def _parse_ljson_v1(lms_dict):
all_points = []
labels = [] # label per group
labels_slices = [] # slices into the full pointcloud per label
offset = 0
connectivity = []
for group in lms_dict['groups']:
lms = group['landmarks']
labels.append(group['label'])
labels_slices.append(slice(offset, len(lms) + offset))
# Create the connectivity if it exists
conn = group.get('connectivity', [])
if conn:
# Offset relative connectivity according to the current index
conn = offset + np.asarray(conn)
connectivity += conn.tolist()
for p in lms:
all_points.append(p['point'])
offset += len(lms)
# Don't create a PointUndirectedGraph with no connectivity
points = _ljson_parse_null_values(all_points)
n_points = points.shape[0]
labels_to_masks = OrderedDict()
# go through each label and build the appropriate boolean array
for label, l_slice in zip(labels, labels_slices):
mask = np.zeros(n_points, dtype=np.bool)
mask[l_slice] = True
labels_to_masks[label] = mask
lmarks = LabelledPointUndirectedGraph.init_from_edges(points, connectivity,
labels_to_masks)
return {'LJSON': lmarks}
def _parse_ljson_v1(lms_dict):
all_points = []
labels = [] # label per group
labels_slices = [] # slices into the full pointcloud per label
offset = 0
connectivity = []
for group in lms_dict["groups"]:
lms = group["landmarks"]
labels.append(group["label"])
labels_slices.append(slice(offset, len(lms) + offset))
# Create the connectivity if it exists
conn = group.get("connectivity", [])
if conn:
# Offset relative connectivity according to the current index
conn = offset + np.asarray(conn)
connectivity += conn.tolist()
for p in lms:
all_points.append(p["point"])
offset += len(lms)
# Don't create a PointUndirectedGraph with no connectivity
points = _ljson_parse_null_values(all_points)
n_points = points.shape[0]
labels_to_masks = OrderedDict()
# go through each label and build the appropriate boolean array
for label, l_slice in zip(labels, labels_slices):
mask = np.zeros(n_points, dtype=np.bool)
mask[l_slice] = True
labels_to_masks[label] = mask
lmarks = LabelledPointUndirectedGraph.init_from_edges(
points, connectivity, labels_to_masks
)
return {"LJSON": lmarks}
def _parse_ljson_v2(lms_dict):
points = _ljson_parse_null_values(lms_dict['landmarks']['points'])
connectivity = lms_dict['landmarks'].get('connectivity')
if connectivity is None and len(lms_dict['labels']) == 0:
return PointCloud(points)
else:
labels_to_mask = OrderedDict() # masks into the pointcloud per label
n_points = points.shape[0]
for label in lms_dict['labels']:
mask = np.zeros(n_points, dtype=np.bool)
mask[label['mask']] = True
labels_to_mask[label['label']] = mask
# Note that we can pass connectivity as None here and the edges will be
# empty.
return LabelledPointUndirectedGraph.init_from_edges(
points, connectivity, labels_to_mask)
def _parse_ljson_v2(lms_dict):
points = _ljson_parse_null_values(lms_dict['landmarks']['points'])
connectivity = lms_dict['landmarks'].get('connectivity')
if connectivity is None and len(lms_dict['labels']) == 0:
return PointCloud(points)
else:
labels_to_mask = OrderedDict() # masks into the pointcloud per label
n_points = points.shape[0]
for label in lms_dict['labels']:
mask = np.zeros(n_points, dtype=np.bool)
mask[label['mask']] = True
labels_to_mask[label['label']] = mask
# Note that we can pass connectivity as None here and the edges will be
# empty.
return LabelledPointUndirectedGraph.init_from_edges(
points, connectivity, labels_to_mask)
def _parse_ljson_v2(lms_dict):
points = _ljson_parse_null_values(lms_dict["landmarks"]["points"])
connectivity = lms_dict["landmarks"].get("connectivity")
if connectivity is None and len(lms_dict["labels"]) == 0:
lmarks = PointCloud(points)
else:
labels_to_mask = OrderedDict() # masks into the pointcloud per label
n_points = points.shape[0]
for label in lms_dict["labels"]:
mask = np.zeros(n_points, dtype=bool)
mask[label["mask"]] = True
labels_to_mask[label["label"]] = mask
# Note that we can pass connectivity as None here and the edges will be
# empty.
lmarks = LabelledPointUndirectedGraph.init_from_edges(
points, connectivity, labels_to_mask)
return {"LJSON": lmarks}
def _parse_ljson_v1(lms_dict):
from menpo.base import MenpoDeprecationWarning
warnings.warn(
'LJSON v1 is deprecated. export_landmark_file{s}() will '
'only save out LJSON v2 files. Please convert all LJSON '
'files to v2 by importing into Menpo and re-exporting to '
'overwrite the files.', MenpoDeprecationWarning)
all_points = []
labels = [] # label per group
labels_slices = [] # slices into the full pointcloud per label
offset = 0
connectivity = []
for group in lms_dict['groups']:
lms = group['landmarks']
labels.append(group['label'])
labels_slices.append(slice(offset, len(lms) + offset))
# Create the connectivity if it exists
conn = group.get('connectivity', [])
if conn:
# Offset relative connectivity according to the current index
conn = offset + np.asarray(conn)
connectivity += conn.tolist()
for p in lms:
all_points.append(p['point'])
offset += len(lms)
# Don't create a PointUndirectedGraph with no connectivity
points = _ljson_parse_null_values(all_points)
n_points = points.shape[0]
labels_to_masks = OrderedDict()
# go through each label and build the appropriate boolean array
for label, l_slice in zip(labels, labels_slices):
mask = np.zeros(n_points, dtype=np.bool)
mask[l_slice] = True
labels_to_masks[label] = mask
return LabelledPointUndirectedGraph.init_from_edges(
points, connectivity, labels_to_masks)
def _parse_ljson_v3(lms_dict):
all_lms = {}
for key, lms_dict_group in lms_dict['groups'].items():
points = _ljson_parse_null_values(lms_dict_group['landmarks']['points'])
connectivity = lms_dict_group['landmarks'].get('connectivity')
# TODO: create the metadata label!
if connectivity is None and len(lms_dict_group['labels']) == 0:
all_lms[key] = PointCloud(points)
else:
# masks into the pointcloud per label
labels_to_mask = OrderedDict()
n_points = points.shape[0]
for label in lms_dict_group['labels']:
mask = np.zeros(n_points, dtype=np.bool)
mask[label['mask']] = True
labels_to_mask[label['label']] = mask
# Note that we can pass connectivity as None here and the edges
# will be empty.
all_lms[key] = LabelledPointUndirectedGraph.init_from_edges(
points, connectivity, labels_to_mask)
return all_lms
def _parse_ljson_v1(lms_dict):
from menpo.base import MenpoDeprecationWarning
warnings.warn('LJSON v1 is deprecated. export_landmark_file{s}() will '
'only save out LJSON v2 files. Please convert all LJSON '
'files to v2 by importing into Menpo and re-exporting to '
'overwrite the files.', MenpoDeprecationWarning)
all_points = []
labels = [] # label per group
labels_slices = [] # slices into the full pointcloud per label
offset = 0
connectivity = []
for group in lms_dict['groups']:
lms = group['landmarks']
labels.append(group['label'])
labels_slices.append(slice(offset, len(lms) + offset))
# Create the connectivity if it exists
conn = group.get('connectivity', [])
if conn:
# Offset relative connectivity according to the current index
conn = offset + np.asarray(conn)
connectivity += conn.tolist()
for p in lms:
all_points.append(p['point'])
offset += len(lms)
# Don't create a PointUndirectedGraph with no connectivity
points = _ljson_parse_null_values(all_points)
n_points = points.shape[0]
labels_to_masks = OrderedDict()
# go through each label and build the appropriate boolean array
for label, l_slice in zip(labels, labels_slices):
mask = np.zeros(n_points, dtype=np.bool)
mask[l_slice] = True
labels_to_masks[label] = mask
return LabelledPointUndirectedGraph.init_from_edges(points, connectivity, labels_to_masks)
def _parse_ljson_v3(lms_dict):
all_lms = {}
for key, lms_dict_group in lms_dict["groups"].items():
points = _ljson_parse_null_values(
lms_dict_group["landmarks"]["points"])
connectivity = lms_dict_group["landmarks"].get("connectivity")
# TODO: create the metadata label!
if connectivity is None and len(lms_dict_group["labels"]) == 0:
all_lms[key] = PointCloud(points)
else:
# masks into the pointcloud per label
labels_to_mask = OrderedDict()
n_points = points.shape[0]
for label in lms_dict_group["labels"]:
mask = np.zeros(n_points, dtype=np.bool)
mask[label["mask"]] = True
labels_to_mask[label["label"]] = mask
# Note that we can pass connectivity as None here and the edges
# will be empty.
all_lms[key] = LabelledPointUndirectedGraph.init_from_edges(
points, connectivity, labels_to_mask)
return all_lms
def asf_importer(filepath, asset=None, **kwargs):
r"""
Importer for the ASF file format.
For images, the `x` and `y` axes are flipped such that the first axis is
`y` (height in the image domain).
Currently only open and closed path types are supported.
Landmark labels:
+---------+
| label |
+=========+
| all |
+---------+
Parameters
----------
filepath : `Path`
Absolute filepath of the file.
asset : `object`, optional
An optional asset that may help with loading. This is unused for this
implementation.
\**kwargs : `dict`, optional
Any other keyword arguments.
Returns
-------
landmarks : `dict` {`str`: :map:`PointCloud`}
Dictionary mapping landmark groups to menpo shapes
References
----------
.. [1] http://www2.imm.dtu.dk/~aam/datasets/datasets.html
"""
with filepath.open("r") as f:
landmarks = f.read()
# Remove comments and blank lines
landmarks = [l for l in landmarks.splitlines() if (l.rstrip() and not "#" in l)]
# Pop the front of the list for the number of landmarks
count = int(landmarks.pop(0))
# Pop the last element of the list for the image_name
image_name = landmarks.pop()
xs = np.empty([count, 1])
ys = np.empty([count, 1])
connectivity = []
# Only unpack the first 7 (the last 3 are always 0)
split_landmarks = [ASFPath(*landmarks[i].split()[:7]) for i in range(count)]
paths = [list(g) for k, g in itertools.groupby(split_landmarks, lambda x: x[0])]
vert_index = 0
for path in paths:
if path:
path_type = path[0].path_type
for vertex in path:
# Relative coordinates, will be scaled by the image size
xs[vert_index, ...] = float(vertex.xpos)
ys[vert_index, ...] = float(vertex.ypos)
vert_index += 1
# If True, isolated point
if not (
vertex.connects_from == vertex.connects_to
and vertex.connects_to == vertex.point_num
):
# Connectivity is defined by connects_from and connects_to
# as well as the path_type:
# Bit 1: Outer edge point/Inside point
# Bit 2: Original annotated point/Artificial point
# Bit 3: Closed path point/Open path point
# Bit 4: Non-hole/Hole point
# For now we only parse cases 0 and 4 (closed or open)
connectivity.append((int(vertex.point_num), int(vertex.connects_to)))
if path and path_type == "0":
connectivity.append((int(path[-1].point_num), int(path[0].point_num)))
connectivity = np.vstack(connectivity)
points = np.hstack([ys, xs])
if asset is not None:
# we've been given an asset. As ASF files are normalized,
# fix that here
points = Scale(np.array(asset.shape)).apply(points)
labels_to_masks = OrderedDict([("all", np.ones(points.shape[0], dtype=np.bool))])
return {
"ASF": LabelledPointUndirectedGraph.init_from_edges(
points, connectivity, labels_to_masks
)
}
def asf_importer(filepath, asset=None, **kwargs):
r"""
Importer for the ASF file format.
For images, the `x` and `y` axes are flipped such that the first axis is
`y` (height in the image domain).
Currently only open and closed path types are supported.
Landmark labels:
+---------+
| label |
+=========+
| all |
+---------+
Parameters
----------
filepath : `Path`
Absolute filepath of the file.
asset : `object`, optional
An optional asset that may help with loading. This is unused for this
implementation.
\**kwargs : `dict`, optional
Any other keyword arguments.
Returns
-------
landmarks : :map:`LabelledPointUndirectedGraph`
The landmarks including appropriate labels if available.
References
----------
.. [1] http://www2.imm.dtu.dk/~aam/datasets/datasets.html
"""
with open(str(filepath), 'r') as f:
landmarks = f.read()
# Remove comments and blank lines
landmarks = [l for l in landmarks.splitlines()
if (l.rstrip() and not '#' in l)]
# Pop the front of the list for the number of landmarks
count = int(landmarks.pop(0))
# Pop the last element of the list for the image_name
image_name = landmarks.pop()
xs = np.empty([count, 1])
ys = np.empty([count, 1])
connectivity = []
# Only unpack the first 7 (the last 3 are always 0)
split_landmarks = [ASFPath(*landmarks[i].split()[:7])
for i in range(count)]
paths = [list(g)
for k, g in itertools.groupby(split_landmarks, lambda x: x[0])]
vert_index = 0
for path in paths:
if path:
path_type = path[0].path_type
for vertex in path:
# Relative coordinates, will be scaled by the image size
xs[vert_index, ...] = float(vertex.xpos)
ys[vert_index, ...] = float(vertex.ypos)
vert_index += 1
# If True, isolated point
if not (vertex.connects_from == vertex.connects_to and
vertex.connects_to == vertex.point_num):
# Connectivity is defined by connects_from and connects_to
# as well as the path_type:
# Bit 1: Outer edge point/Inside point
# Bit 2: Original annotated point/Artificial point
# Bit 3: Closed path point/Open path point
# Bit 4: Non-hole/Hole point
# For now we only parse cases 0 and 4 (closed or open)
connectivity.append((int(vertex.point_num),
int(vertex.connects_to)))
if path_type == '0':
connectivity.append((int(path[-1].point_num),
int(path[0].point_num)))
connectivity = np.vstack(connectivity)
points = np.hstack([ys, xs])
if asset is not None:
# we've been given an asset. As ASF files are normalized,
# fix that here
points = Scale(np.array(asset.shape)).apply(points)
labels_to_masks = OrderedDict(
[('all', np.ones(points.shape[0], dtype=np.bool))])
return LabelledPointUndirectedGraph.init_from_edges(points, connectivity,
labels_to_masks)
