def read_and_prepocess_single_img(filename_queue, shortside_len, is_training):
img_name, img, gtboxes_and_label, num_objects = read_single_example_and_decode(
filename_queue)
img = tf.cast(img, tf.float32)
if is_training:
img, gtboxes_and_label = image_preprocess.short_side_resize(
img_tensor=img,
gtboxes_and_label=gtboxes_and_label,
target_shortside_len=shortside_len,
length_limitation=cfgs.IMG_MAX_LENGTH)
img, gtboxes_and_label = image_preprocess.random_flip_left_right(
img_tensor=img, gtboxes_and_label=gtboxes_and_label)
else:
img, gtboxes_and_label = image_preprocess.short_side_resize(
img_tensor=img,
gtboxes_and_label=gtboxes_and_label,
target_shortside_len=shortside_len,
length_limitation=cfgs.IMG_MAX_LENGTH)
if cfgs.NET_NAME in ['resnet152_v1d', 'resnet101_v1d', 'resnet50_v1d']:
img = img / 255 - tf.constant([[cfgs.PIXEL_MEAN_]])
else:
img = img - tf.constant([[cfgs.PIXEL_MEAN]]) # sub pixel mean at last
return img_name, img, gtboxes_and_label, num_objects
def read_and_prepocess_single_img(filename_queue, shortside_len, is_training):
img_name, img, gtboxes_and_label, num_objects = read_single_example_and_decode(
filename_queue)
img = tf.cast(img, tf.float32)
img = img - tf.constant(cfgs.PIXEL_MEAN)
if is_training:
img, gtboxes_and_label = image_preprocess.short_side_resize(
img_tensor=img,
gtboxes_and_label=gtboxes_and_label,
target_shortside_len=shortside_len)
img, gtboxes_and_label = image_preprocess.random_flip_left_right(
img_tensor=img, gtboxes_and_label=gtboxes_and_label)
#print(img.shape)
img = img + tf.random_normal(
shape=tf.shape(img), mean=0.0, stddev=255.0 * 0.2)
else:
img, gtboxes_and_label = image_preprocess.short_side_resize(
img_tensor=img,
gtboxes_and_label=gtboxes_and_label,
target_shortside_len=shortside_len)
return img_name, img, gtboxes_and_label, num_objects
def preprocess_img(img_plac, gtbox_plac):
'''
:param img_plac: [H, W, 3] uint 8 img. In RGB.
:param gtbox_plac: shape of [-1, 5]. [xmin, ymin, xmax, ymax, label]
:return:
'''
img = tf.cast(img_plac, tf.float32)
# gtboxes_and_label = tf.cast(gtbox_plac, tf.float32)
img, gtboxes_and_label = image_preprocess.short_side_resize(img_tensor=img,
gtboxes_and_label=gtbox_plac,
target_shortside_len=cfgs.IMG_SHORT_SIDE_LEN,
length_limitation=cfgs.IMG_MAX_LENGTH)
img, gtboxes_and_label = image_preprocess.random_flip_left_right(img_tensor=img,
gtboxes_and_label=gtboxes_and_label)
if cfgs.NET_NAME in ['resnet152_v1d', 'resnet101_v1d', 'resnet50_v1d']:
img = img / 255 - tf.constant([[cfgs.PIXEL_MEAN_]])
else:
img = img - tf.constant([[cfgs.PIXEL_MEAN]])
img_batch = tf.expand_dims(img, axis=0)
# gtboxes_and_label = tf.Print(gtboxes_and_label, [tf.shape(gtboxes_and_label)], message='gtbox shape')
return img_batch, gtboxes_and_label
def read_and_prepocess_single_img(filename_queue, shortside_len, is_training):
img_name, img, gtboxes_and_label, num_objects = read_single_example_and_decode(filename_queue)
# img = tf.image.per_image_standardization(img)
img = tf.cast(img, tf.float32)
img = img - tf.constant([103.939, 116.779, 123.68])
if is_training:
img, gtboxes_and_label = image_preprocess.short_side_resize(img_tensor=img, gtboxes_and_label=gtboxes_and_label,
target_shortside_len=shortside_len)
img, gtboxes_and_label = image_preprocess.random_flip_left_right(img_tensor=img, gtboxes_and_label=gtboxes_and_label)
else:
img, gtboxes_and_label = image_preprocess.short_side_resize(img_tensor=img, gtboxes_and_label=gtboxes_and_label,
target_shortside_len=shortside_len)
return img_name, img, gtboxes_and_label, num_objects
def read_and_prepocess_single_img(filename_queue, shortside_len, is_training):
img_name, img, gtboxes_and_label, num_objects = read_single_example_and_decode(filename_queue)
img = tf.cast(img, tf.float32)
img = img - tf.constant([103.939, 116.779, 123.68])
if is_training:
img, gtboxes_and_label = image_preprocess.short_side_resize(img_tensor=img, gtboxes_and_label=gtboxes_and_label,
target_shortside_len=shortside_len)
img, gtboxes_and_label = image_preprocess.random_flip_left_right(img_tensor=img, gtboxes_and_label=gtboxes_and_label)
else:
img, gtboxes_and_label = image_preprocess.short_side_resize(img_tensor=img, gtboxes_and_label=gtboxes_and_label,
target_shortside_len=shortside_len)
return img_name, img, gtboxes_and_label, num_objects
def read_and_prepocess_single_img(raw_dataset, shortside_len, is_training):
img_name, img, gtboxes_and_label, num_objects = read_single_example_and_decode(raw_dataset)
img = tf.cast(img, tf.float32)
img = img - tf.constant(cfgs.PIXEL_MEAN)
if is_training:
img, gtboxes_and_label = image_preprocess.short_side_resize(img_tensor=img, gtboxes_and_label=gtboxes_and_label,
target_shortside_len=shortside_len)
img, gtboxes_and_label = image_preprocess.random_flip_left_right(img_tensor=img,
gtboxes_and_label=gtboxes_and_label)
else:
img, gtboxes_and_label = image_preprocess.short_side_resize(img_tensor=img, gtboxes_and_label=gtboxes_and_label,
target_shortside_len=shortside_len)
return img_name, img, gtboxes_and_label, num_objects
def read_and_preprocess_single_image(filename_queue, shortside_len,
longside_len, is_training):
image_name, image, gt_boxes_and_label = read_single_example_and_decode(
filename_queue)
image = tf.cast(image, tf.float32)
# image = image - tf.constant([103.939, 116.779, 123.68])
if is_training:
image, gt_boxes_and_label = image_preprocess.short_side_resize(
image, gt_boxes_and_label, shortside_len, longside_len)
image, gt_boxes_and_label = image_preprocess.random_flip_left_right(
image, gt_boxes_and_label) # 随机水平翻转
else:
image, gt_boxes_and_label = image_preprocess.short_side_resize(
image, gt_boxes_and_label, shortside_len, longside_len)
# image = tf.divide(image, 255.)
return image_name, image, gt_boxes_and_label
def read_and_prepocess_single_img(filename_queue, shortside_len, is_training):
img_name, img, gtboxes_and_label, num_objects = read_single_example_and_decode(
filename_queue)
img = tf.cast(img, tf.float32)
if is_training:
# prob is 0.3: convert to gray
img = image_preprocess.random_rgb2gray(
img_tensor=img, gtboxes_and_label=gtboxes_and_label)
# rotate with 0.5 prob. and if rotate, if will random choose a theta from : tf.range(-90, 90+16, delta=15)
img, gtboxes_and_label = image_preprocess.random_rotate_img(
img_tensor=img, gtboxes_and_label=gtboxes_and_label)
img, gtboxes_and_label = image_preprocess.short_side_resize(
img_tensor=img,
gtboxes_and_label=gtboxes_and_label,
target_shortside_len=shortside_len,
max_len=cfgs.IMG_MAX_LENGTH)
img, gtboxes_and_label = image_preprocess.random_flip_left_right(
img_tensor=img, gtboxes_and_label=gtboxes_and_label)
img, gtboxes_and_label = image_preprocess.random_flip_up_dowm(
img_tensor=img, gtboxes_and_label=gtboxes_and_label)
else:
img, gtboxes_and_label = image_preprocess.short_side_resize(
img_tensor=img,
gtboxes_and_label=gtboxes_and_label,
target_shortside_len=shortside_len)
if cfgs.MXNET_NORM:
print("Use Mxnet Norm")
img /= 255.0
img = img - tf.constant([[cfgs.MXNET_MEAN]])
img = img / tf.constant(cfgs.MXNET_STD)
else:
img = img - tf.constant([[cfgs.PIXEL_MEAN]]) # sub pixel mean at last
if cfgs.NET_NAME.endswith(('b', 'd')):
print("Note: Use Mxnet ResNet, But Do Not Norm Img like MxNet....")
print('\n')
return img_name, img, gtboxes_and_label, num_objects
def preprocess_img(img_plac, gtbox_plac):
'''
:param img_plac: [H, W, 3] uint 8 img. In RGB.
:param gtbox_plac: shape of [-1, 5]. [xmin, ymin, xmax, ymax, label]
:return:
'''
img = tf.cast(img_plac, tf.float32)
# gtboxes_and_label = tf.cast(gtbox_plac, tf.float32)
img, gtboxes_and_label = image_preprocess.short_side_resize(img_tensor=img,
gtboxes_and_label=gtbox_plac,
target_shortside_len=cfgs.IMG_SHORT_SIDE_LEN,
length_limitation=cfgs.IMG_MAX_LENGTH)
img, gtboxes_and_label = image_preprocess.random_flip_left_right(img_tensor=img,
gtboxes_and_label=gtboxes_and_label)
img = img - tf.constant([[cfgs.PIXEL_MEAN]])
img_batch = tf.expand_dims(img, axis=0)
return img_batch, gtboxes_and_label
def next_batch(dataset_name, batch_size, shortside_len, is_training):
'''
:return:
img_name_batch: shape(1, 1)
img_batch: shape:(1, new_imgH, new_imgW, C)
gtboxes_and_label_batch: shape(1, Num_Of_objects, 5] .each row is [x1, y1, x2, y2, label]
'''
assert batch_size == 1, "we only support batch_size is 1.We may support large batch_size in the future"
if dataset_name not in ['jyzdata', 'DOTA', 'ship', 'ICDAR2015', 'pascal', 'coco', 'DOTA_TOTAL', 'WIDER']:
raise ValueError('dataSet name must be in pascal, coco spacenet and ship')
if is_training:
pattern = os.path.join('/home/work/tfrecord/', dataset_name + '_val.tfrecord')
else:
pattern = os.path.join('/content/drive/', dataset_name + '_test.tfrecord')
print('tfrecord path is -->', os.path.abspath(pattern))
# filename_tensorlist = tf.train.match_filenames_once(pattern)
# filename_queue = tf.train.string_input_producer([pattern])
raw_dataset = tf.data.TFRecordDataset(pattern)
raw_dataset = raw_dataset.map(_parse_image_function)
raw_dataset = raw_dataset.repeat()
raw_dataset = raw_dataset.shuffle(1000)
raw_dataset = tf.data.make_one_shot_iterator(raw_dataset)
# parsed_image_dataset = raw_image_dataset.map(_parse_image_function)
# raw_dataset = tf.python_io.tf_record_iterator(path=pattern)
shortside_len = tf.constant(shortside_len)
shortside_len = tf.random_shuffle(shortside_len)[0]
single_ex = raw_dataset.get_next()
img_name = single_ex['img_name']
img_height = tf.cast(single_ex['img_height'], tf.int32)
img_width = tf.cast(single_ex['img_width'], tf.int32)
img = tf.decode_raw(single_ex['img'], tf.uint8)
img = tf.reshape(img, shape=[img_height, img_width, 3])
gtboxes_and_label = tf.decode_raw(single_ex['gtboxes_and_label'], tf.int32)
gtboxes_and_label = tf.reshape(gtboxes_and_label, [-1, 9])
num_obs = tf.cast(single_ex['num_objects'], tf.int32)
img = tf.cast(img, tf.float32)
img = img - tf.constant(cfgs.PIXEL_MEAN)
if is_training:
img, gtboxes_and_label = image_preprocess.short_side_resize(img_tensor=img, gtboxes_and_label=gtboxes_and_label,
target_shortside_len=shortside_len)
img, gtboxes_and_label = image_preprocess.random_flip_left_right(img_tensor=img,
gtboxes_and_label=gtboxes_and_label)
else:
img, gtboxes_and_label = image_preprocess.short_side_resize(img_tensor=img, gtboxes_and_label=gtboxes_and_label,
target_shortside_len=shortside_len)
# img_name, img, gtboxes_and_label, num_obs = read_and_prepocess_single_img(raw_dataset, shortside_len,
# is_training=is_training)
img_name_batch, img_batch, gtboxes_and_label_batch , num_obs_batch = \
tf.train.batch(
[img_name, tf.to_float(img), gtboxes_and_label, num_obs],
batch_size=batch_size,
capacity=1,
num_threads=1,
dynamic_pad=True)
return img_name_batch, img_batch, gtboxes_and_label_batch, num_obs_batch
