Non max suppression run time issue (Chapter 05) about pytorch-computer-vision-cookbook HOT 7 OPEN

I fixed this with replacing:
ww = detections[supp_inds, 4]
To:
ww = detections[supp_inds, :4]

from pytorch-computer-vision-cookbook.

Hi Tomlac92,

Thanks for the reply, but i doubt your solution is correct. The idea is to average the remaining bounding boxes position, but taking into account the object score of each box. That is exactly what the code does by taking ww=detections[supp_inds, 4]
You are now taking the sum of all boxes x1^2, y1^2, x2^2, y2^2, and them dividing each of the sums by the sum of all x1, y1, x2, y2 values of all boxes, which seems strange to me.

from pytorch-computer-vision-cookbook.

Hey, thanks for the explanations. I didn't sure what exactly this line doing, so I share my first solution, which doesn't give an error, but I can't judge it is correct or not. Before changing that line I was getting error like "tensor a must be the same shape, like tensor b". Tell me, your solution works? Any idea to fix this NMS algorithm will be really helpfull.

from pytorch-computer-vision-cookbook.

Hi Tomlac92,

yes my code works, keep in mind you need to patch the NMS function xyxyh2xywh as well, see the issue i created for this how to patch this method.

from pytorch-computer-vision-cookbook.

I also had idea, that ww needs to be ww. unsqueezed(0), because error I was getting was informing that tensor ww is one dimension smaller than tensor detections[supp_inds, :4], anyway I have not tested it yet. So many thanks for providing a working solution.

from pytorch-computer-vision-cookbook.

@mravendi Can you have a look at this issue?

from pytorch-computer-vision-cookbook.

Thanks @wvalcke. Your fix makes sense since you are basically converting 'ww' into a 2d array using '.view' method.

from pytorch-computer-vision-cookbook.