Skip to content

Latest commit

 

History

History
37 lines (31 loc) · 1.92 KB

FAQ.md

File metadata and controls

37 lines (31 loc) · 1.92 KB
Raw

FAQ for TableTrainNet project.

I decided to write some explanation about the choices I have made for this project.

Inference

Do images need to be beautiful?

Not so much indeed. The pre-trained neural network has been trained over normal, 24-bit depth 3-channels images. It is not simple for it to understand that tables are objects, since for it those are lines that have not so much correlation with others. For this reason it is useful to apply this transformation, that let my NN perform much better.

Is this black box?

What the neural network returns is a list of boxes and scores. In my opinion, it is not wise to use them without understanding what the NN is seeing.

I observed some pattern inside the insure policies I had to analyzed:

  • tables are always single-column;
  • they mostly contain number and text;
  • every page hardly contains more then 4 tables each.

Since I could not prepare a specific dataset for this kind of pages due to lack of time, the inference is not so precise:

  • it mistakes big tables and considering it as a set of ones;
  • scores rarely goes over 80%;

For this reason I decided to interpret boxes:

  1. First step is to consider the first 10 boxes (MAX_NUM_BOXES) with score over 0.4 (MIN_SCORE): this let me consider very imprecise boxes, which have some kind of idea of where the boxes are;
  2. Then merge all vertical overlapping boxes: if a box is overlapping another one there is good probability that both of them are looking at the same table. So merging them - considering the min y_min and the max y_max of the two - let me increase accuracy;
  3. Crop widely: since tables are in single-column with no inline text I could consider only 2 parameters out of 4 (y_min and y_max and not x_min and x_max). This let me reduce inference error.

Then no, it is not a black box at all.