search

dlyldxwl / fssd.pytorch

Pytorch re-implementation of Fssd
67 stars 27 forks source link

any plans to add support for focal loss? (instead of ohem) #6

Closed zhanghengdev closed 5 years ago

zhanghengdev commented 5 years ago

Dear @dlyldxwl , Great work! have you ever tried focal loss in FSSD? It seems to be better than ohem in your programme.

Following repos might help: https://github.com/marvis/pytorch-yolo2/blob/master/FocalLoss.py https://github.com/yhenon/pytorch-retinanet/blob/master/losses.py

I have tried myself, but it seems to be worse than ohem...

Best regards,

dlyldxwl commented 5 years ago

Yes, but not this repo, I will release a project after 1-2 weeks about stronger one-stage detector based on ssd and yolo v3, which integrate much tricks, e.g., mix up, GIOU, focal loss and so on. If you are interested in it, please give more attention.

zhanghengdev commented 5 years ago

Great!

dlyldxwl commented 5 years ago

@ZHANGHeng19931123 Hello~ When I apply label smooth to SSD, mAP drops ~1 point. I think this is because background class produces negative impact, thus I ingore it, but also not work. Have you tried it?

zhanghengdev commented 5 years ago

Sorry, I have never tried label smooth...