def get_matrix_poses(corner_storage, intrisinc_mat):
pairs = get_best_intersected(corner_storage)
best_pair = -1, -1
best_pair_result = -1
for i, j, _ in pairs[:100]:
ids1, ids2 = build_index_intersection(corner_storage[i].ids,
corner_storage[j].ids)
points1 = corner_storage[i].points[ids1]
points2 = corner_storage[j].points[ids2]
E, mask = cv2.findEssentialMat(points1,
points2,
focal=intrisinc_mat[0][0])
if mask.sum() < 10:
continue
F, mask = cv2.findFundamentalMat(points1, points2)
if mask.sum() < 10:
continue
_, R, t, mask = cv2.recoverPose(E,
points1,
points1,
focal=intrisinc_mat[0][0])
if mask.sum() < 10:
continue
corrs = build_correspondences(corner_storage[i], corner_storage[j])
points, ids, _ = triangulate_correspondences(
corrs, eye3x4(), np.hstack((R, t)), intrisinc_mat,
TriangulationParameters(max_reprojection_error=5,
min_triangulation_angle_deg=5.,
min_depth=0.001))
current_result = len(ids) // 20
if current_result > best_pair_result:
best_pair = i, j
best_pair_result = current_result
i, j = best_pair
ids1, ids2 = build_index_intersection(corner_storage[i].ids,
corner_storage[j].ids)
points1 = corner_storage[i].points[ids1]
points2 = corner_storage[j].points[ids2]
E, mask = cv2.findEssentialMat(points1, points2, focal=intrisinc_mat[0][0])
F, mask = cv2.findFundamentalMat(points1, points2)
_, R, t, mask = cv2.recoverPose(E,
points1,
points1,
focal=intrisinc_mat[0][0])
print(f"Chosen frames {i} and {j}")
return (i, view_mat3x4_to_pose(eye3x4())),\
(j, view_mat3x4_to_pose(np.hstack((R, t))))
def calculate_known_views(
intrinsic_mat,
corner_storage: CornerStorage,
min_correspondencies_count=100,
max_homography=0.7,
) -> Tuple[Tuple[int, Pose], Tuple[int, Pose]]:
best_points_num, best_known_views = -1, ((None, None), (None, None))
for i in range(len(corner_storage)):
for j in range(i + 1, min(i + 40, len(corner_storage))):
corresp = build_correspondences(corner_storage[i],
corner_storage[j])
if len(corresp[0]) < min_correspondencies_count:
break
E, mask = cv2.findEssentialMat(corresp.points_1,
corresp.points_2,
cameraMatrix=intrinsic_mat,
method=cv2.RANSAC)
mask = (mask.squeeze() == 1)
corresp = Correspondences(corresp.ids[mask],
corresp.points_1[mask],
corresp.points_2[mask])
if E is None:
continue
# Validate E using homography
H, mask = cv2.findHomography(corresp.points_1,
corresp.points_2,
method=cv2.RANSAC)
if np.count_nonzero(mask) / len(corresp.ids) > max_homography:
continue
R1, R2, T = cv2.decomposeEssentialMat(E)
for view_mat2 in [
np.hstack((R, t)) for R in [R1, R2] for t in [T, -T]
]:
view_mat1 = np.eye(3, 4)
points, _, _ = triangulate_correspondences(
corresp, view_mat1, view_mat2, intrinsic_mat,
triang_params)
print('Try frames {}, {}: {} correspondent points'.format(
i, j, len(points)))
if len(points) > best_points_num:
best_known_views = ((i, view_mat3x4_to_pose(view_mat1)),
(j, view_mat3x4_to_pose(view_mat2)))
best_points_num = len(points)
if best_points_num > 1500:
return best_known_views
return best_known_views
def _test_frame_pair(intrinsic_mat: np.ndarray,
corners_1: FrameCorners,
corners_2: FrameCorners,
param_koeff: float = 1) -> Tuple[int, Optional[Pose]]:
correspondences = build_correspondences(corners_1, corners_2)
if len(correspondences.ids) < 6:
return 0, None
points2d_1 = correspondences.points_1
points2d_2 = correspondences.points_2
essential, essential_inliers = cv2.findEssentialMat(points2d_1, points2d_2, intrinsic_mat, threshold=param_koeff)
homography, homography_inliers = cv2.findHomography(points2d_1, points2d_2, method=cv2.RANSAC)
if len(np.where(homography_inliers > 0)[0]) > len(np.where(essential_inliers > 0)[0]):
return 0, None
if essential.shape != (3, 3):
return 0, None
num_passed, rot, t, mask = cv2.recoverPose(essential, points2d_1, points2d_2,
intrinsic_mat, mask=essential_inliers)
outlier_ids = np.array(
[pt_id for pt_id, mask_elem in zip(correspondences.ids, mask) if mask_elem[0] == 0],
dtype=correspondences.ids.dtype)
inlier_correspondences = _remove_correspondences_with_ids(correspondences, outlier_ids)
if len(inlier_correspondences.ids) < 4:
return 0, None
view_matr_1 = eye3x4()
view_matr_2 = np.hstack((rot, t))
triangulation_params = TriangulationParameters(param_koeff, 1.0 / param_koeff, 0.0)
pts_3d, trianulated_ids, med_cos = triangulate_correspondences(inlier_correspondences, view_matr_1, view_matr_2,
intrinsic_mat, triangulation_params)
if len(pts_3d) < 4:
return 0, None
print(len(inlier_correspondences.ids), len(pts_3d))
return len(pts_3d), view_mat3x4_to_pose(view_matr_2)
def camera_pose(id,
corner_storage,
point_cloud_builder,
intrinsic_mat,
dist_coef=None):
frame_corners = corner_storage[id]
points = frame_corners.points
frame_ids = frame_corners.ids
cloud_ids = point_cloud_builder.ids
ids, frame_indexes, cloud_indexes = np.intersect1d(frame_ids,
cloud_ids,
return_indices=True)
if ids.shape[0] < 4:
return None
cloud_points = point_cloud_builder.points[cloud_indexes]
frame_points = points[frame_indexes]
retval, rvec, tvec, inliers = cv2.solvePnPRansac(cloud_points,
frame_points,
intrinsic_mat, dist_coef)
if retval is None:
return None
retval, rvec, tvec = cv2.solvePnP(cloud_points[inliers],
frame_points[inliers], intrinsic_mat,
dist_coef)
if retval is None:
return None
return view_mat3x4_to_pose(
rodrigues_and_translation_to_view_mat3x4(rvec, tvec)), np.setdiff1d(
ids, inliers)
def calc_known_views(corner_storage, intrinsic_mat, indent=5):
num_frames = len(corner_storage)
known_view_1 = (None, None)
known_view_2 = (None, None)
num_points = -1
for frame_1 in range(num_frames):
for frame_2 in range(frame_1 + indent, num_frames):
corrs = build_correspondences(corner_storage[frame_1],
corner_storage[frame_2])
if len(corrs.ids) < 6:
continue
points_1 = corrs.points_1
points_2 = corrs.points_2
H, mask_h = cv2.findHomography(points_1,
points_2,
method=cv2.RANSAC)
if mask_h is None:
continue
mask_h = mask_h.reshape(-1)
E, mask_e = cv2.findEssentialMat(points_1,
points_2,
method=cv2.RANSAC,
cameraMatrix=intrinsic_mat)
if mask_e is None:
continue
mask_e = mask_e.reshape(-1)
if mask_h.sum() / mask_e.sum() > 0.5:
continue
corrs = Correspondences(corrs.ids[(mask_e == 1)],
points_1[(mask_e == 1)],
points_2[(mask_e == 1)])
R1, R2, t = cv2.decomposeEssentialMat(E)
for poss_pose in [
Pose(R1.T, R1.T @ t),
Pose(R1.T, R1.T @ (-t)),
Pose(R2.T, R2.T @ t),
Pose(R2.T, R2.T @ (-t))
]:
points3d, _, _ = triangulate_correspondences(
corrs, eye3x4(), pose_to_view_mat3x4(poss_pose),
intrinsic_mat, TriangulationParameters(1, 2, .1))
if len(points3d) > num_points:
num_points = len(points3d)
known_view_1 = (frame_1, view_mat3x4_to_pose(eye3x4()))
known_view_2 = (frame_2, poss_pose)
return known_view_1, known_view_2
def select_frames(frames, corner_storage, camera_matrix):
frame1 = (0, view_mat3x4_to_pose(eye3x4()))
frame2 = (-1, view_mat3x4_to_pose(eye3x4()))
mx = 0
for i in range(1, len(frames)):
correspondences = build_correspondences(corner_storage[frame1[0]],
corner_storage[i])
if len(correspondences.ids) < 8:
continue
E, mask = cv2.findEssentialMat(correspondences.points_1,
correspondences.points_2,
camera_matrix,
method=cv2.RANSAC)
if mask is None:
continue
correspondences = _remove_correspondences_with_ids(
correspondences, np.argwhere(mask.flatten() == 0))
if len(correspondences.ids) < 8 or not validate(correspondences, mask):
continue
R1, R2, t = cv2.decomposeEssentialMat(E)
ps = [
Pose(R1.T, R1.T @ t),
Pose(R2.T, R2.T @ t),
Pose(R1.T, R1.T @ (-t)),
Pose(R2.T, R2.T @ (-t))
]
for pose in ps:
points, _, _ = triangulate_correspondences(
correspondences, pose_to_view_mat3x4(frame1[1]),
pose_to_view_mat3x4(pose), camera_matrix, TRIANG_PARAMS)
if len(points) > mx:
frame2 = (i, pose)
mx = len(points)
print(frame1[0], frame2[0])
return frame1, frame2
def find_best_start_poses(frames_n, corner_storage, intrinsic_mat):
best_frames = (0, 0)
best_snd_pose = None
best_pose_score = 0
for first_frame in range(frames_n):
for second_frame in range(first_frame + 5, frames_n):
pose, pose_score = get_pose_with_score(first_frame, second_frame,
corner_storage,
intrinsic_mat)
if pose_score > best_pose_score:
best_frames = (first_frame, second_frame)
best_snd_pose = pose
best_pose_score = pose_score
return best_frames, view_mat3x4_to_pose(eye3x4()), best_snd_pose
def _find_pos_pair(intrinsic_mat: np.ndarray,
corner_storage: CornerStorage,
param_koeff: float = 1) -> Tuple[Tuple[int, Pose], Tuple[int, Pose]]:
best_num_pts = 0
best_frame = 1
best_pose = None
for frame_number in range(1, len(corner_storage)):
print(frame_number)
num_pts, pose = _test_frame_pair(intrinsic_mat, corner_storage[0], corner_storage[frame_number])
if num_pts > best_num_pts:
best_num_pts = num_pts
best_pose = pose
best_frame = frame_number
if best_pose is None:
if param_koeff < 4:
return _find_pos_pair(intrinsic_mat, corner_storage, param_koeff * 2)
raise TrackingError("Failed to initialize tracking")
return (0, view_mat3x4_to_pose(eye3x4())), (best_frame, best_pose)
def initial_camera_views(
corner_storage: CornerStorage, intrinsic_mat: np.ndarray
) -> Tuple[Tuple[int, Pose], Tuple[int, Pose]]:
max_pose = None
max_npoints = 0
maxj = -1
maxi = -1
enum_corner_storage = list(enumerate(corner_storage))
for j, base_frame in tqdm(enum_corner_storage[::32]):
base_frame = corner_storage[0]
for i, frame in enum_corner_storage[j + 1:j + 32]:
pose, npoints = pose_by_frames(base_frame, frame, intrinsic_mat)
print(i, j, npoints)
if npoints > max_npoints:
max_npoints = npoints
maxi = i
maxj = j
max_pose = pose
print(max_npoints)
return (maxj, view_mat3x4_to_pose(eye3x4())), (maxi, max_pose)
def track(self):
step_num = 1
num_of_defined_poses = np.sum([
track_pos_info is not None for track_pos_info in self.tracked_poses
])
while num_of_defined_poses != self.num_of_frames:
unsolved_frames = [
i for i in range(self.num_of_frames)
if self.tracked_poses[i] is None
]
new_cams_info = []
for frame, cam_info in zip(unsolved_frames,
map(self._get_pos, unsolved_frames)):
if cam_info is not None:
new_cams_info.append((frame, cam_info))
if len(new_cams_info) == 0:
print(f'Can not get more camera positions, '
f'{self.num_of_frames - num_of_defined_poses}'
f' frames left without defined camera position')
self.tracked_poses = [
CameraInfo(view_mat3x4_to_pose(eye3x4()), 0)
if pos is None else pos for pos in self.tracked_poses
]
break
best_frame = None
best_new_cam_info = None
for frame, cam_info in new_cams_info:
if best_new_cam_info is None or best_new_cam_info.inliers < cam_info.inliers:
best_new_cam_info = cam_info
best_frame = frame
print('Added camera position for frame ', best_frame)
print('Number of inliers: ', best_new_cam_info.inliers)
self.tracked_poses[best_frame] = best_new_cam_info
self._retriangulate_3d_points(best_frame)
if step_num % 4 == 0:
self._update_camera_poses()
step_num += 1
num_of_defined_poses = np.sum([
tracked_pos_info is not None
for tracked_pos_info in self.tracked_poses
])
print(
f'{num_of_defined_poses}/{self.num_of_frames} camera positions found, {len(self.point_cloud)} points in cloud'
)
self._update_camera_poses()
ids, cloud_points = [], []
for pt_id, clout_pt_info in self.point_cloud.items():
ids.append(pt_id)
cloud_points.append(clout_pt_info.pos)
return list(map(lambda tracked_pos_info: tracked_pos_info.pos, self.tracked_poses)), \
PointCloudBuilder(np.array(ids), np.array(cloud_points))