clovaai / deep-text-recognition-benchmark

Text recognition (optical character recognition) with deep learning methods, ICCV 2019
Apache License 2.0
3.77k stars 1.11k forks source link

CTC beam decoder #118

Closed soldierofhell closed 4 years ago

soldierofhell commented 5 years ago

Here we're using greedy decoding so the performance may be lower. Is there any PyTorch implementation of beam decoder like in tf: https://www.tensorflow.org/api_docs/python/tf/nn/ctc_beam_search_decoder or any other decoder (e.g. with dictionary)?

ku21fan commented 4 years ago

Hello,

Maybe there is no official CTC beam decoder. (maybe you can find unofficial ones in Github).

And, Good point! we also had a question about it. We tested with Tensorflow ctc beam search decoder with various beam_width. However, there was no significant difference.

So, we just followed STR convention (use greedy decoding). In STR field, people usually use greedy decoding in CTC. Below snippet is the part of CRNN paper. 스크린샷_122719_074837_PM

Best.