Table cell content recognition module generates garbage result on unseen data

[therajmaurya]

On unseen table images, Table structure extraction and Table cell bbox detection work very well but the Table cell content recognition generates garbage. Replacing it with any open-source OCR model for extracting cell content works better than the given model. There seems to be an issue with the Table cell content recognition model that does not generalize well.

P.S. When I run the Table cell content recognition on the dataset on which it was trained/evaluated, it works extremely well, but fails to generalise on new unseen table images.