how to get label of segmentations in SynthText

[TriplePool]

Thank you for your work. Could you please release the code of pre-processing of SynthText.

[MhLiao]

@TriplePool I will release the code of pre-processing of SynthText later. I hope the converted annotations are helpful to you currently.

[TriplePool]

@MhLiao Thank you very much.

[TriplePool]

@MhLiao Hi, I have downloaded the converted annotations, and the label of segmentations is a quadrangle that from "wordBB" in SynthText instead of polygon like TotalText. Is it necessary that pre-train the model with SynthText polygons label of word region, if I want to handle with curved text?

[MhLiao]

@TriplePool In my opinion, more accurate annotations would be better. But I did not try that. If you do care about the performance of curved text, you can try it.

[ChChwang]

@MhLiao the char annotation of icdar2013 was used in training?

[TriplePool]

@TriplePool In my opinion, more accurate annotations would be better. But I did not try that. If you do care about the performance of curved text, you can try it.

Got it, thank you

[TriplePool]

@MhLiao the char annotation of icdar2013 was used in training?

As far as I know, there is no char annotation in icdar2013 dataset, is there?

[MhLiao]

@TriplePool @ChChwang The char annotation of icdar2013 was used in training.

[ChChwang]

@MhLiao there is error in maskrcnn_benchmark/data/datasets/synthtext.py, what if "," in labels? see the label of city_112_47.jpg.

[ustczhouyu]

Please help me. After I use other methods to predict the results, I get .txt such as res_img_1.txt, and each line is x1, y1, x2, y2, x3, y3, x4, y4, word. I don't know how to generate S, W, and G results. How to evaluate the end-to-end result? Please help me.