one-to-many head

[MuhabHariri]

Hi

Regarding this table :

I noticed that you declared the number of FLOPs for YOLOv10n as 6.7G, but when I train it, I see that it is 8.7G. I searched in the issues section and found this answer from @jameslahm :

"Thanks for your interest! It seems that you also count the parameters and FLOPs of the one-to-many head. Since the one-to-many head is not needed during inference, this part of parameters and FLOPs can be ignored."

My question is: doesn't the same situation (the one-to-many head not being needed during inference) apply to YOLOv8n, which means the number of FLOPs should be lower for it too?

[matpalm]

it's only v10 that the two heads ( one-to-one and one-to-many ). v8 only uses the one.