Good deep net architectures

Belval / TextRecognitionDataGenerator

A synthetic data generator for text recognition

MIT License

3.22k stars 962 forks source link

Good deep net architectures #71

Open tadas-subonis opened 5 years ago

tadas-subonis commented 5 years ago

Hello,

would anyone be willing to share some architectures that worked for you quite well? I figured that this would be a great place to ask that even though it is (kinda) not directly related to the project :-).

For example, I had some decent success with https://github.com/Holmeyoung/crnn-pytorch/blob/master/models/crnn.py (CRNN)

What other models are worth giving a shot?

Belval commented 5 years ago

Hi,

I've only read of two methods, but there is probably a lot more literature covering the subject than what I read.

CRNN, that has been proven to work
NRTR, that uses attention instead of the Bidirectional RNN used by Shi and al.

Unfortunately, I did not find any implementation of the latter on GitHub, and my version is unfinished.

YacobBY commented 5 years ago

@tadas-subonis https://github.com/clovaai/deep-text-recognition-benchmark This one combined with their CRAFT detector is very good.

tadas-subonis commented 5 years ago

@YacobBY Thanks for the link! It has some really useful stuff.