def _process_mask(mask, ignore_label, image_info):
mask = tf.cast(mask, dtype=tf.float32)
mask = tf.reshape(mask, shape=[1, data['height'], data['width'], 1])
mask += 1
if self._segmentation_resize_eval_groundtruth:
# Resizes eval masks to match input image sizes. In that case, mean IoU
# is computed on output_size not the original size of the images.
image_scale = image_info[2, :]
offset = image_info[3, :]
mask = preprocess_ops.resize_and_crop_masks(
mask, image_scale, self._output_size, offset)
else:
mask = tf.image.pad_to_bounding_box(
mask, 0, 0,
self._segmentation_groundtruth_padded_size[0],
self._segmentation_groundtruth_padded_size[1])
mask -= 1
# Assign ignore label to the padded region.
mask = tf.where(
tf.equal(mask, -1),
ignore_label * tf.ones_like(mask),
mask)
mask = tf.squeeze(mask, axis=0)
return mask
def _parse_eval_data(self, data):
"""Parses data for training and evaluation."""
image, label = self._prepare_image_and_label(data)
# The label is first offset by +1 and then padded with 0.
label += 1
label = tf.expand_dims(label, axis=3)
# Resizes and crops image.
image, image_info = preprocess_ops.resize_and_crop_image(
image, self._output_size, self._output_size)
if self._resize_eval_groundtruth:
# Resizes eval masks to match input image sizes. In that case, mean IoU
# is computed on output_size not the original size of the images.
image_scale = image_info[2, :]
offset = image_info[3, :]
label = preprocess_ops.resize_and_crop_masks(
label, image_scale, self._output_size, offset)
else:
label = tf.image.pad_to_bounding_box(
label, 0, 0, self._groundtruth_padded_size[0],
self._groundtruth_padded_size[1])
label -= 1
label = tf.where(tf.equal(label, -1),
self._ignore_label * tf.ones_like(label), label)
label = tf.squeeze(label, axis=0)
valid_mask = tf.not_equal(label, self._ignore_label)
labels = {
'masks': label,
'valid_masks': valid_mask,
'image_info': image_info
}
# Cast image as self._dtype
image = tf.cast(image, dtype=self._dtype)
return image, labels
def _resize_and_crop_mask(self, mask, image_info, is_training):
"""Resizes and crops mask using `image_info` dict."""
height = image_info[0][0]
width = image_info[0][1]
mask = tf.reshape(mask, shape=[1, height, width, 1])
mask += 1
if is_training or self._resize_eval_groundtruth:
image_scale = image_info[2, :]
offset = image_info[3, :]
mask = preprocess_ops.resize_and_crop_masks(
mask, image_scale, self._output_size, offset)
else:
mask = tf.image.pad_to_bounding_box(
mask, 0, 0, self._groundtruth_padded_size[0],
self._groundtruth_padded_size[1])
mask -= 1
# Assign ignore label to the padded region.
mask = tf.where(tf.equal(mask, -1),
self._ignore_label * tf.ones_like(mask), mask)
mask = tf.squeeze(mask, axis=0)
return mask
def _parse_train_data(self, data):
"""Parses data for training.
Args:
data: the decoded tensor dictionary from TfExampleDecoder.
Returns:
image: image tensor that is preproessed to have normalized value and
dimension [output_size[0], output_size[1], 3]
labels: a dictionary of tensors used for training. The following describes
{key: value} pairs in the dictionary.
image_info: a 2D `Tensor` that encodes the information of the image and
the applied preprocessing. It is in the format of
[[original_height, original_width], [scaled_height, scaled_width]],
anchor_boxes: ordered dictionary with keys
[min_level, min_level+1, ..., max_level]. The values are tensor with
shape [height_l, width_l, 4] representing anchor boxes at each level.
rpn_score_targets: ordered dictionary with keys
[min_level, min_level+1, ..., max_level]. The values are tensor with
shape [height_l, width_l, anchors_per_location]. The height_l and
width_l represent the dimension of class logits at l-th level.
rpn_box_targets: ordered dictionary with keys
[min_level, min_level+1, ..., max_level]. The values are tensor with
shape [height_l, width_l, anchors_per_location * 4]. The height_l and
width_l represent the dimension of bounding box regression output at
l-th level.
gt_boxes: Groundtruth bounding box annotations. The box is represented
in [y1, x1, y2, x2] format. The coordinates are w.r.t the scaled
image that is fed to the network. The tennsor is padded with -1 to
the fixed dimension [self._max_num_instances, 4].
gt_classes: Groundtruth classes annotations. The tennsor is padded
with -1 to the fixed dimension [self._max_num_instances].
gt_masks: Groundtruth masks cropped by the bounding box and
resized to a fixed size determined by mask_crop_size.
gt_segmentation_mask: Groundtruth mask for segmentation head, this is
resized to a fixed size determined by output_size.
gt_segmentation_valid_mask: Binary mask that marks the pixels that
are supposed to be used in computing the segmentation loss while
training.
"""
segmentation_mask = data['groundtruth_segmentation_mask']
# Flips image randomly during training.
if self.aug_rand_hflip:
masks = data['groundtruth_instance_masks']
image_mask = tf.concat([data['image'], segmentation_mask], axis=2)
image_mask, boxes, masks = preprocess_ops.random_horizontal_flip(
image_mask, data['groundtruth_boxes'], masks)
segmentation_mask = image_mask[:, :, -1:]
image = image_mask[:, :, :-1]
data['image'] = image
data['groundtruth_boxes'] = boxes
data['groundtruth_instance_masks'] = masks
image, labels = super(Parser, self)._parse_train_data(data)
image_info = labels['image_info']
image_scale = image_info[2, :]
offset = image_info[3, :]
segmentation_mask = tf.reshape(
segmentation_mask, shape=[1, data['height'], data['width']])
segmentation_mask = tf.cast(segmentation_mask, tf.float32)
# Pad label and make sure the padded region assigned to the ignore label.
# The label is first offset by +1 and then padded with 0.
segmentation_mask += 1
segmentation_mask = tf.expand_dims(segmentation_mask, axis=3)
segmentation_mask = preprocess_ops.resize_and_crop_masks(
segmentation_mask, image_scale, self._output_size, offset)
segmentation_mask -= 1
segmentation_mask = tf.where(
tf.equal(segmentation_mask, -1),
self._segmentation_ignore_label * tf.ones_like(segmentation_mask),
segmentation_mask)
segmentation_mask = tf.squeeze(segmentation_mask, axis=0)
segmentation_valid_mask = tf.not_equal(
segmentation_mask, self._segmentation_ignore_label)
labels.update({
'gt_segmentation_mask': segmentation_mask,
'gt_segmentation_valid_mask': segmentation_valid_mask})
return image, labels
def _parse_train_data(self, data):
"""Parses data for training and evaluation."""
image, label = self._prepare_image_and_label(data)
if self._crop_size:
label = tf.reshape(label,
[data['image/height'], data['image/width'], 1])
# If output_size is specified, resize image, and label to desired
# output_size.
if self._output_size:
image = tf.image.resize(image,
self._output_size,
method='bilinear')
label = tf.image.resize(label,
self._output_size,
method='nearest')
image_mask = tf.concat([image, label], axis=2)
image_mask_crop = tf.image.random_crop(image_mask,
self._crop_size + [4])
image = image_mask_crop[:, :, :-1]
label = tf.reshape(image_mask_crop[:, :, -1],
[1] + self._crop_size)
# Flips image randomly during training.
if self._aug_rand_hflip:
image, _, label = preprocess_ops.random_horizontal_flip(
image, masks=label)
train_image_size = self._crop_size if self._crop_size else self._output_size
# Resizes and crops image.
image, image_info = preprocess_ops.resize_and_crop_image(
image,
train_image_size,
train_image_size,
aug_scale_min=self._aug_scale_min,
aug_scale_max=self._aug_scale_max)
# Resizes and crops boxes.
image_scale = image_info[2, :]
offset = image_info[3, :]
# Pad label and make sure the padded region assigned to the ignore label.
# The label is first offset by +1 and then padded with 0.
label += 1
label = tf.expand_dims(label, axis=3)
label = preprocess_ops.resize_and_crop_masks(label, image_scale,
train_image_size, offset)
label -= 1
label = tf.where(tf.equal(label, -1),
self._ignore_label * tf.ones_like(label), label)
label = tf.squeeze(label, axis=0)
valid_mask = tf.not_equal(label, self._ignore_label)
labels = {
'masks': label,
'valid_masks': valid_mask,
'image_info': image_info,
}
# Cast image as self._dtype
image = tf.cast(image, dtype=self._dtype)
return image, labels