Not able to save the annotated data of table in gt.txt using PPOCRLabel

[Bishow-99]

问题描述 / I am using PPOCRLabel for table annotation, I successfully annotated the table data in the PPOCRLabel tool but it is hard to save the annotated file of the table to gt.txt. Although it creates gt.txt on my system, internally it remains empty even though I successfully saved the annotated data on the PPOCRLabel tool.

运行环境 / Runtime Environment

• OS: Ubuntu 20.04
• Paddle: 2.6.0
• PaddleOCR: 2.8

Please help me to understand, how I can save the table annotated data to a gt.txt file.

[SWHL]

The PPOCRLabel project has been migrated to new repositories → link

Please move to the new PPOCR Label repo and ask questions. Thank you.

[Bishow-99]

Thank you so much @SWHL for your immediate response, and now PPOCRLabel works nicely, @SWHL could you please tell me does PaddleOCR works for an image like below

We can see above it has three borderless columns and multiple rowspan. If we work then what could be our annotation and training strategy?

[SWHL]

You can try your image by this online demo.
If it can satisfy your command, I think you could try to train your own model by refering this docs

[Bishow-99]

Hello @SWHL I have given the below image to online demo.

I got the below one

I understand that we need to train our own model for accurate results. But even after following this docs and training the own model I am getting still inaccurate results. Please click this #13499 where I have described all the processes that I have done to train the table structure recognition model. Although, It correctly identifies HTML structures, but cell coordinates are in the wrong position. So, could you please give me how can we handle these things.

[SWHL]

I think you use too little training data.
I suggest you pre-train with some public table recognition datasets, such as PubTabNet, and then fine-tune with your own labeled data. This may give better results.

[Bishow-99]

@SWHL, I've been fine-tuning the PaddleOCR table structure recognition model, which is already pre-trained on datasets like PubTabNet. Given this, do you think it's sufficient, or should I consider training the model from scratch and then fine-tuning it with my own labeled data? I am still confused.

[SWHL]

I think the training data is not enough, you need to get more data for training the pre-trained model.

[Bishow-99]

Okay @SWHL in conclusion, you're suggesting that I should increase the size of my dataset and use the PaddleOCR pre-trained table structure model to fine-tune with my own data for better performance. Is that correct?

At the same time could you please check my annotation for the particular image #13499

[SWHL]

Okay @SWHL in conclusion, you're suggesting that I should increase the size of my dataset and use the PaddleOCR pre-trained table structure model to fine-tune with my own data for better performance. Is that correct?

Yes

At the same time could you please check my annotation for the particular image #13499

I think it's no problem.

[Bishow-99]

@SWHL Thank you so much for these information

[0225pang]

问题描述 / 我正在使用 PPOCRLabel 进行表格注释，我在 PPOCRLabel 工具中成功注释了表格数据，但很难将表格的注释文件保存到gt.txt。虽然它在我的系统上创建了gt.txt，但即使我在 PPOCRLabel 工具上成功保存了带注释的数据，它在内部仍然为空。

运行环境 / Runtime Environment

• 操作系统： Ubuntu 20.04
• 拨片：2.6.0
• PaddleOCR：2.8

请帮助我了解如何将表注释数据保存到 gt.txt 文件中。

请问您是怎么解决的？