folded corners and staples: dataset blind spot?

[drzraf]

I tried this model on a set of A4 photos with very interesting results. Still, I had multiple items exposing the following defect:

• overzealous crop
• spurious distortion

The condition leading to this are apparently:

• folded corner (or, more exactly partially occlusion of a corner due to another stapled page)
• presence of a staple
• document edge touching the border of the image
• poor illumination

Details:

• distortion glitch (caused by the mere staple?)

• another distortion glitch

• overcrop (caused by illumination?)

• overcrop (caused by the folded edge?)

• overcrop + distrortion

• overcrop + distrortion

• overcrop (document's top border touching the edge of the image?)

• overcrop

Questions:

• I know the dataset is not public (but only available on demand), but I wonder if, from a quick look, you consider the current limitations to be caused by dataset's limitation or rather the chosen NN?
• In the first case, would you suggest retraining on an enriched dataset?
• If yes, do you anticipate that a limited amount of (manually) cropped (containing staples/partial corner occlusion), used with their original counterpart, for training, would significantly improve the overall detection?