Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

The score on the ICDAR2013 dataset is very low, only around 65. #36

Open
ranye92 opened this issue Aug 29, 2020 · 4 comments
Open

The score on the ICDAR2013 dataset is very low, only around 65. #36

ranye92 opened this issue Aug 29, 2020 · 4 comments

Comments

@ranye92
Copy link

ranye92 commented Aug 29, 2020

I tried on several datasets, but this repo only performed well on ICDAR2015. I think ICDAR2013 is much easier than ICDAR2015,but I don't why this happen. @SakuraRiven
Anyone knows or get good score on any other dataset? We can dicuss here, thx!!!!!!!!!!中文最好:)

@SakuraRiven
Copy link
Owner

15数据集文字比较小,尺度较为固定,UNet结构dense predict比较容易。13数据集文字尺度变化比较大,不容易回归,可以尝试FPN分层预测。

loss函数可以CE+Dice一起用

生成label的时候,被切断的部分可以直接置ignore,比现在这种循环找要方便的多

@ranye92
Copy link
Author

ranye92 commented Aug 30, 2020

15数据集文字比较小,尺度较为固定,UNet结构dense predict比较容易。13数据集文字尺度变化比较大,不容易回归,可以尝试FPN分层预测。

loss函数可以CE+Dice一起用

生成label的时候,被切断的部分可以直接置ignore,比现在这种循环找要方便的多

感谢大佬这么快就回复。我刚入门不久,具体细节还是想多问一句,烦请解惑:
第一点中FPN分层预测(我理解是提取不同尺度的feature之后,不进行concat,每个scale都输出一个预测吗?不同scale的结果该如何融合呢?)

@SakuraRiven
Copy link
Owner

就类似于FRCNN那样,不同scale的feature map负责不同的预测,结果直接对应回原图就可以了啊。

这个只是我个人想法,当时这个repo真就写着玩的,当入门看就行,真做项目还是自己改进下

@ranye92
Copy link
Author

ranye92 commented Aug 30, 2020

就类似于FRCNN那样,不同scale的feature map负责不同的预测,结果直接对应回原图就可以了啊。

这个只是我个人想法,当时这个repo真就写着玩的,当入门看就行,真做项目还是自己改进下

收到,感谢!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants