search

open-mmlab / mmocr

OpenMMLab Text Detection, Recognition and Understanding Toolbox
https://mmocr.readthedocs.io/en/dev-1.x/
Apache License 2.0
4.32k stars 747 forks source link

performance of FCENet in ICDAR15 #241

Closed wushilian closed 3 years ago

wushilian commented 3 years ago

Can you upload a model trained by fcenet on ic15？ I have trained FCENet 600 epoch，but only got 80 hmean

jeffreykuang commented 3 years ago

@Zyq-scut

Zyq-scut commented 3 years ago

Hi, we have not train FCENet on IC15 yet, but we will do it soon. To improve proformance, you can test with larger 'img_scale' since text in IC15 are small. And 600 epoches are not enough for FCENet to converge, typically 1500 epoches.

wushilian commented 3 years ago

@Zyq-scut I have trained 1500 epoches，but only get 81 hmean. Here is my config fcenet_r50_fpn_1500e_ic15.txt

Zyq-scut commented 3 years ago

Thanks for your feedback. We are training FCENet on IC15 and locating the problem you mentioned above. If there is any progress, we will synchronize with you.

Zyq-scut commented 3 years ago

Problem solved. Our trained model on ICDAR2015 is available now. Thanks!

kbrajwani commented 3 years ago

@Zyq-scut can you please check the link of fcenet ICDAR2015? The link of trained model ICDAR2015 and CTW1500 is same. Can you please update it? Thanks

Zyq-scut commented 3 years ago

@kbrajwani Hi, it is correct now. Thanks you for feedback.