search

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

无法复现svtr的论文精度，相同的配置文件，训练精度相差30% #11468

Closed wuxiaolianggit closed 3 months ago

wuxiaolianggit commented 8 months ago

大佬们好，我这边按照svtr_tiny的配置文件，使用使用MJSynth和SynthText两个文字识别数据集训练，在IIIT, SVT, IC03, IC13, IC15, SVTP, CUTE数据集上进行评估，但是不管怎么训练svtr_tiny算法的测试评估精度都是60多一点，恳请各位大佬帮忙解答一下，可能是哪些原因导致的呢？

wuxiaolianggit commented 8 months ago

@tink2123 @sdcb @ZeyuChen @haobibo 各位大佬，可以帮我解答一下上面的问题吗？

Haeppypuppy commented 6 months ago

同問~