person segmentation soft label

[xiongzhu666]

Hi, Xuebin, I am coming again for the person segmentation, I prepare the training datasets of high-precision soft label. do you have any suggestion about modify the loss function??? because I think binarized label won‘t achieve the hair strand level precision. thank best regards

[xuebinqin]

Hi, Xiong,

I think you can try bce first. According to my experience, bce usually works well because it often generates relatively smooth boundaries. Of course, it depends on your ground truth. Then you can also try the l2 or other losses used in the image matting tasks such as https://arxiv.org/pdf/1703.03872.pdf BTW, some others also tried human image matting. Their results look not bad. You can take a look at https://github.com/xuebinqin/U-2-Net/issues/111

regards,

On Mon, Dec 20, 2021 at 8:45 AM xiongzhu666 @.***> wrote:

Hi, Xuebin, I am coming again for the person segmentation, I prepare the training datasets of high-precision soft label. do you have any suggestion about modify the loss function??? because I think binarized label won‘t achieve the hair strand level precision. thank best regards

— Reply to this email directly, view it on GitHub https://github.com/xuebinqin/U-2-Net/issues/275, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADSGORNVTJHEUSEQAT4OSYLUR2YGRANCNFSM5KMY35XA . 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/