About loss function

[skeras]

Hi， This is a great project, thanks to the author's open source。 I have some questions about this project‘s loss， how about bce+ssim+iou result？ more better than only use bce or worse then bce？ Thanks

[xuebinqin]

thanks for your interest. The improvements depend on what metric you are using. bce+ssim+iou is designed to supervise with multi-scale patterns, when the model has relatively weak ability in capturing these patterns. u2net is good at multi-scale feature extraction. So the improvements by using this hybrid loss is not that significant in terms of some metrics like maxF. But with and without hybrid loss will produce probability maps with different characteristics. For example, if you are segmenting targets with rigid boundaries, the hybrid loss is suggested. Otherwise you can just use bce for achieving soft boundaries, such as feather, hair, etc.

On Tue, Jan 4, 2022 at 5:00 PM skeras @.***> wrote:

Hi， This is a great project, thanks to the author's open source。 I have some questions about this project‘s loss， how about bce+ssim+iou result？ more better than only use bce or worse then bce？ Thanks

— Reply to this email directly, view it on GitHub https://github.com/xuebinqin/U-2-Net/issues/278, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADSGORL3XOPUAOFTH4642BTUULVQLANCNFSM5LHMROMQ . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.

You are receiving this because you are subscribed to this thread.Message ID: @.***>

-- Xuebin Qin PhD Department of Computing Science University of Alberta, Edmonton, AB, Canada Homepage: https://xuebinqin.github.io/

[skeras]

Thanks a lot.