def compute_class_weights(labels_dir, label_values):
'''
Arguments:
labels_dir(list): Directory where the image segmentation labels are
num_classes(int): the number of classes of pixels in all images
Returns:
class_weights(list): a list of class weights where each index represents
each class label and the element is the class weight for that label.
'''
# TODO: need Test and Debug
image_files = filelist(labels_dir, ifPath=True, extension='.png')
num_classes = len(label_values)
class_pixels = np.zeros(num_classes)
total_pixels = 0.0
for n in range(len(image_files)):
image = cv2.imread(image_files[n], -1)
for index, colour in enumerate(label_values):
class_map = np.all(np.equal(image, colour), axis=-1)
class_map = class_map.astype(np.float32)
class_pixels[index] += np.sum(class_map)
print("\rProcessing image: " + str(n) + " / " + str(len(image_files)),
end="")
sys.stdout.flush()
total_pixels = float(np.sum(class_pixels))
index_to_delete = np.argwhere(class_pixels == 0.0)
class_pixels = np.delete(class_pixels, index_to_delete)
class_weights = total_pixels / class_pixels
class_weights = class_weights / np.sum(class_weights)
return class_weights
def data_statistics():
''' Count (pixel-level) mean and variance of dataset. '''
dir_list = [
'/home/tao/Data/RBDD/train',
] # list of data folder dir.
mean = np.array([0, 0, 0], np.float32) # BGR
std = np.array([0, 0, 0], np.float32) # BGR
total_image = 0
# Statisify mean and std
for folder in dir_list:
image_paths = filelist(folder, ifPath=True)
for path in tqdm(image_paths):
img = cv2.imread(path, cv2.IMREAD_COLOR) # BGR
img = np.float32(img) / 255.0
# images.append(np.expand_dims(img))
mean += np.mean(np.mean(img, 0), 0)
total_image += 1
mean /= total_image # get pixel-level BGR mean
# Statisify std
# for folder in dir_list:
# image_paths = filelist(folder, ifPath=True)
# for path in tqdm(image_paths):
# img = cv2.imread(path, cv2.IMREAD_COLOR) # BGR
# img = np.float32(img) / 255.0
# X = img-mean
# X = X*X
# std += np.sum(np.sum(X, axis=0), axis=0)
# std /= total_image
print('mean=', mean)
# print('std=', std)
print('Careful: BGR!')
def convert_image_type(image_dir, new_type, old_type=None):
'''Covert image's type(may be specify the old type).
Args:
new_type: The target type of image conversion(such as: 'png', 'jpg').
'''
image_names = filelist(image_dir, True, old_type)
for name in tqdm(image_names):
img = cv2.imread(name, 1) # 默认BGR模式读，适应Tiff的标签图
os.remove(name)
name = os.path.splitext(name)[0] + '.' + new_type
cv2.imwrite(name, img, [1, 100])
def MorphologicPostTreate(input_dir, output_dir):
'''
形态学后处理
'''
image_names = filelist(input_dir)
kernel = np.ones([7, 7], dtype=np.uint8)
kernel[0, 0], kernel[-1, -1], kernel[0, -1], kernel[-1, 0] = 0, 0, 0, 0
# print(kernel)
for i in tqdm(range(len(image_names))):
img = cv2.imread(input_dir + "/" + image_names[i], -1)
img = bwmorph(img, kernel_1=kernel) # 输出[HW]label
img = np.repeat(np.expand_dims(img, axis=2), 3, axis=2)
cv2.imwrite(output_dir + "/" + image_names[i], img)
def label_pretreat(label_dir, label_values):
'''
Pre-treat the orignal RGB label, to fix the apparent bug.
(Actually it eliminate the wrong colors that are not in class_dict)
By the way, unify the dtype of label imgs to png.
'''
l_names = filelist(label_dir, ifPath=True)
for name in tqdm(l_names):
label = cv2.imread(name, 1) # read in RGB
os.remove(name)
name = os.path.splitext(name)[0] + '.png'
new_label = np.zeros(label.shape, label.dtype)
# Fix the color(stand for a class) by class-info
for color in label_values:
equality = np.equal(label, color)
ind_mat = np.all(equality, axis=-1)
new_label[ind_mat] = color # this color list can be customized
cv2.imwrite(name, new_label) # new_label-BGR(according to class_dict)
def spilt_dataset(rate=0.025, max_num=None):
''' Spilt images(maybe with labels)'''
from shutil import move
# from sklearn.model_selection import train_test_split
# root = '/home/tao/Data/cvpr_road/'
in_floders = [
'/home/tao/Data/cvpr_road/512/train',
'/home/tao/Data/cvpr_road/512/train_labels'
]
out_floders = [
'/home/tao/Data/cvpr_road/512/val',
'/home/tao/Data/cvpr_road/512/val_labels'
]
name_list = []
for d in in_floders:
names = filelist(d)
name_list.append(names)
file_num = len(name_list[0]) # 文件总数
max_num = file_num if max_num is None else max_num
# 按比例抽取
index = np.random.permutation(file_num)
index = index[:int(file_num * rate)]
for d, names in zip(range(len(in_floders)), name_list):
print(in_floders[d])
cnt = 0
for i in index:
# 针对文件的操作
# os.remove(d+'/'+names[i])
move(in_floders[d] + '/' + names[i],
out_floders[d] + '/' + names[i])
cnt += 1
if cnt >= max_num:
break # if set max_num, prioritize the max_num
print('Finished!')