microsoft

/

table-transformer-detection

like 307

Follow

Microsoft 5,292

Hi,

Refer to this notebook to get one-on-one matching results with the Microsoft implementation (this was manually verified).

@nielsr

still some issues with that notebook code. this image produces no detections on hugging face page but does produce detections using notebook code. obviously it is not a table.

Hi @nielsr ,
two backup questions:

1. You provide two notebooks that use the table transformer models, here and here. Both notebooks are quite similar, however, in one you defiine the image preprocessor by yourself, in the other you use the DetrFeatureExtractor. Which one is to prefer?
2. Are there any image preprocessing steps prior to applying both, the detection and the structure recognition, that improve the quality of the results? I am thinking about adding some padding (if yes, how much?), deskewing, greyscale, binarization, etc.?

Any advice would be highly appreciated!

Hi,

The recommended preprocessing functions are defined here. So I would recommend those image processing steps.

Ideally there should be a TableTransformerImageProcessor class in the Transformers library which replicates this, I opened an issue for that: https://github.com/huggingface/transformers/issues/30718.

Thank you very much for your quick response, it was very helpful!