def __getitem__(self, index):
refine_path, input_path, gt_path, i, j, k, f, mod, seq = self.random_data_paths[
index]
fp = open(refine_path, 'rb')
refine = binvox_rw.read_as_3d_array(fp, fix_coords=False).data
refine = refine.astype('float32')
fp = open(input_path, 'rb')
input = binvox_rw.read_as_3d_array(fp, fix_coords=False).data
input = input.astype('float32')
# read shapenet alignent ground truth
fp = open(gt_path, 'rb')
voxel_data = binvox_rw.read_as_3d_array(fp, fix_coords=False).data
gt = np.zeros((2, 128, 128, 128), dtype='float32')
gt[0, :, :, :] = voxel_data < 1
gt[1, :, :, :] = voxel_data
refine = refine[i * 32:(i + 1) * 32, j * 32:(j + 1) * 32,
k * 32:(k + 1) * 32]
input = input[i * 64:(i + 1) * 64, j * 64:(j + 1) * 64,
k * 64:(k + 1) * 64]
gt = gt[:, i * 64:(i + 1) * 64, j * 64:(j + 1) * 64,
k * 64:(k + 1) * 64]
refine = torch.from_numpy(refine[None, :, :, :]).type(
torch.FloatTensor)
input = torch.from_numpy(input[None, :, :, :]).type(torch.FloatTensor)
gt = torch.from_numpy(gt).type(torch.FloatTensor)
return refine, input, gt, i, j, k, f, mod, seq
def read_coords_from_binvox(fixed_points_file,
random_points_file,
edge_file,
offset=0,
f_all=True,
half=False):
with open(fixed_points_file, 'rb') as f, open(random_points_file,
'rb') as r:
fixed = read_as_3d_array(f)
random = read_as_3d_array(r)
f_coords = np.transpose(dense_to_sparse(fixed.data), (1, 0)).copy()
r_coords = np.transpose(dense_to_sparse(random.data), (1, 0)).copy()
if half:
f_idx = []
for i in range(len(f_coords)):
if f_coords[i][0] < 256 and np.random.random() < 0.3:
f_idx.append(i)
r_idx = []
for i in range(len(r_coords)):
if r_coords[i][0] < 256 and np.random.random() < 0.03:
r_idx.append(i)
elif not f_all:
f_idx = np.random.choice(f_coords.shape[0], 900, replace=False)
r_idx = np.random.choice(r_coords.shape[0], 300, replace=False)
edge_list = np.load(edge_file) + offset
if f_all:
return f_coords, r_coords, edge_list
new_edge_list = []
for e in edge_list:
if e[0] in f_idx and e[1] in f_idx:
new_edge_list.append([
np.where(f_idx == e[0])[0][0],
np.where(f_idx == e[1])[0][0]
])
return f_coords[f_idx], r_coords[r_idx], new_edge_list
def demo():
'''
demo on how to use this class.
'''
import binvox_rw
# Constructor
sc = Shape_complete(verbose=True)
# Read demo binvox as (64*64*64) array
with open('demo/occupy.binvox', 'rb') as f:
occ = binvox_rw.read_as_3d_array(f).data
with open('demo/non_occupy.binvox', 'rb') as f:
non = binvox_rw.read_as_3d_array(f).data
# Complete shape
out = sc.complete(occ=occ, non=non, verbose=False)
# Thresholding. Threshold sets to be 0.5
th = 0.5
out[out >= th] = 1
out[out < th] = 0
# Save to file for demo
vox = binvox_rw.Voxels(out, [64, 64, 64], [0, 0, 0], 1, 'xyz')
with open('demo/output.binvox', 'wb') as f:
vox.write(f)
print('Output saved to demo/output.binvox.')
print(
'Please use ./viewvox demo/output.binvox to visualize the result.')
def get_batch(self):
if self.idx >= len(self.data_paths):
self.idx = 0
refine_path, input_path, gt_path, f, mod, seq = self.data_paths[self.idx]
fp = open(refine_path, 'rb')
refine = binvox_rw.read_as_3d_array(fp, fix_coords=False).data
refine = refine.astype('float32')
fp = open(input_path, 'rb')
input = binvox_rw.read_as_3d_array(fp, fix_coords=False).data
input = input.astype('float32')
fp = open(gt_path, 'rb')
voxel_data = binvox_rw.read_as_3d_array(fp, fix_coords=False).data
gt = np.zeros((128, 128, 128), dtype='int64')
gt[:] = voxel_data
refine_batch = []
input_batch = []
gt_batch = []
for i in range(2):
for j in range(2):
for k in range(2):
refine_batch.append(refine[None, None, i * 32:(i + 1) * 32, j * 32:(j + 1) * 32, k * 32:(k + 1) * 32])
input_batch.append(input[None, None, i * 64:(i + 1) * 64, j* 64:(j + 1) * 64, k * 64:(k + 1) * 64])
gt_batch.append(gt[None, i * 64:(i + 1) * 64, j * 64:(j + 1) * 64, k * 64:(k + 1) * 64])
refine_batch = np.concatenate(refine_batch, axis=0)
input_batch = np.concatenate(input_batch, axis=0)
gt_batch = np.concatenate(gt_batch, axis=0)
refine_batch = torch.from_numpy(refine_batch).type(torch.FloatTensor)
input_batch = torch.from_numpy(input_batch).type(torch.FloatTensor)
gt_batch = torch.from_numpy(gt_batch).type(torch.LongTensor)
self.idx += 1
return refine_batch, input_batch, gt_batch, f, mod, seq
def demo():
'''
Publish sample data to ROS
'''
global base_path
base_path = rospkg.RosPack().get_path('mps_shape_completion')
# Read demo binvox as (64*64*64) array
with open(base_path + '/demo/occupy.binvox', 'rb') as f:
occ = binvox_rw.read_as_3d_array(f).data
with open(base_path + '/demo/non_occupy.binvox', 'rb') as f:
non = binvox_rw.read_as_3d_array(f).data
rospy.init_node('shape_demo_loader')
rospy.wait_for_service('complete_shape')
pub = rospy.Publisher('local_occupancy',
numpy_msg(ByteMultiArray),
queue_size=10)
rospy.Subscriber("local_occupancy_predicted", numpy_msg(ByteMultiArray),
callback)
time.sleep(1)
print("Requesting shape completion")
pub.publish(vox_to_msg(occ))
rospy.spin()
def read_voxel(self):
'''
Reads in a triangulated 3D model file (.obj, .stl, etc.), rasterizes
it using 'binvox', and saves the data as 3D numpy array of 0's and 1's
'''
if len(self.filename) != 0:
binvox = self.filename[:self.filename.rfind('.')] + '.binvox'
if not os.path.isfile(binvox):
subprocess.call("./binvox -d "+ str(self.size) +
" " + self.filename, shell = True)
fid = open(binvox, 'r')
model = binvox_rw.read_as_3d_array(fid)
if model.dims[0] != self.size:
os.remove(binvox)
subprocess.call("./binvox -d "+ str(self.size) +
" " + self.filename, shell = True)
fid = open(binvox, 'r')
model = binvox_rw.read_as_3d_array(fid)
self.voxel = 1*model.data
if self.scale != 1:
self.pad_voxel([self.resolution] * 3)
def evaluate_instance(file_name, gtfolder=None):
if 'imview' in file_name or '_1' in file_name:
return [0]
prefl = file_name
names = prefl.split('/')
cat = names[-3]
md5 = names[-2]
gtfl = '%s/%s/%s/model-0.45.binvox' % (gtfolder, cat, md5)
try:
with open(prefl, 'rb') as f:
data = binvox_rw.read_as_3d_array(f)
with open(gtfl, 'rb') as f:
data2 = binvox_rw.read_as_3d_array(f)
except:
print('Error in read data')
print(prefl)
print(gtfl)
return [0]
iouall = data.data | data2.data
iouoverlap = data.data & data2.data
iouthis = np.sum(iouoverlap) / (np.sum(iouall) + 1e-8)
iouthisgt = np.sum(iouoverlap) / (np.sum(data2.data) + 1e-8)
iouthispre = np.sum(iouoverlap) / (np.sum(data.data) + 1e-8)
return [cat, iouthis, iouthisgt, iouthispre]
def demo():
'''
Publish sample data to ROS
'''
# Read demo binvox as (64*64*64) array
with open('demo/occupy.binvox', 'rb') as f:
occ = binvox_rw.read_as_3d_array(f).data
with open('demo/non_occupy.binvox', 'rb') as f:
non = binvox_rw.read_as_3d_array(f).data
msg = ByteMultiArray()
msg.data = (occ.astype(int) - non.astype(int)).flatten().tolist()
msg.layout.dim.append(
MultiArrayDimension(label='x', size=DIM, stride=DIM * DIM * DIM))
msg.layout.dim.append(
MultiArrayDimension(label='y', size=DIM, stride=DIM * DIM))
msg.layout.dim.append(MultiArrayDimension(label='z', size=DIM, stride=DIM))
rospy.init_node('shape_demo_loader')
pub = rospy.Publisher('local_occupancy',
numpy_msg(ByteMultiArray),
queue_size=10)
rospy.Subscriber("local_occupancy_predicted", numpy_msg(ByteMultiArray),
callback)
time.sleep(5)
pub.publish(msg)
rospy.spin()
def __getitem__(self, index):
rgb_path, input32_path, gt32_path, input64_path, gt64_path, f, mod, seq = self.data_paths[index]
if self.train:
im = Image.open(rgb_path)
im = self.dataAugmentation(im) # random crop
else:
im = Image.open(rgb_path)
im = self.validating(im) # center crop
data = self.transforms(im) # scale
data = data[:3, :, :]
fp = open(input32_path, 'rb')
input32 = binvox_rw.read_as_3d_array(fp, fix_coords=False).data
input32 = input32.astype('float32')
input32 = torch.from_numpy(input32[None, :, :, :]).type(torch.FloatTensor)
fp = open(gt32_path, 'rb')
voxel_data = binvox_rw.read_as_3d_array(fp, fix_coords=False).data
gt32 = np.zeros((32, 32, 32), dtype='int64')
gt32[:] = voxel_data
gt32 = torch.from_numpy(gt32).type(torch.LongTensor)
fp = open(input64_path, 'rb')
input64 = binvox_rw.read_as_3d_array(fp, fix_coords=False).data
input64 = input64.astype('float32')
input64 = torch.from_numpy(input64[None, :, :, :]).type(torch.FloatTensor)
fp = open(gt64_path, 'rb')
voxel_data = binvox_rw.read_as_3d_array(fp, fix_coords=False).data
gt64 = np.zeros((64, 64, 64), dtype='int64')
gt64[:] = voxel_data
gt64 = torch.from_numpy(gt64).type(torch.LongTensor)
return data, input32, gt32, input64, gt64, f, mod, seq
def jaccard_similarity(mesh_filepath0,
mesh_filepath1,
grid_size=40,
exact=True):
temp_mesh0_filepath = tempfile.mktemp(suffix=".ply")
temp_mesh1_filepath = tempfile.mktemp(suffix=".ply")
binvox0_filepath = temp_mesh0_filepath.replace(".ply", ".binvox")
binvox1_filepath = temp_mesh1_filepath.replace(".ply", ".binvox")
os.symlink(os.path.abspath(mesh_filepath0), temp_mesh0_filepath)
os.symlink(os.path.abspath(mesh_filepath1), temp_mesh1_filepath)
mesh0 = plyfile.PlyData.read(temp_mesh0_filepath)
minx, maxx = minmax(mesh0['vertex']['x'])
miny, maxy = minmax(mesh0['vertex']['y'])
minz, maxz = minmax(mesh0['vertex']['z'])
# -d: specify voxel grid size (default 256, max 1024)(no max when using -e)
# -e: exact voxelization (any voxel with part of a triangle gets set)(does not use graphics card)
# -bb <minx> <miny> <minz> <maxx> <maxy> <maxz>: force a different input model bounding box
cmd_base = "binvox -pb "
if exact:
cmd_base += "-e "
cmd_base += "-d " + str(grid_size) + " -bb " + str(minx) + " " + str(
miny) + " " + str(minz) + " " + str(maxx) + " " + str(
maxy) + " " + str(maxz)
mesh0_cmd = cmd_base + " " + temp_mesh0_filepath
mesh1_cmd = cmd_base + " " + temp_mesh1_filepath
process = subprocess.Popen(mesh0_cmd.split(" "), stdout=subprocess.PIPE)
command1_output, _ = process.communicate()
process = subprocess.Popen(mesh1_cmd.split(" "), stdout=subprocess.PIPE)
command2_output, _ = process.communicate()
with open(binvox0_filepath, 'r') as mesh0_binvox_file:
mesh0_binvox = binvox_rw.read_as_3d_array(mesh0_binvox_file)
with open(binvox1_filepath, 'r') as mesh1_binvox_file:
mesh1_binvox = binvox_rw.read_as_3d_array(mesh1_binvox_file)
jaccard = _jaccard_distance(mesh0_binvox.data, mesh1_binvox.data)
if os.path.exists(temp_mesh0_filepath):
os.remove(temp_mesh0_filepath)
if os.path.exists(temp_mesh1_filepath):
os.remove(temp_mesh1_filepath)
if os.path.exists(binvox0_filepath):
os.remove(binvox0_filepath)
if os.path.exists(binvox1_filepath):
os.remove(binvox1_filepath)
return jaccard
def get_model(self, material_name, product_name):
'''
Get 3d-Array form of obj. file
'''
with open('model_binvox/%s' % material_name, 'rb') as f:
material_binvox = binvox_rw.read_as_3d_array(f)
material = np.array([material_binvox.data])[0, :, :, :]
with open('model_binvox/%s' % product_name, 'rb') as f:
product_binvox = binvox_rw.read_as_3d_array(f)
product = np.array([product_binvox.data])[0, :, :, :]
return material, product
def make_boundary(self, hx):
boundary = (hx == -2)
all_vox_files = glob.glob('../../Flow-Sculpter/data/train/**/*.binvox')
num_file_try = np.random.randint(2, 6)
for i in xrange(num_file_try):
file_ind = np.random.randint(0, len(all_vox_files))
with open(all_vox_files[file_ind], 'rb') as f:
model = binvox_rw.read_as_3d_array(f)
model = model.data[:, :, model.dims[2] / 2]
model = np.array(model, dtype=np.int)
model = np.pad(model, ((1, 1), (1, 1)),
'constant',
constant_values=0)
floodfill(model, 0, 0)
model = np.greater(model, -0.1)
pos_x = np.random.randint(1, hx.shape[0] - model.shape[0] - 1)
pos_y = np.random.randint(1, hx.shape[1] - model.shape[1] - 1)
boundary[pos_x:pos_x + model.shape[0], pos_y:pos_y +
model.shape[0]] = model | boundary[pos_x:pos_x +
model.shape[0],
pos_y:pos_y +
model.shape[0]]
return boundary
def test(self):
models_dir = '/srv/3d_conv_data/ModelNet10'
categories = [
d for d in os.listdir(models_dir)
if os.path.isdir(os.path.join(models_dir, d))
]
examples = []
subdir = '/' + 'train' + '/'
for category in categories:
for file_name in os.listdir(models_dir + '/' + category + subdir):
if ".binvox" in file_name:
examples.append(models_dir + '/' + category + subdir +
file_name)
subdir = '/' + 'test' + '/'
for category in categories:
for file_name in os.listdir(models_dir + '/' + category + subdir):
if ".binvox" in file_name:
examples.append(models_dir + '/' + category + subdir +
file_name)
for example in examples:
with open(example, 'rb') as f:
model = binvox_rw.read_as_3d_array(f).data
if model.max() == 0:
print example
def binvox_to_step(binvox_file,
voxel_length,
voxel_width,
voxel_height,
application_protocol="AP203"):
"""function used to change binvox file to step file
binvox_file: the binvox file ('chair.binvox' etc.)
voxel_length: the length of one voxel
voxel_width: the width of one voxel
voxel_height: the height of one voxel
application protocol: "AP203" or "AP214IS" or "AP242DIS"
"""
with open(binvox_file, 'rb') as f:
model = binvox_rw.read_as_3d_array(f)
voxel = voxel_to_TopoDS(model, voxel_length, voxel_width, voxel_height)
# initialize the STEP exporter
step_writer = STEPControl_Writer()
Interface_Static_SetCVal("write.step.schema", application_protocol)
# transfer shapes and write file
step_writer.Transfer(voxel, STEPControl_AsIs)
status = step_writer.Write(binvox_file[:-6] + "stp")
if status != IFSelect_RetDone:
raise AssertionError("load failed")
def load_single_Y_vox(vox_path):
with open(vox_path, 'rb') as ff:
vox = binvox_rw.read_as_3d_array(ff)
vox_grid = vox.data.astype(int)
#Data.plotFromVoxels(vox_grid)
return vox_grid
def main(args):
input_dir = args.input_dir
output_dir = args.output_dir
log_dir = args.log_dir
error_log = os.path.join(log_dir, "error.txt")
target_dim = 64
allf = os.listdir(input_dir)
start = time.time()
for i, f in enumerate(allf):
if (i % 1000 == 0):
print("Now checking the {0}-th file. Elapsed time {1:.2f} min.".
format(i, (time.time() - start) / 60))
with open(os.path.join(input_dir, f), "rb") as b:
model = binvox_rw.read_as_3d_array(b)
volume = np.asarray(model.data * 1, dtype=np.float32)
# print("the shape of volume")
# print(volume.shape)
if np.sum(volume) == 0:
print("Empty voxel! Logging...")
log_error(error_log, f)
continue
else:
newv = dilate(volume, target_dim, 0.5)
outpath = os.path.join(output_dir, f.split(".")[0] + ".npy")
np.save(outpath, newv)
return
def buildclanlogo(self):
print("I will build clan logo")
print(self.logoname)
x0=self.logopos[0]
y0=self.logopos[1]
z0=self.logopos[2]
print(x0,y0,z0)
with open(self.logoname, 'rb') as f:
model = binvox_rw.read_as_3d_array(f)
print(model.dims)
print(model.scale)
print(model.translate)
#print(model.data)
for y in range(model.dims[1]):
#print("layer y=",y)
layer_data=model.data[y]
stringlayer=""
for x in range(model.dims[0]):
stringlayer=stringlayer+"\n"
for z in range(model.dims[2]):
if model.data[x][z][y] == True:
stringlayer=stringlayer+'1'
mc.setBlock(x0+x,y0+y,z0+z,89)
else:
stringlayer=stringlayer+'0'
mc.setBlock(x0+x,y0+y,z0+z,block.AIR.id)
#print(stringlayer）
time.sleep(5)
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 convert_bin():
for s in wanted_classes:
directory = 'data/voxels/' + labels[s] + '/'
# find all binvoxes
models = glob('data/managable_objects/' + labels[s] + '/*.binvox')
if not os.path.exists(directory):
os.makedirs(directory)
for m in tqdm(models):
with open(m, 'rb') as f:
try:
model = binvox_rw.read_as_3d_array(f).data
except ValueError:
continue
# remove internals from models
# I think this makes it easier to learn
positions = np.where(model != 0)
new_mod = np.zeros(model.shape)
for i, j, k in zip(*positions):
# identifies if current voxel has an exposed face
if np.sum(model[i - 1:i + 2, j - 1:j + 2, k - 1:k + 2]) < 27:
new_mod[i, j, k] = 1
# save as np array
sio.savemat(directory + m.split('/')[-1][:-7],
{'model': new_mod.astype(np.uint8)})
def gen_binary(path):
dataset = np.zeros((1, 64 * 64 * 64 + 1))
for file in os.listdir(path):
suffix = file.split('.')[-1]
if suffix == 'binvox':
name = file.split('_')
label = np.array([int(name[0])])
print(label)
f = open(path + file, 'rb')
model = binvox_rw.read_as_3d_array(f)
model_num = model.data * 1
model_rot = rotations6(model_num)
for rot_dir in model_rot:
model_vector = np.reshape(rot_dir, (1, 64 * 64 * 64))[0]
data = np.append(label, model_vector)
dataset = np.vstack((dataset, data))
# print(np.shape(model_num))# 1/0
# model_vector = np.reshape(model_num,(1,64*64*64))[0]
#
# data = np.append(label,model_vector)
# dataset=np.vstack((dataset,data))
dataset = dataset[1:, :]
dataset = np.array(dataset, dtype=np.uint8) # convert float64 to uint8
np.random.shuffle(dataset) # shuffle input
return dataset
def __init__(self, mc, binvox_pathname):
self.mc = mc
with open(binvox_pathname, 'rb') as f:
self.model = binvox_rw.read_as_3d_array(f)
print(self.model.dims)
print(self.model.scale)
print(self.model.translate)
def rotate_voxels(rep, angle, fov):
a = binvox_rw.read_as_3d_array(
open("unprojected_voxels/outline_scale_47.binvox", "rb"))
val = a.data
val = tf.convert_to_tensor(np.expand_dims(np.expand_dims(val, 0), -1))
voxel.fov = fov
phi, theta = angle
proj_val = voxel.rotate_voxel(val, phi, theta)
num = np.where(proj_val > 0.5)[0]
if len(num) > 0:
print("found")
fovs_working[fov] = len(num)
proj_val = np.squeeze(proj_val)
proj_val = proj_val > 0.5
proj_imgZ = np.mean(proj_val, 0)
imsave(
'{}/valRotate_phi_{}_theta_{}_fov_{:04d}_Z.png'.format(
rep, phi, theta, fov), proj_imgZ)
save_voxel(
np.squeeze(proj_val),
"{}/valRotate_THETA_{}_PHI_{}_fov_{}_.binvox".format(
rep, theta[0], phi[0], fov))
def mylogo(px, py, pz):
pos.x = px
pos.y = py
pos.z = pz
with open('mylogo.binvox', 'rb') as f:
model = binvox_rw.read_as_3d_array(f)
print(model.dims)
print(model.scale)
print(model.translate)
#print(model.data)
for y in range(model.dims[1]):
print("layer y=", y)
layer_data = model.data[y]
stringlayer = ""
for x in range(model.dims[0]):
stringlayer = stringlayer + "\n"
for z in range(model.dims[2]):
if model.data[x][y][z] == True:
stringlayer = stringlayer + '1'
mc.setBlock(pos.x + x, pos.y + z, pos.z + y,
block.DIAMOND_BLOCK.id)
else:
stringlayer = stringlayer + '0'
mc.setBlock(pos.x + x, pos.y + z, pos.z + y, block.AIR.id)
print(stringlayer)
# mylogo(pos.x,pos.y,pos.z)
def diaoxiang(self):
x0 = self.data[0]
y0 = self.data[1]
z0 = self.data[2]
with open(self.load, 'rb') as f:
model = binvox_rw.read_as_3d_array(f)
#print(model.dims)
#print(model.scale)
#print(model.translate)
#print(model.data)
for y in range(model.dims[1]):
print("layer y=", y)
layer_data = model.data[y]
stringlayer = ""
for x in range(model.dims[0]):
stringlayer = stringlayer + "\n"
for z in range(model.dims[2]):
if model.data[x][y][z] == True:
stringlayer = stringlayer + '1'
mc.setBlock(x0 + x, y0 + 25 + y, z0 + z,
block.STONE.id)
else:
stringlayer = stringlayer + '0'
mc.setBlock(x0 + x, y0 + 25 + y, z0 + z, block.AIR.id)
print(stringlayer)
def __init__(self, filename='map2.binvox'):
with open(filename, 'rb') as f:
self.model = binvox_rw.read_as_3d_array(f)
# the point in airsim coordinate that generate voxel maps
self.center = np.array([0, 0, 0])
# resolution of each voxel
self.res = 0.5
# origin of voxel in airsim coordinate
voxel_origin_x = -int(self.model.dims[0] / 2) * self.res + self.center[0]
voxel_origin_y = -int(self.model.dims[1] / 2) * self.res + self.center[1]
voxel_origin_z = -int(self.model.dims[2] / 2) * self.res + self.center[2]
self.voxel_origin = (voxel_origin_x, voxel_origin_y, voxel_origin_z)
# coordinates of all voxels in airsim coordinate
self.coordinates = np.zeros(self.model.dims + [3])
self.tsdf = np.zeros(self.model.dims + [4])
for i in range(self.model.dims[0]):
for j in range(self.model.dims[1]):
for k in range(self.model.dims[2]):
x = (i - int(self.model.dims[0] / 2)) * self.res + self.center[0]
y = (j - int(self.model.dims[1] / 2)) * self.res + self.center[1]
z = (k - int(self.model.dims[2] / 2)) * self.res + self.center[2]
self.coordinates[i, j, k, :] = np.array([x, y, z])
p = multiprocessing.Pool(multiprocessing.cpu_count())
indices = []
for i in range(self.model.dims[0]):
for j in range(self.model.dims[1]):
for k in range(self.model.dims[2]):
indices.append((i, j, k))
p.map(self.calculate_tsdf, indices)
p.close()
p.join()
def create_voxel_file_list(folderPath, expression, printMessages=True):
path = os.getcwd() + "/" + folderPath
if (printMessages):
print("Loading voxel files from " + path)
fileList = []
for i in os.listdir(path):
if re.match(expression, i):
fileList.append(path + i)
assert(len(fileList) > 0)
if (printMessages):
print("Found " + str(len(fileList)) + " files")
modelList = []
for filePath in fileList:
with open(filePath, 'rb') as f:
geom = binvox_rw.read_as_3d_array(f)
# Don't append any empty files.
if (geom.data.max() > 0):
modelList.append(geom)
else:
print(filePath + " was empty!!")
if (printMessages):
print("Done loading voxel files.")
return modelList
def _load_binvox(self, path):
''' loads voxels saved in binvox format. see also _load_vox '''
if not os.path.exists(path):
raise Exception('path does not exist: ' + path)
with open(path, 'rb') as fin:
voxels = read_as_3d_array(fin)
return torch.from_numpy(voxels.data.astype('uint8'))
def create_occupancy_grid_from_obstacles(obstacles,
mins_xyz,
step_size_xyz,
dims_xyz,
use_binvox=False):
voxel_grid = np.zeros(shape=dims_xyz)
for obstacle in obstacles:
if use_binvox:
vox = binvox_rw.read_as_3d_array(
open(obstacle.mesh_filepath.replace('.ply', '.binvox'), 'r'))
vertices = binvox_to_points(vox)
else:
vertices = read_vertex_points_from_ply_filepath(
obstacle.mesh_filepath)
frame = tf_conversions.fromMsg(obstacle.pose_stamped.pose)
transform = tf_conversions.toMatrix(frame)
vertices_transformed = transform_points(vertices, transform)
if use_binvox:
voxel_grid += add_obstacles_to_reachability_space_full_binvox(
vertices_transformed, mins_xyz, step_size_xyz, dims_xyz)
else:
voxel_grid += add_obstacles_to_reachability_space_full(
vertices_transformed, mins_xyz, step_size_xyz, dims_xyz)
voxel_grid[np.where(voxel_grid > 0)] = 1
return voxel_grid
def generatevoxels(stlfile):
os.system("binvox -cb -pb -d 32 " + str(stlfile))
try:
binvoxpath = stlfile.split(".")[0] + ".binvox"
with open(binvoxpath, 'rb') as f:
model = binvox_rw.read_as_3d_array(f)
voxel_array = model.data
voxel_array = voxel_array * 1 # convert to zero and one
print voxel_array
print voxel_array.shape
#model[False] = 0
#model[True] = 1
#print voxel_array
#print type(voxel_array)
#print voxel_array.shape'
if render == True:
render_in_3d(voxel_array)
voxel_array = np.expand_dims(voxel_array, -1)
os.system("rm " + str(binvoxpath))
master_array.append(voxel_array)
print "Here's how many have been processed: " + str(len(master_array))
if len(master_array) % 500 == 0:
outfile = open("data/thingi10k_" + str(len(master_array)) + ".npy",
"w")
np.save(outfile, np.array(master_array))
except Exception as e:
print e
def saveVoxelsBinvox(pose, shape, fileshape):
if os.path.exists(fileshape):
print('Already exists ' + fileshape)
else:
dict_shape = {}
# Save gt points
m = model.copy()
m.betas[:] = shape
m.pose[:] = pose
dict_shape['points'] = m.r
dict_shape['J_transformed'] = m.J_transformed.r
# Write to .obj file
obj_path = fileshape[:-10] + '.obj'
smpl_utils.save_smpl_obj(obj_path, m, saveFaces=True)
# Voxelize using binvox
call([
os.path.join(BINVOX_PATH, "binvox"), "-e", "-fit", "-d", "128",
"%s" % obj_path
])
# Read the output of binvox
binvox_path = obj_path[:-4] + '.binvox'
with open(binvox_path, 'rb') as f:
binvoxModel = binvox_rw.read_as_3d_array(f)
# Remove intermediate files
call(["rm", obj_path])
call(["rm", binvox_path])
# Save binvox results to mat
dict_shape['voxels'] = binvoxModel.data
dict_shape['voxelsdims'] = binvoxModel.dims
dict_shape['voxelstranslate'] = binvoxModel.translate
dict_shape['voxelsscale'] = binvoxModel.scale
sio.savemat(fileshape, dict_shape, do_compression=True)
print('Saved ' + fileshape)
def getVoxelGrid(binvoxPath):
binvoxObj = binvox_rw.read_as_3d_array(open(binvoxPath, 'rb'))
scale = binvoxObj.scale
translate = binvoxObj.translate
voxelgrid = binvoxObj.data.astype('float32')
dims = binvoxObj.dims
return voxelgrid, dims, scale, translate
def fromMesh(self, fileName, resolution=128):
# Voxelize the model
call('optirun binvox', '-t {}'.format(resolution), fileName)
voxelFile = ''.join(os.path.splitext(fileName)[:-1] + ('.binvox',))
with open(voxelFile, 'rb') as f:
self.voxelModel = bv.read_as_3d_array(f)
self.fromBinaryArray(self.voxelModel.data)
def build_training_example(model_filepath, pose_filepath, single_view_pointcloud_filepath, patch_size):
pc = np.load(single_view_pointcloud_filepath) # Point cloud. Shape is (number of points, 4). R,G,B,Color
#remove 32 bit color channel
pc = pc[:, 0:3]
model_pose = np.load(pose_filepath) # 4x4 homogeneous transform matrix
with open(model_filepath, 'rb') as f:
model = binvox_rw.read_as_3d_array(f)
# import IPython
# IPython.embed()
points = model.data
scale = model.scale
translate = model.translate
dims = model.dims
non_zero_points = points.nonzero()
#get numpy array of nonzero points
num_points = len(non_zero_points[0])
non_zero_arr = np.zeros((4, num_points))
non_zero_arr[0] = non_zero_points[0]
non_zero_arr[1] = non_zero_points[1]
non_zero_arr[2] = non_zero_points[2]
non_zero_arr[3] = 1.0
translate_arr = np.array(translate).reshape(3, 1)
non_zero_arr[0:3, :] = non_zero_arr[0:3, :] + translate_arr
non_zero_arr[0:3, :] = non_zero_arr[0:3, :] / (scale * 4)
#this is needed, to recenter binvox model at origin for some reason
#the translate array does not seem to fully compensate.
non_zero_arr[2, :] -= .09
#this is an easier task, the y value is always the same. i.e the model standing
#up at the origin.
#pc2_out, non_zero_arr1 = self.map_pointclouds_to_world(pc, non_zero_arr, model_pose)
pc2_out, non_zero_arr1 = map_pointclouds_to_camera_frame(pc, non_zero_arr, model_pose)
min_x = pc2_out[0, :].min()
min_y = pc2_out[1, :].min()
min_z = pc2_out[2, :].min()
max_x = pc2_out[0, :].max()
max_y = pc2_out[1, :].max()
max_z = pc2_out[2, :].max()
center = (min_x + (max_x - min_x) / 2.0, min_y + (max_y - min_y) / 2.0, min_z + (max_z - min_z) / 2.0)
#now non_zero_arr and pc points are in the same frame of reference.
#since the images were captured with the model at the origin
#we can just compute an occupancy grid centered around the origin.
x = create_voxel_grid_around_point(pc2_out[0:3, :].T, center, voxel_resolution=.02, num_voxels_per_dim=patch_size)
y = create_voxel_grid_around_point(non_zero_arr1.T[:, 0:3], center, voxel_resolution=.02, num_voxels_per_dim=patch_size)
# viz.visualize_3d(x)
# viz.visualize_3d(y)
# viz.visualize_pointcloud(pc2_out[0:3, :].T)
# viz.visualize_pointclouds(pc2_out.T, non_zero_arr1.T[:, 0:3], False, True)
# import IPython
# IPython.embed()
return x, y
if c==state:
ctr += 1
# if ctr hits max, dump
if ctr==255:
fp.write(chr(state))
fp.write(chr(ctr))
ctr = 0
else:
# if switch state, dump
fp.write(chr(state))
fp.write(chr(ctr))
state = c
ctr = 1
# flush out remainders
if ctr > 0:
fp.write(chr(state))
fp.write(chr(ctr))
if __name__ == '__main__':
#~ import doctest
#~ doctest.testmod()
import numpy as np
import binvox_rw
with open('3D Models/761_hand-olivier_2.binvox', 'rb') as f:
m1 = binvox_rw.read_as_3d_array(f)
print m1.dims
print m1.scale
print m1.data
def build_training_example(binvox_file_path, model_pose_filepath, single_view_pointcloud_filepath, patch_size):
pc = np.load(single_view_pointcloud_filepath)
pc = pc[:, 0:3]
model_pose = np.load(model_pose_filepath)
with open(binvox_file_path, 'rb') as f:
model = binvox_rw.read_as_3d_array(f)
points = model.data
scale = model.scale
translate = model.translate
dims = model.dims
non_zero_points = points.nonzero()
#get numpy array of nonzero points
num_points = len(non_zero_points[0])
non_zero_arr = np.zeros((4, num_points))
non_zero_arr[0] = non_zero_points[0]
non_zero_arr[1] = non_zero_points[1]
non_zero_arr[2] = non_zero_points[2]
non_zero_arr[3] = 1.0
translate_arr = np.array(translate).reshape(3, 1)
#meters to centimeters
scale /= 100
#inches to meters
scale /= 2.54
non_zero_arr[0:3, :] = non_zero_arr[0:3, :] + translate_arr * 1.0/scale
non_zero_arr[0:3, :] = non_zero_arr[0:3, :] * scale
#this is an easier task, the y value is always the same. i.e the model standing
#up at the origin.
#pc2_out, non_zero_arr1 = map_pointclouds_to_world(pc, non_zero_arr, model_pose)
pc2_out, non_zero_arr1 = map_pointclouds_to_camera_frame(pc, non_zero_arr, model_pose)
min_x = pc2_out[0, :].min()
min_y = pc2_out[1, :].min()
min_z = pc2_out[2, :].min()
max_x = pc2_out[0, :].max()
max_y = pc2_out[1, :].max()
max_z = pc2_out[2, :].max()
center = (min_x + (max_x-min_x)/2.0, min_y + (max_y-min_y)/2.0, min_z + (max_z-min_z)/2.0)
# viz.visualize_pointclouds(pc2_out.T, non_zero_arr1.T[:, 0:3], False, True)
# import IPython
# IPython.embed()
#now non_zero_arr and pc points are in the same frame of reference.
#since the images were captured with the model at the origin
#we can just compute an occupancy grid centered around the origin.
x = create_voxel_grid_around_point(pc2_out[0:3, :].T, center, voxel_resolution=.02, num_voxels_per_dim=patch_size)
y = create_voxel_grid_around_point(non_zero_arr1.T[:, 0:3], center, voxel_resolution=.02, num_voxels_per_dim=patch_size)
return x, y