def save2binvox(reconstructed_volume, data_name1, data_name2):
with open(data_name1, "rb") as f:
bvx = brw.read_as_3d_array(f)
bvx.dims = [256, 256, 256]
bvx.data = reconstructed_volume
with open(data_name2, "wb") as f:
brw.write(bvx, f)
def voxelizePart(grid, scale, translate, dims, cloud, labels, partId,
outputPath):
bbmin = np.min(cloud, axis=0)
bbmax = np.max(cloud, axis=0)
center = 0.5 * (bbmax - bbmin)
w1s = np.where(grid == 1)
grid_xyz = [[x, y, z] for x, y, z in zip(w1s[0], w1s[1], w1s[2])]
grid_xyz = np.array(grid_xyz)
grid_xyz_sc = []
for p in grid_xyz:
trans_p = [0, 0, 0]
trans_p[0] = scale * ((1 / scale) * center[0] - 0.5 + float(
(p[0] + 0.5) / dims)) + translate[0]
trans_p[1] = scale * ((1 / scale) * center[1] - 0.5 + float(
(p[1] + 0.5) / dims)) + translate[1]
trans_p[2] = scale * ((1 / scale) * center[2] - 0.5 + float(
(p[2] + 0.5) / dims)) + translate[2]
grid_xyz_sc.append(trans_p)
grid_xyz_sc = np.array(grid_xyz_sc)
#grid_xyz_sc is now in the same coordinate frame as the point-cloud
clf = knc(n_neighbors=1)
clf.fit(cloud, labels)
voxelLabels = clf.predict(grid_xyz_sc)
partIndices = voxelLabels == partId
partVoxelIndices = grid_xyz[partIndices, :]
partvox = np.zeros((dims, dims, dims, 1))
partvox[partVoxelIndices[:, 0], partVoxelIndices[:, 2],
partVoxelIndices[:, 1], 0] = 1
partvox = partvox.astype('int')
partbinvox = binvox_rw.Voxels(partvox, (dims, dims, dims), [0, 0, 0], 1,
'xzy')
partname = 'model_' + str(partId) + '.binvox'
binvox_rw.write(partbinvox, open(os.path.join(outputPath, partname), 'wb'))
def write_voxel(self, filename):
'''
Write the voxel model data into a .binvox file
'''
if len(filename) != 0 and not self.isempty():
fp = open(filename, 'w')
data = self.voxel > 0
dims = list(self.get_voxel_shape())
translate = [0.0, 0.0, 0.0]
scale = 1.0
axis_order = 'xyz'
model = binvox_rw.Voxels(data, dims, translate, scale, axis_order)
binvox_rw.write(model, fp)
def save_binvox_input_output(input_cloud, mean_voxel, binvox_folder):
import binvox_rw
import curvox
cnn_voxel = shape_completion_utils.round_voxel_grid_to_0_and_1(mean_voxel)
partial_vox = curvox.pc_vox_utils.pc_to_binvox_for_shape_completion(
points=input_cloud[:, 0:3], patch_size=40)
completed_vox = binvox_rw.Voxels(cnn_voxel, partial_vox.dims,
partial_vox.translate, partial_vox.scale,
partial_vox.axis_order)
binvox_rw.write(partial_vox,
open(binvox_folder + "network_input.binvox", 'w'))
binvox_rw.write(completed_vox,
open(binvox_folder + "network_output.binvox", 'w'))
def process_binvox(outfile, res, prediction, fn, idx):
# output voxel .binvox file
# computer center and scale
t = time.time()
MIN = np.min(prediction, 0)
MAX = np.max(prediction, 0)
translate = (MIN + MAX) * 0.5
translate = [float(x) for x in translate]
scale = np.max(MAX - MIN)
dims = [res, res, res]
order = 'xzy'
voxel_data = ((prediction - translate) / scale * res + (res - 1.0) / 2.0).T
with open(outfile, 'wb') as fout:
binvox_rw.write(
binvox_rw.Voxels(np.ascontiguousarray(voxel_data), dims, translate,
scale, order), fout)
print(fn, idx, voxel_data.shape, ' took %f[s]' % (time.time() - t))
def cnn_and_pc_to_mesh(observed_pc,
cnn_voxel,
filepath,
mesh_name,
model_pose,
log_pc=False,
pc_name=""):
cnn_voxel = round_voxel_grid_to_0_and_1(cnn_voxel)
temp_pcd_handle, temp_pcd_filepath = tempfile.mkstemp(suffix=".pcd")
os.close(temp_pcd_handle)
pcd = np_to_pcl(observed_pc)
pcl.save(pcd, temp_pcd_filepath)
partial_vox = curvox.pc_vox_utils.pc_to_binvox_for_shape_completion(
points=observed_pc[:, 0:3], patch_size=40)
completed_vox = binvox_rw.Voxels(cnn_voxel, partial_vox.dims,
partial_vox.translate, partial_vox.scale,
partial_vox.axis_order)
# Now we save the binvox file so that it can be passed to the
# post processing along with the partial.pcd
temp_handle, temp_binvox_filepath = tempfile.mkstemp(
suffix="output.binvox")
os.close(temp_handle)
binvox_rw.write(completed_vox, open(temp_binvox_filepath, 'w'))
# This is the file that the post-processed mesh will be saved it.
mesh_file = file_utils.create_file(filepath, mesh_name)
# mesh_reconstruction tmp/completion.binvox tmp/partial.pcd tmp/post_processed.ply
# This command will look something like
cmd_str = "mesh_reconstruction" + " " + temp_binvox_filepath + " " + temp_pcd_filepath \
+ " " + mesh_file + " --cuda"
subprocess.call(cmd_str.split(" "), stdout=FNULL, stderr=subprocess.STDOUT)
# subprocess.call(cmd_str.split(" "))
if log_pc:
pcd_file = file_utils.create_file(filepath, pc_name)
cmd_str = "pcl_pcd2ply -format 0" + " " + temp_pcd_filepath + " " + pcd_file
subprocess.call(cmd_str.split(" "),
stdout=FNULL,
stderr=subprocess.STDOUT)
map_object_to_gt(pcd_file, model_pose)
map_object_to_gt(mesh_file, model_pose)
def main():
if not len(sys.argv) == 2:
print('python binvox_converter.py input_file_folder')
sys.exit(1)
input_file_folder = sys.argv[1]
if not os.path.exists(input_file_folder) or not os.path.isdir(
input_file_folder):
print('[ERROR] Input folder not exists!')
sys.exit(2)
N_VOX = 32
MESH_EXTENSION = '*.off'
folder_path = os.path.join(input_file_folder, MESH_EXTENSION)
mesh_files = glob(folder_path)
for m_file in mesh_files:
file_path = os.path.join(input_file_folder, m_file)
file_name, ext = os.path.splitext(m_file)
binvox_file_path = os.path.join(input_file_folder,
'%s.binvox' % file_name)
if os.path.exists(binvox_file_path):
print('[WARN] %s File: %s exists. It will be overwritten.' %
(dt.now(), binvox_file_path))
os.remove(binvox_file_path)
print('[INFO] %s Processing file: %s' % (dt.now(), file_path))
rc = subprocess.call([
'binvox', '-d',
str(N_VOX), '-e', '-cb', '-rotx', '-rotx', '-rotx', '-rotz', m_file
])
if not rc == 0:
print('[WARN] %s Failed to convert file: %s' % (dt.now(), m_file))
continue
with open(binvox_file_path, 'rb') as file:
v = binvox_rw.read_as_3d_array(file)
v.data = np.transpose(v.data, (2, 0, 1))
with open(binvox_file_path, 'wb') as file:
binvox_rw.write(v, file)
def pcd2binvox(pcd_filepath, binvox_filepath, grid_dim):
'''
This function takes a pcd filepath, and converts it to
a binvox file. This does not voxeliz the point cloud,
it assumes each point can be represented as indices
ex cloud:
pts = [(0,1,0), (0,10,23), ...]
would be converted to a binvox with voxels
(0,1,0) and (0 , 10, 23) occupied
'''
pc = pcl.load(pcd_filepath)
pc_arr = pc.to_array().astype(int)
pc_mask = (pc_arr[:,0], pc_arr[:,1], pc_arr[:,2])
out = np.zeros((grid_dim, grid_dim, grid_dim))
out[pc_mask] = 1
out_int = out.astype(int)
bv = binvox_rw.Voxels(out_int, out_int.shape, (0,0,0),1, 'xyz')
binvox_rw.write(bv, open(binvox_filepath, 'w'))
def write_binvox_file(pred, filename):
with open(filename, 'wb+') as f:
voxel = binvox_rw.Voxels(pred, [32, 32, 32], [0, 0, 0], 1, 'xzy')
binvox_rw.write(voxel, f)
f.close()
voxel_model = pcvu.pc_to_binvox_for_shape_completion(pcd_as_np, patch_size)
# print(voxel_model)
# print(type(voxel_model))
#is a <class 'binvox_rw.binvox_rw.Voxels'>
#Vs. saving this method gives a solid version even at 256, but looks low resolution
# voxel_model = pcvu.voxelize_points(pcd_as_np, [0,0,0], 1., patch_size, [0,0,0])
# voxel_model = binvox_rw.Voxels(voxel_model, [patch_size,patch_size,patch_size], [0, 0, 0], 0, 'xzy')
#bbox_center = [0,0,0]
#voxel_resolution = 1.
# center = [patch_size/2.,patch_size/2.,patch_size/2.]
# voxel_model = pcvu.voxelize_points(pcd_as_np, bbox_center, voxel_resolution, patch_size, center)
# voxel_model = binvox_rw.Voxels(voxel_model, [patch_size,patch_size,patch_size], [0,0,0], 0, 'xzy')
#Just save out a binvox to look at it and make sure it looks right:
#voxel_model = binvox_rw.Voxels(occupancy_3d_array, [resolution, resolution, resolution], [0, 0, 0], 0, 'xzy')
test_binvox_path = "/home/gregkocher/Desktop/aaaaaaaaa.binvox"
with open(test_binvox_path, 'wb') as fp:
binvox_rw.write(voxel_model, fp, fast=True)
viewvox_path = '/home/gregkocher/CLionProjects/viewvox/viewvox'
test_cmd = '{0} {1}'.format(viewvox_path, test_binvox_path)
test_process = subprocess.Popen(test_cmd, shell=True)
#Now convert from voxels to octree.
#Even if following OGN paper and doing input as 2D image, need the octrees as truth for loss functions.
#Using the scripts provided by the OGN paper authors:
#10/2/2017 actually the OGN repo does not have working code for this.
#Their .../tools/ogn_converter.cpp cannot compile because of various issues.
#Their python code does not go binvox->octree, only the reverse direction.
#So have to make our own...
def _SaveBinvox(vol, filepath):
binvoxvol = binvox_rw.Voxels(vol.voxels, vol.voxels.shape, vol.origin, "%g %g %g" % (vol.voxsize[0], vol.voxsize[1], vol.voxsize[2]), 'xzy')
binvox_rw.write(binvoxvol, filepath)
def test(model, model_pose_filepath, partial_view_filepath,
completion_filepath, gt_center_to_upright_filepath):
pc = pcl.load(partial_view_filepath).to_array()
model_pose = np.load(model_pose_filepath)
gt_center_to_upright = np.load(gt_center_to_upright_filepath)
batch_x = np.zeros(
(1, args.PATCH_SIZE, args.PATCH_SIZE, args.PATCH_SIZE, 1),
dtype=np.float32)
batch_x[
0, :, :, :, :], voxel_resolution, offset = build_test_from_pc_scaled(
pc, args.PATCH_SIZE)
mask = get_occluded_voxel_grid(batch_x[0, :, :, :, 0])
#make batch B2C01 rather than B012C
batch_x = batch_x.transpose(0, 3, 4, 1, 2)
pred = model._predict(batch_x)
# Prediction comes in format [batch number, z-axis, patch number, x-axis,
# y-axis].
pred = pred.reshape(1, args.PATCH_SIZE, 1, args.PATCH_SIZE,
args.PATCH_SIZE)
# Convert prediction to format [batch number, x-axis, y-axis, z-axis,
# patch number].
pred_as_b012c = pred.transpose(0, 3, 4, 1, 2)
completed_region = pred_as_b012c[0, :, :, :, 0]
#output is voxel grid from camera_frame of reference.
output = completed_region * mask
output_vox = output > 0.5
# This is the voxel grid in the camera frame of reference.
vox = binvox_rw.Voxels(output > 0.5,
(args.PATCH_SIZE, args.PATCH_SIZE, args.PATCH_SIZE),
(offset[0], offset[1], offset[2]),
voxel_resolution * args.PATCH_SIZE, "xyz")
'''
#go from voxel grid back to list of points.
#4xn
completion = np.zeros((4, len(output_vox.nonzero()[0])))
completion[0] = (output_vox.nonzero()[0] ) * voxel_resolution + offset[0]
completion[1] = (output_vox.nonzero()[1] ) * voxel_resolution + offset[1]
completion[2] = (output_vox.nonzero()[2] ) * voxel_resolution + offset[2]
completion[3] = 1.0
world_to_camera_transform = np.array([[0, 0, 1, -1],[-1, 0, 0, 0], [0, -1, 0, 0], [0, 0, 0, 1]])
transform = np.dot(model_pose.T, world_to_camera_transform)
# 4xn array
completion_rot = np.dot(transform, completion).T
for i in range(3):
completion_rot[:,i] += gt_center_to_upright[i]
center = get_center(completion_rot[:,0:3])
pc_center_in_voxel_grid = (20, 20, 20)
offset = np.array(center) - np.array(pc_center_in_voxel_grid) * voxel_resolution
patch_center_z = (completion_rot[:, 2].max() + completion_rot[:, 2].min())/2.0
patch_center = (0,0,patch_center_z)
voxel_grid = create_voxel_grid_around_point_scaled(completion_rot[:,0:3], patch_center,
voxel_resolution, args.PATCH_SIZE,
(20,20,20))
vox = binvox_rw.Voxels(voxel_grid[:,:,:,0],
(args.PATCH_SIZE, args.PATCH_SIZE, args.PATCH_SIZE),
(offset[0], offset[1], offset[2]),
voxel_resolution * args.PATCH_SIZE,
"xyz")
'''
#if not os.path.exists(completion_filepath):
binvox_rw.write(vox, open(completion_filepath, 'w'))
def _SaveBinvox(vol, filepath):
binvoxvol = binvox_rw.Voxels(
vol.voxels, vol.voxels.shape, vol.origin,
"%g %g %g" % (vol.voxsize[0], vol.voxsize[1], vol.voxsize[2]), 'xzy')
binvox_rw.write(binvoxvol, filepath)
def completion_cb(self, goal):
rospy.loginfo('Received Completion Goal')
self._feedback = pc_pipeline_msgs.msg.CompletePartialCloudFeedback()
self._result = pc_pipeline_msgs.msg.CompletePartialCloudResult()
temp_pcd_handle, temp_pcd_filepath = tempfile.mkstemp(suffix=".pcd")
os.close(temp_pcd_handle)
partial_pc_np = curvox.cloud_conversions.cloud_msg_to_np(
goal.partial_cloud)
pcd = curvox.cloud_conversions.np_to_pcl(partial_pc_np)
pcl.save(pcd, temp_pcd_filepath)
partial_vox = curvox.pc_vox_utils.pc_to_binvox_for_shape_completion(
points=partial_pc_np[:, 0:3], patch_size=self.patch_size)
batch_x = np.zeros(
(1, self.patch_size, self.patch_size, self.patch_size, 1),
dtype=np.float32)
batch_x[0, :, :, :, 0] = partial_vox.data
if self.flip_batch_x:
rospy.loginfo("Flipping Batch X, if performance is poor,\
try setting flip_batch_x=False")
batch_x = batch_x.transpose(0, 2, 1, 3, 4)
batch_x_new = np.zeros_like(batch_x)
for i in range(40):
batch_x_new[0, i, :, :, 0] = batch_x[0, 40 - i - 1, :, :, 0]
batch_x = batch_x_new
else:
rospy.loginfo("NOT Flipping Batch X, if performance is poor,\
try setting flip_batch_x=True")
# output is the completed voxel grid,
# it is all floats between 0,1 as last layer is softmax
# think of this as probability of occupancy per voxel
# output.shape = (X,Y,Z)
output = self.complete_voxel_grid(batch_x)
if self.flip_batch_x:
rospy.loginfo("flipping batch x back")
output_new = np.zeros_like(output)
for i in range(40):
output_new[i, :, :] = output[40 - i - 1, :, :]
output = output_new
output = output.transpose(1, 0, 2)
# mask the output, so above 0.5 is occupied
# below 0.5 is empty space.
output_vox = np.array(output) > 0.5
# Save the binary voxel grid as an occupancy map
# in a binvox file
completed_vox = binvox_rw.Voxels(output_vox, partial_vox.dims,
partial_vox.translate,
partial_vox.scale,
partial_vox.axis_order)
# Now we save the binvox file so that it can be passed to the
# post processing along with the partial.pcd
temp_handle, temp_binvox_filepath = tempfile.mkstemp(
suffix="output.binvox")
os.close(temp_handle)
binvox_rw.write(completed_vox, open(temp_binvox_filepath, 'w'))
# Now we save the binvox file so that it can be passed to the
# post processing along with the partial.pcd
temp_handle, temp_input_binvox_file = tempfile.mkstemp(
suffix="input.binvox")
os.close(temp_handle)
binvox_rw.write(partial_vox, open(temp_input_binvox_file, 'w'))
# This is the file that the post-processed mesh will be saved it.
temp_handle, temp_completion_filepath = tempfile.mkstemp(suffix=".ply")
os.close(temp_handle)
# This command will look something like
# mesh_reconstruction tmp/completion.binvox tmp/partial.pcd tmp/post_processed.ply
cmd_str = self.post_process_executable + " " + temp_binvox_filepath + " " + temp_pcd_filepath \
+ " " + temp_completion_filepath + " --cuda"
subprocess.call(cmd_str.split(" "))
# Now we are going to read in the post-processed mesh, that is a merge
# of the partial view and of the completion from the CNN
mesh = curvox.mesh_conversions.read_mesh_msg_from_ply_filepath(
temp_completion_filepath)
self._result.mesh = mesh
self._as.set_succeeded(self._result)
rospy.loginfo('Finished Msg')
def align_gt_and_partial(gt_file_path,
gt_mesh_file_path,
model_pose_filepath,
single_view_pointcloud_filepath,
model_name,
partial_name ):
model_pose = np.load(model_pose_filepath)
partial_np_pc = load_partial_np_pc(single_view_pointcloud_filepath)
gt_np_pc = load_gt_np_pc(gt_file_path)
gt_mesh = plyfile.PlyData.read(gt_mesh_file_path)
gt_mesh_vertices = np.zeros((gt_mesh['vertex']['x'].shape[0], 4))
gt_mesh_vertices[:, 0] = gt_mesh['vertex']['x']
gt_mesh_vertices[:, 1] = gt_mesh['vertex']['y']
gt_mesh_vertices[:, 2] = gt_mesh['vertex']['z']
partial_np_pc_cf, gt_np_pc_cf, gt_mesh_vertices_cf = map_to_camera_frame(
partial_np_pc, gt_np_pc, gt_mesh_vertices, model_pose)
partial_np_pc_of, camera_to_object_transform = map_to_object_frame(
partial_np_pc, model_pose)
gt_mesh['vertex']['x'] = gt_mesh_vertices_cf[:, 0]
gt_mesh['vertex']['y'] = gt_mesh_vertices_cf[:, 1]
gt_mesh['vertex']['z'] = gt_mesh_vertices_cf[:, 2]
partial_pcd_pc_cf = pcl.PointCloud(np.array(partial_np_pc_cf.transpose()[:, 0:3], np.float32))
gt_pcd_pc_cf = pcl.PointCloud(np.array(gt_np_pc_cf.transpose()[:, 0:3], np.float32))
partial_pcd_pc_of = pcl.PointCloud(np.array(partial_np_pc_of.transpose()[:, 0:3], np.float32))
gt_pcd_pc_of = pcl.PointCloud(np.array(gt_np_pc[:, 0:3], np.float32))
result_folder = results_dir + model_name + "/" + partial_name + "/"
if not os.path.exists(result_folder):
os.makedirs(result_folder)
partial_filepath = result_folder + "c_partial.pcd"
if not os.path.exists(partial_filepath):
pcl.save(partial_pcd_pc_cf, partial_filepath)
gt_filepath = result_folder + "c_gt.pcd"
if not os.path.exists(gt_filepath):
pcl.save(gt_pcd_pc_cf, gt_filepath)
partial_filepath_of = result_folder + "c_partial_of.pcd"
if not os.path.exists(partial_filepath_of):
pcl.save(partial_pcd_pc_of, partial_filepath_of)
gt_filepath_of = result_folder + "c_gt_of.pcd"
if not os.path.exists(gt_filepath_of):
pcl.save(gt_pcd_pc_of, gt_filepath_of)
gt_mesh_filepath = result_folder + "c_gt.ply"
if True: #not os.path.exists(gt_mesh_filepath):
gt_mesh.text = True
gt_mesh.write(open(gt_mesh_filepath, "w"))
completed_filepath = result_folder + "c_completed.pcd"
if not os.path.exists(completed_filepath):
completed_pcd_cf, completed_pcd_of, completed_binvox = get_completion(partial_np_pc, camera_to_object_transform)
pcl.save(completed_pcd_cf, completed_filepath)
pcl.save(completed_pcd_of, completed_filepath.replace(".pcd", "_of.pcd"))
binvox_filepath = result_folder + "c_completed.binvox"
if not os.path.exists(binvox_filepath):
binvox_rw.write(completed_binvox, open(binvox_filepath, 'w'))
camera_to_object_transform_filepath = result_folder + "camera_to_object.npy"
if not os.path.exists(camera_to_object_transform_filepath):
np.save(camera_to_object_transform_filepath, camera_to_object_transform.T)
def AnalyzeOutput_SingleFile(test_ot_path,truth_ot_path,view=True,ignore_truth=False):
"""
Convert the .ot output file to a binvox file and optionally view the 3d model using viewvox.
Also, load the ground truth .ot file to do basic numerical analysis.
test_ot_path - path to output .ot file
truth_ot_path - path to corresponding ground truth .ot file
e.g.
test_ot_path = '/home/gregkocher/CLionProjects/ogn/examples/shape_from_id/shapenet_cars_256/output/0065.ot'
truth_ot_path = '/home/gregkocher/CLionProjects/ogn/data/shapenet_cars/256_l4/0065.ot'
IOU, NMI = AnalyzeOutput_SingleFile(test_ot_path,truth_ot_path,view=True)
"""
#Load the .ot file that is output from the OGN test net, use the authors' scripts to convert to numpy array
#npy_out_path = test_ot_path.replace('.ot','_voxels.npy')
ot, resolution = python_octree.import_ot(test_ot_path)
output = python_octree.octree_to_voxel_grid(ot, resolution)
# print(output)
# print(output.sum())
# print(output.size)
#np.save(npy_out_path,output)
#Now use binvox_rw python module to convert the numpy array to .binvox file
occupancy_3d_array = output.astype(np.int)
#binvox_rw.Voxels(occupancy_3d_array, data, dims, translate, scale, axis_order)
voxel_model = binvox_rw.Voxels(occupancy_3d_array, [resolution, resolution, resolution], [0, 0, 0], 0, 'xzy')
#Output as a binvox file
test_binvox_path = test_ot_path[:-3] + '.binvox'
with open(test_binvox_path,'wb') as fp:
binvox_rw.write(voxel_model, fp, fast=True)
#binvox_rw.write(voxel_model, fp, fast=False)
IOU = None
NMI = None
if not ignore_truth:
#Also load the ground truth .ot file for comparison
truth_ot, truth_resolution = python_octree.import_ot(truth_ot_path)
truth_output = python_octree.octree_to_voxel_grid(truth_ot, truth_resolution).astype(np.int)
#Now use binvox_rw python module to convert the numpy array to .binvox file
truth_occupancy_3d_array = truth_output.astype(np.int)
#binvox_rw.Voxels(occupancy_3d_array, data, dims, translate, scale, axis_order)
truth_voxel_model = binvox_rw.Voxels(truth_occupancy_3d_array, [truth_resolution, truth_resolution, truth_resolution], [0, 0, 0], 0, 'xzy')
#Output as a binvox file, put it in the test output dir side by side with test output binvox
truth_binvox_path = test_binvox_path.replace('.binvox','_truth.binvox')
with open(truth_binvox_path,'wb') as fp:
binvox_rw.write(truth_voxel_model, fp, fast=True)
#binvox_rw.write(truth_voxel_model, fp, fast=False)
#Now do some basic numerical comparisons of output to ground truth:
#IOU, Jake's mesh Hausdorff....
#intersection_over_union(gt, pr)
IOU = python_octree.intersection_over_union(truth_occupancy_3d_array.flatten(), occupancy_3d_array.flatten())
NMI = metrics.normalized_mutual_info_score(truth_occupancy_3d_array.flatten(), occupancy_3d_array.flatten())
print('IOU', IOU)
print('NMI', NMI)
#print('AMI', metrics.adjusted_mutual_info_score(truth_occupancy_3d_array.flatten(), occupancy_3d_array.flatten()))
#Optionally view the binvox files using viewvox
if view:
viewvox_path = '/home/gregkocher/CLionProjects/viewvox/viewvox'
test_cmd = '{0} {1}'.format(viewvox_path,test_binvox_path)
test_process = subprocess.Popen(test_cmd, shell=True)
if not ignore_truth:
truth_cmd = '{0} {1}'.format(viewvox_path, truth_binvox_path)
truth_process = subprocess.Popen(truth_cmd, shell=True)
return IOU, NMI
rec.update(recall.item())
for j in xrange(len(f)):
modname = mod[j]
squence = seq[j]
# save .binvox
binvox = os.path.join(outdir, modname + '_' + squence + '.binvox')
pred = prediction3[j, :, :, :]
voxel_data = pred.cpu().data.squeeze().numpy().astype('bool')
dims = [64, 64, 64]
translate = [0, 0, 0]
scale = 1.0
axis_order = 'xzy'
with open(binvox, 'wb') as fout:
binvox_rw.write(
binvox_rw.Voxels(np.ascontiguousarray(voxel_data), dims,
translate, scale, axis_order), fout)
print(
'[%d/%d] mean iou img32: %f , mean iou input32: %f , mean iou input64: %f, precision: %f, recall: %f, time: %f'
% (i, len(dataset) / opt.batchSize, iou1.avg, iou2.avg, iou3.avg,
pre.avg, rec.avg, time.time() - t0))
iou1.reset()
iou2.reset()
iou3.reset()
pre.reset()
rec.reset()
network.eval()
with torch.no_grad():
for i, data in enumerate(dataloader_test, 0):
t0 = time.time()
