Hi @MhLiao , Sorry to bother you.
I have tried to reproduce the text recognition model which only uses segmentation branch. I modified the source code and used the weights which you offered to test in IC13/IC15/SVT/IIIT5K. But the sequence accuracy without lexicon is much lower than the paper gives. I use SynthText to train the modified recognition model from scratch, and the results are closed to source weights. Sequence accuracy | Icdar2015 | Icdar2013 | Iiit5k | svtpaper | 0.7380 | 0.9230 | 0.9400 | 0.8720
source weights | 0.4889 | 0.7620 | 0.7993 | 0.7399
From scratch | 0.4875 | 0.7935 | 0.7840 | 0.7183
I also tried to add a class for non-alphanumeric character and PPM module as the paper says, but they don't work.
Did you add any other module in text recognition task or did I miss something?
Hi @MhLiao , Sorry to bother you. I have tried to reproduce the text recognition model which only uses segmentation branch. I modified the source code and used the weights which you offered to test in IC13/IC15/SVT/IIIT5K. But the sequence accuracy without lexicon is much lower than the paper gives. I use SynthText to train the modified recognition model from scratch, and the results are closed to source weights.
Sequence accuracy | Icdar2015 | Icdar2013 | Iiit5k | svt paper | 0.7380 | 0.9230 | 0.9400 | 0.8720 source weights | 0.4889 | 0.7620 | 0.7993 | 0.7399 From scratch | 0.4875 | 0.7935 | 0.7840 | 0.7183
I also tried to add a class for non-alphanumeric character and PPM module as the paper says, but they don't work. Did you add any other module in text recognition task or did I miss something?