Does anyone try the SOLO model for thin, dotted line? about solo HOT 6 OPEN

@pvthuy What's your loss functions? I suggest using Focal Loss as mask prediction loss. And did you mean the class scores are low or mask prediction after sigmoid?

from solo.

@WXinlong Thanks for your reply.

I used similar settings as in this. It seemed that the loss function is FocalLoss?
My dataset has only two classes (foreground & background) so I set num_classes=2.

I attached the log and a sample prediction here:

log.txt

from solo.

@pvthuy If you didn't change the loss part of that config file, then you were using dice loss for mask prediction. However, for your case, Focal Loss might be better. We extended SOLO to instance contour detection. You can refer to Section 7 of our paper.

BTW, for the current model, you can set a smaller mask threshold (e.g. 0.2 instead of 0.5 in default).

from solo.

@WXinlong Thank you very much!

I changed dice loss to focal loss, and mask threshold 0.5 to 0.2.
The predictions are slightly better, although not good as I expected. The prediction masks are very large compared to actual objects.

Anw, thanks for your help.

from solo.

I tried it for the text,i keep the origin para,and it looks good!

from solo.

hello ,

Can you please shed some light on how to change the loss function to focal-lossas suggested previously.
in the config files I've noticed the following :
loss_ins=dict(
type='DiceLoss',
use_sigmoid=True,
loss_weight=3.0),
loss_cate=dict(
type='FocalLoss',
use_sigmoid=True,
gamma=2.0,
alpha=0.25,
loss_weight=1.0)),

can you please explain exeactly what needs to be changed?

• my case has both thin objects and "normal" objects, if changing the loss function should help for thin objects, won't it hurt the results for the normal objects? ( and if so , can you suggest an alternative ? )

thanks,
Omer

from solo.