search

yuantailing / ctw-baseline

Baseline methods for [CTW dataset](https://ctwdataset.github.io/)
MIT License
329 stars 88 forks source link

Question when running detection_perf.py #29

Open wjp0408 opened 5 years ago

wjp0408 commented 5 years ago

I just use train set as "trainval set" + val set as "test set". And when I run detection_perf.py,I got this output: image It's weird that APs of 店，路，车 are all 0.0 %. So it's means that there're some error heppened? Or these characters didn't appear in the val set? Or the other reasons... ? Thanks for your help. ：）

yuantailing commented 5 years ago

We use train+val to generate cates.json. Our trained models use this order to map character catetory into ID (i.e., the #class output by CNN). If you are using trained model, you should generate cates.json using train+val, and then test performance on val. e.g., If the model outputs ID=4, it means '店' according to cates.json. But if you generate cate.json using trainset only, you may translate 'ID=4' into another character catetory, not '店'.

wjp0408 commented 5 years ago

Thanks for your reply ! And I think I got the correct result with your help : image

Thanks.

peppapeppapeppapeppa commented 5 years ago

Thanks for your reply ! And I think I got the correct result with your help : image

您好 我想问下您是怎么解决这个问题的，我也遇到了这个问题 非常期待您的回答 谢谢