happyharrycn / actionformer_release

Code release for ActionFormer (ECCV 2022)
MIT License
425 stars 77 forks source link

K-way Binary Classification vs. Multi-class Classification #139

Open WenliangGuo opened 1 day ago

WenliangGuo commented 1 day ago

Thanks for your wonderful work! My question is that why you choose to treat a K-way classification as K binary classification for action label prediction. Is this an experimental result that binary focal loss outperforms CE loss, or there are some other motivation behind it?

tzzcl commented 1 day ago

With the label assignment strategy in ActionFormer, positive samples will be much smaller than positive samples, thus we choose K-way binary classification along with focal loss.

WenliangGuo commented 1 day ago

Thanks for your answer!