How to process png image about ocr_tablenet

[Sunnycheey]

It turns out that the shape of my png file is (x, y), but the shape of bmp file provided in dataset is (x, y, 3), so How can I expand the png file dim?

[tomassosorio]

Hi @Sunnycheey ,

The easiest way is to convert the GREY image to an RGB.

[Sunnycheey]

Thank you for your quick reply. Actually, I have test some image from icdar dataset, and I found that the eval result is far from the report result (much worse than the value presented in the paper), have I missing something?

[tomassosorio]

So, for trainning I used the same configurations as the paper with the exception of two things, training with images of size 1024 by 1024, since that would take a lot of VRAM which I did not have available at the time, and this was not also performed:

There are two computation graphs which require training. Each training sample is a tuple of a document image, table mask and column mask. With each training tuple, the two graphs are computed at-least twice. In the initial phase of training, the table branch and column branch are computed in the ratio of 2:1. With each training tuple, the table branch of the computational graph is computed twice, and then the column branch of the model is computed.

Regarding the predicting part, the paper does not explain how they perform it and is there where is the sauce in my opinion. So you could start from where I left I try to improve it.