def extract_training(self, path):
"""
This method extract features/labels of a subset of the training points. It ensures a balanced choice between
classes.
:param path: path where the ply files are located.
:return: features and labels
"""
# Get all the ply files in data folder
ply_files = [f for f in listdir(path) if f.endswith('.ply')]
# Initiate arrays
training_features = np.empty((0, 12))
training_labels = np.empty((0,))
# Loop over each training cloud
for i, file in enumerate(ply_files):
# Load Training cloud
cloud_ply = read_ply(join(path, file))
points = np.vstack((cloud_ply['x'], cloud_ply['y'], cloud_ply['z'])).T
labels = cloud_ply['class']
# Initiate training indices array
training_inds = np.empty(0, dtype=np.int32)
# Loop over each class to choose training points
for label, name in self.label_names.items():
# Do not include class 0 in training
if label == 0:
continue
# Collect all indices of the current class
label_inds = np.where(labels == label)[0]
# If you have not enough indices, just take all of them
if len(label_inds) <= self.num_per_class:
training_inds = np.hstack((training_inds, label_inds))
# If you have more than enough indices, choose randomly
else:
random_choice = np.random.choice(len(label_inds), self.num_per_class, replace=False)
training_inds = np.hstack((training_inds, label_inds[random_choice]))
# Gather chosen points
training_points = points[training_inds, :]
# Compute features for the points of the chosen indices and place them in a [N, 4] matrix
vert, line, plan, sphe = compute_features(training_points, points, self.radius)
cosn, sinn, cost, sint, omnivariance, anisotropy, eigenentropy, change_curvature = compute_add_features(training_points, points, self.radius)
features = np.vstack((vert.ravel(), line.ravel(), plan.ravel(), sphe.ravel(), cosn.ravel(), sinn.ravel(), cost.ravel(), sint.ravel(), omnivariance.ravel(), anisotropy.ravel(), eigenentropy.ravel(), change_curvature.ravel())).T
# Concatenate features / labels of all clouds
training_features = np.vstack((training_features, features))
training_labels = np.hstack((training_labels, labels[training_inds]))
return training_features, training_labels
def basic_point_choice(num_per_class, radius, label_names):
# Initiate arrays
training_features = np.empty((0, 4))
training_labels = np.empty((0, ))
# Loop over each training cloud
for i in range(1, 4):
# Load Training cloud
cloud_path = '../data/Lille_street_{:d}.ply'.format(i)
cloud_ply = read_ply(cloud_path)
points = np.vstack((cloud_ply['x'], cloud_ply['y'], cloud_ply['z'])).T
labels = cloud_ply['labels']
# Initiate training indices array
training_inds = np.empty(0, dtype=np.int32)
# Loop over each class to choose training points
for label, name in label_names.items():
# Do not include class 0 in training
if label == 0:
continue
# Collect all indices of the current class
label_inds = np.where(labels == label)[0]
# If you have not enough indices, just take all of them
if len(label_inds) <= num_per_class:
training_inds = np.hstack((training_inds, label_inds))
# If you have more than enough indices, choose randomly
else:
random_choice = np.random.choice(len(label_inds),
num_per_class,
replace=False)
training_inds = np.hstack(
(training_inds, label_inds[random_choice]))
# Compute features for the points of the chosen indices
verticality, linearity, planarity, sphericity = compute_features(
points[training_inds, :], points, radius)
features = np.vstack((verticality.ravel(), linearity.ravel(),
planarity.ravel(), sphericity.ravel())).T
# Concatenate features / labels of all clouds
training_features = np.vstack((training_features, features))
training_labels = np.hstack((training_labels, labels[training_inds]))
return training_features, training_labels
def extract_test(self, path):
"""
This method extract features of all the test points.
:param path: path where the ply files are located.
:return: features
"""
# Get all the ply files in data folder
ply_files = [f for f in listdir(path) if f.endswith('.ply')]
# Initiate arrays
test_features = np.empty((0, 4))
# Loop over each training cloud
for i, file in enumerate(ply_files):
print("File {}: {}".format(i, file))
# Load Training cloud
cloud_ply = read_ply(join(path, file))
points = np.vstack(
(cloud_ply['x'], cloud_ply['y'], cloud_ply['z'])).T
# Compute features only one time and save them for further use
#
# WARNING : This will save you some time but do not forget to delete your features file if you change
# your features. Otherwise you will not compute them and use the previous ones
#
# Name the feature file after the ply file.
feature_file = file[:-4] + '_features.npy'
feature_file = join(path, feature_file)
# If the file exists load the previously computed features
if exists(join(path, feature_file)):
features = np.load(feature_file)
# If the file does not exist, compute the features (very long) and save them for future use
else:
vert, line, plan, sphe = compute_features(
points, points, self.radius)
features = np.vstack(
(vert.ravel(), line.ravel(), plan.ravel(), sphe.ravel())).T
np.save(feature_file, features)
# Concatenate features of several clouds
# (For this minichallenge this is useless as the test set contains only one cloud)
test_features = np.vstack((test_features, features))
return test_features
def basic_training_and_test():
# Parameters
# **********
#
radius = 0.5
num_per_class = 500
label_names = {
0: 'Unclassified',
1: 'Ground',
2: 'Building',
3: 'Vegetation',
4: 'Barriers',
5: 'Cars',
6: 'Signage'
}
# Collect training features / labels
# **********************************
#
print('Collect Training Features')
t0 = time.time()
training_features, training_labels = basic_point_choice(
num_per_class, radius, label_names)
t1 = time.time()
print('Done in %.3fs\n' % (t1 - t0))
# Train a random forest classifier
# ********************************
#
print('Training Random Forest')
t0 = time.time()
clf = RandomForestClassifier()
clf.fit(training_features, training_labels)
t1 = time.time()
print('Done in %.3fs\n' % (t1 - t0))
# Test
# ****
#
print('Compute testing features')
t0 = time.time()
# Load cloud as a [N x 3] matrix
cloud_path = '../data/Lille_street_test.ply'
cloud_ply = read_ply(cloud_path)
points = np.vstack((cloud_ply['x'], cloud_ply['y'], cloud_ply['z'])).T
# Compute features only one time and save them for further use
#
# WARNING : This will save you some time but do not forget to delete your features file if you change
# your features. Otherwise you will not compute them and use the previous ones
#
feature_file = '../data/Lille_street_test_features.npy'
if exists(feature_file):
features = np.load(feature_file)
else:
verticality, linearity, planarity, sphericity = compute_features(
points, points, radius)
features = np.vstack((verticality.ravel(), linearity.ravel(),
planarity.ravel(), sphericity.ravel())).T
np.save(feature_file, features)
t1 = time.time()
print('Done in %.3fs\n' % (t1 - t0))
print('Test')
t0 = time.time()
predictions = clf.predict(features)
t1 = time.time()
print('Done in %.3fs\n' % (t1 - t0))
print('Save predictions')
t0 = time.time()
np.savetxt('../data/Lille_street_test_preds.txt', predictions, fmt='%d')
t1 = time.time()
print('Done in %.3fs\n' % (t1 - t0))
print('Save predictions ply')
t0 = time.time()
write_ply('../Lille_street_test_preds.ply', [points, predictions],
['x', 'y', 'z', 'labels'])
t1 = time.time()
print('Done in %.3fs\n' % (t1 - t0))