preprocessor.py

# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================

"""Preprocess images and bounding boxes for detection.

We perform two sets of operations in preprocessing stage:
(a) operations that are applied to both training and testing data,
(b) operations that are applied only to training data for the purpose of
    data augmentation.

A preprocessing function receives a set of inputs,
e.g. an image and bounding boxes,
performs an operation on them, and returns them.
Some examples are: randomly cropping the image, randomly mirroring the image,
                   randomly changing the brightness, contrast, hue and
                   randomly jittering the bounding boxes.

The preprocess function receives a tensor_dict which is a dictionary that maps
different field names to their tensors. For example,
tensor_dict[fields.InputDataFields.image] holds the image tensor.
The image is a rank 4 tensor: [1, height, width, channels] with
dtype=tf.float32. The groundtruth_boxes is a rank 2 tensor: [N, 4] where
in each row there is a box with [ymin xmin ymax xmax].
Boxes are in normalized coordinates meaning
their coordinate values range in [0, 1]

Important Note: In tensor_dict, images is a rank 4 tensor, but preprocessing
functions receive a rank 3 tensor for processing the image. Thus, inside the
preprocess function we squeeze the image to become a rank 3 tensor and then
we pass it to the functions. At the end of the preprocess we expand the image
back to rank 4.
"""

import sys
import tensorflow as tf

from tensorflow.python.ops import control_flow_ops

from object_detection.core import box_list
from object_detection.core import box_list_ops
from object_detection.core import keypoint_ops
from object_detection.core import standard_fields as fields


def _apply_with_random_selector(x, func, num_cases):
  """Computes func(x, sel), with sel sampled from [0...num_cases-1].

  Args:
    x: input Tensor.
    func: Python function to apply.
    num_cases: Python int32, number of cases to sample sel from.

  Returns:
    The result of func(x, sel), where func receives the value of the
    selector as a python integer, but sel is sampled dynamically.
  """
  rand_sel = tf.random_uniform([], maxval=num_cases, dtype=tf.int32)
  # Pass the real x only to one of the func calls.
  return control_flow_ops.merge([func(
      control_flow_ops.switch(x, tf.equal(rand_sel, case))[1], case)
                                 for case in range(num_cases)])[0]


def _apply_with_random_selector_tuples(x, func, num_cases):
  """Computes func(x, sel), with sel sampled from [0...num_cases-1].

  Args:
    x: A tuple of input tensors.
    func: Python function to apply.
    num_cases: Python int32, number of cases to sample sel from.

  Returns:
    The result of func(x, sel), where func receives the value of the
    selector as a python integer, but sel is sampled dynamically.
  """
  num_inputs = len(x)
  rand_sel = tf.random_uniform([], maxval=num_cases, dtype=tf.int32)
  # Pass the real x only to one of the func calls.

  tuples = [list() for t in x]
  for case in range(num_cases):
    new_x = [control_flow_ops.switch(t, tf.equal(rand_sel, case))[1] for t in x]
    output = func(tuple(new_x), case)
    for j in range(num_inputs):
      tuples[j].append(output[j])

  for i in range(num_inputs):
    tuples[i] = control_flow_ops.merge(tuples[i])[0]
  return tuple(tuples)


def _random_integer(minval, maxval, seed):
  """Returns a random 0-D tensor between minval and maxval.

  Args:
    minval: minimum value of the random tensor.
    maxval: maximum value of the random tensor.
    seed: random seed.

  Returns:
    A random 0-D tensor between minval and maxval.
  """
  return tf.random_uniform(
      [], minval=minval, maxval=maxval, dtype=tf.int32, seed=seed)


def normalize_image(image, original_minval, original_maxval, target_minval,
                    target_maxval):
  """Normalizes pixel values in the image.

  Moves the pixel values from the current [original_minval, original_maxval]
  range to a the [target_minval, target_maxval] range.

  Args:
    image: rank 3 float32 tensor containing 1
           image -> [height, width, channels].
    original_minval: current image minimum value.
    original_maxval: current image maximum value.
    target_minval: target image minimum value.
    target_maxval: target image maximum value.

  Returns:
    image: image which is the same shape as input image.
  """
  with tf.name_scope('NormalizeImage', values=[image]):
    original_minval = float(original_minval)
    original_maxval = float(original_maxval)
    target_minval = float(target_minval)
    target_maxval = float(target_maxval)
    image = tf.to_float(image)
    image = tf.subtract(image, original_minval)
    image = tf.multiply(image, (target_maxval - target_minval) /
                        (original_maxval - original_minval))
    image = tf.add(image, target_minval)
    return image


def flip_boxes(boxes):
  """Left-right flip the boxes.

  Args:
    boxes: rank 2 float32 tensor containing the bounding boxes -> [N, 4].
           Boxes are in normalized form meaning their coordinates vary
           between [0, 1].
           Each row is in the form of [ymin, xmin, ymax, xmax].

  Returns:
    Flipped boxes.
  """
  # Flip boxes.
  ymin, xmin, ymax, xmax = tf.split(value=boxes, num_or_size_splits=4, axis=1)
  flipped_xmin = tf.subtract(1.0, xmax)
  flipped_xmax = tf.subtract(1.0, xmin)
  flipped_boxes = tf.concat([ymin, flipped_xmin, ymax, flipped_xmax], 1)
  return flipped_boxes


def retain_boxes_above_threshold(
    boxes, labels, label_scores, masks=None, keypoints=None, threshold=0.0):
  """Retains boxes whose label score is above a given threshold.

  If the label score for a box is missing (represented by NaN), the box is
  retained. The boxes that don't pass the threshold will not appear in the
  returned tensor.

  Args:
    boxes: float32 tensor of shape [num_instance, 4] representing boxes
      location in normalized coordinates.
    labels: rank 1 int32 tensor of shape [num_instance] containing the object
      classes.
    label_scores: float32 tensor of shape [num_instance] representing the
      score for each box.
    masks: (optional) rank 3 float32 tensor with shape
      [num_instances, height, width] containing instance masks. The masks are of
      the same height, width as the input `image`.
    keypoints: (optional) rank 3 float32 tensor with shape
      [num_instances, num_keypoints, 2]. The keypoints are in y-x normalized
      coordinates.
    threshold: scalar python float.

  Returns:
    retained_boxes: [num_retained_instance, 4]
    retianed_labels: [num_retained_instance]
    retained_label_scores: [num_retained_instance]

    If masks, or keypoints are not None, the function also returns:

    retained_masks: [num_retained_instance, height, width]
    retained_keypoints: [num_retained_instance, num_keypoints, 2]
  """
  with tf.name_scope('RetainBoxesAboveThreshold',
                     values=[boxes, labels, label_scores]):
    indices = tf.where(
        tf.logical_or(label_scores > threshold, tf.is_nan(label_scores)))
    indices = tf.squeeze(indices, axis=1)
    retained_boxes = tf.gather(boxes, indices)
    retained_labels = tf.gather(labels, indices)
    retained_label_scores = tf.gather(label_scores, indices)
    result = [retained_boxes, retained_labels, retained_label_scores]

    if masks is not None:
      retained_masks = tf.gather(masks, indices)
      result.append(retained_masks)

    if keypoints is not None:
      retained_keypoints = tf.gather(keypoints, indices)
      result.append(retained_keypoints)

    return result


def _flip_masks(masks):
  """Left-right flips masks.

  Args:
    masks: rank 3 float32 tensor with shape
      [num_instances, height, width] representing instance masks.

  Returns:
    flipped masks: rank 3 float32 tensor with shape
      [num_instances, height, width] representing instance masks.
  """
  return masks[:, :, ::-1]


def random_horizontal_flip(
    image,
    boxes=None,
    masks=None,
    keypoints=None,
    keypoint_flip_permutation=None,
    seed=None):
  """Randomly decides whether to mirror the image and detections or not.

  The probability of flipping the image is 50%.

  Args:
    image: rank 3 float32 tensor with shape [height, width, channels].
    boxes: (optional) rank 2 float32 tensor with shape [N, 4]
           containing the bounding boxes.
           Boxes are in normalized form meaning their coordinates vary
           between [0, 1].
           Each row is in the form of [ymin, xmin, ymax, xmax].
    masks: (optional) rank 3 float32 tensor with shape
           [num_instances, height, width] containing instance masks. The masks
           are of the same height, width as the input `image`.
    keypoints: (optional) rank 3 float32 tensor with shape
               [num_instances, num_keypoints, 2]. The keypoints are in y-x
               normalized coordinates.
    keypoint_flip_permutation: rank 1 int32 tensor containing keypoint flip
                               permutation.
    seed: random seed

  Returns:
    image: image which is the same shape as input image.

    If boxes, masks, keypoints, and keypoint_flip_permutation is not None,
    the function also returns the following tensors.

    boxes: rank 2 float32 tensor containing the bounding boxes -> [N, 4].
           Boxes are in normalized form meaning their coordinates vary
           between [0, 1].
    masks: rank 3 float32 tensor with shape [num_instances, height, width]
           containing instance masks.
    keypoints: rank 3 float32 tensor with shape
               [num_instances, num_keypoints, 2]

  Raises:
    ValueError: if keypoints are provided but keypoint_flip_permutation is not.
  """
  def _flip_image(image):
    # flip image
    image_flipped = tf.image.flip_left_right(image)
    return image_flipped

  if keypoints is not None and keypoint_flip_permutation is None:
    raise ValueError(
        'keypoints are provided but keypoints_flip_permutation is not provided')

  with tf.name_scope('RandomHorizontalFlip', values=[image, boxes]):
    result = []
    # random variable defining whether to do flip or not
    do_a_flip_random = tf.random_uniform([], seed=seed)
    # flip only if there are bounding boxes in image!
    do_a_flip_random = tf.logical_and(
        tf.greater(tf.size(boxes), 0), tf.greater(do_a_flip_random, 0.5))

    # flip image
    image = tf.cond(do_a_flip_random, lambda: _flip_image(image), lambda: image)
    result.append(image)

    # flip boxes
    if boxes is not None:
      boxes = tf.cond(
          do_a_flip_random, lambda: flip_boxes(boxes), lambda: boxes)
      result.append(boxes)

    # flip masks
    if masks is not None:
      masks = tf.cond(
          do_a_flip_random, lambda: _flip_masks(masks), lambda: masks)
      result.append(masks)

    # flip keypoints
    if keypoints is not None and keypoint_flip_permutation is not None:
      permutation = keypoint_flip_permutation
      keypoints = tf.cond(
          do_a_flip_random,
          lambda: keypoint_ops.flip_horizontal(keypoints, 0.5, permutation),
          lambda: keypoints)
      result.append(keypoints)

    return tuple(result)


def random_pixel_value_scale(image, minval=0.9, maxval=1.1, seed=None):
  """Scales each value in the pixels of the image.

     This function scales each pixel independent of the other ones.
     For each value in image tensor, draws a random number between
     minval and maxval and multiples the values with them.

  Args:
    image: rank 3 float32 tensor contains 1 image -> [height, width, channels]
           with pixel values varying between [0, 1].
    minval: lower ratio of scaling pixel values.
    maxval: upper ratio of scaling pixel values.
    seed: random seed.

  Returns:
    image: image which is the same shape as input image.
  """
  with tf.name_scope('RandomPixelValueScale', values=[image]):
    color_coef = tf.random_uniform(
        tf.shape(image),
        minval=minval,
        maxval=maxval,
        dtype=tf.float32,
        seed=seed)
    image = tf.multiply(image, color_coef)
    image = tf.clip_by_value(image, 0.0, 1.0)

  return image


def random_image_scale(image,
                       masks=None,
                       min_scale_ratio=0.5,
                       max_scale_ratio=2.0,
                       seed=None):
  """Scales the image size.

  Args:
    image: rank 3 float32 tensor contains 1 image -> [height, width, channels].
    masks: (optional) rank 3 float32 tensor containing masks with
      size [height, width, num_masks]. The value is set to None if there are no
      masks.
    min_scale_ratio: minimum scaling ratio.
    max_scale_ratio: maximum scaling ratio.
    seed: random seed.

  Returns:
    image: image which is the same rank as input image.
    masks: If masks is not none, resized masks which are the same rank as input
      masks will be returned.
  """
  with tf.name_scope('RandomImageScale', values=[image]):
    result = []
    image_shape = tf.shape(image)
    image_height = image_shape[0]
    image_width = image_shape[1]
    size_coef = tf.random_uniform([],
                                  minval=min_scale_ratio,
                                  maxval=max_scale_ratio,
                                  dtype=tf.float32, seed=seed)
    image_newysize = tf.to_int32(
        tf.multiply(tf.to_float(image_height), size_coef))
    image_newxsize = tf.to_int32(
        tf.multiply(tf.to_float(image_width), size_coef))
    image = tf.image.resize_images(
        image, [image_newysize, image_newxsize], align_corners=True)
    result.append(image)
    if masks:
      masks = tf.image.resize_nearest_neighbor(
          masks, [image_newysize, image_newxsize], align_corners=True)
      result.append(masks)
    return tuple(result)


def random_rgb_to_gray(image, probability=0.1, seed=None):
  """Changes the image from RGB to Grayscale with the given probability.

  Args:
    image: rank 3 float32 tensor contains 1 image -> [height, width, channels]
           with pixel values varying between [0, 1].
    probability: the probability of returning a grayscale image.
            The probability should be a number between [0, 1].
    seed: random seed.

  Returns:
    image: image which is the same shape as input image.
  """
  def _image_to_gray(image):
    image_gray1 = tf.image.rgb_to_grayscale(image)
    image_gray3 = tf.image.grayscale_to_rgb(image_gray1)
    return image_gray3

  with tf.name_scope('RandomRGBtoGray', values=[image]):
    # random variable defining whether to do flip or not
    do_gray_random = tf.random_uniform([], seed=seed)

    image = tf.cond(
        tf.greater(do_gray_random, probability), lambda: image,
        lambda: _image_to_gray(image))

  return image


def random_adjust_brightness(image, max_delta=0.2):
  """Randomly adjusts brightness.

  Makes sure the output image is still between 0 and 1.

  Args:
    image: rank 3 float32 tensor contains 1 image -> [height, width, channels]
           with pixel values varying between [0, 1].
    max_delta: how much to change the brightness. A value between [0, 1).

  Returns:
    image: image which is the same shape as input image.
    boxes: boxes which is the same shape as input boxes.
  """
  with tf.name_scope('RandomAdjustBrightness', values=[image]):
    image = tf.image.random_brightness(image, max_delta)
    image = tf.clip_by_value(image, clip_value_min=0.0, clip_value_max=1.0)
    return image


def random_adjust_contrast(image, min_delta=0.8, max_delta=1.25):
  """Randomly adjusts contrast.

  Makes sure the output image is still between 0 and 1.

  Args:
    image: rank 3 float32 tensor contains 1 image -> [height, width, channels]
           with pixel values varying between [0, 1].
    min_delta: see max_delta.
    max_delta: how much to change the contrast. Contrast will change with a
               value between min_delta and max_delta. This value will be
               multiplied to the current contrast of the image.

  Returns:
    image: image which is the same shape as input image.
  """
  with tf.name_scope('RandomAdjustContrast', values=[image]):
    image = tf.image.random_contrast(image, min_delta, max_delta)
    image = tf.clip_by_value(image, clip_value_min=0.0, clip_value_max=1.0)
    return image


def random_adjust_hue(image, max_delta=0.02):
  """Randomly adjusts hue.

  Makes sure the output image is still between 0 and 1.

  Args:
    image: rank 3 float32 tensor contains 1 image -> [height, width, channels]
           with pixel values varying between [0, 1].
    max_delta: change hue randomly with a value between 0 and max_delta.

  Returns:
    image: image which is the same shape as input image.
  """
  with tf.name_scope('RandomAdjustHue', values=[image]):
    image = tf.image.random_hue(image, max_delta)
    image = tf.clip_by_value(image, clip_value_min=0.0, clip_value_max=1.0)
    return image


def random_adjust_saturation(image, min_delta=0.8, max_delta=1.25):
  """Randomly adjusts saturation.

  Makes sure the output image is still between 0 and 1.

  Args:
    image: rank 3 float32 tensor contains 1 image -> [height, width, channels]
           with pixel values varying between [0, 1].
    min_delta: see max_delta.
    max_delta: how much to change the saturation. Saturation will change with a
               value between min_delta and max_delta. This value will be
               multiplied to the current saturation of the image.

  Returns:
    image: image which is the same shape as input image.
  """
  with tf.name_scope('RandomAdjustSaturation', values=[image]):
    image = tf.image.random_saturation(image, min_delta, max_delta)
    image = tf.clip_by_value(image, clip_value_min=0.0, clip_value_max=1.0)
    return image


def random_distort_color(image, color_ordering=0):
  """Randomly distorts color.

  Randomly distorts color using a combination of brightness, hue, contrast
  and saturation changes. Makes sure the output image is still between 0 and 1.

  Args:
    image: rank 3 float32 tensor contains 1 image -> [height, width, channels]
           with pixel values varying between [0, 1].
    color_ordering: Python int, a type of distortion (valid values: 0, 1).

  Returns:
    image: image which is the same shape as input image.

  Raises:
    ValueError: if color_ordering is not in {0, 1}.
  """
  with tf.name_scope('RandomDistortColor', values=[image]):
    if color_ordering == 0:
      image = tf.image.random_brightness(image, max_delta=32. / 255.)
      image = tf.image.random_saturation(image, lower=0.5, upper=1.5)
      image = tf.image.random_hue(image, max_delta=0.2)
      image = tf.image.random_contrast(image, lower=0.5, upper=1.5)
    elif color_ordering == 1:
      image = tf.image.random_brightness(image, max_delta=32. / 255.)
      image = tf.image.random_contrast(image, lower=0.5, upper=1.5)
      image = tf.image.random_saturation(image, lower=0.5, upper=1.5)
      image = tf.image.random_hue(image, max_delta=0.2)
    else:
      raise ValueError('color_ordering must be in {0, 1}')

    # The random_* ops do not necessarily clamp.
    image = tf.clip_by_value(image, 0.0, 1.0)
    return image


def random_jitter_boxes(boxes, ratio=0.05, seed=None):
  """Randomly jitter boxes in image.

  Args:
    boxes: rank 2 float32 tensor containing the bounding boxes -> [N, 4].
           Boxes are in normalized form meaning their coordinates vary
           between [0, 1].
           Each row is in the form of [ymin, xmin, ymax, xmax].
    ratio: The ratio of the box width and height that the corners can jitter.
           For example if the width is 100 pixels and ratio is 0.05,
           the corners can jitter up to 5 pixels in the x direction.
    seed: random seed.

  Returns:
    boxes: boxes which is the same shape as input boxes.
  """
  def random_jitter_box(box, ratio, seed):
    """Randomly jitter box.

    Args:
      box: bounding box [1, 1, 4].
      ratio: max ratio between jittered box and original box,
      a number between [0, 0.5].
      seed: random seed.

    Returns:
      jittered_box: jittered box.
    """
    rand_numbers = tf.random_uniform(
        [1, 1, 4], minval=-ratio, maxval=ratio, dtype=tf.float32, seed=seed)
    box_width = tf.subtract(box[0, 0, 3], box[0, 0, 1])
    box_height = tf.subtract(box[0, 0, 2], box[0, 0, 0])
    hw_coefs = tf.stack([box_height, box_width, box_height, box_width])
    hw_rand_coefs = tf.multiply(hw_coefs, rand_numbers)
    jittered_box = tf.add(box, hw_rand_coefs)
    jittered_box = tf.clip_by_value(jittered_box, 0.0, 1.0)
    return jittered_box

  with tf.name_scope('RandomJitterBoxes', values=[boxes]):
    # boxes are [N, 4]. Lets first make them [N, 1, 1, 4]
    boxes_shape = tf.shape(boxes)
    boxes = tf.expand_dims(boxes, 1)
    boxes = tf.expand_dims(boxes, 2)

    distorted_boxes = tf.map_fn(
        lambda x: random_jitter_box(x, ratio, seed), boxes, dtype=tf.float32)

    distorted_boxes = tf.reshape(distorted_boxes, boxes_shape)

    return distorted_boxes


def _strict_random_crop_image(image,
                              boxes,
                              labels,
                              masks=None,
                              keypoints=None,
                              min_object_covered=1.0,
                              aspect_ratio_range=(0.75, 1.33),
                              area_range=(0.1, 1.0),
                              overlap_thresh=0.3):
  """Performs random crop.

  Note: boxes will be clipped to the crop. Keypoint coordinates that are
  outside the crop will be set to NaN, which is consistent with the original
  keypoint encoding for non-existing keypoints. This function always crops
  the image and is supposed to be used by `random_crop_image` function which
  sometimes returns image unchanged.

  Args:
    image: rank 3 float32 tensor containing 1 image -> [height, width, channels]
           with pixel values varying between [0, 1].
    boxes: rank 2 float32 tensor containing the bounding boxes with shape
           [num_instances, 4].
           Boxes are in normalized form meaning their coordinates vary
           between [0, 1].
           Each row is in the form of [ymin, xmin, ymax, xmax].
    labels: rank 1 int32 tensor containing the object classes.
    masks: (optional) rank 3 float32 tensor with shape
           [num_instances, height, width] containing instance masks. The masks
           are of the same height, width as the input `image`.
    keypoints: (optional) rank 3 float32 tensor with shape
               [num_instances, num_keypoints, 2]. The keypoints are in y-x
               normalized coordinates.
    min_object_covered: the cropped image must cover at least this fraction of
                        at least one of the input bounding boxes.
    aspect_ratio_range: allowed range for aspect ratio of cropped image.
    area_range: allowed range for area ratio between cropped image and the
                original image.
    overlap_thresh: minimum overlap thresh with new cropped
                    image to keep the box.

  Returns:
    image: image which is the same rank as input image.
    boxes: boxes which is the same rank as input boxes.
           Boxes are in normalized form.
    labels: new labels.

    If masks, or keypoints is not None, the function also returns:

    masks: rank 3 float32 tensor with shape [num_instances, height, width]
           containing instance masks.
    keypoints: rank 3 float32 tensor with shape
               [num_instances, num_keypoints, 2]
  """
  with tf.name_scope('RandomCropImage', values=[image, boxes]):
    image_shape = tf.shape(image)

    # boxes are [N, 4]. Lets first make them [N, 1, 4].
    boxes_expanded = tf.expand_dims(
        tf.clip_by_value(
            boxes, clip_value_min=0.0, clip_value_max=1.0), 1)

    sample_distorted_bounding_box = tf.image.sample_distorted_bounding_box(
        image_shape,
        bounding_boxes=boxes_expanded,
        min_object_covered=min_object_covered,
        aspect_ratio_range=aspect_ratio_range,
        area_range=area_range,
        max_attempts=100,
        use_image_if_no_bounding_boxes=True)

    im_box_begin, im_box_size, im_box = sample_distorted_bounding_box

    new_image = tf.slice(image, im_box_begin, im_box_size)
    new_image.set_shape([None, None, image.get_shape()[2]])

    # [1, 4]
    im_box_rank2 = tf.squeeze(im_box, squeeze_dims=[0])
    # [4]
    im_box_rank1 = tf.squeeze(im_box)

    boxlist = box_list.BoxList(boxes)
    boxlist.add_field('labels', labels)

    im_boxlist = box_list.BoxList(im_box_rank2)

    # remove boxes that are outside cropped image
    boxlist, inside_window_ids = box_list_ops.prune_completely_outside_window(
        boxlist, im_box_rank1)

    # remove boxes that are outside image
    overlapping_boxlist, keep_ids = box_list_ops.prune_non_overlapping_boxes(
        boxlist, im_boxlist, overlap_thresh)

    # change the coordinate of the remaining boxes
    new_labels = overlapping_boxlist.get_field('labels')
    new_boxlist = box_list_ops.change_coordinate_frame(overlapping_boxlist,
                                                       im_box_rank1)
    new_boxes = new_boxlist.get()
    new_boxes = tf.clip_by_value(
        new_boxes, clip_value_min=0.0, clip_value_max=1.0)

    result = [new_image, new_boxes, new_labels]

    if masks is not None:
      masks_of_boxes_inside_window = tf.gather(masks, inside_window_ids)
      masks_of_boxes_completely_inside_window = tf.gather(
          masks_of_boxes_inside_window, keep_ids)
      masks_box_begin = [0, im_box_begin[0], im_box_begin[1]]
      masks_box_size = [-1, im_box_size[0], im_box_size[1]]
      new_masks = tf.slice(
          masks_of_boxes_completely_inside_window,
          masks_box_begin, masks_box_size)
      result.append(new_masks)

    if keypoints is not None:
      keypoints_of_boxes_inside_window = tf.gather(keypoints, inside_window_ids)
      keypoints_of_boxes_completely_inside_window = tf.gather(
          keypoints_of_boxes_inside_window, keep_ids)
      new_keypoints = keypoint_ops.change_coordinate_frame(
          keypoints_of_boxes_completely_inside_window, im_box_rank1)
      new_keypoints = keypoint_ops.prune_outside_window(new_keypoints,
                                                        [0.0, 0.0, 1.0, 1.0])
      result.append(new_keypoints)

    return tuple(result)


def random_crop_image(image,
                      boxes,
                      labels,
                      masks=None,
                      keypoints=None,
                      min_object_covered=1.0,
                      aspect_ratio_range=(0.75, 1.33),
                      area_range=(0.1, 1.0),
                      overlap_thresh=0.3,
                      random_coef=0.0,
                      seed=None):
  """Randomly crops the image.

  Given the input image and its bounding boxes, this op randomly
  crops a subimage.  Given a user-provided set of input constraints,
  the crop window is resampled until it satisfies these constraints.
  If within 100 trials it is unable to find a valid crop, the original
  image is returned. See the Args section for a description of the input
  constraints. Both input boxes and returned Boxes are in normalized
  form (e.g., lie in the unit square [0, 1]).
  This function will return the original image with probability random_coef.

  Note: boxes will be clipped to the crop. Keypoint coordinates that are
  outside the crop will be set to NaN, which is consistent with the original
  keypoint encoding for non-existing keypoints.

  Args:
    image: rank 3 float32 tensor contains 1 image -> [height, width, channels]
           with pixel values varying between [0, 1].
    boxes: rank 2 float32 tensor containing the bounding boxes with shape
           [num_instances, 4].
           Boxes are in normalized form meaning their coordinates vary
           between [0, 1].
           Each row is in the form of [ymin, xmin, ymax, xmax].
    labels: rank 1 int32 tensor containing the object classes.
    masks: (optional) rank 3 float32 tensor with shape
           [num_instances, height, width] containing instance masks. The masks
           are of the same height, width as the input `image`.
    keypoints: (optional) rank 3 float32 tensor with shape
               [num_instances, num_keypoints, 2]. The keypoints are in y-x
               normalized coordinates.
    min_object_covered: the cropped image must cover at least this fraction of
                        at least one of the input bounding boxes.
    aspect_ratio_range: allowed range for aspect ratio of cropped image.
    area_range: allowed range for area ratio between cropped image and the
                original image.
    overlap_thresh: minimum overlap thresh with new cropped
                    image to keep the box.
    random_coef: a random coefficient that defines the chance of getting the
                 original image. If random_coef is 0, we will always get the
                 cropped image, and if it is 1.0, we will always get the
                 original image.
    seed: random seed.

  Returns:
    image: Image shape will be [new_height, new_width, channels].
    boxes: boxes which is the same rank as input boxes. Boxes are in normalized
           form.
    labels: new labels.

    If masks, or keypoints are not None, the function also returns:

    masks: rank 3 float32 tensor with shape [num_instances, height, width]
           containing instance masks.
    keypoints: rank 3 float32 tensor with shape
               [num_instances, num_keypoints, 2]
  """

  def strict_random_crop_image_fn():
    return _strict_random_crop_image(
        image,
        boxes,
        labels,
        masks=masks,
        keypoints=keypoints,
        min_object_covered=min_object_covered,
        aspect_ratio_range=aspect_ratio_range,
        area_range=area_range,
        overlap_thresh=overlap_thresh)

  # avoids tf.cond to make faster RCNN training on borg. See b/140057645.
  if random_coef < sys.float_info.min:
    result = strict_random_crop_image_fn()
  else:
    do_a_crop_random = tf.random_uniform([], seed=seed)
    do_a_crop_random = tf.greater(do_a_crop_random, random_coef)

    outputs = [image, boxes, labels]
    if masks is not None:
      outputs.append(masks)
    if keypoints is not None:
      outputs.append(keypoints)

    result = tf.cond(do_a_crop_random,
                     strict_random_crop_image_fn,
                     lambda: tuple(outputs))
  return result


def random_pad_image(image,
                     boxes,
                     min_image_size=None,
                     max_image_size=None,
                     pad_color=None,
                     seed=None):
  """Randomly pads the image.

  This function randomly pads the image with zeros. The final size of the
  padded image will be between min_image_size and max_image_size.
  if min_image_size is smaller than the input image size, min_image_size will
  be set to the input image size. The same for max_image_size. The input image
  will be located at a uniformly random location inside the padded image.
  The relative location of the boxes to the original image will remain the same.

  Args:
    image: rank 3 float32 tensor containing 1 image -> [height, width, channels]
           with pixel values varying between [0, 1].
    boxes: rank 2 float32 tensor containing the bounding boxes -> [N, 4].
           Boxes are in normalized form meaning their coordinates vary
           between [0, 1].
           Each row is in the form of [ymin, xmin, ymax, xmax].
    min_image_size: a tensor of size [min_height, min_width], type tf.int32.
                    If passed as None, will be set to image size
                    [height, width].
    max_image_size: a tensor of size [max_height, max_width], type tf.int32.
                    If passed as None, will be set to twice the
                    image [height * 2, width * 2].
    pad_color: padding color. A rank 1 tensor of [3] with dtype=tf.float32.
               if set as None, it will be set to average color of the input
               image.

    seed: random seed.

  Returns:
    image: Image shape will be [new_height, new_width, channels].
    boxes: boxes which is the same rank as input boxes. Boxes are in normalized
           form.
  """
  if pad_color is None:
    pad_color = tf.reduce_mean(image, reduction_indices=[0, 1])

  image_shape = tf.shape(image)
  image_height = image_shape[0]
  image_width = image_shape[1]

  if max_image_size is None:
    max_image_size = tf.stack([image_height * 2, image_width * 2])
  max_image_size = tf.maximum(max_image_size,
                              tf.stack([image_height, image_width]))

  if min_image_size is None:
    min_image_size = tf.stack([image_height, image_width])
  min_image_size = tf.maximum(min_image_size,
                              tf.stack([image_height, image_width]))

  target_height = tf.cond(
      max_image_size[0] > min_image_size[0],
      lambda: _random_integer(min_image_size[0], max_image_size[0], seed),
      lambda: max_image_size[0])

  target_width = tf.cond(
      max_image_size[1] > min_image_size[1],
      lambda: _random_integer(min_image_size[1], max_image_size[1], seed),
      lambda: max_image_size[1])

  offset_height = tf.cond(
      target_height > image_height,
      lambda: _random_integer(0, target_height - image_height, seed),
      lambda: tf.constant(0, dtype=tf.int32))

  offset_width = tf.cond(
      target_width > image_width,
      lambda: _random_integer(0, target_width - image_width, seed),
      lambda: tf.constant(0, dtype=tf.int32))

  new_image = tf.image.pad_to_bounding_box(
      image, offset_height=offset_height, offset_width=offset_width,
      target_height=target_height, target_width=target_width)

  # Setting color of the padded pixels
  image_ones = tf.ones_like(image)
  image_ones_padded = tf.image.pad_to_bounding_box(
      image_ones, offset_height=offset_height, offset_width=offset_width,
      target_height=target_height, target_width=target_width)
  image_color_paded = (1.0 - image_ones_padded) * pad_color
  new_image += image_color_paded

  # setting boxes
  new_window = tf.to_float(
      tf.stack([
          -offset_height, -offset_width, target_height - offset_height,
          target_width - offset_width
      ]))
  new_window /= tf.to_float(
      tf.stack([image_height, image_width, image_height, image_width]))
  boxlist = box_list.BoxList(boxes)
  new_boxlist = box_list_ops.change_coordinate_frame(boxlist, new_window)
  new_boxes = new_boxlist.get()

  return new_image, new_boxes


def random_crop_pad_image(image,
                          boxes,
                          labels,
                          min_object_covered=1.0,
                          aspect_ratio_range=(0.75, 1.33),
                          area_range=(0.1, 1.0),
                          overlap_thresh=0.3,
                          random_coef=0.0,
                          min_padded_size_ratio=None,
                          max_padded_size_ratio=None,
                          pad_color=None,
                          seed=None):
  """Randomly crops and pads the image.

  Given an input image and its bounding boxes, this op first randomly crops
  the image and then randomly pads the image with background values. Parameters
  min_padded_size_ratio and max_padded_size_ratio, determine the range of the
  final output image size.  Specifically, the final image size will have a size
  in the range of min_padded_size_ratio * tf.shape(image) and
  max_padded_size_ratio * tf.shape(image). Note that these ratios are with
  respect to the size of the original image, so we can't capture the same
  effect easily by independently applying RandomCropImage
  followed by RandomPadImage.

  Args:
    image: rank 3 float32 tensor containing 1 image -> [height, width, channels]
           with pixel values varying between [0, 1].
    boxes: rank 2 float32 tensor containing the bounding boxes -> [N, 4].
           Boxes are in normalized form meaning their coordinates vary
           between [0, 1].
           Each row is in the form of [ymin, xmin, ymax, xmax].
    labels: rank 1 int32 tensor containing the object classes.
    min_object_covered: the cropped image must cover at least this fraction of
                        at least one of the input bounding boxes.
    aspect_ratio_range: allowed range for aspect ratio of cropped image.
    area_range: allowed range for area ratio between cropped image and the
                original image.
    overlap_thresh: minimum overlap thresh with new cropped
                    image to keep the box.
    random_coef: a random coefficient that defines the chance of getting the
                 original image. If random_coef is 0, we will always get the
                 cropped image, and if it is 1.0, we will always get the
                 original image.
    min_padded_size_ratio: min ratio of padded image height and width to the
                           input image's height and width. If None, it will
                           be set to [0.0, 0.0].
    max_padded_size_ratio: max ratio of padded image height and width to the
                           input image's height and width. If None, it will
                           be set to [2.0, 2.0].
    pad_color: padding color. A rank 1 tensor of [3] with dtype=tf.float32.
               if set as None, it will be set to average color of the randomly
               cropped image.
    seed: random seed.

  Returns:
    padded_image: padded image.
    padded_boxes: boxes which is the same rank as input boxes. Boxes are in
                  normalized form.
    cropped_labels: cropped labels.
  """
  image_size = tf.shape(image)
  image_height = image_size[0]
  image_width = image_size[1]
  if min_padded_size_ratio is None:
    min_padded_size_ratio = tf.constant([0.0, 0.0], tf.float32)
  if max_padded_size_ratio is None:
    max_padded_size_ratio = tf.constant([2.0, 2.0], tf.float32)
  cropped_image, cropped_boxes, cropped_labels = random_crop_image(
      image=image,
      boxes=boxes,
      labels=labels,
      min_object_covered=min_object_covered,
      aspect_ratio_range=aspect_ratio_range,
      area_range=area_range,
      overlap_thresh=overlap_thresh,
      random_coef=random_coef,
      seed=seed)

  min_image_size = tf.to_int32(
      tf.to_float(tf.stack([image_height, image_width])) *
      min_padded_size_ratio)
  max_image_size = tf.to_int32(
      tf.to_float(tf.stack([image_height, image_width])) *
      max_padded_size_ratio)

  padded_image, padded_boxes = random_pad_image(
      cropped_image,
      cropped_boxes,
      min_image_size=min_image_size,
      max_image_size=max_image_size,
      pad_color=pad_color,
      seed=seed)

  return padded_image, padded_boxes, cropped_labels


def random_crop_to_aspect_ratio(image,
                                boxes,
                                labels,
                                masks=None,
                                keypoints=None,
                                aspect_ratio=1.0,
                                overlap_thresh=0.3,
                                seed=None):
  """Randomly crops an image to the specified aspect ratio.

  Randomly crops the a portion of the image such that the crop is of the
  specified aspect ratio, and the crop is as large as possible. If the specified
  aspect ratio is larger than the aspect ratio of the image, this op will
  randomly remove rows from the top and bottom of the image. If the specified
  aspect ratio is less than the aspect ratio of the image, this op will randomly
  remove cols from the left and right of the image. If the specified aspect
  ratio is the same as the aspect ratio of the image, this op will return the
  image.

  Args:
    image: rank 3 float32 tensor contains 1 image -> [height, width, channels]
           with pixel values varying between [0, 1].
    boxes: rank 2 float32 tensor containing the bounding boxes -> [N, 4].
           Boxes are in normalized form meaning their coordinates vary
           between [0, 1].
           Each row is in the form of [ymin, xmin, ymax, xmax].
    labels: rank 1 int32 tensor containing the object classes.
    masks: (optional) rank 3 float32 tensor with shape
           [num_instances, height, width] containing instance masks. The masks
           are of the same height, width as the input `image`.
    keypoints: (optional) rank 3 float32 tensor with shape
               [num_instances, num_keypoints, 2]. The keypoints are in y-x
               normalized coordinates.
    aspect_ratio: the aspect ratio of cropped image.
    overlap_thresh: minimum overlap thresh with new cropped
                    image to keep the box.
    seed: random seed.

  Returns:
    image: image which is the same rank as input image.
    boxes: boxes which is the same rank as input boxes.
           Boxes are in normalized form.
    labels: new labels.

    If masks, or keypoints is not None, the function also returns:

    masks: rank 3 float32 tensor with shape [num_instances, height, width]
           containing instance masks.
    keypoints: rank 3 float32 tensor with shape
               [num_instances, num_keypoints, 2]

  Raises:
    ValueError: If image is not a 3D tensor.
  """
  if len(image.get_shape()) != 3:
    raise ValueError('Image should be 3D tensor')

  with tf.name_scope('RandomCropToAspectRatio', values=[image]):
    image_shape = tf.shape(image)
    orig_height = image_shape[0]
    orig_width = image_shape[1]
    orig_aspect_ratio = tf.to_float(orig_width) / tf.to_float(orig_height)
    new_aspect_ratio = tf.constant(aspect_ratio, dtype=tf.float32)
    def target_height_fn():
      return tf.to_int32(
          tf.round(
              tf.to_float(orig_height) * orig_aspect_ratio / new_aspect_ratio))
    target_height = tf.cond(
        orig_aspect_ratio >= new_aspect_ratio,
        lambda: orig_height,
        target_height_fn)
    def target_width_fn():
      return tf.to_int32(
          tf.round(
              tf.to_float(orig_width) * new_aspect_ratio / orig_aspect_ratio))
    target_width = tf.cond(
        orig_aspect_ratio <= new_aspect_ratio,
        lambda: orig_width,
        target_width_fn)

    # either offset_height = 0 and offset_width is randomly chosen from
    # [0, offset_width - target_width), or else offset_width = 0 and
    # offset_height is randomly chosen from [0, offset_height - target_height)
    offset_height = _random_integer(0, orig_height - target_height + 1, seed)
    offset_width = _random_integer(0, orig_width - target_width + 1, seed)
    new_image = tf.image.crop_to_bounding_box(
        image, offset_height, offset_width, target_height, target_width)

    im_box = tf.stack([
        tf.to_float(offset_height) / tf.to_float(orig_height),
        tf.to_float(offset_width) / tf.to_float(orig_width),
        tf.to_float(offset_height + target_height) / tf.to_float(orig_height),
        tf.to_float(offset_width + target_width) / tf.to_float(orig_width)
    ])

    boxlist = box_list.BoxList(boxes)
    boxlist.add_field('labels', labels)

    im_boxlist = box_list.BoxList(tf.expand_dims(im_box, 0))

    # remove boxes whose overlap with the image is less than overlap_thresh
    overlapping_boxlist, keep_ids = box_list_ops.prune_non_overlapping_boxes(
        boxlist, im_boxlist, overlap_thresh)

    # change the coordinate of the remaining boxes
    new_labels = overlapping_boxlist.get_field('labels')
    new_boxlist = box_list_ops.change_coordinate_frame(overlapping_boxlist,
                                                       im_box)
    new_boxlist = box_list_ops.clip_to_window(new_boxlist,
                                              tf.constant(
                                                  [0.0, 0.0, 1.0, 1.0],
                                                  tf.float32))
    new_boxes = new_boxlist.get()

    result = [new_image, new_boxes, new_labels]

    if masks is not None:
      masks_inside_window = tf.gather(masks, keep_ids)
      masks_box_begin = tf.stack([0, offset_height, offset_width])
      masks_box_size = tf.stack([-1, target_height, target_width])
      new_masks = tf.slice(masks_inside_window, masks_box_begin, masks_box_size)
      result.append(new_masks)

    if keypoints is not None:
      keypoints_inside_window = tf.gather(keypoints, keep_ids)
      new_keypoints = keypoint_ops.change_coordinate_frame(
          keypoints_inside_window, im_box)
      new_keypoints = keypoint_ops.prune_outside_window(new_keypoints,
                                                        [0.0, 0.0, 1.0, 1.0])
      result.append(new_keypoints)

    return tuple(result)


def random_black_patches(image,
                         max_black_patches=10,
                         probability=0.5,
                         size_to_image_ratio=0.1,
                         random_seed=None):
  """Randomly adds some black patches to the image.

  This op adds up to max_black_patches square black patches of a fixed size
  to the image where size is specified via the size_to_image_ratio parameter.

  Args:
    image: rank 3 float32 tensor containing 1 image -> [height, width, channels]
           with pixel values varying between [0, 1].
    max_black_patches: number of times that the function tries to add a
                       black box to the image.
    probability: at each try, what is the chance of adding a box.
    size_to_image_ratio: Determines the ratio of the size of the black patches
                         to the size of the image.
                         box_size = size_to_image_ratio *
                                    min(image_width, image_height)
    random_seed: random seed.

  Returns:
    image
  """
  def add_black_patch_to_image(image):
    """Function for adding one patch to the image.

    Args:
      image: image

    Returns:
      image with a randomly added black box
    """
    image_shape = tf.shape(image)
    image_height = image_shape[0]
    image_width = image_shape[1]
    box_size = tf.to_int32(
        tf.multiply(
            tf.minimum(tf.to_float(image_height), tf.to_float(image_width)),
            size_to_image_ratio))
    normalized_y_min = tf.random_uniform(
        [], minval=0.0, maxval=(1.0 - size_to_image_ratio), seed=random_seed)
    normalized_x_min = tf.random_uniform(
        [], minval=0.0, maxval=(1.0 - size_to_image_ratio), seed=random_seed)
    y_min = tf.to_int32(normalized_y_min * tf.to_float(image_height))
    x_min = tf.to_int32(normalized_x_min * tf.to_float(image_width))
    black_box = tf.ones([box_size, box_size, 3], dtype=tf.float32)
    mask = 1.0 - tf.image.pad_to_bounding_box(black_box, y_min, x_min,
                                              image_height, image_width)
    image = tf.multiply(image, mask)
    return image

  with tf.name_scope('RandomBlackPatchInImage', values=[image]):
    for _ in range(max_black_patches):
      random_prob = tf.random_uniform([], minval=0.0, maxval=1.0,
                                      dtype=tf.float32, seed=random_seed)
      image = tf.cond(
          tf.greater(random_prob, probability), lambda: image,
          lambda: add_black_patch_to_image(image))

    return image


def image_to_float(image):
  """Used in Faster R-CNN. Casts image pixel values to float.

  Args:
    image: input image which might be in tf.uint8 or sth else format

  Returns:
    image: image in tf.float32 format.
  """
  with tf.name_scope('ImageToFloat', values=[image]):
    image = tf.to_float(image)
    return image


def random_resize_method(image, target_size):
  """Uses a random resize method to resize the image to target size.

  Args:
    image: a rank 3 tensor.
    target_size: a list of [target_height, target_width]

  Returns:
    resized image.
  """

  resized_image = _apply_with_random_selector(
      image,
      lambda x, method: tf.image.resize_images(x, target_size, method),
      num_cases=4)

  return resized_image


def _compute_new_static_size(image,
                             min_dimension,
                             max_dimension):
  """Compute new static shape for resize_to_range method."""
  image_shape = image.get_shape().as_list()
  orig_height = image_shape[0]
  orig_width = image_shape[1]
  orig_min_dim = min(orig_height, orig_width)
  # Calculates the larger of the possible sizes
  large_scale_factor = min_dimension / float(orig_min_dim)
  # Scaling orig_(height|width) by large_scale_factor will make the smaller
  # dimension equal to min_dimension, save for floating point rounding errors.
  # For reasonably-sized images, taking the nearest integer will reliably
  # eliminate this error.
  large_height = int(round(orig_height * large_scale_factor))
  large_width = int(round(orig_width * large_scale_factor))
  large_size = [large_height, large_width]
  if max_dimension:
    # Calculates the smaller of the possible sizes, use that if the larger
    # is too big.
    orig_max_dim = max(orig_height, orig_width)
    small_scale_factor = max_dimension / float(orig_max_dim)
    # Scaling orig_(height|width) by small_scale_factor will make the larger
    # dimension equal to max_dimension, save for floating point rounding
    # errors. For reasonably-sized images, taking the nearest integer will
    # reliably eliminate this error.
    small_height = int(round(orig_height * small_scale_factor))
    small_width = int(round(orig_width * small_scale_factor))
    small_size = [small_height, small_width]
    new_size = large_size
    if max(large_size) > max_dimension:
      new_size = small_size
  else:
    new_size = large_size
  return tf.constant(new_size)


def _compute_new_dynamic_size(image,
                              min_dimension,
                              max_dimension):
  """Compute new dynamic shape for resize_to_range method."""
  image_shape = tf.shape(image)
  orig_height = tf.to_float(image_shape[0])
  orig_width = tf.to_float(image_shape[1])
  orig_min_dim = tf.minimum(orig_height, orig_width)
  # Calculates the larger of the possible sizes
  min_dimension = tf.constant(min_dimension, dtype=tf.float32)
  large_scale_factor = min_dimension / orig_min_dim
  # Scaling orig_(height|width) by large_scale_factor will make the smaller
  # dimension equal to min_dimension, save for floating point rounding errors.
  # For reasonably-sized images, taking the nearest integer will reliably
  # eliminate this error.
  large_height = tf.to_int32(tf.round(orig_height * large_scale_factor))
  large_width = tf.to_int32(tf.round(orig_width * large_scale_factor))
  large_size = tf.stack([large_height, large_width])
  if max_dimension:
    # Calculates the smaller of the possible sizes, use that if the larger
    # is too big.
    orig_max_dim = tf.maximum(orig_height, orig_width)
    max_dimension = tf.constant(max_dimension, dtype=tf.float32)
    small_scale_factor = max_dimension / orig_max_dim
    # Scaling orig_(height|width) by small_scale_factor will make the larger
    # dimension equal to max_dimension, save for floating point rounding
    # errors. For reasonably-sized images, taking the nearest integer will
    # reliably eliminate this error.
    small_height = tf.to_int32(tf.round(orig_height * small_scale_factor))
    small_width = tf.to_int32(tf.round(orig_width * small_scale_factor))
    small_size = tf.stack([small_height, small_width])
    new_size = tf.cond(
        tf.to_float(tf.reduce_max(large_size)) > max_dimension,
        lambda: small_size, lambda: large_size)
  else:
    new_size = large_size
  return new_size


def resize_to_range(image,
                    masks=None,
                    min_dimension=None,
                    max_dimension=None,
                    align_corners=False):
  """Resizes an image so its dimensions are within the provided value.

  The output size can be described by two cases:
  1. If the image can be rescaled so its minimum dimension is equal to the
     provided value without the other dimension exceeding max_dimension,
     then do so.
  2. Otherwise, resize so the largest dimension is equal to max_dimension.

  Args:
    image: A 3D tensor of shape [height, width, channels]
    masks: (optional) rank 3 float32 tensor with shape
           [num_instances, height, width] containing instance masks.
    min_dimension: (optional) (scalar) desired size of the smaller image
                   dimension.
    max_dimension: (optional) (scalar) maximum allowed size
                   of the larger image dimension.
    align_corners: bool. If true, exactly align all 4 corners of the input
                   and output. Defaults to False.

  Returns:
    A 3D tensor of shape [new_height, new_width, channels],
    where the image has been resized (with bilinear interpolation) so that
    min(new_height, new_width) == min_dimension or
    max(new_height, new_width) == max_dimension.

    If masks is not None, also outputs masks:
    A 3D tensor of shape [num_instances, new_height, new_width]

  Raises:
    ValueError: if the image is not a 3D tensor.
  """
  if len(image.get_shape()) != 3:
    raise ValueError('Image should be 3D tensor')

  with tf.name_scope('ResizeToRange', values=[image, min_dimension]):
    if image.get_shape().is_fully_defined():
      new_size = _compute_new_static_size(image, min_dimension,
                                          max_dimension)
    else:
      new_size = _compute_new_dynamic_size(image, min_dimension,
                                           max_dimension)
    new_image = tf.image.resize_images(image, new_size,
                                       align_corners=align_corners)

    result = new_image
    if masks is not None:
      new_masks = tf.expand_dims(masks, 3)
      new_masks = tf.image.resize_nearest_neighbor(new_masks, new_size,
                                                   align_corners=align_corners)
      new_masks = tf.squeeze(new_masks, 3)
      result = [new_image, new_masks]

    return result


def scale_boxes_to_pixel_coordinates(image, boxes, keypoints=None):
  """Scales boxes from normalized to pixel coordinates.

  Args:
    image: A 3D float32 tensor of shape [height, width, channels].
    boxes: A 2D float32 tensor of shape [num_boxes, 4] containing the bounding
      boxes in normalized coordinates. Each row is of the form
      [ymin, xmin, ymax, xmax].
    keypoints: (optional) rank 3 float32 tensor with shape
      [num_instances, num_keypoints, 2]. The keypoints are in y-x normalized
      coordinates.

  Returns:
    image: unchanged input image.
    scaled_boxes: a 2D float32 tensor of shape [num_boxes, 4] containing the
      bounding boxes in pixel coordinates.
    scaled_keypoints: a 3D float32 tensor with shape
      [num_instances, num_keypoints, 2] containing the keypoints in pixel
      coordinates.
  """
  boxlist = box_list.BoxList(boxes)
  image_height = tf.shape(image)[0]
  image_width = tf.shape(image)[1]
  scaled_boxes = box_list_ops.scale(boxlist, image_height, image_width).get()
  result = [image, scaled_boxes]
  if keypoints is not None:
    scaled_keypoints = keypoint_ops.scale(keypoints, image_height, image_width)
    result.append(scaled_keypoints)
  return tuple(result)


# pylint: disable=g-doc-return-or-yield
def resize_image(image,
                 masks=None,
                 new_height=600,
                 new_width=1024,
                 method=tf.image.ResizeMethod.BILINEAR,
                 align_corners=False):
  """See `tf.image.resize_images` for detailed doc."""
  with tf.name_scope(
      'ResizeImage',
      values=[image, new_height, new_width, method, align_corners]):
    new_image = tf.image.resize_images(image, [new_height, new_width],
                                       method=method,
                                       align_corners=align_corners)
    result = new_image
    if masks is not None:
      num_instances = tf.shape(masks)[0]
      new_size = tf.constant([new_height, new_width], dtype=tf.int32)
      def resize_masks_branch():
        new_masks = tf.expand_dims(masks, 3)
        new_masks = tf.image.resize_nearest_neighbor(
            new_masks, new_size, align_corners=align_corners)
        new_masks = tf.squeeze(new_masks, axis=3)
        return new_masks

      def reshape_masks_branch():
        new_masks = tf.reshape(masks, [0, new_size[0], new_size[1]])
        return new_masks

      masks = tf.cond(num_instances > 0,
                      resize_masks_branch,
                      reshape_masks_branch)
      result = [new_image, masks]

    return result


def subtract_channel_mean(image, means=None):
  """Normalizes an image by subtracting a mean from each channel.

  Args:
    image: A 3D tensor of shape [height, width, channels]
    means: float list containing a mean for each channel
  Returns:
    normalized_images: a tensor of shape [height, width, channels]
  Raises:
    ValueError: if images is not a 4D tensor or if the number of means is not
      equal to the number of channels.
  """
  with tf.name_scope('SubtractChannelMean', values=[image, means]):
    if len(image.get_shape()) != 3:
      raise ValueError('Input must be of size [height, width, channels]')
    if len(means) != image.get_shape()[-1]:
      raise ValueError('len(means) must match the number of channels')
    return image - [[means]]


def one_hot_encoding(labels, num_classes=None):
  """One-hot encodes the multiclass labels.

  Example usage:
    labels = tf.constant([1, 4], dtype=tf.int32)
    one_hot = OneHotEncoding(labels, num_classes=5)
    one_hot.eval()    # evaluates to [0, 1, 0, 0, 1]

  Args:
    labels: A tensor of shape [None] corresponding to the labels.
    num_classes: Number of classes in the dataset.
  Returns:
    onehot_labels: a tensor of shape [num_classes] corresponding to the one hot
      encoding of the labels.
  Raises:
    ValueError: if num_classes is not specified.
  """
  with tf.name_scope('OneHotEncoding', values=[labels]):
    if num_classes is None:
      raise ValueError('num_classes must be specified')

    labels = tf.one_hot(labels, num_classes, 1, 0)
    return tf.reduce_max(labels, 0)


def rgb_to_gray(image):
  """Converts a 3 channel RGB image to a 1 channel grayscale image.

  Args:
    image: Rank 3 float32 tensor containing 1 image -> [height, width, 3]
           with pixel values varying between [0, 1].

  Returns:
    image: A single channel grayscale image -> [image, height, 1].
  """
  return tf.image.rgb_to_grayscale(image)


def ssd_random_crop(image,
                    boxes,
                    labels,
                    masks=None,
                    keypoints=None,
                    min_object_covered=(0.0, 0.1, 0.3, 0.5, 0.7, 0.9, 1.0),
                    aspect_ratio_range=((0.5, 2.0),) * 7,
                    area_range=((0.1, 1.0),) * 7,
                    overlap_thresh=(0.0, 0.1, 0.3, 0.5, 0.7, 0.9, 1.0),
                    random_coef=(0.15,) * 7,
                    seed=None):
  """Random crop preprocessing with default parameters as in SSD paper.

  Liu et al., SSD: Single shot multibox detector.
  For further information on random crop preprocessing refer to RandomCrop
  function above.

  Args:
    image: rank 3 float32 tensor contains 1 image -> [height, width, channels]
           with pixel values varying between [0, 1].
    boxes: rank 2 float32 tensor containing the bounding boxes -> [N, 4].
           Boxes are in normalized form meaning their coordinates vary
           between [0, 1].
           Each row is in the form of [ymin, xmin, ymax, xmax].
    labels: rank 1 int32 tensor containing the object classes.
    masks: (optional) rank 3 float32 tensor with shape
           [num_instances, height, width] containing instance masks. The masks
           are of the same height, width as the input `image`.
    keypoints: (optional) rank 3 float32 tensor with shape
               [num_instances, num_keypoints, 2]. The keypoints are in y-x
               normalized coordinates.
    min_object_covered: the cropped image must cover at least this fraction of
                        at least one of the input bounding boxes.
    aspect_ratio_range: allowed range for aspect ratio of cropped image.
    area_range: allowed range for area ratio between cropped image and the
                original image.
    overlap_thresh: minimum overlap thresh with new cropped
                    image to keep the box.
    random_coef: a random coefficient that defines the chance of getting the
                 original image. If random_coef is 0, we will always get the
                 cropped image, and if it is 1.0, we will always get the
                 original image.
    seed: random seed.

  Returns:
    image: image which is the same rank as input image.
    boxes: boxes which is the same rank as input boxes.
           Boxes are in normalized form.
    labels: new labels.

    If masks, or keypoints is not None, the function also returns:

    masks: rank 3 float32 tensor with shape [num_instances, height, width]
           containing instance masks.
    keypoints: rank 3 float32 tensor with shape
               [num_instances, num_keypoints, 2]
  """
  def random_crop_selector(selected_result, index):
    """Applies random_crop_image to selected result.

    Args:
      selected_result: A tuple containing image, boxes, labels, keypoints (if
                       not None), and masks (if not None).
      index: The index that was randomly selected.

    Returns: A tuple containing image, boxes, labels, keypoints (if not None),
             and masks (if not None).
    """
    i = 3
    image, boxes, labels = selected_result[:i]
    selected_masks = None
    selected_keypoints = None
    if masks is not None:
      selected_masks = selected_result[i]
      i += 1
    if keypoints is not None:
      selected_keypoints = selected_result[i]

    return random_crop_image(
        image=image,
        boxes=boxes,
        labels=labels,
        masks=selected_masks,
        keypoints=selected_keypoints,
        min_object_covered=min_object_covered[index],
        aspect_ratio_range=aspect_ratio_range[index],
        area_range=area_range[index],
        overlap_thresh=overlap_thresh[index],
        random_coef=random_coef[index],
        seed=seed)

  result = _apply_with_random_selector_tuples(
      tuple(
          t for t in (image, boxes, labels, masks, keypoints) if t is not None),
      random_crop_selector,
      num_cases=len(min_object_covered))
  return result


def ssd_random_crop_pad(image,
                        boxes,
                        labels,
                        min_object_covered=(0.1, 0.3, 0.5, 0.7, 0.9, 1.0),
                        aspect_ratio_range=((0.5, 2.0),) * 6,
                        area_range=((0.1, 1.0),) * 6,
                        overlap_thresh=(0.1, 0.3, 0.5, 0.7, 0.9, 1.0),
                        random_coef=(0.15,) * 6,
                        min_padded_size_ratio=(None,) * 6,
                        max_padded_size_ratio=(None,) * 6,
                        pad_color=(None,) * 6,
                        seed=None):
  """Random crop preprocessing with default parameters as in SSD paper.

  Liu et al., SSD: Single shot multibox detector.
  For further information on random crop preprocessing refer to RandomCrop
  function above.

  Args:
    image: rank 3 float32 tensor containing 1 image -> [height, width, channels]
           with pixel values varying between [0, 1].
    boxes: rank 2 float32 tensor containing the bounding boxes -> [N, 4].
           Boxes are in normalized form meaning their coordinates vary
           between [0, 1].
           Each row is in the form of [ymin, xmin, ymax, xmax].
    labels: rank 1 int32 tensor containing the object classes.
    min_object_covered: the cropped image must cover at least this fraction of
                        at least one of the input bounding boxes.
    aspect_ratio_range: allowed range for aspect ratio of cropped image.
    area_range: allowed range for area ratio between cropped image and the
                original image.
    overlap_thresh: minimum overlap thresh with new cropped
                    image to keep the box.
    random_coef: a random coefficient that defines the chance of getting the
                 original image. If random_coef is 0, we will always get the
                 cropped image, and if it is 1.0, we will always get the
                 original image.
    min_padded_size_ratio: min ratio of padded image height and width to the
                           input image's height and width. If None, it will
                           be set to [0.0, 0.0].
    max_padded_size_ratio: max ratio of padded image height and width to the
                           input image's height and width. If None, it will
                           be set to [2.0, 2.0].
    pad_color: padding color. A rank 1 tensor of [3] with dtype=tf.float32.
               if set as None, it will be set to average color of the randomly
               cropped image.
    seed: random seed.

  Returns:
    image: Image shape will be [new_height, new_width, channels].
    boxes: boxes which is the same rank as input boxes. Boxes are in normalized
           form.
    new_labels: new labels.
  """
  def random_crop_pad_selector(image_boxes_labels, index):
    image, boxes, labels = image_boxes_labels

    return random_crop_pad_image(
        image,
        boxes,
        labels,
        min_object_covered=min_object_covered[index],
        aspect_ratio_range=aspect_ratio_range[index],
        area_range=area_range[index],
        overlap_thresh=overlap_thresh[index],
        random_coef=random_coef[index],
        min_padded_size_ratio=min_padded_size_ratio[index],
        max_padded_size_ratio=max_padded_size_ratio[index],
        pad_color=pad_color[index],
        seed=seed)

  new_image, new_boxes, new_labels = _apply_with_random_selector_tuples(
      (image, boxes, labels),
      random_crop_pad_selector,
      num_cases=len(min_object_covered))
  return new_image, new_boxes, new_labels


def ssd_random_crop_fixed_aspect_ratio(
    image,
    boxes,
    labels,
    masks=None,
    keypoints=None,
    min_object_covered=(0.0, 0.1, 0.3, 0.5, 0.7, 0.9, 1.0),
    aspect_ratio=1.0,
    area_range=((0.1, 1.0),) * 7,
    overlap_thresh=(0.0, 0.1, 0.3, 0.5, 0.7, 0.9, 1.0),
    random_coef=(0.15,) * 7,
    seed=None):
  """Random crop preprocessing with default parameters as in SSD paper.

  Liu et al., SSD: Single shot multibox detector.
  For further information on random crop preprocessing refer to RandomCrop
  function above.

  The only difference is that the aspect ratio of the crops are fixed.

  Args:
    image: rank 3 float32 tensor contains 1 image -> [height, width, channels]
           with pixel values varying between [0, 1].
    boxes: rank 2 float32 tensor containing the bounding boxes -> [N, 4].
           Boxes are in normalized form meaning their coordinates vary
           between [0, 1].
           Each row is in the form of [ymin, xmin, ymax, xmax].
    labels: rank 1 int32 tensor containing the object classes.
    masks: (optional) rank 3 float32 tensor with shape
           [num_instances, height, width] containing instance masks. The masks
           are of the same height, width as the input `image`.
    keypoints: (optional) rank 3 float32 tensor with shape
               [num_instances, num_keypoints, 2]. The keypoints are in y-x
               normalized coordinates.
    min_object_covered: the cropped image must cover at least this fraction of
                        at least one of the input bounding boxes.
    aspect_ratio: aspect ratio of the cropped image.
    area_range: allowed range for area ratio between cropped image and the
                original image.
    overlap_thresh: minimum overlap thresh with new cropped
                    image to keep the box.
    random_coef: a random coefficient that defines the chance of getting the
                 original image. If random_coef is 0, we will always get the
                 cropped image, and if it is 1.0, we will always get the
                 original image.
    seed: random seed.

  Returns:
    image: image which is the same rank as input image.
    boxes: boxes which is the same rank as input boxes.
           Boxes are in normalized form.
    labels: new labels.

    If masks, or keypoints is not None, the function also returns:

    masks: rank 3 float32 tensor with shape [num_instances, height, width]
           containing instance masks.
    keypoints: rank 3 float32 tensor with shape
               [num_instances, num_keypoints, 2]

  """
  aspect_ratio_range = ((aspect_ratio, aspect_ratio),) * len(area_range)

  crop_result = ssd_random_crop(image, boxes, labels, masks, keypoints,
                                min_object_covered, aspect_ratio_range,
                                area_range, overlap_thresh, random_coef, seed)
  i = 3
  new_image, new_boxes, new_labels = crop_result[:i]
  new_masks = None
  new_keypoints = None
  if masks is not None:
    new_masks = crop_result[i]
    i += 1
  if keypoints is not None:
    new_keypoints = crop_result[i]
  result = random_crop_to_aspect_ratio(
      new_image,
      new_boxes,
      new_labels,
      new_masks,
      new_keypoints,
      aspect_ratio=aspect_ratio,
      seed=seed)

  return result


def get_default_func_arg_map(include_instance_masks=False,
                             include_keypoints=False):
  """Returns the default mapping from a preprocessor function to its args.

  Args:
    include_instance_masks: If True, preprocessing functions will modify the
      instance masks, too.
    include_keypoints: If True, preprocessing functions will modify the
      keypoints, too.

  Returns:
    A map from preprocessing functions to the arguments they receive.
  """
  groundtruth_instance_masks = None
  if include_instance_masks:
    groundtruth_instance_masks = (
        fields.InputDataFields.groundtruth_instance_masks)

  groundtruth_keypoints = None
  if include_keypoints:
    groundtruth_keypoints = fields.InputDataFields.groundtruth_keypoints

  prep_func_arg_map = {
      normalize_image: (fields.InputDataFields.image,),
      random_horizontal_flip: (fields.InputDataFields.image,
                               fields.InputDataFields.groundtruth_boxes,
                               groundtruth_instance_masks,
                               groundtruth_keypoints,),
      random_pixel_value_scale: (fields.InputDataFields.image,),
      random_image_scale: (fields.InputDataFields.image,
                           groundtruth_instance_masks,),
      random_rgb_to_gray: (fields.InputDataFields.image,),
      random_adjust_brightness: (fields.InputDataFields.image,),
      random_adjust_contrast: (fields.InputDataFields.image,),
      random_adjust_hue: (fields.InputDataFields.image,),
      random_adjust_saturation: (fields.InputDataFields.image,),
      random_distort_color: (fields.InputDataFields.image,),
      random_jitter_boxes: (fields.InputDataFields.groundtruth_boxes,),
      random_crop_image: (fields.InputDataFields.image,
                          fields.InputDataFields.groundtruth_boxes,
                          fields.InputDataFields.groundtruth_classes,
                          groundtruth_instance_masks,
                          groundtruth_keypoints,),
      random_pad_image: (fields.InputDataFields.image,
                         fields.InputDataFields.groundtruth_boxes),
      random_crop_pad_image: (fields.InputDataFields.image,
                              fields.InputDataFields.groundtruth_boxes,
                              fields.InputDataFields.groundtruth_classes),
      random_crop_to_aspect_ratio: (fields.InputDataFields.image,
                                    fields.InputDataFields.groundtruth_boxes,
                                    fields.InputDataFields.groundtruth_classes,
                                    groundtruth_instance_masks,
                                    groundtruth_keypoints,),
      random_black_patches: (fields.InputDataFields.image,),
      retain_boxes_above_threshold: (
          fields.InputDataFields.groundtruth_boxes,
          fields.InputDataFields.groundtruth_classes,
          fields.InputDataFields.groundtruth_label_scores,
          groundtruth_instance_masks,
          groundtruth_keypoints,),
      image_to_float: (fields.InputDataFields.image,),
      random_resize_method: (fields.InputDataFields.image,),
      resize_to_range: (fields.InputDataFields.image,
                        groundtruth_instance_masks,),
      scale_boxes_to_pixel_coordinates: (
          fields.InputDataFields.image,
          fields.InputDataFields.groundtruth_boxes,
          groundtruth_keypoints,),
      flip_boxes: (fields.InputDataFields.groundtruth_boxes,),
      resize_image: (fields.InputDataFields.image,
                     groundtruth_instance_masks,),
      subtract_channel_mean: (fields.InputDataFields.image,),
      one_hot_encoding: (fields.InputDataFields.groundtruth_image_classes,),
      rgb_to_gray: (fields.InputDataFields.image,),
      ssd_random_crop: (fields.InputDataFields.image,
                        fields.InputDataFields.groundtruth_boxes,
                        fields.InputDataFields.groundtruth_classes,
                        groundtruth_instance_masks,
                        groundtruth_keypoints,),
      ssd_random_crop_pad: (fields.InputDataFields.image,
                            fields.InputDataFields.groundtruth_boxes,
                            fields.InputDataFields.groundtruth_classes),
      ssd_random_crop_fixed_aspect_ratio: (
          fields.InputDataFields.image,
          fields.InputDataFields.groundtruth_boxes,
          fields.InputDataFields.groundtruth_classes,
          groundtruth_instance_masks,
          groundtruth_keypoints,),
  }

  return prep_func_arg_map


def preprocess(tensor_dict, preprocess_options, func_arg_map=None):
  """Preprocess images and bounding boxes.

  Various types of preprocessing (to be implemented) based on the
  preprocess_options dictionary e.g. "crop image" (affects image and possibly
  boxes), "white balance image" (affects only image), etc. If self._options
  is None, no preprocessing is done.

  Args:
    tensor_dict: dictionary that contains images, boxes, and can contain other
                 things as well.
                 images-> rank 4 float32 tensor contains
                          1 image -> [1, height, width, 3].
                          with pixel values varying between [0, 1]
                 boxes-> rank 2 float32 tensor containing
                         the bounding boxes -> [N, 4].
                         Boxes are in normalized form meaning
                         their coordinates vary between [0, 1].
                         Each row is in the form
                         of [ymin, xmin, ymax, xmax].
    preprocess_options: It is a list of tuples, where each tuple contains a
                        function and a dictionary that contains arguments and
                        their values.
    func_arg_map: mapping from preprocessing functions to arguments that they
                  expect to receive and return.

  Returns:
    tensor_dict: which contains the preprocessed images, bounding boxes, etc.

  Raises:
    ValueError: (a) If the functions passed to Preprocess
                    are not in func_arg_map.
                (b) If the arguments that a function needs
                    do not exist in tensor_dict.
                (c) If image in tensor_dict is not rank 4
  """
  if func_arg_map is None:
    func_arg_map = get_default_func_arg_map()

  # changes the images to image (rank 4 to rank 3) since the functions
  # receive rank 3 tensor for image
  if fields.InputDataFields.image in tensor_dict:
    images = tensor_dict[fields.InputDataFields.image]
    if len(images.get_shape()) != 4:
      raise ValueError('images in tensor_dict should be rank 4')
    image = tf.squeeze(images, squeeze_dims=[0])
    tensor_dict[fields.InputDataFields.image] = image

  # Preprocess inputs based on preprocess_options
  for option in preprocess_options:
    func, params = option
    if func not in func_arg_map:
      raise ValueError('The function %s does not exist in func_arg_map' %
                       (func.__name__))
    arg_names = func_arg_map[func]
    for a in arg_names:
      if a is not None and a not in tensor_dict:
        raise ValueError('The function %s requires argument %s' %
                         (func.__name__, a))

    def get_arg(key):
      return tensor_dict[key] if key is not None else None
    args = [get_arg(a) for a in arg_names]
    results = func(*args, **params)
    if not isinstance(results, (list, tuple)):
      results = (results,)
    # Removes None args since the return values will not contain those.
    arg_names = [arg_name for arg_name in arg_names if arg_name is not None]
    for res, arg_name in zip(results, arg_names):
      tensor_dict[arg_name] = res

  # changes the image to images (rank 3 to rank 4) to be compatible to what
  # we received in the first place
  if fields.InputDataFields.image in tensor_dict:
    image = tensor_dict[fields.InputDataFields.image]
    images = tf.expand_dims(image, 0)
    tensor_dict[fields.InputDataFields.image] = images

  return tensor_dict