def process_chunk(vertex_chunk, label_chunk, has_labels, xyz, rgb, labels):
xyz_full = np.ascontiguousarray(
np.array(vertex_chunk.values[:, 0:3], dtype='float32'))
rgb_full = np.ascontiguousarray(
np.array(vertex_chunk.values[:, 4:7], dtype='uint8'))
if has_labels:
labels_full = label_chunk.values.squeeze()
else:
labels_full = None
if voxel_width > 0:
if has_labels > 0:
xyz_sub, rgb_sub, labels_sub, objets_sub = libply_c.prune(
xyz_full, voxel_width, rgb_full, labels_full,
np.zeros(1, dtype='uint8'), n_class, 0)
labels = np.vstack((labels, labels_sub))
del labels_full
else:
xyz_sub, rgb_sub, l, o = libply_c.prune(
xyz_full, voxel_width, rgb_full, np.zeros(1,
dtype='uint8'),
np.zeros(1, dtype='uint8'), 0, 0)
xyz = np.vstack((xyz, xyz_sub))
rgb = np.vstack((rgb, rgb_sub))
else:
xyz = xyz_full
rgb = xyz_full
labels = labels_full
return xyz, rgb, labels
def read_semantic3d_format2(data_file, n_class, file_label_path, voxel_width,
ver_batch):
"""read the format of semantic3d.
ver_batch : if ver_batch>0 then load the file ver_batch lines at a time.
useful for huge files (> 5millions lines)
voxel_width: if voxel_width>0, voxelize data with a regular grid
n_class : the number of class; if 0 won't search for labels (test set)
implements batch-loading for huge files
and pruning"""
xyz = np.zeros((0, 3), dtype='float32')
rgb = np.zeros((0, 3), dtype='uint8')
labels = np.zeros((0, n_class + 1), dtype='uint32')
#---the clouds can potentially be too big to parse directly---
#---they are cut in batches in the order they are stored---
i_rows = 0
while True:
try:
head = None
if ver_batch > 0:
print("Reading lines %d to %d" % (i_rows, i_rows + ver_batch))
vertices = np.genfromtxt(data_file,
delimiter=' ',
max_rows=ver_batch,
skip_header=i_rows)
#if i_rows > 0:
# head = i_rows-1
#vertices = pd.read_csv(data_file
# , sep=' ', nrows=ver_batch
# , header=head).values
else:
#vertices = np.genfromtxt(data_file, delimiter=' ')
vertices = np.pd.read_csv(data_file, sep=' ',
header=None).values
break
except (StopIteration, pd.errors.ParserError):
#end of file
break
if len(vertices) == 0:
break
xyz_full = np.ascontiguousarray(
np.array(vertices[:, 0:3], dtype='float32'))
rgb_full = np.ascontiguousarray(
np.array(vertices[:, 4:7], dtype='uint8'))
del vertices
if n_class > 0:
#labels_full = pd.read_csv(file_label_path, dtype="u1"
# , nrows=ver_batch, header=head).values.squeeze()
labels_full = np.genfromtxt(file_label_path,
dtype="u1",
delimiter=' ',
max_rows=ver_batch,
skip_header=i_rows)
if voxel_width > 0:
if n_class > 0:
xyz_sub, rgb_sub, labels_sub, objets_sub = libply_c.prune(
xyz_full, voxel_width, rgb_full, labels_full,
np.zeros(1, dtype='uint8'), n_class, 0)
labels = np.vstack((labels, labels_sub))
else:
xyz_sub, rgb_sub, l, o = libply_c.prune(
xyz_full, voxel_width, rgb_full, np.zeros(1,
dtype='uint8'),
np.zeros(1, dtype='uint8'), 0, 0)
del xyz_full, rgb_full
xyz = np.vstack((xyz, xyz_sub))
rgb = np.vstack((rgb, rgb_sub))
i_rows = i_rows + ver_batch
print("Reading done")
if n_class > 0:
return xyz, rgb, labels
else:
return xyz, rgb
def main():
parser = argparse.ArgumentParser(description='Large-scale Point Cloud Semantic Segmentation with Superpoint Graphs')
parser.add_argument('--ROOT_PATH', default='datasets/s3dis')
parser.add_argument('--dataset', default='s3dis')
#parameters
parser.add_argument('--compute_geof', default=1, type=int, help='compute hand-crafted features of the local geometry')
parser.add_argument('--k_nn_local', default=20, type=int, help='number of neighbors to describe the local geometry')
parser.add_argument('--k_nn_adj', default=5, type=int, help='number of neighbors for the adjacency graph')
parser.add_argument('--voxel_width', default=0.03, type=float, help='voxel size when subsampling (in m)')
parser.add_argument('--plane_model', default=1, type=int, help='uses a simple plane model to derive elevation')
parser.add_argument('--use_voronoi', default=0.0, type=float, help='uses the Voronoi graph in combination to knn to build the adjacency graph, useful for sparse aquisitions. If 0., do not use voronoi. If >0, then is the upper length limit for an edge to be kept. ')
parser.add_argument('--ver_batch', default=5000000, type=int, help='batch size for reading large files')
args = parser.parse_args()
#path to data
if args.ROOT_PATH[-1]=='/':
root = args.ROOT_PATH
else:
root = args.ROOT_PATH+'/'
if not os.path.exists(root + 'features_supervision'):
os.makedirs(root + 'features_supervision')
#list of subfolders to be processed
if args.dataset == 's3dis':
folders = ["Area_1/", "Area_2/", "Area_3/", "Area_4/", "Area_5/", "Area_6/"]
n_labels = 13
elif args.dataset == 'sema3d':
folders = ["train/", "test_reduced/", "test_full/"]
n_labels = 8
elif args.dataset == 'vkitti':
folders = ["01/", "02/","03/", "04/","05/", "06/"]
n_labels = 13 #number of classes
elif args.dataset == 'custom_dataset':
folders = ["train/", "test/"]
n_labels = 10 #number of classes
else:
raise ValueError('%s is an unknown data set' % args.dataset)
pruning = args.voxel_width > 0
#------------------------------------------------------------------------------
for folder in folders:
print("=================\n "+folder+"\n=================")
data_folder = root + "data/" + folder
str_folder = root + "features_supervision/" + folder
if not os.path.isdir(data_folder):
raise ValueError("%s does not exist" % data_folder)
# os.mkdir(data_folder)
if not os.path.isdir(str_folder):
os.mkdir(str_folder)
if args.dataset == 's3dis':
files = [os.path.join(data_folder, o) for o in os.listdir(data_folder)
if os.path.isdir(os.path.join(data_folder,o))]
elif args.dataset == 'sema3d':
files = glob.glob(data_folder + "*.txt")
elif args.dataset == 'vkitti':
files = glob.glob(data_folder + "*.npy")
if (len(files) == 0):
continue
#raise ValueError('%s is empty' % data_folder)
n_files = len(files)
i_file = 0
for file in files:
file_name = os.path.splitext(os.path.basename(file))[0]
if args.dataset=='s3dis':
data_file = data_folder + file_name + '/' + file_name + ".txt"
str_file = str_folder + file_name + '.h5'
elif args.dataset=='sema3d':
file_name_short = '_'.join(file_name.split('_')[:2])
data_file = data_folder + file_name + ".txt"
label_file = data_folder + file_name + ".labels"
str_file = str_folder + file_name_short + '.h5'
elif args.dataset=='vkitti':
data_file = data_folder + file_name + ".npy"
str_file = str_folder + file_name + '.h5'
i_file = i_file + 1
print(str(i_file) + " / " + str(n_files) + "---> "+file_name)
if os.path.isfile(str_file):
print(" graph structure already computed - delete for update...")
else:
#--- build the geometric feature file h5 file ---
print(" computing graph structure...")
#--- read the data files and compute the labels---
if args.dataset == 's3dis':
xyz, rgb, labels, objects = read_s3dis_format(data_file)
if pruning:
n_objects = int(objects.max()+1)
xyz, rgb, labels, objects = libply_c.prune(xyz, args.voxel_width, rgb, labels, objects, n_labels, n_objects)
#hard_labels = labels.argmax(axis=1)
objects = objects[:,1:].argmax(axis=1)+1
else:
#hard_labels = labels
objects = objects
elif args.dataset=='sema3d':
has_labels = (os.path.isfile(label_file))
if (has_labels):
xyz, rgb, labels = read_semantic3d_format(data_file, n_labels, label_file, args.voxel_width, args.ver_batch)
else:
xyz, rgb = read_semantic3d_format(data_file, 0, '', args.voxel_width, args.ver_batch)
labels = np.array([0])
objects = np.array([0])
is_transition = np.array(False)
elif args.dataset == 'vkitti':
xyz, rgb, labels = read_vkitti_format(data_file)
if pruning:
xyz, rgb, labels, o = libply_c.prune(xyz.astype('f4'), args.voxel_width, rgb.astype('uint8'), labels.astype('uint8'), np.zeros(1, dtype='uint8'), n_labels, 0)
#---compute nn graph-------
n_ver = xyz.shape[0]
print("computing NN structure")
graph_nn, local_neighbors = compute_graph_nn_2(xyz, args.k_nn_adj, args.k_nn_local, voronoi = args.use_voronoi)
if args.dataset=='s3dis':
is_transition = objects[graph_nn["source"]]!=objects[graph_nn["target"]]
elif args.dataset=='sema3d' and has_labels:
#sema has no object, we make them ourselves with label inpainting
hard_labels = np.argmax(labels[:,1:], 1)+1
no_labels = (labels[:,1:].sum(1)==0).nonzero()
hard_labels[no_labels] = 0
is_transition = hard_labels[graph_nn["source"]]!=hard_labels[graph_nn["target"]] * (hard_labels[graph_nn["source"]]!=0) \
* (hard_labels[graph_nn["target"]]!=0)
edg_source = graph_nn["source"][(is_transition==0).nonzero()].astype('uint32')
edg_target = graph_nn["target"][(is_transition==0).nonzero()].astype('uint32')
edge_weight = np.ones_like(edg_source).astype('f4')
node_weight = np.ones((n_ver,),dtype='f4')
node_weight[no_labels] = 0
print("Inpainting labels")
dump, objects = libcp.cutpursuit2(np.array(hard_labels).reshape((n_ver,1)).astype('f4'), edg_source, edg_target, edge_weight, node_weight, 0.01)
is_transition = objects[graph_nn["source"]]!=objects[graph_nn["target"]]
elif args.dataset=='vkitti':
#we define the objects as the constant connected components of the labels
hard_labels = np.argmax(labels, 1)
is_transition = hard_labels[graph_nn["source"]]!=hard_labels[graph_nn["target"]]
dump, objects = libply_c.connected_comp(n_ver \
, graph_nn["source"].astype('uint32'), graph_nn["target"].astype('uint32') \
, (is_transition==0).astype('uint8'), 0)
if (args.compute_geof):
geof = libply_c.compute_geof(xyz, local_neighbors, args.k_nn_local).astype('float32')
geof[:,3] = 2. * geof[:,3]
else:
geof = 0
if args.plane_model: #use a simple palne model to the compute elevation
low_points = ((xyz[:,2]-xyz[:,2].min() < 0.5)).nonzero()[0]
reg = RANSACRegressor(random_state=0).fit(xyz[low_points,:2], xyz[low_points,2])
elevation = xyz[:,2]-reg.predict(xyz[:,:2])
else:
elevation = xyz[:,2] - xyz[:,2].min()
#compute the xy normalized position
ma, mi = np.max(xyz[:,:2],axis=0,keepdims=True), np.min(xyz[:,:2],axis=0,keepdims=True)
xyn = (xyz[:,:2] - mi) / (ma - mi + 1e-8) #global position
write_structure(str_file, xyz, rgb, graph_nn, local_neighbors.reshape([n_ver, args.k_nn_local]), \
is_transition, labels, objects, geof, elevation, xyn)
