            for (TSequenceFeatureObservations::const_iterator it_obs =
                     observations.begin();
                 it_obs != observations.end(); ++it_obs)
            {
                const TFeatureID feat_id = it_obs->id_feature;
                const TCameraPoseID frame_id = it_obs->id_frame;

                if (jac_iter->J_frame_valid && jac_iter->J_point_valid)
                {
                    const CMatrixFixedNumeric<double, ObsDim, FrameDof>&
                        J_frame = jac_iter->J_frame;
                    const CMatrixFixedNumeric<double, ObsDim, PointDof>&
                        J_point = jac_iter->J_point;
                    const pair<TCameraPoseID, TLandmarkID> id_pair =
                        make_pair(frame_id, feat_id);

                    Matrix_FxP tmp(UNINITIALIZED_MATRIX);
                    tmp.multiply_AtB(J_frame, J_point);

                    // W[ids] = J_f^T * J_p
                    // Was: W[id_pair] = tmp;
                    const pair<WMap::iterator, bool>& retInsert =
                        W.insert(make_pair(id_pair, tmp));
                    ASSERT_(retInsert.second == true);
                    W_entries[feat_id].push_back(
                        retInsert.first);  // Keep the iterator

                    // Y[ids] = W[ids] * H_p^{-1}
                    Y[id_pair].multiply_AB(
                        tmp, V_inv[feat_id - num_fix_points]);
                }
                ++jac_iter;
            }

            mrpt::aligned_std_map<pair<TCameraPoseID, TLandmarkID>, Matrix_FxF>
                YW_map;
            for (size_t i = 0; i < H_f.size(); ++i)
                YW_map[std::pair<TCameraPoseID, TLandmarkID>(i, i)] = U_star[i];

            CVectorDouble delta(
                len_free_frames + len_free_points);  // The optimal step
            CVectorDouble e(len_free_frames);

            profiler.enter("Schur.build.reduced.frames");
            for (size_t j = 0; j < num_free_frames; ++j)
                ::memcpy(
                    &e[j * FrameDof], &eps_frame[j][0],
                    sizeof(e[0]) * FrameDof);  // e.slice(j*FrameDof,FrameDof) =
            // AT(eps_frame,j);

            for (WMap::iterator Y_ij = Y.begin(); Y_ij != Y.end(); ++Y_ij)
            {
                const TLandmarkID point_id = Y_ij->first.second;
                const size_t i =
                    Y_ij->first.second - num_fix_points;  // point index
                const size_t j =
                    Y_ij->first.first - num_fix_frames;  // frame index

                const vector<WMap::iterator>& iters =
                    W_entries[point_id];  //->second;

                for (size_t itIdx = 0; itIdx < iters.size(); itIdx++)
                // for (size_t k=0; k<num_free_frames; ++k)
                {
                    const WMap::iterator& W_ik = iters[itIdx];
                    const TLandmarkID k = W_ik->first.first - num_fix_frames;

                    Matrix_FxF YWt(UNINITIALIZED_MATRIX);  //-(Y_ij->second) *
                    //(W_ik->second).T();
                    YWt.multiply_ABt(Y_ij->second, W_ik->second);
                    // YWt*=-1.0; // The "-" sign is taken into account below:

                    const pair<TCameraPoseID, TLandmarkID> ids_jk =
                        pair<TCameraPoseID, TLandmarkID>(j, k);

                    auto it = YW_map.find(ids_jk);
                    if (it != YW_map.end())
                        it->second -= YWt;  // += (-YWt);
                    else
                        YW_map[ids_jk] = YWt * (-1.0);
                }

                CArrayDouble<FrameDof> r;
                Y_ij->second.multiply_Ab(eps_point[i], r);
                for (size_t k = 0; k < FrameDof; k++)
                    e[j * FrameDof + k] -= r[k];
            }
            profiler.leave("Schur.build.reduced.frames");

            profiler.enter("sS:ALL");
            profiler.enter("sS:fill");

            VERBOSE_COUT << "Entries in YW_map:" << YW_map.size() << endl;

            CSparseMatrix sS(len_free_frames, len_free_frames);

            for (auto it = YW_map.begin(); it != YW_map.end(); ++it)
            {
                const pair<TCameraPoseID, TLandmarkID>& ids = it->first;
                const Matrix_FxF& YW = it->second;
                sS.insert_submatrix(
                    ids.first * FrameDof, ids.second * FrameDof, YW);
            }
            profiler.leave("sS:fill");

            // Compress the sparse matrix:
            profiler.enter("sS:compress");
            sS.compressFromTriplet();
            profiler.leave("sS:compress");

            try
            {
                profiler.enter("sS:chol");
                if (!ptrCh.get())
                    ptrCh.reset(new CSparseMatrix::CholeskyDecomp(sS));
                else
                    ptrCh.get()->update(sS);
                profiler.leave("sS:chol");

                profiler.enter("sS:backsub");
                CVectorDouble bck_res;
                ptrCh->backsub(e, bck_res);  // Ax = b -->  delta= x*
                ::memcpy(
                    &delta[0], &bck_res[0],
                    bck_res.size() * sizeof(bck_res[0]));  // delta.slice(0,...)
                // = Ch.backsub(e);
                profiler.leave("sS:backsub");
                profiler.leave("sS:ALL");
            }
            catch (CExceptionNotDefPos&)
            {
                profiler.leave("sS:ALL");
                // not positive definite so increase mu and try again
                mu *= nu;
                nu *= 2.;
                stop = (mu > 999999999.f);
                continue;
            }

            profiler.enter("PostSchur.landmarks");

            CVectorDouble g(len_free_frames + len_free_points);
            ::memcpy(
                &g[0], &e[0],
                len_free_frames *
                    sizeof(
                        g[0]));  // g.slice(0,FrameDof*(num_frames-num_fix_frames))
            // = e;

            for (size_t i = 0; i < num_free_points; ++i)
            {
                Array_P tmp =
                    eps_point[i];  // eps_point.slice(PointDof*i,PointDof);

                for (size_t j = 0; j < num_free_frames; ++j)
                {
                    WMap::iterator W_ij;
                    W_ij = W.find(
                        make_pair(
                            TCameraPoseID(j + num_fix_frames),
                            TLandmarkID(i + num_fix_points)));

                    if (W_ij != W.end())
                    {
                        // tmp -= W_ij->second.T() *
                        // delta.slice(j*FrameDof,FrameDof);
                        const Array_F v(&delta[j * FrameDof]);
                        Array_P r;
                        W_ij->second.multiply_Atb(v, r);  // r= A^t * v
                        tmp -= r;
                    }
                }
                Array_P Vi_tmp;
                V_inv[i].multiply_Ab(tmp, Vi_tmp);  // Vi_tmp = V_inv[i] * tmp

                ::memcpy(
                    &delta[len_free_frames + i * PointDof], &Vi_tmp[0],
                    sizeof(Vi_tmp[0]) * PointDof);
                ::memcpy(
                    &g[len_free_frames + i * PointDof], &eps_point[i][0],
                    sizeof(eps_point[0][0]) * PointDof);
            }
            profiler.leave("PostSchur.landmarks");

            // Vars for temptative new estimates:
            TFramePosesVec new_frame_poses;
            TLandmarkLocationsVec new_landmark_points;

            profiler.enter("add_se3_deltas_to_frames");
            add_se3_deltas_to_frames(
                frame_poses, delta, 0, len_free_frames, new_frame_poses,
                num_fix_frames);
            profiler.leave("add_se3_deltas_to_frames");

            profiler.enter("add_3d_deltas_to_points");
            add_3d_deltas_to_points(
                landmark_points, delta, len_free_frames, len_free_points,
                new_landmark_points, num_fix_points);
            profiler.leave("add_3d_deltas_to_points");

            vector<Array_O> new_residual_vec(num_obs);
            vector<double> new_kernel_1st_deriv(num_obs);

            profiler.enter("reprojectionResiduals");
            double res_new = mrpt::vision::reprojectionResiduals(
                observations, camera_params, new_frame_poses,
                new_landmark_points, new_residual_vec,
                INV_POSES_BOOL,  // are poses inverse?
                use_robust_kernel, kernel_param,
                use_robust_kernel ? &new_kernel_1st_deriv : nullptr);
            profiler.leave("reprojectionResiduals");

            MRPT_CHECK_NORMAL_NUMBER(res_new);

            has_improved = (res_new < res);

            if (has_improved)
            {
                // rho = (res-)/ (delta.array()*(mu*delta + g).array() ).sum();
                // Good: Accept new values
                VERBOSE_COUT << "new total sqr.err=" << res_new
                             << " avr.err(px):" << std::sqrt(res / num_obs)
                             << "->" << std::sqrt(res_new / num_obs) << endl;

                // swap is faster than "="
                frame_poses.swap(new_frame_poses);
                landmark_points.swap(new_landmark_points);
                residual_vec.swap(new_residual_vec);
                kernel_1st_deriv.swap(new_kernel_1st_deriv);

                res = res_new;

                profiler.enter("compute_Jacobians");
                ba_compute_Jacobians<INV_POSES_BOOL>(
                    frame_poses, landmark_points, camera_params, jac_data_vec,
                    num_fix_frames, num_fix_points);
                profiler.leave("compute_Jacobians");

                // Reset to zeros:
                H_f.assign(num_free_frames, Matrix_FxF());
                H_p.assign(num_free_points, Matrix_PxP());
                eps_frame.assign(num_free_frames, arrF_zeros);
                eps_point.assign(num_free_points, arrP_zeros);

                profiler.enter("build_gradient_Hessians");
                ba_build_gradient_Hessians(
                    observations, residual_vec, jac_data_vec, H_f, eps_frame,
                    H_p, eps_point, num_fix_frames, num_fix_points,
                    use_robust_kernel ? &kernel_1st_deriv : nullptr);
                profiler.leave("build_gradient_Hessians");

                stop = norm_inf(g) <= eps;
                // mu *= max(1.0/3.0, 1-std::pow(2*rho-1,3.0) );
                mu *= 0.1;
                mu = std::max(mu, 1e-100);
                nu = 2.0;
            }
            else
            {
                VERBOSE_COUT << "no update: res vs.res_new " << res << " vs. "
                             << res_new << endl;
                mu *= nu;
                nu *= 2.0;
                stop = (mu > 1e9);
            }

            profiler.leave("COMPLETE_ITER");
        } while (!has_improved && !stop);

        if (stop) break;

    }  // end for each "iter"

// *Warning*: This implementation assumes inverse camera poses: inverse them at
// the entrance and at exit:
#ifdef USE_INVERSE_POSES
    profiler.enter("invert_poses");
    for (size_t i = 0; i < num_frames; i++) frame_poses[i].inverse();
    profiler.leave("invert_poses");
#endif

    profiler.leave("bundle_adj_full (complete run)");

    return res;
    MRPT_END
}