def _resize_image(self, image, bboxes=None):
"""
We need to resize image and bounding boxes when the biggest side
dimension is bigger than `self._image_max_size` or when the smaller
side is smaller than `self._image_min_size`.
Then, using the ratio we used, we need to properly scale the bounding
boxes.
Args:
image: Tensor with image of shape (H, W, 3).
bboxes: Tensor with bounding boxes with shape (num_bboxes, 5).
where we have (x_min, y_min, x_max, y_max, label) for each one.
Returns:
image: Tensor with scaled image.
bboxes: Tensor with scaled (using the same factor as the image)
bounding boxes with shape (num_bboxes, 5).
scale_factor: Scale factor used to modify the image (1.0 means no
change).
"""
if self._fixed_resize:
resized = resize_image_fixed(image,
self._image_fixed_height,
self._image_fixed_width,
bboxes=bboxes)
else:
resized = resize_image(
image,
bboxes=bboxes,
min_size=self._image_min_size,
max_size=self._image_max_size,
)
return resized["image"], resized.get("bboxes"), resized["scale_factor"]
def _resize_image(self,
image_array,
boxes_array=None,
min_size=None,
max_size=None):
image = tf.placeholder(tf.float32, image_array.shape)
feed_dict = {
image: image_array,
}
if boxes_array is not None:
boxes = tf.placeholder(tf.float32, boxes_array.shape)
feed_dict[boxes] = boxes_array
else:
boxes = None
resized = resize_image(image,
bboxes=boxes,
min_size=min_size,
max_size=max_size)
with self.test_session() as sess:
resized_dict = sess.run(resized, feed_dict=feed_dict)
return (
resized_dict['image'],
resized_dict.get('bboxes'),
resized_dict.get('scale_factor'),
)
def _resize_image(self, image, bboxes=None):
"""
We need to resize image and bounding boxes when the biggest side
dimension is bigger than `self._image_max_size` or when the smaller
side is smaller than `self._image_min_size`.
Then, using the ratio we used, we need to properly scale the bounding
boxes.
Args:
image: Tensor with image of shape (H, W, 3).
bboxes: Tensor with bounding boxes with shape (num_bboxes, 5).
where we have (x_min, y_min, x_max, y_max, label) for each one.
Returns:
image: Tensor with scaled image.
bboxes: Tensor with scaled (using the same factor as the image)
bounding boxes with shape (num_bboxes, 5).
scale_factor: Scale factor used to modify the image (1.0 means no
change).
"""
resized = resize_image(
image, bboxes=bboxes, min_size=self._image_min_size,
max_size=self._image_max_size
)
return resized['image'], resized.get('bboxes'), resized['scale_factor']
def _resize_image(self, image_array, boxes_array=None, min_size=None,
max_size=None):
image = tf.placeholder(tf.float32, image_array.shape)
feed_dict = {
image: image_array,
}
if boxes_array is not None:
boxes = tf.placeholder(tf.float32, boxes_array.shape)
feed_dict[boxes] = boxes_array
else:
boxes = None
resized = resize_image(
image, bboxes=boxes, min_size=min_size, max_size=max_size
)
with self.test_session() as sess:
resized_dict = sess.run(resized, feed_dict=feed_dict)
return (
resized_dict['image'],
resized_dict.get('bboxes'),
resized_dict.get('scale_factor'),
)
