def create_attribute_vectors (scene_splits, scaler, kmeanspp, classifiers):
scenes = []
attribute_vectors = []
for subdir, images in scene_splits.iteritems():
# skip root directory
if len(images) == 0:
continue
# extract scene name
scene_name = subdir[subdir.rfind('/') + 1:]
for image in images:
descriptors = fe.extract(image, settings.scale_size)
kmeanspp_predicted = kmeanspp.predict(
scaler.transform(descriptors)
)
histogram = bow.generate_histogram(kmeanspp_predicted)
# predict attributes of all present classifiers
# (87 for asymmetric, 102 for symmetric)
attribute_vector = []
for attribute, classifier_list in classifiers.iteritems():
attribute_vector.extend(classifier_list[5].predict(histogram))
# store scene and attribute vector
scenes.append(scene_name)
attribute_vectors.append(attribute_vector)
return [attribute_vectors, scenes]
def create (images, computed_feature_vectors, scaler):
feature_vectors = []
step_mod = len(images) * 0.05 # print progress in steps of 5 percent
done = 0
print('[K-MEANS++] SIFT feature extraction:\n[K-MEANS++] 0%', end = '')
sys.stdout.flush()
for image in images:
if image in computed_feature_vectors:
descriptors = computed_feature_vectors[image]
else:
try:
descriptors = fe.extract(settings.images_dir + image, settings.scale_size)
except Exception as e:
print('Error while computing SIFT for image {}.'.format(image))
continue
computed_feature_vectors[image] = descriptors
fe.unroll_and_append_descriptors(descriptors, feature_vectors)
done += 1
if done % step_mod == 0:
print(' {}%'.format((100 * done) / len(images)), end = '')
sys.stdout.flush()
if done < len(images):
print(' 100%')
print('[K-MEANS++] Number of \'SIFTed\' images: {}'.format(done))
print('[K-MEANS++] Total number of sift descriptors: {}'
.format(len(feature_vectors))
)
sys.stdout.flush()
# do k-means clustering (512 clusters, k-means++, 4 cores, show information)
kmeans = cluster.KMeans(n_clusters = settings.class_count, init = 'k-means++',
n_jobs = 4
)
# 10 SIFT descriptors per images (approximately)
kmeans.fit(scaler.transform(
random.sample(feature_vectors, int(10 * done))
))
# dump k-means clustering for further usage
with open(settings.filepaths['k-means++'], 'wb') as f:
pe.dump(kmeans, f)
return kmeans
def create (images, computed_feature_vectors):
feature_vectors = []
step_mod = len(images) * 0.05 # print progress in steps of 5 percent
done = 0
print('[SCALER] SIFT feature extraction:\n[SCALER] 0%', end = '')
sys.stdout.flush()
for image in images:
if image in computed_feature_vectors:
descriptors = computed_feature_vectors[image]
else:
try:
descriptors = fe.extract(settings.images_dir + image, settings.scale_size)
except Exception as e:
print('Error while computing SIFT for image {}.'.format(image))
continue
computed_feature_vectors[image] = descriptors
fe.unroll_and_append_descriptors(descriptors, feature_vectors)
done += 1
if done % step_mod == 0:
print(' {}%'.format((100 * done) / len(images)), end = '')
sys.stdout.flush()
if done < len(images):
print(' 100%')
print('[SCALER] Number of \'SIFTed\' images: {}'.format(done))
print('[SCALER] Total number of SIFT descriptors: {}'
.format(len(feature_vectors))
)
sys.stdout.flush()
# calibrate min-max-scaler
min_max_scaler = preprocessing.MinMaxScaler()
# 10 SIFT descriptors per image (approximately)
min_max_scaler = min_max_scaler.fit(
random.sample(feature_vectors, int(10 * done))
)
# dump min-max-scaler for further usage
with open(settings.filepaths['scaler'], 'wb') as f:
pe.dump(min_max_scaler, f)
return min_max_scaler
def select_random_images (size, candidates, computed_feature_vectors):
selected_candidates = []
i = 0
while i < size:
image = settings.images_dir + choice(candidates)
# avoid duplicates
if image in selected_candidates:
continue;
if image in computed_feature_vectors:
descriptors = computed_feature_vectors[image]
else:
try:
descriptors = fe.extract(image, settings.scale_size)
computed_feature_vectors[image] = descriptors
except Exception as e:
print('Error while computing SIFT for image {}'.format(image))
continue
selected_candidates.append(image)
i += 1
sys.stdout.flush()
return selected_candidates
