为什么使用tf.image.sample_distorted_bounding_box边界框和裁剪的图片不匹配？

Question

I want get a bounding box and clipped picture according to the bounding box , so I used tf.image.sample_distorted_bounding_box . But I failed, What I did wrong? My result looks like

The bounding box and the clipped picture according to the bounding box does not match.

My code:

with tf.Session() as sess:         
    boxes = tf.constant([[[0.05, 0.05, 0.9, 0.7], [0.35, 0.47, 0.5, 0.56]]])

    image_float = tf.image.convert_image_dtype(img_data, tf.float32) # uint8 -> float

    # resize image
    image_small = tf.image.resize_images(image_float, [180, 267], method=0)

    # Generate a single distorted bounding box.
    begin, size, bbox_for_draw = tf.image.sample_distorted_bounding_box(
        tf.shape(image_small),
        bounding_boxes=boxes,
        min_object_covered=0.1)

    # Draw the bounding box in an image summary.
    image_with_box = tf.image.draw_bounding_boxes(tf.expand_dims(image_small, 0),
                                                  bbox_for_draw)

    tf.summary.image('images_with_box', image_with_box)

    # Employ the bounding box to distort the image.
    distorted_image = tf.slice(image_small, begin, size)

    plt.figure(figsize = (30, 20))
    plt.subplot(3, 1, 1)
    plt.title("image with a random box")
    plt.imshow(image_with_box[0].eval())

    plt.subplot(3, 1, 2)
    plt.title("destorted image")
    plt.imshow(distorted_image.eval())
plt.show()

Answer 1

That is because each call to .eval() triggers a new run of the graph and therefore produces a new random bounding box.

To have consistent outputs you need to run operators simultaneously, eg

res = sess.run([image_with_box[0], distorted_image])