Some unwanted area detected as table

[skwskwskwskw]

With reference to the image above, those above date is also detected as table, when I hope it's not detected. Is there anyway to refine the table detection model?

By the way, I am using the transformer model released here: https://huggingface.co/microsoft/table-transformer-detection

[bsmock]

My first thought would be: is it possible for you to create a simple post-processing rule to filter these unwanted tables out when they occur?

If you want the model itself to suppress the unwanted tables, you will need to fine-tune the detection model on additional labeled examples where only the tables you want are labeled as tables. If the unwanted tables are all very similar in appearance, you will not need very many examples to teach the model to ignore the unwanted tables.

For any questions about how to fine-tune the Hugging Face model, I recommend posting on Hugging Face specifically.

Best, Brandon

[skwskwskwskw]

Thanks. Yay. Your first suggestion coincides with what I thought.

Just wondering what's the difference between transformer pretrained model and your pretrained model in terms of weights?

[bsmock]

There is no difference between the two currently. The Hugging Face model was converted directly from the one in this repo. (This could change in the future, though).