ML Community Day is November 9! Join us for updates from TensorFlow, JAX, and more Learn more


Generate a randomly distorted bounding box for an image deterministically.

Bounding box annotations are often supplied in addition to ground-truth labels in image recognition or object localization tasks. A common technique for training such a system is to randomly distort an image while preserving its content, i.e. data augmentation. This Op, given the same seed, deterministically outputs a randomly distorted localization of an object, i.e. bounding box, given an image_size, bounding_boxes and a series of constraints.

The output of this Op is a single bounding box that may be used to crop the original image. The output is returned as 3 tensors: begin, size and bboxes. The first 2 tensors can be fed directly into tf.slice to crop the image. The latter may be supplied to tf.image.draw_bounding_boxes to visualize what the bounding box looks like.

Bounding boxes are supplied and returned as [y_min, x_min, y_max, x_max]. The bounding box coordinates are floats in [0.0, 1.0] relative to the width and the height of the underlying image.

The output of this Op is guaranteed to be the same given the same seed and is independent of how many times the function is called, and independent of global seed settings (e.g. tf.random.set_seed).

Example usage:

image = np.array([[[1], [2], [3]], [[4], [5], [6]], [[7], [8], [9]]])
bbox = tf.constant(
  [0.0, 0.0, 1.0, 1.0], dtype=tf.float32, shape=[1, 1, 4])
seed = (1, 2)
# Generate a single distorted bounding box.
bbox_begin, bbox_size, bbox_draw = (
    tf.shape(image), bounding_boxes=bbox, seed=seed))
# Employ the bounding box to distort the image.
tf.slice(image, bbox_begin, bbox_size)
<tf.Tensor: shape=(2, 2, 1), dtype=int64, numpy=
# Draw the bounding box in an image summary.
colors = np.array([[1.0, 0.0, 0.0], [0.0, 0.0, 1.0]])
  tf.expand_dims(tf.cast(image, tf.float32),0), bbox_draw, colors)
<tf.Tensor: shape=(1, 3, 3, 1), dtype=float32, numpy=
         [9.]]]], dtype=float32)>

Note that if no bounding box information is available, setting use_image_if_no_bounding_boxes = true will assume there is a single implicit bounding box covering the whole image. If use_image_if_no_bounding_boxes is false and no bounding boxes are supplied, an error is raised.

image_size A Tensor. Must be one of the following types: uint8, int8, int16, int32, int64. 1-D, containing [height, width, channels].
bounding_boxes A Tensor of type float32. 3-D with shape [batch, N, 4] describing the N bounding boxes associated with the image.
min_object_covered A Tensor of type float32. The cropped area of the image mu