use_space_char parameter not working

[ThanhLV97]

use_space_char parameter not working when training a custom text recognition. Can you please tell me why? and How to resolve this problem? Tks

[tink2123]

"use_space_char" automatically adds a space character to the dictionary:

PaddleOCR/ppocr/data/imaug/label_ops.py

Lines 123 to 124 in 0438b40

	if use_space_char:
	self.character_str.append(" ")

not working when training a custom text recognition.

Can you elaborate on what went wrong while training? Why do you say it doesn't work?

[LDOUBLEV]

If you expect to recognize spaces and spaces are annotated, you need to set use_space_char to True during training.
Otherwise, it is not needed.
'

PaddleOCR/configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml

Line 19 in 0438b40

use_space_char: true

'

[trucdtt]

@LDOUBLEV I have the same problems. I tried to train English text recognition model followed by readme file (https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.5/doc/doc_en/recognition_en.md). With the ICDAR2015 dataset and en_PP-OCRv3_rec.yml config, predicted results are still no space.

[LDOUBLEV]

@LDOUBLEV I have the same problems. I tried to train English text recognition model followed by readme file (https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.5/doc/doc_en/recognition_en.md). With the ICDAR2015 dataset and en_PP-OCRv3_rec.yml config, predicted results are still no space.

Because the space character is not one of character list of the icdar2015 dataset.

[trucdtt]

There has been confusion about annotating space in ICDAR 2015 and ICDAR 2019. First, I trained with ICDAR 2019 and en_PP-OCRv3_rec.yml config, so the predicted results are no space. After that, I tried to re-train with ICDAR 2015, text after recognition has space between text. I checked ground truth in 2 datasets and found that space characters are only annotated in ICDAR 2015.

[ThanhLV97]

I have a sample confusion. Can paddle's team confirm that ? @LDOUBLEV thanks