search

tianzhi0549 / FCOS

FCOS: Fully Convolutional One-Stage Object Detection (ICCV'19)
https://arxiv.org/abs/1904.01355
Other
3.28k stars 630 forks source link

about bceloss #13

Closed GGjerry01 closed 5 years ago

GGjerry01 commented 5 years ago

why use becloss to do centerness prediction?

tianzhi0549 commented 5 years ago

@GGjerry01 Because the center-ness is in the range [0, 1]. We also tried L2 and it yielded the same performance. But you can still try to use other loss functions, maybe you will find a better one.

GGjerry01 commented 5 years ago

thanks reply. if centerness target is 0.5, then net predict centerness is 0.5. so the loss should be 0,but if i use bceloss, the loss is not 0, so i am very confusing about this loss. otherwise if i misunderstanding the centerness function

------------------ 原始邮件 ------------------ 发件人: "Tian Zhi"notifications@github.com; 发送时间: 2019年4月23日(星期二) 下午3:26 收件人: "tianzhi0549/FCOS"FCOS@noreply.github.com; 抄送: "皮皮桂"534501712@qq.com; "Mention"mention@noreply.github.com; 主题: Re: [tianzhi0549/FCOS] about bceloss (#13)

@GGjerry01 Because the center-ness is in the range [0, 1]. We also tried L2 and it yielded the same performance. But you can still try to use other loss functions, maybe you will find a better one.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

tianzhi0549 commented 5 years ago

@GGjerry01 The loss value is not important, we only care about if it is the minima. If you want its minima being 0, please try KL divergence instead.

GGjerry01 commented 5 years ago

ok thanks very much

------------------ 原始邮件 ------------------ 发件人: "Tian Zhi"notifications@github.com; 发送时间: 2019年4月23日(星期二) 下午3:43 收件人: "tianzhi0549/FCOS"FCOS@noreply.github.com; 抄送: "皮皮桂"534501712@qq.com; "Mention"mention@noreply.github.com; 主题: Re: [tianzhi0549/FCOS] about bceloss (#13)

@GGjerry01 The loss value is not important, we only care about if it is the minima. If you want its minima being 0, please try KL divergence instead.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

soldier828 commented 3 years ago

why use becloss to do centerness prediction?

The same question. After some search, I found a keyword 'bce soft target'. In official definition of bceloss, the target is a float between 0 and 1(namely [0,1]).
So maybe the author actually want to use soft label, and bceloss just can achieve it.

tianzhi0549 commented 3 years ago

@soldier828 Yes, it can be understood as the BCE with soft targets.