How to create dataset to train table recognition model?

[johntoma]

Hello! I intend to create a new dataset using my own set of table images to train one of the table recognition models, but I don't know where to start or what tools I should be using. I also need to include empty cells in the labelling as they are very important for my use case. I know there is a tool "PPOCRLabelv2" that can do some kind of automatic labelling, but when I try to use the "Table Recognition" feature, it doesn't seem to label the cells, but rather the lines of text for some reason.

Can anyone give me some guidance on where I should start, or maybe some examples I can take a look at?

Thanks!

[TasneemVKhan]

Hi, even we intend to create our set of table images to train the ppstructure model.

I want to label my own dataset and train the table recognition model but haven't found a proper path for it.

I was able to label a few of my table images but the application gives an error and shuts itself down. Haven't been successful.

Use this link to help you succeed in your goal. If any errors you can ask for help

[johntoma]

Yeah I am aware of that tool. Unfortunately I'm also experiencing some issues with the tool crashing. I've submitted issues related to that but haven't received replies from the PaddleOCR team yet.

[TasneemVKhan]

Yeah I have faced the same issue. Refer to this [link] (#8015 (reply in thread)). Ongoing discussion regarding the same issue of application crashing

[AkshatHotCode]

Did you get any reply from them, I am too trying to fine-tune the model with my own set of data, but am unable to process it. Can anyone of you help me with the same?

[TasneemVKhan]

PPOCRLabel is the annotations tool used to generate the dataset for table recognition.

PPOCRLabel has many bugs, even I haven't been successful in it

[AkshatHotCode]

I am able to annotate the dataset on my machine, and i am getting the final output too as Lable.txt.
Now can you help me how can we use this annotations to fine tune the model? https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.6/doc/doc_en/table_recognition_en.md In this link, the format of dataset is different of what we are getting from PPOCRLabel.

[TasneemVKhan]

Are you able to generate a gt.txt file ?

Use train.py and whichever config file you want to use make changes in it and run it through the cmd

[AkshatHotCode]

I was able to get Label.txt.

[SivaKanishka]

@AkshatHotCode, were you able to identify a way to use Lable.txt for fine tuning the table detection model. I feel there are some missing pieces in the documentation.

[johntoma]

@AkshatHotCode @TasneemVKhan Hello, when you annotated your dataset, did you also label the empty cells as well? It's still unclear to me whether or not we must label the empty cells for the training data. I've submitted two tickets about this but unfortunately the paddle team arent responding to them.

[Shreejan-git]

Annotating the whole table (including the blank part as well) and letting the paddleocrlabel to generate gt file gives completely messed HTML. I do not know what should we exactly do!