def vlad_vector_batch(input_path, output_path, vocabulary):
vladL = []
file_names = list_all_files(input_path)
for f in file_names:
vec = vlad_vector(vector_path=input_path + f, vocabulary=vocabulary)
vladL.append(vec)
output_path = output_path + '_vlad.txt'
savetxt(output_path, asarray(vladL), delimiter=",", fmt='%.4f')
def vlad_vector_batch(input_path, output_path, vocabulary):
vladL = []
file_names = list_all_files(input_path)
for f in file_names:
vec = vlad_vector(vector_path=input_path + f, vocabulary=vocabulary)
vladL.append(vec)
output_path = output_path + '_vlad.txt'
savetxt(output_path, asarray(vladL), delimiter=",", fmt='%.4f')
def sift_image_batch(input_path, output_path, params=None):
# Process all images in the input_path by SIFT, e.g. windows processing
# Return those images whose SIFT vector is invalid, i.e. empty
image_names = list_all_files(input_path, onlyImage=True)
res_path_L = []
for img in image_names:
res_path = sift_image(input_path=input_path, image_name=img, output_path=output_path, params=params)
res_path_L.append(res_path)
return res_path_L
def batch_all_images(input_image_path, annotation_path, output_parent_path,
unit_ratio_list, overlap_ratio, target, target_pos_path, vladVector, target_count=20, pca=True,
k=30, max_iter=30, preVLAD=False, voca_path=None, dataset_mode=False, overlap_threshold=0.5):
global_sift_path = '%sglobal_sift.txt' % (output_parent_path)
image_name_list = get_target_pos_names(input_path=target_pos_path, target=target, target_count=target_count)
sift_path_L = []
if preVLAD:
delete_file(file_path=global_sift_path)
for image_name in image_name_list:
metadata_path = '%s%s.xml' % (annotation_path, image_name)
win_sift_path = batch_one_image_pre_VLAD(input_image_path, image_name, metadata_path, output_parent_path,
unit_ratio_list, overlap_ratio)
sift_path_L.append(win_sift_path)
append_file(dest_file=global_sift_path, input_path=win_sift_path)
print "----------pre-VLAD Done"
else:
all_dir = list_all_files(input_path=output_parent_path + 'windows/', onlyDir=True)
for d in all_dir:
sift_path_L.append("%s/windows/%s/temp_sift/" % (output_parent_path, d))
print "----------pre-VLAD is enabled"
if voca_path is None or not isfile(voca_path):
print "~~~~~~~Learning vocabulary by the sift vectors of all windows of all images"
vector_matrix = None
if pca:
vector_matrix = pca_dataset(input_path=global_sift_path)
vocabulary = learn_vocabulary(input_path=global_sift_path,
k=k, max_iter=max_iter, single_file=True, vector_matrix=vector_matrix)
save_matrix(v=vocabulary, output_path=voca_path)
print "~~~~~~~Learning vocabulary done"
elif vladVector:
print "~~~~~~~Loading existing vocabulary"
vocabulary = load_matrix(input_path=voca_path)
if vladVector:
for i in xrange(len(image_name_list)):
image_name = image_name_list[i]
output_path = "%swindows/%s/%s" % (output_parent_path, image_name, image_name)
print "\t======Creating VLAD vectors"
vlad_vector_batch(input_path=sift_path_L[i],
output_path=output_path, vocabulary=vocabulary)
print "\t======VLAD Done for", image_name
else:
print "##########No VLAD vector generated...."
if dataset_mode:
print "^^^^^^^^^^Generate data set for global windows and VLAD"
global_X_path = output_parent_path + "global_X.txt"
global_Y_path = output_parent_path + "global_Y.txt"
delete_file(global_X_path)
delete_file(global_Y_path)
for img_name in image_name_list:
img_window_path = "%s/windows/%s/%s_windows.txt" % (output_parent_path, img_name , img_name)
img_vlad_path = "%s/windows/%s/%s_vlad.txt" % (output_parent_path, img_name , img_name)
metadata_path = '%s%s.xml' % (annotation_path, img_name)
batch_one_image_dataset(global_X_path, global_Y_path, img_window_path,
img_vlad_path, metadata_path, target, overlap_threshold=overlap_threshold)
print "\tData set done for", img_name
print "....................All done"
def sift_image_batch(input_path, output_path, params=None):
# Process all images in the input_path by SIFT, e.g. windows processing
# Return those images whose SIFT vector is invalid, i.e. empty
image_names = list_all_files(input_path, onlyImage=True)
res_path_L = []
for img in image_names:
res_path = sift_image(input_path=input_path,
image_name=img,
output_path=output_path,
params=params)
res_path_L.append(res_path)
return res_path_L
def learn_vocabulary(k, max_iter, single_file, vector_matrix=None, input_path=None):
if vector_matrix is None:
if single_file:
vector_matrix = read_feature_vector(input_path)
else:
file_names = list_all_files(input_path)
vector_matrix = read_feature_vector(input_path + file_names[0])
for f in xrange(1, len(file_names)):
vector = read_feature_vector(input_path + file_names[f])
if vector is not None:
vector_matrix = append(vector_matrix, vector, axis=0)
return mykmeans_plus(data=vector_matrix,k=k, max_iter=max_iter)[0]
def learn_vocabulary(k,
max_iter,
single_file,
vector_matrix=None,
input_path=None):
if vector_matrix is None:
if single_file:
vector_matrix = read_feature_vector(input_path)
else:
file_names = list_all_files(input_path)
vector_matrix = read_feature_vector(input_path + file_names[0])
for f in xrange(1, len(file_names)):
vector = read_feature_vector(input_path + file_names[f])
if vector is not None:
vector_matrix = append(vector_matrix, vector, axis=0)
return mykmeans_plus(data=vector_matrix, k=k, max_iter=max_iter)[0]
def batch_all_images(input_image_path,
annotation_path,
output_parent_path,
unit_ratio_list,
overlap_ratio,
target,
target_pos_path,
vladVector,
target_count=20,
pca=True,
k=30,
max_iter=30,
preVLAD=False,
voca_path=None,
dataset_mode=False,
overlap_threshold=0.5):
global_sift_path = '%sglobal_sift.txt' % (output_parent_path)
image_name_list = get_target_pos_names(input_path=target_pos_path,
target=target,
target_count=target_count)
sift_path_L = []
if preVLAD:
delete_file(file_path=global_sift_path)
for image_name in image_name_list:
metadata_path = '%s%s.xml' % (annotation_path, image_name)
win_sift_path = batch_one_image_pre_VLAD(input_image_path,
image_name, metadata_path,
output_parent_path,
unit_ratio_list,
overlap_ratio)
sift_path_L.append(win_sift_path)
append_file(dest_file=global_sift_path, input_path=win_sift_path)
print "----------pre-VLAD Done"
else:
all_dir = list_all_files(input_path=output_parent_path + 'windows/',
onlyDir=True)
for d in all_dir:
sift_path_L.append("%s/windows/%s/temp_sift/" %
(output_parent_path, d))
print "----------pre-VLAD is enabled"
if voca_path is None or not isfile(voca_path):
print "~~~~~~~Learning vocabulary by the sift vectors of all windows of all images"
vector_matrix = None
if pca:
vector_matrix = pca_dataset(input_path=global_sift_path)
vocabulary = learn_vocabulary(input_path=global_sift_path,
k=k,
max_iter=max_iter,
single_file=True,
vector_matrix=vector_matrix)
save_matrix(v=vocabulary, output_path=voca_path)
print "~~~~~~~Learning vocabulary done"
elif vladVector:
print "~~~~~~~Loading existing vocabulary"
vocabulary = load_matrix(input_path=voca_path)
if vladVector:
for i in xrange(len(image_name_list)):
image_name = image_name_list[i]
output_path = "%swindows/%s/%s" % (output_parent_path, image_name,
image_name)
print "\t======Creating VLAD vectors"
vlad_vector_batch(input_path=sift_path_L[i],
output_path=output_path,
vocabulary=vocabulary)
print "\t======VLAD Done for", image_name
else:
print "##########No VLAD vector generated...."
if dataset_mode:
print "^^^^^^^^^^Generate data set for global windows and VLAD"
global_X_path = output_parent_path + "global_X.txt"
global_Y_path = output_parent_path + "global_Y.txt"
delete_file(global_X_path)
delete_file(global_Y_path)
for img_name in image_name_list:
img_window_path = "%s/windows/%s/%s_windows.txt" % (
output_parent_path, img_name, img_name)
img_vlad_path = "%s/windows/%s/%s_vlad.txt" % (output_parent_path,
img_name, img_name)
metadata_path = '%s%s.xml' % (annotation_path, img_name)
batch_one_image_dataset(global_X_path,
global_Y_path,
img_window_path,
img_vlad_path,
metadata_path,
target,
overlap_threshold=overlap_threshold)
print "\tData set done for", img_name
print "....................All done"
