search

FangShancheng / ABINet

Read Like Humans: Autonomous, Bidirectional and Iterative Language Modeling for Scene Text Recognition
Other
420 stars 72 forks source link

ABINet has language bias? #85

Closed light42 closed 1 year ago

light42 commented 1 year ago

When I use malay/indonesian sample it has bad performance especially on blurred images. While the sample on huggingspace could be correctly predicted eventhough some images should've been hard to guess even by humans.

FangShancheng commented 1 year ago

you should fine-tune your models on malay/indonesian datasets well (better to pretraining your language model).