search

dingwoai / FairMOT-BDD100K

A fork of FairMOT used to do MOT on BDD100K Dataset
MIT License
27 stars 9 forks source link

'id_loss' doesnt reduce at all #4

Closed RamkumarSakthivel closed 3 years ago

RamkumarSakthivel commented 3 years ago

I am training hrnet-32 model from scratch on the BDD100k MOT dataset, but the id_loss doesn't reduce at all. It stays around 11 even after 10 epochs. Does anyone have any idea why?

dingwoai commented 3 years ago

reID is a classification task in this work. So it is difficult to reduce id_loss, especially with so many different IDs in a driving scene, which is the case in BDD100k MOT dataset.

RamkumarSakthivel commented 3 years ago

Got it, reID is a crossentropy loss. Thanks !