*Question/Issue: i have a question that if i use the cut point like cut_point = mobile_net_v2.get_layer(‘block_13_expand_relu’) where H and W are 1/16 from the input, does it mean the center of an object (from labeled data) that going to be trained is also 16x16 or it will stay 8x8 ? thankyou *
Yes that’s right. We use the cut point to decide the architecture, which implies the reduction amount, and then all data, images and bounding boxes, are projected accordingly.
The only constraints we have are
- that output is square, i.e. height == width, though that constraint could additionally be removed with some work.
- the reduction doesn’t result in a height or width <1 at the output. e.g. you had a 1/16 reduction and were feeding a 8x8 input.