fangshancheng / conv-ensemble-str Goto Github PK

I want to evaluate this model, but there is no weights to test.
If weights and char-set can be offered, it would be very grateful. Thank you.

First of all, Thank you for sharing this great work.
The pre-trained model generates random character when testing with the demo images. Also, I tried with a back/white image demo1.jpg

python demo.py --path=data/demo.jpg --checkpoint=PATH_TO_PRETRAINED_MODEL

demo.jpg: 000VC00VVzhWWAWVWAWWWWAhVh0hWW
demo1.jpg: HHyIyHypYyIYyZZyZLHyyHyyIYyIHZ (simple black and white)
demo2.jpg: YYvmZ4mv5ZSmvvmZoZS0mYjIu4ZSZS
demo3.jpg: 111k11GGkYG11k1k1Y1Y1GDG111111
demo4.png: u0yJyDDDDGD5DyyJHyJyDJ7yJquJDJ

update:
Without this --checkpoint flag, everything works fine

python demo.py --path=data/demo.jpg

Thanks

Could you please share the speed and your device infomation?

Hello
Have you evaluated the performance between the beam search and greedy search?

您好,
非常感谢你开源的代码以及训练好的模型! 我用您训练的模型以及demo.py评估IIIT5K时(没有lexicon), 发现word accuracy只有83.7%, 当我只考虑IIIT5K中的全为字符并且至少包含3个字符的word时, word-accuracy =86.13%, 跟论文中的86.7%不一致. 想请问一下, 您的开源的模型是在Syn90k+SynText上训练的吗? 评估iiit5k时,只考虑IIIT5K中的全为字符并且至少包含3个字符的word吗?