def calibrate_intrinsic(args):
paths = setup_paths(args.paths)
setup_logging(args.runtime.log_level, [], log_file=paths.log_file)
info(pformat_struct(args))
image_path = os.path.expanduser(args.paths.image_path)
info(f"Finding images in {image_path}")
camera_images = find_camera_images(image_path,
args.paths.cameras,
args.paths.camera_pattern,
matching=False)
image_counts = {
k: len(files)
for k, files in zip(camera_images.cameras, camera_images.filenames)
}
info("Found camera directories with images {}".format(image_counts))
board_names, boards = split_dict(
find_board_config(image_path, args.paths.boards))
info("Loading images..")
images = image.detect.load_images(camera_images.filenames,
prefix=camera_images.image_path,
j=args.runtime.num_threads)
image_sizes = map_list(common_image_size, images)
info({
k: image_size
for k, image_size in zip(camera_images.cameras, image_sizes)
})
cache_key = struct(boards=boards,
image_sizes=image_sizes,
filenames=camera_images.filenames)
detected_points = detect_boards_cached(boards,
images,
paths.detections,
cache_key,
j=args.runtime.num_threads)
cameras, errs = calibrate_cameras(boards,
detected_points,
image_sizes,
model=args.camera.distortion_model,
fix_aspect=args.camera.fix_aspect,
has_skew=args.camera.allow_skew,
max_images=args.camera.limit_intrinsic)
for name, camera, err in zip(camera_images.cameras, cameras, errs):
info(f"Calibrated {name}, with RMS={err:.2f}")
info(camera)
info("")
info(f"Writing single calibrations to {paths.calibration_file}")
export_single(paths.calibration_file, cameras, camera_images.cameras,
camera_images.filenames)
def setup_paths(paths):
output_path = paths.image_path or paths.output_path
temp_folder = pathlib.Path(output_path).joinpath("." + paths.name)
temp_folder.mkdir(exist_ok=True, parents=True)
return struct(output=output_path,
temp=str(temp_folder),
calibration_file=path.join(output_path,
f"{paths.name}.json"),
log_file=str(temp_folder.joinpath("log.txt")),
detections=str(temp_folder.joinpath("detections.pkl")))
def matching_points(points, board, cam1, cam2):
points1, points2 = points._index[cam1], points._index[cam2]
matching = []
for i, j in zip(points1._sequence(0), points2._sequence(0)):
row1, row2, ids = common_entries(i, j)
matching.append(
struct(points1=row1.points,
points2=row2.points,
object_points=board.points[ids],
ids=ids))
return transpose_structs(matching)
def initialise_poses(pose_table):
# Find relative transforms between cameras and rig poses
camera = estimate_relative_poses(pose_table, axis=0)
board = estimate_relative_poses_inv(pose_table, axis=2)
# solve for the rig transforms cam @ rig @ board = pose
# first take inverse of both sides by board pose
# cam @ rig = board_relative = pose @ board^-1
board_relative = multiply_tables(pose_table, expand(inverse(board), [0, 1]) )
# solve for unknown rig
expanded = broadcast_to(expand(camera, [1, 2]), board_relative)
times = relative_between_n(expanded, board_relative, axis=1, inv=True)
return struct(times=times, camera=camera, board=board)
def initialise_poses(pose_table, camera_poses=None):
# Find relative transforms between cameras and rig poses
camera = estimate_relative_poses(pose_table, axis=0)
if camera_poses is not None:
info("Camera initialisation vs. supplied calibration")
report_poses("camera", camera_poses, camera.poses)
camera = Table.create(poses=camera_poses,
valid=np.ones(camera_poses.shape[0],
dtype=np.bool))
board = estimate_relative_poses_inv(pose_table, axis=2)
# solve for the rig transforms cam @ rig @ board = pose
# first take inverse of both sides by board pose
# cam @ rig = board_relative = pose @ board^-1
board_relative = multiply_tables(pose_table,
expand(inverse(board), [0, 1]))
# solve for unknown rig
expanded = broadcast_to(expand(camera, [1, 2]), board_relative)
times = relative_between_n(expanded, board_relative, axis=1, inv=True)
return struct(times=times, camera=camera, board=board)
def extract_points(frame_dets):
points, mask = fill_sparse(num_points, frame_dets.corners,
frame_dets.ids)
return struct(points=points, valid=mask)
def pad_points(board):
points = board.adjusted_points.astype(np.float64)
return struct(points=np.pad(points,
[(0, padded_points - points.shape[0]),
(0, 0)]),
valid=np.arange(padded_points) < board.num_points)
def valid_pose(t):
return struct(poses=t, valid=True)
def sparse_points(points):
ids = np.flatnonzero(points.valid)
return struct(corners=points.points[ids], ids=ids)
def fill_sparse_tile(n, values, ids, tile):
assert tile.shape == values.shape[1:]
dense = np.broadcast_to(np.expand_dims(tile, 0), (n, *tile.shape)).copy()
dense[ids] = values
mask = np.full(n, False)
mask[ids] = True
return dense, mask
def sparse_points(points):
ids = np.flatnonzero(points.valid)
return struct(corners=points.points[ids], ids=ids)
invalid_pose = struct(poses=np.eye(4), num_points=0, valid=False)
def valid_pose(t):
return struct(poses=t, valid=True)
def extract_pose(points, board, camera):
detections = sparse_points(points)
poses = board.estimate_pose_points(camera, detections)
return valid_pose(rtvec.to_matrix(poses))._extend(num_points=len(detections.ids))\
if poses is not None else invalid_pose
def map_table(f, point_table, boards, cameras):
