I have a question regarding the different criteria for the loss calculation in DETR and for the Hungarian matching. For e.g.,
for calculating the label cost, in the Hungarian matcher, 1 - proba[target class] is used while for the label loss, NLL is used. Is there a reason for doing this? I am interested to know this because, is there any numerical stability issue if, for e.g., the same/different criteria are used for loss calculation and Hungarian matching?
Thanks for the wonderful work!
I have a question regarding the different criteria for the loss calculation in DETR and for the Hungarian matching. For e.g., for calculating the label cost, in the Hungarian matcher,
1 - proba[target class]is used while for the label loss,NLLis used. Is there a reason for doing this? I am interested to know this because, is there any numerical stability issue if, for e.g., the same/different criteria are used for loss calculation and Hungarian matching?