Closed vfbsilva closed 5 years ago
Thanks a lot. I did it a run and found i very cool. Still I had some issues when identifying multiple persons like here: [https://pasteboard.co/IvRb8Aw.png](image 1) Do you have any clue of what might be going on?
What I see from the image you shared is that FLIR and your dataset might have very different data distribution. Considering the fact that neural networks cannot extrapolate, I would suggest you take the weights I shared above and finetune on your dataset, or preferably train from scratch if applicable.
I also highly recommend you to go over these steps
Yes I do agree I'm doing it right now I just got curious cause I don't know the kind of data you have used for training but I`m by no means complaining btw I'm super grateful it did work like a charm as a proof of concept
Em qua, 4 de set de 2019 às 15:50, Enes Özipek notifications@github.com escreveu:
What I see from the image you shared is that FLIR and your dataset might have very different data distribution. Considering the fact that neural networks cannot extrapolate, I would suggest you take the weights I shared above and finetune on your dataset, or preferably train from scratch if applicable.
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/enesozi/object-detection/issues/1?email_source=notifications&email_token=AAGVSFX4WCXGF2NTWFRBYFTQH77ORA5CNFSM4ITJ4DTKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD54S2GA#issuecomment-528035096, or mute the thread https://github.com/notifications/unsubscribe-auth/AAGVSFQXIUU6BUR4JUQHYSLQH77ORANCNFSM4ITJ4DTA .
I would like to give it a run but I'm out of a powerful GPU atm.