SO the YOLO V3 and RetinaNet both uses the Feature pyramids which look something like this:enter image description here (except b and e which have one output)

I'm just confuse how the predictions and training is done? Do we have to give EACH feature map a different Y label? IF yes, how is that possible? We need to have N different ground truth in my opinion. (Also ther'll be 3 different losses I think?)

If not, then how are these done at once?

There is a lot of confusion on these networks because I am not able to get my head around How are y-labels provided, trained and predicted in YOLOv3 and RetinaNet . Everything will make sense about loss, multioutputs and all if I know this one thing.

  • $\begingroup$ great question man, in short they do a heck of a complicated things to map boxes to anchors and then to tensors. Moreover each of the different approaches use a different strategy to map anchors, so the answer to your question is not short. $\endgroup$
    – JVGD
    Feb 6, 2021 at 10:06
  • 2
    $\begingroup$ Long would do too ;) $\endgroup$
    – Deshwal
    Feb 9, 2021 at 7:24


You must log in to answer this question.