def test_stitch():
vertices = np.zeros((10, 3), np.float32)
vertices[:, 0] = np.arange(10)
# Make 3 and 6 duplicates of 2 and 4, respectively
vertices[3] = vertices[2]
vertices[6] = vertices[4]
faces = [[0, 1, 2], [3, 4, 5], [6, 7, 8],
[7, 8, 4]] # <- duplicate face (different vertex order, and 4=6)
# After mapping away from dupe vertices
remapped_faces = [[0, 1, 2], [2, 4, 5], [4, 7, 8],
[7, 8, 4]] # duplicated face (different vertex order)
remapped_faces = np.array(remapped_faces)
# After dropping dupe rows
remapped_faces[(remapped_faces > 6)] -= 1
remapped_faces[(remapped_faces > 3)] -= 1
# Drop last face (duplicated)
remapped_faces = remapped_faces[:-1, :]
reduced_vertices = list(vertices)
del reduced_vertices[9] # wasn't referenced to begin with
del reduced_vertices[6] # was duplicated
del reduced_vertices[3] # was duplicated
reduced_vertices = np.asarray(reduced_vertices)
mesh = Mesh(vertices, faces)
mesh.stitch_adjacent_faces()
assert (mesh.faces == remapped_faces).all()
assert (mesh.vertices_zyx == reduced_vertices).all()
def test_smoothing_hexagon(self):
"""
Try 'smoothing' a simple 2D hexagon, which is an easy case to understand.
"""
# This map is correctly labeled with the vertex indices
_ = -1
hexagon = [[[_, _, _, _, _, _, _], [_, _, 0, _, 1, _, _],
[_, _, _, _, _, _, _], [_, 2, _, 3, _, 4, _],
[_, _, _, _, _, _, _], [_, _, 5, _, 6, _, _],
[_, _, _, _, _, _, _]]]
hexagon = 1 + np.array(hexagon)
original_vertices_zyx = np.transpose(hexagon.nonzero())
faces = [[3, 1, 4], [3, 4, 6], [3, 6, 5], [3, 5, 2], [3, 2, 0],
[3, 0, 1]]
mesh = Mesh(original_vertices_zyx, faces)
#mesh.serialize('/tmp/hexagon.obj')
mesh.laplacian_smooth(1)
#mesh.serialize('/tmp/hexagon-smoothed.obj')
# Since vertex 3 is exactly centered between the rest,
# it's location never changes.
assert (mesh.vertices_zyx[3] == original_vertices_zyx[3]).all()
def test_tiny_array():
"""
Tiny arrays trigger an exception in skimage, so they must be padded first.
Verify that they can be meshified (after padding).
"""
one_voxel = np.ones((1, 1, 1), np.uint8)
_mesh = Mesh.from_binary_vol(one_voxel, [(0, 0, 0), (1, 1, 1)],
method='skimage')
_mesh = Mesh.from_binary_vol(one_voxel, [(0, 0, 0), (1, 1, 1)],
method='ilastik')
def test_smoothing_trivial(self):
vertices_zyx = np.array([[0.0, 0.0, 0.0], [0.0, 0.0, 1.0],
[0.0, 0.0, 2.0]])
# This "face" is actually straight line,
# which makes it easy to see what's going on
faces = np.array([[0, 1, 2]])
mesh = Mesh(vertices_zyx, faces)
average_vertex = vertices_zyx.sum(axis=0) / 3
mesh.laplacian_smooth(1)
assert (mesh.vertices_zyx == average_vertex).all()
def test_trim(binary_vol_input):
mesh = Mesh.from_binary_vol(binary_vol_input[0])
mesh.fragment_origin = np.array([25, 25, 25])
mesh.fragment_shape = np.array([50, 50, 50])
num_faces = len(mesh.faces)
mesh.trim()
assert len(mesh.faces) < num_faces
mesh = Mesh.from_binary_vol(binary_vol_input[0])
mesh.fragment_origin = np.array([125, 125, 125])
mesh.fragment_shape = np.array([50, 50, 50])
mesh.trim()
assert len(mesh.faces) == 0
def test_compress(binary_vol_input):
binary_vol, data_box, _nonzero_box = binary_vol_input
mesh_orig = Mesh.from_binary_vol(binary_vol, data_box)
uncompressed_size = mesh_orig.normals_zyx.nbytes + mesh_orig.vertices_zyx.nbytes + mesh_orig.faces.nbytes
mesh = copy.deepcopy(mesh_orig)
size = mesh.compress('lz4')
assert size < uncompressed_size
assert (mesh.faces == mesh_orig.faces).all()
assert (mesh.vertices_zyx == mesh_orig.vertices_zyx).all()
assert (mesh.normals_zyx == mesh_orig.normals_zyx).all()
# Draco is lossy, so we can't compare exactly.
# Just make sure the arrays are at least of the correct shape.
size = mesh.compress('draco')
assert size < uncompressed_size
assert (mesh.faces.shape == mesh_orig.faces.shape)
assert (mesh.vertices_zyx.shape == mesh_orig.vertices_zyx.shape)
assert (mesh.normals_zyx.shape == mesh_orig.normals_zyx.shape)
mesh = copy.deepcopy(mesh_orig)
mesh.fragment_shape = np.asarray(data_box[1])
mesh.fragment_origin = np.asarray(data_box[0])
size = mesh.compress('custom_draco')
assert size < uncompressed_size
assert (mesh.faces.shape == mesh_orig.faces.shape)
assert (mesh.vertices_zyx.shape == mesh_orig.vertices_zyx.shape)
assert (mesh.normals_zyx.shape == mesh_orig.normals_zyx.shape)
def test_basic(self):
# Pretend the data was downsampled and translated,
# and therefore the mesh requires upscaling and translation
data_box = np.array(self.data_box)
data_box += 1000
nonzero_box = self.nonzero_box + 1000
FACTOR = 2
data_box *= FACTOR
nonzero_box *= FACTOR
mesh = Mesh.from_binary_vol(self.binary_vol, data_box)
assert mesh.vertices_zyx.dtype == np.float32
mesh_box = np.array(
[mesh.vertices_zyx.min(axis=0),
mesh.vertices_zyx.max(axis=0)])
assert (mesh_box == nonzero_box
).all(), f"{mesh_box.tolist()} != {nonzero_box.tolist()}"
serialized = mesh.serialize(fmt='obj')
unserialized = mesh.from_buffer(serialized, 'obj')
assert len(unserialized.vertices_zyx) == len(mesh.vertices_zyx)
serialized = mesh.serialize(fmt='drc')
unserialized = mesh.from_buffer(serialized, 'drc')
assert len(unserialized.vertices_zyx) == len(mesh.vertices_zyx)
serialized = mesh.serialize(fmt='ngmesh')
unserialized = mesh.from_buffer(serialized, 'ngmesh')
assert len(unserialized.vertices_zyx) == len(mesh.vertices_zyx)
def test_blockwise(self):
data_box = np.array(self.data_box)
blocks = []
boxes = []
for z in range(0, 100, 20):
for y in range(0, 100, 20):
for x in range(0, 100, 20):
OVERLAP = 1
box = np.asarray([(z, y, x), (z + 20, y + 20, x + 20)],
dtype=int)
box[0] -= OVERLAP
box[1] += OVERLAP
box = np.maximum(box, 0)
box = np.minimum(box, 1 + data_box[1])
block = self.binary_vol[box_to_slicing(*box)]
if block.any():
blocks.append(block)
boxes.append(box)
mesh = Mesh.from_binary_blocks(blocks, boxes)
data_box = np.array(self.data_box)
mesh_box = np.array(
[mesh.vertices_zyx.min(axis=0),
mesh.vertices_zyx.max(axis=0)])
assert (mesh_box == self.nonzero_box
).all(), f"{mesh_box.tolist()} != {self.nonzero_box.tolist()}"
def process_body(body_id):
with resource_mgr_client.access_context( input_config["server"], True, 1, 0 ):
tar_bytes = fetch_tarfile(server, uuid, tsv_instance, body_id)
sv_meshes = Mesh.from_tarfile(tar_bytes, concatenate=False)
sv_meshes = {int(os.path.splitext(name)[0]): m for name, m in sv_meshes.items()}
total_body_vertices = sum([len(m.vertices_zyx) for m in sv_meshes.values()])
decimation = min(1.0, max_body_vertices / total_body_vertices)
try:
_process_sv = partial(process_sv, decimation, decimation_lib, max_sv_vertices, output_format)
if num_procs <= 1:
output_table = [*starmap(_process_sv, sv_meshes.items())]
else:
output_table = compute_parallel(_process_sv, sv_meshes.items(), starmap=True, processes=num_procs, ordered=False, show_progress=False)
cols = ['sv', 'orig_vertices', 'final_vertices', 'final_decimation', 'effective_decimation', 'mesh_bytes']
output_df = pd.DataFrame(output_table, columns=cols)
output_df['body'] = body_id
output_df['error'] = ""
write_sv_meshes(output_df, output_config, output_format, resource_mgr_client)
except Exception as ex:
svs = [*sv_meshes.keys()]
orig_vertices = [len(m.vertices_zyx) for m in sv_meshes.values()]
output_df = pd.DataFrame({'sv': svs, 'orig_vertices': orig_vertices})
output_df['final_vertices'] = -1
output_df['final_decimation'] = -1
output_df['effective_decimation'] = -1
output_df['mesh_bytes'] = -1
output_df['body'] = body_id
output_df['error'] = str(ex)
return output_df.drop(columns=['mesh_bytes'])
def test_blockwise_simple(self):
"""
Simple test case to manually explore the output
of marching cubes as computed in blocks without halo.
"""
_ = 0
img = [[_, _, _, _, _, 1, _, _], [_, 1, _, _, _, _, _, _],
[_, _, 1, 1, 1, 1, 1, _], [_, 1, 1, 1, 1, 1, 1, _],
[_, 1, 1, 1, 1, 1, 1, _], [_, 1, 1, 1, 1, 1, 1, _],
[_, 1, 1, 1, 1, 1, 1, _], [_, _, _, _, _, _, _, _]]
vol = np.zeros((3, 8, 8), dtype=bool)
vol[1] = img
blocks = (vol[:, 0:4, 0:4], vol[:, 0:4, 4:8], vol[:, 4:8,
0:4], vol[:, 4:8,
4:8])
starts = [[0, 0, 0], [0, 0, 4], [0, 4, 0], [0, 4, 4]]
starts = np.array(starts)
boxes = np.zeros((4, 2, 3), np.uint32)
boxes[:, 0, :] = starts
boxes[:, 1, :] = starts + (3, 4, 4)
_mesh = Mesh.from_binary_blocks(blocks[3:4], boxes[3:4], stitch=False)
def test_pickling_empty(self):
mesh = Mesh(np.zeros((0, 3), np.float32), np.zeros((0, 3), np.uint32))
pickled = pickle.dumps(mesh)
unpickled = pickle.loads(pickled)
assert len(unpickled.vertices_zyx) == 0
assert len(unpickled.faces) == 0
def test_smoothing_X(self):
"""
This just exercises the code on our standard X-shaped
test object, but doesn't verify the results.
Uncomment the serialize() lines to inspect the effects manually.
"""
mesh = Mesh.from_binary_vol(self.binary_vol, self.data_box)
#mesh.serialize('/tmp/x-unsmoothed.obj')
mesh.simplify(0.2)
mesh.laplacian_smooth(5)
#mesh.serialize('/tmp/x-simplified-smoothed.obj')
mesh = Mesh.from_binary_vol(self.binary_vol, self.data_box)
mesh.laplacian_smooth(5)
mesh.simplify(0.2)
def setUp(self):
self.vertexes_1 = np.array([[0, 0, 0], [0, 1, 0], [0, 1, 1]])
self.faces_1 = np.array([[2, 1, 0]])
self.vertexes_2 = np.array([[0, 0, 1], [0, 2, 0], [0, 2, 2]])
self.faces_2 = np.array([[2, 1, 0], [1, 2, 0]])
self.vertexes_3 = np.array([[1, 0, 1], [1, 2, 0], [1, 2, 2]])
self.faces_3 = np.array([[1, 2, 0]])
self.mesh_1 = Mesh(self.vertexes_1, self.faces_1)
self.mesh_2 = Mesh(self.vertexes_2, self.faces_2)
self.mesh_3 = Mesh(self.vertexes_3, self.faces_3)
self.mesh_4 = Mesh(np.zeros((0, 3), np.float32),
np.zeros((0, 3), np.uint32)) # Empty mesh
def test_pickling(self):
mesh = Mesh.from_binary_vol(self.binary_vol)
pickled = pickle.dumps(mesh)
unpickled = pickle.loads(pickled)
# It's not easy to verify that unpickled is identical,
# since draco may re-order vertices and faces.
# The validity of our draco encoding functions is tested elsewhere,
# so here we just check for vertex/face count
assert len(mesh.vertices_zyx) == len(unpickled.vertices_zyx)
assert len(mesh.faces) == len(unpickled.faces)
def tiny_meshes():
vertexes_1 = np.array([[0, 0, 0], [0, 1, 0], [0, 1, 1]])
faces_1 = np.array([[2, 1, 0]])
vertexes_2 = np.array([[0, 0, 1], [0, 2, 0], [0, 2, 2]])
faces_2 = np.array([[2, 1, 0], [1, 2, 0]])
vertexes_3 = np.array([[1, 0, 1], [1, 2, 0], [1, 2, 2]])
faces_3 = np.array([[1, 2, 0]])
mesh_1 = Mesh(vertexes_1, faces_1)
mesh_2 = Mesh(vertexes_2, faces_2)
mesh_3 = Mesh(vertexes_3, faces_3)
mesh_4 = Mesh(np.zeros((0, 3), np.float32), np.zeros(
(0, 3), np.uint32)) # Empty mesh
return mesh_1, mesh_2, mesh_3, mesh_4
def test_basic(binary_vol_input):
binary_vol, data_box, nonzero_box = binary_vol_input
# Pretend the data was downsampled and translated,
# and therefore the mesh requires upscaling and translation
data_box = np.array(data_box)
data_box += 1000
nonzero_box = nonzero_box + 1000
FACTOR = 2
data_box *= FACTOR
nonzero_box *= FACTOR
mesh = Mesh.from_binary_vol(binary_vol, data_box)
assert mesh.vertices_zyx.dtype == np.float32
mesh_box = np.array(
[mesh.vertices_zyx.min(axis=0),
mesh.vertices_zyx.max(axis=0)])
assert (mesh_box == nonzero_box
).all(), f"{mesh_box.tolist()} != {nonzero_box.tolist()}"
serialized = mesh.serialize(fmt='obj')
unserialized = mesh.from_buffer(serialized, 'obj')
assert len(unserialized.vertices_zyx) == len(mesh.vertices_zyx)
serialized = mesh.serialize(fmt='drc')
unserialized = mesh.from_buffer(serialized, 'drc')
assert len(unserialized.vertices_zyx) == len(mesh.vertices_zyx)
mesh = Mesh.from_binary_vol(binary_vol,
data_box,
fragment_shape=np.asarray(data_box[1]),
fragment_origin=np.asarray(data_box[0]))
serialized = mesh.serialize(fmt='custom_drc')
unserialized = mesh.from_buffer(serialized, 'drc')
assert len(unserialized.vertices_zyx) == len(mesh.vertices_zyx)
serialized = mesh.serialize(fmt='ngmesh')
unserialized = mesh.from_buffer(serialized, 'ngmesh')
assert len(unserialized.vertices_zyx) == len(mesh.vertices_zyx)
def test_solid_array(self):
"""
Solid volumes can't be meshified. An empty mesh is returned instead.
"""
box = [(0, 0, 0), (3, 3, 3)]
solid_volume = np.ones((3, 3, 3), np.uint8)
mesh = Mesh.from_binary_vol(solid_volume, box)
assert mesh.vertices_zyx.shape == (0, 3)
assert mesh.faces.shape == (0, 3)
assert mesh.normals_zyx.shape == (0, 3)
assert (mesh.box == box).all()
def test_normals_implementations(binary_vol_input):
"""
Compare the numpy-based and numba-based normals computation implementations.
"""
binary_vol, data_box, _nonzero_box = binary_vol_input
try:
from vol2mesh.normals import (compute_face_normals,
compute_face_normals_numba,
compute_face_normals_numpy,
compute_vertex_normals,
compute_vertex_normals_numba,
compute_vertex_normals_numpy)
except ImportError:
pytest.skip("numba not installed")
mesh = Mesh.from_binary_vol(binary_vol, data_box)
face_normals_default = compute_face_normals(mesh.vertices_zyx,
mesh.faces,
normalize=True)
face_normals_numba = compute_face_normals_numba(mesh.vertices_zyx,
mesh.faces,
normalize=True)
face_normals_numpy = compute_face_normals_numpy(mesh.vertices_zyx,
mesh.faces,
normalize=True)
assert np.allclose(face_normals_numba, face_normals_default)
assert np.allclose(face_normals_numba, face_normals_numpy)
vertex_normals_default = compute_vertex_normals(
mesh.vertices_zyx,
mesh.faces,
weight_by_face_area=False,
face_normals=face_normals_default)
vertex_normals_numba = compute_vertex_normals_numba(
mesh.vertices_zyx,
mesh.faces,
weight_by_face_area=False,
face_normals=face_normals_numba)
vertex_normals_numpy = compute_vertex_normals_numpy(
mesh.vertices_zyx,
mesh.faces,
weight_by_face_area=False,
face_normals=face_normals_numpy)
assert np.allclose(vertex_normals_numba, vertex_normals_default)
assert np.allclose(vertex_normals_numba, vertex_normals_numpy)
def test_compress(self):
mesh_orig = Mesh.from_binary_vol(self.binary_vol, self.data_box)
uncompressed_size = mesh_orig.normals_zyx.nbytes + mesh_orig.vertices_zyx.nbytes + mesh_orig.faces.nbytes
mesh = copy.deepcopy(mesh_orig)
size = mesh.compress('lz4')
assert size < uncompressed_size
assert (mesh.faces == mesh_orig.faces).all()
assert (mesh.vertices_zyx == mesh_orig.vertices_zyx).all()
assert (mesh.normals_zyx == mesh_orig.normals_zyx).all()
# Draco is lossy, so we can't compare exactly.
# Just make sure the arrays are at least of the correct shape.
size = mesh.compress('draco')
assert size < uncompressed_size
assert (mesh.faces.shape == mesh_orig.faces.shape)
assert (mesh.vertices_zyx.shape == mesh_orig.vertices_zyx.shape)
assert (mesh.normals_zyx.shape == mesh_orig.normals_zyx.shape)
def process_sv(decimation, decimation_lib, max_sv_vertices, output_format, sv: int, mesh: Mesh):
try:
orig_vertices = len(mesh.vertices_zyx)
if orig_vertices == 0:
final_decimation = 1.0
else:
final_decimation = min(decimation, max_sv_vertices / len(mesh.vertices_zyx))
if decimation_lib == "openmesh":
mesh.simplify_openmesh(final_decimation)
elif decimation_lib == "fq-in-memory":
mesh.simplify(decimation, True)
elif decimation_lib == "fq-via-disk":
mesh.simplify(decimation, False)
else:
raise AssertionError()
final_vertices = len(mesh.vertices_zyx)
effective_decimation = final_vertices / orig_vertices
mesh_bytes = mesh.serialize(fmt=output_format)
return sv, orig_vertices, final_vertices, final_decimation, effective_decimation, mesh_bytes
except Exception as ex:
raise RuntimeError(f"Failed processing SV {sv}: {type(ex)}") from ex
def test_empty_mesh(self):
"""
What happens when we call functions on an empty mesh?
"""
mesh = Mesh(np.zeros((0, 3), np.float32), np.zeros((0, 3), int))
mesh.simplify(1.0)
assert len(mesh.vertices_zyx) == len(mesh.normals_zyx) == len(
mesh.faces) == 0
mesh.simplify(0.1)
assert len(mesh.vertices_zyx) == len(mesh.normals_zyx) == len(
mesh.faces) == 0
mesh.laplacian_smooth(0)
assert len(mesh.vertices_zyx) == len(mesh.normals_zyx) == len(
mesh.faces) == 0
mesh.laplacian_smooth(2)
assert len(mesh.vertices_zyx) == len(mesh.normals_zyx) == len(
mesh.faces) == 0
mesh.stitch_adjacent_faces()
assert len(mesh.vertices_zyx) == len(mesh.normals_zyx) == len(
mesh.faces) == 0
mesh.serialize(fmt='obj')
assert len(mesh.vertices_zyx) == len(mesh.normals_zyx) == len(
mesh.faces) == 0
mesh.serialize(fmt='drc')
assert len(mesh.vertices_zyx) == len(mesh.normals_zyx) == len(
mesh.faces) == 0
mesh.compress()
concatenate_meshes((mesh, mesh))
assert len(mesh.vertices_zyx) == len(mesh.normals_zyx) == len(
mesh.faces) == 0
def test_normals_guarantees(self):
"""
Member functions have guarantees about whether normals are present or absent after the function runs.
- simplify(): Always present afterwards
- laplacian_smooth(): Always present afterwards
- stitch_adjacent_faces(): Present afterwards IFF they were present before.
"""
data_box = np.array(self.data_box)
FACTOR = 2
data_box *= FACTOR
mesh_orig = Mesh.from_binary_vol(self.binary_vol, data_box)
mesh = copy.deepcopy(mesh_orig)
assert mesh.normals_zyx.shape[0] > 1
# Verify normals are always present after simplification,
# Regardless of whether or not they were present before,
# or if simplification was even performed.
mesh.simplify(1.0)
assert mesh.normals_zyx.shape[0] > 1
mesh.simplify(0.5)
assert mesh.normals_zyx.shape[0] > 1
mesh.drop_normals()
mesh.simplify(0.5)
assert mesh.normals_zyx.shape[0] > 1
# Verify normals are always present after smoothing,
# Regardless of whether or not they were present before,
# or if smoothing was even performed.
mesh = copy.deepcopy(mesh_orig)
mesh.laplacian_smooth(0)
assert mesh.normals_zyx.shape[0] > 1
mesh.laplacian_smooth(2)
assert mesh.normals_zyx.shape[0] > 1
mesh.drop_normals()
mesh.laplacian_smooth(2)
assert mesh.normals_zyx.shape[0] > 1
# Verify that the presence or absence of normals is the SAME after stitching,
# Whether or not stitching had any effect.
# no stitching, keep normals
mesh = copy.deepcopy(mesh_orig)
stitching_performed = mesh.stitch_adjacent_faces()
assert not stitching_performed
assert mesh.normals_zyx.shape[0] > 1
# no stitching, no normals in the first place
mesh.drop_normals()
stitching_performed = mesh.stitch_adjacent_faces()
assert not stitching_performed
assert mesh.normals_zyx.shape[0] == 0
# stitching, generate normals
mesh = copy.deepcopy(mesh_orig)
duplicated_mesh = concatenate_meshes([mesh, mesh])
assert duplicated_mesh.normals_zyx.shape[0] > 1
stitching_performed = duplicated_mesh.stitch_adjacent_faces()
assert stitching_performed
assert duplicated_mesh.normals_zyx.shape[0] > 1
# stitching, no normals in the first place
mesh = copy.deepcopy(mesh_orig)
duplicated_mesh = concatenate_meshes([mesh, mesh])
duplicated_mesh.drop_normals()
stitching_performed = duplicated_mesh.stitch_adjacent_faces()
assert stitching_performed
assert duplicated_mesh.normals_zyx.shape[0] == 0