PaddlePaddle / PaddleOCR

Awesome multilingual OCR toolkits based on PaddlePaddle (practical ultra lightweight OCR system, support 80+ languages recognition, provide data annotation and synthesis tools, support training and deployment among server, mobile, embedded and IoT devices)
https://paddlepaddle.github.io/PaddleOCR/
Apache License 2.0
42.44k stars 7.66k forks source link

FCENet效果巨差 #8894

Closed chc-stars closed 1 year ago

chc-stars commented 1 year ago

请提供下述完整信息以便快速定位问题/Please provide the following information to quickly locate the problem

chc-stars commented 1 year ago

请问下圈中的thr指的是什么? image

andyjiang1116 commented 1 year ago

训练数据有多少呢?有加载预训练模型吗

andyjiang1116 commented 1 year ago

请问下圈中的thr指的是什么? image

thr是得分阈值

chc-stars commented 1 year ago

没有,数据集就是CTW,训练1000,测试500.使用DB和EAST都能达到90以上,并且80多epoch就能收敛

andyjiang1116 commented 1 year ago

用的是最新版2.6分支的代码训练的吗?有没有改动配置啥的?

LDOUBLEV commented 1 year ago

发布的模型配置文件是 configs/det/det_r50_vd_dcn_fce_ctw.yml ,参考:https://github.com/PaddlePaddle/PaddleOCR/blob/release%2F2.6/doc/doc_ch/algorithm_det_fcenet.md

chc-stars commented 1 year ago

image 这种召回率很低是什么原因啊,我的数据集中每张图只有一个检测目标,并且是用四边形的四个点进行标注的,我看到召回率低的原因可能时,正样本被标为负样本,还有就是正负样本不均衡,但是我用DB,EAST网络检测的很好。我想问下是不是FCE网络是不是只针对弯曲文本,水平文本检测能力若,还是在标注是坐标点数太少的原因?

chc-stars commented 1 year ago

image zher这是在我自己数据集上跑500个epoch的精度,训练集上的测试,总共2700张,测出来的智勇1000张。 image

github-actions[bot] commented 1 year ago

This issue has been automatically marked as stale because it has not had recent activity. It will be closed in 7 days if no further activity occurs. Thank you for your contributions.