search

wangpengnorman / CLPD_dataset

43 stars 2 forks source link

two question about your paper #4

Open meixitu opened 4 years ago

meixitu commented 4 years ago

Hi @wangpengnorman ,

Thanks for you great paper and great data.
I have two question about your paper, I would appreciate  If you can give me some clue.

1. In CCPD dataset, most of the License plate have same province, how do you reduce the big data bias in training?
2. In table VI, you compared with other algorithms,
    How do you get the accuracy with other algorithms, such as [13]Moran? retraining with CCPD or just use the pretrained model?

[13] Canjie Luo, Lianwen Jin, and Zenghui Sun. Moran: A multi-object rectified attention network for scene text recognition. Pattern Recognition, 90:109–118, 2019

Thanks Meixitu

17666107783 commented 4 years ago
  1. If it is only for the CCPD test set, because their province distribution is basically the same, direct training, but if you want to be practical, then you have to synthesize a large number of fake license plates from other provinces to supplement the training set.
  2. It is necessary to retrain MORAN with the CCPD training set.
17666107783 commented 4 years ago

有什么问题直接给我发邮件： zhanglinjiang@mail.nwpu.edu.cn 基本在5分钟内回复，github不是我的账号，我一般很久才来看一下，谢谢

meixitu commented 4 years ago

@17666107783 , Thanks for your help. It seems synthesize data is a MUST for real use.

Thanks Meixitu