bool FundamentalMatrixSolver::ComputeFundamentalMatrices(
std::vector<Matrix3d>& fundamental_matrices) {
// Clear the output.
fundamental_matrices.clear();
// Determine a fundamental matrix for each pair of images.
for (const auto& pair_data : matched_image_data_) {
Matrix3d fundamental_matrix;
if (ComputeFundamentalMatrix(pair_data.feature_matches_,
fundamental_matrix)) {
fundamental_matrices.push_back(fundamental_matrix);
} else {
VLOG(1) << "Failed to compute funamental matrix.";
}
}
// Return whether or not we computed any fundamental matrices.
if (fundamental_matrices.empty()) {
VLOG(1) << "Unable to compute a fundamental matrix for any of the input "
"image pairs.";
return false;
}
return true;
}
bool BundlerApp::EstimateRelativePose(int i1, int i2,
camera_params_t &camera1,
camera_params_t &camera2)
{
MatchIndex list_idx;
if (i1 < i2)
list_idx = GetMatchIndex(i1, i2);
else
list_idx = GetMatchIndex(i2, i1);
std::vector<KeypointMatch> &matches = m_matches.GetMatchList(list_idx);
int num_matches = (int) matches.size();
double f1 = m_image_data[i1].m_init_focal;
double f2 = m_image_data[i2].m_init_focal;
double E[9], F[9];
std::vector<int> inliers;
if (!m_optimize_for_fisheye) {
inliers =
EstimateEMatrix(m_image_data[i1].m_keys, m_image_data[i2].m_keys,
matches,
4 * m_fmatrix_rounds, // 8 * m_fmatrix_rounds,
m_fmatrix_threshold * m_fmatrix_threshold,
f1, f2, E, F);
} else {
/* FIXME */
inliers =
EstimateEMatrix(m_image_data[i1].m_keys, m_image_data[i2].m_keys,
matches,
4 * m_fmatrix_rounds, // 8 * m_fmatrix_rounds,
m_fmatrix_threshold * m_fmatrix_threshold,
f1, f2, E, F);
}
if ((int) inliers.size() == 0)
return false;
int num_inliers = (int) inliers.size();
printf(" Found %d / %d inliers (%0.3f%%)\n", num_inliers, num_matches,
100.0 * num_inliers / num_matches);
/* Estimate a homography with the inliers */
std::vector<KeypointMatch> match_inliers;
for (int i = 0; i < num_inliers; i++) {
match_inliers.push_back(matches[inliers[i]]);
}
int num_match_inliers = (int) match_inliers.size();
double H[9];
std::vector<int> Hinliers =
EstimateTransform(m_image_data[i1].m_keys, m_image_data[i2].m_keys,
match_inliers, MotionHomography,
128 /*m_homography_rounds*/,
6.0 /*m_homography_threshold*/, H);
printf(" Found %d / %d homography inliers (%0.3f%%)\n",
(int) Hinliers.size(), num_inliers,
100.0 * Hinliers.size() / num_inliers);
bool initialized = false;
if ((int) Hinliers.size() > 0) {
matrix_print(3, 3, H);
printf("\n");
if ((double) Hinliers.size() / num_inliers >= 0.75 /*0.85*/) {
KeypointMatch &match0 = matches[Hinliers[0]];
v2_t p10 = v2_new(m_image_data[i1].m_keys[match0.m_idx1].m_x,
m_image_data[i1].m_keys[match0.m_idx1].m_y);
v2_t p20 = v2_new(m_image_data[i2].m_keys[match0.m_idx2].m_x,
m_image_data[i2].m_keys[match0.m_idx2].m_y);
double R1[9], t1[3], R2[9], t2[3];
bool success =
DecomposeHomography(H, f1, f2, R1, t1, R2, t2, p10, p20);
if (success) {
printf("[BundleTwoFrame] Using homography "
"for initialization\n");
/* Decide which solution to use */
double F1h[9], F2h[9];
ComputeFundamentalMatrix(f1, f2, R1, t1, F1h);
ComputeFundamentalMatrix(f1, f2, R2, t2, F2h);
double F1hT[9], F2hT[9];
matrix_transpose(3, 3, F1h, F1hT);
matrix_transpose(3, 3, F2h, F2hT);
int num_inliers1 = 0, num_inliers2 = 0;
for (int i = 0; i < num_match_inliers; i++) {
const KeypointMatch &match = match_inliers[i];
const Keypoint &k1 = m_image_data[i1].m_keys[match.m_idx1];
const Keypoint &k2 = m_image_data[i2].m_keys[match.m_idx2];
v3_t rt = v3_new(k1.m_x, k1.m_y, 1.0);
v3_t lft = v3_new(k2.m_x, k2.m_y, 1.0);
double r1a = fmatrix_compute_residual(F1h, lft, rt);
double r1b = fmatrix_compute_residual(F1hT, rt, lft);
double r2a = fmatrix_compute_residual(F2h, lft, rt);
double r2b = fmatrix_compute_residual(F2hT, rt, lft);
if (r1a < m_fmatrix_threshold && r1b < m_fmatrix_threshold)
num_inliers1++;
if (r2a < m_fmatrix_threshold && r2b < m_fmatrix_threshold)
num_inliers2++;
}
initialized = true;
double *R, *t;
printf(" H1: %d inliers, H2: %d inliers\n",
num_inliers1, num_inliers2);
if (num_inliers1 > num_inliers2) {
R = R1;
t = t1;
} else {
R = R2;
t = t2;
}
memcpy(camera2.R, R, sizeof(double) * 9);
matrix_transpose_product(3, 3, 3, 1, R, t, camera2.t);
matrix_scale(3, 1, camera2.t, -1.0, camera2.t);
}
}
}
if (!initialized) {
KeypointMatch &match = matches[inliers[0]];
v2_t p1 = v2_new(m_image_data[i1].m_keys[match.m_idx1].m_x / f1,
m_image_data[i1].m_keys[match.m_idx1].m_y / f1);
v2_t p2 = v2_new(m_image_data[i2].m_keys[match.m_idx2].m_x / f2,
m_image_data[i2].m_keys[match.m_idx2].m_y / f2);
double R[9], t[3];
int success = find_extrinsics_essential(E, p1, p2, R, t);
if (!success) {
return false;
}
memcpy(camera2.R, R, sizeof(double) * 9);
matrix_transpose_product(3, 3, 3, 1, R, t, camera2.t);
matrix_scale(3, 1, camera2.t, -1.0, camera2.t);
}
return true;
}
