长文本检测的问题

[liyandongpy]

测试结果，短文本效果还行，长文本效果很差。

[mad-fogs]

Guess u might get expected result if u remove the '--polygon' option. It is not the problem of this repo.

[SylviaHolic]

As for me, I train the model from scratch and DB works very well on long texts. Texts in the public scene text datasets are mostly short english words. So the released model has limitations in long texts. Also variance and bias in datasets matters a lot. It is reasonable that you cannot detect long texts well without training fully with your own dataset.

[shaohailin]

How do you train the model from scratch? what's the training command? @SylviaHolic @MhLiao

[futureisatyourhand]

As for me, I train the model from scratch and DB works very well on long texts. Texts in the public scene text datasets are mostly short english words. So the released model has limitations in long texts. Also variance and bias in datasets matters a lot. It is reasonable that you cannot detect long texts well without training fully with your own dataset.

hello, how do you ideal with the issue? How many epoch do you train? Thanks. I find the problem, the model tests well aster training 3-6 epochs , just for short text. The long text is worse.