detectRecog / CCPD

[ECCV 2018] CCPD: a diverse and well-annotated dataset for license plate detection and recognition
MIT License
2.19k stars 564 forks source link

Reproduce the results of the paper #69

Open shuangzixing opened 4 years ago

shuangzixing commented 4 years ago
I tried to reproduce the results of the paper according to the training method of the paper, but failed. The results are as follows: base blur challenge db fn rotate tilt weather
98.04 22.74 33.75 31.92 27.82 56.86 47.53 95.25

I guess other sub-datasets (CCPD-DB, CCPD-Blur, CCPD-FN, CCPD-Rotate, CCPD-Tilt, CCPD-Challenge) in CCPD are not only used for test, maybe they are all split to train/val set? Please help me handle this!

NadimintiSaiSirisha commented 2 years ago

Did you train the model? If yes, what images did you use for training it?

123Vincent2018 commented 2 years ago

the almostly same with u

1649832697(1)

2021LsPKU commented 2 years ago

我也遇到了类似的问题,在用作者的训练方法复现他的论文的时候除了base数据集的准确率可以达到,其他都有很大的差距。希望作者可以解答一下?

123Vincent2018 commented 2 years ago

使用的都是它给的ccpd_base的作为训练,怎么可能拟合到它的测试数据上去,测试数据的样本和base的数据差别还是很大的,由他给的训练数据去测测试数据,很明显会无法胜任,但是对比作者在github上给的使用SSD512+HC的基准方法来看,还是可以大致达到的,至于论文里的精度为什么这么高确实不太理解,有可能是加了一部分测试数据到训练集中

------------------ 原始邮件 ------------------ 发件人: "detectRecog/CCPD" @.>; 发送时间: 2022年4月14日(星期四) 中午11:05 @.>; @.**@.>; 主题: Re: [detectRecog/CCPD] Reproduce the results of the paper (#69)

我也遇到了类似的问题,在用作者的训练方法复现他的论文的时候除了base数据集的准确率可以达到,其他都有很大的差距。希望作者可以解答一下?

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>

2021LsPKU commented 2 years ago

使用的都是它给的ccpd_base的作为训练,怎么可能拟合到它的测试数据上去,测试数据的样本和base的数据差别还是很大的,由他给的训练数据去测测试数据,很明显会无法胜任,但是对比作者在github上给的使用SSD512+HC的基准方法来看,还是可以大致达到的,至于论文里的精度为什么这么高确实不太理解,有可能是加了一部分测试数据到训练集中 ------------------ 原始邮件 ------------------ 发件人: "detectRecog/CCPD" @.>; 发送时间: 2022年4月14日(星期四) 中午11:05 @.>; @.**@.>; 主题: Re: [detectRecog/CCPD] Reproduce the results of the paper (#69) 我也遇到了类似的问题,在用作者的训练方法复现他的论文的时候除了base数据集的准确率可以达到,其他都有很大的差距。希望作者可以解答一下? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>

但是论文中所说的就是这样的方法,我自己搭了一个网络,base达到了99.6左右的准确率,weather98.1,rotate91,tilt87.06,然后fn,db都在四十多,用的相同的方法,然后我训练的他的网络得到的结果更差db只有二十多....我不太明白论文里面的数据是怎么来的

123Vincent2018 commented 2 years ago

确实离谱

---Original--- From: @.> Date: Thu, Apr 14, 2022 11:30 AM To: @.>; Cc: @.**@.>; Subject: Re: [detectRecog/CCPD] Reproduce the results of the paper (#69)

使用的都是它给的ccpd_base的作为训练,怎么可能拟合到它的测试数据上去,测试数据的样本和base的数据差别还是很大的,由他给的训练数据去测测试数据,很明显会无法胜任,但是对比作者在github上给的使用SSD512+HC的基准方法来看,还是可以大致达到的,至于论文里的精度为什么这么高确实不太理解,有可能是加了一部分测试数据到训练集中 … ------------------ 原始邮件 ------------------ 发件人: "detectRecog/CCPD" @.>; 发送时间: 2022年4月14日(星期四) 中午11:05 @.>; @.@.>; 主题: Re: [detectRecog/CCPD] Reproduce the results of the paper (#69) 我也遇到了类似的问题,在用作者的训练方法复现他的论文的时候除了base数据集的准确率可以达到,其他都有很大的差距。希望作者可以解答一下? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>

但是论文中所说的就是这样的方法,我自己搭了一个网络,base达到了99.6左右的准确率,weather98.1,rotate91,tilt87.06,然后fn,db都在四十多,用的相同的方法,然后我训练的他的网络得到的结果更差db只有二十多....我不太明白论文里面的数据是怎么来的

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>

xiamangfu commented 1 year ago

我是真复现不出来,感觉DB不可能到90多。有人弄出来过吗?