Training on handwriting recognition --- any major change compared to STR setup?

[staghado]

Hi there,

First of all, thank you for your excellent work.
I was experimenting with PARSeq on handwritten text recognition lately. I have kept most of the parameters unchanged except for max_label_length and charsets.
I'm getting relatively good results, but i'm trying to improve them if there is room for improvement.
So I wanted to know what do you think are potential hyperparameters (PARSeq architecture, datastet augmentation, optimizer params, scheduler params, ...etc) that could improve the specific case of handwriting recognition without scaling up the model too much?

And also, was there a specific reason why PARSeq was not trained for handwriting?

Thanks in advance! :)

[staghado]

Any updates on this ?

[baudm]

PARSeq was not trained on handwriting data because handwriting recognition is considered a distinct area of research. PARSeq focuses on scene text recognition, but as the demo shows, can also recognize handwriting somehow.

For the hyperparameters, just use an automated tuning system like Ray Tune. The hyperparameters depend largely on the data that you want to use.