When you look at the YOLOv1 paper and corresponding implementations it is always mentioned that
for every grid cell, we predict B bounding boxes (usually two).
Then we use IoU to choose the bounding box out of the B predicted ones, which fits best the real one. All the other B-1 boxes are neglected.
My question is: Why do we even predict B bounding boxes? Isn’t this a waste of resources and may even decrease performance compared to only predicting one box? It is also not like „achor-boxes“ which were introduced in the YOLOv2 model.
