Dimension of the output tensors

Course Q&A Deep Learning Specialization Convolutional Neural Networks
1

Hi,

there is something unclear to me in the 3rd exercise docstring. This is the 1st assignment of the 3rd week. It is written that

Note: The “None” dimension of the output tensors has obviously to be less than max_boxes.

But, I think it should be less than or equal. For example, there could be 2 sets of boxes, and each set 3 boxes. Let’s suppose the sets do not overlap but the boxes in the set do (see the image below). Those boxes could overlap with iou >= iou_threshold, hence the output would be 2 boxes (nout_box = 2). But the max_boxes could be set to 1, 2, or 3 because it is not known in advance how the boxes overlap.
If max_boxes = 2 than nout_box = max_boxes and if max_boxes = 3 than nout_box < max_boxes. If max_boxes = 1, we would have
nout_box = max_boxes. I guess in the latter scenario only the box with the biggest score would be left.

image
image694×346 437 Bytes

Am I right?

Also, why do we need a tf.Variable for max_boxes_tensor? In the documentation for tf.image.non_max_suppression() the requirement is a tensor. I tested with tf.constant(max_boxes, dtype=‘int32’) and I got “All tests passed!”. Is max_boxes_tensor used as a training parameter later?

Henrikh

3

I think your guess is right.

my guess is that less than or equal is probably correct.

At the bottom of API description, we can see,

image
image1790×257 21.8 KB

4

Thank you for your clarification!