Annotation and pipeline for lexicon free custom dataset

[AumXIV]

Hello, I’m working on license plate which is lexicon free. Now I understand about Synth90k dataset which is lexicon. And my annotation is shown below. How can I use this dataset for CRNN?

I wonder if the Chinese recognition example could help, but I can’t download the dataset. Is the Chinese dataset use the same pipeline with Synth90k? And the same tfrecord?

I can also write char_dict_th.json and ord_map_th.json from my char_dict_th.txt by using local_utils/establish_char_dict.py to generate the json file. But the ord_map_th.json have repeatedly number 32, I don’t know that it can use or not. Please give me some suggestions. Thanks. :)

[MaybeShewill-CV]

@AumXIV 1.If you want use new dataset. You can remove ./data/char_dict/*.json. The char_dict.json and ord_map.json file will be automatically generated when you make the tfrecords. 2. As for the repetition in order_map.json that may due to the repetition in your char_dict.txt file. The old_map.json file records the character's ord number:)

[AumXIV]

Okay, thanks. Here, I try the tfrecord and some errors are appearing. :)

[MaybeShewill-CV]

@AumXIV Label index should be a number sequence. You should follow the synth90k's dataset format to reconstruct your dataset. As for dataset's format this may help #354

[AumXIV]

So you mean I must have a lexicon file? How about let the lexicon file be like the label of each image?

Become like this?

[MaybeShewill-CV]

@AumXIV The second col in annotation file should be the row index in lexicon file:)

[AumXIV]

From

Become

Right? :)

[MaybeShewill-CV]

@AumXIV Right. But I'm not sure if the row index begin with 0 or 1. You may check it yourself:)

[AumXIV]

Okay, now I can generate tfrecord and train the model. Thank you lots. :) ><

[GarhwalRifles]

@AumXIV @MaybeShewill-CV so row index begin with 0 right??

[XinyuDu]

@AumXIV @MaybeShewill-CV so row index begin with 0 right??

yes, the row index begin with 0.
You can find label index 0 in annotation.txt of synth90k's dataset:
./2919/3/160_0_0.jpg 0